text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
International Journal on Computational Sciences & Applications (IJCSA) Vol.4, No.5, October 2014
Integrating Fuzzy and Ant Colony System for
Fuzzy Vehicle Routing Problem with Time
Windows
Sandhya and V.Katiyar
Department of Computer Science and Engineering, Maharishi Markandeswar University,
Ambala, Haryana,India.
Abstract
In this paper fuzzy VRPTW with an uncertain travel time is considered. Credibility theory is used to model
the problem and specifies a preference index at which it is desired that the travel times to reach the
customers fall into their time windows. We propose the integration of fuzzy and ant colony system based
evolutionary algorithm to solve the problem while preserving the constraints. Computational results for
certain benchmark problems having short and long time horizons are presented to show the effectiveness of
the algorithm. Comparison between different preferences indexes have been obtained to help the user in
making suitable decisions
Keywords:
VRPTW, Fuzzy sets, Credibility theory, ant colony system, uncertainty, stochastic simulation.
1.Introduction
Transportation is one important component of logistics. Efficient utilization of vehicles directly
affects the logistic cost as 40% to 50% money is spent on transportation. Moreover proper
utilization of vehicles is also important from environmental view point. The use of automated
route planning and scheduling can lead to huge savings in transportation cost ranging from 5% to
20% [25]. So efficient routing of fleets is a crucial issue for companies and this can be formulated
as vehicle routing problem.Vehicle Routing Problem was first introduced by Danting and
Rameser [26] in 1959. Till now many variants of the problem have been proposed [2]. Vehicle
Routing Problem with Time Windows is one flavor of VRP. In VRPTW a least cost route from
central depot to all customers is to be designed in such a way that all customers are visited by
homogenous vehicles once and only once while preserving the capacity and time window
constraints. In this problem it is assumed that the time to travel from one customer to another is
equivalent to distance between the customers. Moreover in these types of problems, all
parameters are assumed to deterministic. In real life scenarios these assumptions donot hold
because of varying road conditions, link failures, rush hours, congestion etc. which lead to
uncertainties in data. Most of the algorithms developed for deterministic problems do not work in
these situations. In this paper VRPTW with fuzzy travel time is considered. This uncertainty is
handled using fuzzy logic. The fuzzy travel time is represented by triangular fuzzy number.
DOI:10.5121/ijcsa.2014.4506
73
International Journal on Computational Sciences & Applications (IJCSA) Vol.4, No.5, October 2014
Credibility theory is used to model the problem. Improved Ant Colony System (IACS) [4] is then
used to find the most efficient route for the problem. The main goal of this paper is to develop an
algorithm that can provide the user with a route having minimum distance with uncertain travel
time at a desired preference index. The rest of the paper is organized as follows.Literature review
is presented in section 2. Section 3 discusses the essential basics of fuzzy theory. A model based
on credibility theory is then proposed in section 4. Section 5 presents improved ACS that is used
to solve the model. Results based on experimentations are discussed in section 6. Finally section
7 presents the conclusions and scope for future work.
2.Literature Review
VRP is concerned with finding the minimum set of routes, starting and ending at the central
depot, for homogenous vehicles to serve number of customers with demands for a goods such that
capacity constraint is preserved.Latest taxonomy on VRP can be found in [2]. The VRPTW
considers the time window for each customer in which service has to be provided. A vehicle can
arrive before the starting time of the window but it cannot arrive after the closing time of window.
A taxonomy can be found on [5,6].VRPTW is NP-hard problem. Many exact and metaheuristics
algorithms have been proposed for solving VRPTW [7,8,9,10]. A categorized bibliography of
metaheuristics and their extensions can be found in [1].But these became infeasible for dynamic
problems.In [11] a mixed integer linear programming approach and a nearest neighbor, branch
and cut algorithm is used to solve the problem for time dependent VRPTW.However this model
does not follow FIFO property.Ichoua et al [12] proposed a tabu search approach to solve
TDVRPTW, where customers are characterized by soft time windows. The model presented
satisfies the FIFO property.In another approach Donati et al. [13] use multi ant colony and local
search improvement approach to update the slack or the feasible time delays. The travel times are
analyzed by discretizing the time space thus satisfying the FIFO property. However these models
fail when uncertainties arises in various parameters.A detailed summary for various uncertain
parameters for VRPTW can be found in [19].Because of lack of data or due to extreme
complexity of the problem it requires subjective judgment. Fuzzy set theory provides meaningful
methodologies to handle uncertainty, vagueness and ambiguity.In [14] triangular fuzzy numbers
are used to represent fuzzy travel time and a route construction method is proposed to solve the
problem.A fuzzy optimization model using imperialist competitive algorithm is presented for
fuzzy vehicle routing problem with time window in [15].Cao Erbao et al [16] use fuzzy
credibility theory to model the vehicle routing problem with fuzzy demand. It uses integration of
stochastic simulation and differential evolution algorithms to solve the same model. However it
requires a lot of parameters to be taken care of.Yongshuang Zheng and Baodinf Liu [17] present
an integration of fuzzy simulation and genetic algorithms to design a hybrid intelligent algorithm
for solving fuzzy vehicle routing model. In [18] two new types of credibility programming
models including fuzzy chance constraint programming and fuzzy chance-constraint goal
programming are presented to model fuzzy VRP with fuzzy travel time. In [3] fuzzy concepts and
genetic algorithm are used for the solution multi-objective VRP. Finally J.Brito et al [20]
proposes a GRASP meta heuristic to solve the VRPTW in which travel time is uncertain. A
chance constraint model is build using credibility approach to solve the problem. However the
proposed algorithm appears to be inefficient because in GRASP metaheuristic each restart is
independent of the previous one.On the other hand the stochastic element of ACO allows to build
74
International Journal on Computational Sciences & Applications (IJCSA) Vol.4, No.5, October 2014
variety of different solutions.In this paper VRPTW with fuzzy travel time is considered.
Credibility theory is used to build a chance constraint programming model for the problem. An
improved ant colony [4] metaheuristic is then used to obtain the optimal routes .Moreover using
this approach a decision maker can also evaluate different planning scenarios to choose the best
alternative with desired confidence level.
3.Fuzzy Credibility Theory
In this section, some basic concepts and results about fuzzy measure theory are summarized. The
term fuzzy logic was introduced by Lofti A. Zadeh [21]. In contrast to conventional logic, fuzzy
logic allows the intermediate values to be defined between conventional evaluations like true or
false. Fuzzy numbers are the numbers that possess fuzzy properties. In this paper we have
represented the fuzzy travel time between two customers as the triangular fuzzy number. A
triangular fuzzy number is represented by triplet TFN=(a, b, c) where b is the mean value (i.e.
mode) and and the left and right extremes of its spread. Its membership function is
=
0
0
<
≤≤
≤≤
>
(1)
Zadeh [23] proposed the concept of possibility measure for fuzzy variables as a counterpart of
probability theory in crisp sets.Concept of the fuzzy set, fuzzy variable, possibility measure,
necessity theory are available in [22].Liu [24] proposed the credibility as the average of
possibility and necessity. Possibilityof an event is measured by most favorable cases only in
contrast to probability of an event where all favorable cases are measured. Let be nonempty set
and P () be the power set. Each element in P is called an event. Also denotes an empty set. In
order to present an axiomatic definition of possibility, it is necessary to assign a number Pos{A}
to each event A, which indicates the possibility that A will occur [16].
Axiom 3.1: Pos() =1 (Normality Axiom)
Axiom 3.2:Pos()=0 (Non negativity Axiom)
Axiom 3.3: For each ⋀
"
Pos ⋃"
#$ %=⋁ #$ '()
(Maximality Axiom)
Figure 2.1 shows the possibility of fuzzy event {X< = x0}
Pos(X<=x0)
x
75
International Journal on Computational Sciences & Applications (IJCSA) Vol.4, No.5, October 2014
Figure 2.1 Possibility of fuzzy event
The explicit expression for possibility Pos(X<=x0) is given by:
Pos * ≤ + = ,
1
< +
≤ + ≤
.
0
+ <
(2)
Necessity for an event A is defined as the impossibility of complement of that event Aci.e
Nec{A}=1-Pos {Ac}. Fig 2.2 shows the necessity of fuzzy event {X<=x0}
1
Nec{X<=x0}=1-Pos(X<=x0 )
xo
x
Figure 2.2 Necessity of fuzzy event
The explicit expression for possibility Nec(X<=x0) is given by:
Nec * ≤ + = ,
1
.
0
< +
≤ + ≤
+ <
(3)
The credibility for event A is Cr{A}=1/2(Pos{A}+Nec{A}).}). The credibility measure signifies
the credibility how the solution satisfies the constraints. The explicit expression for credibility in
case of triangular fuzzy measure is:
Cr * ≤ + =
1
. /0
/
.
/
0
< +
≤ + ≤
≤ + ≤
+ <
(4)
4.Chance Constraint Model for Fuzzy VRPTW
We define a Fuzzy VRPTW as follows:
Given a set of n geographically distributed customers requiring services within a specific time
period from set of k homogenous vehicles stationed at a central depot with known demands of
customers having uncertain travel time between customers but lying within known ranges with
76
International Journal on Computational Sciences & Applications (IJCSA) Vol.4, No.5, October 2014
most likely values being known then the objective of fuzzy VRPTW is to find the route with
minimum distance with specified confidence level that will meet all customers’ time windows
requirements assuming that:
1. Each vehicle starts and ends its tour at central depot indexed by 0.
2. Each vehicle has fixed identical capacity.
3. Each customer is to be visited once and only once by only one vehicle.
4. Each customer has a predefined time window in which it has to be served [ei , li].
5. The demand of the customer is fixed and is assumedthat it will not exceed the vehicle
capacity.
6. The travel time between each customer is not fixed and it is assumed to be expressed as
triangular fuzzy number 25
34 =(2 6$ , 2 6/ , 2 68 ).
Let D[i] be the departure time of vehicle v from customer i, then arrival time A[j]at j will be the
summation of departure time from previous customer i and the fuzzy travel time from node i to j.
A[j] = D[i] +29
34
(5)
The time to begin service will be maximum of arrival time or the opening time.
(6)
TS[j] =max (A[j], ej)
Because of fuzzy travel time, arrival time and time to begin service at next customer will also be
fuzzy. However the service time S[i] and opening of the time windows e[i] are crisp numbers that
are special case of triangular fuzzy numbers where the three defining numbers are equal. We
obtain the credibility that the time to begin service at next customer does not exceed its close time
to be
0
Cr(TS[j] ≤lj) =
AB CD>6@E
/∗CD 6@G CD>6@E
AB /∗CD>6@G 0CD>6@K
/∗CL>6@K CD>6@G
1
:; <6 < =>?@$
:; <6 ≥ =>?@$ IJ <6 < =>?@/
:; <6 ≥ =>?@/ IJ <6 < =>?@8
:; <6 ≥ =>?@8
(7)
As we know that if the travel time to next customer is smaller than its closing time then the
chances of serving that customer by that vehicle will grow. That is greater the difference between
the closing time and the maximum travel time greater is the chance to serve that customer and if
this difference is small that customer may not be served. Therefore, at some preference index Cr*,
the solution r verifies the fuzzy constraint of service times within the corresponding time window
if: Cr(TS[r] ≤lj).≥ Cr*.The goal is to determine the value of MN ∗ which will result in a route
having minimum distance. For this stochastic simulation is done. Thus the objective of
corresponding chance constraint model of fuzzy VRPTW using credibility theory is as follows:
R
R
Q Q
= O:I ∑S
Q#$ ∑ #+ ∑6#+ 6 6
(8)
77
International Journal on Computational Sciences & Applications (IJCSA) Vol.4, No.5, October 2014
subject to:
Q
∑RQ#$ ∑R ,6#+ 6
=1
Q
Q
∑R6#$ +6
=1 and ∑R #$ +
=1
(9)
Q
∑R6#+ 6
- ∑R #+ 6 Q =0
(10)
∑RW#$ ∑R6#+ MN=>N@
(13)
∑R #$ J
≤ TQ
Q 5
∑R #+ ∑R6#+ 6
234
+ ) 6 + V2 ≤ <
≤ <6 > MN ∗
(11)
(12)
(14)
Following indices and notations are considered:
•
i=0,1,2,…,n are the customer indexes with 0 denoting base station.
•
k=1,2,…,m are the vehicles.
•
di is the demand of the customer.
•
Qk is the capacity of the vehicle.
•
cij cost of moving from node i to j, expressed in terms of distance from customer i to j.
•
sij is the service time at customer i.
•
[ei, li] is the time window of customer i.
•
tij is the fuzzy travel time between i and j.
•
wtiis the waiting time at customer i.
•
TS[i] is the fuzzy time to begin service.
Q
Q
A binary variable 6
is introduced such that 6
=1 if vehicle k travels directly from customer i to
customer j and otherwise 0. The objective function (8) seeks to minimize the total traveled
distance whereas constraint (9), specifies that every customer is visited by one vehicle and splitting
of deliveries are forbidden.Eq (10), states that each tour starts and ends at depot indexed 0.
Capacity constraint is preserved in Eq (12) and it ensure that demand of each customer does not
exceeds vehicle capacity.Eq. (13), preserves the constraints and ensuring that sum of fuzzy
travelling time to the customer, service time and waiting time are less than the closing time of
customer’s window. Eq (14), preserves time to begin service for route r is within a specified
preference index.
5.Proposed Solution Approach
In this section, first, stochastic simulation will be used to calculate the total distance. Then ant
colony heuristic is used to obtain the least cost route plan with the best value of dispatcher
preference index Cr*.
78
International Journal on Computational Sciences & Applications (IJCSA) Vol.4, No.5, October 2014
a.Stochastic Simulation for additional distance
In this paper, as travel time between each customer is uncertain and they are represented as
triangular fuzzy numbers, so algorithms for deterministic problems cannot be applied for the fuzzy
VRPTW. Moreover real values of the travel time will be known after reaching the customer.
However uncertain travel time can be considered deterministic by stochastic simulation. We
summarize the algorithm as follows:
Step 1: Simulate actual travel time for each customer by following process:
Step 1.1: Generate a random number x within the left and right boundaries of a triangular
fuzzy travel time between each customer and calculate its membership u(x).
Step 1.2: Generate a random number r ∈ 0,1.
Step 1.3: Compare r and u(x): if r<u(x), use x as the actual travel time, otherwise generate x
and r again and compare them until they satisfy the condition r<u(x).
Step 2: Calculate total distance by moving along the planned route.
Step 3: Repeat step 1 and 2 N times and calculate the average total distance.
b.Route Construction using Ant Colony Optimization
To obtain the best solution, enhance ant colony optimization of [4] is applied. The algorithm works
as under.
Step 1: Initialize the pheromone matrix Z 6 = 1[ 6 and place O ants at depot and the set of
unvisited.
customers as
Step 2: Repeat while all the customers
are not marked as visited.
Step 2.1 Start an ant and mark depot as the current location.
Step 2.2 From the current location (i) choose the customer (j) to be visited by a pseudorandom proportional rule given by:
?=\
a
c
N]O ^_Z 6 ` . _b 6 ` d ,
e
:; N ≤ N+
(2ℎgNV:)g
(15)
whereN ∈ >0,1@is a uniformly generated random number, and b 6 is the visibility of customer j
and defined by
b 6 = 1m
j
l
> hM 6 + V26 i + h<6 − g6 i @
(16)
Here M 6 is the cost of travelling from node i to j which is distance from node i to j. In our case as
travel time are uncertain and their values will be known after reaching the customers so it may be
possible that vehicle arrives at acustomer but cannot serve it because of expiry of time window.
79
International Journal on Computational Sciences & Applications (IJCSA) Vol.4, No.5, October 2014
g6 − 6 :; g6 > 6
is the waiting time at location j before service can be started
0
(2ℎgNV:)g
and<6 − 6 , 6 < <6 , i.e. the difference between the latest arrival time ljand actual arrival time aj at
customer j is the measure of urgency of customer j to be served.
<)( e ∈ oQ is the customer selected according to the probability pq given by
Also V26 = n
pq =
t
_rsB ` ._usB `
t
v
∑B∈w _rsB ` ._usB `
x
v
(17)
yz is the set of customers which will be successfully visited from the current location by the same
vehicle without violating the following time and capacity constraints.
(i) Time constraints: 6 ≤ <6 i.e. arrival at {|} customer must be up to the closing
time<6 of that customer with
6 = O6$ , g6$ + =6$ + 26$,6 and
Cr(TS[r] ≤lj).≥ Cr*
Q Q
(ii) Capacity constraints:h 6
J 6 ≤ TQ i
(18)
(19)
Step 2.3 Mark (j) as the current location and update the set of unvisited customers~ →
~ – ?. Also update capacity and current time of the ant k.
Step 2.4 From the current location again find oQ . If oQ = ~<< then go to 2.1 else
repeat the 2.2.
Step 3: Out of O paths find the best path to be followed with minimum total travelled distance and
further try for improving that route by applying local search.
Step 4: Update the pheromone matrix by the global pheromone updation rule:
Z 6 = 1 − Z 6 + ∑Q#$ ΔZ 6
(20)
where ∈ 0, 1is a constant that controls the speed of evaporation of pheromone. is number of
routes in the current best solution. The deposited pheromone ΔZ 6 on the links is given byΔZ 6 =
T/. Here T is the constant, is the tour length of current solution.
Step 5: Repeat the step 2, 3 for desired number of iteration and provide the best obtained solution.
6.Computational Experience
The proposed algorithm has been encoded in MATLAB 8.0. This study uses the dataset of J.Brito
et al [20] which was generated from the example of Zheng and Liu [17]. Two types of
80
International Journal on Computational Sciences & Applications (IJCSA) Vol.4, No.5, October 2014
experimental conditions based on the time windows are generated. We assume that there are 18
customers with short time horizons and with long time horizons. Customer labeled 0 is assumed
as depot. In each experiment demand for each customer, distance and fuzzy travel time between
each customer is same as that of J.Brito et al. Start time and the closing time for each customer in
long time horizon is assumed to be same but for dataset with short time period the total opening
duration is assumed to be 100 for each customer. The relative parameters and their values that
were used during implementation are listed in table6.1 .
Table 6.1 Simulation Parameters
Sr. No.
Simulation Parameter
Parameter Value
1
Number of customers
18
2
Number of iterations
1000
3
Capacity of Vehicles(Q)
1000
4
Initial Pheromone value for all arcs
1
5
Α
2.5
6
Β
.7
7
Γ
.5
8
Ρ
0.2
9
q0
0.5
10
No. of Ants
11
11
Credibility
.8
12
Depot Close time
5000
13
Service Time
15
Comparison of results obtained by our algorithm,J.Brito[20] and Zheng et al [17] on various
parameters are presented in table 6.2
81
International Journal on Computational Sciences & Applications (IJCSA) Vol.4, No.5, October 2014
Table 6.2 Comparison Table
Factors
Zheng et al[17]
J.Brito et al[20]
Our Approach
Meta Heuristics
Genetic
GRASP
ANT COLONY SYSTEM
Total Iterations
10,000
-
1000
Time
Consumed
10 hours
1 minute
3 minutes
Vehicle used
4
3
3
Total Distance
457.5
365.5
373
Vehicle Routes
R1:0-16-17-18-5-0
R1:0-17-18-16-1514-12-13-0
R1:0-1- 2-3-18-17-13– 8-0
R2:0-10-12-13-14-15-80
R3:0-1-2-3-0
R2:0-2-1-3-4-6-8-0
R2:0-4-5-6-10-11-7-9-0
R3:0-16-15-14-12-0
R3:0-10-9-11-7-5-0
R4:0-9-6-7-11-4-0
Loads of
Vehicles
590,815,440,640
795,930,760
955,840,550
Robust
Less
More
More
One can observe that our algorithm produces effective results than Zheng et al [17]and
comparable results with GRASP. Moreover proposed algorithm is more robust than [17] in terms
of utilization of vehicles. To evaluate the importance of dispatcher preference Cr varied with the
interval of 0 to 1 with a step of 0.1. The average computational results of 10 times are given in fig
6.1
Aggregated Cost
Influence of prefrence index
Cr
Fig 6.1 Influence of preference index
82
International Journal on Computational Sciences & Applications (I
(IJCSA)
JCSA) Vol.4, No.5, October 2014
From above figure on can conclude that if decision maker is risk lover it will choose lower values
of Crr whereas if he is risk adverse he will go for higher values with plan having higher cost.
Figure 6.2 shows the effect of fuzzy travel time on time windows for long duration problems with
a confidence level of 0.8.
Time
Missed Time Windows
customers
Arrival_Time
Close_Time
Fig 6.2 Missed time windows for long horizon problems with Cr=0.8
One can note that if we are travelling with the same fuzzy sspeed as given in [19]] all customers are
served within their time windows. But for short horizon problems if Cr<=0.4
r<=0.4 some customers are
suffered as shown in fig 6.3.. For example for customer 4 and 5 the closing time window is around
900 but the actual arrival time of the vehicle is around 1000.
Time
Missed Time Windows at (.3)
Customers
A_T
(.3)
C_T
Fig 6.3 Missed time w
windows for short problems with Cr=0.3
If we set the Cr>=0.5
r>=0.5 every customer can be served within it
itss time boundaries as shown in fig6.4
fig
83
International Journal on Computational Sciences & Applications (I
(IJCSA)
JCSA) Vol.4, No.5, October 2014
Time
Missed Time Windows at (0.5)
Customers A_T(.5)
C_T
Fig 6.4 Missed time windows for short problems wi
with Cr=0.5
7.Conclusions
Deterministic assumptions of the VRPTW make it unsuitable for real world environment. In this
paper, VRPTW with fuzzy travel time is considered to capture the real life scenario.
scenario Our goal is
to construct efficient and reliable routes for the problem. We propose a chance constrained model
for the problem using fuzzy credibility theory where fuzzy travel time is represented as triangular
fuzzy number. Additionally, stochastic
tochastic simulation is done to get the fuzzy travel time and then ant
colony optimization algorithm is used to get the optimal solution for the problem in reasonable
time. We apply
ly our solution approach to problems having short and long duration of time
windows. It was concluded that the proposed approach performs well in each structure of the
problem and provides improved results on [17] and comparable results with [20]
[20
approach.Further
Further comparison between different confidence levels is done to show the influence
on total distance and it was concluded that higher values of preference index leads to higher cost.
This comparison helps a decision maker in choosing different values for confidence level to get
different results based on cheapness and robustness
robustness.
References
1.
2.
3.
Gendreau, Michel, et al. Metaheuristics for the vehicle routing problem and its extensions: A
categorized bibliography. Springer US, 2008.
Eksioglu, Burak, ArifVolkanVural, and Arnold Reisman. "The vehicle routing problem: A taxonomic
review." Computers & Industrial Engineering 57.4 (2009): 1472
1472-1483.
Gupta, Radha, Bijendra Singh, and Dhaneshwar Pandey. "Multi
"Multi-objective
objective fuzzy vehicle routing
problem: a casee study." International Journal of Contemporary Mathematical Sciences 5.29 (2010):
1439-1454.
84
International Journal on Computational Sciences & Applications (IJCSA) Vol.4, No.5, October 2014
4.
Bansal Sandhya, Rajeev Goel, and C. Mohan. "Use of Ant Colony System in Solving Vehicle Routing
Problem with Time Window Constraints." Proceedings of the Second International Conference on Soft
Computing for Problem Solving (SocProS 2012), December 28-30, 2012. Springer India, 2014.
5. Soo, Raymond Kuo Yang, and Yong HaurTay. "A survey on the progress of research on vehicle
routing problem with time window constraints." Symposium on Progress in Information and
Communication Technology (SPICT). 2009.
6. Laporte, Gilbert. "Fifty years of vehicle routing." Transportation Science 43.4 (2009): 408-416.
7. Laporte, G. and Y. Nobert (1997). Exact algorithms for the vehicle routing problem.Surveys in
Combinatorial Optimization, North-Holland, Amsterdam 132, 147_184.
8. R. Baldacci, E. Bartolini, A. Mingozzi, R. Roberti, 2010, “An exact solution framework for a broad
class of vehicle routing problems”, Computational management Time, vol7(3), 229-268.
9. Lysgaard, Jens. "Clarke & Wright’s Savings Algorithm." Department of Management Science and
Logistics, The Aarhus School of Business (1997).
10. Bräysy, Olli, and Michel Gendreau. "Metaheuristics for the vehicle routing problem with time
windows." Report STF42 A 1025 (2001).
11. Malandraki, Chryssi, and Mark S. Daskin. "Time dependent vehicle routing problems: Formulations,
properties and heuristic algorithms." Transportation science 26.3 (1992): 185-200.
12.I choua, Soumia, Michel Gendreau, and Jean-Yves Potvin. "Vehicle dispatching with time-dependent
travel times." European journal of operational research144.2 (2003): 379-396.
13. Donati, Alberto V., et al. "Time dependent vehicle routing problem with a multi ant colony system."
European journal of operational research 185.3 (2008): 1174-1191.
14. Demirel, Tufan, Nihan Cetin Demirel, and BelginTasdelen. "Time Dependent Vehicle Routing
Problem with Fuzzy Traveling Times Under Different Traffic Conditions." Journal of Multiple-Valued
Logic & Soft Computing 14 (2008).
15. Wang, Geng-jia, Yuan-Biao Zhang, and Jia-Wei Chen. "A novel algorithm to solve the vehicle routing
problem with time windows: Imperialist competitive algorithm." Advances in Information Sciences
and Service Sciences 3.5 (2011): 108-116.
16. Erbao, Cao, and Lai Mingyong. "A hybrid differential evolution algorithm to vehicle routing problem
with fuzzy demands." Journal of computational and applied mathematics 231.1 (2009): 302-310.
17. Zheng, Yongshuang, and Baoding Liu. "Fuzzy vehicle routing model with credibility measure and its
hybrid intelligent algorithm." Applied Mathematics and Computation 176.2 (2006): 673-683.
18. Goncalves, Gilles, Tiente Hsu, and JianXu. "Vehicle routing problem with time windows and fuzzy
demands: an approach based on the possibility theory."International Journal of Advanced Operations
Management 1.4 (2009): 312-330.
19. Brito, Julio, José Andrés Moreno, and José L. Verdegay. "Fuzzy Optimization in Vehicle Routing
Problems." IFSA/EUSFLAT Conf.. 2009.
20. Brito, J., et al. "Fuzzy Vehicle Routing Problem with Time Windows."Proceedings of IPMU. Vol. 8.
2008.
21 .L.A. Zadeh, Fuzzy Sets, Information and Control 8 (1965) 338_353.
22. A. Kaufman, Introduction to the Theory of Fuzzy Subsets, Academic Press, New York, 1975.
23 .L.A. Zadeh, Fuzzy sets as a basis for a theory of possibility, Fuzzy Sets and Systems 1 (1978) 3_28.
24 B. Liu, Uncertain Theory: An Introduce to its Axiomatic Foundations, Springer, Berlin, 2004.
25. Gemeinschaften, Europäische, ed. White paper-European transport policy for 2010: time to decide.
Office for Official Publications of the European Communities, 2001.
26. Dantzig, George B., and John H. Ramser. "The truck dispatching problem."Management science 6.1
(1959): 80-91.
85
| 9cs.NE
|
Prediction and Control of Projectile Impact Point using Approximate
Statistical Moments
arXiv:1710.00289v1 [cs.SY] 1 Oct 2017
Cenk Demir1 , Abhyudai Singh2
Abstract— In this paper, trajectory prediction and control
design for a desired hit point of a projectile is studied.
Projectiles are subject to environment noise such as wind effect
and measurement noise. In addition, mathematical models of
projectiles contain a large number of important states that
should be taken into account for having a realistic prediction.
Furthermore, dynamics of projectiles contain nonlinear functions such as monomials and sine functions. To address all these
issues we formulate a stochastic model for the projectile. We
showed that with a set of transformations projectile dynamics
only contains nonlinearities of the form of monomials. In the
next step we derived approximate moment dynamics of this
system using mean-field approximation. Our method still suffers
from size of the system. To address this problem we selected
a subset of first- and second-order statistical moments and we
showed that they give reliable approximations of the mean and
standard deviation of the impact point for a real projectile.
Finally we used these selected moments to derive a control law
that reduces error to hit a desired point.
I. I NTRODUCTION
In a broad sense, a projectile is a ranged weapon that
moves through the air in the presence of external forces
and endures its motion because of its own inertia. Begining
from Aristotle and later Galileo, projectile trajectory prediction and its impact points have been studied [1], [2]. The
goal of these studies is to provide a realistic mathematical model in order to explain the behavior of projectiles.
Mathematical modeling allows us to understand the motion
of projectiles while requiring a much lower cost than experimental analysis. However, mathematical models become
rapidly convoluted when all the parameters of the system are
considered (e.g. atmospeheric conditions, air density, fuel,
etc.). Hence, forming an optimal model, that contains only
critical information about a projectile, is essential. Various
approximations, methods, and assumptions are used to obtain
reliable models [3]–[6]. In this paper, we define a novel
model to predict impact points for ballistic targets. Predicting
these points is essential in order to avoid hitting constrained
areas [7], and to create effective defense systems for those
points [8].
In order to reach to a reasonable Impact Point Prediction (IPP), several deterministic and stochastic approaches
are studied, such as various types of Kalman Filters [9],
Maximum Likelihood Estimator [10], and Stochastic Model
Predictive Control [7]. Recent works have also included the
1 C. Demir is with Department of Electrical and Computer Engineering,
University of Delaware, Newark, DE USA 19716. [email protected]
2 A. Singh is with the Department of Electrical and Computer Engineering, Biomedical Engineering, Mathematical Sciences, Center for Bioinformatics and Computational Biology, University of Delaware, Newark, DE
USA 19716. [email protected]
wind effect which is crucial to obtaining realisctic IPP [11].
However, it is highly inaccessible to receive information
about wind instantaneously and hastily because of limited
sensor accuracy [12]. Furthermore, the deviations of wind
speed and direction are random which makes it unrealizable
to predict the future evolution of wind [13]. This noise also
shifts the impact points of projectiles which makes classic
IPP model erroneous.
In this paper, the effect of wind is modelled as a stochastic
process. In the presence of such randomness, statistical
moments are reliable tools that give us useful information
about the mean and variance of impact points. However, it
is not always possible to determine statistical moments for
highly nonlinear systems because of the unclosed moment
dynamics which means that higher order moments appear
in lower ones. To interpret moments in this case, various
approximation techniques are used which are called closure techniques (see, e.g., [14]–[18]). Unfortunately, these
methods are mainly developed to deal with higher order
moments of monomial form. For instance, how to approximate skewness as a function of mean and variance. However,
IPP contains nonlinearities of the trigonometric form due to
its nature. Here we show that by using Euler formula we
can transform a system to new coordinates. The transformed
system can be modelled as classic monomial form with a
change of variables. In the next step, we derive moment
dynamics for the transformed system. These dynamics are
free of trigonometric functions, yet they are still unclosed.
Hence, we apply mean-field approximation to close them
[19]. Mean-field approximation gives reliable results in the
limit of weak correlation between states of the system, which
is the case here due to the presence of independent noise
terms in different states of IPP [20].
The ultimate aim of aerospace studies is to control projectile around its desired trajectory which is named as projectile
guidance. To do so, many guidance laws are developed
[21], [22]. In between them, propotional navigation guidance
(PNG) is the most used and well performed when the target
is stationary [23]. PNG is applied to a projectile by changing
forces that act on it. Such changes can be obtained by using
configurations of canards, wings or tails [12]. Change of
configurations is achieved through various control strategies
[7], [24], [25]. In this paper, we used feedback control to
move along a projectile in a desired trajectory. In the next,
we start our analysis by defining projectile dynamic models.
II. E XACT DYNAMIC M ODEL OF P ROJECTILE
To model a projectile motion we first need to define proper
coordinates. Projectile dynamics contain two frames. One
TABLE I. Summary of notation used
Parameter
x, y, z
θ, ψ, φ
p, q, r
u, v, w
X, Y, Z
L, M, N
I
sθ , cθ , tθ
V
D
ρ
CNA
vw , ww
CLP , CM Q
RM CM
RM CP
CY P A
Cx0
W1 , W2 , W3
α, β
uci , vci , wci
Di
Si
Li , D i
CL , CD
Mi
λi
rxi , ryi , rzi
θE , ψE
Kp , Kψ , Kθ , Kφ
ep , eq , er
Description
Location of projectile
Roll, pitch, yaw angles
Angular velocity components
Projectile velocity components
Force components of projectile
Physical moment components of projectile
Inertia matrix
Sine, cosine and tangent of θ
Total velocity
Projectile characteristic length
Atmospheric density
Normal force aerodynamic coefficient
Wind components
Roll and pitch damping aerodynamic coefficients
Distance from center of mass to center of
Magnus pressure
Distance from center of mass to center of pressure
Magnus force aerodynamic coefficient
Zero yaw axial force aerodynamic coefficient
Zero-mean white noise components
Angle of attack and side slip angle
ith canard velocity
ith canard characteristic length
ith canard surface area
ith canard lift and drag force
Canard aerodynamic lift and drag coefficients
ith canard Mach number
ith canard angle
ith canard distance from center of gravity
Yaw and pitch angle error
Control gains
Control inputs
of them is the inertial frame which is often construed as
the Earth coordinate frame. The other one is the projectile
referenced frame which is the body frame. In these frames,
position states are denoted by (x, y, z), and orientation states
are denoted by (ψ, φ, θ). Note that position (x, y, z) and
orientation (φ, θ, ψ) are independent parameters, and called
six degrees of freedom [26], [27] which can also be seen in
Figure 1.
The position states (x, y, z) give information about location of the projectile in the Earth coordinate which is
portrayed as
of Euler angles (φ̇, θ̇, ψ̇)
p
φ̇
q = LB
E θ̇ .
r
ψ̇
(2)
This transformation is a non-orthogonal transformation because the axes on which we measured Euler angle derivatives
are non-orthogonal. For this reason, it is necessary to define
LB
E which is the matrix that allow us to separate Euler
angle derivatives to orthonormal components [29]. Thus,
from taking inverse of LB
E , we can define the dynamics of
φ̇, θ̇, ψ̇
1 sφ tθ cφ tθ p
φ̇
(3)
θ̇ = 0 cφ −sφ q .
cφ
sφ
0
r
ψ̇
cθ
cθ
The body velocity components depend on the acting force
on the projectile (X, Y, Z); explicitly gravitational force,
aerodynamic steady state force, aerodynamic Magnus force,
and canard lifting force [30]. Further, velocity components
(u, v, w) are
X
0 −r q u
u̇ m
0 −p v .
v̇ = Y − r
(4)
m
Z
−q
p
0
w
ẇ
m
Moreover, angular velocity depends on the physical moments
that acting on a projectile, (L, M, N ). These moments are
aerodynamic steady state moment, aerodynamic unsteady
moment, aerodynamic Magnus moment, and the moment of
the canard [30]. The connection between angular velocity
and physical moment is given by
0 −r q p
ṗ
L
−1
M − r
q̇ = I
q ,
0 −p I
ṙ
r
N
−q p
0
(5)
⃗I
φ
cθ cψ
ẋ
ẏ = cθ sψ
ż
−sθ
sθ cψ cφ + sψ sφ u
sθ sψ cφ − cψ sφ v .
cθ cφ
w
(1)
This equation shows that position states (x, y, z) depend on
the body translational velocities (u, v, w). Furthermore x, y, z
depend on angles between projectile and Earth frame origin
(Figure 1). The matrix in equation (1) is known as rotation
matrix. This matrix is created by the standard aerospace
rotation sequence [28]. It allows us to represent the speed
of projectile in all directions in the Earth coordinate system.
The orientation states of the projectile, (φ, θ, ψ), are known
as Euler’s angles, they are called specifically: roll, pitch, and
yaw. The inertial angular rates (p, q, r) are the time derivative
sφ sθ cψ − sψ cφ
sφ sθ sψ + cψ cφ
sφ cθ
θ
ψ
-z
y
⃗
I
Impact Point
⃗
x
I
Fig. 1. Model schematic of a projectile in inertial coordinate frame. A projectile is fully identified by its position
in inertial coordinates (x, y, z) and its orientation (φ, θ, ψ).
By knowing these 6 degrees of freedom one can predict the
impact points, i.e. the point where the projectile hits.
where I is inertia matrix of
IXX
I = IY X
IZX
a projectile
IXY
IY Y
IZY
IXZ
IY Z .
IZZ
(6)
The force and the moment terms in equations (4) and (5)
are defined as the following: the gravitational force is the
force that pulls projectile towards the center of the Earth
[31]. Aerodynamic force and physical moment are the effect
of pressure and shear stress on the body of projectile [32].
This force can be split into lift force which is opposite to
gravitational force, and drag force which is perpendicular to
gravitational force and opposite to the lateral velocity of a
projectile. After these definitions, the aerodynamic moment
can be constructed for any point on the body by using
these forces. In addition, canard lifting force and moment
are the aerodynamic force and moment over the canard
surface. The last phenomena on the projectile is Magnus
force and moment which is seen in the spinning projectiles.
The pressure difference between opposite sides of a spining
projectile creates this force and moment [26].
Equations (1)-(6) are full description of dynamics for a
projectile [33], and it is clear that these dynamics are highly
nonlinear. Hence different variations of this model are used
by considering different assumptions. In the next section, we
briefly review the current approximations.
III. L INEAR M ODEL AND M ODIFIED L INEAR M ODEL OF
P ROJECTILE
In order to decrease the complexity of projectile dynamics,
researchers came up with different assumptions [33], [34].
One famous framework which is built based on series of
simplifying assumptions is known as projectile linear theory.
This theory allows us to define analytic solution of projectile
[35], however it usually generates considerable error on impact point prediction [36]. To overcome this issue, projectile
modified linear model is interpreted with relaxing some of
the assumptions. Specifically,
• The pitch angle, θ, is not small
sin (θ) 6= θ, cos (θ) 6= 1.
•
(7)
Projectile roll rate and pitch angle are not assumed
constant
p 6= p0 , θ 6= θ0 .
(8)
After these two changes projectile modified linear model is
ẋ = cos (θ)D,
D
ẏ = cos (θ)Dψ + ṽ,
V
D cos (θ)
ż = −D sin (θ) +
w̃,
V
D
φ̇ = p̃,
V
D
θ̇ = q̃,
V
D
ψ̇ =
r̃,
V cos (θ)
3
πρD Cx0 V
Dg sin (θ)
V̇ = −
−
,
8m
V
(9)
(10)
(11)
(12)
(13)
(14)
(15)
πρD3 CNA (ṽ − ṽw )
ṽ˙ = −
− Dr̃,
8m
Dg cos θ
πρD3 CNA (w̃ − w̃w )
+ Dq̃ +
,
w̃˙ = −
8m
V
4
5
πρD V CDD
πρD CLP p̃
p̃˙ =
+
,
8Ixx
16Ixx
4
πρD5 CM Q q̃
πρD RM CM CY P A p̃(ṽ − ṽw )
+
q̃˙ =
16V Iyy
16Iyy
πρD3 RM CP CNA (w̃ − w̃w )
Ixx Dp̃r̃
+
−
,
8Iyy
Iyy V
5
4
πρD CM Q r̃
πρD RM CM CY P A p̃(w̃ − w̃w )
+
r̃˙ =
16V Iyy
16Iyy
3
πρD RM CP CNA (ṽ − ṽw )
Ixx Dp̃q̃
−
+
8Iyy
Iyy V
(16)
(17)
(18)
(19)
(20)
[36]. Note that in this paper, a superscript ∼ is used to
describe components in the fixed frame.
All of these projectile models are deterministic. However,
in reality, motion of projectiles is not deterministic, and has
noise components. This calls for a new model for a projectile.
In the next section, we will introduce a new stochastic model
for a projectile with stochastic noise effect.
IV. S TOCHASTIC M ODIFIED L INEAR M ODEL OF
P ROJECTILE
We start with introducing the general Stochastic Differential Equation (SDE). A typical equation of the form of SDE
can be written as
dx = f (x, t)dt + g(x, t)dW t ,
(21)
where x = [x1 x2 ... xn ]T
∈ Rn is the
state vector. The system dynamics are f (x, t) =
[f1 (x, t) f2 (x, t) ... fn (x, t)]T ∈ Rn × [0, ∞) → Rn , and
g(x, t) = [g1 (x, t) g2 (x, t) ... gn (x, t)]T ∈ Rn × [0, ∞) →
Rn . W t = [W1 W2 ...Wn ] is a n-dimensional Wiener process
whose mean is zero.
Projectiles contain miscellaneous physical components
that are subject to noise. Various noise sources include measurement noise, sensor noise, wind effect and, etc. Consistent
with previous studies, we model these noise terms as state
independent white noise [11], [37]–[40]
dx = (cos (θ)D)dt + a1 dW1 ,
D
dy = (cos (θ)Dψ + ṽ)dt + a2 dW2 ,
V
D cos (θ)
dz = (−D sin (θ) +
w̃)dt + a3 dW3 ,
V
where a1 , a2 and a3 are constant, and
hdWi dWi i = dt,
hdWi dWj i = 0, i 6= j
(22)
(23)
(24)
(25)
In the rest of this paper, hi denotes the expected value.
W1 , W2 and W3 can be understood as the effect of noise on
different coordinates (Figure 2). After these modifications,
dynamics of projectile is in the form of SDE equation (21),
thus we call this model Projectile Stochastic Modified Linear
Model (PSMLM).
A. Statistical Moment Dynamics of PSMLM
The moment dynamics of a general SDE in the form of
equation (21) can be derived by using Itø formula
We start our analysis by writing the first statistical moment
dynamics,
n
n
n
X
∂ 2 x[m]
∂x[m]
1 XX
dhx i
(gg T )ij
fi
+
,
=
dt
∂xi
2 i=1 j=1
∂xi ∂xj
i=1
(26)
(35)
[m]
T
where [m] = [m1 m2 ...mn ] [41]. The sum of mj is the
order of the moment. If f (x, t) and g(x, t) are linear, then
moment dynamics can be written compactly as
dµ
= c + Aµ,
(27)
dt
where µ contains all the moment of the system up to order
M
n
X
mj .
(28)
M≡
j=1
The vector c, and the matrix A are determined by using
f (x, t) and g(x, t). To solve the linear equation (27) is
effortless because the desired moment order is always combination of higher or the same order moments. However, when
f (x, t) or g(x, t) are nonlinear, it is not simple to determine
all the moments, and (27) needs new configuration. This is
dµ
= c + Aµ + B µ̄,
(29)
dt
where µ̄ only contains the higher moments than the desired
ones.
This problem is a fundamental problem of statistical moments determination when f (x, t) or g(x, t) are nonlinear.
To overcome this fundamental problem, we used well-known
mean-field closure technique [42], [43]. This approach is
convenient for the systems that have computational complexity or high-dimension [44]. The basic idea of this moment
closure technique is to define higher order moments as the
product of the moments of individuals. For instance
m
m
j
mi
j
i
hxm
i xj i ≈ hxi ihxj i,
D
dhδ + i
= j hδ + q̃i
dt
V0
dhe−jθ i
D
= −j hδ − q̃i
dt
V0
D
dhxi
= ( (hδ + i + hδ − i)
dt
2
(36)
(37)
These moment dynamics depend on the second order moment dynamics hδ + q̃i and hδ − q̃i. In the next step, we add
dynamic of these two moment to our system which are
dhδ + q̃i
πρD4 RM CM CY P A +
D
hδ p̃(ṽ − ṽw )i
= j hδ + q̃ 2 i +
dt
V0
16V0 Iyy
πρD5 CM Q +
πρD3 RM CP CNA +
+
hδ q̃i +
hδ (w̃ − w̃w )i
16Iyy
8Iyy
Ixx D +
hδ p̃r̃i,
(38)
−
Iyy V0
dhδ − q̃i
πρD5 CM Q −
D
hδ q̃i
= −j hδ − q̃ 2 i +
dt
V0
16Iyy
πρD4 RM CM CY P A −
+
hδ p̃(ṽ − ṽw )i
16V0 Iyy
3
πρD RM CP CNA −
Ixx D −
+
hδ (w̃ − w̃w )i −
hδ p̃r̃i.
(39)
8Iyy
Iyy V0
These dynamics depend on the third order moments hδ + q̃ 2 i
and hδ − q̃ 2 i. The moment dynamics of this system are not
closed in the sense that any close set of moments depends
on higher order moments. To overcome this issue, mean-field
closure technique is used
hδ + q̃ 2 i ≈ hδ + ihq̃ 2 i, hδ − q̃ 2 i ≈ hδ − ihq̃ 2 i.
(40)
In order to calculate mean, we need to add second order
moment dynamics in the next.
mi , mj ∈ [0, 1, 2, . . .]. (30)
1) Approximation of mean of PSMLM using mean-field:
It is handy to use (26) to find moments of the system. For
example for state x,
D
cos (θ)dt + a1 dW1 .
(31)
2
This dynamic is nonlinear because of cosine term. If we use
Euler’s relation,
dx =
eiθ + e−iθ
eiθ − e−iθ
, sin θ =
.
2
2i
Thus, (31) can be rewritten as
cos θ =
dx = (
D jθ
(e + e−jθ )dt + dW1 .
2
(32)
(33)
To proceed from (33), we define two new states δ + = ejθ
and δ − = e−jθ
D
D
dδ + = j δ + ,
dδ − = −j δ −
(34)
V0
V0
Fig. 2. Our approximation technique captures the mean
behavior of projectile. The bold lines show the mean
trajectory of a projectile obtained from numerical simulations
and mean-field approximation. The approximation results are
indistinguishable from numerical results, showing significant
performance of our method.
9200
10000
Mean-field Standard Deviation
Data Standard Deviation
9100
9000
9500
y-coordinates (ft)
y-coordinates (ft)
8900
8800
8700
8600
9000
8500
8500
8400
8000
Mean-field Standard deviation
Data Standard Deviation
Impact Points
8300
8200
1250
1300
1350
1400
1450
7500
1200
1500
1250
1300
x-coordinates (ft)
2) Second order moments of PSMLM: This system has
14 states, which means there exist 14 first order statistical
moments, and 105 second order statistical moments. This
number of equations make running different methods and
finding optimal solutions in real time impossible. Hence, here
we only consider a few number of second order moments and
we approximate the rest as functions of first order moments.
Namely we add the 14 equations of the form of a sole
state, for instance
dhx2 i
= D(hxδ + i + hxδ − i) + a1 .
(41)
dt
Moreover, we include hxδ + i, hxδ − i and the moments that
show up in the moment dynamics of coordinates. The rest
of moments are approximated,
hδ + q̃i ≈ hδ + ihq̃i, hδ − q̃i ≈ hδ − ihq̃i,
+
hδ p̃ṽi ≈ hδ ihp̃ṽi, hδ p̃ṽi ≈ hδ ihp̃ṽi,
hδ + p̃r̃i ≈ hδ + ihp̃r̃i, hδ − p̃r̃i ≈ hδ − ihp̃r̃i.
−
1400
1450
1500
1550
1600
x-coordinates (ft)
Fig. 3. Our method successfully predicts the standard
deviation of the impact points. Every cross represents
impact point when a projectile have noise component. The
blue and red ellipse are the standard deviation of x and
y coordinates by using mean-field approximation and data
simulations, respectively.
+
1350
−
(42)
(43)
(44)
By applying these closures, we have closed set of first
order moment dynamics and selected sub-set of second
order moments of PSMLM. Solving these equations give
us approximate time trend of the mean and the standard
deviation.
V. S IMULATION R ESULTS OF P ROJECTILE P REDICTION
U SING M EAN - FIELD
In this paper, fin stabilized projectile is used with following initial conditions for location, x = 0 ft, y = 0 ft,
and z = 0 ft which is origin of reference frame. Speed
terms are considered as ũ = 400 ft/s, ṽ = 0 ft/s, and
w̃ = 0 ft/s. Angles in reference coordinates system are
φ = 2.9 rad, θ = 0.267 rad, and ψ = −0.007 rad. Angular
velocities are p̃ = 399.7 rad/sec, q̃ = 0.43 rad/sec, and
Fig. 4. Randomness in starting point increase the errors
in impact points, hence it vitiates the accuracy of the
missile. Mean field approximation is capable of giving
accurate estimations of error even in the presence of random
initial conditions.
r̃ = −1.54 rad/sec. Physical parameters are selected as air
density of ρ = 0.00238 slug/(cu ft), and gravity constant
of g = 32.174 ft/s2 . The rest of physical parameters are
chosen as in [35], i.e. reference diameter is 0.343521 ft,
weight of the projectile is m = 0.0116 slug. Aerodynamic
coefficients are CX0 = 0.279, Cdd = 2.672, Clp = −0.042,
Cna = 2.329, Cypa = −0.295, Cmq = −1.800. Distances
from center of mass for aerodynamic moment and force are
Rmcp = −0.1657 ft, and Rmcm = −0.1677 ft. Moments of
inertia are Ixx = 2.85 × 10−5 slug/ft2 , Iyy = 2.72 × 10−5
slug/ft2 [35]. In addition, wind speed can be taken from
weather cast agencies, here we used ṽw = 15 ft/s and
w̃w = 15 ft/s.
After using these initial conditions and parameters, we
simulated PSMLM. The result of this simulation can be seen
in Figure 2. As it is clear from this figure, our approximation
is able to predict the mean behavior of projectile through
the time successfully. In the next, we analyzed the impact
points. Figure 3 shows that our method is able to predict
standard deviation of impact points around their mean with
small error.
TABLE II. Initial Values of Probability Distribution
Parameter
x
y
z
φ
θ
ψ
Mean
0
0
0
2.9
0.267
-0.007
S.D.
3
3
0
1
0.017
0.002
Parameter
u
v
w
p
q
r
S.D.
400
5
5
399.7
0.43
-1.54
Mean
2
0.01
0.001
3
0.01
0.01
Moreover, in reality the initial condition may change
because of the randomness in nature such as wind, and
topology of the terrain. To address such uncertainty projectile
prediction is performed for a distribution of initial conditions
[7], [45], [46]. We implemented this source of uncertainty
in our numerical simulations by choosing a random initial
+
Canard
1
−β
VI. D ESIGN
OF
+
P ROJECTILE C ONTROL L AW
2
1
2
2 πDi
ρ(u2ci + vci
+ wci
)
CL , i = 1, 2, 3, 4,
2
4
and the drag force of canard is [26]
Li =
Di =
2
1
2
2 πDi
ρ(u2ci + vci
+ wci
)
CD , i = 1, 2, 3, 4,
2
4
where Si =
πDi2
4 ,
(45)
(46)
M2
Canard
3
M3
CD = cd0 (Mi ) +
cd2 (Mi )α2i
+
(48)
We have 4 canards in different locations. Two of them
are located parallel to missile body in y-direction. The force
terms for these two are
Xci
Yci
Zci
L √ wci 2 − Di √ 2uci 2
i u2ci +wci
uci +wci
0
=
ci
−Li √ 2uci 2 − Di √ w
2
2
uci +wci
uci +wci
, i = 1, 3. (50)
The other two are located parallel to missile body in zdirection. Their force terms are
(
Xci
Yci
Zci
)
+
Canard
4
-
++ +
++ +
+ -+
++
F3
u
v
w
q
r
θ
ψ
φ
M4
eφ
eθ
eψ
xd yd zd
p
F4
x
y
z
Σ
e1
e2
e3
pitch
&
yaw
θE
ψΕ
Fig. 5. The controller schematic for following the desired
projectile by changing the angle of canards. The control
process starts with getting information about the current
states of a projectile (x, y, z, α, β, u, v, w, φ, p). The states
(x, y, z) are subtracted from desired states (xd , yd , zd ) to find
the location error. According to this error, yaw and pitch
angle errors are calculated by using velocity states (u, v, w).
Then, yaw and pitch angle errors and states (p, φ) are used
to control system through a feedback. Canards’ angle are
updated using output of controller, angle of attack and side
slip angle.
(47)
ci (Mi )CL2 .
Here Mi is Mach number on which aerodynamic coefficients
depend [25]. Mach number is the ratio between speed of
sound and the speed of air vehicles [29]. In addition, αi is
the angle of attack of each canard which depends on the
location of that canard and projectile angle of attack. For
each canard, this angle can be defined as
wci
αi = λi ± tan−1 (
).
(49)
uci
)
−β
β
and coefficients are defined as [25]
CL = clα (Mi )αi ,
(
+
F2
Canard
2
α
In this section, we focus on controlling a projectile. To
apply control, four canards are used whose characteristics
are exactly the same. These canards allow us to adjust
forces and physical moments that act on a projectile [25].
Controlling these forces and moments are generally obtained
by manipulating canard angles (λ1 , λ2 , λ3 , λ4 ). Here, we
design a feedback control with the subjective of following a
desired trajectory (e.g. the trajectory with minumum flight
time [12]) by using these angles. In the next, we define
canard properties and feedback control law with details.
The four symmetric canards are added into projectile. The
lift force of canard is [26]
α
F1
M1
PROJECTILE
DYNAMICS
α
Controller
condition drawn from a distribution introduced in Table II.
Figure 4 shows that our method is still capable of predicting
impact points. As expected standard deviation increases in
presence of stochastic starting point.
−L √ vci 2 − Di √ u2 ci 2
i u2ci +vci
uci +vci
Li √ u2 ci 2 − Di √ v2ci 2
, i = 2, 4.
=
uci +vci
uci +vci
0
(51)
Moreover, physical moments of these 4 canards are defined
as
0
−rzi ryi Xci
Lci
Mci = rzi
0
−rxi Yci ,
(52)
Zci
Nci
−ryi rxi
0
where rxi , ryi and rzi are the distance from center of gravity
for each canard in each direction. Then, these forces and
moments are added into the system dynamic model directly.
To design a controller for a projectile by using feedback
control law, a desired trajectory is required. In this paper, we
assume that the desired trajectory is pre-defined. The error
terms (e1 , e2 , e3 ) between the desired trajectory and actual
trajectory are the difference along x, y and z coordinates
[25]. According to these error terms, yaw and pitch angle
errors are described as
w̃
e3
θE = tan−1 ( ) − tan−1 ( ),
ũ
e1
−1 ṽ
−1 e2
ψE = − tan ( ) + tan ( )
ũ
e1
(53)
(54)
−1 v
( u ) are angle of attack (α) and
where tan−1 ( w
u ) and tan
side slip angle (β). By using these errors, the control law is
eφ = Kp p + Kφ φ,
e θ = K θ θE ,
eψ = Kψ ψE ,
(55)
(56)
where Kp , Kφ , Kθ , Kψ are feedback control gains [25].
Finally, the control is applied by updating the angle of each
canard [25].
λ1 = eθ − eφ ,
λ3 = eθ + eφ ,
λ2 = eψ + eφ ,
λ4 = eψ − eφ .
(57)
(58)
The entire control design schematic is shown in Figure 5.
VII. S IMULATION R ESULTS FOR I MPLEMENTING THE
C ONTROLLER
We assumed that the aerodynamic coefficients of canards
(clα , cd0 , cd2 , ci ) are constant. Also, the distance of each
canards from center of gravity is described as the following.
TABLE III. Distance from center of gravity for each canards
Parameter
rx1
rx2
rx3
rx4
Value (f t)
0.474
0.474
0.474
0.474
Parameter
ry1
ry2
ry3
ry4
Value
0.102
0
-0.102
0
Parameter
rz1
rz2
rz3
rz4
Value
0
0.102
0
-0.102
Because of usage of identical canards, all the canard
wing areas are equal to Si = 0.02104 f t2 . The velocities
of canards are not the same because of angular rates and
position of each canard [25]. However, these velocities can
be calculated by considering location of canards and angular
velocities in each iteration. In control process, we selected
the control gain parameters as Kp = −2, Kφ = −1.5, Kθ =
0.01, Kψ = 0.015. By using these control parameters,
change of standard deviation and impact points are shown in
Figure 6. Our control law successfully reduced the variance
of impact points. Hence, it increases the reliability and
accuracy of the missile. However, the error between the
simulation results and mean-field approximation is small, but
not negligible. One way to reduce this error is to add more
moment dynamics to our analysis. Future work will quantify
bounds on error of estimation to find the optimal number
of dynamics needed to be added to reach to a desired error.
Such bounds on approximation error of moments is recently
developed for simple dynamic systems [47], [48].
9200
9100
y-coordinates (ft)
9000
Mean-field Standard deviation
Standard Deviation
Impact Points
Uncontrolled Standard Deviation
8900
8800
8700
8600
8500
8400
8300
8200
1250
1300
1350
1400
1450
1500
x-coordinates (ft)
Fig. 6. By controlling a projectile, standard deviation of
impact points location reduces considerably. The control
law successfully rejected the contribution of the noise and
made the projectile to follow the desired path. This results
in lower deviation of impact point.
VIII. C ONCLUSION
In this paper, we used SDEs to model projectiles under
noise effect. Next, we applied Euler’s formula to deduce
nonlinearities of the trigonometric to monomial form. Then,
we employ mean-field approximation to obtain closed form
equations describing mean and standard deviation of the
system. Our approximation gives reliable results in predicting
time evolution of projectile and characteristics of impact
points. Finally, we proposed a control scheme to reduce the
errors in impact points.
Furthermore, while the aim of a projectile is to hit the
exact target point, it also evades to hit constrained areas. For
this purpose, skewness, and kurtosis can be used to avoid
hitting those areas by changing the shape of distribution of
impact points. Further research will study the higher order
moments of projectile such as skewness and kurtosis. Finally
in this work we assumed that the controller is built in the
projectile. Sometimes we need to give a new control law
to the projectile through a transmission channel. Prospect
research will merge dynamics of the projectile with random
discrete transmission events modelled as renewal transitions
to address this requirement [49]–[51].
ACKNOWLEDGMENT
AS is supported by the National Science Foundation Grant
ECCS-1711548.
R EFERENCES
[1] E. Hussey, Physics: books III and IV. Clarendon Aristotle Series,
1983, vol. 2.
[2] R. H. Naylor, “Galileo’s theory of projectile motion,” Isis, vol. 71,
no. 4, pp. 550–570, 1980.
[3] D. N. Mavris, D. A. DeLaurentis, O. Bandte, and M. A. Hale, “A
stochastic approach to multi-disciplinary aircraft analysis and design,”
in 36th Aerospace Sciences Meeting & Exhibit, Reno, NV, 1998.
[4] A. Charters, “The linearized equations of motion underlying the
dynamic stability of aircraft, spinning projectiles, and symmetrical
missiles,” 1955.
[5] B. J. Guidos and G. R. Cooper, “Linearized motion of a fin-stabilized
projectile subjected to a lateral impulse,” Journal of spacecraft and
rockets, vol. 39, no. 3, pp. 384–391, 2002.
[6] L. Zhao and Y. Jia, “Finite-time attitude stabilisation for a class of
stochastic spacecraft systems,” IET Control Theory & Applications,
vol. 9, no. 8, pp. 1320–1327, 2015.
[7] J. Rogers, “Stochastic model predictive control for guided projectiles
under impact area constraints,” Journal of Dynamic Systems, Measurement, and Control, vol. 137, no. 3, p. 034503, 2015.
[8] M. Finn, “Value of area defense impact-point prediction in a two-layer
defense with perfect attackers and defenders. final report,” Institute for
Defense Analyses, Alexandria, VA (USA), Tech. Rep., 1986.
[9] A. Fieee, S. Immediata, L. Timmoneri, M. Meloni, and D. Vigilante,
“Comparison of recursive and batch processing for impact point prediction of ballistic targets,” in IEEE International Radar Conference,
2005, pp. 121–126.
[10] A. Farina, L. Timmoneri, and D. Vigilante, “Classification and launchimpact point prediction of ballistic target via multiple model maximum
likelihood estimator (mm-mle),” in IEEE Conference on Radar, 2006,
pp. 802–806.
[11] T. Yuan, Y. Bar-Shalom, P. Willett, and D. Hardiman, “Impact point
prediction for thrusting projectiles in the presence of wind,” IEEE
Transactions on Aerospace and Electronic Systems, vol. 50, no. 1, pp.
102–119, 2014.
[12] R. Yanushevsky, Modern missile guidance. CRC Press, 2007.
[13] J. F. Manwell, J. G. McGowan, and A. L. Rogers, Wind energy
explained: theory, design and application. John Wiley & Sons, 2010.
[14] M. Soltani, C. A. Vargas-Garcia, and A. Singh, “Conditional moment
closure schemes for studying stochastic dynamics of genetic circuits,”
IEEE transactions on biomedical circuits and systems, vol. 9, no. 4,
pp. 518–526, 2015.
[15] C. H. Lee, K.-H. Kim, and P. Kim, “A moment closure method for
stochastic reaction networks,” The Journal of chemical physics, vol.
130, no. 13, p. 134107, 2009.
[16] C. S. Gillespie, “Moment-closure approximations for mass-action
models,” IET systems biology, vol. 3, no. 1, pp. 52–58, 2009.
[17] C. Kuehn, “Moment closurea brief review,” in Control of selforganizing nonlinear systems. Springer, 2016, pp. 253–271.
[18] A. Singh and J. P. Hespanha, “Approximate moment dynamics for
chemically reacting systems,” IEEE Transactions on Automatic Control, vol. 56, no. 2, pp. 414–418, 2011.
[19] V. M. Tsarenko and A. M. Yaglom, Second-Order Modeling Of
Turbulent Diffusion in an Atmospheric Surface Layer. Springer New
York, 1992, pp. 39–58.
[20] K. R. Ghusinga, M. Soltani, A. Lamperski, S. Dhople, and A. Singh,
“Approximate moment dynamics for polynomial and trigonometric
stochastic systems,” arXiv preprint arXiv:1703.08841, 2017.
[21] W. Zhang, J. Yang, and J. Zhan, “Analysis on guidance law performance of semi-active radar guided surface-to-air missile,” Journal of
Projectiles, Rockets, Missiles and Guidance, vol. 6, p. 007, 2008.
[22] P. Metz, “Terminal guidance method and a guided missile operating
according to this method,” Feb. 4 1986, uS Patent 4,568,040.
[Online]. Available: https://www.google.com/patents/US4568040
[23] H. Wang, D. Cao, and X. Wang, “The stochastic sliding mode variable
structure guidance laws based on optimal control theory,” Journal of
Control Theory and Applications, vol. 11, no. 1, pp. 86–91, 2013.
[24] M. Gross and M. Costello, “Impact point model predictive control of
a spin-stabilized projectile with instability protection,” Proceedings of
the Institution of Mechanical Engineers, Part G: Journal of Aerospace
Engineering, vol. 228, no. 12, pp. 2215–2225, 2014.
[25] M. F. Costello, “Potential field artillery projectile improvement using
movable canards.” Military Academy West Point NY, Tech. Rep.,
1997.
[26] R. McCoy, Modern exterior ballistics: The launch and flight dynamics
of symmetric projectiles. Schiffer Pub., 1999.
[27] M. J. Amoruso, “Euler angles and quaternions in six degree of freedom
simulations of projectiles,” Army Armament Research Development
and Engineering Center Picatinny Arsenal NJ Armament Engineering
Directorate, Tech. Rep., 1996.
[28] F. Fresconi, I. Celmins, and S. I. Silton, “Theory, guidance, and flight
control for high maneuverability projectiles,” Army Research Lab
Aberdeen Proving MD Weapons and Materials Research Directorate,
Tech. Rep., 2014.
[29] R. F. Stengel, Exploring the Flight Envelope. Princeton University
Press, 2015, pp. 29–146.
[30] M. Costello, “Extended range of a gun launched smart projectile using
controllable canards,” Shock and Vibration, vol. 8, no. 3-4, pp. 203–
213, 2001.
[31] M. V. Cook, Flight dynamics principles: a linear systems approach to
aircraft stability and control. Butterworth-Heinemann, 2012.
[32] J. D. Anderson Jr, Fundamentals of aerodynamics. Tata McGraw-Hill
Education, 2010.
[33] B. Etkin, Dynamics of atmospheric flight. Courier Corporation, 2012.
[34] C. H. Murphy, “Free flight motion of symmetric missiles,” Army
Ballistic Research Lab Aberdeen Proving Ground MD, Tech. Rep.,
1963.
[35] G. Cooper and M. Costello, “Flight dynamic response of spinning
projectiles to lateral impulsive loads,” Journal of Dynamic Systems,
Measurement, and Control, vol. 126, no. 3, pp. 605–613, 2004.
[36] L. C. Hainz and M. Costello, “Modified projectile linear theory
for rapid trajectory prediction,” Journal of Guidance Control and
Dynamics, vol. 28, no. 5, pp. 1006–1014, 2005.
[37] T. Yuan, Y. Bar-Shalom, P. Willett, D. Hardiman, and H. A. AMRDEC,
“Impact point prediction for short range thrusting projectiles,” in Proc.
of SPIE Vol, vol. 7698, 2010, pp. 76 981G–1.
[38] R. G. Hutchins and A. San Jose, “Imm tracking of a theater ballistic
missile during boost phase,” in SPIE Proceedings on Signal and Data
Processing of Small Targets, vol. 3373, 1998, pp. 528–537.
[39] Angle-only tracking and prediction of boost vehicle position, vol. 1481,
1991.
[40] J. M. Maley, “Optimal estimation with two process models and no
measurements,” Army Research Lab Aberdeen Proving Ground MD
Weapons and Materials Reseach Directorate, Tech. Rep., 2015.
[41] J. P. Hespanha and A. Singh, “Stochastic models for chemically reacting systems using polynomial stochastic hybrid systems,” International
Journal of robust and nonlinear control, vol. 15, no. 15, pp. 669–689,
2005.
[42] A. Bobbio, M. Gribaudo, and M. Telek, “Mean field methods in performance analysis,” in Fifth International Conference on Quantitative
Evaluation of Systems, QEST, 2008, pp. 215–224.
[43] S. Chibbaro and J.-P. Minier, “Stochastic modelling of polydisperse
turbulent two-phase flows,” in Stochastic Methods in Fluid Mechanics.
Springer, 2014, pp. 39–86.
[44] M. D. Vrettas, M. Opper, and D. Cornford, “Variational meanfield algorithm for efficient inference in large systems of stochastic
differential equations,” Physical Review E, vol. 91, no. 1, p. 012148,
2015.
[45] D. Ollerenshaw and M. Costello, “Model predictive control of a direct
fire projectile equipped with canards,” Journal of Dynamic Systems,
Measurement, and Control, vol. 130, no. 6, p. 061010, 2008.
[46] J. Rogers, M. Costello, and G. Cooper, “Design considerations for
stability of liquid payload projectiles,” Journal of Spacecraft and
Rockets, vol. 50, no. 1, pp. 169–178, 2013.
[47] K. R. Ghusinga, C. A. Vargas-Garcia, A. Lamperski, and A. Singh,
“Exact lower and upper bounds on stationary moments in stochastic
biochemical systems,” Physical Biology, vol. 14, 2017.
[48] A. Lamperski, K. R. Ghusinga, and A. Singh, “Stochastic optimal
control using semidefinite programming for moment dynamics,” in
Decision and Control (CDC), 2016 IEEE 55th Conference on. IEEE,
2016, pp. 1990–1995.
[49] M. Soltani and A. Singh, “Stochastic analysis of linear time-invariant
systems with renewal transitions,” in 2017 American Control Conference (ACC), 2017, pp. 1734–1739.
[50] ——, “Control design and analysis of a stochastic network control
system,” arXiv preprint:1704.00236, 2017.
[51] ——, “Moment-based analysis of stochastic hybrid systems with
renewal transitions,” Automatica, vol. 84, pp. 62–69, 2017.
| 3cs.SY
|
Fast Estimation of the Median Covariation Matrix with
Application to Online Robust Principal Components
arXiv:1504.02852v5 [math.ST] 9 Jul 2016
Analysis
Hervé Cardot, Antoine Godichon-Baggioni
Institut de Mathématiques de Bourgogne,
Université de Bourgogne Franche-Comté,
9, rue Alain Savary, 21078 Dijon, France
July 12, 2016
Abstract
The geometric median covariation matrix is a robust multivariate indicator of dispersion which can be extended without any difficulty to functional data. We define
estimators, based on recursive algorithms, that can be simply updated at each new
observation and are able to deal rapidly with large samples of high dimensional data
without being obliged to store all the data in memory. Asymptotic convergence properties of the recursive algorithms are studied under weak conditions. The computation
of the principal components can also be performed online and this approach can be
useful for online outlier detection. A simulation study clearly shows that this robust
indicator is a competitive alternative to minimum covariance determinant when the
dimension of the data is small and robust principal components analysis based on
projection pursuit and spherical projections for high dimension data. An illustration
on a large sample and high dimensional dataset consisting of individual TV audiences
measured at a minute scale over a period of 24 hours confirms the interest of considering the robust principal components analysis based on the median covariation matrix.
All studied algorithms are available in the R package Gmedian on CRAN.
Keywords. Averaging, Functional data, Geometric median, Online algorithms, Online
principal components, Recursive robust estimation, Stochastic gradient, Weiszfeld’s algorithm.
1
Introduction
Principal Components Analysis is one of the most useful statistical tool to extract information by reducing the dimension when one has to analyze large samples of multivariate or
functional data (see e.g. Jolliffe (2002) or Ramsay and Silverman (2005)). When both the
dimension and the sample size are large, outlying observations may be difficult to detect
automatically. Principal components, which are derived from the spectral analysis of the
1
covariance matrix, can be very sensitive to outliers (see Devlin et al. (1981)) and many
robust procedures for principal components analysis have been considered in the literature
(see Hubert et al. (2008), Huber and Ronchetti (2009) and Maronna et al. (2006)).
The most popular approaches are probably the minimum covariance determinant estimator (see Rousseeuw and van Driessen (1999)) and the robust projection pursuit (see
Croux and Ruiz-Gazen (2005) and Croux et al. (2007)). Robust PCA based on projection
pursuit has been extended to deal with functional data in Hyndman and Ullah (2007) and
Bali et al. (2011). Adopting another point of view, robust modifications of the covariance matrix, based on projection of the data onto the unit sphere, have been proposed in
Locantore et al. (1999) (see also Gervini (2008) and Taskinen et al. (2012)).
We consider in this work another robust way of measuring association between variables, that can be extended directly to functional data. It is based on the notion of median
covariation matrix (MCM) which is defined as the minimizer of an expected loss criterion
based on the Hilbert-Schmidt norm (see Kraus and Panaretos (2012) for a first definition in
a more general M -estimation setting). It can be seen as a geometric median (see Kemperman (1987) or Möttönen et al. (2010)) in the particular Hilbert spaces of square matrices
(or operators for functional data) equipped with the Frobenius (or Hilbert-Schmidt) norm.
The MCM is non negative and unique under weak conditions. As shown in Kraus and
Panaretos (2012) it also has the same eigenspace as the usual covariance matrix when the
distribution of the data is symmetric and the second order moment is finite. Being a spatial
median in a particular Hilbert space of matrices, the MCM is also a robust indicator of
central location, among the covariance matrices, which has a 50 % breakdown point (see
Kemperman (1987) or Maronna et al. (2006)) as well as a bounded gross sensitivity error
(see Cardot et al. (2013)).
The aim of this work is twofold. It provides efficient recursive estimation algorithms
of the MCM that are able to deal with large samples of high dimensional data. By this
recursive property, these algorithms can naturally deal with data that are observed sequentially and provide a natural update of the estimators at each new observation. Another
advantage compared to classical approaches is that such recursive algorithms will not require to store all the data. Secondly, this work also aims at highlighting the interest of
considering the median covariation matrix to perform principal components analysis of
high dimensional contaminated data.
Different algorithms can be considered to get effective estimators of the MCM. When
the dimension of the data is not too high and the sample size is not too large, Weiszfeld’s
algorithm (see Weiszfeld (1937) and Vardi and Zhang (2000)) can be directly used to
estimate effectively both the geometric median and the median covariation matrix. When
both the dimension and the sample size are large this static algorithm which requires
to store all the data may be inappropriate and ineffective. We show how the algorithm
developed by Cardot et al. (2013) for the geometric median in Hilbert spaces can be adapted
to estimate recursively and simultaneously the median as well as the median covariation
matrix. Then an averaging step (Polyak and Juditsky (1992)) of the two initial recursive
2
estimators of the median and the MCM permits to improve the accuracy of the initial
stochastic gradient algorithms. A simple modification of the stochastic gradient algorithm
is proposed in order to ensure that the median covariance estimator is non negative. We
also explain how the eigenelements of the estimator of the MCM can be updated online
without being obliged to perform a new spectral decomposition at each new observation.
The paper is organized as follows. The median covariation matrix as well as the recursive estimators are defined in Section 2. In Section 3, almost sure and quadratic mean
consistency results are given for variables taking values in general separable Hilbert spaces.
The proofs, which are based on new induction steps compared to Cardot et al. (2013), allow
to get better convergence rates in quadratic mean even if this new framework is much more
complicated because two averaged non linear algorithms are running simultaneously. One
can also note that the techniques generally employed to deal with two time scale Robbins
Monro algorithms (see Mokkadem and Pelletier (2006) for the multivariate case) require
assumptions on the rest of the Taylor expansion and the finite dimension of the data that
are too restrictive in our framework. In Section 4, a comparison with some classic robust
PCA techniques is made on simulated data. The interest of considering the MCM is also
highlighted on the analysis of individual TV audiences, a large sample of high dimensional
data which, because of its dimension, can not be analyzed in a reasonable time with classical robust PCA approaches. The main parts of the proofs are described in Section 5.
Perspectives for future research are discussed in Section 6. Some technical parts of the
proofs as well as a description of Weiszfeld’s algorithm in our context are gathered in an
Appendix.
2
Population point of view and recursive estimators
Let H be a separable Hilbert space (for example H = Rd or H = L2 (I), for some closed
interval I ⊂ R). We denote by h., .i its inner product and by k·k the associated norm.
We consider a random variable X that takes values in H and define its center m ∈ H
as follows:
m := arg min E [kX − uk − kXk] .
u∈H
(1)
The solution m ∈ H is often called the geometric median of X. It is uniquely defined
under broad assumptions on the distribution of X (see Kemperman (1987)) which can be
expressed as follows.
Assumption 1. There exist two linearly independent unit vectors (u1 , u2 ) ∈ H 2 , such that
for u ∈ {u1 , u2 }.
Var(hu, Xi) > 0,
If the distribution of X − m is symmetric around zero and if X admits a first moment
that is finite then the geometric median is equal to the expectation of X, m = E [X].
Note however that the general definition (1) does not require to assume that the first order
moment of kXk is finite since |E [kX − uk − kXk] | ≤ kuk.
3
2.1
The (geometric) median covariation matrix (MCM)
We now consider the special vector space, denoted by S(H), of d×d matrices when H = Rd ,
or for general separable Hilbert spaces H, the vector space of linear operators mapping
H → H. Denoting by {ej , j ∈ J} an orthonormal basis in H, the vector space S(H)
equipped with the following inner product:
hA, BiF =
X
hAej , Bej i
(2)
j∈J
is also a separable Hilbert space. In S(Rd ), we have equivalently
hA, BiF = tr AT B ,
(3)
where AT is the transpose matrix of A. The induced norm is the well known Frobenius
norm (also called Hilbert-Schmidt norm) and is denoted by k.kF .
When X has finite second order moments, with expectation E [X] = µ, the covariance
matrix of X, E (X − µ)(X − µ)T can be defined as the minimum argument, over all the
elements belonging to S(H), of the functional Gµ,2 : S(H) → R,
h
i
2
2
Gµ,2 (Γ) = E (X − µ)(X − µ)T − Γ F − (X − µ)(X − µ)T F .
Note that in general Hilbert spaces with inner product h., .i, operator (X − µ)(X − µ)T
should be understood as the operator u ∈ H 7→ hu, X − µi(X − µ). The MCM is obtained
by removing the squares in previous function in order to get a more robust indicator of
"covariation". For α ∈ H, define Gα : S(H) → R by
Gα (V ) := E
(X − α)(X − α)T − V
F
− (X − α)(X − α)T
F
.
(4)
The median covariation matrix, denoted by Γm , is defined as the minimizer of Gm (V ) over
all elements V ∈ S(H). The second term at the right-hand side of (4) prevents from having
to introduce hypotheses on the existence of the moments of X. Introducing the random
variable Y := (X − m)(X − m)T that takes values in S(H), the MCM is unique provided
that the support of Y is not concentrated on a line and Assumption 1 can be rephrased as
follows in S(H),
Assumption 2. There exist two linearly independent unit vectors (V1 , V2 ) ∈ S(H)2 , such
that
Var(hV, Y iF ) > 0, for V ∈ {V1 , V2 }.
We can remark that Assumption 1 and Assumption 2 are strongly connected. Indeed, if
Assumption 1 holds, then Var(hu, Xi) > 0 for u ∈ {u1 , u2 }. Consider the rank one matrices
V1 = u1 uT1 and V2 = u2 uT2 , we have hV1 , Y iF = hu1 , X − mi2 which has a strictly positive
variance when the distribution of X has no atom. More generally Var(hV1 , Y iF ) > 0
unless there is a scalar a > 0 such that P [hu1 , X − mi = a] = P [hu1 , X − mi = −a] = 12
(assuming also that P [X − m = 0] = 0).
4
Furthermore it can be deduced easily that the MCM, which is a geometric median in
the particular Hilbert spaces of Hilbert-Schmidt operators, is a robust indicator with a
50% breakdown point (see Kemperman (1987)) and a bounded sensitive gross error (see
Cardot et al. (2013)).
We also assume that
Assumption 3. There is a constant C such that for all h ∈ H and all V ∈ S(H)
i
h
−1
T
(a) : E (X − h)(X − h) − V F ≤ C.
h
i
−2
(b) : E (X − h)(X − h)T − V F ≤ C.
This assumption implicitly forces the distribution of (X −h)(X −h)T to have no atoms.
It is more "likely" to be satisfied when the dimension d of the data is large (see Chaudhuri
(1992) and Cardot et al. (2013) for a discussion). Note that it could be weakened as in
Cardot et al. (2013) by allowing points, necessarily different from the MCM Γm , to have
strictly positive masses. Considering the particular case V = 0, Assumption 3(a) implies
that for all h ∈ H,
1
≤ C,
(5)
E
kX − hk2
and this is not restrictive when the dimension d of H is equal or larger than 3.
Under Assumption 3(a), the functional Gh is twice Fréchet differentiable, with gradient
(X − h)(X − h)T − V
∇Gh (V ) = −E
.
(6)
k(X − h)(X − h)T − V kF
and Hessian operator, ∇2h G(V ) : S(H) → S(H),
"
!#
1
(Y
(h)
−
V
)
⊗
(Y
(h)
−
V
)
F
∇2h G(V ) = E
IS(H) −
.
kY (h) − V kF
kY (h) − V kF 2
(7)
where Y (h) = (X − h)(X − h)T , IS(H) is the identity operator on S(H) and A ⊗F B(V ) =
hA, V iF B for any elements A, B and V belonging to S(H).
Furthermore, Γm is also defined as the unique zero of the non linear equation:
∇Gm (Γm ) = 0.
Remarking that previous equality can be rewritten as follows,
1
(X − m)(X − m)T
iE
,
Γm = h
1
k(X − m)(X − m)T − Γm kF
E k(X−m)(X−m)
T −Γ k
m
(8)
(9)
F
it is clear that Γm is a bounded, symmetric and non negative operator in S(H).
As stated in Proposition 2 of Kraus and Panaretos (2012), operator Γm has an important stability property when the distribution of X is symmetric, with finite second moment,
5
h
i
i.e E kXk2 < ∞. Indeed, the covariance operator of X, Σ = E (X − m)(X − m)T ,
which is well defined in this case, and Γm share the same eigenvectors: if ej is an eigenvector of Σ with corresponding eigenvalue λj , then Γm ej = λ̃j ej , for some non negative
value λ̃j . This important result means that for Gaussian and more generally symmetric
distribution (with finite second order moments), the covariance operator and the median
covariation operator have the same eigenspaces. Note that it is also conjectured in Kraus
and Panaretos (2012) that the order of the eigenfunctions is also the same.
2.2
Efficient recursive algorithms
We suppose now that we have i.i.d. copies X1 , . . . , Xn , . . . of random variables with the
same law as X.
For simplicity, we temporarily suppose that the median m of X is known. We consider
a sequence of (learning) weights γn = cγ /nα , with cγ > 0 and 1/2 < α < 1 and we define
the recursive estimation procedure as follows
(Xn+1 − m)(Xn+1 − m)T − Wn
k(Xn+1 − m)(Xn+1 − m)T − Wn kF
1
= Wn −
W n − Wn+1 .
n+1
Wn+1 = Wn + γn
(10)
W n+1
(11)
This algorithm can be seen as a particular case of the averaged stochastic gradient algorithm studied in Cardot et al. (2013). Indeed, the first recursive algorithm (10) is a
stochastic gradient algorithm,
(Xn+1 − m)(Xn+1 − m)T − Wn
E
|F
n = ∇Gm (Wn )
k(Xn+1 − m)(Xn+1 − m)T − Wn kF
where Fn = σ(X1 , . . . , Xn ) is the σ-algebra generated by X1 , . . . , Xn whereas the final
estimator W n is obtained by averaging the past values of the first algorithm. The averaging
step (see Polyak and Juditsky (1992)), i.e. the computation of the arithmetical mean of the
past values of a slowly convergent estimator (see Proposition 3.4 below), permits to obtain
a new and efficient estimator converging at a parametric rate, with the same asymptotic
variance as the empirical risk minimizer (see Theorem 3.1 below).
In most of the cases the value of m is unknown so that it also required to estimate the
median. To build an estimator of Γm , it is possible to estimate simultaneously m and Γm by
considering two averaged stochastic gradient algorithms that are running simultaneously.
For n ≥ 1,
mn+1 = mn + γn(m)
Xn+1 − mn
kXn+1 − mn k
1
(mn − mn+1 )
n+1
(Xn+1 − mn )(Xn+1 − mn )T − Vn
= V n + γn
k(Xn+1 − mn )(Xn+1 − mn )T − Vn kF
1
=Vn−
V n − Vn+1 ,
n+1
mn+1 = mn −
(12)
Vn+1
(13)
V n+1
6
(14)
where the averaged recursive estimator mn+1 of the median m is controlled by a sequence
(m)
(m)
of descent steps γn . The learning rates are generally chosen as follows, γn = cm n−α ,
where the tuning constants satisfy cm ∈ [2, 20] and 1/2 < α < 1.
Note that by construction, even if Vn is non negative, Vn+1 may not be a non negative
matrix when the learning steps do not satisfy
γn
≤ 1.
k(Xn+1 − mn )(Xn+1 − mn )T − Vn kF
Projecting Vn+1 onto the closed convex cone of non negative operators would require to
compute the eigenvalues of Vn+1 which is time consuming in high dimension even if Vn+1 is
a rank one perturbation to Vn (see Cardot and Degras (2015)). We consider the following
simple approximation to this projection which consists in replacing in (13) the descent step
γn by a thresholded one,
γn,pos = min γn ,
(Xn+1 − mn )(Xn+1 − mn )T − Vn
F
(15)
which ensures that Vn+1 remains non negative when Vn is non negative. The use of these
modified steps and an initialization of the recursive algorithm (13) with a non negative
matrix (for example V0 = 0) ensure that for all n ≥ 1, Vn and V n are non negative.
2.3
Online estimation of the principal components
It is also possible to approximate recursively the q eigenvectors (unique up to sign) of
Γm associated to the q largest eigenvalues without being obliged to perform a spectral
decomposition of V n+1 at each new observation. Many recursive strategies can be employed
(see Cardot and Degras (2015) for a review on various recursive estimation procedures of
the eigenelements of a covariance matrix). Because of its simplicity and its accuracy, we
consider the following one:
uj,n
1
uj,n+1 = uj,n +
V n+1
− uj,n , j = 1, . . . , q
(16)
n+1
kuj,n k
combined with an orthogonalization by deflation of u1,n+1 , . . . uq,n+1 . This recursive algorithm is based on ideas developed by Weng et al. (2003) that are related to the power
method for extracting eigenvectors. If we assume that the q first eigenvalues λ1 > · · · > λq
are distinct, the estimated eigenvectors u1,n+1 , . . . uq,n+1 , which are uniquely determined
up to sign change, tend to λ1 u1 , . . . , λq uq .
Once the eigenvectors are computed, it is possible to compute the principal components
as well as indices of outlyingness for each new observation (see Hubert et al. (2008) for a
review of outliers detection with multivariate approaches).
2.4
Practical issues, complexity and memory
The recursive algorithms (13) and (14) require each O(d2 ) elementary operations at each
update. With the additional online estimation given in (16) of the q eigenvectors associated
7
to the q largest eigenvalues, O(qd2 ) additional operations are required. The orthogonalization procedure only requires O(q 2 d) elementary operations.
Note that the use of classical Newton-Raphson algorithms for estimating the MCM (see
Fritz et al. (2012)) can not be envisaged for high dimensional data since the computation
or the approximation of the Hessian matrix would require O(d4 ) elementary operations.
The well known and fast Weiszfeld’s algorithm requires O(nd2 ) elementary operations for
each sample with size n. However, the estimation cannot be updated automatically if the
data arrive sequentially. Another drawback compared to the recursive algorithms studied
in this paper is that all the data must be stored in memory, which is of order O(nd2 )
elements whereas the recursive technique require an amount of memory of order O(d2 ).
The performances of the recursive algorithms depend on the values of tuning parameters
cγ , cm and α. The value of parameter α is often chosen to be α = 2/3 or α = 3/4. Previous
empirical studies (see Cardot et al. (2013) and Cardot et al. (2010)) have shown that, thanks
to the averaging step, estimator mn performs well and is not too sensitive to the choice of
cm , provided that the value of cm is not too small. An intuitive explanation could be that
here the recursive process is in some sense "self-normalized" since the deviations at each
iteration in (10) have unit norm and finding some universal values for cm is possible. Usual
values for cm and cγ are in the interval [2, 20]. When n is fixed, this averaged recursive
algorithm is about 30 times faster than the Weiszfeld’s approach (see Cardot et al. (2013)).
3
Asymptotic properties
When m is known, W n can be seen as an averaged stochastic gradient estimator of the
geometric median in a particular Hilbert space and the asymptotic weak convergence of
such estimator has been studied in Cardot et al. (2013). They have shown that:
Theorem 3.1. (Cardot et al. (2013), Theorem 3.4).
If assumptions 1-3(a) hold, then as n tends to infinity,
√
n W n − Γm
N (0, ∆)
−1
−1
2
2
stands for convergence in distribution
is
h and ∆ = ∇m (Γm ) i Ψ ∇m (Γm )
(Y (m)−Γm )⊗F (Y (m)−Γm )
the limiting covariance operator, with Ψ = E
.
kY (m)−Γ k 2
where
m F
As explained in Cardot et al. (2013), the estimator W n is efficient in the sense that it
has the same asymptotic distribution as the empirical risk minimizer related to Gm (V ) (see
for the derivation of its asymptotic normality in Möttönen et al. (2010) in the multivariate
case and Chakraborty and Chaudhuri (2014) in a more general functional framework).
Using the delta method for weak convergence in Hilbert spaces (see Dauxois et al. (1982)
or Cupidon et al. (2007)), one can deduce, from Theorem 3.1, the asymptotic normality of
the estimated eigenvectors of W n . It can also be proven (see Godichon-Baggioni (2016)),
under Assumptions 1-3, that there is a positive constant K such that for all n ≥ 1,
h
i K
2
E W n − Γm F ≤ .
n
8
Note finally that non asymptotic bounds for the deviation of W n around Γm can be derived
readily with the general results given in Cardot et al. (2016).
The more realistic case in which m must also be estimated is more complicated because
V n depends on mn which is also estimated recursively with the same data. We first state
the strong consistency of the estimators Vn and V n .
Theorem 3.2. If assumptions 1-3(b) hold, we have
lim kVn − Γm kF = 0
n→∞
a.s.
and
lim
n→∞
V n − Γm
F
=0
a.s.
The obtention of the rate convergence of the averaged recursive algorithm relies on a
fine control of the asymptotic behavior of the Robbins-Monro algorithms, as stated in the
following proposition.
Theorem 3.3. If assumptions 1-3(b) hold, there is a positive constant C 0 , and for all
β ∈ (α, 2α), there is a positive constant Cβ such that for all n ≥ 1,
h
i C0
E kVn − Γm k2F ≤ α ,
n
h
i C 00
E kVn+1 − Γm k4F ≤ β .
n
The obtention of an upper bound for the rate of convergence at the order four of the
Robbins-Monro algorithm is crucial in the proofs. Furthermore, the following proposition
ensures that the exhibited rate in quadratic mean is the optimal one.
Proposition 3.4. Under assumptions 1-3(b), there is a positive constant c0 such that for
all n ≥ 1,
h
i
c0
E kVn − Γm k2F ≥ α .
n
Finally, the following theorem is the most important theoretical result of this work. It
shows that, in spite of the fact that it only considers the observed data one by one, the
averaged recursive estimation procedure gives an estimator which has a classical parametric
√
n rate of convergence in the Hilbert-Schmidt norm.
Theorem 3.5. Under Assumptions 1-3(b), there is a positive constant K 0 such that for
all n ≥ 1,
h
i K0
2
E V n − Γm F ≤
.
n
Assuming the eigenvalues of Γm are of multiplicity one, it can be deduced from Theorem 3.5 and Lemma 4.3 in Bosq (2000), the convergence in quadratic mean of the eigenvectors of V n towards the corresponding (up to sign) eigenvector of Γm .
9
4
An illustration on simulated and real data
A small comparison with other classical robust PCA techniques is performed in this section
considering data in relatively high dimension but samples with moderate sizes. This permits to compare our approach with classical robust PCA techniques, which are generally
not designed to deal with large samples of high dimensional data. In our comparison, we
have employed the following well known robust techniques: robust projection pursuit (see
Croux and Ruiz-Gazen (2005) and Croux et al. (2007)), minimum covariance determinant
(MCD, see Rousseeuw and van Driessen (1999)) and spherical PCA (see Locantore et al.
(1999)). The computations were made in the R language (R Development Core Team
(2010)), with the help of packages pcaPP and rrcov. For reproductible research, our
codes for computing the MCM have been posted on CRAN in the Gmedian package. We
will denote by MCM(R) the recursive estimator V n defined in (14) and MCM(R+) its non
negative modification whose learning weights are defined in (15).
If the size of the data n × d is not too large, an effective way for estimating Γm is to
employ Weiszfeld’s algorithm (see Weiszfeld (1937) and Vardi and Zhang (2000) as well
the Supplementary file for a description of the algorithms in our particular situation). The
estimate obtained thanks to Weiszfeld’s algorithm is denoted by MCM(W) in the following.
Note that other optimization algorithms which may be preferred in small dimension (see
Fritz et al. (2012)) have not been considered here since they would require the computation
of the Hessian matrix, whose size is d4 , and this would lead to much slower algorithms.
Note finally that all these alternative algorithms do not admit a natural updating scheme
when the data arrive sequentially so that they should be completely ran again at each new
observation.
4.1
Simulation protocol
Independent realizations of a random variable Y ∈ Rd are drawn, where
Y
= (1 − O(δ))X + O(δ),
(17)
is a mixture of two distributions and X, O and are independent random variables. The
random vector X has a centered Gaussian distribution in Rd with covariance matrix [Σ]`,j =
min(`, j)/d and can be thought as a discretized version of a Brownian sample path in
[0, 1]. The multivariate contamination comes from , with different rates of contamination
controlled by the Bernoulli variable O(δ), independent from X and , with P(O(δ) = 1) = δ
and P(O(δ) = 0) = 1 − δ. Three different scenarios (see Figure 1) are considered for the
distribution of :
• The elements of vector are d independent realizations of a Student t distribution
with one degree of freedom. This means that the first moment of Y is not defined
when δ > 0.
• The elements of vector are d independent realizations of a Student t distribution
10
0
10
20
30
40
50
1
-2
-1
0
1
0
-1
-2
-2
-1
0
1
2
Reverse time BM
2
Student 2 df
2
Student 1 df
0
10
20
30
40
50
0
10
20
30
40
Figure 1: A sample of n = 20 trajectories when d = 50 and δ = 0.10 for the three different
contamination scenarios: Student t with 1 degree of freedom, Student t with 2 degrees of
freedom and reverse time Brownian motion (from left to right).
with two degrees of freedom. This means that the second moment of Y is not defined
when δ > 0.
• The vector is distributed as a "reverse time" Brownian motion. It has a Gaussian
centered distribution, with covariance matrix [Σ ]`,j = 2 min(d − `, d − j)/d. The
covariance matrix of Y is (1 − δ)Σ + δΣ .
For the averaged recursive algorithms, we have considered tuning coefficients cm = cγ =
2 and a speed rate of α = 3/4. Note that the values of these tuning parameters have not
been particularly optimised. We have noted that the simulation results were very stable,
and did not depend much on the value of cm and cγ for cm , cγ ∈ [1, 20].
The estimation error of the eigenspaces associated to the largest eigenvalues is evaluated
by considering the squared Frobenius norm between the associated orthogonal projectors.
Denoting by Pq the orthogonal projector onto the space generated by the q eigenvectors of
b q an estimation,
the covariance matrix Σ associated to the q largest eigenvalues and by P
we consider the following loss criterion,
T
b
b
b
R(Pq , Pq ) = tr Pq − Pq
Pq − Pq
h
i
b q Pq .
= 2q − 2tr P
(18)
11
50
0.500
0.200
0.050
0.020
0.005
Oracle
MCD
MCM(W)
MCM(R+)
MCM(R)
SphPCA
PP
Figure 2: Estimation errors (at a logarithmic scale) over 500 Monte Carlo replications,
for n = 200, d = 50 with no contamination (δ = 0). MCM(W) stands for the estimation
performed by the Weiszfeld’s algorithm whereas MCM(R) denotes the averaged recursive
approach and MCM(R+) its non negative modification (see equation 15).
b q , Pq ) ≤ 2q and R(P
b q , Pq ) = 2q means that the eigenspaces
Note that we always have R(P
generated by the true and the estimated eigenvectors are orthogonal.
4.2
Comparison with classical robust PCA techniques
We first compare the performances of the two estimators of the MCM based on the
Weiszfeld’s algorithm and the recursive algorithms (see (14)) with more classical robust
PCA techniques.
We generated samples of Y with size n = 500 and dimension d ∈ {50, 200}, over 500
replications. Different levels of contamination are considered : δ ∈ {0, 0.02, 0.05, 0.10, 0.20}.
For both dimensions d = 50 and d = 200, the first eigenvalue of the covariance matrix of
X represents about 81 % of the total variance, and the second one about 9 %.
The median errors of estimation of the eigenspace generated by the first two eigenvectors
(q = 2), according to criterion (18), are given in Table 1 for non contaminated data (δ = 0).
b q , Pq ) is drawn for the different approaches
The distribution of the estimation error R(P
in Figure 2 when the dimension is not large (d = 50). As expected, the "Oracle", which
is the classical PCA in this situation, provides the best estimations of the eigenspaces.
Then, the MCD and the median covariation matrix, estimated by the Weiszfeld algorithm
12
2.000
0.500
0.200
0.050
0.020
0.005
Oracle
MCD
MCM(W)
MCM(R+)
MCM(R)
SphPCA
PP
PCA
Figure 3: Estimation errors (at a logarithmic scale) over 500 Monte Carlo replications,
for n = 200, d = 50 and a contamination by a t distribution with 2 degrees of freedom
with δ = 0.02. MCM(W) stands for the estimation performed by the Weiszfeld’s algorithm whereas MCM(R) denotes the averaged recursive approach and MCM(R+), its non
negative modification with learning steps as in (15).
13
d=50
d=200
PCA
0.0156
0.0148
MCD
0.0199
-
MCM(W)
0.0208
0.0200
MCM(R+)
0.0211
0.0209
MCM(R)
0.0243
0.0246
SphPCA
0.0287
0.0275
PP
0.0955
0.0895
b q , Pq ) with a dimension
Table 1: Median estimation errors, according to criterion R(P
q = 2, for non contaminated samples of size n = 200, over 500 Monte Carlo experiments.
or the modified MCM(R+) recursive estimator, behave well and similarly. Note that when
the dimension gets larger, the MCD cannot be used anymore and the MCM is the more
effective robust estimator of the eigenspaces.
When the data are contaminated, the median errors of estimation of the eigenspace
generated by the first two eigenvectors (q = 2), according to criterion (18), are given in
b q , Pq ) is drawn for the
Table 2. In Figure 3, the distribution of the estimation error R(P
different approaches.
We can make the following remarks. At first note that even when the level of contamination is small (2% and 5%), the performances of classical PCA are strongly affected
by the presence of outlying values in such (large) dimensions. When d = 50, the MCD
algorithm and the MCM estimation provide the best estimations of the original two dimensional eigenspace, whereas when d gets larger (d = n = 200), the MCD estimator can not
be used anymore (by construction) and the MCM estimators, obtained with Weiszfeld’s
and the non negative recursive algorithm, remain the most accurate. We can also remark
that the recursive MCM algorithms, which are designed to deal with very large samples,
performs well even for such moderate sample sizes (see also Figure 3). The modification
of the descent step suggested in (15), which corresponds to estimator MCM(R+), permits
to improve the accuracy the initial MCM estimator, specially when the noise level is not
small. The performances of the spherical PCA are slightly less accurate whereas the median error of the robust PP is always the largest among the robust estimators. When, the
contamination is highly structured temporally and the level of contamination is not small
(contamination by a reverse time Brownian motion, with δ = 0.20), the behavior of the
MCM is different from the other robust estimators and, with our criterion, it can appear
as less effective. However, one can think that we are in presence of two different populations with completely different multivariate correlation structure and the MCD completely
ignores that part of the data, which is not necessarily a better behavior.
4.3
Online estimation of the principal components
We now consider an experiment in high dimension, d = 1000, and evaluate the ability
of the recursive algorithms defined in (16) to estimate recursively the eigenvectors of Γm
associated to the largest eigenvalues. Note that due to the high dimension of the data and
limited computation time, we only make comparison of the recursive robust techniques with
the classical PCA. For this we generate growing samples and compute, for each sample size
the approximation error of the different (fast) strategies to the true eigenspace generated
14
t 1 df
δ
2%
5%
10%
20%
Method
PCA
PP
MCD
Sph. PCA
MCM (Weiszfeld)
MCM (R+)
MCM (R)
PCA
PP
MCD
Sph. PCA
MCM (Weiszfeld)
MCM (R+)
MCM (R)
PCA
PP
MCD
Sph. PCA
MCM (Weiszfeld)
MCM (R+)
MCM (R)
PCA
PP
MCD
Sph. PCA
MCM (Weiszfeld)
MCM (R+)
MCM (R)
3.13
0.086
0.022
0.028
0.021
0.022
0.026
3.82
0.090
0.022
0.029
0.023
0.025
0.029
3.83
0.107
0.023
0.033
0.025
0.030
0.035
3.84
0.110
0.025
0.037
0.030
0.044
0.050
t 2 df
d = 50
1.04
0.097
0.021
0.029
0.021
0.022
0.025
1.91
0.103
0.023
0.031
0.023
0.024
0.027
1.96
0.108
0.022
0.033
0.026
0.027
0.032
2.02
0.135
0.026
0.038
0.030
0.036
0.041
inv. B.
t 1 df
0.698
0.090
0.021
0.027
0.021
0.024
0.027
0.862
0.093
0.021
0.033
0.028
0.035
0.037
1.03
0.099
0.023
0.054
0.059
0.089
0.088
1.19
0.138
0.026
0.140
0.174
0.255
0.251
3.95
0.085
–
0.027
0.021
0.023
0.026
3.96
0.089
–
0.029
0.022
0.024
0.028
3.96
0.088
–
0.031
0.023
0.027
0.032
3.96
0.091
–
0.034
0.026
0.038
0.042
t 2 df
d = 200
1.87
0.094
–
0.030
0.022
0.023
0.027
1.98
0.098
–
0.031
0.023
0.024
0.028
1.99
0.101
–
0.033
0.024
0.027
0.031
2.01
0.122
–
0.037
0.028
0.032
0.037
inv. B.
0.731
0.084
–
0.028
0.022
0.025
0.028
0.910
0.087
–
0.034
0.030
0.039
0.040
1.10
0.097
–
0.057
0.056
0.086
0.086
1.25
0.137
–
0.150
0.181
0.256
0.256
b q , Pq ) with a dimension
Table 2: Median estimation errors, according to criterion R(P
q = 2, for datasets with a sample size n = 200, over 500 Monte Carlo experiments.
15
by the q eigenvectors associated to the q largest eigenvalues of Γm .
We have drawn in Figure 4, the evolution of the mean (over 100 replications) approximation error R(Pq , P̂q ), for a dimension q = 3, as a function of the sample size
for samples contaminated by a 2 degrees of freedom Student t distribution with a rate
δ = 0.1. An important fact is that the recursive algorithm which approximates recursively
the eigenelements behaves very well and we can see nearly no difference between the spectral decomposition of V n (denoted by MCM in Figure 4) and the estimates produced with
the sequential algorithm (16) for sample sizes larger than a few hundreds. We can also
note that the error made by the classical PCA is always very high and does not decrease
with the sample size.
4.4
Robust PCA of TV audience
The last example is a high dimension and large sample case. Individual TV audiences are
measured, by the French company Médiamétrie, every minutes for a panel of n = 5422
people over a period of 24 hours, d = 1440 (see Cardot et al. (2012) for a more detailed
presentation of the data). With a classical PCA, the first eigenspace represents 24.4% of
the total variability, whereas the second one reproduces 13.5% of the total variance, the
third one 9.64% and the fourth one 6.79%. Thus, more than 54% of the variability of the
data can be captured in a four dimensional space. Taking account of the large dimension
of the data, these values indicate a high temporal correlation.
Because of the large dimension of the data, the Weiszfeld’s algorithm as well as the
other robust PCA techniques can not be used anymore in a reasonable time with a personal
computer. The MCM has been computed thanks to the recursive algorithm given in (14)
in approximately 3 minutes on a laptop in the R language (without any specific C routine).
As seen in Figure 5, the first two eigenvectors obtained by a classical PCA and the
robust PCA based on the MCM are rather different. This is confirmed by the relatively
large distance between the two corresponding eigenspaces, R(Pb2P CA , Pb2M CM ) = 0.56. The
first robust eigenvector puts the stress on the time period comprised between 1000 minutes
and 1200 minutes whereas the first non robust eigenvector focuses, with a smaller intensity,
on a larger period of time comprised between 600 and 1200 minutes. The second robust
eigenvector differentiates between people watching TV during the period between 890 and
1050 minutes (negative value of the second principal component) and people watching TV
between minutes 1090 and 1220 (positive value of the second principal component). Rather
surprisingly, the third and fourth eigenvectors of the non robust and robust covariance
matrices look quite similar (see Figure 6).
5
Proofs
We give in this Section the proofs of Theorems 3.2, 3.3 and 3.5. These proofs rely on
several technical Lemmas whose proofs are given in the Supplementary file.
16
5.000
2.000
0.500
0.200
0.050
0.005
0.020
Estimation error (R)
PCA
Oracle
MCM
MCM−update
0
500
1000
1500
2000
n
b q )) with d = 1000 and q =
Figure 4: Estimation errors of the eigenspaces (criterion R(P
3 for classical PCA, the oracle PCA and the recursive MCM estimator with recursive
estimation of the eigenelements (MCM-update) and with static estimation (based on the
spectral decomposition of V n ) of the eigenelements (MCM).
17
PC 2
0.00
-0.06
0.00
-0.04
0.02
0.04
TV audience
PCA
MCM
0.02
0.04
PCA
MCM
-0.02
0.06
0.06
0.08
PC 1
0
200
400
600
800
1000 1200 1400
0
minutes
200
400
600
800
1000 1200 1400
minutes
Figure 5: TV audience data measured the 6th September 2010, at the minute scale. Comparison of the principal components of the classical PCA (black) and robust PCA based on
the Median Covariation Matrix (red). First eigenvectors on the left, second eigenvectors
on the right.
18
PC 4
0.06
0.06
0.08
PC 3
0.00
-0.06
-0.04
-0.04
-0.02
-0.02
0.00
0.02
TV audience
0.02
0.04
0.04
PCA
MCM
PCA
MCM
0
200
400
600
800
1000
1200
1400
0
minutes
200
400
600
800
1000
1200
1400
minutes
Figure 6: TV audience data measured the 6th September 2010, at the minute scale. Comparison of the principal components of the classical PCA (black) and robust PCA based
on the MCM (red). Third eigenvectors on the left, fourth eigenvectors on the right.
19
5.1
Proof of Theorem 3.2
Let us recall the Robbins-Monro algorithm, defined recursively by
Vn+1 = Vn + γn
(Xn+1 − mn ) (Xn+1 − mn )T − Vn
(Xn+1 − mn ) (Xn+1 − mn )T − Vn
F
= Vn − γn Un+1 ,
with Un+1 := −
(Xn+1 −mn )(Xn+1 −mn )T −Vn
k(Xn+1 −mn )(Xn+1 −mn )T −Vn kF
. Since Fn := σ (X1 , ..., Xn ), we have E [Un+1 |Fn ] =
∇Gmn (Vn ). Thus ξn+1 := ∇mn G(Vn ) − Un+1 , (ξn ) is a sequence of martingale differences
adapted to the filtration (Fn ). Indeed, E [ξn+1 |Fn ] = ∇Gmn (Vn ) − E [Un+1 |Fn ] = 0. The
algorithm can be written as follows
Vn+1 = Vn − γn ∇Gmn (Vn ) + γn ξn+1 .
Moreover, it can be considered as a stochastic gradient algorithm because it can be decomposed as follows:
Vn+1 = Vn − γn (∇Gmn (Vn ) − ∇Gmn (Γm )) + γn ξn+1 − γn rn ,
(19)
with rn := ∇Gmn (Γm ) − ∇Gm (Γm ). Finally, linearizing the gradient,
Vn+1 − Γm = IS(H) − γn ∇2m G(Γm ) (Vn − Γm ) + γn ξn+1 − γn rn − γn rn0 − γn δn ,
(20)
with
rn0 := ∇2mn G (Γm ) − ∇2m G (Γm ) (Vn − Γm ) ,
δn := ∇Gmn (Vn ) − ∇Gmn (Γm ) − ∇2mn G (Γm ) (Vn − Γm ) .
The following lemma gives upper bounds of these remainder terms. Its proof is given
in the Supplementary file.
Lemma 5.1. Under assumptions 1-3(b), we can bound the three remainder terms. First,
kδn kF ≤ 6C kVn − Γm k2F .
(21)
In the same way, for all n ≥ 1,
krn kF ≤ 4
√
q
C + C kΓm kF kmn − mk .
(22)
Finally, for all n ≥ 1,
rn0 F
q
3/4
≤ 12 C kΓm kF + C
kmn − mk kVn − Γm kF .
20
(23)
We deduce from decomposition (33) that for all n ≥ 1,
kVn+1 − Γm k2F = kVn − Γm k2F − 2γn hVn − Γm , ∇Gmn (Vn ) − ∇Gmn (Γm )iF
+ γn2 k∇Gmn (Vn ) − ∇Gmn (Γm )k2F
+ γn2 kξn+1 k2F + 2γn hVn − Γm − γn (∇Gmn (Vn ) − ∇Gmn (Γm )) , ξn+1 iF
+ γn2 krn k2F − 2γn hrn , Vn − Γm iF − 2γn2 hrn , ξn+1 − ∇Gmn (Vn ) + ∇Gmn (Γm )iF .
Note that for all h ∈ H and V ∈ S(H) we have k∇Gh (V )kF ≤ 1. Furthermore, krn kF ≤ 2
and kξn+1 kF ≤ 2. Using the fact that (ξn ) is a sequence of martingale differences adapted
to the filtration (Fn ),
h
i
E kVn+1 − Γm k2F |Fn ≤ kVn − Γm k2F − 2γn hVn − Γm , ∇mn G (Vn ) − ∇mn G (Γm )iF
+ 28γn2 − 2γn hrn , Vn − Γm iF .
Let αn = n−β , with β ∈ (1 − α, α), we have
h
i
E kVn+1 − Γm k2F |Fn ≤ (1 + γn αn ) kVn − Γm k2F − 2γn hVn − Γm , ∇mn G (Vn ) − ∇mn G (Γm )iF
(24)
γn
+ 28γn2 +
krn k2F .
αn
Moreover, applying Lemma 5.1 and Theorem 5.1 in Godichon-Baggioni (2016), we get
for all positive constant δ,
!
1+δ
(ln
n)
a.s.
krn k2F = O kmn − mk2 = O
n
Thus, since 2γn hVn − Γm , ∇mn G (Vn ) − ∇mn G (Γm )iF ≥ 0, the Robbins-Siegmund Theorem (see Duflo (1997) for instance) ensures that kVn − Γm kF converges almost surely to a
finite random variable and
X
γn hVn − Γm , ∇mn G (Vn ) − ∇mn G (Γm )iF < +∞ a.s.
n≥1
Furthermore, by induction, inequality (24) becomes
!
∞
h
i
h
i
Y
2
E kVn+1 − Γm kF ≤
(1 + γk αk ) E kV1 − Γm k2F + 28
k=1
+
∞
Y
k=1
∞
Y
k=1
!
(1 + γk αk )
∞
X
γk2
k=1
∞
i
X
γk h
E krk k2F .
(1 + γk αk )
αk
!
k=1
Since β < α, applying Theorem 4.2 in Godichon-Baggioni (2016) and Lemma 6.1, there is
a positive constant C0 such that
∞
∞
i
X
X
γk h
E krk k2F = C0
k −α−1−β < +∞.
αk
k=1
k=1
21
h
i
Thus, there is a positive constant M such that for all n ≥ 1, E kVn − Γm k2F ≤ M . Since
mn converges almost surely to m, one can conclude the proof of the almost sure consistency
of Vn with the same arguments as in the proof of Theorem 3.1 in Cardot et al. (2013) and
the convexity properties given in the Section B of the supplementary file.
Finally, the almost sure consistency of V n is obtained by a direct application of
Topelitz’s lemma (see e.g. Lemma 2.2.13 in Duflo (1997)).
5.2
Proof of Theorem 3.3
The proof of Theorem 3.3 relies on properties of the p-th moments of Vn for all p ≥ 1
given in the following three Lemmas. These properties enable us, with the application
of Markov’s inequality, to control the probability of the deviations of the Robbins Monro
algorithm from Γm .
Lemma 5.2. Under assumptions 1-3(b), for all integer p, there is a positive constant Mp
such that for all n ≥ 1,
i
h
E kVn − Γm k2p
F ≤ Mp .
Lemma 5.3. Under assumptions 1-3(b), there are positive constants C1 , C10 , C2 , C3 such
that for all n ≥ 1,
h
i
h
i
C2
0 1−α
sup
E kVk − Γm k4 ,
E kVn − Γm k2 ≤ C1 e−C1 n
+ α + C3
n
E(n/2)+1≤k≤n−1
where E(x) is the integer part of the real number x.
Lemma 5.4. Under assumptions 1-3(b), for all integer p0 ≥ 1, there are a rank np0 and
positive constants C1,p0 , C2,p0 , C3,p0 , cp0 such that for all n ≥ np0 ,
h
i
h
i C 0
i
C2,p0 h
C3,p0
− 1−α
1,p
E kVn+1 − Γm k4F ≤ 1 − cp0 γn n p0 E kVn − Γm k4F + 3α + 2α E kVn − Γm k2F +
1−α .
3α−3
n
n
p0
n
We can now prove Theorem 3.3.
Let us choose an integer p0 such that p0 > 3/2. Thus, 2 + α − 3 1−α
≥ 3α, and
p0
applying Lemma 5.4, there are positive constants C1,p0 , C2,p0 , cp0 and a rank np0 such that
for all n ≥ np0 ,
h
i
h
i C 0
i
C2,p0 h
− 1−α
1,p
E kVn+1 − Γm k4F ≤ 1 − cp0 γn n p0 E kVn − Γm k4F + 3α + 2α E kVn − Γm k2F .
n
n
(25)
1−α
1−α
0
0
Let us now choose β ∈ (α, 2α) and p such that p > 2α−β . Note that 3α −β > α + p0 .
One can check that there is a rank n0p0 ≥ np0 such that for all n ≥ n0p0 ,
1
1
+ C3 2β+1
≤ 1,
2
(n + 1)β−α
n + 1 β
C1,p0 + C2,p0
− 1−α
0
1 − cp0 γn n p
+ 23α
≤ 1.
n
(n + 1)3α−β
0
1−α
(n + 1)α C1 e−C1 n
+
22
With the help of a strong induction, we are going to prove the announced results, that
is to say that there are positive constants Cp0 , Cβ such that 2Cp0 ≥ Cβ ≥ Cp0 ≥ 1 and
Cp0 ≥ 2α+1 C2 (with C2 defined in Lemma 5.3), such that for all n ≥ 1,
h
i C0
p
E kVn − Γm k2F ≤ α ,
n
h
i C
β
E kVn − Γm k4F ≤ β .
n
First, let us choose Cp0 and Cβ such that
n
h
io
2
α
Cp0 ≥ max
k
E
kV
−
Γ
k
,
m
k
F
0
k≤np0
Cβ ≥ max
0
k≤np0
β
k E
Vn0 0 − Γm
p
4
.
F
Thus, for all k ≤ n0p0 ,
h
i C0
p
E kVk − Γm k2F ≤ α ,
k
h
i C
β
E kVk − Γm k4F ≤ β .
k
We suppose from now that n ≥ n0p0 and that previous inequalities are verified for all
k ≤ n − 1. Applying Lemma 5.2 and by induction,
n h
io
h
i
C2
0 1−α
sup
E kVk − Γm k4F
E kVn+1 − Γm k2F ≤ C1 e−C1 n
+ α + C3
n
E((n+1)/2)+1≤k≤n
Cβ
C2
−C10 n1−α
≤ C1 e
+ α + C3
sup
n
kβ
E((n+1)/2)+1≤k≤n
0
≤ C1 e−C1 n
1−α
+
Cβ
C2
+ C3 2 β β .
α
n
n
Since 2Cp0 ≥ Cβ ≥ Cp0 ≥ 1 and since Cp0 ≥ 2α+1 C2 , factorizing by
h
E kVn+1 −
Γm k2F
i
Cp0
(n+1)α ,
1
β 2Cp0
+
C
2
3
nα
nβ
α
Cp0
Cp 0
Cp 0
n
C3 2β+1
α
−C10 n1−α
−α
≤
(n
+
1)
C
e
+
2
+
1
(n + 1)α
n+1
2(n + 1)α (n + 1)β−α (n + 1)α
Cp0
Cp0
1 Cp 0
1
0 1−α
≤
C1 (n + 1)α e−C1 n
+
+ C3 2β+1
α
α
β−α
(n + 1)
2 (n + 1)
(n + 1)α
(n + 1)
Cp 0
1
1
0 1−α
≤ (n + 1)α C1 e−C1 n
.
+ + C3 2β+1
β−α
2
(n + 1)α
(n + 1)
0
1−α
≤ Cp0 C1 e−C1 n
+ Cp0 2−α−1
By definition of n0p0 ,
Cp 0
.
(26)
(n + 1)α
In the same way, applying Lemma 5.4 and by induction,
h
i
h
i C 0
i
C2,p0 h
− 1−α
1,p
E kVn+1 − Γm k4F ≤ 1 − cp0 γn n p0 E kVn − Γm k4F + 3α + 2α E kVn − Γm k2F
n
n
C
0
0 Cp 0
C
C
− 1−α
β
1,p
2,p
≤ 1 − cp0 γn n p0
+ 3α + 2α α .
n
n
n
nβ
h
i
E kVn+1 − Γm k2F ≤
23
Since Cβ ≥ Cp0 ≥ 1, factorizing by
Cβ
,
(n+1)β
h
i
Cβ
− 1−α Cβ
E kVn+1 − Γm k4F ≤ 1 − cp0 γn n p0
+ C1,p0 + C2,p0 3α
β
n
n
β
1−α
C
C
Cβ
n+1
−
β
1,p0 + C2,p0
≤ 1 − cp0 γn n p0
+ 23α
β
3α−β
n
n
(n + 1)
(n + 1)β
!
n + 1 β
C1,p0 + C2,p0
Cβ
− 1−α
0
p
+ 23α
≤
1 − cp0 γn n
.
3α−β
n
(n + 1)
(n + 1)β
By definition of n0p0 ,
h
i
E kVn+1 − Γm k4F ≤
Cβ
,
(n + 1)β
(27)
which concludes the induction and the proof.
5.3
Proof of Theorem 3.5
In order to prove Theorem 3.5, we first recall the following Lemma.
Lemma 5.5 (Godichon-Baggioni (2016)). Let Y1 , ..., Yn be random variables taking values
in a normed vector space such that for all positive constant q and for all k ≥ 1, E [kYk kq ] <
∞. Then, for all real numbers a1 , ..., an and for all integer p, we have
" n
!p
p#
n
X
X
p p1
E
(28)
ak Yk
≤
|ak | (E [kYk k ])
k=1
k=1
We can now prove Theorem 3.5. Let us rewrite decomposition (34) as follows
∇2m G (Γm ) (Vn − Γm ) =
Tn Tn+1
−
+ ξn+1 − rn − rn0 − δn ,
γn
γn
(29)
with Tn := Vn −Γm . As in Pelletier (2000), we sum these equalities, apply Abel’s transform
and divide by n to get
∇2m G (Γm )
V n − Γm
1
=
n
n
T1 Tn+1 X
−
+
Tk
γ1
γn+1
k=2
1
1
−
γk
γk−1
−
n
X
k=1
δk −
n
X
k=1
rk −
n
X
rk0
k=1
We now bound the quadratic
mean of each term at the right-hand side of previous equality.
2
T1
1
First, we have n2 E γ1
= o n1 . Applying Theorem 3.3,
F
1
E
n2
"
Tn+1
γn
2
F
#
1 C 0 c−2
γ
≤ 2 −α = o
n n
1
.
n
−1
Moreover, since γk−1 − γk−1
≤ 2αcγ−1 k α−1 , the application of Lemma 5.5 and Theorem 3.3
24
+
n
X
k=1
!
ξk+1
.
gives
1
E
n2
n
X
2
γk−1
−
−1
γk−1
Tk
k=2
F
n
X
≤ 1
n2
γk−1
−
−1
γk−1
!
r h
i 2
2
E kTk kF
k=2
1
0
≤ 2 4α2 c−2
γ C
n
1
=O
n2−α
1
=o
,
n
n
X
k=2
1
!2
k 1−α/2
since α < 1. In the same way, since kδn kF ≤ 6C kTn k2F , applying Lemma 5.5 and Theorem 3.3 with β > 1,
!
2
n
n r h
i 2
X
X
1
1
2
E
E kδk kF
δk ≤ 2
n2
n
k=1
k=1
F
!
n r h
i 2
36C 2 X
≤
E kTk k4F
n2
k=1
!2
n
36C 2 Cβ X 1
≤
n2
k β/2
k=1
1
=O
nβ
1
=o
,
n
√
p
C + C kΓm kF . Since krn kF ≤ D kmn − mk, and since there
Moreover, let D := 12
h
i
is a positive constant C 00 such that for all n ≥ 1, E kmn − mk2 ≤ C 00 n−1 ,
n
1 X
E
rk
n2
k=1
2
F
n
X
≤ 1
n2
≤
≤
!
r h
i 2
2
E krk kF
k=1
D2
n2
n
X
!
r h
i
2
E kmn − mk
k=1
D2 C 00
n2
1
=O
.
n
n
X
1
1/2
k
k=1
!2
Since krn0 kF ≤ C0 kmn − mk kVn − Γm k2F with C0 := 12 C
25
p
kΓm kF + C 3/4 , Cauchy-
Schwarz’s inequality and Lemma 5.5 give
!
2
n r h
n
i 2
X
X
1
1
2
0
E
E krn0 kF
rn ≤ 2
n2
n
k=1
k=1
F
!
r h
i 2
E kmn − mk2 kVn − Γm k2F
C2
≤ 20
n
n
X
C2
≤ 20
n
n h
i 1 h
i 1
X
4
4
E kVn − Γm k4F
E kmn − mk4
k=1
!2
.
k=1
Applying Theorem 4.2 in Godichon-Baggioni (2016) and Theorem 3.3,
!2
p √
2
n
n
X
C02 Cβ K2 X
1
1
0
E
rn ≤
n2
n2
k β/4+1/2
k=1
k=1
F
1
=O
n1+β/2
1
=o
,
n
h
i
since β > 0. Finally, one can easily check that E kξn+1 k2F ≤ 1, and since (ξn ) is a
sequence of martingale differences adapted to the filtration (Fn ),
!
2
n
n
n
n
h
i
X
X
X
X
1
1
E
E kξk+1 k2F + 2
ξk+1 = 2
E [hξk+1 , ξk0 +1 iF ]
n2
n
0 =k+1
k=1
k=1
k=1
k
F
!
n
n
n
iE i
i
hD
h
X
X
1 X h
E kξk+1 k2F + 2
E ξk+1 , E ξk0 +1 Fk0
= 2
n
F
0
=
1
n2
k=1
n
X
k=1 k =k+1
h
i
E kξk+1 k2F
k=1
1
≤ .
n
Thus, there is a positive constant K such that for all n ≥ 1,
h
2i K
E ∇2m G (Γm ) V n − Γm F ≤ .
n
Let λmin be the smallest eigenvalue of ∇2m G (Γm ). We have, with Proposition B.1 in the
supplementary file, that λmin > 0 and the announced result is proven,
h
i
K
2
E V n − Γm F ≤ 2
.
λmin n
6
Concluding remarks
The simulation study and the illustration on real data indicate that performing robust
principal components analysis via the median covariation matrix, which can bring new
26
information compared to classical PCA, is an interesting alternative to more classical robust
principal components analysis techniques. The use of recursive algorithms permits to
perform robust PCA on very large datasets, in which outlying observations may be hard
to detect. Another interest of the use of such sequential algorithms is that estimation of the
median covariation matrix as well as the principal components can be performed online
with automatic update at each new observation and without being obliged to store all
the data in memory. A simple modification of the averaged stochastic gradient algorithm
is proposed that ensures non negativeness of the estimated covariation matrices. This
modified algorithms has better performances on our simulated data.
A deeper study of the asymptotic behaviour of the recursive algorithms would certainly
deserve further investigations. Proving the asymptotic normality and obtaining the limiting
variance of the sequence of estimators V n when m is unknown would be of great interest.
This is a challenging issue that is beyond the scope of the paper and would require to
study the joint weak convergence of the two simultaneous recursive averaged estimators of
m and Γm .
The use of the MCM could be interesting to robustify the estimation in many different
statistical models, particularly with functional data. For example, it could be employed
as an alternative to robust functional projection pursuit in robust functional time series
prediction or for robust estimation in functional linear regression, with the introduction of
the median cross-covariation matrix.
Acknowledgements. We thank the company Médiamétrie for allowing us to illustrate
our methodologies with their data. We also thank Dr. Peggy Cénac for a careful reading
of the proofs.
A
Estimating the median covariation matrix with Weiszfeld’s
algorithm
Suppose we have a fixed size sample X1 , . . . , Xn and we want to estimate the geometric
median.
The iterative Weiszfeld’s algorithm relies on the fact that the solution m∗n of the following optimization problem
n
X
min
kXi − µk
µ∈H
satisfies, when
m∗n
i=1
6= Xi , for all i = 1, . . . , n
m∗n =
n
X
wi (m∗n ) Xi
i=1
where the weights wi (x) are defined by
wi (x) =
kXi − xk−1
.
n
X
−1
kXj − xk
j=1
27
Weiszfeld’s algorithm is based on the following iterative scheme. Consider first a pilot
(e+1)
estimator m
b (0) of m. At step (e), a new approximation m
bn
to m is given by
m
b (e+1)
n
=
n
X
wi m
b (e)
Xi .
n
(30)
i=1
(e+1)
(e)
The iterative procedure is stopped when m
bn
−m
b n ≤ , for some precision known
in advance. The final value of the algorithm is denoted by m
b n.
b (e) has been calculated at
The estimator of the MCM is computed similarly. Suppose Γ
b (e+1) to Γm is defined by
step (e), then at step (e + 1), the new approximation Γ
b (e+1)
Γ
=
n
n
X
b (e) (Xi − m
b n )(Xi − m
b n )T .
Wi Γ
(31)
i=1
b (e+1) − Γ
b (e)
The procedure is stopped when Γ
≤ , for some precision fixed in advance.
F
Note that by construction, this algorithm leads to an estimated median covariation
matrix that is always non negative.
B
Convexity results
In this section, we first give and recall some convexity properties of functional Gh . The
following one gives some information on the spectrum of the Hessian of G.
Proposition B.1. Under assumptions 1-3(b), for all h ∈ H and V ∈ S(H), S(H) admits
an orthonormal basis composed of eigenvectors of ∇2h G(V ). Let us denote by {λh,V,i , i ∈ N}
the set of eigenvalues of ∇2h G(V ). For all i ∈ N,
0 ≤ λh,V,i ≤ C.
Moreover, there is a positive constant cm such that for all i ∈ N,
0 < cm ≤ λm,Γm ,i ≤ C.
Finally, by continuity, there are positive constants , 0 such that for all h ∈ B (m, ) and
V ∈ B (Γm , 0 ), and for all i ∈ N,
1
cm ≤ λh,V,i ≤ C.
2
The proof is very similar to the one in Cardot et al. (2013) and consequently it is not
given here. Furthermore, as in Cardot et al. (2016), it ensures the local strong convexity
as shown in the following corollary.
Corollary B.2. Under assumptions 1-3(b), for all positive constant A, there is a positive
constant cA such that for all V ∈ B (Γm , A) and h ∈ B (m, ),
h∇h G(V ) − ∇h G(Γm ), V − Γm iH ≥ cA kV − Γm k2F .
28
Finally, the following lemma gives an upper bound on the remainder term in the Taylor’s
expansion of the gradient.
Lemma B.3. Under assumptions 1-3(b), for all h ∈ H and V ∈ S(H),
∇Gh (V ) − ∇Gh (Γm ) − ∇2h G (Γm ) (V − Γm )
F
≤ 6C kV − Γm k2F .
(32)
Proof of Lemma B.3. Let δV,h := ∇Gh (V ) − ∇Gh (Γm ) − ∇2h G (Γm ) (V − Γm ), since
R1
∇Gh (V ) − ∇Gh (Γm ) = 0 ∇2h G (Γm + t (V − Γm )) (V − Γm ) dt, we have
Z
kδV,h kF =
1
∇2h G (Γm + t (V − Γm )) ((V − Γm ) dt − ∇2h G (Γm ) (V − Γm )
0
1
Z
F
∇2h G (Γm + t (V − Γm )) ((V − Γm ) − ∇2h G (Γm ) (V − Γm )
≤
0
F
dt.
As in the proof of Lemma 5.1 in Cardot et al. (2016), under assumptions 1-3(b), one can
check that for all h ∈ H, and t ∈ [0, 1],
∇2h G (Γm + t (V − Γm )) ((V − Γm ) − ∇2h G (Γm ) (V − Γm )
F
≤ 6C kV − Γm k2F ,
which concludes the proof.
C
Decompositions of the Robbins-Monro algorithm and proof
of Lemma 5.1
Let us recall that the Robbins-Monro algorithm is defined recursively by
Vn+1 = Vn + γn
(Xn+1 − mn ) (Xn+1 − mn )T − Vn
(Xn+1 − mn ) (Xn+1 − mn )T − Vn
F
= Vn − γn Un+1 ,
T
(Xn+1 −mn )(Xn+1 −mn ) −Vn
. Let us remark that ξn+1 := ∇mn G(Vn ) −
k(Xn+1 −mn )(Xn+1 −mn )T −Vn kF
Un+1 , (ξn ) is a sequence of martingale differences adapted to the filtration (Fn ) and the
algorithm can be written as follows
with Un+1 := −
Vn+1 = Vn − γn (∇Gmn (Vn ) − ∇Gmn (Γm )) + γn ξn+1 − γn rn ,
(33)
with rn := ∇Gmn (Γm ) − ∇Gm (Γm ). Finally, let is consider the following linearization of
the gradient,
Vn+1 − Γm = IS(H) − γn ∇2m G(Γm ) (Vn − Γm ) + γn ξn+1 − γn rn − γn rn0 − γn δn ,
with
rn0 := ∇2mn G (Γm ) − ∇2m G (Γm ) (Vn − Γm ) ,
δn := ∇Gmn (Vn ) − ∇Gmn (Γm ) − ∇2mn G (Γm ) (Vn − Γm ) .
29
(34)
Proof of Lemma 5.1. The bound of kδn k is a corollary of Lemma B.3.
Bounding krn k
Let us recall that for all h ∈ H, Y (h) := (X − h) (X − h)T . We now define for all
h ∈ H the random function ϕh : [0, 1] −→ S(H) defined for all t ∈ [0, 1] by
Y (m + th) − Γm
.
kY (m + th) − Γm kF
i
h
Note that rn = E ϕmn −m (0) − ϕmn −m (1) Fn . Thus, by dominated convergence,
ϕh (t) :=
krn kF ≤ sup E
h
0
ϕm
(t)
n −m
t∈[0,1]
F
i
Fn .
Moreover, one can check that for all h ∈ H,
h (X − m − th)T
(X − m − th) hT
−
kY (m + th) − Γm kF
kY (m + th) − Γm kF
D
E
Y (m + th) − Γm
+ Y (m + th) − Γm , h (X − m − th)T
F kY (m + th) − Γm k3
F
Y (m + th) − Γm
T
.
+ Y (m + th) − Γm , (X − m − th) h F
kY (m + th) − Γm k3F
ϕ0h (t) = −
We now bound each term on the right-hand side of previous equality. First, applying
Cauchy-Schwarz’s inequality and using the fact that for all h, h0 ∈ H, hh0T F = khk kh0 k,
E
h (X − m − th)T
F
kY (m + th) − Γm kF
Thus, since E
h
kX − m − thk
kY (m + th) − Γm kF
#
" p
kY (m + th)kF
≤ khk E
kY (m + th) − Γm kF
"
#
"
#!
p
kΓm kF
1
≤ khk E
+E p
.
kY (m + th) − Γm kF
kY (m + th) − Γm kF
1
kY (m+th)−Γm kF
E
≤ khk E
i
≤ C,
h (X − m − th)T
F
kY (m + th) − Γm kF
q
√
≤ khk C kΓm k + C .
F
(35)
In the same way,
"
(X − m − th) hT F
E
kY (m + th) − Γm kF
#
q
√
≤ khk C kΓm kF + C .
30
(36)
Applying Cauchy-Schwarz’s inequality,
"
E
D
Y (m + th) − Γm , h (X − m − th)T
Thus, since E
h
1
kY (m+th)−Γm kF
i
kY (m + th) − Γm kF
E
F
#
h (X − m − th)T
F
kY (m + th) − Γm kF
kX − m − thk
≤ khk E
kY (m + th) − Γm kF
" p
#
kY (m + th)kF
≤ khk E
.
kY (m + th) − Γm kF
≤ E
kY (m + th) − Γm k3F
≤ C, and since for all positive constants a, b,
√
a+b ≤
√
a+
√
b,
" p
#
"
#
"
#!
p
kY (m + th)kF
kΓm kF
1
khk E
≤ khk E
+E p
kY (m + th) − Γm kF
kY (m + th) − Γm kF
kY (m + th) − Γm kF
q
√
≤ khk C kΓm kF + C .
Finally,
"
D
E
Y (m + th) − Γm , h (X − m − th)T
E
kY (m + th) − Γm kF
F
#
kY (m + th) − Γm k3F
q
√
≤ khk C kΓm kF + C ,
(37)
"
E
Y (m + th) − Γm , (X − m − th) hT
kY (m + th) − Γm kF
F
kY (m + th) − Γm k3F
#
q
√
≤ khk C kΓm kF + C .
(38)
Applying inequalities (35) to (38) with h = mn − m, the announced result is proven,
q
√
krn kF ≤ 4
C + C kΓm kF kmn − mk .
Bounding krn0 k
For all h ∈ H and V ∈ S(H), we define the random function ϕh,V : [0, 1] −→ S(H)
such that for all t ∈ [0, 1],
!
1
(Y (m + th) − Γm ) ⊗F (Y (m + th) − Γm )
IS(H) −
ϕh,V (t) :=
(V ) .
kY (m + th) − Γm kF
kY (m + th) − Γm k2F
h
i
Note that rn0 = E ϕmn −m,Vn −Γm (1) − ϕmn −m,Vn −Γm (0) Fn . By dominated convergence,
rn0
F
≤ sup E
t∈[0,1]
h
ϕ0mn −m,Vn −Γm (t)
F
i
Fn .
Moreover, as for the bound of krn k, one can check, with an application of Cauchy-Schwarz’s
31
inequality, that for all h ∈ H, V ∈ S(H), and t ∈ [0, 1],
ϕ0h,V (t) ≤ 6
+6
kY (m + th) − Γm kF hT (X − m − th)
kY (m + th) − Γm kF h(X − m − th)T
kY (m + th) −
≤ 12
F
kV kF
F
k(Y (m + th) − Γm ) ⊗F (Y (m + th) − Γm ) (V )kF
kY (m + th) − Γm k3F
h(X − m − th)T
F
kY (m + th) − Γm k2F
Γm k5F
kV kF .
Finally,
"
E
h(X − m − th)T
kY (m + th) −
F
2
Γm kF
#
kV kF
"
#
khk kX − m − thk
≤E
kV kF
kY (m + th) − Γm k2F
"
#
p
kΓm kF
≤ khk kV kF E
kY (m + th) − Γm k2F
#
"
1
+ khk kV kF E
3/2
kY (m + th) − Γm kF
q
≤ C kΓm kF + C 3/4 khk kV kF .
(39)
Then the announced result follows from an application of inequality (39) with h = mn − m
and V = Vn − Γm ,
q
3/4
0
kmn − mk kVn − Γm kF .
rn ≤ 12 C kΓm kF + C
D
Proofs of Lemma 5.2, 5.3 and 5.4
Proof of Lemma 5.2. Using decomposition (33),
kVn+1 − Γm k2F = kVn − Γm k2F − 2γn hVn − Γm , ∇Gmn (Vn ) − ∇Gmn (Γm )iF
+ γn2 k∇Gmn (Vn ) − ∇Gmn (Γm )k2F
+ γn2 kξn+1 k2F + 2γn hVn − Γm − γn (∇Gmn (Vn ) − ∇Gmn (Γm )) , ξn+1 iF
+ γn2 krn k2F − 2γn hrn , Vn − Γm iF − 2γn2 hrn , ξn+1 − ∇Gmn (Vn ) + ∇Gmn (Γm )iF .
Note that for all h ∈ H and V ∈ S(H) we have k∇Gh (V )kF ≤ 1. Moreover, krn kF ≤ 2 and
kξn+1 kF ≤ 2. Since for all h ∈ H, Gh is a convex function, we get with Cauchy-Schwarz’s
inequality,
kVn+1 − Γm k2F a ≤ kVn − Γm k2F + 36γn2 + 2γn hξn+1 , Vn − Γm iF − 2γn hrn , Vn − Γm iF .
(40)
√
p
Let C 0 := 4
C + C kΓm kF , let us recall that krn kF ≤ C 0 kmn − mk. We now prove
by induction
that for all
h
i integer p ≥ 1, there is a positive constant Mp such that for all
2p
n ≥ 1, E kVn − Γm kF ≤ Mp .
32
The case p = 1 has been studied in the proof of Theorem 3.2. Let p ≥ 2 and suppose
from now that for all k ≤ p − 1, there is a positive constant Mk such that for all n ≥ 1,
h
i
E kVn − Γm k2k
≤ Mk .
F
h
i
Bounding E kVn − Γm k2p
F .
Let us apply inequality (40), for all p ≥ 2 and use the fact that (ξn ) is a sequence of
martingales differences adapted to the filtration (Fn ),
h
i
h
p i
2
2
E kVn+1 − Γm k2p
≤
E
kV
−
Γ
k
+
36γ
+
2γ
kr
k
kV
−
Γ
k
n
m F
n
n F
n
m F
n
F
p
p−k
X p
k
2
2
E (2γn hVn − Γm , ξn+1 iF ) kVn − Γm kF + 36γn + 2γn krn kF kVn − Γm kF
.
+
k
k=2
(41)
Let us denote by (∗) the second term on the right-hand side of inequality (41). Applying
Cauchy-Schwarz’s inequality and since kξn+1 kF ≤ 2,
n
p−k
X
p
k
2
2
(∗) =
E (2γn hVn − Γm , ξn+1 i) kVn − Γm kF + 36γn + 2γn krn kF kVn − Γm kF
k
k=2
p
p−k
X
p 2k k
k
2
2
≤
2 γn E kVn − Γm kF kVn − Γm kF + 36γn + 2γn krn kF kVn − Γm kF
.
k
k=2
With the help of Lemma E.1,
(∗) ≤
p
X
p
h
i
i X
h
22k 3p−k−1 36p−k γn2p−k E kVn − Γm kkF
22k 3p−k−1 γnk E kVn − Γm kF2p−k +
k=2
k=2
+
p
X
i
h
p
p−k
p+k p−k−1 p
2 3
γn E krn kF kVn − Γm kF .
k=2
Applying Cauchy-Schwarz’s inequality,
p
X
p
i X
h
i
h
p+1−k
p−1
2k p−k−1 k
kV
−
Γ
k
=
2
3
γ
E
kV
−
Γ
k
22k 3p−k−1 γnk E kVn − Γm k2p−k
n
m F
n
m F
n
F
k=2
k=2
≤
p
X
22k 3p−k−1 γnk
r h
ir h
i
2(p−1)
2(p+1−k)
E kVn − Γm kF
E kVn − Γm kF
.
k=2
By induction,
p
X
22k 3p−k−1 γnk E
p
h
i X
p
p
2p−k
kVn − Γm kF
≤
22k 3p−k−1 γnk Mp−1 Mp+1−k
k=2
k=2
= O γn2 .
33
(42)
In the same way, applying Cauchy-Schwarz’s inequality and by induction,
p
X
2k p−k−1
2 3
36p−k γn2p−k E
h
kVn −
Γm kkF
i
=
k=2
p
X
h
i
22k 3p−k−1 36p−k γn2p−k E kVn − Γm kF kVn − Γm kk−1
F
k=2
≤
p
X
22k 3p−k−1 36p−k γn2p−k
p p
M1 Mk−1
k=2
= O γn2 ,
(43)
since p ≥ 2. Similarly, since krn kF ≤ 2 and since p ≥ 2, applying Cauchy-Schwarz’s
inequality and by induction,
p
X
2p+k 3p−k−1 γnp E
h
krn kp−k
F
kVn −
Γm kpF
i
≤
k=2
p
X
22p 3p−k−1 γnp E kVn − Γm kpF
k=2
≤
p
X
22p 3p−k−1 γnp
p p
M1 Mp−1
k=2
= O γn2 .
(44)
Finally, applying inequalities (42) to (44), there is a positive constant A01 such that for all
n ≥ 1,
" p
#
p−k
X p
k
2
2
E
(2γn hVn − Γm , ξn+1 iF ) kVn − Γm kF + 36γn + 2γn krn kF kVn − Γm kF
≤ A01 γn2 .
k
k=2
(45)
We now denote by (∗∗) the first term at the right-hand side of inequality (41). With the
help of Lemma E.1 and applying Cauchy-Schwarz’s inequality,
p h
i X
h
i
k
p
2p
E 36γn2 + 2γn hrn , Vn − Γm iF kVn − Γm kF2p−2k
(∗∗) ≤ E kVn − Γm kF +
k
k=1
p
i X p
h
i
h
2p
.
2k−1 E 36k γn2k + 2k γnk krn kkF kVn − Γm kkF kVn − Γm k2p−2k
≤ E kVn − Γm kF +
F
k
k=1
Moreover, let
p
i
X
p k−1 h k 2k
(∗ ∗ ∗) :=
2 E 36 γn + 2k γnk krn kkF kVn − Γm kkF kVn − Γm k2p−2k
F
k
k=1
p
p
h
i X
i
X p
p 2k−1 k h
2p−2k
k−1 k 2k
=
2 36 γn E kVn − Γm kF
+
2
γn E krn kkF kVn − Γm kF2p−k .
k
k
k=1
k=1
By induction,
p
p
i X
X
p k−1 k 2k h
p k−1 k 2k
2p−2k
2 36 γn E kVn − Γm kF
=
2 36 γn Mp−k
k
k
k=1
k=1
= O γn2 .
34
Moreover,
p
p
i
i X
X
p 2k−1 k h
p 2k−1 k h
2p−k
k
2
γn E krn kkF kVn − Γm kF2p−k
=
2
γn E krn kF kVn − Γm kF
k
k
k=2
k=1
h
i
+ 2pγn E krn kF kVn − Γm k2p−1
.
F
Applying Cauchy-Schwarz’s inequality and by induction, since krn kF ≤ 2,
p
p
i
i X
X
p 3k−1 k h
p 2k−1 k h
2p−k
k
2
γn E kVn − Γm kF2p−k
≤
2
γn E krn kF kVn − Γm kF
k
k
k=2
k=2
p
X
p
p 3k−1 k p
2
γn Mp+1−k Mp−1
≤
k
k=2
= O γn2 .
Moreover, applying Theorem 4.2 in Godichon-Baggioni (2016) and Hölder’s inequality,
since krn kF ≤ C 0 kmn − mk,
h
i
h
i
2p−1
0
≤
2C
pγ
E
km
−
mk
kV
−
Γ
k
2pγn E krn kF kVn − Γm k2p−1
n
n
n
m
F
F
i 2p−1
h
i 1 h
2p
2p
≤ 2C 0 pγn E kmn − mk2p
E kVn − Γm k2p
F
1
i 2p−1
Kp2p h
2p
.
≤ 2C pγn 1/2 E kVn − Γm k2p
F
n
0
Finally,
1
1
n
h
i 2p−1
io
Kp2p h
Kp2p
2p
2p
0
2C pγn 1/2 E kVn − Γm kF
≤ 2C 0 pγn 1/2 max 1, E kVn − Γm k2p
F
n
n
1
i
h
Kp2p
.
≤ 2C 0 pγn 1/2 1 + E kVn − Γm k2p
F
n
Thus, there are positive constants A000 , A001 such that
h
i
1
1
00
(∗∗) ≤ 1 + A000 α+1/2 E kVn − Γm k2p
F + A1 α+1/2 .
n
n
(46)
Finally, thanks to inequalities (45) and (46), there are positive constants A00 , A01 such
that
h
h
i
i
1
1
2p
0
0
E kVn+1 − Γm kF ≤ 1 + A0 α+1/2 E kVn − Γm k2p
F + A1 α+1/2
n
n
n
n
n
h
i
Y
X
Y
1
1
1
0
0
≤
1 + A00 α+1/2 E kV1 − Γm k2p
+
1
+
A
0 α+1/2 A1 α+1/2
F
k
j
k
k=1
k=1 j=k+1
h
X
∞
∞
∞
i
Y
Y
1
1
1
2p
0
0
≤
1 + A0 α+1/2 E kV1 − Γm kF +
1 + A0 α+1/2
A01 α+1/2
k
j
k
j=1
k=1
k=1
≤ Mp ,
which concludes the induction and the proof.
35
Proof of Lemma 5.3. Let us define the following linear operators:
αn := IS(H) − γn ∇2m G(Γm ),
βn :=
n
Y
αk =
k=1
n
Y
IS(H) − γk ∇2m G(Γm ) ,
k=1
β0 := IS(H) .
Using decomposition (34) and by induction, for all n ≥ 1,
Vn − Γm = βn−1 (V1 − Γm ) + βn−1 Mn − βn−1 Rn − βn−1 Rn0 − βn−1 ∆n ,
(47)
with
Mn :=
Rn0 :=
n−1
X
γk βk−1 ξk+1 ,
Rn :=
k=1
n−1
X
γk βk−1 rk0 ,
∆n :=
k=1
n−1
X
k=1
n−1
X
γk βk−1 rk ,
γk βk−1 δk .
k=1
We now study the asymptotic behavior of the linear operators βn and βn−1 βk−1 . As in
Cardot et al. (2013), one can check that there are positive constants c0 , c1 such that for all
integers k, n ≥ 1 with k ≤ n − 1,
kβn−1 kop ≤ c0 e−λmin
Pn
k=1
γn
βn−1 βk−1
,
op
≤ c1 e−λmin
Pn
j=k
γj
,
(48)
where k.kop is the usual spectral norm for linear operators. We now bound the quadratic
mean of each term in decomposition (47).
Step 1: the quasi deterministic term βn−1 (V1 − Γm ).
Applying inequality (48), there is a positive constant c00 such that
h
i
h
i
E kβn−1 (V1 − Γm )k2F ≤ kβn−1 k2op E kV1 − Γm k2F
h
i
Pn
≤ c0 e−2λmin k=1 γn E kV1 − Γm k2F
h
i
0 1−α
≤ c0 e−c0 n E kV1 − Γm k2F .
(49)
This term converges exponentially fast to 0.
Step 2: the martingale term βn−1 Mn .
Since (ξn ) is a sequence of martingale differences adapted to the filtration (Fn ),
h
i n−1
X h
2
E kβn−1 Mn kF =
E βn−1 βk−1 γk ξk+1
2
F
i
2
βn−1 βk−1 γk ξk+1 F
i
+2
=
h
E
+2
n−1
X
n−1
X
n−1
X
k=1 k0 =k+1
k=1
=
n−1
X
γk γk 0 E
γk γk 0 E
βn−1 βk−1 ξk+1 , βn−1 βk−1
0 ξk 0 +1
F
k=1 k0 =k+1
k=1
n−1
X
n−1
X
h
E
βn−1 βk−1 γk ξk+1
2
F
i
k=1
36
.
βn−1 βk−1 ξk+1 , βn−1 βk−1
0 E [ξk 0 +1 |Fk 0 ]
F
Moreover, as in Cardot et al. (2016), Lemma E.2 ensures that there is a positive constant
C10 such that for all n ≥ 1,
h
i C0
E kβn−1 Mn k2F ≤ α1 .
(50)
n
Step 3: the first remainder
√ termpβn−1 Rn .
C + C kΓm kF kmn − mk,
Remarking that krn kF ≤ 4
h
n−1
X
i
E kβn−1 Rn k2F ≤ E
!2
γk βn−1 βk−1
op
krk kF
k=1
≤ 16
√
2
n−1
q
X
C + kΓm kF E
γk βn−1 βk−1
!2
op
kmk − mk
.
k=1
Applying Lemma 4.3 and Theorem 4.2 in Godichon-Baggioni (2016),
h
E
kβn−1 Rn k2F
i
≤ 16
√
2
q
C + C kΓm kF
n−1
X
γk
βn−1 βk−1 op
!
r h
i 2
2
E kmk − mk
k=1
≤ 16
√
2
q
C + C kΓm kF K1
n−1
X
γk
βn−1 βk−1 op
k=1
1
!2
.
k 1/2
Applying inequality (48),
h
i
E kβn−1 Rn k2F ≤ 16
√
p 2
C + C Γm K1
n−1
X
γk e−
Pn
γk e −
Pn
j=k
γj
≤ 16
p 2
C + C Γm K1
n
X
!2
k 1/2
k=1
√
1
j=k
k=1
γj
1
k 1/2
!2
.
Splitting the sum into two parts and applying Lemma E.2, we have
2
2
E(n/2)
q
h
i
Pn
X
√
1
E kβn−1 Rn k2F ≤ 32
C + C kΓm kF K1
γk e− j=k γj 1/2
k
k=1
2
2
n
q
Pn
X
√
1
C + C kΓm kF K1
γk e− j=k γj 1/2
+ 32
k
k=E(n/2)+1
1
=O
.
n
Thus, there is a positive constant C20 such that for all n ≥ 1,
h
i C0
E kβn−1 Rn k2F ≤ 2 .
n
(51)
Step 4: the second remainder term βn−1 Rn0 .
Let us recall that for all n ≥ 1, krn0 kF ≤ 12D kmn − mk kVn − Γm kF with D :=
37
C
p
kΓm kF + C 3/4 . Thus,
n−1
h
i
X
2
E βn−1 Rn0 F ≤ E
γk βn−1 βk−1
!2
op
rk0
F
k=1
n−1
X
≤ 144D2 E
!2
γk βn−1 βk−1
op
kmk − mk kVk − Γm kF
.
k=1
Applying Lemma 4.3 in Godichon-Baggioni (2016),
h
E
2
βn−1 Rn0 F
i
≤ 144D
2
n−1
X
γk
βn−1 βk−1 op
!
r h
i 2
2
2
E kmk − mk kVk − Γm kF
.
k=1
h
i
Thanks to Lemma 5.2, there is a positive constant M2 such that for all n ≥ 1, E kVn − Γm k4F ≤ M2 .
Thus, applying Cauchy-Schwarz’s inequality and Theorem 4.2 in Godichon-Baggioni (2016),
h
E
βn−1 Rn0
2
F
i
≤ 144D2
n−1
X
γk βn−1 βk−1
h
i 1
i 1 h
4
4
4
4
E
km
E
kV
−
Γ
k
−
mk
m F
k
k
op
!2
k=1
p
≤ 144D M2 K2
2
n−1
X
γk
βn−1 βk−1 op
k=1
!2
1
k 1/2
.
As in step 3, splitting the sum into two parts, one can check that there is a positive constant
C100 such that for all n ≥ 1,
h
i C 00
2
E βn−1 Rn0 F ≤ 1 .
(52)
n
Step 5: the third remainder term: βn−1 ∆n
Since kδn kF ≤ 6C kVn − Γm k2F , applying Lemma 4.3 in Godichon-Baggioni (2016),
!2
n−1
h
i
X
E kβn−1 ∆n k2F ≤ E
γk βn−1 βk−1 op kδk kF
k=1
≤ 36C 2 E
n−1
X
!2
γk βn−1 βk−1
op
kVk − Γm k2F
k=1
≤ 36C 2
n−1
X
γk βn−1 βk−1
!
r h
i 2
E kVk − Γm k4F
.
op
k=1
h
Γm k4F
Thanks to Lemma 5.2, there is a positive constant M2 such that for all n ≥ 1, E kVn −
Thus, splitting the sum into two parts and applying inequalities (48) and Lemma E.2, there
38
i
≤ M2 .
are positive constant c00 , C20 such that for all n ≥ 1,
2
E(n/2)
h
i
Pn
X
E kβn−1 ∆n k2 ≤ 72C 2 M22
γk e− j=k γj
F
k=1
+ 72C 2
sup
E(n/2)+1≤k≤n−1
≤ C20
sup
E(n/2)+1≤k≤n−1
n h
io
E kVk − Γm k4F
n
X
2
γk e−
Pn
j=k
γj
k=E(n/2)+1
n h
io
0 1−α
E kVk − Γm k4F
+ O e−2c0 n
.
Thus, there is a positive constant C00 such that for all n ≥ 1,
h
i
n h
io
0 1−α
+ C20
sup
E kβn−1 ∆n k2F ≤ C00 e−2c0 n
E kVk − Γm k4F .
(53)
E(n/2)+1≤k≤n−1
Conclusion:
Applying Lemma E.1 and decomposition (47), for all n ≥ 1,
h
i
h
i
h
i
h
i
2
2
2
2
E kVn − Γm kF ≤ 5E kβn−1 (V1 − Γm )kF + 5E kβn−1 Mn kF + 5E kβn−1 Rn kF
h
i
h
i
2
+ 5E βn−1 Rn0 F + 5E kβn−1 ∆n k2F .
Applying inequalities (49) to (53), there are positive constants C1 , C10 , C2 , C3 such that for
all n ≥ 1,
h
i
h
i
C2
0 1−α
sup
E kVk − Γm k4F .
E kVn − Γm k2 ≤ C1 e−C1 n
+ α + C3
n
E(n/2)+1≤k≤n−1
Proof of Lemma 5.4. Let us define Wn := Vn − Γm − γn (∇Gmn (Vn ) − ∇Gmn (Γm )) and
use decomposition (33),
kVn+1 − Γm k2F = kWn k2F + γn2 kξn+1 k2F + γn2 krn k2F + 2γn hξn+1 , Vn − Γm iF + 2γn2 hξn+1 , ∇Gmn (Vn )iF
− 2γn2 hrn , ∇Gmn (Vn ) − ∇Gmn (Γm )iF − 2γn hrn , Vn − Γm iF .
Since kξn+1 kF ≤ 2, krn kF ≤ 2 and the fact that for all h ∈ H, V ∈ S(H), ∇h G(V ) ≤ 1,
we get with an application of Cauchy-Schwarz’s inequality
kVn+1 − Γm k2F ≤ kWn k2F + 2γn hξn+1 , Vn − Γm iF + 2γn krn kF kVn − Γm kF + 20γn2 .
Thus, since (ξn ) is a sequence of martingale differences adapted to the filtration (Fn ), and
since kWn k2F ≤ 1 + C 2 c2γ kVn − Γm k2F (this inequality follows from Proposition B.1 and
from the fact that for all h ∈ H, Gh is a convex application),
h
i
h
i
h
i
E kVn+1 − Γm k4F ≤ E kWn k4F + 2γn E krn kF kWn k2F kVn − Γm kF
h
i
+ 40 1 + C 2 c2γ γn2 E kVn − Γm k2F
h
i
h
i
+ 4γn2 E hξn+1 , Vn − Γm i2F + 400γn4 + 40γn3 E krn kF kVn − Γm k2F
i
h
+ 4γn2 E krn k2F kVn − Γm k2F .
39
Since kξn+1 kF ≤ 2 and krn kF ≤ 2, applying Cauchy-Schwarz’s inequality, there are positive
constants C10 , C20 such that for all n ≥ 1,
h
i
h
i
h
i C0
i
C20 h
1
E kVn+1 − Γm k4F ≤ E kWn k4F +2γn E krn kF kWn k2F kVn − Γm kF + 3α
+ 2α
E kVn − Γm k2F .
n
n
(54)
We now bound the two first terms at the right-hand side of inequality (54).
h
i
Step 1: bounding E kWn k4F .
R1
Since ∇Gmn (Vn )−∇Gmn (Γm ) = 0 ∇2mn G (Γm + t (Vn − Γm )) (Vn − Γm ) dt, applying Proposition B.1, one can check that
kWn k2 = kVn − Γm k2F − 2γn hVn − Γm , ∇Gmn (Vn ) − ∇Gmn (Γm )iH + γn2 k∇Gmn (Vn ) − ∇Gmn (Γm )k2F
≤ 1 + C 2 γn2 kVn − Γm k2F − 2γn hVn − Γm , ∇Gmn (Vn ) − ∇Gmn (Γm )iH .
Since for all h ∈ H, Gh is a convex application, kWn k2F ≤ 1 + c2γ C 2 kVn − Γm k2F . Let p0
be a positive integer. We now introduce the sequence of events An,p0 n∈N defined for all
n ≥ 1 by
n
o
1−α
An,p0 := ω ∈ Ω, kVn (ω) − Γm kF ≤ n p0 , and kmn (ω) − mk ≤ ,
(55)
with defined in Proposition B.1. For the sake of simplicity, we consider that 0 defined in
Proposition B.1 verifies 0 ≤ 1. Applying Proposition B.1, let
Bn : = h∇Gmn (Vn ) − ∇Gmn (Γm ), Vn − Γm iF 1An,p0 1{kVn −Γm k ≤0 }
F
Z 1
=
∇2mn G (Γm + t (Vn − Γm )) (Vn − Γm ) , Vn − Γm F 1{kVn −Γm k
F ≤
0
0
} 1An,p0 dt
1
≥ cm kVn − Γm k2F 1{kVn −Γm k ≤0 } 1An,p0 .
F
2
(56)
In the same way, since Gmn is convex, let
Bn0 : = h∇Gmn (Vn ) − ∇Gmn (Γm ), Vn − Γm iF 1An,p0 1{kVn −Γm k >0 }
F
Z 1
=
∇2mn (Γm + t (Vn − Γm )) (Vn − Γm ) , Vn − Γm 1{kVn −Γm k
0
F >
0
Z
≥
0
kVn −Γm kF
0
} 1An,p0 dt
∇2mn (Γm + t (Vn − Γm )) (Vn − Γm ) , Vn − Γm 1{kVn −Γm k >0 } 1An,p0 dt
F
Applying Proposition B.1,
Bn0
Z
≥
0
0
kVn −Γm kF
1
cm kVn − Γm k2F 1{kVn −Γm k >0 } 1An,p0 dt
F
2
0 cm
kVn − Γm k2F 1{kVn −Γm k >0 } 1An,p0
F
2 kVn − Γm kF
0
cm − 1−α
≥
n p0 kVn − Γm k2F 1{kVn −Γm k >0 } 1An,p0 .
F
2
≥
40
(57)
0
− 1−α
There is a rank n0p0 such that for all n ≥ n0p0 , we have c2m n p ≤ 12 cm . Thus, applying
inequalities (56) and (57), for all n ≥ n0p0 ,
0 cm
− 1−α
2
0
p
kVn − Γm k2F 1An,p0 .
γn n
kWn kF 1An,p0 ≤ 1 −
2
Thus, there are a positive constant cp0 and a rank np0 such that for all n ≥ np0 ,
2 h
h
i
i
0 cm
− 1−α
4
0
p
E kWn kF 1An,p0 ≤ 1 −
γn n
E kVn − Γm k4F 1An,p0
2
i
h
− 1−α
(58)
≤ 1 − 2cp0 γn n p0 E kVn − Γm k4F .
i
h
Now, we must get an upper bound for E kWn k4F 1Acn,p0 . Since kWn k2F ≤ 1 + c2γ C 2 kVn − Γm k2F
and since there is a positive constant c0 such that for all n ≥ 1,
kVn − Γm kF ≤ kV1 − Γm kF +
n
X
γk ≤ c0 n1−α
k=1
we have
h
i
i
h
4
4
2 2 2
c
c
E kVn − Γm kF 1An,p0
E kWn kF 1An,p0 ≤ 1 + cγ C
2
≤ 1 + c2γ C 2 c40 n4−4α P Acn,p0
h
1−α i
2
≤ 1 + c2γ C 2 c40 n4−4α P [kmn − mk ≥ ] + P kVn − Γm kF ≥ n p0
.
Applying Markov’s inequality, Theorem 4.2 in Godichon-Baggioni (2016) and Lemma 5.2,
i
i
h
h
2q
2p00
h
i
−
mk
E
kV
−
Γ
k
E
km
n
n
m
F
2
E kWn k4F 1Acn,p0 ≤ 1 + c2γ C 2 c40 n4−4α
+
2q 1−α
2p00
n p0
2
2
Kp00
00
4−4α−2q 1−α
p0 .
≤ 2p00 1 + c2γ C 2 c40 n4−4α−p + 1 + c2γ C 2 c40 Mq n
4−α
Taking p00 ≥ 4 − α and q ≥ p0 2(1−α)
,
h
i
1
4
.
(59)
E kWn kF 1Acn,p0 = O
n3α
Thus, applying inequalities (58) and (59), there are positive constants cp0 , C1,p0 and a rank
np0 such that for all n ≥ np0 ,
h
h
i
i C 0
− 1−α
1,p
E kWn k4F ≤ 1 − 2cp0 γn n p0 E kVn − Γm k4F + 3α .
(60)
n
h
i
Step 2: bounding 2γn E krn kF kWn k2F kVn − Γm kF .
Since kWn k2F ≤ 1 + c2γ C 2 kVn − Γm k2F , applying Lemma E.1, let
h
i
Dn : = 2γn E krn kF kWn k2F kVn − Γm kF
h
i
≤ 2 1 + c2γ C 2 γn E krn kF kVn − Γm k3F
h
i 1
h
i
1−α
2
2
− 1−α
1 + c2γ C 2 γn n p0 E krn k2F kVn − Γm k2F + cp0 γn n p0 E kVn − Γm k4F
≤
cp0
2
h
i
h
i
1−α
1−α
2
3 p0
− p0
4
4
2 2 4
0
≤ 2 1 + cγ C
γn n
E krn kF + cp γn n
E kVn − Γm kF .
cp0
41
√
p
Since krn kF ≤
C + C kΓm kF kmn − mkF and applying Theorem 4.2 in GodichonBaggioni (2016),
4
q
h
h
i
i
√
2
− 1−α
3 1−α
2 2 4
Dn ≤ 2 1 + cγ C
C + C kΓm kF γn n p0 E kmn − mk4 + cp0 γn n p0 E kVn − Γm k4F
cp0
4
q
h
i
√
2
− 1−α
3 1−α 1
2 2 4
≤ 2 K2 1 + cγ C
C + C kΓm kF γn n p0 2 + cp0 γn n p0 E kVn − Γm k4F
n
cp0
h
i
1
− 1−α
.
(61)
= cp0 γn n p0 E kVn − Γm k4F + O
0
2+α−3(1−α)/p
n
Step 3: Conclusion.
Applying inequalities (54), (60) and (61), there are a rank np0 and positive constants
cp0 , C1,p0 , C2,p0 , C3,p0 such that for all n ≥ np0 ,
i
i
h
i C 0 C 0 h
h
C3,p0
− 1−α
1,p
2,p
.
E kVn+1 − Γm k4F ≤ 1 − cp0 γn n p0 E kVn − Γm k4F + 3α + 2α E kVn − Γm k2F +
2+α−3 1−α
n
n
p0
n
E
Some technical inequalities
First, the following lemma recalls some well-known inequalities.
Lemma E.1. Let a, b, c be positive constants. Then,
a2 b2 c
+
,
2c
2
c a2
a≤ + .
2 2c
Moreover, let k, p be positive integers and a1 , ..., ap be positive constants. Then,
k
p
p
X
X
aj ≤ pk−1
akj .
ab ≤
j=1
j=1
The following lemma gives the asymptotic behavior for some specific sequences of descent steps.
Lemma E.2. Let α, β be non-negative constants such that 0 < α < 1, and (un ), (vn ) be
two sequences defined for all n ≥ 1 by
cu
cv
vn := β ,
un := α ,
n
n
with cu , cv > 0. Thus, there is a positive constant c0 such that for all n ≥ 1,
E(n/2)
X
e−
Pn
j=k
uj
1−α
uk vk = O e−c0 n
,
(62)
k=1
n
X
e−
Pn
j=k
uj
k=E(n/2)+1
where E(.) is the integer part function.
42
uk vk = O (vn ) ,
(63)
Proof of Lemma E.2. We first prove inequality (62). For all n ≥ 1,
E(n/2)
E(n/2)
e−
X
Pn
j=k
uj
uk vk = cu cv
k=1
X
e−
Pn
j=k
uj
k=1
E(n/2)
≤ cu cv
X
−cu
e
1
k α+β
Pn
1
j=k j α
.
k=1
Moreover, for all k ≤ E(n/2),
cu
n
X
n 1
1
≥ cu
α
j
2 nα
j=k
≥
cu 1−α
n
.
2
Thus,
E(n/2)
X
e−
Pn
j=k
uj
uk vk ≤ cu cv ne−
cu 1−α
n
2
.
k=1
We now prove inequality (63). With the help of an integral test for convergence,
n
X
n
X
1
uj = cu
jα
j=k
j=k
Z n+1
1
≥ cu
dt
tα
k
cu
(n + 1)1−α − k −α .
≥
1−α
Thus,
n
X
e−
Pn
j=k
uj
1−α
uk vk ≤ cu cv e−(n+1)
k=E(n/2)+1
n
X
ek
1−α
k −α−β
k=E(n/2)+1
With the help of an integral test for convergence, there is a rank nu,v (for sake of simplicity,
we consider that nu,v = 1) such that for all n ≥ nu,v ,
Z n+1
n
X
1−α
k1−α −α−β
e
k
≤
et t−α−β dt
E(n/2)+1
k=E(n/2)+1
1 h t1−α −β in
≤
e
t
+β
1−α
E(n/2)+1
Z n+1
(n+1)1−α (n+1)−β
=e
+o
Z
n
1−α
et
E(n/2)+1
et
1−α
t−α−β dt ,
E(n/2)+1
since α < 1. Thus,
n
X
ek
1−α
1−α −β
k −α−β = O en n
.
k=E(n/2)+1
As a conclusion, we have
n
X
e−
Pn
j=k
uj
1−α
1−α
uk vk = O e−(n+1) +n vn
k=E(n/2)+1
= O (vn ) .
43
t−1−β dt
!
References
Bali, J.-L., Boente, G., Tyler, D.-E., and Wang, J.-L. (2011). Robust functional principal
components: a projection-pursuit approach. The Annals of Statistics, 39:2852–2882.
Bosq, D. (2000). Linear processes in function spaces, volume 149 of Lecture Notes in
Statistics. Springer-Verlag, New York. Theory and applications.
Cardot, H., Cénac, P., and Chaouch, M. (2010). Stochastic approximation to the multivariate and the functional median. In Lechevallier, Y. and Saporta, G., editors, Compstat
2010, pages 421–428. Physica Verlag, Springer.
Cardot, H., Cénac, P., and Godichon-Baggioni, A. (2016). Online estimation of the geometric median in Hilbert spaces: non asymptotic confidence balls. The Annals of Statistics
(to appear).
Cardot, H., Cénac, P., and Monnez, J.-M. (2012). A fast and recursive algorithm for
clustering large datasets with k-medians. Computational Statistics and Data Analysis,
56:1434–1449.
Cardot, H., Cénac, P., and Zitt, P.-A. (2013). Efficient and fast estimation of the geometric
median in Hilbert spaces with an averaged stochastic gradient algorithm. Bernoulli,
19:18–43.
Cardot, H. and Degras, D. (2015). Online principal components analysis: which algorithm
to choose ? Technical report, arXiv:1511.03688.
Chakraborty, A. and Chaudhuri, P. (2014). The spatial distribution in infinite dimensional
spaces and related quantiles and depths. The Annals of Statistics, 42:1203–1231.
Chaudhuri, P. (1992). Multivariate location estimation using extension of R-estimates
through U -statistics type approach. Ann. Statist., 20(2):897–916.
Croux, C., Filzmoser, P., and Oliveira, M. (2007). Algorithms for projection-pursuit robust
principal component analysis. Chemometrics and Intelligent Laboratory Systems, 87:218–
225.
Croux, C. and Ruiz-Gazen, A. (2005). High breakdown estimators for principal components: the projection-pursuit approach revisited. J. Multivariate Anal., 95:206–226.
Cupidon, J., Gilliam, D., Eubank, R., and Ruymgaart, F. (2007). The delta method for
analytic functions of random operators with application to functional data. Bernoulli,
13:1179–1194.
44
Dauxois, J., Pousse, A., and Romain, Y. (1982). Asymptotic theory for principal components analysis of a random vector function: some applications to statistical inference.
Journal of Multivariate Analysis, 12:136–154.
Devlin, S., Gnanadesikan, R., and Kettenring, J. (1981). Robust estimation of dispersion
matrices and principal components. J. Amer. Statist. Assoc., 76:354–362.
Duflo, M. (1997). Random iterative models, volume 34 of Applications of Mathematics (New
York). Springer-Verlag, Berlin. Translated from the 1990 French original by Stephen S.
Wilson and revised by the author.
Fritz, H., Filzmoser, P., and Croux, C. (2012). A comparison of algorithms for the multivariate L1 -median. Comput. Stat., 27:393–410.
Gervini, D. (2008). Robust functional estimation using the median and spherical principal
components. Biometrika, 95(3):587–600.
Godichon-Baggioni, A. (2016). Estimating the geometric median in Hilbert spaces with
stochastic gradient algorithms; Lp and almost sure rates of convergence. J. of Multivariate Analysis, 146:209–222.
Huber, P. and Ronchetti, E. (2009). Robust Statistics. John Wiley and Sons, second
edition.
Hubert, M., Rousseeuw, P., and Van Aelst, S. (2008). High-breakdown robust multivariate
methods. Statistical Science, 13:92–119.
Hyndman, R. and Ullah, S. (2007). Robust forecasting of mortality and fertility rates: A
functional data approach. Computational Statistics and Data Analysis, 51:4942–4956.
Jolliffe, I. (2002). Principal Components Analysis. Springer Verlag, New York, second
edition.
Kemperman, J. H. B. (1987). The median of a finite measure on a Banach space. In
Statistical data analysis based on the L1 -norm and related methods (Neuchâtel, 1987),
pages 217–230. North-Holland, Amsterdam.
Kraus, D. and Panaretos, V. M. (2012). Dispersion operators and resistant second-order
functional data analysis. Biometrika, 99:813–832.
Locantore, N., Marron, J., Simpson, D., Tripoli, N., Zhang, J., and Cohen, K. (1999).
Robust principal components for functional data. Test, 8:1–73.
Maronna, R. A., Martin, R. D., and Yohai, V. J. (2006). Robust statistics. Wiley Series in
Probability and Statistics. John Wiley & Sons, Ltd., Chichester. Theory and methods.
Mokkadem, A. and Pelletier, M. (2006). Convergence rate and averaging of nonlinear twotime-scale stochastic approximation algorithms. Ann. Appl. Probab., 16(3):1671–1702.
45
Möttönen, J., Nordhausen, K., and Oja, H. (2010). Asymptotic theory of the spatial
median. In Nonparametrics and Robustness in Modern Statistical Inference and Time
Series Analysis: A Festschrift in honor of Professor Jana Jurec̆ková, volume 7, pages
182–193. IMS Collection.
Pelletier, M. (2000). Asymptotic almost sure efficiency of averaged stochastic algorithms.
SIAM J. Control Optim., 39(1):49–72.
Polyak, B. and Juditsky, A. (1992). Acceleration of stochastic approximation. SIAM J.
Control and Optimization, 30:838–855.
R Development Core Team (2010). R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria. ISBN 3-900051-07-0.
Ramsay, J. O. and Silverman, B. W. (2005). Functional Data Analysis. Springer, New
York, second edition.
Rousseeuw, P. and van Driessen, K. (1999). A fast algorithm for the minimum covariance
determinant estimator. Technometrics, 41:212–223.
Taskinen, S., Koch, I., and Oja, H. (2012). Robustifying principal components analysis
with spatial sign vectors. Statist. and Probability Letters, 82:765–774.
Vardi, Y. and Zhang, C.-H. (2000). The multivariate L1 -median and associated data depth.
Proc. Natl. Acad. Sci. USA, 97(4):1423–1426.
Weiszfeld, E. (1937). On the point for which the sum of the distances to n given points is
minimum. Tohoku Math. J., 43:355–386.
Weng, J., Zhang, Y., and Hwang, W.-S. (2003). Candid covariance-free incremental principal component analysis. IEEE Trans. Pattern Anal. Mach. Intell., 25:1034–1040.
46
| 10math.ST
|
Learning ELM network weights using linear
discriminant analysis
Philip de Chazal1, Jonathan Tapson1 and André van Schaik1
1
The MARCS Institute, University of Western Sydney, Penrith NSW 2751, Australia
{p.dechazal, j.tapson, a.vanschaik}@uws.edu.au
Abstract. We present an alternative to the pseudo-inverse method for determining the hidden to output weight values for Extreme Learning Machines
performing classification tasks. The method is based on linear discriminant
analysis and provides Bayes optimal single point estimates for the weight values.
Keywords: Extreme learning machine, Linear discriminant analysis, Hidden to
output weight optimization, MNIST database
1
Introduction
The Extreme Learning Machine (ELM) is a multi-layer feedforward neural network
that offers fast training and flexible non-linearity for function and classification tasks.
Its principal benefit is that the network parameters are calculated in a single pass during the training process [1]. In its standard form it has an input layer that is fully connected to a hidden layer with non-linear activation functions. The hidden layer is fully
connected to an output layer with linear activation functions. The number of hidden
units is often much greater than the input layer with a fan-out of 5 to 20 hidden units
per input frequently used. A key feature of ELMs is that the weights connecting the
input layer to the hidden layer are set to random values. This simplifies the requirements for training to one of determining the hidden to output unit weights, which can
be achieved in a single pass. By randomly projecting the inputs to a much higher dimensionality, it is possible to find a hyperplane which approximates a desired regression function, or represents a linear separable classification problem [2].
A common way of calculating the hidden to output weights is to use the MoorePenrose pseudo-inverse applied to the hidden layer outputs using labelled training
data. In this paper we present an alternative method for hidden to output weight calculation for networks performing classification tasks. The advantage of our method over
the pseudo-inverse method is that the weights are the best single point estimates from
a Bayesian perspective for a linear output stage. Using the same network architecture
and same random values for the input to hidden layer weights, we applied the two
weight calculation methods to the MNIST database and demonstrated that our method
offers a performance advantage.
2
Methods
If we consider a particular sample of input data x k ∈ L×1 where k is a series index and K is the length of the series, then the forward propagation of the local signals through the network can be described by:
M
L
(2)
(1)
yn ,k = ∑ wnm
g ∑ wml
xl ,k
=
m 1=
l 1
(1)
Where y k ∈ N ×1 is the output vector corresponding to the input vector x k , l and
L are the input layer index and number of input features respectively, m and M are
the hidden layer index and number of hidden units respectively, and n and N are the
output layer index and number of output units respectively. w(1) and w(2) are the
weights associated with the input to hidden layer and the hidden to output layer linear
sums respectively. g ( ) is the hidden layer non-linear activation function.
With ELM, w(1) are assigned randomly which simplifies the training requirements
to task of optimisation of the w(2) only. The choice of linear output neurons further
simplifies the optimisation problem of w(2) to a single pass algorithm.
The weight optimisation problem for w(2) can be stated as
M
L (1)
(2)
yn ,k = ∑ wnm
am ,k where am ,k = g ∑ wml
xl ,k .
m =1
l =1
(2)
(2)
,
We can restate this as matrix equation by using W ∈ N × M with elements wnm
and A ∈ M × K in which each column contains outputs of the hidden unit at one instant
in the series ak ∈ M ×1 , and the output Y ∈ N × K where each column contains output
of the network at one instance in the series as follows:
Y = WA .
(3)
The optimisation problem involves determining the matrix W given a series of
desired outputs for Y and a series of hidden layer outputs A .
We represent the desired outputs for y k using the target vectors t k ∈ N ×1 where tn ,k
has value 1 in the row corresponding to the desired class and 0 for the other N-1 elements. For example t k = [0,1, 0, 0] indicates the desired target is class 2 (of four
T
classes). As above we can restate the desired targets using a matrix T ∈ N × K where
each column contains the desired targets of the network at one instance in the series.
Substituting T in for the desired outputs for Y , the optimization problem involves
solving the following linear equation for W :
T = WA .
2.1
(4)
Output weight calculation using the pseudo-inverse
In ELM literature W is often determined by the taking the Moore-Penrose pseudo-inverse A + ∈ K × M of A [3]. If the rows of are A are linearly independent (which
normally true if K>M) then W maybe calculated using
(
W = TA + where A + = A T AA T
)
-1
.
(5)
This minimises the sum of square error between networks outputs Y and the desired outputs T , i.e.
T2
A + minimises Y −=
K
N
∑∑ ( y
k 1=
n 1
=
n ,k
− tn , k )
2
(6)
We refer to the pseudo-inverse method for output weight calculation as PI-ELM.
We note that in cases where the classification problem is ill-posed it may be necessary
to regularize this solution, using standard methods such as Tikhonov regularization
(ridge regression).
2.2
Output weight calculation using linear discriminant analysis
In this paper we develop an alternative approach to estimating W based on a maximum likelihood estimator assuming a linear model. We refer to it as the LDA-ELM
method as it is equivalent to applying linear discriminant analysis to the hidden layer
outputs. Our presentation is based on the notation of Ripley [4].
For an N-class problem Bayes’ rule states that the posterior probability of the nth
class pn is related to its prior probability π k and its class density function f n ( d, θ n )
by
pn =
π n f n ( d, θ n )
N
∑ π z f z ( d, θ z )
(7)
z =1
where d is the input data vector (in our case the hidden layer output), and θ n are the
parameters of the class density function.
The class densities are modelled with a multi-variate Gaussian model with common covariance Σ and class dependent mean vectors μ k . Given an input vector ak
the class density is
f n ( ak , θn = μn , Σ ) =
( 2π )
− M2
Σ
− 12
T
exp − 12 ( ak − μ n ) Σ −1 ( ak − μ n )
(8)
We set the dimension of the Gaussian model equal to the number of hidden units
so that ak is as defined above for the hidden unit output and hence Σ ∈ M × M and
μ n ∈ M ×1 .
To begin with, the training data is partitioned according to the class membership so
N
that
we
K = ∑ K n labelled
have
data
vectors
of
hidden
unit
outputs,
n =1
M ×1
a(n)
,k =
1..K n where all members a
k ∈
(n)
belong to class n .
For a given set of hidden unit output data and class membership a likelihood function l ( θ ) is formed using
=
l (θ) l=
( θ1 , θ2 ,..., θ N )
Kn
N
∏∏ π
=
n 1=
k 1
n
f n ( a(kn ) , θ n )
(9)
Our aim is to find values of θ n that maximise l ( θ ) for given set of training data.
Equivalently we can maximise the value of the log-likelihood:
=
=
L ( θ1 , θ 2 ,..., θ N ) log
( l ( θ1 , θ2 ,..., θ N ) )
∑∑ log ( f ( a
N
Kn
=
n 1=
k 1
n
(n)
k
)
, θ n ) + ∑ K n log (π n ) (10)
N
=
n 1
Substituting our multi-variate Gaussian model for f n ( ak , θ n ) we get
L=
( θ1 , θ2 ,..., θ N ) L=
( μ1 ,..., μ N , Σ )
∑∑ ( −
N
Kn
=
n 1=
k 1
M
2
)
log ( 2π ) − 12 log ( Σ ) − 12 ( a(kn ) − μ n ) Σ −1 ( a(kn ) − μ n ) + ∑ K n log (π n ).
T
N
(11)
=
n 1
This is maximized when
Kn
μ n = ∑ a(kn ) K n , and Σ=
k =1
Kn
∑∑ ( a
N
=
n 1=
k 1
(n)
k
− μ k )( a(kn ) − μ k )
T
K.
(12)
Having determined the μ n ’s and Σ from the training data we now need to find the
values for W . We begin by substituting (8) into (7), bringing the π n into exponential
function and removing the common numerator and denominator term ( 2π )
− M2
Σ
− 12
,
giving us
T
exp − 12 ( ak − μ n ) Σ −1 ( ak − μ n ) + log (π n )
.
pn = N
T
exp − 12 ( ak − μ z ) Σ −1 ( ak − μ z ) + log (π z )
∑
z =1
After expanding the
(13)
− 12 ( ak − μ n ) Σ−1 ( ak − μ n ) terms and removing the
T
− 12 aTk Σ −1ak from the numerator and denominator we get
pn =
exp ( yn )
(14)
N
∑ exp ( ya )
a =1
where
yn = log (π n ) + μTn Σ −1ak − 12 μTn Σ−1μ n
(15)
Classification is performed by choosing the class with the highest value of pn . As
pn in (14) is a monotonic function of yn in (15) we can use either function when
deciding our final class. We choose to use yn defined in (15) as it is a linear function
of the input data vector ak and it can be used to determine W for our network as follows:
log (π 1 ) − 12 μ1T Σ −1μ1 log (π 2 ) − 12 μT2 Σ −1μ 2 log (π N ) − 12 μTN Σ −1μ N
W=
. (16)
μ1T Σ −1
μT2 Σ −1
μTN Σ −1
Note that W ∈ N × M +1 , as a constant term has been introduced into the hidden to
output layer weights (the first row of W ). If we want to determine the posterior probabilities then we use (14) applied to the network outputs.
Summary of method
In summary calculating W proceeds as follows
(i) Partitioned the hidden unit output data according to the class membership
N
L×1
so that we have K = ∑ K n labelled data vectors, a(n)
,k =
1..K n where
k ∈
n =1
all members a(n) belong to class n
Kn
(ii) Calculate μ n = ∑ a(kn ) K n and Σ=
k =1
Kn
∑∑ ( a
N
=
n 1=
k 1
(n)
k
− μ k )( a(kn ) − μ k )
T
K
(iii) Set the prior probabilities π n
(iv) Calculate
log (π 1 ) − 12 μ1T Σ −1μ1 log (π 2 ) − 12 μT2 Σ −1μ 2 log (π N ) − 12 μTN Σ −1μ N
W=
μ1T Σ −1
μT2 Σ −1
μTN Σ −1
To classify new data we
(i) Calculate the network output y in response to the hidden layer output a is
1
y = W
a
(ii) (Optional) Calculate the posterior probabilities
pn =
exp ( yn )
N
∑ exp ( y )
a =1
a
(iii) The final decision of the network is the output with the highest value of yn
or, equivalently, pn .
Combining Classifiers
Equation (14) provides an easy way to combine the outputs of multiple classifiers.
Once the posterior probabilities are calculated for each class for each classifier we can
form a combined posterior probability and choose the class with the highest combined
posterior probability. There are any schemes for doing this [5] with unweighted averaging across the posterior probability outputs being one of the most simple schemes.
3
Experiments
We applied the LDA-ELM and PI-ELM weight calculation method to the MNIST
handwritten digit recognition problem [6]. Authors JT and AvS have previously reported good classification results using ELM on this database [2]. The database has
60,000 training and 10,000 testing examples. Each example is a 28*28 pixel 256 level
grayscale image of a handwritten digit between 0 and 9. The 10 classes are approximately equally distributed in the training and testing sets.
The ELM algorithms were applied directly to the unprocessed images and we
trained the networks by providing all data in batch mode. The random values for the
input layer weights were uniformly distributed between -0.5 and 0.5. The prior probabilities for the 10 classes for LDA-ELM were each set to 0.1.
In order to perform a direct comparison of the two methods we used the following
protocol:
For fan-out of 1 to 20 hidden units per input, repeat 200 times
(i) Assign random values to the input layer weights and determine the hidden
layer outputs for the 60,000 training data examples.
(ii) Determine PI-ELM network weights using data from (i).
(iii) Determine LDA-ELM network weights using data from (i).
(iv) Evaluate both networks on the 10,000 test data examples and store results.
We averaged the results for the 200 repeats of the experiment for each fan-out and
compared the misclassification rates. These results are shown in Fig. 1 and Table 1.
7
Error (%)
6
5
LDA-ELM
4
PI-ELM
3
2
0
5
10
15
20
Fan-out
Fig. 1. The error rate of the LDA-ELM and the PI-ELM on the MNIST database for fan-out
varying between 1 and 20. All results at each fan-out are averaged from 200 repeats of the
experiment.
The results show that the LDA-ELM outperforms the PI-ELM at every fan-out
value. The average performance benefit was a 3.1% decrease in the error rate of
LDA-ELM with a larger benefit at smaller fan-out values. Table 2 below shows that
there is little extra computational requirement for the LDA-ELM method.
Table 1. The error rate (%) and percentage improvement of the LDA-ELM over the PI-ELM on
the MNIST database. Results averaged from 200 repeats.
1
2
3
4
5
6
Fanout
7
8
9
10
12
15
20
PI-ELM
7.20 5.21 4.32 3.80 3.45 3.20 3.00 2.86 2.74 2.63 2.49 2.31 2.15
LDA-ELM
6.84 5.03 4.17 3.68 3.35 3.11 2.92 2.78 2.66 2.55 2.42 2.25 2.08
% improvement 4.9 3.5 3.3 3.1 2.9 3.0 2.6 2.9 2.8 2.7 2.6 2.6 3.3
Table 2. Computation times (in seconds). The elapsed time is shown for training the PI-ELM
and LDA-ELM networks on the 60,000 images from MNIST database and testing on the
10,000 images using MATLAB R2012a code running on 2012 Sony Vaio Z series laptop with
an Intel i7-2640M 2.8GHz processor and 8GB RAM.
Fan-out
1
3
4
5
6
7
8
9
10
12
15
20
6.2 13.9
24.9
37.8
53.3
68.5
88.2
111
136
162
228
339
630
LDA-ELM 6.2 13.9
25.2
38.1
54.0
69.5
90.6
113
140
167
238
357
702
PI-ELM
2
The last experiment we performed investigated combining multiple networks using
the LDA-ELM by averaging posterior probabilities. We investigated using an ensemble number between 1 and 20 and repeated the training and testing 10 times at each
ensemble number. We then averaged the results which are shown below in Fig. 2.
2.1
2.0
Error (%)
1.9
1.8
1.7
1.6
1.5
0
5
10
15
20
Ensemble
Fig. 2. The error rate of the LDA-ELM on the MNIST database at a fan-out of 20 with the
ensemble number varying between 1 and 20. The result at each ensemble number is averaged
from 10 repeats of the experiment.
The results shown in Fig. 2 demonstrate the benefit of combining multiple LDAELM networks on the MNIST database. Combining two networks reduced the error
rate from 2.08% to 1.86% and adding more networks further reduced the error. The
best error rate was 1.69% achieved when 20 networks were combined.
4
Discussion
The results on the MNIST database shown in Fig. 1 suggest that there is a performance benefit to be gained by using the LDA-ELM output weight calculation over the
PI-ELM method. As there is only a small extra computation overhead we believe it is
a viable alternative to the pseudo-inverse method especially at small fan-out values.
Another benefit of the LDA-ELM is the ability to combine outputs from networks
by combining the posterior probabilities estimates of the individual networks. When
we applied this to the MNIST database we were able to reduce the error rate to 1.7%.
This result is comparable to the best performance of most other 2 and 3 layer neural
networks processing the raw data [7]. Further work will include comparing the two
weight calculation methods on other publicly available databases such as abalone and
iris data sets [8].
5
Conclusion
We have presented a new method for weight calculation for hidden to output
weights for ELM networks performing classification tasks. The method is based on
linear discriminant analysis and requires a modest amount of extra calculation time
compared to the pseudo-inverse method (<12% for a fan-out ≤ 20). When applied to
the MNIST database the average misclassification rate improvement was 3.1% in
comparison to the pseudo-inverse method for identically configured and initialized
networks.
6
Bibliography
1. G.B. Huang, Q.Y. Zhu and C.K. Siew, “Extreme Learning Machine: Theory and Applications,” Neurocomputing, vol. 70, pp. 489–501, 2006.
2. J. Tapson and A. van Schaik, “Learning the pseudoinverse solution to network weights,”
Neural Networks, vol. 45, pp. 94–100, Sep. 2013.
3. R. Penrose and J. A. Todd, “On best approximate solutions of linear matrix equations,”
Mathematical Proceedings of the Cambridge Philosophical Society, vol. 52, pp 17-19,
1956.
4. B.D. Ripley, Pattern Recognition and Neural Networks, Cambridge Univ. Press, 1996.
5. L.I. Kuncheva, Combining Pattern Classifiers: Methods and Algorithms, Wiley Press,
2004.
6. Y. LeCun and C. Cortes, “The MNIST database of handwritten digits”, Available at:
http://yann.lecun.com/exdb/mnist.
7. Y. LeCun, L. Bottou, Y. Bengio and P. Haffner, “Gradient-Based Learning Applied to
Document Recognition,” Proceedings of the IEEE, vol. 86(11), pp. 2278-2324, Nov. 1998.
8. Data sets from the UCI Machine Learning Repository, http://archive.ics.uci.edu/ml/.
| 9cs.NE
|
Recurrent Orthogonal Networks and Long-Memory Tasks
Mikael Henaff
New York University, Facebook AI Research
arXiv:1602.06662v2 [cs.NE] 15 Mar 2017
Arthur Szlam
Facebook AI Research
MBH 305@ NYU . EDU
ASZLAM @ FACEBOOK . COM
Yann LeCun
New York University, Facebook AI Research
Abstract
Although RNNs have been shown to be powerful tools for processing sequential data, finding
architectures or optimization strategies that allow them to model very long term dependencies
is still an active area of research. In this work,
we carefully analyze two synthetic datasets originally outlined in (Hochreiter & Schmidhuber,
1997) which are used to evaluate the ability of
RNNs to store information over many time steps.
We explicitly construct RNN solutions to these
problems, and using these constructions, illuminate both the problems themselves and the way in
which RNNs store different types of information
in their hidden states. These constructions furthermore explain the success of recent methods
that specify unitary initializations or constraints
on the transition matrices.
1. Introduction
Recurrent Neural Networks (RNNs) are powerful models
which are naturally suited to processing sequential data.
They can maintain a hidden state which encodes information about previous elements in the sequence. For a classical version of RNN (Elman, 1990), at every timestep, the
hidden state is updated as a function of both the input and
the current hidden state. In theory, this recursive procedure
allows these models to store complex signals for arbitrarily
long timescales.
However, in practice RNNs are considered difficult to train
due to the so-called vanishing and exploding gradient problems (Bengio et al., 1994). These problems arise when the
Proceedings of the 33 rd International Conference on Machine
Learning, New York, NY, USA, 2016. JMLR: W&CP volume
48. Copyright 2016 by the author(s).
YANN @ CS . NYU . EDU
spectral norm of the transition matrix is significantly different than 1, or due to the non-linear transition functions. If
the spectral norm of the transition matrix is greater than 1,
the gradients will grow exponentially in magnitude during
backpropagation, which is known as the exploding gradient problem. If the spectral norm is less than 1, the gradients will vanish exponentially quickly, which is known as
the vanishing gradient problem. Recently, a simple strategy of clipping gradients has been introduced, and has
proved effective in addressing the exploding gradient problem (Mikolov, 2012). The problem of vanishing gradients
has shown itself to be more difficult, and various strategies
have been proposed over the years to address it. One very
successful approach, known as Long Short-Term Memory
(LSTM) units (Hochreiter & Schmidhuber, 1997), has been
to modify the architecture of the hidden units by introducing gates which explicitly control the flow of information
as a function of both the state and the input. Specifically,
the signal stored in a hidden unit must be explicitly erased
by a forget gate and is otherwise stored indefinitely. This
allows information to be carried over long periods of time.
LSTMs have become very successful in applications to language modeling, machine translation, and speech recognition (Zaremba et al., 2014; Sutskever et al., 2014; Graves
et al., 2013). Other methods have been proposed to deal
with learning long-term dependencies, such as adding a
separate contextual memory (Mikolov et al., 2015), stabilizing activations (Krueger & Mimesevic, 2015) or using more sophisticated optimization schemes (Martens &
Sutskever, 2011). Two recent methods propose to directly
address the vanishing gradient problem by either initializing or parameterizing the transition matrix with orthogonal
or unitary matrices (Arjovsky et al., 2015; Le et al., 2015).
These works have used a set of synthetic problems (originally outlined in (Hochreiter & Schmidhuber, 1997) or
variants thereof) for testing the ability of methods to learn
long-term dependencies. These synthetic problems are
Recurrent Orthogonal Networks and Long-Memory Tasks
designed to be pathologically difficult, and require models to store information over very long timescales (hundreds of timesteps). Different approaches have solved these
problems to varying degrees of success. In (Martens &
Sutskever, 2011), the authors report that their HessianFree optimization based method solves the addition task for
T = 200 timesteps. The authors of (Krueger & Mimesevic, 2015) reported that their method beat the chance baseline for the adding task in 8/9 cases for T = 400. In (Le
et al., 2015), the IRNN is reported to solve the addition task
for T = 300. The method proposed in (Arjovsky et al.,
2015) is able to solve the copy task for up to T = 500
timesteps, and is able to completely solve the addition task
for up to T = 400 timesteps, and partially solves it for
T = 750.
In this work we analyze these “long-memory” tasks, and
construct explicit RNN solutions for them. The solutions
illuminate both the tasks, and provide a theoretical justification for the success of recent approaches using orthogonal
initializations of, or unitary constraints on, the transition
matrix of the RNN. In particular, we show that a classical
Elman RNN with no transition non-linearity and random
orthogonal initialization is with high probability close to
our explicit RNN solution to the sequence memorization
task, and the same network architecture with identity initialization is close to the explicit solution to the addition
task. We verify experimentally that initializing correctly
(i.e. random orthogonal or identity) is critical for success
on these tasks. Finally, we show how l2 pooling can be used
to allow a model to “choose” between a random-orthogonal
or identity-like memory.
There are several other works which have studied the properties of orthogonal matrices in relation to neural networks.
The work of (Saxe et al., 2013) gives exact solutions to the
learning dynamics of deep linear networks, and based on
this analysis, suggests an orthogonal initialization scheme
to accelerate learning. The authors of (White et al., 2004)
and (Ganguli et al., 2008) study the ability of linear RNNs
(with orthogonal and generic transition matrices, respectively) to store scalar sequences in their hidden state, and
show that the memory capacity scales with the number of
hidden units. Our work complements theirs by providing a
related analysis for discrete input sequences.
2. Architectures
We review some recurrent neural network (RNN) architectures for processing sequential data, and discuss the
modifications we use for the long memory problems. We
fix the following notation: input sequences are denoted
x0 , x1 , ..., xt , ..., and output sequences are denoted by
y0 , y1 , ..., yt , ... .
2.1. sRNN
An sRNN (Elman, 1990) consists of a d × d transition matrix V , an M × d decoder matrix W (where M is the output dimension), a d × N encoder matrix U (where N is
the input dimension), and a bias b. If either the output or
input is categorical, M (respectively N ) is the number of
classes, and we use a one-hot representation. As the sRNN
ingests a sequence, it keeps running updates to a hidden
state h, and using the hidden state and the decoder matrix,
produces outputs y:
ht = σ(U xt + V ht−1 + b)
yt = W ht
(1)
where xt , yt , ht are the input, output and hidden state respectively at time t. While there have been great improvements in the training of sRNNs since their introduction, and
while they have been shown to be powerful models in tasks
such as language modeling (Mikolov, 2012), it can still be
difficult to train generic sRNNs to use information about
inputs from hundreds of timesteps previous for computing the current output (Bengio et al., 1994; Pascanu et al.,
2013).
In the following, we will use a simplification of the sRNNs
that makes them in some sense less powerful models, but
makes it easier to train them to solve simple long-memory
tasks. Namely, by placing the non-linearity between the
input and hidden state, rather than between the hidden state
and output, we obtain RNNs with linear transitions (or, in
the case of categorical inputs, not using a non-linearity at
all). We call these LT-RNNs. The update equations are
then:
ht = σ(U xt + b) + V ht−1
yt = W ht
(2)
Finally, note that by appropriately scaling the weights and
biases, an sRNN can be made to approximate a LT-RNN,
but not the other way around (and of course the optimization may never find this scaling).
2.2. LSTM
The LSTM of (Hochreiter & Schmidhuber, 1997) is an architecture designed to improve upon the sRNN with the
introduction of simple memory cells with a gating architecture. In this work we use the architecture originally
proposed in (Hochreiter & Schmidhuber, 1997). For each
memory cell, the network computes the output of four
gates: an update gate, input gate, forget gate and output
gate. The outputs of these gates are:
Recurrent Orthogonal Networks and Long-Memory Tasks
3.1. Copying Problem
i = σ(Ui xt + Vi ht−1 )
f = σ(Uf xt + Vf ht−1 )
o = σ(Uo xt + Vo ht−1 )
(3)
g = tanh(Ug xt + Vg ht−1 )
The cell state is then updated as a function of the input and
the previous state:
ct = f
ct−1 + i
g
(4)
Finally, the hidden state is computed as a function of the
cell state and the output gate:
ht = o
tanh(ct )
(5)
A relatively common variation of the original LSTM
involves adding so-called “peephole connections” (Gers
et al., 2003) which allows information to flow from the cell
state to the various gates. This variant was originally designed to measure or generate precise time intervals, and
has proven successful for speech recognition and sequence
generation (Graves, 2013; Graves et al., 2013).
2.3. LT-RNN with l2 pooling
Below we will consider LT-RNNs initialized with either
random orthogonal transition matrices, or identity transitions, and we will see that there is a large difference in
behavior between these initializations. However, we can
set up an architecture where a random orthogonal initialization behaves much closer to an identity initialization by
using an l2 pooling layer at the output. If we feed both the
pooled and unpooled hidden layer to the decoder, the model
can choose whether it wants an identity-like or randomorthogonal like representation. We fix a pool size k, and
then the update equations for this model are:
ht = σ(U xt + b) + V ht−1
yt = WI ht + WP Pk (ht )
(6)
where if h is the kd dimensional vector h = [h1 , ..., hkd ]T ,
then P (h) is the d dimensional vector defined by
v
u
ki
u X
P (h)i = t
h2j
This task tests the network’s ability to recall information
seen many time steps previously. We follow the same setup
as (Arjovsky et al., 2015), which we briefly outline here.
Let A = {ai }K
i=1 be a set of K symbols, and pick numbers
S and T . The input consists of a T + 2S length vector of
categories, starting with S entries sampled uniformly from
{ai }K
i=1 which are the sequence to be remembered. The
next T − 1 inputs are set to aK+1 , which is a blank category. The following (single) input is aK+2 , which represents a delimiter indicating that the network should output
the initial S entries of the input. The last S inputs are set
to aK+1 . The required output sequence consists of T + S
entries of aK+1 , followed by the first S entries of the input
sequence in exactly the same order. The task is to minimize the average cross-entropy of the predictions at each
time step, which amounts to remembering a categorical sequence of length S for T time steps.
3.1.1. A SOLUTION MECHANISM
We can write out an LT-RNN solution for this problem. We
will write out descriptions for U, V, W from equation (2).
Note that since the inputs are categorical, we assume that
no non-linearity is used. Fix a number d. For each j in
{1, ..., d}, pick a random integer lj drawn uniformly from
{1, ..., T + S}, and let
cos(2lj π/(T + S)) sin(2lj π/(T + S))
Qj =
.
− sin(2lj π/(T + S)) cos(2lj π/(T + S))
Now define Q, and then V from (2) by
Q1 0 · · · 0
0 Q2 · · · 0
Q 0
,
V
=
.
Q= .
..
0 1
..
. 0
0
0
Qd
So V is a (2d + 1) × (2d + 1) block diagonal matrix. Note
that iterating Q “spins” each of the Qi at different rates, but
they all synchronize at multiples of S +T . Thus Q acts as a
“clock” with period S+T . Now set Ũ to be a 2d×K matrix
with columns sampled uniformly from the unit sphere, and
form U by appending two zero columns to Ũ and then one
extra row, with −1/S for each entry between 1 and K, −1
for the K + 1 entry, and T + S + 1 for the K + 2 entry.
Schematically,
j=k(i−1)+1
3. Tasks
In this section we describe tasks from (Hochreiter &
Schmidhuber, 1997; Arjovsky et al., 2015; Le et al., 2015)
which involve dependencies over very long timescales
which are designed to be pathologically hard for the sRNN.
U=
Ũ
− S1
0
−1
0
.
S+T +1
Finally, set W = U T , except scale the K + 1 column by
S + 1, zero out the K + 2 column, and also zero out the
entries below Ũ .
Recurrent Orthogonal Networks and Long-Memory Tasks
This gives
T
Ũ
W = 0
0
0
−(S + 1) .
0
Now we will show how the RNN operates, starting with
a high-level overview. The last dimension of the hidden
state divides the state space into two regions, one where
the model outputs the blank symbol and the other where it
outputs one of the first K symbols in the dictionary. The
model begins in the first region and remains there until it
encounters the delimiter symbol, which sends it into the
second. Now, the symbols in the input sequence are all
encoded in the hidden state and a rotation is applied with
each timestep. A key result is that rotation by powers of Q
“hides” a symbol encoded in the hidden state, i.e. decorrelates its current representation from its original one. Due
to the periodicity of Q, after T + S timesteps, the different
symbols in the input sequence will surface from the hidden
state one at a time, in the order in which they were seen,
while the other symbols in the sequence, whose representations have rotations applied to them, remain hidden. This
causes the output units of each of the symbols to fire in the
correct order.
We now give a more precise description, for which we need
a little more notation. Denote by h̃ the first 2d coordinates
of h, and by h2d+1 the last coordinate, and denote by uj
the jth column of Ũ . Then the RNN works as follows,
initialized with hidden state 0:
• After the first S inputs, we have
h̃ =
S
X
S−j
Q
Figure 1. Success percentage of the mechanism from 3.1.1 to the
copy problem for T = 500, computed over 500 trials with d =
128.
• The sum continues to cycle, giving ij as output for
each following j up to S.
PS
We now briefly argue that uTi1 j=2 Q1−j uij is small when
d is large enough w.r.t. S and K. We will repeatedly use
that the variance of a sum of independent, mean-zero random variables grows as the sum of the variances. Denote
by uji the pair of coordinates of the jth column uj of Ũ
corresponding to the ith block; since u are uniform on the
sphere, we expect
||uji ||2 ∼ 1/d.
Since for each fixed p ∈ {1, T + S}, over the choices of li
in the definition of Qi , the Qpi are independent, uTji Qpi uji
has mean zero, and since
uTj Qp uj =
uij .
d
X
uTji Qpi uji
i=1
j=1
we expect
• For the next T inputs, only h2d+1 changes, incrementing by −1 at each step. Note that WK+1 has so far
been the best match to h because it has large negative
last component.
• At time T + S when the aK+2 token is seen, h2d+1 is
set positive, which ensures the blank symbol will not
be output.
Moreover, since the Qp uj are uniform on the sphere,
|uTj Qp uj 0 |2 ∼ 1/d
for j 0 6= j. Similarly, we expect
uTi1
• At time T + S + 1,
h̃ = ui1 +
|uTj Qp uj |2 ∼ 1/d.
S
X
2
1−j
Q
uij
∼ (S − 1)/d.
j=2
S
X
Q1−j uij .
j=2
We argue below that if d is large
PS enough w.r.t. S and
K, with high probability, uTi1 j=2 Q1−j uij is small,
and so multiplication with W has max value at i1 .
Thus we can fix a small number , say = .1, and
choose
PSd large enough so that with high probability
|wiT1 j=2 Qj−S uij |2 < , even though wiT ui = 1. Finally, there is a weak dependence on K here; for fixed
and K it is exponentially unlikely (in d) that the nearest
neighbor i0 6= i1 is close enough to ui to interfere.
Recurrent Orthogonal Networks and Long-Memory Tasks
This solution mechanism suggests that a random orthogonal matrix (chosen, for example, via QR decomposition of
a Gaussian matrix) is a good starting point for solving this
task. The construction above is invariant to rotations; and
we can always find a basis so that a given orthogonal matrix has the block form above in that basis. Thus all that is
necessary is for the descent to nudge the eigenvalues of the
orthogonal matrix to be S + T roots of unity, and then it
already has the basic form of the construction above. This
also gives a good explanation for the performance of the
models used for the copy problem in (Arjovsky et al., 2015)
is added to the hidden state, as xj [1] ≤ 1. On the other
hand, if xj [2] = 1, then exactly xj [2] is added to h.
This mechanism has been known (at least implicitly, although we don’t know if it has been written down explicitly
before) at least since (Hochreiter & Schmidhuber, 1997),
and it can be seen as a very simple LSTM model, with the
following gates:
i = xj [1]
f =1
o=1
g = (xj [2] − 1)+
(7)
Finally, note that although we used the setup of (Arjovsky
et al., 2015), the construction can be modified to solve
problems 2a and 2b in (Hochreiter & Schmidhuber, 1997)
and no non-linearity in Equation (5).
3.1.2. SOLUTION MECHANISM EXPERIMENTS
3.3. Comparison between the tasks
Since the construction of the copy mechanism is randomized, we provide an experiment to show how the solution
degrades as a function of K (the dictionary size) and S (the
length of the sequence to be remembered). There is not a
strong dependence on T (the length of time to remember
the sequence). Figure 1 shows the number of successes
over 500 runs with d = 128.
Note that the 1 × 1 matrix V in the mechanism for the
adding problem is the “identity”. We can build a more redundant solution by using a larger identity matrix. We can
describe the identity using the same block structure as the
matrix Q defined for the copy task; namely each lj = 0.
On the other hand, the Q for the copy task acts as a “clock”
that synchronizes after a fixed number of steps T + S. It
is important for the mechanism we described that the clock
looks random at any time between 1 and S + T . For example, if we had instead used the same lj in each block Qj ,
the mechanism would not succeed. The transition matrices
for the addition task and the copy task are thus opposites
in the sense that for addition, all the lj are the same (i.e. a
δ mass on the unit circle), and for copy, the lj are as uniformly distributed on the unit circle as possible.
3.1.3. VARIABLE L ENGTH C OPY P ROBLEM
Note that the solution mechanism for the copy problem
above depends on having a fixed location for regurgitating
the input. In the experiments below, we also discuss a variant of the copy task, where the symbol to indicate that the
memorized sequence must be output is randomly located
in S + 1, S + T ; this can be considered a variant of task 2c
in (Hochreiter & Schmidhuber, 1997). We do not know a
bounded in T explicit LT-RNN or sRNN solution for this
variable length problem (although the above solution using
a multiplicative RNN instead of an sRNN, and keeping d2
extra hidden variables to track the power of V solves it).
3.2. Adding Problem
The adding problem requires the network to remember two
marked numbers in a long sequence and add them. Specifically, the input consists of a two dimensional sequence
{x1 , ..., xT }. The first coordinate xj [1] is uniformly sampled between 0 and 1, and the second coordinate is 0 at each
j save two; in these two entries, xj [2] = 1. The required
output is xj1 [1] + xj2 [1], where xji [2] = 1.
3.2.1. A SOLUTION MECHANISM
This problem has a simple, explicit solution using a LTRNN with a ReLU non-linearity and a one dimensional
hidden state. Namely: set U = [11], b = −1, W = 1,
and V = 1. At each time step j, if xj [2] = 0 then nothing
In the experiments below, we will show that it is hard for
an LT-RNN to learn the adding task when its transition matrix is initialized as a random orthogonal matrix but easy
when initialized with the identity, and vice-versa for the
copy task. One way to get a “unified” solution is to use l2
pooling, as in 6. Then when initialized with a matrix with
lj distributed uniformly, the decoder can choose to use the
pooled hiddens (which through away the phase, and so appear identity-like) for the adding task, or use the raw hiddens, which are clock-like.
4. Experiments
4.1. Impact of Initialization
Based on the above analysis, we hypothesize that an LTRNN with random orthogonal initialization (denoted LTORNN) should perform well on the sequence memorization problem, and an LT-RNN with identity initialization
(denoted LT-IRNN) should perform well on the addition
task. To test this, we conducted the following experiment
on both the copy and addition task for different timescales.
Recurrent Orthogonal Networks and Long-Memory Tasks
c opy ta s k, T = 200
0.20
c opy ta s k, T = 500
0.20
LSTM
LT-IRNN
LT-ORNN
LSTM
LT-IRNN
LT-ORNN
0.15
Cros s Entropy
Cros s Entropy
0.15
0.10
0.05
0.00
0.10
0.05
0
200
400
600
ite ra tions (hundre ds )
800
0.00
1000
0
200
400
600
ite ra tions (hundre ds )
800
1000
Figure 2. Results for the copy task.
a dd ta s k, T = 200
0.5
a dd ta s k, T = 400
0.5
LSTM
LT-IRNN
LT-ORNN
0.4
0.4
MSE
0.3
MSE
MSE
LSTM
LT-IRNN
LT-ORNN
0.4
0.3
0.3
a dd ta s k, T = 750
0.5
LSTM
LT-IRNN
LT-ORNN
0.2
0.2
0.2
0.1
0.1
0.1
0.0
0
200
400
600
800
1000
1200
ite ra tions (hundre ds )
1400
1600
1800
0.0
0
200
400
600
800
1000
1200
ite ra tions (hundre ds )
1400
1600
1800
0.0
0
200
400
600
800
1000
1200
ite ra tions (hundre ds )
1400
1600
1800
Figure 3. Results for the addition task.
For each task and timescale, we trained 8 LT-ORNNs and
8 LT-IRNNs with different random seeds. The transformation matrices for all models were intialized using
√ a Gaussian distribution with mean 0 and variance 1/ n (where
n is the number of incoming connections to each hidden
unit). For LT-ORNNs, we then projected the transition matrix to its nearest orthogonal matrix by setting its singular
values to 1.
In all experiments, we used RMSProp to train our networks with a fixed learning rate and a decay rate of 0.9.
In preliminary experiments we tried different learning rates
in {1, 10−1 , 10−2 , 10−3 , 10−4 , 10−5 } and chose the largest
one for which the loss did not diverge, for the LT-RNN’s we
used 10−4 .
We also included LSTMs in all our experiments as a baseline. We used the same method as for LT-RNN to pick the
learning rate, and ended up with 10−3 .
For all experiments, we normalized the gradients with respect to hidden activations by 1/T , where T denotes the
number of timesteps. In preliminary experiments, we also
found that for LT-RNN models the activations frequently
exploded whenever the largest singular value of the transition matrix became much greater than 1. Therefore,
we adopted a simple activation clipping strategy where
we rescaled activations to to have magnitude l whenever
their magnitude exceeded l. In our experiments we chose
l = 1000.
Figure 2 shows the results on the copy task for the LSTM,
LT-ORNN and LT-IRNN. All networks are trained with 80
hidden units. We see that the LSTM has difficulty beating the baseline performance of only outputting the empty
symbol; however it does eventually converge to the solution
(this is not shown in the figure). However, the LT-ORNN
solves the task almost immediately. We note that this behavior is similar to that of the uRNN in (Arjovsky et al.,
2015), which is paramaterized in a way that makes it easy
to recover the explicit solution described above. The LTIRNN is never able to find the solution.
Figure 3 shows the results of the addition task for T =
200, 400 and 750 timesteps. All networks are trained with
128 hidden units. For T = 750, we trained a single LTORNN and LT-IRNN due to time constraints. In contrast to
the copy task, here the LT-IRNN is able to efficiently solve
Recurrent Orthogonal Networks and Long-Memory Tasks
a dd ta s k, T = 400
0.5
0.4
LSTM
LT-IRNN
LT-ORNN
LT-ORNN-POOL
0.4
MSE
0.3
MSE
0.3
0.2
0.2
0.1
0.1
0.0
0
200
400
600
800
1000
1200
ite ra tions (hundre ds )
1400
1600
0
500
0.05
2000
LSTM
LT-IRNN
LT-ORNN
LT-ORNN-POOL
0.15
c ros s e ntropy
0.10
1000
1500
ite ra tions (hundre ds )
c opy ta s k, T = 500
0.20
LSTM
LT-IRNN
LT-ORNN
LT-ORNN-POOL
0.15
0.00
0.0
1800
c opy ta s k, T = 200
0.20
c ros s e ntropy
a dd ta s k, T = 750
0.5
LSTM
LT-IRNN
LT-ORNN
LT-ORNN-POOL
0.10
0.05
0
200
400
600
ite ra tions (hundre ds )
800
1000
0.00
0
200
400
600
ite ra tions (hundre ds )
800
1000
Figure 4. Results for copy and addition task with pooling architectures. Note that the LSTM will eventually solve the copy task, but the
LT-IRNN will not.
the problem whereas the LT-ORNN is only able to solve it
after a very long time, or not at all. The LSTM is also able
to easily solve the task, which is consistent with the original work of (Hochreiter & Schmidhuber, 1997) where the
authors report solving the task for up to 1000 timesteps. We
note that this LSTM baseline differs from that of (Arjovsky
et al., 2015; Le et al., 2015) where it is reported to have
more difficulty solving the addition task. We hypothesize
that this difference is due to the use of different variants of
the LSTM architecture such as peephole connections.
4.2. Pooling Experiments
We next ran a series of experiments to examine the effect
of feeding pooled outputs to the decoder, to see if we could
obtain good performance on both the copy and addition
tasks with a single architecture and initialization. In these
experiments, we added a soft penalty on the transition matrix V to keep it orthogonal throughout training. Specifi-
cally, at every iteration we applied one step of stochastic
gradient descent to minimize the loss ||V T V − I||, evaluated at m random points on the unit sphere. Note that
this requires O(md2 ) operations and a regular update requires O(T md2 ) operations, so adding this soft constraint
has negligible computational overhead. In our experiments
we set m = 50, which was the same at the minibatch size.
In all pooling experiments we used a pool size and stride
of 2. The results are shown in Figure 4. The LT-ORNN
with pooling is easily able to solve the copy task for both
timescales, and approximately solves the addition task for
both timescales as well, even though convergence is slower
than the LT-IRNN. Its success on the copy task is not surprising, since by zeroing out the matrix WP in Equation 6
it can solve the problem with the same solution as the regular LT-ORNN. The good performance on the adding task
is somewhat more interesting. To gain insight into how the
network stores information in a stable manner while having
Recurrent Orthogonal Networks and Long-Memory Tasks
va ria ble le ng th c opy ta s k, T = 100
0.5
LSTM
LT-IRNN
LT-ORNN
LT-ORNN-POOL
c ros s e ntropy
0.4
0.3
0.2
0.1
0.0
0
200
400
600
800
ite ra tions (hundre ds )
1000
1200
Figure 5. Results for the variable length copy task.
an (approximately) orthogonal transition matrix, we plotted
the activations of its hidden states over time as it processes
an input sequence. This is displayed in Figure 6. We observe relatively constant activations until the first marked
number is encountered, which triggers oscillatory patterns
along certain dimensions. When the second marked number is seen, existing oscillations are amplified and new ones
emerge. This suggests that the network stores information stably through the radius of its hidden state’s rotations
along different 2-dimensional subspaces. The information
is then recovered as the phase is discarded though the pooling operation. Thus the model can have “uniform” clocklike oscillations that are perceived as δ-like after the pooling.
4.3. Variable Length Copy Task
Having seen the stark impact of initialization on the performance of LT-IRNNs and LT-ORNNs for the copy and
addition task, and its mitigation through the addition of a
pooling layer, we then tested all the models on a problem
for which we did not have a (roughly fixed size) solution
mechanism, namely the variable length copy task. Figure 5 shows the performance of an LT-IRNN, LT-ORNN,
LT-ORNN with l2 pooling, and LSTM (each with 80 hidden units) on the variable length copy task with T = 100
timesteps. Even though the number of timesteps is significantly less than in other tasks, none of the LT-RNNs are
able to beat the chance baseline, whereas the LSTM is able
to solve the task even though its convergence is slow. This
experiment is a classic example of how a detail of construction of a synthetic benchmark can favor a model in a way
that fails to generalize to other tasks.
Figure 6. Activation patterns of pooling network. The two marked
numbers to be added occur at positions 264 and 463.
5. Conclusion
In this work, we analyzed two standard synthetic long-term
memory problems and provided explicit RNN solutions for
them. We found that the (fixed length T ) copy problem can
be solved using an RNN with a transition matrix that is a
T + S root of the identity matrix I, and whose eigenvalues
are well distributed on the unit circle, and we remarked that
random orthogonal matrices almost satisfy this description.
We also saw that the addition problem can be solved with
I as a transition matrix. We showed that correspondingly,
initializing with I allows a linear-transition RNN to easily
be optimized for solving the addition task, and initializing
with a random orthogonal matrix allows easy optimization
for the copy task; but that flipping these leads to poor results. This suggests an optimization difficulty in transitioning between oscillatory and steady dynamics, which can
be mitigated by adding an l2 pooling layer that allows the
model to easily choose between the two regimes. Finally,
our experiment with the variable length copy task illustrates
that although synthetic benchmarks can be useful for evaluating specific capabilities of a given model, success does
not necessarily generalize across different tasks, and novel
model architectures should be evaluated on a broad set of
benchmarks as well as natural data.
References
Arjovsky, M, Shah, A, and Bengio, Y. Unitary Evolution Recurrent Neural Networks, September 2015. URL
http://arxiv.org/abs/1511.06464.
Bengio, Y., Simard, P., and Frasconi, P. Learning long-term
dependencies with gradient descent is difficult. IEEE
Transactions on Neural Networks, 5(2):157–166, 1994.
Recurrent Orthogonal Networks and Long-Memory Tasks
Elman, Jeffrey L. Finding structure in time. COGNITIVE
SCIENCE, 14(2):179–211, 1990.
Ganguli, Surya, Huh, Dongsung, and Sompolinsky, Haim.
Memory traces in dynamical systems. 105(48):18970–
18975, 2008. doi: 10.1073/pnas.0804451105.
Gers, Felix A., Schraudolph, Nicol N., and Schmidhuber, Jürgen. Learning precise timing with lstm
recurrent networks. J. Mach. Learn. Res., 3:115–
143, March 2003. ISSN 1532-4435. doi: 10.1162/
153244303768966139. URL http://dx.doi.org/
10.1162/153244303768966139.
Graves, Alex. Generating Sequences with Recurrent
Neural Networks, September 2013. URL http://
arxiv.org/abs/1308.0850.
Graves, Alex, Mohamed, Abdel-rahman, and Hinton,
Geoffrey E.
Speech recognition with deep recurrent neural networks. In IEEE International Conference on Acoustics, Speech and Signal Processing,
ICASSP 2013, Vancouver, BC, Canada, May 26-31,
2013, pp. 6645–6649, 2013. doi: 10.1109/ICASSP.2013.
6638947. URL http://dx.doi.org/10.1109/
ICASSP.2013.6638947.
Hochreiter, S and Schmidhuber, J. Long short-term memory. Neural Computation, 8(9):1735–1780, 1997.
Krueger, D and Mimesevic, R. Regularizing RNNs by
Stabilizing Activations, September 2015. URL http:
//arxiv.org/abs/1511.08400.
Langley, P. Crafting papers on machine learning. In Langley, Pat (ed.), Proceedings of the 17th International Conference on Machine Learning (ICML 2000), pp. 1207–
1216, Stanford, CA, 2000. Morgan Kaufmann.
Le, Quoc, Jaitly, N, and Hinton, G. A Simple Way to
Initialize Recurrent Networks of Rectified Linear Units,
September 2015. URL http://arxiv.org/abs/
1504.00941.
Martens, J and Sutskever, I. Learning recurrent neural networks with hessian-free optimization. In Proceedings of
the 28th International Conference on Machine Learning,
ICML 2011, Bellevue, Washington, USA, June 28 - July
2, 2011, pp. 1033–1040, 2011.
Mikolov, T, Joulin, A, and Chopra, S et. al. Learning
Longer Memory in Recurrent Neural Networks, September 2015. URL http://arxiv.org/abs/1412.
7753.
Mikolov, Tomáš. Statistical Language Models Based
on Neural Networks.
PhD thesis, 2012.
URL
http://www.fit.vutbr.cz/research/
view_pub.php?id=10158.
Pascanu, Razvan, Mikolov, Tomas, and Bengio, Yoshua.
On the difficulty of training recurrent neural networks.
In Proceedings of the 30th International Conference on
Machine Learning, ICML 2013, Atlanta, GA, USA, 1621 June 2013, pp. 1310–1318, 2013.
Saxe, Andrew M., McClelland, James L., and Ganguli, Surya. Exact solutions to the nonlinear dynamics of learning in deep linear neural networks, 2013.
URL http://arxiv.org/abs/1312.6120. cite
arxiv:1312.6120.
Sutskever, Ilya, Vinyals, Oriol, and Le, Quoc V. Sequence
to sequence learning with neural networks. In Advances
in Neural Information Processing Systems 27: Annual
Conference on Neural Information Processing Systems
2014, December 8-13 2014, Montreal, Quebec, Canada,
pp. 3104–3112, 2014.
White, O, Lee, D, and Sompolinky, H. Short-term memory
in orthogonal neural networks. Physical Review Letters,
92, 2004. ISSN 14.
Zaremba, W, Sutskever, I, and Vinyals, O. Recurrent
Neural Network Regularization, September 2014. URL
http://arxiv.org/abs/1409.2329.
| 9cs.NE
|
MINT: Mutual Information based Transductive
Feature Selection for Genetic Trait Prediction
arXiv:1310.1659v1 [cs.LG] 7 Oct 2013
Dan He1 , Irina Rish1 , David Haws1 , Simon Teyssedre2 , Zivan Karaman2, Laxmi
Parida1
1
2
1
IBM T.J. Watson Research, Yorktown Heights, NY, USA
{dhe, rish, dhaws, parida}@us.ibm.com
Limagrain Europe, Chappes Research Center, CS 3911, 63720 Chappes, France
{simon.teyssedre, zivan.karaman}@limagrain.com
Introduction
Whole genome prediction of complex phenotypic traits using high-density genotyping
arrays has recently attracted a lot of attention, as it is relevant for the fields of plant and
animal breeding and genetic epidemiology. Given a set of biallelic molecular markers,
such as SNPs, with genotype values encoded as 0, 1, 2 on a collection of plant, animal
or human samples, the goal is to predict the values of certain traits, usually highly
polygenic and quantitative, by modeling simultaneously all marker effects, unlike the
traditional GWAS. rrBLUP [9,14] has been used widely for trait prediction where it
builds a linear model by fitting all the genotypes and the coefficient computed for each
marker can be considered as a measure of the importance of the marker. The underlying
hypothesis of normal distribution of marker effects is well suited for highly polygenic
traits, and as the computations are fast and robust, it is one of the most used models
in whole genome prediction. Other popular predictive models are Elastic-Net, Lasso,
Ridge Regression[13,2] , Bayes A, Bayes B[9], Bayes Cπ[7], and Bayesian Lasso [8,10],
etc.
As the number of genotypes is generally much bigger than the number of samples, the predictive models suffer from the “curse of dimensionality”. The “curse of
dimensionality” problem not only affects the computational efficiency of the learning
algorithms, but can also lead to poor performance, mainly because of the correlation
among markers. Feature selection [15,4,6] has been considered as a successful solution
for this problem, where a subset of important features are selected and the predictive
models are trained only on these features. A popular criterion for feature selection is
called Max-Relevance and Min-Redundancy (MRMR) [11] where the selected features
are maximally relevant to the class value and simultaneously minimally dependent on
each other. The method mRMR [11] has been proposed which greedily selects features
that maximize the relevance while minimize the redundancy. mRMR has been applied
successfully on various applications [5,1,16].
Transductive learning, first introduced by Vapnik [3], assumes test data for predictor
variables (here markers) are available to the learning algorithms (the target variable
values for test samples are of course unknown). Therefore the models are built on both
the training and test data, and usually lead to better predictive performance on the
test data. In this work, we proposed a transductive feature selection method MINT
based on information theory. MINT applies the MRMR criterion and integrates the
test data in a natural way in the feature selection process. A dynamic programming
2
algorithm is developed to speed up the selection process. Our experiments on both
simulated and real data show that MINT generally achieves similar or better results
than mRMR does which relies on training data only. To our knowledge, this is the first
transductive feature selection method based on the MRMR criterion.
2
Methods
A popular criterion for feature selection is Max-Relevance and Min-Redundancy (MRMR)
[11]. Max-Relevance searches for features satisfying the Equation 1, which measures the
mean value of all mutual information values between individual feature xi and class
variable c.
1 X
max D(S, c), D =
I(xi ; c)
(1)
|S|
xi ∈S
where S are the selected features, and I(xi ; c) is the mutual information between xi
and c.
However, feature selection just based on max-relevance tends to select features
that have high redundancy, namely the correlations of the selected features tend to
be big. If we remove some of the features that are highly correlated with other features, the respective class-discriminative power would not change much. Therefore,
Min-Redundancy is proposed to select mutually exclusive features:
X
1
I(xi , xj )
(2)
min R(S), R =
|S|2
xi ,xj ∈S
An operator Φ(D, R) is defined to combine D and R from the above two equations
where D and R are optimized at the same time:
max Φ(D, R), Φ = D − R
(3)
In this work, based on the MRMR criterion, we proposed a novel method MINT
(Mutual INformation based Transductive feature selection), which targets feature
selection with both the training data and the unlabeled test data. We developed a
dynamic programming based greedy algorithm to efficiently select features.
We observe that the MRMR criterion has two components, one for maximum relevance and one for minimum redundancy and that the two components are independent.
Maximum relevance requires calculation of the mutual information between the selected
features and the target variable. As in transductive learning, the target variable values
of the test samples are not available, this component remains untouched. Minimum
redundancy, on the other hand, calculates the mutual information among all the selected features and the target variable values are not involved. Therefore we can make
the method transductive by including all the test samples in this component to help
improve the estimation of mutual information.
We applied the same incremental search strategy used in [11] to effectively find
the near-optimal features defined by Φ() in equation 3. The incremental algorithm
works as the following: Assume feature set Sm−1 is already generated and contains
m − 1 features. The m-th feature needs to be selected from the set X − Sm−1 , which
maximizes the following objective function:
3
h
; ctraining ) −
maxxj ∈X−Sm−1 I(xtraining
j
1
m−1
X
I(xjtraining+test ; xitraining+test )
i
(4)
xi ∈Sm−1
where xtraining
denotes the j-th feature vector including only the training data, xtraining+test
j
j
denotes the j-th feature vector including both the training and test data, ctraining denotes the class value vector including only the training data, I(xi , xj ) is the mutual
information between xi and xj .
We next propose an efficient greedy algorithm to incrementally select the features
P based on a dynamic programming strategy. Our motivation is that the operation
xi ∈Sm−1 I(xj ; xi ) (for simplicity, we ignore the superscripts of “training” and “test”)
need not be re-conducted for every xj . Since the features are added in an incremental
manner, the differences between Sm−1 and Sm−2 is just the (m − 1)-th feature. Therefore, we do not need to re-compute the sum
between xj and
P
P of the mutual information
xi where 1 ≤ i ≤ m − 2. The two sums xi ∈Sm−2 I(xj ; xi ) and xi ∈Sm−1 I(xj ; xi ) are
P
just different by I(xj ; xm−1 ). Therefore, we can save this sum xi ∈Sm−1 I(xj ; xi ) at
every step and reuse them in the next step. The complexity of this dynamic programming algorithm is O(N M ), where N is the number of selected features and M is the
number of total features.
3
Experimental Results
We compare the predictive performance of rrBLUP [9,14] on the full set of variables versus its performance on the subsets of variables of different size selected by mRMR [11]
and MINT, referred to as ”mRMR + rrBLUP” and ”MINT + rrBLUP”, respectively.
Similar results, not included due to space restrictions, were obtained when applying
some other predictive methods to features selected by mRMR and MINT. In all experiments, we perform 10-fold cross-validations and measure the average coefficient of
determination r2 (computed as the square of Pearson’s correlation coefficient) between
the true and predicted outputs; higher r2 indicates better performance.
Simulated Data. As our method is based on the MRMR criterion, we experiment
with different levels of relevance and redundancy, and show that the performance of
MINT relies on both components. We randomly simulate 10 different data sets for each
parameter settings, and report average results. First, we simulate a 200-dimensional
target variable vector Y following multivariate uniform distribution Y ∼ U (0, 1), and
then we simulate the features as F = Y + e, where e ∼ N (0, δ 2 ) is a 200-dimensional
noise vector following a multivariate normal distribution. Thus, the large-δ features are
noisy (“bad”), while lower-δ features are less noisy (“good”). We simulate 100 “good”
features with e ∼ N (0, 100) and 1900 “bad” features with e′ ∼ N (0, 1000). The results
are shown in Table 1, Case one. In this case the feature selection methods work well,
but as the good features are randomly simulated and they have low redundancy, the
performances of mRMR+rrBLUP and MINT+rrBLUP are almost identical.
Next, we again simulate a target variable vector Y ∼ U (0, 1). For the design matrix
X, we simulate 50 “seed” features F = Y + e with e ∼ N (0, 500). Then for each
seed feature, we simulate 9 “duplicate” features as F ′ = F + e′ where e′ ∼ N (0, 100).
We consider all these 500 features as “good” features. We also simulate 4500 “bad”
features F ′′ = Y + e′′ with e′′ ∼ N (0, 1000). Therefore the good and bad features
4
are still relatively easy to be distinguished and there are large redundancies among the
good features. The results are shown in Table 1, Case two. MINT+rrBLUP consistently
outperforms mRMR+rrBLUP, due to the redundancy we introduced in the good feature
set; both methods outperform rrBLUP.
Table 1. Performance (average r2 over 10-fold CV) of rrBLUP on the full set of features
vs. MINT+rrBLUP and mRMR+rrBLUP, on simulated data for two different cases.
Case
rrBLUP
Number of
MINT
mRMR
(all features) selected features + rrBLUP +rrBLUP
50
0.940
0.938
One
0.845
100
0.958
0.962
150
0.281
0.280
250
0.376
0.363
Two
0.187
350
0.434
0.411
450
0.432
0.414
550
0.456
0.448
Real data. Next, we compare the same methods on the two Maize data sets Dent
and Flint [12] and show the results in Tables 2. Each data set, Dent and Flint, has
three phenotypes, thus we have six phenotypes overall. Dent has 216 samples and
30,027 features and Flint has 216 samples and 29,094 features. We vary the number of
selected features as 100, 200, 300, 400 and 500. It is obvious that both mRMR+rrBLUP
and MINT+rrBLUP outperform rrBLUP significantly, indicating feature selection in
general is able to improve the performance of the predictive model. On the other
hand, in almost all data sets, MINT outperforms mRMR consistently, illustrating the
effectiveness of transduction.
Table 2. Performance (average r2 over 10-fold CV) of rrBLUP (all features) vs.
mRMR+rrBLUP and MINT+rrBLUP on Maize (Dent and Flint) and n is the number
of selected features.
Data Set (Dent)
rrBLUP (all features) Heritability n=100 n=200 n=300 n=400 n=500
Pheno 1 (mRMR+rrBLUP)
0.426 0.514 0.526 0.563 0.536
0.439
0.952
Pheno 1 (MINT+rrBLUP)
0.623 0.662 0.668 0.653 0.663
Pheno 2 (mRMR+rrBLUP)
0.584 0.591 0.603 0.619 0.629
0.410
0.932
Pheno 2 (MINT+rrBLUP)
0.687 0.674 0.678 0.667 0.669
Pheno 3 (mRMR+rrBLUP)
0.502 0.523 0.521 0.515 0.498
0.228
0.791
Pheno 3 (MINT+rrBLUP)
0.517 0.515 0.514 0.536 0.537
Data Set (Flint)
rrBLUP (all features) Heritability n=100 n=200 n=300 n=400 n=500
Pheno 1 (mRMR+rrBLUP)
0.493 0.472 0.476 0.495 0.508
0.275
0.954
Pheno 1 (MINT+rrBLUP)
0.627 0.599 0.606 0.588 0.601
Pheno 2 (mRMR+rrBLUP)
0.340 0.407 0.411 0.408 0.405
0.255
0.643
Pheno 2 (MINT+rrBLUP)
0.480 0.500 0.498 0.486 0.491
Pheno 3 (mRMR+rrBLUP)
0.199 0.250 0.288 0.275 0.271
0.047
0.355
Pheno 3 (MINT+rrBLUP)
0.320 0.350 0.361 0.352 0.350
4
Conclusions
In this work, we proposed a transductive feature selection method MINT based on
information theory where the test data is integrated in a natural way into a greedy
5
Table 3. Performance of Feature Selection
Data Set MINT+rrBLUP (n=500) mRMR+rrBLUP (n=500) rrBLUP
Data 1
0.14
0.015
0.018
Data 2
0.303
0.110
0.204
Data 3
0.563
0.536
0.607
Data 4
0.369
0.235
0.31
Data 5
0.364
0.201
0.22
Data 6
0.276
0.148
0.17
Lasso Elastic Net
0.018
0.019
0.129
0.146
0.540
0.545
0.255
0.261
0.212
0.214
0.15
0.162
SVR num. of samples
0.079
220
0.234
319
0.615
1217
0.306
671
0.174
532
0.184
620
feature selection process. A dynamic programming algorithm is developed to speed
up the greedy selection. Our experiments on both simulated and real data show that
MINT is generally a better method than the inductive feature selection method mRMR.
What’s more, MINT is not restricted to genetic trait prediction problems but is a
generic feature selection model.
References
1. Y. Cai, J. He, X. Li, K. Feng, L. Lu, K. Feng, X. Kong, and W. Lu. Prediction of protein
subcellular locations with feature selection and analysis. Protein and Peptide Letters,
17(4):464–472, 2010.
2. Scott Shaobing Chen, David L. Donoho, Michael, and A. Saunders. Atomic decomposition
by basis pursuit. SIAM Journal on Scientific Computing, 20:33–61, 1998.
3. A. Gammerman, V. Vovk, and V. Vapnik. Learning by transduction. In Proceedings of the
Fourteenth Conference on Uncertainty in Artificial Intelligence, pages 148–155. Morgan
Kaufmann Publishers Inc., 1998.
4. I. Guyon and A. Elisseeff. An introduction to variable and feature selection. The Journal
of Machine Learning Research, 3:1157–1182, 2003.
5. T. Huang, X.H. Shi, P. Wang, Z. He, K.Y. Feng, L.L. Hu, X. Kong, Y.X. Li, Y.D. Cai,
and K.C. Chou. Analysis and prediction of the metabolic stability of proteins based
on their sequential features, subcellular locations and interaction networks. PLoS One,
5(6):e10972, 2010.
6. A. Jain and D. Zongker. Feature selection: Evaluation, application, and small sample performance. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 19(2):153–
158, 1997.
7. K Kizilkaya, RL Fernando, and DJ Garrick. Genomic prediction of simulated multibreed
and purebred performance using observed fifty thousand single nucleotide polymorphism
genotypes. Journal of animal science, 88(2):544–551, 2010.
8. Andrés Legarra, Christele Robert-Granié, Pascal Croiseau, François Guillaume, Sébastien
Fritz, et al. Improved lasso for genomic selection. Genetics research, 93(1):77, 2011.
9. T.H.E. Meuwissen, B.J. Hayes, and M.E. Goddard. Prediction of total genetic value using
genome-wide dense marker maps. Genetics, 157:1819–1829, 2001.
10. Trevor Park and George Casella. The bayesian lasso. Journal of the American Statistical
Association, 103:681–686, June 2008.
11. H. Peng, F. Long, and C. Ding. Feature selection based on mutual information criteria
of max-dependency, max-relevance, and min-redundancy. Pattern Analysis and Machine
Intelligence, IEEE Transactions on, 27(8):1226–1238, 2005.
12. Renaud Rincent, Denis Laloë, Stéphane Nicolas, Thomas Altmann, Dominique Brunel,
Pedro Revilla, Victor M Rodriguez, J Moreno-Gonzalez, A Melchinger, Eva Bauer, et al.
6
13.
14.
15.
16.
Maximizing the reliability of genomic selection by optimizing the calibration set of reference individuals: Comparison of methods in two diverse groups of maize inbreds (zea
mays l.). Genetics, 192(2):715–728, 2012.
Robert Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal
Statistical Society, Series B, 58:267–288, 1994.
J.C. Whittaker, R. Thompson, and M.C. Denham. Marker-assisted selection using ridge
regression. Genet Res, 75:249–252, 2000.
Y. Yang and J.O. Pedersen. A comparative study on feature selection in text categorization. In Machine Learning-International Workshop then Conference-, pages 412–420.
Morgan Kaufmann Publishers, INC., 1997.
Y. Zhang, C. Ding, and T. Li. Gene selection algorithm by combining relieff and mrmr.
BMC genomics, 9(Suppl 2):S27, 2008.
| 5cs.CE
|
Bayesian Network Learning via Topological Order
Young Woong Park
∗1
and Diego Klabjan
†2
1
2
College of Business, Iowa State University, Ames, IA 50011, USA
Department of Industrial Engineering and Management Sciences, Northwestern
University, Evanston, IL 60208, USA
arXiv:1701.05654v2 [stat.ML] 20 Aug 2017
August 22, 2017
Abstract
We propose a mixed integer programming (MIP) model and iterative algorithms based on topological orders to solve optimization problems with acyclic constraints on a directed graph. The proposed
MIP model has a significantly lower number of constraints compared to popular MIP models based
on cycle elimination constraints and triangular inequalities. The proposed iterative algorithms use
gradient descent and iterative reordering approaches, respectively, for searching topological orders.
A computational experiment is presented for the Gaussian Bayesian network learning problem, an
optimization problem minimizing the sum of squared errors of regression models with L1 penalty over
a feature network with application of gene network inference in bioinformatics.
1
Introduction
Directed graph G is a directed acyclic graph (DAG) or acyclic digraph if G does not contain a directed
cycle. In this paper, we consider a generic optimization problem over a directed graph with acyclic
constraints, which require the selected subgraph to be a DAG.
Let us consider a complete digraph G. Let m be the number of nodes in digraph G, Y ∈ Rm×m a
decision variable matrix associated with the arcs, where Yjk is related to arc (j, k), supp(Y ) ∈ {0, 1}m×m
the 0-1 (adjacency) matrix with supp(Y )jk = 1 if Yjk 6= 0, supp(Y )jk = 0 otherwise, G(supp(Y )) the
sub-graph of G defined by supp(Y ), and let A be the collection of all acyclic subgraphs of G. Then, we
can write the optimization problem with acyclic constraints as
G(supp(Y )) ∈ A,
min F (Y ) s.t.
Y
(1)
where F is a function of Y .
Acyclic constraints (or DAG constraints) appear in many network structured problems. The maximum
acyclic subgraph problem (MAS) is to find a subgraph of G with maximum cardinality while the subgraph
satisfies acyclic constraints. MAS can be written in the form of (1) with F (Y ) = −ksupp(Y )k0 . Although
exact algorithms were proposed for a superclass of cubic graphs [10] and for general directed graphs [22],
most of the works have focused on approximations [9, 20] or inapproximability [18] of either MAS or the
minimum feedback arc set problem (FAS). FAS of a directed graph G is a subgraph of G that creates a
DAG when the arcs in the feedback arc set are removed from G. Note that MAS is closely related to FAS
and is dual to the minimum FAS. Finding a feedback arc set with minimum cardinality is N P-complete
in general [24]. However, minimum FAS is solvable in polynomial time for some special graphs such
as planar graphs [28] and reducible flow graphs [33], and a polynomial time approximation scheme was
developed for a special case of minimum FAS, where exactly one arc exists between any two nodes (called
tournament) [25].
DAGs are also extensively studied in Bayesian network learning. Given observational data with m
features, the goal is to find the true unknown underlying network of the nodes (features) while the selected
arcs (dependency relationship between features) do not create a cycle. In the literature, approaches are
∗ [email protected]
† [email protected]
1
classified into three categories: (i) score-based approaches that try to optimize a score function defined
to measure fitness, (ii) constraint-based approaches that test conditional independence to check existence
of arcs between nodes (iii) and hybrid approaches that use both constraint and score-based approaches.
Although there are many approaches based on the constraint-based or hybrid approaches, our focus is
solving (1) by means of score-based approaches. For a detailed discussion of constraint-based and hybrid
approaches and models for undirected graphs, the reader is referred to Aragam and Zhou [1] and Han et
al. [19].
For estimating the true network structure by a score-based approach, various functions have been
used as different functions give different solutions and behave differently. Many works focus on penalized
least squares, where penalty is used to obtain sparse solutions. Popular choices of the penalty term
include BIC [26], L0 -penalty [5, 37], L1 -penalty [19], and concave penalty [1]. Lam and Bacchus [26] use
minimum-description length as a score function, which is equivalent to BIC. Chickering [5] proposes a
two-phase greedy algorithm, called greedy equivalence search, with the L1 norm penalty. Van de Geer
and Bühlmann [37] study the properties of the L0 norm penalty and show positive aspects of using L0
regularization. Raskutti and Uhler [34] use a variant of the L0 norm. They use cardinality of the selected
subgraph as the score function where the subgraphs not satisfying the Markov assumption are penalized
with a very large penalty. Aragam and Zhou [1] introduce a generalized penalty, which includes the
concave penalty, and develop a coordinate descent algorithm. Han et al. [19] use the L1 norm penalty
and propose a Tabu search based greedy algorithm for reduced arc sets by neighborhood selection in the
pre-processing step.
With any choice of a score function, optimizing the score function is computationally challenging,
because the number of possible DAGs of G grows super exponentially in the number of nodes m and
learning Bayesian networks is also shown to be N P-complete [4]. Many heuristic algorithms have been
developed based on greedy hill climbing [5, 21, 19] or coordinate descent [14], or enumeration [34] when the
score function itself is the main focus. There also exist exact solution approaches based on mathematical
programming. One of the natural approaches is based on cycle prevention constraints, which are reviewed
in Section 2. The model is covered in Han et al. [19] as a benchmark for their algorithm, but the MIP
based approach does not scale; computational time increases drastically as data size increases and the
underlying algorithm cannot solve larger instances. Baharev et al. [2] studied MIP models for minimum
FAS based on triangle inequalities and set covering models. Several works have been focused on the
polyhedral study of the acyclic subgraph polytopes [3, 15, 17, 27]. In general, MIP models have gotten
relatively less attention due to the scalability issue.
In this paper, we propose an MIP model and iterative algorithms based on the following well-known
property of DAGs [6].
Property 1. A directed graph is a DAG if and only if it has a topological order.
A topological order or topological sort of a DAG is a linear ordering of all of the nodes in the graph
such that the graph contains arc (u, v) if and only if u appears before v in the order [7]. Suppose that Z
is the adjacency matrix of an acyclic graph. Then, by sorting the nodes of acyclic graph G(Z) based on
the topological order, we can create a lower triangular matrix from Z, where row and column indices of
the lower triangular matrix are in the topological order. Then, any arc in the lower triangular matrix can
be used without creating a cycle. By considering all arcs in the lower triangular matrix, we can optimize
F in (1) without worrying to create a cycle. This is an advantage compared to arc-based search, where
acyclicity needs to be examined whenever an arc is added. Although the search space of topological
orders is very large, a smart search strategy for a topological order may lead to a better algorithm than
the existing arc-based search methods. Node orderings are used for Bayesian Network learnings based on
Markov chain Monte Carlo methods [8, 13, 30] as alternatives to network structure based approaches.
The proposed MIP assigns node orders to all nodes and add constraints to satisfy Property 1. The
iterative algorithms search over the topological order space by moving from one topological order to
another order. The first algorithm uses the gradient to find a better topological order and the second
algorithm uses historical choice of arcs to define the score of the nodes.
With the proposed MIP model and algorithms for (1), we consider a Gaussian Bayesian network
learning problem with L1 penalty for sparsity, which is discussed in detail in Section 4. Out of many
possible models in the literature, we pick the L1 -penalized least square model from recently published
2
work of Han et al. [19], which solves the problem using a Tabu search based greedy algorithm. The
algorithm is one of the latest algorithms based on arc search and is shown to be scalable when m is
large. Further, their score function, L1 penalized least squares, is convex and can be solved by standard
mathematical optimization packages. Hence, we select the score function from Han et al. [19] and use
their algorithm as benchmark. In the computational experiment, we compare the performance of the
proposed MIP model and algorithms against the algorithm in Han et al. [19] and other available MIP
models for synthetic and real instances.
Our contributions are summarized in the following.
1. We consider a general optimization problem with acyclic constraints and propose an MIP model
and iterative algorithms for the problem based on the notion of topological orders.
2. The proposed MIP model has significantly less constraints than the other MIP models in the literature, while maintaining the same order of the number of variables. The computational experiment
shows that the proposed MIP model outperforms the other MIP models when the subgraph is
sparse.
3. The iterative algorithms based on topological orders outperform when the subgraph is dense. They
are more scalable than the benchmark algorithm of Han et al. [19] when the subgraph is dense.
In Section 2, we present the new MIP model along with two MIP models in the literature. In Section 3,
we present two iterative algorithms based on different search strategies for topological orders. The Gaussian Bayesian network learning problem with L1 -penalized least square is introduced and computational
experiment are presented in Sections 4 and 5, respectively.
In the rest of the paper, we use the following notation.
J = {1, · · · , m} = index set of the nodes
J k = J \ {k} = index set of the nodes excluding node k, k ∈ J
Z = supp(Y ) ∈ {0, 1}m×m
π = topological order
Given π, we define πk = q to denote that the order of node k is q. For example, given three nodes a, b, c,
and topological order b − c − a, we have πa = 3, πb = 1, and πc = 2. With this notation, if πj > πk , then
we can add an arc from j to k.
2
MIP Formulations based on Topological Order
In this section, we present three MIP models for (1). The first and second models, denoted as MIPcp and
MIPin, respectively, are models in the literature for similar problems with acyclic constraints. The third
model, denoted as MIPto, is the new model we propose based on Property 1.
A popular mathematical programming based approach for solving (1) is the cutting plane algorithm,
which is well-known for the traveling salesman problem formulation. Let C be the set of all possible cycles
and Cl ∈ C be the set of the arcs defining a cycle. Let H(supp(Y ), Cl ) be a function that counts the
number of selected arcs in supp(Y ) from Cl . Then, (1) can be solved by
MIPcp
min F (Y ) s.t.
Y
H(supp(Y ), Cl ) ≤ |Cl | − 1, Cl ∈ C,
(2)
which can be formulated as an MIP. Note that (2) has exponentially many constraints due to the cardinality of C. Therefore, it is not practical to pass all cycles in C to a solver. Instead, the cutting
plane algorithm starts with an empty active cycle set C A and iteratively adds cycles to C A . That is, the
algorithm iteratively solves
min F (Y ) s.t.
Y
H(supp(Y ), Cl ) ≤ |Cl | − 1, Cl ∈ C A ,
(3)
with the current active set C A , detects cycles from the solution, and adds the cycles to C A . The algorithm
terminates when there is no cycle detected from the solution of (3). One of the drawbacks of the cutting
3
plane algorithm based on (3) is that in the worst case we can add all exponentially many constraints.
In fact, Han et al. [19] study the same model and concluded that the cutting plane algorithm does not
scale.
Baharev et al. [2] recently presented MIP models for the minimum feedback arc set problem based
on linear ordering and triangular inequalities, where the acyclic constraints presented were previously
used for cutting plane algorithms for the linear ordering problem [16, 29]. For any F , we can write the
following MIP model based on triangular inequalities presented in [2, 16, 29].
MIPin
min
F (Y )
(4a)
s.t.
Z = supp(Y ),
(4b)
Zqj + Zjk − Zqk ≤ 1,
1 ≤ q < j < k ≤ m,
(4c)
− Zqj − Zjk + Zqk ≤ 0,
1 ≤ q < j < k ≤ m,
(4d)
Zjk ∈ {0, 1},
1≤j<k≤m
(4e)
Note that Zjk is not defined for all j ∈ J k and k ∈ J. Instead of having a full matrix of binary variables,
the formulation only uses lower triangle of the matrix using the fact that Zjk + Zkj = 1. We can also
use this technique to any of the MIP models presented in this paper. However, for ease of explanation,
we will use the full matrix, while the computational experiment is done with the reduced number of
binary variables. Therefore, the cutting plane algorithm with MIPcp should be more scalable than the
implementation in Han et al. [19], which has twice more binary variables.
Baharev et al. [2] also provides a set covering based MIP formulation. The idea is similar to MIPcp.
In the set covering formulation, each row and column represents a cycle and an arc, respectively. Similar
to MIPcp, existence of exponentially many cycles is a drawback of the formulation and Baharev et al. [2]
use the cutting plane algorithm.
Next, we propose an MIP model based on Property 1. Although MIPin uses significantly less constraints than MIPcp, MIPin still has O(m3 ) constraints which grows rapidly in m. On the other hand,
the MIP model we propose has O(m2 ) variables and O(m2 ) constraints. In addition to Z, let us define
decision variable matrix O ∈ {0, 1}m×m .
1 if πk = q,
Okq =
k ∈ J, q ∈ J
0 otherwise,
Then, we have the following MIP model.
MIPto
min F (Y )
s.t.
(5a)
Z = supp(Y ),
Zjk − mZkj ≤
(5b)
X
r(Okr − Ojr ),
j ∈ J k , k ∈ J,
(5c)
j ∈ J k , k ∈ J,
(5d)
k ∈ J,
(5e)
q ∈ J,
(5f)
r∈J
Zjk + Zkj ≤ 1,
X
Okq = 1,
q∈J k
X
Okq = 1,
k∈J q
Z, O ∈ {0, 1}m×m , Y unrestricted
(5g)
The key constraint in (5) is (5c). Recall that Zjk indicates which node comes first in the topological
order and Okr stores the exact location in the order. With these definitions, (5c) forces correct values of
Zjk and Zkj by comparing the order difference. Recall that we can reduce the number of binary variables
4
Zjk ’s by plugging Zjk + Zkj = 1, but we keep the full matrix notation for ease of explanation. We next
show that (5) correctly solves (1).
Proposition 1. An optimal solution to (5) is an optimal solution to (1).
Proof. By Property 1, any DAG has a corresponding topological order. Let π ∗ be the topological order
defined by an optimal solution Y ∗ for (1). Note that (5e) and (5f) define a topological order. Hence,
it suffices to show that (5) gives a DAG given π ∗ . Note that the right hand side of (5c) measures the
difference in the topological order between nodes j and k. If the value is positive,
it implies πk > πj .
P
Consider (5c) for j1 and j2 with πj∗2 > πj∗1 . When j = j1 and k = j2 , we have r∈J r(Oj2 r − Oj1 r ) > 0
in
P (5c) and at most one of zj1 j2 and zj2 j1 can be 1 by (5d). When j = j2 and k = j1 , we have
r∈J r(Oj1 r − Oj2 r ) < 0 in (5c) and we must have zj1 j2 = 1 by the left hand side of (5c). Therefore, we
have correct value zj1 j2 = 1 when πj∗2 > πj∗1 . This completes the proof.
In Table 1, we compare the MIP models introduced in this section. Although all three MIP models
have O(m2 ) binary variables, MIPto has more binary variables than MIPcp and MIPin due to Okq ’s.
MIP models MIPin and MIPto have polynomially many constraints, whereas MIPcp has exponentially
many constraints. MIPto has the smallest number of constraints among the three MIP models. In
the computational experiment, we use a variation of the cutting plane algorithm for MIPcp as it has
exponentially many constraints. For MIPin and MIPto, we do not use a cutting plane algorithm.
Name
MIPcp
MIPin
MIPto
Reference
(2)
(4)
(5)
# binary variables
O(m2 )
O(m2 )
O(m2 )
# constraints
exponential
O(m3 )
O(m2 )
Table 1: Number of binary variables and constraints of MIP models
3
Algorithms based on Topological Order
Although the MIP models introduced in Section 2 guarantee optimality, the execution time for solving
an integer programming problem can be exponential in problem size. Further, the execution time could
increase drastically in m, as all of the models require at least O(m2 ) binary variables and O(m2 ) constraints. In order to deal with larger graphs, we propose iterative algorithms for (1) based on Property 1.
Observe that, if we are given a topological order of the nodes, then Z and O are automatically determined
in (5). In other words, we can easily obtain a subset of the arcs such that all of the arcs can be used
without creating a cycle. Let R̄ be the determined adjacency matrix given topological order π̄. In detail,
we set
R̄jk = 1 if π̄j > π̄k , R̄jk = 0 otherwise.
Let adj(π̄) be the function generating R̄ given input topological order π̄. If we are given π̄, then we can
generate R̄ by adj(π̄), and solving (1) can be written as
min F (Y ) s.t. R̄ ≥ supp(Y ).
Y
(6)
Note that (6) has acyclic constraint R̄ ≥ supp(Y ), not R̄ = supp(Y ). The inequality is needed when we
try to obtain a sparse solution, i.e., only a subset of the arcs is selected among all possible arcs implied
by R̄. As long as we satisfy the inequality, Y forms an acyclic subgraph. Hence, R̄ can be different
from adjacency matrix supp(Y ) in an optimal solution of (6), and any arc (j, k) such that R̄jk = 1
can be selected without creating a cycle. For this reason, we call R̄ an adjacency candidate matrix.
The algorithms proposed later in this section solve (6) by providing different π̄ and R̄ = adj(π̄) in each
iteration. In fact, (6) is separable into m sub problems if F is separable. Let Yk and Zk be the k th
columns of Y and R, respectively, for node k. Then, solving
min Fk (Yk ) s.t.
Yk
5
R̄k ≥ supp(Yk ),
(7)
P
for all k ∈ J gives the same solution as solving (6) if F is separable as F (Y ) = k∈J Fk (Yk ).
In Section 3.1, a local improvement algorithm for a given topological order is presented. The algorithm
swaps pairs of nodes in the order. In both of the iterative algorithms proposed in Sections 3.2 and 3.3,
we use the local improvement algorithm presented in the following section.
3.1
Topological Order Swapping Algorithm
Algorithm 1 tries to improve the solution by swapping the topological order. In each iteration, the
algorithm determines the nodes to swap that have order s1 and s2 in Line 3, where s2 = s1 + 1 implies
that we select two nodes which are neighbors in the current topological order. Then in Line 4, the actual
node indices k1 and k2 such that πk1 = s1 and πk2 = s2 are detected. The condition in Line 5 is to
avoid meaningless computation when Y ∗ is sparse. If |Yk∗2 k1 | > 0, we know for sure that Yk∗2 k1 = 0 after
swapping the orders of k1 and k2 and thus we will get a different solution. However, if Yk∗2 k1 = 0, we
will still have Yk∗2 k1 = 0 after the swap forced by the new order. In Line 6, we create a new topological
order π̄ by swapping nodes k1 and k2 in π ∗ . After obtaining adjacency candidate matrix R̄ in Line 7, we
solve (6) with R̄. It is worth noting that, if F is separable, we only need to solve (7) with k = k1 and k2
because the values of R̄ are the same with R∗ except for k1 and k2 as the order difference was 1 in π ∗ . In
Line 9, we update the best solution if the new solution is better. The iterations continue until there is no
improvement in the past m iterations, which implies that we would swap the same nodes if we proceed
after this iteration. Algorithm 1 is illustrated by the following toy example.
Algorithm 1 TOSA (Topological Order Swapping Algorithm)
Input: Y 0 , R0 , π 0
Output: Best solution Y ∗ , R∗ , π ∗
1: (Y ∗ , R∗ , π ∗ ) ← (Y 0 , R0 , π 0 ), t ← 0
2: While there is an improvement in the past m iterations
3:
t ← t + 1, s1 ← (t mod (m − 1)) + 1, s2 ← s1 + 1
4:
(k1 , k2 ) ← node indices satisfying πk1 = s1 and πk2 = s2
5:
If |Yk∗2 k1 | > 0 then
6:
π̄ ← π ∗ , π̄k1 = s2 , π̄k2 = s1
7:
R̄ ← adj(π̄)
8:
Ȳ ← solve (6) with R̄
9:
If F (Ȳ ) < F (Y ∗ ) then update (Y ∗ , R∗ , π ∗ )
10:
End if
11: End While
Example. Consider a graph with m =
corresponding order 3 − 1 − 2 − 4,
0
0
0
0
Y0 =
0
0
0.4 0.8
4 nodes. Let us assume that inputs are π 0 = (2, 3, 1, 4) with
0.5 0
0
0.5 0
, and R0 = 0
0
0 0
0.1 0
1
0
0
0
1
1
1
0
1
0
0
.
0
0
In iteration 1, t = 1, s1 = 1, s2 = 2, k1 = 3, and k2 = 1. Hence, we are swapping nodes 3 and 1. Since
∗
|Y13
| = 0.5 > 0, π̄ = (1, 3, 2, 4) is created in Line 6, where the associated order is 1 − 3 − 2 − 4. If π̄
gives an improved objective function value, then π ∗ is updated in Line 9. Let us assume that π ∗ is not
∗
| = 0, Lines 6 - 9 are not
updated. In iteration 2, t = 2, s1 = 2, s2 = 3, k1 = 1, and k2 = 2. Since |Y21
executed.
3.2
Iterative Reorering Algorithm
We propose an iterative reordering algorithm based on Property 1, which solves (6) in each iteration
aiming to optimize (1). In each iteration of the algorithm, all nodes are sorted based on scores defined
6
by (i) merit scores of the arcs, (ii) historical choice of the arcs (used as weights), (iii) and some random
components. Then the sorted node order is directly used as a topological order. The selected arcs by the
topological order give updates on arc weights. Let us first define notation.
ν = uniform random variable on [νlb , νub ], νlb < 1 < νub
ρjk = pre-determined merit score of arc (j, k) for j ∈ J, k ∈ J
wjk = weight of arc (j, k) for j ∈ J, k ∈ J
ck = score of node k, k ∈ J
The range [νlb , νub ] of the uniform random variable ν balances the randomness and structured scores.
Note that ρjk should be determined based on the data and the characteristic of the problem considered,
where larger ρjk implies that arc (j, k) is attractive. Based on the arc merit scores ρ, the score for node
k, k ∈ J, is defined as
X
wjk ρjk , k ∈ J,
ck = ν ·
(8)
j∈J k
which can be interpreted as a weighted summation of ρjk ’s multiplied by perturbation random number
ν. Hence, nodes with high scores are attractive. Initially, all arcs have equal weights and the weights are
updated in each iteration based on the topological order in the iteration. If R̄ = adj(π̄) is the adjacency
candidate matrix in the iteration, then, the weights are updated by
wjk = wjk + 1,
if R̄jk = 1.
(9)
The overall algorithmic framework is summarized in Algorithm 2. In Line 1, weights wjk ’s are initialized to 1 and t̄, which counts the number of iterations without a best solution update, is initialized. Also,
a random order π ∗ of the nodes is generated, and the corresponding solution becomes the best solution.
In each iteration, first node scores ck ’s are calculated (Line 3), then topological order π̄ is obtained by
sorting the nodes, and finally adjacency candidate matrix R̄ is generated (Line 4). Then, in Lines 5 and
6, solution Ȳ is obtained by solving (6) with R̄ and the best solution is updated if available. In Lines 7
- 10, TOSA is executed if the current solution is within a certain percentage α from the best solution.
Lines 11 and 12 update t̄, and Line 13 updates wjk ’s. This ends the iteration and the algorithm continues
until π̄ is converged or there is no update of the best solution in the last t∗ iterations. Algorithm 2 is
illustrated by the following toy example.
Algorithm 2 IR (Iterative Reordering)
Input: Merit score ρ ∈ Rm×m , termination parameter t∗ , TOSA execution parameter α
Output: Best solution Y ∗ , R∗ , π ∗
1: wjk ← 1, π ∗ ← a random order, R∗ ← adj(π ∗ ), π̄ ← π ∗ , Y ∗ ← solve (6) with R∗ , t̄ ← 0
2: While (i) π̄ is not convergent or (ii) t̄ < t∗
3:
Calculate score ck by (8)
4:
π̄ ← sort nodes with respect to ck , R̄ ← adj(π̄)
5:
Ȳ ← solve (6) with R̄
6:
If F (Ȳ ) < F (Y ∗ ) then update (Y ∗ , R∗ , π ∗ )
7:
If F (Ȳ ) < F (Y ∗ ) · (1 + α)
8:
(Y 0 , R0 , π 0 ) ← TOSA(Ȳ , R̄, π̄),
9:
If F (Y 0 ) < F (Y ∗ ) then update (Y ∗ , R∗ , π ∗ )
10:
End If
11:
If (Y ∗ , R∗ , π ∗ ) is updated then t̄ ← 0
12:
Else t̄ ← t̄ + 1
13:
Update weights by (9)
14: End While
Example. Consider a graph with m = 3 nodes. In the current iteration, let us assume that we are given
7
0.5
0 1 2
0.2 and w = 1 0 1 .
0
2 1 0
P
P
Note
P that we have j∈J 1 wj1 ρj1 = 0.2 · 1 + 0.3 · 2 = 0.8, j∈J 2 wj2 ρj2 = 0.5 · 1 + 0.3 · 1 = 0.8, and
j∈J 3 wj3 ρj3 = 0.4 · 2 + 0.2 · 1 = 1. If random numbers (ν) are 0.9, 1.1, 0.8 for nodes 1,2, and 3,
respectively, then by (8), c1 = 0.9 · 0.8 = 0.72, c2 = 1.1 · 0.8 = 0.88, and c3 = 0.8 · 1 = 0.8. Then in
Line 4, we obtain π̄ = (3, 1, 2), with corresponding order 2 − 3 − 1, and R̄ = [0, 1, 1; 0, 0, 0; 0, 1, 0]. After
obtaining Ȳ and updating the best solution in Lines 5-12, the weights are updated by (9) as follows.
0 1 2
0 2 3
wnew = 1 0 1 + R̄ = 1 0 1
2 1 0
2 2 0
0 0.5
ρ = 0.2 0
0.3 0.3
This ends the current iteration.
3.3
Gradient Descent Algorithm
In this section, we propose a gradient descent algorithm based on Property 1. The algorithm iteratively
executes: (i) moving toward an improving direction by gradients, (ii) DAG structure is recovered and
topological order is obtained by a projection step. The algorithm is based on the standard gradient
descent framework while the projection step takes care of the acyclicity constraints by generating a
topological order from the current (possibly cyclic) solution matrix. In order to distinguish the solutions
with and without the acyclicity property, we use the following notation.
U t ∈ Rm×m = decision variable matrix without acyclicity requirement in iteration t
Y t ∈ Rm×m = decision variable matrix satisfying G(supp(Y t )) ∈ A in iteration t
Let γ t be the step size in iteration t, ∇F (Y t ) be the derivative of F at Y t , and Gt ∈ Rm×m be a weight
matrix that weighs each element. We assume k∇F (Y t )k∞ ≤ M1 for a constant M1 , where k · k∞ is the
uniform (infinity) norm. The update formula
U t = Y t − γ t ∇F (Y t ) ◦ Gt , t ≥ 0,
(10)
updates Y t based on the weighted gradient, where ◦ represents the entrywise or Hadamard product of
the two matrices. Given topological order π t , we define Gt as
t
1 πk
,
Gtjk = 1 + t
πk
j ∈ Jk
(11)
to balance gradients of the nodes with different orders (small and large values of πkt ). For nodes k1 and
k2 with πkt 1 = 1 and πkt 2 = m, most of the gradients for node k1 are zero and most of the gradients for
node k2 are nonzero. Weight (11) tries to adjust this gap. Note that we have 2 ≤ Gtjk ≤ e for any large
m. Since U t may not satisfy acyclic constraints, in order to obtain a DAG, the algorithm needs to solve
the projection problem
Y ∗ = argminY kY − U t k22
s.t.
G(supp(Y )) ∈ A,
(12)
where k · k2 is the L2 norm.
Proposition 2. If U t is arbitrary, then optimization problem (12) is N P-hard.
Proof. Recall that feedback arc set is N P-complete [24] and maximum acyclic subgraph is the dual of the
feedback arc set problem. With U t = 1, (12) becomes the weighted maximum acyclic subgraph problem.
Therefore, (12) is N P-complete.
Because solving (12) to optimality does not guarantee an optimal solution for (1), we use a greedy
strategy to solve (12). The greedy algorithm, presented in Algorithm 3, sequentially determines and fixes
the topological order of a node where in each iteration the problem is solved optimally given the currently
8
fixed nodes and corresponding orders. The detailed derivations of the algorithm and the proof that each
iteration is optimal, given already fixed node orders, are available in Appendix A. In other words, we show
that Line 3 is ‘locally’ optimal, i.e., it selects the best next node given that
q + 1, q + 2, · · · , m
P the order
t 2
is fixed. In each iteration, in Line 3, the algorithm first calculates score j∈J¯(Ūjk
) for each node k in J¯
and picks node k ∗ with the minimum value. Then, in Line 4, the order of the selected node is fixed to q.
The fixed node is then excluded from the active set J¯ and iterate q is decreased by 1 in Line 5. At the
end of the algorithm, we can determine Ȳ based on the order π̄ determined and (18) in Appendix A. We
illustrate Algorithm 3 by the following example.
Algorithm 3 Greedy
Input: U t ∈ Rm×m
Output: Ȳ feasible to (12), topological order π̄
1: q ← m, J¯ ← J
2: While J¯ 6= ∅
o
nX
t 2
(Ujk
)
3:
k ∗ = argmink∈J¯
j∈J¯
π̄k∗ = q
J¯ ← J¯ \ {k ∗ }, q ← q − 1
End While
7: Determine Ȳ by (18) in Appendix A
4:
5:
6:
Example. Consider a graph with m = 3
following.
0
Ut = 4
5
nodes. Given U t , the algorithm returns Ȳ presented in the
1
0
2
2
2
0
0
Ȳ = 4
5
0
0
0
0
2
0
Algorithm 3 starts with q = 3 and J¯ = {1, 2, 3}. In iteration 1, node 2 is selected to have π2 = 3 based
on argmin{42 + 52 , 12 + 22 , 22 + 22 }. Then, set J¯ and integer q are updated to J¯ = {1, 3} and q = 2.
In iteration 2, node 3 is selected to have π3 = 2 based on argmin{52 , 22 }. Then, set J¯ and integer q are
updated to J¯ = {1} and q = 1. In iteration 3, node 1 is selected. Hence, we have node order 1-3-2 and
we obtain Ȳ presented above with objective function value kȲ − U t k22 = 12 + 22 + 22 = 9.
The overall gradient descent algorithm for (1) is presented in Algorithm 4. In Line 1, the algorithm
generates a random order π ∗ and obtain corresponding R∗ and Y ∗ and save them as the best solution.
In each iteration of the loop, Lines 3-6 follow the standard gradient descent algorithm. The weighted
gradient H t is calculated in Line 3, and the step size is determined in Line 4 based on the ratio between
t
t
maxj∈J k ,k∈J |Hjk
| and maxj∈J k ,k∈J |Yjk
|. In Line 5, the solution is updated based on the weighted
gradient and, in Line 6, the greedy algorithm is used to obtain the projected solution and the topological
order. Observe that we do not directly use the projected solution. This is because the projected solution
is not necessarily optimal given π t+1 . Hence, in Line 7, a new solution Y t+1 is obtained based on π t+1 .
In Lines 9 - 12, T OSA is executed if the current solution is within a certain percentage from the current
best solution. Lines 13 and 14 update t̄ and Line 15 copies Y ∗ to Y t+1 if t̄ ≥ t∗2 in order to focus on the
solution space near Y ∗ . The algorithm continues until Y t is convergent or t̄ ≥ t∗1 .
In gradient based algorithms, it is common to have γ t depend only on t, but in our case dependency
on H t and Y t is justifiable since we multiply the gradient by Gt . We next show the convergence of Y t
in Algorithm 4 when t∗1 = t∗2 = ∞. This makes the algorithm not to terminate unless Y t has converged
and modification of Y t in Line 15 is not executed. Further, we assume the following for the analysis.
t
t
Assumption 1. For any non-zero element Yjk
, j, k ∈ J, of Y t in any iteration t, we assume ε < |Yjk
|<
M2 , where ε is a small positive number and M2 is a large enough number.
Note that Assumption 1 is a mild assumption, as ignoring near-zero values of Y t happens in practice
anyway due to finite precision. For notational convenience, let Lt = γ t ∇F (Y t ) ◦ Gt = γ t H t be the second
9
Algorithm 4 GD (Gradient Descent)
Input: Parameters t∗1 and t∗2 , TOSA execution parameter α
Output: Best solution Y ∗ , R∗ , π ∗
1: t ← 1, t̄ ← 0, π ∗ ← a random order, R∗ ← adj(π ∗ ), Y ∗ ← solve (6) with R∗
2: While (i) Y t is not convergent or (ii) t̄ < t∗
1
3:
H t ← ∇F (Y t ) ◦ Gt , Gt defined in (11)
t
k∞ √
t
4:
γ t ← kH
kY t k∞
t
t
t t
5:
U ←Y −γ H
6:
π t+1 ← Greedy(U t )
7:
Y t+1 ← solve (6) with Rt+1 = adj(π t+1 )
8:
If F (Y t+1 ) < F (Y ∗ ) then (Y ∗ , R∗ , π ∗ ) ← (Y t+1 , Rt+1 , π t+1 )
9:
If F (Ȳ ) < F (Y ∗ ) · (1 + α)
10:
(Y 0 , R0 , π 0 ) ← T OSA(Y t+1 , Rt+1 , π t ),
11:
If F (Y 0 ) < F (Y ∗ ) then (Y ∗ , R∗ , π ∗ ) ← (Y 0 , R0 , π 0 ), (Y t+1 , Rt+1 , π t+1 ) ← (Y 0 , R0 , π 0 )
12:
End If
13:
If (Y ∗ , R∗ , π ∗ ) is updated then t̄ ← 0
14:
Else t̄ ← t̄ + 1
15:
If t̄ ≥ t∗2 then Y t+1 ← Y ∗
16:
t←t+1
17: End While
term in (10). Then, U t can be written as U t = Y t − Lt = Y t − γ t H t . In the following lemma, we show
that the node orders converge.
√
2
Lemma 1. If t is sufficiently large satisfying t > √ 2(M12e)
, then π t = π t+1 .
ε(
M2 +ε /m−M2 )
Proof. Recall that Y t is obtained by solving (20) and we know the corresponding node order π t . Let
k1 , k2 , · · · , km be the node indices defined based on π t . In other words, node k1 appears first, followed by
nodes k2 , k3 , and so on in the topological order π t . In the proof, we show that there is no change in the
√
2
node order when the condition t > √ 2(M12e)
is met, where M1 and M2 are the upper bounds
ε(
M2 +ε /m−M2 )
for k∇F (Y t )k∞ and kY t k∞ , respectively, as assumed. We first derive
t
t
kL k∞ = γ H
t
∞
1 kH t k∞ t
= √
H
t kY t k∞
∞
2
1 kH t k∞
1 (M1 e)2
√
,
≤√
<
ε
t kY t k∞
t
(13)
where the last inequality holds since (i) kY t k∞ > ε by Assumption 1, (ii) kH t k∞ = k∇F (Y t ) ◦ Gt k∞ ≤
M1 e because k∇F (Y t )k∞ ≤ M1 by the assumption and kGt k∞ ≤ e, where e is natural number.
Now let us consider q = q̄ in Algorithm 3 to decide node order q̄ in iteration t + 1 and assume
πkt r = πkt+1
for r = m, m − 1, · · · , q̄ − 1. Note that we currently have J¯ = {k1 , k2 , · · · , kq̄ }.
r
2 2
P
P
P
t
t
1. For kq̄ , we derive j∈J¯(Ujk
)2 = j∈J¯(Yjk
− Ltjkq̄ )2 = j∈J¯(Ltjkq̄ )2 < m (M√1tεe) , where the
q̄
q̄
t
second equality holds since Yjk
= 0 for all j ∈ J¯ since πkq̄ = q̄ and no arc can be used to the nodes
q̄
¯ and the last inequality holds due to (13) and |J|
¯ ≤ m.
in J,
2. For all other nodes kr ∈ J¯ \ {kq̄ }, we derive
P
P
t
2
=
(Y t r − Ltjkr )2
j∈J¯(Ujkr )
Pj∈J¯ jk
P
P
t
=
(Y t r )2 + j∈J¯(Ltjkr )2 − 2 j∈J¯ Yjk
· Ltjkr
r
Pj∈J¯ jk
P
t
2
t
t
) − 2 j∈J¯ Yjkr · Ljkr
>
j∈J¯(Y
Pjkr
t
> ε2 − 2 j∈J¯ Yjk
· Ltjkr
r
P
t
≥ ε2 − 2 j∈J¯ |Yjk
· Ltjkr |
m
>
2
e)
√1
ε2 − 2M2 m(M
tε
10
t
where the fourth line holds due to |Yjk
| > ε by Assumption 1, and the sixth line holds due to
r
(M1 e)2
t
t
¯
|Y | ≤ M2 , |J| ≤ m, and |L | < √
by (13).
jk
jkr
tε
(M1 e)2 2
P
t
2
√
< ε2 −
Combining the two results for kq̄ and kr ∈ J¯ \ {kq̄ }, we obtain
j∈J¯(Ujkq̄ ) < m
tε
2
P
e)
t
√1
2M2 m(M
< j∈J¯(Ujk
)2 , for any r ∈ {1, 2, · · · , q̄ − 1}, where the second inequality holds due to the
r
tε
√
2
condition t > √ 2(M12e)
. The result implies that we must have πkt+1
= πkt q̄ = q̄ by Line 3 in
q̄
ε(
M2 +ε /m−M2 )
Algorithm 3.
Note that, when q̄ = m, we have J = J¯ and the assumption of πkt r = πkt+1
for r = m, m − 1, · · · , q̄ − 1
r
automatically holds. By iteratively applying the above derivation technique from q = m to 1, we can
show that π t = π t+1 .
When (6) is solved with the identical node orders, the resulting solutions are equivalent. Hence, the
following proposition holds.
Proposition 3. In Algorithm 4, Y t converges in t.
4
Estimation of Gaussian Bayesian Networks
In this section, we introduce the Gaussian Bayesian network learning problem, which follows the form of
(1). The goal is to learn or estimate unknown structure between the nodes of a graph, where the error is
normally distributed. The network can be estimated by optimizing a score function, testing conditional
independence, or a mix of the two, as described in Section 1. Among the three categories, we select the
score based approach with the L1 -penalized least square function recently studied in Han et al. [19].
Let X ∈ Rn×m be a data set with n observations and m features. Let I = {1, · · · , n} and J =
{1, · · · , m} be the index set of observations and features, respectively. For each k ∈ J, we build a
regression model in order to explain feature k using a subset of variables in J k . In other words, we set
feature k as the response variable and a sparse subset of J k as explanatory variables of the regression
model for variable k. In order to obtain a subset of J k , the LASSO penalty function is added. Considering
regression models for all k ∈ J together, the problem can be represented on a graph. Each feature is
a node in the graph, and the directed arc from node j to node k represents explanatory and response
variable relationship between node j and k. The goal is to minimize the sum of penalized SSE over all
regression models for k ∈ J, while the selected arcs do not create a cycle.
Let βjk , j ∈ J k , k ∈ J, be the coefficient of attribute j for dependent variable k. Then the problem
can be written as
XX
X
1 XX
|βjk | s.t. G(supp(β)) ∈ A,
(14)
(xik −
βjk xij )2 + λ
min
β n
k
k
k∈J i∈I
j∈J
k∈J j∈J
which P
follows
Pthe form of (1). In Han et al. [19], individual weights are used for the penalty term,
i.e., λ k∈J j∈J k wjk |βjk |, however, in the computational experiment, we set all weights equal to 1 for
simplicity.
Let Zjk = 1 if attribute j is used for dependent variable k and Zjk = 0 otherwise. Then, we can
formulate MIPto for (14) as
min
X
XX
1 XX
(xik −
βjk xij )2 + λ
|βjk |
n
k
k
k∈J i∈I
j∈J
s.t. |βjk | ≤ M Zjk ,
Zjk − mZkj ≤
X
(15a)
k∈J j∈J
r(Okr − Ojr ),
j ∈ J k , k ∈ J,
(15b)
j ∈ J k , k ∈ J,
(15c)
j ∈ J k , k ∈ J,
(15d)
r∈J
Zjk + Zkj ≤ 1,
11
X
Okq = 1,
k ∈ J,
(15e)
Okq = 1,
q ∈ J,
(15f)
q∈J
X
k∈J
Z, O ∈ {0, 1}m×m ,
(15g)
j ∈ J k ∪ {0}, k ∈ J,
βjk not restricted,
(15h)
where M is a large constant. Note that (15b) is the linear constraint corresponding to Z = supp(Y ) in
(5). Similarly, (15a), (15b), (4c) - (4e) and (15h) can be used to formulate MIPin for (14). For MIPcp,
(15a), (15b), (15h), and the constraints in (2) can be used for (14).
Note that M in (15b) plays an important role in computational efficiency and optimality. If M is
too small, the MIP model cannot guarantee optimality. If M is too large, the solution time can be as
large as enumeration. The algorithm for getting a valid value for M in Park and Klabjan [31] can be
used. However, the valid value of big M for multiple linear regression is often too large [31]. For (15),
we observed that a simple heuristic presented in Section 5 works well.
In each iteration of IR (Algorithm 2) and GD (Algorithm 4), we are given topological order π̄ and
matrix R̄ = adj(π̄). Let S k = {j ∈ J k |R̄jk = 1} be the set of selected candidate arcs for dependent
variable k. Given fixed R̄, (14) is separable into m LASSO linear regression problems
min
X
X
1X
|βjk |,
βjk xij )2 + λ
(xik −
n
k
k
i∈I
5
k ∈ J.
(16)
j∈S
j∈S
Computational Experiment
For all computational experiments, a server with two Xeon 2.70GHz CPUs and 24GB RAM is used.
Although there are many papers studying Bayesian network learning with various error measures and
penalties, here we focus on minimizing the LASSO type objective (SSE and penalty) and we picked one
of the latest paper of Han et al. [19] with the same objective function as benchmark.
The MIP models MIPcp, MIPin, and MIPto are implemented with CPLEX 12.6 in C#. For MIPcp,
instead of implementing the original cutting plane algorithm, we use CPLEX Lazy Callback, which is
similar to the cutting plane algorithm. Instead of solving (3) to optimally from scratch in each iteration,
we solve (2) with Lazy Callback, which allows updating (adding) constraints (cycle prevention constraints)
in the process of the branch and bound algorithm whenever an integer solution with cycles is found. Given
a solution with the cycles, we detect all cycles and add cycle prevention constraints for the detected cycles.
For MIPcp, MIPin, and MIPto, we set big M as follows. Given λ, we solve (14) without acyclic
constraints. Hence, we are allowed to use all arcs in J k for each model k ∈ J. Then, we obtain the
estimated upper bound for big M by
M =2
max |βjk | .
(17)
j∈J k ,k∈J
We observed that the above formula gives large enough big M for all cases in the following experiment.
In Appendix B, we present comparison of regression coefficients of implanted network (DAG) with big M
values by (17). The result shows that the big M value in (17) is always valid for all cases considered.
We compare our algorithms and models with the algorithm in Han et al. [19], which we denote as
DIST here. Their algorithm starts with neighborhood selection (NS), which filters unattractive arcs and
removes them from consideration. The procedure is specifically developed for high dimensional variable
selection when m is much larger than n. In our experiment, many instances considered are not high
dimensional and some have dense solutions. Further, by filtering arcs, there exists a probability that an
arc in the optimal solution can be removed. Hence, we deactivated the neighborhood selection step of
their original algorithm, where the R script of the original algorithm is available on the journal website.
For GD and IR, the algorithms are written in R [32]. We use glmnet package [12] function glmnet for
solving LASSO linear regression problems in (16). For IR, we use parameters α = 0.01, t∗ = 10, νlb = 0.8,
12
and νub = 1.2. For GD, we use parameters α = 0.01, t∗1 = 10, and t∗2 = 5. Because both GD and IR
start with a random solution, they perform different with different random solutions. Further, since we
observe that the execution time of GD and IR are much faster than DIST, we decided to run GD and
IR with 10 different random seeds and report the best solution. To emphasize the number of different
random seeds for GD and IR, we use the notation GD10 and IR10 in the rest of the section.
We first test all algorithms with synthetic instances generated using R package pcalg [23]. Function
randomDAG is used to generate a DAG and function rmvDAG is used to generate multivariate data with
the standard normal error distribution. First, a DAG is generated by randomDAG function. Next, the
generated DAG and random coefficients are used to create each column (with standard normal error
added) by rmvDAG function which uses linear regression as the underlying model. After obtaining the
data matrix from the package, we standardize each column to have zero mean with standard deviation
equal to one. The DAG used to generate the multivariate data is considered as the true structure or
true arc set while it may not be the optimal solution for the score function. The random instances are
generated for various parameters described in the following.
m: number of features (nodes)
n: number of observations
s: expected number of true arcs per node
d: expected density of the adjacency matrix of the true arcs
By changing the ranges of the above parameters, three classes of random instances are generated.
Sparse data sets: The expected total number of true arcs is controlled by s and most of the instances
have a sparse true arc set. We use n ∈ {100, 200, 300}, m ∈ {20, 30, 40, 50}, and s ∈ {1, 2, 3} to
generate 10 instances for each (n, m, s) triplet. This yields a total of 360 random instances.
Dense data sets: The expected total number of true arcs is controlled by d and most of the instances
have a dense true arc set compared with the sparse data sets. We use n ∈ {100, 200, 300}, m ∈
{20, 30, 40, 50}, and d ∈ {0.1, 0.2, 0.3} to generate 10 instances for each (n, m, d) triplet, and thus
we have a total of 360 random instances.
High dimensional data sets: The instances are high dimensional (m ≥ n) and very sparse. The
expected total number of true arcs is controlled by s. We use n = 100 and m ∈ {100, 150, 200},
and s ∈ {0.5, 1, 1.5} to generate 10 instances for each (m, s) pair, which yields a total of 90 random
instances.
We use four λ values differently defined for each data set in order to cover the expected number of arcs
with the four λ values. For each sparse instance, we solve (14) with λ ∈ {1, 0.5, 0.1, 0.05}. For dense data
sets, a wide range of λ values are needed to obtain selected arc sets that have similar cardinalities with the
true arc sets. Hence, for each dense instance, instead of fixed values over all instances in the set, we use
λ values based on expected density d: λ = λ0 · 10−(10·d−1) , where λ0 ∈ {1, 0.1, 0.01, 0.001}. For each high
dimensional instance, we use λ ∈ {1, 0.8, 0.6, 0.4}. Observe that the expected densities of the adjacency
matrices vary across the three data sets. The sparse instances have expected densities between 0.02 and
0.15, the dense instances have expected densities between 0.1 and 0.3, and the high dimensional instances
have expected densities between 0.002 and 0.015. Hence, different ranges of λ values are necessary.
For all of the results presented in this section, we present the average performance by n, m, s, d and
λ. For example, the result for n = 100 and m = 20 are the averages of 120 and 90 instances, respectively.
In all of the comparisons, we use the following metrics.
time: computation time in seconds
δsol : relative gap (%) from the best objective value among the compared algorithms or models. For
example, if we compare the three MIP models, δsol of an MIP model is the relative gap from the
best of the three objective function values obtained by the MIP models.
kzk0 : number of arcs selected (number of nonzero regression coefficients βjk ’s)
In comparing the performance metrics, we use plot matrices. In each Figure 1, 2, 3, 5, and 6, multiple
bar plots form a matrix. The rows of the plot matrix correspond to performance metrics and the columns
stand for parameters used for result aggregation. For example, the left top plot in Figure 1 shows
13
execution times of the algorithms where the results are aggregated by n (the number of observations),
because the first row and first column of the plot matrix in Figure 1 are for execution times and n,
respectively.
In Section 5.1, we compare the performance of iterative algorithms GD10 and IR10 and the benchmark
algorithm DIST. In Section 5.2, we compare the performance of MIP models MIPcp, MIPin, and MIPto.
We also compare all models and algorithms with a subset of the synthetic instances in Section 5.2. Finally,
in Section 5.4, we solve a popular real instance of Sachs et al. [35] in the literature.
5.1
Comparison of Iterative Algorithms
Time (seconds)
In this section, we compare the performance of GD10, IR10, and DIST by time, δsol , and kzk0 for each of
the three data sets.
In Figure 1, the result for the sparse data sets is presented. The bar plot matrix presents the performance measures aggregated by n, m, s, and λ.
150
150
150
150
100
100
100
100
50
50
50
50
0
0
100
𝛿𝑠𝑜𝑙
0
0
0
20
300
30
40
50
1
2
0.05
3
1.2%
1.2%
1.2%
1.2%
1.0%
1.0%
1.0%
1.0%
0.8%
0.8%
0.8%
0.8%
0.6%
0.6%
0.6%
0.6%
0.4%
0.4%
0.4%
0.4%
0.2%
0.2%
0.2%
0.2%
0.0%
0.0%
0.0%
100
𝑧
200
200
300
20
30
40
2
3
0.05
400
400
400
400
300
300
300
300
200
200
200
200
100
100
100
100
0
0
100
200
By n (# obs)
300
30
40
50
1
By m (# nodes)
GD10
2
3
By s (# arcs per node)
IR10
1
0.1
0.5
1
0.5
1
0
0
20
0.5
0.0%
1
50
0.1
0.05
0.1
By λ (penalty)
DIST
Figure 1: Performance of GD10, IR10, and DIST (sparse data)
The computation time of all three algorithms increases in increasing m and decreasing λ, where
the computation time of DIST increases faster than the other two. The computation time of DIST is
approximately 10 times faster than the GD10 time when λ = 1, but 2 times slower when λ = 0.05.
With increasing n, the computation times of all algorithms stay the same or decrease. This can be seen
counter-intuitive because larger instances do not increase time. However, a larger number of observations
can make predictions more accurate and could reduce search time for unattractive subsets. Especially,
the computation time of DIST decreases in increasing n. We think this is because more observations
give better local selection in the algorithm when adding and removing arcs. The number of selected arcs
(kzk0 ) of GD10 and IR10 is greater than DIST for all cases because
the topological order based algorithms
are capable of using the maximum number of arcs m(m−1)
, while arc selection based algorithms, such
2
as DIST, are struggling to select many arcs without violating acyclic constraints. In terms of the solution
quality, all algorithms have δsol less than 1.2% and perform good. However, we observe several trends. As
14
Time (seconds)
λ decreases (required to select more arcs), GD10 and IR10 start to outperform. We also observe that, as
the problem requires to select more arcs (increasing m, increasing s, and decreasing λ), GD10 and IR10
perform better. As n increases, δsol of GD10 and DIST decrease, whereas δsol of IR10 increases.
The result for the dense data sets is presented in Figure 2. The bar plot matrix presents the performance measures aggregated by n, m, d, and λ0 . Recall that, for the dense data set, we solve (14) with
λ = λ0 · 10−(10·d−1) and λ0 ∈ {1, 0.1, 0.01, 0.001}. For simplicity of presenting the aggregated result, we
use λ0 in the plot matrix, while λ values are used for actual computation.
800
800
800
800
600
600
600
600
400
400
400
400
200
200
200
200
0
0
100
𝛿𝑠𝑜𝑙
200
30
40
50
0.1
0.2
0.001
0.3
50%
50%
50%
50%
40%
40%
40%
40%
30%
30%
30%
30%
20%
20%
20%
10%
0.47%
0.51%
0.26%
0.46%
0.32%
0.49%
0%
10%
0.06%
0.40%
0.20%
0.72%
0.15%
0.59%
0.47%
10%
0.75%
200
300
20
30
40
0.85%
0.38%
0.59%
0.15%
0.47%
10%
0.2
0.3
600
600
600
600
500
500
500
500
400
400
400
400
300
300
300
0 200
300
200
200
200
100
100
100
100
0
0
0
100
200
By n (# obs)
300
20
30
40
50
GD10
0.2
0.3
By d (density of solution matrix)
IR10
1
0.57%
0.49%
0.39%
0.47%
0.56%
0.33%
0.63%
0.12%
0.37%
0.001
0.01
0.1
1
0.001
0.01
0.1
1
0
0.1
By m (# nodes)
0.1
0%
0.1
50
0.01
20%
0.08%
0%
0%
100
𝑧
0
0
20
300
By λ0 (penalty)
DIST
Figure 2: Performance of GD10, IR10, and DIST (dense data)
The computation time of all three algorithms again increases in increasing m and decreasing λ, where
the computation time of DIST increases faster than the other two. Compare to the result for the sparse
data sets, the execution times are all larger for the dense data set. The number of selected arcs (kzk0 )
of GD10 and IR10 is again greater than DIST for all cases, where kzk0 is twice larger for GD10 and IR10
when m or d is large, or λ is small. In terms of the solution quality, GD10 and IR10 outperform in most
of the cases, while δsol values of DIST increase fast in changing m, d, and λ. The values of δsol for all
algorithms are larger than the sparse data sets result. GD10 and IR10 are better for most of the cases. In
general, we again observe that GD10 and IR10 perform better when the problem requires to select more
arcs.
The result for the high dimensional data sets is presented in Figure 3. The bar plot matrix presents
the performance measures aggregated by m, s, and λ, while n is excluded from the matrix as we fixed n
to 100.
The computation time of all three algorithms again increases in increasing m and decreasing λ.
However, unlike the previous two sets, the computation times of GD10 and IR10 increase faster than
DIST. This is due to the efficiency of the topological order based algorithms. When a very small portion
of the arcs should be selected in the solution, topological orders are not informative. For example, consider
a graph with three nodes A, B, C and assume that only one arc, (A, B), is selected due to a large penalty.
For this case, three topological orders A − B − C, A − C − B, and C − A − B can represent the selected
arc. The third row of Figure 3 shows that kzk0 of the three algorithms are very similar, while DIST has
15
Time (seconds)
600
600
600
500
500
500
400
400
400
300
300
300
200
200
200
100
100
100
0
0
100
𝛿𝑠𝑜𝑙
0
200
0
0.5
1
1.5
0.4
0.5%
0.5%
0.5%
0.4%
0.4%
0.4%
0.3%
0.3%
0.3%
0.2%
0.2%
0.2%
0.1%
0.1%
0.1%
0.0%
0.0%
100
𝑧
150
150
200
1
1.5
0.4
400
400
300
300
300
200
200
200
100
100
100
0
0
150
200
1
0.6
0.8
1
0.6
0.8
1
0
0.5
By m (# nodes)
0.8
0.0%
0.5
400
100
0.6
1
0.4
1.5
By s (# arcs per node)
GD10
IR10
By λ (penalty)
DIST
Figure 3: Performance of GD10, IR10, and DIST (high dimensional data)
much smaller values for the previous two data sets. This implies that arc based search by DIST does
not have difficulties preventing cycles and the algorithm can decide whether to include arcs easier. The
comparison of δsol values also show that arc based search is competitive. Although all algorithms have
δsol values less than 0.5%, we find clear evidence that the performance of IR10 decreases in increasing
m and s and decreasing λ. Although the δsol values of GD10 and DIST are similar, considering the fast
computing time of DIST, we recommend to use DIST for very sparse high dimensional data.
In Figure 4, we present combined results of all three data sets by relating solution densities and δsol .
Observe that, for each quadruplet of n, m, s(or d), λ, we have results from 10 random instances for each
algorithm. Value δsol is the average of the 10 results for each quadruplet and for each algorithm, AvgDen
is the average density of the adjacency matrices of the 10 results and the three algorithms. In Figure 4,
we present a scatter plot of ln(1 + AvgDen) and ln(1 + 100 · δsol ). Each point in the plot is the average of
10 results by an algorithm and each algorithm has 324 points displayed 1 . The numbers in the parenthesis
along the axes are the corresponding values of δsol and AvgDen. In the plot, we first observe that the
algorithms perform similarly when the solutions are sparse and the δsol values have large variance when
the solutions are dense. When the log transformed solution density is less than 2, the average δsol values
of GD10, IR10, and DIST are 0.06%, 0.15%, and 0.04%, respectively. However, the solution quality of
DIST drastically decreases as the solutions become denser. This makes sense because sparse solutions
can be efficiently searched by arc-based search, while dense solutions are not easy to obtain by adding
or removing arcs one by one. This also explains the relatively small and large δsol values for dense and
spares solutions, respectively, by the topological order based algorithms. Between GD10 and IR10, we do
not observe a big difference.
1 324 = (3 · 4 · 3 · 4) + (3 · 4 · 3 · 4) + (1 · 3 · 3 · 4), where the three terms are for the three data sets and each term is obtained
by multiplying the number of parameters n, m, s(d), and λ, respectively.
16
Log transformed δsol
5 (147.4%)
4 (53.6%)
3 (19.1%)
2 (6.4%)
1 (1.7%)
0
0
0.5
(0.006)
1
(0.017)
1.5
(0.035)
2
(0.064)
2.5
(0.11)
3
(0.19)
3.5
(0.32)
4
(0.54)
Log transformed average subgraph density
GD10
IR10
DIST
Figure 4: Scatter plot of δsol and average solution densities
5.2
Comparison of MIP Models
In this section, we compare the performance of MIPto, MIPin, and MIPcp using time, δsol , and kzk0 and
the following additional metric.
δIP : the optimality gap (%) obtained by CPLEX within allowed 15 · m seconds
Due to scalability issues of all models, we only use the sparse data with m = 20, 30, 40. We also limit
n = 100. For all instances, we use the 15 · m seconds time limit for CPLEX. For example, we have time
limit of 300 seconds for instances with m = 20.
The result is presented in Figure 5. Comparing the time of all models with the time limit for CPLEX,
we observe that MIPin and MIPcp were able to terminate with optimality for several instances when
m = 20 and λ = 1. This implies that MIPin and MIPcp are efficient when the problem is small and the
number of selected arcs kzk0 is small. However, in general, δIP values tend to be consistent with different
models, while they increase in increasing m and s and in decreasing λ for all three MIP models. The
execution times of all models increase in increasing m and s, and in decreasing λ. The same trend can be
found for δIP for all models. By comparing δsol values, we observe that MIPin is best when m = 20 and
30. However, the performance of MIPin drops drastically as m and s increase and λ decreases. Actually,
MIPin fails to obtain a reasonably good solution within the time limit for several instances. This gives
large δsol values and increases the average. The δsol values of MIPto are smaller than MIPcp when m and
s are small, while MIPcp outperforms when m = 40 or s = 3.
5.3
Comparison of all MIP Models and Algorithms
In Figure 6, we compare all models and algorithms for selected sparse instances with n = 100 and m ∈
{20, 30, 40}, which were used to test MIP models. In the plot matrix, we show the average computation
time and δsol (gap from the best objective value among the six models and algorithms) by m, s, and λ.
Note that δsol values of a few MIPin results are not fully displayed in the bar plots due to their large
values. Instead, the actual numbers are displayed next to the corresponding bar. The result shows that
MIP models spent more time while the solution qualities are inferior in general. The values of δsol for
the MIP models are competitive only when λ is large, which requires sparse solution. However, even for
this case, MIP models spend longer time than the algorithms. Hence, ignoring the benefit of knowing
and guaranteeing optimality by the MIP models, we conclude that the algorithms perform better for all
cases.
The primary reason for inferior performance of the MIP models is the difficulty of solving integer
programming problems. Further, all of the MIP models have at least O(m2 ) binary variables and O(m2 )
17
Time (seconds)
600
600
600
500
500
500
400
400
400
300
300
300
200
200
200
100
100
100
0
0
20
𝛿𝐼𝑃
30
0
1
2
3
30%
30%
30%
25%
25%
25%
20%
20%
20%
15%
15%
15%
10%
10%
10%
5%
5%
5%
0%
0%
20
𝛿𝑠𝑜𝑙
40
30
2
3
8%
8%
8%
6%
6%
6%
4%
4%
4%
2%
2%
2%
0%
0%
20
30
By m (# nodes)
2
3
1
0.05
0.1
0.5
1
0.5
1
0.05
By s (# arcs per node)
MIPto
0.5
0%
1
40
0.1
0%
1
40
0.05
MIPin
0.1
By λ (penalty)
MIPcp
Figure 5: Performance of MIPto, MIPin, and MIPcp (sparse data with n = 100 and m ∈ {20, 30, 40})
constraints and the problem complexity grows fast. Finally, large values of big M in (15b) make the
problem even more difficult. Due to non-tight values of big M, fathoming does not happen frequently
in the branch and bound procedure. Hence, at least for sparse Gaussian Bayesian network learning, the
MIP models may not be the best option unless other complicated constraints, which cannot be easily
dealt with iterative algorithms, are needed.
5.4
Real Data Example
In this section, we study the flow cytometry data set from Sachs et al. [35] by solving (14). The data set
has been studied in many works including Friedman et al. [11], Shojaie and Michailidis [36], Fu and Zhou
[14], and Aragam and Zhou [1]. The data set is often used as a benchmark as the casual relationships
(underlying DAG) are known. It has n = 7466 cells obtained from multiple experiments with m = 11
measurements. The known structure contains 20 arcs. For the experiment, we standardize each column
to have zero mean with standard deviation equal to one.
In Table 2, we compare the performance of the three algorithms GD10, IR10, and DIST for various
values of λ. The MIP models are excluded due to scalability issue 2 . We compare the previously
used performance measures execution time, solution cardinalities (kzk0 ), and solution quality (δsol ). In
addition, we also compare sensitivities (true positive ratio) of the solutions. By comparing with the
known structure with 20 arcs, we calculate directed true positive (dTP) and undirected true positive
(uTP). If arc (j, k) is in the known structure, dTP counts only if arc (j, k) is in the algorithm’s solution,
whereas uTP counts either of arcs (j, k) or (k, j) is in the algorithm’s solution.
The solution times of the three algorithms are all within a few seconds, as m is small. Also, the
2 Although m is small, with n =7466 observations, the MIP models have at least 7466 · 11 · 2 = 164,252 continuous
variables and 7466 · 11 = 82,126 constraints just for the residual terms. Combined with the complexity increment due to
the binary variables and acyclicity constraints, it was not feasible to obtain a reasonable solution by any of the MIP models.
18
Time (seconds)
700
600
500
400
300
200
100
0
700
600
500
400
300
200
100
0
20
30
𝛿𝑠𝑜𝑙
40
≈
4.0%
700
600
500
400
300
200
100
0
1
2
3
≈
4.0%
9.2%
0.05
4.0%
7.8%
3.0%
3.0%
3.0%
2.0%
2.0%
2.0%
1.0%
1.0%
1.0%
0.0%
0.0%
20
30
40
MIPin
1
0.1
0.5
1
9.0%
2
3
0.05
By s (# arcs per node)
MIPto
0.5
0.0%
1
By m (# nodes)
≈
0.1
MIPcp
GD10
By λ (penalty)
IR10
DIST
Figure 6: Performance of all models and algorithms (sparse data with n = 100 and m ∈ {20, 30, 40})
solution cardinalities (kzk0 ) are similar. The best δsol value among the three algorithms in each row is
in boldface. We observe that GD10 provides the best solution (smallest δsol ) in most cases and IR10 is
the second best. The δsol values of DIST increase as λ increases. This is consistent with the findings in
Section 5.1. Note that the density of the underlying structure is 20/112 = 0.165, which is dense. This
explains the good performance of GD10 and IR10. On the other hand, even though GD10 provides the
best objective function values for most of the cases, the dTP and uTP values of GD10 are not always the
best. The highest value among the three algorithms in each row is in boldface. While δsol values of DIST
are the largest among the three algorithms, dTP and uTP values are the best in some cases. When λ is
small, DIST tends to have higher dTP and uTP. However, as λ increases, GD10 gives the best dTP and
uTP values. To further improve the prediction power, we may need weighting features or observations.
λ
0.5
0.45
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
time
14.6
13.7
14.3
16.0
15.1
14.4
16.0
16.4
17.4
17.7
kzk0
9
11
11
13
13
16
16
21
24
28
GD10
δsol
0.00%
0.00%
0.00%
0.00%
0.00%
0.00%
0.00%
0.00%
0.00%
0.32%
dTP
0.22
0.27
0.27
0.23
0.23
0.38
0.31
0.33
0.25
0.39
uTP
0.56
0.55
0.55
0.54
0.46
0.56
0.56
0.57
0.50
0.50
time
10.8
10.9
14.8
12.4
14.3
14.8
14.6
15.6
11.0
11.5
kzk0
9
13
13
17
17
20
22
25
28
30
IR10
δsol
0.22%
0.30%
0.36%
0.42%
0.21%
0.22%
0.31%
0.32%
0.26%
0.00%
dTP
0.22
0.15
0.15
0.12
0.24
0.25
0.32
0.32
0.07
0.10
uTP
0.56
0.46
0.46
0.41
0.41
0.50
0.55
0.52
0.43
0.47
time
5.8
8.5
7.0
8.6
8.0
6.7
6.8
9.5
9.1
10.1
kzk0
9
9
13
15
17
20
21
23
27
32
DIST
δsol
0.38%
0.49%
0.61%
0.73%
0.84%
0.94%
1.12%
1.26%
1.45%
1.87%
dTP
0.56
0.56
0.38
0.33
0.29
0.30
0.33
0.30
0.26
0.22
uTP
0.56
0.56
0.46
0.47
0.47
0.50
0.52
0.52
0.44
0.41
Table 2: Performance on the real data set from Sachs et al. [35]
From Table 2, we observe that a slight change in solution quality δsol affects the final selection of DAG
significantly. Also, the best objective function value does not necessarily give the highest true positive
value since the L1 -norm penalized least square (14) may not be the best score function. In Figure 7,
we present the graphs of known casual interactions (a), estimated subgraph by GD10 (b), IR10 (c), and
DIST (d). All graphs are obtained with λ = 0.25, but the numbers of arcs are different (see Table 2). In
fact, the difference between δsol values of IR10 and DIST is less than 1% while the subgraphs have only
10 common arcs.
19
(a)
(b)
(c)
(d)
Figure 7: Known DAG (a) and estimated subgraphs with λ = 0.25 by GD10 (b), IR10 (c), and DIST (d)
6
Conclusion
We propose an MIP model and iterative algorithms based on topological order. Although the computational experiment is conducted for Gaussian Bayesian network learning, all the proposed model and
algorithms are applicable for problems following the form in (1). While many MIP models and algorithms
are designed based on arc search, using topological order provides some advantages that improve solution
quality and algorithm efficiency.
1. DAG constraints (acyclicity constraints) are automatically satisfied when arcs from high order nodes
to low order nodes are used.
2. In applying the concept for MIP, a lower number of constraints is needed (O(m2 )) whereas arc
based modeling can have exponentially many constraints in the worst case.
3. In applying the concept in designing iterative algorithms, one of the biggest merits is the capability
of utilizing the maximum number of arcs possible ( m(m−1)
), while arc based algorithms struggle
2
with using all possible arcs.
The proposed MIP model has the smallest number of constraints while the number of binary variables
is in the same order with the already known MIP models. It performs as good as a cutting plane algorithm.
The proposed iterative algorithms get the biggest benefit when the solution matrix is dense. The result
presented in Section 5.1 clearly indicates that the topological order based algorithms outperform when
the density of the resulting solution is high. On the other hand, arc-based search algorithms, represented
by DIST in our experiment, can be efficient when the desired solutions are very sparse.
Comparing all models and algorithms used in the experiment, we observe that the MIP models are not
competitive or scale well compared to the heuristic algorithms except for small instances. The experiment
shows that the solution times of the MIP models are significantly affected by the number of nodes m.
For Gaussian Bayesian network learning, we observe that large n could also decrease the MIP model
efficiency even when m is small (Section 5.4). Among the iterative algorithms, we recommend DIST for
very sparse high dimensional data and GD10 and IR10 for dense data. Among the two topological order
based algorithms, GD10 performs slightly better and is more stable.
References
[1] B. Aragam and Q. Zhou. Concave penalized estimation of sparse Gaussian Bayesian networks.
Journal of Machine Learning Research, 16:2273–2328, 2015.
[2] A. Baharev, H. Schichl, and A. Neumaier. An exact method for the minimum feedback arc set
problem. Technical report, 2015. Available from http://www.mat.univie.ac.at/∼neum/.
[3] G. Bolotashvili, M. Kovalev, and E. Girlich. New facets of the linear ordering polytope. SIAM
Journal on Discrete Mathematics, 12(3):326–336, 1999.
20
[4] D. M. Chickering. Learning Bayesian networks is NP-complete. In Learning from data, pages 121–
130. Springer, 1996.
[5] D. M. Chickering. Optimal structure identification with greedy search. Journal of Machine Learning
Research, 3(Nov):507–554, 2002.
[6] W. J. Cook, W. H. Cunningham, W. R. Pulleyblank, and A. Schrijver. Combinatorial Optimization.
John Wiley & Sons, Inc., New York, NY, USA, 1998.
[7] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein. Introduction to algorithms. MIT press,
2009.
[8] B. Ellis and W. H. Wong. Learning causal Bayesian network structures from experimental data.
Journal of the American Statistical Association, 103(482):778–789, 2008.
[9] G. Even, J. S. Naor, B. Schieber, and M. Sudan. Approximating minimum feedback sets and
multicuts in directed graphs. Algorithmica, 20(2):151–174, 1998.
[10] H. Fernau and D. Raible. Exact algorithms for maximum acyclic subgraph on a superclass of cubic
graphs. In International Workshop on Algorithms and Computation, pages 144–156. Springer, 2008.
[11] J. Friedman, T. Hastie, and R. Tibshirani. Sparse inverse covariance estimation with the graphical
Lasso. Biostatistics, 9(3):432–441, 2008.
[12] J. Friedman, T. Hastie, and R. Tibshirani. Regularization paths for generalized linear models via
coordinate descent. Journal of Statistical Software, 33(1):1–22, 2010.
[13] N. Friedman and D. Koller. Being Bayesian about network structure. A Bayesian approach to
structure discovery in Bayesian networks. Machine Learning, 50(1):95–125, Jan 2003.
[14] F. Fu and Q. Zhou. Learning sparse causal Gaussian networks with experimental intervention:
regularization and coordinate descent. Journal of the American Statistical Association, 108(501):288–
300, 2013.
[15] M. X. Goemans and L. A. Hall. The strongest facets of the acyclic subgraph polytope are unknown. In
International Conference on Integer Programming and Combinatorial Optimization, pages 415–429.
Springer, 1996.
[16] M. Grötschel, M. Jünger, and G. Reinelt. A cutting plane algorithm for the linear ordering problem.
Operations Research, 32(6):1195–1220, 1984.
[17] M. Grötschel, M. Jünger, and G. Reinelt. On the acyclic subgraph polytope. Mathematical Programming, 33(1):28–42, 1985.
[18] V. Guruswami, R. Manokaran, and P. Raghavendra. Beating the random ordering is hard: Inapproximability of maximum acyclic subgraph. In 49th Annual IEEE Symposium on Foundations of
Computer Science, pages 573–582. IEEE, 2008.
[19] S. W. Han, G. Chen, M.-S. Cheon, and H. Zhong. Estimation of directed acyclic graphs through
two-stage adaptive Lasso for gene network inference. Journal of the American Statistical Association,
pages 1–48, 2016.
[20] R. Hassin and S. Rubinstein. Approximations for the maximum acyclic subgraph problem. Information Processing Letters, 51(3):133–140, 1994.
[21] D. Heckerman, D. Geiger, and D. M. Chickering. Learning Bayesian networks: The combination of
knowledge and statistical data. Machine learning, 20(3):197–243, 1995.
[22] R. Kaas. A branch and bound algorithm for the acyclic subgraph problem. European Journal of
Operational Research, 8(4):355–362, 1981.
21
[23] M. Kalisch, M. Mächler, D. Colombo, M. H. Maathuis, and P. Bühlmann. Causal inference using
graphical models with the R package pcalg. Journal of Statistical Software, 47(11):1–26, 2012.
[24] R. M. Karp. Reducibility among combinatorial problems. Springer, 1972.
[25] C. Kenyon-Mathieu and W. Schudy. How to rank with few errors. In Proceedings of the thirty-ninth
annual ACM symposium on Theory of computing, pages 95–103. ACM, 2007.
[26] W. Lam and F. Bacchus. Learning Bayesian belief networks: An approach based on the mdl principle.
Computational Intelligence, 10(3):269–293, 1994.
[27] J. Leung and J. Lee. More facets from fences for linear ordering and acyclic subgraph polytopes.
Discrete Applied Mathematics, 50(2):185–200, 1994.
[28] C. Lucchesi and D. Younger. A minimax theorem for directed graphs. Journal of the London
Mathematical Society, 17(3):369–374, 1978.
[29] J. E. Mitchell and B. Borchers. Solving linear ordering problems with a combined interior
point/simplex cutting plane algorithm. In High Performance Optimization, pages 349–366. Springer,
2000.
[30] T. Niinimäki, P. Parviainen, and M. Koivisto. Structure discovery in Bayesian networks by sampling
partial orders. The Journal of Machine Learning Research, 17(1):2002–2048, 2016.
[31] Y. W. Park and D. Klabjan. Subset selection for multiple linear regression via optimization. Technical
report, 2013. Available from http://www.klabjan.dynresmanagement.com.
[32] R Core Team. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria, 2016.
[33] V. Ramachandran. Finding a minimum feedback arc set in reducible flow graphs. Journal of
Algorithms, 9(3):299–313, 1988.
[34] G. Raskutti and C. Uhler. Learning directed acyclic graphs based on sparsest permutations. arXiv
preprint arXiv:1307.0366, 2013.
[35] K. Sachs, O. Perez, D. Pe’er, D. A. Lauffenburger, and G. P. Nolan. Causal protein-signaling
networks derived from multiparameter single-cell data. Science, 308(5721):523–529, 2005.
[36] A. Shojaie and G. Michailidis. Penalized likelihood methods for estimation of sparse high-dimensional
directed acyclic graphs. Biometrika, 97(3):519–538, 2010.
[37] S. Van de Geer and P. Bühlmann. L0-penalized maximum likelihood for sparse directed acyclic
graphs. The Annals of Statistics, 41(2):536–567, 2013.
APPENDIX
A
Greedy Algorithm for Projection Problem
In this section, we present the detail derivations and proofs of Algorithm 3 (greedy). The algorithm
sequentially determines topological order by optimizing the projection problem given an already fixed
order up to the iteration point. Solving the projection problem is N P-complete and our algorithm may
not give a global optimal solution. However, the result of this section shows that Algorithm 3 gives an
optimal choice of the next node to have fixed order given pre-fixed orders.
We start by describing some properties of Y ∗ in the following three lemmas. For the following lemmas,
let πj∗ represent the topological order of node j defined by Y ∗ .
∗
∗
∗
t
Lemma 2. For any Yjk
, we must have either Yjk
= 0 or Yjk
= Ujk
.
22
∗
∗
t
Proof. For a contradiction, let us assume that there exist indices q and r such that Yqr
6= 0 and Yqr
6= Uqr
.
∗
t
Let us create a new solution Ȳ such that Ȳ = Y except Ȳqr = Uqr . Note that Ȳ is a feasible solution to
∗
(12) because supp(Ȳ ) ≤ supp(Y ∗ ) element-wise since Yqr
6= 0. Further, we have kY ∗ − U t k2 > kȲ − U t k2
∗
t 2
t 2
∗
∗
because (Yqr − Uqr ) > 0 = (Ȳqr − Uqr ) and Ȳ = Y except Yqr
6= Ȳqr . This contradicts optimality of
∗
Y .
∗
∗
Note that Lemma 2 implies that solving (12) is essentially choosing between 0 and Ujk
for Yjk
. This
selection is also based on the following property.
∗
t
= Ujk
.
Lemma 3. If πj∗ > πk∗ , then Yjk
∗
t
Proof. For a contradiction, let us assume that there exist indices q and r such that Yqr
6= Uqr
while
∗
∗
∗
t
πq > πr . Let us create a new solution Ȳ such that Ȳ = Y except Ȳqr = Uqr .
∗
1. If Yqr
6= 0, then Ȳ is a DAG since supp(Ȳ ) ≤ supp(Y ∗ ) element-wise.
∗
2. If Yqr
= 0, then arc (q, r) can be used in the solution without creating a cycle because πq∗ > πr∗ .
Hence, Ȳ is a DAG.
Therefore, Ȳ is a feasible solution to (12). However, it is easy to see that kȲ − U t k2 < kY ∗ − U ∗ k2
∗
t 2
t 2
because (Yqr
− Uqr
) > 0 = (Ȳqr − Uqr
) . This contradicts optimality of Y ∗ .
Given the topological order by Y ∗ , let Jˆk = {j ∈ J k |πj∗ > πk∗ } be the subset of J k such that the
nodes in Jˆk are earlier than node k. Combining Lemmas 2 and 3, we conclude that Y ∗ has the following
structure.
(
t
if j ∈ Jˆk ,
Ujk
∗
k∈J
(18)
Yjk =
0
if j ∈ J \ Jˆk ,
Further, we can calculate node k’s contribution to the objective function value without explicitly using
Y ∗.
Lemma 4. For each node k ∈ J, it contributes
X
t 2
(Ujk
)
j∈J\Jˆk
to the objective function value for (12). In other words, the contribution of node k is the squared sum of
t
Ujk
for nodes with πj∗ < πk∗ .
Proof. For node k, we can derive
P
P
P
∗
t 2
∗
t 2
t 2
j∈J (Yjk − Ujk ) =
j∈J\Jˆk (Yjk − Ujk ) =
j∈J\Jˆk (Ujk ) ,
∗
t
where both equal signs are due to (18). The first equality is due to Yjk
= Ujk
for j ∈ Jˆk and the second
∗
k
equality holds since Yjk = 0 for j ∈ J \ Jˆ .
We next detail the derivation of the greedy algorithm presented in Algorithm 3. Let J¯ ⊆ J be the
index set of yet-to-be-ordered nodes and J¯c = J \ J¯ be the index set of the nodes that have already been
ordered. The procedure is equivalent to iteratively solving
n
o
X
t 2
k ∗ = argmink∈J¯ min
(Yjk − Ujk
)
,
(19)
Yk
j∈J
where Yk = [Y1k , Y2k , · · · , Ymk ] ∈ Rm×1 is the column in Y corresponding to node k. Set J¯ is updated
¯ after solving (19). We propose an algorithm to solve (19) based on the
by J¯ = J¯ \ {k ∗ } and πk∗∗ = |J|
¯ we solve
properties of Y ∗ described in Lemmas 2 - 4. Given J,
nX
o
t 2
k ∗ = argmink∈J¯
(Ujk
) .
(20)
j∈J¯
23
Next, we show that solving (20) gives an optimal solution Ȳjk∗ , j ∈ J k to (19). We can actually replicate
the properties of Y ∗ for Ȳjk∗ .
t
k
Lemma 5. An optimal solution to (19) must have either Ȳjk∗ = 0 or Ȳjk∗ = Ujk
∗ for all j ∈ J .
t
¯
Lemma 6. An optimal solution to (19) must have Ȳjk∗ = Ujk
∗ for j ∈ J \ J.
P
P
t
2
t
2
Lemma 7. An optimal solution to (19) must satisfy j∈J (Ȳjk∗ − Ujk
∗) =
j∈J¯(Ujk∗ )
The proofs are omitted as they are P
similar to the proofs of Lemmas
2 - 4, respectively. Note that, by
P
t
2
t
2
Lemma 7, we show the equivalence of j∈J (Yjk∗ − Ujk
and j∈J¯(Ujk
∗)
∗ ) . This result also holds for
¯
each term, k ∈ J, of the argmin function in (19). Hence, the following lemma holds.
Lemma 8. Solving (20) is equivalent to solving (19).
Observe that (20) is used in Line 3 of the greedy algorithm in Algorithm 3. Hence, by property (19),
the algorithm gives an optimal choice of node to be fixed with the pre-fixed topological order.
B
Summary Statistics for Maximum Coefficients
In this section, we show that the heuristic formula (17) for selecting big M gives reasonable and large
enough values. Note that, when creating a synthetic instance, we used a random DAG to generate multivariate data. Although the optimal DAG for the penalized least squares are unknown, with appropriate
penalty constants, we can use the implanted DAG to obtain coefficients estimation. That is, we calculate
B̂ =
max
j∈J k ,k∈K
|β̂jk |
by using the implanted DAG. The B̂ value is then compared with the big M value in (17) for all instances
used for MIP models (sparse data with n = 100 and m ∈ {20, 30, 40}). We calculate minimum, average,
and maximum of B̂ and M by m, s, and λ. The result is presented in Table 3. Note that we have B̂ < M
for all columns Min, Avg, and Max and for all rows. Although only the summary statistics are presented
in Table 3, we observed that M is greater than B̂ for all cases considered.
By
Param Value
20
m
30
40
1
s
2
3
0.05
0.1
λ
0.5
1
Min
0.17
0.18
0.18
0.17
0.23
0.32
0.62
0.60
0.41
0.17
B̂
Avg
0.65
0.62
0.63
0.53
0.63
0.74
0.82
0.78
0.59
0.35
Max
1.48
0.95
1.03
0.79
0.94
1.48
1.48
0.96
0.74
0.49
Min
0.33
0.36
0.36
0.33
0.45
0.63
1.24
1.19
0.82
0.33
M
Avg
1.34
1.32
1.33
1.13
1.33
1.53
1.95
1.54
1.14
0.69
Max
3.73
3.10
3.30
2.04
3.54
3.73
3.73
2.35
1.45
0.97
Table 3: Comparison of maximum coefficients and big M (sparse data with n = 100 and m ∈ {20, 30, 40})
24
| 8cs.DS
|
Under consideration for publication in Theory and Practice of Logic Programming
1
arXiv:0712.3830v1 [cs.PL] 26 Dec 2007
TCHR: a framework for tabled CLP
TOM SCHRIJVERS∗ and BART DEMOEN
Dept. of Computer Science, K.U.Leuven, Belgium
(e-mail: {toms,bmd}@cs.kuleuven.ac.be)
DAVID S. WARREN
Dept. of Computer Science, State University of New York at Stony Brook, USA
(e-mail: [email protected])
submitted 25 September 2006 ; revised 19 July 2007; accepted 19 Dec 2007
Abstract
Tabled Constraint Logic Programming is a powerful execution mechanism for dealing
with Constraint Logic Programming without worrying about fixpoint computation. Various applications, e.g in the fields of program analysis and model checking, have been
proposed. Unfortunately, a high-level system for developing new applications is lacking,
and programmers are forced to resort to complicated ad hoc solutions.
This papers presents TCHR, a high-level framework for tabled Constraint Logic Programming. It integrates in a light-weight manner Constraint Handling Rules (CHR), a
high-level language for constraint solvers, with tabled Logic Programming. The framework
is easily instantiated with new application-specific constraint domains. Various high-level
operations can be instantiated to control performance. In particular, we propose a novel,
generalized technique for compacting answer sets.
KEYWORDS: Constraint Logic Programming, Constraint Handling Rules, tabled execution
1 Introduction
The notion of tabled Constraint Logic Programming (CLP) originates from the
constraint databases community (Kanellakis et al. 1995). In an ordinary database,
data is stored in relations of atomic values. Constraint databases generalize atomic
values to constraint variables: a field is restricted to a range of values rather than
a single value. This allows for more compact representations than explicitly enumerating the atomic values. DATALOG, a formalism for reasoning about ordinary
databases and queries in particular, is generalized to DATALOGD for this purpose.
Just as DATALOG is a restricted form of Logic Programming, DATALOGD is a restricted form of Constraint Logic Programming. The restrictions enforce programs
to have finite interpretations.
∗ Research Assistant of the fund for Scientific Research - Flanders (Belgium)(F.W.O. - Vlaanderen)
2
T. Schrijvers et al.
Due to the finiteness properties, queries on DATALOGD programs can be resolved by bottom-up computation rather than the usual top-down goal-directed
computation of CLP. The former has the advantage that it terminates for DATALOGD
programs, whereas the latter may get stuck in infinite loops. However, a goaldirected approach usually obtains the desired result much faster and uses less space.
For this reason, Toman (Toman 1997b) proposed a compromise: tabling. Tabling
is an LP technique for improving the termination properties of the goal-directed
approach through the memoization of intermediate results. By generalizing tabled
DATALOGD to Tabled CLP, we benefit from both the generalized expressivity of
CLP and the improved termination properties of tabling.
A number of different applications have been proposed for tabled DATALOGD
and Tabled CLP. Toman himself considers it an alternative approach to implementing abstract interpretation (Toman 1997a): constraints abstract concrete values and
tabling takes care of fixpoints. Various applications in the context of model checking
have been developed (Mukund et al. 2000; Du et al. 2000; Pemmasani et al. 2002):
constraints impose restrictions on parameters in parametrized models while tabling
takes care of cycles in the model graphs.
The above establishes a clear need for Tabled CLP, but let us consider the availability of Tabled CLP systems. It turns out that a user-friendly and comprehensive system for developing new Tabled CLP applications is missing completely. The
above-mentioned model checking applications (e.g. (Mukund et al. 2000; Du et al. 2000;
Pemmasani et al. 2002)) have adapted an existing tabled logic programming system, XSB (Warren et al. 2005), with constraint programming facilities in various
ad hoc and laborious ways.
At first, XSB developers resorted to interfacing with foreign language libraries
or implementing constraint solvers in XSB itself with a close coupling of constraint
solver and application as a consequence. For instance, the initial feasibility study of a
real-time model checking system used a meta interpreter written in XSB to deal with
constraints (see (Mukund et al. 2000)). The subsequent full system implements
an interface between XSB and the POLINE polyhedra-based constraint solver library, and passes around handles to the constraint store in the XSB program (see
(Du et al. 2000)). At a later stage this real time model checking application used
distance bound matrices implemented in XSB itself (see (Pemmasani et al. 2002)).
In an attempt to facilitate the use of constraints, XSB was extended with attributed variables (Cui and Warren 2000a). Attributed variables (Holzbaur 1992)
are Prolog language feature widely used for implementing constraint solvers. It allows to associate data with unbound variables, manipulate it at will and also to
interrupt the unification of these variables. Unfortunately, constraint solvers are
complex programs and even with attributed variables it can be a daunting task to
implement them.
In order to substantially lower the threshold for tabled CLP, a high-level formalism is needed for writing new constraint solvers and integrating them in a tabled
logic programming system. In this work we present such a formalism: Tabled Constraint Handling Rules, or TCHR for short.
TCHR is a high-level framework for developing new constraint solvers in a tabled
TCHR: a framework for tabled CLP
3
logic programming environment. It integrates Constraint Handling Rules (CHR)
(Frühwirth 1998), an established high-level formalism for writing new constraint
solvers, and tabled logic programming. The framework offers a number of default
operations that can be specialized by instantiations to control both semantics and
performance.
A practical implementation of the framework is presented: the integration of
K.U.Leuven CHR in XSB. The integration shows how a tabled constraint logic
programming system can be obtained from a Constraint Logic Programming and
a tabled logic programming system with little impact on either. Although we have
chosen XSB as our particular tabled logic programming system, we believe that our
ideas readily apply to other table-based LP systems.
In summary, the major contributions of this work are:
• a high-level framework for developing new constraint solvers in a tabled logic
programming system,
• a practical implementation of the framework in terms of K.U.Leuven CHR
and XSB, and
• a novel, generalized approach for answer set reduction.
The CHR-XSB integration, we believe, combines both the bottom-up and topdown fixpoint computations, the superior termination properties of XSB and the
constraint programming capabilities of CHR. This combined power enables programmers to easily write highly declarative programs that are easy to maintain
and extend.
Overview The rest of this text is structured as follows. First, in Sections 2
and 3 we provide basic technical background knowledge on tabled execution of
Constraint Logic Programs and Constraint Handling Rules, respectively.
Section 4 outlines our contribution: a framework for tabled CLP system integrated in terms of SLG and Constraint Handling Rules. Subsequent sections discuss
in more detail the different options and operations of the framework: call abstraction
(Section 6), answer projection (Section 7) and answer set optimization (Section 8).
Finally, Section 9 discusses related and possible future work, and Section 10
concludes.
But first we end this introduction with a small motivating example from the domain of model checking:
Example 1 Data-independent systems (Wolper 1986) manipulate data variables
over unbounded domains but have a finite number of control locations. Such systems
can be modeled as extended finite automata (Sarna-Starosta and Ramakrishnan 2003):
finite automata with guards on the transitions and variable mapping relations between source and destination locations. They are useful for modelling and subsequently checking e.g. buffers and protocols.
A simple example of such a data-independent system, modeled in CLP(FD), is:
edge(a,b,Xa,Xb) :- Xa < 10, Xb = Xa.
edge(b,a,Xb,Xa) :- Xb > 0, Xa = Xb + 1.
edge(b,c,Xb,Xc) :- Xb > 3, Xc = Xb.
4
T. Schrijvers et al.
This system has three control locations a,b,c each with one variable, respectively
Xa,Xb,Xc. Each edge/4 clause represents an edge in the system: a possible transition from one control location (the source) to another (the destination). The inequality constraint in each clause guards the transition, and the equality constraint
relates the source variable to the destination variable (the variable mapping).
Suppose that we are interested in whether location c is reachable from location
a, and for which values of the parameter Xa. Let us define a reachability predicate:
reach(A,A,X).
reach(A,C,X) :edge(A,B,X,NX),
reach(B,C,NX).
Then our reachability question is captured by the query ?- reach(a,c,X). In order to answer this query, tabling is required to avoid the non-termination trap of
the a-b cycle in the graph. At the same time, constraints allow a compact symbolic representation of the infinite search space for X. Without a good interaction
between both tabling and constraints, we would not be able to obtain as concise a
solution as 0 < X < 10 with so little effort.
2 Tabled Constraint Logic Programs
In this section we cover the basics of tabled Constraint Logic Programming. First,
the syntax of Constraint Logic Programs is presented in Section 2.1. Next, Section
2.2 explains about the constraints part of CLP: the constraint domain. Finally,
Section 2.3 presents the (operational) semantics SLGD of Tabled Constraint Logic
Programs.
2.1 Syntax of Constraint Logic Programs
A Constraint Logic Program consists of a number of rules, called clauses, of the
form:
H:-C, L1 , . . . ,Ln .
where n ≥ 0 and H is an atom, C is a constraint and L1 , . . . , Ln are literals. A
literal is either an atom A or a negated atom ¬A.
H is called the head of the clause and C, L1 , . . . , Ln is called the body. The comma
“,” is called conjunction as it corresponds with logical conjunction in the semantics
of Constraint Logic Programs.
The atoms are constructed from predicate symbols p/n and variables. Their
meaning is defined by the Constraint Logic Programming itself. The syntax and
semantics of constraints is defined by the constraint domain D (see Section 2.2).
If all the literals in the body are positive, the clause is a definite clause. A normal
clause is a clause that may also contain negative literals. A definite Constraint Logic
Program consists of definite clauses only, while a normal Constraint Logic Program
TCHR: a framework for tabled CLP
5
has normal clauses. From now one we will only consider definite programs and
address them as programs for short.
2.2 Constraint Domains
A constraint solver is a (partial) executable implementation of a constraint domain.
A constraint domain D consists of a set Π of constraint symbols, a logical theory
T and for every constraint symbol c/n ∈ Π a tuple of value sets hV1 , . . . , Vn i.
A primitive constraint is constructed from a constraint symbol c/n and for every
argument position i (1 ≤ i ≤ n) either a variable or a value from the corresponding
value set Vi , similar to the way an atom is constructed in a logic program.
A constraint is of the form c1 ∧ . . . ∧ cn where n ≥ 0 and c1 , . . . , cn are primitive
constraints. Two distinct constraints are true and false. The former always holds
and the latter never holds. The empty conjunction of constraints is written as true.
The logical theory T determines what constraints hold and what constraints do
not hold. Typically, we use D to also refer specifically to T . For example D |= c
means that under the logical theory T of constraint domain D the constraint c
holds.
A valuation θ for a constraint C is a variable substitution that maps the variables
in vars(C) onto values of the constraint domain D. If θ is a valuation for C, then it
is a solution for C if Cθ holds in the constraint domain D, i.e. D |= Cθ. A constraint
C is satisfiable if it has a solution; otherwise it is unsatisfiable. Two constraints C1
and C2 are equivalent, denoted D |= C1 ↔ C2 , if and only if they have the same
solutions.
A constraint domain of particular interest is the Herbrand domain H. Its only
constraint symbol is term equality = /2, which ranges over Herbrand terms. Plain
Logic Programming can be seen as a specialized form of Constraint Logic Programming over the Herbrand domain.
Two problems associated with a constraint C are the solution problem, i.e. determining a particular solution, and the satisfaction problem, i.e. determining whether
there exists at least one solution. An algorithm for determining the satisfiability of a
constraint is called a constraint solver. Often a solution is produced as a by-product.
A general technique used by many constraint solvers is to repeatedly rewrite a constraint into an equivalent constraint until a solved form is obtained. A constraint
in solved form has the property that it is clear whether it is satisfiable or not.
See (Marriott and Stuckey 1998) for a more extensive introduction to constraint
solvers.
2.3 Semantics of Constraint Logic Programs
In a survey of Constraint Logic Programming (CLP) (Jaffar and Maher 1994) various forms of semantics are listed for Constraint Logic Programs: logic semantics
based on Clark completion(Clark 1987), fixpoint semantics (Jaffar and Lassez 1987)
as well as a new framework for top-down and bottom-up operational semantics.
6
T. Schrijvers et al.
The CLP fixpoint semantics are defined, in the usual way, as the fixpoint of an
extended immediate consequence operator.
Definition 1 (CLP Immediate Consequence Operator )
The one-step consequence function TPD for a CLP program P with constraint domain D is defined as:
¯
T D (I) = {p(d)|
p(x̄) ← c, b1 , . . . , bn ∈ P,
P
∃v.v is a valuation on D :
D |= v(c),
¯
v(x̄) = d,
∀i : 1 ≤ i ≤ n ⇒ v(bi ) ∈ I}
A goal-directed execution strategy, SLGD , using tabling for the above CLP fixpoint semantics has been developed by Toman in (Toman 1997b). This SLGD semantics encompasses the best of both top-down and bottom-up operational semantics: it is goal-directed like top-down evaluation and has the favorable termination
properties like bottom-up evaluation.
2.3.1 Basic SLGD Semantics
The SLGD semantics makes two assumptions about the constraint domain D.
Firstly, D includes a projection operation that returns a disjunction of constraints:
¯T C is used to state that Cj is one of the dis¯T C = W Ci . The notation Cj ∈ ∃
∃
i
juncts in this disjunction. Secondly, it is assumed that a relation ≤D is provided.
This relation should be at least as strong as implication, i.e.
∀C1 , C2 : C1 ≤D C2 ⇒ D |= C1 → C2
SLGD is formulated in terms of four resolution (or rewriting) rules, listed in
Table 1. These rules either expand existing tree nodes or create new root nodes.
There are four different kinds of tree nodes: root(G; C), body(G; B1 , . . . , Bk ; C),
goal(G; B, C ′ ; B2 . . . , Bk ; C) and ans(G; A) where G is an atom, B1, . . . , Bk are
literals, and C, C ′ , A are constraints1 in D.
An SLGD tree is built from a root(G; C) node using the resolution rules. An
SLGD forest is a set of SLGD trees. The meaning of the different resolution rules
is the following:
• The Clause Resolution rule expands a root node: for every matching clause
head a body node is created containing the clause’s body literals.
• If there is at least one literal in a body node, it is expanded by the Query
Projection rule into goal nodes. This rule selects a literal to be resolved.
The given Query Projection rule implements a left-to-right selection strategy,
which is common to most LP systems, including XSB. However, any other
strategy is valid as well. The current constraint store C is projected onto
the selected literal’s variables, yielding only the constraints relevant for that
1
Also known as constraint stores.
TCHR: a framework for tabled CLP
7
literal. As the projection yields a disjunction of constraints, one goal node is
created for every disjunct.
• If there is no literal in a body node, it is expanded by the Answer Projection
rule into a number of answer nodes. For this purpose the current constraint
store C is projected onto the goal’s variables, retaining only those constraints
relevant to the goal. In this way variables local to the chosen clause’s body
are eliminated.
• A goal node is expanded into new body nodes by the Answer Propagation rule.
This rule substitutes the selected literal by its answers: the selected literal’s
answer constraint stores are incorporated in the current store.
Note that the Answer Propagation and Answer Projection rules cooperate: whenever a new answer is produced, it is propagated to all the nodes that have already
been resolved using answers from this tree. Also the Answer Propagation rule is
responsible for creating new SLGD trees: when no tree with a root node that subsumes the goal (B, C ′ ) to be resolved can be found, a node root(B, C ′ ) is created
to start a separate tree.
Finally, a query in the SLGD formalism is a tuple (G, C, P ) where vars(C) ⊆
vars(G) and all arguments of G are variables. The SLGD resolution rules are used
for query evaluation as follows:
1. create an SLGD forest containing a single tree {root(G, C)},
2. expand the leftmost node using the resolution rules as long as they can be
applied, and
3. return the set ans(G, C) as the answers for the query.
Definition 2
(Answer Set) The answer set ans(G, C) is the set of all A such that ans(G; A) ∈
slg(G, C), where slg(G, C) is the SLGD tree rooted at root(G, C).
2.3.2 An Example
Let us consider the following very simple CLP program P :
p(X) :- X = 1-Y, q(Y).
p(X) :- X = 0.
q(X) :- true, p(X).
The constraint domain is that of domain integers. The supported basic constraint
=/2 is equality of arithmetic expressions.
Figure 1 depicts the SLGD forest for the query (p(U ); true; P ). The full arrows
represent the SLGD tree branches, whereas the dashed arrow indicates the start of
a new tree and the dotted arrows indicate the propagation of new answers. Each
arrow is labeled with its step number.
The answer set ans(p(U ), true) consists of two answers: U = 0 and U = 1.
8
T. Schrijvers et al.
Parent
Children
Conditions
Clause Resolution
8
body(G; B11 , . . . , Bk11 ; C ∧ θ ∧ D1 )
>
>
<
..
root(G; C)
.
>
>
:
l
body(G; B1 , . . . , Bkl l ; C ∧ θ ∧ Dl )
for all 0 < i ≤ l such that
G′ → Di , B1i , . . . , Bki i
and θ ≡ (G = G′ )
and C ∧ θ ∧ Di is satisfiable
Query Projection
8
goal(G; B1 , C1 ; B2 . . . , Bk ; C)
>
>
<
..
body(G; B1 , . . . , Bk ; C)
.
>
>
:
goal(G; B1 , Cl ; B2 . . . , Bk ; C)
¯B C
for all Ci ∈ ∃
1
Answer Propagation
8
body(G; B2 , . . . , Bk ; C ∧ θ ∧ A1 )
>
>
<
..
goal(G; B1 , C1 ; B2 . . . , Bk ; C)
.
>
>
:
body(G; B2 , . . . , Bk ; C ∧ θ ∧ Al )
for all Ai ∈ ans(B ′ , C ′ )
where θ ≡ (B ′ = B1 )
and C1 ∧ θ ≤D C ′
and C ∧ θ ∧ Ai is satisfiable
Answer Projection
8
ans(G; A1 )
>
>
<
..
body(G; ; C)
.
>
>
:
ans(G; Al )
¯G C
for all Ai ∈ ∃
Table 1. SLGD resolution rules
Parent
Children
Optimized Query Projection
8
goal(G; B1 , C1 ; B2 . . . , Bk ; C)
>
>
<
..
body(G; B1 , . . . , Bk ; C)
.
>
>
:
goal(G; B1 , Cl ; B2 . . . , Bk ; C)
Conditions
¯ B C → C1 ∨ . . . ∨ Cl
D |= ∃
1
for some C1 , . . . , Cl
Table 2. Optimized Query Projection for SLGD resolution
2.3.3 SLGD Optimizations
Several optimizations to the rewriting formulas have been proposed by Toman, of
which one, Query Projection, is of particular interest to us. The optimization allows
for more general goals than strictly necessary to be resolved. In this way fewer goals
have to be resolved, as distinct specific queries can be covered by the same general
goal.
Table 2 lists the modified Query Projection rule, called Optimized Query Projection.
A second important optimization is a modified version of the answer set definition:
9
TCHR: a framework for tabled CLP
root(p(U ); true)
TTTT
TTT(1:CR)
TTTT
TTTT
T*
ggg
gggg
g
g
g
g
gggg
s ggg
g
(1:CR)
body(p(U ); ; U = 0)
body(p(U ); q(Y ); U = 1 − Y )
(6:AP j)
(2:QP )
goal(p(U ); q(Y ), true; ; U = 1 − Y )
ans(p(U ); U = 0)
WWWWW
WWW(13:AP
WWWW p)
F
(9:AP p)
WWWW
F
WWW+
F
F
body(p(U ); ; U = 1 − Y ∧ Y =F0) body(p(U ); ; U = 1 − Y ∧ Y = 1)
f
C
F (3)
F
(10:Apj)
(14:AP j)
F
F
F
ans(p(U ); U = 1)
F ans(p(U ); U = 0)
F
(11)
F
F
F"
F
F
(7)
root(q(Y ); true)
(4:CR)
(13)
body(q(U ); p(U ); true)
(9)
(5:QP )
goal(q(U ); p(Y ), true; ; true)
gggg
gggg
g
g
g
gggg(7:AP p)
s ggg
u g
body(q(U ); ; U = 0)
TTTT
TTT(11:AP
TTTT p)
TTTT
T*
- body(q(U ); ; U = 1)
(8:AP j)
ans(q(U ); U = 0)
(12:AP j)
ans(q(U ); U = 1)
Fig. 1. Example SLGD forest
Definition 3
(Optimized Answer Set) The optimized answer set of the query (G, C, P ), denoted
ans(G, C), is the set of all A such that ans(G; A) ∈ slg(G, C) and no A′ is already
in ans(G, C) for which A ≤D A′ .
This alternative definition allows for answers to be omitted if they are already
entailed by earlier more general answers. While logically the same answers are
entailed, the set of answers is smaller with the new definition.
Note that SLG, the operational semantics of tabled Logic Programming, is in
fact a specialized form of the SLGD semantics for the Herbrand domain. Several
implementations of SLG exist, including XSB. The topic of this paper, the integration of CHR with tabled execution, is in effect an implementation of SLGD for
arbitrary D defined by a CHR program.
In (Toman 1996) Toman has also extended his work to a goal-directed execution
10
T. Schrijvers et al.
strategy for CLP programs with negation. This extension realizes the well-founded
semantics. An implementation of this extension is not covered by our work. It
imposes additional requirements on the constraint solver: a finite representation
of the negation of any constraint should exist. Moreover, the detection of loops
through negation requires a more complicated tabling mechanism.
3 Constraint Handling Rules
In this section we give a brief overview of Constraint Handling Rules (CHR) (Frühwirth 1998;
Frühwirth and Abdennadher 2003).
3.1 Syntax of CHR
We use two disjoint sets of predicate symbols for two different kinds of constraints:
built-in (pre-defined) constraint symbols which are solved by a given constraint
solver, and CHR (user-defined) constraint symbols which are defined by the rules
in a CHR program. There are three kinds of rules:
Simplification rule:
Propagation rule:
Simpagation rule:
N ame @ H <=> C | B,
N ame @ H ==> C | B,
N ame @ H \ H ′ <=> C | B,
where Name is an optional, unique identifier of a rule, the head H, H ′ is a non-empty
comma-separated conjunction of CHR constraints, the guard C is a conjunction of
built-in constraints, and the body B is a goal. A query is a conjunction of built-in
and CHR constraints. A trivial guard expression “true |” can be omitted from a
rule. The head of a simplification rule is called a removed head, as the rule replaces
its head by its body. Similarly, the head of a propagation rule is called a kept head,
as the rule adds its body in the presence of its head. Simpagation rules abbreviate
simplification rules of the form N ame @ H, H ′ <=> C | H, B, i.e. H is a kept
head and H ′ a removed head. A CHR program P consists of an ordered set of CHR
rules.
3.2 Operational Semantics of CHR
The formal operational semantics of CHR is given in terms of a state transition
system in Figure 2. The program state is an indexed 4-tuple hG, S, B, T in . The
first part of tuple, the goal G is the multiset of constraints to be rewritten to solved
form. The CHR constraint store S is the multiset of identified CHR constraints
that can be matched with rules in the program P . An identified CHR constraint
c#i is a CHR constraint c associated with some unique integer i, the constraint
identifier. This number serves to differentiate among copies of the same constraint.
We introduce the functions chr (c#i) = c and id (c#i) = i, and extend them to
sequences, sets and multisets of identified CHR constraints in the obvious manner,
e.g. chr (S) = {c | c#i ∈ S}.
The built-in constraint store B is the conjunction of all built-in constraints that
TCHR: a framework for tabled CLP
11
1. Solve h{c} ⊎ G, S, B, T in solve hG, S, c ∧ B, T in where c is a built-in constraint.
2. Introduce h{c} ⊎ G, S, B, T in introduce hG, {c#n} ⊎ S, B, T i(n+1) where c
is a CHR constraint.
3. Apply hG, H1 ⊎ H2 ⊎ S, B, T in apply hC ⊎ G, H1 ⊎ S, θ ∧ B, T ′ in where there
exists a (renamed apart) rule in P of the form
r @ H1′ \ H2′ ⇐⇒ g | C
and a matching substitution θ such that chr(H1 ) = θ(H1′ ), chr(H2 ) = θ(H2′ ) and
¯ B (θ ∧ g). In the result T ′ = T ∪ {id(H1 ) ++ id(H2 ) ++ [r]}. It
Db |= B → ∃
should hold that T ′ 6= T .
Fig. 2. The transition rules of the operational semantics of CHR
have been passed to the underlying solver. Since we will usually have no information
about the internal representation of B, we will model it as an abstract logical
conjunction of constraints. The propagation history T is a set of sequences, each
recording the identities of the CHR constraints that fired a rule, and the name of the
rule itself. This is necessary to prevent trivial non-termination for propagation rules:
a propagation rule is allowed to fire on a set of constraints only if the constraints
have not been used to fire the same rule before. Finally, the counter n represents
the next free integer that can be used to number a CHR constraint.
Given an initial query G, the initial program state is: hG, ∅, true, ∅i1 .
The rules of a program P are applied to exhaustion on this initial program
state. A rule is applicable, if its head constraints are matched by constraints in the
current CHR store one-by-one and if, under this matching, the guard of the rule is
implied by the built-in constraints in the goal. Any of the applicable rules can be
applied, and the application cannot be undone, it is committed-choice (in contrast
to Prolog). When a simplification rule is applied, the matched constraints in the
current CHR store are replaced by the body of the rule; when a propagation rule is
applied, the body of the rule is added to the goal without removing any constraints.
3.3 Implementation of CHR
This high-level description of the operational semantics of CHR leaves two main
sources of non-determinism: the order in which constraints of a query are processed
and the order in which rules are applied.2 As in Prolog, almost all CHR implementations execute queries from left to right and apply rules top-down in the textual
order of the program. This behavior has been formalized in the so-called refined
semantics that was also proven to be a concretization of the standard operational
semantics (Duck et al. 2004).
2
The nondeterminism due to the wake-up order of delayed constraints and multiple matches for
the same rule are of no relevance for the programs discussed here.
12
T. Schrijvers et al.
In this refined semantics of actual implementations, a CHR constraint in a query
can be understood as a procedure that goes efficiently through the rules of the
program in the order they are written, and when it matches a head constraint of
a rule, it will look for the other, partner constraints of the head in the constraint
store and check the guard until an applicable rule is found. We consider such a
constraint to be active. If the active constraint has not been removed after trying
all rules, it will be put into the constraint store. Constraints from the store will
be reconsidered (woken) if newly added built-in constraints constrain variables of
the constraint, because then rules may become applicable if their guards are now
implied.
The refined operational semantics is implemented by all major CHR systems,
among which the K.U.Leuven CHR system (Schrijvers and Demoen 2004). This
system is currently available in three different Prolog systems (hProlog (Demoen 2004),
SWI-Prolog (Wielemaker 2004) and XSB) and it serves as the basis of our integration with tabled execution in this paper.
The K.U.Leuven CHR system (Schrijvers and Demoen 2004) is based on the general compilation schema of CHR by Holzbaur (Holzbaur and Frühwirth 2000). For
this paper (Section 5) it is relevant to know that the CHR constraint store is implemented as a global updatable term, containing identified constraints, in this context
also called suspended constraints, grouped by their functor. Each suspended constraint c#i is represented as a suspension term, including the following information:
• The constraint c itself.
• The constraint identifier i.
• The continuation goal, executed on reactivation. This goal contains the suspension itself as an argument and it is in fact a cyclic term.
• The part of the propagation history T containing for each propagation rule
the tuple of identifiers of other constraints that this constraint has interacted
with.
Variables involved in the suspended constraints behave as indexes into the global
store: they have the suspensions attached to them as attributes. Because we aim
towards a light-weight integration of CHR with tabled Logic Programming, we
do not question these established representation properties, but consider them as
something to cope with.
We refer the interest reader to (Schrijvers 2005) for more details on CHR implementation.
3.4 CHR for Constraint Solving
The CHR language is intended as a language for implementing constraint solvers.
A CHR program P is a constraint solver for the constraint domain DP whose
constraint symbols ΠP are the CHR and built-in constraint symbols. The constraint
theory TP of the program consists of the built-in constraint theory together with the
declarative meaning of the CHR rules. The declarative meaning of a simplification
TCHR: a framework for tabled CLP
13
rule of the form Hr <=> G | B is:
∀x̄.∃ȳ.G → (H ↔ ∃z̄.B)
where x̄ = vars(H)∪, ȳ = vars(G)\vars(H) and z̄ = vars(B)\(vars(H)∪vars(G)).
Similarly, the declarative meaning of a propagation rule of the form H ==> G | B
is:
∀x̄.∃y.G → (H → ∃z̄B)).
Value sets are not explicitly defined by the CHR program, but they exist implicitly
in the intention of the programmer.
See (Frühwirth and Abdennadher 2003) for an extensive treatment of CHR for
writing constraint solvers.
4 The TCHR Framework
The main challenge of introducing CHR in XSB is the integration of CHR constraint solvers with the backward chaining fixpoint computation of SLG resolution
according to the SLGD semantics of the previous section.
A similar integration problem has been solved in (Cui and Warren 2000a), which
describes a framework for constraint solvers written with attributed variables for
XSB. The name Tabled Constraint Logic Programming (TCLP) is coined in that
publication, though it is not formulated in terms of SLGD resolution. Porting CHR
to XSB was already there recognized as important future work.
CHR is much more convenient for developing constraint solvers than attributed
variables, because of its high-level nature. This advantage should be carried over to
the tabled context, making tabled CHR a more convenient paradigm than TCLP
with attributed variables. Indeed, we will show how the internal details presented
in the current section can be hidden from the user.
In (Cui and Warren 2000a) the general TCLP framework specifies three operations to control the tabling of constraints: call abstraction, entailment checking of
answers and answer projection. These operations correspond with the optimization
to Query Projection, the projection in Answer Projection and the compaction of
the ans(G; C) set. It is left to the constraint solver programmer to implement these
operations for his particular solver.
In the following we formulate these operations in terms of CHR. The operations
are covered in significant detail as the actual CHR implementation and the encoding
of the global CHR constraint store are taken into account.
4.1 General Scheme of the TCHR Implementation
The objective of TCHR is to implement SLGD semantics for an arbitrary constraint
domain D which is implemented as a CHR constraint solver. For this purpose we
have both an SLGH implementation3 , i.e. the SLG implementation of XSB, and an
SLDD implementation, i.e. the CHR implementation of XSB, at our disposal.
3
SLG is a special case of SLGD where D is the Herbrand constraint domain H.
14
T. Schrijvers et al.
Hence, we aim for the simplest and least intrusive solution. That is:
1. We use the unmodified CHR implementation for constraint solving.
2. We use the unmodified SLG implementation for tabled execution.
3. At the intersection of point 1 and point 2 we transform back and forth between
the CHR D constraints and an encoding of these as H constraints.
The advantages to this lightweight approach are twofold. Firstly, it is straightforward to realize the full expressivity of SLGD within an existing system. Secondly,
it does not affect existing programs or their performance. On the downside we note
that TCHR performance and, in particular, constant factors involved are not optimal. However, CHR on its own does not aim towards performance in the first place,
but rather towards being a highly expressive formalism for experimenting with new
constraint solvers. Similarly, we see the TCHR framework as a highly expressive
prototyping system for exploring new applications. It does offer some high-level
means to affect performance, and when the resulting performance is simply not
good enough, one may decide to reimplement the established high-level approach
in a lower-level language.
Now we look at our solution in more detail. As points 1 and 2 leave the system
untouched, we only have to consider implementing point 3, translating between
constraint encodings.
First let us consider the different kinds of nodes used in SLGD . Of the tree nodes,
only the root and answer nodes are manifestly represented by SLGH implementations like XSB, in respectively call and answer tables. Hence these two nodes require
the constraint store to be in H encoding form. The other two nodes, the goal and
body nodes, are implicit in the execution mechanism. So here we are free to use the
form that suits us best.
With these formats for the nodes in mind, we consider one by one the different
resolution rules:
Clause Resolution The rule is depicted again below with each constraint annotated with its type of encoding: H for Herbrand encoding and CHR for natural
CHR encoding. The constraint store C is initially Herbrand encoded in the root
node and has to be decoded into its natural CHR form for solving C ∧ θ ∧ Di with
the CHR solver. The CHR solver either fails, if the conjunction is not satisfiable,
or returns a simplified form of the conjunction.
8
1
)
body(G; B11 , . . . , Bk11 ; CCHR ∧ θ ∧ DCHR
>
>
<
..
root(G; CH )
.
>
>
:
l
l
l
)
body(G; B1 , . . . , Bkl ; CCHR ∧ θ ∧ DCHR
∀i.0 < i ≤ l such that
i
G′ → DCHR
, B1i , . . . , Bki i
and θ ≡ (G = G′ )
i
and CCHR ∧ θ ∧ DCHR
is
satisfiable
Optimized Query Projection This rule directly starts with a constraint store
in the natural CHR constraint form, and projects it onto the first literal and
subsequently generalizes it. A CHR program does not normally come with such a
combined projection & generalization operation, so one will have to be supplied
by the TCHR framework: the call abstraction. Section 6 discusses what kind of
generic projection operation the TCHR framework implements.
TCHR: a framework for tabled CLP
15
8
1
goal(G; B1 , CCHR
; B2 . . . , Bk ; CCHR )
>
>
<
¯ B C → C1 ∨ . . . ∨ Cl
..
D |= ∃
1
body(G; B1 , . . . , Bk ; CCHR )
.
>
for
some
C1 , . . . , C l
>
:
l
goal(G; B1 , CCHR
; B2 . . . , B k ; CCHR )
Answer Propagation The answers consumed by this rule have to be decoded
from Herbrand form for the implication check and the satisfiability check. A
CHR program does not normally come with an implication check, so one will
have to be supplied here by the TCHR framework. This is covered together with
call abstraction in Section 6.
i
8
body(G; B2 , . . . , Bk ; CCHR ∧ θ ∧ A1CHR )
>
>
<
..
1
goal(G; B1 , CCHR
; B2 . . . , Bk ; CCHR )
.
>
>
:
body(G; B2 , . . . , Bk ; CCHR ∧ θ ∧ AlCHR )
′
∀AH ∈ ans(B ′ , CH
)
where θ ≡ (B ′ = B1 )
1
′
and CCHR
∧ θ ≤D CCHR
i
and CCHR ∧ θ ∧ AH is satisfiable
Answer Projection Again a projection is performed on the CHR constraint representation. This instance of of projection we call answer projection. Like answer
projection, it is to be supplied by the framework. In Section 7 the details of this
operation within the framework are elaborated.
8
ans(G; A1H )
>
>
<
..
¯G C
body(G; ; CCHR )
for all AiCHR ∈ ∃
.
>
>
:
ans(G; AlH )
Having established what new operations and mappings to include in the framework, we should consider how these are to be incorporated into the existing SLGH
system XSB. Recall that we did intend not to modify the system to incorporate
our encoding/decoding and projection operations in order to keep the integration
light-weight. Neither do we want to encumber the programmer with this tedious
and rather low-level task. Instead we propose an automatic source-to-source transformation based on a simple declaration to introduce these operations.
The source-to-source transformation maps the SLGD program P onto the SLGH
program P ′ . In the mapping every predicate p/n ∈ P is considered independently,
and mapped onto three predicates p/n, tabled p/(n+ 2), original p/n ∈ P ′ . T maps
the SLGD program P onto the SLGH program P ′ = T(P ).
We outline the high-level transformation for a single predicate p/2:
:- table p/2.
p(X,Y) :- Body.
The three resulting predicates are:
p(X,Y) :encode_store(CurrentStoreEncoding),
call_abstraction([X,Y],CurrentStoreEncoding,AbstractStoreEncoding),
empty_store,
tabled_p(X,Y,AbstractStoreEncoding,AnswerStoreEncoding),
decode_store(CurrentStoreEncoding),
decode_store(AnswerStoreEncoding).
16
T. Schrijvers et al.
:- table tabled_p/4.
tabled_p(X,Y,StoreEncoding,NStoreEncoding) :decode_store(StoreEncoding),
original_p(X,Y),
encode_store(StoreEncoding1),
empty_store,
project([X,Y],StoreEncoding1,NStoreEncoding).
original_p(X,Y) :- Body.
The new predicate p/2 is a front to the actual tabled predicate tabled p/4.
This front allows the predicate to be called with the old calling convention where
the constraint store is implicit, i.e. in the natural CHR form. Thanks to this front
the transformation is modular: we do not have to modify any existing calls to
the predicate, either in other predicates’ bodies, its own body Body or in queries.
The auxiliary predicate encode store/1 returns a Herbrand encoding of the current (implicit) constraint store and the call abstraction/3 predicate projects
the Herbrand encoded store onto the call arguments. Then the implicit constraint
store is emptied with empty store/0 so as not to interfere with the tabled call,
which has the Herbrand encoded stores as manifest answers. Finally, the predicate decode store/1 decodes the Herbrand encoding and adds the resulting CHR
constraint store to the implicit store. In p/2 this predicate is called twice: first to
restore the current constraint store and then to add to it the answer constraint
store of the tabled call.
The tabled p/4 predicate is the tabled predicate. In its body the encoded input
store is decoded again, then the original predicate code original p/2 is run, the
resulting store is encoded again and projected onto the call arguments. Again the
implicit store is emptied so as not to interfere with the caller.
Note that this is only a high-level outline of the mapping. In practice the scheme
is specialized for the concrete operations. This is discusses later as we discuss each
of the operations in detail.
The above transformation of both predicates and queries can be fully transparent.
All the user has to do is to indicate what predicates have to be tabled, i.e. add a
declaration of the form
:- table_chr p(_,chr) with Options.
p(X,Y) :- ...
meaning that the predicate p/2 should be tabled, its first argument is an ordinary
Prolog term and its second argument is a CHR constraint variable. An (optional)
list of additional options Options may be provided to control the transformation:
encoding(EncodingType)
17
TCHR: a framework for tabled CLP
Section 5 studies two alternative encodings of the Herbrand constraint store. This
option allows the user to choose between them.
projection(PredName )
The projection applied in the Answer Projection rule is addressed in Section 7.
This projection is realized as a call to a projection predicate that reduces the
constraint store to its projected form.
canonical form(PredName )
answer combination(PredName )
These two options relate to optimizations of the answer set, based on Definition
3 and a novel generalization of this principle. It is discussed in Section 8.
Figure 3 summarizes the different steps in handling a call to a tabled predicate.
call
abstract call
no
in table?
project
answer
execute
call
yes
yes
store in
table
yes
no
entailed?
more
answers?
no
combine answer
with call
Fig. 3. Tabled call flowchart
5 Herbrand Constraint Store Encodings
In this section we present two alternative Herbrand constraint store encodings. An
encoding must have the following properties:
• The encoding has to be suitable for passing it as an argument in a predicate
and for storing it in an answer table.
• It should be possible to convert from the natural CHR constraint form (see
Section 3.3) and back, for insertion into the call table and retrieval from the
answer table.
The essential aspects of the ordinary CHR constraint store implementation have
been covered in Section 3.3. Two different Herbrand constraint store encodings
18
T. Schrijvers et al.
that are based on this ordinary form have been explored: the suspension encoding
and the goal encoding. The former is based on state copying and the latter on
recomputation. A discussion of their respective merits and weaknesses as well as an
evaluation follow in Sections 5.1 and 5.2 respectively.
One implicit aspect of CHR execution under the refined operational semantics is
the order in which constraints are processed. Ordering information is not maintained
explicitly. Without any additional support, it is not straightforward to maintain this
ordering information for tabled constraints. However, in the spirit of tabling, the
declarative meaning of a program rather than its operational behavior is of importance. For that reason we shall not attempt to realize the ordering of the refined
operational semantics. From the user’s point of view, the CHR constraints behave
according to the theoretical operational semantics and no assumptions should be
made about ordering.
5.1 Suspension Encoding
This encoding aims at keeping the tabled encoding as close as possible to the ordinary form. The essential issue is to retain the propagation history of the constraints.
In that way no unnecessary re-firing of propagation rules occurs after the constraints
have been retrieved from the table.
However, it is not possible to just store the ordinary constraint suspensions in
the table as they are. Fortunately, attributed variables themselves can be stored
in tables (see (Cui and Warren 2000b)), but two other aspects have to be taken
into account. Firstly, these suspensions are cyclic terms that the tables cannot
handle. This can be dealt with by breaking the cycles upon encoding and resetting
them during decoding. Secondly, the constraint identifiers have to be replaced by
fresh ones during decoding, as multiple calls would otherwise create multiple copies
of the same constraints all with identical identifiers. Finally, after decoding, the
constraints have to be activated again in order to solve them together with the
already present constraints. This is done by simply calling their continuation goals.
Example 2 Let us consider the following program:
:- constraint a/0, b/0.
r1 @ a ==> b.
:- chr_table p.
p :- a.
and the query ?- p. After having fired rule r1, the suspension of an a constraint
looks like:
Sa = suspension(42,reactivate a(Sa),[r1-[42]])
where 42 is the identifier, reactivate a(Sa) is the continuation goal and [r1-[42]]
is the propagation history, which has recorded that rule r1 has fired with only the
TCHR: a framework for tabled CLP
19
constraint itself. The other suspension in the store would be for a b constraint:
Sb = suspension(43,reactivate b(Sb),[])
The suspension encoding for this store of two constraints would look like:
[ Sa / ID1 / suspension(ID1,reactivate_a(Sa),[[r1-[ID1]]])
, Sb / ID2 / suspension(ID2,reactivate_b(Sb),[])]
Upon decoding we simply unify ID1 and ID2 with fresh identifiers, and S1 and S2
with their corresponding suspension terms. The resulting well-formed suspension
terms are placed in the implicit CHR constraint store and finally the continuation
goals of both suspensions are called.
5.2 Goal Encoding
The goal encoding aims at keeping the information in the table in as simple a form
as possible: for each suspended constraint only the goal to impose this constraint
is retained in the table. It is easy to create this goal from a suspension and easy to
merge this goal back into another constraint store: it needs only to be called.
Whenever it is necessary the goal creates a suspension with a fresh unique identifier and inserts it into the constraint store.
The only information that is lost in this encoding is the propagation history. This
may lead to multiple propagations for the same combination of head constraints.
For this to be sound, a further restriction on the CHR rules is required: they should
behave according to set semantics, i.e. the presence of multiple identical constraints
should not lead to different answers modulo identical constraints.
Example 3 The goal encoding of the above example is:
[a, b]
and the decoding procedure simply calls a and b.
5.3 Evaluation
To measure the relative performance of the two presented encodings, consider the
following two programs:
prop
simp
:- constraints a/1.
:- constraints a/1, b/1.
a(0) <=> true.
b(0) <=> true.
a(N) ==> N > 0
b(N) <=> N > 0
| M is N - 1, a(M).
| a(N), M is N - 1, b(M).
p(N) :- a(N).
p(N) :- b(N).
20
T. Schrijvers et al.
Table 3. Evaluation of the two tabled store encodings.
program
prop
simp
no tabling
runtime space
10
0
10
0
encoding(suspension)
runtime
space
150
2,153,100
109
1,829,100
encoding(goal)
runtime
space
1,739 270,700
89 270,700
For both programs the predicate p(N) puts the constraints a(1)...a(N) in the constraint store. The prop program uses a propagation rule to achieve this while the
simp program uses an auxiliary constraint b/1. The non-tabled version of the query
p(N) has time complexity O(N ) for both the simp and the prop program.
The two possible encodings for the answer constraint store can be specified in
the tabling declaration as follows:
:- table_chr p(_) with [encoding(suspension)].
and
:- table_chr p(_) with [encoding(goal)].
Table 3 gives the results for the query p(400), both untabled and tabled using
the two encodings: runtime in milliseconds and space usage of the tables in bytes.
For both programs the answer table contains the constraint store with the 400 a/1
constraints.
Most of the space overhead is due to the difference in encoding: a suspension
contains more information than a simple call. However, the difference is only a
constant factor. The only part of a suspension in general that can have a size greater
than O(1) is the propagation history. In the prop program every a/1 constraint’s
history is limited to remembering that the propagation rule has been used once.
For the simp program the propagation history is always empty.
The runtime of the prop version with the suspension encoding is considerably
better than that of the version with the goal encoding. In fact, there is a complexity
difference. When the answer is retrieved from the table for the suspension encoding, the propagation history prevents re-propagation. Hence, answer retrieval takes
O(N ) time. However, for the goal encoding every constraint a(I) from the answer
will start propagating and the complexity of answer retrieval becomes O(N 2 ).
On the other hand, for simp the propagation history plays no role. The runtime
overhead is mostly due to the additional overhead of the pre- and post-processing
of the suspension encoding as opposed to the simpler form of the goal encoding. In
comparison, without tabling the query takes only 10 milliseconds for both programs.
6 Call Abstraction
In the call abstraction operation we combine the projection and generalization
operations of the Optimized Query Projection rule in the SLGD semantics.
The idea of both steps is to reduce the number of distinct SLGD trees, and hence
the number of tables. When a predicate is called with many different call patterns, a
TCHR: a framework for tabled CLP
21
table is generated for each such call pattern. Thus it is possible that the information
for one strongly constrained call is present many times in tables for different less
constrained call patterns. This duplication in the tables can be avoided by using
call abstraction to obtain a smaller set of call patterns.
The projection reduces the context of the predicate call, i.e. the constraint store,
to the constraints relevant for the call. In this way, two calls to p(X), respectively
with constraint stores {X > 5, Y > 7} and {X > 5, Z < 3} both yield the same
projected call store { X > 5 }. The subsequent generalization step goes even further, e.g. by relaxing bounds to the reference value 0 both constraint stores {X >
5} and {X > 10} become {X > 0}. Hence, call abstraction effectively is a means to
control the number of tables. At the level of SLGH , call abstraction means not passing certain bindings to the call. For example, p(q,A) can be abstracted to p(Q,A).
This goal has then to be followed by Q = q to ensure that only the appropriate
bindings for A are retained.
For SLGD , call abstraction can be generalized from bindings to constraints: abstraction means removing some of the constraints on the arguments. Consider for
example the call p(Q,A) with constraint Q =< N on Q. This call can be abstracted
to p(Q’,A), followed by Q’=Q to reintroduce the constraint.
Abstraction is particularly useful for those constraint solvers for which the number of constraints on a variable can be much larger than the number of different
bindings for that variable. Consider for example a finite domain constraint solver
with constraint domain/2, where the first argument is a variable and the second
argument the set of its possible values. If the variable has a domain of size n (i.e. it
contains n different values), the variable can take as many as 2n different domain/2
constraints, one for each subset of values. Thus many different tables would be
needed to cover every possible call pattern.
Varying degrees of abstraction are possible, depending on the particular constraint system or application. Full constraint abstraction, i.e. the removal of all
constraints from the call, is generally the only option for CHR, for the following
reasons:
• CHR rules do not require constraints to be on variables. They can be exclusively on ground terms or atoms as well. This is useful for various reasons. By
encoding constraint variables as ground terms, particular solving algorithms
can be used more conveniently or efficiently, e.g. the equation solving algorithm union-find has optimal time-complexity when using ground elements
(Schrijvers and Frühwirth 2006).
It is not straightforward to automatically define abstraction for ground terms
as these are not necessarily passed in as arguments but can just as well be created inside the call. Hence there is no explicit link with the call environment,
while such a link is needed for call abstraction. As such, only “no abstraction”
or full constraint abstraction seem suitable for CHR.
• Full constraint abstraction is preferable when the previously mentioned table
blow-up is likely.
• In order to reuse existing answers, existing calls are considered in the Answer
22
T. Schrijvers et al.
Propagation rule. These previous calls are compared to the new call using
the implication check ≤D . Unfortunately, such an implication check does not
come with the CHR solver. A special case of this subsumption-based tabling
is where ≤D is taken to be ↔, i.e. equivalence-based, or variant-based in
SLGH terminology, tabling. Unfortunately, even establishing the equivalence
of constraint stores is not directly supported by CHR solvers.
However, if the call constraint store is empty, i.e. true, this problem disappears: true implies true independent of the constraint domain.
Moreover, it may be costly to sort out what constraints should be passed in to
the call or abstracted away. Hence often full abstraction is cheaper than partial
abstraction. For instance, consider a typical propagation-based finite domain constraint solver with binary constraints only. The constraint graph for a number of
such finite domain constraints has a node for every variable involved in a constraint
and an edge between variables involved in the same constraint. Any additional constraint imposed on a variable in a component of the graph may affect the domain
of all other variables in the same component. Hence, call abstraction on a subset
of the variables involves a costly transitive closure of reachability in the constraint
graph.
Let us now revisit the transformation scheme of Section 4.1 for a predicate p/2,
and specialize it for full call abstraction:
p(X,Y) :encode_store(StoreEncoding),
empty_store,
tabled_p(X,Y1,NStoreEncoding),
decode_store(StoreEncoding),
decode_store(NStoreEncoding),
Y1 = Y.
:- table tabled_p/3.
tabled_p(X,Y,NStoreEncoding) :original_p(X,Y),
encode_store(StoreEncoding1),
empty_store,
project([X,Y],StoreEncoding1,NStoreEncoding).
original_p(X,Y) :- Body.
As we know that the call-abstracted constraint store is empty, we no longer need
to pass it as an argument to the tabled p predicate and decode it there. The only
effect of the call abstraction is then to replace the constraint variable Y with a
fresh variable Y1. This is necessary to prevent any constraints on Y from being
reachable through attributes on Y. The unification Y=Y1 at the end of p/2 is then
a specialization of the substitution θ ≡ (p(X, Y 1) = p(X, Y )) that appears in the
Answer Propagation rule.
TCHR: a framework for tabled CLP
23
7 Answer Projection
In most constraint domains, the same logical answer can be represented in many
different ways. For example, consider the predicate p/1.
p(X) :- X > 5.
p(X) :- X > 5, Y > 0.
Both X > 5 and X > 5/ Y > 0 represent the same answer to the call p(X)
concerning X. Constraints that do not relate to the call arguments, like Y > 0, are
meaningless outside of the call. The local variable Y is existentially quantified, and
cannot be further constrained to introduce unsatisfiability at a later stage.
It is the purpose of projection to restrict constraints to a set of variables of interest, and to eliminate other variables as much as possible. In our setting, the variables
of interest are the call arguments. For projection to be sound, already present but
not yet detected unsatisfiability should not be removed. A sufficient, but not necessary condition is for the constraint system to be complete, i.e. unsatisfiability is
detected immediately.
Projection is important in the context of tabling, because it may give logically
equivalent answers the same syntactical form. When two answers have the same
syntactical form, the are recognized as duplicates and only one is retained in the
table. A vital application of projection is when a predicate with an infinite number
of different answers may be turned into one with just a finite number of answers
by discarding the constraints on local variables.
Example 4 Consider this program:
path(From,To,X) :edge(From,To,X).
path(From,To,X) :path(From,Between,X), path(Between,To,X).
edge(a,a,X) :leq(X,Y),
leq(Y,1).
leq(X,X)
leq(X,Y) \ leq(Y,X)
leq(X,Y) \ leq(X,Y)
leq(X,Y) , leq(Y,Z)
<=>
<=>
<=>
==>
true.
X = Y.
true.
leq(X,Z).
It defines a path/3 predicate that expresses reachability in a graph represented by
edge/3 predicates. The first two arguments of both predicates are edges (origin and
destination) and the third is a constraint variable. Along every edge in the graph
some additional constraints may be imposed on this variable. In our example, the
graph consists of a single loop from edge a to itself. This loop imposes two less-thanor-equal-to constraints: leq(X,Y), leq(Y,1). The variable Y is a local variable and
the fourth rule for leq/2 derives that leq(X,1) also holds.
24
T. Schrijvers et al.
The query ?- path(A,B,X) determines the different paths. There are an infinite
number of paths in our simple graph, one for each non-zero integer n. A path for
n takes the loop n times. For every time the loop is taken a new variable Yi is
created and two more constraints leq(X,Yi) and leq(Yi,1) are added. Through
the propagation rule also an leq(X,1) is added for each time the loop is taken. The
second simpagation rule however removes all but one copy of this last constraint.
Even though there are an infinite number of answers, the constraints involving
the local variables Yi are of no interest and only the single leq(X,1) is relevant.
In general constraint projection onto a set of variables transforms a constraint
store into another constraint store in which only variables of the given set are involved. The form of the resulting constraint store strongly depends on the particular
constraint solver and its computation may involve arbitrary analysis of the original
constraint store.
We propose what we believe is an elegant CHR-based approach to projection. It
consists of a compact and high level notation.
The user declares the use of the CHR-based approach to projection as follows:
:- table_chr p(_,chr) with [projection(P redN ame)].
and implements the projection as a number of CHR rules that involve the special
P redN ame/1 constraint. This constraint has as its argument the set of variables
to project on.
The source-to-source transformation generates the predicate tabled p based on
this declaration:
:- table tabled_p/3.
tabled_p(X,Y,NStoreEncoding) :original_p(X,Y),
P redN ame([X,Y]),
encode_store(NStoreEncoding),
empty_store.
When no projection operation is supplied, the default action is to return the constraint store unmodified.
To implement the projection simpagation rules can be used to decide what constraints to remove. A final simplification rule at the end can be used to remove the
projection constraint from the store.
The following example shows how to project away all leq/2 constraints that
involve arguments not contained in a given set Vars:
project(Vars) \ leq(X,Y) <=>
\+ (member(X,Vars),member(Y,Vars)) | true.
project(Vars)
<=> true.
Besides removal of constraints more sophisticated operations such as weakening
are possible. E.g. consider a set solver with two constraints: in/2 that requires an
25
TCHR: a framework for tabled CLP
element to be in a set and nonempty/1 that requires a set to be non-empty. The
rules for projection could include the following weakening rule:
project(Vars) \ in(Elem,Set) <=>
member(Set,Vars),
\+ member(Elem,Vars) | nonempty(Set).
8 Answer Set Optimization
In this section we consider various ways for reducing the size of the answer set.
First, in Section 8.1 we consider the subsumption-based technique proposed by
Toman. This leads us to a sidetrack in Section 8.2 where we outline a technique
for dynamic programming through answer subsumption. Section 8.3 continues with
the main story: it is established that answer subsumption is suboptimal for general
constraint domains and a generalized approach is proposed instead. In Section 8.4
we relax the soundness condition of answer reduction and speculate on applications
to program analysis. Finally, in Section 8.5, we evaluate the two main approaches.
8.1 Answer Subsumption
Some of the answers computed for a tabled predicate may be redundant and so
need not be saved. The property is exploited by Definition 3, the Optimized Answer
Set definition. In terms of SLGH , consider for example that the answer p(a,X) is
already in the table of predicate p/2. Now a new answer, p(a,b) is found. This
new answer is redundant as it is covered by the more general p(a,X) that is already
in the table. Hence it is logically valid to not record this answer in the table, and
to simply discard it. This does not affect the soundness or completeness of the
procedure.
We can extend the idea of answer subsumption to CHR constraints. This path
length computation will serve as an illustration:
Example 5
dist(A,B,D) :- edge(A,B,D1), leq(D1,D).
dist(A,B,D) :- dist(A,C,D1), edge(C,B,D2), leq(D1 + D2, D).
Suppose appropriate rules for the leq/2 constraint are in the above program, where
leq means less-than-or-equal. The semantics are that dist(A,B,D) holds if there
is a path from A to B of length less than or equal to D. In other words, D is an upper
bound on the length of a path from A to B.
If the answer dist(n1,n2,D) :- leq(d1, D) is already in the table and a new
answer dist(n1,n2,D) :- leq(d2, D), where d1 =< d2, is found, then this new
answer is redundant. Hence it can be discarded. This does not affect the soundness,
since logically the same answers are covered.
A strategy for establishing implication is provided by the following property:
∀i ∈ {0, 1} : C1−i → Ci ⇐⇒ C0 ∧ C1 ↔ Ci
(8.1)
26
T. Schrijvers et al.
for any logical formulas C0 and C1 . In particular, consider C0 and C1 to be a
previous answer constraint store and a newly computed one. The strategy then is
as follows. At the end of the tabled predicate’s execution a previous answer store
C0 is merged with a new answer store C1 . After merging, the store is simplified and
propagated to C by the available rules of the CHR program P. This combines the
two answers into a new one. This mechanism can be used to check entailment of
one answer by the other: if the combined answer store S is equal to one of the two,
then that answer store entails the other.
A practical procedure is the following:
:- table tabled_p/3.
tabled_p(X,Y,NStoreEncoding) :original_p(X,Y),
project([X,Y]),
encode_store(StoreEncoding),
( previous_answer(p(X,Y,PrevStoreEncoding),AnswerID),
decode_store(PrevStoreEncoding),
encode_store(Conjunction),
( Conjunction == PrevStoreEncoding ->
del_answer(AnswerID),
fail
;
Conjunction \== StoreEncoding
) ->
fail
;
NStoreEncoding = StoreEncoding
),
empty_store.
After computing, projecting and Herbrand encoding a new answer store C1 , we
look at previous answer stores C0 . We assume that there is a built-in predicate
previous answer/3 for this purpose, that backtracks over previous answers and
also provides a handle AnswerID to the returned answer. As the previous answer
store is still in Herbrand encoding, we decode it. This has the simultaneous effect of
adding it to the new implicit CHR constraint store that is still in place, i.e. it computes C0 ∧ C1 . This resulting conjunction is Herbrand encoded for further comparison. Syntactical equality (≡) is used as a sound approximation of the equivalence
check for the first equivalence sign (↔) in the Formula 8.1. If the conjunction equals
the previous answer PrevStoreEncoding, then that previous answer is implied by
the new answer and hence obsolete. We use the built-in predicate del answer to
erase it from the answer table and we backtrack over alternative previous answers.
Otherwise, if the conjunction does not equal the new answer, then neither implies
the other and we also backtrack over alternative previous answers. However, if the
conjunction equals the new answer StoreEncoding, that means it is implied by the
TCHR: a framework for tabled CLP
27
previous answer. Hence we fail, ignoring further alternative previous answers. If on
the other hand, the resulting answer is not implied by any previous answers, then
it is a genuinely new answer and is stored in the answer table.
Example 6 Consider again the dist/3 example, and assume that the answer
stores {leq(7,D}, {leq(3,D))} and {leq(5,D))} are successively produced for
the query ?- dist(a,b,D). When the first answer, {leq(7,D} is produced, there
are no previous answers, so it makes its way into the answer table. For the second
answer, {leq(3,D))} there is already a previous answer {leq(7,D}, so both are
conjoined. The following rule leq/2 rule simplifies the conjunction to retain the
more general answer:
leq(N1,D) \ leq(N2,D) <=> N1 >= N2 | true.
Hence, the resulting solved form of the conjunction is {leq(D,7)}, or in other words
the previous answer. In other words, this previous answer is implied by the new
answer. So it is deleted from the answer table and the new answer is recorded.
Finally, following the same procedure we discover that the third answer is already
implied by the second one. So the final answer set contains just the second answer.
Note that the dist/3 program would normally generate an infinite number of
answers for a cyclic graph, logically correct but not terminating. However, if it is
tabled with answer subsumption, it does terminate for non-negative weights. Not
only does it terminate, it only produces one answer, namely dist(n1,n2,D) :leq(d,D) with d the length of the shortest path. Indeed, the predicate only returns
the optimal answer.
The syntactical equality check on the Herbrand encoding is in general only an
approximation of a proper equivalence check. An option for the table chr declarations allows to improve its effectiveness: canonical form(PredName ) specifies the
name of the predicate that should compute the (approximate) canonical form of
the Herbrand encoded answer constraint store. This canonical form is used to check
equivalence of two constraint stores.
Example 7 Both [leq(1,X),leq(X,3)] and [leq(X,3),leq(1,X)] are permutations of the same Herbrand constraint store encoding. Obviously, based on a simple
syntactic equality check they are different. However, they can both be reduced to
the same canonical form, e.g. with the help of the Prolog built-in sort/2.
We refer to (Schrijvers et al. 2006) for a more elaborated discussion of the property 8.1 and an alternative, more elaborate implementation of the implication checking strategy in CHR.
In contrast to our generic approach above, the traditional approach in CLP is for
the solver to provide a number of predefined ask constraints (Saraswat and Rinard 1990),
i.e. subsumption checks for primitive constraints. These primitive ask constraints
can then be combined to form more complicated subsumption checks (Duck et al. 2004).
We have avoided this approach because it puts a greater burden on the constraint
solver implementer, who has to provide the implementation of the primitive ask con-
28
T. Schrijvers et al.
straints. In future work, we could incorporate user-defined ask constraints in our
generic approach for greater programmer control over performance and accuracy of
subsumption tests.
8.2 Dynamic Programming through Answer Subsumption
The technique used in the dist/3 program is to replace the computation of the
exact distance of a path with the computation of an upper bound on the distance
via constraints. Then, by tabling the predicate and performing answer subsumption,
the defining predicate has effectively been turned into an optimizing one, computing
the length of the shortest path. It is a straightforward yet powerful optimization
technique that can be applied to other defining predicates as well, turning them
into optimizing (dynamic programming) predicates with a minimum of changes.
In comparison, the usual approach consists in explicitly computing the list of all
answers, e.g. using Prolog’s findall/3 meta-programming built-in, and in processing this list of answers. Guo and Gupta (Guo and Gupta 2004) have added a specific
feature to tabled execution to realize this dynamic programming functionality. In
adding support for CHR to tabling, we get this functionality for free.
8.3 General Answer Compaction
Definition 3 yields a sound approach for reducing the size of answer tables. However,
we have discovered that it is only a special case of what is really possible. Therefore,
we propose the following generalized definition of answer sets, compacted answer
set, which covers all sound approaches for reducing the answer set size.
Definition 4 (Compacted Answer Set )
A compacted answer set of the query (G, C, P ), denoted ans(G, C), is a set such
that:
• No new fully instantiated (i.e. ground) answers are introduced:
∀A, θ : (A ∈ ans(G, C)) ∧ (D ⊢ Aθ)
=⇒
∃A′ , θ′ : (ans(G; A′ ) ∈ slg(G, C)) ∧ (A′ θ′ ≡ Aθ)
(8.2)
• All fully instantiated answers are covered:
∀A′ , θ′ : (ans(G; A′ ) ∈ slg(G, C)) ∧ (D ⊢ A′ θ)
=⇒
∃A, θ : (A ∈ ans(G, C)) ∧ (Aθ ≡ A′ θ′ )
(8.3)
• The answer set is more compact than the individual answers:
#ans(G, C) ≤ #{A′ |ans(G; A′ ) ∈ slg(G, C)}
(8.4)
where A and A′ are constraint stores and θ and θ′ are valuations.
Note that an optimized answer set is a special instance of a compacted answer
set and certainly, for Herbrand constraints, it is an optimal strategy, because:
H |= ∀H0 , H1 , H : H ↔ H0 ∨ H1 =⇒ ∃i ∈ {0, 1} : H ↔ Hi
(8.5)
TCHR: a framework for tabled CLP
29
where H0 , H1 , H are conjunctions of Herbrand equality constraints. In other words,
for finding a single Herbrand constraint that covers two given ones, it is sufficient
to considering those two.
Unfortunately, a similar property does not hold for all constraint domains: a
single constraint store may be equivalent to the disjunction of two others, while it
is not equivalent to either of the two. For example, leq(X , Y ) ∨ leq(Y , X ) ↔ true
and yet we have that neither leq(X , Y ) ↔ true nor leq(Y , X ) ↔ true.
Nevertheless, checking whether one answer subsumes the other is a rather convenient strategy, since it does not require any knowledge on the particularities of
the used constraint solver. That makes it a good choice for the default strategy
for CHR answer subsumption. Better strategies may be supplied for particular constraint solvers through the option answer combination(PredName ). It specifies the
name of the predicate that returns the disjunction of two given answer stores, or
fails if it cannot find one.
Example 8 Consider a simple interval-based solver, featuring constraints of the
form X ∈ [L, U ], where X is a constraint variable and L and U are integers, and
the rules:
X ∈ [L,U] ==> L =< U.
X ∈ [L1,U1], X ∈ [L2,U2] <=> X ∈ ([L1,U1] ∩ [L2,U2]).
For this solver, the subsumption approach merges two constraints X ∈ [L1 , U1 ]
and X ∈ [L2 , U2 ] iff [L1 , U1 ] ⊆ [L2 , U2 ] or [L2 , U2 ] ⊆ [L1 , U1 ]. However, it fails to
work for e.g. X ∈ [1, 3] and X ∈ [2, 4]. Nevertheless there is a single constraint form
that covers both: X ∈ [1, 4]. An optimal answer combinator in this case is one that
returns the union of two overlapping intervals. This also captures the subsumption
approach. If the intervals do not overlap, there is no single constraint that covers
both without introducing new answers.
Note that the idea of general answer compaction is not specific implementation
of constraints, and, in particular, should apply to non-CHR constraint solvers too.
8.4 Relaxed Answer Compaction Semantics
For some applications the soundness condition of answer generalization can be relaxed. An example in regular Prolog would be to have two answers p(a,b) and
p(a,c) and to replace the two of them with one answer p(a,X). This guarantees
(for positive programs) that no answers are lost, but it may introduce extraneous
answers. In other words, property 8.3 is preserved while property 8.2 is not. A similar technique is possible with constrained answers. While this approach is logically
unsound, it may be acceptable for some applications if only answer coverage is
required.
An example is the use of the least upper bound (lub) operator to combine answers
in the tabled abstract interpretation setting of (Codish et al. 1998). There is often
30
T. Schrijvers et al.
a trade-off between accuracy and efficiency in space and time. By exploiting this
trade-off abstract interpretation can remain feasible in many circumstances.
Toman has explored in (Toman 1997a) the use of CLP for program analysis and
compared it to abstract interpretation. In his proposal constraints serve as the abstractions of concrete values, and bottom-up computation or tabling is necessary to
reach a fixpoint over recursive program constructs. He notes that the CLP approach
is less flexible than actual abstract interpretation because it lacks flexible control
over the accuracy/efficiency trade-off. We believe that our proposal for relaxed answer compaction could function as a lub or widening operator to remedy this issue,
making Toman’s program analysis technique more practical. This remains to be
explored in future work.
8.5 Evaluation: A Shipment Problem
We evaluate the usefulness of the two proposed answer set optimization approaches
based on a shipment problem.
Problem statement: There are N packages available for shipping using trucks.
Each package has a weight and some constraints on the time to be delivered. Each
truck has a maximum load and a destination. Determine whether there is a subset
of the packages that can fully load a truck destined for a certain place so that all
the packages in this subset are delivered on time. (from (Cui 2000))
The problem is solved by the truckload program:
The truckload Program
:- constraints leq/2.
leq(X,X)
leq(N1,N2)
leq(N1,X) \
leq(X,N1) \
leq(X,Y) \
leq(X,Y) ,
<=>
<=>
leq(N2,X) <=>
leq(X,N2) <=>
leq(X,Y) <=>
leq(Y,Z) ==>
true.
number(N1), number(N2) | N1 =< N2.
number(N1), number(N2), N1 > N2 | true.
number(N1), number(N2), N1 < N2 | true.
true.
leq(X,Z).
truckload(0,0,_,_).
truckload(I,W,D,T) :I > 0,
I1 is I - 1,
truckload(I1,W,D,T).
truckload(I,W,D,T) :I > 0,
pack(I,Wi,D,T),
W1 is W - Wi,
W1 >= 0,
I1 is I - 1,
truckload(I1,W1,D,T).
% do not include pack I
% include pack I
TCHR: a framework for tabled CLP
31
pack(30,29,chicago,T) :- leq(19,T),leq(T,29).
pack(29,82,chicago,T) :- leq(20,T),leq(T,29).
pack(28,24,chicago,T) :- leq(8,T),leq(T,12).
%...
pack(3,60,chicago,T) :- leq(4,T),leq(T,29).
pack(2,82,chicago,T) :- leq(28,T),leq(T,29).
pack(1,41,chicago,T) :- leq(27,T),leq(T,28).
Packages are represented by a constraint database: clauses of pack/4, e.g.
pack(3,60,chicago,T) :- leq(4,T),leq(T,29).
means that the third package weights 60 pounds, is destined for Chicago and has to
be delivered between the 4th and the 29th day. The truckload/4 predicate computes the answer to the problem, e.g. :- truckload(30,100,chicago,T) computes
whether a subset of the packages numbered 1 to 30 exists to fill up a truck with a
maximum load of 100 pounds destined for Chicago. The time constraints are captured in the bound on the constraint variable T. There may be multiple answers to
this query, if multiple subsets exist that satisfy it.
We have run the program in four different modes:
• No Tabling: the program is run as is without tabling.
• Tabling - Plain: to avoid the recomputation of subproblems in recursive
calls the truckload/4 predicate is tabled with:
:- table_chr truckload(_,_,_,chr)
with [encoding(goal)].
• Tabling - Sorted: the answer store is canonicalized by simple sorting such
that permutations are detected to be identical answers:
:- table_chr truckload(_,_,_,chr)
with [encoding(goal),
canonical_form(sort)].
• Tabling - Combinator: we apply the custom answer combinator proposed
in Example 8: two answers with overlapping time intervals are merged into
one answer with the union of the time intervals. This variant is declared as:
:- table_chr truckload(_,_,_,chr)
with [encoding(goal),
answer_combination(interval_union)].
with interval union/3 the custom answer combinator.
Table 4 contains the runtime results of running the program in the four different modes for different maximum loads. Runtime is in milliseconds and has been
obtained on an Intel Pentium 4 2.00 GHz with 512 MB of RAM, with XSB 6.1
running on Linux 2.6.18. For the modes with tabling the space usage, in kilobytes,
32
T. Schrijvers et al.
No Tabling
Load
100
200
300
400
500
<1
160
2,461
12,400
> 5 min.
Plain
100
461
1,039
1,500
1,541
Tabling
Sorted Combinator
100
100
461
451
1,041
971
1,510
1,351
1,541
1,451
Table 4. Runtime results for the truckload program
Load
100
200
300
400
500
Plain
286
979
1,799
2,308
2,449
Tabling
Sorted Combinator
286
279
956
904
1,723
1,584
2,202
2,054
2,365
2,267
Table 5. Space usage for the truckload program
of the tables and number of unique answers have been recorded as well, in Table 5
and Table 6 respectively.
It is clear from the results that tabling has an overhead for small loads, but that
it scales much better. Both the modes with the canonical form and the answer
combination have a slight space advantage over plain tabling which increases with
the total number of answers. There is hardly any runtime effect for the canonical
form, whereas the answer combination mode is faster with increasing load.
In summary, canonicalization of the answer store and answer combination can
have a favorable impact on both runtime and table space depending on the particular problem.
9 Related and Future Work
The theoretical background for this paper, SLGD resolution, was realized by Toman
in (Toman 1997b). Toman establishes soundness, completeness and termination
properties for particular classes of constraint domains. While he has implemented
a prototype implementation of SLGD resolution for evaluation, no practical and
fully-fledged implementation in a Prolog system was done.
load
100
200
300
400
500
plain
324
2,082
4,721
5,801
4,972
tabling
sorted combinator
324
283
2,069
1,686
4,665
3,543
5,751
4,449
4,935
4,017
Table 6. Number of tabled answers for the truckload program
TCHR: a framework for tabled CLP
33
Various ad hoc approaches to using constraints in XSB were used in the past
by Ramakrishnan et al., such as a meta-interpreter (Mukund et al. 2000), interfacing with a solver written in C (Du et al. 2000) and explicit constraint store management in Prolog (Pemmasani et al. 2002). However, these approaches are quite
cumbersome and lack the ease of use and generality of CHR.
The most closely related implementation work that this paper builds on is (Cui and Warren 2000a),
which presents a framework for constraint solvers written with attributed variables.
Attributed variables are a much cruder tool for writing constraint solvers though.
Implementation issues such as constraint store encoding and scheduling strategies
that are hidden by CHR become the user’s responsibility when she programs with
attributed variables. Also in the tabled setting, the user has to think through all
the integration issues of the attributed variables solver. For CHR we have provided generic solutions that work for all CHR constraint solvers and more powerful
features can be accessed through parametrized options.
Guo and Gupta propose a technique for dynamic programming with tabling
(Guo and Gupta 2004) that is somewhat similar to the one proposed here. During
entailment checking a particular argument in a new answer is compared with the
value in the previous answer. Either one is kept depending on the optimization
criterion. Their technique is specified for particular numeric arguments whereas
ours is for constraint stores and as such more general. Further investigation of our
technique is certainly necessary to establish the extent of its applicability.
Part of this work was previously published at the International Conference of
Logic Programming (Schrijvers and Warren 2004) and the Colloquium on Implementation of Constraint and Logic Programming Systems (Schrijvers et al. 2003).
In (Schrijvers et al. 2003) we briefly discuss two applications of CHR with tabling in
the field of model checking. The integration of CHR and XSB has shown to make
the implementation of model checking applications with constraints significantly
easier. The next step in the search for applications is to explore more expressive
models to be checked than are currently viable with traditional approaches.
Further applications should also serve to improve the currently limited performance assessment of CHR with tabling. The shipment problem has given us some
indication of improved performance behavior in practice, but theoretical reasoning
indicates that slow-downs are a possibility as well.
The global CHR store has proven to be one of the main complications in tabling
CHR constraints. For particular CHR programs it is possible to replace the global
data structure with localized, distributed ones. Assessment (Sarna-Starosta and Ramakrishnan 2007)
of this approach has shown to be very promising.
Partial abstraction and subsumption are closely related. The former transforms a
call into a more general call while the latter looks for answers to more general calls,
but if none are available still executes the actual call. We still have to look at how
to implement partial abstraction and the implications of variant and subsumption
based tabling (Rao et al. 1996).
Finally, better automatic techniques for entailment testing, such as those of
(Schrijvers et al. 2006), and for projection should be investigated in the context
of SLGD .
34
T. Schrijvers et al.
10 Conclusion
We have presented a high-level framework for tabled CLP, based on a light-weight
integration of CHR with a tabled LP system. Tabling-related problems that have
to be solved time and again for ad hoc constraint solver integrations are solved
once and for all for CHR constraint solvers. Solutions have been formulated for
call abstraction, tabling constraint stores, answer projection, answer combination
(e.g. for optimization), and answer set optimization. Hence integrating a particular
CHR constraint solver requires much less knowledge of implementation intricacies
and decisions can be made on a higher level.
If performance turns out to be a bottleneck, once the high-level integration is stable and well-understood, then its implementation may be specialized in a lower-level
language using the TCHR implementation as its specification. Our novel contribution, generalized answer set compaction, may certainly contribute towards that
end.
Finally, we would like to mention that an XSB release, number 2.7, with the
presented CHR system integrated with tabling is publicly available since December
30, 2004 (see http://xsb.sf.net).
Acknowledgements
We are grateful to Beata Sarna-Starosta, Giridhar Pemmasani and C.R. Ramakrishnan for the interesting discussions and help on applications of tabled execution
and constraints in the field of model checking.
We thank the anonymous reviewers for their helpful comments.
References
Clark, K. L. 1987. Negation as Failure. In Logic and Databases, H. Gallaire and
J. Minker, Eds. Plenum Press, New York, 293–322.
Codish, M., Demoen, B., and Sagonas, K. 1998. Semantic-based Program Analysis
for Logic-based Languages Using XSB. International Journal of Software Tools for
Technology Transfer 2, 1 (Jan.), 29–45.
Cui, B. 2000. A System for Tabled Constraint Logic Programming. Ph.D. thesis, State
University of New York at Stony Brook.
Cui, B. and Warren, D. S. 2000a. A System for Tabled Constraint Logic Programming.
In CL 2000: Proceedings of the 1st International Conference on Computational Logic,
J. W. Lloyd, V. Dahl, U. Furbach, M. Kerber, K.-K. Lau, C. Palamidessi, L. M. Pereira,
Y. Sagiv, and P. J. Stuckey, Eds. Lecture Notes in Computer Science, vol. 1861. Springer
Verlag, London, UK, 478–492.
Cui, B. and Warren, D. S. 2000b. Attributed Variables in XSB. In Electronic Notes in
Theoretical Computer Science, I. Dutra et al., Eds. Vol. 30. Elsevier, 67–80.
Demoen, B. 2004. hProlog. http://www.cs.kuleuven.be/˜bmd/hProlog/.
Du, X., Ramakrishnan, C. R., and Smolka, S. A. 2000. Tabled Resolution + Constraints: A Recipe for Model Checking Real-Time Systems. In IEEE Real Time Systems
Symposium. Orlando, Florida, 175–184.
TCHR: a framework for tabled CLP
35
Duck, G. J., Garcı́a de la Banda, M., and Stuckey, P. J. 2004. Compiling Ask
Constraints. In ICLP’04: Proceedings of the 20th International Conference on Logic
Programming. Lecture Notes in Computer Science, vol. 3132. Springer Verlag, St-Malo,
France, 105–119.
Duck, G. J., Stuckey, P. J., Garcı́a de la Banda, M., and Holzbaur, C. 2004. The
Refined Operational Semantics of Constraint Handling Rules. In ICLP’04: Proceedings
of the 20th International Conference on Logic Programming. Lecture Notes in Computer
Science, vol. 3132. Springer Verlag, St-Malo, France, 90–104.
Frühwirth, T. 1998. Theory and practice of constraint handling rules. Journal of Logic
Programming 37, 1–3 (October), 95–138.
Frühwirth, T. and Abdennadher, S. 2003. Essentials of Constraint Programming.
Cognitive Technologies. Springer Verlag.
Guo, H.-F. and Gupta, G. 2004. Simplifying Dynamic Programming via Tabling. In
Proc. Sixth International Symposium on Practical Aspects of Declarative Languages,
P. V. Hentenryck, Ed. Lecture Notes in Computer Science, vol. 3819. Springer Verlag,
163–177.
Holzbaur, C. 1992. Metastructures vs. Attributed Variables in the Context of Extensible
Unification. Tech. Rep. TR-92-23, Austrian Research Institute for Artificial Intelligence,
Vienna, Austria.
Holzbaur, C. and Frühwirth, T. 2000. A Prolog Constraint Handling Rules Compiler and Runtime System. Special Issue Journal of Applied Artificial Intelligence on
Constraint Handling Rules 14, 4 (April), 369–388.
Jaffar, J. and Lassez, J.-L. 1987. Constraint Logic Programming. In POPL ’87: Proceedings of the 14th ACM SIGACT-SIGPLAN symposium on Principles of programming
languages. ACM Press, New York, NY, USA, 111–119.
Jaffar, J. and Maher, M. J. 1994. Constraint Logic Programming: A Survey. Journal
of Logic Programming 19/20, 503–581.
Kanellakis, P. C., Kuper, G. M., and Revesz, P. Z. 1995. Constraint query languages.
In Selected papers of the 9th annual ACM SIGACT-SIGMOD-SIGART symposium on
Principles of database systems. Academic Press, Inc., Orlando, FL, USA, 26–52.
Marriott, K. and Stuckey, P. J. 1998. Programming with Constraints: an Introduction. MIT Press.
Mukund, M., Ramakrishnan, C. R., Ramakrishnan, I. V., and Verma, R. 2000.
Symbolic Bisimulation using Tabled Constraint Logic Programming. In International
Workshop on Tabulation in Parsing and Deduction. Vigo, Spain, 1–9.
Pemmasani, G., Ramakrishnan, C. R., and Ramakrishnan, I. V. 2002. Efficient
Model Checking of Real Time Systems Using Tabled Logic Programming and Constraints. In International Conference on Logic Programming. Lecture Notes in Computer
Science. Springer, Copenhagen, Denmark, 405–410.
Rao, P., Ramakrishnan, C. R., and Ramakrishnan, I. V. 1996. A Thread in Time
Saves Tabling Time. In Joint International Conference and Symposium on Logic Programming. 112–126.
Saraswat, V. A. and Rinard, M. 1990. Concurrent constraint programming. In POPL
’90: Proceedings of the 17th ACM SIGPLAN-SIGACT symposium on Principles of programming languages. ACM Press, New York, NY, USA, 232–245.
Sarna-Starosta, B. and Ramakrishnan, C. R. 2003. Constraint-Based Model Checking of Data-Independent Systems. In 5th International Conference on Formal Engineering Methods, ICFEM 2003, J. S. Dong and J. Woodcock, Eds. Lecture Notes in
Computer Science, vol. 2885. Springer-Verlag, 579–598.
Sarna-Starosta, B. and Ramakrishnan, C. R. 2007. Compiling Constraint Handling
36
T. Schrijvers et al.
Rules for Efficient Tabled Evaluation. In PADL’07: Ninth International Symposium on
Practical Aspects of Declarative Languages, M. Hanus, Ed. Lecture Notes in Computer
Science. Springer Verlag, 170–184.
Schrijvers, T. 2005. Analyses, Optimizations and Extensions of Constraint Handling
Rules. Ph.D. thesis, Department of Computer Science, K.U.Leuven, Leuven, Belgium.
Schrijvers, T. and Demoen, B. 2004. The K.U.Leuven CHR system: Implementation
and application. In First Workshop on Constraint Handling Rules: Selected Contributions, T. Frühwirth and M. Meister, Eds. Ulm, Germany, 1–5.
Schrijvers, T., Demoen, B., Duck, G., Stuckey, P., and Frühwirth, T. 2006. Automatic Implication Checking for CHR Constraints. In Electronic Notes in Theoretical
Computer Science. Vol. 147. 93–111.
Schrijvers, T. and Frühwirth, T. 2006. Optimal Union-Find in Constraint Handling
Rules. Theory and Practice of Logic Programming 6, 1&2, 213–224.
Schrijvers, T. and Warren, D. S. 2004. Constraint handling rules and tabled execution. In ICLP’04: Proceedings of the 20th International Conference on Logic Programming, B. Demoen and V. Lifschitz, Eds. Lecture Notes in Computer Science, vol. 3132.
Springer Verlag, St-Malo, France, 120–136.
Schrijvers, T., Warren, D. S., and Demoen, B. 2003. CHR for XSB. In CICLOPS
2003: Proceedings of the Colloquium on Implementation of Constraint and LOgic Programming Systems, R. Lopes and M. Ferreira, Eds. University of Porto, Mumbai, India,
7–20.
Toman, D. 1996. Computing the Well-founded Semantics for Constraint Extensions of
Datalog. In Proceedings of CP’96 Workshop on Constraint Databases. Number 1191 in
Lecture Notes in Computer Science. Cambridge, MA, USA, 64–79.
Toman, D. 1997a. Constraint Databases and Program Analysis Using Abstract Interpretation. In Constraint Databases and Their Applications, Second International Workshop
on Constraint Database Systems (CDB ’97), V. Gaede, A. Brodsky, O. Günther, D. Srivastava, V. Vianu, and M. Wallace, Eds. Lecture Notes in Computer Science, vol. 1191.
Springer Verlag, 246–262.
Toman, D. 1997b. Memoing Evaluation for Constraint Extensions of Datalog. Constraints: An International Journal, Special Issue on Constraints and Databases 2, 3/4
(December), 337–359.
Warren, D. S. et al. 2005. The XSB Programmer’s Manual: version 2.7, vols. 1 and 2.
http://xsb.sf.net.
Wielemaker, J. 2004. SWI-Prolog release 5.4.0. http://www.swi-prolog.org/.
Wolper, P. 1986. Expressing interesting properties of programs in propositional temporal
logic. In POPL ’86: Proceedings of the 13th ACM SIGACT-SIGPLAN symposium on
Principles of programming languages. ACM Press, New York, NY, USA, 184–193.
| 6cs.PL
|
arXiv:1501.04010v1 [cs.NE] 16 Jan 2015
Coevolutionary intransitivity in games:
A landscape analysis
Hendrik Richter
HTWK Leipzig University of Applied Sciences
Faculty of Electrical Engineering & Information Technology
Postfach 301166, D–04251 Leipzig, Germany
Email: [email protected]
January 19, 2015
Abstract
Intransitivity is supposed to be a main reason for deficits in coevolutionary progress and inheritable superiority. Besides, coevolutionary dynamics is characterized by interactions yielding subjective fitness, but
aiming at solutions that are superior with respect to an objective measurement. Such an approximation
of objective fitness may be, for instance, generalization performance. In the paper a link between rating–
and ranking–based measures of intransitivity and fitness landscapes that can address the dichotomy between subjective and objective fitness is explored. The approach is illustrated by numerical experiments
involving a simple random game with continuously tunable degree of randomness.
1
Introduction
Despite earlier promises and optimism, using coevolutionary algorithms (CEAs) for evolving candidate
solutions towards an optimum remains a complicated and almost arcane matter with generally unclear
prospects of success [13, 15, 20]. This is prominently caused by a defining feature of coevolution. CEAs
are driven by fitness that originates from interaction of candidate solutions with other candidate solutions.
In other words, the fitness obtained from these interactions is subjective as it depends on which candidate
solutions are actually interacting and when in coevolutionary run–time the interactions take place. The
interactions can be understood as to constitute tests, which leads to labeling such kinds of coevolutionary
problems as test–based problems [15, 17]. Test–based problems particularly occur in game playing or
game–like contexts, for instance in situations where the players’ strategies are subject to (competitive)
coevolutionary optimum finding [15, 19]. It has been argued that in games with players and strategies the
player space can be understood a phenotypic according to the framework of fitness landscapes [10, 16], while
the strategy space is genotypic [1, 14]. This view is adopted in the following discussion.
Notwithstanding that coevolutionary dynamics is induced by subjective fitness, the aim of using a CEA
is identifying candidate solutions that are superior in a more general sense. Hence, in coevolution next to
the fitness resulting from (a limited number of) tests, a second notion of fitness is helpful. Such a fitness
generalizing subjective fitness occurs in test–based problems in different forms. For games with players and
strategies, there is usually no absolute quality measurement, or an absolute quality measurement would
require to evaluate all possible test cases, which is computationally infeasible [4, 5, 19]. To circumvent
this problem and enable experimental studies of the relationships between subjective fitness and absolute
quality measurements, number games have been proposed, for instance minimal substrates [18, 21]. These
artificial problem settings postulate an absolute quality, called objective fitness [7]. In this line of reasoning,
coevolutionary dynamics can be understood as aiming at progress in objective fitness by proxy of subjective
fitness. Consequently, a main difficulty in designing CEAs stems from the question of how well subjective
1
fitness represents objective fitness. In analogy to the (postulated) objective fitness of number games, for
games with players and strategies all general quality measurements of subjective fitness are interpretable as
objective fitness. This implies that for game playing there may only be an approximation of objective fitness
and different approximations are possible, for instance different instances of generalization performance [4, 5].
Put another way, this interpretation suggests that in game playing there are different degrees of objective
fitness.
Application examples of CEAs have frequently reported experiments showing mediocre performance,
mostly attributed to coevolutionary intransitivity [3, 9, 21]. Generally speaking, intransitivity occurs when
superiority relations are cyclic. Such cyclic superiority relations have consequences for coevolutionary dynamics as intransitivity may occurs across subsequent generations. In such a case, it may be that all solutions
at generation k + 1 are better than at k, and that the same applies for k + 2 with respect to k + 1. This,
however, does not imply that the solutions at k + 2 are strictly better than at k. Cyclic superiority relations
that occur across generations are connoted as coevolutionary (dynamic) intransitivity. In this paper, the
problem of coevolutionary intransitivity is linked to the dichotomy between subjective and objective fitness.
This is done by combining a rating– and ranking–based measuring approach of intransitivity proposed by
Samothrakis et al. [19] with a framework of codynamic fitness landscapes recently suggested [18]. Codynamic
fitness landscapes enable to analyse the relationship between objective and subjective fitness for all possible
solutions of the coevolutionary search process. The landscape approach proposed particularly explores how
coevolutionary intransitivity is related to the objective–versus–subjective–fitness issue. The remainder of
the paper is structured as follows. In the next section, the concept of codynamic landscapes composed of
objective and subjective fitness is briefly recalled. In Sec. 3, intransitivity is discussed. For the discussion, a
simple random game is introduced, where the degree of randomness can be continuously tuned. It is shown
that intransitivity can be characterized by different types of intransitivity measures. Numerical experiments
with the simple random game are presented in Sec. 4, and Sec. 5 concludes the paper with a summary.
2
Coevolution, codynamic landscapes and number games
This section focuses attention on an approach recently suggested [18] that is useful for understanding coevolutionary dynamics through codynamic fitness landscapes. Such landscapes allow studying the relationship
between objective and subjective fitness which, in turn, mainly determines coevolutionary dynamics.
We define objective fitness as the triple of search space S with search space points s ∈ S, neighborhood
structure n(s) and fitness function fobj (s). The objective fitness landscape can be considered as to describe
the optimization problem to be solved by the CEA. This problem solving is based on coevolutionary interactions between potential solutions which yields subjective fitness. Hence, subjective fitness can be viewed as
the way the CEA perceives the problem posed by the objective fitness. From this, it appears to be sensible
to assume that the subjective landscape possesses the same search space and neighborhood structure, but
has a fitness function fsub that more or less strongly deviates from the objective fitness fobj . This can be
seen as the subjective fitness usually overestimating or underestimating objective fitness. Moreover, for a
coevolutionary run, the deviation between objective and subjective fitness is dynamic. In other word, the
coevolutionary dynamics dynamically deforms the subjective fitness landscape. In the following, the link between subjective and objective fitness is exemplified for a number game, which is also called a coevolutionary
minimal substrate [18, 21].
In this number game a population P of players is considered that inhabits the search spaces S. The
search space is one–dimensional and real–valued. At each instance of the game k = 0, 1, 2, . . ., the players
of P (k) may have possible values s ∈ S. An objective fitness function is defined over the search space, that
is fobj (s), which consequently casts an objective fitness landscape. The subjective fitness of P (k) is the
result of an interactive number game. Therefore, for each calculation of the subjective fitness fsub (s) for a
player from P , a sample σ(E) of evaluators from E ⊆ P is randomly selected. This sample is statistically
independent from the sample for the next calculation. Denote µ the size of the sample σ(E) out of λ
evaluators, with µ ≤ λ. The number game further defines that the fitness fsub (s) with respect to the sample
2
σ(E) is calculated by counting the (averaged) number of members in σ(E) that have a smaller objective
fitness fobj (σi (E)), i = 1, 2, . . . , µ, than the objective fitness fobj (s) [18]:
µ
1X
eval(s, σi (E))
fsub (s) =
µ
with
eval(s, σi ) =
i=1
1
0
if fobj (s) > fobj (σi )
.
otherwise
(1)
Note that the number game considered postulates objective fitness and defines by Eq. (1) how subjective
fitness is obtained by a coevolutionary interaction. In the next section, this perspective of subjective and
objective fitness is applied to games with players and strategies. A population of players is engaged to
evaluate their subjective fitness by interaction with other players.
3
Static and coevolutionary intransitivity in games
A relation R is called intransitive over a set S if for three elements {s1 , s2 , s3 } ∈ S the relation (s1 Rs2 ) ∧
(s2 Rs3 ) does not always imply s1 Rs3 . An instance of intransitivity is superiority relations that are cyclic,
which in its most obvious and purest form appears in game playing. Cyclic superiority relations here mean
that for three players and three strategies s1 , s2 and s3 , the player using s1 wins against s2 , and s2 wins
against s3 , but s1 loses against s3 . A simple example is a rock–paper–scissor game, where “paper” wins over
“rock”, “scissor” wins over “paper”, but “scissor” loses against “rock”. Thus, “paper”, “rock” and “scissor”
are possible strategies a player can adopt in this game. Note that this kind of intransitivity is a feature of
the preference in a single round of the game. Hence, such an intransitivity is static (and actually game–
induced) and has no immediate link to (co-)evolutionary dynamics. Consequently, the next question is how
these superiority relations resemble situations with coevolutionary intransitivity and can be understood by
the dichotomy between subjective and objective fitness. To obtain (co-)evolutionary dynamics, the players
need to adjust their strategies, and the game needs to be played for more than one round. In other words,
studying iterated games is also interesting as it serves to juxtapose static (game–induced) intransitivity with
coevolutionary (dynamic and search–induced) intransitivity.
One way to build a relationship between game results and fitness is to apply a rating system. Examples
of rating systems are the Elo system to evaluate chess players [8, 11], or the Bradley–Terry–Luce model of
paired comparison [2, 12]. Recently, it has been shown that this methodology is also useful for analyzing
coevolutionary intransitivity [19]. A rating system creates a probabilistic model based on past game results
that can be seen as a predictor of future results. Most significantly, the rating system also imposes a
(temporal) ranking of the players. In the following, these ideas are applied to a simple random game where
the degree of randomness can be tuned. The game consists of players using a strategy to perform against all
other players once, called a round robin tournament. The game outcome, which can be interpreted as payoff,
is subject to the players’ ratings and random. Given that the game has N players, there are N (N2−1) games in
a single round robin. Define a percentage prand of games that have a random result to obtain prand · N (N2−1)
games whose outcome is chance with a predefined distribution. The remaining games end deterministically
according to the rating difference between the players. Thus, such a game falls into the category of perfect
and incomplete information. Viewed over a series of round robin tournaments, this interaction between
rating–based determinism and random chance creates temporal “rating triangles”, where a player scores
high results (and has a high rating) over a certain time, but may also lose against a nominal weaker (low–
rating) player, which over time may or may not show the same characteristics towards a third player. Such
a behavior complies with coevolutionary (dynamic and actually search–induced) intransitivity. In addition,
the game can also reproduce rock–paper–scissor–like intransitivities. For N players the maximal number of
static intransitivities is
(N − 2)(N − 1)N
N
#intra max =
=
,
(2)
3
6
see [6, 19]. Any three players form a triangle of cyclic superiority relations if they each win one game
against the other two. Thus, for N not very large, the (average) number of actual static intransitivities
3
#
1
2
3
4
5
rt(0)
1600
1600
1600
1600
1600
Table 1: Results of the simple random game for three instance
#1 #2 # 3 #4 #5 sc(0) rt(1) #1 #2 # 3 #4
x
1
1
1
1
4
1630
x
0
1
1
0
x
1
0
1
2
1600
1
x
1
1
0
0
x
1
1
2
1600
0
0
x
0
0
1
0
x
1
2
1600
0
0
1
x
0
0
0
0
x
0
1570
0
1
1
0
of the game
#5 sc(1)
1
3
0
3
0
0
1
2
x
2
rt(2)
1642
1615
1570
1600
1573
gp
3.5
2.5
1
2
1
rank
1
2
4.5
3
4.5
can be determined by enumeration and gives a static intransitivity measure called the intransitivity index
(itx) [6, 19]. As an alternative, Samothrakis et al. [19] suggested to use a difference measure based on
Kullback–Leibler divergence (kld) between the prediction made by the rating system and the actual outcome
to measure static intransitivity. Both quantities itx and kld are subjects of the numerical experiments
reported in the next section.
A game in a coevolutionary setting involves finding the strategy a player should adopt to score best
according to a given understanding of performance. This clearly implies that the performance measurement
should generalize a single round robin tournament. Thus, if there are several instances of round robin
tournaments, the overall results can also be accounted for by generalization performance [4, 5]. Each instance
of a round robin can be scaled to a generation of coevolutionary run–time. Generalization performance is
defined as mean score of a solution in all possible test cases. Because considering all possible test cases may
be computationally infeasible, Chong et al. [4, 5] used a statistical approach involving confidence bounds to
estimate the amount of needed test cases for a given error margin. Given this understanding, and assuming
that all strategies are equally likely to be selected as test strategies, the generalization performance of
strategy i is:
K
1 X
gpi =
sci (k),
(3)
K
k=1
where sci (k) is the score the i-th strategy yields in the k–th instance of a round robin tournament. The needed
number of instances K depends on the bounds given by Chong et al. [4, 5]. Generalization performance also
builds a relationship between actual game results and fitness and therefore can be seen as an alternative to
a rating system.
As an example of the simple random game assume that there are five players (N = 5), denoted as #1 to
#5, which each act upon a unknown, but internally adjusting strategy. Further assume that as an a–priori
evaluation all players are considered equal and have the same rating, say rt(0) = 1600. These players are
now engaged in a round robin tournament, where a win scores 1 and a loss counts 0. The results achieved
depend on the pre–game rating and random. Assume that these results were scored, see Tab. 1, column
3–7. While player #1 wins all games and #5 loses all games, the players #2–#4 build a rock–paper–scissor
triangle of cyclic superiority relations. These results somehow violate the expectations established by the
initial (pre–game) rating, which could have been met by all players winning 2 out of 4 games. Furthermore,
the results show clearly that the game cannot be completely deterministic with respect to the a–priori
evaluation. In other words, a rating and ranking approach subsumes the game history and can only be
a predictor of future game results. The quality of prediction depends on the percentage of random game
results. Also, from the results of the round robin, it is not evident whether player #1 was as successful
as it was because its strategy was (objectively) good, or because the strategies of the other players were
(objectively) poor. All, the round robin gives is a comparison.
The a–priori rating were indicating that all players have the same rank (and hence there is no ranking
difference between them), but the results were showing otherwise. Hence, the round robin tournament
updates the rating, producing an after–game rating. According to the Elo system [8, 11], which is adopted
4
sc(
1
4
0
2
3
here, this is done via first calculating the expected outcome
exi (k) =
1
X
j
1+
10(rtj (k)−rti (k))/400
.
(4)
The quantity exi (k) summarizes winning probabilities of player #i with respect to all other players in round
k. For a single game, that is j = 1, the quantity exi is the expected winning probability for the player #i
winning against player #j. For the example with all players having the same rating, the expected outcome
is also the same, namely exi = 2. The rating of the players is updated according to the difference between
expectation and actual score:
rti (k + 1) = rti (k) + K(sci (k) − exi (k)).
(5)
The K is called the K–factor, which tunes the sensitivity between the rating and the results of a single
round robin tournament. Using K = 15, the new rating rt(1) gives differences between the players, see
Tab. 1, column 9. The best player #1 is now ranked highest, the poorest player #5 is ranked lowest.
Also note that the players #2–#4 engaged in the game–induced intransitivity triangle still share the same
rating as before. Now assume the next round of the game. The strategies of the players have been adjusted,
supposedly by a (competitive) coevolutionary search process. See the results in Tab. 1, column 10–14. The
outcome generally confirms the impression of the first round with player #1 still being strong. Violating
the expectations are the good scores of the players #2 and #5, and the poor score of player #3. Note that
in this round the players #1, #2, and #5 as well as the players #2, #4, and #5 form a rock–paper–scissor
triangle of game–induced intransitivity. Contrary to the first round, these players now neither have the same
score, nor the same pre–game or after–game rating. This after–game rating rt(2) is calculated according to
Eq. (5) and given in Tab. 1, column 16.
As there are now two instances of the round robin tournament, an first estimation of the generalized
performance gp can be obtained by averaging the scores according to Eq. (3). The results are given in Tab.
1, column 17. Conforming with this account of quality evaluation, player #1 is best, and can be ranked first,
followed by players #2 and #4. The ranking with respect to generalization performance is given in Tab. 1,
column 18. Note that the ranking gives average ranks for tied ranks as for player #3 and #5 (gp = 1 leads
to rank = 4.5), which preserves the sum over all ranks. Further note that this ranking is almost equal to
the ranking according to ratings, with the exception of players #3 and #5 which have a very similar but
not equal rating.
Now, the simple random game is interpreted according to the landscape view of subjective and objective
fitness. Recall that subjective fitness is associated with fitness gained by individuals through interaction
with others. According to this view, a round robin tournament yields subjective fitness: fsub = sc(k).
Objective fitness, in turn, generalizes subjective fitness in terms of an absolute quality measurement. Possible
candidates are the rating fobj = rt(k) or generalization performance fobj = gp(k). Defining subjective and
objective fitness in such a way also gives raise to reformulating coevolutionary intransitivity. Generally
speaking, coevolutionary intransitivity involves cycling of (objective) solution quality. This cycling may be
caused by subjective fitness not adequately representing objective fitness. Hence, subjective fitness may
drive evolution into search space regions visited before but evaluated differently, or generally into directions
not favorable. Hence, coevolutionary (dynamic) intransitivity can be understood as temporal mismatches
in order between subjective and objective fitness. Consider again the example of the game whose results are
given in Tab. 1. Suppose another instance of the round robin is played, and (omitting the specific results)
the scores are in Tab. 1, column 19. The rating rt(k) is considered to be objective, while sc(k) is subjective.
For player #1, the rating is rt1 = 1600 ≤ 1630 ≤ 1642, while its score is sc1 = 4 > 3 > 1, which is a
temporal mismatch between objective and subjective fitness. If we were to suppose for a moment that the
rating declared as objective fitness is indeed the quantity to achieve in coevolutionary search, and if a CEA
were to use score for guiding this search, then player #1 would likely be misguided. On the other hand,
for player #2, the rating rt2 = 1600 ≤ 1630 ≤ 1642, and the score sc2 = 2 < 3 < 4 show a match between
5
6
2000
6
1900
5
5
1800
4
3
1700
gp(k)
rat(k)
sc(k)
4
3
1600
2
2
1500
1
0
0
1
1400
2
4
6
8
10
12
14
16
18
1300
0
20
2
4
6
8
10
k
12
14
16
18
0
0
20
2
4
6
8
k
(a)
14
16
18
20
12
14
16
18
20
(c)
2000
6
1900
5
12
k
(b)
6
10
5
1800
4
3
1700
gp(k)
rat(k)
sc(k)
4
3
1600
2
2
1500
1
0
0
1
1400
2
4
6
8
10
12
14
16
18
20
1300
0
2
4
6
8
10
k
(d)
12
14
16
k
(e)
18
20
0
0
2
4
6
8
10
k
(f)
Fig. 1. Scores, ratings, and generalization performance for two percentages of random games: (a)–(c),
prand = 0.01; (d)–(f), prand = 0.75.
subjective and objective fitness. These conditions can be reformulated employing a ranking function with
tied ranks and gives a measure of coevolutionary intransitivity. Hence, the player–wise temporal mismatch
(ptm) can be defined as the average number of rank mismatches: rank(fobj (k), fobj (k + 1)), fobj (k + 2)) 6=
rank(fsub (k), fsub (k + 1)), fsub (k + 2)) for each player for a given number of instances of round robins.
An alternative measure of coevolutionary intransitivity that is related to the ptm just discussed stems
from the fact that coevolutionary selection is based on comparing subjective fitness values. Hence, the fitness
ranking within one instance of the game (or one generation) gives indication as to what direction is preferred.
If difficulties in the search process are caused by how well subjective fitness represents objective fitness,
then the difference in the ranking according to subjective fitness and the ranking according to objective
fitness for each instance is also a suitable measure of coevolutionary intransitivity. Therefore, the quantity
|rank(fsub (k)) − rank(fobj (k))| over all players for each instance k is another measure of coevolutionary
intransitivity and is called collective ranking difference (crd). The observations discussed so far suggest
some relationships, namely that game–induced, static intransitivity has ambiguous effect on coevolutionary
progress, and that coevolutionary, dynamic intransitivity can be expressed as ranking differences between
objective and subjective fitness. In other words, the quantities ptm and crd may be useful as measures of
coevolutionary intransitivity. All these relationships can be studied by numerical experiments, which are
the topic of next section.
6
0.018
0.018
0.018
prand
prand
0.016
prand
0.016
0.016
0.85
0.85
0.014
0.85
0.014
0.50
0.012
0.014
0.50
0.012
0.008
itx
0.01
itx
0.01
itx
0.01
0.008
0.20
0.006
0.008
0.20
0.006
0.004
0.004
0.10
0.002
0.002
0.01
4
6
8
10
12
14
0.01
0
0
16
500
1000
1500
2000
sc
2500
3000
0.01
0
0
3500
(b)
0.2
0.2
N = 32
0.1
<itx>
0.2
<itx>
0.25
0.15
0.15
0.2
0.25
0.3
0.35
N = 24
0.15
N = 16
0.05
N=8
0.4
0.45
0.5
0
0
kld
(d)
N = 32
0.1
N = 16
N=8
0.15
1
N = 24
0.05
N = 16
0.1
0.8
N=8
N = 32
0.1
N = 24
0.05
0.6
(c)
0.25
0.05
0.4
gp
0.25
0
0
0.2
rt
(a)
<itx>
0.10
0.002
2
0.20
0.006
0.004
0.10
0
0
0.50
0.012
0.01
0.02
0.03
0.04
0.05
0.06
ptm(sc,rt)
(e)
0.07
0.08
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
ptm(sc,gp)
(f)
Fig. 2. Intransitivity measure itx versus score = sc(k), rating rt(k) and generalization performance gp(k)
(a–c) as scatter plots; relationships between the intransitivity measure itx, kld, and ptm (d–f)
4
Numerical experiments
The report of experimental results starts with the time evolution of the simple random game introduced
in the last section. Fig. 1 shows the scores, ratings and generalization performances for two different
percentages of random results (prand = 0.01 and prand = 0.75) and N = 6 players and 20 instances of
a round robin. The experiments are initialized with ratings rt(0) being slightly spread around the value
rt(0) = 1600. The randomly determined chance to win or to lose is evenly distributed. The figures show
curves for a single instance of the randomly determined part of the game outcomes. Hence, the curves are
not meant to be statistically significant, but for illustrating typical behaviour only. It can be seen that for
a low number of random outcome (Fig. 1a,b,c) the scores each player achieves are mainly defined by the
initial rating. Small differences in the initial rating are amplified and lead to well–sorted long–term rankings
of both the rating and generalization performance. For a large number of random game results (Fig. 1d,e,f)
the scores are almost purely chance which is evenly distributed. Consistently, ratings and generalization
performances tend to approach the expected value implied by the underlying distribution for all players alike.
The next experiments address the relationships between static (game–induced) intransitivity expressed by
the measure itx and quantities representing subjective as well as objective fitness. The scatter plots given in
Fig. 2a,b,c are for N = 16 players and five levels of prand . The experimental setup includes a repetition of
each run for 100 times for 1000 instances of the round robin, where the first 200 instances are discarded to
omit transients. Note that this gives a sufficient number of instance according to the bounds of generalization
7
2
prand
1.6
0.20
0.05
0.10
1.2
0.50
1
0.8
0.20
0.6
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
0.50
1
0.8
0.20
0.4
0.10
0.10
0.2
0.01
0.01
0.4
1.2
0.6
0.2
0.2
0.85
1.4
crd(sc,gp)
crd(sc,gp)
0.15
0.1
1.6
0.85
1.4
0.50
0.2
0
0
prand
1.8
0.85
0.25
<itx>
2
prand
1.8
0
0.4
2
0.5
0.6
0.7
crd(sc,gp)
0.8
0.9
1
0.01
0
0.4
1.1
0.5
0.6
0.7
max sc
(a)
0.8
0.9
1
1.1
max gp
(b)
(c)
2.5
0.07
0.06
0.25
2
0.05
0.15
1.5
N = 32
1
N = 32
0.1
ptm(sc,rt)
crd(sc,gp)
<itx>
0.2
0.04
0.03
N = 24
N = 32
0.02
N = 16
N = 24
N = 24
0.5
0.05
N=8
N = 16
N = 16
0.01
N=8
N=8
0
0
0.5
1
1.5
2
2.5
0
0
crd(sc,gp)
(d)
0.2
0.4
0.6
0.8
1
1.2
1.4
crd(sc,rt)
(e)
1.6
1.8
0
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
crd(sc,rt)
(f)
Fig. 3. Intransitivity measure crd versus time average itx, max sc, and max pg (a–c) as scatter plots;
relationships between the intransitivity measure crd, itx, and ptm (d–f)
performance [4, 5]. Hence, the results can be seen as statistically significant. In fact, the 99% confidence
intervals are so small that they are not depicted in the figures. It can be seen that for each level of
randomness in the game, we obtain a distinct level of static intransitiviy measured by itx, where rising prand
also increases itx. What is interesting is that there is almost no variation in score, rating or generalization
performance for a given level of intransitivity itx. This indicates that static intransitivity seems to have little
influence on neither subjective nor objective fitness. This is particularly visible for low levels of prand where
these is a clear sorting according to the range of fitness a given player achieved but this is not connected to
differences in the itx. A next experiment explores the relations between the index based measure itx and the
probabilistically motivated measure kld, see Fig. 2d. Here as well as in the following figure the results are
for four different number of players N (N = 8, 16, 24, 32) and five levels of random results prand ; the time–
average itx, denoted hitxi, is given, the quantities are normalized according to the number of players. It can
be seen that between both quantities there is a proportional relationship, which on the one hand generally
confirms the result in [19], but does not show that itx is more brittle that kld. For the studied game it can be
concluded that both quantities are interchangeable. Next, the relationship between the static intransitivity
and the dynamic intransitivity measure player–wise temporal mismatch (ptm) is studied, Fig. 2e,f. For the
mismatch ptm based on rating as objective fitness (Fig. 2e) there is a linear relationship, at least for small
values of prand , for the mismatch ptm based on generalization performance (Fig. 2f), no sensible conclusions
about relations can be drawn. Finally, we focus on the rank–based dynamic intransitivity measure collective
ranking difference (crd), see Fig. 3. The Fig. 3a,b,c again shows scatter plots, now for the crd based on
8
score as subjective fitness and generalization performance as objective fitness for N = 16 players and five
levels of prand . It can be seen that although different prand give different itx, there is no difference in the
crd, Fig. 3a. However, there is a almost linear relation between the measure crd and max sc and max
gp, at least for lower levels of prand . This can be interpreted as the crd scaling with the time evolution of
the subjective and objective fitness, but not with the time evolution of static intansitivity, compare to Fig.
2a,b,c, which does not show such a characteristics. For the time–averages, there is a scaling for different
number of players and different levels of randomness prand , see Fig. 3d. In Fig. 3e, the relation between
the crd based on rating as objective fitness (crd(sc,rt)) and the crd based on generalization performance as
objective fitness (crd(sc,gp)) is shown. It can be seen that both quantities scale piece-wise linear for prand
not very large, which allows to conclude that both quantities account for the same intransitivity properties.
Finally, the relation between the crd and ptm is shown, Fig. 3f. It can be seen that the ptm scales weaker
than the crd, particularly for a small number of players and high randomness, and it can be conjectured
that the crd is a more meaningful coevolutionary intransitivity measure than the ptm.
5
Conclusions
This paper is a contribution to the ongoing discussion about the effect of intransitivities on coevolutionary
progress. An approach was presented that allowed to link a rating– and ranking–based measuring approach
of intransitivity [19] with a framework of fitness landscapes to enable analyzing the relationship between
objective and subjective fitness. For experimentally illustrating the approach a simple random game with
continuously tunable degree of randomness was proposed. Apart from the random, the game results depend
on the ratings of the players, which reflect the past success of each player. Thus, the game proposed
characterizes many real–world games as their outcome is also a function of chance as well as of predictions
based on game history.
For studying the effect of intransitivity, measures were explored. In extension of existing static intransitivity measures, dynamic measures that can account for coevolutionary intransitivity were proposed. These
measure are based on rankings between subjective and objective fitness and it was shown that coevolutionary intransitivity can be understood as a ranking problem, and hence be accounted for by ranking statistics.
To enlarge the scope of the presented approach, as a next step the intransitivity measures could be studied
for other types of game for instance social games as the iterated prisoner’s dilemma or board game such as
Othello.
References
[1] T. Antal, H. Ohtsuki, J. Wakeley, P. Taylor, M. A. Nowak, Evolutionary game dynamics in phenotype
space. Proc. Nat. Acad. Sci. 106: 8597–8600, 2009.
[2] R. A. Bradley, M. E. Terry, Rank analysis of incomplete block designs, I. the method of paired comparisons. Biometrika 39: 324–345, 1952.
[3] E. D. de Jong, Intransitivity in coevolution. In: X. Yao et al. (eds.), Parallel Problem Solving from
Nature–PPSN VIII, Springer–Verlag, Berlin Heidelberg New York, 843–851, 2004.
[4] S. Y. Chong, P. Tino, X. Yao, Measuring generalization performance in coevolutionary learning. IEEE
Trans. Evolut. Comp. 12: 479–505, 2008.
[5] S. Y. Chong, P. Tino, D. C. Ku, X. Yao, Improving generalization performance in coevolutionary
learning. IEEE Trans. Evolut. Comp. 16: 70–85, 2012.
[6] O. Frank, F. Harary, Cluster inference by using transitivity indices in empirical graphs. Jour. Am.
Statist. Assoc. 77: 835–840, 1982.
9
[7] E. D. de Jong, Objective fitness correlation. In: H. Lipson (ed.), Proc. Genetic and Evolutionary
Computation Conference, GECCO 2007, ACM, New York, 440–447, 2007.
[8] A. E. Elo, The Rating of Chess Players, Past and Present. Batsford, London, 1978.
[9] P. Funes, E. Pujals, Intransitivity revisited coevolutionary dynamics of numbers games. In: H. G.
Beyer, U. M. O’Reilly (eds.), Proc. Genetic and Evolutionary Computation Conference, GECCO 2005,
Morgan Kaufmann, San Francisco, CA, 515–521, 2005.
[10] L. Kallel, B. Naudts, C. R. Reeves, Properties of fitness functions and search landscapes. In: L. Kallel,
B. Naudts, A. Rogers (eds.), Theoretical Aspects of Evolutionary Computing, Springer–Verlag, Berlin,
Heidelberg, New York, 177–208, 2001.
[11] A. N. Langville, C. D. Meyer, Who’s #1? The Science of Rating and Ranking. Princeton University
Press, Princeton, NJ, 2012.
[12] R. D. Luce, Individual Choice Behaviours: A Theoretical Analysis. John Wiley, New York, 1959.
[13] T. Miconi, Why coevolution doesn’t “work”: Superiority and progress in coevolution. In: L. Vanneschi,
et al. (eds.), EuroGP 2009, Springer–Verlag, Berlin Heidelberg New York, 49–60, 2009.
[14] M. A. Nowak, C. E. Tarnita, T. Antal, Evolutionary dynamics in structured populations. Phil. Trans.
R. Soc. B 365: 19–30, 2010.
[15] E. Popovici, A. Bucci, R. P. Wiegand, E. D. de Jong, Coevolutionary principles. In: G. Rozenberg, T.
Bäck, J. N. Kok (eds.), Handbook of Natural Computing, Springer–Verlag, Berlin Heidelberg New York,
987–1033, 2010.
[16] H. Richter, A. P. Engelbrecht, Recent Advances in the Theory and Application of Fitness Landscapes,
Springer–Verlag, Berlin Heidelberg New York, 2014.
[17] H. Richter, Fitness landscapes that depend on time. In: H. Richter, A. P. Engelbrecht (eds.), Recent
Advances in the Theory and Application of Fitness Landscapes, Springer–Verlag, Berlin Heidelberg New
York, 265–299, 2014.
[18] H. Richter, Codynamic fitness landscapes of coevolutionary minimal substrates. In: C. A. Coello Coello
(ed.) Proc. IEEE Congress on Evolutionary Computation, IEEE CEC 2014, IEEE Press, Piscataway,
NJ, 2692–2699, 2014.
[19] S. Samothrakis, S. M. Lucas, T. P. Runarsson, D. Robles, Coevolving game–playing agents: Measuring
performance and intransitivities. IEEE Trans. Evolut. Comp. 17: 213–226, 2013.
[20] R. P. T. van Wijngaarden, E. D. de Jong, Evaluation and diversity in co-evolution. In: G. Rudolph
et al. (eds.), Parallel Problem Solving from Nature–PPSN X, Springer–Verlag, Berlin Heidelberg New
York, 631–640, 2008.
[21] R. A. Watson, J. B. Pollack, Coevolutionary dynamics in a minimal substrate. In: L. Spector et al.
(eds.), Proc. Genetic and Evolutionary Computation Conference, GECCO 2001, Morgan Kaufmann,
San Francisco, CA, 702–709, 2001.
10
| 9cs.NE
|
arXiv:1711.10185v1 [cs.AI] 28 Nov 2017
Hyper-dimensional computing for a visual
question-answering system that is trainable
end-to-end
Guglielmo Montone
Laboratoire Psychologie de la Perception
Université Paris Descartes
75006 Paris, France
[email protected]
J.Kevin O’Regan
Laboratoire Psychologie de la Perception
Université Paris Descartes
75006 Paris, France
[email protected]
Alexander V. Terekhov
Laboratoire Psychologie de la Perception
Université Paris Descartes
75006 Paris, France
[email protected]
Abstract
In this work we propose a system for visual question answering. Our architecture
is composed of two parts, the first part creates the logical knowledge base given the
image. The second part evaluates questions against the knowledge base. Differently
from previous work, the knowledge base is represented using hyper-dimensional
computing. This choice has the advantage that all the operations in the system,
namely creating the knowledge base and evaluating the questions against it, are
differentiable, thereby making the system easily trainable in an end-to-end fashion.
1
Introduction
Visual Question Answering (VQA) or Visual Turing Test are terms that refer to a task in which
a machine is provided with a picture and a question about a picture and the machine is asked to
return the answer to the question. Such tasks have become popular in the last year thanks to the
emergence of several datasets containing images, with associated questions and answers [8, 1].
Classical deep neural network approaches face these tasks by training architectures composed of
several parts. In these architectures there are often a RNN (often an LSTM) for encoding the question
and for producing the answer, and a CNN for analyzing the image [2, 7]. The main idea behind such
architectures is to project the question and the image into a relatively low dimensional space where
the two are compared. These approaches give impressive results when the question is about simple
properties of the image like "What is the color of the bus?". However the results become worse when
the question is more complex and involves verifying one or more relations among objects in the
picture ("Is the fruit in the plate different from the one in the basket?").
Another group of approaches builds architectures with a perceiver and an evaluator [5]. The perceiver
receives the image as input and has to build the knowledge base relative to the input. The evaluator,
executes the question against the knowledge base to produce an answer. The disadvantage of such
approaches is that the computations performed by the different parts of the system are of very different
nature, leading to cumbersome architectures that are complex to train in an end-to-end fashion.
The main contribution of this work is to show that, at least in the simple case we propose, it is possible
to build an architecture where the evaluator performs computations of the same kind as the perceiver.
It is for this reason that the architecture can be easily trained in an end-to-end fashion. In particular
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
Figure 1: Some images in the dataset
in our architecture we will exploit the properties of hyper-dimensional computing (HD-computing)
[4]. It is well known in fact that, by defining vectors in a high-dimensional space (HD-vectors), it is
possible to store and then retrieve information from these vectors with very simple operations [4]. We
will use HD-vectors to store the values of logical constants, and also the values of relations among
these constants that will constitute the knowledge base of the system. In our architecture a FFW
network will be asked to associate to each image in the input the HD-vector describing the image.
The HD-vector returned as output can then be queried to extract answers to given questions, with
such questions consisting in sets of differentiable operations on the HD-vectors. The parameters of
the network will be updated with a gradient descent procedure in order to minimize the number of
wrong answers. In the present paper we prove that this architecture can be successfully trained on a
simple VQA task.
The paper is organized as follows, in section 2 we present the dataset, in 3 we show how it is possible
to encode the knowledge base in an HD-vector and then to retrieve information from it. Finally in
section 4 we describe the training procedure and comment on the results of the tests.
2
Dataset
The dataset is composed of 28x28 RGB images. In each image there are two geometrical figures of
four possible colors. The geometrical figures can be in four different positions in the image, namely
top-left, top-right, bottom-left, bottom-right. The colors and shapes used are the following:
• colors: red, green, magenta, orange;
• shapes: circle, square, triangle, cross.
All possible images containing two figures were created. A sample of the images in the dataset is
presented in Figure 1. The number of total images was 3072. 30% of these images were used as test
set, and the rest of the images were used as training set.
3
HD representation of the knowledge base
Each picture in the dataset can be described in terms of the following set of natural language concepts:
"position", "color", "shape", "top-left", "top-right", "bottom-left", "bottom-right", "red", "green",
"magenta", "orange", "circle", "square", "triangle", "cross". For example the first picture of Figure
1 can be described by the following sentence: "In position top-left there is a shape of type square
and color magenta; and in position top-right there is a shape of type square and color green". In
the following we show how by using a method known as HD-computing, it is possible to store and
retrieve the information contained in this sentence in one vector [3, 6].
We will take the HD-vectors to be 1000 dimensional vectors where each component is randomly
2
chosen to be 1 or -1. We will label HD-vectors using letters in italics, like k, m, n. Let’s define two
operations over the HD-vectors: entangle and grouping. Entangle is defined as follows:
⊗:
k =m⊗n
⇔
ki = XOR(mi , ni ) ∀i ∈ {1, . . . , 1000}
(1)
−1
Please notice that XOR = (XOR) , so we will sometimes refer to the previous operation also as
dis-entangle. Grouping is defined as follows:
⊕:
k =m⊕n
⇔
ki = m i + ni
∀i ∈ {1, . . . , 1000}
(2)
Also it will be useful to define a distance between two vectors. We choose as distance the following
function:
P
(mi · ni )
(3)
cos(m, n) = p P i 2
P
( i mi ) · ( i n2i )
We associate each of the concepts listed in the previous paragraph with a random HD-vector. The
HD-vector representing a concept will have the same name as the concept but it will be written in
italics. So for example to the concept "color" will correspond the HD-vector color.
To each image in the dataset we can associate an HD-vector in the following way that we will illustrate
with an example. Consider the first picture in Figure 1. The natural language description of the
picture is the following: "In the position top-left there is a shape of type square and color magenta.
And in the position top-right there is a shape of type square and color green". To such a description
we will associate the following HD-vector:
m = top−lef t ⊗ (shape ⊗ square ⊕ color ⊗ magenta)
⊕ top−right ⊗ (shape ⊗ square ⊕ color ⊗ green)
(4)
It can be shown that the information stored in the vector m can be retrieved. For example to retrieve
the information about the shape in position "top-left" we can dis-entangle the vector m with the
vector top−lef t and then with the vector shape. The resulting vector will be much closer to the
vector square than to any of the other three vectors: triangle, circle and cross.
Following the previous example we associated to each image I j in the dataset an HD-vector mj ,
building the following dataset:
D = {I j , mj }3072
(5)
j=1
3.1
Querying the knowledge base
The vectors mj just defined contain information about the pictures and can be queried to retrieve some
of the properties of the picture. Querying an HD-vector consists in applying a set of operations in a
specific sequence. In the following we will present the queries we used to train our architecture. The
queries will be presented at first in natural language followed by the corresponding set of operations
in the space of HD-vectors. To do so let’s first define the set Positions:
P ositions = {top−lef t, top−right, bottom−lef t, bottom−right}
In the following the questions.
Question 1: "Is there a circle in the picture?"
X
cos(circle , shape ⊗ pos ⊗ m) > 0.5,
(6)
pos ∈ P ositions.
(7)
pos ∈ P ositions.
(8)
pos
Question 2: "Is there the color green?"
X
cos(green , color ⊗ pos ⊗ m) > 0.5,
pos
Question 3: "Is there a magenta triangle?"
X
cos(magenta , color⊗pos⊗m)·cos(triangle , shape⊗pos⊗m) > 0.25,
pos ∈ P ositions.
pos
(9)
Question 4: "Is there a square in the bottom-left?":
cos(square , shape ⊗ bottom−lef t ⊗ m) > 0.5.
3
(10)
Question 5: "Is the shape in position top-left the same as the one in top-right?":
cos( shape ⊗ top−lef t ⊗ m, shape ⊗ top−right ⊗ m) > 0.5.
(11)
Each of the previous questions has a positive or a negative answer for each of the pictures in the
dataset. So to each picture I j we will associate five numbers, namely q1j , q2j , q3j , q4j , q5j . The number
q1j will be equal to 1 if the answer of Question 1 for the j−th picture is true, otherwise qj1 will be 0.
We will use these values during the training of the network as target values. So let’s add these values
to the dataset D previously defined:
D = {I j , mj , q1j , q2j , q3j , q4j , q5j }3072
(12)
j=1
4
Training and Testing
We trained a FFW network with two hidden layers of 200 rectified linear units each. The network
was asked to return as output the HD-vector describing the picture. For this reason the output of the
network was a layer of 1000 nodes with hyperbolic tangent as activation function.
The network was trained in order to associate to each input image an HD-vector such that when the
vector was queried, it returned the correct information about the image. For this reason the error
function that we minimized during the training was composed of a term for each of the questions.
In particular for each equation ( 7, 8, 9, 10, 11 ) we took the left side of the equation and computed
its value for each image in the dataset by substituting in the equation the term m with the output
of the network net(I j ) (here net is the function implemented by the network). The result of the
computation was forced to be closer to 1 if the answer to the question for the picture I j was true,
otherwise the result of the computation was forced to be closer to 0. Let’s consider for example
Question 1 and its equation (eq. 7). Here we defined the error term E1 as:
!2
X X
j
j
E1 =
cos circle , shape ⊗ pos ⊗ net(I ) − q1
(13)
j
pos
In the same way we defined an error term for each of the questions in the previous paragraph. We
called E2 , E3 , E4 , E5 the error terms relative to Question 2, Question 3, Question 4 and Question 5.
The error function that we minimized during the training was the following sum:
5
X
E=
Ei
(14)
i=1
We developed two kinds of test for the network. In a first test the network was asked to answer the
questions used during the training, but on new data. In the second test the network was asked to
answer new questions that were not used in the training. In the first test the network was tested on
30% of the dataset, on data that were not used during the training. For each example in the test set we
evaluated the answer to each of the questions by substituting the output of the net net(I j ) with the
vector m in the equation representing the question, assuming the answer to be true if the inequality
was respected, false otherwise. The network reached an accuracy of 100% on all the questions. On
the second experiment we tested the network on new questions. In particular we asked questions
similar to question 7, but relative to the three other shapes: "square", "triangle" and "cross". In these
cases we obtained accuracy values of 72%, 69% and 60%. Interestingly the worst performance was
obtained on the "cross" shape, which was the shape that was never used in any of the questions in the
training.
5
Conclusions
In this paper we presented an architecture for VQA that uses HD-computing to encode the knowledge
base and evaluating a query against it. This choice makes the system easy to train in an end-to-end
fashion. The system proved to work very well and to generalize well on a simple task. Further
experiments are needed to test the architecture on more challenging and natural benchmarks.
Acknowledgments
This work was funded by ERC Advanced Grant Number 323674 “FEEL” and ERC Proof of Concept
Grant Number 692765 "FeelSpeech".
4
References
[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International
Conference on Computer Vision, pages 2425–2433, 2015.
[2] Haoyuan Gao, Junhua Mao, Jie Zhou, Zhiheng Huang, Lei Wang, and Wei Xu. Are you talking
to a machine? dataset and methods for multilingual image question. In Advances in Neural
Information Processing Systems, pages 2296–2304, 2015.
[3] Aditya Joshi, Johan Halseth, and Pentti Kanerva. Language recognition using random indexing.
arXiv preprint arXiv:1412.7026, 2014.
[4] Pentti Kanerva. Hyperdimensional computing: An introduction to computing in distributed
representation with high-dimensional random vectors. Cognitive Computation, 1(2):139–159,
2009.
[5] Jayant Krishnamurthy and Thomas Kollar. Jointly learning to parse and perceive: Connecting
natural language to the physical world. Transactions of the Association for Computational
Linguistics, 1:193–206, 2013.
[6] Abbas Rahimi, Pentti Kanerva, and Jan M. Rabaey. A robust and energy-efficient classifier
using brain-inspired hyperdimensional computing. In Proceedings of the 2016 International
Symposium on Low Power Electronics and Design. ACM, 2016.
[7] Mengye Ren, Ryan Kiros, and Richard Zemel. Image question answering: A visual semantic
embedding model and a new dataset. Proc. Advances in Neural Inf. Process. Syst, 1(2):5, 2015.
[8] Licheng Yu, Eunbyung Park, Alexander C Berg, and Tamara L Berg. Visual madlibs: Fill in the
blank description generation and question answering. In Proceedings of the IEEE International
Conference on Computer Vision, pages 2461–2469, 2015.
5
| 9cs.NE
|
Imperial College London
arXiv:1502.06096v1 [cs.NE] 21 Feb 2015
Department of Computing
Reinforcement Learning in a Neurally Controlled Robot
Using Dopamine Modulated STDP
by
Richard Evans
Submitted in partial fulfilment of the requirements for the MSc Degree in
Advanced Computing of Imperial College London
September 2012
Contents
1 Introduction
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Project Aims . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Background
2.1 Neurons . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.1 Biological Neurons . . . . . . . . . . . . . . . . .
2.1.2 Hodgkin-Huxley Model . . . . . . . . . . . . . .
2.1.3 Izhikevich Model . . . . . . . . . . . . . . . . . .
2.2 Models of Neural Networks . . . . . . . . . . . . . . . .
2.2.1 Artificial Neural Networks . . . . . . . . . . . . .
2.2.2 Spiking Neural Networks . . . . . . . . . . . . .
2.3 Reinforcement Learning in Machine Learning . . . . . .
2.3.1 Markov Algorithm’s . . . . . . . . . . . . . . . .
2.3.2 Eligibility Traces . . . . . . . . . . . . . . . . . .
2.3.3 Continuous Parameter Spaces . . . . . . . . . . .
2.4 Reinforcement Learning in the Brain . . . . . . . . . . .
2.4.1 BCM Theory . . . . . . . . . . . . . . . . . . . .
2.4.2 Spike-timing Dependent Plasticity . . . . . . . .
2.4.3 Dopamine Modulated STDP . . . . . . . . . . .
2.4.4 Time-Based DA-modulated STDP . . . . . . . .
2.4.5 Plasticity and Stability . . . . . . . . . . . . . .
2.5 Neural encoding . . . . . . . . . . . . . . . . . . . . . .
2.6 Taxis . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.7 Spiking Neural Network Controlled Robots . . . . . . .
2.7.1 Robot Training using Genetic Algorithm’s . . . .
2.7.2 Robot Training using STDP . . . . . . . . . . . .
2.7.3 Robot Training using Reward Modulated STDP
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3 Methodology
3.1 Environment . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Robot and Sensors . . . . . . . . . . . . . . . . . . . . . .
3.3 Neural Model . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.1 Numerical Approximation . . . . . . . . . . . . . .
3.4 Phasic Activity . . . . . . . . . . . . . . . . . . . . . . . .
3.5 Sensorimotor Encoding . . . . . . . . . . . . . . . . . . . .
3.5.1 Sensor to Neuron Encoding . . . . . . . . . . . . .
3.5.2 Motor Velocity Calculation . . . . . . . . . . . . .
3.6 Network Architecture . . . . . . . . . . . . . . . . . . . .
3.7 Plasticity . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.8 Moving Dopamine Response from Unconditioned-Stimulus
3.9 Network Stability . . . . . . . . . . . . . . . . . . . . . . .
3.10 Taxis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.11 Exploration . . . . . . . . . . . . . . . . . . . . . . . . . .
i
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
to Conditioned
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
2
2
2
3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4
4
4
5
5
6
6
7
7
7
9
9
10
10
10
11
14
14
15
15
16
16
16
16
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
Stimulus .
. . . . . .
. . . . . .
. . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
18
18
20
22
22
22
23
23
23
23
25
27
29
30
32
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4 Results and Discussion
4.1 Orbiting Behaviour . . . . . . . . . . . . . .
4.1.1 Experimental Set-up . . . . . . . . .
4.1.2 Results and Discussion . . . . . . . .
4.2 Food Attraction Learning . . . . . . . . . .
4.2.1 Experimental Set-up . . . . . . . . .
4.2.2 Results and Discussion . . . . . . . .
4.3 Food-Poison Plasticity Learning . . . . . . .
4.3.1 Experimental Set-up . . . . . . . . .
4.3.2 Results and Discussion . . . . . . . .
4.4 Dopamine Response to Secondary Stimulus.
4.4.1 Experimental Set-up . . . . . . . . .
4.4.2 Results and Discussion . . . . . . . .
4.5 Secondary Behaviour Learning . . . . . . .
4.5.1 Experimental Set-up . . . . . . . . .
4.5.2 Results and Discussion . . . . . . . .
4.6 Dual Behaviour Learning . . . . . . . . . .
4.6.1 Experimental Set-up . . . . . . . . .
4.6.2 Results and Discussion . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
33
33
33
34
38
38
38
42
42
42
45
45
46
48
48
49
52
52
52
5 Conclusions
55
5.1 Reinforcement Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.2 Biological Plausibility and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6 Future Work
57
ii
Abstract
Recent work has shown that dopamine-modulated STDP can solve many of the issues associated with
reinforcement learning, such as the distal reward problem. Spiking neural networks provide a useful technique
in implementing reinforcement learning in an embodied context as they can deal with continuous parameter
spaces and as such are better at generalizing the correct behaviour to perform in a given context.
In this project we implement a version of DA-modulated STDP in an embodied robot on a food foraging
task. Through simulated dopaminergic neurons we show how the robot is able to learn a sequence of
behaviours in order to achieve a food reward. In tests the robot was able to learn food-attraction behaviour,
and subsequently unlearn this behaviour when the environment changed, in all 50 trials. Moreover we show
that the robot is able to operate in an environment whereby the optimal behaviour changes rapidly and so
the agent must constantly relearn. In a more complex environment, consisting of food-containers, the robot
was able to learn food-container attraction in 95% of trials, despite the large temporal distance between the
correct behaviour and the reward. This is achieved by shifting the dopamine response from the primary
stimulus (food) to the secondary stimulus (food-container).
Our work provides insights into the reasons behind some observed biological phenomena, such as the
bursting behaviour observed in dopaminergic neurons. As well as demonstrating how spiking neural network
controlled robots are able to solve a range of reinforcement learning tasks.
Acknowledgements
I would like to thank my supervisor Prof. Murray Shanahan. Whose support and guidance helped greatly
in allowing me to conduct research into neural robotics.
1
Chapter 1
Introduction
1.1
Motivation
The ability of a robot to operate autonomously is a highly desirable characteristic and has applications in
a wide range of areas, such as space exploration, self-driving cars, cleaning robot and assistive robotics.
Looking at the animal kingdom, a key property that a large variety of animals possess is that they are
able to learn which behaviours to perform in order to receive a reward or avoid unpleasant situations, for
example in foraging for food or avoiding prey. This is also a useful property for robots to have, where we
want the robots to achieve some goal but the exact sequence of behaviours needed to achieve this goal may
be highly complex or change over time. This learning paradigm is known as reinforcement learning and has
been researched in the context of machine learning for many years. Despite this, even simple animals are
able to outperform robots in the vast majority of real world reinforcement learning problems.
In recent times some of the underlying neural processes by which the brain can solve these reinforcement
learning tasks has started to be revealed. Changes in connection strength between neurons has long been
thought to be the key process by which animals are able to learn. A key process that is believed to control
how the connection weights are modified is that of dopamine-modulated spike timing dependent plasticity
(DA-modulated STDP) [1]. Using spiking neural networks, that implement these models of learning and
plasticity, to control robots has two main advantages. Firstly, it is hoped that these neurally inspired control
algorithms will, at least in some domains, be able to outperform their classical machine learning counterparts.
This is due to the vast arrays of problems that real brains can deal with and their implicit ability to deal with
continuous environmental domains. Secondly, by implementing the current models of neurological learning
in an embodied context we can gain greater insight into the range of behaviour that can be explained by
these models, and also where these models fail.
1.2
Project Aims
In this project we will attempt to solve a reinforcement learning task of food foraging and poison avoidance
using a robot controlled with a spiking neural network. By incorporating several aspects of the brain that
have been observed we aim to show how a robot can learn attraction and avoidance behaviours in a dynamic
environment. By designing an environment that requires a sequence of behaviours before food is reached
we aim to show that the robot can learn by propagating the reward signal (dopamine) to reward-predicting
stimulus. In this way we aim to show how a robot is able to learn sequences of behaviour over larger
time-scales than have previously been demonstrated.
This work builds on the models of Izhikevich [1] and others in providing the neurological models that
will be used to control the robot. The use of these models to control a robot in a foraging/avoidance task
is based on the work of Chorley & Seth [2]. We aim to show that through several key extensions, such
as incorporating the dopamine signal into the network directly, as well as using a more generic network
architecture, our robot is able to deal with a much wider variety of problems, learn over longer time-scales
and learn sequences of behaviour that previous implementations wouldn’t have been able to learn. These
are vital properties for an agent interacting in the real world and a key step towards a fully autonomous
agent that can deal with the wide variety of problems scenarios that the real world provides.
The mechanisms of plasticity and stability play a key role in many properties of the brain, such as
learning, memory and attention. Through implementing a neurally inspired system that incorporates many
of the current neural models we aim to show how well these models are able to explain some of the learning
behaviours that are seen in the animal world.
2
1.3
Thesis Outline
The remainder of this report is structured into five chapters. In chapter 2 we review current research and
relevant literature in the context of reinforcement learning in neurally controlled robots. Chapter 3 presents
the techniques, models and methodology that was used in this project. The architecture of the robot,
set-up of the environment and the spiking neural network control architecture is discussed. The results of
evaluating our robot in a wide range of scenarios is discussed in chapter 4, along with a discussion of which
properties of the robot allow it to achieve these results. In chapter 5 we relate the results to the wider
context of reinforcement learning and neurological models and discuss their implications. Finally in chapter
6 we present possible future work, based on the results of this paper.
3
Chapter 2
Background
2.1
Neurons
The underlying biological structure of neurons, and models that have been used to simulate them, are
discussed in this section.
2.1.1
Biological Neurons
The brain consists of a vast connected network of neurons. These neurons can be further subdivided into
different types, however the general architecture is the same in almost all neurons. Figure 2.1 shows the basic
structure of a typical neuron. Within the neuron communication takes place via the electrical properties of
Dendrites
Cell Body (Soma)
Axon
Myelin Sheath
Axon Hillock
Axon Terminal
Synapse
Nucleus
Figure 2.1: The structure of a neuron.
the cell, figure 2.2 shows the membrane potential of a neuron plotted over time whilst receiving a constant
input current. In its resting state the membrane potential of the interior of a neuron sits at approximately
-70mV. Each neuron receives inputs from several other neurons via its dendrites. When one of these presynaptic neurons fires it has the effect of raising or lowering the membrane potential of the post-synaptic
neuron. Once the membrane potential at the axon hillock (see figure 2.1) reaches a critical threshold,
voltage gated ion channels open along the axon which causes the neuron to rapidly depolarize. This is
referred to as the neuron spiking or firing. After this wave of depolarization reaches the axon terminal it is
converted to a chemical signal that is passed across the synapse to the next neuron. Once this depolarization
reaches a critical threshold a second voltage gated ion channel is opened which causes the neuron to rapidly
repolarize before returning to its resting potential. This repolarization overshoots the resting potential
causing hyperpolarization. After firing the neuron enters a refractory period during which it cannot fire.
The two main types of neurons that will be used in this paper are excitatory and inhibitory neurons.
Excitatory neurons have the effect of raising the membrane potential of the post-synaptic neuron, as such
increasing its likelihood of firing. Whereas inhibitory neurons have the effect of decreasing the membrane
potential of the post-synaptic neuron, decreasing its likelihood of firing.
4
Membrane Potential of a Single Neuron Over Time
30
20
Membrane Potential (mV)
10
0
−10
−20
−30
−40
−50
Firing Threshold
−60
Resting Potential
−70
Hyperpolarization
80
90
100
110
120
130
140
150
160
170
180
Time (ms)
Figure 2.2: The membrane potential of a neuron receiving constant input current plotted over time.
The synapse between neurons has an associated weight or strength. This is the degree to which the
post-synaptic membrane potential is raised or lowered following a pre-synaptic spike. The mechanisms by
which neurons communicate have several useful properties. Firstly, communication within the cell is fast,
via the use of voltage gated ion channels. Secondly, by communicating in an all or nothing fashion to the
post-synaptic cell, the network of neurons is able to represent highly complex non-linear functions. Research
has shown that spiking neural networks can represent any real-valued function to within a specified degree
of accuracy [3].
2.1.2
Hodgkin-Huxley Model
In 1952 Hodgkin and Huxley developed an accurate model of the membrane potential of a neuron [4] based
on modelling the ion channels within the neuron. The equation is defined as:
C
X
dv
=−
Ik + I
dt
(2.1)
k
where
C = The capacitance of the neuron,
v = The membrane potential of the neuron,
Ik = The various ionic currents that pass through the cell,
I = The external current coming from pre-synaptic neurons,
t = Time.
The ionic currents are modelled using further differential equations, the details of which are not included here
but a full definition can be found in Hodgkin & Huxley [4]. The Hodgkin-Huxley equation is very biologically
accurate, however it incurs a heavy computational cost and as such is infeasible in most practical situations.
2.1.3
Izhikevich Model
Several attempts [5, 6, 7] have been made to formulate a model that provides a good compromise between
biological accuracy and computational feasibility. One of the best models in terms of computational efficiency
and biological accuracy was formulated by Izhikevich [8]. In the Izhikevich model the membrane potential
5
is modelled as:
dv
= 0.04v 2 + 5v + 140 − u + I
(2.2)
dt
Where v is the membrane potential and u is the recovery variable that determines the refractory period of
the neuron and is modelled as:
du
= a(bv − u)
(2.3)
dt
Where a and b are parameters of the model. When a spike occurs the membrane potential is reset and the
recovery variable is incremented, formally:
v←c
if v ≥ 30 then
(2.4)
u←u+d
The four variables a,b,c and d have been shown to be able to create neurons that have a wide variety
of behaviours. Figure 2.3 summarizes some of the neuronal types that can be modelled by varying these
parameters.
peak 30 mV
cay
with r
ate a
FS
RS,IB,CH
0.2
8
parameter d
reset d
de
parameter b
reset c
v(t)
RZ
LTS,TC
0.25
sensitivity b
regular spiking (RS)
0 0.02
0.1
IB
4
2
0.05
u(t)
RS
FS,LTS,RZ
-65
parameter a
CH
TC
-55
-50
parameter c
intrinsically bursting (IB)
chattering (CH)
fast spiking (FS)
thalamo-cortical (TC)
resonator (RZ)
low-threshold spiking (LTS)
v(t)
I(t)
thalamo-cortical (TC)
20 mV
40 ms
-63 mV
-87 mV
Figure 2.3: An overview of some types of neurons that can be modelled with the Izhikevich model1 .
2.2
Models of Neural Networks
In this section the different approaches that have been used to model networks of connected neurons are
discussed.
2.2.1
Artificial Neural Networks
The first computational model of a neural network that was developed was the artificial neural network
(ANN). The ANN consists of a set of processing units, or neurons. Each neuron computes the weighted
sum of its inbound connections, which may be other neurons or external inputs. This value is then passed
through an activation function, the result of which is then passed to the next set of neurons or as an output
of the system. The most common activation function that is used is the sigmoid function as this provides a
differentiable approximation to the all or nothing processing that is used in real neurons. Figure 2.4 shows
the model of a single neuron in an ANN.
1 Electronic
version of the figure and reproduction permissions are freely available at www.izhikevich.com
6
inputs
x1
weights
w1j
x2
w2j
x3
w3j
Σ
φ
oj
output
transfer
function
...
...
xn
activation
function
wnj
Figure 2.4: A perceptron that computes the weighted sum over the inputs x with weights w, then passes
this value through the activation function to produce a binary output.
Initially these ANN’s consisted of a single layer of neurons [9, 10]. These ANN’s, known as perceptrons,
are limited in that they can only represent linear functions of their inputs. For example in a classification
task, the data must be linearly separable for a single layer network to be able to classify all data correctly.
To combat these problems a multilayer perceptron was introduced [11], where outputs from previous layers
serve as inputs to the next layer.
The development of an efficient training algorithm for the multilayer perceptron, known as back-propagation,
which updates the weights of the network to minimize the squared error on a set of training examples has
allowed ANN’s to be used in a wide variety of fields.
The biggest limiting factor in artificial neural networks, especially as a model of the brain, is that they
can’t process information encoded in the time domain.
2.2.2
Spiking Neural Networks
A more realistic model of biological neural networks was developed with spiking neural networks (SNN).
These have the advantage of being able to process temporally distributed data, as well as having a large
memory capacity. Unlike ANN’s, which are commonly directed acyclic graphs, SNN’s are more amenable to
networks with cycles in them. These cycles allow spiking neural networks to have a form of working memory.
The individual neurons in a spiking neural network can be modelled in a variety of ways, several of these
models were outlined in section 2.1. Some methods by which SNN’s can be trained are outlined in section
2.4.
2.3
Reinforcement Learning in Machine Learning
Reinforcement learning is the process by which an agent can automatically learn the correct behaviour in
order to maximize its performance. Reinforcement learning is an important learning paradigm for a robot
that interacts with the real world, in this scenario the robot is often not explicitly told what the correct
action is in a given environment state, it must be inferred from rewards or punishments that are received
after an action is performed. In the standard reinforcement learning set-up the agent is given a reward when
it performs an action or is in a specific state, and it seeks to maximize this reward over time. Sutton & Barto
give a good history of reinforcement learning algorithms [12], the key approaches relevant to this paper are
outlined below.
2.3.1
Markov Algorithm’s
The majority of work on reinforcement learning have modelled the problem as a Markov decision process.
The agent can visit a finite number of states, in each state the agent can perform a finite number of actions
that transform the environment into a new state. The state that is reached after performing an action only
depends on the previous state, and the action. This agent-environment interaction is shown in figure 2.5.
An agent’s policy π is defined as the probability of choosing an action in a given state: π(s, a) = p(a|s, π).
An important function in reinforcement learning is the Q-function, this gives the expected future reward
given a state and action, defined as:
Qπ (s, a) = E[
∞
X
γ k rt+k+1 |π, st = s, at = a]
k=0
7
(2.5)
Agent
state
st
reward
rt
rt+1
st+1
action
at
Environment
Figure 2.5: The agent-environment interaction. Given a state st and reward rt at time t the agent must
decide which action at to perform. This action will update the state st+1 and result in a further reward
rt+1 . Figure from Sutton & Barto [12].
Where γ determines the importance we place on rewards that are closer in time than ones that are more
distant. If we know the Q-function then the optimal policy becomes trivial, we pick the action with highest
expected reward. Most RL algorithms are concerned with finding the value of the Q-function.
An underlying feature of most RL algorithms is the concept of temporal difference (TD) learning. Where
the difference between the predicted reward and received reward is used to update the agents policy. RL
algorithms are divided into off-line and on-line learning. Off-line algorithms assume that a series of states
and actions generated as in figure 2.5 will eventually reach a terminating state. Learning is only performed
once this terminating state is reached. On-line algorithms, however, aim to continuously update the agents
estimate of the Q-function as the agent explores the environment. On-line algorithms provide the advantage
of not having to wait until the end of an episode to incorporate new information and, as is the case in this
paper, an agent interacting in the real world will not always have a clearly defined termination condition
that causes an episode to end.
Within the set of on-line RL algorithms there is a further distinction between on-policy and off-policy
algorithms. On-policy algorithms attempt to learn the optimal Q-function and policy, whilst simultaneously
following their own estimate of the optimal policy. Note that the agents estimate of the optimal policy is
likely to be different to the actual optimal policy, as such this method can get stuck in local minima and not
converge to the optimum policy. Off-policy algorithms on the other hand attempt to learn the optimal policy
whilst not necessarily following it, choosing actions with a probability related to their expected reward. This
has the advantage that the agent can further explore the state space, however this comes at the cost of a
possible reduced reward compared to the on-policy methods, due to the exploration of sub-optimal states.
In this paper we will demonstrate a mechanism by which off-policy learning can be achieved using spiking
neural networks.
The best known on-line algorithm is Sarsa, the outline of which is given in Algorithm 1. The key step is
that the observed reward r and the predicted Q-value at the next time step are used to estimate the correct
Q-value: r + γQ(s0 , a0 ). This is compared against the predicted Q-value, Q(s, a), to give a prediction error
which can be used to update the estimate of Q(s, a).
Algorithm 1 Sarsa
Initialize Q(s, a) arbitrarily
Repeat (for each episode)
Initialize s
Choose a from s using policy derived from Q
Repeat (for each step of episode):
Take action a, observe r, s0
Choose a0 from s0 using policy derived from Q
Q(s, a) ← Q(s, a) + α[r + γQ(s0 , a0 ) − Q(s, a)]
s ← s0 ; a ← a0 ;
until s is terminal
The best known algorithm for off-line learning is Q-learning, the pseudocode for which is given in Algorithm 2. The key difference is that the algorithm takes the maximum Q-value in the next state, over all
actions.
8
Algorithm 2 Q-Learning
Initialize Q(s, a) arbitrarily
Repeat (for each episode)
Initialize s
Choose a from s using policy derived from Q with exploration
Repeat (for each step of episode):
Take action a, observe r, s0
Q(s, a) ← Q(s, a) + α[r + γmaxa0 Q(s0 , a0 ) − Q(s, a)]
s ← s0
until s is terminal
2.3.2
Eligibility Traces
Both the Q-Learning and Sarsa algorithm as described only have one state of lookahead. They only look
at the reward at the next time step when updating the Q-function. Due to this we often have to revisit a
state many times before the algorithm will converge. We encounter further problems if we relax the Markov
property, for example if the reward received for an action is not received until several states later. In this case
it can be shown that the algorithms above may fail to converge to the correct Q-function. What we would like
to be able to do is incorporate the reward from all future states in our update function. This is accomplished
by introducing the concept of an eligibility trace. The eligibility trace for a state keeps a weighted track
of visits to that state, in this way we can update previously visited states when we receive a new reward,
the concept of the eligibility trace becomes important when discussing the dynamics of the reinforcement
learning in the brain, and in our implementation of DA-modulated STDP. Figure 2.6 demonstrates how the
eligibility trace changes over time, as the state is repeatedly visited.
Time
Figure 2.6: The eligibility trace for a state over time, as it is repeatedly visited.
We can incorporate the eligibility trace into the Sarsa algorithm (referred to as Sarsa(λ)), if we define
e(s, a) as the eligibility trace for the state s and action a then the Q-function update becomes:
Q(s, a) ← Q(s, a) + αe(s, a)[r + γQ(s0 , a0 ) − Q(s, a)]
(2.6)
Where α is the step size; the amount that we update the Q values by, at each time step. γ determines the
importance we place on more recent rewards over future rewards. Q(s0 , a0 ) is the Q value for the next state
and action. Note that this update is applied to all states and actions, unlike in the previous algorithm where
it was restricted only to the current state-action pair. The full Sarsa(λ) procedure is outlined in Algorithm
3.
2.3.3
Continuous Parameter Spaces
So far our discussion has been limited to reinforcement learning in discrete state and action spaces. The
simplest way of applying RL to a continuous state space is simply to divide the continuous space up such
that it becomes a discrete one. However there are two main issues with this, firstly we may end up with
a very large parameter space, especially if we wish to have a fine grained representation of the continuous
parameter space. This will mean that the RL algorithm will take a long time to converge on the optimal
solution as the probability of visiting each state will be low. The second issue is that we lose the ability
to generalize, it often the case that knowing the expected future reward for a particular state will tell us
something about the expected reward for nearby states.
In the continuous case the Q-function becomes a continuous valued function over s and a. If we treat
each update from the discrete RL algorithm as an example input-output of this function. Then our problem
becomes one of generalizing from specific inputs. This is a standard machine learning problem, which can
9
Algorithm 3 Sarsa(λ)
Initialize Q(s, a) arbitrarily and e(s, a) = 0, for all s,a
Repeat (for each episode)
Initialize s, a
Repeat (for each step of episode):
Take action a, observe r, s0
Choose a0 from s0 using policy derived from Q
δ ← r + γQ(s0 , a0 ) − Q(s, a)
e(s, a) ← e(s, a) + 1
For all s,a:
Q(s, a) ← Q(s, a) + αδe(s, a)
e(s, a) ← γλe(s, a)
s ← s0 ; a ← a0
until s is terminal
be solved through, among other things, neural networks, decision trees, genetic algorithms and gradient
descent.
Another method of performing reinforcement learning in a continuous parameter space is to use spiking
neural networks, a full description of spiking neural networks was given in section 2.2.2. Spiking neural
networks offer the advantage that, if constructed correctly, they can generalize to learn the correct action in
a previously unseen state. Reinforcement learning was first investigated, and algorithms developed, before
it was fully understood how the brain performed similar tasks. As will be shown in the following section the
underlying processes used by the brain has many correlates with these reinforcement learning algorithms.
In this paper we implement a spiking neural network version of reinforcement learning, how the underlying
processes in our implementation compare to standard machine learning algorithms for reinforcement learning
are discussed in section 5.1.
2.4
Reinforcement Learning in the Brain
In recent times it has been shown that the brain implements a reinforcement learning system similar to the
TD-Learning proposed by Sutton & Barto [12]. The mechanisms by which this happens are outlined below.
2.4.1
BCM Theory
In 1981 Bienenstock, Cooper and Munro developed a model by which synapses can increase and decrease over
time [13]. This slow increasing and decreasing of synaptic weights is known as long term potentiation (LTP)
and long term depression (LTD) respectively. In BCM theory the amount that a synaptic weight changes
is dependent on the product of the presynaptic firing rate and a non-linear function φ of the postsynaptic
activity. For low post-synaptic firing rates φ is negative and for high post-synaptic firing rates φ is positive.
The effect of this is that if the pre-synaptic neuron is causing a lot of firing in the post-synaptic neuron then
the synapse will be potentiated.
Some evidence for the BCM model has been seen in the brain, however it cannot account for synaptic
potentiation based on spike timings, which has also been observed in the brain.
2.4.2
Spike-timing Dependent Plasticity
The main mechanism through which the connections in the brain are modified over time is believed to be
spike-timing dependent plasticity [14, 15]. The core concept is that neuron connection strength is increased
when pre-synaptic activity is followed by post-synaptic activity, and weakened when post-synaptic activity
is followed by pre-synaptic activity. The synaptic weight update rule under STDP is defined as:
(
+
A+ e−∆t/τ
if ∆t ≥ 0
∆w =
(2.7)
− ∆t/τ −
−A e
if ∆t < 0
Where ∆w is the weight update, ∆t is the time difference between the pre and post synaptic neurons
firing and A+ ,A− ,τ + and τ − are constants that define how STDP is applied over time. These constants are
normally chosen so that long term depression is favoured over long term potentiation to prevent uncontrollable growth of w. Figure 2.7 shows how ∆w changes with respect to ∆t, this shows the window over which
STDP can act. Large values of ∆t will have negligible effect on ∆w.
10
∆w
∆t
Figure 2.7: Graph showing how the weight update ∆w relates to the ∆t = tpost − tpre parameter.
There are two main update schemes that are followed when calculating the change in weight ∆w. In the
all-to-all scheme all pre-synaptic and post-synaptic spikes within the STDP window are considered when
updating the weight. In the nearest-neighbour scheme only the most recent pre or post synaptic spike is
considered. In practice these methods are often equivalent, as the probability of multiple firings for a specific
neuron within the STDP window is low. STDP can be thought of as an associative learning mechanism. If
two inputs to the neural network are often presented together, and in sequence, then the connections from
the first input to the second input will become strengthened.
Experiments have provided evidence for both STDP and the BCM model in the brain. Therefore any
model of synaptic weight modification will need to account for both these processes. Izhikevich [16] has
shown that under the nearest-neighbour mechanism it is possible for the observed properties of the BCM
model to be generated through STDP. As such it may be that the underlying mechanism for synaptic
potentiation/depression in the brain is indeed STDP.
2.4.3
Dopamine Modulated STDP
STDP alone cannot account for the ability of an animal to perform reinforcement learning as it has no
concept of reward. It has been observed that dopamine can modulate synaptic plasticity when it is received
within a 15-25s window [17]. Dopamine in the brain is modulated by way of dopaminergic neurons, these
being neurons that use dopamine as their neurotransmitter. The majority (90%) of the dopaminergic neurons
are contained in the ventral part of the mesencephalon [18]. Within this region the most important area is
the ventral tegmental area (VTA) which is vital in the reward circuitry of the brain. Figure 2.8 shows the
main dopamine pathways in the brain. The VTA has connections to the nucleus accumbens, which plays an
important role in the perception of pleasure, as well as to the hippocampus which has been associated with
planning complex behaviour, decision making and moderating social behaviour [19]. Whenever the VTA is
stimulated it produces a burst of spiking activity which in turn raises the level of dopamine in the nucleus
accumbens and the hippocampus.
Striatum
Frontal
cortex
Substantia
nigra
Nucleus
accumbens
VTA
Hippocampus
Figure 2.8: The main dopamine pathways in the human brain.
11
Dopaminergic neurons are characterized as having two different firing patterns. In the absence of any
stimulus they exhibit a slow (1-8Hz) firing rate [20], this is known as background firing. When stimulated the
dopaminergic neurons exhibit burst firing [21]. Burst firing is where neurons fire in very rapid bursts, followed
by a period of inactivity. This behaviour is illustrated in figure 2.9. One important, especially in the context
of this paper, aspect of burst firing is that for many central synapses bursts facilitate neural transmitter
release whereas single spikes do not [22]. This means that stimulus induced firing of dopaminergic neurons
will result in a spike in the level of dopamine, whereas background firing will not not have a significant effect
on the level of dopamine. This is a useful property for Pavlovian learning and its implications are explored
further in section 3.8.
Membrane Potential
Burst Spiking Behaviour of a Neuron
Time
Figure 2.9: An example of a neuron exhibiting bursting behaviour, the neuron fires in bursts followed by
periods of inactivity.
Matsumoto & Hikosaka [23] have recently shown, via experiments on monkeys, that the dopaminergic
neurons do not function as a homogeneous group as previously thought. Some neurons exhibit a reward
predicting behaviour, with increased firing when a reward is received, and a drop in activity for negative
stimulus. However, other dopamine neurons have been found to exhibit an increased response for both a
negative and a positive stimulus. The spatial location of these two groups is distinct and it is likely that
they play different roles in learning in the different ways that they are connected to the rest of the brain,
though this mechanism is not yet understood fully.
In order for dopamine to be able to modulate synaptic plasticity over relatively large time-scales on
the order of seconds it has been shown that synapses must have a form of synaptic tag that remembers
the activity of the two neurons over a short time [24, 25]. The precise chemical process that underlies this
synaptic tagging is not yet known. Research has shown [26] that this synaptic tag, and long term potentiation
are dissociable. That is, the induction of synaptic potentiation creates the potential for a lasting change
in synaptic efficacy, but does not commit to this change. This is exactly what is needed for reinforcement
learning, in standard machine learning algorithms for reinforcement learning the eligibility trace itself does
not modify the value function (estimate of reward for a given environment state), but it is modified through
a combination of the eligibility trace and received reward.
Izhikevich [1] has shown that dopamine modulated STDP can perform many aspects of reinforcement
learning and can accurately model many of the features of learning observed in the brain. One of the key
problems in reinforcement learning is that the receipt of a reward is often delayed after the action, or neural
pattern, that caused the reward. This is known as the distal reward problem [27]. STDP as described above
works over the millisecond time-scale and so another mechanism is needed to account for this distal reward
problem. This is solved in DA modulated STDP by the inclusion of a variable that acts as a synaptic tag and
has a direct correlation with the eligibility trace used in machine learning. This synaptic tag is incremented
in proportion to the amount of STDP that would normally be applied at a synapse and then decays over
time, it is defined as:
ċ = −c/τc + ST DP (τ )δ(t − tpre/post )
(2.8)
Where τc defines the decay rate of the eligibility trace, and hence the time-scale over which DA modulation
can work. δ(t) is the Dirac delta function and has the effect of step increasing c when STDP takes place.
The value of c over time looks very similar to the eligibility trace of machine learning as shown in figure 2.6.
This variable is then used to modify the synaptic weight using:
ṡ = cd
(2.9)
Where s is the synapse strength and d is the current level of dopamine. One important feature of this
model is that it can deal with random firings in between the stimulus and the reward, this is due to the low
12
probability of randomly occurring coincident firings between the stimulus and reward. As such the eligibility
trace maintains the record of the pre-post firing that we wish to strengthen. Figure 2.10 shows an example
of how the network is able to respond to a delayed reward.
500 ms
pre
post
rew
ard
(de
l
aye
db
y1
-3
sec
)
extracellular dopamine, d(t)
eligibility trace, c(t)
synaptic strength, s(t)
Figure 2.10: Figure from Izhikevich [1] showing how an eligibility trace can cope with a delayed reward.
When the pre then post synaptic neurons coincidentally fire, a delayed reward is received in the form of a
spike in dopamine d. This coincident firing increases the eligibility trace c, which slowly decays. It is still
positive when the reward is received, and as such the synaptic strength s is increased.
Using this model it has been shown that a network of neurons is able to correctly identify a conditioned
stimulus embedded in a stream of equally salient stimuli, even when the reward is delayed enough such that
several random stimuli occur in between the CS and the reward. After training, the learnt response shows
up as an increased firing when the CS is presented compared with the other stimuli, as in figure 2.11.
Figure 2.11: The response of a conditioned stimulus S1 after being subject to DA modulated STDP, as
compared with several unconditioned stimulus. Figure reproduced from Izhikevich [1].
In Pavlovian conditioning [28] repeated pairing of a conditioned stimulus with a reward-producing unconditioned stimulus can move the response from the US to the CS. By the introduction of a set of dopamine
releasing neurons similar to those in the ventral tegmental area in the brain, it was shown [1] that the
DA modulated STDP model can exhibit this response shifting behaviour. The US was initially connected
with maximum weights to the dopamine neurons, simulating an US. When a CS was repeatedly presented
preceding this US there was an increase in the connections between the CS and the dopamine neurons, the
dopamine response to the US stimulus was also reduced. This effect is similar to the effect observed in vivo
in monkeys and rats [29]. A modified version of this DA-modulated STDP process is used in this paper
to control a robot. The ability to move the dopamine response is an important property when multiple,
sequential, behaviours need to be learnt before a reward is received.
13
The above model demonstrates the ability to model several features needed for reinforcement learning.
However, the dopaminergic response to a stimulus does not correlate directly to the the difference between
the expected reward and received reward that is crucial to classical machine learning RL algorithms. This
should manifest itself as a negative (drop in neural activity) response when a reward is expected, but not
received. This dip in neural activity has been observed in vivo [30]. As the model developed by Izhikevich [1]
did not include any sort of working memory it did not produce this dip in activity.
2.4.4
Time-Based DA-modulated STDP
Chorley & Seth [31] have incorporated the “dual-path” model [32, 33] into the DA modulated STDP model
proposed by Izhikevich [1]. In the dual path model there are two pathways between a stimulus and dopaminergic neurons, an excitatory pathway and a time-dependent inhibitory pathway. The consequence of this is
discussed in more depth, in the context of the new hybrid model, below. The new model accounts for several
key features of DA responses observed in the brain, as well as being closer to the standard machine learning
algorithms. These features, as defined by Chorley & Seth [31], are:
• DA neurons display phasic activation in response to unexpected rewards [34, 35].
• These DA neurons display phasic responses to reliably reward-predicting stimuli, yet do not respond
to stimuli which are themselves predicted by earlier stimuli [36].
• Reward-related DA responses reappear if a previously predictable reward occurs unexpectedly [36].
• DA neurons display a brief dip in activity at precisely the time of an expected reward, if that reward
is omitted [30].
An overview of the network architect that was used is given in figure 2.12. The key feature that allows
the network to display a dip in DA firing when a reward is expected, but not received, is that that the PFC
has a basic working memory. This is implemented as a sequence of distinct neural firing patterns that occur
in sequence after a stimulus is presented. This means that when an untrained stimulus (CS) is followed
by a reward producing stimulus (US), the corresponding phasic release of DA will strengthen connections
between the current PFC time-dependent firing pattern and the STR. Therefore the next time the CS is
presented, and after the same time period, there will be an increase in firing in the STR and hence an
increased inhibition of the DA cells.
Figure 2.12: The network architecture used by Chorley & Seth [31], red lines represent excitatory connections
and blue represent inhibitory connections. When neurons in the DA module fire then dopamine is released
which causes STDP of the PFC→STR and SEN→INT pathways. The mean firing rate of the STR module
is also modulated by the amount of dopamine.
2.4.5
Plasticity and Stability
One of the problems with Hebbian learning through STDP is that it can cause runaway feedback loops. For
example, if two neurons fire in sequence then their synaptic weight will be increased. This means that in
future they are more likely to fire in sequence, and as such they are more likely to have their synaptic weight
further increased. Without any kind of stability mechanism this process will quickly cause the weights to
go towards infinity. One simple method that is often used [1, 31] is to restrict the range of values that the
synaptic weights can take, for example limiting an excitatory synapse to the range 0-4mV. This prevents the
14
synaptic weight from becoming negative, and hence becoming an inhibitory synapse, whilst also preventing
the synaptic weight increasing forever.
This synaptic weight capping does not, however, prevent synapses from becoming quickly potentiated
to their maximum allowable value. In the brain it has been shown that highly active neurons have their
excitability decreased over time, and highly inactive neurons have their excitability level increased [37]. This
is achieved by processes that modify the tonic level of Ca2+ within the neurons. This raises or lowers the
membrane potential of the neuron, and hence increases or decreases the neurons excitability. It has also
been shown that within populations of neurons a global mechanism for maintaining homeostatic plasticity is
used [38]. This mechanism has the effect that if the firing rate of a population of neurons is increased then
the strength of all synapses within the population is decreased, and equivalently if the firing rate decreases
then the strength of all synapses is increased.
A consequence of these homeostatic mechanisms is that they can foster competition between competing
behaviours. For example, if a group of neurons has the possibility of representing two distinct behaviours
and the neurons corresponding to behaviour 1 are highly active, whereas the neurons for behaviour 2 aren’t
as active. Then, due to the high firing rate, all the synaptic weights will be reduced, which will have the
result of further reducing the activity of behaviour 2. Since connected neurons with high firing rates will tend
to increase their synaptic weights faster through STDP, the neurons corresponding to behaviour 1 will have
their synaptic weights increased faster than behaviour 2. This process will continue until only the neurons
corresponding to behaviour 1 are firing.
2.5
Neural encoding
In the brain there needs to be a mechanism to convert external stimulus into a neural firing pattern, and
to convert a neural firing pattern into an output of the network. There are three main neural encoding
mechanisms that have been observed in the brain, these being rate coding, temporal coding and population
coding.
In rate coding the strength of the stimulus or output is encoded in the firing rate of a population of
neurons. A stronger stimulus emits a faster firing rate, and equivalently a faster firing rate will result in a
stronger output of the network, such as joint movement. This encoding mechanism was first observed by
Adrian & Zotterman in 1926 [39]. Rate coding is useful when an increase in the value of a stimulus should
result in a correlated increase in the response to the stimulus.
Population coding involves the stimulation of different neurons for different values of the external stimulus.
For example, the direction of movement of the eye could be represented by a network of 180 neurons with
each neuron coding for a direction range of 2 degrees. In practice there is usually some overlap between
the neural firing with a Gaussian distribution in firings being common, this allows a wider range of values
to be represented as well as being more robust to noise. Population coding is used in various parts of the
brain, for example in coding the direction of movement of an observed object [40]. Population coding has
the advantage of being able to react to changes in the external stimulus faster than rate coding.
In temporal coding the external stimulus is represented by the precise spike timings in a population of
neurons, in this way a greater range of parameters can be represented by the same group of neurons than
could be represented with rate coding [41].
In the context of learning taxis behaviour for a simple robot rate coding of sensor inputs/motor outputs
is an appropriate mechanism as it allows the strength of response to an input stimulus to be correlated with
the strength of the input stimulus to be learnt more easily.
2.6
Taxis
In the animal world, one of the simplest forms of behaviour is that of taxis. Where taxis is defined as motion
towards or away from a directed stimulus. Examples include phototaxis, movement directed by light, and
chemotaxis, movement directed by a chemical gradient. This behaviour has advantages such as food foraging
and poison avoidance.
Through a series of thought experiments Braitenberg [42] showed how taxis behaviour, as well as more
complex behaviours, could be implemented in a simple robot. The robots were endowed with two sensors
and two motors, attraction behaviour can then be achieved through connecting the left sensor the right
motor, and the right sensor to the left motor. Avoidance can be achieved through connecting the left sensor
to the left motor and the right sensor to the right motor. A version of these simple Braitenberg vehicles will
be used in this paper to demonstrate reinforcement learning.
15
2.7
Spiking Neural Network Controlled Robots
A variety of research has been conducted into using SNN’s to control robots. Initially this was done by
manually setting the weights of the network [43, 44]. For example Lewis et al. [44] used a very small network
of 4 neurons to control an artificial leg. The complexity of SNN’s means that manually setting the synaptic
weights is only feasible for very small networks. Various mechanisms for training the weights of a SNN have
been used, these are outlined below.
2.7.1
Robot Training using Genetic Algorithm’s
One such solution to the problem of determining the synaptic weights is to use genetic algorithms to determine
the weights. Genetic algorithms learn the best solution to a problem by using a population of possible
solutions. These possible solutions are then evaluated to give some measure of the fitness of each solution
in solving the given problem. Highly fit solutions are then combined to create new individuals which are
replaced with unfit solutions and the process is repeated.
Hagras et al. [45] have used genetic algorithms to train a SNN controlled robot such that it exhibits wall
following behaviour. Evolving SNN’s for robot control work well when the environment is mostly static and
the optimal solution does not change over time. In this case the training can be done off-line in a simulated
environment with a large number of robots. However in dynamic environments or where we are restricted
to online learning with a single robot then GA’s do not perform as well.
2.7.2
Robot Training using STDP
Another solution to determining the synaptic weights is to modify the weights over time using STDP as
described in section 2.4.2. Bouganis & Shanahan [46] effectively used STDP to train a robotic arm to
associate the direction of movement of the end effector with the corresponding motor commands. This
was achieved by a training stage during which the motors were randomly stimulated along with the actual
direction of movement of the end effector. STDP was able to modify the connection weights in the network
such that if you subsequently stimulated the neurons corresponding to a desired end effector direction of
movement, at a given position, then the arm would move in that direction. This is an effective method for
learning associations, however it is limited in that it cannot cope with goal oriented tasks where an agent
has to maximize some reward. It also requires that the action and response happen simultaneously, in many
real world situations this is not the case. For example if we want the robot to reach out and touch an object
there will be a delay between the initial movement, and the act of touching the object.
2.7.3
Robot Training using Reward Modulated STDP
Several experiments have been conducted into training an SNN controlled robot using reward-modulated
STDP. A common learning task used to demonstrate reinforcement learning is that of the Morris water
maze. This learning paradigm consists of placing the agent (originally a rat) in a non-transparent liquid, in
the environment there is a hidden submerged platform. The agent must find this platform in order to escape
the unpleasant experience. When the experiment is run with a rat the platform is initially found by chance.
But in subsequent runs the rat learns to navigate directly towards the platform.
Vasilaki et al. [47] used a spiking neural network to control a robot in the Morris water maze task. The
robot was able to learn to successfully learn to navigate directly towards the hidden platform when placed
close to the platform. The level of reward was simulated as an external parameter that was artificially set
when the platform was found. The robot only received a reward when it reached the platform, as such the
robot was only able to learn the correct behaviour within a small radius around the hidden platform due to
the eligibility trace decaying over larger distances. In this paper the level of reward will be embedded within
the network and as such allow the robot to learn over a much larger time frame.
In the context of learning taxis behaviour using reward modulated STDP Chorley & Seth [2] implemented
a simple Braitenberg vehicle [42] that was able to learn to distinguish between two stimuli, one of which
would elicit a reward. The experimental set-up of the robot is shown in figure 2.13. It is important to note
that in the experimental set-up the robot is already wired for taxis attraction behaviour, with each sensor
only being connected to the opposite motor. The learning phase consists of learning the strength of the
taxis response to two different stimuli. The robot was able to learn to elicit a stronger taxis response to the
reward-giving stimuli, and it was also able to change its behaviour when the reward was switched from one
stimuli to another. Once again the level of reward was simulated as an external parameter rather than being
embedded in the network itself.
In this paper we will extend the work of Chorley in several key ways, these being:
16
Figure 2.13: The experimental set-up used by Chorley & Seth [2].
• In our implementation the robot will not be explicitly wired for attraction behaviour, the robot will
have the ability to learn either attraction or avoidance behaviour.
• By incorporating the level of dopamine into the network directly, we will aim to show how sequences
of behaviour can be learnt.
• Through the use of a form of hunger our implementation will be able to deal with situations in which
there is no fixed optimal behaviour for an environment, and a dynamic set of behaviours must be
learnt.
17
Chapter 3
Methodology
In this paper a simple SNN controlled Braitenberg vehicle was simulated in environments consisting of
food, poison and containers. The robot was subject to DA-modulated STDP and had to learn the correct
behaviour to collect food items. The precise implementation details are outlined in this chapter.
Due to the fact that robot and environment are simulated, all distances and sizes are relative and do not
have a real world unit associated with them, by convention we will take the default unit to be centimetres.
3.1
Environment
The learning task that was used to test reinforcement learning using spiking neural networks was that of
food collection and poison avoidance in a simulated environment by a simple robot. There are four types of
objects that can exist in the environment, these being:
• Food - Food items are discs with a radius of 2.4cm, when a food item is collected it is replaced by
another food item in a random location.
• Poison - Poison is identical to food, except that it induces a negative dopamine response (see section
3.7).
• Food Container - A food container is a disc with radius of 14cm which contains a single food item
randomly located within it. Food items cannot be sensed from outside a food container, other containers
cannot be sensed from within a container.
• Empty Container - This is identical to a food container except that is does not contain any food items,
as well as being sensed differently by the robot (see section 3.2).
The environment has no walls, and has the characteristic of a torus. That is, if the robot moves over the
right edge of the environment it will reappear at the left of the environment and the same for the top and
bottom ’edges’ of the environment. In this way the robot does not have to deal with collision avoidance.
There are two main variations in the environment. In the simplest case (figure 3.1(a)) the environment
consists of randomly located food objects. Whenever the robot reaches an item of food the piece of food
is moved to another random location within the environment. This environment set-up will be referred
to as the food-only environment. In the second type of environment food is contained within containers
(figure 3.1(b)). The food cannot be sensed until the robot is within a container, other containers cannot
be sensed when the robot is within a container. This forces the robot to have to learn to turn towards
food containers, something that does not directly give it a reward. When food is collected the food and
container are moved to a random location in the environment. This environment will be referred to as
the food-container environment. An extension to this environment includes empty containers which do not
contain food items.
18
Food
100
90
80
70
60
50
40
30
20
10
10
20
30
40
50
60
70
80
90
100
Environment Wraps Around
(a) Food-Only Environment
300
250
200
Food
Container
150
100
Food
50
50
100
150
200
250
300
(b) Food-Container Environment
Figure 3.1: (a) shows a 100cmx100cm environment containing randomly placed food. (b) shows a
300cmx300cm environment with food items that are positioned within containers. Food cannot be sensed
from outside the container, other containers cannot be sensed when inside a container.
19
3.2
Robot and Sensors
The robot consists of a circular body with two wheels positioned 1cm apart on either side of the robot.
To ensure that the robot is always moving, the wheel motor velocities are restricted to the range 25cm/s 31.2cm/s. This gives the robot a minimum turning circle of radius 4.5cm. The robot is equipped with two
types of sensor, range-sensors and touch-sensors. The range-sensors come in pairs, one left and one right,
each covers a range of π/2 from directly in front of the robot to the left and right as in figure 3.2.
Sensor Range
Sensors
1cm
Wheels
Figure 3.2: Structure of the robot.
The touch-sensors elicit an instantaneous response that lasts for 1ms whenever the robot comes into
contact with the object for which they are sensitive. The response of the range-sensors is linear with the
distance to the object. So an object directly in front of the robot will elicit a sensor response of one, whereas
an object at the edge of the sensor’s range will produce a response of zero in the sensor. If there is more
than one object in a sensor’s range then only the closest object is sensed. This prevents the robot being
distracted by objects further away before responding to closer ones.
A winner-takes-all mechanism was implemented between the two range-sensors. Such that if both sensors
detect an object at any given time, then only the sensor with the strongest value elicits a response. This
winner-takes-all mechanism is useful in situations where two food items are located at equal distances to the
left and right of the robot, as illustrated in figure 3.3. Without a winner-takes-all mechanism both sensors
will sense food items and as such the robot will increase its left motor velocity to turn right, it will also
increase its right motor velocity to turn left. The effect of which will be that the robot drives straight between
the two food items. It has been shown that a winner-takes-all mechanism is biologically plausible [48]. For
simplicity this winner-takes-all mechanism was implemented in the sensors, an alternative implementation
would be to implement it directly in the wiring of the robot’s neural network.
Table 3.1 gives an overview of all the sensors that the robot can have. In the food-only environment the
robot has two food range-sensors and a food touch-sensor. In some environments the food items are replaced
with poison, to demonstrate the ability of the robot to modify its behaviour. In this case the same food
range and touch sensors are used for sensing poison.
In the environment with food containers the robot has an additional container range-sensor, this sensor
reacts to both food-containers and empty-containers. It also has a food-container touch-sensor, this sensor
only elicits a response when a food-container is entered (no response is elicited whilst the robot is inside the
container). Equivalently it has an empty-container sensor, which only reacts to entering containers which
don’t contain food items. This set-up is useful in demonstrating that the robot can differentiate between
the reward predicting stimulus of the food-container touch-sensor, and the non-reward predicting stimulus
of the empty-container touch-sensor.
20
Direction
of Motion
Food
Figure 3.3: Without a winner-takes-all mechanism both sensors will respond equally and the robot will
increase both left and right motor speeds, ultimately driving straight between the food items.
Object Sensed
Type
Range
Food/Poison
Containers
Food
Food-Containers
Empty-Containers
Range
Range
Touch
Touch
Touch
30cm
60cm
0cm
0cm
0cm
Available in FoodOnly Environment
Yes
No
Yes
No
No
Available
in
FoodContainer Environment
Yes
Yes
Yes
Yes
Yes
Table 3.1: The different types of sensor that the robot can have, and in which environment the robot is
equipped with these sensors.
21
3.3
Neural Model
The Izhikevich model was used to simulate the individual neurons. An overview of this model is given in
section 2.1.3. The dynamics of the neuron are modelled by two differential equations:
dv
= 0.04v 2 + 5v + 140 − u + I
dt
(3.1)
Where I is the inbound current from pre-synaptic spikes (weighted by the strength of the synapse), v is the
membrane potential and u is the recovery variable that determines the refractory period of the neuron and
is modelled as:
du
= a(bv − u)
(3.2)
dt
The parameters of the model were set for excitatory neurons as:
a = 0.2, b = 0.02, c = −65 + 15r2 , d = 8 − 6r2
(3.3)
Where r is a random variable in the range [0,1]. This produces neurons with a regular spiking pattern (see
figure 2.3). For inhibitory neurons the parameters were set to:
a = 0.1, b = 0.2, c = −65 + 15r2 , d = 8 − 6r2
(3.4)
Which results in neurons which exhibit fast spiking behaviour. These neural spiking patterns are consistent
with observed behaviour in the brain, and previous neural models [8].
3.3.1
Numerical Approximation
The differential equations given by the Izhikevich model were simulated using the Euler method of numerical
approximation. Given a differential equation such that y 0 (t) = f (t, y(t)), if we know the value of y(t) for
some time t then, by the Euler method, we can approximate the value of y(t + δt) for some small time step
δt as:
y(t + δt) = y(t) + δt.f (t, y(t))
(3.5)
This formula can be applied repeatedly to approximate y(t) over time. In our simulations we used δt =0.5ms
for approximating v(t) and δt =1ms for approximating u(t). This is consistent with Izhikevich’s original
paper [8].
3.4
Phasic Activity
The properties of individual neurons mean that if a group of neurons is interconnected and then stimulated
with a low baseline current then the group of neurons will tend to fire in phase; synchronous firing followed
by synchronous inactivity. The frequency with which the neurons oscillate depends on various parameters
of the neurons and network, phasic activity has been observed in the brain over a wide range of frequencies
ranging from 0.02Hz to >100Hz [49]. In our model the inputs and outputs to the network are updated every
70ms (corresponding to an update frequency of 14Hz), in the following sections we will describe the exact
method by which the sensors and motors are encoded/decoded. For simplicity this 14Hz update frequency
was explicitly implemented outside of the neural network.
One possible way in which this oscillatory property could be embedded in the network directly would be
to have a separate group of interconnected inhibitory neurons with a baseline level of firing such that the
firing rate of the group oscillates at 14Hz. If this inhibitory population is connected to the relevant sensor
and motor neurons then it would force them to fire with 14Hz frequency as well, and as such simulate the
update model that was used in this paper.
The dynamics of STDP mean that, without phasic activity, two populations of connected neurons will
have their synaptic weights depressed over time. There is just as much probability of a pre-synaptic neuron
firing before or after a post-synaptic neuron. As the parameters of STDP were set to favour long term
depression, the synapses would be depressed. Through the use of phasic activity the network exhibits clear
sense-respond behaviour. For example, if the left sensor is stimulated and, for the rest of the active phase
of the neurons, the right motor neurons are active, then we can say the right motor action was a response
to the left sensor stimulus. The eligibility trace for left sensor to right motor synapses would be high at
this point, as the left sensor neurons fired before the motor neurons. As such, if a reward was subsequently
received then the robot would learn attraction behaviour.
22
3.5
Sensorimotor Encoding
3.5.1
Sensor to Neuron Encoding
Neuron Index
Both the food and container range-sensors return results in the range [0,1]. Every 70ms the corresponding sensor neurons are stimulated for 1ms with a current taken from a Poisson distribution with mean of
sensor value ∗ 30mA. The effect this stimulus current has on the firing rate of a population of sensor neurons
is shown in figure 3.4.
Neuron Firings Over Time
20
10
500
1000
1500
2000
Time (ms)
2500
3000
3500
4000
Figure 3.4: The firings in a group of 20 sensor neurons as the sensor value is linearly increased from zero to
one. The neurons are stimulated every 70ms.
The food touch-sensor and the container touch-sensor elicit an instantaneous response in the corresponding neurons. This is achieved by stimulating them for 1ms with a current taken from a Poisson distribution
with a mean of 12mA.
3.5.2
Motor Velocity Calculation
The simulation is updated every 70ms, the total firings during this 70ms for the left and right motor neurons
(fleft and fright respectively) are first calculated. The motor velocities are then calculated, for example for
the left motor, using the following formula:
if fleft > fright
vmax
vmin
if fleft < fright
(3.6)
vleft =
vmax +vmin
otherwise.
2
Where vleft is the output velocity of the left motor and vmax and vmin are the maximum and minimum
allowable velocities of the motors, which are set to 31.2cm/s and 25cm/s respectively. The equation for the
right motor is equivalent, with left and right terms switched. Note that this formula implements a winnertakes-all mechanism between the motors. If one group of motor neurons has a greater firing rate then the
corresponding motor will be set to maximum velocity, and the other motor set to its minimum velocity. This
helps with robot exploration as it reduces the probability of the robot driving in a straight line. Also this
winner-takes-all mechanism helps with learning attraction/avoidance behaviours as it amplifies the result of
any slight difference in synaptic weights between the sensors and motors.
The robot was restricted to run at half-speed whilst in food-containers or empty-containers, such that
its maximum and minimum velocities were set to 15.6cm/s and 12.5cm/s. This was to increase the time
between the reward predicting stimulus and the reward, as discussed in section 3.8.
3.6
Network Architecture
The robot was controlled via a spiking neural network consisting of connected groups of neurons. Figure 3.5
shows the connections in the neural network for the robot in the food-only environment. This is a subset
of the connections for the robot in the food-container environment which is shown in figure 3.6. All rangesensor neurons are connected to motor neurons with 85% probability, that is, there is an 85% chance that
any individual range-sensor neuron is connected to any individual motor neuron. All touch-sensor neurons
are connected to dopaminergic neurons with 10% probability. The synaptic weights between food touchsensor neurons and dopaminergic neurons is fixed and set to 3mV. This high synaptic weight forces a strong
response in the dopaminergic neurons whenever food is eaten. Unless otherwise specified all other synapses
are plastic and have their initial weights set to zero.
In addition to the neurons specified in these architectural diagrams there is also a group of inhibitory
neurons, this group is connected to all other neurons with 10% probability and its synaptic connections are
initialized to random values in the range [0,-3] mV. This has the effect of slightly dampening the activity of
the network.
All of the neuron groups, with the exception of the dopaminergic neurons, consist of 20 neurons. The
dopaminergic group consists of 40 neurons. The reason that the dopaminergic neuron group is larger is that
23
Left Food
Sensor
Right Food
Sensor
Food Touch
Sensor
External Inputs
Connection subject
to plasticity.
Connection with
fixed weights.
Left
Food Sensor
Neurons
Right
Food Sensor
Neurons
Food Touch
Sensor
Neurons
Spiking Neural
Network
Left Motor
Neurons
Right Motor
Neurons
Left Motor
Right Motor
Dopamine
Neurons
External Outputs
Figure 3.5: The architecture of the robot’s neural network for the food-only environment.
Connection
subject to
plasticity.
Range Sensors
Touch Sensors
Connection
with fixed
weights.
Left Food
Sensor
Right Food
Sensor
Left
Container
Sensor
Right
Container
Sensor
Food Touch
Sensor
Food-Container
Touch Sensor
Empty-Container
Touch Sensor
Left
Food Sensor
Neurons
Right
Food Sensor
Neurons
Left
Container
Neurons
Right
Container
Neurons
Food Touch
Sensor
Neurons
Food-Container
Touch Sensor
Neurons
Empty-Container
Touch Sensor
Neurons
Left Motor
Neurons
Right Motor
Neurons
Left Motor
Right Motor
Dopamine
Neurons
External
Inputs
Spiking
Neural
Network
External
Outputs
Figure 3.6: The architecture of the robot’s neural network for the food-container environment.
24
learning to elicit a dopamine response when a reward predicting stimulus (food-container touch-sensor) is
received relies heavily on random firings in the dopaminergic neuron group. A larger number of neurons was
found to increase the learning speed. The conductance delay for all synapses was set to 1ms, meaning that it
would take 1ms for a pre-synaptic spike to propagate into a change in membrane potential of a post-synaptic
neuron.
The spiking neural network for the food-only robot consists of 160 neurons, the distribution of these
neurons across the different neural groups is shown in table 3.2. The network for the food-container robot
consists of 240 neurons in total, the distribution of the neurons for this robot is shown in figure 3.3.
Neuron Range
1-20
21-40
41-60
61-80
81-100
101-140
141-160
Neuron Group
Left Food Sensor
Right Food Sensor
Left Motor
Right Motor
Food Touch-Sensor
Dopamine Neurons
Inhibitory Neurons
Table 3.2: The organization of the neurons for the robot in the food-only environment.
Neuron Range
1-20
21-40
41-60
61-80
81-100
101-120
121-140
141-160
161-180
181-220
221-240
Neuron Group
Left Food Sensor
Right Food Sensor
Left Container Sensor
Right Container Sensor
Left Motor
Right Motor
Food Touch-Sensor
Food-Container Touch-Sensor
Empty-Container Touch-Sensor
Dopamine Neurons
Inhibitory Neurons
Table 3.3: The organization of the neurons for the robot in the food-container environment.
3.7
Plasticity
Synaptic plasticity was implemented using dopamine modulated STDP, an outline of which is given in section
2.4.3. The value of the function STDP (t) at time t is calculated as:
(
+
A+ e−∆t/τ
if ∆t ≥ 0
STDP (t) =
(3.7)
−
−A− e∆t/τ
if ∆t < 0
Where ∆t is the time difference between the pre and post synaptic neurons firing. The nearest-neighbour
update scheme was used, such that only the most recent pre/post synaptic spike was considered when
calculating STDP (t) for a specific synapse. The optimal parameters for the model were found by running
the robot in the food-only environment for 1000 seconds, 15 times. This was repeated for a large range of
parameters and the value that achieved the best score (as measures by number of food items collected) was
used. These parameters were:
• A+ = 0.1
• A− = 0.15
• τ + = 0.02s
• τ − = 0.11s
The strength of STDP is plotted against the difference in pre and post synaptic firing, ∆t, in figure
3.7. There is a stronger response when the post-synaptic neuron fires before the pre-synaptic neuron (∆t
25
is negative) and as such long term depression is preferred over long term potentiation. This means that
uncorrelated firing between two populations of neurons will tend to decay the synaptic weights to zero and
only correlated firings will have the effect of increasing the associated synaptic weights over time. This is
useful in helping the robot learn the correct behaviour and ignore any noise in the neuron firings.
STDP Strength vs Spike Timing
0.1
STDP (mV)
0.05
0
−0.05
−0.1
−0.15
−500
−400
−300
−200
−100
0
100
Δ t (ms)
200
300
400
500
Figure 3.7: The Strength of STDP is shown against the difference in spike timings, where ∆t = tpost − tpre .
Each synapse maintains an eligibility trace, c. The value of which is updated according to:
˙ = −c(t)/τc + STDP (t)δ(t − t
c(t)
pre/post )
(3.8)
Where τc is the decay rate of the eligibility trace and is set to 0.476s, δ is the Dirac delta function and ensures
that the STDP only effects the eligibility trace when either the pre or post synaptic neuron fires. Figure
3.8 demonstrates the eligibility trace of a single synapse after simultaneous pre-synaptic then post-synaptic
firing. The eligibility trace becomes negligible after a maximum of about one second after coincident firing.
This means that any action the robot is currently performing will be forgotten after this time. See section
3.8 for details of how this memory is extended further by moving the dopamine response.
Eligibility Trace Over Time
0.09
0.08
0.07
Eligibility Trace
0.06
0.05
0.04
0.03
0.02
0.01
0
−0.01
0
500
1000
1500
Time (ms)
2000
2500
3000
3500
Figure 3.8: The eligibility trace for a synapse. The pre-synaptic neuron fires after 30ms, the post-synaptic
neuron fires after 32ms. No more firing occurs and the eligibility trace decays.
Every millisecond the synaptic weights are then updated according to:
ṡ = cd
26
(3.9)
Where s is the synapse strength and d is the current level of dopamine. The level of dopamine is increased
by 0.0035µM for every dopaminergic neuron that fires. This has the effect of increasing the level of dopamine
by approximately 0.12µM every time a food item is reached and the majority of the dopaminergic neurons
fire. During experiments which contain poison, the poison has the effect of reducing the level of dopamine
by 0.0035µM for every dopaminergic neuron that fires. The consequence of this is that synapses that were
highly active in the lead up to the poison, and as such have a high eligibility trace, will have their synaptic
weight reduced. A delay of 5ms between a dopaminergic neuron firing and the corresponding increase in
dopamine was used. This gives enough time for the eligibility trace of synapses projecting to dopaminergic
neurons to be updated, before the corresponding spike in dopamine.
The baseline level of dopamine in the system was set to -0.0004µM, meaning that in the absence of any
firing by the dopaminergic neurons the level of dopamine would decay to this negative value. This has the
effect of working like hunger, if the robot does not collect any food for a long time then strong synaptic
weights will become weaker and very weak synaptic weights will become stronger. This allows the robot to
change its behaviour when its current strategy isn’t being effective in collecting food.
A constant current of 3.65mV was supplied to the dopaminergic neurons, this induced a background
firing rate of 0.4Hz. For values higher than this the robot was not able to learn correctly, this is talked about
in more depth in section 3.8.
As described in section 2.4.3 dopaminergic neurons have two patterns of activity, background firing and
stimulus induced burst firing. The result of which is that that background firing does not have any significant
effect on the level of dopamine. To simulate this the level of dopamine was only increased if more than five
dopaminergic neurons fired in a 1ms window. The effect this has is discussed more in section 3.8.
3.8
Moving Dopamine Response from Unconditioned-Stimulus to
Conditioned Stimulus
In many scenarios an agent will need to perform a sequence of behaviours in order to get a reward. If
dopamine is only released when the reward is achieved then plasticity will only occur at this time. The
efficacy of DA-modulated STDP drops off exponentially and is negligible after about 1000ms, therefore
if the reward predicting behaviour takes place more than 1000ms before the reward is received then this
behaviour will not be learnt. For example, in the food-container environment, with the robot run at halfspeed, the average time between entering a food-container and the food being collected is about 800ms.
Therefore food-container attraction behaviour would take a long time to learn, if at all, if dopamine was
only released when the food was collected. If the robot needed to learn a third behaviour, such as picking
up a key to open the container, then it is likely that the temporal difference between this behaviour and the
actual reward would be too great for it to ever be learnt.
A simple method to deal with this problem is for the agent to learn to elicit a dopamine response when it
receives a reward predicting stimulus. For example, the food-container touch-sensor is often stimulated prior
to a food item being collected and hence a spike in dopamine being received. If the robot learnt to produce
a dopamine spike when the food-container touch-sensor is stimulated then the robot should be able to learn
food-container attraction behaviour much faster, as there wouldn’t be a long delay between performing the
behaviour and receiving a dopamine response. It can be seen how this process could be repeated such that
a long chain of behaviours could be learnt.
This moving of dopamine response is achieved by connecting the container touch-sensor neurons to the
dopaminergic neurons and allowing these synapses to be plastic. Figure 3.9 shows an example of how the
robot is able to strengthen food-container touch-sensor to dopaminergic neuron connections. The firings
of just two neurons are plotted, one from the food-container touch-sensor group (index 141) and one from
the dopaminergic group (index 181). At 1000ms the robot enters a food container and the food-container
touch-sensor neuron fires. Purely by chance the dopaminergic neuron fires 20ms later due to background
firing. This causes the eligibility trace between these two neurons to jump up. 800ms later the robot reaches
the food item and all the dopaminergic neurons fire, causing a spike in the level of dopamine. This spike
in dopamine, combined with the still positive eligibility trace, increase the weight of the synapse. Since
the synaptic weight is higher this increases the probability that the dopaminergic neuron will fire after the
food-container touch-sensor neuron in future, and have its weight further increased.
The ability of the robot to learn this dopamine response relies on a combination of the background firing
rate of the dopaminergic neurons and on the strength of long term potentiation as compared with long term
depression. If the background firing is too low then the chance of a dopaminergic neuron firing soon after
a food-container touch-sensor neuron will be small and as such the robot will take a long time to learn,
if at all. If the background firing rate is too high, and the level of long term depression is also high then
random uncorrelated firings between the food-container sensor neurons and the dopaminergic neurons will,
27
Eligibility Trace (mV)
Shift of Dopamine Example
Eligibility Trace
0.08
0.06
0.04
0.02
0
0
500
1000
1500
Time (ms)
Firings
2000
2500
3000
0
500
1000
1500
Time (seconds)
Dopamine
2000
2500
3000
0
500
1000
1500
Time (seconds)
Synaptic Strength
2000
2500
3000
0
500
1000
1500
Time (seconds)
2000
2500
3000
Neuron Index
200
150
100
50
0
Dopamine (µM)
0.15
0.1
0.05
0
Synaptic Weight (mV)
−0.05
0.08
0.06
0.04
0.02
0
Figure 3.9: The activity of two neurons is plotted, one (index 141) from the food-container sensor group,
and one (index 181) from the dopaminergic neuron group. At 1000ms the robot enters the food container
and the food-container touch-sensor neuron spikes, by chance the dopaminergic neuron fires soon after which
raises the eligibility trace. At 1800ms the robot collects the food item and the level of dopamine spikes.
This dopamine spike, combined with the positive eligibility trace, increase the weight of the synapse from
the food-container touch-sensor neuron and the dopaminergic neuron.
28
on average, cause the synapses’ eligibility traces to become negative and cancel out any potentiation of the
synapses.
If we consider what would happen if we did not threshold the number of neurons required for the level
of dopamine to be raised (or equivalently if dopaminergic neurons in the brain produced bursting activity
during background firing) then we can see that non-reward predicting stimulus can also have their synaptic
weights, for synapses projecting to dopaminergic neurons, potentiated. This is due to the feedback loop
that can be created. For example, if by chance an empty-container touch-sensor neuron fires before a
dopaminergic neuron. The eligibility trace will increase, and the robot will immediately receive an increase
in dopamine (due to the dopaminergic neuron firing), this will increase the probability that the next time
the empty-container sensor neuron is triggered, the dopaminergic neuron fires afterwards, resulting in its
synaptic weight increasing even further.
If the parameters of the model were altered such that plasticity happened at a much slower rate then
this feedback loop could be avoided, the small increase in synaptic weight in the previous example would not
alter the probability of the neurons firing in a correlated manner much. Therefore it would be likely that
the dopaminergic neuron would at some point fire before the sensor neuron and cause the synaptic weight
to decrease again. However the advantage of the above method of thresholding the dopaminergic firing rate
needed to increase the level of dopamine is that the robot is more robust in its learning, as well as being
able to learn much faster.
3.9
Network Stability
STDP is an inherently unstable mechanism, once a synaptic weight has been increased then the probability of
the pre-synaptic neuron firing before the post-synaptic neuron becomes increased and as such the probability
of the synaptic weight potentiating further is increased. This feedback mechanism can lead to synapses
quickly becoming potentiated to their maximum value. To combat this two mechanisms were used. The
synaptic weights of excitatory synapses was restricted to the range [0,4] mV.
In addition, a local dampening mechanism was used based on the observation that, in the brain, a highly
active group of neurons have their excitability decreased over time, which was discussed in section 2.4.5.
The synapses were divided into three distinct groups, the synapses belonging to these three groups being:
• Food Taxis Group - Food range-sensor neuron to motor neuron synapses.
• Container Taxis Group - Container range-sensor neuron to motor neuron synapses.
• Container Dopamine Group - Container touch-sensor neuron to dopaminergic neuron synapses.
For each of these groups, if the mean synaptic weight becomes potentiated by more than 2mV (half the
maximum value) then all synapses within the group have their weights reduced by 0.1mV. This prevents a
feedback loop such that all the synaptic weights become potentiated to their maximum value. The other
benefit of this approach is that it can lead to competition between behaviours within the synaptic group,
figure 3.10 demonstrates how this competition can occur between the food attraction and food avoidance
behaviours.
29
Synaptic Competition Example
Synaptic Weight (mV)
3.5
3
Food attraction
synapses
Food avoidance
synapses
2.5
2
1.5
2
3
4
5
6
Time (seconds)
7
8
9
Figure 3.10: An example of how synaptic weight dampening can lead to competition between behaviours.
Both the synaptic weights for food attraction and for food avoidance behaviour are being potentiated, but the
food attraction synaptic weights are potentiating faster. At 6 seconds the average synaptic weight over both
groups reaches the maximal allowable value, and all weights are reduced by 0.1mV. The food attraction
synaptic weights then increase at a faster rate, and as such the next time the maximal average synaptic
weight is reached the food attraction behaviour has increased by more than 0.1mV and the food avoidance
behaviour has increased by less than 0.1. In this way the strength of the food avoidance behaviour becomes
reduced over time, whilst the strength of the food attraction behaviour is increased.
3.10
Taxis
The robot’s architecture is constructed such that it has the potential for taxis behaviour, for example food
attraction or poison avoidance. This is achieved through having two sensors at the front (left and right) of
the robot, which are wired to two motors. The wiring of the robot does not predispose it to prefer attraction
or avoidance behaviour, there is an equal number of connections from each sensor to each motor. The
acquisition of attraction/avoidance behaviour is left up to DA-modulated STDP to strengthen the synaptic
connections corresponding to the behaviour to be learnt.
Figure 3.11 shows how strong connections between opposite sensors and motors can cause food attraction
taxis behaviour.
30
(a) Left Sensor Active, Right Motor Active
(b) Right Sensor Active, Left Motor Active
(c) Left Sensor Active, Right Motor Active
Figure 3.11: Demonstration of how wiring the left sensor to right motor, and right sensor to left motor can
cause food attraction taxis. In (a) the left sensor is active, so the robot turns forward and left, in (b) the
right sensor is active so the robot turns forward and right. Finally in (c) the robot turns and collects the
food item.
31
3.11
Exploration
To ensure that the robot explores the environment the robot needs to have some kind of exploratory behaviour. This was implemented by, every 70ms, randomly choosing either the right or left motor. Then for
the remaining 70ms this motor’s neurons were stimulated with a current taken from a Poisson distribution
with a mean of 2.35mA.
One possible way in which this exploration could be implemented directly in the network, would be to
have the motor neurons connected such that they are oscillating at 14Hz (see section 3.4 for details of how
oscillating behaviour can occur). If a winner-takes-all mechanism was then implemented between the left
and right motors, this would have the effect that one population of neurons would be active whilst the
motor neurons are phasically active. When the motor neurons become phasically inactive then it would be
easier for the winner-takes-all mechanism to switch the active group. This network would display the same
properties as the implemented exploration motor stimulation.
Once the synaptic weights between the sensors and the motors have increased enough then the current to
the motors from the sensors will override the exploration current. The effect of this is that before the robot
learns anything it will perform a random walk. After learning the robot will only perform a random walk
whenever it can’t sense anything. Also, if the environment changes, causing the synaptic weights between
sensors and motors to drop, then the robot will go back to performing a random walk. The path of the robot
in an environment with no food, after having run for 30 seconds is shown in figure 3.12. As can be seen the
robot is able to effectively explore the environment.
100
Path of the robot performing a random walk.
90
80
y position (cm)
70
60
50
40
30
20
10
20
40
60
x position (cm)
80
100
Figure 3.12: The path taken by the robot for 30 seconds in an environment with no food. The random motor
stimulation causes the environment to be explored effectively.
32
Chapter 4
Results and Discussion
4.1
Orbiting Behaviour
An interesting property of the robot and environment is that there is no such thing as a fixed set of optimal
synaptic weights. In almost all scenarios in the environment the best set of synaptic weights would be to
have all left sensor to right motor and right sensor to left motor synapses set to their maximum weight and
all other sensor to motor synapses set to zero. This will cause the robot to turn the quickest towards food.
However, because the robot has a non-zero turning circle, in some situations the robot can orbit a food
item indefinitely. Figure 4.1 shows the path of a robot using these synaptic weights when it approaches a
food item at 90 degrees. At each time step the robot senses food with its left sensor and so turns to the
left, however it never reaches the food. The optimal robot needs to determine when it is orbiting food, and
modify its behaviour accordingly.
100
Path of Robot When Orbiting Food
90
80
y position(cm)
70
60
50
Food
40
Robot Path
Orbits Food
30
20
10
Start Position
20
40
60
x position(cm)
80
100
Figure 4.1: The path of the robot is shown in blue, with the robot’s position being plotted every 70ms. Once
the robot can sense the food with its left sensor then it turns left, it continues to sense the food with its
left sensor and so continues turning left. However it cannot turn fast enough to reach the food and ends up
orbiting it.
4.1.1
Experimental Set-up
To demonstrate the ability of the robot to learn to modify its behaviour when it gets stuck in one of these
orbits the robot was placed in an environment with one food item. The synaptic weights between opposite
sensors/motors were set to their maximum allowable value, and all other sensor to motor synapses were set
to zero. This forces the robot to initially perform food attraction behaviour. The robot was positioned in
such a way that it would start out orbiting the food item as shown in figure 4.2.
In the first scenario none of the robot’s synapses are plastic, it does not learn. The robot was simulated
for 40 seconds, this was repeated 20 times. In the second scenario the robot’s synapses were modified using
33
Starting Setup for Orbiting Environment
100
90
80
y position(cm)
70
60
50
Food
40
30
20
Robot Start Position
and orientation
10
20
40
60
x position(cm)
80
100
Figure 4.2: The starting position and set-up of the environment to force the robot to start orbiting.
DA-modulated STDP, as such the robot was allowed to learn. The robot was again simulated for 40 seconds,
for 20 trials.
4.1.2
Results and Discussion
In none of the trials with learning disabled did the robot manage to break out of its orbiting behaviour.
With learning enabled the robot managed to learn to stop orbiting the food, and was able to collect the food
item, in all 20 trials. On average it took the robot 5.1 seconds to stop orbiting the food. The distance to
the food for the first 10 seconds of one such trial is plotted in figure 4.3. This figure shows the robot was
able to modify its behaviour to stop orbiting after less than 6 seconds, the food item is then collected at 8
seconds. Once the food is collected it is replaced at a random location, and the robot collects the food one
more time.
Distance From Robot to Food Over Time
50
45
40
Distance(cm)
35
30
25
20
Robot Stops
Orbiting Food
15
10
Food Items
Collected
5
0
0
1
2
3
4
5
Time (seconds)
6
7
8
9
10
Figure 4.3: The distance between the robot and the food item is plotted over time, the synaptic weights are
initially set to force food attraction behaviour and plasticity is turned on. The robot starts out orbiting the
food, but after 5.8 seconds is able to unlearn the orbiting behaviour and stops orbiting the food.
To see how this relearning is achieved the mean synaptic weights between both sensors and both motors
over time are plotted in figure 4.4. There are several interesting properties that are shown. Firstly the
synaptic weights between the left sensor and the right motor are slowly decaying whilst the robot is orbiting
34
the food. This is because the baseline level of dopamine in the system is a small negative value, therefore
highly active synapses with strong weights will have their synaptic weight reduced due to their high eligibility
trace. This has the effect of reducing the strength of any behaviour that the robot is currently performing,
in this case turning left when the left food sensor is activated. The negative dopamine can be thought of as
analogous to hunger, if whatever the robot is currently doing isn’t giving it food then it has an incentive to
change its behaviour.
Notice that the right sensor to left motor synaptic weights remain strong. This is due to the fact that
the right sensor neurons are not being stimulated and as such the eligibility trace remains small. Whilst the
robot is orbiting the synaptic weights between the left sensor and the left motor are also increased. This
is because the left motor neurons are still being randomly stimulated by the exploration method described
in section 3.11. This causes some random firings in the left motor population that are uncorrelated with
the left sensor neuron firings. The parameters of STDP used favour long term depression over long term
potentiation and as such the eligibility trace for these synapses will be negative. Given a negative eligibility
trace, a negative dopamine value will increase the synaptic weights.
Mean Sensor to Motor Synaptic Weights for Orbiting Environment
Mean Synaptic Weight
Mean Synaptic Weight
Mean Synaptic Weight
Mean Synaptic Weight
Left Sensor −> Left Motor
3
2
1
0
0
1
2
3
4
5
6
Time(seconds)
Left Sensor −> Right Motor
7
8
9
10
3
2
1
0
Food
Collected
Orbiting Stopped
0
1
2
3
0
1
2
3
0
1
2
3
4
5
6
Time(seconds)
Right Sensor −> Left Motor
7
8
9
10
4
7
8
9
10
7
8
9
10
3
2
1
0
5
6
Time(seconds)
Right Sensor −> Right Motor
3
2
1
0
4
5
Time(seconds)
6
Figure 4.4: The mean synaptic weights between the left and right sensor/motor groups is shown for a robot
in an environment with one food item. The robot is initially orbiting the food, at 14 seconds this orbiting
is unlearned. At 35,37 and 39 seconds the robot collects a food item.
Figure 4.5 shows the neural firing rate of both the sensor neuron groups and the motor neuron groups
for the same trial. This shows that as the synaptic weights between the left sensor and right motor are
decreased, the firing rate of the right motor also decreases. Equivalently, as the synaptic weights between
the left sensor and the left motor are increased, so is the firing rate of the left motor neurons. After 5.8
seconds the combination of increased synaptic weights to the left motor and the random exploration stimulus
of the motors means that the firing rate in the left motor is greater than the firing rate in the right motor
35
and the robot turns away from the food item. The robot then performs a random walk until it can again
sense the food. Once the food is collected notice that the left sensor to right motor synaptic weights are
increased (figure 4.4), and the left sensor to left motor synaptic weights are decreased as the robot starts to
relearn food attraction behaviour.
Neuron Firing Rate in Orbiting Environment
−4
x 10
Mean Firing Rate (Hz)
Left Sensor
Right Sensor
Left Motor
Right Motor
2
1
0
0
1
2
3
4
5
Time (seconds)
6
7
8
9
10
Figure 4.5: The mean firing rate of neurons in the sensor and motor groups over a single trial in the orbiting
environment. After 5.8 seconds the left motor firing has increased enough, and the right motor firing has
decreased enough, that the robot stops orbiting the food.
To confirm that the ability to relearn, when the current behaviour isn’t working, relies on the baseline
level of dopamine being negative, the experiment was re-run with the baseline level of dopamine set to
0.0004µM. In none of these 20 trials did the robot manage to escape from orbiting the food. Figure 4.6
shows the synaptic weights for one of these trials. As this figure shows, a positive level of background
dopamine has the effect of reinforcing any behaviour that the robot is currently performing, as shown by the
increase in weights between the left sensor and the right motor.
36
Synaptic Weights in Orbiting Environment (positive baseline dopamine)
Mean Synaptic Weight
Mean Synaptic Weight
Mean Synaptic Weight
Mean Synaptic Weight
Left Sensor −> Left Motor
3
2
1
0
0
2
4
6
0
2
4
6
0
2
4
6
0
2
4
6
8
10
12
Time(seconds)
Left Sensor −> Right Motor
14
16
18
20
8
14
16
18
20
8
14
16
18
20
14
16
18
20
3
2
1
0
10
12
Time(seconds)
Right Sensor −> Left Motor
3
2
1
0
10
12
Time(seconds)
Right Sensor −> Right Motor
3
2
1
0
8
10
12
Time(seconds)
Figure 4.6: The synaptic weights between sensors and motors when the baseline level of dopamine is positive,
and set to 0.0004µM in the orbiting environment. This positive dopamine acts to reinforce the current
behaviour, and the left sensor to right motor synaptic weights increase.
37
4.2
Food Attraction Learning
In this section we test the ability of the robot to learn food-attraction behaviour by modifying it’s synaptic
weights.
4.2.1
Experimental Set-up
A 100cmx100cm environment consisting of 20 randomly placed food items was constructed. When a food item
is collected it is moved to a random position in the environment. The robot was placed in the environment
and simulated for 1000 seconds with plasticity taking place between the sensor neurons and the motor
neurons. This process was repeated for 50 trials, each time resetting the sensor to motor synaptic weights
to zero.
For comparison a robot with no connections between sensor neurons and motor neurons and no plasticity
was run for 50 trials of 1000 seconds each in the environment. The effect of this is that the robot performs
a random walk, and only collects food by randomly driving into it.
4.2.2
Results and Discussion
The mean food collection rate over time for the robot with learning enabled and the random walk robot is
shown in figure 4.7. The robot with learning enabled was able to rapidly learn to turn towards and collect
food items, and this high collection rate remained stable for the rest of the trial. On average the robot took
about 200 seconds for the full food attraction behaviour to be learnt and stabilized.
Food Collection Rate in Food−Only Environment
Collection Rate (collections per second)
1.8
1.6
1.4
1.2
Robot With Learning Disabled
Robot With Learning Enabled
1
0.8
0.6
0.4
0.2
0
0
100
200
300
400
500
600
Time (seconds)
700
800
900
1000
Figure 4.7: The food collection rate over time for a robot in a 100cmx100cm environment with 20 randomly
placed food items.
Table 4.1 shows the mean food collected over the 50 trials and whether the robot was able to correctly
learn food attraction behaviour. The robot was deemed to have learnt attraction behaviour if at the end of
the trial the synapses for food attraction (left sensor to right motor, and right sensor to left motor) were
on average more than 0.5mV, as well as being on average more than 10% higher than the synapses for food
avoidance. These values being high enough that exploration behaviour was overridden, and different enough
that food attraction was preferred. The robot was able to learn this behaviour in all of the trials with
learning enabled.
Robot Type
Learning Disabled
Learning Enabled
Mean Food Collected
269
1418
Std. Deviation Food Collected
16.1
77.1
% Correct
0
100
Table 4.1: This table shows the total amount of food collected for the robots, averaged over the 50 trials as
well as what percentage of robots that were able to correctly learn food attraction behaviour.
Figure 4.8 shows the mean weights between the sensors and motors, averaged over the 50 trials. The
synaptic weights between opposite sensors/motors are rapidly increased and reach the maximum weight after
38
about 100 seconds on average. This causes the robot to perform food attraction behaviour, as reflected by
the increase in food collection in figure 4.7.
Robot Sensor to Motor Connection in Food-Only Enviroment
Mean Synaptic Weight
Mean Synaptic Weight
Mean Synaptic Weight
Mean Synaptic Weight
Left Sensor −> Left Motor
2
1
0
0
100
200
300
0
100
200
300
0
100
200
300
0
100
200
300
400
500
600
Time(seconds)
Left Sensor −> Right Motor
700
800
900
1000
400
700
800
900
1000
400
700
800
900
1000
700
800
900
1000
2
1
0
500
600
Time(seconds)
Right Sensor −> Left Motor
2
1
0
500
600
Time(seconds)
Right Sensor −> Right Motor
2
1
0
400
500
600
Time(seconds)
Figure 4.8: These graphs show the mean synaptic weights between the sensor and motor neurons averaged
over 50 trials in the food-only environment. The left to right, and right to left connection weights are rapidly
increased to their maximum value, causing food-attraction behaviour.
To demonstrate how the robot is able to learn to perform food attraction behaviour, the mean eligibility
trace for synapses between the left sensor and both motors, for the first 30 seconds in one of the trials, has
been plotted in figure 4.9. At the start the synaptic weights are set to zero and the robot performs a random
walk, notice that the eligibility traces fluctuate up and down, but neither is consistently higher. When the
robot collects a food item it is more likely that it was recently turning towards the food item, and as such
the left sensor and right motor neurons will have been active and correlated (left sensor stimulated, followed
by right motor neurons being active). This means that the eligibility trace is likely to be higher for the left
sensor to right motor when a food item is collected. This can start to be seen at 16 seconds, but is most
noticeable after 21 seconds. The synaptic weights are only significantly modified when the dopamine level is
high (when food is collected), so the left sensor to right motor connections will become strengthened. This
in turn means that the robot is more likely to turn towards food items and as such the left sensor to right
motor connection strength is even more likely to be increased. This action reinforcement is why the food
attraction synaptic connections quickly become saturated to their maximum value.
39
Sensor to Motor Eligibility Traces in Food−Only Environment
0.15
Left Sensor−>Left Motor
Left Sensor−>Right Motor
Food Collected
0.1
Eligibility Trace
0.05
0
−0.05
−0.1
−0.15
−0.2
0
5
10
15
Time (seconds)
20
25
30
Figure 4.9: The mean synaptic eligibility trace for the left sensor to left motor neurons, and the left sensor
to right sensor neurons. The vertical lines display when food is collected. Initially the weights are zero and
the robot performs a random walk. The probability that the eligibility trace will be higher for the left sensor
to left motor connections when food is achieved acts to increase these connections. This acts as a feedback
loop as the robot is more likely to turn towards the food and so the left sensor to left motor eligibility trace
is even more likely to be positive.
Notice that, from figure 4.8, the left sensor to left motor and right sensor to right motor synaptic weights
also increase, but they level off at a lower value. The reason for this is the negative baseline level of dopamine.
Figure 4.10 shows the eligibility trace for the left sensor to motor synapses for a single trial, as well as the
synaptic weights for these synapses. Initially the synaptic weights between the left sensor and the left motor
are small, as such the firing of the left sensor neurons and the left motor neurons are uncorrelated, causing
a negative eligibility trace. The negative baseline level of dopamine acts to increase the synaptic weights
of synapses with negative eligibility trace. As the synaptic weights between the left sensor and left motor
increase, the neurons start to fire in a more correlated manner (stimulating the left motor causes some firings
in the left motor). This in turn causes the eligibility trace to rise. Eventually, at around 200 seconds, the
eligibility trace rises to a point where the positive dopamine spikes caused by food act to cancel out the
negative baseline level of dopamine. At this point the left sensor to left motor synaptic weights stabilize.
The left sensor to right motor synapses are also subject to this process. However, the main cause of
potentiation in these synapses comes from the large spikes in dopamine when food is collected. These large
spikes in dopamine, which occur when the left sensor to right motor eligibility traces are highest, override
any effects the small negative baseline dopamine has.
The final synaptic weights for all sensor neuron to motor neuron synapses for a single trial is shown in
figure 4.11. There are a couple of interesting properties to note about these weights. Firstly, a lot of synapses
are either fully potentiated (and have their weight set to 4mV) or are fully depressed (and have a weight
of 0mV). This is due to the feedback mechanism inherent in STDP, if a synaptic weight starts to become
potentiated then it has a greater chance of being further potentiated in future.
The second interesting property is that individual motor neurons have differentiated themselves to either
only respond to the left sensor, or only respond to the right sensor. This occurs because all the sensor
neurons are stimulated at once, so if a synapse from a left sensor neuron to a left motor neuron becomes
potentiated then it is likely that the left motor neuron will fire after all the left sensor neurons are stimulated,
and as such all synapses from the left sensor to that left motor neuron will become potentiated. Once the
robot has learnt attraction behaviour, then a common sequence of neural firings is:
Left Sensor->Right Motor->Right Sensor->Left Motor->Left Sensor->Right Motor...
As can be seen, once a motor neuron is reacting strongly to the left sensor it is more likely to fire just before
the right sensor is stimulated. This acts to depress any connections from the right motor, and keep the
neuron differentiated.
40
Sensor to Motor Eligibility Trace for Food−Only Environment
Left Sensor−>Left Motor
Left Sensor−>Right Motor
0.2
Eligibility Trace (mV)
0.15
0.1
0.05
0
−0.05
−0.1
−0.15
0
50
100
150
200
250
300
Time (seconds)
350
400
450
500
(a) Eligibility Trace for Single Trial
Robot Sensor to Motor Connections in the Food−Only Environment (Single Trial)
Mean Synaptic Weight (mV)
2.5
2
Left Sensor −> Left Motor
Left Sensor −> Right Motor
1.5
1
0.5
0
0
50
100
150
200
250
300
Time(seconds)
350
400
450
500
(b) Synaptic Weights for Single Trial
Figure 4.10: These graphs show the eligibility trace and synaptic weights between the left sensor neurons and
both groups of motor neurons for the first 500 seconds in a single trial in the food-only environment. Initially
the left sensor to left motor eligibility traces are highly negative, this drives up their synaptic weights. After
200 milliseconds the left sensor to left motor eligibility traces have risen enough that their synaptic weights
are stable and do not increase.
41
Post−Synaptic Neuron Index
Sensor to Motor Synaptic Weights after Training (Single Trial)
No Connection Exists
Between Neurons
45
3.5
50
3
55
2.5
60
2
65
1.5
70
1
75
0.5
80
5
10
15
20
25
30
Pre−Synaptic Neuron Index
35
40
0
Figure 4.11: The synaptic weights between sensors and motors, after training in the food-only environment,
is plotted. Motor neuron indices are on the left axis and range from 41-60 (left motor) and 61-80 (right
motor). Sensor neuron indices are on the bottom axis and range from 1-21 (left sensor) and 21-40 (right
sensor). Most neurons are either fully potentiated (black) or fully depressed (white). Motor neurons have
differentiated themselves to either respond to left sensor neurons or right sensor neurons, as evident by the
horizontal banding.
4.3
Food-Poison Plasticity Learning
Once the robot has learnt a behaviour, it should be able to modify its behaviour if the environment changes.
This is demonstrated by replacing food items with poison once the robot has learnt food attraction behaviour.
4.3.1
Experimental Set-up
The robot was placed in a 100cmx100cm environment consisting of 20 food items for 1000 seconds, giving
the robot enough time to learn food attraction behaviour. The food items were then replaced with poison,
which produces a negative dopamine response, and the robot was run for a further 1000 seconds. Each trial
was run 50 times. As a comparison a robot with learning disabled was also run in this environmental set-up
for 50 trials.
4.3.2
Results and Discussion
The collection rate over time for the dynamic food-poison environment is shown in figure 4.12. The robot is
able to unlearn the food attraction behaviour almost immediately.
To see how such a rapid relearning takes place it is useful to examine the eligibility trace at the point
that poison is introduced, the mean eligibility trace for connections from the left sensor to the left and right
motors, over the period where food is switched with poison, is shown in figure 4.13. Due to the fact that the
robot has learnt food attraction behaviour there are strong connections between the left sensor and the right
motor. This means that the eligibility trace between these two sets of neurons is very high, as a stimulation
of the left sensor always causes a response in activity in the right motor. This high eligibility trace means
that when poison is introduced, and a negative spike in dopamine is received, there is a rapid decrease in the
weights between the left sensor and right motor. This is shown in figure 4.14(a) where the mean synaptic
weights between sensors and motors are plotted, averaged over all 50 trials. There is also some reduction in
synaptic weights for the left sensor to left motor and right sensor to right motor synapses. This is because
a few of these synaptic connections have become potentiated to their maximum value and as such have a
high eligibility trace, the drop in weights from these saturated connections is greater than the increase in
synaptic weight from uncorrelated connections.
The other thing to note about the food collection rate in this environment is that after poison is introduced
the rate of food collection is only marginally less than for the robot performing a random walk, this is reflected
in the synaptic weights for avoidance only being marginally higher, as can be seen in figure 4.14(b). This
42
Food Collection Rate in Food−Poison Environment
1.6
Robot With Learning Disabled
Robot With Learning Enabled
Collection Rate (collections per second)
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0
200
400
600
800
1000
1200
Time (seconds)
1400
1600
1800
2000
Figure 4.12: The collection rate over time for robots in a 100cmx100cm environment with 20 food items.
After 1000 seconds these food items are replaced with poison. The learning robot quickly unlearns food
attraction behaviour once poison is introduced.
Sensor to Motor Eligibility Trace for Food−Poison Environment (Single Trial)
Food Collected
Poison Collected
Left Sensor−>Left Motor
Left Sensor−>Right Motor
0.35
0.3
Eligibility Trace (mV)
0.25
0.2
0.15
0.1
0.05
0
−0.05
−0.1
−0.15
990
995
1000
Time (seconds)
1005
1010
1015
Figure 4.13: The mean eligibility trace for synapses from the left sensor to the left and right motors is shown.
Food items are replaced with poison at 1000ms. Before the poison is introduced the eligibility trace for the
left sensor to right motor is very high. Poison is collected at exactly the point at which the left sensor to
right motor eligibility traces are highest. After poison has been collected, and food attraction unlearnt, the
eligibility trace fluctuates around zero.
43
Robot Sensor to Motor Connections in the Food−Poison Environment
2.5
Left Sensor −> Left Motor
Left Sensor −> Right Motor
Right Sensor −> Left Motor
Right Sensor −> Right Motor
Mean Synaptic Weight
2
1.5
1
0.5
0
0
200
400
600
800
1000
1200
Time(seconds)
1400
1600
1800
2000
(a) Synaptic Weights in the Food-Poison Environment
Robot Sensor to Motor Connections in the Food−Poison Environment (Close−Up)
Left Sensor −> Left Motor
Left Sensor −> Right Motor
Right Sensor −> Left Motor
Right Sensor −> Right Motor
Mean Synaptic Weight
0.7
0.65
0.6
0.55
0.5
1000
1100
1200
1300
1400
1500
1600
Time(seconds)
1700
1800
1900
2000
(b) Synaptic Weights in the Food-Poison Environment (Close-Up)
Figure 4.14: The mean synaptic weights between sensors and motors, averaged over the 50 trials in the foodpoison environment is shown. (a) shows that after poison is introduced the synaptic weights are rapidly
reduced. (b) shows a close-up of the last 1000 seconds, this shows that on average the food avoidance
synaptic weights are very slightly higher than those for food attraction.
44
is due to the fact that the robot is not rewarded for avoiding poison (in terms of an increase in dopamine),
therefore the synapses corresponding to poison avoidance will not be potentiated. As such all the sensor
to motor synapses end up with their weights at very low values. In this situation the exploration stimulus
will override any current provided by the sensor to motor connections. One possible way to get the robot to
learn food avoidance behaviour would be to have a positive baseline level of dopamine. Though, as has been
shown in section 4.1, this would prevent the robot escaping orbiting behaviour, so more research would be
needed to find a solution that works for both problems.
4.4
Dopamine Response to Secondary Stimulus.
In this section the ability of the robot to learn to elicit a positive dopamine response to a new stimulus that
is repeatedly paired with a dopamine inducing stimulus is tested.
4.4.1
Experimental Set-up
To demonstrate this reward predicting dopamine response, an environment consisting of food items in containers was used, see section 3.6 for detail in how the architecture of the robot was changed for this environment. The environment is 300cmx300cm and consists of 12 containers which contain food, and 12 containers
which contain no food, this environment is visualized in figure 4.15. The robot has a different sensor for
detecting if it is entering a food containing container, or an empty container. However, the left and right
container sensors cannot distinguish between the two container types. At the start of the experiment the
sensor to motor connections were set for food and container attraction behaviour. Figure 4.16 shows how
the synaptic weights were set to force attraction behaviour. The result of this attraction behaviour is that
the robot will visit both empty and food containing containers equally often, when inside a food container
the robot will turn towards and consume the food item.
300
Environment with Food Containers and Empty Containers
250
Food-Container
Food
y location (cm)
200
Empty-Container
150
100
50
50
100
150
200
x location (cm)
250
300
Figure 4.15: The environmental set-up to demonstrate secondary stimulus dopamine learning, consisting of
food items, food-containers, and empty-containers.
The synaptic weights between the two container touch-sensor neurons and the dopaminergic neurons
were initially set to zero and plasticity was turned on for these connections. The robot was placed in the
environment and run for a 3000 second trial. This trial was repeated 50 times.
45
Left
Food Sensor
Neurons
Right
Food Sensor
Neurons
Left
Container
Neurons
Synapses
Set to
Maximum
Right
Container
Neurons
Synapses
Set to Zero
Left Motor
Neurons
Right Motor
Neurons
Figure 4.16: The starting weights for the robot to force attraction behaviour.
4.4.2
Results and Discussion
In all 50 trials the food-container touch-sensor to dopaminergic neuron synapses were potentiated by more
than four times that of the empty-container touch-sensor to dopaminergic neuron synapses by the end of the
trial. Figure 4.17 shows the mean synaptic weight over time for the food container to dopaminergic neuron
synapses and for the empty container to dopaminergic neuron synapses. As can be seen, the synapses
between the reward predicting stimulus and the dopaminergic neurons are potentiated until levelling off at a
maximum value. The non-reward predicting stimulus, by contrast, has very little potentiation in its synaptic
connections to the dopaminergic neurons. These remaining stable at just above zero.
Connection Strength Between Container Sensors and Dopamine Neurons
0.35
Food Container −> Dopamine
Empty Container −> Dopamine
Mean Synaptic Weight
0.3
0.25
0.2
0.15
0.1
0.05
0
0
500
1000
1500
Time (seconds)
2000
2500
3000
Figure 4.17: The mean synaptic weight over time between food-container touch-sensor neurons and dopaminergic neurons, as well as between empty-container touch-sensor neurons and dopaminergic neurons over 50
trials.
To see the effect that dopamine thresholding (described in section 3.8) has on the dynamics of the system,
the experiment was re-run without requiring five dopaminergic neurons to fire in order to raise the level of
dopamine. The effect of this is that background firing in the dopaminergic neurons can raise the level of
dopamine. With dopamine thresholding switched off the food-container touch-sensor to dopaminergic neuron
synaptic weights at the end of the trial were higher than the empty-container touch-sensor to dopaminergic
neuron synaptic weights in only 80% of the trials, as compared with 100% with thresholding enabled. The
synaptic weights for the experiment without thresholding are plotted in figure 4.18.
Note that there is in essence two processes that are happening. For the first 400 seconds both sets of
synapses are increased. To see why both sets of synaptic weights are increased, and why the reward predicting
set is increased more, it is useful to plot the eligibility trace for a single synapse. In figure 4.19 the eligibility
trace, firing rate, and synaptic strength is plotted for a single synapse from an empty-container neuron to a
single dopaminergic neuron. At 1300ms the container neuron happens to fire just before the dopaminergic
neuron purely by chance, due to the background firing rate. This causes the eligibility trace to jump up, as
the second neuron is a dopaminergic neuron its firing immediately causes a spike in the level of dopamine
46
Connection Strength Between Container Sensors and Dopamine Neurons (No Threshold)
0.25
Mean Synaptic Weight
0.2
Food Container −> Dopamine
Empty Container −> Dopamine
0.15
0.1
0.05
0
0
500
1000
1500
Time (seconds)
2000
2500
3000
Figure 4.18: The synaptic weights between the empty-container sensor and dopaminergic neurons, and
between the food-container sensor and dopaminergic neurons is shown. In this experiment there was no
dopamine threshold, such that a single dopaminergic neuron spiking could raise the level of dopamine.
(which would not happen with dopamine thresholding). The higher dopamine level combined with the high
eligibility trace cause the synaptic weight to increase. This could be described as a double feedback loop,
the increase in synaptic weight means that the two neurons are more likely to fire in sequence in future,
meaning that their eligibility trace is more likely to increase. While an increase in synaptic strength means
that the dopaminergic neuron is more likely to fire, and hence increase the level of dopamine, at exactly the
point that will raise the strength of the synapse the most.
If we consider what happens when a reward predicting stimulus (food-container touch-sensor) neuron
coincidentally fires before a dopaminergic neuron, then not only do we have the same increase in dopamine
and synaptic weight as described above. It is also likely that there is a very large spike in dopamine within
1000ms (due to a food item being collected), which will have the effect of raising the synaptic weight even
more since the eligibility trace will still be positive. This explains why the rate of increase of synaptic weights
from the reward predicting stimulus to dopaminergic neurons is faster than for the non-reward predicting
stimulus.
The second thing to note about figure 4.18 is that after about 400 seconds the synaptic weights of the
empty-container touch-sensor to dopaminergic synapses start to decrease. This is due to the homeostatic
mechanism described in section 3.9, the combined synaptic weights for all synapses leading to dopaminergic
neurons is limited. Once this limit is reached then all synapses leading to dopaminergic neurons have their
weight reduced. The food container to dopaminergic synapses will then increase their weight faster than the
empty container to dopaminergic synapses, so the next time the weights are reduced the empty container
synapses will be reduced by more than they increased in this time interval.
Without dopamine thresholding and synaptic weight dampening there would be nothing to stop the
empty-container touch-sensor to dopaminergic neuron synapses potentiating to their maximum value. By
using both of these techniques the robot is more robust and able to correctly learn to produce a spike in
dopamine for a reward-predicting stimulus, but not for a non reward-predicting stimulus.
47
Eligibility Trace
0.05
0
Neuron Index
Dopamine (µM)
Synaptic Weight
Eligibility Trace
0.1
0
500
1000
1500
2000
Time (ms)
Firings
2500
3000
3500
4000
0
500
1000
1500
2000
Time (seconds)
Dopamine
2500
3000
3500
4000
0
500
1000
1500
2000
Time (seconds)
Synaptic Strength
2500
3000
3500
4000
0
500
1000
1500
2000
Time (seconds)
2500
3000
3500
4000
3
2
1
0
0.02
0.01
0
0.06
0.04
0.02
0
Figure 4.19: An example of how empty-container to dopaminergic neuron synapses can become potentiated.
Neuron 1 is a empty-container sensor neuron, and neuron 2 is a dopaminergic neuron. At 1300ms they
coincidentally fire, the corresponding spike in dopamine causes the synaptic weight to increase.
4.5
Secondary Behaviour Learning
In this section we explore the ability of the robot to learn a secondary behaviour. This being a behaviour
that, in and of itself, does not give a direct reward, but when succeeded by another behaviour can lead to
a reward. The behaviour to be learnt is that of food-container attraction, driving towards food-containers
is an advantageous behaviour, as entering a food-container often leads to food being collected if the robot
subsequently performs food-attraction behaviour.
4.5.1
Experimental Set-up
An environment consisting of 17 food-containers, each containing a single item of food, was created. The
food-containers are positioned randomly and are moved to a random location whenever their associated
food item is collected. Five different robots were run in this environment for 50 trials of 2000 seconds each.
All robots, except the benchmark robot, have their initial sensor to motor synaptic weights set for food
attraction behaviour, no preference for container attraction/avoidance is set. The result is that initially all
the robots wander randomly whilst outside containers. Whilst inside containers the robots will turn towards
and collect food items.
Note that unlike in the food-only environment, this environment has a fixed optimal robot. This is
because the size of the food-containers means that the robot cannot get stuck orbiting them. Also, whilst in
the food-container, if the robot starts orbiting a food item then its orbit will take it out of the container. As
such orbiting isn’t a problem in this environment and we can set up a fixed optimal robot as a benchmark.
The dynamics of the five robots is:
• Learning Disabled - Plasticity is switched off for all synapses. The result is that the robot wanders
randomly whilst outside containers. Whilst inside containers the robot turns and collects food items.
• No Container Dopamine - Connections between container touch-sensors and dopaminergic neurons are
set to zero, and plasticity is switched off for these synapses. As such no dopamine response is received
when the container touch-sensor is stimulated. Plasticity is enabled for container range-sensor to motor
synapses.
48
• Container Dopamine Learning Enabled - Connections between container touch-sensors and dopaminergic neurons are initially set to zero. Plasticity is enabled for these connections, allowing them to become
potentiated and for the robot to learn a dopamine response to entering a food-container. Container
range-sensor to motor synapses are plastic.
• Fixed High Container Dopamine - Connections between container touch-sensors and dopaminergic
neurons are set to their maximum value and plasticity is switched off for these synapses. This forces
a spike in dopamine whenever the robot enters a food-container. Container range-sensor to motor
synapses are plastic.
• Benchmark - This robot has plasticity switched off. Food attraction synapses are set to their maximal
value as well as food-container attraction synapses. All other sensor to motor synapses are set to zero.
4.5.2
Results and Discussion
In figure 4.20 the average score over time for each of the five robots is plotted. Correspondingly, the total
food collected as well as the percentage of robots that had learnt food-container attraction behaviour by the
end of the trial is shown in table 4.2. A robot was deemed to have learnt container-attraction behaviour if,
at the end of the trial, the left container range-sensor to right motor neuron synapses where stronger than
0.5mV on average, as well as being more than 10% stronger on average than the left container range-sensor to
left motor neuron synapses. The equivalent also had to be true for the right container range-sensor synapses
for the robot to be deemed correct.
Food Collection Rate in Food−Container Environment
Benchmark
Fixed High Container−Dopamine
Container−Dopamine Learning Enabled
No Container Dopamine
Learning Disabled
Collection Rate (collections per second)
0.5
0.4
0.3
0.2
0.1
0
0
200
400
600
800
1000
Time (seconds)
1200
1400
1600
1800
2000
Figure 4.20: The averaged score over time for five different robots run in a 300cmx300cm environment with
20 food-items contained in containers.
There are several interesting properties shown in these figures. Firstly, note that all the robots managed
to perform better than the robot with learning disabled that performed a random walk whilst outside
the containers. This means that even the robot which did not receive any dopamine when the container
touch-sensors were stimulated (No Container Dopamine) was able to learn some food-container attraction
behaviour, from table 4.2 we can see that it managed to learn container attraction in 50% of cases. This is
due to the fact that there is on average 800ms between entering a container and food being collected. This
is small enough that DA-modulated STDP can still have a small effect on potentiating the synapses that
cause container attraction behaviour.
The next thing to notice is that both the robot with fixed strong weights between food-container sensors
and dopaminergic neurons (Fixed High Container Dopamine), and the robot that has to learn to poten49
Robot Type
Learning Disabled
No Container-Dopamine
Container-Dopamine
Learning Enabled
Fixed High ContainerDopamine
Benchmark
Mean Food Collected
224
348.5
446.7
Std. Deviation Food Collected
32.8
87.5
103.4
% Correct
0
50
96
529.7
152.5
94
808.4
21.5
100
Table 4.2: This table shows the total amount of food collected for the three robots, averaged over the 50
trials.
tiate these synapses (Container Dopamine Learning Enabled ), perform better than the robot without any
dopamine response to the food-container touch-sensor. They are both able to learn food-container attraction
behaviour in about 95% of cases. This shows that by being able to learn to elicit a dopamine response to a
reward-predicting stimulus helps in being able to learn a sequence of behaviours needed for the agent to get
a reward.
As expected, the robot which starts with its food-container touch-sensor to dopaminergic neurons already
fully potentiated is able to learn food-container attraction behaviour faster than the robot which has to learn
to potentiate these neurons before stimulating the food-container touch-sensor will elicit a dopamine response.
Figure 4.21 shows the mean synaptic weight between the food-container touch-sensor and dopaminergic
neurons for the robot with container to dopamine plasticity enabled. As can be seen the robot is able to
correctly potentiate these synaptic weights. As these synapses potentiate the robot is able to learn foodcontainer attraction behaviour, this is shown by the fact that food collection rate for this robot slowly rises
until it is equal to the robot that had its container touch-sensor to dopaminergic synapses strong from the
start.
Food−Container to Dopaminergic Neurons Synaptic Weight
Mean Synaptic Weight (mV)
0.2
0.15
0.1
0.05
0
0
200
400
600
800
1000
1200
Time (seconds)
1400
1600
1800
2000
Figure 4.21: The average synaptic weight between the food-container touch-sensor neurons and the dopaminergic neurons, over the 50 trials, is shown for the Container Dopamine Learning Enabled robot.
The last thing to note about the food collection rate in figure 4.20 is that none of the robots with learning
enabled are able to match the food collection rate of a hard-coded optimal robot. To see why this is the case,
the synaptic weights for container range-sensor to motor synapses have been plotted in figure 4.22 for the
robot with a fixed high container touch-sensor dopamine response (Fixed High Container-Dopamine). The
mean synaptic weights for container attraction plateau at a value less than the maximum possible. With
85% connection probability and a maximum single synapse weight of 4mV, the maximum mean synaptic
weight between a group of sensor neurons and a group of motor neurons is 3.4mV. The reason that they
level off at a lower value is due to the motor neuron differentiation that was discussed in section 4.2.2 (figure
4.11). Some motor neurons have differentiated to only respond to container-avoidance behaviour and as such
will not become potentiated.
In addition the synapses for container avoidance have also had their synaptic weights increased, though to
a lesser extent, this is due to the negative baseline level of dopamine. This mechanism was discussed in more
depth in section 4.2.2. These two mechanisms are needed for the robot to be able to modify its behaviour if
the environment changes. As a direct consequence of this ability to cope with a dynamic environment, the
50
robot does not perform as well as a robot with a fixed, optimal, set of weights for this environment. The
greater difference in synaptic weights allows the benchmark robot to respond to very weak sensory input,
from the edge of the sensors range, and as such perform taxis from further away.
Container Range Sensor to Motor Synaptic Weights for Fixed High Container−Dopamine Robot
1.6
Mean Synaptic Weight (mV)
1.4
Left Sensor −> Left Motor
Left Sensor −> Right Motor
Right Sensor −> Left Motor
Right Sensor −> Right Motor
1.2
1
0.8
0.6
0.4
0.2
0
0
200
400
600
800
1000
1200
Time(seconds)
1400
1600
1800
2000
Figure 4.22: Mean synaptic weights between container sensor neurons and motor neurons for the Fixed High
Container-Dopamine robot in the food-container environment.
It is interesting to look at what happens in the case where the robot does not learn the correct behaviour.
Figure 4.23 shows the mean synaptic weights over time from the left container sensor to both motors, for a
single trial of the Fixed High Container Dopamine robot, in which the robot failed to learn correctly. What
can be seen from this figure is that if the synaptic weights from sensor neurons to both motors are close in
value and both strong then the synaptic weights are increased and decreased in a synchronous fashion. The
reason for this is that if the left sensor is connected strongly to both motors, then stimulating the left sensor
will have the same effect on all motor neurons, they will fire in response. This means that the eligibility
traces between the sensor and the two motors will all increase together and as such the synaptic weights will
increase or decrease together. In this way, once the synapses for avoidance and attraction have both become
potentiated, it is hard for the robot to learn to only increase the synaptic weights needed for attraction
behaviour.
Left Container Sensor Synaptic Weights for Failed Learning
Mean Synaptic Weight
3
2.5
2
1.5
Left Sensor −> Left Motor
Left Sensor −> Right Motor
1
0.5
0
0
200
400
600
800
1000
Time(seconds)
1200
1400
1600
1800
2000
Figure 4.23: The synaptic weights between the left container sensor and the left and right motor neurons
is shown for an example where the robot was unable to learn container attraction behaviour. The weights
have become synchronized and rise and fall together.
The cause for both sets of synaptic weights to become high in the first place is that it is still possible
51
for the robot to collect food items even if it isn’t performing container attraction behaviour. For example,
if the robot was just to constantly drive straight then purely by chance it will run into a food container and
as such have the synaptic weights between all sensors and all motors increased. Due to random exploration
it may happen that this will occur a few times in a row, at which point the robot will get ’stuck’ performing
this sub-optimal behaviour.
4.6
Dual Behaviour Learning
Finally, we demonstrate the ability of the robot to learn two different, sequential, behaviours at the same
time. These behaviours being food-attraction and container-attraction.
4.6.1
Experimental Set-up
The same environmental set-up as in section 4.5 was used; 20 food containers randomly positioned in a
300cmx300cm environment. Three different robots were run in this environment for 50 trials, with 5000
seconds for each trial. The three robots were:
• Random Walk - This robot had all sensor to motor connections set to zero and plasticity switched off,
food is only collected by randomly driving into it.
• Benchmark - This robot has all synaptic weights for food-attraction and container-attraction behaviour
set to their maximum value. All other sensor to motor synaptic weights were set to zero, and plasticity
was switched off.
• Learning Enabled - Initially all sensor to motor synapses are set to zero, plasticity is enabled for these
synapses. In addition food-container touch-sensor to dopaminergic neuron synapses were initially set
to zero, and plasticity was enabled for these synapses.
4.6.2
Results and Discussion
The average food collection rate over time for the three robots is shown in figure 4.24. The collection rate
for the learning robot, after 5000 seconds, is comparable to the robots with food-taxis behaviour hard-coded
in the previous section (see figure 4.20). The learning robot was able to learn both food-attraction and
container-attraction behaviour (the same definition of correctness as in sections 4.2.2 and 4.5 was used) in
82% of the trials. This is slightly worse than the 94% figure achieved when food-attraction behaviour was
hard-coded. This is to be expected, as there will be longer periods between food collection and as such there
is more chance that random fluctuations in synaptic weight cause two sensor to motor groups to become
synchronized, as described in section 4.5.
In figure 4.25 the mean synaptic weights for the learning robot, averaged over the 50 trials, is shown.
The first behaviour to be learnt is food-attraction, as shown by the quick potentiation of the corresponding
synapses. After this, the container-dopamine response and container-attraction behaviour are learnt in
parallel. With the container-attraction behaviour being reinforced by the corresponding increase in release
of dopamine.
52
Food Collection Rate in Food−Container Environment (Dual Behaviour Learning)
0.5
0.45
Collection Rate (collections per second)
0.4
0.35
0.3
0.25
0.2
Benchmark
Learning Enabled
Random Walk
0.15
0.1
0.05
0
0
500
1000
1500
2000
2500
3000
Time (seconds)
3500
4000
4500
5000
Figure 4.24: The collection rate over time, averaged over all trials, for each of the robots is shown. The
robot with learning enabled increases its collection rate over time.
53
Food Sensor to Motor Connections in the Food−Container Environment (Dual Learning)
Mean Synaptic Weight (mV)
2.5
Left Food Sensor −> Left Motor
Left Food Sensor −> Right Motor
Right Food Sensor −> Left Motor
Right Food Sensor −> Right Motor
2
1.5
1
0.5
0
0
500
1000
1500
2000
2500
Time(seconds)
3000
3500
4000
4500
5000
4500
5000
(a) Food Sensor to Motor Synaptic Weights
Container Sensor to Motor Connections in the Food−Container Environment (Dual Learning)
Mean Synaptic Weight (mV)
2.5
Left Container Sensor −> Left Motor
Left Container Sensor −> Right Motor
Right Container Sensor −> Left Motor
Right Container Sensor −> Right Motor
2
1.5
1
0.5
0
0
500
1000
1500
2000
2500
3000
Time(seconds)
3500
4000
(b) Container Sensor to Motor Synaptic Weights
Food−Container to Dopaminergic Synaptic Weights in the Food−Container Environment (Dual Learning)
Mean Synaptic Weight (mV)
0.25
0.2
0.15
0.1
0.05
0
0
500
1000
1500
2000
2500
3000
Time (seconds)
3500
4000
4500
5000
(c) Food-Container Touch-Sensor to Dopaminergic Neurons Synaptic Weights
Figure 4.25: Synaptic weights for the dual learning robot in the food-container environment. (a) shows the
mean synaptic weights between the food range-sensors and the motors. The synapses for food-attraction
behaviour are quickly potentiated. (b) shows the mean synaptic weights between the food-container rangesensors and the motors. The synapses for container-attraction behaviour are potentiated, but at a slower
rate. (c) shows the mean synaptic weight between the food-container touch-sensor and the dopaminergic
neurons. Once food-attraction behaviour is learnt, these synapses are slows potentiated.
54
Chapter 5
Conclusions
5.1
Reinforcement Learning
Comparing the properties of our implementation with classical reinforcement learning algorithms (an outline
of which is given in section 2.3) we can see that several properties of our model have correlates with the
TD-Error reinforcement learning algorithms. The most obvious of which is the eligibility trace that is
stored within the synapses. This directly correlates with the eligibility trace as it is defined in standard
reinforcement learning algorithms such as Sarsa(λ). The role of this eligibility trace in both situations is to
allow the agent to learn much faster, and over a greater time frame.
If the case of our novel exploration strategy whereby we randomly stimulate the motors with a small
current, this can be considered as analogous to off-policy reinforcement learning algorithms. Whereby the
probability of choosing an action is related to the expected reward. In our implementation, as the strength
of a behaviour increases, and as such the confidence that it will produce the best reward increases, the
probability of performing exploratory behaviour decreases.
In reinforcement learning algorithms the value function is propagated through the state space. For
example, initially only the goal state will have a positive value associated with it. After repeated exploration
of the environment then the states close to the goal state will have positive value function. This is similar to
the way in which our network is able to learn a dopaminergic neural response for stimulus that often precede
food collection.
The key difference between our implementation and reinforcement learning algorithms is that in standard
reinforcement learning algorithms the value function is updated based on the difference between expected
and received reward. For example if the value function predicts a reward of 0.8 and the agent subsequently
receives a reward of 0.4, then the value function is decremented. The value function can be thought of as
analogous to the strength of connections to dopaminergic neurons. The model used in this paper is not able
to produce this negative response when a trained reward-predicting stimulus is subsequently stimulated but
not paired with a reward. For example, if a robot is trained in the food-container environment and then we
remove the food items from the food-containers. The network will still elicit a dopamine response when the
food-container touch-sensor is stimulated and as such food-container attraction behaviour will not become
unlearnt. This is the main limitation of our implementation and possible solutions are discussed in section
6.
Our implementation provides a method of reinforcement learning in continuous parameter spaces. Standard reinforcement learning algorithms are designed for state and actions spaces that have been discretized.
Being able to operate with continuous spaces directly allows an agent to generalize much more easily when
a previously unseen environmental state is encountered.
5.2
Biological Plausibility and Conclusions
In this paper we have shown that by allowing the baseline level of dopamine to be negative the robot is
able to relearn when its current behaviour is not producing a reward. In the brain the level of dopamine
cannot become negative, though it may be that there is another type of pain neurotransmitter that acts in
the reverse way to dopamine and as such has the effect of reducing highly active synapses, though whether
or not one exists is currently unknown.
One consequence of our model was that over time motor neurons tended to differentiate themselves to
only respond to a single sensor. This is an interesting property as it provides a mechanism as to how a group
of neurons in the brain can self organize into functionally different subgroups.
55
In the brain it has been shown that dopaminergic neurons have two different states of activity, these
being background firing and bursting activity when stimulated, such that background firing does not result
in a significant level of dopamine increase [22]. In our model we have shown that by separating dopaminergic
neuron activity into these two distinct groups the robot is able to learn a dopamine response for a rewardpredicting stimulus much more reliably. This provides one possible explanation for why dopaminergic neurons
have these two different firing states.
56
Chapter 6
Future Work
Several mechanisms that our robot used were explicitly programmed, separate to the neural network itself.
It would be advantageous to provide an implementation of these within the spiking neural network. Such
that they can be subject to plasticity and modified in response to environmental changes along with the rest
of the network. These mechanisms, and a possible SNN implementation are given below:
• Phasic Activity - The sensor neurons were stimulated at a rate of 14Hz, rather than being given a
constant input current. EEG recordings show that the brain also exhibits neural oscillatory behaviour.
A group of interconnected inhibitory neurons when provided constant input current will generate
oscillatory firing. This group of neurons could be used a control group and connected to the sensory
neurons, forcing them to fire in an oscillatory manner.
• Winner-Takes-All - We explicitly implemented a winner takes all mechanism between sensors and
between motors. One way in which this could possibly be implemented directly in the SNN is shown
in figure 6.1.
• Exploration - We explicitly stimulated the motors to provide exploratory behaviour. However, if the
phasic activity and winner-takes-all mechanisms were implemented directly in the network as described
above then no explicit exploration behaviour would be needed. The combination of background firing
and winner-takes-all would ensure that only one set of motor neurons was active at any one time,
whilst the phasic activity would allow the active motor to optionally switch during the inactive phase
of the neurons.
Simulus 1
Stimulus 2
Stimulus 1
Neurons
Stimulus 2
Neurons
Inhibitory
Group 1
Inhibitory
Group 2
External Inputs
Excitatory Neurons
Inhibitory Neurons
Figure 6.1: An example of how a winner-takes-all mechanism can be implemented in a SNN. Each set of
stimulus neurons acts to inhibit the other until the network reaches a stable state where only one group is
active.
The main limitation in the approach taken in this paper is that the robot cannot unlearn an already learnt
dopamine response. To be able to do this a fuller SNN implementation of the TD-error algorithm would need
to be implemented, whereby the dopamine response would correspond to the difference between expected
reward and perceived reward. To be able to do this requires the SNN to have a short-term memory. Chorley
& Seth [31] have shown a network architecture where the level of dopamine corresponds to the difference
between perceived and expected reward through the use of a model of short-term memory. Future work
would incorporate this network architecture into the robot to allow sequential behaviours to be unlearnt.
Another obvious extension to the work in this paper would be to implement the control architecture in
a real, rather than simulated, robot. In this case the robot would need to be able to deal with collision
57
avoidance, which could possibly be solved through a similar reinforcement learning mechanism as described
in this paper.
58
Bibliography
[1] Izhikevich, E. (2007) Solving the distal reward problem through linkage of stdp and dopamine signaling.
Cerebral Cortex , 17, 2443–2452.
[2] Chorley, P. and Seth, A. (2008) Closing the sensory-motor loop on dopamine signalled reinforcement
learning. Proceedings of the 10th international conference on Simulation of Adaptive Behavior: From
Animals to Animats, Berlin, Heidelberg, pp. 280–290, SAB ’08, Springer-Verlag.
[3] Iannella, N. and Back, A. (2001) A spiking neural network architecture for nonlinear function approximation. Neural Networks, 14, 933 – 939.
[4] Hodgkin, A. and Huxley, A. (1952) A quantitative description of membrane current and its application
to conduction and excitation in nerve. The Journal of Physiology, 117, 500–544.
[5] Brunel, N. and van Rossum, M. (2007) Lapicques 1907 paper: from frogs to integrate-and-fire. Biological
Cybernetics, 97, 337–339.
[6] FitzHugh, R. (1961) Impulses and physiological states in theoretical models of nerve membrane. Biophysical J., 1, 445–466.
[7] Hindmarsh, J. and Rose, R. (1984) A model of neuronal bursting using three coupled first order differential equations. Proceedings of the Royal Society of London. Series B. Biological Sciences, 221,
87–102.
[8] Izhikevich, E. (2003) Simple model of spiking neurons. Neural Networks, IEEE Transactions on, 14,
1569 – 1572.
[9] Rosenblatt, F. (1962) Principles in neurodynamics. Spartan Books.
[10] Widrow, B., Stearns, S., and Burgess, J. (1986) Adaptive signal processing edited by bernard widrow
and samuel d. stearns. The Journal of the Acoustical Society of America, 80, 991–992.
[11] Werbos, P. (1974) Beyond Regression: New Tools for Prediction and Analysis in the Behavioral Sciences.
Ph.D. thesis, Harvard University, Cambridge, MA.
[12] Sutton, R. and Barto, A. (1998) Reinforcement Learning: An Introduction (Adaptive Computation and
Machine Learning). The MIT Press.
[13] Bienenstock, E., L., C., and P., M. (1982) Theory for the development of neuron selectivity: orientation
specificity and binocular interaction in visual cortex. The Journal of Neuroscience, 2, 32–48.
[14] Markram, H., Lbke, J., Frotscher, M., and Sakmann, B. (1997) Regulation of synaptic efficacy by
coincidence of postsynaptic aps and epsps. Science, 275, pp. 213–215.
[15] Song, S. and Abbott, L. (2001) Cortical development and remapping through spike timing-dependent
plasticity. Neuron, 32, 339 – 350.
[16] Izhikevich, E. and Desai, N. (2003) Relating stdp to bcm. Neural Computation, 15, 1511–1523.
[17] Otmakhova, N. and Lisman, J. (1996) D1/d5 dopamine receptor activation increases the magnitude of
early long-term potentiation at ca1 hippocampal synapses. The Journal of Neuroscience, 16, 7478–7486.
[18] Chinta, S. and Andersen, J. (2005) Dopaminergic neurons. The International Journal of Biochemistry
& Cell Biology, 37, 942 – 946.
[19] Miller, E., Freedman, D., and Wallis, J. (2002) The prefrontal cortex: categories, concepts and cognition.
Philos.Trans.R.Soc.B-Biol.Sci., 357, 1123–1136.
59
[20] Ping, H. and Shepard, P. (1996) Apamin-sensitive ca2+-activated k+ channels regulate pacemaker
activity in nigral dopamine neurons. Neuroreport, 7, 809–814.
[21] Overton, P. and Clark, D. (1997) Burst firing in midbrain dopaminergic neurons. Brain Research Reviews, 25, 312 – 334.
[22] Lisman, J. (1997) Bursts as a unit of neural information: making unreliable synapses reliable. Trends
in Neurosciences, 20, 38 – 43.
[23] Matsumoto, M. and Hikosaka, O. (2009) Two types of dopamine neuron distinctly convey positive and
negative motivational signals. Nature, 459, 837–841.
[24] Frey, U. and Morris, R. (1997) Synaptic tagging and long-term potentiation. Nature; Nature, 385,
533–536.
[25] Ppper, M., Kempter, R., and Leibold, C. (2011) Synaptic tagging, evaluation of memories, and the
distal reward problem. Learning & Memory, 18, 58–70.
[26] Redondo, R. and Morris, R. (2011) Making memories last: the synaptic tagging and capture hypothesis.
Nature reviews.Neuroscience, 12, 17.
[27] Hull, C. (1943) Principles of behavor: an introduction to behavior theory. Appleton-Century.
[28] Pavlov, I. (1927) Conditioned reflexes. Oxford University Press.
[29] Pan, W., Schmidt, R., Wickens, J., and Hyland, B. (2005) Dopamine cells respond to predicted events
during classical conditioning: Evidence for eligibility traces in the reward-learning network. The Journal
of Neuroscience, 25, 6235–6242.
[30] Ljungberg, T., Apicella, P., and Schultz, W. (1991) Responses of monkey midbrain dopamine neurons
during delayed alternation performance. Brain Research, 567, 337 – 341.
[31] Chorley, P. and Seth, A. (2011) Dopamine-signalled reward predictions generated by competitive excitation and inhibition in a spiking neural network model. Frontiers in Computational Neuroscience,
5.
[32] Brown, J., Bullock, D., and Grossberg, S. (1999) How the basal ganglia use parallel excitatory and
inhibitory learning pathways to selectively respond to unexpected rewarding cues. The Journal of Neuroscience, 19, 10502–10511.
[33] Tan, C. and Bullock, D. (2008) A local circuit model of learned striatal and dopamine cell responses
under probabilistic schedules of reward. The Journal of Neuroscience, 28, 10062–10074.
[34] Schultz, W. and Romo, R. (1990) Dopamine neurons of the monkey midbrain: contingencies of responses
to stimuli eliciting immediate behavioral reactions. Journal of Neurophysiology, 63, 607–624.
[35] Schultz, W. (1998) Predictive reward signal of dopamine neurons. Journal of Neurophysiology, 80, 1–27.
[36] Ljungberg, T., Apicella, P., and Schultz, W. (1992) Responses of monkey dopamine neurons during
learning of behavioral reactions. Journal of Neurophysiology, 67, 145–163.
[37] Turrigiano, G. (1999) Homeostatic plasticity in neuronal networks: the more things change, the more
they stay the same. Trends in Neurosciences, 22, 280 –.
[38] Turrigiano, G., Leslie, K., Desai, N., Rutherford, L., and S.B., N. (1998) Activity-dependent scaling of
quantal amplitude in neocortical neurons. Nature, 391, 892.
[39] Adrian, E. and Zotterman, Y. (1926) The impulses produced by sensory nerve endings: Part ii: The
response of a single end organ. Journal of Physiology, 61, 151–171.
[40] Maunsell, J. and Van Essen, D. (1983) Functional properties of neurons in middle temporal visual
area of the macaque monkey. i. selectivity for stimulus direction, speed, and orientation. Journal of
Neurophysiology, 49, 1127–1147.
[41] Zador, S. and Charles, A. (1995) The enigma of the brain. Current Biology, 5.
[42] Braitenberg, V. (1984) Vehicles : experiments in synthetic psychology. MIT Press.
60
[43] Indiveri, G. (1999) Neuromorphic analog vlsi sensor for visual tracking: circuits and application examples. Circuits and Systems II: Analog and Digital Signal Processing, IEEE Transactions on, 46, 1337
–1347.
[44] Lewis, M., Etienne-Cummings, R., Cohen, A., and Hartmann, M. (2000) Toward biomorphic control using custom avlsi cpg chips. Robotics and Automation, 2000. Proceedings. ICRA ’00. IEEE International
Conference on, vol. 1, pp. 494 –500 vol.1.
[45] Hagras, H., Pounds-Cornish, A., Colley, M., Callaghan, V., and Clarke, G. (2004) Evolving spiking
neural network controllers for autonomous robots. Robotics and Automation, 2004. Proceedings. ICRA
’04. 2004 IEEE International Conference on, april-1 may, vol. 5, pp. 4620 – 4626 Vol.5.
[46] Bouganis, A. and Shanahan, M. (2010) Training a spiking neural network to control a 4-dof robotic arm
based on spike timing-dependent plasticity. Neural Networks (IJCNN), The 2010 International Joint
Conference on, july, pp. 1 –8.
[47] Vasilaki, E., Fremaux, N., Urbanczik, R., Senn, W., and Gerstner, W. (2009) Spike-based reinforcement
learning in continuous state and action space: when policy gradient methods fail. PLoS computational
biology, 5, 1000586.
[48] Gupta, A. and Long, L. (2009) Hebbian learning with winner take all for spiking neural networks. Neural
Networks, 2009. IJCNN 2009. International Joint Conference on, june, pp. 1054 –1060.
[49] Timofeev, M. and Bazhenov, I. (2006) Thalamocortical oscillations. Scholarpedia, 1, 1319.
61
| 9cs.NE
|
International Journal in Foundations of Computer Science & Technology (IJFCST), Vol. 3, No.6, November 2013
A COMPARATIVE STUDY ON REMOTE TRACKING OF
PARKINSON’S DISEASE PROGRESSION USING DATA
MINING METHODS
Peyman Mohammadi1, Abdolreza Hatamlou 2 and Mohammad Masdari3
1
Department of Computer Engineering, Science and Research Branch, Islamic Azad
University, West Azerbaijan, Iran
2
Islamic Azad University, Khoy Branch, Iran
3
Department of Computer Engineering, Urmia Branch, Islamic Azad University, Urmia,
Iran
ABSTRACT
In recent years, applications of data mining methods are become more popular in many fields of medical
diagnosis and evaluations. The data mining methods are appropriate tools for discovering and extracting
of available knowledge in medical databases. In this study, we divided 11 data mining algorithms into five
groups which are applied to a dataset of patient’s clinical variables data with Parkinson’s Disease (PD) to
study the disease progression. The dataset includes 22 properties of 42 people that all of our algorithms
are applied to this dataset. The Decision Table with 0.9985 correlation coefficients has the best accuracy
and Decision Stump with 0.7919 correlation coefficients has the lowest accuracy.
KEYWORDS
Data Mining, Knowledge Discovery, Pattern Recognition, Parkinson’s disease
1. INTRODUCTION
The Parkinson’s Disease (PD) is a chronic neurological disorder with an unknown etiology,
which usually affects people over 50 years old [1]. It was reported that the Parkinson, beyond
many unknown cases, has affected one million people [2]. It is expected that as the age of earth’s
population increases, the number of patients with PD also increases.
The number of Parkinson’s disease (PD) patients is estimated to be 120–180 out of every 100,000
people, although the percentage (and hence the number of affected people) is increasing as the
life expectancies increase. The causes of Parkinson’s disease (PD) are unknown; however,
researches have shown that the degradation of the dopaminergic neurons affects the dopamine
production [3].The PD symptoms include limb tremor (especially at rest), muscles stiffness and
movement slowness, difficulty in walking, balance and coordination (especially when beginning
of motion), difficulty in eating and swallowing, and vocal disorder and mood disorders [4, 5, 6].
This disease causes voice disorder in 90% of patients [7]. Dysarthria is a motor speech disorder
that indicates inability in expression properly.
It’s observable in patients with PD to have hypokinetic dysarthria, and its classical symptoms
include reduced vocal loudness (hypophonia), monopitch, disruption of voice quality, and
abnormally fast rate of speech [8]. For decades, researchers have strived to understand more
71
International Journal in Foundations of Computer Science & Technology (IJFCST), Vol. 3, No.6, November 2013
about this disease and therefore to find some methods for successfully limiting its symptoms
which are most commonly periodic muscle tremor and/or rigidity. Other symptoms such as
akinesia, bradykinesia, and dysarthria may occur in later stages of PD [9].
In recent years, development and innovation in fusion data equipment such as sensors in scientific
and medicine areas has led to production of massive data values. The modern medicine produces
massive amounts of data stored in databases whose use of these large values of data is timeconsuming and in some cases it is impossible. In this regard, we can use some data mining
methods to perform this difficult task. Of course, all data mining methods have their own
limitations and constraints and are our task to compare methods and algorithms to choose the best
and most useful method.
The term used as “Data Mining”, was introduced in the 1990s; however data mining is the
progress and improvement of a field with a long history [10]. As datasets are growing in size,
applications and complexity, Direct hands-on data analysis has increasingly been augmented with
indirect, automatic data processing. This has been achieved by other discoveries in computer
science such as artificial neural networks (ANNs), classification and clustering algorithms plans.
Data mining techniques and algorithms has been accomplished for genetic algorithm (GA) in the
1950s, and for decision trees (DTs) in the 1960s. Also it’s a supported vector machine (SVM) in
1990s methods [11]. The first use of data mining techniques in health information systems (HIS)
and clinical application was fulfilled with the expert systems developed since the 1970s [12].
Given to knowledge, experience and the sense of physicians, medicine decisions and reviews are
done, and because of massive amount of medicine data which are kept in medical clinics and
health centers, the data mining methods can be appropriate tools for discovering available
knowledge in this database to help the physicians with diagnosis, treatment and tracking process.
The rest of this paper is organized as follows. First explained data mining techniques and
algorithms used in this study in the next section; and data mining application in healthcare field
are described in Section 3. Then the data mining algorithms are categorized and applied to a
special dataset in Section 4. The results are analysed and discussed in Section 5. Finally, some
concluding remarks and proofs outcome as well as future research lines are presented in Section
6.
2. D ATA MINING
Database pertaining to trade, agriculture, cyberspace, details of phone calls, medical data and etc.
are collected and stored very rapidly with the development of information technology, data
collection and production methods. So, since three decades ago, the human was thinking about a
new method to access hidden data in this huge and large database because the traditional systems
were not able to do so. Competition in economic, scientific, political, military fields and the
importance of access to useful information among a high amount of data without human
intervention caused the data analysis science or data mining to be established.
The concept of data mining has been developed in the late 1980s, which went further
during 10 years later [13]. Data mining is the process of discovering meaningful new
correlations, patterns and trends by sifting and investigating through large amounts of data
stored in repositories, using pattern recognition technologies as well as statistical and
mathematical techniques (Gartner Group). As like as mining gold through huge rocks and large
amount of soils, data mining is extracting beneficial knowledge from bulky data sets. But
they aren’t entirely the same; there is a fine point which is data mining but it isn’t just searching
and gaining knowledge, it is to obtain knowledge by analysing and reconsidering [14].
72
International Journal in Foundations of Computer Science & Technology (IJFCST), Vol. 3, No.6, November 2013
Many people consider data mining as equivalent of another commonly used term, Knowledge
Discovery from Data, or the KDD. Alternatively, others view data mining as simply an
indispensable step in the process of knowledge discovery [12].
Typically, the steps of knowledge discovery (KD) are divided into 7 phases. Four primary steps
are pre-processing data and different forms of data preparation for data mining. As mentioned
above, data mining as a fifth steps, is one of the necessary steps of this process. Finally, two last
steps have the task of identifying useful patterns and displaying it to user. Below there is a brief
description of these steps:
Data cleaning: clean incompatible data and noises.
Data integration: integrate multiple sources, if any.
Data selection: restore data related to analysis and evaluate database.
Data transformation: transform or modulate data into searchable forms by summarizing
or aggregating.
Data mining: process of applying intelligent algorithms to exploiting data patterns.
Pattern evaluation: identify related patterns.
Knowledge presentation: present extracted knowledge to users using presentation
techniques and visual modeling.
Therefore data mining is a step in the KDD process consisting of a particular enumeration of
patterns over the data, subjected to some computational limitations. The term pattern goes beyond
its traditional concept and notion to include structures or models in the data warehouses.
Historical data are used to discover regularities and improve future decisions [15].
The goals of data mining are to divide it into two general types as predictive and descriptive [12].
By application of the predictive type, we understand that application of the model which
attempts to ‘predict’ the value that a certain variable may take ,given what we know at present
[16]. Descriptive data mining characterizes the generic attributes of the data in the database [12].
3. D ATA MINING IN H EALTHCARE
Today, in medical areas, data collection about different diseases is very important. Medical
centers with different purposes are to collect these data. To survey these data and to obtain useful
results and patterns in relation to disease is one of the objectives of the use of these data.
Collected data volume is very high and we must use data mining techniques to obtain desired
patterns and results among massive volume of data.
Medical and health areas are one of the most important sections in industrial societies. The
extraction of knowledge among massive volume of data related to diseases and people medical
records using data mining process can be lead to identifying the laws governing the creation,
development and epidemic diseases. We can allow expertise and health area staffs to access these
valuable data according to the environmental factors in order to identify diseases causes,
anticipate and treat the diseases. Finally, this means extending the life and comforting people in
the community.
For example some medicine applications of data mining are:
Predicting health care costs.
Determination of disease treatment.
Analyzing and processing medical images.
73
International Journal in Foundations of Computer Science & Technology (IJFCST), Vol. 3, No.6, November 2013
Analyzing health information system (HIS).
Effect of drugs on the disease and its side effects.
Predicting the success rate of medicine operations like surgeries.
Diagnosing and predicting of most kind of diseases such as cancer.
In a study published in 2011, Cheng-Ding Chang et al. used data mining techniques for predicting
the incidental multi-disease (hypertension and hyperlipidemia). They presented a two-phase
analysis procedure to simultaneously predict both hypertension and hyperlipidemia. Firstly they
chose six data mining approaches, and then they used these algorithms to select the individual
risk factors of these two diseases, afterward determined the common risk factors using the voting
principle. Next, they used the Multivariate Adaptive Regression Splines (MARS) method to
construct a multiple predictive model for hypertension and hyperlipidemia. This study used data
from a physical examination center database in Taiwan that included 2048 subjects. The proposed
multi-disease predictor method in this study has a classification accuracy rate of 93.07% [17].
Also Ömer Eskidere and two other colleagues studied the performance of Support Vector
Machines (SVM), Least Square Support Vector Machines (LS-SVM), Multilayer Perceptron
Neural Network (MLPNN), and General Regression Neural Network (GRNN) regression
methods in application to remote tracking of Parkinson’s disease progression. It is found that
using LS-SVM regression with log-transformation, the estimated motor-UPDRS is within 4.87
points (out of 108) for log-FS-4 feature set and total-UPDRS is within 6.18 points (out of 176) for
Log-FS-4 feature set of the clinicians’ estimation [18].
In one of the last-mentioned ascertainments in current year (2013), Daniel Ansari et al. enquired
about Artificial Neural Networks (ANNs) as nonlinear pattern recognition techniques that can be
used as a tool in medical decision making. Based on ANNs, by using clinical and
histopathological data from 84 patients a flexible nonlinear survival model was designed. After
the use of ANNs for predicting survival in pancreatic cancer, the C-index was 0.79 for the ANN
and 0.67 for Cox regression [19].
4. P ROPOSED METHODS AND IMPLEMENTATION
Data mining has wide and common uses in classification and recognition of the problems of
health systems. In this section, we describe the materials and methods which are used in current
research. Depending on the situation and desired outcome, there are various types of data
mining algorithms that you can use. All experiments described in this paper were performed using
libraries from Weka 3.7.9 machine learning environment. Before building models, the data set
were randomly split into two subsets: 75% (n=4406) of the data for training set and 25%
(n=1469) of the data for test set and all of data mining methods apply to total-UPDRS property.
4.1. Data Set
Parkinson’s telemonitoring dataset was created by Athanasios Tsanas and Max Little of the
University of Oxford, in collaboration with 10 medical centers in the US and Intel Corporation
which developed the AHTD telemonitoring device to record the speech signals [6]. Dataset has
been made available online at UCI machine learning archive very recently in October 2009. This
dataset consists of around 200 recordings per patient from 42 people (28 men and 14 women)
with early-stage PD which makes a total of 5875 voice recordings. Each patient is recorded with
phonations of the sustained vowel /a/. The dataset contains 22 attributes including subject
number, subject age, subject gender, time interval from baseline recruitment data, motor-UPDRS,
total-UPDRS, and 16 biomedical voice measures (vocal features).The vocal features in the dataset
74
International Journal in Foundations of Computer Science & Technology (IJFCST), Vol. 3, No.6, November 2013
are diverse, some of them based on traditional measures (Jitter, shimmer, HNR and NHR), and
some (RPDE, DFA, PPE) are based on nonlinear dynamical systems theory. In the dataset,
Motor-UPDRS and total-UPDRS were assessed at baseline (onset of trial), three-and-six-month
Table 1. Description of the features and UPDRS scores of the Parkinson’s telemonitoring dataset.
trial periods, but the voice recordings were obtained at weekly intervals. Hence both MotorUPDRS and total-UPDRS linearly interpolated [6]. Table.1 represents baseline, after three and six
months, UPDRS scores with the feature labels and short explanations for the measurement along
with some basic statistics of the dataset.
4.2. Functions
4.2.1. Simple Linear Regression (SLR)
The simple linear regression model y = α + βx + e, with measurement error (errors-in-variables
model) has been subject to extensive research in the statistical literature for over a century. It is
well known that the simple linear regression model with error in the predictor x is not identifiable
without any extra information in the normal case. In statistics, simple linear regression is
the least squares estimator of a linear regression model with a single explanatory variable. In
other words, simple linear regression fits a straight line through the set of n points in such a way
that makes the sum of squared residuals of the model (that is, vertical distances between the
points of the data set and the fitted line) as small as possible [20].
So in the simple linear regression model, the observations yi are generated according to:
yi xi ei
, i=1, …, n.
75
International Journal in Foundations of Computer Science & Technology (IJFCST), Vol. 3, No.6, November 2013
where α is the intercept parameter, β the slope parameter, xi the explanatory variables and
finally ei the error terms. If α=0 we obtain the simple linear regression model without intercept.
yi xi ei
, i=1, …, n.
You can see a simple linear regression model in Figure.1 that shows this model’s parameters.
Figure 1. Simple Linear Regression Model.
4.2.2. Multi-Layer Perceptron (MLP)
An MLP is based on single perceptron that was introduced by Rosenblatt [21].The network
contains an input layer, a hidden layer and an output layer. Each neuron at one layer receives a
weighted sum from all neurons in the previous layer and provides an input to all neurons of the
later layer [22].
Figure 2. MLP structure that used.
However there is merely a single neuron in the output layer. The architecture of the Neural
Network, used in this research, is the multilayer perceptron network architecture with 22input
nodes, 13 hidden nodes in our approach. We showed the MLP structure used in this study in.
The number of input nodes is determined by the ultimate data; the number of hidden nodes is
determined through trial and error; and the number of output nodes is represented as a rage,
demonstrating the disease classification. Each neuron at one layer receives a weighted sum from
all neurons in the previous layer and provides an input to all neurons of the later layer.
76
International Journal in Foundations of Computer Science & Technology (IJFCST), Vol. 3, No.6, November 2013
4.2.3. SMOreg
The Sequential Minimal Optimization (SMO) algorithm on classification tasks defined on sparse
data sets has been shown to be an effective method for training SVM. SMOreg implements a
sequential minimal optimization algorithm for training a support vector regression model. This
implementation globally replaces all missing values and transforms nominal attributes into binary
ones [23]. It also normalizes all attributes by default. We ‘normalized’ training dataset used the
‘polynomial’ kernel: K(x, y) = <x, y>^p or K(x, y) = (<x, y>+1) ^p for kernel and used
‘RegSMO’ to optimizing.
4.3. Rules
4.3.1. M5Rules
Generate a decision list for regression problems used to divide and to conquer. It builds a model
tree using M5 and makes the “best” leaf into a rule in each iteration of progress. The method for
generating rules from model trees, called M5-Rules, is straightforward and works as follows; a
tree learner is applied to the full training dataset and a pruned tree is learned. Then, the best
branch is made into a rule and the tree is discarded.
All instances covered by the rule are removed from the dataset. The process is applied recursively
to the remaining instances and terminates when all instances are covered by one or more rules.
This is a basic divide-and-conquer strategy for learning rules; however, instead of building a
single rule, as it is usually done, we built a full model tree at each stage, and make its “best”
branch into a rule. In contrast to PART (partial decision trees), which employs the same strategy
for categorical prediction, M5-Rules builds full trees instead of partially explored trees. Building
partial trees leads to a greater computational efficiency, and does not affect the size and accuracy
of the resulting rules [24, 25, 26, 27 and 28]. In simulating this mode we set the minimum number
of instances to allow at a leaf node to 4.0.
4.3.2. Decision Table
The Decision Table (DT) is a symbolic way to express the test experts or design experts
knowledge in a compact form [29].A decision table includes a hierarchical table. These tables
consider that each entry in a higher level table gets broken down by the values of a pair of
additional attributes to form another table. The structure is similar to dimensional stacking [30].
For DT implementation we used “Best First” method to find good attribute combinations for the
decision table. For example in Table.2, the final row tells us to classify the applicant as Heart
Failure (HF) if gender is Female, age is more than 42 and blood pressure isn’t important.
77
International Journal in Foundations of Computer Science & Technology (IJFCST), Vol. 3, No.6, November 2013
4.4. Trees
4.4.1. M5P
This is an algorithm for generating M5 model trees [31]. M5 builds a tree for a given instance, to
predict numeric values. When the input attributes can be either discrete or continuous the
algorithm requires the output attribute to be numeric. For a given instance the tree is traversed
from top to bottom until a leaf node is reached. At each node of the tree a decision is made to
follow a particular branch based on a test condition on the attribute associated with that node.
Each leaf node has a linear regression model associated with the form:
w0 w1 a1 ... wk a k
Based on some of the input attributes a1 , a 2 ,..., a k in the instance and whose respective weights
w0 , w1 ,..., wk are calculated using standard regression. The tree is called a model tree as the leaf
nodes contain a linear regression model to generate the predicted output [32]. We start with a set
of training instances to build a model tree, using the M5 algorithm. The tree is built using a
divide-and-conquer method. The M5P Model Tree algorithm in WEKA is available in the Java
class ‘‘weka.classifiers.trees.M5P”. The minimum number of instances to allow at a leaf node
was 2.0 and after simulation number of rules was 170.
4.4.2. REPTree
This algorithm is a fast decision tree learner. It is also based on C4.5 algorithm and can cause
classification (discrete outcome) or regression trees (continuous outcome). It uses information
gain/variance to building a regression/decision tree using and prunes it using reduced-error
pruning (with back-fitting). Only sorts values for numeric attributes once. Missing values are
dividing by splitting the corresponding instances into pieces [33].
In implement of REPTree we set the maximum tree depth to -1, it means there was no restriction.
The minimum proportion of the variance on all the data that needs to be presented at a node was
0.001, which is used for splitting, to be performed in regression trees. After implementation we
have a tree whose size is 675.
4.4.3. Decision Stump
The decision stump is the simplest special case of a decision tree. Each decision stump consists of
a single decision node and two prediction leaves. Each decision node is a rule that checks the
presence or absence of specified n-gram of command sequence. The boosted decision stumps are
created using a weighted voting of the decision stumps [34]. It’s usually used in conjunction with
a boosting algorithm, does regression or classification. Missing is treated as a separate value.
4.5. Lazy
4.5.1. IBk
This is the K-nearest neighbour classifier. We used IB1 which is the simplest instance-based
learning algorithm. It gives a test instance, and to find the training instance closest to the given
test instance uses a simple distance measure, and predicts the same class as this training instance.
If more than one instance is the same smallest distance (multiple) to the test instance, the first one
found is used.
78
International Journal in Foundations of Computer Science & Technology (IJFCST), Vol. 3, No.6, November 2013
The similarity function used here is:
n
Similarity ( x, y )
f (x , y )
i
i
i 1
Where the instances are described by n attributes and it’s 22 in our case. We
define f ( xi , yi ) ( xi yi ) 2 for numeric-valued attributes and f ( xi , y i ) ( xi y i ) for Boolean
and symbolic-valued attributes. The difference of IBI and nearest neighbour algorithm is that IBI
normalizes its attributes ranges, processes instances incrementally, and unlike nearest neighbour
algorithm has a simple policy for tolerating missing values [35]. In this model we used “Linear
NN Search” algorithm for searching nearest neighbour which is implementing the brute force to
search algorithm for nearest neighbour search in 1 nearest neighbour(s) for classification.
4.5.1. LWL
Lazy learning methods defer training data process until a query needs to be answered. This
algorithm to assign instance weights uses an instance-based. A good choice for classification is
Naive Bayes. This implementation surveys a form of lazy learning, locally weighted learning, that
uses locally weighted training to average, interpolate between, extrapolate from, or otherwise
combine training data [36]. In implementation of LWL model we used all of the neighbours in
weighting function which is linear function and “Linear NN Search” algorithm for searching
nearest neighbour which is implementing the brute force to search algorithm with Euclidean
distance (or similarity) function.
4.6. Meta
4.6.1. Regression by Discretization
A class for a regression scheme that utilizes any distribution classifier on a version of the data that
has the class attribute discretized. The predicted value based on the predicted probabilities for
each interval is the expected value of the mean class value for each discretized interval. This class
now also supports conditional density estimation by building a univariate density estimator from
the target values in the training data, weighted by the class probabilities. To implement this
model, J48 with C4.5 algorithm was adopted due to it abilities to apply negotiation strategies; and
set Number of bins for discretization to 10 and estimator type for density estimating was
histogram estimator. Our created J48 pruned tree size is 315 and it has 158 leaves node.
5. R ESULTS AND D ISCUSSION
We report the results from eleven methods used to study the usefulness of the data mining
algorithms for remote tracking of Parkinson’s disease progression. The data set, used in the
detection of PD progression with eleven data mining methods, was divided into two subsets as
Training and Test data. There were a total of 4406 pieces of data, exclusively used for training
and 1469 pieces of data exclusively used for testing. The classification accuracies obtained by this
study for PD dataset are presented in Table.3 You can compare and assimilate results of the
algorithms in correlation coefficient, Mean absolute error, Root mean squared error, relative
absolute error (in percentage) and root relative squared error (in percentage). In statistics,
the Mean Absolute Error (MAE) is a quantity used to measure how close forecasts or predictions
are to the eventual outcomes and the Root Mean Square Error (RMSE) is a frequently used
measure of the differences between values predicted by a model or an estimator and the values
actually observed.
79
International Journal in Foundations of Computer Science & Technology (IJFCST), Vol. 3, No.6, November 2013
Table 3. Experiential results of classification models.
Figure.3 summarized correlation coefficient of each algorithm in categories. We notice Decision
Table gave higher accuracy of correlation coefficient on our dataset (99.85%) in Rules category
and M5Rules algorithm has comparatively a good accuracy equivalent to 99.67%.
Figure 3. Correlation coefficient of classification models.
The relative absolute error is very similar to the relative squared error in the sense that it is also
relative to a simple predictor, which is just the average of the actual values. In this case, though,
the error is just the total absolute error instead of the total squared error. Thus, the relative
absolute error takes the total absolute error and normalizes it by dividing by the total absolute
error of the simple predictor and the root relative squared error is relative to what it would have
been if a simple predictor had been used. More specifically, this simple predictor is just the
average of the actual values. Thus, the relative squared error takes the total squared error and
normalizes it by dividing by the total squared error of the simple predictor. By taking the square
root of the relative squared error one reduces the error to the same dimensions as the quantity
being predicted. In Figure.4 we show these two parameters in a chart.
80
International Journal in Foundations of Computer Science & Technology (IJFCST), Vol. 3, No.6, November 2013
Figure 4. Relative absolute error and Root relative squared error of classification models.
5. C ONCLUSION AND F UTURE WORKS
This experimental study compares classification performance of different eleven data mining
algorithms via using of a Parkinson’s Telemonitoring dataset. This dataset comprises of 22
attributes with various ranges of values. Early detection of any kind of disease is an important
factor and for PD, remote tracking of UPDRS using voice measurements would facilitate clinical
monitoring of elderly people and increase the chances of its early diagnosis.
Used data mining techniques are Simple Linear Regression (SLR), Multi-Layer Perceptron
(MLP), SMOreg, M5Rules, Decision Table, M5P, REPTree, Decision Stump, IBk, LWL and
Regression by Discretization models. The best approach for remote Parkinson’s Telemonitoring
is Decision Table model also M5Rules in Rules category has a good result with correlation
coefficient of 0.9967. Mathematical equations from different data mining techniques for
calculating experimental results were derived.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
Tsanas, A., Little, M. A., McSharry, P. E., & Ramig, L. O. (2010). Enhanced classical dysphonia
measures and sparse regression for telemonitoring of Parkinson’s disease progression. International
Conference on Acoustics, Speech and Signal Processing, 594–597.
Lang, A. E., & Lozano, A. M. (1998). Parkinson’s disease – First of two parts. New England Journal
Medicine, 339, 1044–1053.
Höglinger GU, Rizk P,MurielMP, Duyckaerts C, OertelWH, Caille I, et al. Dopamine depletion
impairs precursor cell proliferation in Parkinson disease. Nat Neurosci 2004; 7:726–35.
Sakar, C. O., & Kursun, O. (2009). Telediagnosis of Parkinson’s disease using measurements of
dysphonia.Journal of Medical Systems., 34(4), 591–599.
Skodda, S., & Schlegel, U. (2008). Speech rate and rhythm in Parkinson’s Disease. Movement
Disorders, 23, 985–992.
Tsanas, A., Little, M. A., McSharry, P. E., & Ramig, L. O. (2010). Accurate telemonitoring of
Parkinson’s disease progression by non-invasive speech tests. IEEE Transactions on Biomedical
Engineering, 57, 884–893.
Little, M. A., McSharry, P. E., Hunter, E. J., & Ramig, L. O. (2009). Suitability of dysphonia
measurements for telemonitoring of Parkinson’s disease. IEEE Transactions on Biomedical
Engineering, 56(4), 1015–1022.
81
International Journal in Foundations of Computer Science & Technology (IJFCST), Vol. 3, No.6, November 2013
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
Skodda, S., Rinsche, H., & Schlegel, U. (2009). Progression of dysprosody in Parkinson’s disease
over time – A longitudinal study. Movement Disorders, 24(5), 716–722.
Aziz, T. Z., Peggs, D., Sambrook, M. A., & Crossman, A. R., Lesion of the subthalamic nucleus
for the alleviation of 1-methyl-4-phenyl-1,2,3,6-tetrahydropyridine (MPTP)-induced parkinsonism
in the primate. Movement Disorders, 6(4), 228–292, 1992.
A.A. Walter, “Data Mining Industry: Emerging Trends and New Opportunities”, Massachusetts
Institute of Technology, pp. 13-15, 2000.
Kantardzic, Mehmed. Data Mining: Concepts, Models, Methods, and Algorithms. John Wiley &
Sons, 2003.
H. Jiawei and K. Micheline, Data Mining: Concepts and Techniques, vol. 2, Morgan Kaufmann
Publishers, 2006.
Zafarani,R, Jashki, M.A, Baghi,H.R , Ghorbani,A., 2008, A Novel Approach for Social Behavior
Analysis of the Blogosphere, springer-Verlag Berlin Heidelberg, S. Bergler (Ed.): Canadian AI,
356–367.
F. S. Gharehchopogh, Peyman Mohammadi and Parvin Hakimi. Application of Decision Tree
Algorithm for Data Mining in Healthcare Operations: A Case Study. International Journal of
Computer Applications 52(6):21-26, August 2012.
T. M. Mitchell, "Machine learning and data mining," Communications of the ACM, vol.42, pp.30-36,
1999.
Florin Gorunescu, Data Mining: Concepts, Models and Techniques, Intelligent Systems Reference
Library, Vol. 12, Springer Publication, 2011.
Cheng-Ding Chang, Chien-Chih Wang, Bernard C. Jiang, Using data mining techniques for multidiseases prediction modeling of hypertension and hyperlipidemia by common risk factors, Expert
Systems with Applications, Volume 38, Issue 5, May 2011, Pages 5507-5513, ISSN 0957-4174.
Ömer Eskidere, Figen Ertaş, Cemal Hanilçi, A comparison of regression methods for remote tracking
of Parkinson’s disease progression, Expert Systems with Applications, Volume 39, Issue 5, April
2012, Pages 5523-5528, ISSN 0957-4174.
Daniel Ansari, Johan Nilsson, Roland Andersson, Sara Regnér, Bobby Tingstedt, Bodil Andersson,
Artificial neural networks predict survival from pancreatic cancer after radical surgery, The American
Journal of Surgery, Volume 205, Issue 1, January 2013, Pages 1-7, ISSN 0002-9610,
10.1016/j.amjsurg.2012.05.032.
Magne Thoresen, Petter Laake, On the simple linear regression model with correlated measurement
errors, Journal of Statistical Planning and Inference, Volume 137, Issue 1, 1 January 2007, Pages 6878, ISSN 0378-3758.
K., Warwick, Artificial intelligence: The basics. Taylor & Francis, 2011.
F. S. Gharehchopogh and Peyman Mohammadi. A Case Study of Parkinson’s Disease Diagnosis
using Artificial Neural Networks. International Journal of Computer Applications 73(19):1-6, July
2013.
Alex J. Smola, Bernhard Scholkopf (1998). A Tutorial on Support Vector Regression. NeuroCOLT2
Technical Report Series - NC2-TR-1998-030.
Goodwin L, VanDyne M, Lin S, Talbert S. Data mining Issues and opportunities for building nursing
knowledge. J Biomed Inform 2003;36:379–88.
Braha D, Shmilovici A. Data mining for improving a cleaning process in the semiconductor industry.
IEEE Trans Semiconductor Manuf 2002;15.
Fayyad UM, Uthurusamy R. Evolving data mining into solutions for insights. Commun ACM
2002;45:28–31.
Zhou ZH. Three perspectives of data mining. Artif Intell 2003;143:139–46.
Mattison R. Data warehousing: strategies, technologies and techniques statistical analysis, SPSS Inc.
WhitePapers; 1997.
R. Welland, Decision Tables and Computer Programming, Heyden & Son Inc., 1981.
Kohavi, R. The power of decision tables, in: Lavrac, N., Wrobel, S., (Eds.), Machine Learning:
Proceedings of the Eighth European Conference on Machine Learning ECML95, Lecture Notes in
Artificial Intelligence, Springer Verlag, 914, Berlin, Heidelberg, NY, pp. 174–189, 1995.
Qinlan, J. R. (1992). Learning with continuous classes. InProceedings of the Australian joint
conference on artificial intelligence (pp. 343–348).
Polumetla, A. (2006).Machine learning methods for the detection of RWIS sensor malfunctions.
Master Thesis. University of Minnesota (p. 30).
82
International Journal in Foundations of Computer Science & Technology (IJFCST), Vol. 3, No.6, November 2013
[33] Mohmad Badr Al Snousy, Hesham Mohamed El-Deeb, Khaled Badran, Ibrahim Ali Al Khlil, Suite of
decision tree-based classification algorithms on cancer gene expression data, Egyptian Informatics
Journal, Volume 12, Issue 2, July 2011, Pages 73-82, ISSN 1110-8665.
[34] Yoav F., Robert ES. A decision theoretic generalization of on-line learning and an application to
boosting. Journal of Computer and System Sciences 1997; 55(1):119–39.
[35] Aha, D., and D. Kibler (1991) "Instance-based learning algorithms", Machine Learning, vol.6, pp. 3766.
[36] Vapnik, V. & Bottou, L. (1993). Local algorithms for pattern recognition and dependencies
estimation. Neural Computation5 (6): 893–909.
Authors
Peyman Mohammadi is an M.Sc. student in Computer Engineering Department,
Science and Research Branch, Islamic Azad University, West Azerbaijan, Iran. His
interested research areas are in the Artificial Neural Networks, Data Mining and
Machine Learning techniques.
83
| 5cs.CE
|
arXiv:1705.10998v1 [cs.AI] 31 May 2017
The Atari Grand Challenge Dataset
Vitaly Kurin
Visual Computing Institute
RWTH Aachen University
Aachen, Germany
[email protected]
Sebastian Nowozin, Katja Hofmann
Machine Intelligence and Perception Group
Microsoft Research
Cambridge, UK
{Sebastian.Nowozin, Katja.Hofmann}@microsoft.com
Lucas Beyer, Bastian Leibe
Visual Computing Institute
RWTH Aachen University
Aachen, Germany
{beyer, leibe}@vision.rwth-aachen.de
Abstract
Recent progress in Reinforcement Learning (RL), fueled by its combination, with
Deep Learning has enabled impressive results in learning to interact with complex
virtual environments, yet real-world applications of RL are still scarce. A key
limitation is data efficiency, with current state-of-the-art approaches requiring
millions of training samples. A promising way to tackle this problem is to augment
RL with learning from human demonstrations. However, human demonstration
data is not yet readily available. This hinders progress in this direction. The
present work addresses this problem as follows. We (i) collect and describe a large
dataset of human Atari 2600 replays – the largest and most diverse such data set
publicly released to date (ii) illustrate an example use of this dataset by analyzing
the relation between demonstration quality and imitation learning performance,
and (iii) outline possible research directions that are opened up by our work.
1
Introduction
In Reinforcement Learning (RL), an agent learns, by trial and error, to perform a task in an initially
unknown environment. Recently, this research area has seen dramatic progress in complex interactive
tasks in virtual environments Mnih et al. [2015, 2016], Silver et al. [2016], Schulman et al. [2015],
largely driven by combinations of RL with deep learning. Yet, despite this recent progress, real-world
applications are still largely lacking.
In the RL setup, agents learn to solve a task completely from scratch. This causes one of the key
limitations of state-of-the-art deep RL approaches — data inefficiency. In comparison to autonomous
agents, humans have a lot of prior information about the world. Every day, we base our decisions on
knowledge about culture and social relationships, our own experience, and information we get from
experience of others. As a result, when learning a new task, people are more effective and need less
actions to master it. During training, the agent executes a lot of actions that a human never would.
This is ineffective and renders the application of RL in complex or potentially dangerous environments
infeasible. A possible solution to the problem is to learn from human demonstrations Schaal [1997],
Ng and Russell [2000], Abbeel and Ng [2004], Monfort et al. [2017], Hester et al. [2017].
Up until now, the RL community has been focused on building environments — test beds for RL
models. Today, RL approaches can be trained and compared on diverse tasks in environments
such as ALE Bellemare et al. [2013], OpenAI Gym Brockman et al. [2016] or Microsoft’s Project
Malmo Johnson et al. [2016]. But there are few publicly available datasets of human demonstrations
of tasks in these environments. This lack hampers the progress of research on learning from
human demonstration. Examples such as ImageNet Deng et al. [2009] in Computer Vision and
Switchboard Godfrey et al. [1992] in Speech Recognition have shown that datasets can catalyze
research progress. In order to accelerate research in learning from demonstration, we release, describe,
and illustrate the use of the Atari Grand Challenge dataset1 — a dataset of human Atari 2600 replays.
Our contributions are: (i) We collect, analyze, and release to the research community the largest
and most diverse dataset of human Atari 2600 replays to date. The dataset comprises ∼9.7 million
frames (∼45 hours) of game play for five games - an order of magnitude larger than previous datasets.
(ii) We illustrate one use of this dataset by analyzing the relation between demonstration quality and
imitation learning performance. (iii) We discuss research directions that are opened up by our work.
2
Background
This section outlines key concepts and notation used throughout the paper. We operate in a usual
Reinforcement Learning (RL) setup: an agent acts in an environment in response to state observations,
and learns from a reward signal that reflects an abstract notion of consequences of the actions taken.
This setup can be formulated as Markov Decision Process (MDP), defined by a tuple hS, A, R, T i,
where S is the set of states, A is the set of actions, R(s, a) is the reward function, and T (s, a, s0 ) is
the transition function that returns a probability over states, given a state and an action: p(s0 |s, a),
s, s0 ∈ S, a ∈ A. In each iteration of its interaction with the environment, the agent observes the
state, takes an action and gets some reward for the transition. An agent’s behavior is characterized by
a policy π(s), a function that returns an action given a state. The policy can be stochastic.
Our experimental analysis (section 5), is based on a recently proposed imitation algorithm suggested
in Hester et al. [2017], in turn based on Q-learning Watkins and Dayan [1992] and in particular
DDQN Van Hasselt et al. [2016]. We next give an overview of the relevant approaches.
Q-learning Watkins and Dayan [1992] is a “model-free” RL algorithm. It centers on learning
to approximate the so-called action-value function Q(s, a). The Q-function reflects the expected
discounted cumulative value of taking a particular action a in state s, and following the particular
policy thereafter. The optimal action-value function Q∗ (s, a) should satisfy the Bellman equation:
Q∗ (s, a) = Es0 R(s, a) + γ max
Q∗ (s0 , a0 )|s, a ,
0
a ∈A
(1)
where γ is a discount factor that trades off immediate versus longer term rewards. The optimal policy
π ∗ (s) is the policy which takes the best possible decision on each time step:
π ∗ (s) = argmax Q∗ (s, a).
(2)
a∈A
DQN Mnih et al. [2015] is a variant of Q-learning that uses a neural network (called Deep Q-Network)
to approximate Q-values. The network Q(s, ·; θ) returns action values for all the actions available
given the current state. A separate target network is used to compute Q-values in training updates, as
well as a so-called replay memory of past experience for minibatch sampling. Both were shown to
improve training stability and resulted in breakthrough results when learning to play Atari games.
Double DQN Van Hasselt et al. [2016] is an extension of DQN which decouples the selection and
value estimate of actions in the max operator, which was shown to result in more accurate Q-value
approximations both theoretically and in practice. The learning objective looks as follows:
2
JDQ (Q) = R(s, a) + γQ(s0 , a0max ; θ0 ) − Q(s, a; θ) ,
where a0max = argmaxa0 ∈A Q(s0 , a0 ; θ).
1
http://atarigrandchallenge.com/data
2
(3)
Recent work Hester et al. [2017] suggests an approach to imitation learning that combines the Double
DQN objective and a large margin classification loss aimed at keeping a learned policy close to
demonstrated behavior:
J(Q) = JDQ (Q) + λ1 JE (Q) + λ2 JL2 (Q),
(4)
where JL2 (Q) is the L2 regularization, JE (Q) is the supervised learning loss:
JE (Q) = max Q(s, a) + l(s, aE , a) − Q(s, aE ),
a∈A
(5)
and l(s, aE , a) is the large margin classification loss that returns some positive number if the expert’s
action aE 6= a and zero otherwise. The large margin classification loss prevents the learner from
over-estimating Q-values for previously unseen states. The Q-values for actions taken by an expert
are forced to be a margin higher than those of the unseen ones.
All the methods described above can also be applied completely off-line on data collected in the
process of a human’s interaction with the environment.
3
Constructing the Atari Grand Challenge dataset
This section details our approach to collecting the Atari Grand Challenge dataset. All described tools
are made public with the data set.
3.1
Collecting the dataset
We collected our dataset using a web application built around Javatari2 , an Atari 2600 emulator
written in JavaScript. Given an initial state and the full sequence of human inputs, the emulator is
completely deterministic, and we can avoid the excessive burden of saving screenshots of the game at
every single time step while it is being played. Instead, we can only record the initial state and player
inputs, and generate the dataset of images offline by playback. This makes data collection at large
scale feasible with limited resources.
After processing, for each time step we have a screenshot of the game, the action taken at that time
step, the reward, the current score, and the information about the time step being terminal or not.
Since incomplete (non-terminal) episodes still carry useful information, we save the episode each
time a player closes the application tab or browser window as well as when the game ends.
All the Atari-related functionality is entirely processed on the client side — within a browser. We
support all the major web browsers: Google Chrome, Mozilla Firefox, Microsoft Edge, Safari. The
server is built using Flask3 . It is only responsible for saving the data and, later, for loading the data
when replaying. All the data is saved in a PostgreSQL database. The replay process is automated
with Selenium4 .
Our case is a good example of gamified crowdsoursing — using people’s desire to play to do
useful things. In order to engage people more, we added two progress bars: one compares players’
performance with the best human player result, the other shows the same comparison with DQN
performance taken from Mnih et al. [2015].
3.2
Dataset post-processing
There are two steps of dataset post-processing. First, we try to eliminate the differences between the
Javatari emulator and the ALE. The only difference we have found is that the states we get in Javatari
are vertically shifted by several pixels in comparison to ALE states. We eliminate this difference by
shifting the states as in ALE and padding them with zeroes at the bottom and top borders.
2
http://javatari.org/
http://flask.pocoo.org/
4
https://github.com/SeleniumHQ
3
3
Table 1: Atari Grand Challenge per-game statistics
episodes
frames
gameplay, hrs
worst score
best score
Space Invaders
Q*bert
Ms.Pacman
Video Pinball
Montezuma’s revenge
445
2,056,741
9.52
5
3,355
659
1,599,453
7.40
25
41,425
384
1,779,771
8.23
10
29,311
211
1,526,215
7.05
100
67,150
668
2,717,676
12.57
100
27,900
During the first frames that we record, the emulated Atari memory is not fully initialized and we
might get an excessively large score, which is not correct. We therefore fix the first several frames’
rewards to zero. We are not also interested in games where the person did not interact with the
application more than just opening and closing it. We filter these cases by simply removing games
with a final score of zero. The post-processing is fully automated, and new data can be processed
with the code provided.
4
Properties of the Atari Grand Challenge dataset
4.1
Description of the dataset
In this section we briefly describe what the Atari Grand Challenge dataset consists of and we
show some of its properties that we deem particularly relevant to research on learning from human
demonstrations.
4.1.1
Scale
The dataset consists of human replays for five popular Atari 2600 games: Video Pinball, Q*bert,
Space Invaders, Ms.Pacman and Montezuma’s Revenge. The choice of the games is not random: we
want to vary the level of difficulty according to the results in Mnih et al. [2015]. The DQN was able
to play the first game significantly better than human players, the results for the second and the third
were comparable to human performance and the latter two were very hard for DQN.
The Atari Grand Challenge dataset consists of 2367 game episodes with positive final score, that is
∼9.7 million frames or ∼45 hours of playing time at 60 frames per second. Table 1 shows per-game
statistics of the dataset and Fig. 1 shows sample screenshots.
Figure 1: Sample screenshots from the dataset (from left to right, the games shown are: Space
Invaders, Q*bert, Ms. Pacman, Video Pinball, Montezuma’s Revenge)
4.1.2
Diversity
Since the data is collected in the wild, some of the players were good, some of them were bad. As
a result, Fig. 2 shows that the Atari Grand Challenge dataset is quite diverse in terms of the final
score distribution. From the episode final score and the time played, we can already make some
assumptions about the different players’ level of expertise. What else can we do to show the player
diversity quantitatively?
It is natural to assume that the more experienced a player is, the more effective he will be. A new
player cannot achieve a challenging reward when an experienced player can. Fig. 3 shows that all the
players have equal access to the rewards (at least for the games in question). From the comparison of
“advanced” and “expert” groups we can see, that “expert” players are faster in achieving the rewards:
the rightmost column data points look more shifted to the left. Given that the final score for the
“advanced” group is higher, they achieve more in shorter periods of time.
4
Figure 2: Atari Grand Challenge dataset human demonstrators score dependency on time.
4.1.3
Extensibility
Currently the dataset comprises five Atari 2600 games, but it can be easily extended by adding a new
game. We do not only publish the dataset, but the code for data collection as well5 . Any of the Atari
2600 non-paddle games available in ALE Bellemare et al. [2013] can be added within few hours of
work. We do not support paddle games like Breakout or Pong since it is almost impossible to exactly
repeat the noisy controller to collect the states off-line.
5
5.1
Influence of data quality on the imitation learning performance
Experiments description
As shown in Section 4, we have replays of good players as well as replays of bad players. In this
section we show how the dataset can be used to study how the demonstrator expertise can influence
the performance of imitation learning.
In this experiment, we filter the training data by a minimum score. We train our model on the frames
of the episodes with final score above a threshold: 50 percentile, 75 percentile (top 25% of the data)
and 95 percentile (top 5%). We also train the model on the whole dataset.
We train the model completely off-line as in Eq. 5, but we do not use the regularization term
from Hester et al. [2017] since we have more data to train on. We use λ1 = 1.0 and l = 0.8 for
non-expert actions as suggested in the same paper. We use the same network architecture as in Mnih
et al. [2015] and train over 106 iterations. We use the Adam Kingma and Ba [2014] optimizer with
learning rate = 0.00025 and β1 = β2 = 0.95. The target network update interval is 10, 000,
mini-batch size is 32, and training is run for one million updates. The code is written in Chainer Tokui
et al. [2015] within the Malmopy6 framework. Since we do not have any frameskip during data
collection, we use frameskip coefficient k = 1. We normalize reward values by dividing the raw
rewards by the largest reward value observed in our data. There were no negative rewards in our case.
The experimentation code can be found on our github page7 .
5
Figure 3: Reward and the frame when the reward was obtained show that all the players have equal
access to rewards, but more experienced players are more efficient. All the game episodes are
divided as follows: “novice” (below 50 percentile of all the scores), “average” (between 50 and 75),
“advanced” (between 75 and 95) and “expert” (above 95 percentile).
5.2
Results
During training we evaluate the performance of the model on 100 episodes every 100,000 mini-batch
updates. After training we take the model with the best average results and re-evaluate it on 100 games
and report the average score and standard error of the mean. Table 2 is in line with our hypothesis:
the higher the filter value for the data, the better the performance. The sub-par performance of the
imitation model on our dataset in three out of five games can be explained by looking at Table 3: our
data has lower and more diverse human demonstrator scores than those of Hester et al. [2017]. At the
same time, in Video Pinball, where our data has better human scores, the model performs better.
5
https://github.com/yobibyte/atarigrandchallenge/
https://github.com/Microsoft/malmo-challenge
7
https://github.com/yobibyte/atarigrandchallenge/
6
6
Table 2: Average score ± standard error of the mean (out of 100 games) for models trained on subsets
of the data filtered by a score. The first four rows use the offline part of the imitation algorithm
(without regularization) from Hester et al. [2017] to train on our data. Top 5% means that the training
data consists of the episodes with a final score higher or equal than the 95 percentile score. Evaluation
is performed with the -greedy policy, = 0.05. Mnih et al. [2015] reports the standard deviation
of the scores, but we do not have information about the number of the episodes and thus we cannot
report their SEM.
Space
Invaders
Q*bert
Ms.Pacman
Video
Pinball
Montezuma’s
revenge
Imitation All data
Imitation top 50%
Imitation top 25%
Imitation top 5%
Imitation Hester et al. [2017]
125 ± 9.94
90 ± 8.64
127 ± 9.69
144 ± 12.40
n/a
146 ± 14.87
127 ± 13.80
179 ± 17.01
545 ± 107.19
5,133.8
250 ± 18.09
308 ± 20.66
271 ± 22.15
418 ± 19.98
692.4
8,823±745.26
11,216±801.53
24,351±2,084.38
17,775±16.10
10,655.5
7 ± 4.32
4 ± 1.97
22 ± 8.11
36 ± 7.98
576.3
DQN Mnih et al. [2015]
DDQN Van Hasselt et al. [2016]
Random uniform
1,976
2,628.7
156±9.46
10,596
11,020.8
162±17.6
2,311
1,241.3
211±11.35
42,684
367,823.7
30,368±3,723.87
0
42
0±0
Table 3: Comparison of human scores in Hester et al. [2017] and the Atari Grand Challenge dataset.
Hester et al. [2017]
Space Invaders
Q*bert
Ms. Pacman
Video Pinball
Montezuma’s revenge
6
Worst score
Best score
80700
31781
8409
32300
n/a
99450
55021
32420
34900
Atari Grand Challenge dataset
#transitions
Worst score
Best score
#transitions
75472
21896
10051
17949
5
25
10
100
100
3355
41425
29311
67150
27900
2,056,741
1,599,453
1,779,771
1,526,215
2,717,676
Related work
There are two directions of RL research which are working on leveraging demonstration data for
training an autonomous agent: Inverse Reinforcement Learning (IRL) and Imitation Learning. The
former group addresses scenarios where there is no access to the reward function. It is true that in RL
tasks the goal is often underspecified, and sometimes it is hard to provide a reward that represents all
the useful information from expert’s demonstration. The general idea is to approximate the reward
function and learn a policy using this approximation Ng and Russell [2000], Abbeel and Ng [2004].
Whilst IRL can benefit from the Atari Grand Challenge dataset by ignoring the reward information,
Imitation Learning is the direct benefactor of our dataset. Imitation learning exploits the reward
information to learn an action-value function, or directly a policy. Schaal [1997] uses a pre-trained
model to speed up training, and has an interesting comparison of pre-training influence on modelfree and model-based RL. The paper notes that model-based learning benefits more from using
demonstration data. The latest work on Learning from Demonstration shows that model-free RL can
also greatly benefit from using human player data Hester et al. [2017], Subramanian et al. [2016],
Hosu and Rebedea [2016].
The datasets collected for the learning from demonstration research described above are either small,
or not available for public use. The Atari Grand Challenge dataset is the largest and the most diverse
in terms of the types of the games as well as amount and types of human players release so far.
Up until now, the RL community has been mostly focusing on building the environments for training
autonomous agents: ALE Bellemare et al. [2013], OpenAI’s gym Brockman et al. [2016] and
Universe8 , Microsoft’s Malmo project Johnson et al. [2016]. The final goal of operating in these
environments is in maximizing the final score. Even if we train our models off-line, it sounds
reasonable to evaluate the performance within such an environment. That is why, in Table 4 below
we describe the datasets coupled with interactive environments.
8
https://github.com/openai/universe
7
Table 4: Learning from demonstration datasets comparison. The replay data for Hosu and Rebedea
[2016] has not been published, but there are Montezuma’s Revenge and Private Eye checkpoints —
saved states of the environment that can be used for continuing the episode.
Atari Grand Challenge
Udacity self-driving dataset9
Hosu and Rebedea [2016]
Hester et al. [2017]
Domain
Tasks
Size (transitions)
Atari 2600
Driving simulator
Atari 2600
Atari 2600
5
1
1
42
∼9.7 mil.
8086 (x3 cameras)
∼1.2 mil.
∼1 mil.
Open
Diverse in
player expertise
3
3
(7)
7
3
7
7
7
Atari 2600 games have recently begun to take a similar role as experimentation ground for RL
research as MNIST has taken for computer vision and many implementations of RL algorithms have
been evaluated on such games. Therefore, it is much easier to compare leveraging the human behavior
data with pure RL implementations or even combining them.
7
Discussion and Future work
Our work opens up a wide range of follow-up work on benefits and uses of human demonstrations
for effectively and efficiently learning to interact with complex environments.
7.1
Extending the Atari Grand Challenge dataset
The Atari Grand Challenge website is still on-line and people keep playing. We plan to update the
dataset in the future as more data becomes available. The most important development of the dataset
is to collect more data of “professional players” who achieve higher scores. As we have shown, the
data quality affects the final performance dramatically, it will be a good improvement, when we do
that.
7.2
Exploiting the Atari Grand Challenge dataset
Video games are a perfect testing ground for evaluating hypotheses and learning how we can use
human data to achieve higher sample efficiency and make the RL training process faster. So, our
future research will focus on improving sample efficiency of RL algorithms by leveraging the data of
diverse quality.
In this paper we have shown just one of the possible dataset applications: how data quality influences
the final performance of imitation learning Hester et al. [2017]. We hope that researchers in machine
learning, game AI and maybe even cognitive science, can find something useful for their own research
purposes. We find the following applications particularly appealing.
Recently, Inverse Reinforcement Learning and Imitation Learning have regained popularity Ho
and Ermon [2016], Baram et al. [2016], Hester et al. [2017]. Our dataset has a direct impact on
this kind of research. It is interesting to check if can we take something useful out of bad players
data. Even experienced players make mistakes. But throwing out this data can waste potentially
important information. Shiarlis et al. [2016] investigates this topic in the Inverse Reinforcement
Learning domain. It might be interesting to see a similar approach for Atari 2600 in Learning from
Demonstration domain.
The frameskip coefficient has been shown to be very important in RL Braylan et al. [2015], Sharma
et al. [2017], Lakshminarayanan et al. [2016]. It would be interesting to investigate frameskip, using
the human data that we can extract from the Atari Grand Challenge dataset.
Curriculum learning has proven useful in RL Leibfried et al. [2016]. Having data of players with
different expertise, we can investigate curriculum learning with respect to this.
We have also seen attempts to investigate how humans learn to play Atari Tsividis et al. [2017]. Our
dataset might be interesting for this kind of research.
9
https://github.com/udacity/self-driving-car-sim
8
In conclusion, we release a dataset of human Atari 2600 replays of five games, that is ∼9.7 Million
frames or ∼45 hours of game play time. We are describing its main properties, i.e. scale and diversity.
We show that in order to achieve high performance, it is more important to collect data of players
with a high level of expertise, than to collect a lot of low-skilled data. We plan to update the dataset in
the future by adding “professional” Atari 2600 players data. We release the code for data collection
as well, which gives the opportunity for everybody to extend the dataset. We also show some of the
possible research directions the Atari Grand Challenge dataset could be used in. We hope that our
release will catalyze the research in sample-efficiency of RL and learning from human demonstration.
Acknowledgments
Vitaly Kurin would like to thank Microsoft Research Cambridge for hosting him during the project
and for the Microsoft Azure for Research grant. The authors would also like to thank Paulo Peccin,
Javatari creator, for the emulator and useful discussions.
References
P. Abbeel and A. Y. Ng. Apprenticeship learning via inverse reinforcement learning. In Proceedings
of the twenty-first international conference on Machine learning, page 1. ACM, 2004.
N. Baram, O. Anschel, and S. Mannor. Model-based adversarial imitation learning. arXiv preprint
arXiv:1612.02179, 2016.
M. G. Bellemare, Y. Naddaf, J. Veness, and M. Bowling. The arcade learning environment: An
evaluation platform for general agents. Journal of Artificial Intelligence Research, 47:253–279,
jun 2013.
A. Braylan, M. Hollenbeck, E. Meyerson, and R. Miikkulainen. Frame skip is a powerful parameter
for learning to play atari. In Workshops at the Twenty-Ninth AAAI Conference on Artificial
Intelligence, 2015.
G. Brockman, V. Cheung, L. Pettersson, J. Schneider, J. Schulman, J. Tang, and W. Zaremba. Openai
gym, 2016.
J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical
image database. In IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255.
IEEE, 2009.
J. J. Godfrey, E. C. Holliman, and J. McDaniel. Switchboard: Telephone speech corpus for research
and development. In Acoustics, Speech, and Signal Processing, 1992. ICASSP-92., 1992 IEEE
International Conference on, volume 1, pages 517–520. IEEE, 1992.
T. Hester, M. Vecerik, O. Pietquin, M. Lanctot, T. Schaul, B. Piot, A. Sendonaris, G. Dulac-Arnold,
I. Osband, J. Agapiou, J. Z. Leibo, and A. Gruslys. Learning from demonstrations for real world
reinforcement learning. arXiv preprint arXiv:1704.03732, 2017.
J. Ho and S. Ermon. Generative adversarial imitation learning. In Neural Information Processing
Systems, pages 4565–4573, 2016.
I.-A. Hosu and T. Rebedea. Playing atari games with deep reinforcement learning and human
checkpoint replay. arXiv preprint arXiv:1607.05077, 2016.
M. Johnson, K. Hofmann, T. Hutton, and D. Bignell. The malmo platform for artificial intelligence
experimentation. In International Joint Conference on Artificial Intelligence (IJCAI), page 4246,
2016.
D. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980,
2014.
A. S. Lakshminarayanan, S. Sharma, and B. Ravindran. Dynamic frame skip deep Q network. CoRR,
abs/1605.05365, 2016. URL http://arxiv.org/abs/1605.05365.
9
F. Leibfried, N. Kushman, and K. Hofmann. A deep learning approach for joint video frame and
reward prediction in atari games. arXiv preprint arXiv:1611.07078, 2016.
V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King,
D. Kumaran, D. Wierstra, S. Legg, and D. Hassabis. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015.
V. Mnih, A. P. Badia, M. Mirza, A. Graves, T. P. Lillicrap, T. Harley, D. Silver, and K. Kavukcuoglu.
Asynchronous methods for deep reinforcement learning. CoRR, abs/1602.01783, 2016. URL
http://arxiv.org/abs/1602.01783.
M. Monfort, M. Johnson, A. Oliva, and K. Hofmann. Asynchronous data aggregation for training end
to end visual control networks. In International Conference on Autonomous Agents & Multiagent
Systems, pages 530–537. International Foundation for Autonomous Agents and Multiagent Systems,
2017.
A. Y. Ng and S. J. Russell. Algorithms for inverse reinforcement learning. In International Conference
on Machine Learning, pages 663–670, 2000.
S. Schaal. Learning from demonstration. Neural Information Processing Systems, pages 1040–1046,
1997.
J. Schulman, S. Levine, P. Moritz, M. I. Jordan, and P. Abbeel. Trust region policy optimization.
CoRR, abs/1502.05477, 2015. URL http://arxiv.org/abs/1502.05477.
S. Sharma, A. S. Lakshminarayanan, and B. Ravindran. Learning to repeat: Fine grained action
repetition for deep reinforcement learning. CoRR, abs/1702.06054, 2017. URL http://arxiv.
org/abs/1702.06054.
K. Shiarlis, J. Messias, and S. Whiteson. Inverse reinforcement learning from failure. In International Conference on Autonomous Agents & Multiagent Systems, pages 1060–1068. International
Foundation for Autonomous Agents and Multiagent Systems, 2016.
D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. Van Den Driessche, J. Schrittwieser,
I. Antonoglou, V. Panneershelvam, M. Lanctot, S. Dieleman, D. Grewe, J. Nham, N. Kalchbrenner,
I. Sutskever, T. Lillicrap, M. Leach, K. Kavukcuoglu, T. Graepel, and D. Hassabis. Mastering the
game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016.
K. Subramanian, C. L. Isbell Jr, and A. L. Thomaz. Exploration from demonstration for interactive
reinforcement learning. In International Conference on Autonomous Agents & Multiagent Systems,
pages 447–456. International Foundation for Autonomous Agents and Multiagent Systems, 2016.
S. Tokui, K. Oono, S. Hido, and J. Clayton. Chainer: a next-generation open source framework
for deep learning. In Proceedings of Workshop on Machine Learning Systems (LearningSys) in
The Twenty-Ninth Annual Conference on Neural Information Processing Systems, 2015. URL
http://learningsys.org/papers/LearningSys_2015_paper_33.pdf.
P. A. Tsividis, T. Pouncy, J. L. Xu, J. B. Tenenbaum, and S. J. Gershman. Human learning in
atari. "The AAAI 2017 Spring Symposium on Science of Intelligence: Computational Principles of
Natural and Artificial Intelligence", 2017.
H. Van Hasselt, A. Guez, and D. Silver. Deep reinforcement learning with double q-learning. In
Association for the Advancement of Artificial Intelligence, pages 2094–2100, 2016.
C. J. Watkins and P. Dayan. Q-learning. Machine learning, 8(3-4):279–292, 1992.
10
| 2cs.AI
|
Sparse Maximum-Entropy Random Graphs
with a Given Power-Law Degree Distribution
Pim van der Hoorn1 , Gabor Lippner2 , and Dmitri Krioukov1,2,3
1
Northeastern University, Department of Physics
Northeastern University, Department of Mathematics
3
Northeastern University, Departments of Electrical&Computer Engineering
arXiv:1705.10261v3 [math.PR] 10 Oct 2017
2
October 12, 2017
Abstract
Even though power-law or close-to-power-law degree distributions are ubiquitously observed in a great variety of large real networks, the mathematically satisfactory treatment
of random power-law graphs satisfying basic statistical requirements of realism is still
lacking. These requirements are: sparsity, exchangeability, projectivity, and unbiasedness. The last requirement states that entropy of the graph ensemble must be maximized
under the degree distribution constraints. Here we prove that the hypersoft configuration
model (HSCM), belonging to the class of random graphs with latent hyperparameters,
also known as inhomogeneous random graphs or W -random graphs, is an ensemble of
random power-law graphs that are sparse, unbiased, and either exchangeable or projective. The proof of their unbiasedness relies on generalized graphons, and on mapping the
problem of maximization of the normalized Gibbs entropy of a random graph ensemble,
to the graphon entropy maximization problem, showing that the two entropies converge
to each other in the large-graph limit.
Keywords: Sparse random graphs, Power-law degree distributions, Maximum-entropy
graphs
PACS: 89.75.Hc, 89.75.Fb, 89.70.Cf
MSC: 05C80, 05C82, 54C70
1
Contents
1 Introduction
1.1 Hypersoft configuration model (HSCM)
1.2 Properties of the HSCM . . . . . . . . .
1.3 Unbiasedness and the maximum-entropy
1.4 Main results . . . . . . . . . . . . . . . .
1.5 Exchangeability and projectivity . . . .
1.6 Other remarks . . . . . . . . . . . . . .
1.7 Paper organization . . . . . . . . . . . .
. . . . . . .
. . . . . . .
requirement
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3
3
4
5
7
8
9
10
2 Background information and definitions
2.1 Graph ensembles and their entropy . . . . . . . . . . . . . . . . . . . . . . . .
2.1.1 Maximum-entropy graphs with a given degree sequence (CM) . . . . .
2.1.2 Maximum-entropy graphs with a given expected degree sequence (SCM)
2.2 Sparse graphs with a given degree distribution . . . . . . . . . . . . . . . . .
2.3 Maximum-entropy graphs with hypersoft constraints . . . . . . . . . . . . . .
2.3.1 Graphon-based graph ensembles . . . . . . . . . . . . . . . . . . . . .
2.3.2 Bernoulli and graphon entropies . . . . . . . . . . . . . . . . . . . . .
2.3.3 Dense maximum-entropy graphs with a given degree distribution . . .
2.3.4 Rescaled graphon entropy of sparse graphs . . . . . . . . . . . . . . .
2.3.5 Sparse maximum-entropy graphs with a given degree distribution . . .
2.3.6 Sparse power-law hypersoft configuration model (sparse HSCM) . . .
10
10
11
11
11
12
12
13
13
15
15
16
3 Results
3.1 Main result . . . . . . . . . . . . . . . . . . . . . . .
3.2 The limit of the degree distribution in the HSCM . . .
3.3 The limit of the expected average degree in the HSCM
3.4 HSCM maximizes graphon entropy . . . . . . . . . . .
3.5 Graphon entropy scaling and convergence . . . . . .
3.6 Gibbs entropy scaling and convergence . . . . . . . .
.
.
.
.
.
.
16
16
17
18
18
19
20
.
.
.
.
.
.
.
.
.
.
.
21
21
26
27
29
31
32
32
34
36
36
37
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4 Proofs
4.1 The classical limit approximation of the Fermi-Dirac graphon . .
4.2 Proofs for node degrees in the HSCM . . . . . . . . . . . . . . . . .
4.2.1 Technical results on Poisson couplings and concentrations
4.2.2 Proof of Theorem 3.2 . . . . . . . . . . . . . . . . . . . .
4.2.3 Proof of Theorem 3.3 . . . . . . . . . . . . . . . . . . . .
4.3 Proofs for graphon entropy . . . . . . . . . . . . . . . . . . . . .
4.3.1 Proof of Proposition 3.4 . . . . . . . . . . . . . . . . . . .
4.3.2 Proof of Theorem 3.5 . . . . . . . . . . . . . . . . . . . .
4.4 Proof of Theorem 3.6 . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.1 Averaging W by a partition of An . . . . . . . . . . . . .
4.4.2 Constructing the partition . . . . . . . . . . . . . . . . . .
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
Introduction
Random graphs have been used extensively to model a variety of real networks. Many of these
networks, ranging from the Internet and social networks to the brain and the universe, have
broad degree distributions, often following closely power laws [1–3], that the simplest random
graph model, the Erdős-Rényi random graphs [4–6] with Poisson degree distributions, does
not reproduce. To resolve this disconnect, several alternative models have been proposed and
studied. The first one is the configuration model (CM), random graphs with a given degree
sequence [7, 8]. This model is a microcanonical ensemble of random graphs. Every graph in
the ensemble has the same fixed degree sequence, e.g., the one observed in a snapshot of a real
network, and every such graph is equiprobable in the ensemble. The ensemble thus maximizes
Gibbs entropy subject to the constraint that the degree sequence is fixed. Yet given a real
network snapshot, one cannot usually trust its degree sequence as some “ultimate truth” for
a variety of reasons, including measurement imperfections, inaccuracies, and incompleteness,
noise and stochasticity, and most importantly, the fact that most real networks are dynamic
both at short and long time scales, growing often by orders of magnitude over years [9, 10, 1–3].
These factors partly motivated the development of the soft configuration model (SCM),
random graphs with a given expected degree sequence, first considered in [11, 12], and later
corrected in [13–16], where it was shown that this correction yields a canonical ensemble of
random graphs that maximize Gibbs entropy under the constraint that the expected degree
sequence is fixed. In statistics, canonical ensembles of random graphs are known as exponential random graphs (ERGs) [17]. In [18, 19] it was shown that the sparse CM and SCM are
not equivalent, but they are equivalent in the case of dense graphs [20]. Yet the SCM still
treats a given degree sequence as a fixed constraint, albeit not as a sharp but soft constraint.
This constraint is in stark contrast with reality of many growing real networks, in which the
degree of all nodes constantly change, yet the shape of the degree distribution and the average
degree do not change, staying essentially constant in networks that grow in size even by orders
of magnitude [9, 10, 1–3]. These observations motivated the development of the hypersoft
configuration model [21–24].
1.1
Hypersoft configuration model (HSCM)
In the HSCM neither degrees nor even their expected values are fixed. Instead the fixed
properties are the degree distribution and average degree. The HSCM with a given average
degree and power-law degree distribution is defined by the exponential measure µ on the real
line R
µ = eαx ,
x ∈ R,
where α > 1 is a constant, and by the Fermi-Dirac graphon W : R2 → [0, 1]
W (x, y) =
1
ex+y
+1
.
The volume-form measure µ establishes then probability measures
µn =
µ|An
= α eα(x−Rn )
µ(An )
3
(1)
on intervals An = (−∞, Rn ], where
Rn =
1
n
log 2 ,
2
β ν
β =1−
1
,
α
(2)
and ν > 0 is another constant. The constants α > 1 and ν > 0 are the only two parameters
of the model. The HSCM random graphs of size n are defined by (W, An , µn ) via sampling n
i.i.d. points x1 , . . . , xn on An according to measure µn , and then connecting pairs of points i
and j at sampled locations xi and xj by an edge with probability W (xi , xj ).
An alternative equivalent definition is obtained by mapping (W, An , µn ) to (WI,n , I, µI ),
where I = [0, 1], µI = 1, and
WI,n (x, y) =
1
n
β2ν
1
,
x, y ∈ I,
(3)
(xy) α + 1
In this definition, xi s are n i.i.d. random variables uniformly distributed on the unit interval
[0, 1], and vertices i and j are connected with probability WI,n (xi , xj ).
Yet another equivalent definition, perhaps the most familiar and most frequently used one,
is given by (WP,n , P, µP ), where interval P = [βν, ∞), and measure µP on P is the Pareto
distribution
µP = α (βν)α x−γ , x ∈ P,
1
WP,n (x, y) = νn
, x, y ∈ P.
xy + 1
In this Pareto representation, the expected degree of a vertex at coordinate x is proportional
to x [15, 16].
Compared to the SCM where only edges are random variables while the expected degrees
are fixed, the HSCM introduces another source of randomness—and hence entropy—coming
from expected degrees that are also random variables. One obtains a particular realization
of an SCM from the HSCM by sampling xi s from their fixed distribution and then freezing
them. Therefore the HSCM is a probabilistic mixture of canonical ensembles, SCM ERGs, so
that one may call the HSCM a hypercanonical ensemble given that latent variables x in the
HSCM are called hyperparameters in statistics [25].
1.2
Properties of the HSCM
We prove in Theorem 3.2 that the distribution of degrees D in the power-law HSCM ensemble
defined above converges—henceforth convergence always means the n → ∞ limit, unless
mentioned otherwise—to
P (D = k) = α(βν)α
Γ(k − α, βν)
Γ(k − α)
= α(βν)α
(1 − P (k − α, βν)) ,
k!
Γ(k + 1)
(4)
where Γ(a, x) is the upper incomplete Gamma function, and P (a, x) is the regularized lower
incomplete Gamma function. Since P (a, x) ∼ (ex/a)a for a ex, while Γ(k − α)/Γ(k + 1) ∼
k −(α+1) for k α, we get
P (D = k) ∼ α(βν)α k −γ ,
4
γ = α + 1.
(5)
4
Pim van der Hoorn, Gabor Lippner, Dmitri Krioukov
10
10
0
10
10
-10
10
-15
10
Theory
HSCM n=104
HSCM n=104
HSCM n=105
HSCM n=105
HSCM n=106
Internet
-5
0
10
2
10
4
0
Theory
10
10
6
HSCM n=106
-5
10
-10
10
-15
10
0
10
1
10
2
10
3
10
4
10
5
Fig. 1 Degree distribution in the HSCM, theory vs. simulations, and in the Internet. The theory curve
Figure
1:panel
Degree
distribution
the
theory
vs.data
simulations,
and
in the
in the left
is Eq. (5)
with α = 1.1 (γ in
= 2.1)
andHSCM,
ν = 4.92. The
simulation
shown by symbols
is averaged
over
100
random
graphs
for
each
graph
size
n.
All
the
graphs
are
generated
according
to
the
HSCM
with
the
Internet. The theory curve in the left panel is Eq. (4) with α = 1.1 (γ = 2.1) and4 ν =5 same
4.92.
α = 1.1 and ν = 4.92. The average degrees, averaged over 100 random graphs, in the graphs of size 10 , 10 , and
6
The
simulation
data
shown
by
symbols
is
averaged
over
100
random
graphs
for
each
graph
10 , are 1.73, 2.16, and 2.51, respectively. The Internet data comes from CAIDA’s Archipelago measurements of
the Internet
at the are
Autonomous
System
level [34]. to
Thethe
number
of nodes
the same
averageαdegree
in and
the
size
n. Alltopology
the graphs
generated
according
HSCM
withandthe
= 1.1
Internet graph are 23, 752 and 4.92. The right panel shows the theoretical degree distribution curve in Eq. (5) with4
να=
average
degrees, averaged
overHSCM
100 graphs
random
graphs,
in with
the the
graphs
size
10 ,
= 4.92.
2.0 and The
ν = 10.0
versus simulations
of 100 random
of different
sizes
same αofand
ν. The
5 , and 106 , are 1.73, 2.16, and4 2.51,
10
respectively.
Internet
data comes from CAIDA’s
average
degrees in the graphs of size 10 , 105 , and
106 , are 9.96, The
9.98, and
10.0, respectively.
Archipelago measurements of the Internet topology at the Autonomous System level [26]. The
number of nodes and the average degree in the Internet graph are 23, 752 and 4.92. The right
the limit, is constant δ = ν/2 if α = 2. In this case, the HSCM can be equivalently defined as a
panel shows the theoretical degree distribution curve in Eq. (4) with α = 2.0 and ν = 10.0
growing model of labeled graphs, converging to soft preferential attachment [22]—for n = 1, 2, . . .,
versus
simulations of 100 random HSCM graphs of different
sizes with the same α and ν. The
the location xn of node n belongs to A4n ’s increment,
xn ∈ Bn = An \ An−1 (A0 = ∅), and
5 , and 106 , are
average
degrees
in
the
graphs
of
size
10
,
10
9.96,
andµ̃10.0,
respectively.
sampled from µ restricted to this increment, i.e., from the probability9.98,
measure
n = µ|B /µ(Bn ) =
n
α eα(x−Rn +Rn−1 ) ; node n is then connected to existing nodes i = 1 . . . n − 1 with probability given
by also
(2). For
theinexact
equivalence,
forthe
each
n, between
the original
equilibrium
HSCM
definitionto
We
prove
Theorem
3.3 that
expected
average
degree in
the ensemble
converges
with ordered xi s, xi < xi+1 , and this growing definition, slightly more care, ensuring that the
joint distribution of xi s is exactly the sameEin[D]
both
cases, must be taken using basics properties
= ν.
(6)
of Poisson point processes [33]. Specifically, in the equilibrium formulation with a given Rn , the
number
of nodes
cannot
be fixed in
to the
n, but
must behas
sampled
from
Poisson
distribution
with
That
is, the
degree
distribution
ensemble
a power
tailthe
with
exponent
γ, while
the
mean n. Alternatively, for a given fixed n, the right boundary Rn of interval An cannot be fixed,
expected average degree is fixed to constant ν that does not depend on n, Fig. 1.
but must be a random variable Rn = (1/2) log(2Vn ), where Vn is a random variable sampled from
the Gamma distribution with shape n and rate δ = ν/2. Node n is then placed at xn = Rn ,
1.3
theofmaximum-entropy
requirement
while Unbiasedness
the coordinates ofand
the rest
n − 1 nodes are sampled
from µn defined by this random
Rn , and then labeled in the increasing order. In the growing graph formulation, the coordinate
While
thethe
average
degree
HSCM by
converges
distribution
2xi
xn+1 of
n + 1’th
node in
is the
determined
vn+1 = to
vn a+ constant,
V , where and
v0 =the
0, vdegree
, and
i = (1/2)e
converges
to
a
power
law,
the
power-law
HSCM
is
certainly
just
one
of
an
infinite
number
V is a random variable sampled from the exponential distribution with rate δ = ν/2. All theof
other
possess these
two properties.
OnePoisson
example
is random
graphs [27]
three models
options that
are equivalent
realizations
of the same
point
processhyperbolic
on R with measure
µ
that
have
constanttoaverage
degree
and power-law
larger
and also
rate δ,
converging
the binomial
sampling
with bothdegree
n and distribution,
Rn fixed [33]. but
Thehave
projective
map πn7→of
the projectivity
definition
above,
mapslimit.
graphsIsGthe
their subgraphs
induced
by
numbers
triangles,
non-zero
clustering
in the
an unbiased
model
n0 in
n to HSCM
0
nodes
i
=
1
.
.
.
n
.
We
note
that
even
though
the
growing
version
of
the
model
is
not
exchangeable
of random power-law graphs with constant average degree? That is, are the HSCM random
since itcharacterized
relies on labeling
of nodes
theproperties
increasingand
order
their coordinates,
it is is
nevertheless
graphs
by only
these in
two
no of
others?
Or colloquially,
the HSCM
equivalent
to
the
equilibrium
HSCM
with
this
ordered
labeling,
because
the
joint
distribution
of
the model of “maximally random” power-law graphs with a constant average degree. This
node coordinates, and the linking probability as a function of these coordinates are the same in
question can be formally answered by checking whether the HSCM satisfies the maximumthe two formulations [33]. This observation suggests that there might exist a less trivial projective
entropy
requirement.
map such
that the model is both projective and exchangeable if α = 2. If α 6= 2, then the HSCM,
A
discrete
distribution
i =equivalent
1, 2, . . ., istosaid
satisfy the
maximum-entropy
requirement
even with ordered
labeling,
soft to
preferential
attachment,
but it is equivalent
to
P pisi ,not
¯
subject
to constraints
= (dis)appearance
1, 2, . . ., where fofr sedges
are some
realexisting
functions
of states
i fir = frate
r , r of
its adjusted
version with iapcertain
between
vertices
[22].i,
Here we note that the HSCM is the zero-clustering limit [38] of random hyperbolic graphs [39],
where the α = 2 case corresponds to the uniform density of points in the hyperbolic space Hd , and
5
and f¯r sPare a collection of real numbers, if the Gibbs/Shannon entropy of the distribution
S = − i pi log pi is maximized subject to these constraints [28]. This entropy-maximizing
distribution is known to be always unique, belonging to the exponential family of distributions, and it can be derived from the basic consistency axioms: uniqueness and invariance
with respect to a change of coordinates, system independence and subset independence [29–
31]. Since entropy is the unique measure of information satisfying the basic requirements
of continuity, monotonicity, and system/subset independence [32], the maximum-entropy requirement formalizes the notion of encoding into the probability distribution pi describing a
stochastic system, all the available information about the system given to us in the form of
the constraints above, and not encoding any other information not given to us. Since the
entropy-maximizing distribution is unique, any other distribution necessarily but possibly implicitly introduces biases by encoding some additional ad-hoc information, and constraining
some other system properties, concerning which we are not given any information, to some
ad-hoc values. Clearly, such uncontrolled information injection into a model of a system may
affect the predictions one may wish to make about the system using the model, and indeed
it is known that given all the available information about a system, the predictive power
of a model that describes the system is maximized by the maximum-entropy model [29–31].
Perhaps the best illustration of this predictive power is the predictive power of equilibrium
statistical mechanics, which can be formulated almost fully in terms of the maximum-entropy
principle [28].
To illustrate the maximum-entropy requirement in application to random graphs, suppose
we are to define a random graph ensemble, and the only available information is that these
random graphs must have n nodes and m edges. From the purely probabilistic perspective,
any random graph ensemble satisfying these constraints—random m-stars or m-cycles, for
instance, if m < n—would be an equally good one. Yet there is only one unique ensemble
that satisfies not only these constraints but also the maximum-entropy requirement. This
ensemble is Gn,m because in Gn,m any graph with n nodes and m edges is equally likely,
so that the probability distribution on the set of all graphs with n nodes and m edges is
uninform, and without further constraints, the uniform distribution is the maximum-entropy
n
2 ) , the number of such
distribution on the state space, which in this case is i = 1, . . . , (m
graphs. Random m-stars or m-cycles, while satisfying the constraints, inject into the model,
in this case explicitly, additional information about the graph structure that was not given to
us. Clearly, predictions based on random m-stars versus Gn,m may be very different, as the
first model trivially predicts that m-stars occur with probability 1, while they appear with a
nearly zero probability in Gn,m if n and m are large.
A slightly less trivial example is the SCM. In this case the given information is that the
n
expected degrees of nodes i = 1, . . . , n must be ki ∈ R+ , and the state space is all the 2( 2 )
graphs on n nodes. As shown in [13–16], the unique entropy-maximizing ensemble satisfying
these constraints is given by random graphs in P
which nodes i and j are connected with
probabilities pij = 1/ (kn/κ
κ
+
1),
where
k
=
i ki , and κi s are the unique solution of
Pi j
the system of n equations j pij = ki . The popular Chung-Lu model [11, 12] is different in
that the connection probability there is pCL
ij = min (ki kj /kn, 1), which can be thought of as a
classical-limit approximation to the entropy-maximizing
Fermi-Dirac pij above. While the CL
P
ensemble also satisfies the desired constraints j pCL
=
ki (albeit only for sequences ki s.t.
ij
ki kj /kn ≤ 1), it does not satisfy the maximum-entropy requirement, so that it injects, in this
case implicitly, some additional information into the ensemble, constraining some undesired
6
properties of graphs in the ensemble to some ad-hoc values. Since the undesired information
injection is implicit in this case, it may be quite difficult to detect and quantify all the biases
introduced into the ensemble.
1.4
Main results
The main result of this paper is the proof in Theorem 3.1 that the HSCM is unbiased, that is,
that the HSCM random graphs maximize the Gibbs entropy of random graphs whose degree
distribution and average degree converge to (4,6).
The first difficulty that we face in proving this result is how to properly formulate the
entropy-maximization problem under these constraints. Indeed, we are to show that the
probability distributions Pn that the HSCM defines on the set of n-sized graphs Gn maximizes
the graph entropy
X
S[Pn0 ] = −
Pn0 (Gn ) log Pn0 (Gn )
Gn
across all the distributions Pn0 that define random graph ensembles with the degree distributions and average degrees converging to (4,6). These constraints are quite different than
the SCM constraints, for example, because for any fixed n, we do not have a fixed set of
constraints or sufficient statistics. Instead of introducing such sufficient statistics, e.g., expected degrees converging to a desired Pareto distribution, and proceeding from there, we
show in Section 2 that the problem of graph entropy maximization under these constraints is
equivalent to a graphon entropy maximization problem [33], i.e., to the problem of finding a
graphon W that maximizes graphon entropy
ZZ
0
σn [W ] =
H W 0 (x, y) dµn (x) dµn (y),
A2n
where H(p) = −p log p − (1 − p) log(1 − p) is the entropy of a Bernoulli random variable with
success probability p, across all the graphons W 0 that satisfy the constraint
Z
W 0 (x, y) dµn (y) = κn (x),
(n − 1)
An
√
where κn (x) ≈ νn e−x is the expected degree of a node at coordinate x in the power-law
HSCM. We then prove in Proposition 3.4 that the unique solution to this graphon entropy
maximization problem is given by the Fermi-Dirac graphon W in (1). The fact that the
Fermi-Dirac graphon is the unique solution to this graphon entropy maximization problem
is a reflection of the basic fact in statistical physics that the grand canonical ensemble of
Fermi particles, which are edges of energy x + y in our case, is the unique maximum-entropy
ensemble with fixed expected values of energy and number of particles [34], in which the
probability to find a particle at a state with energy x + y is given by (1).
Yet the solutions to the graph and graphon entropy maximization problems yield
equivan
∗
lent random graph ensembles only if the rescaled graph entropy S [Pn ] = S[Pn ]/ 2 converges
to the graphon entropy σn [W ]. Here we face another difficulty that since our ensembles are
sparse, both S ∗ [Pn ] and σn [W ] converge to zero, so that we are actually to prove that the two
entropies converge to each other faster than either of them converges to zero. To this end we
prove in Theorems 3.5 and 3.6 that both the graphon and graph entropies converges to zero
7
as σn [W ], S ∗ [Pn ] ∼ ν log(n)/n. The key result then, also in Theorem 3.6, is the proof that
if divided by the scaling factor of log(n)/n, the difference between the graphon and graph
entropies vanishes in the limit, limn→∞ (n/ log n)|S ∗ [Pn ] − σn [W ]| → 0, meaning that the two
entropies do indeed converge to each other faster than to zero.
The combination of graphon (1) being the entropy maximizer, and the convergence of the
rescaled graph entropy to the entropy of this graphon, implies the main result in Theorem 3.1
that the power-law HSCM is a graph entropy maximizer subject to the degree distribution
and average degree constraints (4,6).
1.5
Exchangeability and projectivity
In addition to the natural, dictated by real-world network data, requirements of constant,
i.e., independent of graphs size n, average degree and power-law degree distribution, as well
as the maximum-entropy requirement, dictated by the basic statistical considerations, a reasonable model of real networks must also satisfy two more requirements: exchangeability and
projectivity.
Exchangeability takes care of the fact that node labels in random graph models are usually
meaningless. Even though node labels in real networks often have some network-specific
meaning, such as autonomous system numbers in the Internet [9], node labels in random graph
models can be, and usually are, random integer indices i = 1, 2, . . .. A random graph model is
exchangeable if for any permutation σ of node indices i, the probabilities of any two graphs G
and Gσ given by adjacency matrices Gi,j and Gσ(i),σ(j) are the same, P (G) = P (Gσ ) [35, 36].
A random graph model is projective if there exists a map πn7→n0 from graphs of size n
to graphs of size n0 < n such that the probability of graphs in the model satisfies P (Gn0 ) =
P (πn7→n0 (Gn )) [37, 38]. If this condition is satisfied, then it is easy to see that the same
model admits a dual formulation as an equilibrium model of graphs of a fixed size, or as a
growing graph model [39]. If this requirement is not satisfied, then as soon as one node is
added to a graph, e.g., due to the growth of a real network that this graph represents, then the
resulting bigger graph is effectively sampled from a different distribution corresponding to the
model with different parameters, necessarily affecting the structure of its existing subgraphs,
a clearly unrealistic scenario. As the simplest examples, Gn,p is projective (the map πn7→n0
simply selects any subset of n nodes consisting of n0 nodes), but Gn,k/n with constant k is
not. In the first case, one can realize Gn,p by growing graphs adding nodes one at a time, and
connecting each new node to all existing nodes with probability p, while in the second case
such growth is impossible since the existing edges in the growing graphs must be removed
with probability 1/n for the resulting graphs to be samples from Gn,k/n for each n.
The HSCM random graphs are manifestly exchangeable as any graphon-based ensemble [40, 41]. Here we note that the fact that these graphs are both sparse and exchangeable
is not by any means in conflict with the Aldous-Hoover theorem [35, 42] that states that
the limit graphon, mapped to the unit square, of any exchangeable sparse graph family is
necessarily zero. Indeed, if mapped to unit square, the limit HSCM graphon WI,n (3) is zero
as well. We also note that the convergence of WI,n to zero does not mean that the ensemble
converges to infinite empty graphs. In fact, the expected degree distribution and average
degree in the ensemble converge to (4,6) in the limit, as stated above.
If α = 2, the HSCM ensemble is also projective, but only with a specific labeling of
nodes breaking exchangeability. This can be seen by observing that the density of points on
α/2 1−α/2
intervals An , δn = n/µ(An ) = α β 2 ν
n
, and consequently on the whole real line R
8
in the limit, is constant δ = ν/2 if α = 2. In this case, the HSCM can be equivalently defined
as a model of growing labeled graphs as follows: for n = 1, 2, . . ., the location xn of new node
n belongs to An ’s increment, xn ∈ Bn = An \ An−1 (A0 = ∅), and sampled from µ restricted
to this increment, i.e., from the probability measure µ̃n = µ|Bn /µ(Bn ) = α eα(x−Rn +Rn−1 ) .
Having its location sampled, new node n then connects to all existing nodes i = 1 . . . n − 1
with probability given by (1).
This growing model is equivalent to the original equilibrium HSCM definition in Section 1.1
only asymptotically. However, the exact equivalence, for each n, between the equilibrium
HSCM with ordered xi s, xi < xi+1 , and its growing counterpart can be also achieved by
ensuring that the joint distribution of xi s is exactly the same in both cases, using basic
properties of Poisson point processes [39]. Specifically, the equilibrium definition in Section 1.1
must be adjusted by making the right boundary Rn of interval An not a fixed function of
n (2), but a random variable Rn = (1/2) log(2Vn ), where Vn is a random variable sampled from
the Gamma distribution with shape n and rate δ = ν/2. Node n is then placed at random
coordinate xn = Rn , while the coordinates of the rest of n − 1 nodes are sampled from
probability measure µn —measure µ restricted to the random interval An = (−∞, Rn ]—and
then labeled in the increasing order of their coordinates. The growing model definition must be
also adjusted: the coordinate xn+1 of the n+1’th node is determined by vn+1 = vn +V , where
v0 = 0, vi = (1/2)e2xi , and V is a random variable sampled from the exponential distribution
with rate δ = ν/2. One can show that coordinates xi , both for finite and infinite n, in both
the equilibrium and growing HSCM models defined this way, are equivalent realizations of
the same Poisson point process on R with measure µ and rate δ, converging to the binomial
sampling with Rn fixed to (2) [39].
The projective map πn7→n0 in the projectivity definition above, simply maps graphs Gn
to their subgraphs induced by nodes i = 1 . . . n0 . We note that even though the growing
HSCM is not exchangeable since it relies on labeling of nodes in the increasing order of their
coordinates, it is nevertheless equivalent to the equilibrium HSCM with this ordered labeling,
because the joint distribution of node coordinates, and the linking probability as a function
of these coordinates are the same in both the equilibrium and growing HSCM definitions [39].
This observation suggests that there might exist a less trivial projective map such that the
HSCM is both projective and exchangeable at the same time.
1.6
Other remarks
We note that thanks to its projectiveness, the power-law HSCM was shown in [24] to be
equivalent to a soft version of preferential attachment, a model of growing graphs in which
new nodes connect to existing nodes with probabilities proportional to the expected degrees
of existing nodes. It is well-known that similar to the HSCM, the degree distribution and
average degree in graphs grown according to preferential attachment do not essentially change
either as graphs grow [43, 44]. If α = 2, then the equivalence between the HSCM and soft
preferential attachment is exact. If α 6= 2, then the HSCM, even with ordered labeling, is not
equivalent to soft preferential attachment, but it is equivalent to its adjusted version with a
certain rate of (dis)appearance of edges between existing vertices [24].
We also note that the HSCM is the zero-clustering limit [45] of random hyperbolic
graphs [27], where the α = 2 case corresponds to the uniform density of points in the hyperbolic space Hd , and where xi s are the radial coordinates of nodes i in the spherical coordinate
system of the hyperboloid model of Hd . These coordinates can certainly not be negative,
9
but the expected fraction of nodes with negative coordinates in the HSCM is negligible:
α/2
µn (R− ) = β 2 ν/n
→ 0. In the zero-clustering limit, the angular coordinates of nodes in
d
H are ignored in the hyperbolic graphon [27], which becomes equivalent to (1).
As a final introductory remark we note that among the rigorous approaches to sparse exchangeable graphs, the HSCM definition is perhaps closest to graphon processes and graphexes
in [46–48]. In particular in [48], where the focus is on graph convergence to well-defined limits,
two ensembles are considered. One ensemble, also appearing in [47], is defined by any graphon
W : R2+ → [0, 1] and any measure µ on R+ (R+ can be replaced with any measure space).
Graphs of a certain expected size, which is a growing function of time t > 0, are defined
by sampling points as the Poisson point process on R+ with intensity tµ, then connecting
pairs of points with the probability given by W , and finally removing isolated vertices. The
other ensemble is even more similar to the HSCM. It is still defined by W and µ on R+ , but
the location of vertex n on R+ is sampled from µn = µ|An /µ(An ), where An s are finite-size
intervals growing with n whose infinite union covers the whole R+ . The latter ensemble is
not exchangeable, but both ensembles are shown to converge to properly stretched graphons
defined by W , yet only if the expected average degree grows to infinity in the limit. The
HSCM definition is different—in particular all n vertices of n-sized graphs are sampled from
the same µn —ensuring exchangeability, and allowing for an explicit control of the degree
distribution and average degree, which can be constant, but making the problem of graph
convergence difficult. We do not further discuss graph convergence here, leaving it, as well as
the generalization of the results to arbitrary degree distributions, for future publications.
1.7
Paper organization
In the next Section 2 we first review in more detail the necessary background information
and provide all the required definitions. In Section 3 we formally state all the results in the
paper, while Section 4 contains all the proofs of these results.
2
2.1
Background information and definitions
Graph ensembles and their entropy
A graph ensemble is a set of graphs G with probability measure P on G. The Gibbs entropy
of the ensemble is
X
S[P ] = −
P (G) log P (G)
(7)
G∈G
Note that this is just the entropy of the random variable G with respect to the probability
measure P . When Gn is a graph of size n, sampled from G according to measure P , we
write S[Gn ] instead of S[P ]. Given a set of constraints, e.g., in the form of graph properties
fixed to given values, the maximum-entropy ensemble is given by P ∗ that maximizes S[P ]
across all measures P that satisfy the constraints. These constraints can be either sharp
(microcanonical) or soft (canonical), satisfied either exactly or on average, respectively. The
simplest example of the constrained graph property is the number of edges, fixed to m, in
graphs of size n. The corresponding microcanonical
and canonical maximum-entropy ensem
bles are Gn,m and Gn,p with p = m/ n2 , respectively. The P ∗ is respectively the uniform and
exponential Boltzmann distribution P (G) = e−H(G) /Z with Hamiltonian H(G) = λm(G),
10
where m(G) is the number of edges in graph G, and the Lagrange multiplier λ is given by
p = 1/ eλ + 1 [13].
When the constraints are given by the degrees of nodes, instead of the number of edges,
we have the following characterization of the microcanonical and canonical ensemble.
2.1.1
Maximum-entropy graphs with a given degree sequence (CM)
Given a degree sequence dn = d1 . . . dn , the microcanonical ensemble of graphs that have this
degree sequence is the configuration model (CM) [7, 8]. The entropy-maximizing P ∗ is uniform
on the set of all graphs that have degree sequence dn .
2.1.2
Maximum-entropy graphs with a given expected degree sequence (SCM)
If the sharp CM constraints are relaxed to soft constraints, the result is the canonical ensemble
of the soft configuration model (SCM) [14, 15]. Given an expected degree sequence kn , which
in contrast to CM’s dn , does not have to be a graphical sequence of non-negative integers, but
can be any sequence of non-negative real numbers, the SCM is defined by connecting nodes
i and j with probabilities
pij =
ki =
1
, where Lagrange multipliers λi are the solution of
+1
pij , i = 1 . . . n.
(8)
eλi +λj
X
(9)
i<j
The entropy-maximizing P ∗ is the Boltzmann distribution with Hamiltonian H(G) =
where di (G) is the degree of node i in graph G [14, 15].
2.2
P
i λi di (G),
Sparse graphs with a given degree distribution
Let p(k), k = 0, 1, 2, . . ., be a probability density function with finite mean. Denote by D the
corresponding random variable, and consider a sequence of graph ensembles that maximize
Gibbs entropy under the constraint that for all k
lim P (Dn = k) = P (D = k) ,
n→∞
(C1)
where Dn is the degree of a uniformly chosen node in the ensemble of graphs of size n. In other
words, this is a maximum-entropy ensemble of graphs whose degree distribution converges to
p(k).
In addition to the degree distribution constraint, we also want our graphs to be sparse.
The most common definition of sparseness seems to be that the number of edges is o(n2 ), so
that the expected average degree can be unbounded. In contrast, here we use the term sparse
to mean that the expected degree converges to the expected value of D:
lim E [Dn ] = E [D] := ν.
n→∞
(C2)
Constraint (C2) implies that the number of edges is O(n). We note that, in general, (C2) does
not follow from (C1), since convergence in distribution does not necessarily imply convergence
in expectation.
11
We also note that constraints (C1,C2) are neither sharp nor soft, since they deal only with
the n → ∞ limits of the degree distribution and expected degree. We call these constraints
hypersoft, since as we will see below, random graphs satisfying these constraints can be realized
as random graphs with Lagrange multipliers that are not parameters but hyperparameters in
the statistics terminology [25].
2.3
Maximum-entropy graphs with hypersoft constraints
Similar to the case of random graphs with a given (expected) degree sequence, we are to
determine the distribution P that satisfies (C1) and (C2), and maximizes the Gibbs entropy.
However, this task poses the question of what it means to maximize entropy under these
limit constraints. In particular, unlike an ensemble of graphs with a given (expected) degree
sequence, we are no longer dealing with a set of graphs of fixed size, but with a sequence of
graphs (Gn )n≥1 of varying sizes. To answer this question, and to give a proper definition of
entropy maximization under hypersoft constraints, we consider graphon-based ensembles of
graphs.
2.3.1
Graphon-based graph ensembles
In the simplest case, a graphon is a symmetric integrable function W : [0, 1]2 → [0, 1].
Graphons, or more precisely graphon equivalence classes, consisting of all functions Wσ such
that W (x, y) = Wσ (σ(x), σ(y)) under all measure preserving transformations σ : [0, 1] →
[0, 1], are well-defined limits of dense graph families [49, 41]. One can think of the interval
[0, 1] as the continuum limit of node indices i, and of W as the limit of graphs’ adjacency
matrices. Equivalently, W (x, y) is the probability that there exists the edge between “nodes”
x and y. Graphons are an application to graphs of a class of earlier results on exchangeable
arrays in statistics [35, 50], and are better known as the connection probability in random
graphs with latent parameters in sociology [51–54] and network science [21, 22], also known
in graph theory as inhomogeneous random graphs [55]. Here we use the term graphon to refer
to any symmetric function W : A × A → [0, 1], for some A ⊆ R.
Let µ be a probability measure on A ⊆ R and W a graphon. Then the standard graphonbased ensemble GW,µ,A known as W -random graphs [40] is the ensemble of random n-sized
graphs Gn defined by first i.i.d. sampling n node coordinates xn = x1 , . . . , xn ∈ A according
to µ, and then connecting every node pair (i, j), independently, with probability W (xi , xj ).
To be able to satisfy the hypersoft constraints we generalize this ensemble as follows. Let
µ be a measure on R, W a graphon, and letS A = (A1 , A2 , . . .), An ⊆ An+1 , be an infinite
sequence of growing subsets of R such that n≥1 An = R and µ(An ) < ∞. We then define
the graphon ensemble GW,µ,A = (GW,µn ,n )n≥1 to be random graphs (Gn )n≥1 defined by the
graphon W and measures
µn =
µ|An
,
µ(An )
which are the probability measures on An associated with µ. To sample a graph Gn from this
ensemble, n i.i.d. points xi ∈ An are first sampled according to measure µn , and then pairs
of nodes i and j are connected by an edge with probability W (xi , xj ).
We remark that if An = A = [0, 1] and µ is the uniform measure, we are in the most
classical settings of W -random graphs [40]. In the case of an arbitrary A and measure µ, our
12
random graph ensemble is similar to a model considered recently in [48], Section 2.4. There,
a growing graph construction is considered where Gn is created from Gn−1 by sampling the
coordinate xn of node n according to measure µn , and then connecting it to all existing nodes
i, independently with probability W (xn , xi ). The main difference between such sampling and
the one above is that in the former case, the coordinates xi of different nodes i = 1 . . . n are
sampled from different measures µi , thus breaking exchangeability, while in the latter case,
all xi s are sampled from the same measure µn .
2.3.2
Bernoulli and graphon entropies
Given the coordinates xn , edges in our random graphs are independent Bernoulli random
variables, albeit with different success probabilities W (xi , xj ). The conditional Bernoulli
entropy of random graphs with fixed coordinates xn is thus
X
S[Gn |xn ] =
H (W (xi , xj )) ,
i<j
where H is the Bernoulli entropy,
H(p) = −p log p − (1 − p) log(1 − p),
0 ≤ p ≤ 1.
The graphon entropy [50, 41, 56–59] of W with respect to µ on A is defined as
ZZ
σ[W, µ, A] =
H (W (x, y)) dµ(x) dµ(y),
(10)
(11)
A2
which we write as σ[W, µ], when µ has support A and no confusion arises. In addition, if Gn
is a graph in the graphon ensemble GW,µn ,n , we write σ[Gn ] for σ[W, µn , An ].
Since for any two discrete random variables X and Y , the expectation of the conditional
entropy of X given Y is a lower bound for X’s entropy, S[X] ≥ E [S[X|Y ]], the graphon
entropy definition implies
n
n
S[Gn ] ≥ E [S[Gn |xn ]] =
E [H(W (x, y))] =
σ[Gn ],
(12)
2
2
The graphon entropy is thus a lower bound for the rescaled Gibbs entropy defined as
S ∗ [Gn ] =
S[Gn ]
.
n
(13)
2
Before we give our definition of the maximum-entropy ensemble of sparse graphs, it is
instructive to consider the case of dense graphs.
2.3.3
Dense maximum-entropy graphs with a given degree distribution
Consider a sequence (dn )n≥1 of degree sequences dn = (d1 . . . dn )n , (di )n ≥ (di+1 )n , for which
there exist constants 0 < c1 ≤ c2 such that c1 n ≤ (di )n ≤ c2 n for any (di )n ∈ dn . Now let
(Gn )n≥1 be a sequence of microcanonical ensembles of random graphs, that is, CMs, defined
by dn . If there exists a function f : [0, 1] → (0, 1) such that for any c ∈ (0, 1)
Dn
lim P
≤ c = P (f (U ) ≤ c) ,
(14)
n→∞
n
13
where Dn is the degree of a random node in random graph Gn ∈ Gn or equivalently, a
uniformly sampled (di )n ∈ dn , and U is a uniform random variable on [0, 1], then it was
proven in [20] that the limit of the CM sequence (Gn )n≥1 is given by the graphon
W (x, y) =
1
eg(x)+g(y)
+1
,
(15)
where g(x) is such that
Z
Z
Z 1
WA (x, y) dµ(y) =
W (x, y) dy =
f (x) =
A
A
0
1
dµ(y),
ex+y + 1
(16)
where A is the image of g and µ = g −1 (functions f and g are continuous and strictly increasing
almost everywhere on [0, 1] [20]).
Some important observations are in order here. First, we note that (14) is very similar to
(C1), with the exception that (14) implies that the degrees of all nodes are O(n), so that the
graphs are dense. In particular, in our graphs we have that Dn = o(n) so that Dn /n → 0.
Second, consider the problem of maximizing the graphon entropy under the constraint
given by (16). We will show in Proposition 3.4 that the solution to this problem is given
by (15), where g is defined by (16). Hence the graphon (15) obtained in [20] maximizes the
graphon entropy under the constraint (16) imposed by the limit f of the sequence of rescaled
degree sequences (dn /n)n≥1 .
Third, Theorem D.5 in [41] states that in dense W -random graph ensembles GW,µ,A
lim S ∗ [Gn ] = σ[W, µ, A],
n→∞
(17)
meaning that the rescaled Gibbs entropy of W -random graphs converges to the graphon
entropy of W .
Given f , this result suggests to consider the family of W -random graph ensembles G̃n
defined by the Fermi-Dirac graphon (15) with g given by (16), which we call the dense
hypersoft configuration model (HSCM). The distribution of rescaled degrees Dn /n in these
HSCM graphs converges to f , cf. (14), while the limit of these graphs is also W since the
limit of any dense W -random graphs is W [60]. That is, the limit of HSCM ensembles G̃n and
the limit of CM ensembles Gn with any (dn /n)n≥1 converging to f , are the same Fermi-Dirac
graphon W (15). Since the two ensembles have the same graphon limit, their rescaled Gibbs
entropies converge to the same value, equal, thanks to (17), to the graphon entropy, even
though for any finite n the two ensembles are quite different.
Fourth, if we replace the sequence of degree sequences (dn )n≥1 with a sequence of expected
degree sequences (kn )n≥1 converging upon rescaling to f , and then replace the sequence of
CMs with the corresponding sequence of SCMs, then the limit of this SCM sequence is the
same graphon W (15) [20], so that the rescaled SCM Gibbs entropy also converges to the
same graphon entropy, squeezed, for any finite n, between the CM entropy [7, 61, 62] and the
HSCM entropy (12). In other words, dense CM, SCM, and HSCM are all equivalent in the
limit, versus the sparse case where the equivalence is broken [18, 19] since the graphon is zero
in the limit.
The key point here however is that the rescaled degree distribution in the dense HSCM
converges to a well-defined limit, i.e., satisfies the hypersoft constraints (14), and that the
HSCM Gibbs entropy converges to the graphon entropy. Therefore if we define a maximumentropy ensemble under given hypersoft constraints f to be an ensemble that: 1) satisfies these
14
constraints, i.e., has a degree distribution converging to f in the limit, 2) maximizes graphon
entropy under these constraints given by (16), and 3) has rescaled Gibbs entropy converging
to the graphon entropy, then the dense HSCM ensemble is trivially such an ensemble. In
addition, this ensemble is the unique maximum-entropy hypersoft ensemble in the dense
case [20].
These observations instruct us to extend this definition of maximum-entropy hypersoft
dense graphs to sparse graphs, where we naturally replace the dense hypersoft constraints (14)
with sparse hypersoft constraints (C1,C2). However, things become immediately less trivial
in this case. In particular, we face the difficulty that since the limit graphon of any sparse
exchangeable graph ensemble is zero according to the Aldous-Hoover theorem [35, 42], the
entropy of this graphon is zero as well. Since this entropy is zero, |S ∗ [Gn ]−σ[Gn ]| → 0 in (17)
does not necessarily imply that the rescaled Gibbs entropy converges to the graphon entropy.
We address this difficulty next.
2.3.4
Rescaled graphon entropy of sparse graphs
Consider again the generalized graphon ensemble GW,µn ,n of random graphs defined in Section 2.3.1, and their graphon entropy defined as
ZZ
σ[Gn ] =
H (W (x, y)) dµn (x) dµn (y).
(18)
A2n
If W 6= {0, 1} everywhere, then for any finite n, σ[Gn ] is positive, but if the ensemble is
sparse, then limn→∞ σ[Gn ] = 0.
To address this problem we rescale the graphon entropy σ[Gn ] such that upon rescaling
it converges to a positive constant. That is, let an be a sequence such that
lim an σ[Gn ] = a ∈ (0, ∞).
n→∞
(19)
This rescaling does not affect the graphon entropy maximization problem, because maximizing
σ[Gn ] for every n as a functional of W under a given constraint is equivalent to maximizing
an σ[Gn ] under the same constraint.
Upon this rescaling, we see that the rescaled Gibbs entropy converges to the graphon
entropy, generalizing (17), if
lim an |S ∗ [Gn ] − σ[Gn ]| = 0,
n→∞
(20)
in which case an S ∗ [Gn ] converges to a. This condition implies that the rescaled Gibbs entropy
S ∗ [Gn ] converges to the graphon entropy σ[Gn ] faster than either of them converge to zero.
2.3.5
Sparse maximum-entropy graphs with a given degree distribution
With the graphon rescaling in the previous section, we can now define a graphon ensemble
GW,µn ,n to be a maximum-entropy ensemble under the sparse hypersoft constraints (C1,C2)
if:
1) the degree distribution and expected degree in GW,µn ,n converge to (C1,C2);
2) graphon entropy σ[Gn ] of GW,µn ,n is maximized for every n under the constraint imposed
by (C1,C2); and
15
3) (20) holds, with an given by (19).
Our main result (Theorem 3.1) is that the sparse power-law hypersoft configuration model,
defined next, is a maximum-entropy model under hypersoft constraints.
2.3.6
Sparse power-law hypersoft configuration model (sparse HSCM)
The sparse power-law HSCM(α, ν) is defined as the graphon ensemble GW,µn ,n = GW,µ,A , Section 2.3.1, with
W (x, y) =
1
,
+1
µ = eαx , α > 1,
(21)
ex+y
(22)
An = (−∞, Rn ],
1
n
1
Rn = log 2 , ν > 0, β = 1 − ,
2
νβ
α
µ|An
= α eα(x−Rn ) .
µn =
µ(An )
(23)
(24)
(25)
The dense power-law HSCM is recovered from the above definition by setting ν = ν̃n, where
ν̃ is a constant, in which case Rn = R = −(1/2) log ν̃β 2 , An = A = (−∞, R], and µn = µ =
α eα(x−R) .
3
Results
In this section we formally state our results, and provide brief overviews of their proofs appearing in subsequent sections. The main result is Theorem 3.1, stating that the HSCM(α, ν)
defined in Section 2.3.6 is a maximum-entropy model under hypersoft power-law degree distribution constraints, according to the definition in Section 2.3.5. This result follows from
Theorems 3.2-3.6 and Proposition 3.4. Theorems 3.2,3.3 establish the limits of the degree distribution and expected average degree in the HSCM(α, ν). Proposition 3.4 states that HSCM’s
graphon uniquely maximizes the graphon entropy under the constraints imposed by the degree
distribution. Theorem 3.5 establishes proper graphon rescaling and the limit of the rescaled
graphon. Finally, the most critical and involved Theorem 3.6 proves that the rescaled Gibbs
entropy of the HSCM converges to its rescaled graphon entropy.
3.1
Main result
Let Y be a Pareto random variable with shape α > 1, scale νβ > 0, β = 1 − 1/α, so that Y ’s
probability density function is
PY (y) = α (νβ)α y −γ , y ≥ νβ, γ = α + 1,
(
(νβ)α y −α if y ≥ νβ
P (Y > y) =
1
otherwise.
and
Let D be a discrete random variable with probability density function
k
Y −Y
P (D = k) = E
e
, k = 0, 1, 2, . . . ,
k!
16
(26)
(27)
(28)
which is the mixed Poisson distribution with mixing parameter Y [63]. Then it follows that
E [D] = E [Y ] = ν,
(29)
and since Y is a power law with exponent γ, the tail of D’s distribution is also a power law
with the same exponent [63]. In particular, P (D = k) is given by (4). Therefore if D is the
degree of a random node in a random graph ensemble, then graphs in this ensemble are sparse
and have a power-law degree distribution.
Our main result is:
Theorem 3.1. For any α > 1 and ν > 0, HSCM(α, ν) is a maximum entropy ensemble of
random graphs under the hypersoft constraints (C1,C2) with P (D = k) and ν defined by (2629).
3.2
The limit of the degree distribution in the HSCM
The degree Dn of a random node i in a random HSCM(α, ν) graph of size n, conditioned on the
node coordinates xn = x1 . . . xn , is the sum of n − 1 independent Bernoulli random variables
with success probabilities W (xi , xj ), j 6= i. The distribution ofPthis sum can be approximated
by the mixed Poisson distribution with the mixing parameter j6=i W (xi , xj ). Therefore after
first integrating over xj with j 6= i and then over xi , the distribution of Dn is approximately
the mixed Poisson distribution
(κn (X))k −κn (X)
P (Dn = k) = E
e
, so that E [Dn ] = E [κn (X)],
k!
where the random variable X has density µn (25), and the mixing parameter κn (x) is the
expected degree of a node at coordinate x:
κn (x) = (n − 1)wn (x),
Z
wn (x) =
W (x, y) dµn (y),
(30)
(31)
An
where An is given by (23).
In the n → ∞ limit, the distribution of the expected degree κn (X) of X converges to the
Pareto distribution with shape α and scale νβ. To prove this, we use the observation that
the mass of measure µn is concentrated towards the right end of the interval An = (−∞, Rn ],
where Rn 1 for large n. Therefore, not only the contributions coming from negative x, y
are negligible, but we can also approximate the Fermi-Dirac graphon W (x, y) in (21) with its
classical limit approximation
c (x, y) = e−(x+y)
W
(32)
on R2+ . In addition, the expected degree function wn (x) can be approximated with w
bn (x)
defined by
(
ωn e−x if 0 ≤ x ≤ Rn
w
bn (x) =
, where
(33)
0
otherwise
Z Rn
ν 1
1
1 − e−(α−1)Rn
2
ωn =
e−x dµn (x) =
=
+
o
n− 2 ,
(34)
βeRn
n
0
17
so that the expected degree of a node at coordinate x can be approximated by
1
1
κ
bn (x) = nw
bn (x) = e−x (νn) 2 + o n 2
.
To see that κ
bn (X) converges to a Pareto random variable, note that since X has density
µn , it follows that for all t > νβ
nω α
nωn
n
P (b
κn (X) > t) = P X < log
= e−αRn
t
t
= (νβ)α t−α (1 + o(1)) = P (Y > t) (1 + o(1)),
where Y is a Pareto-distributed random variable (27). We therefore have the following result,
the full proof of which can be found in Section 4.2.2:
Theorem 3.2 (HSCM(α, ν) satisfies (C1)). Let α > 1, ν > 0, and Dn be the degree of a
uniformly chosen vertex in the HSCM(α, ν) graphs of size n. Then, for each k = 0, 1, 2, . . .,
lim P (Dn = k) = P (D = k) ,
n→∞
where P (D = k) is given by (28).
3.3
The limit of the expected average degree in the HSCM
The expected degree of a random node in n-sized HSCM graphs is, for any fixed i,
X
E [Dn ] =
E [W (Xi , Xj )] = (n − 1)E [W (X, Y )] ,
(35)
j6=i
where X and Y are independent random variables with distribution µn . Approximating
c (x, y) on R2 , and using e2Rn = n/νβ 2 , we have
W (x, y) with W
+
Rn
ZZ
Rn
Z
n
0
−x
e
c (x, y) dµn (x) dµn (y) = n
W
2
dµn (x) =
0
2
n
−(α−1)Rn
1
−
e
= ν + o(1).
β 2 e2Rn
All other contributions are shown to also vanish in the n → ∞ limit in Section 4.2.3,
where the following theorem is proved:
Theorem 3.3 (HSCM satisfies (C2)). Let α > 1, ν > 0, and Dn be the degree of a uniformly
chosen vertex in the HSCM(α, ν) graphs of size n. Then
lim E [Dn ] = ν.
n→∞
3.4
HSCM maximizes graphon entropy
Let A ⊆ R be some interval, µ a measure on A with µ(A) < ∞, and suppose some µ-integrable
function w : A → R is given. Consider the graphon entropy maximization problem under the
constraint
Z
w(x) =
W (x, y) dµ(y).
(36)
A
That is, the problem is to find a symmetric function W ∗ that maximizes graphon entropy
σ[W, µ, A] in (11) and satisfies the constraint above, for fixed A and µ.
18
We note that this problem is a “continuous version” of the Gibbs entropy maximization
problem in the SCM ensemble in Section 2.1.2. The following proposition, which we prove in
Section 4.3.1, states that the solution to this problem is a “continuous version” of the SCM
solution (8):
Proposition 3.4 (HSCM Maximizes Graphon Entropy). Suppose there exists a solution W ∗ to
the graphon entropy maximization problem defined above. Then this solution has the following
form:
1
W ∗ (x, y) = λ(x)+λ(y)
,
(37)
e
+1
where function λ is µ-almost-everywhere uniquely defined on A by (36).
This proposition proves that for each n, the HSCM(α, ν)’s Fermi-Dirac graphon (21) maximizes the graphon entropy under constraint (31), because A and µ in the HSCM(α, ν) are
chosen such that λ(x) = x. This is always possible as soon as λ(x) is invertible, cf. Section 2.3.3. For each n, interval An (23) and measure µn (25) in the HSCM(α, ν) can be mapped
to [0, 1] and 1, respectively, in which case λn (x) = Rn + (1/α) log x, leading to (3). In other
words, node coordinates x in the original HSCM(α, ν) definition in Section 2.3.6, and their
coordinates x̃ in its equivalent definition with An = [0, 1] and µn = 1 are related by
x̃ = eα(x−Rn ) .
3.5
Graphon entropy scaling and convergence
To derive the rate of convergence of HSCM’s graphon entropy to zero, it suffices to consider
c (32) to W (21) on R2 . Its Bernoulli entropy (10) is
the classical limit approximation W
+
c (x, y) = (x + y)W
c (x, y) − (1 − W
c (x, y)) log(1 − W
c (x, y)).
H W
c (Rn , Rn ) → 0, the second
Since the most of the mass of µn is concentrated near Rn and W
term is negligible. Integrating the first term over [0, Rn ], we get
Rn
ZZ
c (x, y) dµn (x) dµn (y) = α2 e−2αRn
(x + y)W
Rn
ZZ
(x + y)e(α−1)(x+y) dx dy
0
0
2Rn
ν
n
log n
= 2 2Rn + O(n−1 ) = log 2 + O(n−1 ) = ν
+ O(n−1 ),
β e
n
νβ
n
(38)
from which we obtain the proper scaling as log(n)/n. All further details behind the proof of
the following theorem are in Section 4.3.2.
Theorem 3.5 (Graphon Entropy Convergence). Let σ[Gn ] be the graphon entropy (18) in
the HSCM(α, ν) ensemble with any α > 1 and ν > 0. Then, as n → ∞,
n σ[Gn ]
− ν = O (1/ log n) .
log n
This theorem implies that σ[Gn ] goes to zero as log(n)/n, while n σ[Gn ]/ log n goes to ν
as 1/ log n.
19
3.6
Gibbs entropy scaling and convergence
The last part of Theorem 3.1 is to prove that the rescaled Gibbs entropy (13) of the HSCM
converges to its graphon entropy (18) faster than the latter converges to zero.
The graphon entropy is a trivial lower bound for the rescaled Gibbs entropy, Section 2.3.2,
so the problem is to find an appropriate upper bound for the latter converging to the graphon
entropy. To identify such an upper bound, we rely on an argument similar to [41]. Specifically,
we first partition An into m intervals It that induce a partition of A2n into rectangles Ist =
Is ×It , s, t = 1 . . . m. We then approximate the graphon by its average value on each rectangle.
Such approximation brings in an error term on each rectangle. We then show that the Gibbs
entropy is upper-bounded by the entropy of the averaged graphon, plus the sum of entropies
of indicator random variables Mi which take value t if the coordinate xi of node i happens to
fall within interval It . The smaller the number of intervals m, the smaller the total entropy
of these random variables Mi , but the larger the sum of the error terms coming from graphon
averaging, because rectangles Ist are large. The smaller they are, the smaller the total error
term, but the larger the total entropy of the Mi ’s. The crux of the proof is to find a “sweet
spot”—the right number of intervals guaranteeing the proper balance between these two
types of contributions to the upper bound, which we want to be tighter than the rate of the
convergence of the graphon entropy to zero.
This program is executed in Section 4.4, where we prove the following theorem:
Theorem 3.6 (Gibbs Entropy Convergence). Let σ[Gn ] be the graphon entropy (18) and
S ∗ [Gn ] be the rescaled Gibbs entropy (13) in the HSCM(α, ν) ensemble with any α > 1 and
ν > 0. Then
n
lim
|S ∗ [Gn ] − σ[Gn ]| = 0, and
n→∞ log n
2 S[Gn ]
lim
= ν.
n→∞ n log n
We remark that this theorem implies that
S[Gn ] ∼
ν
n log n,
2
which is the leading term of the (S)CM Gibbs entropy obtained in [23]. It is also instructive to
compare this scaling of Gibbs entropy with its scaling in dense ensembles with limn→∞ σ[Gn ] =
σ[W, µ, A] = σ ∈ (0, ∞) [41]:
S[Gn ] ∼
σ 2
n .
2
Finally, it is worth mentioning that even though we use the Fermi-Dirac graphon W (21) to
define our W -random graphs, the same convergence results could be obtained for W -random
graphs defined by any other graphon W 0 such that
lim n E W (X, Y ) − W 0 (X, Y ) = 0,
n→∞
with X and Y having density µn . In fact, to establish the required limits, we use the classic instead of W . Therefore there exists a vast equivalence
cal limit approximation graphon W
class of W -random graphs defined by graphons W 0 that all have the same limit degree distribution (C1) and average degree (C2), and whose rescaled Gibbs entropy converges to
20
the graphon entropy of the Fermi-Dirac W . However, it follows from Proposition 3.4 that
among all these ensembles, only the W -random graph ensemble defined by the Fermi-Dirac
graphon (21) uniquely maximizes the graphon entropy (11) for each n, which, by our definition of maximum-entropy ensembles under hypersoft constraints, is a necessary condition for
graph entropy maximization.
4
Proofs
In this section we provide the proofs of all the results stated in the previous section. In
Section 4.1 we begin with some preliminary results on the accuracy of the approximation of
c . In the same section
the Fermi-Dirac graphon W (21) by the classical limit approximation W
we also establish results showing that the main contribution of the integration with respect to
µn is on the positive part [0, Rn ] of the interval An as defined by (23). In particular, we show
that all contributions coming from the negative part of this interval, i.e., R− , are o(n−1 ),
which means that for all our results the negative part of the support of our measure µn is
negligible. We then proceed with proving Theorems 3.2 and 3.3 in Section 4.2. The proofs
of Proposition 3.4 and Theorem 3.5 can be found in Section 4.3. Finally, the convergence of
the rescaled Gibbs entropy to the graphon entropy (Theorem 3.6) is given in Section 4.4.
4.1
The classical limit approximation of the Fermi-Dirac graphon
We will use e−(x+y) as an approximation to the graphon W to compute all necessary limits.
To be precise we define
c (x, y) = min{e−(x+y) , 1}
W
(39)
c converge to zero as n tends to
and show that differences between the integrals of W and W
c , we could have also worked with the integral expressions
infinity. Note that, instead of W
involving W , which might have led to better bounds. However, these integrals tend to evaluate
c are much easier to
to combinations of hypergeometric functions, while the integrals of W
evaluate and are sufficient for our purposes.
c (x, y) we need to consider separately, the intervals (−∞, 0] and
By the definition of W
(0, Rn ]. Since graphons are symmetric functions, this leads to the following three different
cases:
I)
− ∞ < x, y ≤ 0
II)
− ∞ < y ≤ Rn and 0 < x ≤ Rn ,
III)
0 < x, y ≤ Rn . (40)
c ≤ 1 and
For case I) we note that W, W
ZZ 0
dµn (y) dµn (x) = O n−α .
(41)
−∞
c , only the
With this we obtain the following result, which shows that for both W and W
2
integration over (0, Rn ] , i.e. case III, matters.
Lemma 4.1.
Z Z Rn
ZZ
c (x, y) dµn (y) dµn (x) −
W
−∞
Rn
c (x, y) dµn (y) dµn (x) = O n− α+1
2
W
0
c with W .
and the same result holds if we replace W
21
Proof. First note that
−Rn
Z
Rn
Z
Z
−Rn
c (x, y) dµn (y) dµn (x) ≤
W
−∞
−∞
We show that
(42)
−∞
0
Z
dµn (y) = O n−α .
Z
−Rn
Rn
c (x, y) dµn (y) dµn (x) = O n− α+1
2
W
,
(43)
0
which together with (41) and (42) implies the first result. The result for W follows by noting
c.
that W ≤ W
We split the integral (43) as follows
Z
0
Z
Rn
c (x, y) dµn (y) dµn (x)
W
−Rn
0
Z
0
Z
=
−x
0
Z
Rn
dµn (y) dµn (x) +
−Rn
−Rn
0
For the first integral we compute
Z
Z 0 Z −x
dµn (y) dµn (x) = αe−2αRn
−Rn
Z
0
= e−2αRn
e−(x+y) dµn (y) dµn (x).
−x
0
e−αx − 1 eαx dx
−Rn
αRn + e−αRn − 1 = O log(n)n−α
Finally, the second integral evaluates to
Z
α −2αRn 0 (α−1)Rn
e
e
− e−(α−1)x e(α−1)x dx
β
−Rn
!
α+1
α −2αRn e(α−1)Rn − 1
= e
− Rn = O n− 2 ,
β
α−1
from which the result follows since α > (α + 1)/2.
c ).
We can show a similar result for H(W ) and H(W
Lemma 4.2.
Z Z Rn
ZZ
c
H W (x, y) dµn (y) dµn (x) −
−∞
Rn
c (x, y) dµn (y) dµn (x) = n ,
H W
0
where
−α
O (log(n)n )
n = O log(n)2n−2
O n− 2+α
2
if 1 < α < 2,
if α = 2,
if α > 2,
c with W .
Moreover, the same result holds if we replace W
22
c . For this we split the interval An into three parts
Proof. We will first prove the result for W
(−∞, −Rn ], [−Rn , 0] and (0, Rn ] and show that the integrals on all ranges other than [0, Rn ]2
are bounded by a term that scales as n .
Since H(p) ≤ log(2) for all 0 ≤ p ≤ 1 it follows from (41) and (42) that, for all α > 1,
0
ZZ
c (x, y) dµn (y) dµn (x) = O n−α = o(n ),
H W
−∞
Rn Z −Rn
Z
−∞
c (x, y) dµn (y) dµn (x) = O n−α = o(n ).
H W
−∞
c , we only need to consider the integration over (0, Rn ] ×
Hence, using the symmetry of W
(−Rn , 0].
First we compute
c (x, y) = e−(x+y) (x + y) − (1 − e−(x+y) ) log 1 − e−(x+y)
H W
and observe that
− (1 − e−z ) log 1 − e−z ≤ e−2z
for all large enough z.
(44)
Now let δ > 0 be such that (44) holds for all z ≥ δ. Then δ < Rn for sufficiently large n and
we split the integration as follows
Z
0
Rn
Z
0
c (x, y) dµn (y) dµn (x)
H W
−Rn
δ
Z
Z
0
−Rn
0
Rn
Z
c (x, y) dµn (y) dµn (x) +
H W
=
0
Z
c (x, y) dµn (y) dµn (x).
H W
−Rn
δ
The first integral is O(n−α ). For the second we note that x + y > δ for all y > δ − x and
hence
Z Rn Z 0
c (x, y) dµn (y) dµn (x)
H W
−Rn
δ
Z
Rn
Z
δ−x
≤
Z
Rn
Z
0
log(2) dµn (y) dµn (x) +
−Rn
δ
δ
e−2(x+y) dµn (y) dµn (x).
δ−x
For the second integral we obtain,
−α
Z Rn Z 0
O (log(n)n ) if 1 < α < 2,
if α = 2,
e−2(x+y) dµn (y) dµn (x) = O log(n)2n−2
2+α
δ
δ−x
−
O n 2
if α > 2,
while for the first integral we have
Z
Rn
Z
δ−x
−2αRn
Z
log(2) dµn (y)dµn (x) ≤ αe
δ
−Rn
δ
23
Rn
eαδ dx = O log(n)n−α .
Therefore we conclude that
Z
−α
O (log(n)n ) if 1 < α < 2,
0
c (x, y) dµn (y)dµn (x) = O log(n)n−2
if α = 2,
H W
−∞
O n− 4+α
2
if α > 2,
Rn Z
0
c.
which yields the result for W
For W we first compute that
(x + y)ex+y
H(W (x, y)) = log 1 + ex+y −
1 + ex+y
x+y
= log 1 + e
− (x + y) + (x + y) 1 −
ex+y
1 + ex+y
(x + y)
−(x+y)
= log 1 + e−(x+y) +
≤
log
1
+
e
+ (x + y)e−(x+y) .
1 + ex+y
c (x, y)) and noting that log (1 + e−z ) ≤ e−2z for large
Comparing this upper bound to H(W
c.
enough z, the result follows from the computation done for W
With these two lemmas we now establish two important results on the approximations of
c (X, Y )
W . The first shows that if X and Y are independent with distribution µn , then W
−1
converges in expectation to W (X, Y ), faster than n .
Proposition 4.3. Let X, Y be independent with density µn and α > 1. Then, as n → ∞,
h
i
c (X, Y ) = O n− α+1
2
E W (X, Y ) − W
Proof. Since
h
i ZZ
c (X, Y ) =
E W (X, Y ) − W
Rn
c (x, y) dµn (y) dµn (x),
W (x, y) − W
−∞
c (x, y) ≤ 1 it follows from Lemma 4.1 that it is enough to consider the
and W (x, y) − W
integral
Z Z Rn
c (x, y) dµn (y) dµn (x).
W (x, y) − W
0
For this we note that
c (x, y) ≤ e−2(x+y) .
W (x, y) − W
Hence we obtain
Z Z Rn
0
c (x, y) dµn (y) dµn (x)
W (x, y) − W
−α
Z Rn
2
O (n )
≤
e−2x dµn (x) = O log(n)2 n−2
0
O n−2
Since all these terms are o(n−(α+1)/2 ), the result follows.
24
if 1 < α < 2
if α = 2
if α > 2.
c (X, Y )) converges in expectation to H(W (X, Y )), faster than
Next we show that also H(W
n−1 .
Proposition 4.4. Let X, Y be independent with density µn and α > 1. Then, as n → ∞,
−α
O (log(n)n ) if 1 < α < 2
i
h
c (X, Y )
= O log(n)3 n−2
E H (W (X, Y )) − H W
if α = 2
−2
O log(n)n
if α > 2.
Proof. Similar to the previous proof we now use Lemma 4.2 to show that it is enough to
consider the integral
Z Z Rn
c (x, y) dµn (y)dµn (x).
H (W (x, y)) − H W
0
Define
δW (x, y) =
1
,
+ ex+y
e2(x+y)
c (x, y) = W (x, y) + δW (x, y). Now fix x, y. Then, by the Taylor-Lagrange
and note that W
c (x, y) such that
theorem, there exists a W (x, y) ≤ cW (x, y) ≤ W
c (x, y) = H 0 (cW (x, y)) δW (x, y)
H (W (x, y)) − H W
c (x, y)
≤ H 0 (W (x, y)) + H 0 W
δW (x, y).
Next we compute that
H 0 (W (x, y)) = (x + y),
c (x, y) = log ex+y − 1 ,
H0 W
and
where log(ex+y − 1) ≤ (x + y) for all x + y ≥ 1. We now split the integral and bound it as
follows
Z Z Rn
c (x, y) dµn (y) dµn (x)
H (W (x, y)) − H W
0
Z 1 Z 1−x
c (x, y) dµn (y) dµn (x)
=
H (W (x, y)) − H W
0
0
Z 1 Z Rn
c (x, y) δW (x, y) dµn (y) dµn (x)
+
H (W (x, y)) − H W
0
Rn
Z
1−x
Z Rn
c (x, y) δW (x, y) dµn (y) dµn (x)
H (W (x, y)) − H W
+
1
Z
0
1 Z 1−x
≤ log(2)
0
Z 1Z
+2
0
Z
dµn (y) dµn (x)
0
Rn
(x + y)δW (x, y) dµn (y) dµn (x)
1−x
Rn Z Rn
+2
(x + y)δW (x, y) dµn (y) dµn (x)
1
0
25
Z
1 Z 1−x
≤ log(2)
ZZ
0
Rn
(x + y)δW (x, y) dµn (y) dµn (x).
dµn (y) dµn (x) + 4
0
0
The first integral is O(log(n)n−α ), while for the second we have
Z Z Rn
4
(x + y)δW (x, y) dµn (y) dµn (x)
0
−α
Z Rn
2
O (log(n)n ) if 1 < α < 2
e−2x dµn (x) = O log(n)3 n−2
≤ 8Rn
if α = 2
0
−2
O log(n)n
if α > 2.
Comparing these scaling to the ones from Lemma 4.2 we see that the former are dominating,
which finishes the proof.
4.2
Proofs for node degrees in the HSCM
In this section we give the proofs of Theorem 3.2 and Theorem 3.3. Denote by Di the degree
of node i and recall that Dn is the degree of a node, sampled uniformly at random. Since the
node labels are interchangeable we can, without loss of generality, consider D1 for Dn .
For Theorem
3.3 we
h
i use (35). We show that if X and Y are independent with distribution
c
µn , then E W (X, Y ) → ν. The final result will then follow from Proposition 4.3.
PnThe proof of Theorem 3.2 is more involved. Given the coordinates X1 , . . . Xn we have Dn =
j=2 W (X1 , Xj ). We follow the strategy from [64, Theorem 6.7] and construct a coupling
between Dn and a mixed Poisson random variable Pn , with mixing parameter nw
bn (X), where
w
bn (x) is given by (33) and X has distribution µn .
In general, a coupling between two random variables X and Y consists of a pair of new
b Yb ), with some joint probability distribution, such that X
b and Yb have
random variables (X,
the same marginal probabilities as, respectively X and Y . The advantage of a coupling is
that we can tune the joint distribution to fit our needs. For our proof we construct a coupling
b n , Pbn ), such that
(D
b n 6= Pbn = 0.
lim P D
n→∞
b n and Pbn have the same distribution as, respectively Dn and Pn we have
Hence, since D
b n = k, D
b n 6= Pbn ,
P (Dn = k) = P (Pn = k) + P D
from which it follows that |P (Dn = k) − P (Pn = k) | → 0. Finally, we show that
lim P (nw
bn (X) > k) = P (Y > k) ,
n→∞
where Y is a Pareto random variable with shape νβ and scale α, i.e. it has probability density
PY (y) given by (26). This implies that the mixed Poisson random variable Pn with mixing
parameter nw
bn (X) converges to a mixed Poisson D with mixing parameter Y , which proves
the result.
Before we give the proofs of the two theorems we first establish some technical results
needed to construct the coupling, required for Theorem 3.2, the proof of which is given in
Section 4.2.2.
26
4.2.1
Technical results on Poisson couplings and concentrations
We first establish a general result for couplings between mixed Poisson random variables,
where the mixing parameters converge in expectation.
Lemma 4.5. Let Xn and Yn be random variables such that,
lim E [|Xn − Yn |] = 0.
n→∞
Then, if Pn and Qn are mixed Poisson random variables with, respectively, parameters Xn
and Yn ,
P (Pn 6= Qn ) = O (E [|Xn − Yn |]) ,
and in particular
lim P (Pn 6= Qn ) = 0.
n→∞
Proof. Let an = E [|Xn − Yn |] and define the event An = {|Xn − Yn | ≤
lim P (|Xn − Yn | >
n→∞
√
√
an }. Then, since
E [|Xn − Yn |]
= 0,
√
n→∞
an
an ) ≤ lim
it is enough to show that
lim P (Pn 6= Qn , An ) = 0.
n→∞
√
Take P̂n to be mixed Poisson with parameter Xn + an . In addition, let Vn and Zn be mixed
√
√
Poisson with parameter min{Xn + an − Yn , 0} and an , respectively. Then, since on An we
√
have Xn + an > Yn we get, using Markov’s inequality,
P (Pn 6= Qn , An ) ≤ P Pn 6= Qn , Pn = P̂n , An + P P̂n 6= Pn , An
= P P̂n 6= Qn , An + P (Zn 6= 0, An ) = P (Vn 6= 0, An ) + P (Zn 6= 0, An )
√
√
≤ P (Vn ≥ 1, An ) + P (Zn ≥ 1, An ) ≤ E [|Xn + an − Yn |] + an
√
√
≤ E [|Xn − Yn |] + 2 an = O( an ).
Since an → 0 by assumption, this finishes the proof.
c (X, Y ) converges in expectation to w
Next we show that W
bn (X) when X, Y have distribution µn . We also establish an upper bound on the rate of convergence, showing that it
converges faster than n−1 .
Lemma 4.6. Let X, Y be independent random variables with density µn . Then, as n → ∞,
h
i
1+α
c (X, Y ) − w
E W
bn (X) = O log(n)n−α + n− 2
27
Proof. Recall that
(
ωn e−x if 0 ≤ x ≤ Rn
, where
w
bn (x) =
0
otherwise
Z Rn
1 − e−(α−1)Rn
ωn =
e−x dµn (x) =
.
βeRn
0
Hence, it follows that
h
Z
i
c (X, Y ) − w
E W
bn (X)
0
Z
Rn
=
c (x, y) dµn (y) dµn (x)
W
−∞ −∞
Z Rn Z Rn
+
−∞
0
c (x, y) − w
W
bn (x) dµn (y) dµn (x),
where the first integral is O(n−(α+1)/2 ) by Lemma 4.1.
To deal with the second integral we first compute
Rn
Z
Z
−x
dµn (y) dµn (x) = αRn e−2αRn = O log(n)n−α .
−∞
0
Therefore we have
Z Rn Z Rn
c (x, y) − w
W
bn (x) dµn (y)dµn (x)
−∞
0
Rn
Z
Z
−x
=
Z
Rn
Z
Rn
dµn (y)dµn (x) +
0
−∞
≤ O log(n)n
−α
Rn
Z
Z
Rn
+
0
= O log(n)n−α + ωn
Z
0
Rn
e−(x+y) − ωn e−x dµn (y)dµn (x)
−x
0
e−x e−y − ωn dµn (y)dµn (x)
e−y − ωn dµn (y)
0
We proceed to compute the last integral and show that it is O(n−(α+1)/2 ). For this we note
that since
e−y ≤ ωn ⇐⇒ y ≤ − log(ωn ),
we have
Z Rn
e−y − ωn dµn (y) =
0
Z
− log(ωn )
(ωn − e−y ) dµn (y) +
Z
Rn
(e−y − ωn ) dµn (y)
− log(ωn )
0
For the first integral we compute
Z
− log(ωn )
(ωn − e
−y
ωn1−α
ωn1−α
ωn
1
−
−
+
α
α−1
α
α−1
−αR
n
ωn e
e−αRn
ωn e−αRn
−
− ωn e−αRn =
−
.
α−1
β
β
−αRn
) dµn (y) = αe
0
=
e−αRn
β
28
Similar calculations yield
Z Rn
(e−y − ωn ) dµn (y) =
− log(ωn )
and hence,
Z Rn
e−Rn
ω 1−α e−αRn
− ωn − n
,
β
α−1
ωn e−αRn
ω 1−α e−αRn
1 −Rn
e
+ e−αRn − ωn −
− n
β
β
α−1
1 −Rn
≤
e
+ e−αRn − ωn .
β
e−y − ωn dµn (y) =
0
We now use this last upper bound, together with
ωn =
1 − e−(α−1)Rn
1 −Rn
=
e
− e−αRn ,
R
n
βe
β
to obtain
Z
ωn
Rn
ωn −Rn
+ e−αRn − ωn2
e
β
2
1
1
= 2 e−2Rn + e−α2Rn − 2 e−Rn − e−αRn
β
β
1+α
2 −(α+1)Rn
= 2e
= O n− 2 ,
β
e−y − ωn dµn (y) ≤
0
from which the result follows.
4.2.2
Proof of Theorem 3.2
b n , Pbn ), between Dn and Pn such that
We start with constructing the coupling (D
b n 6= Pbn = 0.
lim P D
n→∞
(45)
First, let Ij be the indicator that the edge (1, j) is present in Gn . Let Xn = X1 , . . . , Xn denote
the coordinates of the nodes. Then, conditioned on these, we have that
PnIj are independent
Bernoulli random variables with probability
W (X1 , Xj ), while D1 = j=2 Ij . Now, let Qn
P
be a mixed Poisson with parameter nj=2 W (X1 , Xj ). Then, see for instance [64, Theorem
b 1, Q
b n ) of D1 and Qn , such that
2.10], there exists a coupling (D
n
X
b 1 6= Q
b n Xn ≤
P D
W (X1 , Xj )2 .
j=2
Therefore, we get that
ZZ
b
b
P D1 6= Qn ≤ (n − 1)
Rn
W (x, y)2 dµn (y) dµn (x)
−(α−1)
Z Rn
2
O n
≤n
e−2x dµn (x) = O log(n)n−1
−∞
O n−1
−∞
29
if 1 < α < 2
.
if α = 2
if α > 2
Next, since X1 and Xj are independent for all 2 ≤ j ≤ n we use Proposition 4.3 together
with Lemma 4.6 to obtain
n
X
E
W (X1 , Xj ) − nw
bn (X1 )
j=2
≤ (n − 1)E [|W (X1 , X2 ) − w
bn (X1 )|] + E [w
bn (X1 )]
h
i
h
i
c (X1 , X2 ) + (n − 1)E W
c (X1 , X2 ) − w
bn (X1 ) + E [w
≤ (n − 1)E W (X1 , X2 ) − W
bn (X1 )]
α−1
= O n− 2 .
from which it follows that
lim E
n→∞
n
X
W (X1 , Xj ) − nw
bn (X1 ) = 0.
(46)
j=2
Now let Pbn be a mixed Poisson random variable with mixing parameter nw
bn (X1 ). Then by
(46) and Lemma 4.5
bn = 0
lim P Pbn 6= Q
n→∞
and (45) follows from
bn =
b n 6= Q
b n + P Pbn 6= Q
bn .
P D
6 Pbn ≤ P D
As a result we have that
b n = k − P Pbn = k
lim |P (Dn = k) − P (Pn = k)| = lim P D
n→∞
n→∞
b n = k, D
b n 6= Pbn ≤ lim P D
b n 6= Pbn = 0,
= lim P D
n→∞
n→∞
(47)
We will now prove that if X has density µn , then for any t > 0
lim |P (nw
bn (X) > t) − P (Y > t)| = 0.
n→∞
(48)
A sequence {Pn }n≥1 of mixed Poisson random variables, with mixing parameters Zn converges
to a mixed Poisson random variable P with mixing parameter Z, if Zn converge to Z in
distribution, see for instance [63]. Therefore, since D is mixed Poisson with mixing parameter
Y , (48) implies
lim P (Pn = k) = P (D = k)
n→∞
which, combined with (47), yields
lim P (Dn = k) = P (D = k) .
n→∞
30
To establish (48), first define
n = e−(α−1)Rn − e−2αRn
so that
nw
bn (x) =
√
n
−(α−1)Rn
1
−
e
= νn e−x (1 − n ) .
βeRn
Next, for all 0 ≤ x ≤ Rn we have that
νβ (1 − n ) ≤ nw
bn (x) ≤
√
νn (1 − n )
and hence
P X < log
P (nŵn (X) > t) = 1
0
Moreover, for any νβ (1 − n ) ≤ t ≤
√
nωn
t
if νβ (1 − n ) ≤ t ≤
if t < νβ (1 − n )
else.
√
νn (1 − n )
νn (1 − n )
nω Z log( nωt n )
n
P (nw
bn (X) > t) = P X < log
=
dµn (x)
t
−∞
α
α
νβ
−αRn nωn
−Rn α −α
(1 − n )α .
=e
= nωn e
t =
t
t
(49)
Now, fix t < νβ. Then, for large enough n it holds that t < νβ(1 − n ) in which case (48)
holds trivially, since both probabilities are 1. Hence we can assume, without loss of generality
√
that t ≥ νβ. Then for n large enough such t ≤ νn (1 − n ), it follows from (49) and (27)
that
α
α
νβ
νβ
|P (nw
bn (X) > t) − P (Y > t)| =
(1 − n )α −
t
t
α
νβ
=
(1 − (1 − n )α ) ≤ 1 − (1 − n )α ,
t
which converges to zero as n → ∞ and hence proves (48).
4.2.3
Proof of Theorem 3.3
First using Lemma 4.1, we compute that
Z Z Rn
h
i
c (X, Y ) = (n − 1)
c (x, y) dµn (y) dµn (x)
(n − 1)E W
W
ZZ
−∞
Rn
= (n − 1)
α−1
e−(x+y) dµn (y) dµn (x) + O n− 2
0
α−1
(n − 1) −2Rn
−αRn 2
=
e
1
−
e
+
O
n− 2
β2
α−1
α−1
2
1
=ν 1− 2
1 − e−αRn + O n− 2
= ν + O n− 2 ,
β n
31
where the last line follows since −1 < −α/2 < −(α − 1)/2.
Next recall (35)
E [Dn ] =
n
X
E [W (X1 , Xj )] = (n − 1)E [W (X, Y )] .
j=2
Therefore, using Lemma 4.6 we have
|E [Dn ] − ν| = |(n − 1)E [W (X, Y )] − ν|
h
i
h
i
c (X, Y ) + (n − 1) E W
c (X, Y ) − ν
≤ (n − 1)E W (X, Y ) − W
α−1
≤ O log(n)n−(α−1) + O n− 2 ,
which yields the result.
4.3
Proofs for graphon entropy
Here we derive the properties of the graphon entropy σ of our graphon W given by (21). We
first give the proof of Proposition 3.4, and then that of Theorem 3.5.
4.3.1
Proof of Proposition 3.4
Recall that, given a measure µ on some interval A ⊆ R with µ(A) < ∞ and a µ-integrable
function w(x), we consider the problem of maximizing σ[W, µ] under the constraint (36),
Z
w(x) =
W (x, y) dµ(y).
A
In particular we need to show that if a solution exists, it is given by (37). Therefore, suppose
there exists at least one graphon W which satisfies the constraint. For this we use the
technique of Lagrange multipliers from variation calculus [65, 66].
To set up the framework, let W denote the space of symmetric functions W : A×A → [0, 1]
which satisfy
ZZ
W (x, y) dµ(y) dµ(x) < ∞.
A2
Observe that W is a convex subset of the Banach space WR , of all symmetric, µ-integrable,
functions W : A × A → R and that the function w is an element of L1 (−∞, R) with respect
to the measure µ, which is also Banach. Denote the latter space by L1A,µ . We slightly abuse
notation and write σ for the functional
ZZ
H(W (x, y)) dµ(y) dµ(x) = σ[W, µ],
W 7→
A2
and define the functional F : WR → L1A,µ by
Z
F(W )(x) = w(x) −
W (x, y) dµ(y).
A
32
Then we need to solve the following Euler-Lagrange equation for some Lagrange multiplier
functional λ : L1A,µ → R,
∂
(σ(W ) − Λ ◦ F(W )) = 0,
∂W
with respect to the Fréchet derivative. By Riesz Representation Theorem, we have that for
any functional Λ : L1A,µ → R, there exists a λ ∈ L∞ , uniquely defined on A, such that
Z
Λ(f ) =
λ(x)f (x) dµ(x).
A
Hence our Euler-Lagrange equation becomes
Z
Z
∂
W (x, y) dµ(y) dµ(x) = 0.
λ(x) w(x) −
σ(W ) −
∂W
A
A
(50)
In addition, since W is symmetric we have
ZZ
ZZ
1
λ(x)W (x, y) dµ(y)dµ(x) =
(λ(x) + λ(y))W (x, y) dµ(y) dµ(x)
2 A2
A2
and hence, by absorbing the factor 1/2 into λ we can rewrite our Euler-Lagrange equation as
Z
ZZ
∂
σ(W ) −
λ(x)w(x) dµ(x) +
(λ(x) + λ(y))W (x, y) dµ(y) dµ(x) = 0.
(51)
∂W
A
A2
For the two derivatives we have
W (x, y)
∂σ(W )
= log(1 − W (x, y)) − log(W (x, y)) = − log
,
∂W
1 − W (x, y)
Z
ZZ
∂
λ(x)w(x) dµ(x) +
(λ(x) + λ(y))W (x, y) dµ(y) dµ(x) = λ(x) + λ(y).
∂W
A
A2
(52)
Hence, we need to solve the equation
W (x, y)
− log
= λ(x) + λ(y),
1 − W (x, y)
which gives (37).
There is, however, a small technicality related to the computation of the derivative (52).
This is caused by the fact that H is only defined on [0, 1]. In particular, for W1 , W2 ∈ W, it
could be that W1 + W2 > 1 on some subset C ⊆ A × A and hence H(W1 + W2 ) is not well
defined. To compute the Fréchet derivative we need H(W1 + εW2 ) to be well defined for any
W1 , W2 ∈ W and some, small enough, ε > 0. To this end we fix a 0 < δ < 1 and define
δ
,
Hδ (x) = H (1 − δ)x +
2
which is just H, stretched out to the interval (−δ/2(1 − δ), 1 + δ/2(1 − δ)). Similarly we define
σδ , using Hδ and consider the corresponding graphon entropy maximization problem. Then
33
we can compute ∂σδ /∂W , by taking a W 0 ∈ W and ε > 0 such that W +εW 0 < 1+δ/2(1−δ)).
Then Hδ (W + εW 0 ) is well-defined and using the chain rule we obtain
∂Hδ (W + εW 0 )
(1 − δ)(W + εW 0 ) + δ/2
0
,
= −(1 − δ)W log
∂ε
1 − (1 − δ)(W + εW 0 ) − δ/2
from which it follows that
∂σδ (W )
= −(1 − δ) log
∂W
(1 − δ)W + δ/2
1 − (1 − δ)W − δ/2
.
Therefore we have the following equation
(1 − δ)(W + δ/2
= λ(x) + λ(y),
−(1 − δ) log
1 − (1 − δ)W − δ/2
which leads to a solution Wδ of the form
λ(x)+λ(y)
1 + e (1−δ)
.
Wδ (x, y) =
λ(x)+λ(y)
(1−δ)
(1 − δ) 1 + e
1−
δ
2
Since δ < 1 it follows, using elementary algebra, that Wδ (x, y) ∈ [0, 1] for all x, y ∈ [0, 1].
From this we conclude that Wδ ∈ W. Moreover Wδ converges to (37) as δ → 0. Since σδ → σ
as δ → 0, we obtain the graphon that maximizes the entropy σ, where the function λ is
determined by the constraint (36).
For uniqueness, suppose there exist two solutions, λ1 and λ2 in L1A,µ , to (36) and for which
the graphon entropy is maximized. Let f (x) = λ1 (x) − λ2 (x) ∈ L1A,µ , so that λ1 = λ2 + f .
Since λ1 satisfies (50) it follows that, due to linearity of the derivative,
Z
∂
f (x)W (x, y) dµ(y) dµ(x) = 0.
∂W A
Now since
∂
∂W
Z
f (x)W (x, y) dµ(y) dµ(x) = f (x),
A
it follows that f = 0, µ almost everywhere on A and hence λ1 = λ2 , µ almost everywhere on
A.
4.3.2
Proof of Theorem 3.5
First note that Proposition 4.4 implies that the difference in expectation between H(W )
c ) converges to zero faster than log(n)/n, so that for the purpose of Theorem
and H(W
c . Hence we are left to show that the rescaled entropy
3.5 we can approximate W with W
c
nσ[W , µn ]/ log(n) converges to ν. By Lemma 4.2 the integration over all regimes except
[0, Rn ]2 goes to zero faster than log(n)/n and therefore we only need to consider integration
over [0, Rn ]2 . The main idea for the rest of the proof is that in this range
c (x, y) ≈ (x + y)W
c (x, y).
H W
34
Let us first compute the integral over [0, Rn ]2 of the right hand side in the above equation.
Z Z Rn
c (x, y) dµn (y) dµn (x)
(x + y)W
0
Z Z Rn
2 −2αRn
(x + y)e(α−1)(x+y) dy dx
=α e
0
=
2e−2(1+α)Rn
(α − 1)Rn e2αRn + e(α+1)Rn − e2Rn − e2αRn
β 2 (α − 1)
2Rn
2
= 2 e−2Rn + e−(α+1)Rn − 2
e−2αRn + e−2Rn
β
β (α − 1)
−2R
n
2Rn e
+ O(n−1 ) = νn−1 log(n) + O(n−1 )
=
β2
which implies that
n
lim
n→∞ log(n)
Rn
ZZ
c (x, y) dµn (y) dµn (x) = ν.
(x + y)W
(53)
0
Next we show that
Z Z Rn
n
c (x, y) − (x + y)W
c (x, y) dµn (y) dµn (x) = O log(n)−1 ,
H W
log(n)
0
(54)
which, together with (53), gives
n
lim
n→∞ log(n)
Rn
ZZ
c (x, y) dµn (y) dµn (x) = ν.
H W
(55)
0
We compute that
x+y
e
−1
−(x+y)
−(x+y)
c
H W (x, y) = e
(x + y) − (1 − e
) log
ex+y
c (x, y) − (1 − e−(x+y) ) log 1 − e−(x+y) .
= (x + y)W
(56)
Note that −(1 − e−z ) log (1 − e−z ) ≤ e−z for all z ≥ 0. Hence, it follows from (56) that, on
[0, Rn ]2 ,
c (x, y) − (x + y)W
c (x, y) = − 1 − e−(x+y) log 1 − e−(x+y) ≤ e−(x+y) = W
c (x, y),
H W
so that by Theorem 3.3,
Z Z Rn
n
c (x, y) − (x + y)W
c (x, y) dµn (y) dµn (x)
H W
log(n)
0
Z Z Rn
n
c (x, y) dµn (y) dµn (x) = O log(n)−1 .
≤
W
log(n) 0
To conclude, we have
Z Z Rn
c
c (x, y) dµn (y)dµn (x)
σ[W , µn ] =
H W
−∞
35
Rn
ZZ
=
c (x, y) dµn (y)dµn (x) + O n−α + log(n)n− α+1
2
H W
0
and hence, using (55),
c , µn ]
nσ[W
= ν.
n→∞
log(n)
lim
4.4
Proof of Theorem 3.6
In this section we first formalize the strategy behind the proof of Theorem 3.6, briefly discussed
in Section 3.6. This strategy relies on partitioning the interval An into non-overlapping
subintervals. We then construct a specific partition satisfying certain requirements, and finish
the proof of the theorem.
4.4.1
Averaging W by a partition of An
We follow the strategy from [41]. First recall that for a graph Gn generated by the HSCM,
n
n
S[Gn ] ≥ E [S[Gn |xn ]] =
E [H(W (x1 , x2 ))] =
σ[Gn ],
2
2
and hence
S ∗ [Gn ] ≥ σ[Gn ],
where S ∗ [Gn ] = S[Gn ]/ n2 denotes the normalized Gibbs entropy.
Therefore, the key ingredient is to find a matching upper bound. For this we partition
the range An = (−∞, Rn ] of our probability measure µn (x) into intervals and approximate
W (x, y) by its average over the box in which x and y lie.
To be more precise, let mn be any increasing sequence of positive integers, {ρt }0≤t≤mn be
such that
ρ0 = −∞,
and ρmn = Rn ,
and consider the partition of (−∞, Rn ] given by
It = (ρt−1 , ρt ]
for 1 ≤ t ≤ mn .
Now define Jn (x) = t ⇐⇒ x ∈ It , and let Mi be the random variable Jn (Xi ), where
Xi has density function µn for any vertex i. The value of Mi equal to t indicates that
vertex i happens to lie within interval It . Denoting the vector of these random variable by
Mn = M1 , . . . , Mn , and their entropy by S[Mn ], we have that
S[Gn ] ≤ S[Mn ] + S[Gn |Mn ]
Z Z Rn
n
f (x, y)) dµn (y) dµn (x),
≤ nS[M ] +
H(W
2
−∞
f (x, y) is the average of W over the square IJ (x) × IJ (y) . That is,
where W
n
n
Z
Z
fn (x, y) = 1
W
W (u, v) dµn (v) dµn (u),
µx,y IJn (x) IJn (y)
36
(57)
with
Z
Z
µx,y =
dµn (v) dµn (u),
IJn (x)
IJn (y)
the measure of the box to which (x, y) belongs.
fn approximates W . More specifically,
The first step in our proof is to investigate how well W
fn , µn ]| scales, depending on the
we want to understand how the difference |σ[W, µn ] − σ[W
specific partition. Note that for any partition ρt of An into mn intervals we have that
S[M ] = −
mn
X
P (M = t) log (P (M = t)) ≤ log(mn ),
t=1
where the upper bound is achieved on the partition which is uniform according to measure
µn . Since
n nS[M ]
2S[M ]
=
,
n
log(n) 2
log(n)(1 − 1/n)
it is enough to find a partition ρt , with log(mn ) = o(log(n)), such that
n
fn , µn ]| = 0,
|σ[W, µn ] − σ[W
lim
n→∞ log(n)
(58)
This then proves Theorem 3.6, since
n
n
lim
|S ∗ [Gn ] − σ[Gn ]| = lim
(S ∗ [Gn ] − σ[Gn ])
n→∞ log(n)
n→∞ log(n)
n
n nS[M ]
fn , µn ]|
+
|σ[W, µn ] − σ[W
log(n) n2
log(n)
≤ lim
n→∞
≤ lim
n→∞
4.4.2
!
2 log(mn )
n
fn , µn ]|
+
|σ[W, µn ] − σ[W
log(n)(1 − 1/n) log(n)
= 0.
Constructing the partition
We will take I1 = (−∞, −Rn ] and partition the remaining interval [−Rn , Rn ] into log(n)2
equal parts. To this end, let n be sufficiently large so that n ≥ νβ 2 , take mn = dlog(n)2 e + 1,
and define the partition ρt by
ρ0 = −∞,
ρ1 = −Rn ,
and ρt = ρt−1 +
2Rn
mn − 1
for all t = 2, . . . , mn .
Note that log(mn ) = O(log log n) = o(log(n) so that all that is left, is to prove (58). In
addition,
Z −Rn Z Rn
n
f (x, y)) dµn (y) dµn (x) = O log(n)−1 n1−α ,
H(W
log(n) −∞ −∞
f with W . Hence it follows that in order to establish (58)
and the same holds if we replace W
we only need to consider the integral over the square [−Rn , Rn ] × [−Rn , Rn ]. That is, we need
to show
Z Z Rn
n
f (x, y)
H (W (x, y)) − H W
dµn (y) dµn (x) = 0,
(59)
lim
n→∞ log(n)
−Rn
37
f , µn ], based on the mean value theorem which states
For this we compare σ[W, µn ] and σ[W
that
for any a ≤ b,
∃ c ∈ [a, b],
such that H(b) − H(a) = H 0 (c) (a − b).
Here |H 0 (c)| = | log(c) − log(1 − c)|, and due to the symmetry of H we get, for any 0 < a ≤
c ≤ b < 1,
min{H(a), H(b)}
|H 0 (c)| ≤
.
(60)
min{a, b, 1 − a, 1 − b}
Note that
0<
min
u∈IMn (x) , v∈IMn (y)
fn (x, y) ≤
W (u, v) ≤ W
max
u∈IMn (x) , v∈IMn (y)
W (u, v) < 1,
(61)
for all x, y ∈ [−Rn , Rn ]. In addition, for all x, y ∈ [−Rn , Rn ] and (u, v) ∈ IMn (x) × IMn (y) we
have |x + y − u − v| ≤ 2Rn /mn ≤ 2/ log(n) and thus |1 − exp(x + y − u − v)| ≤ 3|x + y − u − v|
by the mean value theorem. Therefore we have
u+v−x−y
|e
− 1| |1 − ex+y−u−v |
|eu+v − ex+y |
≤ min
,
|W (x, y) − W (u, v)| =
(1 + ex+y )(1 + eu+v )
1 + eu+v
1 + eu+v
≤ 3|u + v − x − y| min{W (u, v), 1 − W (u, v)},
By symmetry we obtain a similar upper bound with W (x, y) instead of W (u, v) and hence
we conclude
|W (x, y) − W (u, v)| ≤ 3|u + v − x − y| min{W (x, y), 1 − W (x, y), W (u, v), 1 − W (u, v)}. (62)
Then, for any x + y ≤ u + v there are c, d such that x + y ≤ c + d ≤ u + v, so by (60) and
(62) we get
|H(W (x, y)) − H(W (u, v))| = |H 0 (W (c, d))(W (x, y) − W (u, v))|
min{H(W (x, y)), H(W (u, v))} |W (x, y) − W (u, v)|
≤
min{W (x, y), 1 − W (x, y), W (u, v), 1 − W (u, v)}
≤ 3|x + y − u − v| · min{H(W (x, y)), H(W (u, v))}.
(63)
Next, for the partition ρt we have
|It | =
2Rn
2Rn
log(n) − log(νβ 2 )
1
≤
=
≤
,
2
2
mn − 1
log(n)
log(n)
log(n)
for t ≥ 2. In addition, (61) implies that for (x, y) ∈ It × Is , with t, s ≥ 2,
fn (x, y), W (x, y) ≤ W (ρt , ρs ),
W (ρt−1 , ρs−1 ) ≤ W
fn (x, y) = W (xs,t , ys,t ) on Is × It .
and thus there is a pair xs,t , ys,t ∈ Is × It such that W
Therefore, using (63), we get
fn (x, y))| = |H(W (x, y)) − H(W (xs,t , ys,t ))|
|H(W (x, y)) − H(W
≤ 3|x + y − xs,t − ys,t |H(W (x, y))
38
≤ 3(|It | + |Is |)H(W (x, y)) ≤
6
H(W (x, y))
log(n)
Finally, integrating the above equation over the whole square [−Rn , Rn ] × [−Rn , Rn ] we
obtain
Z Z Rn
6
f (x, y)
H (W (x, y)) − H W
dµn (y) dµn (x) ≤
σ[W, µn ],
log(n)
−Rn
which proves (59) since
lim
n→∞
n
6n
fn , µn ]| ≤ lim
|σ[W, µn ] − σ[W
σ[W, µn ] = 0
n→∞ (log n)2
log(n)
thanks to Theorem 3.5.
Acknowledgments This work was supported by the ARO grant No. W911NF-16-1-0391
and by the NSF grant No. CNS-1442999.
References
[1] S Boccaletti, V Latora, Y Moreno, M Chavez, and D.-U. Hwanga. Complex Networks:
Structure and Dynamics. Phys Rep, 424:175–308, 2006. doi:10.1016/j.physrep.2005.
10.009.
[2] M E J Newman. Networks: An Introduction. Oxford University Press, Oxford, 2010.
[3] Albert-László Barabási. Network science. Cambridge University Press, Cambridge, UK,
2016.
[4] Ray Solomonoff and Anatol Rapoport. Connectivity of random nets. B Math Biophys,
13(2):107–117, 1951. doi:10.1007/BF02478357.
[5] E. N. Gilbert. Random Graphs. Ann Math Stat, 30(4):1141–1144, 1959. doi:10.1214/
aoms/1177706098.
[6] P Erdős and A Rényi. On Random Graphs. Publ Math, 6:290–297, 1959.
[7] Edward A Bender and E Rodney Canfield. The asymptotic number of labeled graphs
with given degree sequences. J Comb Theory, Ser A, 24(3):296–307, 1978. doi:10.1016/
0097-3165(78)90059-6.
[8] M Molloy and B Reed. A Critical Point for Random Graphs With a Given Degree
Sequence. Random Struct Algor, 6:161–179, 1995.
[9] Amogh Dhamdhere and Constantine Dovrolis. Twelve Years in the Evolution of the
Internet Ecosystem. IEEE/ACM Trans Netw, 19(5):1420–1433, 2011. doi:10.1109/
TNET.2011.2119327.
[10] M. E. J. Newman. Clustering and preferential attachment in growing networks. Phys
Rev E, 64(2):025102, 2001. doi:10.1103/PhysRevE.64.025102.
39
[11] Fan Chung and Linyuan Lu. Connected Components in Random Graphs with Given
Expected Degree Sequences. Ann Comb, 6(2):125–145, 2002. doi:10.1007/PL00012580.
[12] Fan Chung and Linyuan Lu. The average distances in random graphs with given expected degrees. Proc Natl Acad Sci USA, 99(25):15879–82, 2002. doi:10.1073/pnas.
252631999.
[13] J Park and M E J Newman. Statistical Mechanics of Networks. Phys Rev E, 70:66117,
2004. doi:10.1103/PhysRevE.70.066117.
[14] Ginestra Bianconi. The entropy of randomized network ensembles. EPL, 81(2):28005,
2008. doi:10.1209/0295-5075/81/28005.
[15] Diego Garlaschelli and Maria Loffredo. Maximum likelihood: Extracting unbiased information from complex networks. Phys Rev E, 78(1):015101, 2008. doi:10.1103/
PhysRevE.78.015101.
[16] Tiziano Squartini and Diego Garlaschelli. Analytical maximum-likelihood method to
detect patterns in real networks. New J Phys, 13(8):083001, 2011. doi:10.1088/
1367-2630/13/8/083001.
[17] Paul W Holland and Samuel Leinhardt. An Exponential Family of Probability Distributions for Directed Graphs. J Am Stat Assoc, 76(373):33–50, 1981. doi:10.1080/
01621459.1981.10477598.
[18] Kartik Anand and Ginestra Bianconi. Entropy Measures for Networks: Toward an
Information Theory of Complex Topologies. Phys Rev E, 80:045102(R), 2009. doi:
10.1103/PhysRevE.80.045102.
[19] Tiziano Squartini, Joey de Mol, Frank den Hollander, and Diego Garlaschelli. Breaking
of Ensemble Equivalence in Networks. Phys Rev Lett, 115(26):268701, 2015. doi:10.
1103/PhysRevLett.115.268701.
[20] Sourav Chatterjee, Persi Diaconis, and Allan Sly. Random graphs with a given degree
sequence. Ann Appl Probab, 21(4):1400–1435, 2011. doi:10.1214/10-AAP728.
[21] Guido Caldarelli, Andrea Capocci, P. De Los Rios, and Miguel Angel Muñoz. Scale-Free
Networks from Varying Vertex Intrinsic Fitness. Phys Rev Lett, 89(25):258702, 2002.
doi:10.1103/PhysRevLett.89.258702.
[22] Marián Boguñá and Romualdo Pastor-Satorras. Class of Correlated Random Networks
with Hidden Variables. Phys Rev E, 68:36112, 2003. doi:10.1103/PhysRevE.68.036112.
[23] Kartik Anand, Dmitri Krioukov, and Ginestra Bianconi. Entropy distribution and condensation in random networks with a given degree distribution. Phys Rev E, 89(6):062807,
2014. doi:10.1103/PhysRevE.89.062807.
[24] Konstantin Zuev, Fragkiskos Papadopoulos, and Dmitri Krioukov. Hamiltonian dynamics of preferential attachment. J Phys A Math Theor, 49(10):105001, 2016. doi:
10.1088/1751-8113/49/10/105001.
40
[25] Michael Evans and Jeffrey S. Rosenthal. Probability and statistics: The science of uncertainty. W.H. Freeman and Co, New York, 2009.
[26] Kimberly Claffy, Young Hyun, Ken Keys, Marina Fomenkov, and Dmitri Krioukov. Internet Mapping: From Art to Science. In 2009 Cybersecurity Appl Technol Conf Homel
Secur, 2009. doi:10.1109/CATCH.2009.38.
[27] Dmitri Krioukov, Fragkiskos Papadopoulos, Maksim Kitsak, Amin Vahdat, and Marián
Boguñá. Hyperbolic Geometry of Complex Networks. Phys Rev E, 82:36106, 2010.
doi:10.1103/PhysRevE.82.036106.
[28] E T Jaynes. Information Theory and Statistical Mechanics. Phys Rev, 106(4):620–630,
1957. doi:10.1103/PhysRev.106.620.
[29] J. Shore and R. Johnson. Axiomatic derivation of the principle of maximum entropy
and the principle of minimum cross-entropy. IEEE Trans Inf Theory, 26(1):26–37, 1980.
doi:10.1109/TIT.1980.1056144.
[30] Y. Tikochinsky, N. Z. Tishby, and R. D. Levine. Consistent Inference of Probabilities
for Reproducible Experiments. Phys Rev Lett, 52(16):1357–1360, 1984. doi:10.1103/
PhysRevLett.52.1357.
[31] John Skilling. The Axioms of Maximum Entropy, In Maximum-Entropy and Bayesian
Methods in Science and Engineering, pages 173–187. Springer Netherlands, Dordrecht,
1988. doi:10.1007/978-94-009-3049-0_8.
[32] Claude Edmund Shannon. A Mathematical Theory of Communication. Bell Syst Tech
J, 27(3):379–423, 1948. doi:10.1002/j.1538-7305.1948.tb01338.x.
[33] Dmitri Krioukov.
Clustering Implies Geometry in Networks.
116(20):208302, 2016. doi:10.1103/PhysRevLett.116.208302.
Phys Rev Lett,
[34] Jagat Narain Kapur. Maximum-entropy models in science and engineering. Wiley, New
Delhi, 1989.
[35] David J Aldous. Representations for partially exchangeable arrays of random variables.
J Multivar Anal, 11(4):581–598, 1981. doi:10.1016/0047-259X(81)90099-3.
[36] Persi Diaconis and Svante Janson. Graph Limits and Exhcangeable Random Graphs.
Rend di Matemtica, 28:33–61, 2008.
[37] Olav Kallenberg. Foundations of modern probability. Springer, New York, 2002.
[38] Cosma Rohilla Shalizi and Alessandro Rinaldo. Consistency under sampling of exponential random graph models. Ann Stat, 41(2):508–535, 2013. doi:10.1214/12-AOS1044.
[39] Dmitri Krioukov and Massimo Ostilli. Duality between equilibrium and growing networks. Phys Rev E, 88(2):022808, 2013. doi:10.1103/PhysRevE.88.022808.
[40] László Lovász and Balázs Szegedy. Limits of dense graph sequences. J Comb Theory,
Ser B, 96(6):933–957, 2006. doi:10.1016/j.jctb.2006.05.002.
41
[41] Svante Janson. Graphons, cut norm and distance, couplings and rearrangements. NYJM
Monogr, 4, 2013.
[42] Douglas N. Hoover. Relations on Probability Spaces and Arrays of Random Variables.
Technical report, Institute for Adanced Study, Princeton, NJ, 1979.
[43] S N Dorogovtsev, J F F Mendes, and A N Samukhin. Structure of Growing Networks with Preferential Linking. Phys Rev Lett, 85(21):4633–4636, 2000. doi:10.1103/
PhysRevLett.85.4633.
[44] P L Krapivsky, S Redner, and F Leyvraz. Connectivity of Growing Random Networks.
Phys Rev Lett, 85(21):4629–4632, 2000. doi:10.1103/PhysRevLett.85.4629.
[45] Rodrigo Aldecoa, Chiara Orsini, and Dmitri Krioukov. Hyperbolic graph generator.
Comput Phys Commun, 196:492–496, 2015. doi:10.1016/j.cpc.2015.05.028.
[46] François Caron and Emily B. Fox. Sparse graphs using exchangeable random measures.
2014. arXiv:1401.1137.
[47] Victor Veitch and Daniel M. Roy. The Class of Random Graphs Arising from Exchangeable Random Measures. 2015. arXiv:1512.03099.
[48] Christian Borgs, Jennifer T Chayes, Henry Cohn, and Nina Holden. Sparse exchangeable
graphs and their limits via graphon processes. 2016. arXiv:1601.07134.
[49] László Lovász. Large Networks and Graph Limits. American Mathematical Society,
Providence, RI, 2012.
[50] David J Aldous. Exchangeability and related topics, In Ecole d’Ete de Probabilites
de Saint-Flour XIII, pages 1–198. Springer, Berlin, Heidelberg, 1983. doi:10.1007/
BFb0099421.
[51] David D McFarland and Daniel J Brown. Social distance as a metric: A systematic
introduction to smallest space analysis, In Bonds of Pluralism: The Form and Substance
of Urban Social Networks, pages 213–252. John Wiley, New York, 1973.
[52] Katherine Faust. Comparison of methods for positional analysis: Structural and general equivalences. Soc Networks, 10(4):313–341, 1988. doi:10.1016/0378-8733(88)
90002-0.
[53] J. M. McPherson and J. R. Ranger-Moore. Evolution on a Dancing Landscape: Organizations and Networks in Dynamic Blau Space. Soc Forces, 70(1):19–42, 1991.
doi:10.1093/sf/70.1.19.
[54] Peter D Hoff, Adrian E Raftery, and Mark S Handcock. Latent Space Approaches to
Social Network Analysis. J Am Stat Assoc, 97(460):1090–1098, 2002. doi:10.1198/
016214502388618906.
[55] Béla Bollobás, Svante Janson, and Oliver Riordan. The phase transition in inhomogeneous random graphs. Random Struct Algor, 31(1):3–122, 2007. doi:10.1002/rsa.
20168.
42
[56] Hamed Hatami, Svante Janson, and Balázs Szegedy. Graph properties, graph limits and
entropy. 2013. arXiv:1312.5626.
[57] Sourav Chatterjee and S.R.S. Varadhan. The large deviation principle for the Erds-Rényi
random graph. Eur J Comb, 32(7):1000–1017, 2011. doi:10.1016/j.ejc.2011.03.014.
[58] Sourav Chatterjee and Persi Diaconis. Estimating and understanding exponential random graph models. Ann Stat, 41(5):2428–2461, 2013. doi:10.1214/13-AOS1155.
[59] Charles Radin and Lorenzo Sadun. Singularities in the Entropy of Asymptotically Large
Simple Graphs. J Stat Phys, 158(4):853–865, 2015. doi:10.1007/s10955-014-1151-3.
[60] Christian Borgs, Jennifer T Chayes, László Lovász, Vera T. Sós, and Katalin Vesztergombi. Convergent sequences of dense graphs I: Subgraph frequencies, metric properties
and testing. Adv Math, 219(6):1801–1851, 2008. doi:10.1016/j.aim.2008.07.008.
[61] Ginestra Bianconi. Entropy of Network Ensembles. Phys Rev E, 79:36114, 2009.
[62] A Barvinok and J A Hartigan. The number of graphs and a random graph with a given
degree sequence. Random Struct Algor, 42(3):301–348, 2013. doi:10.1002/rsa.20409.
[63] Jan Grandell. Mixed Poisson processes. Chapman & Hall/CRC, London, UK, 1997.
[64] Remco van der Hofstad. Random graphs and complex networks. Cambridge University
Press, Cambridge, UK, 2016.
[65] Jean Pierre Aubin. Applied functional analysis. John Wiley & Sons, Inc, New York,
2000.
[66] I. M. Gelfand and S. V. Fomin. Calculus of variations. Dover Publications, New York,
2000.
43
| 10math.ST
|
arXiv:1208.4035v1 [cs.PL] 16 Aug 2012
QIRAL: A High Level Language for Lattice
QCD Code Generation
Denis Barthou
Gilbert Grosdidier
University of Bordeaux,
Bordeaux, France
[email protected]
Laboratoire de l’Accélerateur Linéaire,
Orsay, France
[email protected]
Michael Kruse
Olivier Pene
INRIA Saclay / ALCHEMY,
Orsay, France
[email protected]
Laboratoire de Physique Théorique,
Orsay, France
[email protected]
Claude Tadonki
Mines ParisTech/CRI - Centre de Recherche en Informatique,
Fontainebleau, France
[email protected]
Abstract
Quantum chromodynamics (QCD) is the theory of subnuclear physics, aiming at modeling the strong nuclear force, which is responsible for the interactions of nuclear particles.
Lattice QCD (LQCD) is the corresponding discrete formulation, widely used for simulations. The computational demand for the LQCD is tremendous. It has played a role in the
history of supercomputers, and has also helped defining their future. Designing efficient
LQCD codes that scale well on large (probably hybrid) supercomputers requires to express
many levels of parallelism, and then to explore different algorithmic solutions. While algorithmic exploration is the key for efficient parallel codes, the process is hampered by the
necessary coding effort.
We present in this paper a domain-specific language, QIRAL, for a high level expression
of parallel algorithms in LQCD. Parallelism is expressed through the mathematical structure of the sparse matrices defining the problem. We show that from these expressions and
from algorithmic and preconditioning formulations, a parallel code can be automatically
generated. This separates algorithms and mathematical formulations for LQCD (that belong to the field of physics) from the effective orchestration of parallelism, mainly related
to compilation and optimization for parallel architectures.
1
Introduction
Quantum Chromodynamics (QCD) is the theory of strong subnuclear interactions[13]. Lattice
QCD (LQCD) is the numerical approach to solve QCD equations. LQCD simulations are
extremely demanding in terms of computing power, and require large parallel and distributed
machines. At the heart of the simulation, there is an inversion problem:
Ax = b,
(1)
where A is a large sparse (also sparse and implicit) matrix, called Dirac matrix, b is a known
vector and x is the unknown. To model reality accurately, A should be of size around 232 × 232
elements at least. Current simulations on supercomputers handle sizes up to 224 × 224 elements. High performance LQCD codes on multicores, multinode and hybrid (using GPUs)
architectures are quite complex to design: performance results from the interplay between the
algorithms chosen to solve the inversion and the parallelism orchestration on the target architecture. Exploring the space of algorithms able to solve the inversion, such as mixed precision
or communication-avoiding algorithms, aggressive preconditioning, deflation techniques or any
1
QIRAL: A High-Level Language for LQCD Code Generation
Barthou et al.
combination of these is essential in order to reach higher levels of performance. Many handtuned, parallel libraries dedicated to LQCD, such as Chroma [5] or QUDA [2] propose building
blocks for these algorithms, so as to ease their development. Parallelism does not stem from
the algorithm themselves but from the structure of the sparse matrices involved in the computation. However, there are a number of shortcomings to the library approach: First, any
significant evolution of the hardware requires to tune the library to the new architecture. This
can lead to change the grain of parallelism (as an adaptation to cache sizes for instance) or
change the data layout. Then, sparse matrices for LQCD, in particular the matrix A, have
a very regular structure and library functions take advantage of it. Combining different algorithms, such as preconditioners, leads to structural changes in matrices that are not supported
by these libraries. These two limitations hinder considerably the time necessary to develop a
parallel/distributed code.
We propose in this paper a high-level Domain-Specific Language, QIRAL, to model LQCD
problems and enable automatic parallel code generation. This novel approach makes possible
the exploration and test of new algorithms for LQCD. Sparse matrices in LQCD can be structured using algebra operators on dense matrices and the key idea is to use this structure to
express parallelism. This extends ideas proposed in SPIRAL [8], the library generator for DSP
algorithms to more complex codes and matrices. The problem formulation is first presented in
Section 2, then the language is described in Section 3 and its implementation in Section 4.
2
Lattice QCD Description
The principle of LQCD is to discretize space-time and describe the theory on the resulting
lattice. The lattice has to contain a very large number of sites since it has to be fine grained and
describe large enough volumes. LQCD exists since 1974. Huge progress in hardware, software
and algorithms have been achieved, but it is not enough to really sit on the parameters of nature.
The light quarks, “u” and “d” are still described by quarks heavier than in nature while the
heavy “b” quark is described as lighter. This implies systematic errors in our results. To break
this limitation, several orders of magnitudes will be needed in the computing power and related
resources. It will demand new hardware with several level of parallelism, and make coding more
and more complicated. Our goal is to provide tools helping to face this new situation.
The heaviest part in the LQCD calculation is to generate a large Monte-Carlo sample of very
large files (field configurations) according to an algorithm named “Hybrid Monte-Carlo”. It is a
Markovian process, every step of which takes several hours on the most powerful computers. The
algorithm is complex but it spends most of its time in inverting very large linear systems which
depend on the field configuration. Typically one will deal with matrices with billions of lines and
columns. The second heaviest part in the calculation is to compute ”quark propagators” which
again boils down to solving large linear systems of the same type[7]. Therefore we will concentre
on this task. The matrices involved in these computation represent 4D stencil computations,
where each vertex is updated by the value of its 8 neighbours. Their structure is therefore very
regular, statically known, and it is used in the following section.
3
A High-Level Domain-Specific Language
QIRAL is a high level language for the description of algorithms and for the definition of
matrices, vectors and equations specific to LQCD. The objective of the algorithmic part is
to define algorithms independently of the expression of sparse matrices used in LQCD. The
2
QIRAL: A High-Level Language for LQCD Code Generation
Barthou et al.
Dirac = IL⊗C⊗S
+ 2 ∗ i ∗ κ ∗ µ ∗ IL⊗C ⊗ γ5
M
−d
((J
⊗ IC ) ∗
U (d)[s]) ⊗ (IS + γ[d])
L
s∈L
M
d
+κ∗
((JL ⊗ IC ) ∗
U (−d)[s]) ⊗ (IS − γ[d])
d∈D
s∈L
+κ∗
X
d∈D
X
Figure 1: Definition of Dirac matrix on a Lattice L in QIRAL.
objective of the system of definitions and equations is to define properties and structure sparse
matrices in order to be able to find parallelism. While the algorithmic part uses straightforward
operational semantics, the equational part defines a rewriting system.
For the sake of simplicity, QIRAL is a subset of LATEX, meaning that the QIRAL input can
be either compiled into a pdf file, for rendering purposes, or compiled into executable codes.
Algorithms and definitions correspond to different predefined LATEX environments.
Variables, Types: Variables and constants used in QIRAL are vectors (denoted by the type
V) of any length, matrices of any size (M), complex (C) or real numbers (R). Besides, counted
loops are indexed by variables of type index, iterating over a domain (denoted indexset).
Index variables are potentially multidimensional, and integers are a particular case of index
value. Functions of any number of argument of these types can be also defined. The size of a
vector is defined through the size of its index set. The size of an index set can be left undefined.
If V1 is the vector and IS is an index set, V1[IS] denotes the subvector indexed only by IS.
Matrices and vectors cannot be manipulated or defined element-wise. Instead, matrices and
vectors are built using either constant, predefined values such as identity IIS (for the identity
on an index set IS), or operators such as +, −, ∗ and the transposition, conjugate, direct sum
and tensor product. The tensor product ⊗ and direct sum ⊕ are defined by:
A ⊗ B = [aij B]ij ,
A⊕B =
A
0
0
B
.
These two operators define parallel operations and help defining the sparse matrices of LQCD.
Definition and Equations: All LQCD knowledge is given as a system of definition (for
matrices and vectors) and a system of equations on these values. In particular, the structure
of sparse matrices, explicitly defining where non-zero elements are located, is given in these
definitions. This structure is then propagated in the algorithms. Figure 1 defines the sparse
Dirac matrix. Declarations of index sets L, C, S, D are not shown here, as well as the declaration
of the constant matrices used here.
The direct sums indexed by the vertices of the lattice (L) define block-diagonal matrices.
These diagonals are then shifted by d or −d columns by a permutation matrix JLd (predefined).
Algorithms, Preconditioners: Many algorithms proposed for solving Equation 1 are
Krylov methods. Figure 2 presents two algorithms in QIRAL. The first is a variant of the
conjugate gradient and is representative of iterative methods. The second one, Schur complement method, is a preconditioner: Computing x, solution of Equation 1 is achieved by
computing two solutions to smaller problems.
The keyword Match helps to define the algorithm as a rewriting rule: Whenever the statement in the Match condition is found, it can be rewritten by the algorithm. For the SCHUR
3
QIRAL: A High-Level Language for LQCD Code Generation
Barthou et al.
preconditioner, this rewriting can be performed at the condition that the requirement is fulfilled: P1 AP1t has to be invertible. Using identities defined through the equation system, the
rewriting system proves here that indeed, the requirement is valid, when matrix A is Dirac
matrix and P1 is a projection matrix keeping only vertices of the lattice with even coordinates
(even-odd preconditioning). QIRAL is expressive enough to represent different variants of conAlgo 1: Conjugate Gradient, Normal
Resolution [CGNR]
Input
: A ∈ M, b ∈ V, ǫ ∈ R
Output : x ∈ V
Match
: x = A−1 ∗ b
Var
: r, p, Ap, z ∈ V, α, β, nr , nz , nz1 ∈ R
r = b ;
z = A† ∗ r;
p =z ;
x =0 ;
nz = (z | z) ;
nr = (r | r) ;
while (nr > ǫ) do
Ap = A ∗ p;
α = nz /(Ap | Ap) ;
x =x+α∗p ;
r = r − α ∗ Ap ;
z = A† ∗ r ;
nz1 = (z | z) ;
β = nz1 /(nz ) ;
p= z+β∗p ;
nz = nz1 ;
nr = (r | r) ;
Algo 2: Definition of Schur complement
method [SCHUR]
Input
Output
Match
Var
:
:
:
:
A, P1 , P2 ∈ M, b ∈ V
x ∈ V
x = A−1 ∗ b
v1 , v2 , x1 , x2 ∈ V, D11 , D12 , D21 , D22 ∈ M
Require : invertible(P1 ∗ A ∗ P1t )
D21 = P2 ∗ A ∗ P1t ;
D11 = P1 ∗ A ∗ P1t ;
D22 = P2 ∗ A ∗ P2t ;
D12 = P1 ∗ A ∗ P2t ;
v1 = P1 ∗ b ;
v2 = P2 ∗ b ;
−1
−1
x2 = (D22 − D21 ∗ D11 ∗ D12 )−1 ∗ (v2 − D21 ∗ D11 ∗ v1 ) ;
−1
x1 = D11 ∗ (v1 − D12 ∗ x2 ) ;
t ∗ x + Pt ∗ x ;
x = P1
1
2
2
Figure 2: Two algorithms. On the left: A variant of the conjugate gradient method; on the
right: A preconditioner, the Schur complement method.
jugate gradient, BiCGSTAB, methods with restart (such as GCR), and other methods proposed
by physicists.
4
From QIRAL to parallel code
The QIRAL compiler takes as input a file describing algorithms and equations and generates
parallel code. The phases are described in 3 and presented in detail in the following.
Figure 3: Overview of QIRAL compilation chain
Rewriting System Generation: From its current LATEX implementation, the QIRAL input
is translated into a set of rewriting rules and equations, using Maude framework [3]. Maude
is a multi-purpose rewriting system, handling equational rewriting rules and reflection. More
4
QIRAL: A High-Level Language for LQCD Code Generation
Barthou et al.
precisely, definitions are translated by a first phase, tex2maude into equations (or conditional
equations) and algorithms into rewriting rules. The output file contains domain specific definitions (e.g. the Dirac operator from Figure 1), the desired program (in our case x = Dirac−1 b),
and all algorithms and preconditioners we want to apply.
Definitions have a semantic similar to a rewriting rule: the left hand side is rewritten into
the right hand side. To ensure the existence of a normal form, the system of equation has to be
convergent and confluent. So far, our implementation does not use a tool such as an automatic
Church-Rosser Checker [4] . Algorithms are translated into conditional rules, applied is their
prerequisite are checked. The left hand side of the rule correspond to the Match clause and the
algorithm itself is the right hand side.
This rewriting system is merged with another one defining general algebraic properties, code
generation rewriting rules and code optimizations. The main phase of the QIRAL compiler is
therefore described through rewriting system, following works such as Stratego [11]. QIRAL
is a static strongly type language, vectors and matrices are defined by their index sets. Type
checking is the first analysis achieved by this main phase.
Applying Algorithms and Simplifications: The user provides the list of algorithms to
compose and the initial program (here, we focused on the equation x = Dirac−1 b). Exploring
different algorithms and sequences of algorithms boils down to change the list given to the QIRAL compiler. Checking that algorithm requirements is automatically achieved by the rewriting
system, using the equational theory provided through the system of definitions. The equational
system, using both equations coming from LQCD definitions and algebraic properties involving
the different operators, simplifies terms that are equal to zero.
The result is a program where the matrix A of the algorithms has been replaced either by
the Dirac matrix, or by a matrix obtained through transformation by preconditioners.
Loop generation and parallelization: At this step, the initial statement has been replaced
by the algorithm statements, directly using the Dirac matrix or a preconditioned version. Assignment statements are vector assignments: Values for all vertices of the lattice can be modified
in one statement. Matrices are still described using tensor products and direct sums.
Loops are obtained by transforming all indexed sums, products into loops or sequences. For
instance, the direct sum operator indexed by the lattice L in the definition of Dirac matrix, in
Figure 1 is transformed into a parallel loop over all elements of the lattice. This loop is parallel,
due to the meaning of ⊕. As the lattice is 4D, either 4 nested loops are created, or one single,
linearized loop is created. The choice depends on a parameter in the QIRAL compiler.
Some usual compiler transformations are then used, computing dependences, fusioning
loops, applying scalar promotion and other optimizations. At the end of this phase, an OpenMP
code is produced, essentially by identifying for parallel loops the set of private variables. These
transformations are driven by rewriting strategies, using reflection in Maude.
Matching Library Calls: The resulting code still uses some high level operators on dense
matrices, such as tensor products on dense matrices, matrix-vector product. These operators
are then replaced by library calls. For this step, it is sufficient to define for each library the
expression it computes, as a rewriting rule. We defined LION, a set of hand-written library
functions used for validation purposes.
Finally, post-processing phase, mostly syntactic rewriting, transforms this output into a C
function. This function is compiled with a program skeleton and a runtime that initializes
data, calls the generated code and stores the result. Custom functions can also be defined here
5
QIRAL: A High-Level Language for LQCD Code Generation
Barthou et al.
...
while(nr > epsilon) {
#pragma omp parallel for private(ID30,ID31,ID43,ID44,ID56)
for(iL = 0; iL < L; iL ++) {
ID31 = spnaddspn(matmulspn(tensor(U[uup(iL, dt)], gmsubgm(gmdiag(c1), G(dt))), p[sup(IDX, dt)]),
spnaddspn(matmulspn(tensor(U[uup(iL, dz)], gmsubgm(gmdiag(c1), G(dz))), p[sup(IDX, dz)]),
spnaddspn(matmulspn(tensor(U[uup(iL, dy)], gmsubgm(gmdiag(c1), G(dy))), p[sup(IDX, dy)]),
matmulspn(tensor(U[uup(iL, dx)], gmsubgm(gmdiag(c1), G(dx))),
p[sup(IDX, dx)])))) ;
ID30 = cplmulspn(kappa, ID31) ;
ID44 = spnaddspn(matmulspn(tensor(U[udn(iL, dt)], gmaddgm(gmdiag(c1), G(dt))), p[sdn(IDX, dt)]),
spnaddspn(matmulspn(tensor(U[udn(iL, dz)], gmaddgm(gmdiag(c1), G(dz))), p[sdn(IDX, dz)]),
spnaddspn(matmulspn(tensor(U[udn(iL, dy)], gmaddgm(gmdiag(c1), G(dy))), p[sdn(IDX, dy)]),
matmulspn(tensor(U[udn(iL, dx)], gmaddgm(gmdiag(c1), G(dx))), p[sdn(IDX, dx)])))) ;
ID43 = cplmulspn(kappa, ID44) ;
...
Figure 4: Sample output C code produced for CGNR algorithm
since undefined operations from the QIRAL input appear as functions calls. One application
is to call BLAS routines instead of letting QIRAL implement them. For instance one defines
a rule that says (C = A ∗ B) = dgemm(A, B, C). At this stage, we are more concerned about
the correctness of the output rather than the efficiency of the code, which will be the purpose
of further steps. Automatically generated codes for different algorithms and preconditionings
show different convergence speed, as shown in Figure 5.
Figure 5: Convergence speed for a lattice of size 44 for different methods: Conjugate Gradient Normal Error (CGNE), Normal Resolution (CGNR), Biconjugate Gradient (BiCGSTAB),
Modified conjugate gradient with preconditioning (MCR1).
5
Conclusion
This paper has presented an overview of QIRAL, a high level language for automatic parallel
code generation of Lattice QCD codes. The language is based on algorithmic specifications and
on mathematical definitions of mathematical objects used in the computation.
6
QIRAL: A High-Level Language for LQCD Code Generation
Barthou et al.
The contribution of this short paper is to show that a high-level representation makes
possible the automatic generation of complex LQCD parallel code. Parallelism directly stems
from the structure of sparse, regular matrices used in LQCD. While the initial matrix represents
a stencil computation, QIRAL is able to manipulate more complex structures obtained through
preconditioning for instance, unlike Pochoir [10]. The approach, similar to the one proposed
by Ashby et al. [1] enables the user to define new equations and domain-specific definitions.
The QIRAL compiler is able to keep such information through transformations resulting from
preconditioners or algorithms. The tensor product and direct sum operators are translated into
parallel loops and lead to OpenMP code generation. This way, algorithmic exploration, the
key for higher levels of performance, can be freed from the constraints and costs of parallel
tuning. Besides generation of distributed codes with communications is within reach. For
heterogeneous architectures, such as Cell or GPUs, further work for automatic data-layout
optimization is required, in order to reach levels of performance of previous works (such as
[12, 6, 9] for the CELL BE and [2] for the GPU).
References
[1] T. Ashby, A. Kennedy, and M. O’Boyle. Cross Component Optimisation in a High Level CategoryBased Language. In Marco Danelutto, Marco Vanneschi, and Domenico Laforenza, editors, EuroPar Parallel Processing, volume 3149 of LNCS, pages 654–661. Springer, 2004.
[2] M.A. Clark, R. Babich, K. Barros, R.C. Brower, and C. Rebbi. Solving lattice QCD systems of
equations using mixed precision solvers on GPUs. Computer Physics Communications, 181(9):1517
– 1528, 2010.
[3] Manuel Clavel, Fransisco Durán, Steven Eker, Patrick Lincoln, Narciso Martı́-Oliet, José Meseguer,
and Jose F. Quesada. The Maude System. In Intl. Conf. on Rewriting Techniques and Applications,
pages 240–243, London, UK, 1999. Springer-Verlag.
[4] Francisco Durán and José Meseguer. A church-rosser checker tool for conditional order-sorted
equational maude specifications. In Rewriting Logic and Its Applications, volume 6381 of LNCS,
pages 69–85. Springer, 2010.
[5] Robert G. Edwards and Balint Joo.
The Chroma Software System for Lattice QCD.
NUCL.PHYS.PROC., 140:832, 2005.
[6] Khaled Z. Ibrahim and Francois Bodin. Implementing Wilson-Dirac operator on the cell broadband
engine. In Intl. Conf. on Supercomputing, pages 4–14, New York, NY, USA, 2008. ACM.
[7] Martin Lüscher. Local coherence and deflation of the low quark modes in lattice qcd. J. of High
Energy Physics, 2007(07):081, 2007.
[8] Markus Püschel, Franz Franchetti, and Yevgen Voronenko. Encyclopedia of Parallel Computing,
chapter Spiral. Springer, 2011.
[9] Claude Tadonki, Gilbert Grodidier, and Olivier Pene. An efficient CELL library for lattice quantum
chromodynamics. SIGARCH Comput. Archit. News, 38:60–65, January 2011.
[10] Yuan Tang, Rezaul Alam Chowdhury, Bradley C Kuszmaul, Chi-Keung Luk, and Charles E Leiserson. The pochoir stencil compiler. Symp. on Parallelism in algorithms and architectures, page
117, 2011.
[11] Eelco Visser and Zine el Abidine Benaissa. A core language for rewriting. Electronic Notes in
Theoretical Computer Science, 15(0):422 – 441, 1998. Rewriting Logic and its Applications.
[12] Pavlos Vranas, Matthias A. Blumrich, Dong Chen, Alan Gara, Mark Giampapa, Philip Heidelberger, Valentina Salapura, James C. Sexton, Ron Soltz, and Gyan Bhanot. Massively parallel
quantum chromodynamics. IBM J. of Research and Development, pages 189–198, 2008.
[13] Frank Wilczek. What QCD Tells Us About Nature – and Why We Should Listen. NUCL.PHYS.A,
663:3, 2000.
7
| 6cs.PL
|
Forward Stochastic Reachability Analysis for Uncontrolled
Linear Systems using Fourier Transforms
Abraham P. Vinod
Electrical & Comp. Eng.
University of New Mexico
Albuquerque, NM 87131, USA
arXiv:1610.04550v3 [cs.SY] 13 Feb 2017
[email protected]
Baisravan
HomChaudhuri
Electrical & Comp. Eng.
University of New Mexico
Albuquerque, NM 87131, USA
[email protected]
ABSTRACT
We propose a scalable method for forward stochastic reachability analysis for uncontrolled linear systems with affine disturbance. Our method uses Fourier transforms to efficiently
compute the forward stochastic reach probability measure
(density) and the forward stochastic reach set. This method
is applicable to systems with bounded or unbounded disturbance sets. We also examine the convexity properties
of the forward stochastic reach set and its probability density. Motivated by the problem of a robot attempting to
capture a stochastically moving, non-adversarial target, we
demonstrate our method on two simple examples. Where
traditional approaches provide approximations, our method
provides exact analytical expressions for the densities and
probability of capture.
CCS Concepts
•Theory of computation → Stochastic control and
optimization; Convex optimization; •Computing methodologies → Control methods; Computational control theory;
Keywords
Stochastic reachability; Fourier transform; Convex optimization
1.
INTRODUCTION
Reachability analysis of discrete-time dynamical systems
with stochastic disturbance input is an established tool to
provide probabilistic assurances of safety or performance
and has been applied in several domains, including motion
planning in robotics [1, 2], spacecraft docking [3], fishery
management and mathematical finance [4], and autonomous
survelliance [5]. The computation of stochastic reachable
and viable sets has been formulated within a dynamic programming framework [4, 6] that generalizes to discrete-time
stochastic hybrid systems, and suffers from the well-known
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than
ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from [email protected].
HSCC ’17, April 18–20, 2017, Pittsburgh, PA, USA
c 2017 ACM. ISBN 978-1-4503-4590-3/17/04. . . $15.00
DOI: http://dx.doi.org/10.1145/3049797.3049818
Meeko M. K. Oishi
Electrical & Comp. Eng.
University of New Mexico
Albuquerque, NM 87131, USA
[email protected]
curse of dimensionality [7]. Recent work in computing stochastic reachable and viable sets aims to circumvent these computational challenges, through approximate dynamic programming [8–10], Gaussian mixtures [9], particle filters [3,
10], and convex chance-constrained optimization [3,5]. These
methods have been applied to systems that are at most 6dimensional [8] – far beyond the scope of what is possible
with dynamic programming, but are not scalable to larger
and more realistic scenarios.
We focus in particular on the forward stochastic reachable set, defined as the smallest closed set that covers all
the reachable states. For LTI systems with bounded disturbances, established verification methods [11–13] can be
adapted to overapproximate the forward stochastic reachable set. However, these methods return a trivial result
with unbounded disturbances and do not address the forward stochastic reach probability measure, which provides
the likelihood of reaching a given set of states.
We present a scalable method to perform forward stochastic reachability analysis of LTI systems with stochastic dynamics, that is, a method to compute the forward stochastic
reachable set as well as its probability measure. We show
that Fourier transforms can be used to provide exact reachability analysis, for systems with bounded or unbounded
disturbances. We provide both iterative and analytical expressions for the probability density, and show that explicit
expressions can be derived in some cases.
We are motivated by a particular application: pursuit of
a dynamic, non-adversarial target [14]. Such a scenario may
arise in e.g., the rescue of a lost first responder in a building on fire [15], capture of a non-aggressive UAV in an urban environment [16], or other non-antagonistic situations.
Solutions for an adversarial target, based in a two-person,
zero-sum differential game, can accommodate bounded disturbances with unknown stochasticity [17–21], but will be
conservative for a non-adversarial target. We seek scalable
solutions that synthesize an optimal controller for the nonadversarial scenario, by exploiting the forward reachable set
and probability measure for the target. We analyze the convexity properties of the forward stochastic reach probability
density and sets, and propose a convex optimization problem to provide the exact probabilistic guarantee of success
and the corresponding optimal controller.
The main contributions of this paper are: 1) a method to
efficiently compute the forward stochastic reach sets and the
corresponding probability measure for linear systems with
uncertainty using Fourier transforms, 2) the convexity prop-
erties of the forward stochastic reach probability measure
and sets, and 3) a convex formulation to maximize the probability of capture of a non-adversarial target with stochastic
dynamics using the forward stochastic reachability analysis.
The paper is organized as follows: We define the forward stochastic reachability problem and review some properties from probability theory and Fourier analysis in Section 2. Section 3 formulates the forward stochastic reachability analysis for linear systems using Fourier transforms and
provides convexity results for the probability measure and
the stochastic reachable set. We apply the proposed method
to solve the controller synthesis problem in Section 4, and
provide conclusions and directions for future work in Section 5.
2.
PRELIMINARIES AND PROBLEM FORMULATION
In this section, we review some properties from probability theory and Fourier analysis relevant for our discussion
and setup the problems. For detailed discussions on probability theory, see [22–25], and on Fourier analysis, see [26].
We denote random vectors with bold case and non-random
vectors with an overline.
2.1
Preliminaries
A random vector w ∈ Rp is defined in a probability space
(W, σ(W), Pw ). Given a sample space W, the sigma-algebra
σ(W) provides a collection of measurable sets defined over
W. The sample space can be either countable (discrete random vector w) or uncountable (continuous random vector
w). In this paper, we focus only on absolutely continuous random variables. For an absolutely continuous random
vector, the probability measure defines a probability density function ψw : Rp → R such that
given a (Borel) set
R
B ∈ σ(W), we have Pw {w ∈ B} = B ψw (z̄)dz̄. Here, dz̄ is
short for dz1 dz2 . . . dzp .
We will use the concept of support to define the forward
stochastic reach set. The support of a random vector is the
smallest closed set that will occur almost surely. Formally,
the support of a random vector w is a unique minimal closed
set supp(w) ∈ σ(W) such that 1) Pw {w ∈ supp(w)} = 1,
and 2) if D ∈ σ(W) such that Pw {w ∈ D} = 1, then
supp(w) ⊆ D [22, Section 10, Ex. 12.9]. Alternatively,
denoting the Euclidean ball of radius δ centered at z̄ as
Ball(z̄, δ), we have (1) which is equivalent to (2) via [27,
Proposition 19.3.2],
(
)
Z
supp(w) = z̄ ∈ W|∀δ > 0,
ψw (z̄)dz̄ > 0
(1)
Ball(z̄,δ)
= W \ {z̄ ∈ W|∃δ > 0, ψw (z̄) = 0 a.e. in Ball(z̄, δ)}
(2)
For a continuous ψw , (2) is the support of the density [28,
Section 8.8]. Denoting the closure of a set using cl(·),
supp(w) = support(ψw ) = cl({z̄ ∈ W|ψw (z̄) > 0}).
(3)
The characteristic function (CF) of a random vector w ∈
Rp with probability density function ψw (z̄) is
h
i
Ψw (ᾱ) , Ew exp j ᾱ> w
Z
>
=
ej ᾱ z̄ ψw (z̄)dz̄ = F {ψw (·)} (−ᾱ)
(4)
Rp
where F {·} denotes the Fourier transformation operator
and ᾱ ∈ Rp . Given a CF Ψw (ᾱ), the density function can
be computed as
ψw (z̄) = F −1 {Ψw (·)} (−z̄)
p Z
>
1
e−j ᾱ z̄ Ψw (ᾱ)dᾱ
=
2π
p
R
(5)
where F −1 {·} denotes the inverse Fourier transformation
operator and dᾱ is short for dα1 dα2 . . . dαp .
We define the Ld (Rp ) spaces, 1 ≤ d < ∞, of measurable real-valued functions with finite Ld norm. The Ld
1/d
R
norm of a density ψw is kψw kd , Rd |ψw (z̄)|d dz̄
where
| · | denotes the absolute value. Here, L1 (Rp ) is the space
of absolutely integrable functions, and L2 (Rp ) is the space
of square-integrable functions. The Fourier transformation
is defined for all functions in L1 (Rp ) and all functions in
L2 (Rp ). Since probability densities are, by definition, in
L1 (Rp ), CFs exist for every probability density [26, Section
1]. Let w1 , w2 ∈ Rp be random vectors with densities ψw1
and ψw2 and CFs Ψw1 and Ψw2 respectively. By definition,
ψw1 , ψw2 ∈ L1 (Rp ). Let z̄, z̄1 , z̄2 , ᾱ, ᾱ1 , ᾱ2 ∈ Rp , β̄ ∈ Rn .
P1) If x = w1 + w2 , then ψx (z̄) = ψw1 (·) ∗ ψw2 (·) (z̄)
and Ψx (ᾱ) = Ψw1 (ᾱ)Ψw2 (ᾱ) [24, Section 21.11]. Also,
supp(x) ⊆ cl(supp(w1 )⊕supp(w2 )) [28, Lemma 8.15].
Here, ∗ denotes convolution and ⊕ Minkowski sum.
P2) If x = F w1 + G where F ∈ Rp×n , G ∈ Rn are matrices, Ψx (β̄) = exp j β̄ > G Ψw1 (F > β̄) (from [24, Section 22.6] and [26, Equation 1.5]).
>
P3) If w1 and w2 are independent vectors, then x = [w>
1 w2 ]
has probability density ψx (ȳ) = ψw1 (z̄1 )ψw2 (z̄2 ), ȳ =
>
[z̄1> z̄2> ] ∈ R2p and CF Ψx (γ̄) = Ψw1 (ᾱ1 )Ψw2 (ᾱ2 ), γ̄ =
>
[ᾱ1> ᾱ2> ] ∈ R2p [24, Section 22.4].
>
P4) The marginal probability density of any group of k
components selected from the random vector w1 is obtained by setting the remaining p − k Fourier variables
in the CF to zero [24, Section 22.4].
An additional assumption of square-integrability of the
probability density of the random variable w3 ∈ Rp results
in ψw3 ∈ L1 (Rp ) ∩ L2 (Rp ). Along with Properties P1-P4,
ψw3 satisfies the following property:
P5) The Fourier transform preserves the inner product in
L2 (Rp ) [26, Theorem 2.3]. Given a square-integrable
function h(z̄) with Fourier transform H(ᾱ) = F {h(·)} (ᾱ)
and a square-integrable probability density ψw3 ,
p Z
Z
†
1
F {ψw3 (·)} (ᾱ)
ψw3 (z̄)† h(z̄)dz̄ =
2π
p
p
R
R
× H(ᾱ)dᾱ
Here, † denotes complex conjugation.
Lemma 1. For square-integrable ψw3 and h,
p Z
Z
1
ψw3 (z̄)h(z̄)dz̄ =
Ψw3 (ᾱ)H(ᾱ)dᾱ.
2π
Rp
Rp
(6)
Proof: Follows from Property P5, (5), and [24, Section 10.6]. Since probability densities are real functions,
†
†
ψw3 (z̄) = ψw3 (z̄) and F {ψw3 (·)} (ᾱ) = Ψw3 (ᾱ).
2.2
Problem formulation
Consider the discrete-time linear time-invariant system,
x[t + 1] = Ax[t] + Bw[t]
(7)
n
with state x[t] ∈ X ⊆ R , disturbance w[t] ∈ W ⊆ Rp ,
and matrices A, B of appropriate dimensions. Let x̄0 ∈ X
be the given initial state and T be the finite time horizon.
The disturbance set W is an uncountable set which can be
either bounded or unbounded, and the random vector w[t] is
defined in a probability space (W, σ(W), Pw ). The random
vector w[t] is assumed to be absolutely continuous with a
known density function ψw . The disturbance process w[·] is
assumed to be a random process with an independent and
identical distribution (IID).
The dynamics in (7) are quite general and includes affine
noise perturbed LTI discrete-time systems with known statefeedback based inputs. An additional affine term in (7)
can include affine noise perturbed LTI discrete-time systems
with known open-loop controllers. For time τ ∈ [1, T ],
x[τ ] = Aτ x̄0 + Cn×(τ p) W
with Cn×(τ p) = [B AB A B . . . A
τ −1
2
>
>
>
(8)
B] ∈ R
n×(τ p)
and
>
W = [w [τ − 1] w [τ − 2] . . . w [0]] as a random vect=τ −1
tor defined by the sequence of random vectors {w[t]}t=0
.
For any given τ , the random vector W is defined in the
τ
product space (W τ , σ(W τ ), PW ) where W τ = t=0 W and
Qτ
Qτ
PW = t=0 Pw , ψW = t=0 ψw by the IID assumption of
the random process w[·]. From (8), the state x[·] is a random
process with the random vector at each instant x[t] defined
0
in the probability space (X , σ(X ), Pt,x̄
x ) where the probat,x̄0
bility measure Px
is induced from PW . We denote the
random process originating from x̄0 as ξ[·; x̄0 ] where for all t,
>
ξ[t; x̄0 ] = x[t], and let Z̄ = [z̄ > [τ − 1] z̄ > [τ − 2] . . . z̄ > [0]] ∈
τp
R denote a realization of the random vector W .
An iterative method for the forward stochastic reachability analysis (FSR analysis) is given in [1] [29, Section
10.5]. However, for systems perturbed by continuous random variables, the numerical implementation of the iterative
approach becomes erroneous for larger time instants due to
the iterative numerical evaluation of improper integrals, motivating the need for an alternative implementable approach.
×
Problem 1. Given the dynamics (7) with initial state x̄0 ,
construct analytical expressions at time instant τ for
1. the smallest closed set that covers all the reachable
states (i.e., the forward stochastic reach set), and
2. the probability measure over the forward stochastic reach
set (i.e., the forward stochastic reach probability measure)
that do not require an iterative approach.
We are additionally interested in applying the forward
stochastic reachable set (FSR set) and probability measure
(FSRPM) to the problem of capturing a non-adversarial target. Specifically, we seek a convex formulation to the problem of capturing a non-adversarial target. This requires convexity of the FSR set and concavity of the objective function
defined on the probability of successful capture.
Problem 2. For a finite time horizon, find a) a convex formulation for the maximization of the probability of capture
of a non-adversarial target with known stochastic dynamics
and initial state, and b) the resulting optimal controller that
a deterministic robot must employ when there is a non-zero
probability of capture.
Problem 2.a. Characterize the sufficient conditions for logconcavity of the FSRPM and convexity of the FSR set.
3.
FORWARD STOCHASTIC REACHABILITY ANALYSIS
The existence of forward stochastic reach probability density (FSRPD) for systems of the form (7) has been demonstrated in [29, Section 10.5]. For any τ ∈ [1, T ], the probability of the state reaching a set S ∈ σ(X ) at time τ starting
0
at x̄0 is defined using the FSRPM Pτ,x̄
,
x
Z
0
Pτ,x̄
{x[τ ] ∈ S} =
ψx (ȳ; τ, x̄0 )dȳ, ȳ ∈ Rn .
(9)
x
S
Since the disturbance set W is uncountable, we focus on the
computation of the FSRPD ψx , and use (9) to link it to the
FSRPM. We have discussed the countable case in [1].
We define the forward stochastic reach set (FSR set) as
the support of the random vector ξ[τ ; x̄0 ] = x[τ ] at τ ∈
[1, T ] when the initial condition is x̄0 ∈ X . From (3), for a
continuous FSRPD,
FSReach(τ, x̄0 ) = cl({ȳ ∈ X |ψx (ȳ; τ, x̄0 ) > 0}) ⊆ X . (10)
0
Lemma 2. S ∩ FSReach(τ, x̄0 ) = ∅ ⇒ Pτ,x̄
{x[τ ] ∈ S} = 0.
x
Proof: Follows from (2).
Note that when the disturbance set W is unbounded, the
definition of the FSR set (10) might trivially become Rn .
Also, for uncountable W, the probability of the state taking a particular value is zero, and therefore, the superlevel
sets of the FSRPD do not have the same interpretation as
in the countable case [1]. However, given the FSRPD, we
can obtain the likelihood that the state of (7) will reach a
particular set of interest via (9) and the FSR set via (10).
3.1
Iterative method for reachability analysis
We extend the iterative approach for the FSR analysis
proposed in [1] for a nonlinear discrete-time systems with
discrete random variables to a linear discrete-time system
with continuous random variables. This discussion, inspired
in part by [29, Section 10.5], helps to develop proofs presented later.
Assume that the system matrix A of (7) is invertible. This
assumption holds for continuous-time systems which have
been discretized via Euler method. For τ ∈ [0, T − 1], we
have from (7) and Property P1,
ψx (ȳ; τ + 1, x̄0 ) = ψAx (·; τ, x̄0 ) ∗ ψBw (·) (ȳ)
(11)
with
ψAx (ȳ; τ, x̄0 ) = |A|−1 ψx (A−1 ȳ; τ, x̄0 )
−1
(12)
ψAx (ȳ; τ, x̄0 ) = (det A) ψx (A−1 ȳ; τ, x̄0 ) from [23, Example
8.9] for τ ≥ 1, ψAx (ȳ; 0, x̄0 ) = δ(ȳ − Ax̄0 ) where δ(·) is the
Dirac-delta function [30, Chapter 5], and ψBw as the probability density of the random vector Bw. We use Property
P2 and (5) to obtain ψBw [26, Corollary 1]. Equation (11) is
a special case of the result in [29, Section 10.5]. We extend
the FSR set computation presented in [1] in the following
lemma.
Lemma 3. For τ ∈ [1, T ], closed disturbance set W, and
the system in (7) with initial condition x̄0 , FSReach(τ, x̄0 ) ⊆
A(FSReach(τ − 1, x̄0 )) ⊕ BW = {Aτ x̄0 } ⊕ Cn×(τ p) W τ .
Proof: Follows from (7), (8), and Property P1.
Lemma 3 allows the use of existing reachability analysis
schemes designed for bounded non-stochastic disturbance
models [11–13] for overapproximating FSR sets. Also, (10)
and (11) provide an iterative method for exact FSR analysis.
Note that (11) is an improper integral which must be
solved iteratively. For densities whose convolution integrals
are difficult to obtain analytically, we would need to rely
on numerical integration (quadrature) techniques. Numerical evaluation of multi-dimensional improper integrals is
computationally expensive [31, Section 4.8]. Moreover, the
quadratures in this method will become increasingly erroneous for larger values of τ ∈ [1, T ] due to the iterative
definition. These disadvantages motivate the need to solve
Problem 1 — an approach that provides analytical expressions of the FSRPD, and thereby reduce the number of
quadratures required. The iterative method performs well
with discrete random vectors as in [1] because discretization
for computation can be exact, however, this is clearly not
true when the disturbance set is uncountable.
3.2
Efficient reachability analysis via characteristic functions
We employ Fourier transformation to provide analytical
expressions of the FSRPD at any instant τ ∈ [1, T ]. This
method involves computing a single integral for the time instant of interest τ as opposed to the iterative approach in
Subsection 3.1. We also show that for certain disturbance
distributions like the Gaussian distribution, an explicit expression for the FSRPD can be obtained.
By Property P3 and the IID assumption on the random
process w[·], the CF of the random vector W is
ΨW (ᾱ) =
t=τ
Y−1
Ψw (ᾱt )
(13)
t=0
>
where ᾱ = [ᾱ0> ᾱ1> . . . ᾱτ>−1 ] ∈ R(τ p) , ᾱt ∈ Rp for all τ ∈
[0, τ −1]. As seen in (8), the random vector W concatenates
the disturbance random process w[t] over t ∈ [0, τ − 1].
Theorem 1. For any time instant τ ∈ [1, T ] and an initial
state x̄0 ∈ X , the FSRPD ψx (·; τ, x̄0 ) of (7) is given by
>
Ψx (ᾱ; τ, x̄0 ) = exp j ᾱ> (Aτ x̄0 ) ΨW (Cn×(τ
(14)
p) ᾱ)
ψx (ȳ; τ, x̄0 ) = F −1 {Ψx (ᾱ; τ, x̄0 )} (−ȳ)
(15)
where ȳ ∈ X , ᾱ ∈ Rn×1 .
Proof: Follows from Property P2, (5), and (8).
Theorem 1 provides an analytical expression for the FSRPD. Theorem 1 holds even if we relax the identical distribution assumption on the random process w[t] to a timevarying independent disturbance process, provided Ψw[t] (·)
is known for all t ∈ [0, τ − 1]. Using Property P2, Theorem 1
can also be easily extended to include affine noise perturbed
LTI discrete-time systems with known open-loop controllers.
Note that the computation of the FSRPD via Theorem 1
does not require gridding of the state space, hence mitigating
the curse of dimensionality associated with the traditional
gridding-based approaches. When the CF Ψx (ᾱ; τ, x̄0 ) has
the structure of known Fourier transforms, Theorem 1 can
be used to provide explicit expressions for the FSRPD (see
Proposition 1). In systems where the inverse Fourier transform is not known, the evaluation of (15) can be done via
any quadrature techniques that can handle improper integrals. Alternatively, the improper integral can be approximated by the quadrature of an appropriately defined proper
integral [31, Chapter 4]. For high-dimensional systems, performance is affected by the scalability of quadrature schemes
with dimension. However, Theorem 1 still requires only a
single n-dimensional quadrature for any time instant of interest τ ∈ [1, T ]. On the other hand, the iterative method
proposed in Subsection 3.1 requires τ quadratures, each ndimensional, resulting in higher computational costs and
degradation in accuracy as τ increases.
One example of a CF with known Fourier transforms arises
in Gaussian distributions. We use Theorem 1 to derive an
explicit expression for the FSRPD of (7) when perturbed
by a Gaussian random vector. Note that the FSRPD in this
case can also be computed using the well-known properties
on linear combination of Gaussian random vectors [23, Section 9] or the theory of Kalman-Bucy filter [32].
Proposition 1. The system trajectory of (7) with initial
condition x̄0 and noise process w ∼ N (µ̄w , Σw ) ∈ Rp is
ξ[τ ; x̄0 ] ∼ N (µ̄[τ ], Σ[τ ])
(16)
where τ ∈ [1, T ] and
µ̄[τ ] = Aτ x̄0 + Cn×(τ p) (1̄τ ×1 ⊗ µ̄w ),
Σ[τ ] = Cn×(τ p) (Iτ ⊗
>
Σw )Cn×(τ
p) .
(17)
(18)
Proof: For ᾱ ∈ Rp , the CF of a multivariate Gaussian
random vector w is [23, Section 9.3]
ᾱ> Σw ᾱ
.
(19)
Ψw (ᾱ) = exp j ᾱ> µ̄w −
2
From the IID assumption of w[·], Property P3, and (19), the
CF of W is
t=τ
Y−1
ᾱ> Σw ᾱt
ΨW (ᾱ) =
exp j ᾱt> µ̄w − t
2
t=0
ᾱ> (Iτ ⊗ Σw )ᾱ
= exp j ᾱ> (1̄τ ×1 ⊗ µ̄w ) −
2
where ᾱ = (ᾱ0 , ᾱ1 , . . . , ᾱτ −1 ) ∈ R(τ p) with ᾱt ∈ Rp . Here,
1̄p×q ∈ Rp×q is a matrix with all entries as 1, and In is the
identity matrix of dimension n. By (15) and (19), W ∼
N (1̄τ ×1 ⊗ µ̄w , Iτ ⊗ Σw ) [26, Corollary 1.22]. From (14), we
see that for β̄ ∈ Rn ,
>
Ψx (β̄; τ, x̄0 ) = exp j β̄ > (Aτ x̄0 ) ΨW (Cn×(τ
p) β̄))
= exp j β̄ > (Aτ x̄0 + Cn×(τ p) (1̄τ ×1 ⊗ µ̄w )) ×
!
>
β̄ > Cn×(τ p) (Iτ ⊗ Σw )Cn×(τ
p) β̄
exp −
.
(20)
2
Equation (20) is the CF of a multivariate Gaussian random
vector [26, Corollary 1.22], and we obtain µG [τ ] and ΣG [τ ]
using (19).
Depending on the system dynamics and time instant of
interest τ , we can have rank(Cn×(τ p) ) < n. In such cases,
the support of the random vector x[τ ] will be restricted to
sets of lower dimension than n [33, Section 8.5], and certain marginal densities can be Dirac-delta functions. For
example, we see that turning off the effect of disturbance in
(7) (setting B = 0 ⇒ Cn×(τ p) = 0 in Theorem 1) yields
ψx (ȳ; τ, x̄0 ) = δ(ȳ − Aτ x̄0 ), the trajectory of the corresponding
deterministic system. We have used the relation
F δ(ȳ − ȳ0 ) (ᾱ) = exp (j ᾱ> y¯0 ) [30, Chapters 5,6].
Theorem 1 and (10) provide an analytical expression for
the FSRPD and the FSR set respectively, and thereby solve
Problem 1 for any density function describing the stochastics
of the perturbation w[t] in (7).
3.3
Convexity results for reachability analysis
For computational tractability, it is useful to study the
convexity properties of the FSRPD and the FSR sets. We
define the random vector wB = Bw with density ψwB .
Lemma 4. [25, Lemma 2.1] If ψw is a log-concave distribution, then ψwB is a log-concave distribution.
Theorem 2. If ψw is a log-concave distribution, and A in
(7) is invertible, then the FSRPD ψx (ȳ; τ, x̄0 ) of (7) is logconcave in ȳ for every τ ∈ [1, T ].
Proof: We prove this theorem via induction. First, we
need to show that the base case is true, i.e, we need to show
that ψx (ȳ; 1, x̄0 ) is log-concave in ȳ. From (7) and Lemma 4,
we have a log-concave density ψx (ȳ; 1, x̄0 ) = ψwB (ȳ − Ax̄0 )
since affine transformations preserve log-concavity [34, Section 3.2.4]. Assume for induction, ψx (ȳ; τ, x̄0 ) is log-concave
in ȳ for some τ ∈ [1, T ]. We have log-concave ψAx (ȳ; τ, x̄0 )
from (12). Since convolution preserves log-concavity [34,
Section 3.5.2], Lemma 4 and (11) complete the proof.
Corollary 1. If ψw is a log-concave distribution, FSReach(τ, x̄0 )
of the system (7) is convex for every τ ∈ [1, T ], x̄0 ∈ X .
Proof: Follows from (10) and [25, Theorem 2.5].
Theorem 2 and Corollary 1 solve Problem 2.a.
4.
REACHING A NON-ADVERSARIAL TARGET WITH STOCHASTIC DYNAMICS
In this section, we will leverage the theory developed in
this paper to solve Problem 2 efficiently.
We consider the problem of a controlled robot (R) having
to capture a stochastically moving non-adversarial target,
denoted here by a goal robot (G). The robot R has controllable linear dynamics while the robot G has uncontrollable
linear dynamics, perturbed by an absolutely continuous random vector. The robot R is said to capture robot G if the
robot G is inside a pre-determined set defined around the
current position of robot R. We seek an open-loop controller
(independent of the current state of robot G) for the robot R
which maximizes the probability of capturing robot G within
the time horizon T . The information available to solve this
problem are the position of the robots R and G at t = 0, the
deterministic dynamics of the robot R, the perturbed dynamics of the robot G, and the density of the perturbation.
We consider a 2-D environment, but our approach can be
easily extended to higher dimensions. We perform the FSR
analysis in the inertial coordinate frame.
We model the robot R as a point mass system discretized
in time,
x̄R [t + 1] = x̄R [t] + BR ūR [t]
(21)
with state (position) x̄R [t] ∈ R2 , input ūR [t] ∈ U ⊆ R2 ,
input matrix BR = Ts I2 and sampling time Ts . We define an open-loop control policy ūR [t] = πopen,x̄R [0] [t] where
πopen,x̄R [0] [t] depends on the initial condition, that is,
πopen,x̄R [0] : [0, T − 1] → U is a sequence of control actions
for a given initial condition x̄R [0]. Let M denote the set of
all feasible control policies πopen,x̄R [0] . From (8),
x̄R [τ + 1] = x̄R [0] + (1̄1×τ ⊗ BR )π̄τ , τ ∈ [0, T − 1]
(22)
>
>
>
with the input vector π̄τ = [ū>
R [τ − 1] ūR [τ − 2] . . . ūR [0]] ,
(2τ )
, and ūR [t] = πopen,x̄R [0] [t].
π̄τ ∈ Mτ ⊆ R
We consider two cases for the dynamics of the robot G:
1) point mass dynamics, and 2) double integrator dynamics,
both discretized in time and perturbed by an absolutely continuous random vector. In the former case, we presume that
the velocity is drawn from a bivariate Gaussian distribution,
xG [t + 1] = xG [t] + BG,PM v G [t]
v G [t] ∼ N (µ̄vG , ΣG ).
(23a)
(23b)
The state (position) is the random vector xG [t] in the probt,x̄ [0]
ability space (X , σ(X ), PxGG ) with X = R2 , disturbance
matrix BG,PM = BR , and x̄G [0] as the known initial state
of the robot G. The stochastic velocity v G [t] ∈ R2 has mean
vector µ̄vG , covariance matrix ΣG and the CF with ᾱ ∈ R2
is given in (19). In the latter case, acceleration in each direction is an independent exponential random variable,
xG [t + 1] = AG,DI xG [t] + BG,DI a[t]
(24a)
(a[t])x ∼ Exp(λax ), (a[t])y ∼ Exp(λay )
(24b)
" 2 #
Ts
1 Ts
2
, BG,DI = I2 ⊗
AG,DI = I2 ⊗
.
0 1
Ts
The state (position and velocity) is the random vector xG [t]
t,x̄ [0]
in the probability space (XDI , σ(XDI ), PxGG ) with XDI =
R4 and x̄G [0] as the known initial state of the robot G.
The stochastic acceleration a[t] = [(a[t])x (a[t])y ]> ∈ R2+ =
[0, ∞) × [0, ∞) has the following probability density and CF
(z̄ = [z1 z2 ]> ∈ R2+ = [0, ∞) × [0, ∞), ᾱ = [α1 α2 ]> ∈ R2 ),
ψa (z̄) = λax λay exp (−λax z1 − λay z2 )
λax λay
.
Ψa (ᾱ) =
(λax − jα1 )(λay − jα2 )
(25)
(26)
The CF Ψa (ᾱ) is defined using Property P3 and the CF of
the exponential given in [22, Section 26].
Formally, the robot R captures robot G at time τ if xG [τ ] ∈
CaptureSet(x̄R [τ ]). In other words, the capture region of
the robot R is the CaptureSet(ȳ) ⊆ R2 when robot R is at
ȳ ∈ R2 . The optimization problem to solve Problem 2 is
ProbA :
maximize
subject to
CapturePrπ̄ (τ, π̄τ ; x̄R [0], x̄G [0])
(τ, π̄τ ) ∈ [1, T ] × Mτ
where the decision variables are the time of capture τ and the
control policy π̄, and the objective function CapturePrπ̄ (·)
gives the probability of robot R capturing robot G. By (22),
an initial state x̄R [0] and the control policy π̄t determines a
unique x̄R [τ ] for every τ . Using this observation, we define
the objective function CapturePrπ̄ (·) in (27). We obtain
ψxG in (27) using our solution to Problem 1, Theorem 1.
G [0]
CapturePrπ̄ (τ, π̄τ ; x̄R [0], x̄G [0]) = CapturePrx̄R (τ, x̄R [τ ]; x̄G [0]) = Pτ,x̄
{xG [τ ] ∈ CaptureSet(x̄R [τ ])}
xG
Z
ψxG (ȳ; τ, x̄G [0])dȳ.
=
(27)
CaptureSet(x̄R [τ ])
Problem ProbA is equivalent (see [34, Section 4.1.3]) to
maximize
ProbB :
subject to
CapturePrx̄R (τ, x̄R [τ ]; x̄G [0])
τ ∈ [1, T ]
x̄R [τ ] ∈ ReachR (τ ; x̄R [0])
where the decision variables are the time of capture τ and
the position of the robot R x̄R [τ ] at time τ . From (22), we
define the reach set for the robot R at time τ as
ReachR (τ ; x̄R [0]) = ȳ ∈ X |∃π̄τ ∈ Mτ s.t. x̄R [τ ] = ȳ .
Several deterministic reachability computation tools are available for the computation of ReachR (τ ; x̄R [0]), like MPT [35]
and ET [12]. We will now formulate Problem ProbB as a
convex optimization problem based on the results developed
in Subsection 3.3.
Lemma 5. [11] If the input space U is convex, the forward
reach set ReachR (τ ; x̄R [0]) is convex.
Proposition 2. If ψw is a log-concave distributions and
CaptureSet(ȳ) is convex for all ȳ ∈ X , then
CapturePrx̄R (τ, ȳ; x̄G [0]) is log-concave in ȳ for all τ .
Proof: From Theorem 2, we know that ψx (ȳ; τ, x̄R [0])
is log-concave in ȳ for every τ . The proof follows from (27)
since the integration of a log-concave function over a convex
set is log-concave [34, Section 3.5.2].
Remark 1. The densities ψv and ψa are log-concave since
multivariate Gaussian density and exponential distribution
(gamma distribution with shape parameter p = 1) are logconcave, and log-concavity is preserved for products [25, Sections 1.4, 2.3] [34, Section 3.5.2].
For any τ ∈ [1, T ], Proposition 2 and Lemma 5 ensure
ProbC :
minimize
subject to
− log(CapturePrx̄R (τ, x̄R [τ ]; x̄G [0]))
x̄R [τ ] ∈ ReachR (τ ; x̄R [0])
is convex with the decision variable x̄R [τ ]. Problem ProbC
is an equivalent convex optimization problem of the partial
maximization with respect to x̄R [τ ] of Problem ProbB since
we have transformed the original objective function with
a monotone function to yield a convex objective and the
constraint sets are identical [34, Section 4.1.3].
We solve Problem ProbB by solving Problem ProbC for
each time instant τ ∈ [1, T ] to obtain x̄∗R [τ ] and compute
the maximum of the resulting finite set to get (τ ∗ , x̄∗R [τ ∗ ]).
Since Problem ProbB could be non-convex, this approach
ensures a global optimum is found. Note that in order to
prevent taking the logarithm of zero, we add an additional
constraint to Problem ProbC
CapturePrx̄R (τ, ·; x̄G [0]) ≥ .
(28)
The constraint (28) does not affect its convexity ( is a small
positive number) from Proposition 2 and the fact that logconcave functions are quasiconcave. Quasiconcave functions
have convex superlevel sets [34, Sections 3.4, 3.5].
Using the optimal solution of Problem ProbB, we can
compute the open-loop controller to drive the robot R from
x̄R [0] to x̄∗R [τ ∗ ] by solving Problem ProbD. Defining CR =
(1̄1×(τ ∗ −1) ⊗ BR ) from (22),
minimize
ProbD :
subject to
Jπ (π̄τ ∗ )
π̄τ ∗
CR π̄τ ∗
∈ Mτ ∗
= x̄R [τ ∗ ] − x̄R [0]
where the decision variable is π̄τ ∗ . The objective function
Jπ (π̄τ ∗ ) = 0 provides a feasible open-loop controller, and
∗
∗
Jπ (π̄τ ∗ ) = π̄τ>∗ R̄π̄τ ∗ , R̄ ∈ R(2τ )×(2τ ) provides an open-loop
controller policy that minimizes the control effort while ensuring that maximum probability of robot R capturing robot
G is achieved. Solving the optimization problems ProbB and
ProbD answers Problem 2.
Our approach to solving Problem 2 is based on our solution to Problem 1, the Fourier transform based FSR analysis, and Problem 2.a, the convexity results of the FSRPD
and the FSR sets presented in this paper. In contrast, the
iterative approach for the FSR analysis, presented in Subsection 3.1, would yield erroneous CapturePrx̄R (τ, ·; x̄G [0])
for larger values of τ due to the heavy reliance on quadrature techniques. Additionally, the traditional approach of
dynamic programming based computations [4] would be prohibitively costly for the large FSR sets encountered in this
problem due to unbounded disturbances. The numerical implementation of this work is discussed in Subsection 4.3.
4.1
Robot G with point mass dynamics
We solve Problem ProbB for the system given by (23).
Here, the disturbance set is W = R2 .
Lemma 6. For the system given in (23) and initial state
of the robot G as x̄G [0] ∈ R2 , FSReachG (τ, x̄G [0]) = R2 for
every τ ∈ [1, T ].
Proof: Follows from Proposition 1 and (10).
Proposition 1 provides the FSRPD and Lemma 6 provides
the FSR set for the system (23). The probability of successful capture of the robot G can be computed using (27) since
the FSRPD ψxG (·; τ, x̄G [0]) is available.
We implement the problem with the following parameters:
0.5 0.8
Ts = 0.2, T = 20, µ̄vG = [1.3 0.3]> , ΣG =
,
0.8 2
>
>
2
x̄G [0] = [−3 0] , x̄R [0] = [−3 − 2] and U = [1, 2] . The
capture region of the robot R is a box centered about the
position of the robot ȳ with edge length 2a (a = 0.25) and
edges parallel to the axes — CaptureSet(ȳ) = Box(ȳ, a), a
convex set. We use Jπ (π̄) = 0 in Problem ProbD.
Figure 1 shows the evolution of the mean position of the
robot G and the optimal capture position for the robot R
at time instants 4, 5, 8, 14, and 20. The contour plots of
ψxG (·; τ, x̄G [0]) are rotated ellipses since ΣE is not a diagonal
matrix. From (17), the mean position of the robot G moves
in a straight line µG [τ ], as it is the trajectory of (23a) when
the input is always µ̄vG . The optimal time of capture is τ ∗ =
5, the optimal capture position is x̄∗R [τ ∗ ] = [−1.8 0]> , and
Time = 4
CapturePr∗x̄R = 0.1571
Time = 5
CapturePr∗x̄R
= 0.219
Time = 6
CapturePr∗x̄R = 0.2124
Figure 2: Solution to Problem ProbC for robot G dynamics in (23), and validation of CapturePrx̄R (τ, x̄∗R [τ ]; x̄G [0])
via Monte-Carlo simulations.
The optimal capture
time is τ ∗ = 5 and the likelihood of capture is
CapturePrx̄R (τ ∗ , x̄∗R [τ ∗ ]; x̄G [0]) = 0.219.
the corresponding probability of robot R capturing robot G
is 0.219. Note that at this instant, the reach set of the robot
R does not cover the current mean position of the robot G,
µ̄[τ ∗ ] = [−1.7 0.3]> (Figure 1b). While the reach set covers
the mean position of robot G at the next time instant t = 6,
the uncertainty in (23) causes the probability of successful
capture to further reduce (Figure 1c). Counterintuitively,
attempting to reach the mean µG [τ ] is not always best. Figure 2 shows the optimal capture probabilities obtained when
solving Problem ProbC for the dynamics (23).
4.2
Time = 14
CapturePr∗x̄R = 0.1049
Time = 20
CapturePr∗x̄R = 0.0624
Robot G with double integrator dynamics
We now consider a more complicated capture problem,
in which the disturbance is exponential (hence tracking the
mean has little relevance because it is not the mode, the
global maxima of the density), and the robot dynamics are
more realistic. We solve Problem ProbB for the system given
by (24). Here, the disturbance set is W = R2+ . Based on the
mean of the stochastic acceleration a[t], the mean position
of robot G has a parabolic trajectory due to the double
integrator dynamics, as opposed to the linear trajectory seen
in Subsection 4.1. Also, in this case, we do not have an
explicit expression for the FSRPD like Proposition 1. Using
Theorem 1, we obtain an explicit expression for the CF of
the FSRPD. We utilize Lemma 1 to evaluate CapturePr(·).
Analogous to Lemma 6 and Proposition 1, we characterize
the FSR set in Lemma 7 and the FSRPD in Proposition 3.
We use Lemma 3 to obtain an overapproximation of the FSR
set due to the unavailability of FSRPD to use (10).
Lemma 7. For the system given in (24) with initial state
x̄G [0] ∈ R4 of the robot G, we have FSReachG (τ, x̄G [0]) ⊆
{AτG,DI x̄G [0]} ⊕ R4+ for every 2 ≤ τ ≤ T , and
FSReachG (1, x̄G [0]) ⊆ {AG,DI x̄G [0]} ⊕ BG,DI R2+ .
Figure 1: Snapshots of optimal capture positions of the
robots G and R when G has point mass dynamics (23).
The blue line shows the mean position trajectory of robot G
µG [τ ], the contour plot characterizes ψxG (·; τ, x̄G [0]), the
blue box shows the reach set of the robot R at time τ
ReachR (τ, x̄R [0]), and the red box shows the capture region
centered at x̄∗R [τ ] CaptureSet(x̄∗R [τ ]).
>
2τ
4
Proof: For the dynamics in (24), C4×(2τ
) R+ = R+
>
since the rank of C4×(2τ ) is 4 for every τ ≥ 2, and elements
>
>
of C4×(2τ
) are nonnegative. For τ = 1, C4×(2τ ) = BG,DI .
Lemma 3 completes the proof.
Time = 1
CapturePr∗x̄R = 0
(Infeasible)
Time = 2
CapturePr∗x̄R = 0.6044
Figure 4: Solution to Problem ProbC for robot G dynamics in (24), and validation of CapturePrx̄R (τ, x̄∗R [τ ]; x̄G [0])
via Monte-Carlo simulations.
The optimal capture
time is τ ∗ = 2 and the capture probability is
CapturePrx̄R (τ ∗ , x̄∗R [τ ∗ ]; x̄G [0]) = 0.6044.
Proposition 3. The CF of the FSRPD of the robot G for
dynamics (24) is
ΨxG (β̄; τ, x̄G [0]) = exp(j β̄ > (AτG,DI x̄G [0]))×
Time = 3
CapturePr∗x̄R = 0.3885
τY
−1
t=0
λax λay
(λax − j ᾱ2t )(λay − j ᾱ2t+1 )
(29)
>
(2τ )
where ᾱ = C4×(2τ
and β̄ ∈ R4 . The FSRPD of the
) β̄ ∈ R
robot G is ψxG (x̄; τ, x̄G [0]) = F −1 ΨxG (·; τ, x̄G [0]) (−x̄).
Time = 6
CapturePr∗x̄R = 0.0495
Proof: Apply Theorem 1 to the dynamics (24).
To solve Problem ProbB, we define CapturePrx̄R (·) as in
(27). Since we are interested in just the position of robot
G, we require only the marginal density of the FSRPD over
the position subspace of robot G, ψxpos
. By Property P4, we
G
have for γ̄ = [γ1 γ2 ] ∈ R2 ,
>
Ψpos
xG (γ̄; τ, x̄G [0]) = ΨxG ([γ1 0 γ2 0] ; τ, x̄G [0]).
(30)
Unlike the case with Gaussian disturbance, explicit expressions for the FSRPD ψxG or its marginal density ψxpos
are
G
unavailable since the Fourier transform (29) is not standard.
Lemma 8. ψBG,DI a , ψxG ∈ L1 (R4 ) ∩ L2 (R4 ).
Time = 9
CapturePr∗x̄R = 0.0091
Figure 3: Snapshots of optimal capture positions of the
robots G and R when G has double integrator dynamics (24).
The blue line shows the mean position trajectory of robot
G µG [τ ], the contour plot characterizes ψxpos
(·; τ, x̄G [0])
G
via Monte-Carlo simulation, the blue box shows the reach
set of the robot R at time τ , ReachR (τ, x̄R [0]), and
the red box shows the capture region centered at x̄∗R [τ ],
CaptureSet(x̄∗R [τ ]).
Proof: (For ψBG,DI a ) By Hölder’s inequality [22, Section 19], ψa ∈ L1 (R2 ) ∩ L2 (R2 ). We also have
ψBG,DI a (z1 , z2 , z3 , z4 ) = δ(z3 − Ts2z4 )δ(z1 − Ts2z2 )ψz24 (z2 , z4 )
where z 24 = [z2 z4 ]> = Ts a ∈ R2 and ψz24 (z2 , z4 ) =
Ts−2 ψa ( Tz2s , Tz4s ) from (12). For i = {1, 2}, kψBG,DI a ki =
kψz24 ki = Ts2−2i kψa ki < ∞ completing the proof.
(For ψxG ) Via induction using (11) (similar to the proof
of Theorem 2). Note that functions in L1 (R4 ) ∩ L2 (R4 ) are
closed under convolution [26, Theorem 1.3].
Lemma 9. ψxpos
(x̄; τ, x̄G [0]) ∈ L1 (R2 ) ∩ L2 (R2 ).
G
Proof: For i = {1, 2}, we have from (30), kψxpos
k =
G i
kψxG ki , and from Lemma 8, kψxG ki < ∞.
Similar to Subsection 4.1, we define a convex capture
region CaptureSet(ȳR ) = Box(ȳR , a) ⊆ R2 where ȳR ∈
R2 is the state of the robot R. We define h(ȳ; ȳR , a) =
1Box(ȳR ,a) (ȳ) as the indicator function corresponding to a 2D box centered at ȳR with edge length 2a > 0 with h(ȳ) = 1
if ȳ ∈ CaptureSet(ȳR ) and zero otherwise. The Fourier
transform of h is a product of sinc functions shifted by ȳR
(follows from Property P2 and [30, Chapter 13])
H(γ̄; ȳR , a) = F {h(·; ȳR , a)}(γ̄)
>
= 4a2 exp (−j ȳR
γ̄)
sin(aγ1 ) sin(aγ2 )
.
γ1 γ2
(31)
Clearly, h is square-integrable, and from Lemmas 1 and 9, we
define CapturePrx̄R (·) in (33). Equation (33) is evaluated
using (29), (30), and (31). We use (33) as opposed (32)
due to the unavailability of an explicit expression for ψxpos
.
G
The numerical evaluation of the inverse Fourier transform of
Ψpos
xG to compute (32) will require two quadratures, resulting
in a higher approximation error as compared to (33).
We implement the problem with the following parameters:
Ts = 0.2, T = 9, a = 0.25, λax = 0.25, λay = 0.45, x̄G [0] =
[1.5 0 − 0.5 2]> , x̄R [0] = [2.5 0]> , and U = [−1.5, 1.5] ×
[1, 4]. We use Jπ (π̄) = 0 in Problem ProbD.
Figure 3 shows the evolution of the mean position of the
robot G and the optimal capture position for the robot R
at time instants 1, 2, 3, 6, and 9. For every τ ∈ [1, T ], the
contour plots of ψxpos
(·; τ, x̄G [0]) were estimated via MonteG
Carlo simulation since evaluating ψxpos
(·; τ, x̄G [0]) via (5)
G
over a grid is computationally expensive. Note that the
mean position of the robot G does not coincide with the
(·; τ, x̄G [0]) in contrast to the problem discussed
mode of ψxpos
G
in Subsection 4.1. The optimal time of capture is at τ ∗ = 2,
the optimal capture position is x̄∗R [τ ∗ ] = [1.9 0.55]> , and
the corresponding probability of robot R capturing robot G
is 0.6044 (Figure 3b). Figure 4 shows the optimal capture
probabilities obtained when solving Problem ProbC for the
dynamics (24), and the validation of the results.
4.3
Numerical implementation and analysis
All computations in this paper were performed using MATLAB on an Intel Core i7 CPU with 3.4GHz clock rate and
16 GB RAM. The MATLAB code for this work is available
at http://hscl.unm.edu/files/code/HSCC17.zip.
We solved Problem ProbC using MATLAB’s built-in functions — fmincon for the optimization, mvncdf to compute
the objective (27) for the case in Subsection 4.1, integral to
compute the objective (33) for the case in Subsection 4.2,
and max to compute the global optimum of Problem ProbB.
In both the sections, we used MPT for the reachable set calculation and solved Problem ProbD using CVX [36]. Using
Lemma 2, the FSR sets restrict the search while solving
Problem ProbC. All geometric computations were done in
the facet representation. We computed the initial guess for
the optimization of Problem ProbC by performing Euclidean
projection of the mean to the feasible set using CVX [34, Section 8.1.1]. Since computing the objective was costly, this
operation saved significant computational time. The MonteCarlo simulation used 500, 000 particles. No offline computations were done in either of the cases.
The overall computation of Problem ProbB and ProbD
for the case in Subsection 4.1 took 5.32 seconds for T = 20.
Since Proposition 1 provides explicit expressions for the FSRPD, the evaluation of the FSRPD for any given point
ȳ ∈ X takes 1.6 millseconds on average. For the case in
Subsection 4.2, the overall computation took 488.55 seconds
(∼ 8 minutes) for T = 9. The numerical evaluation of the
improper integral (33) is the major cause of increase in runtime. The evaluation of the FSRPD for any given point
ȳ ∈ X using (5) takes about 10.5 seconds, and the runtime
and the accuracy depend heavily on the point ȳ as well as
the bounds used for the integral approximation. However,
the evaluation of CapturePrx̄R (·) using (33) is much faster
(0.81 seconds) because H(γ̄; ȳR , a) is a decaying, 2-D sinc
function (decaying much faster than the CF).
The decaying properties of the integrand in (33) and CFs
in general permits approximating the improper integrals in
(5) and (33) by as a proper integral with suitably defined
finite bounds. The tradeoff between accuracy and computational speed, common in quadrature techniques, dictates the
choice of the bound. A detailed analysis of various quadrature techniques, their computational complexity, and their
error analysis can be found in [31, Chapter 4].
5.
CONCLUSIONS AND FUTURE WORK
This paper provides a method for forward stochastic reachability analysis using Fourier transforms. The method is applicable to uncontrolled stochastic linear systems. Fourier
transforms simplify the computation and mitigate the curse
of dimensionality associated with gridding the state space.
We also analyze several convexity results associated with the
FSRPD and FSR sets. We demonstrate our method on the
problem of controller synthesis for a controlled robot pursuing a stochastically moving non-adversarial target.
Future work includes exploration of various quadrature
techniques like particle filters for high-dimensional quadratures, and extension to a model predictive control framework
and to discrete random vectors (countable disturbance sets).
Multiple pursuer applications will also be investigated.
6.
ACKNOWLEDGEMENTS
The authors thank Prof. M. Hayat for discussions on
Fourier transforms in probability theory and the reviewers
for their insightful comments.
This material is based upon work supported by the National Science Foundation, under Grant Numbers CMMI1254990, CNS-1329878, and IIS-1528047. Any opinions,
findings, and conclusions or recommendations expressed in
this material are those of the authors and do not necessarily
reflect the views of the National Science Foundation.
7.
REFERENCES
[1] Baisravan HomChaudhuri, Abraham P. Vinod, and
Meeko M. K. Oishi. Computation of forward
stochastic reach sets: Application to stochastic,
dynamic obstacle avoidance. In Proc. American
Control Conf., 2017. (accepted).
[2] Nick Malone, Kendra Lesser, Meeko Oishi, and Lydia
Tapia. Stochastic reachability based motion planning
for multiple moving obstacle avoidance. In Proc.
Hybrid Syst.: Comput. and Control, pages 51–60, 2014.
[3] Kendra Lesser, Meeko Oishi, and R. Scott Erwin.
Stochastic reachability for control of spacecraft
relative motion. In Proc. IEEE Conf. on Decision and
Control, pages 4705–4712, 2013.
[4] Sean Summers and John Lygeros. Verification of
discrete time stochastic hybrid systems: A stochastic
reach-avoid decision problem. Automatica,
46(12):1951–1961, 2010.
[5] Nikolaos Kariotoglou, Davide M Raimondo, Sean
Summers, and John Lygeros. A stochastic reachability
Z
CapturePrx̄R (τ, x̄R [τ ]; x̄G [0]) =
R2
=
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
ψxpos
(x̄; τ, x̄G [0])h(x̄; x̄R [τ ], a)dx̄
G
1
2π
framework for autonomous surveillance with
pan-tilt-zoom cameras. In European Control Conf.,
pages 1411–1416, 2011.
Alessandro Abate, Maria Prandini, John Lygeros, and
Shankar Sastry. Probabilistic reachability and safety
for controlled discrete time stochastic hybrid systems.
Automatica, 44(11):2724–2734, 2008.
Alessandro Abate, Saurabh Amin, Maria Prandini,
John Lygeros, and Shankar Sastry. Computational
approaches to reachability analysis of stochastic
hybrid systems. In Proc. Hybrid Syst.: Comput. and
Control, pages 4–17, 2007.
Nikolaos Kariotoglou, Sean Summers, Tyler Summers,
Maryam Kamgarpour, and John Lygeros.
Approximate dynamic programming for stochastic
reachability. In European Control Conf., pages
584–589, 2013.
Nikolaos Kariotoglou, Kostas Margellos, and John
Lygeros. On the computational complexity and
generalization properties of multi-stage and stage-wise
coupled scenario programs. Syst. and Control Lett.,
94:63–69, 2016.
Giorgio Manganini, Matteo Pirotta, Marcello Restelli,
Luigi Piroddi, and Maria Prandini. Policy search for
the optimal control of Markov Decision Processes: A
novel particle-based iterative scheme. IEEE Trans.
Cybern., pages 1–13, 2015.
Michal Kvasnica, Bálint Takács, Juraj Holaza, and
Deepak Ingole. Reachability analysis and control
synthesis for uncertain linear systems in MPT. IFAC
Symp. on Robust Control D., 48(14):302–307, 2015.
Alex A. Kurzhanskiy and Pravin Varaiya. Ellipsoidal
toolbox. Technical Report UCB/EECS-2006-46, EECS
Department, University of California, Berkeley, 2006.
Antoine Girard. Reachability of uncertain linear
systems using zonotopes. In Proc. Hybrid Syst.:
Comput. and Control, pages 291–305, 2005.
Geoffrey Hollinger, Sanjiv Singh, Joseph Djugash, and
Athanasios Kehagias. Efficient multi-robot search for a
moving target. Int’l J. Robotics and Research,
28(2):201–219, 2009.
Vijay Kumar, Daniela Rus, and Sanjiv Singh. Robot
and sensor networks for first responders. IEEE
Pervasive computing, 3(4):24–33, 2004.
Christopher Geyer. Active target search from UAVs in
urban environments. In Proc. IEEE Int’l Conf.
Robotics and Autom., pages 2366–2371, 2008.
Ian Mitchell and Claire J. Tomlin. Level set methods
for computation in hybrid systems. In Proc. Hybrid
Syst.: Comput. and Control, pages 310–323, 2000.
Claire J. Tomlin, John Lygeros, and Shankar Sastry. A
game theoretic approach to controller design for
hybrid systems. Proc. IEEE, 88(7):949–970, 2000.
Claire J. Tomlin, Ian Mitchell, Alexandre M. Bayen,
2 Z
R2
Ψpos
xG (γ̄; τ, x̄G [0])H(γ̄; x̄R [τ ], a)dγ̄.
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
(32)
(33)
and Meeko Oishi. Computational techniques for the
verification of hybrid systems. Proc. IEEE,
91(7):986–1001, 2003.
Olivier Bokanowski, Nicolas Forcadel, and Hasnaa
Zidani. Reachability and Minimal Times for State
Constrained Nonlinear Problems without Any
Controllability Assumption. SIAM J. of Control and
Optimization, 48(7):4292–4316, 2010.
Haomiao Huang, Jerry Ding, Wei Zhang, and Claire J.
Tomlin. Automation-assisted capture-the-flag: A
differential game approach. IEEE Trans. Control Syst.
Technol., 23:1014–1028, 2015.
Patrick Billingsley. Probability and measure. Wiley,
New York, 3 edition, 1995.
John A Gubner. Probability and random processes for
electrical and computer engineers. Cambridge
University Press, New York; Cambridge, 2006.
Harald Cramér. Mathematical methods of statistics
(PMS-9). Princeton university press, 9 edition, 1961.
Sudhakar Dharmadhikari and Kumar Joag-Dev.
Unimodality, convexity, and applications. Elsevier,
1988.
Elias M Stein and Guido L Weiss. Introduction to
Fourier analysis on Euclidean spaces, volume 1.
Princeton University Press, 1971.
Terence Tao. Analysis II. Hindustan Book Agency, 2
edition, 2009.
Jean-Paul Penot. Analysis: From Concepts to
Applications. Springer, 1 edition, 2016.
Andrzej Lasota and Michael C Mackey. Chaos,
fractals, and noise: stochastic aspects of dynamics,
volume 97. Springer Science & Business Media, 2013.
Ron Bracewell. The Fourier transform and its
applications. McGraw-Hill, Inc., 1986.
William H. Press, Saul A. Teukolsky, William T.
Vetterling, and Brian P. Flannery. Numerical recipes:
The art of scientific computing. Cambridge University
Press, New York, NY, USA, 3 edition, 2007.
Peter Dorato, Vito Cerone, and Chaouki Abdallah.
Linear-quadratic control: An introduction. Simon &
Schuster, 1994.
Y.S. Chow and H. Teicher. Probability Theory:
Independence, Interchangeability, Martingales.
Springer Texts in Statistics. Springer New York, 1997.
Stephen P. Boyd and Lieven Vandenberghe. Convex
optimization. Cambridge University Press, Cambridge,
UK ; New York, 2004.
Martin Herceg, Michal Kvasnica, Colin N. Jones, and
Manfred Morari. Multi-Parametric Toolbox 3.0. In
European Control Conf., pages 502–510, 2013.
http://control.ee.ethz.ch/˜mpt.
Michael Grant and Stephen Boyd. CVX: MATLAB
software for disciplined convex programming, version
2.1. http://cvxr.com/cvx, 2014.
| 3cs.SY
|
1
Stochastic Geometry Modeling and Analysis of
Single- and Multi-Cluster Wireless Networks
Seyed Mohammad Azimi-Abarghouyi, Behrooz Makki,
Martin Haenggi, Fellow, IEEE, Masoumeh Nasiri-Kenari, Senior Member, IEEE,
arXiv:1712.08784v1 [cs.IT] 23 Dec 2017
and Tommy Svensson, Senior Member, IEEE
Abstract
This paper develops a stochastic geometry-based approach for the modeling and analysis of singleand multi-cluster wireless networks. We first define finite homogeneous Poisson point processes to model
the number and locations of the transmitters in a confined region as a single-cluster wireless network.
We study the coverage probability for a reference receiver for two strategies; closest-selection, where
the receiver is served by the closest transmitter among all transmitters, and uniform-selection, where
the serving transmitter is selected randomly with uniform distribution. Second, using Matern cluster
processes, we extend our model and analysis to multi-cluster wireless networks. Here, the receivers
are modeled in two types, namely, closed- and open-access. Closed-access receivers are distributed
around the cluster centers of the transmitters according to a symmetric normal distribution and can
be served only by the transmitters of their corresponding clusters. Open-access receivers, on the other
hand, are placed independently of the transmitters and can be served by all transmitters. In all cases,
the link distance distribution and the Laplace transform (LT) of the interference are derived. We also
derive closed-form lower bounds on the LT of the interference for single-cluster wireless networks. The
impact of different parameters on the performance is also investigated.
Index Terms
Stochastic geometry, coverage probability, clustered wireless networks, Poisson point process, Matern
cluster process.
S.M. Azimi-Abarghouyi and M. Nasiri-Kenari are with the Dep. of Electrical Engineering, Sharif University of Technology,
Tehran 11365-9363, Iran. (e-mail: [email protected]; [email protected]). B. Makki and T. Svensson are with the
Dep. of Signals and Systems, Chalmers University of Technology, 412 96 Gothenburg, Sweden. (e-mail: {behrooz.makki,
tommy.svensson}@chalmers.se). M. Haenggi is with the Dept. of Electrical Engineering, University of Notre Dame, IN 46556,
USA. (e-mail: [email protected]). This work has been supported in part by the Research Office of Sharif University of
Technology under grant QB960605 and by the VR Research Link Project "Green Communications" and by the U.S. National
Science Foundation under grant CCF 1525904.
2
I. I NTRODUCTION
Single-cluster wireless networks are composed of a number of nodes distributed inside a finite
region. This spatial setup is an appropriate model for, e.g., various millimeter wave communications use-case scenarios, indoor and ad hoc networks, as promising candidate technologies for
the next generation of wireless networks [1]-[2]. This setup is also useful in situations where
there is a range limit for backhaul links in connecting transmitters to a core network, e.g., cloud
radio access networks [3]. On the other hand, the ever-growing randomness and irregularity in
the locations of nodes in a wireless network has led to a growing interest in the use of stochastic
geometry and Poisson point processes (PPPs) for accurate, flexible, and tractable spatial modeling
and analysis [4]-[9].
In comparison to wireless networks on infinite regions that are mostly modeled by the infinite
homogeneous PPP (HPPP) [10, Def. 2.8], the modeling and performance analysis of single-cluster
wireless networks is more challenging and requires different approaches. The main challenge
is that a finite point process is not statistically similar from different locations, and therefore,
the system performance depends on the receiver location even after averaging over the point
process.
The stochastic geometry-based modeling and analysis of single-cluster wireless networks
modeled as a binomial point process (BPP) [10, Def. 2.11] has been well studied [11]-[21]. In the
BPP model, a fixed and finite number of nodes are distributed independently and uniformly inside
a finite region. Most prior works focus on a setup where the reference receiver is placed at the
center of a circular network region [11]-[15]. Also, considering a circular region, [16] has recently
developed a comprehensive framework for performance characterizations of an arbitrarily-located
reference receiver inside the region under different transmitter selection strategies. Considering
transmitters at a non-zero fixed altitude, disk-shaped networks of unmanned aerial vehicles are
analyzed in [17]. There are also a few studies that present outage probability characterizations
of a fixed link inside an arbitrarily-shaped finite region [18]-[21].
In spite of the usefulness of HPPP for modeling and analysis of coverage-centric and uniform
deployments of nodes [4]-[10], it cannot accurately model user-centric and content-centric deployments, where the nodes may be deployed at the places with high user density [22]-[24]. In
such deployments, it is important to take into account non-uniformity as well as the correlation
that may exist between the locations of the transmitters and receivers. Accordingly, third gener-
3
ation partnership project (3GPP) has considered clustered models in [24]-[25]. Models based on
Poisson cluster processes (PCPs) [10, Sec. 3.4] have recently been studied for heterogeneous [22]
and device-to-device (D2D) networks [23]. In these works, the network follows a Thomas cluster
process (TCP) [10, Def. 3.5]. A PCP model is also proposed and analyzed for heterogeneous
networks in [26]. In [27], clustered ad hoc networks are modeled using the Matern cluster
process (MCP) [10, Def. 3.6] and the TCP, and the performance of a fixed link is analyzed.
Nearest-neighbor and contact distance distributions for the MCP are derived in [28].
In this paper, we develop tractable models for single- and multi-cluster wireless networks. We
define a finite homogeneous Poisson point process (FHPPP) to model nodes in a finite region
and then develop a framework for the analysis of single-cluster wireless networks under two
different strategies. The first approach is referred to as closest-selection where a reference receiver
is served by the closest transmitter in the network. In the second approach, which we refer to
as uniform-selection, a uniformly randomly selected transmitter is connected to the receiver.
These strategies cover a broad range of requirements of wireless networks. For instance, the
closest-selection approach is suitable for cellular networks, while the uniform-selection scheme
is relevant to ad hoc networks.
To model multi-cluster wireless networks consisting of different single-cluster wireless networks, we consider an MCP of transmitters. For the receivers, we consider two types, i) closedaccess receivers, which are located around the cluster centers of transmitters with a symmetric
normal distribution and are allowed to be served only by the transmitters of their corresponding
clusters according to the closest- or uniform-selection strategy, and ii) open-access receivers,
which can be served by all transmitters according to the closest-selection strategy.
We derive exact expressions for the coverage probability of a reference receiver in singleand multi-cluster wireless networks for the different selection strategies and types of receivers.
For each selection strategy and type of receiver in single- and multi-cluster wireless networks,
we characterize the Laplace transform (LT) of the interference. Moreover, as a key step for the
coverage probability analysis, the distributions of the distance from the reference receiver to its
serving transmitter are derived. We also derive tight closed-form lower bounds on the LT of
the interference in the case of single-cluster wireless networks, which are convenient for the
coverage probability analysis.
We investigate the impact of different parameters of the system models on the performance in
terms of the coverage probability and spectral efficiency. In most cases, a higher path loss
4
exponent improves the performance. However, at relatively high distances of the reference
receiver to the center of single-cluster wireless networks, a higher path loss has a degrading
effect on the performance. Also, an increase in the distance of the reference receiver to the
center of the network decreases the chance for coverage. Our analysis reveals that there exists
an optimal distance for the location of the reference receiver from the center of the network
that maximizes the coverage probability. An optimal distance is also observed for the spectral
efficiency. Our evaluation also shows that, for a broad range of parameter settings, our proposed
lower bounds tightly mimic the exact results on the coverage probability.
Our work is different from the state-of-the-art literature, e.g., [10]-[23], [26]-[28], from three
perspectives. First, different from the BPP, which models a fixed number of nodes in a region,
we consider a point process that is suitable for finite regions with a random number of nodes and
allow for arbitrary receiver locations. Second, we comprehensively study multi-cluster wireless
networks using the MCP. In our analysis, we also derive the contact distribution function of the
MCP in a form that is significantly simpler than the one in [28, Thm. 1]. Third, we propose
open-access and closed-access receivers and different transmitter selection strategies.
The rest of the paper is organized as follows. Section II describes the system models. Secion
III proposes the transmitter selection strategies and presents the analytical results for the coverage probability of single-cluster wireless networks, including characterizations for the serving
distance distributions, and the LT of the interference and its corresponding lower bounds. Section
IV presents the analytical results for the coverage probability of multi-cluster wireless networks,
and derives the related serving distance distributions and the LT of the interferences. Section V
presents the numerical results. Finally, Section VI concludes the paper.
II. S YSTEM M ODEL
In this section, we provide a mathematical model of the system, including the spatial distribution of the nodes for single- and multi-cluster wireless networks and the channel model.
A. Spatial Model for Single-Cluster Wireless Networks
Let us define an FHPPP as follows.
Definition 1: We define the FHPPP as Φ = P ∩ A, where P is an HPPP of intensity λ and
A ⊂ R2 .
We consider a single-cluster wireless network as shown in Fig. 1, where the locations of
active transmitters are modeled as an FHPPP. The transmitters are assumed to transmit at the
5
Fig. 1: An illustration of the system model for single-cluster wireless networks.
same power. For simplicity and in harmony with, e.g., [11]-[17], we let A = b(xo , D), where
b(xo , D) represents a disk centered at xo with radius D. However, our theoretical results can
be extended to arbitrary regions A.
Receivers can be located everywhere in R2 . With no loss of generality, we conduct the analysis
at a reference receiver located at the origin o. We further define d = kxo k.
The proposed setup well models a wireless network confined in a finite region, such as indoor
and ad hoc networks.
B. Spatial Model for Multi-Cluster Wireless Networks
An MCP is defined as follows [10, Def. 3.6].
Definition 2: An MCP Φ is a union of offspring points that are located around parent points.
The parent point process is an HPPP Φp with intensity λp , and the offspring point processes (one
per parent) are conditionally independent. Conditioned on x ∈ Φp , the offsprings form an FHPPP
Φx with intensity λ over the disk b(x, D).
We consider a multi-cluster wireless network as shown in Fig. 2, where the locations of active
transmitters are modeled as an MCP.
We consider two types of receivers. The first type, referred to as closed-access receivers,
can be served by only a single cluster of transmitters. This closed-access receiver is distributed
according to a symmetric normal distribution with variance σc2 around the parent point of its
corresponding cluster. Therefore, assuming a closed-access receiver at y and its parent point at
x, kx − yk is Rayleigh distributed with probability density function (PDF)
v
v2
fkx−yk (v) = 2 exp − 2 .
σc
2σc
(1)
The second type, referred to as open-access receivers, considers receivers that are placed independently of the transmitters and can be served by all transmitters.
6
Fig. 2: An illustration of the system model for a finite piece of multi-cluster wireless networks.
The proposed setup well models various use-case scenarios as follows:
1) Clustered Small-Cell Base Stations (BSs): The trend in cellular networks is to deploy
small-cell BSs at the places with high user density, referred to as user-centric networks [22],
[29], as also proposed in 3GPP [24]-[25]. In this way, according to our setup, users who are
likely to be close to a cluster of small-cell BSs can be modeled as closed-access receivers, such
as users at a stadium or a mall. The open-access receivers can model users who are distributed
homogeneously and independently of the small-cell BS locations, such as pedestrians or cars.
2) Cloud BSs: A cloud BS is a distributed multiple-antenna system formed by a number of
simple antenna terminals [30]. In this application, the closed-access receivers can be modeled
as the users who have a license to use a certain BS, while the open-access receivers can model
users with flexibility to access all BSs.
3) Clustered Access Networks: A large building may have a number of WiFi access points as
an access network to meet its users’ demands. The closed-access receivers model the users who
are in a building and use its access network. On the other hand, the open-access receivers are
users who can handoff between access networks of different buildings.
4) Clustered D2D Networks: A device typically has nearby devices in a finite region as a
cluster in a content-centric network, which can have direct communications with each other.
The closest- or uniform-selection strategy can be considered for cellular or ad hoc access to
contents distributed over the devices, respectively.
C. Channel Model
We assume distance-dependent power-law path loss and small-scale Rayleigh fading. Thus,
the received power at the reference receiver from a transmitter located at y is hy kyk−α , where
7
the (common) transmit power is set to 1 with no loss of generality and α > 2 is the path loss
exponent. The sequence {hy } consists of i.i.d. exponential random variables with mean 1.
III. S INGLE -C LUSTER W IRELESS N ETWORKS
In this section, we concentrate on single-cluster wireless networks. To allocate a transmitter
to a reference receiver, we propose selection strategies in Subsection III.A. Then, distance
distributions and coverage probabilities for the selection strategies are derived in Subsections
III.B and III.C, respectively. However, the resulting expressions for the coverage probabilities
are not very easy to use. Hence, we derive a closed-form lower bound for the coverage probability
of each strategy in Subsection III.D.
A. Selection Strategies
1) Closest-selection. Here, a reference receiver is served by the transmitter that provides the
maximum received power averaged over the fading. In our model, this leads to the closestselection strategy, i.e.,
xc = arg
min
kyk,
y∈{Φ|n(Φ)>0}
(2)
where n(·) denotes the number of elements in a set. Suitable for networks with infrastructure such
as downlink cellular networks, this strategy implies that a receiver is served by the transmitter
whose Voronoi cell it resides in [4]-[23].
2) Uniform-selection. Here, the serving transmitter is selected randomly with uniform distribution among all transmitters. This leads to
xu = Unif {Φ | n(Φ) > 0} ,
(3)
where Unif {·} denotes the uniform-selection operation. Uniform-selection models random allocation of receivers to transmitters, which may be the case in networks without infrastructure,
e.g., ad hoc networks and D2D networks [16], [23]. It is also suitable for applications where
the content of interest for a receiver can be available at each transmitter among all transmitters
with equal probability, such as caching networks.
The signal-to-interference-and-noise ratio (SINR) of a reference receiver at the origin can be
expressed as
hxq kxq k−α
, q = {c, u} ,
SINRq =
σ 2 + Iq
(4)
8
where xq is the location of the serving transmitter, Iq =
P
y∈Φ\{xq }
hy kyk−α denotes the
interference, and σ 2 is the noise power. To distinguish between the strategies, q = c and
u represent the case of closest- and uniform-selection strategies, respectively. For notational
simplicity, let us also define Rq = kxq k.
B. Serving Distance Distribution
Considering the closest- and uniform-selection strategies, in this subsection, we derive the
distributions of the distance from the reference receiver to its serving transmitter. These distance
distributions will be used later in the coverage probability analysis.
2 2 2
Let us first define ϕ0 = sin−1 Dd , ϕ1 (r) = cos−1 r +d2dr−D , R1 (θ) = d cos (θ) +
p
p
D2 − d2 sin2 (θ) and R̂1 (θ) = d cos (θ) − D2 − d2 sin2 (θ), and present a lemma on the
intersection area of two circles.
Lemma 1: Consider two circles with radii D and r with centers separated by distance d. The
area of their intersection is given by [31, Eq. (12.76)]
2
q
1
D + d2 − r 2
2
2
−1
+ r ϕ1 (r) −
(r + d)2 − D2 D2 − (r − d)2 .
Bd (r) = D cos
2dD
2
(5)
Considering the closest-selection strategy, the distance from the reference receiver to its nearest
transmitter Rc is larger than r if and only if at least one transmitter exists inside A and there is
no transmitter located within b(o, r) ∩ A. Letting Cr denote the intersection, we have
P(n(Φ ∩ Cr ) = 0 and n(Φ) > 0)
P(n(Φ) > 0)
(a) P(n(Φ ∩ Cr ) = 0)P(n(Φ \ Cr ) > 0)
=
P(n(Φ) > 0)
2
exp (−λ |Cr |) − exp(−λπD2 )
(b) exp (−λ |Cr |) (1 − exp(−λ(πD − |Cr |)))
=
=
,(6)
1 − exp(−λπD2 )
1 − exp(−λπD2 )
P (Rc > r) =
where | · | denotes the area of a region. Also, (a) follows from the fact that the numbers of
points of a PPP in disjoint regions are independent, and (b) is because Rc ≤ D + d. Note when
the intersection is the whole of A, i.e., |Cr | = πD2 , (6) is zero.
According to Fig. 3, which shows illustrations of the intersection, there are two different cases
as follows.
9
Fig. 3: An illustration of the intersection in the case d ≤ D for Subplot (a): D − d ≤ Rc < D + d and Subplot (b):
0 ≤ Rc < D − d, and in the case d > D for Subplot (c): d − D ≤ Rc < d + D.
Case 1: If d ≤ D, then
2
0 ≤ r < D − d,
πr
|Cr | =
Bd (r) D − d ≤ r < D + d,
πD2
r ≥ D + d,
(7)
where Bd (r) is given in (5).
Case 2: If d > D, then
0 ≤ r < d − D,
0
|Cr | =
Bd (r) d − D ≤ r < D + d,
πD2
r ≥ D + d.
(8)
Considering the uniform-selection strategy, on the other hand, the distance from the reference
receiver to a randomly chosen transmitter, i.e., Ru , is less than r if and only if, there is
a transmitter and the transmitter is located within b(o, r) ∩ A. Thus, as each transmitter is
distributed independently and uniformly within A, we have the following cases.
Case 1: If d ≤ D, then
P (Ru ≤ r) =
r2
D2
Bd (r)
πD2
0 ≤ r < D − d,
D − d ≤ r < D + d,
1
r ≥ D + d.
0
0 ≤ r < d − D,
(9)
Case 2: If d > D, then
P (Ru ≤ r) =
Bd (r)
πD2
1
d − D ≤ r < D + d,
r ≥ D + d.
(10)
10
C. Coverage Probability
In this subsection, the distance distribution results obtained in (6)-(8) and (9)-(10) are used
to derive the coverage probability of the reference receiver for the two transmitter selection
strategies. As a key step in the coverage probability derivation of each strategy, we obtain the
LT of the interference (Theorems 1 and 2). For notational simplicity, we define F(s, x) =
x2 2 F1 1, α2 ; 1 + α2 ; − 1s xα where 2 F1 (a, b; c; t) denotes the Gauss hypergeometric function [32].
1) Closest-Selection Strategy:
Theorem 1: Conditioned on Rc , the LT of the interference under the closest-selection strategy
is
LdIc (s|Rc ) =
Ad (s)
B d (s)
if d ≤ D and 0 ≤ Rc < D − d,
if d ≤ D and D − d ≤ Rc < D + d
√
or d > D and d2 − D2 ≤ Rc < d + D,
√
if d > D and d − D ≤ Rc < d2 − D2 ,
C d (s)
where Ad (s), B d (s) and C d (s) are defined as
Z π
d
F(s, R1 (θ))dθ ,
A (s) = exp πλF(s, Rc ) − λ
(11)
(12)
0
Z
B (s) = exp ϕ1 (Rc )λF(s, Rc ) − λ
ϕ1 (Rc )
d
F(s, R1 (θ))dθ ,
(13)
0
Z
C (s) = B (s) exp −λ
d
ϕ0
d
n
o
F(s, R1 (θ)) − F(s, R̂1 (θ)) dθ ,
(14)
ϕ1 (Rc )
and ϕ0 , ϕ1 , R1 and R̂1 are defined in Subsection III.B.
Proof: See Appendix A.
Using the conditional LT of the interference derived in Theorem 1, we can express the coverage
probability of the reference receiver for the closest-selection strategy as
PCc (β) = P(n(Φ) > 0)P(SINRc > β | n(Φ) > 0),
(15)
where β is the minimum required SINR for a coverage. Note that the coverage probability is
zero when there is no transmitter, and the SINRc is only meaningfully defined when there is a
transmitter. Then, from (4) and averaging over the serving distance Rc , we have
Z ∞
hxc r−α
c
2
PC (β) = (1 − exp(−λπD ))
P
> β fRd c (r) dr.
2+I
σ
c
0
(16)
11
where fRd c is the PDF of Rc obtained from (6). The conditional coverage probability given a link
distance r can be expressed as
o
(a) n
hxc r−α
2
2
α
α
σ
+
I
P
σ
+
I
=
E
exp
−βr
>
β
=
P
h
>
βr
c
c
x
c
σ 2 + Ic
= exp −βσ 2 rα LdIc (βrα |r) ,
(17)
where (a) follows from hxc ∼ exp(1). Finally, according to the cases considered in Subsection
III.B and with LdIc given in (11), the coverage probability is obtained as
R D−d
2
2 α
d
α
0 R 2πλr exp(−λπr ) exp (−βσ r ) LIc (βr |r) dr
D+d ∂Bd (r)
PCc (β) = + D−d
λ ∂r exp(−λBd (r)) exp (−βσ 2 rα ) LdIc (βrα |r) dr if d ≤ D,
R
D+d λ ∂Bd (r) exp(−λBd (r)) exp (−βσ 2 rα ) Ld (βrα |r) dr
if d > D.
Ic
∂r
d−D
(18)
In the special case of infinite wireless networks, i.e., D → ∞, the coverage probability (18)
simplifies to the result in [8, Thm. 2].
2) Uniform-Selection Strategy:
Theorem 2: The LT of the interference under the uniform-selection strategy is
E d (s) if d ≤ D,
d
LIu (s) =
F d (s) if d > D,
where E d (s) and F d (s) are defined as
R
R D+d λ ∂Bd (x)
D−d 2πλx
dx
+
dx − 1
2
2 exp
1+sx−α
0
D−d 1+sx−α ∂x
πD exp(−λπD )
E d (s) =
,
R
R
D−d 2πx
D+d
∂Bd (x)
1
1 − exp(−λπD2 )
dx
−α dx +
−α
0
R
D+d
exp
d−D
πD exp(−λπD )
d
F (s) =
R D+d
1 − exp(−λπD2 )
2
2
1+sx
∂B (x)
d
λ ∂x
dx
1+sx−α
D−d 1+sx
(19)
(20)
∂x
−1
∂Bd (x)
1
dx
d−D 1+sx−α ∂x
.
(21)
Proof: See Appendix B.
Here the LT of the interference is independent of the serving distance Ru . Using Theorem 2 and
following the same approach as in (16)-(18), the coverage probability for the uniform-selection
strategy can be expressed as
nR
D−d 2r
−λπD2
(1
−
e
)
exp (−βσ 2 rα ) LdIu (βrα ) dr
D2
0
R
o
D+d 1 ∂Bd (r)
2 α
d
α
PCu (β) =
+ D−d πD2 ∂r exp (−βσ r ) LIu (βr ) dr
(1 − e−λπD2 ) R D+d 1 ∂Bd (r) exp (−βσ 2 rα ) Ld (βrα ) dr
Iu
d−D πD2 ∂r
if d ≤ D,
if d > D.
(22)
12
(a) d ≤ D
(b) d > D
Fig. 4: Outer bounds of A in the two cases d ≤ D and d > D.
D. Lower Bounds on Coverage Probability
Since the results derived for the LT of the interference in Theorems 1 and 2 require intensive
numerical computations, we derive tight lower bounds on the LT of the interference that lead
to more tractable and useful analytical results. Then, using the bounds in (18) and (22), lower
bounds on coverage probability can be also provided. The tightness of the bounds will be verified
with numerical results (Fig. 7).
To obtain the lower bounds, we outer bound the region A by a region that permits closed-form
bounds on the LT of the interference. Note that using a larger region leads to an upper bound
on the interference and, in turn, a lower bound on its LT.
The outer region for the cases with d ≤ D and d > D is shown in Fig. 4. Placing the center
of the sectors at the reference receiver in case d ≤ D, two covering half-circles with radii
√
d + D and D2 − d2 are considered. Also, in the case d > D, we consider the sector with radii
d + D and d − D and the front angle 2ϕ0 entangled between the two tangent lines. However,
for the closest-selection strategy in the case d > D, we can achieve a tighter bound by the
following regions for A\b(o, Rc ), which is the region including interfering transmitters. In the
√
case Rc > d2 − D2 , the sector with the front angle equal to twice the intersection angle, i.e.,
√
2ϕ1 (Rc ), and radii Rc and D + d is considered. In the case Rc < d2 − D2 , we consider two
sectors with the front angle 2ϕ0 and radii Rc and R1 (ϕ1 (Rc )) and with the front angle 2ϕ1 (Rc )
and radii R1 (ϕ1 (Rc )) and D + d.
In the following corollaries, we present the lower bounds on the LT of the interference for
both strategies.
Corollary 1: With closest-selection, the LT of the interference given the serving distance Rc
13
is lower bounded by
Âd (s) if d ≤ D
B̂ d (s) if d ≤ D
d
LIcb (s|Rc ) =
Ĉ d (s) if d > D
d
D̂ (s) if d > D
√
and 0 < Rc ≤ D2 − d2 ,
√
and D2 − d2 ≤ Rc < D + d,
√
and d − D ≤ Rc < d2 − D2 ,
√
and d2 − D2 ≤ Rc < D + d.
where Âd (s), B̂ d (s), Ĉ d (s), and D̂d (s) are defined as
√
1
1
d
2
2
 (s) = exp πλ F(s, Rc ) − F(s, d + D) − F(s, D − d ) ,
2
2
(23)
(24)
π n
o
λ F(s, Rc ) − F(s, d + D) ,
(25)
2
n
o
Ĉ d (s) = exp λ ϕ0 F(s, Rc ) + (ϕ1 (Rc ) − ϕ0 ) F(s, R1 (ϕ1 (Rc ))) − ϕ1 (Rc )F(s, d + D) , (26)
B̂ d (s) = exp
n
o
D̂d (s) = exp λϕ1 (Rc ) F(s, Rc ) − F(s, d + D) .
(27)
Proof: The proof follows the same approach as in Appendix A, except that the disk is
replaced with the regions given in Fig. 4.
Corollary 2: With uniform-selection, the LT of the interference is lower bounded by
Ê d (s) if d ≤ D,
d
LIub (s) =
F̂ d (s) if d > D.
(28)
where Ê d (s) and F̂ d (s) are defined as
n
o
√
2 − d2 ) + F(s, D + d)
exp − λπ
F(s,
D
− exp(−λπD(D + d))
2
d
n
o , (29)
Ê (s) =
√
1
2
2
(1 − exp(−λπD(D + d))) 1 − 2D(D+d) F(s, D − d ) + F(s, D + d)
n
o
exp −λϕ0 F(s, d + D) − F(s, d − D) − exp(−4λdDϕ0 )
n
o .
F̂ d (s) =
1
(1 − exp(−4λdDϕ0 )) 1 − 4dD
F(s, d + D) − F(s, d − D)
(30)
Proof: The proof follows the same approach as in Appendix B, except that the disk is
replaced with the regions given in Fig. 4.
IV. M ULTI -C LUSTER W IRELESS N ETWORKS
In this section, we extend our analysis to multi-cluster wireless networks. Closed-access and
open-access receivers are investigated in Subsections IV.A and IV.B, respectively.
14
A. Closed-Access Receivers
For the analysis, we consider the reference closed-access receiver at the origin and add a cluster
Φxo with intensity λ over the disk b(xo , D), where kxo k has PDF given in (1), to the network.
Thanks to Slivnyak’s theorem [10, Thm. 8.10], this additional cluster and its receiver become the
representative cluster and receiver under expectation over Φ. This means this link’s performance
corresponds to the average performance of all links in any realization of the network. Therefore,
the serving transmitter under the closest-selection strategy is
xc = arg
min
y∈{Φxo |n(Φxo )>0}
kyk,
(31)
while with the uniform-selection strategy, the serving transmitter is found by
xu = Unif {Φxo | n(Φxo ) > 0} .
(32)
Then, the SINR at the origin with distance kxk relative to the serving transmitter is
SINR =
hx kxk−α
,
σ 2 + Iintra + Iinter
(33)
hy kyk−α denotes the intra-cluster interference caused by the transmitters
y∈Φxo \{x}
P
inside the representative cluster, and Iinter =
hy kyk−α represents the inter-cluster interference
where Iintra =
P
y∈Φ
caused by the transmitters outside the representative cluster.
Given the distance between the receiver and the center of the representative cluster, i.e., kxo k,
the distributions of the distances Rc = kxc k and Ru = kxu k are derived in Subsection III.B
for the closest- and uniform-selection strategies, respectively. Also, the LT of the intra-cluster
kx k
o
interference conditioned on kxo k, defined as the function LIintra
, is given by Theorems 1 and 2
for the closest-selection and uniform-selection strategies, relatively. The LT of the inter-cluster
interference is also characterized in the following theorem.
Theorem 3: The LT of the inter-cluster interference Iinter is
ZDn
Z∞n
o
o
LIinter (s) = exp −2πλp
1 − exp (−λf (s, u)) udu +
1 − exp (−λg(s, u)) udu , (34)
0
D
where f (s, u) and g(s, u) are defined as
Z π
f (s, u) =
F(s, R2 (u, θ))dθ,
(35)
0
Z
g(s, u) =
0
ϕ2 (u) n
o
F(s, R2 (u, θ)) − F(s, R̂2 (u, θ)) dθ,
(36)
15
where ϕ2 (u) = sin−1
p
D2 − u2 sin2 (θ).
D
u
, R2 (u, θ) = u cos (θ) +
p
D2 − u2 sin2 (θ) and R̂2 (u, θ) = u cos (θ) −
Proof: See Appendix C.
Using the expressions derived for the intra- and inter-cluster interferences, the coverage
probability under the closest-selection strategy is calculated by deconditioning as
PCca-c (β) = P(n(Φxo ) > 0)P(SINRc > β | n(Φxo ) > 0) = (1 − exp(−λπD2 )) ×
Z ∞Z ∞
hx r−α
P
> β fRv c (r) fkxo k (v) drdv,
σ 2 + Iintra + Iinter
0
0
where the conditional coverage probability is expressed as
v
hx r−α
2 α
P
>
β
=
exp
−βσ
r
LIc (βrα |r) LIinter (βrα ) .
2
σ + Iintra + Iinter
(37)
(38)
Then, according to the cases in Subsections III.B and C, (37) is given by
Z D Z D−v
2
ca-c
2πλre−λπr exp −βσ 2 rα LvIc (βrα |r) LIinter (βrα ) fkxo k (v) drdv
PC (β) =
Z
D
Z
0
0
D+v
∂Bv (r) −λBv (r)
e
exp −βσ 2 rα LvIc (βrα |r) LIinter (βrα ) fkxo k (v) drdv
∂r
0
D−v
Z ∞ Z D+v
∂Bv (r) −λBv (r)
e
exp −βσ 2 rα LvIc (βrα |r) LIinter (βrα ) fkxo k (v) drdv.
λ
+
∂r
D
v−D
+
λ
(39)
Finally, the coverage probability with the uniform-selection strategy can be derived with the
same procedure as in (37)-(39). It is given by
Z D Z D−v
2r
ca-u
−λπD2
exp −βσ 2 rα LvIu (βrα ) LIinter (βrα ) fkxo k (v) drdv
PC (β) = (1 − e
)
2
D
0
0
Z D Z D+v
1 ∂Bv (r)
+
exp −βσ 2 rα LvIu (βrα ) LIinter (βrα ) fkxo k (v) drdv
2
∂r
0
D−v πD
Z ∞ Z D+v
v
1 ∂Bv (r)
2 α
α
α
exp −βσ r LIu (βr ) LIinter (βr ) fkxo k (v) drdv . (40)
+
2
∂r
D
v−D πD
B. Open-Access Receivers
Here, we only consider the closest-selection strategy. Note that there is no uniform-selection
strategy in this case, since the number of transmitters is infinite. Therefore, the serving transmitter
is
xt = arg min kyk.
y∈Φ
(41)
The distribution of the distance from the reference open-access receiver at the origin, i.e., Rt =
kxt k, is given in the following theorem.
16
Theorem 4: The cumulative distribution function (CDF) of Rt is
2
1 − exp −2πλp (1 − exp (−λπr2 )) (D−r)
2
R D+r
if 0 ≤ r < D,
+ D−r (1 − exp (−λBu (r))) udu
FRt (r) =
2
1 − exp −2πλp (1 − exp (−λπD2 )) (r−D)
2
R D+r
if r ≥ D.
+ r−D (1 − exp (−λBu (r))) udu
(42)
Proof: See Appendix D.
Note that the contact distribution function of MCP is also derived in [28] using the probability
generating functional (PGFL) of PCPs in [10, Cor. 4.13]. However, our approach is more tractable
and leads to the result (42), which is much easier to numerically evaluate than [28, Thm. 1].
By taking the derivative of FRt (r), using Leibniz integral rule, and some simplifications, the
PDF of Rt can be obtained as
R
2
D+r
∂B
(r)
u
2πλp πλre−λπr (D − r)2 + D−r λ ∂r
e−λBu (r) udu
2
× exp −2πλp (1 − exp (−λπr2 )) (D−r)
2
R D+r
+ D−r (1 − exp (−λBu (r))) udu
fRt (r) =
R D+r ∂Bu (r) −λB (r)
u
2πλp r−D λ ∂r e
udu
(r−D)2
× exp −2πλp (1 − exp (−λπD2 )) 2
R D+r
+ r−D (1 − exp (−λBu (r))) udu
if 0 ≤ r < D,
(43)
if r ≥ D.
The SINR at the reference open-access receiver located at the origin is
hxt Rt −α
SINRt = 2
,
σ + It
where It =
P
(44)
hy kyk−α denotes the total interference caused by all transmitters except for
y∈Φ\{xt }
the serving transmitter. The LT of the total interference It conditioned on the serving distance
Rt is characterized in the following theorem.
17
Theorem 5: Conditioned on Rt , the LT of the total interference It is
R D−Rt
exp −2πλp 0
{1 − Au (s)} udu
R √D2 +Rt2
R D+Rt
u
√ 2 2 {1 − C u (s)} udu
+
{1
−
B
(s)}
udu
+
D−R
t
D +Rt
R
∞
+ D+Rt {1 − exp (−λg(s, u))} udu
if 0 ≤ Rt < D,
√
(45)
LIt (s|Rt ) =
R D2 +Rt2
u
exp −2πλp Rt −D
{1 − B (s)} udu
D+Rt
+ R√
{1 − C u (s)} udu
D2 +Rt2
R∞
if Rt ≥ D,
+ D+Rt {1 − exp (−λg(s, u))} udu
where Au (s), B u (s) and C u (s) are defined for a u in (12), (13) and (14), respectively. Also,
g(s, u) is given in (36).
Proof: See Appendix E.
Using Theorem 5 and (43) and the same procedure as in Subsection III.C, the coverage probability
of the reference open-access receiver in multi-cluster wireless networks is found as
Z ∞
oa
exp −βσ 2 rα LIt (βrα |r) fRt (r)dr.
PC (β) =
(46)
0
V. N UMERICAL R ESULTS AND D ISCUSSION
In this section, we provide analytical results for specific scenarios of single- and multi-cluster
wireless networks. In addition, we discuss the results and provide key design insights.
A. Single-Cluster Wireless Networks
We consider a scenario of single-cluster wireless networks in which the transmitters are
distributed according to an FHPPP with intensity λ = 0.01 m−2 in a disk with radius D = 15 m
and evaluate the coverage probability results derived in Subsection III.C. Also, we evaluate
the spectral efficiency, defined as τ = E {log(1 + SINR)}. The variance of the additive white
Gaussian noise is set to σ 2 = 0.0001. We further define the normalized (relative) distance δ =
d
.
D
In the following, we study the impact of the path loss exponent and the distance of the receiver
from the center of the disk on the coverage probability and the spectral efficiency. We also
investigate the tightness of the bounds derived in Subsection III.D.
Effect of path loss exponent: The coverage probability as a function of the minimum required
SINR β is plotted in Fig. 5 for the closest- and uniform-selection strategies, considering δ =
2
3
18
and
4
3
and α = 3 and 4. It is observed that the coverage probability is improved by increasing
the path loss exponent. However, with the uniform-selection, a higher path loss exponent has a
degrading effect when β is lower than −5 dB. That is because the SINR exhibits a tradeoff as
α increases. The power of both the desired and the interfering signals decrease, which can lead
to an increase or decrease in the SINR, depending on β and the other parameters. At a target
coverage probability of 0.8, the (horizontal) gap between uniform- and closest-selection is about
20 dB for α = 3 and 25 dB for α = 4 in the case δ =
2
3
and 12 dB for α = 3 and 18 dB for
α = 4 in the case δ = 43 .
Effect of receiver distance from the center: The coverage probability as a function of the
normalized distance δ is studied in Fig. 6 for both selection strategies, α = 4 and β = −5
and 0 dB. It is observed that, depending on β, there is an optimal value for the distance of the
receiver, about 2D for uniform-selection at β = −5 dB and 0.8D for other cases, in terms of
the coverage probability. This is due to the fact that the SINR has a tradeoff since the power of
both the desired and the interfering signals decrease as the distance of the receiver to the center
of the disk increases.
Tightness of the bounds: The tightness of the bounds on coverage probability derived in
Subsection III.D is evaluated in Fig. 7 in the cases with different selection strategies and α = 4
and δ =
2
3
and 43 . As observed, the bounds tightly approximate the performance in a broad range
of SINR thresholds β and for different positions of the receiver inside or outside the disk.
Spectral efficiency: The spectral efficiency as a function of δ is shown in Fig. 8 for both
selection strategies and α = 3 and 4. We analytically obtain the spectral efficiency from the
coverage probability as, e.g., [8, Thm. 3]1
Z ∞
Z
q
P(log (1 + SINRq ) > t) dt =
τ =
0
0
∞
1 PCq (t)
dt bits/channel use, q = {c, u} .
ln2 t + 1
(47)
As observed, closest-selection achieves a much higher spectral efficiency than uniform-selection
for different receiver distances. Also, there is a crossing point, whereby the spectral efficiency
improves as α increases before reaching a distance for the receiver location outside the disk.
Then, at higher distances, the reverse happens. This is intuitive because, at lower distances,
the power of the interfering signals, which decreases with a higher path loss exponent, is the
1
Note that this is an upper bound on the actual spectral efficiency, since this formulation assumes that the transmitter knows
all the fading coefficients between all other transmitters and the receiver, which would be a very generous assumption. A tighter
lower bound could be found using the approach described in [33].
19
1
c, δ = 2/3, α = 4
c, δ = 2/3, α = 3
u, δ = 2/3, α = 4
u, δ = 2/3, α = 3
c, δ = 4/3, α = 4
c, δ = 4/3, α = 3
u, δ = 4/3, α = 4
u, δ = 4/3, α = 3
0.9
Coverage Probability
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
−50
−30
−10
β (dB)
10
30
50
Fig. 5: Coverage probability as a function of the SINR threshold β with δ =
2
3
and 43 . c and u denote closest- and
uniform-selection, respectively.
1
c, β = −5 dB
c, β = 0 dB
u, β = −5 dB
u, β = 0 dB
0.9
Coverage Probability
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
1
2
δ
3
4
5
Fig. 6: Coverage probability as a function of normalized distance δ with α = 4.
dominant factor on the SINR. On the other hand, at higher distances, the power of the desired
signal, which increases with a smaller path loss exponent, dominates. Also, there may be an
optimal value for the distance of the receiver in terms of the spectral efficiency. This is the result
of the coverage probability behavior with the distance.
B. Multi-Cluster Wireless Networks
Here, we consider a multi-cluster wireless network where transmitters form an MCP with
intensity λ = 0.01 m−2 inside disks, all with equal radius D = 15 m, whose centers follow an
HPPP with intensity λp = 0.0004 m−2 . Closed-access receivers are distributed with normal distribution with variance σc2 around the cluster centers of transmitters. We further define normalized
standard deviation δc =
σc
.
D
In the following, the effects of the path loss exponent and variance of
the normal distribution of closed-access receivers are assessed on the theoretical results derived
in Subsections IV.A and IV.B.
20
1
c, Ex, δ = 2/3
c, LB, δ = 2/3
c, Ex, δ = 4/3
c, LB, δ = 4/3
u, Ex, δ = 2/3
u, LB, δ = 2/3
u, Ex, δ = 4/3
u, LB, δ = 4/3
0.9
Coverage Probability
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
−50
−30
−10
β (dB)
10
30
50
Fig. 7: On the tightness of the coverage probability lower bounds, α = 4. Ex and LB denote the exact result and
Spectral Efficiency (bits/channel use)
the lower bound, respectively.
4
c, α = 4
c, α = 3
u, α = 4
u, α = 3
3.5
3
2.5
2
1.5
1
0.5
0
0
1
2
δ
3
4
5
Fig. 8: Spectral efficiency as a function of normalized distance δ.
1
CA, c, α = 4
CA, c, α = 3
CA, u, α = 4
CA, u, α = 3
OA, c, α = 4
OA, c, α = 3
Coverage Probability
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
−50
−30
−10
β (dB)
10
30
50
Fig. 9: Coverage probability of the closed-access (CA) and open-access (OA) receiver as a function of threshold β
with δc = 23 .
Effect of path loss exponent: In Fig. 9, the coverage probability as a function of β is plotted for
the closest- and uniform-selection strategies in the case of closed-access receivers with δc = 32 . In
this figure, the case of open-access receivers with the closest-selection strategy is also considered.
Here, the results are presented for α = 3 and 4. As observed, a higher path loss exponent improves
21
1
c, β = −5 dB
c, β = 0 dB
u, β = −5 dB
u, β = 0 dB
Coverage Probability
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
1
2
δc
3
4
5
Fig. 10: Coverage probability of the closed-access receiver as a function of normalized standard deviation δc with
α = 4.
the coverage probability in practical minimum required SINRs.
Effect of variance of the normal distribution: The coverage probability as a function of the
normalized standard deviation δc is shown in Fig. 10. Here, the results are presented for the
closed-access receivers, considering α = 4 and β = −5 and 0 dB. It is observed that the
coverage probability decreases as the standard deviation (or variance) of the normal distribution
increases. It is intuitive because the probability of the event that the receiver is located farther
from the representative transmitters increases with σc .
VI. C ONCLUSION
In this paper, we developed a comprehensive tractable framework for the modeling and analysis of single- and multi-cluster wireless networks. Suitable for different wireless applications,
we considered two strategies for a reference receiver to select a serving transmitter over a
single-cluster wireless network: closest-selection and uniform-selection. Considering two types of
receivers—closed-access receivers and open-access receivers—we then extended our modeling to
multi-cluster wireless networks that are composed of distributed single-cluster wireless networks.
Using tools from stochastic geometry, we derived exact expressions for the coverage probability in the cases with different transmitter selection strategies and types of receivers in
single- and multi-cluster wireless networks. We also proposed tight closed-form expressions
bounding the coverage probability in the case of single-cluster wireless networks. Our analysis
revealed that a higher path loss exponent improves the performance except when the receiver
is outside and relatively far from its cluster of transmitters. In addition, although an increase in
the distance of the receiver to the center of the cluster typically degrades the performance, there
22
exists a location with optimal performance for the receiver. Also, the closest-selection strategy
significantly outperforms the uniform-selection strategy.
A PPENDIX A
P ROOF OF T HEOREM 1
The LT of the interference, given the serving distance Rc , is calculated as
X
−α
d
LIc (s|Rc ) = E exp −s
hy kyk
| n(Φ) > 0
y∈Φ\{xc }
(
)
)
(
Y
Y
1
(a)
=E
exp −shy kyk−α | n(Φ) > 0 = E
−α | n(Φ) > 0
1
+
skyk
y∈Φ\{xc }
y∈Φ\{xc }
Z
1
(b)
dy ,
(48)
= exp −λ
1−
1 + skyk−α
A\b(o,Rc )
where (a) is obtained by hy ∼ exp(1) and (b) follows from the PGFL of the PPP [10, Thm. 4.9]
and the fact that interfering nodes are farther away than Rc . Having a non-empty intersection
between A and b(o, Rc ), to compute (48), there are two types for d < D (Case 1) and two types
for d > D (Case 2) to convert from Cartesian to polar coordinates. In fact, each type denotes a
special form of A\b(o, Rc ) that can be represented by polar coordinates uniquely.
Case 1: We have two different types for the case d < D as follows.
Type 1: If A ∩ b(o, Rc ) = b(o, Rc ) as given in Fig. 2(b), i.e., 0 ≤ Rc < D − d, then
Z2π RZ1 (θ)
LdIc (s|Rc ) = exp −λ
1−
0
1
1 + sx−α
Z2π ZRc
xdxdθ + λ
1−
0
0
1
1 + sx−α
xdxdθ .(49)
0
This can be simplified to
!
Z 2π Z R1 (θ)
x
x
LdIc (s|Rc ) = exp 2πλ
α dx − λ
α dxdθ
1 + xs
1 + xs
0
0
0
Z π
(c)
= exp πλF(s, Rc ) − λ
F(s, R1 (θ))dθ ,
Z
Rc
(50)
0
where (c) follows from replacing xα with u and calculating the corresponding integral based on
the formula [32, (3.194.1)] which uses the Gauss hypergeometric function 2 F1 .
Type 2: If A ∩ b(o, Rc ) 6= b(o, Rc ) as given in Fig. 2(a), i.e., D − d ≤ Rc < D + d, then
!
Z ϕ1 (Rc ) Z R1 (θ)
1
1−
LdIc (s|Rc ) = exp −λ
xdxdθ
1 + sx−α
−ϕ1 (Rc ) Rc
Z ϕ1 (Rc ) n
o
= exp −λ
F(s, R1 (θ)) − F(s, Rc ) dθ .
(51)
0
23
Case 2: As observed from Fig. 2(c) for the case d > D, we have two different types of the
LT of the interference depending on whether or not, at an angle, the lower boundary of A within
the two tangent lines crossing the origin is included in the boundary of A\b(o, Rc ).
√
Type 1: Since the intersection angle ϕ1 (r) at r = d2 − D2 is equal to the angle of the
√
tangent lines ϕ0 , if d − D < Rc < d2 − D2 , we have
Z −ϕ1 (Rc ) Z R1 (θ)
1
d
1−
LIc (s|Rc ) = exp −λ
xdxdθ
1 + sx−α
R̂1 (θ)
−ϕ0
Z ϕ1 (Rc ) Z R1 (θ)
Z ϕ0 Z R1 (θ)
1
1
1−
1−
+
xdxdθ +
xdxdθ
1 + sx−α
1 + sx−α
−ϕ1 (Rc ) Rc
ϕ1 (Rc ) R̂1 (θ)
ϕZ
1 (Rc )
Zϕ0 n
o
n
o
= exp −λ
F(s, R1 (θ)) − F(s, R̂1 (θ)) dθ − λ
F(s, R1 (θ)) − F(s, Rc ) dθ .(52)
0
ϕ1 (Rc )
√
Type 2: If d2 − D2 < Rc < d + D, then
Z ϕ1 (Rc ) Z
d
LIc (s|Rc ) = exp −λ
−ϕ1 (Rc )
R1 (θ)
1−
Rc
1
1 + sx−α
!
xdxdθ
ϕ1 (Rc ) n
Z
= exp −λ
o
F(s, R1 (θ)) − F(s, Rc ) dθ .
(53)
0
A PPENDIX B
P ROOF OF T HEOREM 2
The LT of the interference is
X
−α
d
LIu (s) = E exp −s
hy kyk
| n(Φ) > 0
y∈Φ\{xu }
(
)
(
Y
Y
(a)
−α
=E
exp −shy kyk
| n(Φ) > 0 = E
)
1
| n(Φ) > 0
1 + skyk−α
y∈Φ\{x }
y∈Φ\{xu }
(Z
(b)
=E
u
)
n(Φ)−1
∞
1
fR (x)dx
| n(Φ) > 0
1 + sx−α u
0
Z ∞
k−1
∞
2
2 k
1
(c) X exp(−λπD )(λπD )
=
fR (x)dx
k!(1 − exp(−λπD2 ))
1 + sx−α u
0
k=1
=
(1 −
1
R∞
exp(−λπD2 ))
0
1
f (x)dx
1+sx−α Ru
Z
∞
X
exp(−λπD2 )(λπD2 )k
k=0
k!
0
∞
×
1
fR (x)dx
1 + sx−α u
k
!
2
− exp(−λπD )
24
exp(−λπD2 )
=
exp λπD2
R∞ 1
(1 − exp(−λπD2 )) 1+sx−α fRu (x)dx
(d)
Z∞
1
fR (x)dx
1 + sx−α u
!
!
−1 ,
(54)
0
0
where fRd u is the PDF of Ru obtained from the distributions in (9)-(10) and (a) is found by
hy ∼ exp(1). Also, (b) follows from the fact that conditioned on n(Φ), Φ is a BPP where the
distance kyk of each point is i.i.d. with distribution fRd u , (c) follows from P(n(Φ) = k | n(Φ) >
0) =
exp(−λπD2 )(λπD2 )k
,
k!(1−exp(−λπD2 ))
and (d) is found by the moment-generating function (MGF) of a Poisson
random variable with mean λπD2 and some simplifications. According to (9)-(10), there are two
cases to compute the integral:
Case 1: If d ≤ D, then
Z D−d
Z D+d
Z ∞
1
1
1
2x
1 ∂Bd (x)
fRu (x)dx =
dx +
dx.
−α
−α
2
−α
1 + sx
1 + sx D
πD2 ∂x
0
D−d 1 + sx
0
Case 2: If d > D, then
Z ∞
Z D+d
1
1
1 ∂Bd (x)
fRu (x)dx =
dx.
−α
−α
1 + sx
πD2 ∂x
0
d−D 1 + sx
(55)
(56)
A PPENDIX C
P ROOF OF T HEOREM 3
The LT of the inter-cluster interference is
!)
(
)
(
X X
Y Y
hy kyk−α
=E
exp −shy kyk−α
LIinter (s) = E exp −s
x∈Φp y∈Φx
(
1
1 + skyk−α
x∈Φp y∈Φx
)
Y Y
1
1 + skz + xk−α
x∈Φp y∈Φx
x∈Φp z∈Φo
Z
Y
1
(c)
=E
dz
exp −λ
1−
1 + skz + xk−α
b(o,D)
(a)
=E
Y Y
(b)
=E
x∈Φp
(d)
Z
= exp −λp
R2
Z
1 − exp −λ
1−
b(o,D)
1
1 + skz + xk−α
dz
dx ,
(57)
where (a) follows from hy ∼ exp(1), and (b) comes from the fact that y ∈ Φx can be written
as y = z + x, where z is taken from the FHPPP Φo with intensity λ over b(o, D) and x ∈ Φp .
Also, both (c) and (d) follow from the PGFL of the PPP. In order to convert the inner integral
from Cartesian to polar coordinates, there are two cases:
25
Case 1: If kxk ≤ D, then
Z
Z 2π Z R2 (kxk,θ)
1
1
1−
f (s, kxk) =
1−
dz =
xdxdθ
1 + sx−α
1 + skz + xk−α
0
0
b(o,D)
Z π
F(s, R2 (kxk, θ))dθ,
(58)
=
0
p
where R2 (kxk, θ) = D2 − kxk2 sin2 (θ) + kxk cos (θ).
Case 2: If kxk > D, then
1−
Z
g(s, kxk) =
b(o,D)
1
1 + skz + xk−α
ϕ2Z(kxk) R2Z
(kxk,θ)
1
1−
1 + sx−α
dz =
xdxdθ
−ϕ2 (kxk) R̂2 (kxk,θ)
ϕ2 (kxk) n
Z
o
F(s, R2 (kxk, θ)) − F(s, R̂2 (kxk, θ)) dθ,
=
(59)
0
where ϕ2 (kxk) = sin−1
D
kxk
and R̂2 (kxk, θ) = kxk cos (θ) −
p
D2 − kxk2 sin2 (θ). Therefore,
converting the outer integral from Cartesian to polar coordinates, the final result can be written
as
ZDn
Z∞n
o
o
LIinter (s) = exp −2πλc
1 − exp (−λf (s, u)) udu +
1 − exp (−λg(s, u)) udu .(60)
0
D
A PPENDIX D
P ROOF OF T HEOREM 4
Let us define Rx = min kyk, which is the distance from the reference receiver to its closest
y∈Φx
transmitter in the cluster with parent point x ∈ Φp . Then, the CDF of Rt = min Rx is
x∈Φp
(a)
= 1 − lim E P min Rx > r | n(Ψρ )
ρ→∞
x∈Ψρ
))
FRt (r) = P (Rt < r) = 1 − P min Rx > r
x∈Φp
( (
Y
(b)
= 1 − lim E E
P (Rx > r) | n(Ψρ )
ρ→∞
x∈Ψρ
( (
(c)
= 1 − lim E E
ρ→∞
Y
P (n(Φx ) = 0) P (Rx > r | n(Φx ) = 0)
x∈Ψρ
))
+P (n(Φx ) > 0) P (Rx > r | n(Φx ) > 0) | n(Ψρ )
26
( (
))
(d)
Y
= 1 − lim E E
ρ→∞
exp(−λπD2 ) + (1 − exp(−λπD2 ))E(x) | n(Ψρ )
x∈Ψρ
(
n(Ψρ ) )
Z
1
(e)
= 1 − lim E
exp(−λπD2 ) + (1 − exp(−λπD2 ))E(x)dx
ρ→∞
πρ2 b(o,ρ)
!!
Z
1
(f )
= 1 − lim exp −λp πρ2 1 − 2
exp −λπD2 + 1 − exp −λπD2 E(x)dx
ρ→∞
πρ
b(o,ρ)
Z
= 1 − exp −λp
1 − exp −λπD
2
(1 − E(x))dx ,
(61)
R2
where Ψρ = Φp ∩ b(o, ρ) and (a) follows from lim Ψρ = Φp . Then, (b) is due to the facts that
ρ→∞
Rx is a function of x ∈ Ψρ and Ψρ conditioned on n(Ψρ ) is a BPP. Also, (c) is obtained by
conditioning on the existence of a transmitter inside each cluster, and (d) comes from Subsection
III.B, whereby E(x) is defined as
E(x) =
exp (−λ |Cx |) − exp(−λπD2 )
1 − exp(−λπD2 )
(62)
where |Cx | is the area of the intersection between b(o, r) and b(x, D). Also, note that
P (Rx > r | n(Φx ) = 0) = 1, for 0 < r < ∞. In addition, (e) follows from the fact that the
points x are i.i.d. uniformly distributed on b(o, ρ), and (f ) is found by the MGF of the number
of points of Ψρ , which is Poisson with mean λp πρ2 .
According to Fig. 3, there are two cases for |Cx |:
Case 1: If kxk ≤ D, then
2
πr
|Cx | =
Bkxk (r)
πD2
Case 2: If kxk > D, then
0
|Cx | =
Bkxk (r)
πD2
kxk < D − r,
kxk > r − D and kxk ≥ D − r,
(63)
kxk ≤ r − D,
kxk > r + D,
kxk > r − D and kxk ≤ r + D,
(64)
kxk ≤ r − D,
where Bkxk (r) is given in (5). Therefore, converting (61) from Cartesian to polar coordinates
according to (62)-(64), the CDF of Rt is found as
Z D−r
Z
2
FRt (r) = 1 − exp −2πλp
1 − exp(−λπr ) udu +
0
Z
D+r
D
(1 − exp (−λBu (r))) udu
D−r
Z
∞
(1 − exp (−λBu (r))) udu +
+
D
D+r
(1 − exp (0)) udu ,
(65)
27
if 0 ≤ r < D,
Z
FRt (r) = 1 − exp −2πλp
r−D
(1 − exp(−λπD2 )) (1 − 0) udu
0
D
Z
Z
(1 − exp (−λBu (r))) udu +
r−D
Z ∞
(1 − exp (0)) udu ,
+
D+r
(1 − exp (−λBu (r))) udu
+
D
(66)
D+r
if D ≤ r < 2D, and
Z
FRt (r) = 1 − exp −2πλp
D
(1 − exp(−λπD2 )) (1 − 0) udu
0
Z
r−D
Z
2
+
D
Z ∞
+
(1 − exp(−λπD )) (1 − 0) udu +
(1 − exp (0)) udu ,
D+r
(1 − exp (−λBu (r))) udu
r−D
(67)
D+r
if r ≥ 2D. Here, though derived differently, (66) and (67) leads to the same expression. The
final result can be obtained by some simplifications.
A PPENDIX E
P ROOF OF T HEOREM 5
The LT of the total interference, conditioned on the serving distance Rt , is
(
!)
X
(a)
LIt (s|Rt ) = E exp −s
hy kyk−α
y∈Φ\b(o,Rt )
(
)
(
Y
(b)
exp −shy kyk−α
=E
)
1
=E
1 + skyk−α
x∈Φp y∈Φx \b(o,Rt )
x∈Φp y∈Φx \b(o,Rt )
Z
Y
1
(c)
=E
exp −λ
1−
dy
1 + skyk−α
b(x,D)\b(o,Rt )
x∈Φp
Z
Z
1
(d)
= exp −λp
1 − exp −λ
1−
dy
dx , (68)
1 + skyk−α
R2
b(x,D)\b(o,Rt )
Y
Y
Y
where (a) follows from the fact that all transmitters with distance more than Rt are interferers,
(b) comes from hy ∼ exp(1), and both of (c) and (d) follow from the PGFL. According to
Theorems 1 and 3, there are the following cases for the inner integral.
Case 1: If kxk ≤ D and kxk < D − Rt , then
Z
1
exp −λ
1−
dy = Akxk (s).
−α
1 + skyk
b(x,D)\b(o,Rt )
(69)
28
Case 2: If kxk ≤ D and kxk > Rt − D and kxk ≥ D − Rt , then
Z
1
dy = B kxk (s).
exp −λ
1−
−α
1 + skyk
b(x,D)\b(o,Rt )
p
Case 3: If kxk > D and kxk > Rt2 + D2 and kxk ≤ Rt + D, then
Z
1
dy = C kxk (s).
exp −λ
1−
−α
1 + s||y||
b(x,D)\b(o,Rt )
p
Case 4: If kxk > D and kxk > Rt − D and kxk ≤ Rt2 + D2 , then
Z
1
1−
exp −λ
dy = B kxk (s).
−α
1 + skyk
b(x,D)\b(o,Rt )
Case 5: If kxk > Rt + D, then
Z
1
1−
exp −λ
dy = exp (−λg(s, kxk)) .
1 + skyk−α
b(x,D)\b(o,Rt )
Case 6: If kxk < Rt − D, then
Z
1
dy = 0.
exp −λ
1−
1 + skyk−α
b(x,D)\b(o,Rt )
Therefore, converting the outer integral from Cartesian to polar coordinates according to
(74), the LT can be given by
Z
LIt (s|Rt ) = exp −2πλp
{1 − Au (s)} udu
Z √ 2 2
D
D +Rt
{1 − B u (s)} udu +
+
D−Rt
D+Rt
(72)
(73)
(74)
(69)-
{1 − B u (s)} udu
D
Z
+ √
(71)
D−Rt
0
Z
(70)
Z
u
∞
{1 − exp (−λg(s, u))} udu ,
{1 − C (s)} udu +
D2 +Rt2
(75)
D+Rt
if 0 ≤ Rt < D,
Z
Rt −D
LIt (s|Rt ) = exp −2πλp
{1 − exp(0)} udu
Z √ 2 2
0
Z
D
Z
D +Rt
{1 − B u (s)} udu +
+
Rt −D
D+Rt
+ √
{1 − B u (s)} udu
D
Z
u
∞
{1 − C (s)} udu +
D2 +Rt2
{1 − exp (−λg(s, u))} udu
if D ≤ Rt < 2D, and
Z
LIt (s|Rt ) = exp −2πλp
Z √D2 +Rt2
Rt −D
{1 − exp(0)} udu +
D+Rt
+ √
u
Z
∞
{1 − C (s)} udu +
D2 +Rt2
(76)
{1 − B u (s)} udu
Rt −D
0
Z
,
D+Rt
{1 − exp (−λg(s, u))} udu ,
D+Rt
if Rt ≥ 2D. The final result can be obtained by some simplifications.
(77)
29
R EFERENCES
[1] J. G. Andrews, S. Buzzi, W. Choi, S. Hanly, A. Lozano, A. C. K. Soong, and J. C. Zhang, "What will 5G be?," IEEE J.
Sel. Areas Commun., vol. 32, no. 6, pp. 1065-1082, Jun. 2014.
[2] F. Boccardi, R. W. Heath, A. Lozano, T. L. Marzetta, and P. Popovski, "Five disruptive technology directions for 5G," IEEE
Commun. Mag., vol. 52, no. 2, pp. 74-80, Feb. 2014.
[3] M. Peng, C. Wang, V. Lau, and H. V. Poor, "Fronthaul-constrained cloud radio access networks: insights and challenges,"
IEEE Wireless Commun., vol. 22, no. 2, pp. 152-160, April 2015.
[4] M. Haenggi, J. G. Andrews, F. Baccelli, O. Dousse, and M. Franceschetti, "Stochastic geometry and random graphs for the
analysis and design of wireless networks," IEEE J. Sel. Areas Commun., vol. 27, no. 7, pp. 1029-1046, Sep. 2009.
[5] H. Elsawy, E. Hossain, and M. Haenggi, "Stochastic geometry for modeling, analysis, and design of multi-tier and cognitive
cellular wireless networks: A survey," IEEE Commun. Surveys Tuts., vol. 15, no. 3, pp. 996-1019, 3rd quarter 2013.
[6] H. ElSawy, A. K. Sultan-Salem, M. S. Alouini, and M. Z. Win, "Modeling and Analysis of Cellular Networks Using
Stochastic Geometry: A Tutorial," IEEE Commun. Surveys Tuts., vol. 19, no. 1, pp. 167-203, 1st quarter 2017.
[7] J. G. Andrews, R. K. Ganti, M. Haenggi, N. Jindal, and S. Weber, "A primer on spatial modeling and analysis in wireless
networks," IEEE Commun. Mag., vol. 48, no. 11, pp. 156-163, Nov. 2010.
[8] J. G. Andrews, F. Baccelli, and R. K. Ganti, "A tractable approach to coverage and rate in cellular networks," IEEE Trans.
Commun., vol. 59, no. 11, pp. 3122-3134, Nov. 2011.
[9] J. G. Andrews, A. K. Gupta, and H. S. Dhillon, "A primer on cellular network analysis using stochastic geometry," arXiv
preprint, Apr. 2016, available online: arxiv.org/abs/1604.03183.
[10] M. Haenggi, Stochastic Geometry for Wireless Networks. Cambridge University Press, 2012.
[11] D. Torrieri and M. C. Valenti, "The outage probability of a finite ad hoc network in Nakagami fading," IEEE Trans.
Commun., vol. 60, no. 11, pp. 3509-3518, Nov. 2012.
[12] K. Venugopal, M. C. Valenti, and R. W. Heath, "Interference in finite-sized highly dense millimeter wave networks," IEEE
ITA’2015, San Diego, USA, Feb. 2015.
[13] M. Afshang and H. S. Dhillon, "Optimal geographic caching in finite wireless networks," Proc. IEEE SPAWC’2016,
Edinburgh, Scotland, July 2016.
[14] V. V. Chetlur and H. S. Dhillon, "Downlink coverage probability in a finite network of unmanned aerial vehicle (UAV)
base stations," Proc. IEEE SPAWC’2016, Edinburgh, Scotland, July 2016.
[15] S. Srinivasa and M. Haenggi, "Distance Distributions in Finite Uniformly Random Networks: Theory and Applications,"
IEEE Trans. Veh. Technol., vol. 59, no. 2, pp. 940-949, Feb. 2010.
[16] M. Afshang and H. S. Dhillon, "Fundamentals of Modeling Finite Wireless Networks using Binomial Point Process", IEEE
Trans. Wireless Commun., vol. 16, no. 5, pp. 3355-3370, May 2017.
[17] V. V. Chetlur and H. S. Dhillon, "Downlink Coverage Analysis for a Finite 3D Wireless Network of Unmanned Aerial
Vehicles", IEEE Trans. Commun., vol. 65, no. 10, pp. 4543-4558, July 2017.
[18] J. Guo, S. Durrani, and X. Zhou, "Outage probability in arbitrarily-shaped finite wireless networks," IEEE Trans. Commun.,
vol. 62, no. 2, pp. 699-712, Feb. 2014.
[19] J. Guo, S. Durrani, and X. Zhou, "Performance analysis of arbitrarily-shaped underlay cognitive networks: Effects of
secondary user activity protocols," IEEE Trans. Commun., vol. 63, no. 2, pp. 376-389, Feb. 2015.
[20] Z. Khalid and S. Durrani, "Distance Distributions in Regular Polygons," IEEE Trans. Veh. Technol., vol. 62, no. 5, pp.
2363-2368, June 2013.
30
[21] M. C. Valenti, D. Torrieri, and S. Talarico, "A direct approach to computing spatially averaged outage probability," IEEE
Commun. Lett., vol. 18, no. 7, pp. 1103-1106, July 2014.
[22] M. Afshang and H. S. Dhillon, "Poisson cluster process based analysis of HetNets with correlated user and base station
locations," arXiv preprint, Dec. 2016, available online: https://arxiv.org/abs/1612.07285.
[23] M. Afshang, H. S. Dhillon, and P. H. J. Chong, "Modeling and performance analysis of clustered device-to-device networks,"
IEEE Trans. Wireless Commun., vol. 15, no. 7, pp. 4957-4972, July 2016.
[24] "3rd generation partnership project; technical specification group radio access network; scenarios and requirements for
small cell enhancements for E-UTRA and E-UTRAN (release 13)," 3GPP TR 36.932 V13.0.0, 2015.
[25] "Further advancements for E-UTRA physical layer aspects," 3GPP TR 36.814, 2010.
[26] V. Suryaprakash, J. Møller, and G. Fettweis, "On the Modeling and Analysis of Heterogeneous Radio Access Networks
Using a Poisson Cluster Process," IEEE Trans. Wireless Commun., vol. 14, no. 2, pp. 1035-1047, Feb. 2015.
[27] R. K. Ganti and M. Haenggi, "Interference and Outage in Clustered Wireless Ad Hoc Networks," IEEE Trans. Wireless
Commun., vol. 55, no. 9, pp. 4067-4086, Sept. 2009.
[28] M. Afshang, C. Saha, and H. S. Dhillon, "Nearest-Neighbor and Contact Distance Distributions for Matern Cluster Process,"
IEEE Commun. Lett., vol. 21, no. 12, pp. 2686-2689, Dec. 2017.
[29] N. Deng, W. Zhou, and M. Haenggi, "Heterogeneous Cellular Network Models with Dependence", IEEE J. Sel. Areas
Commun., vol. 33, no. 10, pp. 2167-2181, Oct. 2015.
[30] T. Flanagan, Creating cloud base stations with TI’s KeyStone multicore architecture, Texas Instruments White Paper, TX,
USA, Oct., 2011, available online: http://www.ti.com/lit/wp/spry183/spry183.pdf.
[31] Z. Han and K. Liu, Resource Allocation for Wireless Networks: Basics, Techniques, and Applications. Cambridge University
Press, 2008.
[32] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and Products, 7th ed. Academic Press, 2007.
[33] G. George, R. K. Mungara, A. Lozano, and M. Haenggi, "Ergodic Spectral Efficiency in MIMO Cellular Networks", IEEE
Trans. Wireless Commun., vol. 16, pp. 2835-2849, May 2017.
| 7cs.IT
|
arXiv:1211.7306v6 [math.AC] 23 Jun 2016
ON POLYNOMIALS WITH GIVEN HILBERT FUNCTION AND
APPLICATIONS
ALESSANDRA BERNARDI, JOACHIM JELISIEJEW, PEDRO MACIAS MARQUES, KRISTIAN
RANESTAD
Abstract. Using Macaulay’s correspondence we study the family of Artinian Gorenstein
local algebras with fixed symmetric Hilbert function decomposition. As an application we
give a new lower bound for cactus varieties of the third Veronese embedding. We discuss
the case of cubic surfaces, where interesting phenomena occur.
Introduction
Macaulay established a correspondence between polynomials and Artinian local Gorenstein
algebras. In particular, any polynomial is a dual socle generator of an Artinian local Gorenstein
algebra. In this paper we interpret the Hilbert function of the algebra as a Hilbert function of
the corresponding polynomial, and give a description of the set of polynomials with given symmetric Hilbert function decomposition, in a fixed polynomial ring. We consider polynomials f
in a divided power ring S = Kdp [x1 , . . . , xn ], and a polynomial ring T = K[y1 , . . . , yn ] acting
on S by contraction (see Section 0.1). The Artinian local Gorenstein algebra A associated to
f ∈ S is the quotient T /f ⊥ where f ⊥ is the annihilator ideal of f . Thus Spec(T /f ⊥ ) ⊂ Spec(T )
is a local Gorenstein scheme supported at the origin of the affine space Spec(T ).
The application we have in mind is that of apolarity and the dimension of cactus varieties
of cubic forms. Cactus varieties are generalizations of secant varieties.
Definition 1. Let X ⊂ PN be a projective variety. The r-th cactus variety Cactusr (X) of X
is the closure of the union of the linear spaces spanned by length r subschemes on X.
Here we abuse slightly the notation of variety, since the cactus variety is often a reducible
n+3
algebraic set. We are interested in the case when X ≃ Pn is embedded into P( 3 )−1 by the
third Veronese embedding.
Consider, like above, a divided power ring S̄ = Kdp [x0 , . . . , xn ], and a polynomial ring
T̄ = K[y0 , . . . , yn ] acting on S̄ by contraction. A cubic form F ∈ S̄3 up to multiplication
by scalars is a point in P(S̄3 ). The pure cubes form a n-dimensional subvariety V3,n ⊂ P(S̄3 ).
The least length of a subscheme Z ⊂ V3,n whose linear span contains F is called the cactus
rank of F . The closure of the set of cubic forms with cactus rank r is the r-th cactus variety
of V3,n , denoted Cactusr (V3,n ). Via the contraction action, T̄ is the natural homogeneous
coordinate ring on P(S̄1 ), and a Z ⊂ V3,n contains F in its span, if and only if its homogeneous
ideal IZ ⊂ T̄ is contained in F ⊥ . This classical fact is called the apolarity lemma and is the
motivation for defining a subscheme Z ⊂ P(S̄1 ) apolar to F if IZ ⊂ F ⊥ .
2010 Mathematics Subject Classification. Primary 13H10, Secondary 14Q15, 14C05.
Key words and phrases. cactus rank, Artinian Gorenstein local algebra.
1
2
A. BERNARDI, J. JELISIEJEW, P. MACIAS MARQUES, K. RANESTAD
We apply Macaulay’s correspondence to investigate local Gorenstein schemes that are apolar
to F . Our main result is the following lower bound on the dimension of cactus varieties of
cubic forms.
Theorem 2 (Corollary 44). Let n ≥ 8 and 18 ≤ r ≤ 2n + 2 and let V3,n be the third Veronese
embedding of Pn in PN , with N = n+3
3 . Then
(
−1
if r ≥ 18 even,
(rn + r − 1) + r(r−2)(r−16)
48
dim Cactusr (V3,n ) ≥
−
2
if
r ≥ 19 odd.
(rn + r − 1) + (r−1)(r−3)(r−17)
48
Hence, under these assumptions the r-th secant variety of V3,n is strictly contained in the r-th
cactus variety of V3,n .
S
The r-th secant variety of V3,n is σr (V3,n ) = P1 ,...,Pr ∈V3,n hP1 , . . . , Pr i. The fact that the
inclusion of σr (V3,n ) in Cactusr (V3,n ) is strict is a consequence of the inequality
dim σr (V3,n ) ≤ rn + r − 1.
The righ-hand side is the expected dimension of the secant variety, and it is an easy parameter
count that gives an upper bound. The actual dimension of of the secant variety is known
thanks to the Alexander and Hirschowitz Theorem [Alexander, Hirschowitz 1995].
For r ≥ 2n+2 the variety Cactusr (V3,n ) fills the ambient space, see [Bernardi, Ranestad 2012].
Observe that σr (V3,n ) = Cactusr (V3,n ), for r ≤ 13, see [Casnati, Notari 2011] for the cases
r ≤ 11, and [Casnati, Jelisiejew, Notari 2015] for the remaining cases.
The link between the Artinian local Gorenstein algebras and apolar schemes to a homogeneous form F ∈ S̄ is provided by the fact that if f = F (1, x1 , . . . , xn ), then
Spec(K[y1 , . . . , yn ]/f ⊥ ) ⊂ P(S̄1 )
is a local scheme supported at [1 : 0 : · · · : 0] that is apolar to F . We define
Definition 3. The minimal length of a local apolar scheme to F is called the local cactus rank
of F .
That link is strengthened by the following result, where we call the sum of the homogeneous
terms of a polynomial g of degree at most d the degree-d tail of g.
Proposition 4. Let F be a homogeneous polynomial of degree d, and let f = F (1, x1 , . . . , xn ).
Let Γ be a zero-dimensional scheme of minimal length among local schemes supported at
[l] = [1 : 0 : · · · : 0] that are apolar to F . Then Γ is the affine apolar scheme of a polynomial g
whose degree-d tail equals f .
A particularly important problem is to find the cactus rank of a general form in S̄d ,
i.e. the minimal r such that Cactusr (Vd,n ) = P(S̄d ). Our results do not improve previous
known bounds, so this remains a major open problem in this theory. We refer the interested
reader to [Iarrobino, Kanev 1999], [Bernardi, Ranestad 2012], [Buczyńska, Buczyński 2014],
and [Bernardi, Brachat, Mourrain 2014].
Now, the first step in order to be able to compute the cactus rank, is to describe the
structure of minimal apolar schemes. If we start by considering a minimal zero-dimensional
scheme Γ apolar to a form F and decompose it as Γ = Γ1 ∪ · · · ∪ Γk , where each Γi is a scheme
supported on one point, then we take the corresponding decomposition F = F1 + · · · + Fk ,
where Γi is a minimal local scheme apolar to Fi for i = 1, . . . , k. According to Proposition 4,
one would like to find invariants for local apolar Gorenstein schemes, parameterizing the degreed tails of all polynomials that define a scheme with given invariant. Iarrobino’s analysis
3
([Iarrobino 1994]), that we recall in the first section, provides one such discrete invariant: the
symmetric Hilbert function decomposition. Then, if one wants to estimate the dimension of
the sets of polynomials with the same local cactus rank, one needs to understand the structure
of the polynomials with the same symmetric Hilbert function decomposition. This is what we
do in Section 2 by the use of standard and exotic forms, that explain the unlucky behavior that
the number of variables involved in each homogeneous summand of a given polynomial f may
be larger than what is expected from the Hilbert function and its symmetric decomposition.
Motivated by this, we describe the family Ffm of polynomials g whose linear partials and
Hilbert function coincide with those of f , and such that g − f is part of an exotic summand
for g. In Proposition 21 we show that the polynomials in Ffm define isomorphic apolar algebras
and compute in Corollary 23 its dimension. In Proposition 24 we give a decomposition of any
polynomial as a sum of a polynomial in standard form and an exotic summand. In Section 3 we
focus on the local cactus rank by proving Proposition 4 and computing the local cactus rank
of a general cubic surface. Finally in Section 4 we use this description in order to estimate
the dimension of the set of all the polynomials with the same symmetric Hilbert function
decomposition, which allows us to estimate the dimension of the Cactus variety. In particular,
we prove our lower bound for this dimension in Corollary 44.
0.1. Notations. For main applications we consider homogeneous forms in S̄ = Kdp [x0 , . . . , xn ]
and their dehomogenization in S = Kdp [x1 , . . . , xn ]:
πx0 : S̄ → S;
F (x0 , x1 , ..., xn ) 7→ F (1, x1 , ..., xn ).
We consider the action of the polynomial ring T̄ := K[y0 , . . . , yn ] on S̄ by contraction: if
α = (α0 , . . . , αn ) and β = (β0 , . . . , βn ) are multi-indices, then
(
xβ−α if β ≥ α,
α
β
y x =
0
otherwise.
Similarly, we consider the action of the polynomial ring T := K[y1 , . . . , yn ] on S:
T × S → S;
α
(ψ, f ) 7→ ψ(f )
β
defined by this contraction restricted to y , x , when α0 = β0 = 0.
Note that we are using the same notation for ordinary powers in T and divided powers in
S, unlike what is usually done in the literature (for properties of divided power rings see for
instance [Iarrobino, Kanev 1999, Appendix A], where the divided power xβ would be written
as x[β] ).
In characteristic 0 we could have used ordinary differentiation, and therefore, by abuse of
language, we call ψ(f ) ∈ S a partial of a polynomial f ∈ S for any ψ ∈ T .
1. Preliminaries
We begin this section by presenting the Macaulay correspondence between polynomials and
Artinian Gorenstein local rings, which is the starting point of our theory.
1.1. Macaulay correspondence. Let K be an algebraically closed field of characteristic
other than 2, 3 and consider the divided power ring S := Kdp [x1 , . . . , xn ]. Consider the action
of the polynomial ring T := K[y1 , . . . , yn ] on S defined in Subsection 0.1.
Let S1 and T1 be the degree-one parts of S and T respectively. With respect to the action
above (classically known as apolarity), S1 and T1 are natural dual spaces and hx1 , . . . , xn i and
hy1 , . . . , yn i are dual bases. The annihilator of a polynomial f of degree d is an ideal in T
which we denote by f ⊥ ⊂ T . The quotient Tf := T /f ⊥ is a local Artinian Gorenstein ring (see
4
A. BERNARDI, J. JELISIEJEW, P. MACIAS MARQUES, K. RANESTAD
e.g. [Iarrobino, Kanev 1999, Lemma 2.14]). In fact, Tf is finitely generated as a K-module so
it is Artinian. The image of T1 in Tf generates the unique maximal ideal m, so Tf is local.
Furthermore Tf has a 1-dimensional socle, the annihilator of the maximal ideal, namely md ,
so Tf is Gorenstein. In addition, if f is a form, f ⊥ is a homogeneous ideal and therefore Tf is
an Artinian Gorenstein graded local ring.
1.2. Symmetric decomposition of the Hilbert function of a polynomial. We consider
now a polynomial f ∈ S = Kdp [x1 , . . . , xn ] and let f ⊥ ⊂ T = K[y1 , . . . , yn ] be its annihilator with respect to contraction. We shall interpret a Hilbert function for the local Artinian
Gorenstein quotient Tf = T /f ⊥ in terms of a filtration of the space of partials of the polynomial
f . In particular we recall and interpret Iarrobino’s analysis of Hilbert functions on associated
graded algebras of Tf and their symmetric decomposition. We will apply this analysis in the
next section to characterize the polynomials with a given Hilbert function.
The local Artinian Gorenstein quotient ring Tf = T /f ⊥ is naturally isomorphic to
Diff(f ) = {ψ(f ) | ψ ∈ T }
the space of all partials of f , as a T -module.
Following [Iarrobino 1994], we consider Hilbert functions on graded rings associated to two
filtrations of Tf . Let m be the maximal ideal of Tf . The m-adic filtration
Tf = m0 ⊃ m ⊃ m2 ⊃ · · · ⊃ md ⊃ md+1 = 0
where d = deg f , defines an associated graded ring
Tf∗ = ⊕di=0 mi /mi+1
whose Hilbert function we denote by Hf . The Löewy filtration
Tf = (0 : md+1 ) ⊃ (0 : md ) ⊃ · · · ⊃ (0 : m2 ) ⊃ (0 : m) ⊃ 0
induces the following sequence of ideals of Tf∗ : For each a = 0, 1, 2, . . . let
d−a
d+1−a−i
Ca = ⊕d−a
) ∩ mi / (0 : md+1−a−i ) ∩ mi+1 ⊂ Tf∗ ,
i=0 Ca,i = ⊕i=0 (0 : m
and consider the Tf∗ -modules
Qa = Ca /Ca+1 ,
a = 0, 1, 2, . . .
and their respective Hilbert functions
∆Q,a = H Qa .
Each Qa decomposes as a sum Qa = ⊕di=0 Qa,i , where Qa,i = Ca,i /Ca+1,i . We can check from
the definitions that for d ≥ 2 and any a > d − 2, we have Ca = Qa = 0. The following is an
important result on the structure of these modules.
Proposition 5. [Iarrobino 1994, Theorem 1.5] The Tf∗ -modules Qa , a = 0, 1, 2, . . . , satisfy
the following reflexivity condition
∼
Q∨
a,i = Qa,d−a−i .
In particular, the Hilbert function ∆Q,a = H Qa is symmetric about (d − a)/2, and thus the
Hilbert function H(Tf∗ ) of Tf∗ has a symmetric decomposition
X
H(Tf∗ ) =
∆Q,a .
a
The possible symmetric decompositions of the Hilbert function is restricted by the fact that
the partial sums of the symmetric decomposition are Hilbert functions of suitable quotients of
Tf∗ .
5
Corollary 6. [Iarrobino 1994, Section 5B, p. 69] The Hilbert function of Tf∗ /Cα+1 satisfies
H(Tf∗ /Cα+1 ) =
α
X
∆Q,a .
a=0
In particular every partial sum
degree 1.
Pα
∆Q,a is the Hilbert function of a K-algebra generated in
a=0
Iarrobino has listed all possible symmetric decompositions of Hilbert functions of rings Tf∗
with dim Tf∗ ≤ 16, see [Iarrobino 1994, Section 5].
We now interpret the ideal Ca and the module Qa in terms of the space Diff(f ) of partials
of f . This interpretation depends on the isomorphism
τ : T /f ⊥ → Diff(f ),
ψ 7→ ψ(f ),
of T -modules and thus K-vector spaces. Let Diff(f )i be the subspace of Diff(f ) of partials of
degree at most i. The image of (0 : mi ) under the map τ is precisely Diff(f )i−1 , so the Löewy
filtration
(0 : m) ⊂ (0 : m2 ) ⊂ (0 : m3 ) ⊂ · · · ⊂ (0 : md ) ⊂ (0 : md+1 ) = Tf
of Tf is mapped to the degree filtration
K = Diff(f )0 ⊂ Diff(f )1 ⊂ Diff(f )2 ⊂ · · · ⊂ Diff(f )d = Diff(f )
of Diff(f ). Now
(0 : mi )/(0 : mi−1 ) ∼
= (mi−1 /mi )∨ ,
so the integral function
Hf (0) = 1,
Hf (i) = dimK Diff(f )i − dimK Diff(f )i−1 ,
coincides with the Hilbert function of
i = 1, . . . , d,
Tf∗ :
Hf (i) = H(Tf∗ )(i).
On the other hand, the m-adic filtration
Tf ⊃ m ⊃ m2 ⊃ · · · ⊃ md ⊃ md+1 = 0
corresponds to an order filtration on Diff(f ). We call the order of ψ ∈ T the smallest degree
of a non-zero homogeneous term of ψ, and denote it by ord(ψ). We call the order of a partial
f ′ of f the largest order of a ψ ∈ T such that f ′ = ψ(f ). Thus the image
τ (mi ) ⊆ Diff(f )d−i
is simply the space of partials of order at least i of f .
∨
∼
The isomorphism Q∨
a,i = Qa,d−a−i allows us to interpret the vector space Qa,i as parameterizing partials of f of degree i and order d − a − i, modulo partials of lower degree and larger
order.
More precisely, let Diff(f )ai ⊂ Diff(f ) be the subspace of partials of degree at most i and
order at least d − i − a, then
Q∨ ∼
= Diff(f )a / Diff(f )a + Diff(f )a−1 .
a,i
i
i−1
i
So
a−1
a
a
) .
∆f,a (i) = dimK Q∨
a,i = dimK Diff(f )i /(Diff(f )i−1 + Diff(f )i
P
Notation 7. We denote by ∆f the symmetric decomposition Hf = a ∆f,a of the Hilbert
function Hf .
(1)
6
A. BERNARDI, J. JELISIEJEW, P. MACIAS MARQUES, K. RANESTAD
Consider the space of linear forms that are partials of f ,
Lin(f ) := Diff(f )1 ∩ S1 ,
and the linear subspaces
Lin(f )a = {l ∈ S1 | l is a partial of f of order at least d − a − 1} = Diff(f )a1 ∩ S1 .
We easily see that for each a ≥ 0, we have an isomorphism Lin(f )a ≃ Diff(f )a1 /Diff(f )0 , so
∆f,a (1) = dimK Lin(f )a − dimK Lin(f )a−1 . We obtain a canonical flag of subspaces of S1 :
Lin(f )0 ⊆ Lin(f )1 ⊆ · · · ⊆ Lin(f )d−2 = Lin(f ) ⊆ S1 .
Example 8. Let f = x13 x2 + x33 + x42 . Its space of partials is generated by the elements in
the following table, where the generators of each Q∨
a are arranged by degree; next to it, we
have the symmetric decomposition of its Hilbert function:
Generators of the space of partials
degree 0
1
∨
Q0
1 x1 , x2
∨
Q1
x3
Q∨
x4
2
2
x1 x2
x32
x12 ,
x13 ,
3
x12 x2
Hilbert function decomposition
4
f
degree
∆f,0
∆f,1
∆f,2
Hf
=
=
=
=
0
1
0
0
1
1
2
1
1
4
2
2
1
0
3
3
2
0
0
2
4
1
0
0
1
For instance x32 is a partial of order 1, since it is obtained as y3 (f ) = x32 and cannot be attained
0
by a higher order element of T , so it is a generator of Q∨
1,2 . Here we have Lin(f ) = hx1 , x2 i,
1
2
Lin(f ) = hx1 , x2 , x3 i, and Lin(f ) = hx1 , x2 , x3 , x4 i.
In the next section we shall enumerate polynomials f with a given Hilbert function Hf
using this symmetric decomposition ∆f . For this purpose we denote a Hilbert function H by
its values
H = H(0), H(1), . . . , H(d)
and the decomposition ∆f ,
X
H=
∆a ,
a
by its summands
∆ = (∆0 , . . . , ∆d−2 ),
where each ∆a is symmetric around (d − a)/2, i.e. ∆a (i) = ∆a (d − a − i).
By Corollary 6, both H and each partial sum
α
X
∆≤α =
∆a
a=0
are Hilbert functions of K-algebras generated in degree 1, so there are some immediate restrictions on these functions. First, Hilbert functions H and ∆≤α have positive values and satisfy
the Macaulay growth condition (cf. [Macaulay 1927]): If the i-binomial expansion of ∆≤α (i)
is
mj
mi−1
mi
; mi > mi−1 > · · · > mj ≥ j ≥ 1,
+ ··· +
+
∆≤α (i) =
j
i−1
i
then
mi + 1
mi−1 + 1
mj + 1
(2)
∆≤α (i + 1) ≤
+
+ ···+
.
i+1
i
j+1
7
Example 9. For H(1) = 8, H(2) ≥ 5 and 1 + H(1)P+ · · · + H(d − 1) + 1 = 17 the possible
Hilbert functions H and their decompositions H = i ∆i that satisfy the Macaulay growth
conditions are the following:
H
∆0
∆1
=
=
=
1 8
1 7
0 1
H
∆0
∆1
∆2
7 1
7 1 ,
0
H
∆0
∆1
∆2
=
=
=
=
1
1
0
0
8
2
2
4
=
=
=
=
5
3
2
0
1
1
0
0
8
1
5
2
6 1
1 1
5 0
0
2 1
2 1
,
0
H
∆0
∆1
∆2
1
1
=
=
=
=
,
1
1
0
0
H
∆0
∆1
∆2
∆3
=
=
=
=
=
8
2
3
3
5 2
2 2
3 0
0
1
1
0
0
0
8
1
0
4
3
1
1
5
1
0
4
0
1
1
0
0
1 1
1 1
0
.
2. Standard forms and exotic forms
At this point of the analysis we would like to find a precise description of all the polynomials
having the same symmetric Hilbert function decomposition. To this purpose we have firstly to
deal with the fact that the number of variables involved in each homogeneous summand of a
given polynomial f may be larger than what is expected from the Hilbert function. This will
be explained by the appearance of what we will call exotic summands of f . We will analyze
their role in a description of all polynomials that have a given Hilbert function.
Let us start with some examples clarifying the kind of phenomena that we have to treat.
2.1. Standard and “ Exotic ” examples. Let A be a local Artinian Gorenstein algebra. As
explained before it can be represented as a quotient A ≃ T /I where I = f ⊥ for a polynomial
f ∈ S. If the ideal I is fixed, then f is unique up to action by a unit of T , but clearly the
choice of I such that A ≃ T /I is not unique. In this section we wish to shed some light on
how this choice can be made.
Example 10. Consider the ring A = K[ε]/(ε4 ) and polynomials g = x13 and h = x13 + x1 x2
in Kdp [x1 , x2 ]. Then
K[y1 , y2 ]
K[y1 , y2 ]
,
=
A≃
(y14 , y2 )
g⊥
but also
K[y1 , y2 ]
K[y1 , y2 ]
.
=
A≃ 4
2
(y1 , y2 − y1 )
h⊥
Note that both x1 and x2 occur in h, but while x1 is a partial of this polynomial, x2 is not,
since its space of partials is Diff(h) = hh, x12 + x2 , x1 , 1i.
Now consider the ring B = K[y1 , y2 ]/(y14 − y22 , y1 y2 ) and polynomials p = x14 + x22 and
q = x14 + x12 x2 in Kdp [x1 , x2 ]. Then
B=
and
K[y1 , y2 ]
,
p⊥
K[y1 , y2 ]
K[y1 , y2 ]
.
=
(y13 − y1 y2 , y22 )
q⊥
In this case, x2 is an order-one partial of both p and q, since y2 (p) = x2 and (−y2 + y12 )(q) = x2 .
However, x2 occurs in q in degree 3, which may be surprising for a linear form that is a partial
of order one.
B≃
8
A. BERNARDI, J. JELISIEJEW, P. MACIAS MARQUES, K. RANESTAD
As we will see in the remainder of this section, the most common behaviour is the one we
can observe in Example 8: both x1 and x2 are partials of order 3 and they occur in f4 ; x3 and
x4 are partials of orders 2 and 1, and they occur in f3 and f2 , respectively.
2.2. Description of the standard and the “ exotic ” phenomena. Referring the notation
of Example 10, we want to distinguish polynomials like g and p that have a “standard” behavior
from the ones like h and q where either one finds a variable that does not occur in the partials
or a partial whose order does not “match” with the degree of the corresponding variable.
To this end, in Definition 11 we will define standard forms of polynomials. Intuitively, they
correspond to minimal embeddings of algebras, in terms of variables appearing in the related
polynomials f . Let f ∈ S and A = T /f ⊥ . Moreover let
(3)
f = fd + fd−1 + · · · + f0
be the decomposition in homogeneous summands.
In Section 1 we defined the Hilbert function of A and its symmetric decomposition ∆. In
particular, we saw that ∆a (1) = dimK (Lin(f )a / Lin(f )a−1 ), which is space of linear partials
of f of order exactly d − a − 1. Let
ni =
i
X
∆j (1) = dimK Lin(f )i ,
j=0
i.e. the dimension of the space of linear partials of f of order at least d − i − 1. By degree
reasons this space is contained in the space of linear partials of fd−i + fd−i−1 + · · · + fd . But,
as we have seen in Example 10, a linear form may occur in fd−i and be a partial of f of order
less than d − i − 1 or not be a partial of f at all.
First of all let us fix here a basis of linear forms x1 , . . . , xn in S1 that agrees with the
filtration by Lin(f )i :
(4)
Lin(f )0 = hx1 , . . . , xn0 i ⊆ Lin(f )1 = hx1 , . . . , xn1 i ⊆ · · ·
· · · ⊆ Lin(f )d−2 = hx1 , . . . , xnd−2 i ⊆ S1 = hx1 , . . . , xn i.
Definition 11. Let f ∈ S be a polynomial with homogeneous decomposition f = fd + · · · + f0 .
Let ∆ be the symmetric decomposition of the Hilbert function of T /f ⊥ . We say that f ∈ S is
in standard form if
fd−i ∈ Kdp Lin(f )i = Kdp [x1 , . . . , xni ] , for all i,
where x1 , . . . , xn is any choice of basis for S1 as in (4).
We define the linear space of standard forms.
StandardForms = f ∈ P≤d | ∀i fd−i ∈ Kdp Lin(f )i = f ∈ P≤d | ∀i fi ∈ Kdp Lin(f )d−i .
A first important property of standard forms is the following:
Proposition 12. The leading summand of a partial of f of degree d − i and order j lies in
Kdp [Lin(f )i−j ].
Proof. Let g be the leading summand of a partial of f of degree d − i and order j, then any
partial x of degree one of g is a partial of order at least d − i + j − 1 of f and therefore lies in
Lin(f )i−j . Therefore g ∈ Kdp [Lin(f )i−j ].
There may be variables appearing in f that do not show up in the leading summands
of partials of f . It is tempting to call them exotic variables, but we reserve exotic for the
non-standard part of f .
9
Definition 13. Let f = fd + · · · + f0 ∈ S be the homogeneous decomposition of f , and choose
a basis for S1 as in (4). The exotic summand of degree d − i of f is the form
fd−i,∞ ∈ hxni +1 , . . . , xn iKdp [x1 , . . . , xn ]
such that the degree d − i homogeneous summand of f can be written as
fd−i = fd−i,i + fd−i,∞ ,
with fd−i,i ∈ Kdp [x1 , . . . , xni ].
Thus f is in standard form if and only if all its exotic summands are zero.
Example 14. Let us see how the above definitions work in the cases of Example 10. We have
HA = (1, 1, 1, 1) and ∆0 = (1, 1, 1, 1), ∆i = 0 for i > 0. Therefore ni = 1 for all i.
Now we have g = x13 ∈ Kdp [x1 ] so that x13 is in standard form. On the other hand,
x1 x2 ∈
/ Kdp [x1 ] so that h = x13 + x1 x2 is not in standard form. In fact x1 x2 is an exotic
summand for h.
For the ring B, we get HB = (1, 2, 1, 1, 1), with symmetric decomposition ∆0 = (1, 1, 1, 1, 1),
∆1 = (0, 0, 0, 0, 0), ∆2 = (0, 1, 0, 0, 0), ∆i = 0 for i > 2. Therefore n0 = n1 = 1, and ni = 2 for
all i ≥ 2. As above we can check that p is in standard form, but q is not, having x12 x2 as an
exotic summand.
2.3. Existence of standard forms and their presentation. Let A = T /f ⊥ be an Artinian
Gorenstein algebra. One could ask:
(1) if there exists a presentation A ≃ T /g ⊥ with g in standard form.
(2) in case such a presentation exists, whether there are any relations between g and f .
Fortunately, these questions have quite satisfactory answers, as we explain below. We need
some notation.
Notation 15. Let T̂ denote the power series ring obtained by completing T at the ideal of
the origin. In coordinates, T̂ := K[[y1 , . . . , yn ]]. We may interpret S as subset of functionals
on T̂ via the pairing defined by
(
1 if α = β
α
β
y ,x =
0 otherwise.
Note that
(5)
hσ, f i = h1, σ(f )i ,
as seen by decomposing σ and f into monomials. In particular if σ(f ) = 0, then hσ, f i = 0.
Let ϕ : T̂ → T̂ be an automorphism of T̂ . It induces a dual map ϕ∨ : S → S defined by the
condition
(6)
hϕ(σ), f i = hσ, ϕ∨ (f )i ,
for all σ ∈ T, f ∈ P.
Let I be a finite colength ideal of T supported at the origin. Then T /I = T̂ /I. Clearly, the
⊥
quotients T̂ /I and T̂ /ϕ(I) are isomorphic. Moreover if I = f ⊥ , then ϕ(I) = (ϕ−1 )∨ (f ) .
The fundamental result is that for every f we may find ϕ so that ϕ∨ (f ) is in standard form.
It fact we prove that ϕ may be chosen “ with no linear part ”.
We now make this precise.
10
A. BERNARDI, J. JELISIEJEW, P. MACIAS MARQUES, K. RANESTAD
Definition 16. Let M be the unique maximal ideal of T̂ . Let ϕ : T̂ → T̂ be an automorphism.
We say that ϕ is of order (at least) two if
ϕ(yi ) = yi
mod (y1 , . . . , yn )2
for all i.
Remark 17. Every automorphism of T̂ induces a linear action on M/M 2 . The order two
automorphisms are precisely those which act trivially. Thus they form a normal subgroup. In
particular an inverse of an order two automorphism is also of order two.
Theorem 18 (Existence of standard forms). Let f ∈ S be a polynomial with symmetric Hilbert
function decomposition ∆. Then there is an automorphism ϕ : T̂ → T̂ such that ϕ∨ (f ) is in
standard form. Consequently, f = ψ ∨ (g) for an element g ∈ StandardForms and an automorphism ψ = ϕ−1 .
Moreover one can choose ϕ and ψ of order two.
Proof. For the existence of ϕ see [Iarrobino 1994, Theorem 5.3AB]. Take one such ϕ. We will
compose ϕ with a linear
P map to obtain the required order two automorphism.
We have ϕ(yi ) = λij yj + si , where si ∈ (y1 , . . . , yn )2 . Let τ be the linear automorphism
P
of T̂ defined by τ (yi ) = λij yj . Then (ϕ ◦ τ −1 )(yi ) ≡ yi mod M 2 so ϕ ◦ τ −1 is an automorphism of order two. Let g = ϕ∨ (f ). By definition g is in standard form. Since τ is a linear
automorphism of T̂ , the map (τ −1 )∨ is simply a linear transformation of S, i.e. a change of
variables. The definition of being in standard form is coordinate free, so that (τ −1 )∨ preserves
being in standard form. In particular h = (τ −1 )∨ (g) is in standard form.
But h = (ϕ ◦ τ −1 )∨ (f ), so that ϕ ◦ τ −1 is a required automorphism of order two.
It is also important and interesting to see an explicit description of the action ϕ∨ : S → S
α
β
of an automorphism ϕ : T̂ Q
→ T̂ . For
this, recall that S is a divided power ring: x · x =
α+β
α+β
αi +βi
α+β
, where α =
.
α x
αi
Proposition 19. Let ϕ : T̂ → T̂ be an automorphism. Let Di = ϕ(yi ) − yi ∈ T̂ . For a
multi-index α denote Dα = D1α1 · · · Dnαn . Let f ∈ S. Then
∨
ϕ (f ) =
X
α∈Nn
0
n
X
xi · Di (f ) + · · · .
x · D (f ) = f +
α
α
i=1
Proof. See [Jelisiejew 2015, Proposition 1.8].
Example 20. Let us illustrate Theorem 18 in the setup of Example 14. We should find an
f in standard form and an automorphism ϕ of K[y1 , y2 ] such that ϕ∨ (f ) = h = x13 + x1 x2 .
We see that the linear partials of h are spanned by x1 , so if we wish ϕ to
Pbe of order at least
two, we must have f ∈ Kdp [x1 ]. According to Proposition 19, ϕ∨ (f ) = α∈Nn xα · Dα (f ) ,
0
for some elements of order at least two D1 , D2 ∈ K[[y1 , y2 ]]. Since f must have degree three,
(Di Dj )(f ) = 0, for any i and j, so
x13 + x1 x2 = f + x1 D1 (f ) + x2 D2 (f ).
This implies that D1 (f ) = 0 and D2 (f ) = x1 , so we must have D1 = 0 mod (y1 , y2 )4 and
D2 = y12 mod (y1 , y2 )4 . Therefore we can choose ϕ : T̂ → T̂ to be the automorphism defined
by ϕ(y1 ) = y1 and ϕ(y2 ) = y2 + y12 .
11
2.4. Description of exotic summands. For parameterization purposes and dimension counts,
it is interesting to consider families of polynomials yielding isomorphic algebras, or at least sharing the same Hilbert and symmetric decomposition. Given a polynomial f ∈ Kdp [x1 , . . . , xk ]
such that Lin(f ) = hx1 , . . . , xk i, we consider the family
Ffm := g ∈ Kdp [x1 , . . . , xk+m ] | g − f ∈ (xk+1 , . . . , xk+m ),
(7)
Hg = Hf , Lin(g) = Lin(f ) = hx1 , . . . , xk i .
The next result gives a characterisation of the elements in this family. We will use the
notation ls(f ) for the leading summand of a polynomial f , i.e. if f = fd + · · · + f0 is its
decomposition into homogeneous summands, ls(f ) = fd .
Proposition 21. Let f ∈ Kdp [x1 , . . . , xk ] and assume
Lin(f ) = hx1 , . . . , xk i.
Let g ∈ Kdp [x1 , . . . , xk+m ] be any polynomial. Then g ∈ Ffm if and only if there are elements
φ1 , . . . , φm ∈ K[y1 , . . . , yk ] of order at least two such that
X
i1
im
(8)
g=
xk+1
· · · xk+m
· φ1i1 · · · φmim (f ).
i1 ,...,im ≥0
In particular, for each g ∈ Ffm , the algebras K[y1 , . . . , yk+m ]/g ⊥ and K[y1 , . . . , yk ]/f ⊥ are
isomorphic.
Proof. Let l be the dimension of Diff(f ) and choose a basis h1 , . . . , hl for this vector space such
that ls(h1 ), . . . , ls(hl ) are linearly independent. Choose elements ψ1 , . . . , ψl ∈ K[y1 , . . . , yk ]
such that for each i, ψi (f ) = hi . Let g ∈ Ffm and write
g = f + xk+1 g1 + · · · + xk+m gm
in such a way that for each j, we have gj ∈ Kdp [x1 , . . . , xk+j ]. Then for each i,
ψi (g) = ψi (f ) + xk+1 ψi (g1 ) + · · · + xk+m ψi (gm ).
Now Proposition 12 tells us that ls ψi (g) ∈ Kdp [x1 , . . . , xk ] and since ψi (f ) cannot be cancelled by the terms in xk+1 ψi (g1 ) + · · · + xk+m ψi (gm ), we must have
ls ψi (g) = ls ψi (f ) = ls(hi ).
But this implies that ψ1 (g), . . . , ψl (g) form a linearly independent set, and since the dimension
of Diff(g) is also l (since g and f yield algebras with the same Hilbert function), we get
Diff(g) = hψ1 (g), . . . , ψl (g)i.
In addition, we know that the variables xk+1 , . . . , xk+m cannot occur in the leading summand
of g, also by Proposition 12, so the polynomials g1 , . . . , gm have degree at most d − 2. Now
yk+m (g) = gm , so gm is a partial of g, which means that there is some φm ∈ hψ1 , . . . , ψl i such
d+1
that gm = φm (g). Moreover φm has order at least one, because deg gm ≤ d − 2, so φm
(g) = 0.
Denote ĝj = xk+1 g1 + · · · + xk+j gj , and observe that yk+s (ĝj ) = gs , if s ≤ j, and yk+s (ĝj ) = 0,
otherwise. So,
g = f + ĝm−1 + xk+m · φm (g)
= f + ĝm−1 + xk+m · φm f + ĝm−1 + xk+m · φm (g)
2
2
= f + ĝm−1 + xk+m · φm (f + ĝm−1 ) + xk+m
· φm
(g)
12
A. BERNARDI, J. JELISIEJEW, P. MACIAS MARQUES, K. RANESTAD
and iterating this further we get
X
i
i
g=
xk+m
· φm
(f + ĝm−1 ).
i≥0
Applying yk+m−1 to both sides of this equality, we get
X
X i
i
i
i
yk+m−1 (f + ĝm−1 ) =
yk+m−1 (g) =
xk+m · φm
(gm−1 ).
xk+m
· φm
i≥0
i≥0
i
i
Again there must be some φm−1 ∈ hψ1 , . . . , ψl i such that i≥0 xk+m
· φm
(gm−1 ) = φm−1 (g).
So,
X
i
i
g=
xk+m
· φm
(f + ĝm−2 + xk+m−1 gm−1 )
P
i≥0
=
X
i
i
xk+m
· φm
(f + ĝm−2 ) + xk+m−1
i≥0
=
X
X
i
i
xk+m
· φm
(gm−1 )
i≥0
i
xk+m
·
i
φm
(f
+ ĝm−2 ) + xk+m−1 · φm−1 (g).
i≥0
Iterating we get
g=
X
i
i
im
m−1
m−1
im
xk+m
xk+m−1
· φm−1
(f + ĝm−2 ).
φm
im−1 ,im ≥0
Applying the remaining operators yk+m−2 , . . . , yk+1 the same way, we obtain (8). It remains to
show that φ1 , . . . , φm have order at least two. Without loss of generality, suppose that φ1 has
order one, and write φ1 = φ′ + φ′′ , where φ′ ∈ hy1 , . . . , yk i and ord φ′′ ≥ 2. Let t ∈ hx1 , . . . , xk i
be such that φ′ (t) = 1. Since t is a partial of f , there exists η ∈ K[y1 , . . . , yk ] such that η(f ) = t.
Note that (φu η)(f ) = φu (t) is a constant, for all u > 1. So
X
X
i1
im
η(g) =
xk+1
· · · xk+m
· φ1i1 · · · φmim η(f ) = t + xk+1 +
xk+u φu (t),
i1 ,...,im ≥0
u>1
and therefore η(g) is a partial of g of degree one that does not belong to h1, x1 , . . . , xk i, a
contradiction.
For the converse, suppose that g admits a presentation as in (8). Then clearly we have
g − f ∈ (xk+1 , . . . , xk+m ). Let ϕ be the automorphism of K[y1 , . . . , yk+m ] defined by ϕ(yi ) = yi
for 1 ≤ i ≤ k and ϕ(yk+i ) = yk+i + φi , for 1 ≤ i ≤ m. Then by Proposition 19, we see that
g = ϕ∨ (f ). So ϕ induces an isomorphism between K[y1 , . . . , yk+m ]/g ⊥ and K[y1 , . . . , yk ]/f ⊥ ,
which proves the last statement, and shows that Hg = Hf . Finally, if t ∈ Lin(f ) and we take
η ∈ K[y1 , . . . , yk ] such that η(f ) = t, we may apply η to both sides of (8) to get η(g) = t. So
Lin(g) ⊆ Lin(f ) and since they have the same dimension, equality must hold, and g ∈ Ffm .
Remark 22. We can get an alternative proof of Proposition 21 if we take g ∈ Ffm and use
Proposition 12 to see that since the linear partials of g lie in hx1 , . . . , xk i, the leading terms of
partials of g only involve these variables, like we did in the proof. Now if we denote (g ⊥ )∗ the
ideal generated by the initial forms of elements of g ⊥ , we know that (g ⊥ )∗ is the annihilator
of the set of leading summands of all partials of g (see a discussion on this at the beginning
of Section 2 in [Casnati, Notari 2013], but also Proposition 3 in [Emsalem 1978]). Therefore
for each j > k, yj ∈ (g ⊥ )∗ , which means that there is an element Dj ∈ T of order at least two
such that yj − Dj ∈ g ⊥ . If a variable yj1 occurs in a Dj2 , with j1 , j2 > k then we may replace
yj1 by Dj1 in Dj2 . Each time we do this, the minimal degree of yj1 in Dj2 grows. When
this degree exceeds deg g, we may discard the remaining part, so the process eventually ends
13
and we may assume that each Dj ∈ k[y1 , . . . , yk ]. We can then consider the automorphism
ϕ : T̂ → T̂ which sends yj to itself for j ≤ k and yj to yj − Dj for j > k. We can now show
that ϕ∨ (g) = f by showing that for any element φ ∈ T , φ, ϕ∨ (g) = hφ, f i.
Note that in the definition of the family Ffm and the hypotheses of Proposition 21, the
polynomial f need not be in standard form. This result gives us a way of adding exotic
summands to a polynomial without changing the Hilbert function of the algebra it yields. It
also gives us the following result:
Corollary 23. Let f ∈ Kdp [x1 , . . . , xk ] ⊂ Kdp [x1 , . . . , xk+m ] be a polynomial of degree d and
assume Lin(f ) = hx1 , . . . , xk i. Then the family Ffm ⊂ Kdp [x1 , . . . , xk+m ]≤d has dimension
X
dim Ffm = m ·
∆f,a (i)
a≥0
i≤d−a−2
The next result shows that all exotic summands may be obtained in a similar description.
Proposition 24. Let f ∈ S be a polynomial of degree d and choose a basis x1 , . . . , xn for
S1 that agrees with the filtration in (4). Then we can write f = fst + fex such that fst is in
standard form and
X
(9)
fex =
xα · Dα (fst ) .
α>0
α1 =···=αn1 =0
where Dn1 +1 , . . . , Dn ∈ T have order at least two and
(1) d − a ≤ deg Dk (fst ) ≤ d − 2, if Dk (fst ) 6= 0 and na−1 < k ≤ na , for any a ≥ 2;
(2) deg Dk (fst ) ≤ d − 2, if k > nd−2 .
Proof. Applying Theorem 18 and Proposition 19 to f , we know that there exists a polynomial
g in standard form such that
X
(10)
f = ϕ∨ (g) =
xα · Dα (g),
α∈Nn
0
with D1 , . . . , Dn ∈ T of order at least two.
Claim. The basis x1 , . . . , xn also satisfies the filtration in (4) for the polynomial g. To see
this, let t ∈ hx1 , . . . , xn i be a linear partial of g of order j, and η ∈ T an element of order j
such that η(g) = t. If h , i : T × S is the usual pairing, using the rule in (6), we have
φ, ϕ−1 (η) (f ) = ϕ−1 (η) · φ, f = hϕ(φ), ti,
for any φ ∈ T . Since ϕ is of order at least two, this implies ϕ−1 (η) (f ) = t. We also know
that ord ϕ−1 (η) = ord η, so t is also a linear partial of f of order j. So f and g have the same
linear partials and those partials have the same order, which proves the claim.
Now we wish to show that g can be replaced by a polynomial g ′ , also in standard form, so
that we do not need the first operators D1 , . . . , Dn1 . We consider expression (10) , and we start
by factoring the power of any variable, say x1α1 , and the power of the corresponding operator,
D1α1 . Note that D1α1 obviously commutes with the remaining operators D2α2 · · · Dnαn . In
order to move also the term x1α1 to the right side of D2α2 · · · Dnαn , we can apply the rule
xi φ(h) = φ(xi h) − φ(i) (h), for any φ ∈ T and h ∈ S,
14
A. BERNARDI, J. JELISIEJEW, P. MACIAS MARQUES, K. RANESTAD
∂
φ. We obtain
where φ(i) = ∂y
i
X α
(x2 2 · · · xnαn )x1α1 · (D2α2 · · · Dnαn )D1α1 (g) ,
f=
α∈Nn
0
=
X
X
(x2α2 · · · xnαn ) ·
i≥0 α2 ,...,αn
=
X
α2 ,...,αn
(1)
D2 − D2 D1
α2
· · · Dn − Dn(1) D1
(x2α2 · · · xnαn ) · (D2′ ) α2 · · · (Dn′ ) αn (g ′ ),
αn
x1i · D1i (g) ,
P
(1)
where g ′ = i≥0 x1α1 · D1α1 (g) and Di′ = Di − Di D1 . The equality between the first and the
second lines can be checked by hand in a straightforward, even if cumbersome, computation.
(1)
Observe that in this way we get an extra piece Di , but this does not change the fact that the
′
′
new elements D1 , . . . , Dn have order at least two. Repeating this procedure, we can rewrite
X
X
xα · (D′′ )α (g) ,
xα · D′′α (g ′′ ) ,
where
g ′′ =
f=
α∈Nn
0
α1 =···=αn1 =0
n
α∈N0 1
with some modified D1′′ , . . . , Dn′′ ∈ T of order at least two. Since any terms of f and any
terms of g involving only the variables x1 , . . . , xn1 are not exotic (because the choice of basis
x1 , . . . , xn agrees with the filtration in (4) for both f and g) we have that g ′′ is in standard
form.
Now we know that for any k > n1 , Dk′′ (g ′′ ) is a partial of degree at most d − 2, otherwise xk
would occur in the leading summand of f , and all such linear partials belong to hx1 , . . . , xn0 i.
Finally, suppose that for some k we have na−1 < k ≤ na but deg Dk′′ (g ′′ ) < d − a. Then the
term xk Dk′′ (g ′′ ) has degree at most d − a and belongs to Kdp [x1 , . . . , xna ], so it is not part of an
2
d−1 ′′
(g ) have lower degree so are also not
exotic summand. The terms xk2 Dk′′ (g ′′ ), . . . , xkd−1 Dk′′
part of exotic summands. Therefore we can perform another modification as above, with the
variable xk and the corresponding operator Dk′′ . Doing this for every such k, we may replace
g ′′ by some g ′′′ also in standard form, and f will be written as in (9), with fst = g ′′′ .
Example 25. Let f be a polynomial of degree five such that Hf = (1, 1, 1, 1, 1, 1); this is the
minimal possible Hilbert function. Then Hf is symmetric, so that the only possible symmetric
decomposition is ∆f,0 = Hf = (1, 1, 1, 1, 1, 1) and all other ∆f,i equal to zero vectors.
By Theorem 18 we see that f = ϕ∨ (g), where g is in standard form and ϕ is of order at
least two. We have ∆g,i = ∆f,i for all i, so dimK Lin(g)i = 1 for all i. Choose a generator x
for this space. Let g = g5 + · · · + g0 be the decomposition into homogeneous summands. From
the definition of standard form we see that gi ∈ K[x] for all i, i.e. g ∈ K[x]. So we may write
g = a5 x5 + a4 x4 + · · · + a0 for constants ai . Since g has degree five, we have a5 6= 0 and by
changing x we
Pmay assume that a5 = 1.
Now f = α∈Nn xα · Dα (g) , where Di = ϕ(yi ) − yi ∈ M 2 . Then Di Dj Dk (g) = 0 for all
0
P
P
i, j, k and the sum becomes shorter: f = g + i xi · Di (g) + i,j xi xj · Di Dj (g). We see that
deg Di (g) ≤ 3 so that
f = x5 +
n
X
i=1
λi xi · x3 +
n
X
i=1
µi xi x2 +
X
λi λj xi xj x + Q,
i,j
where λi , µi are constants and Q is a polynomial of degree at most two, partially depending on
λi and µi . What is the dimension of possible f obtained this way? Each Di may be chosen as
an element of the square of the maximal ideal of T /(x5 )⊥ , therefore we have (5−2)-dimensional
15
choice. Together with the choice of x, we obtain at most a 4(dim Spec(T )) = 4n-dimensional
family.
3. Apolarity and local cactus rank
We shall now apply our analysis of Hilbert functions of polynomials to apolar subschemes
of a homogeneous form. Recall from 0.1 that we denote by S̄ = Kdp [x0 , . . . , xn ] and T̄ =
K[y0 , . . . , yn ].
Definition 26. A subscheme Z ⊂ P(S̄1 ) is apolar to a form F ∈ S̄ if its homogeneous ideal
IZ ⊂ T̄ is contained in F ⊥ .
Apolarity for a subscheme Z ⊂ P(S̄1 ) to a form F of degree d may be given the following
natural interpretation in terms of the d-uple embedding of Z, the image νd (Z) ⊂ P(S̄d ) where
νd : P(S̄1 ) → P(S̄d ), [l] 7→ [ld ].
Lemma 27. (Apolarity Lemma). A scheme Z ⊂ P(S̄1 ) is apolar to F ∈ S̄d if and only if
[F ] ∈ hνd (Z)i ⊂ P(S̄d ).
Proof. If Z is apolar to F , then (IZ )d ⊆ (F ⊥ )d and we get [F ] ∈ V (IZ )d = hνd (Z)i ⊂ P(S̄d ),
so the “ only if ” part follows. For the “ if ” part, (IZ )e ⊂ (F ⊥ )e = T̄e when e > d, so it remains
to consider Ψ ∈ (IZ )e , for some e ≤ d. In this case, T̄d−e Ψ ⊂ (IZ )d . So if [F ] ∈ hνd (Z)i, then
T̄d−e Ψ ⊂ (F ⊥ )d . But T̄d−e Ψ(F ) = 0 only if Ψ(F ) = 0, so the “ if” part follows also.
We are particularly interested in minimal apolar zero-dimensional subschemes to a form,
their length is called the cactus rank of the form. The closure of the set of forms with a given
cactus rank is called a cactus variety of forms, although it may be reducible. Minimal apolar
zero-dimensional schemes are locally Gorenstein, so our first aim is to describe the forms that
have a given minimal length local Gorenstein scheme. For any form F a minimal apolar zero-dimensional scheme decomposes into local Artinian Gorenstein schemes. This decomposition
corresponds to an additive decomposition of the form F . In particular, the cactus variety is
the join of varieties of forms whose minimal apolar scheme is local.
Let F be a form of degree d in S̄ and [l] ∈ P(S̄1 ). Consider the family Z ⊂ Hilb P(S̄1 ) of
subschemes apolar to F and supported at [l]. We construct a particular subscheme ZF,l ∈ Z
that we will call the natural apolar subscheme of F at [l]. The element [l] ∈ P(S̄1 ) defines
a hyperplane V (l) ⊂ P(T̄1 ). The complement U of this hyperplane is isomorphic to an affine
space Spec(S̄l ). Moreover, we get a homomorphism π = π[l] : S̄ → S̄l corresponding to passing
from the homogeneous coordinate ring S of P(T̄1 ) to the coordinate ring of U .
Choose dual bases hl, l1 , . . . , ln i and hl′ , l1′ , . . . , ln′ i for S̄1 and T̄1 , respectively. Let S̄l =
K[l1 , . . . , ln ] and let T̄l = K[l1′ , . . . , ln′ ]. Then (S̄l )1 and (T̄l )1 are natural dual spaces like S̄1
and T̄1 above, and T̄l is the coordinate ring of the affine space U ′ that contains the point
[l] and is the complement of the hyperplane V (l′ ) ⊂ P(S̄1 ). Given any polynomial g ∈ S̄l we
denote by Zg the subscheme V (g ⊥ ) ⊆ U ′ ⊂ P(S̄1 ).
Definition 28. Let F ∈ S be any form and l ∈ S̄1 a linear form. Take f := π[l] (F ) and
f ⊥ ⊆ T̄l . We define ZF,l to be the subscheme Zf = V (f ⊥ ) ⊆ U ′ ⊂ P(S̄1 ).
Since ZF,l is finite, it is a closed subset of P(S̄1 ). By construction, the support of ZF,l is
[l] ∈ P(S̄1 ).
In the definition of ZF,l we have not used any particular coordinate system. However,
to simplify the following proofs we fix coordinates. First, note that for every lifting l ∈ S̄1
of [l] we have a canonical isomorphism S̄l ≃ S̄/(l − 1). Changing x0 , . . . , xn if necessary we
16
A. BERNARDI, J. JELISIEJEW, P. MACIAS MARQUES, K. RANESTAD
may assume that l = x0 . Then S̄l may be identified with S = Kdp [x1 , . . . , xn ] and T̄l with
T = K[y1 , . . . , yn ]. The homomorphism
π := πx0 : S̄ → S
sends x0 to 1 and xi to itself for i > 0. Note that π induces an isomorphism between the space
S̄d of homogeneous polynomials of degree d and the space (S)≤d of polynomials of degree d.
Furthermore we fix a “dual” homomorphism
π ∗ : T̄ → T
sending y0 to 1 and yi to itself for i > 0.
Let G ∈ S̄ be a homogeneous polynomial, let Ψ ∈ T̄ be a homogeneous differential operator,
and denote g = π(G) and ψ = π ∗ (Ψ). In general, π Ψ(G) and ψ(g) are different polynomials,
but the following lemma gives the basic relation between them.
Lemma 29. Let G ∈ S̄ be a homogeneous polynomial, let Ψ ∈ T̄ be a homogeneous operator, with deg G ≥ deg Ψ, and let d = deg G − deg Ψ. Let g = π(G) and ψ = π ∗ (Ψ), then the
degree-d tails of π Ψ(G) and ψ(g) are equal. Moreover, if G is divisible by x0 deg Ψ , then
π Ψ(G) = ψ(g).
Proof. The first statement is equivalent to saying that the images of π Ψ(G) and ψ(g) are
equal in the linear space S/(S)>d . Therefore, both statements are linear with respect to Ψ and
αn
0
G and it is enough to prove them in the case when G and Ψ are monomials. Let G = xα
0 . . . xn
β0
βn
and Ψ = y0 . . . yn . By definition
(
0 −β0 α1 −β1
xα
x1
· · · xnαn −βn if αi ≥ βi for all i ≥ 0,
0
Ψ(G) =
0
otherwise,
(
1 −β1
n −βn
xα
· · · xα
if αi ≥ βi for all i ≥ 1,
n
1
ψ(g) =
0
otherwise.
We consider two cases. First, suppose α0 ≥ β0 . Then
the conditions ∀i ≥ 0, αi − βi ≥ 0
and ∀i ≥ 1, αi − βi ≥ 0 are equivalent. Thus π Ψ(G) = ψ(g) and so their images in S/(S)>d
agree.
Next, suppose that α0 < β0 . Then Ψ(G) = 0. Suppose ψ(g) 6= 0 in S. Then ψ(g) =
P
1 −β1
xα
. . . xnαn −βn is a monomial of degree i≥1 αi − βi = d − (α0 − β0 ) > d, thus its image in
1
S/(S)>d is zero, i.e. equal to the image of Ψ(G). This finishes the proof of the first claim.
For the second claim, note that by assumption α0 ≥ deg Ψ ≥ β0 , thus the proof of the first
case above applies, giving π(Ψ(G)) = ψ(g).
Corollary 30. Let G ∈ S̄ and Ψ ∈ T̄ be homogeneous polynomials.
ψ = π ∗ (Ψ). If ψ(g) = 0, then Ψ(G) = 0.
Let g = π(G) and
Proof. Let b = deg(Ψ) and G′ := xb0 G. Then g = π(G′ ) and G′ , Ψ, g, ψ satisfy assumptions of
Lemma 29, therefore π Ψ(G′ ) = ψ(g) = 0. Since π is an isomorphism between S̄d and (S)≤d
we have Ψ(G′ ) = 0. But G = y0b (G′ ), so that
Ψ(G) = Ψ y0b (G′ ) = y0b Ψ(G′ ) = 0.
Corollary 31. Let F ∈ S be a homogeneous polynomial and l ∈ S̄1 any linear form. Then the
scheme ZF,l (see Definition 28) is apolar to F .
Proof. Take any homogeneous form Ψ ∈ I(ZF,l ). Then ψ = π ∗ (Ψ) is an element annihilating
f = π(F ). By Corollary 30 we have Ψ(F ) = 0, i.e. Ψ ∈ F ⊥ .
17
Remark 32. It is easy to characterize the d-uple embedding νd (ZF,l ) in a manner similar to
the proof of the Apolarity Lemma 27.
For l ∈ S̄1 , let l⊥ ∩ T̄1 be the subspace of linear forms in T̄1 that annihilate l, and let (l⊥ )e
be its e-th symmetric product. Then (l⊥ )e (F ) ⊂ S̄d−e for e ≤ d is a subspace of partials of F
or degree d − e, and the linear span of νd (ZF,l ) is given by
hνd (ZF,l )i = P ld ⊕ ld−1 · (l⊥ )d−1 (F ) ⊕ · · · ⊕ l · (l⊥ )1 (F ) ⊕ F .
Furthermore νd (ZF,l ) = hνd (ZF,l )i ∩ Vd,n ⊂ P(S̄d ).
The following lemma is a private communication from Jaroslaw Buczyński.
Lemma 33 (Buczyński). If Z is any local scheme in P(S̄1 ) apolar to a homogeneous polynomial
F ∈ S and supported at [l] = [1 : 0 : · · · : 0], then there exists a closed subscheme Z ′ ⊆ Z, apolar
to F , such that Z ′ = ZG,l , for some G ∈ S̄. Moreover, F = Ψ(G) for some Ψ ∈ T̄ .
Proof. By [Buczyńska, Buczyński 2014, Proposition 2.2, Lemma 2.3], the scheme Z contains
a closed Gorenstein subscheme Z ′ apolar to F .
Let g ∈ S be a polynomial such that Z ′ = V (g ⊥ ) and let G ∈ S̄ be a homogenization of g
such that G is divisible by xd0 , where d = deg(F ). Then g = π(G) and Z ′ = ZG,l . Lemma 29
asserts that (G⊥ )s = (IZG,l )s for any s ≤ d, and therefore (G⊥ )s = (IZG,l )s ⊆ (F ⊥ )s . Since F
is of degree d, it follows that G⊥ ⊆ F ⊥ , so F is a partial of G.
We may now prove Proposition 4.
Proof (Proposition 4). By Lemma 33, we may assume that Γ = ZG,l and that F = Ψ(G) for
some homogeneous Ψ ∈ K[y0 , . . . , yn ]. Let g = π(G) and ψ = π ∗ (Ψ). By Lemma 29 the
polynomial f is the degree-d tail of ψ(g).
Clearly Diff ψ(g) ⊆ Diff(g), thus Zψ(g) is a closed subscheme of Zg = Γ. By minimality
of
Γ it is enough to prove that Zψ(g) is a scheme apolar to F . Let ψ ′ be such that ψ ′ ψ(g) = 0
and let Ψ′ ∈ T̄ be a homogeneous polynomial such that ψ ′ = π ∗ (Ψ′ ). Then π ∗ (Ψ′ Ψ) = ψ ′ ψ.
By Corollary 30 we have Ψ′ Ψ(G) = 0. Since F = Ψ(G) we get that Ψ′ (F ) = 0, thus Zψ(g)
is apolar to F .
3.1. The local cactus rank of a general cubic surface. In this subsection we restrict to
characteristic 0. First we present an example of a quartic polynomial whose cubic tail has
more partials than the polynomial itself. Similar examples play a role in our computation of
the local cactus rank of a general cubic surface, the main issue in this section.
Example 34. Let f = x12 x2 + x22 and g = x14 + f . Then Diff(f ) has a basis
f, x12 + x2 , x1 x2 , x1 , x2 , 1,
thus dimK Diff(f ) = 6. On the other hand, Diff(g) is spanned by
g, x13 + x1 x2 , x12 + x2 , x1 , 1,
so that dimK Diff(g) = 5. Notice that y2 − y12 ∈ g ⊥ , so that x2 does not appear in the leading
summand of any partial of g (cf. Proposition 12).
For our computation of the local cactus rank of a general cubic surface V (F ), we need
to translate the generality assumptions on F into properties of its partials. First note the
following algebraic–geometric correspondences for Ψ ∈ T̄1 :
(1) Ψ3 (F ) = 0 if and only if [Ψ] ∈ V (F ) ⊆ P(T̄1 ),
(2) Ψ2 (F ) = 0 if and only if [Ψ] is a singular point of the (hyper)surface V (F ),
(3) Ψ(F ) = 0 if and only if [Ψ] is a cone point of V (F ).
18
A. BERNARDI, J. JELISIEJEW, P. MACIAS MARQUES, K. RANESTAD
Lemma 35. Let F ∈ S be a general cubic form in four variables. Then
(1) The set of [Ψ] ∈ P(T̄1 ) such that the quadric Ψ(F ) has rank less than 4 is an irreducible
surface of degree 4. Furthermore, the quadric Ψ(F ) has rank less than 4 if and only if
there exist a Ψ′ ∈ T̄1 such that (ΨΨ′ )(F ) = 0.
(2) There are no points [Ψ] ∈ V (F ) ⊂ P(T̄1 ), such that Ψ(F ) is a quadric of rank less than
3 i.e. if Ψ33 (F ) = 0 and (Ψ2 Ψ3 )(F ) = 0, then Ψ23 (F ) 6= 0.
(3) The cubic surface V (F ) is smooth, i.e. Ψ2 (F ) 6= 0 for every nonzero Ψ ∈ T̄1 .
(4) The cubic surface V (F ) has no Eckhardt points, i.e. no plane section is a cone.
Proof. These facts are classical. For a good recent reference see [Dolgachev 2012, 9.4].
Proposition 36. Let F ∈ S̄3 be a general smooth cubic form in four variables. Then the local
cactus rank of F is 7. For every linear form l ∈ S̄1 , the apolar scheme of the dehomogenization
π[l] (F ) has length 8, while if {l = 0} defines a singular curve section of V (F ) ⊂ P(T̄1 ) whose
tangent cone at the singular point is a square, then there is a length 7 scheme supported at
[l] ∈ P(S̄1 ) that is apolar to F .
Proof. Let F ∈ S̄ = Kdp [x0 , x1 , x2 , x3 ] be a general cubic form in the sense of Lemma 35.
We claim that for all non-zero l ∈ S̄1 , the Hilbert function
(11)
HFl = (1, 3, 3, 1).
Indeed, suppose it is not so, then HFl ≤ (1, 3, 2, 1) for some non-zero l ∈ S̄1 and there exists a non-zero linear form Ψ ∈ l⊥ , such that ψ(Fl ) has degree at most one, where ψ =
π ∗ (Ψ). Let Fl = f3 + f2 + f1 + f0 be the decomposition into homogeneous components, then
F = f3 + lf2 + l2 f1 + l3 f0 . Since deg ψ(Fl ) ≤ 1 we have Ψ(f3 ) = 0. But then Ψ(F
) is divisible
by l, so it is a quadric of rank at most 2. On the other hand Ψ(l) = 0, so Ψ2 Ψ(F ) = Ψ3 (F ) = 0,
so [Ψ] ∈ V (F ) ⊂ P(T̄1 ). This contradicts the generality assumption Lemma 35.2 of F .
The cubic surface V (F ) has a one dimensional family of plane cuspidal cubic sections,
and finitely many reducible plane sections that are unions of a smooth conic and a tangent
line. In either case, the tangent cone at the singular point is a square. We pick one such
plane section. After a linear change of coordinates, we may assume that this plane section is
V (F, x0 ), that it is singular at V (x0 , x1 , x2 ) and that V (x0 , x12 ) is the tangent cone, so that
the dehomogenization π[x0 ] (F ) has the form
π[x0 ] (F ) = f + x12 x3 + x32 + x3 l
where f and l are polynomials in Kdp [x1 , x2 ] of degree three and one respectively. The plane
section V (F, x0 ) = V (f3 + x12 x3 , x0 ) where f3 is the cubic summand of f . By Lemma 35.4,
the linear partials of f3 + x21 x3 fill hx1 , x2 , x3 i. For later use we note that y12 is the only
monomial quadric such that y12 (f3 +x21 x3 ) contains x3 . Hence we have (y2 , y3 ) · (y1 , y2 , y3 )(f ) =
h1, x1 , x2 i. Since l − y12 (f ) ∈ Kdp [x1 , x2 ] we can find σ0 ∈ (y2 , y3 )(y1 , y2 , y3 ) such that σ0 (f ) =
l − y12 (f ) mod hx1 i. Then there also exists σ = σ0 + ay13 ∈ (y1 , y2 , y3 )2 such that σ(x41 + f ) =
l − y12 (f ). Clearly σ0 x41 = 0.
Let G = x14 + x0 F . Then y0 (G) = F , hence G⊥ ⊆ F ⊥ . By Lemma 31 we have I(ZG,x0 ) ⊆
⊥
G , so we conclude that I(ZG,x0 ) ⊆ F ⊥ , i.e. that the local Gorenstein scheme ZG,x0 is apolar
to F . We claim that length(ZG,x0 ) ≤ 7, and hence that the local cactus rank of F is at most 7.
We prove the claim, by showing that x12 x3 + x32 + x3 l is an exotic summand for Gx0 :=
π[x0 ] (G) = x14 + f + x12 x3 + x32 + x3 l. For this we consider the partials
y3 (Gx0 ) = x12 + x3 + l,
y12 (Gx0 ) = x12 + x3 + y12 (f ).
19
We have (y3 − y12 )(Gx0 ) = l − y12 (f ) = σ(Gx0 ), so y3 − y12 − σ annihilates Gx0 . If we take
ψ = y12 + σ, then ψ 2 Gx0 = y14 (Gx0 ) = 1 and we may write
Gx0 = x14 + f + x3 ψ(x14 + f ) + x32 ψ 2 (x14 + f )
which shows, by Proposition 19, that x12 x3 + x32 + x3 l is an exotic summand for Gx0 . Thus
Gx0 has the Hilbert function of the binary polynomial x14 + f . The maximal values of this
function is clearly (1, 2, 2, 1, 1), so this proves the claim that length(ZG,x0 ) ≤ 7.
Finally, suppose that there exists a local Gorenstein scheme of length at most 6 apolar to F .
It must be defined by some polynomial g whose cubic tail coincides with Fl for some l. Thus
g has degree at least four, and its Hilbert function is Hg = (1, 2, 1, 1, 1), Hg = (1, 1, 1, 1, 1)
or Hg = (1, 1, 1, 1, 1, 1). In the first two cases, g has degree 4 and the leading summand of
g is a pure power l 4 . Each order-one partial of the cubic summand Fl,3 of Fl is therefore
proportional to l 2 . In particular π ∗ (Ψ)(Fl,3 ) = 0 for some Ψ ∈ l⊥ ⊂ T̄1 , so Ψ(F ) is divisible
by l contradicting, as above, the generality assumption on F .
In the case Hg = (1, 1, 1, 1, 1, 1) we see that Fl is the degree-at-most-three part of g and the
standard form of g is x5 . By Example 25, we have a 4 · 3-dimensional choice of Fl . Together
with the choice of l, we obtain a 16-dimensional variety of possible F , thus such F is not
general.
4. On the dimension of the cactus varieties of cubic forms
In this section we consider polynomials with Hilbert function (1, m − 1, m − 1, 1) and
(1, m − 1, m − 1, 1, 1) and derive lower bounds on the dimension of the Cactus variety of cubic
forms Cactus2m (V3,n ) and Cactus2m+1 (V3,n ), respectively.
The cactus variety Cactusr (V3,n ) of the third Veronese embedding V3,n ⊂ P(S̄3 ) is, according
to Proposition 4, the closure of the family of cubic forms [F ] admitting a decomposition
F = F1 + · · · + Fs and distinct linear forms l1 , . . . , ls ∈ S̄1 and forms G1 , . . . , Gs ∈ S, such that
ZG1 ,l1 ∪ · · · ∪ ZGs ,ls has length at most r, and the dehomogenization fi = π[li ] (F ) of F at li is
the cubic tail of the dehomogenization of Gi at li (see Definition 28).
To get a lower bound on the dimension of the cactus variety, we consider the extreme
opposite to the higher secants, namely linear spaces that intersect V3,n in a local scheme. In
particular we consider the closure
W2m (V3,n ) ⊂ Cactusr (V3,n ) ⊂ P(S̄3 )
of the family of cubic forms [F ], for which there exist a linear form l ∈ S̄1 and a form G ∈ S,
such that f = π[l] (F ) is the cubic tail of g = π[l] (G) and the polynomial g has Hilbert
function (1, m − 1, m − 1, 1). We define W2m+1 (V3,n ) analogously, using the Hilbert function
(1, m − 1, m − 1, 1, 1). In the first case g is itself a cubic polynomial, i.e. its own cubic tail,
while in the second case, g is a quartic polynomial.
To find the dimension of Wr (V3,n ) when r = 2m, we note that it is the union over l ∈ S̄1
of varieties isomorphic to the projectivisation of Vr (n), the family of cubic polynomials f ∈
K[x1 , . . . , xn ] with Hilbert function (1, m−1, m−1, 1). When r = 2m+1, the variety Wr (V3,n )
is union over l of varieties isomorphic to the projectivisation of Vr (n) of Tailsr (3, n), the family
of cubic polynomials f ∈ K[x1 , . . . , xn ], that are tails of polynomials g ∈ K[x1 , . . . , xn ] with
Hilbert function (1, m − 1, m − 1, 1, 1).
Example 37. If f ∈ Kdp [x1 , . . . , xn ] has Hilbert function Hf = (1, m − 1, m − 1, 1), its only
possible symmetric decomposition is
Hf = ∆ = ∆0 = (1, m − 1, m − 1, 1),
20
A. BERNARDI, J. JELISIEJEW, P. MACIAS MARQUES, K. RANESTAD
and therefore deg f = 3, ∆0 (1) = m − 1 and
(n0 , n1 ) = (m − 1, m − 1).
If fα is a general cubic polynomial in Kdp [x1 , . . . , xm−1 ], and fβ is any quadratic polynomial
in hxm , . . . , xn ihx1 , . . . , xm−1 , 1i, then fα is in standard form and fβ is an exotic summand for
fα + fβ (cf. Definition 13). Furthermore Hfα +fβ = ∆ and so fα + fβ ∈ V2m (n). The subspace
hx1 , . . . , xm−1 i = Lin(f )0
is determined by f , and the variety of subspaces hx1 , . . . , xm−1 i ⊂ hx1 , . . . , xn i has dimension
(n − m + 1)(m − 1), so we get
m+2
dim V2m (n) =
+ (2m − 1)(n − m + 1).
3
Notice furthermore that the affine variety V2m (n) is a cone, so that its projectivisation has
dimension one less.
Example 38. If g ∈ Kdp [x1 , . . . , xn ] has Hilbert function Hf = (1, m − 1, m − 1, 1, 1), its only
possible symmetric decomposition is
∆ = (∆0 , ∆1 ) = (1, 1, 1, 1, 1), (0, m − 2, m − 2, 0) ,
and therefore deg g = 4, ∆0 (1) = 1, ∆1 (1) = m − 2 and
(n0 , n1 , n2 ) = (1, m − 1, m − 1).
If fα is a general cubic polynomial in Kdp [x1 , . . . , xm−1 ], and fβ is a cubic polynomial of the
form l0 x12 + l02 + f∞ , where l0 ∈ hxm , . . . , xn i and f∞ ∈ hxm , . . . , xn ihx1 , . . . , xm−1 , 1i, then
x14 + fα is in standard form and fβ is an exotic summand for
g = x14 + fα + fβ = x14 + fα + l0 x12 + l02 + f∞ .
Furthermore
Hg = ∆0 + ∆1 = (1, m − 1, m − 1, 1, 1),
and so fα + fβ ∈ Tails2m+1 (3, n).
The flag of subspaces
hx1 i = Lin(g)0 ⊂ hx1 , . . . , xm−1 i = Lin(g)1 ⊂ hx1 , . . . , xn i
is determined by g, and the variety of such flags has dimension m − 2 + (n − m + 1)(m − 1),
so we get
m+2
+ (2m − 1)(n − m + 1) + n − 1.
dim Tails2m+1 (3, n) =
3
Notice that, since the summand l0 x21 + l02 is quadratic in the form l0 , the affine variety
Tails2m+1 (3, n) is not a cone, so its projectivisation has the same dimension.
We use the Examples 37 and 38 to give a lower bound on the dimension of the union of
linear spaces that intersect V3,n in a local subscheme.
(
Proposition 39. Let 3 ≤ m ≤ n. The union CactusL
2m (V3,n ) of linear spans in P
local subschemes in V3,n of length 2m has dimension
m+2
L
dim Cactus2m (V3,n ) ≥
+ 2m(n − m) + 3m − 2.
3
n+3
3
)−1 of
21
(n+3
3 )−1 of local subschemes
Let 4 ≤ m ≤ n. The union CactusL
2m+1 (V3,n ) of linear spans in P
in V3,n of length 2m + 1 has dimension
m+2
(V
)
≥
+ 2m(n − m) + 3m + n − 2.
dim CactusL
3,n
2m+1
3
L
Proof. Clearly W2m (V3,n ) ⊂ CactusL
2m (V3,n ) and W2m+1 (V3,n ) ⊂ Cactus2m+1 (V3,n ), so we get
the inequality by computing the dimension of these subvarieties.
Let m > 2. W2m (V3,n ) is the union as l varies, of projective varieties whose affine cones are
isomorphic to V2m (n), so W2m (V3,n ) has dimension
dim W2m (V3,n ) ≤ dim V2m (n) − 1 + n
equal to the right hand side in the lemma. Similarly, W2m+1 (V3,n ) is the union, as l varies, of
varieties isomorphic to Tails2m+1 (3, n), so
dim W2m+1 (V3,n ) ≤ dim Tails2m+1 (3, n) + n.
In both cases the right hand side is the dimension of the given parametrization of the variety
Wr (V3,n ). To get equality, we show that the parameterization is generically one to one.
When r is even, we show that for a general [F ] ∈ Wr (V3,n ) there is a unique l such that
ZF,l has length r. When r is odd, we show that there is a unique l such that f = πl (F ) is the
tail of a quartic polynomial gl whose apolar scheme Zgl has length r.
Let r = 2m ≤ 2n, and assume that [F ] ∈ Wr (V3,n ) is general. Let l ∈ S̄1 and f = πl (F )
be the local polynomial of F at l such that f has Hilbert function (1, m − 1, m − 1, 1). Then
F = F3 + lF2 where F3 depends on m − 1 variables. Therefore V (F, l) is a cone inside the
hyperplane V (l). Let [y] be a point in the n − m dimensional linear vertex of V (F, l). Then
y(F3 ) = y(l) = 0. Furthermore, since all partials of Fl of degree 1 are partials of F3 , we have
y 2 (F2 ) = 0. In particular, y 2 (F ) = 0 and y(F ) = l · l′ , so V (F ) is singular at [y] with a tangent
cone of rank 2. On the other hand, if V (F ) is singular at [y] with a tangent cone of rank 2,
then
F = F3 + lF2 + l · l′ x,
where y(x) = 1, y(F3 ) = y(F2 ) = y(l) = y(l′ ) = 0 and F3 has Hilbert function (1, m−1, m−1, 1)
for some m ≤ n.
If r = 2m + 1, and assume that [F ] ∈ Wr (V3,n ) is general. Let l ∈ S̄1 and f = πl (F ) be the
local polynomial of F at l such that f has Hilbert function (1, m, m, 1), but is the cubic tail
of a quartic polynomial g with Hilbert function (1, m − 1, m − 1, 1, 1). When m ≤ n, then
g = α0 l04 + g3 + α1 l02 l′ + α2 (l′ )2 + g2
where g3 and l0 depends on m variables and l′ is a hidden variable for g. Thus
f3 = g3 + α1 l02 l′ ,
F = g3 + α1 l02 l′ + α2 l(l′ )2 + lg2
and V (F, l) is a singular cubic hypersurface with a double point whose tangent cone is a square.
In fact, if m < n, then V (F, l) is a cone with linear vertex of dimension n − m − 1 over such a
singular hypersurface.
In both cases, if m < n, let l1 , . . . , ln−m are general linear forms. If r = 2m, then the
linear section V (F, l1 , . . . , ln−m ) still has a singular point whose tangent cone has rank 2. If
r = 2m + 1, then the linear section V (F3 , l, l1 , . . . , ln−m ) is still a singular cubic hypersurface
inside a (m − 1)-dimensional linear subspace with a non reduced tangent cone at the singular
point. The proof of uniqueness of l may therefore in both cases be reduced to the case, when
n = m.
The following is a classical result.
22
A. BERNARDI, J. JELISIEJEW, P. MACIAS MARQUES, K. RANESTAD
Lemma 40. The set of singular cubic hypersurfaces in Pn , n > 2 whose tangent cone at the
n+3
singular point has rank at most 2, form a subvariety of codimension n−1
+ 1 in P( 3 )−1 ,
2
and the general member in the set has exactly one singular point.
Proof. It suffices to note that the set of singular cubic hypersurfaces in Pn form a hypersurface
n+3
in P( 3 )−1 . The general point in this hypersurface, the discriminant, corresponds to a cubic
hypersurface with a quadratic singularity, i.e. the tangent cone is a quadric of rank n. In the
space of quadrics of rank at most n, the quadrics of rank 2 form a subvariety of codimension
n−1
2 . The two codimensions add up to the codimension in the lemma. For uniqueness it
suffices to fix a quadric q of rank 2 and a point p in its vertex,and notice, by Bertini’s theorem,
that the general cubic hypersurface through for which q is the tangent cone at p is smooth
elsewhere.
Remark 41. Notice that the codimension in lemma is consistent with the dimensions of
W2n (V3,n ). When n = m in the proposition, we get
n+3
n−1
dim W2n,n (V3,n ) =
−1−
− 1.
3
2
For the case r is odd, we show that
Lemma 42. The set of cubic hypersurfaces in Pn , n > 3 with a singular hyperplane section
whose tangent cone at the singular point is a square, form a subvariety of codimension n−2
−1
2
n+3
−1
)
(
3
, and when n > 4, the general member in the set has exactly one such hyperplane
in P
section.
Proof. Assume V (F ) is a general cubic of dimension n − 1 > 2 with a singular hyperplane
section V (F, l) whose tangent cone at the singular point is a square. Let p ∈ V (F, l) be a
singular point and V (l, (l′ )2 ) the tangent cone in V (l). We may choose coordinates x0 , . . . , xn ,
so that l = x0 , l′ = x1 and p = [0 : . . . : 0 : 1]. Then
F = F3 + x21 xn + x0 F2 ,
where F3 ∈ Kdp [x1 , . . . , xn−1 ] and F3 + x21 xn is the cubic form defining
the singular
hyperplane
n+2
section V (F, x0 ) and F2 ∈ Kdp [x0 , . . . , xn ]. Thus F depends on n+1
+
1
+
coefficients.
3
2
Now, p, l and l′ varies in a (n + (n − 1) + (n − 2)) dimensional variety, so we get that cubics
with a singular hyperplane section whose tangent cone at the singular point is a square form
a variety of codimension
n+3
n+1
n+2
n+1
n−2
−
+1+
− 3(n − 1) =
− 1 − 3(n − 1) =
−1
3
3
2
2
2
When n > 4, this codimension is positive.
The forms F , when F3 and F2 vary, define a linear system of cubic hypersurfaces with base
locus supported at p = [0 : . . . : 0 : 1]. The general member is smooth, and the tangent
hyperplane section at p is singular only at p, and the tangent cone at p is a square. If this
hyperplane section is not unique with this property, there is another point q distinct from p
such that the tangent hyperplane section at q also has this property. To count dimensions,
we fix two flags p ∈ Lp ⊂ Hp and q ∈ Lq ⊂ Hq , and consider the space of smooth cubic
hypersurfaces through p and q, whose tangent hyperplanes are Hp and Hq and whose tangent
cones at p and q are squares with support along Lp and Lq respectively. Notice that Hp and
Hq are distinct, while Lp ∩ Hq may equal Lq ∩ Hp . This gives two cases for the dimension
count. These are both similar to the dimension count above and show that the variety of
cubic hypersurfaces with two special points as above, has positive codimension in the variety
23
of cubics with only one such point when n > 4. Therefore the last statement of the lemma
follows.
Remark 43. Notice that codimension in the lemma is consistent with the dimension of
W2n+1 (V3,n ). When n = m in the proposition, we get
n+3
n−2
dim W2n+1,n (V3,n ) =
−1−
+ 1.
3
2
We conclude that the parameterization of Wr (V3,n ) is birational for any even r with
5 < r < 2n + 1, and any odd r with 8 < r ≤ 2n + 1, and hence that the dimension formulas of
Proposition 39 are the dimensions of W2m (V3,n ) and W2m+1 (V3,n ) respectively.
We rewrite the formulas for the dimensions of W2m (V3,n ) and W2m+1 (V3,n ) in terms of the
lengths r = 2m (resp. r = 2m + 1):
(
−1
if 5 < r < 2n + 1, r even,
(rn + r − 1) + r(r−2)(r−16)
48
dim Wr (V3,n ) =
−
2
if
8 < r < 2n + 2, r odd.
(rn + r − 1) + (r−1)(r−3)(r−17)
48
Corollary 44. When 18 ≤ r ≤ 2n + 2, then
(
(rn + r − 1) +
dim Cactusr (V3,n ) ≥
(rn + r − 1) +
r(r−2)(r−16)
−1
48
(r−1)(r−3)(r−17)
−
48
2
if r ≥ 18 even,
if r ≥ 19 odd.
For each possible Hilbert function for local schemes of length r, one may define a variety
analogous to Wr (V3,n ). The dimensions of these varieties are in general not known, and this
remains an obstacle to finding a precise dimension for the cactus variety Cactusr (V3,n ).
Finally we leave an open question: we know that the cactus rank of a general cubic surface
equals the rank, which is 5, while the local cactus rank is 7 (see Proposition 36), but we do
not know whether for a larger number of variables the local cactus rank and the cactus rank
agree.
Question 45. Is the cactus rank of a cubic form in Kdp [x0 , . . . , xn ] always computed locally,
when n ≥ 8 and the cactus rank is at least 18?
Acknowledgements
JJ and KR thank Jaroslaw Buczyński for fruitful discussions and the Homing Plus programme of Foundation for Polish Science, co-financed from European Union, Regional Development Fund for partial support of their mutual visits. PMM thanks Anthony Iarrobino and
Jerzy Weyman for the invitation and hospitality at Northeastern University, and is grateful to
Anthony Iarrobino for having introduced him to this subject and for very fruitful discussions.
He also thanks James Adler for help with language. AB was partially supported by Project
Galaad of INRIA Sophia Antipolis Méditerranée, France, Marie Curie Intra-European Fellowships for Carrer Development (FP7-PEOPLE-2009-IEF): “DECONSTRUCT”, GNSAGA
of INDAM, Mathematical Department Giuseppe Peano of Turin, Italy, and Politecnico of
Turin, Italy. JJ is a doctoral fellow at the Warsaw Center of Mathematics and Computer Science financed by the Polish program KNOW and by Polish National Science Center, project
2014/13/N/ST1/02640 and a member of “Computational complexity, generalised Waring type
problems and tensor decompositions” project within “Canaletto”, the executive program for
scientific and technological cooperation between Italy and Poland, 2013-2015. PMM was partially supported by Fundação para a Ciência e Tecnologia, projects “Geometria Algébrica em
Portugal”, PTDC/MAT/099275/2008, “Comunidade Portuguesa de Geometria Algébrica”,
24
A. BERNARDI, J. JELISIEJEW, P. MACIAS MARQUES, K. RANESTAD
PTDC/MAT-GEO/0675/2012, and sabbatical leave grant SFRH/BSAB/1392/2013, by CIMA
– Centro de Investigação em Matemática e Aplicações, Universidade de Évora, projects PEstOE/MAT/UI0117/2011 and PEst-OE/MAT/UI0117/2014, and by Fundação de Amparo à
Pesquisa do Estado de São Paulo, grant 2014/12558–9. KR was supported by the RCN project
no 239015 “Special Geometries”.
References
[Alexander, Hirschowitz 1995] Alexander, James, Hirschowitz, André: Polynomial interpolation in several variables. J. of Alg. Geom., 4 (1995), no. 2, pp. 201–222.
[Bernardi, Brachat, Mourrain 2014] Bernardi, Alessandra, Brachat, Jérôme, Mourrain, Bernard: A comparison
of different notions of ranks of symmetric tensors, Linear Algebra and its Applications, 460, 2014, pp. 205–230.
[Bernardi, Ranestad 2012] Bernardi, Alessandra, Ranestad, Kristian: On the cactus rank of cubics forms, J.
Symbolic Comput. 50 (2012), pp. 291–297.
[Buczyńska, Buczyński 2014] Buczyńska, Weronika, Buczyński, Jaroslaw: Secant varieties to high degree
Veronese reembeddings, catalecticant matrices and smoothable Gorenstein schemes, J. Algebraic Geom. 23
(2014), no. 1, pp. 63–90.
[Casnati, Notari 2011] Casnati, Gianfranco, Notari, Roberto: On the irreducibility and the singularities of the
Gorenstain locus of the punctual Hilbert scheme of degree 10, J. Pure Appl. Algebra 215 (2011), no. 6, pp.
1243–1254.
[Casnati, Notari 2013] Casnati, Gianfranco, Notari, Roberto: A structure theorem for 2-stretched Gorenstein
algebras, arXiv:1312.2191, to apear in J. Commut. Algebra.
[Casnati, Jelisiejew, Notari 2015] Casnati, Gianfranco, Jelisiejew, Joachim, Notari, Roberto: Irreducibility of
the Gorenstein loci of Hilbert schemes via ray families, Algebra Number Theory 9 (2015), no. 7, pp. 1525–1570.
[Dolgachev 2012] Dolgachev, Igor V.: Classical algebraic geometry: a modern view. Cambridge University
Press, (2012).
[Emsalem 1978] Emsalem, Jacques: Géométrie des points épais, Bull. Soc. Math. France 106, (1978), no. 4,
pp. 399–416.
[Iarrobino 1994] Iarrobino, Anthony: Associated graded algebra of a Gorenstein Artin Algebra, Mem. Amer.
Math. Soc. 107, (1994), no. 514, Amer. Math. Soc. Providence.
[Iarrobino, Kanev 1999] Iarrobino, Anthony, Kanev, Vassil: Power Sums, Gorentein Algebras and Determinantal Loci, Lecture Notes in Mathematics 1721, Springer-Verlag, Berlin Heidelberg New York (1999).
[Jelisiejew 2015] Jelisiejew, Joachim: Classifying local Artinian Gorenstein algebras, arXiv:1511.08007.
[Macaulay 1927] Macaulay, Francis Sowerby: Some properties of enumeration in the theory of modular systems,
Proc. London Math. Soc. S2-26 (1927), no. 1, pp. 531–555.
Dipartimento di Matematica, Università di Trento, via Sommarive 14, I-38123 Povo (Trento),
Italy.
E-mail address: [email protected]
Faculty of Mathematics, Informatics, and Mechanics, University of Warsaw, Banacha 2, 02-097
Warszawa, Poland
E-mail address: [email protected]
Departamento de Matemática, Escola de Ciências e Tecnologia, Centro de Investigação em
Matemática e Aplicações, Instituto de Investigação e Formação Avançada, Universidade de Évora,
Rua Romão Ramalho, 59, P–7000–671 Évora, Portugal
E-mail address: [email protected]
Matematisk institutt, Universitetet i Oslo, PO Box 1053, Blindern, NO-0316 Oslo, Norway
E-mail address: [email protected]
| 0math.AC
|
1
arXiv:1511.05625v1 [cs.NE] 18 Nov 2015
MOEA/D-GM: Using probabilistic graphical models
in MOEA/D for solving combinatorial optimization
problems
Murilo Zangari de Souza1, Roberto Santana2, Aurora Trinidad Ramirez Pozo1, Alexander Mendiburu2
1 DInf - Federal University of Paraná, CP: 19081, CEP 19031-970, Curitiba, Brazil
2 Intelligent Systems Group, University of the Basque Country UPV/EHU, San Sebastián, Spain
Email: {murilo.zangari, auroratrinidad}@gmail.com, {roberto.santana, alexander.mendiburu}@ehu.eus
(PF). Popular strategies include:1 (i) Pareto dominance based;
(ii) Indicator based and (iii) Decomposition based (also called,
scalarization function based) [3], [5].
Zhang and Li [2] proposed a decomposition based algorithm called MOEA/D (multi-objective Evolutionary Algorithm Based on Decomposition) framework, which decomposes a MOP into a number of single-objective scalar optimization subproblems and optimizes them simultaneously in
a collaborative manner using the concept of neighborhood
between the subproblems.
Current research on MOEA/D are various and include
the extension of these algorithms to continuous MOPs with
complicated Pareto sets [6], many-objective optimization problems [5], [7], [8], methods to parallelize the algorithm [9],
incorporation of preferences to the search [10], automatic
adaptation of weight vectors [11], new strategies of selection
and replacement to balance convergence and diversity [3],
hybridization with local searches procedures [12], [13], etc.
However, it has been shown that traditional EA operators
fail to properly solve some problems when certain characterIndex Terms—multi-objective optimization, MOEA/D, proba- istics are present in the problem such as deception [14]. A
main reason for this shortcoming is that, these algorithms do
bilistic graphical models, deceptive functions, EDAs
not consider the dependencies between the variables of the
problem [14]. To address this issue, evolutionary algorithms
I. I NTRODUCTION
that (instead of using classical genetic operators) incorporate
Several real-world problems can be stated as multi-objective machine learning methods have been proposed. These algooptimization problems (MOPs) which have two or more ob- rithms are usually referred to as Estimation of Distribution
jectives to be optimized. Very often these objectives conflict Algorithms (EDAs). In EDAs, the collected information is repwith each other. Therefore, no single solution can optimize resented using a probabilistic model which is later employed
these objectives at the same time. Pareto optimal solutions are to generate (sample) new solutions [14]–[16].
of very practical interest to decision makers for selecting a
EDAs based on modeling the statistical dependencies befinal preferred solution. Most MOPs may have many or even tween the variables of the problem have been proposed as
infinite optimal solutions and it is very time consuming (if not a way to encode and exploit the regularities of complex
impossible) to obtain the complete set of optimal solutions [1]. problems. Those EDAs use more expressive probabilistic
Since early nineties, much effort has been devoted to models that, in general, are called probabilistic graphical
develop evolutionary algorithms for solving MOPs. Multi- models (PGMs) [14]. A PGM comprises a graphical compoobjective evolutionary algorithms (MOEAs) aim at finding a nent representing the conditional dependencies between the
set of representative Pareto optimal solutions in a single run variables, and a set of parameters, usually tables of marginal
[1]–[3].
and conditional probabilities. Additionally, the analysis of the
Different strategies have been used as criteria to maintain a
population of optimal non-dominated solutions, or Pareto set
1 There exist algorithms, such as NSGAIII [4], that combine the ideas from
(PS), and consequently finding an approximated Pareto front both decomposition based and Pareto dominance based.
Abstract—Evolutionary algorithms based on modeling the
statistical dependencies (interactions) between the variables have
been proposed to solve a wide range of complex problems. These
algorithms learn and sample probabilistic graphical models able
to encode and exploit the regularities of the problem. This paper
investigates the effect of using probabilistic modeling techniques
as a way to enhance the behavior of MOEA/D framework.
MOEA/D is a decomposition based evolutionary algorithm that
decomposes an multi-objective optimization problem (MOP) in
a number of scalar single-objective subproblems and optimizes
them in a collaborative manner. MOEA/D framework has been
widely used to solve several MOPs. The proposed algorithm,
MOEA/D using probabilistic Graphical Models (MOEA/D-GM)
is able to instantiate both univariate and multi-variate probabilistic models for each subproblem. To validate the introduced
framework algorithm, an experimental study is conducted on
a multi-objective version of the deceptive function Trap5. The
results show that the variant of the framework (MOEA/DTree), where tree models are learned from the matrices of the
mutual information between the variables, is able to capture
the structure of the problem. MOEA/D-Tree is able to achieve
significantly better results than both MOEA/D using genetic
operators and MOEA/D using univariate probability models in
terms of approximation to the true Pareto front.
2
graphical components of the PGMs learned during the search
can provide information about the problem structure. MOEDAs using PGMs have been also applied for solving different
MOPs [17]–[19].
In the specified literature of MO-EDAs, a number of algorithms that integrate to different extents the idea of probabilistic modeling into MOEA/D have been previously proposed [20]–[24]. Usually, these algorithms learn and sample
a probabilistic model for each subproblem. Most of them
use only univariate models, which are not able to represent
dependencies between the variables.
In this paper, we investigate the effect of using probabilistic
modeling techniques as a way to enhance the behavior of
MOEA/D. We propose a MOEA/D framework able to instantiate different PGMs. The general framework is called MOEA/D
Graphical Models (MOEA/D-GM). The goals of this paper are:
(i) Introduce MOEA/D-GM as a class of MOEA/D algorithms
that learn and sample probabilistic graphical models defined on
discrete domain; (ii) To empirically show that MOEA/D-GM
can improve the results of traditional MOEA/D for deceptive
functions; (iii) To investigate the influence of modeling variables dependencies in different metrics used to estimate the
quality of the Pareto front approximations and (iv) to show,
for the first time, evidence of how the problem structure is
captured in the models learned by MOEA/D-GM.
The paper is organized as follows. In the next section some
relevant concepts used in the paper are introduced. Sections III
and IV respectively explain the basis of MOEA/D and the
EDAs used in the paper. Section V introduces the MOEA/DGM and explains the enhancements required by EDAs in order
to efficiently work within the MOEA/D context. In Section VI
we discuss the class of functions which are the focus of
this paper, multi-objective deceptive functions. We explain the
rationale of applying probabilistic modeling for these functions. Related work is discussed in detail in Section VII. The
experiments that empirically show the behavior of MOEA/DGM are described in Section VIII. The conclusions of our
paper and some possible trends for future work are presented
in Section IX.
II. P RELIMINARIES
Let X = (X1 , . . . , Xn ) denote a vector of discrete random
variables. x = (x1 , . . . , xn ) is used to denote an assignment
to the variables and x ∈ {0, 1}n. A population is represented
as a set of vectors x1 , . . . , xN where N is the size of the
population. Similarly, xij represents the assignment to the jth
variable of the ith solution in the population.
A general MOP can be defined as follows [1]:
minimize (or maximize) F (x) = (f1 (x), ..., fm (x))
(1)
subject to x ∈ Ω
where x = (x1 , x2 , . . . xn )T is the decision variable vector of
size n, Ω is the decision space, F : Ω → Rm consists of m
objective functions and Rm is the objective space.
Pareto optimality is used to define the set of optimal
solutions.
Pareto optimality2 [1]: Let x, y ∈ Ω, x is said to dominate
y if and only if fl (x) ≤ fl (y) for all l ∈ {1, .., m} and
fl (x) < fl (y) for at least one l. A solution x∗ ∈ Ω is called
Pareto optimal if there is no other x ∈ Ω which dominates x∗ .
The set of all the Pareto optimal solutions is called the Pareto
set (PS) and the solutions mapped in the objective space are
called Pareto front (PF), i.e., P F = {F (x)|x ∈ P S}. In many
real life applications, the PF is of great interest to decision
makers for understanding the trade-off nature of the different
objectives and selecting their preferred final solution.
III. MOEA/D
MOEA/D [2] decomposes a MOP into a number (N ) of
scalar single objective optimization subproblems and optimizes them simultaneously in a collaborative manner using the
concept of neighborhood between the subproblems. Each subproblem i is associated to a weight vector λi = (λi1 , ..., λim )T
and the set of all weight vectors is {λ1 , ..., λN }.
Several decomposition approaches have been used in the
MOEA/D [25]. The Weighted Sum and Tchebycheff are two
of the most common approaches used, mainly in combinatorial
problems [2], [25].
P
Let λ = (λ1 , ..., λm )T be a weight vector, where m
l=1 λl =
1 and λl ≥ 0 for all l = 1, ..., m.
Weighted Sum Approach: The optimal solution to the following scalar single-optimization problem is defined as:
minimize (or maximize) g ws (x|λ) =
m
X
λl fl (x)
l=1
subject to x ∈ Ω
(2)
where we use g ws (x|λ) to emphasize that λ is a weight vector
in this objective function. x is a Pareto optimal solution to (1)
if the PF of (1) is convex (or concave).
Tchebycheff approach: The optimal solution to the following
scalar single-optimization problems is defined as:
minimize g te (x|λ, z ∗ ) = max {λl |fl (x) − zl∗ |};
1≤l≤m
(3)
subject to x ∈ Ω
∗ T
) is the reference point, i.e., zl∗ =
where z ∗ = (z1∗ , ..., zm
max{fl (x)|x ∈ Ω} for each l = 1, ..., m. For each Pareto
optimal solution x∗ there exists a weight vector λ such that
x∗ is the optimal solution of (3) and each optimal solution of
(3) is Pareto optimal of (1).
The neighborhood relation among the subproblems is defined in MOEA/D. The neighborhood for each subproblem
i, B(i) = {i1 , ..., iT }, is defined according to the Euclidean
distance between its weight vector and the other weight
vectors. The relationship of the neighbor subproblems is used
for the selection of parent solutions and the replacement of
old solutions (also called update mechanism). The size of the
neighborhood used for selection and replacement plays a vital
role in MOEA/D to exchange information among the subproblems [3], [5]. Moreover, optionally, an external population EP
is used to maintain all Pareto optimal solutions found during
2 This definition of domination is for minimization. For maximization, the
inequalities should be reversed.
3
the search. Algorithm 1 (adapted from [3]) presents the pseudocode of the general MOEA/D which serves as a basis for this
paper.
Algorithm 1: General MOEA/D framework
[P op, Ts , Tr , EP ] ← Initialization() // P op is the initial
population of size N . Set B(i) = Ts and R(i) = Tr , where
B(i) and R(i) are, respectively, the neighborhood size for
selection and replacement. EP is the external population.
2 While a termination condition not met
3
For each subproblem i ∈ 1, ..., N at each generation
4
y ← Variation(B(i))
5
Evaluate F (y) using the fitness function
6
Update_Population(P op, y, R(i))
7
Update_EP (F (y),EP)
8 Return P op, EP
1
1) Initialization: The N weight vectors λ1 , .., λN are set.
The Euclidean distance between any two weight vectors is
computed. For each subproblem i, the set of neighbors for
the selection step B(i) and the update step R(i) are initialized with the Ts and Tr closest neighbors respectively
according to the Euclidean distance. The initial population
P op = x1 , ..., xN is generated in a random way and their
corresponding fitness function F (x1 ), ..., F (xN ) are computed.
The external Pareto EP is initialized with the non-dominated
solutions from the initial population P op.
2) Variation: The reproduction is performed using B(i) =
{i1 , ..., iTs } to generate a new solution y. The conventional
algorithm selects two parent solutions from B(i) and applies
crossover and mutation to generate y.
3) Update Population: This step decides which subproblems should be updated from R(i) = {i1 , ..., iTr }. The current
solutions of these subproblems are replaced by y if y has a
better aggregation function value g te (y|λr ) than g te (xr |λr ),
xr ∈ R(i)
4) Update EP: This step removes from EP the solutions
dominated by y and adds y if no solution dominates it.
A. Improvements on MOEA/D framework
Wang et. al [3] have reported that different problems need
different trade-offs between diversity and convergence, which
can be controlled by the different mechanisms and parameters
of the algorithm. As so far, most proposed MOEA/D versions
adopted the (µ + 1)-selection/variation scheme, which selects
µ individuals from a population and generates 1 offspring [3].
Different strategies for selection and replacement have been
proposed [3], [5]. In this paper, the replacement proposed by
[6] is used. In this strategy, the maximal number of solutions
replaced by a new solution y is bounded by nr , which
should be set to be much smaller than Tr . This replacement
mechanism prevents one solution having many copies in the
population.
IV. E STIMATION OF DISTRIBUTION ALGORITHMS
EDAs [14]–[16] are stochastic population-based optimization algorithms that explore the space of candidate solutions
by sampling a probabilistic model constructed from a set of
selected solutions found so far. Usually, in EDAs, the population is ranked according to the fitness function. From this
ranked population, a subset of the most promising solutions
are selected by a selection operator (such as: Truncation
selection with truncation threshold t = 50%). The algorithm
then constructs a probabilistic model which attempts to estimate the probability distribution of the selected solutions.
Then, according to the probabilistic model, new solutions are
sampled and incorporated into the population, which can be
entirely replaced [26]. The algorithm stops when a termination
condition is met such as the number of generations.
We work with positive distributions denoted by p. p(xI )
denotes the marginal probability for XI = xI . p(xj | xk )
denotes the conditional probability distribution of Xj = xj
given Xk = xk . The set of selected promising solutions
is represented as S. Algorithm 2 presents the general EDA
procedure.
Algorithm 2: General EDA
1
2
3
4
5
6
P op ← Generate N solutions randomly
While a termination condition not met
For each solution xi compute its fitness function F (xi )
Select the most promising solutions S from P op
Build a probabilistic model M of solutions in S
Generate new candidate solutions sampling from M and
add them to P op
The way in which the learning and sampling components
of the algorithm are implemented are also critical for their performance and computational cost [14]. In the next section, the
univariate and tree-based probabilistic models are introduced.
A. Univariate probabilistic models
In the univariate marginal distribution (or univariate probabilistic model), the variables are considered to be independent,
and the probability of a solution is the product of the univariate
probabilities for all the variables:
pu (x) =
n
Y
p(xj )
(4)
j=1
One of the simplest EDAs that use the univariate model is
the univariate marginal distribution algorithm (UMDA) [27].
UMDA uses a probability vector pu (x) as the probabilistic
model, where p(xj ) denotes the univariate probability associated to the corresponding discrete value. In this paper we
focus on binary problems. To learn the probability vector
for these problems, each p(xj ) is set to the proportion of
"1s" in the selected population S. To generate new solutions,
each variable is independently sampled. Random values are
assigned to the variables and following the corresponding
univariate distribution.
The population-based incremental learning (PBIL) [28],
like UMDA, uses the probabilistic model in the form of a
probability vector pu (x). The initial probability of a "1" in
each position p(xj ) is set to 0.5. The probability vector is
4
updated using each selected solution x in S. For each variable,
the corresponding entry in the probability vector is updated by:
p(xj ) = (1.0 − α)p(xj ) + (α ∗ xj )
(5)
where α is the learning rate specified by the user.
To prevent premature convergence, each position of the
probability vector is slightly varied at each generation based
on a mutation rate parameter [26], [28]. Recently, it has been
acknowledged [29] that implementations of the probabilistic
vector update mechanism in PBIL are in fact different, and
produce an important variability in the behavior of PBIL.
Univariate approximations are expected to work well for
functions that can be additively
decomposed into functions of
P
order one (e.g. g(x) =
j xj ). Also, other non additively
decomposable functions can beQsolved P
with EDAs that use
univariate models (e.g. g(x) = j xj + j xj ) [30].
will be able to capture the most relevant bivariate dependencies
between the problem variables.
Details on EDAs that use the tree models can be obtained
from [31], [33]. More details on the use of PGMs for probabilistic modeling in EDAs can be obtained from [14].
C. Optimal Mutation Rate for EDAs
Most of EDAs do not use any kind of stochastic mutation.
However, for certain problems, lack of diversity in the population is a critical issue that can cause EDAs to produce
poor results due to premature convergence. As a remedy to
this problem, in [34], the authors propose the use of Bayesian
priors as an effective way to introduce a mutation operator
into UMDA. Bayesian priors are used for the computation of
the univariate probabilities in such a way that the computed
probabilities will include a mutation-like effect. In this paper
we use the Bayesian prior as a natural way to introduce
B. Tree-based models
mutation in the MOEA/D-GM. In the following this strategy
Different from the EDAs that use univariate models, some is described.
There are two possibilities to estimate the probability of
EDAs can assume dependencies between the decision vari"head"
of a biased coin. The maximum-likelihood estimate
ables. In this case, the probability distribution is represented
counts
the
number of occurrences of each case. With m times
by a PGM.
"head"
in
l
throws, the probability is estimated as p = ml .
Tree-based models [31] are PGMs capable to capture some
pairwise interactions between variables. In a tree model, the The Bayesian approach assumes that the probability of "head"
conditional probability of a variable may only depend on at of a biased coin is determined by an unknown parameter θ.
Starting from an a priori distribution of this parameter. Using
most one other variable, its parent in the tree.
rule the univariate probability is computed as
The probability distribution pT (x) that is conformal with a the Bayesian
m+r
,
and
the so called hyperparameter r has to be chosen
p
=
l+2r
tree model is defined as:
in
advance.
n
Y
To relate the Bayesian prior to the mutation, the authors
p(xj |pa(xj )),
(6)
pT (x) =
used
the following theorem: For binary variables, the exj=1
pectation value for the probability using a Bayesian prior
where pa(Xj ) is the parent of Xj in the tree, and with parameter r is the same as mutation with mutation rate
p(xj |pa(xj )) = p(xj ) when pa(Xj ) = ∅, i.e. Xj is a root µ = r and using the maximum likelihood estimate [34].
l+2r
of the tree.
The bivariate marginal distribution algorithm (BMDA) proV. MOEA/D-GM: A N MOEA/D ALGORITHM WITH
posed in [31] uses a model based on a set of mutually
GRAPHICAL MODELS
3
independent trees (a forest) . In each iteration of the algorithm,
Algorithm 3 shows the steps of the proposed MOEA/Da tree model is created and sampled to generate new candidate
solutions based on the conditional probabilities learned from GM framework. First, the initialization procedure generates N
initial solutions and the external Pareto (EP) is initialized with
the population.
The algorithm Tree-EDA proposed in [32] combines fea- the non-dominated solutions. Within the main while-loop, in
tures from algorithms introduced in [33] and [31]. In Tree- case the termination criteria are not met, for each subproblem
i, a probabilistic model is learned using as base population
EDA, the learning procedure works as follows:
• Step 1: Compute the univariate and bivariate marginal fre- all solutions in the neighborhood B(i). Then, the sampling
quencies pj (xj ) and pjk (xj |xk ) using the set of selected procedure is used to generate a new solution from the model.
The new solutions is used to update the parent population
promising solutions S;
according to an elite-preserving mechanism. Finally, the new
• Step 2: Calculate the matrix of mutual information using
solution is used to update EP as described in Algorithm 1.
the univariate and bivariate frequencies;
A simple way to use probabilistic models into MOEA/D
• Step 3: Calculate the maximum weight spanning tree
is
learning
and sampling a probabilistic model for each scalar
from the mutual information. Compute the parameters of
subproblem
i using the set of closest neighbors as the selected
the model.
population. Therefore, at each generation, the MOEA/D-GM
The idea is that by computing the maximum weight spankeeps N probabilistic models.
ning tree from the matrix of mutual information the algorithm
The EDAs presented in the previous section: (i) UMDA
[27], (ii) PBIL [28] and (iii) Tree-EDA [32] can be instantiated
3 For convenience of notation, this set of mutually independent trees is
in the framework. Also, the genetic operators (crossover and
refereed as Tree.
5
Algorithm 3: MOEA/D-GM
1 [P op, Ts , Tr , EP ] ← Initialization() // P op is the initial population
of size N that is randomly generated. Set B(i) = Ts and R(i) = Tr ,
where B(i) and R(i) are, respectively, the neighborhood size for selection
and replacement. EP is the external population.
2 While a termination condition not met
3
For each subproblem i ∈ 1, ..., N at each generation
4
Learning()
5
Case GA: Choose two parent solutions x1 x2 from B(i)
6
Case UMDA: Learn a probabilistic vector pi using the solutions
from B(i)
7
Case PBIL: Learn an incremental probabilistic vector pi using the
solutions from B(i)
8
Case Tree-EDA: Learn a tree model piT using the solutions from
B(i)
Sampling() // Try up to Ts times to sample a new solution y that is
different from any solution from B(i)
Case GA: Apply the crossover and mutation to generate y
9
10
11
Case PBIL or UMDA: Sample y using the probability vector pi
12
Case Tree-EDA: Sample y using the tree model piT
13
Compute the fitness function F (y)
14
Update_Population(P op, y, R(i))
15
Update_EP (F (y),EP)
16 Return P op, EP
mutation), that are used in the standard MOEA/D, can be
applied using the learning and sampling procedures.
The general MOEA/D-GM allows the introduction of other
types of probabilistic graphical models (PGMs) like Bayesian
networks [35] and Markov networks [36]. Different classes of
PGMs can be also used for each scalar subproblem but we do
not consider this particular scenario in the analysis presented
here.
Moreover, the MOEA/D-GM introduces some particular
features that are explained in the next section.
1) Learning models from the complete neighborhood:
Usually in EDAs, population sizes are large and the size of the
selected population can be as high as the population size. In
MOEA/D, each subproblem has a set of neighbors that plays
a similar role to the selected population4. Therefore, the main
difference between the MOEA/D-GM and the EDAs presented
in section IV is that, in the former, instead of selecting a
subset of individuals based on their fitness to keep a unique
probabilistic model, a probabilistic model is computed over
the neighborhood solutions of each scalar subproblem.
2) Diversity preserving sampling: In preliminary experiments we detected that one cause for early convergence of
the algorithm was that solutions that were already in the
population were newly sampled. Sampling solutions already
present in the population is also detrimental in terms of
efficiency since these solutions have to be evaluated. As a
way to avoid this situation, we added a simple procedure in
which each new sampled solution is tested for presence in the
neighborhood B(i) of each subproblem i.
If a new sampled solution y is equal to a parent solution
from B(i) then the algorithm discards the solution and samples
a new one until a different solution has been found or a
4 Usually, the size of the neighborhood is N × 0.1. Although, as in [6], the
propose MOEA/D has a low probability to select two parent solutions from
the complete population.
maximum number of trials is reached. This procedure is
specially suitable to deal with expensive fitness functions. The
maximum number of tries can be specified by the user. We
call the sampling that incorporates this verification procedure
as diversity preserving sampling (ds). When it is applied
on MOEA/D-GM, to emphasize it, the algorithm is called
MOEA/D-GM-ds.
VI. M ULTI - OBJECTIVE D ECEPTIVE O PTIMIZATION
P ROBLEM
There exists a class of scalable problems where the difficulty
is given by the interactions that arise among subsets of
decision variables. Thus, some problems should require the
algorithm to be capable of linkage learning, i.e., identifying
and exploring interactions between the decision variables to
provide effective exploration. Decomposable deceptive problems [37] have played a fundamental role in the analysis of
EAs. As mentioned before, one of the advantages of EDAs that
use probabilistic graphical models is their capacity to capture
the structure of deceptive functions. Different works in the
literature have proposed EDAs [38]–[41] and MO-EDAs [42],
[43] for solving decomposable deceptive problems.
One example of this class of decomposable deceptive functions is the Trap-k, where k is the fixed number of variables
in the subsets (also called, partitions or building blocks)
[38]. Traps deceive the algorithm away from the optimum if
interactions between the variables in each partition are not
considered. According to [42], that is why standard crossover
operators of genetic algorithms fail to solve traps unless the
bits in each partition are located close to each other in the
chosen representation.
Pelikan et al [42] used a bi-objective version of Trap-k for
analyzing the behavior of multi-objective hierarchical BOA
(hBOA).
The functions ftrap5 (Equation (7)) and finv_trap5 (Equation (8)) consist in evaluating a vector of decision variables
x ∈ {0, 1}n, in which the positions of x are divided into
disjoint subsets or partitions of 5 bits each (n is assumed to
be a multiple of 5). The partition is fixed during the entire optimization run, but the algorithm is not given information about
the partitioning in advance. Bits in each partition contribute to
trap of order 5 using the following functions:
ftrap5 (x) =
l−1 X
5
X
trap5(x5∗k+i )
(7)
k=0 i=1
(
5
trap5(u) =
4−u
finv_trap5 (x) =
l−1 X
5
X
if u = 5
if u < 5
inv_trap5(x5∗k+i )
(8)
k=0 i=1
(
5
inv_trap5(u) =
u−1
if u = 0
if u > 0
where l is the number of building blocks, i.e., n = 5l, and u
is the number of ones in the input string of 5 bits.
6
In the bi-objective problem, ftrap5 and finv_trap5 are conflicting. Thus, there is not one single global optimum solution
but a set of Pareto optimal solutions. Moreover, the amount of
possible solutions grows exponentially with the problem size
[39]. Multi-objective Trap-k has been previously investigated
for Pareto based multi-objective EDAs [42]–[44].
VII. R ELATED WORK
In this section we review a number of related works emphasizing the differences to the work presented in this paper.
Besides, we did not find any previous report of multi-objective
decomposition approaches that incorporate PGMs for solving
combinatorial MOPs.
A. MOEA/D using univariate EDAs
In [21], the multi-objective estimation of distribution algorithm based on decomposition (MEDA/D) is proposed for solving the multi-objective Traveling Salesman Problem (mTSP).
For each subproblem i, MEDA/D uses a matrix Qi to represent
the connection "strength" between cities in the best solutions
found so far. Matrix Qi is combined with a priori information
about the distances between cities of problem i, in a new
matrix P i that represents the probability that the sth and
tth cities are connected in the route of the ith sub-problem.
Although each matrix P i encodes a set of probabilities relevant
for each corresponding TSP subproblem, these matrices can
not be considered PGMs since they do not comprise a graphical component representing the dependencies of the problem.
The type of updates applied to the matrices is more related
to parametric learning than to structural learning as done in
PGMs. Furthermore, this type of "models" resemble more
the class of structures traditionally used for ACO and they
heavily depend on the use of prior information (in the case of
MEDA/D, the incorporation of information about the distances
between cities). Therefore, we do not consider MEDA/D as a
member of the MOEA/D-GM class of algorithms.
Another approach combining the use of probabilistic models
and MOEAD for mTSP is presented in [20]. In that paper, a
univariate model is used to encode the probabilities of each
city of the TSP being assigned to each of the possible positions
of the permutation. Therefore, the model is represented as a
matrix of dimension n × n comprising the univariate probabilities of the city configurations for each position. One main
difference of this hybrid MEDA/D with our approach, and with
the proposal of Zhou et al [21], is that a single matrix is learned
using all the solutions. Therefore, the information contained
in the univariate model combines information from all the
subproblems, disregarding the potential regularities contained
in the local neighborhoods. Furthermore, since the sampling
process from the matrix does not take into account the
constraints related to the permutations, repair mechanisms and
penalty functions are used to "correct" the infeasible solutions.
As a consequence, much of the information sampled from the
model to the solution can be modified by the application of
the repair mechanism.
These previous MEDA/Ds were applied for solving
permutation-based MOPs. The application of EDAs for solving permutation problems is increasing in interest. A number
of EDAs have been also specifically designed to deal with
permutation-based problems [45], [46]. The framework proposed in this paper is only investigated for binary problems.
However, EDAs that deal with the permutation space can be
incorporated into the framework in the future.
In [22], the authors proposed a univariate MEDA/D for solving the (binary) multi-objective knapsack problem (MOKP)
that uses an adaptive operator at the sampling step to preserve
diversity, i.e., prevents the learned probability vector from premature convergence. Therefore, the sampling step depends on
both the univariate probabilistic vector and an extra parameter
"r".
B. MOEA/D using multivariate EDAs
In [23], a decomposition-based algorithm is proposed to
solve many-objective optimization problems. The proposed
framework (MACE-gD) involves two ingredients: (i) a concept
called generalized decomposition, in which the decision maker
can guide the underlying search algorithm toward specific
regions of interest, or the entire Pareto front and (ii) an
EDA based on low-order statistics, namely the cross-entropy
method [47]. MACE-gD is applied on a set of many-objective
continuous functions. The obtained results showed that the
proposed algorithm is competitive with the standard MOEA/D
and RM-MEDA [48]. The class of low-order statistics used
by MACE-gD (Normal univariate models) limit the ability of
the algorithm to capture and represent interactions between
the variables. The univariate densities are updated using an
updating rule as the one originally proposed for the PBIL
algorithm [28].
In [24], the covariance matrix adaptation evolution strategy
(CMA-ES) [49] is used as the probabilistic model of MOEA/D.
Although CMA-ES was introduced and has been developed
in the context of evolutionary strategies, it learns a Gaussian
model of the search. The covariance matrix learned by CMAES is able to capture dependencies between the variables.
However, the nature of probabilistic modeling in the continuous domain is different to the one in the discrete domain.
The methods used for learning and sampling the models are
different. Furthermore, the stated main purpose of the work
presented in [24] was to investigate to what extent the CMAES could be appropriately integrated in MOEA/D and what
are the benefits one could obtain. Therefore, emphasis was put
on the particular adaptations needed by CMA-ES to efficiently
learn and sample its model in this different context. Since these
adaptations are essentially different that the ones required by
the discrete EDAs used in this paper, the contributions are
different.
C. Other MO-EDAs
Pelikan et al. [17] discussed the multi-objective decomposable problems and their difficulty. The authors attempted to review a number of MO-EDAs, such as: multi-objective mixturebased iterated density estimation algorithm (mMIDEA) [50],
multi-objective mixed Bayesian optimization algorithm (mmBOA) [51] and multi-objective hierarchical BOA (mohBOA)
[52]. Moreover, the authors introduced an improvement to
7
mohBOA. The algorithm combines three ingredients: (i) the
VIII. E XPERIMENTS
hierarchical Bayesian optimization algorithm (hBOA) [52],
The proposed MOEA/D-GM framework has been imple(ii) the multi-objective concepts from NSGAII [53] and (iii) mented in C++. In the comparison study, four different instanclustering in the objective space. The experimental study tiations are used. For convenience, we called each algorithm
showed that the mohBOA efficiently solved multi-objective instantiated as: MOEA/D-GA, MOEA/D-UMDA, MOEA/Ddecomposable problems with a large number of competing PBIL and MOEA/D-Tree.
building blocks. The algorithm was capable of effective reTo evaluate the algorithms, the deceptive functions ftrap5
combination by building and sampling Bayesian networks with and f
inv_trap5 (Section VI) are used to compose the bidecision trees, and significantly outperformed algorithms with objective Trap problem with different number of variables
standard variation operators on problems that require effective n ∈ {30, 50, 100}. Equation (9) defines the notation of the
linkage learning.
bi-Trap function.
All MO-EDAs covered in [17] are Pareto-based and they
use concepts from algorithms such as NSGAII and SPEA2
[51]. Since, in the past few years, the MOEA/D framework
bi-T rap(x) = (ftrap5 (x), finv_trap5 (x))
(9)
has been one of the major frameworks to design MOEAs [5],
subject to x ∈ {0, 1}n
incorporating probabilistic graphical models into MOEA/D
seems to be a promising technique to solve scalable deceptive
A. Performance metrics
multi-objective problems.
Martins et al. [43] proposed a new approach for solvThe true PF from bi-Trap is known. Therefore, two perforing decomposable deceptive multi-objective problems. The mance metrics are used to evaluate the algorithms performance:
MO-EDA, called moφGA, uses a probabilistic model based (i) inverted generational distance metric (IGD) [54], and (ii)
on a phylogenetic tree. The moφGA was tested on the the number of true Pareto optimal solutions found by the
multi-objective deceptive functions ftrap5 and finv_trap5 . The algorithm through the generations. The stop condition for all
moφGA outperformed mohBOA in terms of number of func- the algorithms is 5n generations.
tion evaluations to achieve the exact PF, specially when the
Let P ∗ be a set of uniformly distributed Pareto optimal
problem in question increased in size.
solutions along the true P F in the objectives. P be an
A question discussed in [43] is: if such probabilistic model approximated set to the true P F obtained by an algorithm.
can identify the correct correlation between the variables of
IGD-metric [54]: The inverted generational distance from
a problem, the combination of improbable values of variables P ∗ to P is defined as
can be avoided. However, as the model becomes more exPv∈P ∗
d(v, P )
∗
pressive, the computational cost of incurred by the algorithm
(10)
IGD(P , P ) =
|P ∗ |
to build the model also grows. Thus, there is a trade-off
between the efficiency of the algorithm for building models where d(v, P ) is the minimum Euclidean distance between v
and its accuracy. In our proposal, at each generation, N and the solutions in P . If |P ∗ | is large enough IGD(P ∗ , P )
probabilistic graphical models are kept. Therefore, a higher could measure both convergence and diversity of P in a sense.
number of subproblems N can be a drawback for the proposal The lower the IGD(P ∗ , P ), the better the approximation of
in terms of efficiency (time consuming). However, if the P ∗ to the true PF.
proposed MOEA/D-GM has an adequate commitment between
Number of true Pareto optimal solutions: The number of
the efficiency and accuracy (approximated PF) then it is Pareto solutions that composes P that belongs to the P ∗ is
expected to behave satisfactorily.
defined as
|P + | = |P ∗ ∩ P |
(11)
D. Contributions with respect to previous work
We summarize some of the main contributions of our work where |P + | is the cardinality of P + .
with respect to the related research.
Also, we use the statistical test Kruskall-Wallis [55] to rank
• We use, for the first time, a probabilistic graphical model
the algorithms according to the results obtained by the metrics.
within MOEA/D to solve combinatorial MOPs. In this If two or more algorithms achieve the same rank, it means that
case, the previous MOEA/Ds that incorporate probabilis- there is no significant difference between them.
tic models cover only univariate models.
• We investigate a particular class of problems (deceptive
B. Parameters Settings
MOPs) for which there exist extensive evidence about the
The following parameters settings were used in the expericonvenience of using probabilistic graphical models.
mental
study:
• We introduced in MOEA/D the class of probabilistic
models where the structure of the model is learned from
1) Number of subproblems (N ): As in most of MOEA/D
the neighborhood of each solution.
algorithms proposed in the literature [2], [6], [5], the
• We investigate the question of how the problem innumber of subproblems N and their correspondent
weight vectors λ1 , . . . , λN are controlled by a parameter
teractions are kept in the scalarized subproblems and
how these interactions are translated to the probabilistic
H, which makes a wide spread distribution of the weight
m−1
models.
vectors according to N = CH+m−1
. Thus, for the
8
TABLE I: Results from the average IGD and average number of true Pareto optimal solutions |P + | computed form the 30 runs for each
combination (problem size n× Algorithm × preserving sampling). The column |P ∗ | is the total number of true Pareto solutions for each
problem n.
|P ∗ |
n
30
50
100
7
11
21
MOEA/D-GA
Average number of true Pareto optimal solutions |P + |
diversity preserving sampling (ds)
standard sampling
MOEA/D-PBIL
MOEA/D-UMDA
MOEA/D-Tree
MOEA/D-GA
MOEA/D-PBIL
MOEA/D-UMDA
MOEA/D-Tree
4.167
3.867
3.3
3.034
2.567
2.534
2.767
3.034
3.5
6.067
5.134
2.967
6.267
6.634
5.334
4.034
2.867
2.8
4.134
3.2
3.433
6.9
9.5
9.367
1.258
3.4
7.473
1.246
3.158
6.726
0.053
0.501
3.359
average IGD measure IGD(P ∗ , P )
30
50
100
7
11
21
1.076
2.174
4.757
1.764
3.585
7.663
1.853
3.229
6.38
0.418
1.657
4.815
0.371
1.308
4.142
TABLE II: Kruskall Wallis statistical ranking test according to the IGD and |P + | for each combination (problem size n× Algorithm ×
preserving sampling). The first value is the average rank over the 240 runs (i.e., 30 runs × 8 algorithms). The second value (in brackets) is
the final ranking from 8 to 1. If two or more ranks are the same, it means that there is no significant difference between them.
Kruskall-Wallis ranking - number of true Pareto optimal solutions |P + |
n
MOEA/D-GA
standard sampling
MOEA/D-PBIL
MOEA/D-UMDA
30
50
100
139.37 (5.50)
126.48 (5.00)
136.43 (5.50)
184.88 (6.00)
188.30 (6.50)
178.67 (6.00)
195.72 (6.50)
165.55 (6.00)
120.37 (5.00)
30
50
100
134.13 (5.50)
106.88 (3.50)
80.30 (2.50)
187.05 (6.00)
194.43 (6.50)
204.03 (7.00)
195.22 (6.50)
169.37 (6.50)
145.72 (6.00)
MOEA/D-Tree
MOEA/D-GA
66.25 (2.00)
88.15 (3.00)
151.77 (5.50)
59.43 (2.00)
52.65 (2.00)
61.58 (2.00)
diversity preserving sampling (ds)
MOEA/D-PBIL
MOEA/D-UMDA
MOEA/D-Tree
146.68 (6.00)
171.40 (6.00)
162.40 (5.50)
141.87 (6.00)
153.43 (6.00)
128.37 (5.50)
29.80 (2.00)
18.03 (1.50)
24.42 (1.50)
148.67 (6.00)
179.35 (6.50)
195.37 (6.50)
144.38 (6.00)
160.17 (6.00)
162.12 (6.50)
29.80 (2.00)
21.00 (2.00)
38.93 (2.50)
Kruskall-Wallis ranking - IGD measure
65.32 (2.00)
74.47 (2.50)
82.08 (2.50)
59.43 (2.00)
58.33 (2.50)
55.45 (2.50)
10
6
8
Average true Pareto solutions
Average true Pareto solutions
9
7
6
5
4
3
2
1
MOEA/D-Tree standard sampling
MOEA/D-Tree diversity preserving sampling
0
0
50
100
150
200
250
300
350
400
5
4
3
2
1
MOEA/D-GA standard sampling
MOEA/D-GA diversity preserving sampling
0
450
500
Generations
0
50
100
150
200
250
300
350
400
450
500
Generations
(a) MOEA/D-Tree
(b) MOEA/D-GA
+
Fig. 1: Average true Pareto optimal solutions |P | over the generations. Each chart shows a comparison between the standard
sampling and the diversity preserving sampling using the same algorithm.
2)
3)
4)
5)
bi-objective problem, we set H = 200, consequently
N = 201.
Neighborhood size (Ts ): As the number of
selected solutions is crucial for EDAs, we
test a range of neighborhood size values
Ts = (1, 5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 150).
Also, as in [3], Ts = Tr .
Maximal number of replacements by a new solution
(nr ): As in [6], nr = 2.
Scalarization function: We have applied both Weighted
Sum and Tchebycheff approaches. As both achieve very
similar results, only the results with Tchebycheff are
presented in this paper.
Maximum number of tries in diversity preserving sam-
pling procedure: We set this parameter with the same
value of the neighborhood size Ts .
6) Genetic operators from MOEA/D-GA: uniform crossover
and mutation rate= n1 .
7) PBIL learning rate: α = 0.05.
Each combination of (algorithm × parameters setting) is
independently run 30 times for each problem size n.
C. Comparison of the different variants of MOEA/D-GM for
the bi-Trap5
First, the algorithms are evaluated using a fixed neighborhood size value (Ts = 20). Table I presents the average values
of the IGD and the number of true Pareto solutions |P + |
computed using the approximated PF found by the algorithms.
10
10
9
9
8
8
Average true Pareto solutions
Average true Pareto solutions
9
7
6
5
4
3
2
MOEA/D-GA
MOEA/D-PBIL
MOEA/D-UMDA
MOEA/D-Tree
1
0
0
50
100
150
200
7
6
5
4
3
2
MOEA/D-GA
MOEA/D-PBIL
MOEA/D-UMDA
MOEA/D-Tree
1
0
250
0
50
100
150
200
Generations
250
300
350
400
450
500
Generations
(a) n = 50
(b) n = 100
+
Fig. 2: Average true Pareto solutions |P | through the generations for problem size 50 and 100 with Ts = 20
MOEA/D-GA
MOEA/D-PBIL
MOEA/D-UMDA
MOEA/D-Tree
MOEA/D-GA
MOEA/D-PBIL
MOEA/D-UMDA
MOEA/D-Tree
5
14
4.5
12
3.5
Average IGD
Average IGD
4
3
2.5
2
1.5
10
8
6
4
1
2
0.5
0
0
1
5
10
20
30
40
50
60
70
80
90 100 150
1
5
10
20
30
40
50
60
70
Neighborhood Size
Neighborhood Size
(a) n = 50
(b) n = 100
80
90 100 150
Fig. 3: Average IGD with different neighborhood sizes for selection Ts ∈ {1, 5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 150}
Table II presents the results of the Kruskall-Wallis test (at explanation for this behavior is a potential advantageous effect
5% significance level) applied to the results obtained by the introduced by the clustering of the solutions determined by the
algorithms (the same results summarized in Table I). The MOEA/D framework. This benefit would be independent of
best rank algorithm(s) is(are) shown in bold. Figure 1 and 2 the type of models used to represent the solutions. Grouping
show the behavior of the algorithms throughout the generations similar solutions in a neighborhood may allow the univariate
according to the average |P + | obtained.
algorithm to produce some global optima, even if the functions
From the analysis of the results, we can extract the following are deceptive.
conclusions:
According to Table II, the MOEA/D-Tree that uses the D. Influence of the neighborhood size selection Ts for the
diversity preserving mechanism (MOEA/D-Tree-ds) is the learning
only algorithm that has achieved the best rank in all the
The neighborhood size has a direct impact in the search
cases according to both indicators. Moreover, the diversity ability of the MOEA/Ds, which can balance convergence and
preserving sampling has improved the behavior of all the diversity for the target problem. Figure 3 presents the IGD
algorithms (mainly the MOEA/D-GA and MOEA/D-Tree) in values obtained with different neighborhood sizes (Ts ).
terms of the quality of the approximation to the true PF. Figure
As we mentioned before, the number of selected solutions
1 confirms these results. Therefore, the diversity preserving has an impact for most of the EDAs. A multi-variate EDA
sampling has a positive effect in the algorithms for the bi- needs a large set of selected solutions to be able to learn
Trap5 problem, i.e., the algorithms are able to achieve a dependencies between the decision variables [30]. This fact
more diverse set of solutions. Additionally, Figure 2 shows may explain (see Figure 3) that the differences between the
that MOEA/D-Tree-ds achieves better results than the other algorithms are lower as the neighborhood size becomes smaller
algorithms from the first generations.
(e.g., Ts = 5). A remarkable point in these results is that, a
Moreover, all the algorithms can find, at least, one global small Ts is better to solve bi-Trap5 for the algorithms except
optimal solution, i.e., a true Pareto solution. One possible for MOEA/D-Tree-ds. MOEA/D-Tree-ds achieves good results
10
50
3000
’extreme0d70.dat.txt’ matrix
2500
40
50
3000
"extreme0d70.dat.txt" matrix
2500
40
2000
30
2000
30
1500
20
1500
20
1000
10
500
0
0
0
10
20
30
40
1000
10
500
0
50
0
0
(a) Ts = 20, subproblem i = 1
10
20
30
40
50
(b) Ts = 70, subproblem i = 1
50
1800
50
4000
1600
"extreme200d20.dat.txt" matrix
40
1400
"extreme200d70.dat.txt" matrix
3500
40
3000
1200
30
1000
800
20
2500
30
2000
20
1500
600
400
10
1000
10
500
200
0
0
0
10
20
30
40
0
50
0
0
(c) Ts = 20, subproblem i = 201
10
20
30
40
50
(d) Ts = 70, subproblem i = 201
50
3000
"extreme100d20.dat.txt" matrix
2500
40
50
3000
"extreme100d70.dat.txt" matrix
2500
40
2000
30
2000
30
1500
20
1500
20
1000
10
500
0
0
0
10
20
30
40
50
(e) Ts = 20, subproblem i = 100
1000
10
500
0
0
0
10
20
30
40
50
(f) Ts = 70, subproblem i = 100
Fig. 4: Frequency matrices (heat maps) learned by the subproblems i = (1, 100, 201) from the MOEA/D-Tree-ds for problem
size n = 50.
with a large neighborhood size (e.g., 60, 70 and 80). Even
if the neighborhood size has an important influence in the
behavior of the algorithm, this parameter can not be seen in
isolation of other parameters that also influence the behavior
of the algorithm.
E. Analyzing the structure of problems as captured by the tree
model
One of the main benefits of EDAs is their capacity to reveal
a priori unknown information about the problem structure.
Although this question has been extensively studied [40], [56],
[57] in the single-objective domain, the analysis in the multiobjective domain are still few [19], [58], [59]. Therefore,
a relevant question was to determine to what extent is the
structure of the problem captured by the probabilistic models
used in MOEA/D-GM. This is a relevant question since there
is no clue about the types of interactions that could be captured
from models learned for scalarized functions. In this section,
the structures learned from MOEA/D-Tree-ds while solving
different subproblems are investigated.
In each generation, for each subproblem, a tree model is
built according to the bi-variate probabilities obtained from
its selected population. We can represent the tree model
as a matrix Mn×n , where each position Mjk represents a
relationship (pairwise) between two variables j, k. Mjk = 1 if
j is the parent of k in tree model learned, otherwise Mjk = 0.
Figure 4 represents the merge frequency matrices obtained
by the 30 runs. The frequencies are represented using heat
11
maps, where lighter colors indicate a higher frequency. We the structure of the objective functions. The analysis has
have plotted the merge matrices learned from the extreme been conducted for MOEA/D-Tree-ds considering the strucsubproblems and the middle subproblem i.e., i = (1, 100, 201) tures learned with different neighborhood sizes and different
using the problem size n = 50. Because of the results of the subproblems. We have found that even if relatively small
previous section, the frequency matrices for two neighborhood neighborhoods are used (in comparison with the standard
sizes are presented, Ts = (20, 70).
population sizes used in EDAs), the models are able to capture
The matrices clearly show a strong relationship between the the interactions of the functions. One potential application of
subsets of variables "building blocks" of size 5, which shows this finding, is that, we could reuse or transfer models between
that the algorithm is able to learn the structure of the Trap5 subproblems, in a similar way to the application of structural
functions. Notice that, for neighborhood size 70, the MOEA/D- transfer between related problems [60].
Tree-ds was able to capture more accurate structures, which
The scalability of MOEA/D-Tree-ds for the bi-Trap5 probexalts the good results found in accordance with the IGD lem and other multi-objective deceptive functions should
metric. This can be explained by the fact that a higher be investigated. The "most appropriate" model for bi-Trap5
population size reduces the number of spurious correlations should be able to learn higher order interactions (of order
learned from the data.
4). Therefore, PGMs based on Bayesian or Markov networks
Moreover, analyzing the different scalar subproblems, we could be of application in this case.
can see that, the algorithm is able to learn a structure even for
The directions for future work are: (i) Conceive strategies
the middle scalar subproblem i = 100, where the two conflict to avoid learning a model for each subproblem. This would
objectives functions ftrap5 and finv_trap5 compete in every improve the results in terms of computational cost; (ii) Use
partition (building block) of the decomposable problem.
of the most probable configurations of the model to speed up
convergence; (iii) Consider the application of hybrid schemes
IX. C ONCLUSIONS AND FUTURE WORK
incorporating local search that take advantage of the informaIn this paper, a novel and general MOEA/D framework able tion learned by the models; (iv) Evaluate MOEA/D-GM on
to instantiate probabilistic graphical models named MOEA/D- other benchmarks of deceptive MOPs.
GM has been introduced. PGMs are used to obtain a more comprehensible representation of a search space. Consequently,
ACKNOWLEDGMENTS
the algorithms that incorporate PGMs can provide a model
This work has received support by CNPq (Productivity
expressing the regularities of the problem structure as well Grant Nos. 305986/2012-0 and Program Science Without Boras the final solutions. The PGM investigated in this paper ders Nos.: 400125/2014-5), by CAPES (Brazil Government),
takes into account the interactions between the variables by by IT-609-13 program (Basque Government) and TIN2013learning a maximum weight spanning tree from the bi-variate 41272P (Spanish Ministry of Science and Innovation).
probabilities distributions. An experimental study on a biobjective version of a well known deceptive function (biR EFERENCES
Trap5) was conducted. In terms of accuracy (approximating
[1]
C.
C.
Coello,
G.
Lamont,
and D. van Veldhuizen, Evolutionary Algoof the true PF), the results have shown that the instantiation
rithms for Solving Multi-Objective Problems., 2nd ed., ser. Genetic and
from MOEA/D-GM, called MOEA/D-Tree, is significantly
Evolutionary Computation. Berlin, Heidelberg: Springer, 2007.
[2] Q. Zhang and H. Li, “MOEA/D: A multiobjective evolutionary algorithm
better than MOEA/D that uses univariate EDAs and traditional
based on decomposition,” IEEE Transactions on Evolutionary Compugenetic operators.
tation, vol. 11, no. 6, pp. 712–731, 2007.
Moreover, other enhancement has been introduced in
[3] Z. Wang, Q. Zhang, G. Maoguo, and Z. Aimin, “A Replacement Strategy
for Balalancing Convergence and Diversity in MOEA/D,” in Proceedings
MOEA/D. A new simple but effective mechanism of sampling
of the Congress on Evolutionay Computation (CEC). IEEE, 2014, pp.
has been proposed, called diversity preserving sampling. Since
2132–2139.
sampling solutions already present in the neighbor solutions
[4] K. Deb and H. Jain, “An evolutionary many-objective optimization algorithm using reference-point-based nondominated sorting approach, part I:
can have a detrimental effect in terms of convergence, the
solving problems with box constraints,” IEEE Transaction Evolutionary
diversity preserving sampling procedure tries generating those
Computation, vol. 18, no. 4, pp. 577–601, 2014.
solutions that are different from the parent solutions. Accord- [5] K. Li, K. Deb, Q. Zhang, and S. Kwong, “An evolutionary manyobjective optimization algorithm based on dominance and decomposiing to both performance indicators, all the algorithms in the
tion.” IEEE Trans. Evolutionary Computation, no. 5, pp. 694–716, 2015.
comparison have improved their results in terms of diversity
[6] H. Li and Q. Zhang, “Multiobjective Optimization Problems with
to the approximation PF.
Complicated Pareto Sets, MOEA/D and NSGA-II,” IEEE Transaction
on Evolutionary Computation, pp. 284–302, 2008.
An analysis of the influence of the neighborhood size on
the behavior of the algorithms were conducted. In general, [7] H. Ishibuchi, N. Akedo, and Y. Nojima, “A Study on the Specification
of a Scalarizing Function in MOEA/D for Many-Objective Knapsack
increasing the neighborhood size has a detrimental effect.
Problems.” in Proceedings of the 7th International Learning and Intelligent Optimization Conference (LION), ser. Lecture Notes in Computer
Although, this is not always the case for MOEA/D-Tree.
Science, vol. 7997. Springer, 2013, pp. 231–246.
Also, independent of the type of models used to represent
[8] Y.-y. Tan, Y.-c. Jiao, H. Li, and X.-k. Wang, “MOEA/D+ uniform
the solutions, grouping similar solutions in a neighborhood
design: A new version of MOEA/D for optimization problems with
many objectives,” Computers & Operations Research, vol. 40, no. 6,
may allow to produce some global optima, even if the funcpp. 1648–1660, 2013.
tions are deceptive.
[9] A. J. Nebro and J. J. Durillo, “A study of the parallelization of the
Finally, we have also investigated for the first time to what
multi-objective metaheuristic MOEA/D,” in Learning and Intelligent
Optimization. Springer, 2010, pp. 303–317.
extent the models learned by MOEA/D-GMs can capture
12
[10] M. Pilat and R. Neruda, “Incorporating user preferences in MOEA/D
through the coevolution of weights,” in Proceedings of the Genetic and
Evolutionary Computation Conference (GECCO). ACM, 2015, pp. 727–
734.
[11] Y. Qi, X. Ma, F. Liu, L. Jiao, J. Sun, and J. Wu, “MOEA/D with adaptive
weight adjustment,” Evolutionary computation, vol. 22, no. 2, pp. 231–
264, 2014.
[12] A. Alhindi and Q. Zhang, “MOEA/D with tabu search for multiobjective
permutation flow shop scheduling problems,” in Proceedings of the IEEE
Congress on Evolutionary Computation, CEC 2014, Beijing, China, July
6-11, 2014. IEEE, 2014, pp. 1155–1164.
[13] L. Ke, Q. Zhang, and R. Battiti, “Hybridization of decomposition
and local search for multiobjective optimization,” IEEE T. Cybernetics,
vol. 44, no. 10, pp. 1808–1820, 2014.
[14] P. Larrañaga, H. Karshenas, C. Bielza, and R. Santana, “A review on
probabilistic graphical models in evolutionary computation,” Journal of
Heuristics, vol. 18, no. 5, pp. 795–819, 2012.
[15] J. A. Lozano, P. Larrañaga, I. Inza, and E. Bengoetxea, Eds., Towards a
New Evolutionary Computation: Advances on Estimation of Distribution
Algorithms. Springer, 2006.
[16] H. Mühlenbein and G. Paaß, “From recombination of genes to the
estimation of distributions I. Binary parameters,” in Parallel Problem
Solving from Nature - PPSN IV, ser. Lectures Notes in Computer Science,
H.-M. Voigt, W. Ebeling, I. Rechenberg, and H.-P. Schwefel, Eds., vol.
1141. Berlin: Springer, 1996, pp. 178–187.
[17] M. Pelikan, K. Sastry, and D. E. Goldberg, “Multiobjective estimation of
distribution algorithms,” in Scalable Optimization via Probabilistic Modeling: From Algorithms to Applications, ser. Studies in Computational
Intelligence, M. Pelikan, K. Sastry, and E. Cantú-Paz, Eds. Springer,
2006, pp. 223–248.
[18] L. Martí, J. García, A. Berlanga, and J. M. Molina, “Multi-objective
optimization with an adaptive resonance theory-based estimation of
distribution algorithm: A comparative study.” in LION, ser. Lecture
Notes in Computer Science, C. A. C. Coello, Ed., vol. 6683. Springer,
2011, pp. 458–472.
[19] H. Karshenas, R. Santana, C. Bielza, and P. Larrañaga, “Multi-objective
optimization based on joint probabilistic modeling of objectives and
variables,” IEEE Transactions on Evolutionary Computation, vol. 18,
no. 4, pp. 519–542, 2014.
[20] V. A. Shim, K. C. Tan, and K. K. Tan, “A Hybrid Estimation of Distribution Algorithm for Solving the Multi-objective Multiple Traveling
Salesman Problem.” in Proceedings of the Congress on Evolutionary
Computation (CEC). IEEE, 2012, pp. 1–8.
[21] A. Zhou, F. Gao, and G. Zhang, “A decomposition based estimation of
distribution algorithm for multiobjective traveling salesman problems.”
Computers and Mathematics with Applications, vol. 66, no. 10, pp. 1857–
1868, 2013.
[22] B. Wang, X. Hua, and Y. Yuan, “Scale Adaptive Reproduction Operator
for Decomposition based Estimation of Distribution Algorithm,” in Proceedings of the Congress on Evolutionary Computation (CEC). IEEE,
2015, pp. 1–8.
[23] I. Giagkiozis, R. C. Purshouse, and P. J. Fleming, “Generalized decomposition and cross entropy methods for many-objective optimization.”
Inf. Sci., vol. 282, pp. 363–387, 2014.
[24] S. Zapotecas-Martínez, B. Derbel, A. Liefooghe, D. Brockhoff,
H. Aguirre, and K. Tanaka, “Injecting CMA-ES into MOEA/D,” in
Proceedings of the Genetic and Evolutionary Computation Conference
(GECCO). ACM, 2015, pp. 783–790.
[25] I. Giagkiozis, R. C. Purshouse, and P. J. Fleming, “Generalized decomposition,” in 7th International ConferenceEvolutionary on Multi-Criterion
Optimization EMO, ser. Lecture Notes in Computer Science, vol. 7811.
Springer, 2013, pp. 428–442.
[26] M. Pelikan, “Analysis of epistasis correlation on NK landscapes with
nearest neighbor interactions,” in Proceedings of the 13th conference
on Genetic and evolutionary computation (GECCO). ACM, 2011, pp.
1013–1020.
[27] H. Mühlenbein, “The equation for response to selection and its use for
prediction,” Evolutionary Computation, vol. 5, no. 3, pp. 303–346, 1998.
[28] S. Baluja, “Population-based incremental learning: A method for integrating genetic search based function optimization and competitive
learning,” Carnegie Mellon University, Pittsburgh, PA, Tech. Rep. CMUCS-94-163, 1994.
[29] M. Zangari, R. Santana, A. Pozo, and A. Mendiburu, “Not all PBILs are
the same: Unveiling the different learning mechanisms of PBIL variants,”
2015, submmitted to publication.
[30] H. Mühlenbein, T. Mahnig, and A. Ochoa, “Schemata, distributions and
graphical models in evolutionary optimization,” Journal of Heuristics,
vol. 5, no. 2, pp. 213–247, 1999.
[31] M. Pelikan and H. Mühlenbein, “The bivariate marginal distribution
algorithm,” in Advances in Soft Computing - Engineering Design and
Manufacturing, R. Roy, T. Furuhashi, and P. Chawdhry, Eds. London:
Springer, 1999, pp. 521–535.
[32] R. Santana, A. Ochoa, and M. R. Soto, “The mixture of trees factorized
distribution algorithm,” in Proceedings of the Genetic and Evolutionary Computation Conference GECCO. San Francisco, CA: Morgan
Kaufmann Publishers, 2001, pp. 543–550.
[33] S. Baluja and S. Davies, “Using optimal dependency-trees for combinatorial optimization: Learning the structure of the search space,” in
Proceedings of the 14th International Conference on Machine Learning,
D. H. Fisher, Ed., 1997, pp. 30–38.
[34] T. Mahnig and H. Muhlenbein, “Optimal mutation rate using bayesian
priors for estimation of distribution algorithms.” in SAGA, ser. Lecture
Notes in Computer Science, K. Steinhofel, Ed., vol. 2264. Springer,
2001, pp. 33–48.
[35] M. Pelikan, D. E. Goldberg, and E. Cantú-Paz, “BOA: The Bayesian
optimization algorithm,” in Proceedings of the Genetic and Evolutionary
Computation Conference (GECCO), vol. I.
Orlando, FL: Morgan
Kaufmann Publishers, San Francisco, CA, 1999, pp. 525–532.
[36] S. Shakya and R. Santana, Eds., Markov Networks in Evolutionary
Computation. Springer, 2012.
[37] D. E. Goldberg, “Simple genetic algorithms and the minimal, deceptive
problem,” in Genetic Algorithms and Simulated Annealing, L. Davis, Ed.
London, UK: Pitman Publishing, 1987, pp. 74–88.
[38] K. Deb and D. E. Goldberg, “Analyzing deception in trap functions,”
University of Illinois at Urbana-Champaign, Illinois Genetic Algorithms
Laboratory, Urbana, IL, IlliGAL Report 91009, 1991.
[39] K. Sastry, H. A. Abbass, D. E. Goldberg, and D. Johnson, “Substructural niching in estimation of distribution algorithms,” in Proceedings of the Conference on Genetic and Evolutionary Computation
(GECCO). ACM, 2005, pp. 671–678.
[40] C. Echegoyen, J. A. Lozano, R. Santana, and P. Larrañaga, “Exact
Bayesian network learning in estimation of distribution algorithms,” in
Proceedings of the Congress on Evolutionary Computation CEC. IEEE
Press, 2007, pp. 1051–1058.
[41] C. Echegoyen, A. Mendiburu, R. Santana, and J. A. Lozano, “Toward
understanding EDAs based on Bayesian networks through a quantitative
analysis,” IEEE Transactions on Evolutionary Computation, vol. 16,
no. 2, pp. 173–189, 2012.
[42] M. Pelikan, K. Sastry, and D. E. Goldberg, “Multiobjective hBOA,
clustering and scalability,” University of Illinois at Urbana-Champaign,
Illinois Genetic Algorithms Laboratory, Urbana, IL, IlliGAL Report No.
2005005, February 2005.
[43] J. P. Martins, A. H. M. Soares, D. V. Vargas, and A. C. B. Delbem,
“Multi-objective phylogenetic algorithm: Solving multi-objective decomposable deceptive problems,” in Evolutionary Multi-Criterion Optimization. Springer, 2011, pp. 285–297.
[44] M. Pelikan, Hierarchical Bayesian Optimization Algorithm. Toward a
New Generation of Evolutionary Algorithms, ser. Studies in Fuzziness
and Soft Computing. Springer, 2005, vol. 170.
[45] J. Ceberio, E. Irurozki, A. Mendiburu, and J. A. Lozano, “A review on
estimation of distribution algorithms in permutation-based combinatorial
optimization problems,” Progress in Artificial Intelligence, vol. 1, no. 1,
pp. 103–117, 2012.
[46] J. Ceberio, A. Mendiburu, and J. A. Lozano, “The Plackett-Luce ranking
model on permutation-based optimization problems,” in Proceedings of
the Congress on Evolutionary Computation (CEC), IEEE. IEEE press,
2013, pp. 494–501.
[47] Z. Botev, D. P. Kroese, R. Y. Rubinstein, P. L’Ecuyer et al., The Cross
Entropy Method: A Unified Approach To Combinatorial Optimization,
Monte-carlo Simulation. North-Holland Mathematical Library-Elsevier,
2013.
[48] Q. Zhang, A. Zhou, and Y. Jin, “RM-MEDA: A regularity model based
multiobjective estimation of distribution algorithm,” IEEE Transactions
on Evolutionary Computation, vol. 12, no. 1, pp. 41–63, 2008.
[49] N. Hansen and A. Ostermeier, “Adapting arbitrary normal mutation
distributions in evolution strategies: The covariance matrix adaptation,”
in Proceedings of the IEEE International Conference on Evolutionary
Computation, 1996, pp. 312–317.
[50] D. Thierens and P. A. Bosman, “Multi-objective optimization with
iterated density estimation evolutionary algorithms using mixture models,” in Evolutionary Computation and Probabilistic Graphical Models.
13
[51]
[52]
[53]
[54]
[55]
[56]
[57]
[58]
[59]
[60]
Proceedings of the Third Symposium on Adaptive Systems (ISAS-2001),
Cuba, Havana, Cuba, March 2001, pp. 129–136.
M. Laumanns and J. Ocenasek, “Bayesian optimization algorithm for
multi-objective optimization,” in Proceedings of the 7th Parallel Problem
Solving from Nature Conference (PPSN), ser. Lecture Notes in Computer
Science, vol. 2439. Springer, 2002, pp. 298–307.
N. Khan, “Bayesian optimization algorithms for multi-objective and
hierarchically difficult problems,” Master’s thesis, University of Illinois
at Urbana-Champaign, Illinois Genetic Algorithms Laboratory, Urbana,
IL, 2003.
K. Deb, L. Thiele, M. Laumanns, and E. Zitzler, “Scalable multiobjective optimization test problems,” in Proceedings of the Congress
on Evolutionary Computation (CEC), vol. 2. IEEE press, 2002, pp.
825–830.
E. Zitzler and L. Thiele, “Multiobjective evolutionary algorithms: A comparative case study and the strength pareto approach,” IEEE Transactions
on Evolutionary Computation, vol. 3, no. 4, pp. 257–271, 1999.
J. Derrac and S. Garcia, “A Practical Tutorial on the use of Nonparametric Statistical Tests as a Methodology for Comparing Evolutionary and
Swarm Intelligence Algorithms.” Swarm and Evolutionary Computation,
vol. 1, no. 1, pp. 3–18, 2011.
A. E. I. Brownlee, J. McCall, and M. Pelikan, “Influence of selection
on structure learning in Markov network EDAs: An empirical study,” in
Proceedings of the Genetic and Evolutionary Computation Conference
GECCO. ACM, 2012, pp. 249–256.
R. Santana, P. Larrañaga, and J. A. Lozano, “Protein folding in simplified
models with estimation of distribution algorithms,” IEEE Transactions
on Evolutionary Computation, vol. 12, no. 4, pp. 418–438, 2008.
G. Fritsche, A. Strickler, A. Pozo, and R. Santana, “Capturing relationships in multi-objective optimization,” in Proceedings of the Brazilian
Conference on Intelligent Systems (BRACIS 2015), Natal Brazil, 2015,
accepted for publication.
R. Santana, C. Bielza, J. A. Lozano, and P. Larrañaga, “Mining probabilistic models learned by EDAs in the optimization of multi-objective
problems,” in Proceedings of the Genetic and Evolutionary Computation
Conference GECCO. ACM, 2009, pp. 445–452.
R. Santana, A. Mendiburu, and J. A. Lozano, “Structural transfer
using EDAs: An application to multi-marker tagging SNP selection,” in
Proceedings of the 2012 Congress on Evolutionary Computation (CEC).
IEEE Press, 2012, pp. 3484–3491, best Paper Award of 2012 Congress
on Evolutionary Computation.
| 9cs.NE
|
Towards Automatic Resource Bound Analysis for
OCaml
Jan Hoffmann
Ankush Das
Carnegie Mellon University
Shu-Chun Weng
Yale University
November 2, 2016
Abstract
This article presents a resource analysis system for OCaml programs. This system automatically derives worst-case resource bounds for higher-order polymorphic programs with
user-defined inductive types. The technique is parametric in the resource and can derive
bounds for time, memory allocations and energy usage. The derived bounds are multivariate
resource polynomials which are functions of different size parameters that depend on the
standard OCaml types. Bound inference is fully automatic and reduced to a linear optimization problem that is passed to an off-the-shelf LP solver. Technically, the analysis system
is based on a novel multivariate automatic amortized resource analysis (AARA). It builds
on existing work on linear AARA for higher-order programs with user-defined inductive
types and on multivariate AARA for first-order programs with built-in lists and binary trees.
For the first time, it is possible to automatically derive polynomial bounds for higher-order
functions and polynomial bounds that depend on user-defined inductive types. Moreover,
the analysis handles programs with side effects and even outperforms the linear bound
inference of previous systems. At the same time, it preserves the expressivity and efficiency
of existing AARA techniques. The practicality of the analysis system is demonstrated with an
implementation and integration with Inria’s OCaml compiler. The implementation is used
to automatically derive resource bounds for 411 functions and 6018 lines of code derived
from OCaml libraries, the CompCert compiler, and implementations of textbook algorithms.
In a case study, the system infers bounds on the number of queries that are sent by OCaml
programs to DynamoDB, a commercial NoSQL cloud database service.
1 Introduction
The quality of software crucially depends on the amount of resources —such as time, memory,
and energy—that are required for its execution. Statically understanding and controlling the
resource usage of software continues to be a pressing issue in software development. Performance bugs are very common and among the bugs that are most difficult to detect [41, 50]
and large software systems are plagued by performance problems. Moreover, many security
vulnerabilities exploit the space and time usage of software [43, 21].
Developers would greatly profit from high-level resource-usage information in the specifications of software libraries and other interfaces, and from automatic warnings about potentially
high-resource usage during code review. Such information is particularly relevant in contexts of
mobile applications and cloud services, where resources are limited or resource usage is a major
cost factor.
1
2
Towards Automatic Resource Bound Analysis for OCaml
Recent years have seen fast progress in developing frameworks for statically reasoning
about the resource usage of programs. Many advanced techniques for imperative integer
programs apply abstract interpretation to generate numerical invariants. The obtained sizechange information forms the basis for the computation of actual bounds on loop iterations
and recursion depths; using counter instrumentation [27], ranking functions [6, 2, 15, 52],
recurrence relations [4, 1], and abstract interpretation itself [58, 18]. Automatic resource analysis
techniques for functional programs are based on sized types [54], recurrence relations [23],
term-rewriting [9], and amortized resource analysis [35, 42, 30, 51].
Despite major steps forward, there are still many obstacles to overcome to make resource
analysis technologies available to developers. On the one hand, typed functional programs are
particularly well-suited for automatic resource-bound analysis since the use of pattern matching
and recursion often results in a relatively regular code structure. Moreover, types provide
detailed information about the shape of data structures. On the other hand, existing automatic
techniques for higher-order programs can only infer linear bounds [54, 42]. Furthermore,
techniques that can derive polynomial bounds are limited to bounds that depend on predefined
lists and binary trees [33, 30] or integers [15, 52]. Finally, resource analyses for functional
programs have been implemented for custom languages that are not supported by mature tools
for compilation and development [35, 54, 42, 30, 51].
The goal of a long term research effort is to overcome these obstacles by developing Resource
Aware ML (RAML), a resource-aware version of the functional programming language OCaml.
RAML is based on an automatic amortized resource analysis (AARA) that derives multivariate
polynomials that are functions of the sizes of the inputs. In this paper, we report on three main
contributions that are part of this effort.
1. We present the first implementation of an AARA that is integrated with an industrialstrength compiler.
2. We develop the first automatic resource analysis system that infers multivariate polynomial bounds that depend on size parameters of complex user-defined data structures.
3. We present the first AARA that infers polynomial bounds for higher-order functions.
The techniques we develop are not tied to a particular resource but are parametric in the resource
of interest. RAML infers tight bounds for many complex example programs such as sorting
algorithms with complex comparison functions, Dijkstra’s single-source shortest-path algorithm,
and the most common higher-order functions such as (sequences) of nested maps, and folds.
The technique is naturally compositional, tracks size changes of data across function boundaries,
and can deal with amortization effects that arise, for instance, from the use of a functional queue.
Local inference rules generate linear constraints and reduce bound inference to off-the-shelf
linear program (LP) solving, despite deriving polynomial bounds.
To ensure compatibility with OCaml’s syntax, we reuse the parser and type inference engine
from Inria’s OCaml compiler [47]. We extract a type-annotated syntax tree to perform (resource
preserving) code transformations and the actual resource-bound analysis. To precisely model
the evaluation of OCaml, we introduce a novel operational semantics that makes the efficient
handling of function closures in Inria’s compiler explicit. The semantics is complemented by a
new type system that refines function types.
To express a wide range of bounds, we introduce a novel class of multivariate resource
polynomials that map data of a given type to a non-negative number. The set of multivariate
resource polynomials that is available for bound inference depends on the types of input data.
It can be parametric in integers, lengths of lists, or number of particular nodes in an inductive
2 Overview
3
data type. As a special case, a resource polynomial can contain conditional additive factors.
These novel multivariate resource polynomials are a substantial generalization of the resource
polynomials that have been previously defined for lists and binary trees [30]. To deal with
realistic OCaml code, we develop a novel multivariate AARA that handles higher-order functions.
To this end, we draw inspirations from multivariate AARA for first-order programs [30] and linear
AARA for higher-order programs [42]. However, our new solution is more than the combination
of existing techniques. For instance, we infer linear bounds for the curried append function for
lists, which has not been possible previously [42]. Moreover, we address specifics of Inria’s OCaml
compiler such as the evaluation order of function arguments to efficiently avoid function-closure
creation.
We performed experiments on more than 6018 lines of OCaml code. We still do not support
all language features of OCaml and it is thus not straightforward to automatically analyze
complete existing applications. However, the automatic analysis performs well on code that
only uses supported language features. For instance, we applied RAML to OCaml’s standard list
library list.ml: RAML automatically derives evaluation-step bounds for 47 of the 51 top-level
functions. All derived bounds are asymptotically tight.
It is also easy to develop and analyze real OCaml applications if we keep the current capabilities of the system in mind. In Section 9, we present a case study in which we automatically
bound the number of queries that an OCaml program issues to Amazon’s DynamoDB NoSQL
cloud database service. Such bounds are interesting since Amazon charges DynamoDB users
based on the number of queries made to a database.
Our experiments are easily reproducible: The source code of RAML, the OCaml code for the
experiments, and an easy-to-use interactive web interface are available online [29].
2 Overview
Before we describe the technical development, we give a short overview of the challenges and
achievements of our work.
Example Bound Analysis (Running Example). To demonstrate user interaction with RAML,
Figure 1 contains an example bound analysis. The OCaml code in Figure 1 will serve as a running
example in this article. The function abmap is a polymorphic map function for a user-defined
list that contains Acons and Bcons nodes. It takes two functions f and g as arguments and applies
f to data stored in the A-nodes and g to data stored in the B-nodes. The function asort takes a
comparison function and an A-B-list in which the A-nodes contain lists. It then uses quicksort
(the code of quicksort is also automatically analyzed and available online [29]) to sort the lists
in the A-nodes. The B-nodes are left unchanged. The function asort’ is a variation of asort that
raises an exception if it encounters a B-node in the list.
To derive a worst-case resource bound with RAML, the user needs to pick a maximal degree
of the search space of polynomials and a resource metric. In the example analysis in Figure 1
we picked degree 4 and the steps metric which counts the number of evaluation steps in the
big-step semantics. After 0.23 seconds, RAML reports a bound for each of the top-level functions.
The shown output is only an excerpt. In this case, all derived bounds are tight in the sense that
there are inputs for every size that exactly result in the reported number of evaluation steps.
In the derived bound, for abmap RAML assumes that the resource cost of f and g is 0.
So we get a linear bound. In the case of asort we derive a bound which is quadratic in the
maximal length of the lists that are stored in the A-nodes (22K + 13K 2 ) for every A-node in the
4
Towards Automatic Resource Bound Analysis for OCaml
type (’a,’b) ablist =
let
|
|
|
Acons of ’a * (’a,’b) ablist
| Bcons of ’b * (’a,’b) ablist
| Nil
rec abmap f g abs = match abs with
Acons (a,abs’) → Acons(f a, abmap f g abs’)
Bcons (b,abs’) → Bcons(g b, abmap f g abs’)
Nil → Nil
let asort gt abs = abmap (quicksort gt) (fun x → x) abs
let asort’ gt abs = abmap (quicksort gt) (fun _ → raise Inv_arg) abs
let btick = abmap (fun a → a) (fun b → Raml.tick 2.5; b)
Excerpt of the RAML output for analyzing evaluation steps (0.23s run time):
Simplified bound for abmap:
3.00 + 12.00*L + 12.00*N
Simplified bound for asort:
11.00 + 22.00*K*N + 13.00*K^2*N + 13.00*L + 15.00*N
Simplified bound asort’:
13.00 + 22.00*K*N + 13.00*K^2*N + 15.00*N
where
L is the number of Bcons-nodes of the 2nd (3rd) component of the argument
N is the number of Acons-nodes of the 2nd (3rd) component of the argument
K is the maximal number of ::-nodes in the Acons-nodes of the 2nd component
of the argument
Figure 1: The function abmap will serve as a running example in this article. When
deriving the linear bound for abmap, we assume that the higher-order arguments f
and g have no resource consumption. If abmap is applied to concrete functions, like in
asort and asort’ then the cost of the concrete application is bounded. Only the Acons
node contribute to the cubic cost in the bound of asort. Moreover, the number of
Bcons nodes do not contribute to the linear factor in asort’.
list ((22K +13K 2 )N ) plus an additional linear factor that also depends on the number of B-nodes
that are simply traversed (13L + 15N ). For asort’ this linear factor only depends on the number
of A-nodes: RAML automatically deduces that the traversal is aborted in case we encounter a
B-node.
The tick metric can be used to derive bounds on user defined metrics. An instructive example
is the function btick. With the tick metric, RAML derives the bound 2.5L where L is the number
of B-nodes in the argument list. This is a tight bound on the sum of “ticks” that are executed in
an evaluation of btick. Ticks can also be negative to express that resources become available.
Note that RAML does not make guarantees about the precision of the derived bounds. Since
an evaluation-step bound proves termination, bound analysis is an undecidable problem. So
there are many functions for which RAML cannot derive a bound either because no (polynomial)
bound exists or the analysis is not able to find a bound. In these cases, RAML terminates with a
message like “A bound for abmap could not be derived.”
2 Overview
5
Currying and Function Closures. Currying and function closures pose a challenge to automatic resource analysis systems that has not been addressed in the past. To see why, assume that
we want to design a type system to verify resource usage. Now consider for example the curried
append function which has the type append : α list → α list → α list in OCaml. At first glance,
we might say that the time complexity of append is O(n) if n is the length of the first argument.
But a closer inspection of the definition of append reveals that this is a gross simplification. In
fact, the complexity of the partial function call app_par = append ` is constant. Moreover, the
complexity of the function app_par is linear—not in the length of the argument but in the length
of the list ` that is captured in the function closure. We are not aware of any existing approach
that can automatically derive a worst-case time bound for the curried append function. For
example, previous AARA systems would fail without deriving a bound [42, 30].
In general, we have to describe the resource consumption of a curried function f : A 1 → · · · →
A n → A with n expressions c i (a 1 , . . . , a i ) such that c i describes the complexity of the computation
that takes place after f is applied to i arguments a 1 , . . . , a i . In Inria’s OCaml implementation,
the situation is even more complex since the resource usage (time and space) depends on
how a function is used at its call sites. If append is partially applied to one argument then a
function closure is created as expected. However—and this is one of the reasons for OCaml’s
great performance—if append is applied to both of its arguments at the same time then the
intermediate closure is not created and the performance of the function is even better than that
of the curried version since we do not have to create a pair before the application.
To model the resource usage of curried functions accurately we refine function types to
capture how functions are used at their call sites. For example, append can have both of the
following types
α list → α list → α list and [α list, α list] → α list .
The first type implies that the function is partially applied and the second type implies that the
function is applied to both arguments at the same time. Of course, it is possible that the function
has both types (technically we achieve this using let polymorphism). For the second type, our
system automatically derives tight time and space bounds that are linear in the first argument.
However, our system fails to derive a bound for the first type. The reason is that we made the
design decision to not derive bounds that asymptotically depend on data captured in function
closures to keep the complexity of the system at a manageable level.
Fortunately, append belongs to a large set of OCaml functions in the standard library that
are defined in the form let rec f x y z = e. If such a function is partially applied, the only computation that happens is the creation of a closure. As a result, eta expansion does not change
the resource behavior of programs. This means for example that we can safely replace the
expression let app_par = append ` in e with the expression let app_par x = append ` x in e prior
to the analysis. Consequently, we can always use the type [α list, α list] → α list of append that
we can successfully analyze.
The conditions under which functions can be analyzed might look complex at first but they
can be boiled down to a simple principle:
The worst-case resource usage of a function must be expressible as a function of the
sizes of its arguments.
Higher-Order Arguments. The other main challenge with higher-order resource analysis is
functions with higher-order arguments. To a large extent, this problem has been successfully
solved for linear resource bounds in previous work [42]. Basically, the higher-order case is
6
Towards Automatic Resource Bound Analysis for OCaml
reduced to the first-order case if the higher-order arguments are available. It is not necessary to
reanalyze such higher-order functions for every call site since we can abstract the resource usage
with a constraint system that has holes for the constraints of the function arguments. However,
a presentation of the system in such a way mixes type checking with the constraint-based type
inference. Therefore, we chose to present the analysis system in a more declarative way in which
the bound of a function with higher-order arguments is derived with respect to a given set of
resource behaviors of the argument functions.
A concrete advantage of our declarative view is that we can derive a meaningful type for a
function like map for lists even when the higher-order argument is not available. The function
map can have the following types.
(α → β) → α list → β list
[α → β, α list] → β list
Unlike append, the resource usage of map does not depend on the size of the first argument. So
both types are equivalent in our system except for the cost of creating an intermediate closure.
If the higher-order argument is not available then previous systems [42] produce a constraint
system that is not meaningful to a user. An innovation in this work is that we are also able to
report a meaningful resource bound for map if the arguments are not available. To this end,
we assume that the argument function does not consume resources. For example, we report
in the case of map that the number of evaluation steps needed is 11n + 3 and the number of
heap cells needed is 4n + 2 where n is the length of the input list. Such bounds are useful for
two purposes. First, a developer can see the cost that map itself contributes to the total cost of
a program. Second, the time bound for map proves that map is guaranteed to terminate if the
higher-order argument terminates for every input.
In contrast, consider the function rec_scheme : (α list → α list) → α list → β list that is defined as follows.
let rec rec_scheme f l =
match l with | [] → []
| x::xs → rec_scheme f (f l)
let g = rec_scheme tail
Here, RAML is not able to derive an evaluation-step bound for rec_scheme since the number of
evaluation steps (and even termination) depends on the argument f . However, RAML derives
the tight evaluation-step bound 12n + 7 for the function g .
Polynomial Bounds and Inductive Types. Existing AARA systems are either limited to linear
bounds [35, 42] or to polynomial bounds that are functions of the sizes of simple predefined
lists and binary-tree data structures [30]. In contrast, this work presents the first analysis that
can derive polynomial bounds that depend on size parameters of complex user-defined data
structures.
The bounds we derive are multivariate resource polynomials that can take into account
individual sizes of inner data structures. While it is possible to simplify the resource polynomials
in the user output, it is essential to have this more precise information for intermediate results
to derive tight whole-program bounds.
In general, the resource bounds are built of functions that count the number of specific
tuples that one can form from the nodes in a tree-like data structure. In their simplest form (i.e.,
without considering the data stored inside the nodes), they have the form
λa.|{~
a | a i is an A ki -node in a and if i < j then a i <apr e a j }|
3 Setting the Stage
7
where a is an inductive data structure with constructors A 1 , . . . , A m , ~
a = (a 1 , . . . , a n ), and <apr e
denotes the pre-order (tree traversal) on the tree a. We are able to keep track of changes of these
quantities in pattern matches and data construction fully automatically by generating linear
constraints. At the same time, they allow us to accurately describe the resource usage of many
common functions in the same way it has been done previously for simple types [28]. As an
interesting special case, we can also derive conditional bounds that describe the resource usage
as a conditional statement. For instance, for an expression such as
match x with | True → quicksort y | False → y
we derive a bound that is quadratic in the length of y if x is True and constant if x is False.
Effects. Our analysis handles references and arrays by ensuring that resource cost does not
asymptotically depend on values that have been stored in mutable cells. While it has been shown
that it is possible to extend AARA to handle mutable state [17], we decided not to add the feature
in the current system to focus on the presentation of the main contributions. There are still a lot
of possible interactions with mutable state, such as storing functions in references.
3 Setting the Stage
We describe and formalize the new resource analysis using Core RAML, a subset of the intermediate language that we use to perform the analysis. Expressions in Core RAML are in
share-let-normal form, which means that syntactic forms allow only variables instead of arbitrary terms whenever possible without restricting expressivity. We automatically transform
user-level OCaml programs to Core RAML without changing their resource behavior before the
analysis.
Syntax. For the purpose of this article, the syntax of Core RAML expressions is defined by
the following grammar. The actual core expressions also contain constants and operators for
primitive data types such as integer, float, and boolean; arrays and built-in operations for arrays;
conditionals; and free versions of syntactic forms. These free versions are semantically identical
to the standard versions but do not contribute to the resource cost. This is needed for the
resource preserving translation of user-level code to share-let-normal form.
e ::= x | x x 1 · · · x n | C x | λx.e | ref x | ! x | x 1 := x 2 | fail | tick (q)
| match x with C y → e 1 | e 2
| (x 1 , . . . , x n ) | match x with (x 1 , . . . , x n ) → e
| share x as (x 1 , x 2 ) in e | let x = e 1 in e 2 | let rec F in e
F ::= f = λx.e | F 1 and F 2
The syntax contains forms for variables, function application, data constructors, lambda abstraction, references, tuples, pattern matching, and (recursive) binding. For simplicity, we only
allow recursive definitions of functions. In the function application we allow the application
of several arguments at once. This is useful to statically determine the cost of closure creation
but also introduces ambiguity. The type system will determine if an expression like f x 1 x 2 is
parsed as ( f x 1 x 2 ) or ( f x 1 ) x 2 . The sharing expressions share x as (x 1 , x 2 ) in e is not standard
and used to explicitly introduce multiple occurrences of a variable. It binds the free variables
x 1 and x 2 in e. The expression fail is used to model exceptions. The expression tick(q) contains
8
Towards Automatic Resource Bound Analysis for OCaml
a floating point constant q. It can be used with the tick metric to specify a constant cost. A
negative floating point number q means that resources become available.
We focus on this set of language features since it is sufficient to present the main contributions of our work. We sometimes take the liberty to describe examples in user level syntax and
to use features such as built-in data types that are not described in this article.
Big-Step Operational Cost Semantics. The resource usage of RAML programs is defined by a
big-step operational cost semantics. The semantics has three interesting non-standard features.
First, it measures (or defines) the resource consumption of the evaluation of a RAML expression
by using a resource metric that defines a constant cost for each evaluation step. If this cost is
negative then resources are returned. Second, it models terminating and diverging executions
by inductively describing finite subtrees of infinite execution trees. Third, it models OCaml’s
stack-based mechanism for function application, which avoids creation of intermediate function
closures.
The semantics of Core RAML is formulated with respect to a stack (to store arguments
for function application), an environment, and a heap. Let Loc be an infinite set of locations
modeling memory addresses. A heap is a finite partial mapping H : Loc * Val that maps
locations to values. An environment is a finite partial mapping V : Var * Loc from variable
identifiers to locations. An argument stack S ::= · | `::S is a finite list of locations. We assume
every heap H contains a distinguished location Null ∈ dom(H ) such that H (Null) = Null.
The set of RAML values Val is given by
v ::= ` | (`1 , . . . , `k ) | (λx.e,V ) | (C , `)
A value v ∈ Val is either a location ` ∈ Loc, a tuple of locations (`1 , . . . , `k ), a function closure
(λx.e,V ), or a node of a data structure (C , `) where C is a constructor and ` is a location. In a
function closure (λx.e,V ), V is an environment, e is an expression, and x is a variable.
Since we also consider resources like memory that can become available during an evaluation, we have to track the watermark of the resource usage, that is, the maximal number of
resource units that are simultaneously used during an evaluation. To derive a watermark of a
sequence of evaluations from the watermarks of the sub evaluations one has to also take into
account the number of resource units that are available after each sub evaluation.
The big-step operational evaluation rules Figure 2 and Figure 3 are formulated with respect
to a resource metric M . They define an evaluation judgment of the form
S,V, H M ` e ⇓ (`, H 0 ) | (q, q 0 ) .
It expresses the following. If the argument stack S, the environment V , and the initial heap H
are given then the expression e evaluates to the location ` and the new heap H 0 . The evaluation
0
+
of e needs q ∈ Q+
0 resource units (watermark) and after the evaluation there are q ∈ Q0 resource
0
units available. The actual resource consumption is then δ = q − q . The quantity δ is negative if
resources become available during the execution of e.
There are two other behaviors that we have to express in the semantics: failure (i.e., array
access outside array bounds) and divergence. To this end, our semantic judgement not only
evaluates expressions to values but also to an error ⊥ and to incomplete computations expressed
by ◦. The judgement has the general form
S,V, H M ` e ⇓ w | (q, q 0 )
where
w ::= (`, H ) | ⊥ | ◦ .
Intuitively, this evaluation statement expresses that the watermark of the resource consumption
after some number of evaluation steps is q and there are currently q 0 resource units left. A
3 Setting the Stage
S 6= ·
9
H (V (x)) = (λx.e,V 0 )
S,V 0 , H M ` λx.e ⇓ w | (q, q 0 )
S,V, H M ` x ⇓ w | M var ·(q, q 0 )
V (x) = `
·,V, H M ` x ⇓ (`, H ) | M var
(E:VAR )
(E:VAR A PP )
S,V, H M ` e ⇓ ◦ | 0
V (x 1 ):: · · · ::V (x n ),V, H M ` x ⇓ w | (q, q 0 )
app
S,V, H M ` x x 1 · · · x n ⇓ w | M n
S = · ∨ w ∈ {⊥, ◦}
·(q, q 0 )
(E:A BORT )
(E:A PP )
S 6= ·
H (`) = (λx.e,V 0 )
V (x 1 ):: · · · ::V (x n ),V, H M ` x ⇓ ` | (q, q 0 )
S,V 0 , H M ` λx.e ⇓ w | (p, p 0 )
app
S,V, H M ` x x 1 · · · x n ⇓ w | M n
·(q, q 0 )·(p, p 0 )
S,V [x 7→ `], H M ` e ⇓ w | (q, q 0 )
`::S,V, H M ` λx.e ⇓ w | M bind ·(q, q 0 )
H 0 = H , ` 7→ (λx.e,V )
·,V, H M ` λx.e ⇓ (`, H 0 ) | M abs
·,V, H M ` e 1 ⇓ w | (q, q 0 )
(E:A BS B IND )
(E:A BS C LOS )
w ∈ {⊥, ◦}
S,V, H M ` let x = e 1 in e 2 ⇓ w | M 1let ·(q, q 0 )
·,V, H M ` e 1 ⇓ (`, H 0 ) | (q, q 0 )
(E:A PPA PP )
(E:L ET 1)
S,V [x 7→ `], H 0 M ` e 2 ⇓ w | (p, p 0 )
S,V, H M ` let x = e 1 in e 2 ⇓ w | M 1let ·(q, q 0 )·M 2let ·(p, p 0 )
(E:L ET 2)
F , f 1 = λx 1 .e 1 and · · · and f n = λx n .e n
V 0 = V [ f 1 7→ `1 , . . . , f n 7→ `n ]
0
0
H = H , `1 7→ (λx 1 .e 1 ,V ), . . . , `n 7→ (λx n .e n ,V )
S,V 0 , H 0 M ` e 0 ⇓ w | (q, q 0 )
0
S,V, H M ` let rec F in e 0 ⇓ w | M rec ·(q, q 0 )
H 0 = H , ` 7→ (C ,V (x))
·,V, H M `C x ⇓ (`, H 0 ) | M cons
H (V (x)) = (C , `)
(E:C ONS )
S,V [y 7→ `], H M ` e 1 ⇓ w | (q, q 0 )
S,V, H M ` match x with C y → e 1 | e 2 ⇓ w | M 1mat ·(q, q 0 )
H (V (x)) 6= (C , `)
S,V, H M ` e 2 ⇓ w | (q, q 0 )
S,V, H M ` match x with C y → e 1 | e 2 ⇓ w | M 2mat ·(q, q 0 )
(E:M AT 1)
(E:M AT 2)
Figure 2: Rules of the operational big-step semantics (part 1 of 2).
(E:L ET R EC )
10
Towards Automatic Resource Bound Analysis for OCaml
H 0 = H , ` 7→ (V (x 1 ), . . . ,V (x n ))
·,V, H M ` (x 1 , . . . , x n ) ⇓ (`, H 0 ) | M tuple
H (V (x)) = (`1 , . . . , `n )
(E:T UPLE )
S,V [x 1 7→ `1 , . . . , x n 7→ `n ], H M ` ⇓ w | (q, q 0 )
S,V, H M ` match x with (x 1 , . . . , x n ) → e ⇓ w | M matT ·(q, q 0 )
V (x) = `
S,V [x 1 7→ `, x 2 7→ `], H M ` e ⇓ w | (q, q 0 )
S,V, H M ` share x as (x 1 , x 2 ) in e ⇓ w | M share ·(q, q 0 )
H 0 = H , ` 7→ V (x)
·,V, H M ` ref x ⇓ (`, H 0 ) | M ref
(E:R EF )
H 0 = H [V (x 1 ) 7→ V (x 2 )]
·,V, H M ` x 1 := x 2 ⇓ (Null, H 0 ) | M assign
(E:A SSIGN )
(E:M AT T )
(E:S HARE )
` = H (V (x))
S,V, H M ` ! x ⇓ (`, H ) | M dref
(E:DR EF )
S,V, H M ` fail ⇓ ⊥ | M fail
·,V, H M ` tick (q) ⇓ (Null, H ) | M tick (q)
(E:U NDEF )
(E:T ICK )
Figure 3: Rules of the operational big-step semantics (part 2 of 2).
resource metric M : K × N → Q defines the resource consumption in each evaluation step of the
big-step semantics where K is a set of constants. We write M nk for M (k, n) and M k for M (k, 0).
It is handy to view the pairs (q, q 0 ) in the evaluation judgments as elements of a monoid
Q = (Q+0 × Q+0 , ·). The neutral element is (0, 0), which means that resources are neither needed
before the evaluation nor returned after the evaluation. The operation (q, q 0 ) · (p, p 0 ) defines
how to account for an evaluation consisting of evaluations whose resource consumptions are
defined by (q, q 0 ) and (p, p 0 ), respectively. We define
½
(q + p − q 0 , p 0 ) if q 0 ≤ p
(q, q 0 ) · (p, p 0 ) =
(q, p 0 + q 0 − p) if q 0 > p
If resources are never returned (as with time) then we only have elements of the form (q, 0) and
(q, 0) · (p, 0) is just (q + p, 0). We identify a rational number q with an element of Q as follows:
q ≥ 0 denotes (q, 0) and q < 0 denotes (0, −q). This notation avoids case distinctions in the
evaluation rules since the constants K that appear in the rules can be negative. In the semantic
rules we use the notation H 0 = H , ` 7→ v to indicate that ` 6∈ dom(H ), dom(H 0 ) = dom(H ) ∪ {`},
H 0 (`) = v, and H 0 (x) = H (x) for all x 6= `.
For efficiency reasons, Inria’s OCaml compiler evaluates function applications e e 1 · · · e n
from right to left, that is, it starts with evaluating e n . In this way, one can avoid the expensive
creation of intermediate function closures. A naive implementation would create n function
closures when evaluating the aforementioned expression: one for e, one for the application to
the first argument, etc. By starting with the last argument, we are able to put the results of the
evaluation on an argument stack and access them when we encounter a function abstraction
during the evaluation. In this case, we do not create a closure but simply bind the value on the
stack to the name in the abstraction.
4 Stack-Based Type System
11
To model the treatment of function application in the OCaml compiler, we use a stack S on
which we store the locations of function arguments. The only rules that push locations to S are
E:A PP and E:A PPA PP. To pop locations from the stack we modify the leaf rules that can return
a function closure, namely, the rules E:VAR and E:A BS for variables and lambda abstractions:
Whenever we would return a function closure (λx.e,V ) we inspect the argument stack S. If S
contains a location ` then we pop it from the stack S, bind it to the argument x, and evaluate the
function body e in the new environment V [x 7→ `]. This is defined by the rule E:A BS B IND and
indirectly by the rule E:VAR A PP. Another rule that modifies the argument stack is E:L ET 2. Here,
we evaluate the subexpression e 1 with an empty argument stack because the arguments on the
stack when evaluating the let expressions are consumed by the result of the evaluation of e 2 .
The argument stack accurately captures Inria’s OCaml compiler’s behavior to avoid the creation of intermediate function closures. It also extends naturally to the evaluation of expressions
that are not in share-let-normal form. As we will see in Section 6, the argument stack is also
necessary to prove the soundness of the multivariate resource bound analysis.
Another important feature of the big-step semantics, is that it can model failing and diverging
evaluations by allowing partial derivation judgments that can be used to derive the resource
usage after n steps. Technically, this is realized by the rule E:A BORT which can be applied at
any point to abort the current evaluation without additional resource cost. The mechanism of
aborting an evaluation is most visible in the rules E:L ET 1 and E:L ET 2: During the evaluation
of a let expression we have two possibilties. The first possibility is that the evaluation of the
subexpression e 1 is aborted using E:A BORT at some point. We can then apply the rule E:L ET 1
to pass on the resource usage before the abort. The second possibility is that e 1 evaluates to a
location `. We can then apply the E:L ET 2 to bind ` to the variable x and evaluate the expression
e2.
Example Evaluation (Running Example). We use the running example defined in Figure 1
to illustrate how the operational cost semantics works. To this end, we use the metric steps
which assigns cost 1 to every evaluation step and the metric tick which assigns cost 0 to every
evaluation step but Raml.tick(q).
Let abs ≡ Acons ([1;2],Bcons (3, Bcons (4, Nil))) is a A-B-list and let e 1 the expression that
arises by concatenating appending the expression asort (>) abs to the code in Figure 1. Then for
every H and V there exists H 0 and ` such that ·,V, H tick` e 1 ⇓ (`, H 0 ) | (0, 0) and ·,V, H steps` e 1 ⇓
(`, H 0 ) | (186, 0). Moreover, ·,V, H steps` e 1 ⇓ ◦ | (n, 0) for every n < 186.
Let e 2 be the expression that results from appending btick abs to the code in Figure 1. Then
for every H and V there exists H 0 and ` such that ·,V, H tick` e 2 ⇓ (`, H 0 ) | (5, 0) and ·,V, H tick` e 2 ⇓
◦ | (n, 0) for every n ∈ {2.5, 0}.
4 Stack-Based Type System
In this section, we introduce a type system that is a refinement of OCaml’s type system. In this
type system, we mirror the resource-aware type system and introduce some particularities that
explain features of the resource-aware types. For the purpose of this article, we define simple
types as follows.
T ::= unit | X | T ref | T1 ∗ · · · ∗ Tn | [T1 , . . . , Tn ] → T
| µX . 〈C 1 : T1 ∗X n1 , . . . ,C k : Tk ∗X nk 〉
12
Towards Automatic Resource Bound Analysis for OCaml
A (simple) type T is the unit type, an uninterpreted type variable X ∈ X , a type T ref of references
of type T , a tuple type T1 ∗ · · · ∗ Tn , a function type [T1 , . . . , Tn ] → T , or an inductive data type
µX . 〈C 1 : T1 ∗X n1 , . . . ,C k : Tk ∗X nk 〉.
Two parts of this definition are non-standard and deserve further explanation. First, bracket
function types [T1 , . . . , Tn ] → T correspond to the standard function type T1 → · · · → Tn → T .
The meaning of [T1 , . . . , Tn ] → T is that the function is applied to its first n arguments at the
same time. The type T1 → · · · → Tn → T indicates that the function is applied to its first n
arguments one after another. These two uses of a function can result in a very different resource
behavior. For instance, in the latter case we have to create n − 1 function closures. Also we
have n different costs to account for: the evaluation cost after the first argument is present, the
cost of the closure when the second argument is present, etc. Of course, it is possible that a
function is used in different ways in program. We account for that with let polymorphism (see
the following subsection). Also note that [T1 , . . . , Tn ] → T still describes a higher-order function
while T1 ∗ · · · ∗ Tn → T describes a first-order function with n arguments.
Second, inductive types are required to have a particular tree-like form. This makes it
possible to track costs that depend on size parameters of values of such types. It is of course
possible to allow arbritary inductive types and not to track such cost. Such an extension is
straighforward and we do not present it in this article.
We assume that each constructor C ∈ C is part of at most one recursive type. Furthermore
we assume that each recursive type has at least one constructor. For an inductive type T =
µX . 〈C 1 : T1 ∗X n1 , . . . ,C k : Tk ∗X nk 〉 we sometimes write T = 〈C 1 : (T1 , n 1 ), . . . ,C k : (Tk , n k )〉. We
say that Ti is the node type and n i is the branching number of the constructor C i . The maximal
branching number n = max{n 1 , . . . , n k } of the constructors is the branching number of T .
Let Polymorphism and Sharing. Following the design of the resource-aware type system,
our stack-based type system is affine. That means that a variable in a context can be used at
most once in an expression. However, we enable multiple uses of a variable with the sharing
expression share x as (x 1 , x 2 ) in e that denotes that x can be used twice in e using the (different)
names x 1 and x 2 . For input programs we allow multiple uses of a variable x in an expression e in
RAML. We then introduce sharing constructs, and replace the occurrences of x in e with the new
names before the analysis.
Interestingly, this mechanism is closely related to let polymorphism. To see this relation, first
note that our type system is polymorphic but that a value can only be used with a single type
in an expression. In practice, that would mean for instance that we have to define a different
map function for every list type. A simple and well-known solution to this problem that is
often applied in practice is let polymorphism. In principle, let polymorphism replaces variables
with their definitions before type checking. For our map function it would mean to type the
expression [map 7→ e map ]e instead of typing the expression let map = e map in e.
In principle, it would be possible to treat sharing of variables in a similar way as let polymorphism. But if we start from an expression let x = e 1 in e 2 and replace the occurrences of x in
the expression e 2 with e 1 then we also change the resource consumption of the evaluation of e 2
because we evaluate e 1 multiple times. Interestingly, this problem coincides with the treatment
of let polymorphism for expressions with side effects (the so called value restriction).
In RAML, we support let polymorphism for function closures only. Assume we have a
function definition let f = λx.e f in e that is used twice in e. Then the usual approach to enable
the analysis in our system would be to use sharing
let f = λx.e f in share f as ( f 1 , f 2 ) in e 0 .
4 Stack-Based Type System
13
To enable let polymorphism, we will however define f twice and ensure that we only pay once
for the creation of the closure and the let binding:
let f 1 = λx.e f in let f 2 = λx.e f in e 0
The functions f 1 and f 2 can now have different types. This method can cause an exponential
blow up of the size of the expression. It is nevertheless appealing because it enables us to treat
resource polymorphism in the same way as let polymorphism.
Type Judgements. Type judgements have the form
Σ; Γ ` e : T
where Σ = T1 , . . . , Tn is a list of types, Γ : Var * T is a type context that maps variables to types, e
is a core expression, and T is a (simple) type. The intuitive meaning (which is formalized later in
this section) is as follows. Given an evaluation environment that matches the type context Γ and
an argument stack that matches the type stack Σ then e evaluates to a value of type T (or does
not terminate).
The most interesting feature of the type judgements is the handling of bracket function
types [T1 , . . . , Tn ] → T . Even though function types can have multiple forms, a well-typed
expression often has a unique type (in a given type context). This type is derived from the way
a function is used. For instance, we have λ f .λx.λy. f x y : ([T1 , T2 ] → T ) → T1 → T2 → T and
λ f .λx.λy.( f x) y : (T1 → T2 → T ) → T1 → T2 → T , and the two function types are both unique.
A type T of an expression e has a unique type derivation that produces a type judgement
·, Γ ` e : T with an empty type stack. We call this canonical type derivation for e and a closed type
judgement. If T is a function type Σ → T 0 then there is a second type derivation for e that we
call an open type derivation. It derives the open type judgement Σ; Γ ` e : T 0 where |Σ| > 0. The
following lemma can be proved by induction on the type derivations.
Lemma 1. ·; Γ ` e : Σ → T if and only if Σ; Γ ` e : T .
Open and canonical type judgements are not interchangeable. An open type judgement
Σ; Γ ` e : T can only appear in a derivation with an open root of the form Σ0 , Σ; Γ ` e : T , or in a
subtree of a derivation whose root is a closed judgement of the form ·; Γ ` e : Σ00 , Σ → T where
|Σ00 | > 0. In other words, in an open derivation Σ; Γ ` e : T , the expression e is a function that
has to be applied to n > |Σ| arguments at the same time. In a given type context and for a fixed
function type, a well-typed expression has as most one open type derivation.
Type Rules. Figure 4 presents selected type rules of the type system. As usual Γ1 , Γ2 denotes
the union of the type contexts Γ1 and Γ2 provided that dom(Γ1 ) ∩ dom(Γ2 ) = ;. We thus have
the implicit side condition dom(Γ1 ) ∩ dom(Γ2 ) = ; whenever Γ1 , Γ2 occurs in a typing rule.
Especially, writing Γ = x 1 :T1 , . . . , x k :Tk means that the variables x i are pairwise distinct.
There is a close correspondence between the evaluation rules and the type rules in the
sense that every evaluation rule corresponds to exactly one type rule. (We view the two rules
for pattern match and let binding as one rule, respectively.) The type stack is modified by the
rules T:VAR P USH, T:A PP P USH, T:A BS P USH, and T:A BS P OP. For every leaf rule that can return
a function type, such as T:VAR, T:A PP, and T:A PP P USH, we add a second rule that derives the
equivalent open type. The reason becomes clear in the resource-aware type system in Section
6. The rules that directly control the shape of the function types are T:A BS P USH and T:A BS P OP
for lambda abstraction. While the other rules are (deterministically) syntax driven, the rules for
14
Towards Automatic Resource Bound Analysis for OCaml
·; x : T ` x : T
( T:VAR )
Σ; x : Σ → T ` x : T
( T:VAR P USH )
Σ; Γ, x:T1 ` e : T2
T1 ::Σ; Γ ` λx.e : T2
Σ; x : [T1 , . . . , Tn ]→Σ→T, x 1 : T1 , . . . , x n : Tn ` x x 1 · · · x n : T
·; x : [T1 , . . . , Tn ]→T, x 1 :T1 , . . . , x n :Tn ` x x 1 · · · x n : T
T = µX . 〈. . .C : U ∗X n . . .〉
Σ; Γ, y : U ∗T n , ` e 1 : T 0
( T:A PP P USH )
Σ; Γ ` λx.e : T
( T:A PP )
·; Γ ` λx.e : Σ → T
Σ; Γ, x : T ` e 2 : T 0
Σ; Γ, x : T ` match x with C y → e 1 | e 2 : T 0
T = µX . 〈. . .C : U ∗X n . . .〉
n
·; x : U ∗T ` C x : T
( T:C ONS )
·; x 1 : T1 , . . . , x n : Tn ` (x 1 , . . . , x n ) : T
Σ; Γ, x : T ` match x with (x 1 , . . . , x n ) → e : T 0
Σ; Γ, x 1 : T, x 2 : T ` e : T 0
Σ; Γ, x : T ` share x as (x 1 , x 2 ) in e : T 0
·; Γ1 ` e 1 : T1
Σ; Γ2 , x : T1 ` e 2 : T2
Σ; Γ1 , Γ2 ` let x = e 1 in e 2 : T2
( T:S HARE )
Σ; Γ, x:A ` e : B
F , f 1 = λx 1 .e 1 and · · · and f n = λx n .e n
∆ = f 1 : T1 , . . . , f n : Tn
∀i : ·; Γi , ∆ ` λx i .e i : Ti
Σ; Γ0 , ∆ ` e : T
Σ; Γ0 , . . . , Γn ` let rec F in e : T
·; x : T ` ref x : T ref
Σ; x : (Σ → T ) ref ` ! x : T
( T:R EF )
( T:DR EF P USH )
Σ; · ` fail : B
( T:FAIL )
( T:T UPLE )
( T:M AT T )
Σ; Γ ` e : B
( T:L ET )
( T:A BS P OP )
( T:M AT )
T = T1 ∗ · · · ∗ Tn
Σ; Γ, x 1 : T1 , . . . , x n : Tn ` e : T 0
T = T1 ∗ · · · ∗ Tn
( T:A BS P USH )
·; x : T ref ` ! x : T
( T:W EAK )
( T:L ET R EC )
( T:DR EF )
·; x 1 : T ref, x 2 : T ` x 1 := x 2 : unit
·; · ` tick (q) : unit
( T:T ICK )
Figure 4: Rules of the stack-based affine type system.
( T:A SSIGN )
4 Stack-Based Type System
X ∈X
15
` ∈ dom(H )
H Í ` 7→ ` : X
H (`) = `0
H Í `0 7→ a : T
H Í ` 7→ R(a) : T ref
( V:R EF )
( V:TVAR )
H Í Null 7→ () : unit
H (`) = (λx.e,V )
∃ Γ : H Í V : Γ ∧ ·; Γ ` λx.e : Σ → T
H Í ` 7→ (λx.e,V ) : Σ → T
H (`) = (`1 , . . . , `n )
∀i : H Í `i 7→ a i : Ti
H Í ` 7→ (a 1 , . . . , a n ) : T1 ∗ · · · ∗ Tn
B = µX . 〈. . . ,C : T ∗X n , . . .〉
( V:U NIT )
H (`) = (C , `0 )
( V:F UN )
( V:T UPLE )
H Í `0 7→ (a, b 1 , . . . , b n ) : T ∗B n
H Í ` 7→ C (a, b 1 , . . . , b n ) : B
( V:C ONS )
Figure 5: Coinductively relating heap cells to semantic values.
lambda abstraction introduce a choice that shapes functions types. However, there is often only
one possible choice depending on how the abstracted function is used.
As mentioned, the type system is affine and every variable in a context can at most be used
once in the typed expression. Multiple uses have to be introduced explicitly using the rule
T:S HARE. The only exception is the rule T:L ET R EC. Here we allow the use of the context ∆ in
the body of all defined functions. The reason for this is apparent in the resource aware version:
sharing of function types is always possible without any restrictions.
Well-Formed Environments. For each simple type T we inductively define a set T of values
of type T . Our goal here is not to advance the state of the art in denotational semantics but rather
to capture the tree structure of data structures stored on the heap. To this end, we distinguish
mainly inductive types (possible inner nodes of the trees) and other types (leaves). For the
formulation of type soundness, we also require that function closures are well-formed. We
simply interpret polymorphic data with the set of locations Loc.
X = Loc
unit = {()}
T ref = {R(a) | a ∈ T }
Σ → T = {(λx.e,V ) | ∃ Γ : H Í V :Γ ∧ ·; Γ ` λx.e : Σ→T }
T1 ∗· · ·∗ Tn = T1 × · · · × Tn
B = Tr(B ) if B = 〈C 1 :(T1 , n 1 ), . . . ,C n :(Tk , n k )〉
Here, T = Tr(〈C 1 :(T1 , n 1 ), . . . ,C n :(Tk , n k )〉) is the set of trees τ with node labels C 1 , . . . ,C k which
are inductively defined as follows. If i ∈ {1, . . . , k}, a i ∈ Ti , and τ j ∈ T for all 1 ≤ j ≤ n i then
C i (a i , τ1 , . . . , τni ) ∈ T .
If H is a heap, ` is a location, A is a type, and a ∈ A then we write H Í ` 7→ a : A to mean
that ` defines the semantic value a ∈ A when pointers are followed in H in the obvious way.
The judgment is formally defined in Figure 5. For a heap H there may exist different semantic
values a and simple types A such that H Í ` 7→ a : A . However, if we fix a simple type A and a
heap H then there exists at most one value a such that H Í ` 7→ a : A .
Proposition 1. Let H be a heap, ` ∈ Loc, and let A be a simple type. If H Í ` 7→ a : A and
H Í ` 7→ a 0 : A then a = a 0 .
16
Towards Automatic Resource Bound Analysis for OCaml
We write H Í ` : A to indicate that there exists a, necessarily unique, semantic value a ∈ A
so that H Í ` 7→ a : A . An environment V and a heap H are well-formed with respect to a context
Γ if H Í V (x) : Γ(x) holds for every x ∈ dom(Γ). We then write H Í V : Γ. Similarly, an argument
stack S = `1 , . . . , `n is well-formed with respect to a type stack Σ = T1 , . . . , Tn in heap H , written
H Í S : Σ, if H Í `i : Ti for all 1 ≤ i ≤ n.
Note that the rules in Figure 5 are interpreted coinductively. The reason is that in the rule
V:F UN, the location ` can be part of the closure environment V if the closure has been created
with the rule E:L ET R EC. The influence of the coinductive definition on the proofs is minimal
since all proofs in this article are by induction.
Type Preservation. Theorem 1 shows that the evaluation of a well-typed expression in a wellformed environment results in a well-formed environment.
Theorem 1. If Σ; Γ ` e : T , H Í V : Γ, H Í S : Σ, and S,V, H M ` e ⇓ (`, H 0 ) | (q, q 0 ) then H 0 Í V : Γ,
H 0 Í S : Σ, and H 0 Í ` : T .
Theorem 1 is proved by induction on the evaluation judgement.
5 Multivariate Resource Polynomials
In this section we define the set of resource polynomials which is the search space of our
automatic resource bound analysis. A resource polynomial p : T → Q+
0 maps a semantic value
of some simple type T to a non-negative rational number.
An analysis of typical polynomial computations operating on a list [a 1 , . . . , a n ] shows that
they often consist of operations that are executed for every k-tuple (a i 1 , . . . , a i k ) with 1 ≤ i 1 <
· · · < i k ≤ n. The simplest examples are linear map operations that perform some operation for
every a i . Other common examples are sorting algorithms that perform comparisons for every
pair (a i , a j ) with 1 ≤ i < j ≤ n in the worst case.
In this article, we generalize this observation to user-defined tree-like data structures. In
lists of different node types with constructors C 1 ,C 2 and C 3 , a linear computation is for instance
often carried out for all C 1 -nodes, all C 2 -nodes, or all C 1 and C 3 nodes. In general, a typical
polynomial computation is carried out for all tuples (a 1 , . . . , a k ) such that a i is a list element
with constructor C j for some j and a i appears in the list before a i +1 for all i .
As in previous work, which considered binary trees, we will essentially interpret all treelike data structures as lists with different nodes by flattening them in pre-order. As a result,
our resource polynomials only depend on the number of nodes of a certain kind in the tree
but not on structural measures like the height of the tree. To include the height into resource
polynomials in a general way, we would need a way to express a maximum (or a choice) in
the resource polynomials. We leave this for future research in favor of compositionality and
modularity. For compositionality, it is useful that the potential of a data structure is invariant
under changes in the structure of the tree.
Base Polynomials and Indices. In Figure 6, we define for each simple type T a set P(T ) of
functions p : T → N that map values of type T to natural numbers. The resource polynomials
for type T are then given as non-negative rational linear combinations of these base polynomials.
Let B = 〈C 1 : (T1 , n 1 ), . . . ,C m : (Tm , n m )〉 be an inductive type. Let C = [C j 1 , . . . ,C j k ] be a list of
B -constructors and b ∈ B . We inductively define a set τB (C , b) of k-tuples as follow: τB (C , b) is
5 Multivariate Resource Polynomials
17
∀i : p i ∈ P(Ti )
λ a. 1 ∈ P(T )
λ~
a.
Y
p i (a i ) ∈ P(T1 ∗ · · · ∗ Tk )
i =1,...,k
B = 〈C 1 : (T1 , n 1 ), . . . ,C m : (Tm , n m )〉
C = [C j 1 , . . . ,C j k ]
X
Y
p i (a i ) ∈ P(B )
λ b.
∀i : p i ∈ P(T j i )
~
a ∈τB (C ,b) i =1,...,k
Figure 6: Defining the set P(T ) of base polynomials for type T .
∀ j : I j ∈ I (T j )
? ∈ I (T )
(I 1 , . . . , I k ) ∈ I (T1 ∗ · · · ∗ Tk )
B = 〈C 1 : (T1 , n 1 ), . . . ,C m : (Tm , n m )〉
∀i : I j i ∈ I (T j i )
[〈I 1 ,C j 1 〉 , . . . , 〈I k ,C j k 〉] ∈ I (B )
Figure 7: Defining the set I (T ) of indices for type T .
the set of k-tuples (a 1 , . . . , a k ) such that C j 1 (a 1 ,~
b 1 ), . . . ,C j k (a k ,~
b k ) are nodes in the tree b ∈ B
~
~
and C j 1 (a 1 , b 1 ) <pre · · · <pre C j k (a k , b k ) for the pre-order <pre on b.
Like in the lambda calculus, we use the notation λ a. e(a) for the anonymous function that
maps an argument a to the natural number that is defined by the expression e(a). Every set
P(T ) contains the constant function λ a. 1. In the case of an inductive data type B this constant
function arises also for C = [] (one element sum, empty product).
In Figure 7, we inductively define for each simple type T a set of indices I (T ). For tuple types
T1 ∗ · · · ∗ Tk we identify the index ? with the index (?, . . . , ?). Similarly, we identify the index ?
with the index [] for inductive types.
Let T be a base type. For each index i ∈ I (T ), we define a base polynomial p i : T → N as
follows.
p ? (a) = 1
p (I 1 ,...,I k ) (a 1 , . . . , a k ) =
Y
p I j (a j )
j =1,...,k
p [〈I 1 ,C 1 〉,...,〈I k ,C k 〉] (b) =
X
Y
~
a ∈τB ([C 1 ,...,C k ],b) j =1,...,k
p I j (a j )
Examples. To illustrate the definitions, we construct the set of base polynomials for different
data types.
• We first consider the inductive type singleton that has only one constructor without
arguments.
singleton = µX 〈 Nil : unit〉
Then we have
18
Towards Automatic Resource Bound Analysis for OCaml
singleton = { Nil (())}
P(singleton) = {λ a. 1, λ a. 0} .
and
To see why, we first examine the set of tuples T (C ) = τsingleton (C , Nil (())) for different list
of constructors C . If |C | > 1 then T (C ) = ; because the tree Nil (()) does not contain any
tuples of size 2. Thus we have p [〈I 1 ,C 1 〉,...,〈I k ,C k 〉] ( Nil (())) = 0 in this case (empty sum). The
only remaining constructor lists C are [] and [〈?, Nil 〉]. As always p [] ( Nil (())) = 1 (singleton
sum). Furthermore p [〈?, Nil 〉] ( Nil (())) = 1 because τsingleton ([〈?, Nil 〉], Nil (())) = { Nil (())}
and P(unit) = {λ a. 1}.
• Let us now consider the usual sum type
sum(T1 , T2 ) = µX 〈 Left : T1 , Right : T2 〉 ; .
Then sum(T1 , T2 ) = { Left (a) | a ∈ T1 } ∪ { Right (b) | b ∈ T2 }. If we define
σC (p)(C 0 (a))
½
p(a)
0
if C = C 0
otherwise
then P(sum(T1 , T2 )) = {x 7→ 1, x 7→ 0} ∪ {σ Left (p) | p ∈ P(T1 )} ∪ {σ Right (p) | p ∈ P(T2 )}.
• The next example is the list type
list(T ) = µX 〈 Cons : T ∗X , Nil : unit〉 .
Then list(T ) = { Nil (()), Cons (a 1 , Nil (())), . . .} and we write
list(T ) = {[], [a 1 ], [a 1 , a 2 ], . . . | a i ∈ T } .
We have τlist ([〈?, Cons 〉], [a 1 , . . . , a n ]) = {a 1 , . . . , a n } and furthermore
τlist ([〈?, Cons 〉 , 〈?, Cons 〉], [a 1 , . . . , a n ]) = {(a i , a j ) | 1 ≤ i < j ≤ n} .
More generally, let C = [〈?, Cons 〉 , . . . , 〈?, Cons 〉] or C = [〈?, Cons 〉 , . . . , 〈?, Cons 〉 , 〈?, Nil 〉]
for lists of length k and k + 1, respectively. Then τlist (C , [a 1 , . . . , a n ]) = {(a i 1 , . . . , a i k ) | 1 ≤
0
i 1 < · · · < i k ≤ n}. On the other hand,
¡ ¢ τlist (D, [a 1 , . . . , a n ]) = ; if D = 〈?, Nil 〉 ::D for some
P
D 0 6= []. Since ~a ∈τ (C ,[a1 ,...,an ]) 1 = nk and λ a. 1 ∈ P(T ) we have
list
Ã
!
|b|
{λ b.
| n ∈ N} ⊆ P(list(T )) .
n
• Finally consider a list type with two different Cons -nodes (as in the running example in
Figure 1)
list2(T1 , T2 ) = µX 〈 C1 : T1 ∗X , C2 : T2 ∗ X , Nil : unit〉 .
Then we write (similarly as for list(T))
list2(T1 , T2 ) = {[], [a 1 ], [a 1 , a 2 ], . . . | a i ∈ ({C 1 } × T1 ) ∪ ({C 2 } × T2 )} .
Let b = [b 1 , . . . , b n ]. We have for example τlist2 ([〈?, C1 〉], b) = {b 1 , . . . , b n | ∀i ∃a : b i =
( C1 , a)} and τlist2 ([〈?, C1 〉 , 〈?, C2 〉], [b 1 , . . . , b n ]) = {(b i , b j ) | ∀i , j ∃a, a 0 : b i = ( C1 , a)∧b j =
( C2 , a 0 ) ∧ 1 ≤ i < j ≤ n}.
5 Multivariate Resource Polynomials
19
¡|b| ¢
P
If C = [〈?,C 1 〉 , . . . , 〈?,C 1 〉] and |C | = k then ~a ∈τ (C ,b) 1 = kC1 where |b| C1 denotes the
list2
number of C1 -nodes in the list b. Therefore we have
Ã
!
Ã
!
|b|C 1
|b|C 2
{λ b.
| n ∈ N} ⊆ P(list2(T )) and {λ b.
| n ∈ N} ⊆ P(list2(T )) .
n
n
Now consider the set D of constructor lists D such that D contains exactly k 1 elements
S
of the form 〈?,C 1 〉 and k 2 elements of the form 〈?,C 2 〉. If S = D∈D τlist2 (D, b) then
¡|b| C ¢¡|b| C ¢
P
1
2 . This means that such products of binomial coefficients are sums of
~
a ∈S 1 = k
k2
1
base polynomials.
• Coinductive types like stream(T ) = µX 〈 St : T ∗X 〉 are not inhabited in our language since
we interpret them inductively. A data structure of such a type cannot be created since we
allow recursive definitions only for functions.
Spurious Indices. The previous examples illustrate that for some inductive data structures,
different indices encode the same resource polynomial. For example, for the type list(T ) we
have p [〈?, Nil 〉] (a) = p [] (a) = 1 for all lists a. Additionally, some indices encode a polynomial that
is constantly zero. For the type list(T ) this is for example the case for p 〈?, Nil 〉 ::C if |C | > 0. We
call such indices spurious.
In practice, it is not beneficial to have spurious indices in the index sets since they slow
down the analysis without being useful components of bounds. It is straightforward to identify
spurious indices from the data type definition. The index [〈I 1 ,C 1 〉 , . . . , 〈I k ,C k 〉] is for example
spurious if k > 1 and the branching number of C i is 0 for an i ∈ {1, . . . , k − 1}.
Resource Polynomials. A resource polynomial p : T → Q+
0 for a simple type T is a nonnegative linear combination of base polynomials, i.e.,
X
p=
qi · p i
i =1,...,m
for m ∈ N, q i ∈ Q+
0 and p i ∈ P(T ). We write R(T ) for the set of resource polynomials for the base
type T .
Running Example. Consider again our running example from Figure 1. For the function
abmap, we derived the evaluation-step bound 3 + 12L + 12N . It corresponds to the following
resource polynomial. 12p (?,?,[〈?, Acons 〉]) + 12p (?,?,[〈?, Bcons 〉]) + 3p (?,?,[]) .
For the function asort’, we derived the evaluation-step bound 13 + 22K N + 13K 2 N + 15N ,
which corresponds to the resource polynomial
26p (?,[〈[〈?,::〉,〈?,::〉], Acons 〉]) + 35p (?,[〈[〈?,::〉], Acons 〉]) + 15p (?,[〈[], Acons 〉]) + 13p (?,[]) .
Selecting a Finite Index Set. Every resource polynomial is defined by a finite number of base
polynomials. In an implementation, we also have to fix a finite set of indices to make possible
an effective analysis. The selection of the indices to track can be customized for each inductive
data type and for every program. However, we currently allow the user only to select a maximal
degree of the bounds and then track all indices that correspond to polynomials of the same or a
smaller degree.
20
Towards Automatic Resource Bound Analysis for OCaml
6 Resource-Aware Type System
In this section, we describe the resource-aware type system. Essentially, we annotate the simple
type system from Section 4 with resource annotations so that type derivations correspond to
proofs of resource bounds.
Type Annotations. We use the indices and base polynomials to define type annotations and
resource polynomials.
A type annotation for a simple type T is defined to be a family
Q T = (q I )I ∈I (T ) with q I ∈ Q+
0
We write Q(T ) for the set of type annotations for the type T .
An annotated type is a pair (A,Q) where Q is a type annotation for the simple type |A| where
A and |A| are defined as follows.
A ::= unit | X | A ref | A 1 ∗ · · · ∗ A n | 〈[A 1 , . . . , A n ] → B, Θ〉
| µX . 〈C 1 : A 1 ∗X n1 , . . . ,C k : A k ∗X nk 〉
We define |A| to be the simple type T that can be obtained from A by removing all type annotations from function types.
A function type 〈[A 1 , . . . , A n ] → B, Θ〉 is annotated with a set
Θ ⊆ {(Q A ,Q B ) | Q A ∈ Q(|A 1 ∗ · · · ∗ A n |) ∧Q B ∈ Q(|B |)} .
The set Θ can contain multiple valid resource annotations for arguments and the result of the
function.
Potential of Annotated Types and Contexts. Let (A,Q) be an annotated type. Let H be a heap
and let v be a value with H Í ` 7→ a : |A|. Then the type annotation Q defines the potential
X
ΦH (v : (A,Q)) =
q I · p I (a)
I ∈I (T )
where only finitely many q I ’s are non-zero. Usually, we define type annotations Q by only stating
the values of the non-zero coefficients q I .
P
If a ∈ |A| and Q ∈ Q(|A|) is a type annotation then we also write Φ(a : (A,Q)) for I q I ·p I (a).
For use in the type system we need to extend the definition of resource polynomials to type
contexts and stacks. We treat them like tuple types. Let Γ = x 1 :A 1 , . . . , x n :A n be a type context
and let Σ = B 1 , . . . , B m be a list of types. The index set I (Σ; Γ) is defined through
I (Σ; Γ)={(I 1 , . . . , I m , J 1 , . . . , J n ) | I j ∈I (|B j |), J i ∈I (|A i |} .
A type annotation Q for Σ; Γ is a family
Q = (q I )I ∈I (Σ;Γ) with q I ∈ Q+
0.
We denote a resource-annotated context with Σ; Γ;Q. Let H be a heap and V be an environment
with H Í V : Γ where H Í V (x j ) 7→ a x j : |Γ(x j )| . Let furthermore S = `1 , . . . , `m be an argument
stack with H Í S : Σ where H Í `i 7→ b i : |B i | for all i . The potential of Σ; Γ;Q with respect to H
and V is
m+n
m
X
Y
Y
ΦS,V,H (Σ; Γ;Q) =
q~I
p I j (b j )
p I j (a x j )
~
I ∈I (Σ;Γ)
j =1
j =m+1
Here, ~
I = (I 1 , · · · , I m+n ). In particular, if Σ = Γ = · then I (Σ; Γ) = {()} and ΦV,H (Σ; Γ; q () ) = q () . We
sometimes also write q ? for q () .
6 Resource-Aware Type System
21
Folding of Potential Annotations. A key notion in the type system is the folding for potential
annotations that is used to assign potential to typing contexts that result from a pattern match
(unfolding) or from the application of a constructor of an inductive data type (folding). Folding
of potential annotations is conceptually similar to folding and unfolding of inductive data types
in type theory.
Let B = µX . 〈. . . ,C : A∗X n , . . .〉 be an inductive data type. Let Σ be a type stack, Γ, b:B be a
context and let Q = (q I )I ∈I (Σ;Γ,y:B ) be a context annotation. The C -unfolding CCB (Q) of Q with
respect to B is an annotation CCB (Q) = (q I0 )I ∈I (Σ;Γ0 ) for a context Γ0 = Γ, x:A∗B n that is defined by
0
q (I
,(J ,L 1 ,...,L n ))
½
=
q (I ,〈J ,C 〉 :: L 1 ···L n ) + q (I ,L 1 ···L n )
q (I ,〈J ,C 〉 :: L 1 ···L n )
J =0
J 6= 0
Here, L 1 · · · L n is the concatenation of the lists L 1 , . . . , L n .
Lemma 2. Let B = µX . 〈. . . ,C : A∗X n , . . .〉 be an inductive data type. Let Σ; Γ, x:B ;Q be an annotated context, H Í V : Γ, x:B , H Í S : Σ, H (V (x)) = (C , `), and V 0 = V [y 7→ `]. Then H Í V 0 :
Γ, y:A∗B n and ΦS,V,H (Σ; Γ, x:B ;Q) = ΦS,V 0 ,H (Σ; Γ, y:A∗B n ; CCB (Q)).
Sharing. Let Σ; Γ, x 1 :A, x 2 :A;Q be an annotated context. The sharing operation . Q defines
an annotation for a context of the form Σ; Γ, x:A. It is used when the potential is split between
multiple occurrences of a variable. Lemma 3 shows that sharing is a linear operation that does
not lead to any loss of potential.
(i , j )
Lemma 3. Let A be a data type. Then there are natural numbers c k for i , j , k ∈ I (|A|) such that
the following holds. For every context Σ; Γ, x 1 :A, x 2 :A;Q and every H ,V with H Í V : Γ, x:A and
H Í S : Σ it holds that ΦS,V,H (Σ, Γ, x:A;Q 0 ) = ΦS,V 0 ,H (Σ; Γ, x 1 :A, x 2 :A;Q) where V 0 = V [x 1 , x 2 7→
P
(i , j )
0
V (x)] and q (`,k)
= i , j ∈I (A) c k q (`,i , j ) .
(i , j )
The coefficients c k can be computed effectively. We were however not able to derive a
closed formula for the coefficients. The proof is similar as in previous work [31]. For a context
Σ; Γ, x 1 :A, x 2 :A;Q we define . Q to be Q 0 from Lemma 3.
Type Judgements. A resource-aware type judgement has the form
Σ; Γ;Q M ` e : (A,Q 0 )
where Σ; Γ;Q is an annotated context, M is a resource metric, A is an annotated type, and Q 0 is a
type annotation for |A|. The intended meaning of this judgment is that if there are more than
Φ(Σ; Γ;Q) resource units available then this is sufficient to cover the evaluation cost of e under
metric M . In addition, there are at least Φ(v:(A,Q 0 )) resource units left if e evaluates to a value v.
Notations. Families that describe type and context annotations are denoted with upper case
letters Q, P, R, . . . with optional superscripts. We use the convention that the elements of the
families are the corresponding lower case letters with corresponding superscripts, i.e., Q =
(q I )I ∈I and Q 0 = (q I0 )I ∈I .
If Q, P and R are annotations with the same index set I then we extend operations on Q
pointwise to Q, P and R. For example, we write Q ≤ P + R if q I ≤ p I + r I for every I ∈ I . For
K ∈ Q we write Q = Q 0 + K to state that q ? = q ?0 + K ≥ 0 and q I = q I0 for I 6= ? ∈ I . Let Γ = Γ1 , Γ2
be a context, let I = (I 1 , . . . , I k ) ∈ I (Γ1 ) and J = (J 1 , . . . , J ` ) ∈ I (Γ2 ) . We write (I , J ) for the index
22
Towards Automatic Resource Bound Analysis for OCaml
(I 1 , . . . , I k , J 1 , . . . , J ` ) ∈ I (Γ). Let Q be an annotation for a context Σ; Γ1 , Γ2 . For J ∈ I (Γ2 ) we define
Γ
the projection π(J1,J 0 ) (Q) of Q to Γ1 to be the annotation Q 0 for ·; Γ1 with q I0 = q (J ,I ,J 0 ) . In the same
Σ;Γ1
way, we define the annotations πΣJ (Q) for Σ; · and π J
(Q) for Σ; Γ1 .
Cost Free Types. We write Σ; Γ;Q cf ` e : (A,Q 0 ) to refer to cost-free type judgments where cf is
the cost-free metric with cf (K ) = 0 for constants K . We use it to assign potential to an extended
context in the let rule. More info is available in previous work [32].
Subtyping. As usual, subtyping is defined inductively so that types have to be structurally
identical. The most interesting rule is the one for function types:
Θ0 ⊆ Θ
∀i : A 0i <: A i
B <: B 0
〈[A 1 , . . . , A n ] → B, Θ〉 <: 〈[A 01 , . . . , A 0n ] → B 0 , Θ0 〉
(S:F UN )
A function type is a subtype of another function type if it allows more resource behaviors (Θ0 ⊆ Θ).
Result types are treated covariant and arguments are treated contravariant.
Unsurprisingly, our type system does not have principle types. This is to allow the typing of
examples such as rec_scheme from Section 2. In a principle type, we would have to assume the
weakest type for the arguments, that is, function types that are annotated with empty sets of type
annotations. This would mean that we cannot use functions in the arguments. However, it is
possible to derive a principle type 〈Σ → B, Θ〉 for fixed argument types Σ. Here, we would derive
all possible annotations (Q,Q 0 ) ∈ Θ in the function annotation and all possible annotations
(Q,Q 0 ) that appear in function annotations of the result type.
If we take the more algorithmic view of previous work [42] then we can express a principle
type for a function with a set of constraints that has holes for the constraint sets of the higherorder arguments. It is however unclear what such a type means for a user and we prefer a more
declarative view that clearly separates type checking and type inference. An open problem with
constraint based principle types is polymorphism.
Type Rules. Figure 8 and Figure 9 contain the type rules for annotated types. We integrated the
new concepts so that the rules look similar to the rules in previous papers [42, 31, 34].
The rule A:VAR can only be applied if the type stack Σ is empty. It then simply accounts for
the cost M var and passes the potential that is assigned to the variable by the type context to
the result type. If the type stack is not empty then the rule A:VAR P USH has to be applied. In
this case, the variable x must have a function type. We then look up a possible type annotation
for the arguments and the result (P, P 0 ) ∈ Θ in the type context, account for the cost of variable
look-up (M var ) and behave as specified by (P, P 0 ). We do not account for the cost of the “function
application” because is cost is handled in the rules A:A PP and A:A PP P USH.
The rules A:A PP and A:A PP P USH correspond to the simple type rules T:A PP and T:A PP P USH.
app
In A:A PP we assume that the type stack is empty. We account for the cost M n of applying a
function to n arguments and look up valid potential annotations (P, P 0 ) for the function body in
the function annotation Θ. We then require that we have the potential specified by P available
and return potential as specified by P 0 . In the rule A:A PP P USH we account for two applications:
We first account for the function application as in the rule A:A PP. We then assume that the
return type is a function type and apply the arguments that are stored on the type stack Σ as we
do in the rule A:VAR P USH.
The rules A:A BS P USH and A:A BS P OP for lambda abstraction correspond the rules T:A BS P USH
and T:A BS P OP. As in the simple type system we can use them to non-deterministically pop
6 Resource-Aware Type System
Q = Q 0 + M var
(A:VAR )
·; x:B ;Q M ` x : (B,Q 0 )
Γ = x 1 :A 1 , . . . , x n :A n
23
(P, P 0 ) ∈ Θ
Σ;·
π?
(Q) = P + M var
P 0 = Q0
Σ; x: 〈Σ → B, Θ〉 ;Q M ` x : (B,Q 0 )
πΓ? (Q) = P + M n
app
(P, P 0 ) ∈ Θ
Q0 = P 0
·; x: 〈[A 1 , . . . , A n ]→B, Θ〉 , Γ;Q M ` x x 1 · · · x n : (B,Q 0 )
(P, P 0 ) ∈ Θ
Γ = x 1 :A 1 , . . . , x n :A n
app
0
πΣ
πΓ? (Q) = P + M n
? (Q)−q ? +p ? = R
(R, R 0 ) ∈ Θ0
(A:VAR P USH )
(A:A PP )
R0 = Q0
Σ; x: 〈[A 1 , . . . , A n ]→ 〈Σ→B, Θ0 〉 , Θ〉 , Γ;Q M ` x x 1 · · · x n : (B,Q 0 )
Σ; Γ, x:A; P M ` e : (B,Q 0 )
∀ I ,~
J : r (I ,~J ) = p (~J ,I )
Q = R + M bind
(A:A BS P USH )
A::Σ; Γ;Q M ` λx.e : (B,Q 0 )
Q = Q 0 + M abs
∀(P, P 0 ) ∈ Θ : Σ; Γ; R M ` λx.e : (B, P 0 ) ∧ r (~I ,~J ) =
½
p~I
0
(A:A PP P USH )
if ~
J =~
?
otherwise
·; Γ;Q M ` λx.e : (〈Σ → B, Θ〉 ,Q 0 )
B = µX . 〈. . .C : A∗X n . . .〉
Σ; Γ, y:A∗B ; P M ` e 1 : (A 0 , P 0 )
Σ; Γ, x:B ; R M ` e 2 : (A 0 , R 0 )
C
mat
0
0
CB (Q) = P +M1
P =Q
Q = R+M 2mat
R0 = Q0
(A:A BS P OP )
n
Σ; Γ, x:B ;Q M ` match x with C y → e 1 | e 2 : (A 0 ,Q 0 )
(A:M AT )
Q = CCB (Q 0 )+M cons
B = µX . 〈. . .C : A∗X n . . .〉
·; x:A∗B n ;Q M `C x : (B,Q 0 )
Σ; Γ, x 1 :A, x 2 :A; P M ` e : (B,Q 0 )
(A:C ONS )
Q = M share + . (P )
Σ; Γ, x:A;Q M ` share x as (x 1 , x 2 ) in e : (B,Q 0 )
(A:S HARE )
Σ; Γ2 , Γ1 ; P M ` e 1
Σ; Γ2 , x:A; P 0
0
Σ; Γ2 , x:A; R M ` e 2 : (B,Q )
Q = P + M 1let
P 0 = R + M 2let
Σ; Γ2 , Γ1 ;Q M ` let x = e 1 in e 2 : (B,Q 0 )
F , f 1 = λx 1 .e 1 and · · · and f n = λx n .e n
∆ = f 1 :A 1 , . . . , f n :A n
∀i : ·; Γi , ∆; P i M ` λx i .e i : (A i , P i0 )
Σ;Γ0
Σ;Γ0
rec
π~? (Q) = π~? (P ) + M + n·M abs
Σ; Γ0 , ∆; P M ` e : (B,Q 0 )
Σ; Γ0 , . . . , Γn ;Q M ` let rec F in e : (B,Q 0 )
(A:L ET )
(A:L ET R EC )
¦ ¦ ¦
∀ j ∈ I (Σ; ∆):
0
0
j =~
? =⇒ ·; Γ; πΓj (Q) M ` e : (A, πx:A
(Q
))
j 6=~
? =⇒ ·; Γ; πΓj (Q) cf ` e : (A, πx:A
j
j (Q ))
Σ; ∆, Γ;Q M ` e
Σ; ∆, x:A;Q 0
Figure 8: Type rules for annotated types (part 1 of 2).
(B:B IND )
24
Towards Automatic Resource Bound Analysis for OCaml
Q = Q 0 + M tuple
B = A1 ∗ · · · ∗ An
·; x 1 :A 1 , . . . , x n :A n ;Q M `(x 1 , . . . , x n ) : (B,Q 0 )
A = A1 ∗ · · · ∗ An
(A:T UPLE )
Q = P + M matT
Σ; Γ, x 1 :A 1 , . . . , x n :A n ; P M ` e : (B,Q 0 )
Σ; Γ, x:A;Q M ` match x with (x 1 , . . . , x n ) → e : (B,Q 0 )
q ? = q ?0 + M fail
Σ; ·;Q ` fail : (B,Q)’
·; ·;Q ` tick (q) : (unit,Q 0 )
q I0
q ? = q ?0 + M ref
·; x : A;Q M ` ref x : (A ref,Q 0 )
(P, P 0 ) ∈ Θ
q () = q ?0 + M tick (q)
(A:FAIL )
(A:R EF )
½
=
q ? − M dref
0
P 0 = Q0
Σ; x : 〈Σ → A, Θ〉 ref;Q M ` ! x : (A,Q 0 )
q ? = q ?0 + M assign
·; x 1 : A ref, x 2 : A;Q M ` x 1 := x 2 : (unit,Q 0 )
Σ; Γ; P ` e : (B, P 0 )
0
Σ; Γ, x:A;Q M ` e : (B,Q )
(A:W EAK -C)
if I = ?
otherwise
(A:A SSIGN )
(A:W EAK -A)
Σ; Γ;Q M ` e : (B 0 ,Q 0 )
Σ; Γ, x:A 0 ;Q M ` e : (B,Q 0 )
B 0 <: B
0
Σ; Γ;Q M ` e : (B,Q )
A <: A 0
Σ; Γ, x:A;Q M ` e : (B,Q 0 )
(A:DR EF )
(A:DR EF P USH )
Q0 ≤ P 0 + c
Q ≥ P +c
Σ; Γ;Q ` e : (B,Q 0 )
Σ; Γ; πΓ? (Q) M ` e : (B,Q 0 )
(A:T ICK )
·; x : A ref;Q M ` ! x : (A,Q 0 )
dref
πΣ;·
? (Q) = P + M
(A:M AT T )
(A:S UBTYPE -C)
Figure 9: Type rules for annotated types (part 2 of 2).
(A:S UBTYPE -R)
6 Resource-Aware Type System
25
the type stack Σ. When we do so in the rule A:A BS P OP, we create the function annotation Θ
by essentially deriving Σ; Γ; P M ` λx.e : (B, P 0 ) for every (P, P 0 ) ∈ Θ. However, we throw away all
potential that depends on the context Γ and only use the potential that is assigned the arguments
Σ (annotation R).
The rule A:C ONS assigns potential to a new node of an inductive data structure. The additive
shift CCB (Q 0 ) transforms the annotation Q 0 to an annotation Q for the context ·; x:A∗B n . Lemma
2 shows that potential is neither gained nor lost by this operation. The potential Q of the context
has to pay for both the potential Q 0 of the resulting list and the resource cost M cons of the
construction of the new node.
The rule A:M AT shows how to treat pattern matching. The initial potential defined by the
annotation Q of the context Σ; Γ, x:B has to be sufficient to pay the costs of the evaluation of e 1
or e 2 (depending on whether the matched succeeds) and the potential defined by the annotation
Q 0 of the result type. To type the expression e 2 we basically just use the annotation Q (after
paying for the constant match cost). To type the expression e 1 we rely on the additive shift CCB (Q)
that results in an annotation for the context Σ; Γ, y:A∗B n . Again there is no loss of potential (see
Lemma 2). The equalities relate the potential before and after the evaluation of e 1 or e 2 , to the
potential before and after the evaluation of the match operation by incorporating the respective
resource cost for the matching.
The rule A:S HARE uses the sharing operation . P to related the potentials defined by
Σ; Γ, x:A;Q and Σ; Γ, x 1 :A, x 2 :A; P . As with matching, there is no loss of potential (see Lemma 3).
In the rule A:L ET the result of the evaluation of an expression e 1 is bound to a variable x.
The problem that arises is that the resulting annotated context Σ; Γ2 , x:A; R features potential
functions whose domain consists of data that is referenced by x as well as data that is referenced
in the type context Γ2 . This potential has to be related to data that is referenced by Γ2 and the
free variables in e 1 (i.e., the variables in the type context Γ1 ).
To express the relations between mixed potentials before and after the evaluation of e 1 , we
introduce a new auxiliary binding judgement of the from
Σ; ∆, Γ;Q M ` e
Σ; ∆, x:A;Q 0
in the rule B:B IND. The intuitive meaning of the judgement is the following. Assume that e is
evaluated in the context ∆, Γ, FV(e) ∈ dom(Γ), and that e evaluates to a value that is bound to the
variable x. Then the initial potential Φ(Σ; ∆, Γ;Q) is larger than the cost of evaluating e in the
metric M plus the potential of the resulting context Φ(Σ, ∆, x:A;Q 0 ). Lemma 4 formalizes this
intuition.
Lemma 4. Let H Í V :∆, Γ, H Í S : Σ, and Σ; ∆, Γ;Q M ` e
Σ; ∆, x:A;Q 0 .
1. If S,V, H M ` e ⇓ (`, H 0 ) | (p, p 0 ) then ΦS,V,H (∆, Γ;Q) ≥ p + ΦS,V 0 ,H 0 (∆, x:A;Q 0 ) where V 0 =
V [x 7→ `].
2. If S,V, H M ` e ⇓ ρ | (p, p 0 ) then p ≤ ΦS,V,H (Γ;Q).
Formally, Lemma 4 is a consequence of the soundness of the type system (Theorem 2). In
the inductive proof of Theorem 2, we use a weaker version of Lemma 4 in which the soundness
of the type judgements in Lemma 4 is an additional precondition.
The rule A:L ET R EC is similar the rule T:L ET R EC for standard type systems. The cost of the
creation of the n closures is accounted for by n·M abs . It is not difficult to relate this cost to the
number of captured variables in the closure but we refrain from doing so in favor of simplicity.
The initial potential, defined by Σ; Γ0 , . . . , Γn ;Q, only flows into the potential Σ; Γ0 , ∆; P that is
26
Towards Automatic Resource Bound Analysis for OCaml
used to pay for the cost of evaluating the expression e. The potential annotation P i and P i0 that
are used in the typing of the recursive functions are unconstrained. This is not a bug but uses
the fact that P i can only be used to pay for the cost of the closure creation in the rule A:A BS P OP.
In the rule A:FAIL, we only require that the constant potential M fail is available. In contrast to
the other rules we do not relate the initial potential Q with the resulting potential Q 0 . Intuitively,
this is sound because the program is aborted when evaluating the expression fail. A consequence
of the rule T: FAIL is that we can type the expression let x = fail in e with constant initial potential
M fail regardless of the resource cost of the expression e.
In the rule A:T ICK we simply require that M tick (q) constant potential is available.
In the rule A:R EF, we only require the availability of the constant potential M ref . We discard
the remaining potential that is assigned to x by Q. Since references do not carry potential in our
system, q ?0 is the only coefficient in Q 0 . In the rule A:A SSIGN we simply pay for the cost of the
operation (M assign ) and discard the potential that is assigned to the arguments. Since the return
value is (), q ?0 is the only coefficient in Q 0 . In the rule T:AD REF, we again discard the potential of
the arguments and also require that the non-linear coefficients of the annotation of the result
are zero. Again, this is because references do not carry potential.
The structural rules A:W EAK -A, A:W EAK -C, A:S UBTYPE -R, and A:S UBTYPE -C apply to every
expression. In the implementation, they are integrated into the syntax directed rules to enable
automatic type inference. As expected, they are used at the exact same places at which you would
use corresponding rules in a standard type system; for instance, when combining branches
(weakening and subtyping) of match expressions or when constructing inductive data structures
(subtyping). The rule A:W EAK -A relies on the fact that an annotated type remains sound if
we add more potential to the context and remove potential from the result. Similarly, the
rule A:W EAK -C states that we can add variables with arbritary to the type context. The rules
A:S UBTYPE -R and A:S UBTYPE -C are similar to the standard rules of subtyping.
Soundness. Our goal is to prove the following soundness statement for type judgements.
Intuitively, it says that the initial potential is an upper bound on the watermark resource usage,
no matter how long we execute the program.
If Σ; Γ;Q M ` e : (A,Q 0 ) and S,V, H M ` e ⇓ w | (p, p 0 ) then p ≤ ΦS,V,H (Σ; Γ;Q).
To prove this statement by induction, we need to prove a stronger statement that takes into
account the return value and the annotated type (A,Q 0 ) of e. Moreover, the previous statement
is only true if the values in S, V and H respect the types required by Σ and Γ. Therefore, we
adapt our definition of well-formed environments to annotated types. We simply replace the
rule V:F UN in Figure 5 with the following rule. Of course, H Í V : Γ refers to the newly defined
judgment.
H (`) = (λx.e,V )
∃ Γ,Q,Q 0 : H Í V : Γ ∧
·; Γ;Q M ` λx.e : (〈Σ → B, Θ〉 ,Q 0 )
H Í ` 7→ (λx.e,V ) : 〈Σ → B, Θ〉
( V:F UN )
In addition to the aforementioned soundness, Theorem 2 states a stronger property for terminating evaluations. If an expression e evaluates to a value v in a well-formed environment then
the difference between initial and final potential is an upper bound on the resource usage of the
evaluation.
Theorem 2 (Soundness). Let H Í V : Γ, H Í S : Σ, and Σ; Γ;Q M ` e : (B,Q 0 ).
1. If S,V, H M ` e ⇓ (`, H 0 ) | (p, p 0 ) then p ≤ ΦS,V,H (Σ; Γ;Q) and p − p 0 ≤ ΦS,V,H (Σ; Γ;Q) −
ΦH 0 (`:(B,Q 0 )) and H Í ` : B .
7 Implementation and Bound Inference
Input
Program
Parser
Type Inference
OCaml
Bytecode
Typed
OCaml
Syntax
Tree
27
RAML Compiler
RAML Analyzer
Explicit Let
Polymorphism
Multivariate
AARA
Bracket-Type
Inference
Stack-Based
Type Checking
Typed
RAML
Syntax
Tree
LP Solver
Frontend
Resource Type
Interpretation
Share-Let
Normal Form
Resource
Metrics
OCaml - C
Bindings
CLP
Resource
Bounds
Figure 10: Implementation of RAML.
2. If S,V, H M ` e ⇓ ◦ | (p, p 0 ) then p ≤ ΦS,V,H (Σ; Γ;Q).
Theorem 2 is proved by a nested induction on the derivation of the evaluation judgment
and the type judgment Σ; Γ;Q ` e:(B,Q 0 ). The inner induction on the type judgment is needed
because of the structural rules. There is one proof for all possible instantiations of the resource
constants. An sole induction on the type judgement fails because the size of the type derivation
can increase in the case of the function application in which we retrieve a type derivation for the
function body from the well-formed judgement as defined by the (updated) rule V:F UN.
The structure of the proof matches the structure of the previous soundness proofs for type
systems based on AARA [35, 42, 31, 34]. The induction case of many rules is similar to the induction cases of the corresponding rules for multivariate AARA for first-order programs [31] and
linear AARA for higher-order programs [42]. For one thing, additional complexity is introduced
by the new resource polynomials for user-defined data types. We designed the system so that this
additional complexity is dealt with locally in the rules A:M AT, A:C ONS, and A:S HARE. The soundness of these rules follows directly from an application of Lemma 2 and Lemma 3, respectively.
As in previous work [35] the well-formed judgement that captures type derivations enables us to
treat function abstraction and application in a very similar fashion as in the first-order case [31].
The coinductive definition of the well-formedness judgement does not cause any difficulties.
A major novel aspect in the proof is the typed argument stack S : Σ that also carries potential.
Surprisingly, this typed stack is simply treated like a typed environment V : Γ in the proof. It is
already incorporated in the shift and share operations (Lemma 2 and Lemma 3).
We deal with the mutable heap by requiring that array elements do not influence the potential
of an array. As a result, we can prove the following lemma, which is used in the proof of Theorem
2.
Lemma 5. If H Í V :Γ, H Í S : Σ, Σ; Γ;Q M ` e : (B,Q 0 ) and S,V, H M ` e ⇓ (`, H 0 ) | (p, p 0 ) then
ΦS,V,H (Γ;Q) = ΦS,V,H 0 (Γ;Q).
7 Implementation and Bound Inference
Figure 10 shows an overview of the implementation of RAML. It consists of about 12000 lines of
OCaml code, excluding the parts that we reused from Inria’s OCaml implementation. The development took around 8 person months. We found it very helpful to develop the implementation
and the theory in parallel, and many theoretical ideas have been inspired by implementation
challenges.
28
Towards Automatic Resource Bound Analysis for OCaml
We reuse the parser and type inference algorithm from OCaml 4.01 to derive a typed OCaml
syntax tree from the source program. We then analyze the function applications to introduce
bracket function types. To this end, we copy a lambda abstraction for every call site. We still
have to implement a unification algorithm since functions, such as let g = f x, that are defined by
partial application may be used at different call sites. Moreover, we have to deal with functions
that are stored in references.
In the next step, we convert the typed OCaml syntax tree into a typed RAML syntax tree.
Furthermore, we transform the program into share-let-normal form without changing the
resource behavior. For this purpose, each syntactic form has a free flag that specifies whether it
contributes to the cost of the original program. For example, all share forms that are introduced
are free. We also insert eta expansions whenever they do not influence resource usage.
After this compilation phase, we perform the actual multivariate AARA on the program in
share-let-normal form. Resource metrics can be easily specified by a user. We include a metric
for heap cells, evaluation steps, and ticks. Ticks allows the user to flexibly specify the resource
cost of programs by inserting tick commands Raml.tick(q) where q is a (possibly negative)
floating-point number.
In principle, the actual bound inference works similarly as in previous AARA systems [35, 30]:
First, we fix a maximal degree of the bounds and annotate all types in the derivation of the
simple types with variables that correspond to type annotations for resource polynomials of
that degree. Second, we generate a set of linear inequalities, which express the relationships
between the added annotation variables as specified by the type rules. Third, we solve the
inequalities with Coin-Or’s fantastic LP solver CLP. A solution of the linear program corresponds
to a type derivation in which the variables in the type annotations are instantiated according
to the solution. The objective function contains the coefficients of the resource annotation of
the program inputs to minimize the initial potential. Modern LP solvers provide support for
iterative solving that allows us to express that minimization of higher-degree annotations should
take priority.
The type system we use in the implementation significantly differs from the declarative
version we describe in this article. For one thing, we have to use algorithmic versions of the
type rules in the inference in which the non-syntax-directed rules are integrated into the syntaxdirected ones [31]. For another thing, we annotate function types not with a set of type annotations but with a function that returns an annotation for the result type if presented with an
annotation of the return type. The annotations here are symbolic and the actual numbers are
yet to be determined by the LP solver. Function annotations have the side effect of sending constraints to the LP solver. It would be possible to keep a constraint set for the respective function
in memory and to send a copy with fresh variables to the LP solver at every call. However, it is
more efficient to lazily trigger the constraint generation from the function body at every call site
when the function is provided with a return annotation.
To make the resource analysis more expressive, we also allow resource-polymorphic recursion. This means that we need a type annotation in the recursive call that differs from the
annotations in the argument and result types of the function. To infer such types we successively
infer type annotations of higher degree. Details can be found in previous work [32].
For the most part, our constraints have the form of a so-called network (or network-flow)
problem [53]. LP solvers can handle network problems very efficiently and in practice CLP solves
the constraints RAML generates in linear time. Because our problem sizes are large, we can save
memory and time by reducing the number of constraints that are generated during typing. A
representative example of an optimization is that we try to reuse constraint names instead of
producing constraints like p = q.
8 Experimental Evaluation
29
RAML provides two ways of analyzing a program. In main mode RAML derives a bound for
evaluation cost of the main expression of the program, that is, the last expression in the top-level
list of let bindings. In module mode, RAML derives a bound for every top-level let binding that
has a function type.
Apart from the analysis itself, we also implemented the conversion of the derived resource
polynomials into easily-understood polynomial bounds and a pretty printer for RAML types
and expressions. Additionally, we implemented an efficient RAML interpreter that we use for
debugging and to determine the quality of the bounds.
8 Experimental Evaluation
The development of RAML was driven by an ongoing experimental evaluation with OCaml code.
Our goal has been to ensure that the derived bounds are precise, that different programming
styles are supported, that the analysis is efficient, and that existing code can be analyzed. In an
experimental evaluation, we applied our automatic resource bound analysis to 411 functions and
6018 lines of code. The source code of RAML as well as all OCaml files used in the experiments
are available online [29]. The website also provides an easy-to-use interactive web interface that
can be used to experiment with RAML.
Analyzed Code and Limitations. The experiments have been performed with four sets of
source code, extracted OCaml code from Coq specifications in CompCert [46], an OCaml tutorial [48], code from the OCaml standard library, and handwritten code. For the handwritten part,
we mostly implemented classical textbook algorithms and use cases inspired from real-word
applications. The textbook algorithms include algorithms for matrices, graph algorithms, search
algorithms, and classic examples from amortized analysis such as functional queues and binary
counters. The use cases include energy management in an autonomous mobile device and
calling Amazon’s Dynamo DB from OCaml (see Section 9).
OCaml is a complex programming language and RAML does not yet support all language
features of OCaml. This includes modules, object-oriented features, record types, built-in
equality, strings, nested patterns, and calls to native C functions. Therefore it is currently hard to
apply RAML directly to existing OCaml code. However, the support of many of these additional
features is not a theoretical limitation of the analysis. but rather caused by a lack of resources
on the implementation side. If RAML can be applied to existing code then the results are very
satisfactory. For instance, we applied RAML to OCaml’s standard list library list.ml: RAML
automatically derives evaluation-step bounds for 47 of the 51 top-level functions. All derived
bounds are asymptotically tight. The 4 functions that cannot be bounded by RAML all use a
variation of merge sort whose termination (and thus resource usage) depends on an arithmetic
shift which is currently unsupported. The file list.ml consists of 428 lines of code and the analysis
takes 3.2 s on a MacBook Pro.
Also note that our technique extends all previous works on AARA for strict, sequential
evaluation. Thus we can handle all examples that have been previously evaluated. The quality
of the derived bounds is identical to the previous ones and the efficiency of the analysis is
improved.
RAML fails if the resource usage can only be bounded by a measure that depends on a
semantic property of the program or a measure that depends on the difference of the sizes of
two data structures. Loose bounds are often the result of inter-procedural dependencies. For
instance, the worst-case behaviors of two functions f and g might be triggered by different
30
Towards Automatic Resource Bound Analysis for OCaml
let comp f x g = fun z → f x (g z)
let rec walk f xs =
match xs with | [] → (fun z → z)
| x::ys → match x with | Left _ →
fun y → comp (walk f) ys (fun z → x::z) y
| Right l →
let x’ = Right (quicksort f l) in
fun y → comp (walk f) ys (fun z → x’::z) y
let rev_sort f l = walk f l []
RAML output for rev_sort (after 0.68s run time):
10 + 23*K*M + 32*L*M + 20*L*M*Y + 13*L*M*Y^2
where
M is the num. of ::-nodes of the 2nd comp. of the argument
L is the fraction of Right-nodes in the ::-nodes of the 2nd component of the arg.
Y is the maximal number of ::-nodes in the Right-nodes in the ::-nodes of the 2nd
component of the argument
K is the fraction of Left-nodes in the ::-nodes of the 2nd component of the arg.
Figure 11: Modified challenge example from Avanzini et al. [9] and shortened output
of the automatic bound analysis performed by RAML for the function rev_sort. The
derived bound is a tight bound on the number of evaluation steps in the big-step
semantics if we do not take into account the cost of the higher-order argument f.
inputs. However, the analysis would add up the worst-case behaviors of both functions in a
program such as f(a);g(a). Another reason for loose bounds is that a tight bound cannot be
represented by a multivariate resource polynomial.
Example Experiment. To give an impression of the experiments we performed, Figure 11
contains the output of an analysis of a challenging function in RAML. The code is an adoption
of an example that has been recently presented by Avanzini et al. [9] as a function that can not
be handled by existing tools. To illustrate the challenges of resource analysis for higher-order
programs, Avanzini et al. implemented a (somewhat contrived) reverse function rev for lists
using higher-order functions. RAML automatically derives a tight linear bound on the number
of evaluation steps used by rev.
To show more features of our analysis, we modified Avanzini et al.’s rev in Figure 11 by
adding an additional argument f and a pattern match to the definition of the function walk. The
resulting type of walk is
(α → α → bool) → [(β ∗ α list) either list; (β ∗ α list) either list]
→ (β ∗ α list) either list
Like before the modification, walk is essentially the append_reverse function for lists. However,
we assume that the input lists contain nodes of the form Left a or Right b so that b is a list.
During the reverse process of the first list in the argument, we sort each list that is contained in
a Right-node using the standard implementation of quick sort (not given here). RAML derives
the tight evaluation-step bound that is shown in Figure 11. Since the comparison function for
8 Experimental Evaluation
CompCert
31
Metric
#Funs
LOC
Time
#Const
#Lin
#Quad
#Cubic
#Poly
#Fail
steps
heap
tick
steps
243
243
174
164
3218
3218
2144
2740
72.10s
70.36s
64.68s
1300.91s
16
41
19
32
130
112
79
99
60
60
53
7
28
22
19
0
239
239
174
138
4
4
0
26
Asym.
Tight
225
225
160
137
Table 1: Overview of experimental results.
quicksort (argument f) is not available, RAML assumes that it does not consume any resources
during the analysis. If rev_sort is applied to a concrete argument f then the analysis is repeated
to derive a bound for this instance.
CompCert Evaluation. We also performed an evaluation with the OCaml code that is created
by Coq’s code extraction mechanism during the compilation of the verified CompCert compilers [46]. We sorted the files topologically from their dependency requirements, and analyzed 13
files from the top. 1 We could not process the files further down the dependency graph because
they heavily relied on modules which we do not currently support. Using the evaluation-step
metric, we analyzed 164 functions, 2740 LOC in 1300 seconds.
Figure 12 shows example functions from the CompCert code base. As an artifact from the
Coq code extraction, CompCert uses two implementations of the reverse function for lists. The
function rev is a naive quadratic implementation that uses append and the function rev’ is an
efficient tail-recursive linear implementation. RAML automatically derives precise evaluation
step bounds for both functions. As a result, a Coq user who is inspecting the derived bounds for
the extracted OCaml code is likely to spot performance problems resulting from the use of rev.
Summary of Results. Table 1 contains a compilation of the experimental results. The first 3
rows show the results for OCaml libraries, handwritten code, and the OCaml tutorial [48]. The
last column shows the results for CompCert [46]. The column LOC contains the total number of
lines of OCaml code that has been analyzed with the respective metric. Similarly, the column
Time contains the total time of all analyses with this metric. The column #Poly contains the
number of functions for which RAML automatically derived a bound. The columns #Const, #Lin,
#Quad, and #Cubic show the number of derived bounds that are constant, linear, quadratic, and
cubic. Finally, columns #Fail and Asym.Tight contain the number of examples for which RAML
is unable to derive a bound and the number of bounds that are asymptotically tight, respectively.
We also experimented with example inputs to determine the precision of the constant factors
in the bounds. In general, the bounds are very precise and often match the actual worse-case
behavior. However, we did not yet perform a systematic evaluation with automatically generated
example inputs.
The reported numbers result from the analysis of 140 non-trivial functions that are (with a
few exceptions) recursive and higher order. Appendix A contains a short description of every
function that is part of the evaluation, along with its type, the run time of the analysis, and the
derived bounds. The functions have been automatically analyzed using the steps metric that
counts the number of evaluation steps and the heap metric that counts the number of allocated
heap cells. Moreover, we have used the tick metric to add custom cost measures to some of the
functions. These measures vary from program to program and include number of function calls,
1 A list of analyzed files and functions is included in the TR.
32
Towards Automatic Resource Bound Analysis for OCaml
let rec app l m =
match l with
| [] → m
| a :: l1 → a :: (app l1 m)
let rec rev = function →
| [] → []
| x :: l’ → app (rev l’) (x :: [])
let rec rev_append l l’ =
match l with
| [] → l’
| a :: l0 → rev_append l0 (a :: l’)
let rev’ l = rev_append l []
RAML output for rev (0.1s run time; steps metric):
3 + 9.5*M + 4.5*M^2
RAML output for rev’ (0.05s run time; steps metric):
7 + 9*M
Figure 12: Two implementations of rev from CompCert [46]. Both the derived bounds
are tight, one is linear and the other is quadratic.
energy consumption, and amount of data sent to the cloud. Details can be found in the source
code [29].
There are two main reasons for the difference between the runtime of the analysis per
function for the CompCert code (7.9s) and the other evaluated code (0.29s). First, the CompCert
code contains more complex data structures and we thus track more coefficients. Second, there
is a larger percentage of functions for which we cannot derive a bound (15.8% vs. 1.6%). As a
result, RAML looks for bounds of higher degree before giving up. Both leads to a larger number
of constraints to solve for the LP solver. Finally, there are a few outlier functions that cause an
unusually long analysis time. This is possibly due to a performance bug.
In general, the analysis is very efficient. RAML is slowing down if the analyzed program
contains many variables or functions with many arguments. Another source of complexity is
the maximal degree of polynomials in the search space. Depending on the complexity of the
program, the analysis becomes unusable when searching for bounds with maximal degree 7 − 9.
The efficiency could be improved by combining amortized resource analysis with data-flow
analyses and heuristics that predict the parts of the input that cause higher-degree resource
usage.
9 Case Study: Bounds for DynamoDB Queries
Having integrated the analysis with Inria’s OCaml compiler enables us to analyze and compile
real programs. An interesting use case of our resource bound analysis is to infer worst-case
bounds on DynamoDB queries. DynamoDB is a commercial NoSQL cloud database service,
which is part of Amazon Web Services (AWS). Amazon charges DynamoDB users on a combination of number of queries, transmitted fields, and throughput. Since DynamoDB is a NoSQL
9 Case Study: Bounds for DynamoDB Queries
33
service, it is often only possible to retrieve the whole table—which can be expensive for large
data sets—or single entries that are identified by a key value. The DynamoDB API is available
through the Opam package aws. We make the API available to the analysis by using tick functions
that specify resource usage. Since the query cost for different tables can be different, we provide
one function per action and table.
let db_query student_id course_id =
Raml.tick(1.0); Awslib.get_item ...
In the following, we describe the analysis of a specific OCaml application that uses a database
that contains a large table that stores grades of students for different courses. Our first function
computes the average grade of a student for a given list of courses.
let avge_grade student_id course_ids =
let f acc cid =
let (length,sum) = acc in
let grade = match db_query student_id cid with
| Some q → q
| None → raise (Not_found (student_id,cid))
in
(length +. 1.0, sum +. grade)
in
let (length,sum) = foldl f (0.0,0.0) course_ids in
sum /. length
In 0.03s RAML computes the tight bound 1·m where m is the length of the argument course_ids.
We omit the standard definitions of functions like foldl and map. However, they are not built-in
into our systems but the bounds are derived form first principles.
Next, we sort a given list of students based on the average grades in a given list of classes
using quick sort. As a first approximation we use a comparison function that is based on
average_grade.
let geq sid1 sid2 cour_ids =
avge_grade sid1 cour_ids >= avge_grade sid2 cour_ids
This results in O(n 2 m) database queries where n is the number of students and m is the number
of courses. The reason is that there are O(n 2 ) comparisons during a run of quick sort. Since
the resource usage of quick sort depends on the number of courses, we have to make the list of
courses an explicit argument and cannot store it in the closure of the comparison function.
let rec partition gt acc l =
match l with
| [] → let (cs,bs,_) = acc in (cs,bs)
| x::xs → let (cs,bs,aux) = acc in
let acc’ = if gt x aux then (cs,x::bs,aux)
else (x::cs,bs,aux)
in partition gt acc’ xs
let rec qsort gt aux l = match l with | [] → []
| x::xs →
let ys,zs = partition (gt x) ([],[],aux) xs in
append (qsort gt aux ys) (x::(qsort gt aux zs))
let sort_students s_ids c_ids = qsort geq c_ids s_ids
34
Towards Automatic Resource Bound Analysis for OCaml
In 0.31s RAML computes the tight bound n 2 m − nm for sort_students where n is the length of
the argument s_ids and m is the length of the argument c_ids. The negative factor arises from
the translation of the resource polynomials to the standard basis.
Given the alarming cubic bound, we reimplement our sorting function using memoization.
To this end we create a table that looks up and stores for each student and course the grade in
the DynamoDB. We then replace the function db_query with the function lookup.
let lookup sid cid table =
let cid_map = find (fun id → id = sid) table in
find (fun id → id = cid) cid_map
For the resulting sorting function, RAML computes the tight bound nm in 0.87s.
10 Related Work
Our work builds on past research on automatic amortized resource analysis (AARA). AARA has
been introduced by Hofmann and Jost for a strict first-order functional language with built-in
data types [35]. The technique has been applied to higher-order functional programs and user
defined types [42], to derive stack-space bounds [16], to programs with lazy evaluation [51, 56],
to object-oriented programs [36, 39], and to low-level code by integrating it with separation
logic [8]. All the aforementioned amortized-analysis–based systems are limited to linear bounds.
Hoffmann et al. [33, 30, 31] presented a multivariate AARA for a first-order language with built-in
lists and binary trees. Hofmann and Moser [38] have proposed a generalization of this system
in the context of (first-order) term rewrite systems. However, it is unclear how to automate
this system. In this article, we introduce the first AARA that is able to automatically derive
(multivariate) polynomial bounds that depend on user-defined inductive data structures. Our
system is the only one that can derive polynomial bounds for higher-order functions. Even for
linear bounds, our analysis is more expressive than existing systems for strict languages [42].
For instance, we can for the first time derive an evaluation-step bound for the curried append
function for lists. Moreover, we integrated AARA for the first time with an existing industrialstrength compiler.
Type systems for inferring and verifying resource bounds have been extensively studied. Vasconcelos et al. [55, 54] described an automatic analysis system that is based on sized-types [40]
and derives linear bounds for higher-order functional programs. Here we derive polynomial
bounds.
Dal Lago et al. [44, 45] introduced linear dependent types to obtain a complete analysis
system for the time complexity of the call-by-name and call-by-value lambda calculus. Crary
and Weirich [20] presented a type system for specifying and certifying resource consumption.
Danielsson [22] developed a library, based on dependent types and manual cost annotations,
that can be used for complexity analyses of functional programs. The advantage of our technique
is that it is fully automatic.
Classically, cost analyses are often based on deriving and solving recurrence relations. This
approach was pioneered by Wegbreit [57] and is actively studied for imperative languages [3, 7,
25, 5]. These works are not concerned with higher-order functions and bounds do not depend
on user-defined data structures.
Benzinger [11] has applied Wegbreit’s method in an automatic complexity analysis for Nuprl
terms. However, complexity information for higher-order functions has to be provided explicitly. Grobauer [26] reported a mechanism to automatically derive cost recurrences from DML
programs using dependent types. Danner et al. [24, 23] propose an interesting technique to
References
35
derive higher-order recurrence relations from higher-order functional programs. Solving the
recurrences is not discussed in these works and in contrast to our work they are not able to
automatically infer closed-form bounds.
Abstract interpretation based approaches to resource analysis [27, 12, 58, 52, 18] focus on
first-order integer programs with loops. Cicek et al. [19] study a type system for incremental
complexity.
In an active area of research, techniques from term rewriting are applied to complexity
analysis [10, 49, 15]; sometimes in combination with amortized analysis [37]. These techniques
are usually restricted to first-order programs and time complexity. Recently, Avanzini et al. [9]
proposed a complexity preserving defunctionaliztion to deal with higher-order programs. While
the transformation is asymptotically complexity preserving, it is unclear whether this technique
can derive bounds with precise constant factors.
Finally, there exists research that studies cost models to formally analyze parallel programs.
Blelloch and Greiner [13] pioneered the cost measures work and depth. There are more advanced
cost models that take into account caches and IO (see, e.g., Blelloch and Harper [14]), However,
these works do not provide machine support for deriving static cost bounds.
11 Conclusion
We have presented important first steps towards a practical automatic resource bound analysis
system for OCaml. Our three main contributions are (1) a novel automatic resource analysis
system that infers multivariate polynomial bounds that depend on size parameters of userdefined data structures, and (2) the first AARA that infers polynomial bounds for higher-order
functions, and (3) the integration of automatic amortized resource analysis with the OCaml
compiler.
As the title of this article indicates, there are many open problems left on the way to a
resource analysis system for OCaml that can be used in every-day development. In the future, we
plan to improve the bound analysis for programs with side-effects and exceptions. We will also
work on mechanisms that allow user interaction for manually deriving bounds if the automation
fails. Furthermore, we will work on taking into account garbage collection and the runtime
system when deriving time and space bounds. Finally, we will investigate techniques to link
the high-level bounds with hardware and the low-level code that is produced by the compiler.
These open questions are certainly challenging but we now have the tools to further push the
boundaries of practical quantitative software verification.
References
[1] E. Albert, P. Arenas, S. Genaim, M. Gómez-Zamalloa, and G. Puebla. Automatic Inference
of Resource Consumption Bounds. In Logic for Programming, Artificial Intelligence, and
Reasoning, 18th Conference (LPAR’12), pages 1–11, 2012.
[2] E. Albert, P. Arenas, S. Genaim, and G. Puebla. Closed-Form Upper Bounds in Static Cost
Analysis. Journal of Automated Reasoning, pages 161–203, 2011.
[3] E. Albert, P. Arenas, S. Genaim, G. Puebla, and D. Zanardini. Cost Analysis of Java Bytecode.
In 16th Euro. Symp. on Prog. (ESOP’07), pages 157–172, 2007.
36
Towards Automatic Resource Bound Analysis for OCaml
[4] E. Albert, P. Arenas, S. Genaim, G. Puebla, and D. Zanardini. Cost Analysis of ObjectOriented Bytecode Programs. Theor. Comput. Sci., 413(1):142 – 159, 2012.
[5] E. Albert, J. C. Fernández, and G. Román-Díez. Non-cumulative Resource Analysis. In Tools
and Algorithms for the Construction and Analysis of Systems - 21st International Conference,
(TACAS’15), pages 85–100, 2015.
[6] C. Alias, A. Darte, P. Feautrier, and L. Gonnord. Multi-dimensional Rankings, Program
Termination, and Complexity Bounds of Flowchart Programs. In 17th Int. Static Analysis
Symposium (SAS’10), pages 117–133, 2010.
[7] D. E. Alonso-Blas and S. Genaim. On the limits of the classical approach to cost analysis. In
19th Int. Static Analysis Symp. (SAS’12), pages 405–421, 2012.
[8] R. Atkey. Amortised Resource Analysis with Separation Logic. In 19th Euro. Symp. on Prog.
(ESOP’10), pages 85–103, 2010.
[9] M. Avanzini, U. D. Lago, and G. Moser. Analysing the Complexity of Functional Programs:
Higher-Order Meets First-Order. In 29th Int. Conf. on Functional Programming (ICFP’15),
2012.
[10] M. Avanzini and G. Moser. A Combination Framework for Complexity. In 24th International
Conference on Rewriting Techniques and Applications (RTA’13), pages 55–70, 2013.
[11] R. Benzinger. Automated Higher-Order Complexity Analysis. Theor. Comput. Sci., 318(12):79–103, 2004.
[12] R. Blanc, T. A. Henzinger, T. Hottelier, and L. Kovács. ABC: Algebraic Bound Computation
for Loops. In Logic for Prog., AI., and Reasoning - 16th Int. Conf. (LPAR’10), pages 103–118,
2010.
[13] G. E. Blelloch and J. Greiner. A Provable Time and Space Efficient Implementation of NESL.
In 1st Int. Conf. on Funct. Prog. (ICFP’96), pages 213–225, 1996.
[14] G. E. Blelloch and R. Harper. Cache and I/O Efficent Functional Algorithms. In 40th ACM
Symp. on Principles Prog. Langs. (POPL’13), pages 39–50, 2013.
[15] M. Brockschmidt, F. Emmes, S. Falke, C. Fuhs, and J. Giesl. Alternating Runtime and Size
Complexity Analysis of Integer Programs. In Tools and Alg. for the Constr. and Anal. of
Systems - 20th Int. Conf. (TACAS’14), pages 140–155, 2014.
[16] B. Campbell. Amortised Memory Analysis using the Depth of Data Structures. In 18th Euro.
Symp. on Prog. (ESOP’09), pages 190–204, 2009.
[17] Q. Carbonneaux, J. Hoffmann, and Z. Shao. Compositional Certified Resource Bounds. In
36th Conf. on Prog. Lang. Design and Impl. (PLDI’15), 2015.
[18] P. Cerný, T. A. Henzinger, L. Kovács, A. Radhakrishna, and J. Zwirchmayr. Segment Abstraction for Worst-Case Execution Time Analysis. In 24th European Symposium on Programming (ESOP’15), pages 105–131, 2015.
[19] E. Çiçek, D. Garg, and U. A. Acar. Refinement Types for Incremental Computational Complexity. In 24th European Symposium on Programming (ESOP’15), pages 406–431, 2015.
References
37
[20] K. Crary and S. Weirich. Resource Bound Certification. In 27th ACM Symp. on Principles of
Prog. Langs. (POPL’00), pages 184–198, 2000.
[21] S. A. Crosby and D. S. Wallach. Denial of service via algorithmic complexity attacks. In 12th
USENIX Security Symposium (USENIX’12), 2003.
[22] N. A. Danielsson. Lightweight Semiformal Time Complexity Analysis for Purely Functional
Data Structures. In 35th ACM Symp. on Principles Prog. Langs. (POPL’08), pages 133–144,
2008.
[23] N. Danner, D. R. Licata, and R. Ramyaa. Denotational Cost Semantics for Functional
Languages with Inductive Types. In 29th Int. Conf. on Functional Programming (ICFP’15),
2012.
[24] N. Danner, J. Paykin, and J. S. Royer. A Static Cost Analysis for a Higher-Order Language. In
7th Workshop on Prog. Languages Meets Prog. Verification (PLPV’13), pages 25–34, 2013.
[25] A. Flores-Montoya and R. Hähnle. Resource Analysis of Complex Programs with Cost
Equations. In Programming Languages and Systems - 12th Asian Symposiu (APLAS’14),
pages 275–295, 2014.
[26] B. Grobauer. Cost Recurrences for DML Programs. In 6th Int. Conf. on Funct. Prog. (ICFP’01),
pages 253–264, 2001.
[27] S. Gulwani, K. K. Mehra, and T. M. Chilimbi. SPEED: Precise and Efficient Static Estimation
of Program Computational Complexity. In 36th ACM Symp. on Principles of Prog. Langs.
(POPL’09), pages 127–139, 2009.
[28] J. Hoffmann. Types with Potential: Polynomial Resource Bounds via Automatic Amortized
Analysis. PhD thesis, Ludwig-Maximilians-Universität München, 2011.
[29] J. Hoffmann. RAML Web Site. http://raml.co, 2015.
[30] J. Hoffmann, K. Aehlig, and M. Hofmann. Multivariate Amortized Resource Analysis. In
38th ACM Symp. on Principles of Prog. Langs. (POPL’11), 2011.
[31] J. Hoffmann, K. Aehlig, and M. Hofmann. Multivariate Amortized Resource Analysis. ACM
Trans. Program. Lang. Syst., 2012.
[32] J. Hoffmann and M. Hofmann. Amortized Resource Analysis with Polymorphic Recursion and Partial Big-Step Operational Semantics. In Prog. Langs. and Systems - 8th Asian
Symposium (APLAS’10), 2010.
[33] J. Hoffmann and M. Hofmann. Amortized Resource Analysis with Polynomial Potential. In
19th Euro. Symp. on Prog. (ESOP’10), 2010.
[34] J. Hoffmann and Z. Shao. Type-Based Amortized Resource Analysis with Integers and
Arrays. In 12th International Symposium on Functional and Logic Programming (FLOPS’14),
2014.
[35] M. Hofmann and S. Jost. Static Prediction of Heap Space Usage for First-Order Functional
Programs. In 30th ACM Symp. on Principles of Prog. Langs. (POPL’03), pages 185–197, 2003.
38
Towards Automatic Resource Bound Analysis for OCaml
[36] M. Hofmann and S. Jost. Type-Based Amortised Heap-Space Analysis. In 15th Euro. Symp.
on Prog. (ESOP’06), pages 22–37, 2006.
[37] M. Hofmann and G. Moser. Amortised Resource Analysis and Typed Polynomial Interpretations. In Rewriting and Typed Lambda Calculi (RTA-TLCA;14), pages 272–286, 2014.
[38] M. Hofmann and G. Moser. Multivariate Amortised Resource Analysis for Term Rewrite
Systems. In 13th International Conference on Typed Lambda Calculi and Applications
(TLCA’15), pages 241–256, 2015.
[39] M. Hofmann and D. Rodriguez. Automatic Type Inference for Amortised Heap-Space
Analysis. In 22nd Euro. Symp. on Prog. (ESOP’13), pages 593–613, 2013.
[40] J. Hughes, L. Pareto, and A. Sabry. Proving the Correctness of Reactive Systems Using Sized
Types. In 23th ACM Symp. on Principles of Prog. Langs. (POPL’96), pages 410–423, 1996.
[41] G. Jin, L. Song, X. Shi, J. Scherpelz, and S. Lu. Understanding and Detecting Real-World
Performance Bugs. In Conference on Programming Language Design and Implementation
PLDI’12, pages 77–88, 2012.
[42] S. Jost, K. Hammond, H.-W. Loidl, and M. Hofmann. Static Determination of Quantitative
Resource Usage for Higher-Order Programs. In 37th ACM Symp. on Principles of Prog. Langs.
(POPL’10), pages 223–236, 2010.
[43] P. C. Kocher. Timing Attacks on Implementations of Diffie-Hellman, RSA, DSS, and Other
Systems. In Advances in Cryptology - 16th Annual International Cryptology Conference
(CRYPTO’96), pages 104–113, 1996.
[44] U. D. Lago and M. Gaboardi. Linear Dependent Types and Relative Completeness. In 26th
IEEE Symp. on Logic in Computer Science (LICS’11), pages 133–142, 2011.
[45] U. D. Lago and B. Petit. The Geometry of Types. In 40th ACM Symp. on Principles Prog.
Langs. (POPL’13), pages 167–178, 2013.
[46] X. Leroy. Formal Verification of a Realistic Compiler. Communications of the ACM, 52(7):107–
115, 2009.
[47] X. Leroy, D. Doligez, A. Frisch, J. Garrigue, D. Rémy, and J. Vouillon. The OCaml system
release 4.02. Technical report, Institut National de Recherche en Informatique et en Automatique, 2014. http://caml.inria.fr/distrib/ocaml-4.02/ocaml-4.02-refman.pdf.
[48] V. Nicollet. 99 problems (solved) in ocaml. https://ocaml.org/learn/tutorials/
99problems.html.
[49] L. Noschinski, F. Emmes, and J. Giesl. Analyzing Innermost Runtime Complexity of Term
Rewriting by Dependency Pairs. J. Autom. Reasoning, 51(1):27–56, 2013.
[50] O. Olivo, I. Dillig, and C. Lin. Static Detection of Asymptotic Performance Bugs in Collection
Traversals. In Conference on Programming Language Design and Implementation (PLDI’15),
pages 369–378, 2015.
[51] H. R. Simões, P. B. Vasconcelos, M. Florido, S. Jost, and K. Hammond. Automatic Amortised
Analysis of Dynamic Memory Allocation for Lazy Functional Programs. In 17th Int. Conf.
on Funct. Prog. (ICFP’12), pages 165–176, 2012.
References
39
[52] M. Sinn, F. Zuleger, and H. Veith. A Simple and Scalable Approach to Bound Analysis and
Amortized Complexity Analysis. In Computer Aided Verification - 26th Int. Conf. (CAV’14),
page 743–759, 2014.
[53] R. J. Vanderbei. Linear Programming: Foundations and Extensions. Springer US, 2001.
[54] P. Vasconcelos. Space Cost Analysis Using Sized Types. PhD thesis, School of Computer
Science, University of St Andrews, 2008.
[55] P. B. Vasconcelos and K. Hammond. Inferring Costs for Recursive, Polymorphic and HigherOrder Functional Programs. In Int. Workshop on Impl. of Funct. Langs. (IFL’03), pages
86–101. Springer-Verlag LNCS, 2003.
[56] P. B. Vasconcelos, S. Jost, M. Florido, and K. Hammond. Type-Based Allocation Analysis for
Co-recursion in Lazy Functional Languages. In 24th European Symposium on Programming
(ESOP’15), pages 787–811, 2015.
[57] B. Wegbreit. Mechanical Program Analysis. Commun. ACM, 18(9):528–539, 1975.
[58] F. Zuleger, M. Sinn, S. Gulwani, and H. Veith. Bound Analysis of Imperative Programs with
the Size-change Abstraction. In 18th Int. Static Analysis Symp. (SAS’11), pages 280–297,
2011.
40
Towards Automatic Resource Bound Analysis for OCaml
A Experimental Results
A.1 Analyzed Functions
Name
Type
File: WorkingWithLists.raml (99 Problems in OCaml)
last
a list -> ’a option
lastTwo
a list -> (’a * ’a) option
at
int -> ’a list -> ’a option
natAt
nat -> ’a list -> ’a option
LOC
Description
8
10
7
5
length
rev
eqList
isPalindrome
flatten
compress
pack
a list -> int
a list -> ’a list
a list -> ’a list -> bool
a list -> bool
a node list -> ’a list
a list -> ’a list
a list -> ’a list list
5
6
13
2
8
10
16
encode
decode
duplicate
replicate
drop
split
slice
concat
rotate
removeAt
insertAt
constructList
random
min
randSelect
lottoSelect
a list -> (int * ’a) list
(int * ’a) list -> ’a list
a list -> ’a list
a list -> int -> ’a list
a list -> int -> ’a list
a list -> int -> ’a list * ’a list
a list -> int -> int -> ’a list
a list -> ’a list -> ’a list
a list -> int -> ’a list
a list -> int -> ’a list
a list -> int -> ’a -> ’a list
int -> int -> ’a list
int -> int
int -> int -> int
a list -> int -> ’a list
int -> int -> int list
15
17
4
15
7
10
15
4
8
4
4
7
1
4
21
1
snd
fst
map
insert
sort
compare
lengthSort
File: LogicAndCodes.raml
eval2
table2
a * ’b -> ’b
a * ’b -> ’a
(’a -> ’b) -> ’a list -> ’b list
(’a -> ’a -> int) -> ’a -> ’a list -> ’a list
(’a -> ’a -> int) -> ’a list -> ’a list
int -> int -> int
a list list -> ’a list list
4
4
4
4
4
5
4
returns the last element of list
returns last two elements of list
outputs element at a particular location
at implemented using natural numbers
defined inductively
returns length of list
reverses the list
checks equality of two lists
checks if list is a palindrome
flattens a tree into a list
deletes successive duplicates
packs successive duplicates into an inner list
run-length encoding of list
decodes a run-length encoding of list
duplicates each element of list
replicates each element of list n times
drops every n-th element
splits list into two lists
extracts a slice from list
concatenates two lists
rotates a list by n positions
removes a list at n-th position
inserts an element at n-th position
constructs a list from 1st to 2nd element
generates a random number
returns min of two integers
generates a random permutation of list
composes randSelect with constructList
returns second element of a product
returns first element of a product
applies f to every element of list
sort helper
sorts list according to compare function
compares two integers
sorts list of lists according to size of list
int -> bool -> int -> bool -> boolExpr
int -> int -> boolExpr -> (bool * bool *
bool) list
int -> (int * ’a) list -> ’a
7
12
table2 helper
constructs truth table of expression
8
(int * bool) list -> boolExpr -> bool
(int * bool) list -> int list -> boolExpr ->
((int * bool) list * bool) list
13
20
returns element of list corresponding to
key
evaluates a boolean expression
a list -> int
a list -> int -> ’a
a list -> ’b list -> int -> ’a * ’b
float list -> float list -> float -> float list
float list -> float list -> int -> float list
4
6
9
8
4
assoc
eval
tableMake
File: echelon_form.raml
size
getElem
get2Elems
subtract_row_helper
subtract_row
returns size of list
returns i-th element of list
returns i-th element of 2 lists
echelon helper
echelon helper
A Experimental Results
subtract_helper
41
4
echelon helper
concat
tail
float list list -> float list -> int -> float list
list
a list -> ’a list -> ’a list
a list -> int -> ’a list
4
6
hd_helper
reverse_helper
reverse
head
split_helper
split
a list -> int -> ’a list -> ’a list
a list -> ’a list -> ’a list
a list -> ’a list
a list -> int -> ’a list
a list -> int -> int -> ’a list * ’a list
a list -> int -> ’a list * ’a list
6
4
6
14
6
8
subtract
echelon_helper
float list list -> int -> float list list
float list list -> int -> ’a list -> float list
list
float list list -> float list list
3
3
concatenates two lists
returns the list excluding the first i elements
echelon helper
echelon helper
reverses a list
returns the first i elements of list
echelon helper
splits the list at i-th position and returns
the two lists
subtract a row
echelon helper
102
takes a matrix with m rows and n
columns and reduces it to an upper triangular matrix
a list list -> int -> bool
int -> int -> ’a list list -> bool
int * int * ’a list list -> bool
int -> int -> ’a list list -> int * int * ’a list
list
a * ’b * ’c list list -> int -> int -> ’c
int -> int -> int -> int
int list -> int list -> int -> int list
int list list -> int list list -> int -> int list
list
int * int * ’a list list -> int * int * ’b list list
-> bool
int * int * int list list -> int * int * int list
list -> int * int * int list list
int * int * int list list -> int * int * int list
list -> int * int * int list list
a list -> ’a -> ’a list
a list list -> ’a list -> ’a list list
a list list -> ’a list list -> ’a list list
a list list -> ’a list list
int list -> int list -> int
int list -> int list list -> int list list
int list list -> int list list -> int list list
7
4
3
5
matrix helper
matrix helper
matrix helper
matrix helper
4
4
7
7
returns (i,j)-th element of matrix mat
matrix helper
matrix helper
matrix helper
5
matrix helper
49
adds two matrices m1 and m2
49
subtracts two matrices m1 and m2
4
11
4
17
7
4
33
7
7
4
5
matrix helper
matrix helper
50
delete
submat
int -> int list -> int list -> int list
int list -> int list list -> int list -> int list
list
int list list -> int list list -> int list list
int * int * ’a list list -> int * int * ’b list list
-> bool
int * int * int list list -> int * int * int list
list -> int * int * int list list
a list -> int -> ’a list
a list list -> int -> int -> ’a list list
appends x to the end of list l
appends column col to matrix m
matrix helper
takes transpose of matrix m
matrix helper
matrix helper
multiplies matrices m1 and m2 (naive
implementation)
matrix helper
matrix helper
multiplies matrices m1 and m2 after
performing dimensional sanity checks
deletes the i-th element of list
deletes the i-th row and j-th column of
matrix m
File: power_radio.raml
sendmsg msg
main1_events
int list -> unit
event list -> unit
13
15
main2_events
event list -> unit
15
echelon_form
File: matrix.raml
check_lists
check_mat
check_matrix
construct_matrix
getElemMatrix
op
rec_list
rec_mat
check_sanity
plus
minus
append
append_col
transpose_helper
transpose
prod
prod_mat
mult_slow
lineMult
computeLine
mat_mult_jan
check_mult_sanity
mult
7
14
sends a list of integers
sends sensor data as soon as it is produced
stores sensor data in buffer and sends
only at specific events
42
Towards Automatic Resource Bound Analysis for OCaml
main3_events
main4_events
event list -> unit
event list -> unit
15
15
main5_events
event list -> unit
36
File: avanzini.raml
partition
(’a -> bool) -> ’a list -> ’a list * ’a list
9
quicksort
(’a -> ’a -> bool) -> ’a list -> ’a list
18
rev_sort
a -> ’a -> bool -> (’b * ’a list) either list
-> (’b * ’a list) either list
17
a list list -> ’a list
a list list list -> ’a list
a list list list list -to ’a list
8
12
16
concatenates inner lists
concatenates innermost lists
concatenates innermost lists
btree -> int -> btree option
btree -> int -> btree option
15
34
depth-first search for binary trees
breadth-first search for binary trees
a list -> (’a * ’a) list
12
generate ordered pairs from a list
bin list -> bin list
bin list -> Rnat.t -> bin list
bin list -> ’a list -> bin list
15
24
28
increment a binary counter
n increments to a binary counter
n increments to a binary counter
(Rnat.t -> unit) -> Rnat.t -> unit
(’a -> Rnat.t -> ’a) -> ’a -> Rnat.t -> ’a
(’a -> ’a) array -> ’a -> ’a
3
5
12
calls a function f (i) for 0<=i<n
fold for natural numbers
successively apply all functions stored
in an array
File: calculator.raml
add
sub
mult
eval_simpl
nat -> nat -> nat
nat -> nat -> nat
nat -> nat -> nat
expr -> nat
4
8
9
26
eval
expr -> nat
35
recursively add two natural numbers
recursively subtract two nats
recursively multiply two nats
a simple evaluater for arithmetic expressions
a evaluater for arithmetic expressions
without aux. funs
a list -> ’a list * ’a list
(’a -> ’a -> bool) -> ’a list -> ’a list -> ’a
list
(’a -> ’a -> bool) -> Rnat.t -> ’a list -> ’a
list
Rnat.t -> int list list -> int list list
8
11
splits a list in the middle
merges two sorted lists
32
merge sort
45
merge sort for lists of lists
(’a -> bool) -> ’a list -> ’a list * ’a list
(’a -> ’a -> bool) -> ’a list -> ’a list
(int * int) list -> (int * int) list
int list list -> int list list
10
19
28
32
partition for quick sort
quick sort
quick sort for integer pairs
quick sort for lists of lists
int -> binary list -> int
30
exponentiation via squaring
int list -> int list -> int
45
longest common subsequence with dynamic programming
(’a -> ’b) -> (’c -> ’d) -> (’a * ’c) ablist ->
(’b * ’d) ablist
(’a -> ’a -> bool) -> (’a list * ’b) ablist ->
(’a list * ’b) ablist
5
map for AB-lists
29
sort inner lists in A-nodes
File: append_all.raml
append_all
append_all2
append_all3
File: bfs.raml
dfs
bfs
File: rev_pairs.raml
pairs
File: binary_counter.raml
add_one
add_many
add_list
File: array_fun.raml
nat_iterate
nat_fold
apply_all
File: mergesort.raml
split
merge
mergesort
mergesort_list
File: quicksort.raml
partition
quicksort
quicksort_pairs
quicksort_list
File: square_mult.raml
square_mult
File: subsequence.raml
lcs
File: running.raml
abmap
asort
same as main2 with debugging mode
same as main3 with function application
same as main4 but data is sent only after specific time intervals
partitions list l into two depending on
whether the elements satisfy function f
performs quick sort on list l using comp
as comparator
A Experimental Results
asort’
btick
abfoldr
cons_all
File: ocaml_sort.raml
merge
list
File: ocaml_list.raml
length
cons
hd
tl
nth
append
rev_append
rev
flatten
concat
map
mapi
rev_map
iter
iteri
fold_left
fold_right
map2
rev_map2
iter2
fold_left2
fold_right2
for_all
exists
for_all2
exists2
mem
memq
assoc
assq
mem_assoc
mem_assq
remove_assoc
remove_assq
find
find_all
filter
partition
split
combine
merge
chop
stable_sort
sort
fast_sort
43
(’a -> ’a -> bool) -> (’a list * ’b) ablist ->
(’a list * ’c) ablist
(int list * int) ablist -> (int list * int) ablist
(’a -> ’b -> ’b) -> (’c -> ’b -> ’b) -> ’b ->
(’a * ’c) ablist -> ’b
(int * int) ablist -> (’a * int) ablist
29
6
10
sort inner lists in A-nodes; raise exception on B-node
use map to tick 2.5 at every B-node
fold for AB-lists
18
two nested folds
(’a -> ’a -> bool) -> ’a list -> ’a list -> ’a
list
(’a -> ’a -> bool) -> ’a list -> ’a list
9
merge two lists
31
interesting variant of merge sort
a list -> int
a -> ’a list -> ’a list
a list -> ’a
a list -> ’a list
a list -> int -> ’a
a list -> ’a list -> ’a list
a list -> ’a list -> ’a list
a list -> ’a list
a list list -> ’a list
a list list -> ’a list
(’a -> ’b) -> ’a list -> ’b list
(int -> ’a -> ’b) -> ’a list -> ’b list
(’a -> ’b) -> ’a list -> ’b list
(’a -> ’b) -> ’a list -> unit
(int -> ’a -> ’b) -> ’a list -> unit
(’a -> ’b -> ’a) -> ’a -> ’b list -> ’a
(’a -> ’b -> ’b) -> ’a list -> ’b -> ’b
(’a -> ’b -> ’c) -> ’a list -> ’b list -> ’c list
(’a -> ’b -> ’c) -> ’a list -> ’b list -> ’c list
(’a -> ’b -> ’c) -> ’a list -> ’b list -> unit
(’a -> ’b -> ’c -> ’a) -> ’a -> ’b list -> ’c
list -> ’a
(’a -> ’b -> ’c -> ’c) -> ’a list -> ’b list ->
’c -> ’c
(’a -> bool) -> ’a list -> bool
(’a -> bool) -> ’a list -> bool
(’a -> ’b -> bool) -> ’a list -> ’b list ->
bool
(’a -> ’b -> bool) -> ’a list -> ’b list ->
bool
a -> ’b list -> bool
a -> ’b list -> bool
a -> (’b * ’c) list -> ’c
a -> (’b * ’c) list -> ’c
a -> (’b * ’c) list -> bool
a -> (’b * ’c) list -> bool
a -> (’b * ’c) list -> (’b * ’c) list
a -> (’b * ’c) list -> (’b * ’c) list
(’a -> bool) -> ’a list -> ’a
(’a -> bool) -> ’a list -> ’a list
(’a -> bool) -> ’a list -> ’a list
(’a -> bool) -> ’a list -> ’a list * ’a list
(’a * ’b) list -> ’a list * ’b list
a list -> ’b list -> (’a * ’b) list
(’a -> ’a -> int) -> ’a list -> ’a list -> ’a list
int -> ’a list -> ’a list
(’a -> ’a -> int) -> ’a list -> ’a list
(’a -> ’a -> int) -> ’a list -> ’a list
(’a -> ’a -> int) -> ’a list -> ’a list
5
1
3
3
7
4
4
6
7
8
3
5
6
3
5
4
4
12
15
12
12
length of a list
list cons
head of a list
tail of a list
nth element of a list
list append
12
right fold with two lists
3
3
12
check condition for all list elems
check condition for one list elem
as for_all but for two lists
12
as exists but for two lists
3
3
5
5
5
5
5
5
3
7
8
5
5
12
10
6
106
107
107
checks if elem is in list
same as mem but uses equality
lookup for key-value pairs
as assoc but uses equality
like mem but for pairs
like memq but for pairs
filter varient using compare
filter varient using equality
list find
list find that returns all matches
list filter
list partition
split a list of pairs
zip two lists
merge for merge sort
take the first an elements
merge sort
merge sort
merge sort
reverses a list
flattens a list
flattens a list
list map
list map with index
reverse and map
iterate over a list
iterate with index
list fold
list fold
list map with two lists
reverse and map with two lists
iterate over two lists
left fold with two lists
44
Towards Automatic Resource Bound Analysis for OCaml
File: aws.raml
average_grade
int -> int list -> float
17
greater_eq
int -> int -> int list -> bool * int list
19
sort_students
int list -> int list -> int list
52
make_table
int list -> int list -> (int * (int * float) list)
list
(’a -> bool) -> (’a * ’b) list -> ’b
int -> int -> (int * (int * ’a) list) list -> ’a
int -> int list -> (int * (int * float) list) list
-> float * (int * (int * float) list) list
int list -> int -> int -> (int * (int * float)
list) list] -> bool * (int * (int * float) list)
list
int list -> int list -> int list
22
9
12
24
avarage by looking up all grades in DynanmoDB
compare students by looking up grades
at DynamoDB
sort students based on avarage grade
using greater_eq
look up grades in DynamoDB and memoize all the grades in a tables
find a value by looking up the key
look up a grade in a table
avarage grade using a look-up table
28
greater_eq using a look-up table
76
sorting using a look-up table
a list list -> ’a list
9
appendAll2
a list list list -> ’a list
14
appendAll3
a list list list list -> ’a list
18
collapses all elements of a 2D matrix
into a list
collapses all elements of a 3D matrix
into a list
collapses all elements of a 4D matrix
into a list
File: duplicates.raml
eq
remove
nub
int list -> int list -> bool
int list -> int list list -> int list list
int list list -> int list list
15
5
27
checks if two lists are equal
duplicates helper
removes duplicate lists from a list of
lists
File: dyade.raml
multList
int -> int list -> int list
4
dyade
int list -> int list -> int list list
8
multiplies all elements of a list with a
constant
multiplies all elements of two lists to
form a 2D matrix
File: eratosthenes.raml
filter
int -> int list -> int list
8
eratos
int list -> int list
3
int list -> int -> int
int list -> int
int -> int -> int -> int * int
int list -> int list -> int -> int list
int list -> int list -> int list
int -> int -> int -> int * int
int list -> int list -> int -> int list * int
int list -> int list -> int list
int list -> int list -> int list
int list -> int list -> int
int list -> int list -> bool
4
6
7
9
26
2
11
17
8
14
16
bitvector helper
converts bit vector to integer
bitvector helper
bitvector helper
adds two bitvectors
bitvector helper
tree -> int list
int -> int list -> int list
int list -> int list
tree -> int list
8
7
12
20
collapses tree into a list
inserts element in a sorted list
performs insertion sort on list
performs insertion sort after flattening
the tree
find
lookup
average_grade’
greater_eq’
sort_students_efficient
File: PROTOTYPE
File: appendAll.raml
appendAll
File: bitvectors.raml
bitToInt’
bitToInt
sum
add’
add
diff
sub’
sub
mult
compare
leq
File: flatten.raml
flatten
insert
insertionsort
flattensort
File: listsort.raml
deletes all elements in list divisible by
the first argument
runs the sieve of Eratosthenes algorithm on the list
subtracts two bitvectors
multiplies two bitvectors
bitvector helper
compares two bitvectors
A Experimental Results
isortlist
int list list -> int list list
21
performs an insertion sort on list of lists,
where lists are compared lexicographically
4
4
2
15
8
47
returns first line of zeros
lcs helper
computes max of two integers
computes new line recursively
computes length of table
computes longest common subsequence of two lists
a list -> ’a list * ’a list
int list -> int list -> int list
int list -> int list
int list -> int list
9
10
15
19
splits list into two
merges two sorted lists
buggy version of mergesort
correct version of mergesort
int list -> int list
int list -> int list
10
14
helper for selection sort
performs selection sort on list
a -> ’b list * ’c list
a -> ’a list * ’b -> ’a list * ’b
a list -> ’a list * ’b -> ’a list * ’b
2
3
4
5
10
17
returns an empty list
enqueues element into list
enqueues a list of trees into a queue of
trees
dequeue helper
dequeues element from queue
constructs a node of tree
7
bfs helper
4
13
4
performs breadth first search on tree
dfs helper
performs depth first search on tree
6
10
partitions the list for performing quick
sort
performs a mutually recursive implementation of quicksort as presented by
Hongwei Xi
10
10
zips 3 lists together
groups list into list of triples
4
2
6
4
7
8
7
4
4
reverse helper
reverses a list
mutual recursion with g
mutual recursion with f
function in the size change paper
re-implementation of f2
f2 reimplemented
helper
late starting descending parameters
12
split helper
4
11
7
6
splits values according to keys
quicksort helper
performs quicksort on list
sorts all value lists with minSort
File: longestCommonSubsequence.raml
firstline
a list -> int list
right
int list -> int
max
int -> int -> int
newline
int -> int list -> int list
lcstable
int list -> int list -> int list list
lcs
int list -> int list -> int
File: mergesort.raml
msplit
merge
mergesortBuggy
mergesort
File: minsort.raml
findMin
minSort
File: queue.raml
empty
enqueue
enqueues
45
copyover
dequeue
children
a list * ’a list -> ’b list * ’a list
a list * ’a list -> (’a list * ’a list) * ’a list
a * ’b * ’b list * ’b list -> (’a * ’b) * (’b * ’b
* ’b list * ’b list) list
breadth
(’a * ’a * ’a list * ’a list) list * (’a * ’a * ’a
list * ’a list) list -> (’a * ’a) list
startBreadth
a list -> (’a * ’a) list
depth
a * ’a * ’a list * ’a list -> (’a * ’a) list
startDepth
a list -> (’a * ’a) list
File: quicksort_mutual.raml
part
int -> int list -> int list -> int list -> int
list
quicksortMutual
int list -> int list
File: rationalPotential.raml
zip3
a list -> ’b list -> ’c list -> (’a * ’b * ’c) list
group3
a list -> (’a * ’a * ’a) list
File: sizechange.raml
r1
a list -> ’a list -> ’a list
rev
a list -> ’a list
f
a list -> ’a list -> ’a list
g
a list -> ’a list -> ’a list -> ’a list
f2
a list -> ’a list -> ’a list
last
a list -> ’a list
f2’
a list -> ’a list -> ’a list
g3
a list -> ’a list -> ’a list
f3
a list -> ’a list -> ’a list
File: splitandsort.raml
insert
a * int -> (’a list * int) list -> (’a list * int)
list
split
(’a * int) list -> (’a list * int) list
splitqs
int * int list -> int list * int list
quicksort
int list -> int list
sortAll
(int list * ’a) list -> (int list * ’a) list
46
splitAndSort
File: subtrees.raml
subtrees
File: tuples.raml
attach
pairs
pairsAux
pairsSlow
triples
quadruples
File: array_dijkstra.raml
makeGraph
dijkstra
File: CompCert
File: String0.ml
string_dec
prefix
File: Tuples.ml
uncurry
File: Specif.ml
projT1
projT2
value
File: EquivDec.ml
equiv_dec
File: Datatypes.ml
implb
xorb
negb
fst
snd
length
app
coq_CompOpp
coq_CompareSpec2Type
coq_CompSpec2Type
File: Bool.ml
bool_dec
eqb
iff_reflect
File: Ring.ml
bool_eq
File: Peano.ml
plus
max
min
nat_iter
File: List0.ml
hd
tl
in_dec
nth_error
remove
rev
rev_append
rev’
Towards Automatic Resource Bound Analysis for OCaml
(int * int) list -> (int list * int) list
3
splits list according to keys, then sorts
the inner lists
tree -> tree list
16
generates a list of all subtrees of a tree
a -> ’b list -> (’a * ’b) list
4
a list -> (’a * ’a) list
a list -> (’a * ’a) list -> (’a * ’a) list
a list -> (’a * ’a) list
a list -> (’a * ’a * ’a) list
a list -> (’a * ’a * ’a * ’a) list
4
4
4
4
4
attaches the first argument to every element of the list
generates all distinct pairs in list
helper for pairsSlow
slow implementation of pairs
generates all distinct triples in list
generates all distinct quadruples in list
Rnat.t -> (Rnat.t * Rnat.t * int) list -> int
array array
int array array -> Rnat.t -> int array
35
int list -> int list -> bool
int list -> int list -> bool
10
7
a list -> ’b -> ’c -> ’d
4
(’a * ’b) sigT -> ’a
(’a * ’b) sigT -> ’b
a -> ’a option
3
3
3
a -> ’a
3
bool -> bool -> bool
bool -> bool -> bool
bool -> bool
a * ’b -> ’a
a * ’b -> ’b
a list -> nat
a list -> ’a list -> ’a list
comparison -> comparison
comparison -> coq_CompareSpecT
a -> ’b -> comparison ->
coq_CompareSpecT
2
2
2
2
2
4
4
3
3
2
bool -> bool -> bool
bool -> bool -> bool
bool -> reflect
2
2
2
bool -> bool -> bool
2
nat -> nat -> nat
nat -> nat -> nat
nat -> nat -> nat
nat -> (’a -> ’a) -> ’a -> ’a
4
7
7
4
a -> ’a list -> ’a
a list -> ’a
(’a -> ’b -> bool) -> ’b -> ’a list -> bool
a list -> nat -> ’a option
(’a -> ’b -> bool) -> ’a -> ’b list -> ’b list
a list -> ’a list
a list -> ’a list -> ’a list
a list -> ’a list
3
3
3
7
4
3
4
6
64
creates an array based weighted graph
from a list
Dijkstra’s shortest-path algorithm
A Experimental Results
list_eq_dec
map
fold_left
fold_right
existsb
forallb
filter
File: EqNat.ml
beq_nat
File: Compare_dec.ml
le_lt_dec
le_gt_dec
nat_compare
File: BinPosDef.ml
succ
add
add_carry
pred_double
pred
pred_N
mask_rect
mask_rec
succ_double_mask
double_mask
double_pred_mask
pred_mask
sub_mask
sub_mask_carry
sub
mul
iter
pow
square
div2
div2_up
size_nat
size
compare_cont
compare
min
max
eqb
leb
ltb
sqrtrem_step
sqrtrem
sqrt
gcdn
gcd
ggcdn
ggcd
coq_Nsucc_double
coq_Ndouble
coq_lor
coq_land
47
((’a -> ’b) -> bool) -> ’a list -> ’b list ->
bool
(’a -> ’b) -> ’a list -> ’b list
((’a -> ’b) -> ’a) -> ’b list -> ’a -> ’a
((’a -> ’b) -> ’b) -> ’b -> ’a list -> ’b
(’a -> bool) -> ’a list -> bool
(’a -> bool) -> ’a list -> bool
(’a list -> bool) -> ’a list -> ’a list
10
nat -> nat -> bool
10
nat -> nat -> bool
nat -> nat -> bool
nat -> nat -> comparison
7
7
10
positive -> positive
positive -> positive -> positive
positive -> positive -> positive
positive -> positive
positive -> positive
positive -> coq_N
a -> (positive -> ’a) -> ’a -> mask -> ’a
a -> (positive -> ’a) -> ’a -> mask -> ’a
mask -> mask
mask -> mask
positive -> mask
mask -> mask
positive -> positive -> mask
positive -> positive -> mask
positive -> positive -> positive
positive -> positive -> positive
positive -> (’a -> ’a) -> ’a -> ’a
positive -> positive -> positive
positive -> positive
positive -> positive
positive -> positive
positive -> nat
positive -> positive
positive -> positive -> comparison ->
comparison
positive -> positive -> comparison
positive -> positive -> positive
positive -> positive -> positive
positive -> positive -> bool
positive -> positive -> bool
positive -> positive -> bool
(positive -> ’a) -> (’a -> positive) -> positive * mask -> positive * mask
positive -> positive * mask
positive -> positive
nat -> positive -> positive -> positive
positive -> positive -> positive
nat -> positive -> positive -> positive *
(positive * positive)
positive -> positive -> positive * (positive * positive)
coq_N -> coq_N
coq_N -> coq_N
positive -> positive -> positive
positive -> positive -> coq_N
5
17
17
5
10
10
5
5
5
5
10
19
32
38
43
22
5
29
22
5
10
5
10
17
3
4
3
3
3
2
19
24
24
17
24
24
67
85
87
39
50
66
68
4
4
17
25
48
ldiff
coq_lxor
shiftl_nat
shiftr_nat
shiftl
shiftr
testbit_nat
testbit
iter_op
to_nat
of_nat
of_succ_nat
digits2_pos
coq_Zdigits2
File: BinNat.ml
succ_double_n
double_n
succ_n
pred_n
succ_pos_n
add_n
sub_n
mul_n
compare_n
eqb_n
leb_n
ltb_n
min_n
max_n
div2_n
even_n
odd_n
pow_n
square_n
log2_n
size_n
size_nat_n
pos_div_eucl
div_eucl
div
modulo
gcd_n
ggcd_n
coq_lor_n
coq_land_n
ldiff_n
coq_lxor_n
shiftl_nat
shiftr_nat
shiftl_n
shiftr_n
testbit_nat_n
testbit_n
to_nat_n
of_nat_n
iter_n
discr
binary_rect
Towards Automatic Resource Bound Analysis for OCaml
positive -> positive -> coq_N
positive -> positive -> coq_N
positive -> nat -> positive
positive -> nat -> positive
positive -> coq_N -> positive
positive -> coq_N -> positive
positive -> nat -> bool
positive -> coq_N -> bool
((’a -> ’a) -> ’a) -> positive -> ’a -> ’a
positive -> nat
nat -> positive
nat -> positive
positive -> positive
coq_Z -> coq_Z
25
25
6
7
9
9
14
14
5
14
7
4
10
15
coq_N -> coq_N
coq_N -> coq_N
coq_N -> coq_N
coq_N -> coq_N
coq_N -> positive
coq_N -> coq_N -> coq_N
coq_N -> coq_N -> coq_N
coq_N -> coq_N -> coq_N
coq_N -> coq_N -> comparison
coq_N -> coq_N -> bool
coq_N -> coq_N -> bool
coq_N -> coq_N -> bool
coq_N -> coq_N -> coq_N
coq_N -> coq_N -> coq_N
coq_N -> coq_N
coq_N -> bool
coq_N -> bool
coq_N -> coq_N -> coq_N
coq_N -> coq_N
coq_N -> coq_N
coq_N -> coq_N
coq_N -> nat
positive -> coq_N -> coq_N * coq_N
coq_N -> coq_N -> coq_N * coq_N
coq_N -> coq_N -> coq_N
coq_N -> coq_N -> coq_N
coq_N -> coq_N -> coq_N
coq_N -> coq_N -> coq_N -> coq_N *
coq_N
coq_N -> coq_N -> coq_N
coq_N -> coq_N -> coq_N
coq_N -> coq_N -> coq_N
coq_N -> coq_N -> coq_N
coq_N -> nat -> coq_N
coq_N -> nat -> coq_N
coq_N -> coq_N -> coq_N
coq_N -> coq_N -> coq_N
coq_N -> nat -> bool
coq_N -> coq_N -> bool
coq_N -> nat
nat -> coq_N
coq_N -> (’a -> ’a) -> ’a -> ’a
coq_N -> positive option
a -> (coq_N -> ’a -> ’a) -> (coq_N -> ’a
-> ’a) -> ’a
4
4
4
4
4
7
11
29
10
10
5
5
5
5
8
8
10
7
26
18
14
9
37
44
46
46
57
75
24
32
32
32
10
14
13
13
18
18
18
8
9
4
12
A Experimental Results
binary_rec
leb_spec0
ltb_spec0
log2_up
lcm
eqb_spec
b2n
setbit
clearbit
ones
lnot
max_case_strong
max_case
max_dec
min_case_strong
min_case
min_dec
max_case_strong_pd
max_case
max_dec_pd
min_case_strong_pd
min_case
min_dec_pd
a -> (coq_N -> ’a -> ’a) -> (coq_N -> ’a
-> ’a) -> ’a
coq_N -> coq_N -> reflect
coq_N -> coq_N -> reflect
coq_N -> coq_N
coq_N -> coq_N -> coq_N
coq_N -> coq_N -> reflect
bool -> coq_N
coq_N -> coq_N -> coq_N
coq_N -> coq_N -> coq_N
coq_N -> coq_N
coq_N -> coq_N -> coq_N
coq_N -> coq_N -> ((coq_N -> coq_N ->
unit -> ’a) -> ’b) -> (unit -> ’a) -> (unit
-> ’a) -> ’b
coq_N -> coq_N -> ((coq_N -> coq_N ->
unit -> ’a) -> ’b) -> ’a -> ’a -> ’b
coq_N -> coq_N -> bool
coq_N -> coq_N -> ((coq_N -> coq_N ->
unit -> ’a) -> ’b) -> (unit -> ’a) -> (unit
-> ’a) -> ’b
coq_N -> coq_N -> ((coq_N -> coq_N ->
unit -> ’a) -> ’b) -> ’a -> ’a -> ’b
coq_N -> coq_N -> bool
coq_N -> coq_N -> (unit -> ’a) -> (unit
-> ’a) -> ’a
coq_N -> coq_N -> ((coq_N -> coq_N ->
unit -> ’a) -> ’b) -> ’a -> ’a -> ’b
coq_N -> coq_N -> bool
coq_N -> coq_N -> (unit -> ’a) -> (unit
-> ’a) -> ’a
coq_N -> coq_N -> ((coq_N -> coq_N ->
unit -> ’a) -> ’b) -> ’a -> ’a -> ’b
coq_N -> coq_N -> bool
49
14
9
9
41
134
16
2
39
47
19
53
23
25
27
23
25
27
25
27
29
25
27
29
50
Towards Automatic Resource Bound Analysis for OCaml
A.2 Evaluation-Step Bounds
Name
Step Bound
File: WorkingWithLists.raml (99 Problems in OCaml)
last
3 + 9M
lastTwo
3 + 11M
at
3 + 18M
natAt
3 + 6K + 4N
length
9 + 12M
rev
9 + 11M
eqList
5 + L + 17M
isPalindrome
20 + 29M
flatten
−
compress
3 + 18M
pack
21 + 34M
encode
21 + 37M
decode
−
duplicate
3 + 13M
replicate
−
drop
9 + 19M
split
23 + 31M
slice
35 + 19M
concat
3 + 9M
rotate
74 + 43M
removeAt
3 + 19M
insertAt
3 + 22M
constructList
−
random
5
min
9
randSelect
61 + 56.5M + 16.5M 2
lottoSelect
−
snd
5
fst
5
map
3 + 13M
insert
6 + 21M
sort
3 + 7.5M + 10.5M 2
compare
15
lengthSort
23 + 12LM + 36M + 27M 2
File: LogicAndCodes.raml
eval2
17 + 38K + 38L + 14M
table2
117 + 152K + 152L + 56M
assoc
2 + 16M
eval
10 + 11L + 16M + 16M X + 16M Y + 20X + 20Y
tableMake
−
File: echelon_form.raml
size
3 + 11n
getElem
2 + 18n
get2Elems
2 + 19L + 2M
subtract_row_helper
3 + 20L + 2M
subtract_row
12 + 39L + 4M
subtract_helper
3 + 19LM + 28M + 24M Y
concat
3 + 11n_1
tail
8 + 18n
hd_helper
8 + 22M
reverse_helper
3 + 11M
reverse
7 + 11n
head
21 + 33n
split_helper
18 + 11L + 34M
split
24 + 34n
subtract
45 + 43LM + 63M
Analysis Time
#Constraints
0.01
0.01
0.01
0.02
0.01
0.02
0.02
0.02
−
0.03
0.05
0.07
−
0.06
−
0.08
0.12
0.13
0.12
0.19
0.18
0.21
−
0.21
0.21
0.46
−
0.46
0.45
0.46
0.52
0.51
0.45
0.67
41
58
49
48
37
42
62
115
f ai l
64
154
151
f ai l
33
f ai l
61
158
119
33
290
47
105
f ai l
8
15
607
f ai l
6
6
73
116
203
26
1000
0.29
1.62
0.06
0.33
−
1044
5899
66
1864
f ai l
0.01
0.01
0.02
0.02
0.03
0.16
0.03
0.03
0.03
0.03
0.03
0.05
0.06
0.06
0.34
29
41
59
58
147
850
35
44
53
35
41
103
102
110
1632
A Experimental Results
echelon_helper
echelon_form
File: matrix.raml
check_lists
check_mat
check_matrix
construct_matrix
getElemMatrix
op
rec_list
rec_mat
check_sanity
plus
minus
append
append_col
transpose_helper
transpose
prod
prod_mat
mult_slow
lineMult
computeLine
mat_mult_jan
check_mult_sanity
mult
delete
submat
File: power_radio.raml
sendmsg msg
main1_events
main2_events
main3_events
main4_events
main5_events
File: avanzini.raml
partition
quicksort
rev_sort
File: append_all.raml
append_all
append_all2
append_all3
File: bfs.raml
dfs
bfs
File: rev_pairs.raml
pairs
File: binary_counter.raml
add_one
add_many
add_list
File: array_fun.raml
nat_iterate
nat_fold
apply_all
File: calculator.raml
add
51
3 + 43LM Y + 63M Y + 59Y
8 + 43m 2 n + 59m + 63m 2
1.75
1.81
8563
8838
3 + 11LM + 19M
18 + 11LM + 30M
25 + 11LM + 30M
28 + 11LM + 30M
9 + 18mn + 18m
13
3 + 29L + 2M
3 + 31LM + 2M + 19Y
79 + 11LM + 30M + 11RY + 30Y
130 + 53m 1 n 1 + 32m 1 + 11m 2 n 2 + 79m 2
130 + 53m 1 n 1 + 32m 1 + 11m 2 n 2 + 79m 2
6 + 11n
3 + 11mn + 8m + 14l
3 + 16.5LM + 5.5LM 2 + 14M + 11M RY
7 + 16.5mn + 5.5m 2 n + 14m
3 + 16L + 2M
3 + 17L + 2LM + 16LY
14+28m 1 +47.5m 1 m 2 n 2 +5.5m 1 m 22 n 2 +14m 1 m 2 +
0.03
0.05
0.05
0.08
0.05
0.03
0.05
0.15
0.15
0.41
0.52
0.26
0.34
0.55
0.55
0.31
0.37
6.81
161
262
271
349
182
27
78
526
681
1916
1916
34
396
1814
1820
51
380
42274
0.37
0.48
1.11
0.6
1.97
78
476
3901
665
7534
1.71
1.85
48
412
12 + 10n
10 + 9K 0 + 23.00L 0 + 10L 0 Y + 9R 0 + 9Z 0
11 + 10K 0 + 16L 0 + 19L 0 Y + 10R 0 + 26Z 0
11 + 13K 0 + 25L 0 + 23L 0 Y + 5L 0 M Y + 29R 0 + 29Z 0
11 + 13K 0 + 25L 0 + 28L 0 Y + 17R 0 + 29Z 0
18 + 13K 0 + 44.6L 0 + 19L 0 Y + 13R 0 + 13Z 0
0.01
0.09
0.2
1.31
0.42
0.44
38
338
1230
7351
2397
2339
5 + 17n
3 + 20n + 13n 2
10 + 23K M + 32L 0 + 20L 0 Y + 13L 0 Y 2
0.01
0.1
0.71
57
587
17704
7 + 9LM + 15M
7 + 15LM + 18LM Y + 21M
7 + 21LM + 27LM RY + 15LM Y + 21M
0.09
0.16
0.39
1011
4869
13435
18 + 26M
24 + 92M
0.04
0.07
488
1723
3 + 7M + 10M 2
0.03
317
8 + 12L 0
3 + 12L 0 + 30X
5 + 12L 0 + 34Y
0.03
0.03
0.03
155
313
341
3 + 10N
3 + 12N
15 + 21N
0.03
0.02
0.03
55
76
135
3 + 8M
0.05
117
47.5m 2 n 2 + 5.5m 22 n 2 + 14m 2
3 + 6L + 16M
3 + 22LY + 18M
3 + 2LM + 18M + 22M RY + 16M Y
71 + 11LM + 30M + 11RY + 30Y
121 + 13m 1 n 1 + 78m 1 + 27m 1 m 2 n 2 + 16m 1 m 2 +
11m 2 n 2 + 30m 2
2 + 21n
9 + 21mn + 19m
52
sub
mult
eval_simpl
eval
File: mergesort.raml
split
merge
mergesort
mergesort_list
File: quicksort.raml
partition
quicksort
quicksort_pairs
quicksort_list
File: square_mult.raml
square_mult
File: subsequence.raml
lcs
File: running.raml
abmap
asort
asort’
btick
abfoldr
cons_all
File: ocaml_sort.raml
merge
list
File: ocaml_list.raml
length
cons
hd
tl
nth
append
rev_append
rev
flatten
concat
map
mapi
rev_map
iter
iteri
fold_left
fold_right
map2
rev_map2
iter2
fold_left2
fold_right2
for_all
exists
for_all2
exists2
mem
memq
assoc
assq
mem_assoc
Towards Automatic Resource Bound Analysis for OCaml
5 + 2K + 8M
3 + 8M + 4M 2
8 + 8K LN + 5K N + 20L + 22X
3 + 20K LN + 14K N + 14L + 14X
0.03
0.06
0.21
0.22
156
427
9548
10723
5 + 8.5M
3 + 17L + 17M
4 − 6K + 32.5K N + 6K 2
4 − 7.5K + 33.5K N − 7K X + 7.5K 2 + 7K 2 X
0.04
0.04
0.17
0.39
159
407
4654
15055
5 + 19M
3 + 19M + 14M 2
3 + 13M + 23M 2
3 − 7.5LM + 7.5LM 2 + 19.5M + 16.5M 2
0.04
0.08
0.09
0.27
197
1747
2131
8712
9 + 21K 0 + 30L 0
0.08
243
21 + 8L + 50LM + 27M
0.04
900
3 + 12M + 12N
11 + 22K N + 13K 2 N + 13L + 15N
14 + 22K N + 13K 2 N + 7L + 15N
3 + 16L + 14N
3 + 13M + 13N
13 + 17M + 21M N + 26N
0.03
0.14
0.16
0.05
0.05
0.08
477
5656
5655
904
267
1844
3 + 17L + 17M
43 + 30.5M + 8.5M 2
0.05
0.11
404
3066
7 + 10M
4
3
3
14 + 16M
3 + 9M
3 + 9M
7 + 9M
3 + 9LM + 11M
3 + 9LM + 11M
3 + 13M
8 + 19M
9 + 11M
3 + 10M
8 + 16M
3 + 10M
3 + 10M
5 + 17L
12 + 15L
5 + 14M
5 + 14L
5 + 14M
3 + 12M
3 + 12M
5 + 16M
5 + 16L
3 + 18M
3 + 22M
2 + 18M
2 + 22M
3 + 20M
0.02
0.01
0.02
0.02
0.02
0.03
0.02
0.02
0.03
0.04
0.02
0.02
0.02
0.02
0.02
0.02
0.03
0.03
0.03
0.03
0.03
0.03
0.03
0.03
0.03
0.04
0.04
0.04
0.03
0.03
0.04
59
9
14
16
93
77
78
84
201
201
73
92
103
55
74
59
55
142
196
120
128
120
61
61
126
205
67
75
74
84
69
A Experimental Results
mem_assq
3 + 24M
remove_assoc
3 + 21M
remove_assq
3 + 25M
find
2 + 10M
find_all
18 + 22M
filter
18 + 22M
partition
30 + 23M
split
5 + 17M
combine
5 + 13M
merge
3 + 21L + 21M
chop
8 + 16M
stable_sort
f ai l
sort
f ai l
fast_sort
f ai l
File: aws.raml
average_grade
18 + 34M
greater_eq
49 + 68M
sort_students
12 − 34LM + 34LM 2 + 15M
make_table
9 + 25LM + 21M
find
2 + 12M
lookup
8 + 17LM + 17M
average_grade’
21 + 17LM + 17LM Y + 33M
greater_eq’
61 + 34LM + 34LM Y + 66M
sort_students_efficient
f ai l
File: PROTOTYPE
File: appendAll.raml
appendAll
3 + 9n 1 n 2 + 13n 1
appendAll2
3 + 13n 1 n 2 + 18n 1 n 2 n 3 + 18n 1
appendAll3
3 + 18n 1 n 2 + 27n 1 n 2 n 3 n 4 + 13n 1 n 2 n 3 + 18n 1
File: duplicates.raml
eq
5 + 17n 1 + n 2
remove
3 + 21L + 18LY
nub
3 − 9N n + 9N 2 n + 5.5N + 10.5N 2
File: dyade.raml
multList
3 + 15n
dyade
3 + 15n 1 n 2 + 17n 1
File: eratosthenes.raml
filter
3 + 23n
eratos
3 + 4.5n + 11.5n 2
File: bitvectors.raml
bitToInt’
3 + 20M
bitToInt
7 + 20n
sum
33
add’
3 + 51L + 2M
add
8 + 2n 1 + 51n 2
diff
31
sub’
5 + 59L + 2M
sub
14 + 61n 1
mult
3 + 53n 1 n 2 + 30n 1
compare
3 + 28L + 2M
leq
13 + 2n 1 + 28n 2
File: flatten.raml
flatten
3 + 5.5nl + 5.5n 2 l + 24n
insert
6 + 17n
insertionsort
3 + 7.5n + 8.5n 2
flattensort
11 + 13nl + 5.5n 2 l + 8.5n 2 l 2 + 24n
File: listsort.raml
isortlist
3 − 9nl + 9n 2 l + 6.5n + 9.5n 2
File: longestCommonSubsequence.raml
firstline
3 + 10M
53
0.04
0.04
0.04
0.04
0.04
0.05
0.06
0.06
0.06
0.08
0.06
0.32
0.45
0.68
77
92
102
66
312
312
580
83
122
253
83
6606
6606
6606
0.04
0.05
0.32
0.08
0.05
0.08
0.09
0.18
0.64
185
721
8933
949
88
269
1094
2572
19377
0.05
0.2
1.57
203
1565
12785
0.01
0.1
0.51
62
625
3919
0.01
0.05
36
246
0.01
0.04
51
180
0.02
0.02
0.01
0.03
0.03
0.03
0.06
0.06
0.12
0.07
0.07
42
48
59
112
119
47
127
138
477
80
94
0.42
0.02
0.05
2.34
3388
55
181
19153
0.41
2676
0.01
29
54
right
max
newline
lcstable
lcs
File: mergesort.raml
msplit
merge
mergesortBuggy
mergesort
File: minsort.raml
findMin
minSort
File: queue.raml
empty
enqueue
enqueues
copyover
dequeue
children
breadth
Towards Automatic Resource Bound Analysis for OCaml
5
7
3 + 2L + 50M
11 + 10L + 52LM + 25M
23 + 10L + 52LM + 25M
0.01
0.01
0.03
0.15
0.16
14
13
128
865
901
5 + 9.5M
3 + 22L + 22M
−
3 − 38.67M + 47.33M 2
0.03
0.06
−
0.14
112
239
f ai l
869
3 + 21M
8 + 25.5M + 10.5M 2
0.03
0.04
125
219
0.01
0.03
0.09
0.08
0.18
0.26
3.99
4
19
203
226
678
333
10540
4.34
4.33
4.29
10818
18711
18747
0.15
896
0.15
896
0.02
0.02
68
52
0.04
0.04
0.12
0.12
−
0.05
0.09
0.04
0.08
124
130
574
574
f ai l
108
331
124
291
0.13
0.18
0.06
0.26
0.51
0.69
437
623
184
1607
3165
3793
0.12
680
0.06
0.27
0.23
0.29
0.95
2.52
95
1428
1041
1428
5804
15613
3
10
3 + 21M
7 + 14M
14 + 39M
27
18+118B R +116B RT −LM +116LM Y +L 2 M +88M +
118M Y + 49R − RT + RT 2
startBreadth
19 + 57M + 59M 2
depth
27 + 42L + 53LM
startDepth
3 + 15.5M + 26.5M 2
File: quicksort_mutual.raml
part
24+16L+30LY +15L 2 +27M +30M Y +15M 2 +46Y +
15Y 2
quicksortMutual
3 + 16M + 15M 2
File: rationalPotential.raml
zip3
3 + 17L + 2M
group3
3 + 5.67M
File: sizechange.raml
r1
3 + 11M
rev
7 + 11M
f
3 + 11LM + 17.33M + 3.67M 3
g
15 + 11L + 11LM + 17.33M + 11M Y + 3.67M 3
f2
−
last
3 + 14M
f2’
12 + 14L + 14M
g3
3 + 11M
f3
11 + 22L + 11M
File: splitandsort.raml
insert
13 + 23M
split
3 + 11.5M + 11.5M 2
splitqs
7 + 24M
quicksort
3 + 16.5M + 17.5M 2
sortAll
3 + 16.5LM + 17.5L 2 M + 19M
splitAndSort
11 + 47M + 29M 2
File: subtrees.raml
subtrees
3 + 23.5M + 5.5M 2
File: tuples.raml
attach
3 + 13M
pairs
3 + 7M + 12M 2
pairsAux
3 + 8M + 12M 2
pairsSlow
3 + 16.17M + M 2 + 1.83M 3
triples
3 + 30.5M − 14.5M 2 + 8M 3
quadruples
3 + 0.33M + 35.5M 2 − 14.83M 3 + 3M 4
File: array_dijkstra.raml
A Experimental Results
makeGraph
dijkstra
File: CompCert
File: String0.ml
string_dec
prefix
File: Tuples.ml
uncurry
File: Specif.ml
projT1
projT2
value
File: EquivDec.ml
equiv_dec
File: Datatypes.ml
implb
xorb
negb
fst
snd
length
app
coq_CompOpp
coq_CompareSpec2Type
coq_CompSpec2Type
File: Bool.ml
bool_dec
eqb
iff_reflect
File: Ring.ml
bool_eq
File: Peano.ml
plus
max
min
nat_iter
File: List0.ml
hd
tl
in_dec
nth_error
remove
rev
rev_append
rev’
list_eq_dec
map
fold_left
fold_right
existsb
forallb
filter
File: EqNat.ml
beq_nat
File: Compare_dec.ml
le_lt_dec
le_gt_dec
nat_compare
File: BinPosDef.ml
succ
55
35 + 36K + 21N + 1N 2
46 + 33M + 111M 2
0.03
0.11
519
2808
5 + 14L
3 + 14M
0.06
0.05
205
170
−
0.02
106
3
3
2
0.01
0.01
0.01
6
6
4
1
0.01
2
3
5
7
3
3
3 + 7M
3 + 9M
3
3
6
0.01
0.01
0.01
0.01
0.01
0.04
0.05
0.01
0.01
0.01
7
12
13
4
4
72
118
25
25
30
5
5
7
0.01
0.01
0.01
17
12
15
12
0.01
23
3 + 8M
3 + 8L + 2M
3 + 10M
3 + 9M
0.04
0.06
0.05
0.06
117
180
147
64
3
3
3 + 14M
5 + 8M
3 + 15M
3 + 9.5M + 4.5M 2
3 + 9M
7 + 9M
5 + 14L
3 + 11M
3 + 10M
3 + 10M
3 + 12M
3 + 12M
3 + 13M
0.03
0.03
0.04
0.08
0.05
0.1
0.05
0.05
0.08
0.05
0.05
0.07
0.06
0.06
0.07
18
21
91
326
179
423
122
128
223
96
77
65
71
71
176
5 + 8L
0.08
162
3 + 6L + 2M
7 + 8M
5 + 8M
0.07
0.07
0.08
127
133
183
5 + 7N
0.04
124
56
add
add_carry
pred_double
pred
pred_N
mask_rect
mask_rec
succ_double_mask
double_mask
double_pred_mask
pred_mask
sub_mask
sub_mask_carry
sub
mul
iter
pow
square
div2
div2_up
size_nat
size
compare_cont
compare
min
max
eqb
leb
ltb
sqrtrem_step
sqrtrem
sqrt
gcdn
gcd
ggcdn
ggcd
coq_Nsucc_double
coq_Ndouble
coq_lor
coq_land
ldiff
coq_lxor
shiftl_nat
shiftr_nat
shiftl
shiftr
testbit_nat
testbit
iter_op
to_nat
of_nat
of_succ_nat
digits2_pos
coq_Zdigits2
File: BinNat.ml
succ_double_n
double_n
succ_n
pred_n
succ_pos_n
Towards Automatic Resource Bound Analysis for OCaml
7 + 7L + 5M + 7N + 5X
9 + 7L + 5M + 7N + 5X
7 + 7M
5 + 7M
6 + 7M
5
5
6
6
8 + 7M
13 + 7L
5 + 9L + 7M + 7N + 9X
3 + 8L + 8M + 8N + 8X
11 + 9L + 7M + 7N + 9X
3 + 7L + 17LN + 8M + 25N + 10N X
−
−
3 + 9M + 10M N + 24.5N + 8.5N 2
3
3 + 7N
5 + 7M + 7N
3 + 19M + 19N
5 + 9L + 9X
10 + 9L + 9X
16 + 9L + 9X
16 + 9L + 9X
5 + 8L + 8X
16 + 9L + 9X
16 + 9L + 9X
−
5 + 66M + 16M N + 4M 2 + 59N + 4N 2
13 + 66M + 16M N + 4M 2 + 59N + 4N 2
−
−
−
−
6
6
5 + 8L + 2M + 2N + 8X
6 + 16L + 16X
5 + 13L + 3M + 3N + 13X
5 + 13L + 3M + 3N + 13X
8 + 12X
8 + 12X
−
−
5 + 8X
5 + L + 23M + 23N + 7Z
3 + 10M + 13N
−
3 + 21M
3 + 19M
3 + 19M + 19N
2 + 19K + 19L + 7 + 1 + 19R + 7 + 19Z
0.64
0.67
0.05
0.06
0.06
0.04
0.04
0.04
0.04
0.06
0.09
0.74
0.73
0.72
1.03
0.59
32.47
0.98
0.17
0.19
0.2
0.23
0.38
0.33
0.45
0.48
0.27
0.41
0.43
5438
5438
124
174
187
41
41
53
48
197
446
6368
6368
6405
7722
2849
139223
7592
44
169
156
542
1248
1258
2071
2071
454
1283
1283
17.54
18.25
19.67
29.64
29.04
29.79
4.81
4.86
5.12
4.97
5.08
5.09
4.85
5.08
6.42
6.58
5.15
5.3
5.2
5.86
5.44
5.27
0.47
0.54
135517
135529
97408
100787
179760
183139
42
37
1032
1355
1435
1509
136
186
5453
7323
291
1219
187
2652
244
221
542
1146
6
6
11 + 7K
11 + 7L
10 + 7K
1.72
1.75
1.79
1.77
1.75
42
37
162
214
150
A Experimental Results
add_n
sub_n
mul_n
compare_n
eqb_n
leb_n
ltb_n
min_n
max_n
div2_n
even_n
odd_n
pow_n
square_n
log2_n
size_n
size_nat_n
pos_div_eucl
div_eucl
div
modulo
gcd_n
ggcd_n
coq_lor_n
coq_land_n
ldiff_n
coq_lxor_n
shiftl_nat
shiftr_nat
shiftl_n
shiftr_n
testbit_nat_n
testbit_n
to_nat_n
of_nat_n
iter_n
discr
binary_rect
binary_rec
leb_spec0
ltb_spec0
log2_up
lcm
eqb_spec
b2n
setbit
clearbit
ones
lnot
max_case_strong
max_case
max_dec
min_case_strong
min_case
min_dec
max_case_strong_pd
max_case
max_dec_pd
min_case_strong_pd
18 + 7K + 5L + 7R + 5S
17 + 8K + 8L + 8R + 8S
2 + 17K R + 10K S + 25K + 8L + 2 + 7R + 8
18 + 9K + 9L
13 + 8R + 8S
27 + 9K + 9L
27 + 9K + 9L
27 + 9K + 9L
27 + 9K + 9L
6
5
17
−
−
5 + 19K + 19L
9 + 19K + 19L
10 + 7K + 7L
−
2 + 50K L − 12.5K + 99K + 12.5K 2 − 12.5L + 99L +
12.5L 2 + 2 + 15
−
−
−
−
14 + 8K + 8L + 2R + 2S
14 + 16R + 16S
13 + 3K + 3L + 13R + 13S
13 + 3K + 3L + 13R + 13S
12 + 15Z
12 + 15Z
−
−
10 + 9Z
11 + 23K + 23L + 7S
−
3 + 19M
−
4
17 + 11K + 11L
17 + 11K + 11L
33 + 9K + 9L
33 + 9K + 9L
51 + 19K + 26L
−
22 + 8R + 8S
8
−
−
−
−
70 + 18K + 18L
78 + 18K + 18L
86 + 18K + 18L
70 + 18K + 18L
78 + 18K + 18L
86 + 9K + 9L + 9R + 9S
78 + 9K + 9L + 9R + 9S
14 + 18K + 18L + 2 + 62
86 + 18K + 18L
78 + 9K + 9L + 9R + 9S
57
2.48
2.65
14.04
2.36
2.4
2.47
2.47
2.71
2.71
2.39
2.29
2.45
7.98
3.4
2.55
2.49
2.52
38.2
38.13
5562
6514
89507
1373
580
1398
1398
2352
2352
79
53
74
25024
7622
1165
575
182
183270
183578
34.64
40.36
29.84
41.02
7.51
7.53
7.75
7.63
7.59
7.58
16.82
16.99
7.79
7.88
16.25
7.91
16.49
7.77
7.99
8.22
8.35
8.45
8.87
169.2
11.12
11.02
20.78
22.03
23.56
22.39
15.23
15.24
16.73
17.72
18.31
20.11
21.86
32.11
23.53
25.48
183586
183588
101067
183702
1156
1442
1560
1627
164
225
5551
8953
387
1261
2694
246
2893
32
593
593
1416
1416
3417
403044
598
20
10128
10672
5992
11681
12294
12264
12270
12294
12264
12270
12300
52970
12270
12300
58
min_case
min_dec_pd
Towards Automatic Resource Bound Analysis for OCaml
14 + 18K + 18L + 2 + 62
86 + 18K + 18L
35.89
27.49
52970
12270
A Experimental Results
59
A.3 Heap-Allocation Bounds
Name
Heap Bound
File: WorkingWithLists.raml (99 Problems in OCaml)
last
2
lastTwo
4
at
2 + 2M
natAt
2
length
4+M
rev
5 + 4M
eqList
1
isPalindrome
6 + 4M
flatten
−
compress
2 + 4M
pack
14 + 14M
encode
13 + 12M
decode
−
duplicate
2 + 8M
replicate
−
drop
7 + 5M
split
16 + 10M
slice
12 + 6M
concat
4M
rotate
23 + 11M
removeAt
2 + 6M
insertAt
3 + 6M
constructList
−
random
1
min
0
randSelect
21 + 9M + 5M 2
lottoSelect
−
snd
0
fst
0
map
2 + 4M
insert
6 + 5M
sort
2 + 3.5M + 2.5M 2
compare
1
lengthSort
12 + LM + 17M + 3M 2
File: LogicAndCodes.raml
eval2
0
table2
46
assoc
0
eval
0
tableMake
−
File: echelon_form.raml
size
1+n
getElem
2n
get2Elems
1 + 2L
subtract_row_helper
2 + 4M
subtract_row
3 + 2L + 4M
subtract_helper
2 + 2LM + 7M + 4M Y
concat
4n_1
tail
1 + 2n
hd_helper
1 + 6M
reverse_helper
4M
reverse
2 + 4n
head
5 + 10n
split_helper
4 + 4L + 9M
split
7 + 9n
subtract
9 + 6LM + 15M
Analysis Time
#Constraints
0.01
0.02
0.01
0.02
0.01
0.01
0.02
0.02
−
0.03
0.05
0.06
−
0.05
−
0.08
0.11
0.13
0.12
0.21
0.18
0.21
−
0.2
0.2
0.48
−
0.44
0.42
0.46
0.46
0.51
0.46
0.66
31
45
32
36
22
29
42
76
f ai l
44
109
98
f ai l
24
f ai l
39
105
72
25
176
31
84
f ai l
4
5
500
f ai l
1
1
62
94
169
12
876
0.29
1.54
0.05
0.31
−
976
5607
49
1804
f ai l
0.01
0.01
0.02
0.02
0.03
0.16
0.03
0.03
0.03
0.03
0.03
0.04
0.06
0.06
0.31
19
24
38
38
89
778
25
27
34
25
28
65
65
69
1469
60
echelon_helper
echelon_form
File: matrix.raml
check_lists
check_mat
check_matrix
construct_matrix
getElemMatrix
op
rec_list
rec_mat
check_sanity
plus
minus
append
append_col
transpose_helper
transpose
prod
prod_mat
mult_slow
lineMult
computeLine
mat_mult_jan
check_mult_sanity
mult
delete
submat
File: power_radio.raml
sendmsg msg
main1_events
main2_events
main3_events
main4_events
main5_events
File: avanzini.raml
partition
quicksort
rev_sort
File: append_all.raml
append_all
append_all2
append_all3
File: bfs.raml
dfs
bfs
File: rev_pairs.raml
pairs
File: binary_counter.raml
add_one
add_many
add_list
File: array_fun.raml
nat_iterate
nat_fold
apply_all
File: calculator.raml
add
sub
mult
Towards Automatic Resource Bound Analysis for OCaml
6LM Y + 15M Y + 10Y
1 + 6m 2 n + 10m + 15m 2
1.74
1.88
8386
8657
1 + LM + M
2 + LM + 2M
2 + LM + 2M
5 + LM + 2M
2mn + 2m
1
2 + 5L
2 + 5LM + 6Y
4 + LM + 2M + RY + 2Y
12 + 7m 1 n 1 + 2m 1 + m 2 n 2 + 10m 2
12 + 7m 1 n 1 + 2m 1 + m 2 n 2 + 10m 2
6 + 4n
4mn + 12l
10LM + 2LM 2 + 4M RY
2 + 10mn + 2m 2 n
1
2 + 5L
4+8m 1 +15m 1 m 2 n 2 +2m 1 m 22 n 2 +15m 2 n 2 +2m 22 n 2
2 + 6M
6LY + 2M
2 + 2LM + 6M + 6M RY
4 + LM + 2M + RY + 2Y
11 + 3m 1 n 1 + 10m 1 + 7m 1 m 2 n 2 + m 2 n 2 + 2m 2
6n
2 + 6mn + 2m
0.03
0.05
0.05
0.07
0.05
0.03
0.04
0.15
0.15
0.4
0.51
0.26
0.36
0.56
0.57
0.31
0.35
6.78
0.39
0.5
1.17
0.64
2.16
1.88
1.81
135
214
216
292
137
10
45
477
542
1652
1652
24
367
1773
1776
33
63
42184
49
431
3844
534
7270
30
365
6
5 + 6M
8 + 4L 0 Y + 8Z 0
11 + 2K 0 + 2L 0 + 8L 0 Y + 8R 0 + 10Z 0
11 + 2K 0 + 2L 0 + 8L 0 Y + 2R 0 + 10Z 0
12 + 2K 0 + 5.2L 0 + 4L 0 Y + 2R 0 + 2Z 0
0.01
0.01
0.2
0.42
0.43
0.4
22
73
1184
2508
2331
2226
6 + 6n
2 + 7n + 5n 2
7 + 12K M + 16L 0 + 7L 0 Y + 5L 0 Y 2
0.01
0.09
0.7
40
542
17616
3 + 4LM
2 + 8LM Y + 3M
2 + 3LM + 12LM RY + 2M
0.07
0.16
0.36
988
4823
13366
12 + 8M
12 + 24M
0.04
0.08
466
1669
2 − 3M + 5M 2
0.03
290
8 + 6L 0
6L 0 + 14X
1 + 6L 0 + 17Y
0.04
0.04
0.03
145
292
316
1
0
2
0.02
0.02
0.02
45
36
102
4M
2
2 − 2M + 2M 2
0.04
0.04
0.04
110
145
411
A Experimental Results
eval_simpl
eval
File: mergesort.raml
split
merge
mergesort
mergesort_list
File: quicksort.raml
partition
quicksort
quicksort_pairs
quicksort_list
File: square_mult.raml
square_mult
File: subsequence.raml
lcs
File: running.raml
abmap
asort
asort’
btick
abfoldr
cons_all
File: ocaml_sort.raml
merge
list
File: ocaml_list.raml
length
cons
hd
tl
nth
append
rev_append
rev
flatten
concat
map
mapi
rev_map
iter
iteri
fold_left
fold_right
map2
rev_map2
iter2
fold_left2
fold_right2
for_all
exists
for_all2
exists2
mem
memq
assoc
assq
mem_assoc
mem_assq
remove_assoc
61
5 + 4K LN + 2X
11K LN + 4.5K N
0.19
0.22
9500
10688
6 + 5M
4L + 4M
2 + 13K N
2 − 0.5K + 17K N + 0.5K 2
0.04
0.05
0.17
0.37
146
383
4589
14973
6+6∗M
2 + 7M + 5M 2
2 + 17M + 5M 2
2 − LM + LM 2 + 17.5M + 5.5M 2
0.04
0.07
0.09
0.25
178
1700
2063
8649
7+K0
0.08
201
9 + 5L + 9LM + 12M
0.04
815
2 + 4M + 4N
8 + 15K N + 5K 2 N + 4L + 6N
8 + 15K N + 5K 2 N + 6N
2 + 7L + 7N
0
5 + 4M + 4M N + 5N
0.04
0.13
0.13
0.05
0.03
0.08
459
5588
5588
884
242
1772
4L + 4M
32 + 17M + 2M 2
0.04
0.11
381
2997
1+M
4
0
0
4 + 2M
4M
4M
2 + 4M
2 + 4LM
2 + 4LM
2 + 4M
3 + 5M
6 + 4M
1
2 + 1M
0
0
2 + 4L
6 + 4M
1
0
0
1
1
1
1
1 + 2M
1 + 2M
2M
2M
1 + 2M
1 + 2M
2 + 6M
0.02
0.02
0.01
0.01
0.02
0.02
0.02
0.02
0.02
0.03
0.02
0.02
0.02
0.02
0.02
0.03
0.02
0.03
0.03
0.02
0.03
0.03
0.03
0.03
0.03
0.03
0.04
0.04
0.04
0.04
0.04
0.04
0.04
46
7
11
13
69
69
70
73
185
185
62
72
90
45
55
48
44
125
176
104
111
103
49
49
108
169
51
55
57
63
51
55
74
62
Towards Automatic Resource Bound Analysis for OCaml
remove_assq
2 + 6M
find
0
find_all
9 + 8M
filter
9 + 8M
partition
16 + 8M
split
6 + 10M
combine
2 + 6M
merge
5L + 5M
chop
1 + 2M
stable_sort
f ai l
sort
f ai l
fast_sort
f ai l
File: aws.raml
average_grade
7 + 6M
greater_eq
16 + 12M
sort_students
4 − 6LM + 6LM 2 + 11.5M + 13.5M 2
make_table
7 + 9LM + 11M
find
0
lookup
4
average_grade’
10 + 8M
greater_eq’
22 + 16M
sort_students_efficient
f ai l
File: PROTOTYPE
File: appendAll.raml
appendAll
2 + 4n 1 n 2
appendAll2
2 + 8n 1 n 2 n 3 + 2n 1
appendAll3
2 + 2n 1 n 2 + 12n 1 n 2 n 3 n 4 + 2n 1
File: duplicates.raml
eq
1
remove
2 + 5L
nub
2 + 3.5N + 2.5N 2
File: dyade.raml
multList
2 + 4n
dyade
2 + 4n 1 n 2 + 6n 1
File: eratosthenes.raml
filter
2 + 5n
eratos
2 + 3.5n + 2.5n 2
File: bitvectors.raml
bitToInt’
1
bitToInt
2
sum
7
add’
2 + 11M
add
3 + 11n 1
diff
5
sub’
4 + 13M
sub
5 + 13n 1
mult
2 + 11n 1 n 2 + 9n 1
compare
1 + 3M
leq
2 + 3n 1
File: flatten.raml
flatten
2 + 2nl + 2n 2 l + 2n
insert
6 + 4n
insertionsort
2 + 4n + 2n 2
flattensort
4 + 6nl + 2n 2 l + 2n 2 l 2 + 2n
File: listsort.raml
isortlist
2 + 3.5n + 2.5n 2
File: longestCommonSubsequence.raml
firstline
2 + 5M
right
1
max
1
0.04
0.05
0.05
0.05
0.07
0.06
0.06
0.07
0.06
0.33
0.44
0.66
80
55
280
280
531
71
110
227
68
6204
6204
6204
0.04
0.06
0.36
0.08
0.06
0.07
0.11
0.22
0.64
143
625
8765
911
75
226
1009
2384
19074
0.04
0.21
1.57
185
1527
12727
0.01
0.02
0.12
42
90
938
0.01
0.06
23
221
0.01
0.04
30
148
0.01
0.01
0.01
0.03
0.03
0.03
0.06
0.06
0.12
0.06
0.07
22
25
32
67
70
20
70
73
409
48
53
0.42
0.02
0.05
2.43
3353
36
152
19084
0.11
704
0.02
0.01
0.01
51
12
5
A Experimental Results
newline
2 + 7M
lcstable
8 + 5L + 7LM + 10M
lcs
9 + 5L + 7LM + 10M
File: mergesort.raml
msplit
6 + 5M
merge
8L + 8M
mergesortBuggy
−
mergesort
2 − 18.67M + 23.33M 2
File: minsort.raml
findMin
2 + 8M
minSort
4 + 10M + 4M 2
File: queue.raml
empty
6
enqueue
6
enqueues
6M
copyover
4 + 6M
dequeue
10 + 18M
children
24
breadth
12 + 54B R + 54B RT + 54LM Y + 38M + 54M Y + 20R
startBreadth
15 + 25M + 27M 2
depth
16 + 24L + 28LM
startDepth
2 + 10M + 14M 2
File: quicksort_mutual.raml
part
8 + 6L + 8LY + 4L 2 + 10M + 8M Y + 4M 2 + 14Y + 4Y 2
quicksortMutual
2 + 6M + 4M 2
File: rationalPotential.raml
zip3
2 + 7L
group3
2 + 2.33M
File: sizechange.raml
r1
4M
rev
2 + 4M
f
4LM − 1.33M + 1.33M 3
g
4L + 4LM + 1.33M + 4M Y + 1.33M 3
f2
−
last
2 + 6M
f2’
2 + 6L + 6M
g3
4M
f3
2 + 8L + 4M
File: splitandsort.raml
insert
14 + 6M
split
2 + 11M + 3M 2
splitqs
6 + 8M
quicksort
2 + 8M + 6M 2
sortAll
2 + 8LM + 6L 2 M + 8M
splitAndSort
4 + 27M + 9M 2
File: subtrees.raml
subtrees
2 + 9M + 2M 2
File: tuples.raml
attach
2 + 6M
pairs
2 − 3M + 5M 2
pairsAux
−3M + 5M 2
pairsSlow
2 + 0.33M + M 2 + 0.67M 3
triples
2 + 9.67M − 9M 2 + 3.33M 3
quadruples
2 − 4.83M + 14.75M 2 − 7.17M 3 + 1.25M 4
File: array_dijkstra.raml
makeGraph
4 + 2N + N 2
dijkstra
6 + 8.5M + 9.5M 2
File: CompCert
File: String0.ml
63
0.08
0.14
0.15
276
782
806
0.03
0.05
−
0.14
97
212
f ai l
805
0.03
0.04
104
188
0.01
0.03
0.1
0.08
0.15
0.23
3.76
3.99
4.13
4.39
4
12
184
214
647
309
10447
10704
18650
18676
0.15
0.15
843
843
0.02
0.02
52
38
0.05
0.04
0.12
0.12
−
0.04
0.08
0.05
0.08
114
117
545
545
f ai l
96
303
114
267
0.13
0.18
0.05
0.26
0.49
0.66
411
587
161
1556
3101
3688
0.12
651
0.07
0.28
0.25
0.27
1.04
2.95
85
1395
1006
1395
5736
15510
0.03
0.1
437
2607
64
string_dec
prefix
File: Tuples.ml
uncurry
File: Specif.ml
projT1
projT2
value
File: EquivDec.ml
equiv_dec
File: Datatypes.ml
implb
xorb
negb
fst
snd
length
app
coq_CompOpp
coq_CompareSpec2Type
coq_CompSpec2Type
File: Bool.ml
bool_dec
eqb
iff_reflect
File: Ring.ml
bool_eq
File: Peano.ml
plus
max
min
nat_iter
File: List0.ml
hd
tl
in_dec
nth_error
remove
rev
rev_append
rev’
list_eq_dec
map
fold_left
fold_right
existsb
forallb
filter
File: EqNat.ml
beq_nat
File: Compare_dec.ml
le_lt_dec
le_gt_dec
nat_compare
File: BinPosDef.ml
succ
add
add_carry
pred_double
pred
Towards Automatic Resource Bound Analysis for OCaml
A Experimental Results
pred_N
mask_rect
mask_rec
succ_double_mask
double_mask
double_pred_mask
pred_mask
sub_mask
sub_mask_carry
sub
mul
iter
pow
square
div2
div2_up
size_nat
size
compare_cont
compare
min
max
eqb
leb
ltb
sqrtrem_step
sqrtrem
sqrt
gcdn
gcd
ggcdn
ggcd
coq_Nsucc_double
coq_Ndouble
coq_lor
coq_land
ldiff
coq_lxor
shiftl_nat
shiftr_nat
shiftl
shiftr
testbit_nat
testbit
iter_op
to_nat
of_nat
of_succ_nat
digits2_pos
coq_Zdigits2
File: BinNat.ml
succ_double_n
double_n
succ_n
pred_n
succ_pos_n
add_n
sub_n
mul_n
compare_n
65
66
eqb_n
leb_n
ltb_n
min_n
max_n
div2_n
even_n
odd_n
pow_n
square_n
log2_n
size_n
size_nat_n
pos_div_eucl
div_eucl
div
modulo
gcd_n
ggcd_n
coq_lor_n
coq_land_n
ldiff_n
coq_lxor_n
shiftl_nat
shiftr_nat
shiftl_n
shiftr_n
testbit_nat_n
testbit_n
to_nat_n
of_nat_n
iter_n
discr
binary_rect
binary_rec
leb_spec0
ltb_spec0
log2_up
lcm
eqb_spec
b2n
setbit
clearbit
ones
lnot
max_case_strong
max_case
max_dec
min_case_strong
min_case
min_dec
max_case_strong_pd
max_case
max_dec_pd
min_case_strong_pd
min_case
min_dec_pd
Towards Automatic Resource Bound Analysis for OCaml
A Experimental Results
67
A.4 Tick Bounds
Name
Tick Bound
File: WorkingWithLists.raml (99 Problems in OCaml)
last
M
lastTwo
M
at
M
natAt
K
length
M
rev
M
eqList
L
isPalindrome
2M
flatten
−
compress
M
pack
2M
encode
2M
decode
−
duplicate
M
replicate
−
drop
M
split
2M
slice
1+M
concat
M
rotate
3M
removeAt
M
insertAt
M
constructList
−
random
1
min
1
randSelect
1 + 2M + M 2
lottoSelect
−
snd
1
fst
1
map
M
insert
M
sort
0.5M + 0.5M 2
compare
1
lengthSort
LM + 2M + 2M 2
File: LogicAndCodes.raml
eval2
1 + 2K + 2L + M
table2
4 + 8K + 8L + 4M
assoc
M
eval
1 + L + M + M X + M Y + 2X + 2Y
tableMake
−
File: echelon_form.raml
size
n
getElem
n
get2Elems
L
subtract_row_helper
L
subtract_row
L+M
subtract_helper
M + 2M Y
concat
n_1
tail
n
hd_helper
M
reverse_helper
M
reverse
n
head
2n
split_helper
L + 2M
split
2n
subtract
2LM + M
Analysis Time
#Constraints
0.01
0.01
0.01
0.01
0.01
0.01
0.02
0.02
−
0.03
0.05
0.07
−
0.06
−
0.08
0.12
0.13
0.12
0.2
0.18
0.25
−
0.24
0.25
0.52
−
0.48
0.45
0.43
0.45
0.48
0.48
0.69
30
42
29
35
20
26
40
71
f ai l
39
92
82
f ai l
20
f ai l
33
91
61
24
156
27
76
−
3
6
484
−
2
2
60
85
160
10
859
0.29
1.53
0.05
0.32
−
980
5594
50
1809
f ai l
0.01
0.01
0.01
0.02
0.03
0.16
0.03
0.03
0.03
0.03
0.03
0.04
0.06
0.06
0.3
18
23
36
35
84
771
24
26
31
24
26
59
60
62
1453
68
echelon_helper
echelon_form
File: matrix.raml
check_lists
check_mat
check_matrix
construct_matrix
getElemMatrix
op
rec_list
rec_mat
check_sanity
plus
minus
append
append_col
transpose_helper
transpose
prod
prod_mat
mult_slow
lineMult
computeLine
mat_mult_jan
check_mult_sanity
mult
delete
submat
File: power_radio.raml
sendmsg msg
main1_events
main2_events
main3_events
main4_events
main5_events
File: avanzini.raml
partition
quicksort
rev_sort
File: append_all.raml
append_all
append_all2
append_all3
File: bfs.raml
dfs
bfs
File: rev_pairs.raml
pairs
File: binary_counter.raml
add_one
add_many
add_list
File: array_fun.raml
nat_iterate
nat_fold
apply_all
File: calculator.raml
add
sub
Towards Automatic Resource Bound Analysis for OCaml
2LM Y + 4M Y
2m 2 n + 4m 2
1.82
1.81
8369
8639
LM + M
LM + 2M
LM + 2M
LM + 2M
mn + m
1
2M
2LM + Y
LM + 2M + RY + 2Y
4m 1 n 1 + 2m 1 + m 2 n 2 + 5m 2
4m 1 n 1 + 2m 1 + m 2 n 2 + 5m 2
n
mn + l
0.5LM + 0.5LM 2 + M + M RY
0.5mn + 0.5m 2 n + m
M
L + LY
m 1 +2.5m 1 m 2 n 2 +0.5m 1 m 22 n 2 +m 1 m 2 +2.5m 2 n 2 +
0.04
0.05
0.06
0.07
0.05
0.03
0.05
0.14
0.15
0.42
0.51
0.25
0.33
0.63
0.58
0.32
0.37
6.57
133
210
212
287
135
10
42
471
534
1632
1632
20
357
1764
1766
32
347
42169
0.5m 22 n 2 + m 2
M
LY + M
M + M RY + M Y
LM + 2M + RY + 2Y
m 1 n 1 + 5m 1 + m 1 m 2 n 2 + m 1 m 2 + m 2 n 2 + 2m 2
n
mn + m
0.41
0.48
1.09
0.69
1.93
1.76
1.86
44
426
3836
526
7249
27
357
200 + 32n
200L 0 + 32L 0 Y
32L 0 Y + 200Z 0
16L 0 Y + 16L 0 M Y + 200R 0 + 200Z 0
32L 0 Y + 200Z 0
40L 0 + 32L 0 Y
0.01
0.1
0.21
0.31
0.43
0.41
21
302
1177
7254
2317
2207
−0.50M + 0.5 ∗ M 2
0.02
284
1+M
1+M
0.04
0.03
109
145
A Experimental Results
mult
eval_simpl
eval
File: mergesort.raml
split
merge
mergesort
mergesort_list
File: quicksort.raml
partition
quicksort
quicksort_pairs
quicksort_list
File: square_mult.raml
square_mult
File: subsequence.raml
lcs
File: running.raml
abmap
asort
asort’
btick
abfoldr
cons_all
File: ocaml_sort.raml
merge
list
File: ocaml_list.raml
length
cons
hd
tl
nth
append
rev_append
rev
flatten
concat
map
mapi
rev_map
iter
iteri
fold_left
fold_right
map2
rev_map2
iter2
fold_left2
fold_right2
for_all
exists
for_all2
exists2
mem
memq
assoc
assq
mem_assoc
mem_assq
69
0.5M + 0.5M 2
K LN + 0.5K N + L + X
0
0.06
0.23
0.2
409
9498
10678
0
L+M
−0.5K + 0.5K 2
−0.5K + 0.5K 2
0.05
0.05
0.16
0.42
133
378
4567
14946
M
−0.5M + 0.5M 2
−0.5M + 0.5M 2
−0.5M + 0.5M 2
0.05
0.07
0.09
0.26
170
1687
2046
8628
K 0 + 2L 0
0.07
199
2.5L
0.05
877
70
Towards Automatic Resource Bound Analysis for OCaml
remove_assoc
remove_assq
find
find_all
filter
partition
split
combine
merge
chop
stable_sort
sort
fast_sort
File: aws.raml
average_grade
M
greater_eq
2M
sort_students
LM + LM 2
make_table
LM
find
0
lookup
0
average_grade’
0
greater_eq’
0
sort_students_efficient
LM
File: PROTOTYPE
File: appendAll.raml
appendAll
n1 n2 + n1
appendAll2
n 1 n 2 + 2n 1 n 2 n 3 + n 1
appendAll3
n 1 n 2 + 3n 1 n 2 n 3 n 4 + n 1 n 2 n 3 + n 1
File: duplicates.raml
eq
n2
remove
L + LY
nub
−0.5N n + 0.5N 2 n + 0.5N + 0.5N 2
File: dyade.raml
multList
n
dyade
n1 n2 + n1
File: eratosthenes.raml
filter
n
eratos
0.5n + 0.5n 2
File: bitvectors.raml
bitToInt’
M
bitToInt
n
sum
1
add’
2M
add
2n 1
diff
1
sub’
2M
sub
1 + 2n 1
mult
2n 1 n 2 + n 1
compare
M
leq
1 + n1
File: flatten.raml
flatten
0.5nl + 0.5n 2 l + n
insert
n
insertionsort
0.5n + 0.5n 2
flattensort
nl + 0.5n 2 l + 0.5n 2 l 2 + n
File: listsort.raml
isortlist
−0.5nl + 0.5n 2 l + 0.5n + 0.5n 2
File: longestCommonSubsequence.raml
firstline
M
right
1
0.04
0.05
0.31
0.07
0.06
0.05
0.09
0.13
0.72
136
610
8733
900
75
224
1000
2365
19030
0.04
0.2
1.57
184
1525
12724
0.01
0.1
0.52
40
583
3864
0.01
0.05
21
217
0.01
0.03
27
143
0.02
0.01
0.01
0.03
0.03
0.03
0.06
0.06
0.11
0.06
0.07
22
24
18
50
52
16
55
58
387
42
47
0.41
0.01
0.04
2.32
3351
24
144
19074
0.39
2621
0.02
0.02
48
12
A Experimental Results
max
1
newline
3M
lcstable
L + 3LM + M
lcs
1 + L + 3LM + M
File: mergesort.raml
msplit
0.5M
merge
L+M
mergesortBuggy
−
mergesort
−1.5M + 1.5M 2
File: minsort.raml
findMin
M
minSort
1.5M + 0.5M 2
File: queue.raml
empty
1
enqueue
1
enqueues
2M
copyover
M
dequeue
1 + 2M
children
1
breadth
7B R + 7B RT + 7LM Y + 4M + 7M Y + 2R
startBreadth
2.5M + 3.5M 2
depth
2L + 3LM
startDepth
1 − 1.5M + 1.5M 2
File: quicksort_mutual.raml
part
1 + L + 2LY + L 2 + 2M + 2M Y + M 2 + 3Y + Y 2
quicksortMutual
M + M2
File: rationalPotential.raml
zip3
L
group3
0.33M
File: sizechange.raml
r1
M
rev
M
f
LM + 1.67M + 0.33M 3
g
1 + L + LM + 1.67M + M Y + 0.33M 3
f2
−
last
M
f2’
1+L +M
g3
M
f3
1 + 2L + M
File: splitandsort.raml
insert
M
split
0.5M + 0.5M 2
splitqs
M
quicksort
M2
sortAll
L2 M + M
splitAndSort
1.5M + 1.5M 2
File: subtrees.raml
subtrees
0.5M + 0.5M 2
File: tuples.raml
attach
M
pairs
M2
pairsAux
M2
pairsSlow
0.83M + 0.17M 3
triples
1.83M − 1.5M 2 + 0.67M 3
quadruples
−0.67M + 2.75M 2 − 1.33M 3 + 0.25M 4
File: array_dijkstra.raml
makeGraph
dijkstra
File: CompCert
71
0.01
0.08
0.15
0.15
5
272
768
791
0.04
0.05
−
0.14
85
203
f ai l
778
0.04
0.04
93
175
0.01
0.03
0.09
0.07
0.16
0.22
3.54
3.83
4.02
4.33
1
10
183
210
631
294
10413
10663
18631
18656
0.14
0.14
836
836
0.02
0.01
47
33
0.05
0.04
0.12
0.12
−
0.04
0.09
0.04
0.08
113
115
546
546
f ai l
93
302
113
265
0.13
0.17
0.05
0.27
0.49
0.61
397
573
152
1543
3085
3658
0.12
646
0.06
0.24
0.25
0.3
1.05
2.79
82
1391
1003
1391
5728
15498
72
File: String0.ml
string_dec
prefix
File: Tuples.ml
uncurry
File: Specif.ml
projT1
projT2
value
File: EquivDec.ml
equiv_dec
File: Datatypes.ml
implb
xorb
negb
fst
snd
length
app
coq_CompOpp
coq_CompareSpec2Type
coq_CompSpec2Type
File: Bool.ml
bool_dec
eqb
iff_reflect
File: Ring.ml
bool_eq
File: Peano.ml
plus
max
min
nat_iter
File: List0.ml
hd
tl
in_dec
nth_error
remove
rev
rev_append
rev’
list_eq_dec
map
fold_left
fold_right
existsb
forallb
filter
File: EqNat.ml
beq_nat
File: Compare_dec.ml
le_lt_dec
le_gt_dec
nat_compare
File: BinPosDef.ml
succ
add
add_carry
pred_double
Towards Automatic Resource Bound Analysis for OCaml
A Experimental Results
pred
pred_N
mask_rect
mask_rec
succ_double_mask
double_mask
double_pred_mask
pred_mask
sub_mask
sub_mask_carry
sub
mul
iter
pow
square
div2
div2_up
size_nat
size
compare_cont
compare
min
max
eqb
leb
ltb
sqrtrem_step
sqrtrem
sqrt
gcdn
gcd
ggcdn
ggcd
coq_Nsucc_double
coq_Ndouble
coq_lor
coq_land
ldiff
coq_lxor
shiftl_nat
shiftr_nat
shiftl
shiftr
testbit_nat
testbit
iter_op
to_nat
of_nat
of_succ_nat
digits2_pos
coq_Zdigits2
File: BinNat.ml
succ_double_n
double_n
succ_n
pred_n
succ_pos_n
add_n
sub_n
mul_n
73
74
compare_n
eqb_n
leb_n
ltb_n
min_n
max_n
div2_n
even_n
odd_n
pow_n
square_n
log2_n
size_n
size_nat_n
pos_div_eucl
div_eucl
div
modulo
gcd_n
ggcd_n
coq_lor_n
coq_land_n
ldiff_n
coq_lxor_n
shiftl_nat
shiftr_nat
shiftl_n
shiftr_n
testbit_nat_n
testbit_n
to_nat_n
of_nat_n
iter_n
discr
binary_rect
binary_rec
leb_spec0
ltb_spec0
log2_up
lcm
eqb_spec
b2n
setbit
clearbit
ones
lnot
max_case_strong
max_case
max_dec
min_case_strong
min_case
min_dec
max_case_strong_pd
max_case
max_dec_pd
min_case_strong_pd
min_case
min_dec_pd
Towards Automatic Resource Bound Analysis for OCaml
| 6cs.PL
|
Fast Exact k-Means, k-Medians and Bregman Divergence
Clustering in 1D
arXiv:1701.07204v3 [cs.DS] 30 Jun 2017
Allan Grønlund∗
Kasper Green Larsen†
Jesper Sindahl Nielsen
§
Stefan Schneider
Alexander Mathiasen‡
¶
Mingzhou Song
k
Abstract
The k-Means clustering problem on n points is NP-Hard for any dimension d ≥ 2, however, for
the 1D case there exist exact polynomial time algorithms. Previous literature reported an O(kn2 ) time
dynamic programming algorithm that uses O(kn) space. We present a new algorithm computing the
optimal√clustering in only O(kn) time using linear space. For k = Ω(lg n), we improve this even further
to n2O( lg lg n lg k) time. We generalize the new algorithm(s) to work for the absolute distance instead of
squared distance and to work for any Bregman Divergence as well.
∗ Aarhus University. Email: [email protected]. Supported by MADALGO - Center for Massive Data Algorithmics, a Center
of the Danish National Research Foundation.
† Aarhus University. Email: [email protected]. Supported by MADALGO, a Villum Young Investigator Grant and an AUFF
Starting Grant.
‡ Aarhus University. Email: [email protected]. Supported by MADALGO and an AUFF Starting Grant.
§ Aarhus University. Email: [email protected]. Supported by MADALGO.
¶ University of California, San Diego. Email: [email protected]. Supported by NSF grant CCF-1213151 from the
Division of Computing and Communication Foundations. Any opinions, findings and conclusions or recommendations expressed
in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation.
k New Mexico State University. Email: [email protected]
1
Introduction
Clustering is the problem of grouping elements into clusters such that each element is similar to the elements
in the cluster assigned to it and not similar to elements in any other cluster. It is one of, if not the, primary
problem in the area of machine learning known as Unsupervised Learning and no clustering problem is as
famous and widely considered as the k-MeansPproblem: Given a multiset X = {x1 , ..., xn } ⊂ Rd find k
centroids M = {µ1 , ..., µk } ⊂ Rd minimizing x∈X minµ∈M ||x − µ||2 . Several NP-Hardness results exist
for finding the optimal k-Means clustering in general, forcing one to turn towards heuristics. k-Means is
NP-hard even for k = 2 and general dimension [4] and it is also NP-hard for d = 2 and general k [16]. Even
hardness of approximation results exist [15, 8]. In [8] the authors show there exists a ε > 0 such that it is
NP-hard to approximate k-Means to within a factor 1 + ε of optimal, and in [15] it is proved that ε ≥ 0.0013.
On the upper bound side the best known polynomial time approximation algorithm for k-Means has an
approximation factor of 6.357 [3]. In practice, Lloyd’s algorithm is a popular iterative local search heuristic
that starts from some random or arbitrary clustering. The running time of Lloyd’s algorithm is O(tknd)
where t is the number of rounds of the local search procedure. In theory, if Lloyd’s algorithm is run to
convergence to a local minimum, t could be exponential and there is no guarantee on how well the solution
found approximates the optimal solution [6, 20]. Lloyd’s algorithm is often combined with the effective
seeding technique for selecting initial centroids due to [7] that gives an expected O(lg k) approximation ratio
for the initial clustering, which can then improved further by Lloyd’s algorithm.
For the one-dimensional case, the k-Means problem is not NP-hard. In particular, there is an O(kn2 )
time and O(kn) space dynamic programming solution for the 1D case, due to work by [22]. The 1D kMeans problem is encountered surprisingly often in practice, some examples being in data analysis in social
networks, bioinformatics and retail market [5, 13, 18].
It is only natural to try other reasonable distance measures for the data considered and define different
clustering problems. There are many other choices than the sum of squares of the Euclidian distances that
define k-Means. For instance, one could use any Lp norm instead. The special case of p = 1 is known as
k-Medians clustering and has also received considerable attention. The k-Medians problems is also NP-hard
and the best polynomial time approximation algorithms has an approximation factor of 2.633 [8].
In [9] the authors consider and define clustering with Bregman Divergences. Bregman Divergence generalize squared Euclidian distance and thus Bregman Clusterings include the k-Means problem, as well as a
wide range of other clustering problems that can be defined from Bregman Divergences like e.g. clustering
with KullbackLeibler divergence as the cost. Interestingly, the heuristic local search algorithm for Bregman
Clustering [9] is basically the same approach as Lloyd’s algorithm for k-Means. Clustering with Bregman
Divergences is clearly NP-Hard as well since it includes k-Means clustering. We refer the reader to [9] for
more about the general problem. For the 1D version of the problem, [17] generalized the algorithm from
[22] to the k-Medians problems and Bregman Divergences achieving the same O(kn2 ) time and O(kn) space
bounds.
1.1
Our Results
In this paper we give theoretically and practically efficient algorithm for 1D clustering problems, in particular
k-Means.
The k-Means clustering problem in 1D is defined as follows. Given X = {x1 , ..., xn } ⊂ R and k, find
centroids M = {µ1 , ..., µk } ⊂ R minimizing the cost
X
min (x − µ)2
x∈X
µ∈M
The main results of this paper are new fast algorithms for 1D k-Means. First we give an algorithm that
computes the optimal k-Means clustering that runs in O(n lg n + kn) time using optimal O(n) space, or
O(kn) time if the input is already sorted. The algorithm also computes the cost of the optimal clustering
using k ′ clusters for all k ′ ≤ k. This is relevant for instance for model selection of the right k. This is an
1
improvement by a factor of n in time and k in space compared to the existing solution (which also supports
computing the cost for all k ′ ≤ k. The constant factors hidden by the O-notation are small and we expect this
algorithm
to be very efficient in practice. Second, we show how to compute an optimal k-Means clustering in
√
n2O( lg lg n lg k) time using O(n) space. This algorithm is mainly of theoretical interest as we expect constants
to be rather large. As opposed to the O(kn) time algorithm, this algorithm does not compute the optimal
′
′
costs for using
√ k clusters for all k ≤ k.
O( lg lg n lg k)
time algorithm relates to a natural regularized version of k-Means clustering where
The n2
instead of specifying the number of clusters beforehand, we instead specify a cost of using an extra cluster
and then minimize the cost of the clustering plus the cost of the number of clusters used. Formally, the
problem is as follows: Given X = {x1 , ..., xn } ⊂ R and λ, compute the optimal regularized clustering:
X
min (x − µ)2 + λk
arg min
k,M={µ1 ,...,µk } x∈X µ∈M
We show that this problem is solvable in O(n) time if the input is sorted.
In 1D, Lloyd’s algorithm can be implemented to run in O(n lg n + tk lg n) time where t is the number of
rounds, and we expect that our algorithm can compute the optimal clustering for reasonable k in essentially
the same time as Lloyd’s algorithm can approximate it.
The k-Medians problem is to compute a clustering that minimize the sum of absolute distances to the
centroid, i.e. compute M = {µ1 , ..., µk } ⊂ R that minimize
X
min |x − µ|
x∈X
µ∈M
Our algorithms generalize naturally to solve this problem in the same time bounds as for the k-means
problem.
Let f be a differentiable real-valued strictly convex function. The Bregman Divergence Df induced by f
is defined as
Df (x, y) = f (x) − f (y) − ∇f (y)(x − y)
Notice that the Bregman Divergence induced from f (x) = x2 , gives squared Euclidian Distance (k-Means).
Bregman divergences are not metrics since they are not symmetric in general and the triangle inequality is
not necessarily satisfied. They do however have many redeeming qualities, for instance Bregman Divergences
are convex in the first argument, albeit not the second, see [9, 10] for a more comprenhensive treatment.
The Bregman Clustering problem as defined in [9] is to find k centroids M = {µ1 , ..., µk } that minimize
X
min Df (x, µ)
x∈X
µ∈M
where Df is a Bregman Divergence. For our case, where the inputs x, y ∈ R, we assume that computing a
Bregman Divergence, i.e. evaluating f and its derivative, takes constant time. We show that our algorithms
naturally generalize to 1D clustering using any Bregman Divergence to define the cluster cost while still
maintaing the same running time as for k-Means.
Implementation. An independent implementation of the O(n lg n + kn) time algorithm is available in the
R package Ckmeans.1d.dp [21]. The implementation is for k-Means clustering, and uses O(kn) space.
1.2
Outline
In Section 2 we describe the existing O(kn2 ) time algorithm for 1D k-Means clustering that uses O(kn)
space. In Section 3 we show how to compute the same output as the old algorithm
using only O(kn) time
√
and O(n) space. Then we show how to improve the running time to O(n2O( lg lg n lg k) ). Finally, in Section
4 we show how our new algorithms generalizes to different cluster costs than squared Euclidian distance.
2
2
The O(kn2 ) Dynamic Programming Algorithm
In this section, we describe the previous O(kn2 ) time and O(kn) space algorithm presented in [22]. We also
introduce the definitions and notation we use in our new algorithm. We will always assume sorted input
x1 ≤ ... ≤ xn ∈ R. If the input is not sorted, we start by sorting it in O(n lg n) time. We also remark
that there could be many ways of partitioning the point set and computing centroids that achieve the same
cost. This is for instance the case if the input is n identical points. The task at hand is to find any optimal
solution.
P
Let CC(i, j) = jℓ=i (xℓ − µi,j )2 be the cost of grouping xi , ..., xj into one cluster with the optimal choice
Pj
1
of centroid, µi,j = j−i+1
ℓ=i xℓ , the arithmetic mean of the points.
Lemma 1. There is an O(n) space data structure that can compute CC(i, j) in O(1) time for any i ≤ j
using O(n) time preprocessing.
Proof. This is a standard application of prefix sums and works as follows. By definition,
CC(i, j) =
j
X
(xℓ − µi,j )2 =
ℓ=i
j
X
x2ℓ + µ2i,j − 2xℓ µi,j = (j − i + 1)µ2i,j + µi,j
ℓ=i
j
X
ℓ=i
xℓ +
j
X
x2ℓ .
ℓ=i
With access to prefix sum arrays of x1 , . . . , xn and x21 , . . . , x2n both the centroid µi,j and the cost is easily
computed in constant time .
2.1
Algorithm Sketch
The algorithm computes the optimal clustering using i clusters for all prefixes of input points x1 , . . . , xm ,
for m = 1, . . . , n, and for all i = 1, . . . , k using Dynamic Programming as follows.
Let D[i][m] be the cost of optimally clustering x1 , ..., xm into i clusters. For i = 1 the cost of optimally
clustering x1 , ..., xm into one cluster is the cluster cost CC(1, m). That is, D[1][m] = CC(1, m) for all m.
This can be computed in O(n) time by Lemma 1.
For i > 1
m
D[i][m] = min D[i − 1][j − 1] + CC(j, m)
(1)
j=1
Notice that D[i − 1][j − 1] is the cost of optimally clustering x1 , ..., xj−1 into i − 1 clusters and CC(j, m)
is the cost of clustering xj , ..., xm into one cluster. This makes xj the first point in the last and rightmost
cluster. Let T [i][m] be the argument that minimizes (1)
m
T [i][m] := arg min D[i − 1][j − 1] + CC(j, m)
j=1
(2)
It is possible there exists multiple j obtaining same minimal value for (1). To make the optimal clustering
C˜(m) unique, such ties are broken in favour of smaller j.
Notice xT [i][m] is the first point in the rightmost cluster of the optimal clustering. Thus, given T one can
find the optimal solution by standard backtracking:
X̃k = {xT [k][n] , ..., xn },
X̃k−1 = {xT [k−1][T [k][n]−1] , ..., xT [k][n]−1 }
..
.
Here X̃i is the i’th cluster in the optimal clustering. One can naively compute each entry of D and T using
(1) and (2). This takes O(n) time for each cell, thus D and T can be computed in O(kn2 ) time using O(kn)
space. This is exactly what is described in [22].
3
3
New Algorithms
The idea of the first new algorithm is simply to compute the tables D and T faster, by reducing the time to
compute each row of D and T to O(n) time instead of O(n2 ) time. This improvement exploits a monotonicity
property of the values stored in a row of T . This is explained in Section 3.1, resulting in an O(kn) time and
O(kn) space solution, assuming sorted inputs. Section 3.2 then shows how to reduce the space usage to just
O(n) while retaining O(kn) running time.
In Section 3.3 we show that the same property allows us to solve
√
O( lg lg n lg k)
1D k-Means for k = Ω(lg n) in, n2
time and linear space, and solve the regularized version of
1D k-Means in O(n) time.
3.1
Faster Algorithm From Monotone Matrices
In this section we reduce the problem of computing a row of D and T to searching an implicitly defined
n × n matrix of a special form, which allows us to compute each row of D and T in linear time.
Define Ci [m][j] as the cost of the optimal clustering of x1 , . . . , xm using i clusters, restricted to having
the rightmost cluster (largest cluster center) contain the elements xj , . . . , xm . For convenience, we define
Ci [m][j] for j > m as the cost of clustering x1 , . . . , xm into i − 1 clusters, i.e. the last cluster is empty. This
means that Ci satisfies:
Ci [m][j] = D[i − 1][min{j − 1, m}] + CC(j, m)
where by definition CC(j, m) = 0 when j > m (which is consistent with the definition in Section 2). We
have that D[i][m] relates to Ci as follows:
D[i][m] = min Ci [m][j]
j
where ties are broken in favor of smaller j (as defined in Section 2.1).
This means that when we compute a row of D and T , we are actually computing minj Ci [m][j] for all
m = 1, . . . , n. We think of Ci as an n × n matrix with rows indexed by m and columns indexed by j. With
this interpretation, computing the i’th row of D and T corresponds to computing for each row r in Ci ,
the column index c that corresponds to the smallest value in row r. In particular, the entries D[i][m] and
T [i][m] correpond to the value and the index of the minimum entry in the m’th row of Ci respectively. The
problem of finding the minimum value in every row of a matrix has been studied before [1]. First we need
the definition of a monotone matrix.
Definition 1. [1] Let A be a matrix with real entries and let arg min(i) be the index of the leftmost column
containing the minimum value in row i of A. A is said to be monotone if a < b implies that arg min(a) ≤
arg min(b). A is totally monotone if all of its submatrices are monotone.1
In [1], the authors showed the following:
Theorem 1. [1] Finding arg min(i) for each row i of an arbitrary n× m monotone matrix requires Θ(m lg n)
time, whereas if the matrix is totally monotone, the time is O(m) when m > n and is O(m(1 + lg(n/m)))
when m < n.
The fast algorithm for totally monotone matrices is known as the SMAWK algorithm and we will refer
to it by that (cool) name.
Let’s relate this to the 1D k-Means clustering problem. That Ci is monotone means that if we consider the
optimal clustering of the points x1 , . . . , xa with i clusters, then if we start adding more points xa+1 ≤ · · · ≤ xb
after xa , then the first (smallest) point in the last of the i clusters can only increase (move right) in the new
optimal clustering of x1 , . . . , xb . This sounds like it should be true for 1D k-Means and it turns out it is.
Thus, applying the algorithm for monotone matrices, we can fill a row of D and T in O(n lg n) time leading
to an O(kn lg n) time algorithm for 1D k-Means, which is already a great improvement.
1 In
[1] the authors use the maximum instead of the minimum
4
However, as we show below, the matrix Ci induced by the 1D k-Means problem is in fact totally monotone:
Lemma 2. The matrix Ci is totally monotone.
Proof. As [1] remarks, a matrix A is totally monotone if all its 2 × 2 submatrices are monotone. To prove
that Ci is totally monotone, we thus need to prove that for any two row indices a, b with a < b and two
column indices u, v with u < v, it holds that if Ci [a][v] < Ci [a][u] then Ci [b][v] < Ci [b][u].
Notice that these values correspond to the costs of clustering elements x1 , . . . , xa and x1 , . . . , xb , starting
the rightmost cluster with element xv and xu respectively. Since Ci [m][j] = D[i−1][min{j−1, m}]+CC(j, m),
this is the same as proving that
D[i − 1][min{v − 1, m}] + CC(v, a) < D[i − 1][min{u − 1, m}] + CC(u, a) ⇒
D[i − 1][min{v − 1, m}] + CC(v, b) < D[i − 1][min{u − 1, m}] + CC(u, b)
which is true if we can prove that CC(v, b) − CC(v, a) ≤ CC(u, b) − CC(u, a). Rearranging terms, what we
need to prove is that for any a < b and u < v, it holds that:
CC(v, b) + CC(u, a) ≤ CC(u, b) + CC(v, a).
(3)
This is the property known as the concave (concave for short) property [24, 12, 23] and has been used to
significantly speed up algorithms, including Dynamic Programming algorithms, for for other problems. We
start by handling the special case where v > a. In this case, we have by definition that CC(v, a) = 0, thus
we need to show that CC(v, b) + CC(u, a) ≤ CC(u, b). This is the case since any point amongst xu , . . . , xb
is included in at most one of xv , . . . , xb and xu , . . . , xa (since a < v). Thus CC(v, b) + CC(u, a) is the cost
of taking two disjoint and consecutive subsets of the points xu , . . . , xb and clustering the two sets using the
optimal choice of centroid in each. Clearly this cost is less than clustering all the points using one centroid.
We now turn to the general case where u < v ≤ a < b. Let µv,a be the mean of xv , . . . , xa and
µu,b be the mean
P of xu , . . . , xb and assume that µv,a ≤ µu,b (the other case is symmetric). Finally, let
CC(w, z)µ = zℓ=w (xℓ − µ)2 denote the cost of grouping the elements xw , . . . , xz in a cluster with centroid
µ. Split the cost CC(u, b) into the cost of the elements xu , . . . , xv−1 and the cost of the elements xv , . . . , xb
as
CC(u, b) =
b
X
(xℓ − µu,b )2 =
ℓ=u
v−1
X
(xℓ − µu,b )2 +
ℓ=u
b
X
(xℓ − µu,b )2 = CC(u, v − 1)µu,b + CC(v, b)µu,b .
ℓ=v
We trivially get CC(v, b)µu,b ≥ CC(v, b) since CC(v, b) is the cost using the optimal centroid. Secondly,
CC(u, v − 1)µu,b + CC(v, a) ≥ CC(u, v − 1)µv,a + CC(v, a) = CC(u, a)µv,a ≥ CC(u, a)
since µv,a ≤ µu,b and all elements xu , . . . , xv−1 are less than or equal to µv,a (since µv,a is the mean of points
xv , . . . , xa that all are greater than xu , . . . , xv−1 ). Combining the results, we see that:
CC(v, b) + CC(u, a) ≤ CC(v, b)µu,b + CC(u, v − 1)µu,b + CC(v, a) = CC(u, b) + CC(v, a).
This completes the proof.
Theorem 2. Computing an optimal k-Means clustering of a sorted input of size n for takes O(kn) time.
By construction the cost of the optimal clustering is computed for all k ′ ≤ k. If we store the T table the
cluster centers for any k ′ ≤ k can be extracted in O(k ′ ) time.
5
3.2
Reducing Space Usage
In the following, we show how to reduce the space usage to just O(n) while maintaining O(kn) running time
using a space reduction technique of Hirschberg [11]. First observe that each row of T and D only refers to
the previous row. Thus one can clearly “forget”row i − 1 when we are done computing row i. The problem
is that if we don’t store all of T , we cannot backtrack and find the optimal solution. In the following, we
present an algorithm that avoids the table T entirely.
Our key observation is the following: Assume k > 1 and that for every prefix x1 , . . . , xm , we have
computed the optimal cost of clustering x1 , . . . , xm into ⌊k/2⌋ clusters. Note that this is precisely the set
of values stored in the ⌊k/2⌋’th row of D. Assume furthermore that we have computed the optimal cost of
clustering every suffix xm , . . . , xn into k − ⌊k/2⌋ clusters. Let us denote these costs by D̃[k − ⌊k/2⌋][m] for
m = 1, . . . , n. Then clearly the optimal cost of clustering x1 , . . . , xn into k clusters is given by:
n
D[k][n] = min D[⌊k/2⌋][j] + D̃[k − ⌊k/2⌋][j + 1].
j=1
(4)
Our main idea is to first compute row ⌊k/2⌋ of D and row k − ⌊k/2⌋ of D̃ using linear space. From these two,
we can compute the argument j minimizing (4). We can then split the reporting of the optimal clustering
into two recursive calls, one reporting the optimal clustering of points x1 , . . . , xj into ⌊k/2⌋ clusters, and one
call reporting the optimal clustering of xj+1 , . . . , xn into k − ⌊k/2⌋ clusters. When the recursion bottoms
out with k = 1, we can clearly report the optimal clustering using linear space and time as this is just the
full set of points.
From Section 3.1 we already know how to compute row ⌊k/2⌋ of D using linear space: Simply call
SMAWK to compute row i of D for i = 1, . . . , ⌊k/2⌋, where we throw away row i − 1 of D (and don’t
even store T ) when we are done computing row i. Now observe that table D̃ can be computed by taking
our points x1 , . . . , xn and reversing their order by negating the values. This way we obtain a new ordered
sequence of points X̃ = x̃1 ≤ x̃2 ≤ · · · ≤ x̃n where x̃i = −xn−i+1 . Running SMAWK repeatedly for
i = 1, . . . , k − ⌊k/2⌋ on the point set X̃ produces a table D̂ such that D̂[i][m] is the optimal cost of clustering
x̃1 , . . . , x̃m = −xn , . . . , −xn−m+1 into i clusters. Since this cost is the same as clustering xn−m+1 , . . . , xn
into i clusters, we get that the (k − ⌊k/2⌋)’th row of D̂ is identical to the i’th row of D̃ if we reverse the
order of the entries.
To summarize our algorithm for reporting the optimal clustering, do as follows: Let L be an initially
empty output list of clusters. If k = 1, append to L a cluster containing all points. Otherwise (k > 1),
use SMAWK on x1 , . . . , xn and −xn , . . . , −x1 to compute row ⌊k/2⌋ of D and row k − ⌊k/2⌋ of D̃ using
linear space (by evicting row i − 1 from memory when we have finished computing row i) and O(kn) time.
Compute the argument j minimizing (4) in O(n) time. Evict row ⌊k/2⌋ of D and row k − ⌊k/2⌋ of D̃ from
memory. Recursively report the optimal clustering of points x1 , . . . , xj into ⌊k/2⌋ clusters (which appends
the output to L). When this terminates, recursively report the optimal clustering of points xj+1 , . . . , xn into
k − ⌊k/2⌋ clusters. When the algorithm terminates, L contains the optimal clustering of x1 , . . . , xn into k
clusters.
At any given time, our algorithm uses only O(n) space. To see this, first note that we evict all memory
used to compute the value j minimizing (4) before recursing. Furthermore, we complete the first recursive
call (and evict all memory used) before starting the second. Finally, for the recursion, we don’t need to
make a copy of points x1 , . . . , xj . It suffices to remember that we are only working on the subset of inputs
x1 , . . . , xj .
Now let F (n, k) denote the time used by the above algorithm to compute the optimal clustering of n
sorted points into k clusters. Then there is some constant C > 0 such that F (n, k) satisfies the recurrence:
F (n, 1) ≤ Cn,
and for k > 1:
n
F (n, k) ≤ max F (j, ⌊k/2⌋) + F (n − j, k − ⌊k/2⌋) + Cnk.
j=1
6
We claim that F (n, k) satisfies F (n, k) ≤ 3Ckn. We prove the claim by induction in k. The base case k = 1
follows trivially by inspection of the formula for F (n, 1). For the inductive step k > 1, we use the induction
hypothesis to conclude:
F (n, k) ≤
n
max 3Cj⌊k/2⌋ + 3C(n − j)(k − ⌊k/2⌋) + Cnk
j=1
n
≤
max 3Cj⌈k/2⌉ + 3C(n − j)⌈k/2⌉ + Cnk
=
3Cn⌈k/2⌉ + Ckn.
j=1
For k > 1, we have that ⌈k/2⌉ ≤ (2/3)k, therefore:
F (n, k) ≤ 3Cn(2/3)k + Ckn
= 3Ckn.
Which is what we needed to prove.
Theorem 3. Computing an optimal k-Means clustering of a sorted input of size n takes O(kn) time and
uses O(n) space.
Note to compute the cost of the optimal clustering for all k ′ ≤ k we ensure that we never delete the last
column of the cost matrix D which requires an additional O(k) = O(n) space.
3.3
Even Faster Algorithm
In this section we show that the concave property we proved for the cluster costs
yields and algorithm for
√
computing the optimal k-Means clustering for one given k = Ω(lg n) in n2O( lg lg n lg k) time. The result
follows almost directly from [19]. In [19] Schieber gives an algorithm with the aforementioned running time
for the problem of finding the shortest path of fixed length k in a directed acyclic graph with nodes 1, . . . , n
where the weights, w(i, j), satisfy the concave property and are represented as a function that returns the
weight of a given edge in constant time.
Theorem 4 ([19]). Computing a minimum weight path of length k between any
two nodes in a directed
√
acyclic graph of size n where the weights satisfy the concave property takes n2O( lg lg n lg k) time using O(n)
space.
We reduce the 1D k-Means problem to a directed graph problem as follows. Sort the input in O(n lg n)
time and let x1 ≤ x2 ≤ . . . xn denote the sorted input sequence. For each input xi we associate a node
vi and add an extra node vn+1 . Now define the weight of the edge from vi to vj as the cost of clustering
xi , . . . , xj−1 in one cluster, which is CC(i, j − 1). Each edge weight is computed in constant time and by the
proof of Lemma 2, particularly Equation 3, the edge weights satisfy the monge concave property. Finally, to
compute the optimal clustering we use Schiebers algorithm to compute the lowest weight path with k edges
from v1 to vn+1 .
Theorem
5. Computing an optimal k-Means clustering of an input of size n for given k = Ω(lg n) takes
√
O( lg lg n lg k)
time using O(n) space.
n2
It is relevant to briefly consider parts of Schiebers algorithm and how it relates to k-Means clustering, in
particular a regularized version of the problem. Schiebers algorithm relies crucially on algorithms that given
a directed acyclic graph where the weights satisfy the concave property computes a minimum weight path
in O(n) time [23, 14]. Note the only difference in this problem compared to above, is that the search is not
restricted to paths of k edges only.
7
3.3.1
Regularized Clustering
Consider a regularized version of the k-Means clustering problem where we instad of providing the number
of clusters k specify the cost of a cluster and ask to minimize the cost of the clustering plus the penalty λ
for each cluster used. For simplicity, assume all the input points are distinct.
If we set λ = 0 the optimal clustering has cost zero and use a cluster for each input point. If we let λ
increase towards infinity, the optimal number of clusters used in the optimal solution monotonically decrease
towards one (zero clusters is not well defined). Let dmin be the smallest distance between points in the input.
The optimal cost of using n − 1 clusters is then d2min /2. When λ > λn = d2min /2 it is less costly to use
only n − 1 clusters since the added clustering of using one less cluster is smaller than the cost of a cluster.
Letting λ increase again will inevitably lead to a miminum value λn−1 < λn such that for λ > λn−1 only
n − 2 clusters is used in the optimal solution. Following the same pattern λn−1 is the difference between
the optimal cost using n − 2 clusters and n − 1 clusters. Continuing this yields the very interesting event
sequence 0 < λn < · · · < λ1 that encodes the only relevant choices for the regularization parameter. Note
that the O(nk) algorithm actually yields λ1 , . . . , λk since it computes the optimal cost for all k ′ ≤ k.
In the reduction to the directed graph problem, adding a cost of λ for each cluster used corresponds to
adding λ to the weight of each edge. Note that the edge weights clearly still satisfy the concave property.
Thus, solving the regularized version of k-Means clustering correpoonds to finding the shortest path (of any
length) in a directed acyclic graph where the weights satisfy the concave property. By the algorithms in
[23, 14] this takes O(n) time.
Theorem 6. Computing an optimal regularized 1D k-Means clustering of a sorted input of size n takes O(n)
time.
Now notice if we actually use λk−1 as the cost per cluster, or any λ ∈ [λk−1 , λk [ there is an optimal
solution using k clusters which is an optimal k-Means clustering. This means that if the inputs are integers,
we can solve the 1D k-Means problem by a simple application of binary search in O(n lg U ) time where U is
the universe size [2].
4
Extending to More Distance Measures
In the following we show how to generalize our algorithm to Bregman Divergences and sum of absolute
distances while retaining the same running time and space usage.
4.1
Bregman Divergence and Bregman Clustering
In this section we show how our algorithm generalizes to any Bregman Divergence. First, let us remind
ourselves what a Bregman Divergence and a Bregman Clustering is. Let f be a differentiable real-valued
strictly convex function. The Bregman Divergence Df defined by f is defined as
Df (x, y) = f (x) − f (y) − ∇f (y)(x − y)
Bregman Clustering. The Bregman Clustering problem as defined in [9], is to find a clustering, M =
{µ1 , ..., µk }, that minimize
X
min Df (x, µ)
x∈X
µ∈M
Notice that the cluster center is the second argument of the Bregman Divergence. This is important since
Bregman Divergences are not in general symmetric.
For the purpose of 1D clustering, we mention two important properties of Bregman Divergences. For any
Bregman Divergence, the unique element that minimizes the summed distance to a multiset of elements is
8
the mean of the elements, exactly as it was for squared Euclidian distance. This is in one sense the defining
property of Bregman Divergences [9].
The second important is the linear separator property, which is very important for clustering with Bregman Divergences but also very relevavant to Bregman Voronoi Diagrams [9, 10].
Linear Separators For Bregman Divergences. For all Bregman divergences, the locus of points that
are equidistant to two fixed points µ1 , µ2 in terms of a Bregman divergence is given by
{x ∈ X | Df (x, p) = Df (x, q)} = {x ∈ X | x(∇f (µ1 ) − ∇f (µ2 )) = f (µ1 ) − µ1 ∇f (µ1 ) − f (µ2 ) + µ2 ∇f (µ2 )}
which corresponds to a hyperplane. Also, the points µ1 , µ2 sits on either side of the hyperplane and the
Voronoi cells defined using Bregman divergences are connected.
This means, in particular, that between any two points in 1D, µ1 < µ2 , there is a hyperplane (point) h
with µ1 < h < µ2 and all points smaller than h are closer to µ1 and all points larger than h are closer to µ2 .
We capture what we need from this observation in a simple “distance” lemma:
Lemma 3. Given two fixed real numbers µ1 < µ2 , then for any point xr ≥ µ2 , we have Df (xr , µ1 ) >
Df (xr , µ2 ), and for any point xl ≤ µ1 we have Df (xl , µ1 ) < Df (xl , µ2 )
Computing Cluster Costs for Bregman Divergences. Since the mean minizes Bregman Divergences,
the centroids used in optimal clusterings are unchanged compared to the k-Means case. The prefix sums idea
used to implement the data structure used for Lemma 1 generalizes to Bregman Divergences as observed
in [17] (under the name Summed Area Tables). The formula for computing the cost of grouping the points
Pj
1
xi , . . . , xj in one cluster is as follows. Let µi,j = j−i+1
ℓ=i xℓ be the arithmetic mean of the points
xi , . . . , xj , then
CC(i, j) =
j
X
Df (xℓ , µi,j ) =
=
ℓ=i
f (xℓ ) − f (µi,j ) − ∇f (µi,j )(xℓ − µi,j )
ℓ=i
ℓ=i
j
X
j
X
!
f (xℓ )
− (j − i + 1)f (µi,j ) − ∇f (µi,j )
j
X
ℓ=i
xℓ
!
− (j − i + 1)µi,j
!
It follows that the Bregman Divergence cost of a consecutive subset of input points and the centroid can be
computed in in constant time with stored prefix sums for x1 , . . . , xn and f (x1 ), . . . , f (xn ).
Monge Concave - Totally Monotone Matrix. The only properties we used in Section 3.1 to prove the
monge concave property and that the matrix Ci is totally monotone, is that the mean is the minimizer of
the sum of distances to a multiset of points, and that
CC(u, v − 1)µu,b + CC(v, a) ≥ CC(u, v − 1)µv,a + CC(v, a) = CC(u, a)µv,a
when µv,a ≤ µu,b and all elements in xu , . . . , xv−1 ≤ µv,a . This is clearly still true by Lemma 3.
It follows that the algorithms we specified for 1D k-Means generalize to any Bregman Divergence.
4.2
k-Medians - Clustering with sum of absolute values
For the k-Medians problem we replace the the sum of squared Euclidian distances with the sum of absolute
distances. Formally, the k-Medians problem is to compute a clustering, M = {µ1 , ..., µk }, minimizing
X
min |x − µ|
x∈X
µ∈M
Note that in 1D, all Lp norms are the same and reduce to this case. Also note that the minimizing centroid
for a cluster is no longer the mean of the points in that cluster, but the median. To solve this problem, we
9
change the centroid to be the median, and if there an even number of points, we fix the median to be the
exact middle point between the two middle elements, making the choice of centroid unique.
As for Bregman Divergences, we need to show that we can compute the cost CC(i, j) with this new cost
in constant time. Also, we need to compute the centroid in constant time and argue that the cost is monge
moncave which implies the implicit matrix Ci is totally monotone. The arguments are essentially the same,
but for completeness we briefly cover them below.
Computing Cluster Costs for Absolute Distances. Not surprisingly, using prefix sums still allow
x⌊mi,j ⌋ +x⌈mi,j ⌉
constant time computation of CC(i, j). Let mi,j = j+i
2 , and compute the centroid as µi,j =
2
CC(i, j) =
j
X
ℓ=i
|xℓ − µi,j | =
⌊mi,j ⌋
X
ℓ=i
µi,j − xℓ +
j
X
xℓ − µi,j
ℓ=1+⌊mi,j ⌋
which can be computed in constant time with access to a prefix sum table of x1 , . . . , xn . This was also
observed in [17].
Monge Concave - Totally Monotone Matrix. The monge concave and totally monotone matrix argument above for Bregman Divergences (and for squared Euclidian distance) remain valid since first of all,
we still have xu , . . . , xv−1 ≤ µv,a as µv,a is the median of points all greater than xu , . . . , xv−1 . Furthermore,
it still holds that when µv,a ≤ µu,b and all elements xu , . . . , xv−1 are less than or equal to µv,a , then:
CC(u, v − 1)µu,b + CC(v, a) ≥ CC(u, v − 1)µv,a + CC(v, a) = CC(u, a)µv,a
It follows that the algorithms we specified for 1D k-Means generalize to the 1D k-Median problem.
Acknowledgements
We wish to thank Pawel Gawrychowski for pointing out important earlier work on concave property.
References
[1] A. Aggarwal, M. M. Klawe, S. Moran, P. Shor, and R. Wilber. Geometric applications of a matrixsearching algorithm. Algorithmica, 2(1):195–208, 1987.
[2] A. Aggarwal, B. Schieber, and T. Tokuyama. Finding a minimum-weightk-link path in graphs with the
concave monge property and applications. Discrete & Computational Geometry, 12(3):263–280, 1994.
[3] S. Ahmadian, A. Norouzi-Fard, O. Svensson, and J. Ward. Better guarantees for k-means and euclidean
k-median by primal-dual algorithms. CoRR, abs/1612.07925, 2016.
[4] D. Aloise, A. Deshpande, P. Hansen, and P. Popat. Np-hardness of euclidean sum-of-squares clustering.
Machine Learning, 75(2):245–248, 2009.
[5] V. Arnaboldi, M. Conti, A. Passarella, and F. Pezzoni. Analysis of ego network structure in online
social networks. In Privacy, security, risk and trust (PASSAT), 2012 international conference on and
2012 international confernece on social computing (SocialCom), pages 31–40. IEEE, 2012.
[6] D. Arthur and S. Vassilvitskii. How slow is the k-means method? In Proceedings of the Twenty-second
Annual Symposium on Computational Geometry, SCG ’06, pages 144–153, New York, NY, USA, 2006.
ACM.
10
[7] D. Arthur and S. Vassilvitskii. k-means++: The advantages of careful seeding. In Proceedings of
the eighteenth annual ACM-SIAM symposium on Discrete algorithms, pages 1027–1035. Society for
Industrial and Applied Mathematics, 2007.
[8] P. Awasthi, M. Charikar, R. Krishnaswamy, and A. K. Sinop. The hardness of approximation of
euclidean k-means. In L. Arge and J. Pach, editors, 31st International Symposium on Computational
Geometry, SoCG 2015, June 22-25, 2015, Eindhoven, The Netherlands, volume 34 of LIPIcs, pages
754–767. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2015.
[9] A. Banerjee, S. Merugu, I. S. Dhillon, and J. Ghosh. Clustering with bregman divergences. J. Mach.
Learn. Res., 6:1705–1749, Dec. 2005.
[10] J.-D. Boissonnat, F. Nielsen, and R. Nock. Bregman voronoi diagrams. Discrete & Computational
Geometry, 44(2):281–307, 2010.
[11] D. S. Hirschberg. A linear space algorithm for computing maximal common subsequences. Commun.
ACM, 18(6):341–343, June 1975.
[12] D. S. Hirschberg and L. L. Larmore. The least weight subsequence problem. SIAM Journal on Computing, 16(4):628–638, 1987.
[13] O. Jeske, M. Jogler, J. Petersen, J. Sikorski, and C. Jogler. From genome mining to phenotypic microarrays: Planctomycetes as source for novel bioactive molecules. Antonie Van Leeuwenhoek, 104(4):551–567,
2013.
[14] M. M. Klawe. A simple linear time algorithm for concave one-dimensional dynamic programming.
Technical report, Vancouver, BC, Canada, Canada, 1989.
[15] E. Lee, M. Schmidt, and J. Wright. Improved and simplified inapproximability for k-means. Information
Processing Letters, 120:40–43, 2017.
[16] M. Mahajan, P. Nimbhorkar, and K. Varadarajan. The Planar k-Means Problem is NP-Hard, pages
274–285. Springer Berlin Heidelberg, Berlin, Heidelberg, 2009.
[17] F. Nielsen and R. Nock. Optimal interval clustering: Application to bregman clustering and statistical
mixture learning. IEEE Signal Process. Lett., 21:1289–1292, 2014.
[18] D. Pennacchioli, M. Coscia, S. Rinzivillo, F. Giannotti, and D. Pedreschi. The retail market as a complex
system. EPJ Data Science, 3(1):1, 2014.
[19] B. Schieber. Computing a minimum weightk-link path in graphs with the concave monge property.
Journal of Algorithms, 29(2):204 – 222, 1998.
[20] A. Vattani. k-means requires exponentially many iterations even in the plane. Discrete & Computational
Geometry, 45(4):596–616, 2011.
[21] H. Wang and J. Song. Ckmeans.1d.dp: Optimal and fast univariate clustering; R package version 4.0.0.,
2017.
[22] H. Wang and M. Song. Ckmeans. 1d. dp: optimal k-means clustering in one dimension by dynamic
programming. The R Journal, 3(2):29–33, 2011.
[23] R. Wilber. The concave least-weight subsequence problem revisited. Journal of Algorithms, 9(3):418 –
425, 1988.
[24] F. F. Yao. Efficient dynamic programming using quadrangle inequalities. In Proceedings of the Twelfth
Annual ACM Symposium on Theory of Computing, STOC ’80, pages 429–435, New York, NY, USA,
1980. ACM.
11
| 8cs.DS
|
c Copyright notice.
1
Maximum Persistency via Iterative Relaxed
Inference in Graphical Models
arXiv:1508.07902v3 [cs.CV] 3 Feb 2017
Alexander Shekhovtsov, Paul Swoboda, and Bogdan Savchynskyy
Abstract—We consider the NP-hard problem of MAP-inference for undirected discrete graphical models. We propose a polynomial
time and practically efficient algorithm for finding a part of its optimal solution. Specifically, our algorithm marks some labels of the
considered graphical model either as (i) optimal, meaning that they belong to all optimal solutions of the inference problem;
(ii) non-optimal if they provably do not belong to any solution. With access to an exact solver of a linear programming relaxation to the
MAP-inference problem, our algorithm marks the maximal possible (in a specified sense) number of labels. We also present a version
of the algorithm, which has access to a suboptimal dual solver only and still can ensure the (non-)optimality for the marked labels,
although the overall number of the marked labels may decrease. We propose an efficient implementation, which runs in time
comparable to a single run of a suboptimal dual solver. Our method is well-scalable and shows state-of-the-art results on
computational benchmarks from machine learning and computer vision.
Index Terms—Persistency, partial optimality, LP relaxation, discrete optimization, WCSP, graphical models, energy minimization
F
1
I NTRODUCTION
W
E consider the energy minimization or maximum a posteriori (MAP) inference problem for discrete graphical models.
In the most common pairwise case it has the form
X
X
fuv (xu , xv ) , (1)
fv (xv ) +
min Ef (x) := f∅ +
x∈X
v∈V
uv∈E
where minimization is performed over vectors x, containing the
discrete-valued components xv . Further notation is to be detailed
in § 2. The problem has numerous applications in computer vision, machine learning, communication theory, signal processing,
information retrieval and statistical physics, see [18, 47, 29] for
an overview of applications. Even in the binary case, when each
coordinate of x can be assigned two values only, the problem is
known to be NP-hard and is also hard to approximate [27].
Hardness of the problem justifies a number of existing approximate methods addressing it [18]. Among them, solvers addressing
its linear programming (LP) relaxations and in particular, the LP
dual [41, 49, 20], count among the most versatile and efficient
ones. However, apart from some notable exceptions (see the
overview of related work below), approximate methods can not
guarantee neither optimality of their solutions as a whole, nor even
optimality of any individual solution coordinates. That is, if x is a
solution returned by an approximate method and x∗ is an optimal
one, there is no guarantee that x∗v = xv for any coordinate v .
In contrast, our method provides such guarantees for some
coordinates. More precisely, for each component xv it eliminates
•
•
•
Alexander Shekhovtsov is with the Institute for Computer Graphics and
Vision (ICG), Graz University of Technology, Inffeldgasse 16, Graz 8010,
Austria.
E-mail: [email protected].
Paul Swoboda is with the Discrete Optimization Group, Institute of Science
and Technology Austria, Am Campus 1, 3400 Klosterneuburg, Austria.
E-mail: [email protected].
Bogdan Savchynskyy is with Computer Vision Lab, Faculty of Computer
Science, Institute for Artificial Intelligence, Dresden University of Technology, 01062 Dresden, Germany.
E-mail: [email protected].
those of its values (henceforth called labels), which provably can
not belong to any optimal solution. We call these eliminated
labels persistent non-optimal. Should a single label a remain
non-eliminated, it implies that for all optimal solutions x∗ it
holds x∗v = a and the label a is called persistent optimal. Our
elimination method is polynomial and is applicable with any
(approximate) solver for a dual of a linear programming relaxation
of the problem, employed as a subroutine.
1.1 Related Work
A trivial but essential observation is that any method identifying
persistency has to be based on tractable sufficient conditions in
order to avoid solving the NP-hard problem (1).
Dead-end elimination methods (DEE) [10] verify local
sufficient conditions by inspecting a given node and its immediate
neighbors at a time. When a label in the node can be substituted
with another one such that the energy for all configurations of the
neighbors does not increase, this label can be eliminated without
loss of optimality.
A similar principle for eliminating interchangeable labels was
proposed in constraint programming [11]. It’s generalization to
a related problem of Weighted Constraint Satisfaction (WCS) is
known as dominance rules or soft neighborhood substitutability.
However, because the WCS in general considers a bounded “+”
operation, the condition appears to be intractable and therefore
weaker sufficient local conditions were introduced, e.g., [26, 9].
The way [9] selects a local substitute label using equivalence
preserving transforms is related to our method, in which we use
an approximate solution based on the dual of the LP relaxation as
a tentative substitute (or test) labeling.
Although the local character of the DEE methods allows for an
efficient implementation, it also significantly limits their quality,
i.e., the number of found persistencies. As shown in [35, 43, 48],
considering more global criteria may significantly increase the
algorithm’s quality.
The roof dual relaxation in quadratic pseudo-Boolean Optimization (QPBO) [5, 30] (equivalent to pairwise energy minimiza-
2
LP - based
Truncated Model
wuv min(2, |xu − xv |)
Potts Model
λ min(1, |xu − xv |)
Graph cut - based
Courtesy of Kovtun [24]
93.6% (instance not available)
Instance used by Alahari et al. [4]
Kovtun’s method: 1s, 87.6%
Shekhovtsov [35]
LP-windowing: 1.5h, 94%
Ours: 22s, 96.7%
Kovtun’s method: 1s, 0.2%
Kohli et al. [19] (MQPBO)
41s, 0.2%
Swoboda et al. [44]
27min, 89.8%
Ours: 16s, 99.94%
Fig. 1: Progress of partial optimality methods. The top row corresponds to a stereo model with Potts interactions and large aggregating
windows for unary costs used in [24, 4] (instance published by [4]). The bottom row is a more refined stereo model with truncated linear
terms [45] (instance from [1]). The hashed red area indicates that the optimal persistent label in the pixel is not found (but some non-optimal
labels might have been eliminated). Solution completeness is given by the percentage of persistent labels. Graph cut based methods are fast
but only efficient for strong unary terms. LP-based methods are able to determine larger persistent assignments but are extremely slow prior to
this work.
tion with binary variables) has the property that all variables that
are integer in the relaxed solution are persistent. Several generalizations of roof duality to higher-order energies were proposed
(e.g., [2, 21]). The MQPBO method [19] and the generalized
roof duality [50] extend roof duality to the multi-label case by
reducing the problem to binary variables and generalizing the
concept of submodular relaxation [21], respectively. Although
for binary pairwise energies these methods provide a very good
trade-off between computational efficiency and a number of found
persistencies, their efficacy drops as the number of label grows.
Auxiliary submodular problems were proposed in [24, 25]
as a sufficient persistency condition for multilabel energy minimization. In the case of Potts model, the method has a very
efficient specialized algorithm [14]. Although these methods have
shown very good efficacy for certain problem classes appearing in
computer vision, the number of persistencies they find drastically
decreases when the energy does not have strong unary terms (see
Fig. 1).
In contrast to the above methods that technically rely either
on local conditions or on computing a maximum flow (min-cut),
the works [42, 43, 44] and [35] proposed persistency approaches
relying on a general linear programming relaxation. Authors
of [42, 43, 44] demonstrated applicability of their approach to
large-scale problems by utilizing existing efficient approximate
MAP-inference algorithms, while in [35] the large-scale problems
are addressed using a windowing technique. Despite the superior
persistency results, the running time of the approximate-LP-based
methods remained prohibitively slow for practical applications as
illustrated by an example in Fig. 1.
Not only LP-based methods can achieve superior results in
practice, but they are even theoretically guaranteed to do so, as
proven for the method [35, 37]. In this method, the problem of
determining the maximum number of persistencies is formulated
as a polynomially solvable linear program. It is guaranteed to find
a provably larger persistency assignment than most of the above
mentioned approaches. However, solving this linear program for
large scale instances is numerically unstable/intractable and applying it to multiple local windows is prohibitively slow. This poses a
challenge of designing an LP-based method that would be indeed
practical.
1.2
Contribution
In this work we propose a method which solves the same maximum
persistency problem as in [35] and therefore delivers provably
better results than other methods. Similar to [44], our method
requires to iteratively (approximately) solve the linear programming relaxation of (1) as a subroutine. However, our method is
significantly faster than [35, 44] due to a substantial theoretical
and algorithmic elaboration of this subroutine.
We demonstrate the efficiency of our approach on benchmark
problems from machine learning and computer vision. We outperform all competing methods in terms of the number of persistent
labels and method [35] in speed and scalability. On randomly
generated small problems, we show that the set of persistent labels
found using approximate LP solver is close to the maximal one as
established by the (costly and not scalable) method [35].
The present paper is a revised version of [39]. Besides reworked explanations, shortened and clarified proofs, one new technical extension is a more general dual algorithm, with termination
guarantees for a larger class of approximate solvers.
2
W ORK OVERVIEW
This section serves as an overview of our method, where we give
the most general definitions, formulate the maximum persistency
problem and briefly describe a generic method to solve it. This
description, equipped with references to subsequent sections,
should serve as a road map for the rest of the paper.
3
autarky property [5]. When applied to the whole search space X
we obtain its image p(X ) – a potentially smaller search space
containing all optimal labelings.
In what follows we will restrict ourselves to node-wise substitutions, i.e., those defined locally for each node: p(x)u = pu (xu ),
where pu : Xu → Xu . Indeed, already this class of substitutions
covers most existing persistency methods.
xu
1
2
N (u)
v0
3
u
Fig. 2: Dead end elimination / dominance. Variables are shown as
boxes and their possible labels as circles. Label xu = 1 is substituted
with label xu = 3. If for any configuration of neighbors xN (u) the
energy does not increase (only the terms inside {u}∪N (u) contribute
to the difference), label xu = 1 can be eliminated without loss of
optimality.
p u : X u ! Xu
pv
A general substitution we consider is applied to labels in all
nodes simultaneously, as illustrated in Fig. 3 for two variables. We
obtain the following principle for identifying persistencies.
1
2
Proposition 2.2. If p is a strictly improving substitution, then any
optimal solution x∗ of (1) must satisfy (∀v ∈ V) pv (x∗v ) = x∗v .
3
Fig. 3: Simultaneous substitution of labels in two variables. Xu =
Xv = {1, 2, 3}, labels at arrow tails are substituted with labels at
arrow heads. So the joint configuration (1, 2) (dashed) is substituted
with the configuration (3, 3) (solid).
2.1
Notation
In the MAP-Inference Problem (1) we assume (V, E) to be
a directed graph with the set of nodes V and the set of
edges E ⊆ V × V . Let uv denote an ordered pair (u, v) and
N (u) = {v | uv ∈ E ∨ vu ∈ E} stands for the set of neighbors
of u. Each node v ∈ V is associated with a variable xv taking
its values in a finite set of labels Xv . Cost functions or potentials
fv : Xv → R, fuv : Xu × Xv → R are associated with nodes
and edges respectively. Let f∅ ∈ R be a constant term, which
we introduce for the
Q sake of notation. Finally, X stands for the
Cartesian product v∈V Xv and its elements x ∈ X are called
labelings.
We represent all potentials of energy (1) by a single cost vector
f ∈ RI , where the set I enumerates all components of all terms:
I = {∅} ∪ {(u, i) | u ∈ V, i ∈ Xu } ∪ {(uv, ij) | uv ∈ E, i ∈
Xu , j ∈ Xv }.
2.2
Improving Substitutions
We formulate our persistency method in the framework of
(strictly) improving substitutions, called improving mappings in
our previous works [35, 37]. It was shown in [35] that most
existing persistency techniques can be expressed as improving
substitutions. A mapping p : X → X is called a substitution,
if it is idempotent, i.e., p(x) = p(p(x)).
Definition 2.1. A substitution p : X → X is called strictly
improving for the cost vector f if
(∀x | p(x) 6= x) Ef (p(x)) < Ef (x).
Example 2.1. Let us consider the dead-end elimination
(DEE) [10, 13]. It is a test whether a given label in a single node,
e.g., xu = 1 in Fig. 2 can be substituted with another one, e.g.,
xu = 3 in Fig. 2. The change of the energy under this substitution
depends only on the configuration of neighbors xN (u) , and the
value of the change is additive in neighbors, so that it can be
verified for all xN (u) whether the substitution always improves
the energy. If it is so, the label xu = 1 can be eliminated and the
test is repeated for a different label in the reduced problem.
(2)
When a strictly improving substitution is applied to any
labeling x, it is guaranteed that p(x) has equal or better energy.
In particular, strictly improving substitutions generalize the strong
Indeed, otherwise Ef (p(x∗ )) < Ef (x∗ ), which is a contradiction. If pv (i) 6= i, then idempotency implies that label (v, i) is
non-optimal persistent and can be excluded from consideration.
2.3 Verification Problem
Verifying whether a given substitution is strictly improving is an
NP hard decision problem [35]. In order to obtain a polynomial
sufficient condition we will first rewrite (2) as an energy minimization problem and then relax it. To this end we reformulate
Definition 2.1 in an optimization form:
Proposition 2.3. Substitution p is strictly improving iff
min Ef (x) − Ef (p(x)) ≥ 0,
x∈X
p(x) = x for all minimizers.
(3a)
(3b)
Proof. Indeed, condition (3a) is equivalent to (∀x) Ef (x) ≥
Ef (p(x)). Sufficiency: if x 6= p(x), then x is not a minimizer
and Ef (x) > Ef (p(x)). Necessity: for x = p(x) we have that
Ef (x) = Ef (p(x)), therefore from Definition 2.1 it follows that
condition (3a) holds and any x = p(x) is a minimizer, moreover,
for any minimizer x it must be Ef (x) − Ef (p(x)) = 0 and from
Definition 2.1 it follows x = p(x).
In § 3 we will show that the difference of the energies in (3a)
can be represented as a pairwise energy with an appropriately
constructed cost vector g so that there holds
Ef (x) − Ef (p(x)) = Eg (x).
(4)
Therefore, according to Proposition 2.3 the verification of the
strictly improving property reduces to minimizing the energy (4)
and checking that (3b) is fulfilled. To make the verification
problem tractable, we relax it as
min Eg (µ) ≥ 0,
(5a)
p(µ) = µ for all minimizers,
(5b)
µ∈Λ
where Λ is a tractable polytope such that its integer vertices
correspond to labelings (the standard LP relaxation that we use
4
will be defined in § 5), µ is a relaxed labeling and Eg (µ) and
p(µ) are appropriately defined extensions of discrete functions
Eg (x) and p(x), defined in § 3. By construction, the objective
value (5) matches exactly to that of (2) for all integer labelings
µ ∈ Λ ∩ {0, 1}I , which is sufficient for (2) to hold. The sufficient
condition (5) (made precise in Definition 3.2) means that p strictly
improves not only integer labelings but also all relaxed labelings
and therefore such substitutions will be called strictly relaxedimproving for a cost vector f . Assuming Λ is fixed by the context,
let Sf denotes the set of all substitutions p satisfying (5).
(a)
p∈P
v∈V
where P is a class of substitutions.
While maximizing over all substitutions is not tractable, maximizing over the following restricted class is [35]. Assume we are
given a test labeling y , which in our case will be an approximate
solution of the MAP inference (1). We then consider substituting
in each node v a subset of labels with yv .
Definition 2.4 ([35]). A substitution p is in the class of subsetto-one substitutions P 2,y , where y ∈ X , if there exist subsets
Yv ⊆ Xv \{yv } for all v such that
yv , if i ∈ Yv ;
pv (i) =
(7)
i,
if i ∈
/ Yv .
See Figs. 3 and 4 for examples. Note, this class is rather large:
there are 2|Xv |−1 possible choices for pv and most of the existing
methods for partial optimality still can be represented using it [35]
(in particular, methods [25, 14] can be represented using a constant
test labeling, yv = α for all v ).
The restriction to the class P 2,y allows to represent the search
of the substitution that eliminates the maximum number of labels
as the one with the largest (by inclusion) sets Yv of substituted
labels. This allows to propose a relatively simple algorithm.
2.5
Cutting Plane Algorithm
The algorithm is a cutting plane method in a general sense: we
maintain a substitution pt which is in all iterations better or
equal than the solution to (6) and achieve feasibility by iteratively
constraining it.
Initialization: Define the substitution p0 by the sets Yv0 =
Xv \{yv }. It substitutes everything with y and clearly maximizes the objective (6).
Verification: Check whether current pt is strictly relaxedimproving for f by solving relaxed problem (5). If yes,
return pt . If not, the optimal relaxed solution µ∗ corresponds
to the most violated constraint.
Cutting plane: Assign pt+1 to the substitution defined by the
largest sets Yvt+1 such that Yvt+1 ⊆ Yvt and the constraints
Eg (µ∗ ) ≥ 0, p(µ∗ ) = µ∗ , are satisfied. Repeat the verification step.
The steps of this meta-algorithm are illustrated in Fig. 4. It is
clear that when the algorithm stops the substitution pt is strictly
improving, although it could be the identity map that does not
0.5
0.5
(b)
2.4 Maximum Persistency and Subset-to-One Substitutions
The maximum persistency approach [35] consists of finding a
relaxed-improving substitution p ∈ Sf that eliminates the maximal
number of labels:
X
max
{i ∈ Xv | p(i) 6= i} , s.t. p ∈ Sf ,
(6)
y
µ⇤
1
y
(c)
y
Fig. 4: Steps of the discrete cutting-plane algorithm. (a) Starting from
substitution that maps everything to the test labeling y (red), crossed
labels would be eliminated if p passes the sufficient condition. (b) A
relaxed solution µ∗ violating the sufficient condition is found (black).
(c) Substitution p is pruned.
eliminate any labels. The exact specification of the cutting plane
step will be derived in § 4 and it will be shown that this algorithm
solves the maximum persistency problem (6) over P 2,y optimally.
2.6
Work Outline
In § 3 we give a precise formulation of the relaxed condition (5)
and its components. In § 4 we specify details of the algorithm and
prove its optimality. These results hold for a general relaxation
Λ ⊇ M but require to solve linear programs (5a) precisely.
The rest of the paper is devoted to an approximate solution
of the problem (6), i.e. finding a relaxed improving mapping,
which is almost maximum. We consider specifically the standard
LP relaxation and reformulate the algorithm to use a dual solver
for the problem (5a), § 5. We then gradually relax requirements on
the optimality of the dual solver while keeping persistency guarantees, §§ 5 to 7, and propose several theoretical and algorithmic
tools to solve the series of verification problems incrementally and
overall efficiently, § 8. Finally, we provide an exhaustive experimental evaluation in § 9, which clearly demonstrates efficacy of
the developed method.
3
3.1
R ELAXED -I MPROVING S UBSTITUTIONS
Overcomplete Representation
In this section we formally derive the strictly relaxed-improving
sufficient condition (5). To obtain the relaxation we use the standard lifting approach (a.k.a. overcomplete representation [47]), in
which a labeling is represented using the 1-hot encoding. This
lifting allows to linearize the energy function, the substitution and
consequently both the non-relaxed (3) and relaxed (5) improving
substitution criteria.
The lifting is defined by the mapping δ : X → RI :
δ(x)∅ = 1,
δ(x)u (i) = [[xu =i]],
δ(x)uv (i, j) = [[xu =i]][[xv =j]],
(8a)
(8b)
(8c)
where [[·]] is the Iverson bracket, i.e., [[A]] equals 1 if A is
true and 0 otherwise. Using this lifting, we can linearize unary
5
P
P
terms as fv (xv ) = k fv (k)[[xv =k]] = k fv (k)δ(x)v (k) and
similarly for the pairwise terms. This allows to linearize the energy
function Ef and write it as a scalar product Ef (x) = hf, δ(x)i
in RI . The energy minimization problem (1) can then be written
as
min Ef (x) = min hf, δ(x)i = min hf, µi ,
(9)
x∈X
µ ∈M
x∈X
where M = conv δ(X ) is the convex hull of all labelings in the
lifted space, also known as marginal polytope [47]. The last equality in (9) uses the fact that the minimum of a linear function on a
finite set equals the minimum on its convex hull. Expression (9) is
an equivalent reformulation of the energy minimization problem
as a linear program, however over a generally intractable polytope
M.
i.e. all relaxed labels are mapped to the indicator of the label
xu = 3. And the adjoint operator P T acts as follows (cf. (10b)):
Pu> ( fu (1) fu (2) fu (3) )> = ( fu (3) fu (3) fu (3) )> ,
(15)
Pv> ( fu (1) fu (2) fu (3) )>
(16)
=
( fu (1) fu (3) fu (3) )>
>
Similarly, due to (10c) we have, e.g., (Puv
fuv )(1, 2)
>
(Puv fuv )(1, 1) = fuv (3, 1) and so on.
3.3
.
= fuv (3, 3),
Strictly improving substitutions
Let I denote the identity mapping RI → RI . Using Proposition 2.3 and the linear extension [p], we obtain that substitution p
is strictly improving iff the value of
min hf, δ(x) − δ(p(x))i = min hf, (I − [p])δ(x)i
x∈X
3.2
Lifting of Substitutions
Next we show how a substitution p : X → X can be represented
as a linear map in the lifted space RI . This will allow to express
the term Ef (p(x)) as a linear function of δ(x) and hence also to
represent the non-relaxed criterion (3).
Proposition 3.1. Given a substitution p, let P T : RI → RI be
defined by its action on a cost vector f ∈ RI as follows:
(P T f )∅ = f∅ ,
(10a)
(P T f )u (i) = fu (pu (i)),
(10b)
(P T f )uv (i, j) = fuv (pu (i), pv (j))
(10c)
x∈X
(∀x ∈ X ) δ(p(x)) = P δ(x).
Definition 3.2 ([35]). Substitution p is strictly Λ-improving for
the cost vector f ∈ RI (shortly, strictly relaxed-improving, or
p ∈ Sf ) if
minµ∈Λ h(I − [p])> f, µi = 0,
[p]µ∗ = µ∗ for all minimizers.
(11)
Proof. Let x ∈ X . From (10) it follows that EP T f (x) =
Ef (p(x)), which can be expressed as a scalar product
hP T f, δ(x)i = hf, P δ(x)i = hf, δ(p(x))i.
Since this equality holds for all f ∈ RI it follows that P δ(x) =
δ(p(x)).
The expression (11) allows to write the energy of the substituted labeling p(x) as
Ef (p(x)) = hf, δ(p(x))i = hf, P δ(x)i = EP T f (x).
(12)
For this reason, the mapping P is called the linear extension of p
and will be denoted with the symbol [p]. The following example
illustrates how [p] looks in coordinates.
Example 3.1. Consider the substitution p depicted in Fig. 3 and
defined by pu : 1, 2, 3 7→ 3, 3, 3; pv : 1, 2, 3 7→ 1, 3, 3. The
relaxed labeling µ ∈ RI has the structure (µ∅ , µu , µv , µuv ). The
linear extension [p] : RI → RI can be written as a block-diagonal
matrix
!
1
0 0 0
1 0 0
Pu
, where Pu = 0 0 0 , Pv = 0 0 0
(13)
Pv
111
011
and Puv is defined by Puv µuv = Pu µuv Pv , where µuv is shaped
as a 3 × 3 matrix. The action of the block Pu expresses as
µ (1)
u
0
Pu µu (2) = 0 ,
(14)
µu (3)
1
µ∈M
is zero and [p]µ = µ for all minimizers. Note that problem (17)
is of the same form as the energy minimization (9) with the cost
vector g = (I − [p]T )f , as introduced in (4).
The sufficient condition for persistency (5) is obtained by
relaxing the intractable marginal polytope M in (17) to a tractable
outer approximation Λ ⊇ M.
∀ u ∈ V , uv ∈ E , ij ∈ Xuv . Then P satisfies
Puv
x∈X
= minh(I − [p])> f, δ(x)i = min h(I − [p])> f, µi (17)
(18a)
(18b)
In § 5, Λ will be defined as the polytope of the standard LP
relaxation but until then the arguments are general and require
only that Λ ⊇ M. Since Λ includes all integer labelings, it
is a sufficient condition for improving substitution and hence
persistency.
Corollary 3.3. If substitution p is strictly Λ-improving for f
and Λ ⊇ M, then p is strictly improving for f .
The problem (18a) will be called the verification LP and the
decision problem to test for p ∈ Sf , i.e. to verify conditions (18),
will be called the verification problem.
4
4.1
G ENERIC P ERSISTENCE A LGORITHM
Structure of P 2,y
In [35] it was shown that the maximum persistency problem (6)
over the class of substitutions P 2,y can be formulated as a
single linear program, where the substitution is represented using
auxiliary (continuous) variables. Here we take a different approach
based on observing a lattice-like structure of improving substitutions.
Throughout this section we will assume that the test labeling
y ∈ X is fixed. Let us compare two substitutions p and q by
the sets of the labels they eliminate. A substitution p ∈ P 2,y
eliminates all labels in Yv , or equivalently all labels not in pv (Xv ).
Definition 4.1. A substitution p ∈ P 2,y is better equal than a
substitution q ∈ P 2,y , denoted by p ≥ q , if (∀v ∈ V) it holds
pv (Xv ) ⊆ qv (Xv ).
6
Proposition 4.2. Let the partially ordered set (Sf ∩ P 2,y , ≥)
of subset-to-one strictly relaxed-improving substitutions has the
maximum and let it be denoted r. Then r is the unique solution
of (6) with P = P 2,y .
Proof. Since r is the maximum, it holds r ≥ q for all q . From
Definition
4.1 we P
have rv (Xv ) ⊆ qv (Xv ) for all v and thus
P
|r
(X
)|
≤
r is optimal
to (6).
v
v v
v |qv (Xv )|. Therefore
P
P
Additionally, if r 6= q , then it holds v |rv (Xv )| < v |qv (Xv )|
and therefore r is the unique solution to (6).
The existence of the maximum will formally follow from the
correctness proof of the algorithm, Theorem 4.6. A stronger claim,
which is not necessary for our analysis, but which may provide a
better insight is that (Sf ∩ P 2,y , ≥) is a lattice isomorphic to
the lattice of sets with union and intersection operations. This is
seen as follows. If both p and q are strictly-improving then so is
their composition r(x) = p(q(x)), as can be verified by chaining
inequalities (2). In P 2,y the composition satisfies the property
ru (Xu ) = pu (Xu ) ∩ qu (Xu ), and can be identified with the join
of p and q (the least r such that r ≥ p and r ≥ q ). This can be
shown to hold also for Sf ∩ P 2,y . It is this structure that allows to
find the maximum in Sf ∩ P 2,y by a relatively simple algorithm.
4.2
Generic Algorithm
Our generic primal algorithm, displayed in Algorithm 1, represents
a substitution p ∈ P 2,y by the sets Yv of labels to be substituted
with y , via (7). Line 1 initializes these sets to all labels but
y . Line 1 constructs the cost vector of the verification LP in
condition (18). Lines 1 and 1 solve the verification LP and test
whether sufficient conditions (18) are satisfied via the following
reformulation.
Proposition 4.3. For a given substitution p, let O∗ ⊆ Λ denote
the set of minimizers of the verification LP (18) and
Ov∗ := {i ∈ Xv | (∃µ ∈ O∗ ) µv (i) > 0},
(19)
which is the support set of all optimal solutions in node v . Then
p ∈ Sf iff (∀v ∈ V ∀i ∈ Ov∗ ) pv (i) = i. Proof in § A.
Corollary 4.4. For substitution p ∈ P 2,y defined by (7) it holds
p ∈ Sf iff (∀v ∈ V) Ov∗ ∩ Yv = ∅.
In the remainder of the paper we will relate the notation
Ov∗ to O∗ as in (19). The set of optimal solutions O∗ can, in
general, be a d-dimensional face of Λ. We need to determine
in (19) whether an optimal solution µ ∈ O∗ exists such that
the coordinate µv (i) is strictly positive. From the theory of linear
programming [46], it is known that if one takes an optimal solution
µ in the relative interior of O∗ (i.e., if O∗ is a 2D face the relative
interior excludes vertices and edges of O∗ ) then its support set
{i | µv (i) > 0} is the same for all such points and it matches
Ov∗ (19). Therefore, it is practically feasible to find the support sets
Ov∗ by using a single solution found by an interior point/barrier
method (which are known to converge to a “central” point of the
optimal face) or methods based on smoothing [31, 32]. Obtaining
an exact solution by these methods may become computationally
expensive as the size of the inference problem (1) grows. Despite
that, Algorithm 1 is implementable and defines the baseline for its
practically efficient variants solving (6) approximately. These are
developed further in the paper.
Since Line 1 of Algorithm 1 verifies precisely the condition of
Corollary 4.4, the algorithm terminates as soon as p ∈ Sf and
Algorithm 1: Iterative Pruning LP-Primal
1
2
3
4
5
6
7
8
Input: Cost vector f ∈ RI , test labeling y ∈ X ;
Output: Maximum strictly improving substitution p;
(∀v ∈ V) Yv := Xv \{yv };
while true
Construct the verification problem potentials
g := (I − [p])> f with p defined by (7);
O∗ = argminµ∈Λ hg, µi;
(∀v ∈ V) Ov∗ = {i ∈ Xv | (∃µ ∈ O∗ ) µv (i) > 0};
if (∀v ∈ V) Ov∗ ∩ Yv = ∅ then return p;
for v ∈ V do
Pruning of substitutions: Yv := Yv \Ov∗ ;
hence p is strictly improving. In the opposite case, Line 1 prunes
the sets Yv by removing labels corresponding to the support set
Ov∗ of all optimal solutions of the verification LP, which have
been identified now to violate the sufficient condition. These labels
may be a part of some optimal solution to (1) and will not be
eliminated.
To complete the analysis of Algorithm 1 it remains to answer
two questions: i) does it terminate and ii) is it optimal for the
maximum persistency problem (6)?
Proposition 4.5. Algorithm 1 runs in polynomial time and returns
a substitution p ∈ Sf ∩ P 2,y .
Proof. As we discussed above, sets Ov∗ in Line 1 can be found
in polynomial time. At every iteration, if the algorithm has
not terminated yet, at least one of the sets Yv strictly shrinks,
as can be seen by comparing termination condition in Line 1
with pruning
P in Line 1. Therefore the algorithm terminates in
at most
v (|Xv | − 1) iterations. On termination, p ∈ Sf by
Corollary 4.4.
Theorem 4.6. Substitution p returned by Algorithm 1 is the
maximum of Sf ∩ P 2,y and thus it solves (6). Proof in ?? .
It is noteworthy that Algorithm 1 can be used to solve problem (6) with any polytope Λ satisfying M ⊆ Λ, i.e., with
any LP relaxation of (1) that can be expressed in the lifted
space RI . Moreover, in order to use the algorithm with higher
order models one needs merely to (straightforwardly) generalize
the linear extension (10) as done in [37].
The test labeling y can itself be chosen using the approximate
solution of the LP-relaxation, e.g., via the zeroth iteration of the
algorithm with g = f and picking yv from Ov∗ . This choice is
motivated by the fact that a strict relaxed-improving substitution
cannot eliminate the labels from the support set of optimal solutions of the LP relaxation [35], and thus these labels may not be
substituted with anything else.
4.3
Comparison to Previous Work
Substitutions in P 2,y are related to the expansion move algorithm [6] in the following sense. While [6] seeks to improve a
single current labeling x by calculating an optimized crossover
(fusion) with a candidate labeling y , we seek which labels can
be moved with a guaranteed improvement to y for all possible
labelings x.
Algorithm 1 is similar in structure to [43]. The later finds
an improving substitution in a small class P 1,y by incrementally shrinking the set of potentially persistent variables. More
7
specifically, given a test labeling y ∈ X , the all-to-one class of
substitutions P 1,y contains substitutions p, which in every node
v either replace all labels with yv or leaves all labels unchanged.
There are only two possible choices for pv : either i 7→ yv for
all i ∈ Xv or the identity i 7→ i. Methods [24, 42, 43] can be
explained as finding an improving mapping in this class [35]. We
generalize the method [43] to substitutions in P 2,y . The original
sufficient condition for persistency in [43] does not extend to
such substitutions. Even for substitutions in P 1,y it is generally
weaker than condition (18) unless a special reparametrization is
applied [44]. Criterion (18) extends to general substitutions and
does not depend on reparametrization. Similarly to [43], we will
use approximate dual solvers in this more general setting.
In [35, (ε-L1)] problem (6) is formulated as one big linear program. We solve problem (6), and hence also the LP problem [35]
in a more combinatorial fashion w.r.t. to the variables defining the
substitution.
It may seem that solving a series of linear programs rather
than a single one is a disadvantage of the proposed approach.
However, as we show further, the proposed iterative algorithm
can be implemented using a dual, possibly suboptimal, solver for
the relaxed verification problem (18). This turns out to be much
more beneficitial in practice since the verification problems can be
incrementally updated from iteration to iteration and solved overall
very efficiently. This approach achieves scalability by exploiting
available specialized approximate solvers for the relaxed MAP
inference. Essentially, any dual (approximate) solver can be used
as a black box in our method.
the mode 3 relies on the persistency problem reduction introduced
in § 6. Intermediate steps 1 and 2 are considered right after
defining the standard LP relaxation and its dual.
5.1 LP Relaxation
We consider the standard local polytope relaxation [41, 49, 47]
of the energy minimization problem (1) given by the following
primal-dual pair:
(primal)
minhf,
µi
P
µ
(i, j) = µu (i),
uv
j
P
µ
(i,
j) = µv (j),
uv
i
P
i µu (i) = µ∅ ,
µu (i) ≥ 0,
µuv (i, j) ≥ 0,
µ∅ = 1
=
(dual)
max f∅ϕ
ϕuv (i) ∈ R,
ϕvu (j) ∈ R,
ϕ u ∈ R,
fuϕ (i) ≥ 0,
ϕ
fuv
(i, j) ≥ 0,
where f ϕ abbreviates
fuϕ (i) = fu (i) +
P
v∈N (u) ϕuv (i) − ϕu ,
ϕ
fuv (i, j) = fuv (i, j) − ϕuv (i) − ϕvu (j),
P
f∅ϕ = f∅ + u ϕu .
P ERSISTENCY WITH D UAL S OLVERS
Though Algorithm 1 is quite general, its practical use is limited
by the strict requirements on the solver, which must be able to
determine the exact support set of all optimal solutions. However,
finding even a single solution of the relaxed problem with standard
methods like simplex or interior point can be practically infeasible
and one has to switch to specialized solvers developed for this
problem. Although there are scalable algorithms based on smoothing techniques [31, 32], which converge to an optimal solution,
waiting until convergence in each iteration of Algorithm 1 can
make the whole procedure impractical. In general, we would like
to avoid restricting ourselves to certain selected solvers to be
able to choose the most efficient one for a given problem. In the
standard LP relaxation (introduced below in (20)), the number of
primal variables grows quadratically with the number of labels,
while the number of dual variables grows only linearly. It is
therefore desirable to use solvers working in the dual domain,
including suboptimal ones, (e.g. [20, 32, 12, 33], performing
block-coordinate descent) as they offer the most performance for
a limited time budget. Furthermore, fast parallel versions of such
methods have been developed to run on GPU/FPGA [38, 7, 16],
making the LP approach feasible for more vision applications.
We will switch to the dual verification LP and gradually relax
our requirements on the solution returned by a dual solver. This is
done in the following steps:
1. an optimal dual solution;
2. an arc consistent dual point;
3. any dual point.
Our main objective is to ensure in each of these cases that the
found substitution p is strictly improving, while possibly compromising its maximality. The final practical algorithm operating in
(21a)
(21b)
(21c)
The constraints of the primal problem (20) define the local
polytope Λ. The cost vector f ϕ is called a reparametrization of
f . There holds cost equivalence: hf ϕ , µi = hf, µi for all µ ∈ Λ
(as well as Ef = Ef ϕ ), see [49]. Using the reparametrization, the
dual problem (20) can be briefly expressed as
max f∅ϕ s.t. (∀ω ∈ V ∪ E) fωϕ ≥ 0.
5
(20)
ϕ
(22)
ϕ
Note that for a feasible ϕ the value f∅ is a lower bound on the
primal problem (20). In what follows we will assume that ϕ in (22)
additionally satisfies the following normalization: mini fuϕ (i) =
ϕ
(i, j) = 0 for all u, v , which is automatically
0 and minij fuv
satisfied for any optimal solution.
5.2 Expressing Ov∗ in the Dual Domain
Let (µ, ϕ) be a pair of primal and dual optimal solutions to (20).
From complementary slackness we know that if µv (i) > 0 then
the respective dual constraint holds with equality:
µv (i) > 0 ⇒ fvϕ (i) = 0,
(23)
in this case we say that fvϕ (i) is active. The set of such active
dual constraints matches the sets of local minimizers of the
reparametrized problem,
Ov (ϕ) := i ∈ Xv | fvϕ (i) = 0 = argmin fvϕ (i).
(24)
i
From complementary slackness (23) we obtain that
Ov∗ ⊆ Ov (ϕ).
(25)
This inclusion is insufficient for an exact reformulation of Algorithm 1, however it is sufficient for correctness if we make sure
that Yv ∩ Ov (ϕ) = ∅ on termination, i.e., that the substitution p
does not displace labels in Ov (ϕ). Then, by Corollary 4.4, p ∈ Sf
follows.
There always exists an optimal primal solution µ and dual ϕ
satisfying strict complementarity [46], in which case relation (23)
becomes an equivalence:
µv (i) > 0 ⇔ fvϕ (i) = 0.
(26)
8
Algorithm 2: Iterative Pruning Arc Consistency
Xu \Yu
I
1
2
3
4
5
6
7
8
Input: Cost vector f ∈ R , test labeling y ∈ X ;
Output: Strictly improving substitution p;
(∀v ∈ V) Yv := Xv \{yv };
while true
Construct verification problem g := (I − [p])> f with p
defined by (7);
Use dual solver to find ϕ such that g ϕ is arc consistent;
Ov (ϕ) := {i ∈ Xv | gvϕ (i) = 0};
if (∀v ∈ V) Ov (ϕ) ∩ Yv = ∅ then return p;
for v ∈ V do
Pruning of substitutions: Yv := Yv \Ov (ϕ);
It is the case when both µ and ϕ are relative interior points of
the optimal primal, resp. optimal dual, faces [46]. For a relative
interior optimal ϕ, the set of constraints that are satisfied as
equalities f ϕ (i) = 0 is the smallest and does not depend on
the specific choice of such ϕ. Under strict complementarity, (25)
turns into equality Ov∗ = Ov (ϕ), which allows to compute
the exact maximum persistency using a dual algorithm without
reconstructing a primal solution. However finding such ϕ appears
more difficult: e.g. the most efficient dual block-coordinate ascent solvers [20, 12, 8, 33] only have convergence guarantees
(see [20, 33]) allowing to find a sub-optimal solution, satisfying
arc consistency.
Definition 5.1 ([49]). A reparametrized problem f ϕ is called arc
ϕ
consistent if: (i) for all uv ∈ E from fuv
(i, j) being active follows
that fuϕ (i) and fvϕ (j) are active; (ii) for all u ∈ V from fuϕ (i)
active follows that for all v ∈ N (u) there exists a j ∈ Xv such
ϕ
that fuv
(i, j) is active.
An optimal dual solution need not be arc consistent, but
it can be reparametrized without loss of optimality to enforce
arc consistency [49]. Labels that become inactive during this
procedure are not in the support set of primal solutions. In general
the following holds.
Proposition 5.2. Arc consistency is a necessary condition for
relative interior optimality: if Ov (ϕ) = Ov∗ for all v ∈ V then
f ϕ is arc consistent. Proof in (38).
This property is in our favor, since we are ideally interested
in the equality Ov (ϕ) = Ov∗ . Next, we propose an algorithm
utilizing an arc consistent solver and prove that it is guaranteed to
output p ∈ Sf .
5.3
Persistency with an Arc Consistency Solver
We propose Algorithm 2 which is based on a dual solver attaining
the arc consistency condition (differences to Algorithm 1 underlined). If the dual solver (in line 4) finds a relative interior optimal
solution, Algorithm 2 solves (6) exactly. Otherwise it is suboptimal
and we need to reestablish its correctness and termination.
Lemma 5.3 (Termination
of Algorithm 2). Algorithm 2 termiP
nates in at most v (|Xv | − 1) iterations.
Proof. In case the return condition in line 6 is not satisfied, Ov ∩
Yv 6= ∅ for some v and the pruning in line 8 excludes at least one
label from Yv .
Xu \Yu
j
j
j
Xv \Yv
i
Xv \Yv
∆uv (i )
i
∆vu (j)
j
i
v
u
∆uv (i) = minj ∈Xu \Yu guv (i, j )
(a)
u
v
min{guv (i, j ), ∆uv (i) + ∆vu (j )}
(b)
Fig. 5: Illustration for the reduction. Labels Xu \Yu are not displaced
by p hence their associated unary and pairwise costs are zero in g =
(I − [p]T )f . In case (a) the indicated pairwise costs are replaced with
their minimum. In case (b) the value of guv (i, j) can be decreased,
assuming all reductions of type (a) and their symmetric counterparts
are already performed. The amount of decrease matches the value
of the mixed derivative (non-submodularity) associated to i, i0 paired
with j, j 0 .
Lemma 5.4 (Correctness of Algorithm 2). If (∀v ∈ V) Ov (ϕ) ∩
Yv = ∅ holds for an arc consistent dual vector ϕ, then ϕ is
optimal. Proof in (57).
It follows that when Algorithm 2 terminates, the found arc
consistent solution ϕ is optimal, in which case inclusion (25) is
satisfied and the found substitution p is guaranteed to be in Sf .
5.4
Solvers Converging to Arc Consistency
One can see that arc consistency is only required on termination of
Algorithm 2. In the intermediate iterations we may as well perform
the pruning step, line 8, without waiting for the solver to converge.
This motivates the following practical strategy:
• Perform a number of iteration towards finding an arcconsistent dual point ϕ;
• Check whether there are some labels to prune, i.e.,
(∃u)Ou (ϕ) ∩ Yu 6= ∅;
• Terminate if ϕ is arc consistent and there is nothing to
prune; otherwise, perform more iterations towards arc consistency.
If the solver is guaranteed to eventually find an arc consistent
solution, the overall algorithm will either terminate with an arc
consistent and (by Lemma 5.4) optimal ϕ or there will be some
labels to prune. However, we have to face the question what
happens if the dual solver does not find an arc consistent solution
in finite time. In this case the algorithm can be iterating infinitely
with no pruning available. At the same time there is no guarantee
that a pruning step will not occur at some point and thus if we
simply terminate the algorithm we get no persistency guarantees.
Even if the dual solver was guaranteed to converge in a finite
number of iterations, it is in principle possible that the time needed
for a pruning to succeed would be proportional to the time of
convergence, making the whole algorithm very slow. Instead, it is
desirable to guarantee a valid result while allowing only a fixed
time budget for the dual solver. We will overcome this difficulty
with the help of the reduced verification LP presented next.
6
V ERIFICATION P ROBLEM R EDUCTION
Algorithms 1 and 2 iteratively solve verification problems. We can
replace the verification LP solved in step 4 by a simpler, reduced
one, without loss of optimality of the algorithms.
9
Definition 6.1. Let g := (I − [p])> f be the cost vector of the
verification LP. The reduced cost vector ḡ is defined as
ḡv (i) := gv (i), v ∈ V; ḡ∅ = 0;
g uv (i, j) :=
0,
0
∆vu (j) := mini0 ∈Y
/ u guv (i , j),
0
/ v guv (i, j ),
∆uv (i) := minj 0 ∈Y
min{∆vu (j) + ∆uv (i), guv (i, j)},
(27a)
(27b)
i∈
/ Yu , j
i∈
/ Yu , j
i ∈ Yu , j
i ∈ Yu , j
∈
/ Yv ,
∈ Yv ,
∈
/ Yv ,
∈ Yv .
The reduction is illustrated in Fig. 5. Taking into account that
guv (i0 , j 0 ) = 0 for i0 ∈ Xu \Yu , j 0 ∈ Xv \Yv , the reduction can
be interpreted as forcing the inequality
guv (i, j 0 ) + guv (i0 , j) − guv (i, j) − guv (i0 , j 0 ) ≥ 0,
(28)
i.e., the non-negativity of mixed discrete derivatives, for all fourtuples i ∈ Yu , j ∈ Yv , i0 ∈
/ Yv , j 0 ∈
/ Yu . The cost vector ḡ is
therefore a partial submodular truncation of g .
Recall that Algorithm 1 on each iterations prunes all substitutions q ≤ p that do not belong to Sf based on the solutions of the
verification LP. The following theorem reestablishes optimality of
this step with the above reduction.
Theorem 6.2 (Reduction). Let p ∈ P 2,y and ḡ be the corresponding reduced cost vector constructed as in Def. 6.1. Let also
q ∈ P 2,y , q ≤ p. Then q ∈ Sf iff q ∈ Sḡ . Proof in (55).
Procedure 3: dual correct(ϕ,ḡ )
1
2
3
4
5
6
for uv ∈ E do
ϕ
(∀i ∈ Xu ) ϕuv (i) := ϕuv (i) + minij ḡuv
(i, j);
ϕ
(∀i ∈ Xu ) ϕuv (i) := ϕuv (i) + minj ḡuv
(i, j);
ϕ
(∀j ∈ Xv ) ϕvu (j) := ϕvu (j) + mini ḡuv
(i, j);
(∀u ∈ V) ϕu := ϕu + mini guϕ (i);
return ϕ;
/* Normalize */
terminate (lines 11-13). If neither occurs in a certain number of
iterations (stopping condition in line 14), the pruning based on
the currently active labels is executed (line 15). After that the cost
vector ḡ is rebuilt, but the dual solver continues from the last found
dual point (warm start).
The speed-ups will be explained in the next section, they
are not critical for the overall correctness. Now we focus on
the new termination conditions (lines 11-13). A correction step
(line 11) is introduced whose purpose is to move the slacks from
pairwise terms to unary terms so that active labels become more
decisive. This procedure is defined in Procedure 3. The correction
is not intermixed with dual updates but serves as a proxy between
the solver and the termination conditions. It has the following
property.
Lemma 7.1. Output ϕ of Procedure 3 is feasible and satisfies
(∀u ∈ V) min guϕ (i) = 0, (30)
i∈Xu
From Theorem 6.2 and Corollary 4.4 it follows that q ∈ Sf iff
ϕ
ϕ
0
qu (Ou∗ ) = Ou∗ , where Ou∗ are the support sets of optimal solutions (∀uv ∈ E, ij ∈ Xuv ) min ḡuv
ḡuv
(i, j 0 ) = 0. (31)
(i , j) = min
j 0 ∈Xv
i0 ∈Xu
to the reduced verification LP,
ϕ
Moreover, if the input ϕ is feasible, the lower bound f∅ does not
argminµ∈Λ hḡ, µi.
(29)
decrease.
Therefore it is valid for algorithms 1 and 2 to consider this
Proof. Line 2 of Procedure 3 moves a constant from an edge
reduced LP and prune all substitutions q that do not satisfy
ϕ
(i, j) to zero.
to node. This turns the minimum of terms guv
∗
∗
the property qv (Ov ) = Ov . The optimal relaxed solutions and
Lines 3 and 4 turn to zero the minimal pairwise value attached
their support sets can in general differ from those of the original
to each label, which provides (31). Line 5 provides (30). In case
verification LP, however for the purpose of the algorithm it is
of feasibility of the initial ϕ, which implies g ϕ ≥ 0, all values of ϕ
an equivalent replacement potentially affecting only the order in
can only increase during steps 2-4 and hence the unary potentials
which substitutions are pruned.
guϕ remain non-negative. Therefore step 5 can not decrease the
The reduction has the following advantages:
ϕ
lower bound value f∅ .
• subsets of labels Xv \Yv can be contracted to a single repAccording to Lemma 7.1 Procedure 3 can not worsen the lower
resentative label yv , because associated unary and pairwise
bound attained by a dual solver. The following theorem guarantees
costs are equal;
that when no further pruning is possible, the corrected dual point
• It will allow (see § 7) to relax the requirements on ap- constitutes an optimal solution, ensuring persistency.
proximate dual solvers needed to establish termination and
Theorem 7.2. Let ϕ be a dual point for reduced problem ḡ
correctness of the algorithm.
• It is useful for the speed up heuristics (§ 8). In particular, satisfying (30)-(31). Then either
ϕ
1. g∅ = 0, ϕ is dual optimal and δ(y) is primal optimal, or
it is easier to find a labeling with a negative cost since we
2. (∃u ∈ V) Ou (ϕ) ∩ Yu 6= ∅.
have decreased many edge costs. It will be shown that such a
labeling allows for an early stopping of the dual solver and a
Proof. Assume (b) does not hold: (∀u ∈ V) Ou (ϕ) ⊆ Xu \Yu .
pruning of substitution without loss of maximality.
Let us pick in each node u a label zu ∈ Ou (ϕ). As ensured
by (31), for each edge uv there is a label j ∈ Xv such that
ϕ
(zu , j) = 0 and similarly, there exists i ∈ Xu such that
7 P ERSISTENCY WITH A F INITE N UMBER OF D UAL ḡuv
ϕ
ḡ
(i,
zv ) = 0. By partial submodularity of ḡ , we have
U PDATES
uv
We assume that a suboptimal dual solver is iterative and can be
represented by a procedure dual_update, which given a current
dual point ϕ makes a step resulting in a new dual point and a guess
of a primal integer solution x.
In this setting we propose Algorithm 4. In its inner loop, the
algorithm calls dual_update (line 7) checks whether a speedup shortcut is available (line 8) and verifies whether it can already
ϕ
ϕ
ϕ
ϕ
ḡuv
(zu , zv ) + ḡuv
(i, j) ≤ ḡuv
(zu , j) + ḡuv
(i, zv ) = 0. (32)
ϕ
ϕ
ϕ
Therefore, ḡuv
(zu , zv ) ≤ −ḡuv
(i, j) ≤ 0. Hence ḡuv
(zu , zv ) =
0 and it is active. Therefore δ(z) and dual point ϕ satisfy
complementarity slackness conditions and hence they are primalϕ
dual optimal and g∅ = Eḡ (z) = 0 = Eḡ (y).
10
Algorithm 4: Efficient Iterative Pruning
I
1
2
3
4
5
6
7
8
9
10
Input: Problem f ∈ R , test labeling y ∈ X ;
Output: Improving substitution p ∈ P 2,y ∩ Sf ;
(∀u ∈ V) Yu := Xu \{yu };
Set ϕ to the initial dual solution if available;
while true
Apply single node pruning ;
/* speed-up */
Construct reduced verification LP ḡ from f and current
sets Yu , according to Definition 6.1;
repeat
(ϕ, x) := dual_update(ḡ, ϕ);
if Eḡ (x) < 0 then /* speed-up
*/
Apply pruning cut with x;
goto step 4 to rebuild ḡ ;
/* Verification of Optimality
11
12
13
14
15
16
*/
ϕ0 := dual_correct(ḡ, ϕ);
0
Ou := {i | ḡuϕ (i) = 0};
if (∀u ∈ V) Ou ∩ Yu = ∅ then return p by (7);
until any stopping condition (e.g., iteration limit);
Prune: (∀u ∈ V) Yu := Yu \Ou ;
Procedure dual_update (ḡ ,ϕ)
Input: Cost vector ḡ , dual point ϕ;
Output: New dual point ϕ, approximate primal integer
solution x;
loss of optimality of the algorithm. Lemma 8.1 below suggests
to solve a simpler verification LP, minµ∈Λ0 hḡ, µi over a subset
Λ0 of Λ. This does not guarantee to remove all non-improving
substitutions (which implies one has to switch to Λ afterwards),
but can be much more efficient than the optimization over Λ. After
the lemma we provide two examples of such efficient procedures.
Lemma 8.1. Let p ∈ P 2,y and ḡ be defined by (27) (depends on
p). Let q ∈ Sf ∩ P 2,y , q ≤ p, Q = [q]. Let Λ0 ⊆ Λ, Q(Λ0 ) ⊆ Λ0
and O∗ = argminµ∈Λ0 hḡ, µi. Then (∀v ∈ V) qv (Ov∗ ) = Ov∗ .
Proof in ?? .
Note, while Theorem 6.2 is necessary and sufficient for pruning, Lemma 8.1 is only sufficient.
8.2
Pruning of Negative Labelings
Assume we found an integer labeling x such that Eḡ (x) ≤ 0 and
p(x) 6= x. Lemma 8.1 gives an answer, for which nodes v the
label xv can be pruned from the set Yv without loss of optimality.
Define the following restriction of the polytope Λ:
Λx = {µ ∈ Λ | (∀v ∈ V) µ(yv ) + µ(xv ) = 1} ⊆ Λ.
Polytope Λx corresponds to the restriction of Λ to the label set
{yv , xv } in each node v ∈ V . According to Lemma 8.1 we need
to solve the problem
O∗ := argminµ∈Λx hḡ, µi
Theorem 7.3 (Termination and Correctness of Algorithm 4). For
any stopping
condition in line 14, Algorithm 4 terminates in at
P
most v (|Xv | − 1) outer iterations and returns p ∈ Sf .
Proof. When the algorithm has not yet terminated some further
pruning is guaranteed to be possible (compare conditions in
lines 13 and 15). The iteration limit follows. When Algorithm 4
terminates, from Theorem 7.2 it follows that ϕ0 is dual optimal
and hence Ou (ϕ0 ) ⊇ Ou∗ . Therefore, (∀u ∈ V) Ou∗ ∩ Yu = ∅,
which is sufficient for p to be strictly Λ-improving according to
Corollary 4.4.
In [40] we prove that a similar result holds for a TRWS iteration without correction by arguing on complete chain
subproblems instead of individual nodes. The correction might be
needed in case the algorithm does not keep slacks on the nodes,
e.g. for SRMP [22].
The stopping condition in line 14 of Algorithm 4 controls the
aggressiveness of pruning. Performing fewer iterations may result
only in the found p not being the maximum, but in any case it
is guaranteed that the Algorithm 4 does not stall and identifies
a correct persistency. In the case when the solver does have
convergence and optimality guarantees, the time budget controls
the degree of approximation to the maximum persistency.
8
8.1
S PEED - UPS
Inference Termination Without Loss of Maximality
Next, we propose several sufficient conditions to quickly prune
some substitutions without worsening the final solution found
by the algorithm. As follows from Definition 3.2, existence of
a labeling x such that (I − [p])> f, δ(x) ≤ 0 and x 6= p(x) is
sufficient to prove that substitution p is not strictly Λ-improving.
Hence one could consider updating the current substituttion p
without waiting for an exact solution of the inference problem in
line 4. The tricky part is to find labels that can be pruned without
(33)
(34)
and exclude xv from Yv if xv ∈ Ov∗ . Due to the partial submodularity of ḡ the problem (34) is submodular and can be solved by
min-cut/max-flow algorithms [23]. Because x was found to have
non-positive energy, it is necessarily that for some nodes v there
will hold xv ∈ Ov∗ ∩ Yv and therefore some pruning will take
place.
8.3
Single Node Pruning
Let us consider “a single node” polytope Λu,i := {µ ∈ Λ |
µu (yu ) + µu (i) = 1; (∀v 6= u) µv (yv ) = 1}. It is a special
case of Λx when y and x differ in a single node u only and
x
Pu = i. In this case problem (34) amounts to calculating ḡu (xu )+
v∈N (u) ḡuv (xu , yv ) . If this value is non-positive, xu must be
excluded from Yu . The single node pruning can be applied to all
pairs (u, i) exhaustively, but it is more efficient to keep track of
the nodes for which sets Yv have changed (either due to a negative
labeling pruning, active labels pruning in line 15 or the single node
pruning itself) and check their neighbors.
8.4
Efficient Message Passing
The main computational element in dual coordinate ascent solvers
like TRWS or MPLP is passing a message, i.e., an update of
the form mini∈Xu (fuv (i, j) + a(i)). In many practical cases the
message passing for f can be computed in time linear in the
number of labels [15, 28, 3]. This is the case when fuv is a
convex function of i−j (e.g., |i−j|, (i−j)2 ) or a minimum of
few such functions (e.g. Potts model is min(1, |i−j|)). However,
in Algorithm 1 we need to solve the problem with the cost vector
g = (I − P T )f , resp. ḡ (27) if we apply the reduction. It turns
out that whenever there is a fast message passing method for f ,
the same holds for ḡ .
11
Our-CPLEX Our Algorithm 1 (Iterative Relaxed Inference) usOur-TRWS
[43]-CPLEX
[43]-TRWS
ε-L1 [35]
Kovtun[24]
MQPBO
MQPBO-10
ing CPLEX [17].
Our Algorithm 4 using TRW-S [20]. Initial solution uses at most 1000 iterations (or the method
has converged). All speedups.
Method [43] with CLPEX.
Method [43] with TRW-S.
Single LP formulation of the maximum strong
persistency [35] solved with CPLEX.
One-against-all method of Kovtun [25].
Multilabel QPBO [19].
MQPBO with 10 random permutations, accumulating persistency.
TABLE 1: List of Evaluated Methods
Theorem 8.2 (Fast message passing). Message passing for an
edge term ḡuv (27) can be reduced to that for fuv in time O(|Yu |+
|Yv |). Proof in Table 6.
This complexity is proportional to the size of the sets Yu . The
more labels are pruned from sets Yu in the course of the algorithm,
the less work is required.
Note, that contrary to limiting the number of iterations of
a dual solver, described in § 7, the speedups presented in this
section do not sacrifice the persistence maximality (6). In our
experiments for some instances, Algorithm 4 finished before ever
reaching step 15. In such cases the found substitution p ∈ Sf is
the maximum.
9
E XPERIMENTAL E VALUATION1
In the experiments we study how well we approximate the
maximum persistency [35], Table 2; illustrate the contribution of
different speedups, Table 4; give an overall performance comparison to a larger set of relevant methods, Table 3; and provide
a more detailed direct comparison to the most relevant scalable
method [43] using exact and approximate LP solvers, Table 5. As a
measure of persistency we use the percentage of labels eliminated
by the improving substitution p:
P
v∈V |Xv \pv (Xv )|
P
v∈V (|Xv |−1)
9.1
=
P
|Yv |
P v∈V
.
v∈V (|Xv |−1)
(35)
Random Instances
Table 2 gives comparison to [43] and [35] on random instances
generated as in [35] (small problems on 4-connected grid with
uniformly distributed integer potentials for “full” model and of
the Potts type for “Potts” model, all not LP-tight). It can be
seen that our exact Algorithm 1 performs identically to the ε-L1
formulation [35]. Although it solves a series of LPs, as opposed
to a single LP solved by ε-L1, it scales better to larger instances.
Instances of size 20x20 in the ε-L1 formulation are already too
difficult for CPLEX: it takes excessive time and sometimes returns
a computational error. The performance of the dual Algorithm 4
confirms that we loose very little in terms of persistency but gain
significantly in speed.
1
The implementation of our method is available through http://cmp.felk.
cvut.cz/∼shekhovt/persistency
9.2
Benchmark Problems
Table 3 summarizes average performance on the OpenGM MRF
benchmark [18]. The datasets include previous benchmark instances from computer vision [45] and protein structure prediction [29, 51] as well as other models from the literature. Results
per instance are given in § B.
9.3
Speedups
In this experiment we report how much speed improvement was
achieved with each subsequent technique of § 8. The evaluation
in Table 4 starts with a basic implementation (using only a warm
start). The solver is allowed to run at most 50 iterations in the
partial optimality phase until pruning is attempted. We expect that
on most datasets the percentage of persistent labels improves when
we apply the speedups (since they are without loss of maximality).
9.4
Discussion
Tables 2 and 5 demonstrate that Our-TRWS, which is using
a suboptimal dual solver, closely approximates the maximum
persistency [35]. Our method is also significantly faster and scales
much better. The method [43] is the closest contender to ours in
terms of algorithm design. Tables 2, 3 and 5 clearly show that
our method determines a larger set of persistent variables. This
holds true with exact (CPLEX) as well as approximate (TRWS)
solvers. There are two reasons for that as discusssed in § 4.3.
First, we optimize over a larger set of substitutions than [43],
i.e., we identify per-label persistencies while [43] is limited to
the whole-variable persistencies. Second, even in the case of the
whole-variable persistencies the criterion in [43] is in general
weaker than (18) and depends on the initial reparametrization
of the problem. This later difference does not matter for Potts
models [44], the examples Figs. 6 and 8, but does matter, e.g., in
Fig. 9. Although our method searches over a significantly larger
space of possible substitutions, it needs fewer TRW-S iterations
due to speedup techniques. Details on iteration counts can be
found in § B. In the comparison of running time it should be
taken into account that different methods are optimized to a
different degree. Nevertheless, it is clear that the algorithmic
speedups were crucial in making the proposed method much more
practical than [43] and [35] while maintaining high persistency
recall quality.
To provide more insights to the numbers reported, we illustrate
in Figs. 1 and 6 to 9 some interesting cases. Fig. 6 shows “the
hardest” instance of color-seg-n4 family. Identified persistencies allow to fix a single label in most of the pixels, but for
some pixels more than one possible label remains. The remainder
of the problem has the reduced search space p(X ), which can be
passed to further solvers. The tsukuba image Fig. 1 is interesting
because it has appeared in many previous works. The performance
of graph-cut based persistency methods relies very much on strong
unary costs, while the proposed method is more robust. Fig. 7
shows an easy example from object-seg, where LP relaxation
is tight, the dual solver finds the optimal labeling y and our
verification LP confirms that this solution is unique. In Fig. 8
we show a hard instance of mrf-stereo. Partial reason for its
hardness is integer costs, leading to non-uniqueness of the optimal
solution. In Fig. 9, photomontage/pano instance, we report
79% solution completeness, but most of these 79% correspond to
trivial forbidden labels in the problem (very big unary costs). At
the same time other methods perform even worse. This problem
12
Problem family
10x10 Potts-3
10x10 full-3
20x20 Potts-3
20x20 full-3
20x20 Potts-4
20x20 full-4
[43]-CPLEX
0.18s 58.46%
0.24s 2.64%
3.25s 73.95%
2.81s 0.83%
12.45s 23.62%
3.96s 0.01%
[43]-TRWS
0.05s 58.38%
0.09s 1.22%
0.21s 68.49%
0.37s 0.83%
0.39s 18.43%
0.39s 0.01%
ε-L1 [35]
0.05s 72.27%
0.06s 62.90%
0.87s 87.38%
0.95s 72.66%
19.40s 74.28%
21.08s 6.58%
Our-CPLEX
0.18s 72.27%
0.24s 62.90%
2.43s 87.38%
3.03s 72.66%
8.56s 74.28%
12.41s 6.58%
Our-TRWS
0.04s 72.21%
0.05s 62.57%
0.06s 87.38%
0.07s 72.31%
0.08s 73.63%
0.08s 6.58%
TABLE 2: Performance evaluation on random instances of [35]. For each problem family (size, type of potentials and number of labels)
average performance over 100 samples is given. To allow for precise comparison all methods are initialized with the same test labeling y found
by LP relaxation. Our-TRWS closely approximates Our-CPLEX, which matches ε-L1 [35], and scales much better.
Problem family
#I
#L
#V
MQPBO
MQPBO-10 Kovtun[24]
mrf-stereo
3 16-60 > 100000
†
†
1.0s 0.23%
mrf-photomontage 2
5-7 ≤ 514080 93s 22% 866s 16% 0.4s 15.94%
color-seg
3
3-4 ≤ 424720 22s 11%
87s 16% 0.3s
98%
color-seg-n4
9 3-12 ≤ 86400 22s
8% 398s 14% 0.2s
67%
ProteinFolding
21 ≤ 483
≤ 1972 685s
2% 2705s 2% 0.02s 4.56%
object-seg
5
4-8
68160 3.2s 0.01%
†
0.1s 93.86%
[43]-TRWS Our-TRWS
2.5h
13% 117s 73.56%
3.7h
16% 483s 41.98%
1.3h 99.88% 61.8s 99.95%
321s
90% 4.9s 99.26%
48s
18% 9.2s 55.70%
138s 98.19% 2.2s 100%
TABLE 3: Performance on OpenGM benchmarks. Columns #I,#L,#V denote the number of instances, labels and variables respectively. For
each method an average over all instances in a family is reported. † – result is not available (memory / implementation limitation).
Instance
Initialization
Extra time for persistency
(1000 it.)
no speedups
+reduction
+node pruning +labeling pruning +fast msgs
Protein folding 1CKK
8.5s
268s (26.53%) 168s (26.53%) 2.0s (26.53%)
2.0s (26.53%) 2.0s (26.53%)
colorseg-n4 pfau-small
9.3s
439s (88.59%) 230s (93.41%) 85s (93.41%)
76s (93.41%)
19s (93.41%)
TABLE 4: Evaluation of speedups on selected examples: computational time drops, as from left to right we add techniques described in § 6.
1CKK: an example when the final time for persistency is only a fraction of the initialization time. pfau-small: an example when times for
initialization and persistency are comparable; speedups also help to improve the persistency as they are based on exact criteria.
Instance
1CKK
1CM1
1SY9
2BBN
PDB1B25
PDB1D2E
#L #V [43]-CPLEX [43]-TRWS Our-CPLEX
Our-TRWS
≤ 445 38 2503s
0% 46s
0% 2758s 27% 8.5+2s 26.53%
≤ 350 37 2388s
0% 51s
0% 4070s 34% 9+3.9s 29.97%
≤ 425 37 1067s
0% 67s
0% 2629s 51% 11+4.2s 57.98%
≤ 404 37 9777s
0% 5421s
0% 9677s
9% 16+4.3s 14.17%
≤ 81 1972 325s
22% 120s 22% 1599s 84% 4.3+7.3s 87.84%
≤ 81 1328 483s
59% 83s 59% 154s 98% 1.6+1.8s 98.25%
TABLE 5: Comparison to [43] using exact and approximate LP solvers. Examples of hard ProteinFolding instances [29, 51]. For
Our-TRWS the initialization + persistency time is given. An occasionally better persistency of Our-TRWS vs. Our-CPLEX is explained
by different test labelings produced by the CPLEX and TRW-S solvers (unlike in Table 2). The results of ε-L1 [35] wold be identical to
Our-CPLEX, as has been proven and verified on random instances.
has hard interaction constraints. It seems that hard constraints
and ambiguous solutions pose difficulties to all methods including
ours.
10
C ONCLUSIONS AND O UTLOOK
We presented an approach to find persistencies for an exp-APXcomplete problem employing only solvers for a convex relaxation. Using a suboptimal solver for the relaxed problem, we
still correctly identify persistencies while the whole approach
becomes scalable. Our method with an exact solver matches
the maximum persistency [35] and with a suboptimal solver
closely approximates it, outperforming state of the art persistency
techniques [43, 19, 25]. The speedups we have developed allow to
achieve this at a reasonable computational cost making the method
much more practical than the works [35, 43] we build on. In fact,
our approach takes an approximate solver, like TRW-S, and turns
it into a method with partial optimality guarantees at a reasonable
computational overhead.
We believe that many of the presented results can be extended
to higher order graphical models and tighter relaxations. Practical
applicability with other approximate solvers can be explored. A
further research direction that seems promising is mixing different
optimization strategies such as persistency and cutting plane
methods.
ACKNOWLEGEMENT
Alexander Shekhovtsov was supported by the Austrian Science Fund
(FWF) under the START project BIVISION, No. Y729. Paul Swoboda
and Bogdan Savchynskyy were supported by the German Research
Foundation (DFG) within the program “Spatio-/Temporal Graphical
Models and Applications in Image Analysis”, grant GRK 1653.
Bogdan Savchynskyy was also supported by European Research
Council (ERC) under the European Unions Horizon 2020 research
and innovation program (grant agreement No 647769).
R EFERENCES
[1] OpenGM benchmark. http://hci.iwr.uni-heidelberg.de/opengm2/
?l0=benchmark.
13
Kovtun[24]: 1s, 5.6%
[43]-TRWS: 13 min 10.43%
Ours: 10+12s, 93.4%
Ours, remainder
12
11
10
9
8
7
6
5
4
3
2
1
0
Fig. 6: Performance on a hard segmentation problem. The remainder
of the problem visualizes |pu (Xu )| for all pixels. The result of the
improved method [44]-TRWS is the same as [43]-TRWS, because
this is a Potts model.
object-seg
[24]: 0.2s, 80.5%
Ours: 1.4s, 100%
(LP-tight)
Fig. 7: Examples of an easy problem, object-seg. TRWS finds
the optimal solution (zero integrality gap) and therefore our method
as well as [43]-TRWS, [44]-TRWS report 100%.
mrf-stereo
[43]-TRWS: 1h, 38%
Ours: 62+180s, 75%
Ours, remainder
Fig. 8: Examples of a hard stereo problem, ted-gm. Method of
Kovtun[24] gives 0.42% in 2.5s and is therefore not displayed.
The result [44]-TRWS is the same as [43]-TRWS, because this is
a Potts model too.
[2] Adams, W. P., Lassiter, J. B., and Sherali, H. D. (1998). Persistency in 0-1 polynomial programming. Mathematics of Operations
Research, 23(2):359–389.
[3] Aggarwal, A., Klawe, M. M., Moran, S., Shor, P., and Wilber, R.
(1987). Geometric applications of a matrix-searching algorithm.
Algorithmica, 2(1):195–208.
[4] Alahari, K., Kohli, P., and Torr, P. H. S. (2008). Reduce, reuse &
recycle: Efficiently solving multi-label MRFs. In CVPR.
[5] Boros, E. and Hammer, P. L. (2002). Pseudo-Boolean optimization. Discrete Applied Mathematics.
[6] Boykov, Y., Veksler, O., and Zabih, R. (2001). Fast approximate
energy minimization via graph cuts. IEEE Trans. Pattern Anal.
Mach. Intell., 23(11):1222–1239.
[7] Choi, J. and Rutenbar, R. A. (2012). Hardware implementation
of MRF MAP inference on an FPGA platform. In Field Programmable Logic, pages 209–216. IEEE.
[8] Cooper, M., de Givry, S., Sanchez, M., Schiex, T., Zytnicki, M.,
and Werner, T. (2010). Soft arc consistency revisited. Artificial
Intelligence, 174(78):449 – 478.
[9] de Givry, S., Prestwich, S. D., and O’Sullivan, B. (2013). Deadend elimination for weighted CSP. In Schulte, C., editor, CP,
volume 8124 of Lecture Notes in Computer Science, pages 263–
272. Springer.
[10] Desmet, J., Maeyer, M. D., Hazes, B., and Lasters, I. (1992).
The dead-end elimination theorem and its use in protein side-chain
positioning. Nature, 356.
[11] Freuder, E. C. (1991). Eliminating interchangeable values in
constraint satisfaction problems. In Proceedings of the Ninth National Conference on Artificial Intelligence - Volume 1, AAAI’91,
pages 227–233. AAAI Press.
[12] Globerson, A. and Jaakkola, T. (2007). Fixing max-product:
Convergent message passing algorithms for MAP LP-relaxations.
In NIPS.
[13] Goldstein, R. F. (1994). Efficient rotamer elimination applied to
protein side-chains and related spin glasses. Biophysical Journal,
66(5):1335–1340.
[14] Gridchyn, I. and Kolmogorov, V. (2013). Potts model, parametric
maxflow and k-submodular functions. In ICCV.
[15] Hirata, T. (1996). A unified linear-time algorithm for computing
distance maps. Information Processing Letters, 58(3):129 – 133.
[16] Hurkat, S., Choi, J., Nurvitadhi, E., Martınez, J. F., and Rutenbar,
R. A. (2012). Fast hierarchical implementation of sequential treereweighted belief propagation for probabilistic inference. In Field
Programmable Logic.
[17] ILOG, Inc. ILOG CPLEX: High-performance software for
mathematical programming and optimization. See http://www.ilog.
com/products/cplex/.
[18] Kappes, J. H., Andres, B., Hamprecht, F. A., Schnörr, C.,
Nowozin, S., Batra, D., Kim, S., Kausler, B. X., Kröger, T.,
Lellmann, J., Komodakis, N., Savchynskyy, B., and Rother, C.
(2015). A comparative study of modern inference techniques for
structured discrete energy minimization problems. International
Journal of Computer Vision, pages 1–30.
[19] Kohli, P., Shekhovtsov, A., Rother, C., Kolmogorov, V., and Torr,
P. (2008). On partial optimality in multi-label MRFs. In ICML.
[20] Kolmogorov, V. (2006). Convergent tree-reweighted message
passing for energy minimization. PAMI, 28(10).
[21] Kolmogorov, V. (2012). Generalized roof duality and bisubmodular functions. Discrete Applied Mathematics, 160(4-5):416–426.
[22] Kolmogorov, V. (2013). Reweighted message passing revisited.
ArXiv e-prints.
[23] Kolmogorov, V. and Zabin, R. (2004). What energy functions
can be minimized via graph cuts? PAMI, 26(2):147–159.
[24] Kovtun, I. (2003). Partial optimal labeling search for a NP-hard
subclass of (max, +) problems. In DAGM-Symposium, pages 402–
409.
[25] Kovtun, I. (2011). Sufficient condition for partial optimality for
(max, +) labeling problems and its usage. Control Systems and
Computers, 2. Special issue.
[26] Lecoutre, C., Roussel, O., and Dehani, D. E. (2012). WCSP
integration of soft neighborhood substitutability. In Milano, M.,
editor, CP, volume 7514 of Lecture Notes in Computer Science,
pages 406–421. Springer.
[27] Li, M., Shekhovtsov, A., and Huber, D. (2016). Complexity of
discrete energy minimization problems. In European Conference
on Computer Vision, pages 834–852.
[28] Meijster, A., Roerdink, J. B. T. M., and Hesselink, W. H. (2000).
A General Algorithm for Computing Distance Transforms in Linear
Time, pages 331–340. Springer US, Boston, MA.
[29] PIC2011. The probabilistic inference challenge (PIC2011). http:
14
mrf-photomontage
7
Kovtun [24]: 0.5s, 27.5%
[44]-TRWS: 3h, 37.75%
Ours, remainder: 130+390s, 79.2%
6
5
4
3
2
1
0
Fig. 9: Example of a very hard instance, photomontage. The number of views covering a pixel is usually smaller than the total number of
views = number of labels. The 79% by our method mostly originate from the elimination of these redundant labels. The result of [43]-TRWS
gives 27.55% (visually same as Kovtun [24]) in 3.9h. [44]-TRWS improves on this due to choosing the optimal reparametrization [44]. Note
that our partial labeling (the part with one label remaining) is larger.
//www.cs.huji.ac.il/project/PASCAL/.
[30] Rother, C., Kolmogorov, V., Lempitsky, V. S., and Szummer, M.
(2007). Optimizing binary MRFs via extended roof duality. In
CVPR.
[31] Savchynskyy, B., Kappes, J., Schmidt, S., and Schnörr, C.
(2011). A study of Nesterov’s scheme for Lagrangian decomposition and MAP labeling. In CVPR.
[32] Savchynskyy, B., Schmidt, S., Kappes, J. H., and Schnörr, C.
(2012). Efficient MRF energy minimization via adaptive diminishing smoothing. In UAI.
[33] Schlesinger, M. I. and Antoniuk, K. V. (2011). Diffusion
algorithms and structural recognition optimization problems. Cybernetics and Sys. Anal., 47(2):175–192.
[34] Shekhovtsov, A. (2013). Exact and Partial Energy Minimization
in Computer Vision. PhD Thesis CTU–CMP–2013–24, CMP,
Czech Technical University in Prague.
[35] Shekhovtsov, A. (2014a). Maximum persistency in energy
minimization. In CVPR.
[36] Shekhovtsov, A. (2014b). Maximum persistency in energy
minimization. Technical report, Graz University of Technology.
[37] Shekhovtsov, A. (2015). Higher order maximum persistency and
comparison theorems. CVIU. accepted.
[38] Shekhovtsov, A., Reinbacher, C., Graber, G., and Pock, T.
(2016). Solving dense image matching in real-time using discretecontinuous optimization. In Computer Vision Winter Workshop,
page 13.
[39] Shekhovtsov, A., Swoboda, P., and Savchynskyy, B. (2015a).
Maximum persistency via iterative relaxed inference with graphical
models. In CVPR.
[40] Shekhovtsov, A., Swoboda, P., and Savchynskyy, B. (2015b).
Maximum persistency via iterative relaxed inference with graphical
models. CoRR, abs/1508.07902.
[41] Shlezinger, M. (1976). Syntactic analysis of two-dimensional
visual signals in the presence of noise. Cybernetics and Systems
Analysis, 4:113–130. See review [49].
[42] Swoboda, P., Savchynskyy, B., Kappes, J. H., and Schnörr, C.
(2013). Partial optimality via iterative pruning for the Potts model.
In SSVM.
[43] Swoboda, P., Savchynskyy, B., Kappes, J. H., and Schnörr, C.
(2014). Partial optimality by pruning for MAP-inference with
general graphical models. In CVPR.
[44] Swoboda, P., Shekhovtsov, A., Kappes, J. H., Schnorr, C., and
Savchynskyy, B. (2016). Partial optimality by pruning for MAPinference with general graphical models. PAMI, 38(7):1370–1382.
[45] Szeliski, R., Zabih, R., Scharstein, D., Veksler, O., Kolmogorov,
V., Agarwala, A., Tappen, M. F., and Rother, C. (2008). A
comparative study of energy minimization methods for Markov
random fields with smoothness-based priors. PAMI, 30(6).
[46] Vanderbei, R. J. (2001). Linear Programming: Foundations and
Extensions. Department of operations and research and financial
engineering, Princeton university.
[47] Wainwright, M. J. and Jordan, M. I. (2008). Graphical models,
exponential families, and variational inference. Found. Trends
Mach. Learn., 1(1-2):1–305.
[48] Wang, C. and Zabih, R. (2016). Relaxation-based preprocessing
techniques for markov random field inference. In The IEEE
Conference on Computer Vision and Pattern Recognition (CVPR).
[49] Werner, T. (2007). A linear programming approach to max-sum
problem: A review. PAMI, 29(7).
[50] Windheuser, T., Ishikawa, H., and Cremers, D. (2012). Generalized roof duality for multi-label optimization: Optimal lower
bounds and persistency. In ECCV.
[51] Yanover, C., Schueler-Furman, O., and Weiss, Y. (2008). Minimizing and learning energy functions for side-chain prediction.
Jour. of Comp. Biol., 15(7).
[52] Zach, C. (2014). A principled approach for coarse-to-fine MAP
inference. In CVPR.
A PPENDIX A
P ROOFS
Proposition 5.2. Arc consistency is a necessary condition for relative
interior optimality: if Ov (ϕ) = Ov∗ for all v ∈ V then f ϕ is arc
consistent.
Proofs of the Generic Algorithms
Proof. Condition Ov (ϕ) = Ov∗ implies that ϕ satisfies strict complementarity with some primal optimal solution µ. The strict complementarity implies that (∀i ∈ Xv ) (fuϕ (i) = 0 ⇒ µu (i) > 0). By
feasibility of µ, there must hold (∀v ∈ N (u)) (∃j ∈ Xv ) µuv (i, j) >
0. And by using complementary slackness again, it must be that
ϕ
fuv
(i, j) = 0. Similarly, the second condition of arc consistency is
verified. It follows that f ϕ is arc consistent.
Proposition 4.3. For a given substitution p, let O∗ ⊆ Λ denote the
set of minimizers of the verification LP (18) and
Ov∗ := {i ∈ Xv | (∃µ ∈ O∗ ) µv (i) > 0},
(19)
which is the support set of all optimal solutions in node v . Then
p ∈ Sf iff (∀v ∈ V ∀i ∈ Ov∗ ) pv (i) = i.
Proof. Direction ⇒. Let p ∈ Sf . Assume for contradiction that (∃v ∈
V ∃i ∈ Ov∗ ) pv (i) 6= i. Since i ∈ Ov∗ there exists µ ∈ O∗ such that
µv (i) > 0. Its image µ0 = [p]µ has µv (i) = 0 due to pv (i) 6= i by
evaluating the extension (10). This contradicts [p]µ = µ.
Direction ⇐. Now let (∀v ∈ V ∀i ∈ Ov∗ ) pv (i) = i. Clearly,
[p]µ = µ holds for all µ on the support set given by (Ov∗ | v ∈ V),
hence for O∗ . It remains to show that the value of the minimum
in (18) is zero. For µ ∈ O∗ we have [p]µ = µ and the objective
in (18), h(I − [p])T f, µi = hf, µ − [p]µi vanishes.
Proofs of the Reduction
The proof of the reduction Theorem 6.2 and Lemma 8.1 (used in
speed-up heuristics) requires several intermediate results. Recall that
a correct pruning can be done when we have a guarantee to preserve
all strictly improving substitutions q , assuming q ≤ p. Therefore
statements in this section are formulated for such pairs. We will
consider adjustments to the cost vector that preserve the set of strictly
improving substitutions. These adjustments do not in general preserve
optimal solutions to the associated LP relaxation.
Theorem 4.6. Substitution p returned by Algorithm 1 is the maximum of Sf ∩ P 2,y and thus it solves (6).
Lemma A.4. Let q ≤ p. Then q ∈ Sf iff q ∈ Sg for g = (I − [p])T f .
Proof. Let Q = [q], P = [p]. Since q ≤ p there holds P Q = P . It
implies (I − P )(I − Q) = (I − Q). Therefore,
Proof. The two following lemmas form a basis for the proof.
Lemma A.1 ([36], Thm. 3(b)). Let q ∈ Sf ; q ≤ p and let O∗ =
argminµ∈Λ h(I − [p])> f, µi, i.e., as in line 4 of Algorithm 1.
Then (∀µ ∈ O∗ ) [q]µ = µ.
hg, (I − Q)µi = h(I − P )T f, (I − Q)µi
= hf, (I − P )(I − Q)µi = hf, (I − Q)µi.
In the case of substitutions from the class P 2,y , the statement
additionally simplifies as follows.
Assume µ ∈ Λ is such that Qµ 6= µ. Equality (39) ensures that
hg, (I − Q)µi > 0 iff hf, (I − Q)µi > 0. The theorem follows from
definition of Sf , Sg .
To reformulate the condition q ∈ Sf we will use the following
dual characterization.
Corollary A.2. Assume conditions of Lemma A.1 and additionally,
let q ∈ P 2,y and Ov∗ := {i ∈ Xv | (∃µ ∈ O∗ ) µv (i) > 0} then
(∀v ∈ V, ∀i ∈ Ov∗ ) qv (i) = i.
(36)
(39)
(40)
Theorem A.5 (Characterization [36]). Let P = [p]. Then
(∀µ ∈ Λ) hf, P µi ≤ hf, µi
Proof. It follows similarly to Corollary 4.4. Assume for contradiction
that (∃v ∈ V ∃i ∈ Ov∗ ) qv (i) 6= i. Since i ∈ Ov∗ there exists
µ ∈ O∗ such that µv (i) > 0. It’s image µ0 = [q]µ has µv (i) = 0
due to qv (i) 6= i by evaluating the extension (10). This contradicts to
[q]µ = µ, the statement of Lemma A.1.
iff there exists a reparametrization ϕ such that
Lemma A.3. Let pt denote the substitution p computed in line 3 of
Algorithm 1 on iteration t. The algorithm maintains the invariant that
(∀q ∈ Sf ∩ P 2,y ) q ≤ pt .
The following lemma assumes an arbitrary substitution q , not
necessarily in P 2,y and takes as input sets Uu that are subsets
of immovable labels. In the context of Theorem 6.2, we will use
Uu = Xu \Yu .
Proof. We prove by induction. The statement holds trivially for the
first iteration. Assume it is true for the current iteration t. Then for
any q ∈ Sf ∩ P 2,y holds q ≤ pt and therefore Corollary A.2 applies.
We can show that line 8 only prunes substitutions that are not in
Sf ∩ P 2,y as follows.
Let pt+1 be the substitution on the next iteration, i.e. computed
by line 3 after pruning line 8.
Assume for contradiction that ∃q ∈ Sf ∩P 2,y such that q 6≤ pt+1 .
By negating the definition and expanding,
Lemma A.6 (Reduction 1). For a substitution q let Uu ⊆ {i ∈ Xu |
qu (i) = i} for all u ∈ V . Let guv (i, j) = 0 for all (i, j) ∈ Uu × Uv
and let ḡ be defined by
(∃v ∈ V) pt+1
v (Xv ) 6⊆ qv (Xv ),
P Tf ϕ ≤ f ϕ.
ḡv = gv , v ∈ V;
min g (i0 , j), i ∈ Uu , j ∈
/ Uv ,
i0 ∈Uu uv
0
guv (i, j ), i ∈
/ Uu , j ∈ Uv ,
g uv (i, j) = min
j 0 ∈Uv
guv (i, j),
otherwise.
(37a)
⇔ (∃v ∈ V ∃i ∈ Xv ) i ∈ pt+1
v (Xv ) ∧ i 6∈ qv (Xv ),
(37b)
⇔ (∃v ∈ V ∃i ∈ Xv ) pt+1
v (i) = i ∧ qv (i) 6= i.
(37c)
(42)
(43a)
(43b)
Then q ∈ Sg iff q ∈ Sḡ .
Proof. Direction ⇐. Let us verify the following inequality:
(∀ij ∈ Xuv ) guv (qu (i), qv (j)) − ḡuv (qu (i), qv (j))
≤ guv (i, j) − ḡuv (i, j).
If i was pruned in line 8, i ∈ Ov∗ , then it must be that qv (i) = i,
which contradicts to (37c). Therefore
(∃v ∈ V ∃i ∈ Xv \Ov∗ ) pt+1
v (i) = i ∧ qv (i) 6= i.
(41)
(44)
We need to consider only cases where ḡuv (i, j) 6= guv (i, j). Let
i ∈ Uu and j ∈
/ Uv (the remaining case is symmetric). In this case
qu (i) = i. Substituting ḡ we have to prove
(38)
t
t
However, in this case pt+1
v (i) = pv (i) = i and q ≤ p fails to hold,
which contradicts to the assumption of induction. Therefore q ≤ pt+1
holds by induction on every iteration.
By Proposition 4.5 the algorithm terminates and returns a substitution in Sf ∩ P 2,y . By Lemma A.3 the returned substitution p satisfies
p ≥ q for all q ∈ Sf ∩ P 2,y . It is the maximum.
guv (i, qv (j)) − min
guv (i0 , qv (j))
0
i ∈Uu
(45)
≤guv (i, j) − min
guv (i0 , j).
0
i ∈Uu
The left hand side is zero because all the respective components of
g are zero by assumption. At the same time the right hand side is
15
ϕ
ϕ
b + c − a = d¯. We proved that ḡuv
(qu (i), qv (j)) ≤ ḡuv
(i, j). In total,
ϕ
ḡ satisfies all component-wise inequalities, same as g ϕ in (48). By
Theorem A.5, it follows that hḡ, (I − Q)µi ≥ 0. The strict inequality
in case Qµ 6= µ is considered similarly to Lemma A.6.
non-negative since i ∈ Uu . The inequality (44) holds. It implies (by
multiplication with pairwise components of µ and using the equality
of unary components of g and ḡ ) that
(∀µ ∈ Λ) hg, Qµi − hḡ, Qµi ≤ hg, µi − hḡ, µi,
(46)
Theorem 6.2 (Reduction). Let p ∈ P 2,y and ḡ be the corresponding
reduced cost vector constructed as in Def. 6.1. Let also q ∈ P 2,y ,
q ≤ p. Then q ∈ Sf iff q ∈ Sḡ .
where Q = [q]. Note, cost vector ḡ satisfying (46) is called auxiliary
for g in [25, 34]. Inequality (46) is equivalent to
hḡ, (I − Q)µi ≤ hg, (I − Q)µi.
Proof. Let g = (I − P )T f . By Lemma A.4, q ∈ Sf iff q ∈ Sg . We
need to consider only pairwise terms. Let uv ∈ E . Since q ≤ p, if
pu (i) = i then necessarily qu (i) = i. Let p be defined using sets Yu
as in (7). The reduction ḡ in (27) will be composed of reductions by
Lemma A.6 and Lemma A.7.
From g = (I −P )T f we have that for i ∈ Xu \Yu and j ∈ Xv \Yv
guv (i, j) = 0. Conditions of Lemma A.6 are satisfied with Uu =
Xu \Yu . We obtain part of the reduction (27) for cases when i ∈
/ Yu
or j ∈
/ Yv . Let us denote the reduced vector ḡ 0 . Applying Lemma A.7
to it, we obtain ḡ as defined in (27).
(47)
Whenever the left hand side of (47) is strictly positive then so is the
right hand side and therefore from q ∈ Sḡ follows q ∈ Sg .
Direction ⇒. Assume q ∈ Sg . By Theorem A.5, there exist dual
multipliers ϕ such that g ϕ verifies inequality (42), in components:
(∀u ∈ V, ∀i ∈ Xu ) guϕ (qu (i)) ≤ guϕ (i);
ϕ
ϕ
(∀uv ∈ E, ∀ij ∈ Xuv ) guv
(qu (i), qv (j)) ≤ guv
(i, j).
(48)
Let us expand the pairwise inequality in the case i ∈ Uu , j ∈
/ Uv . Let
qv (j) = j ∗ . Using qu (i) = i we obtain
∗
Lemma 8.1. Let p ∈ P 2,y and ḡ be defined by (27) (depends on p).
Let q ∈ Sf ∩ P 2,y , q ≤ p, Q = [q]. Let Λ0 ⊆ Λ, Q(Λ0 ) ⊆ Λ0 and
O∗ = argminµ∈Λ0 hḡ, µi. Then (∀v ∈ V) qv (Ov∗ ) = Ov∗ .
∗
guv (i, j ) − ϕuv (i) − ϕvu (j ) ≤ guv (i, j) − ϕuv (i) − ϕvu (j);
guv (i, j ∗ ) − ϕvu (j ∗ ) ≤ guv (i, j) − ϕvu (j).
(49)
(50)
Proof. Let µ ∈ O∗ . Assume for contradiction that Qµ 6= µ. In this
case, by Theorem 6.2, we have that hḡ, Qµi < hḡ, µi. Since µ ∈ Λ0
and Q(Λ0 ) ⊆ Λ there holds Qµ ∈ Λ0 . It follows that Qµ is a feasible
solution of a better cost than µ which contradicts optimality of µ. It
must be therefore that Qµ = µ. The claim q(Ov∗ ) = Ov∗ follows.
(51)
Termination with arc consistent Solvers
We take min over i ∈ Uu of both sides:
min guv (i, j ∗ ) − ϕvu (j ∗ ) ≤ min guv (i, j) − ϕvu (j).
i∈Uu
i∈Uu
Finally we subtract ϕuv (i) on both sides and obtain
ϕ
ϕ
ḡuv
(i, j ∗ ) ≤ ḡuv
(i, j).
Theorem A.8. Consider the verification LP defined by g = (I −
P T )f . Let g ϕ be an arc-consistent reparametrization and let Yu =
{i | p(i) 6= i}. Then at least one of the two conditions is satisfied:
ϕ
(a) g∅
= 0 and ϕ is dual optimal;
(b) (∃u ∈ V) Ou (ϕ) ∩ Yu 6= ∅.
The case when i ∈
/ Uu , j ∈ Uv is symmetric. In the remaining cases,
ϕ
ḡuv
(i, j) = ḡ(i, j) − ϕuv (i) − ϕvu (j) = g(i, j) − ϕuv (i) − ϕvu (j) =
g ϕ (i, j). In total, ḡ ϕ satisfies all component-wise inequalities that
does g ϕ in (48). By Theorem A.5,
(∀µ ∈ Λ) hḡ, Qµi ≤ hḡ, µi.
(52)
Proof. Assume (b) does not hold: (∀u ∈ V) Ou (ϕ) ⊆ Xu \Yu . For
each node u let us chose a label zu ∈ Ou (ϕ). By arc consistency,
for each edge uv there is a label j ∈ Ov (ϕ) ⊆ Xv \Yv such that
ϕ
guv
(zu , j) is active and similarly, there exists i ∈ Ou (ϕ) ⊆ Xu \Yu
ϕ
such that guv
(i, zv ) is active.
By construction, guv (i0 , j 0 ) = 0 for all i0 j 0 ∈ Xuv \Yuv and
therefore the following modularity equality holds:
We have shown that hḡ, (I − Q)µi ≥ 0. It remains to prove that
the inequality holds strictly when Qµ 6= µ. Since q ∈ Sg , there
holds hg, µi < hQT g, µi. It is necessary that at least one of unary
or pairwise inequalities (48) from the support of µ holds strictly in
which case inequality (52) is also strict.
Lemma A.7 (Reduction 2). For a substitution q and cost vector g let
ḡ = g − ∆+ , where ∆+ ∈ RI+ has zero unary components and its
pairwise components read:
∆+
(53)
uv (i, j) = max 0, guv (i, j) + guv (qu (i), qv (j))
ϕ
ϕ
guv
(zu , zv ) + guv
(i, j)
(56)
= 0 − ϕuv (zu ) − ϕuv (zv )) + 0 − ϕuv (i) − ϕuv (j)
ϕ
ϕ
=guv
(zu , j) + guv
(i, zv ).
ϕ
From guv
(zu , j) being active we have
− guv (i, qv (j)) − guv (qu (i), j) .
ϕ
ϕ
guv
(zu , j) ≤ guv
(i, j).
Then q ∈ Sg iff q ∈ Sḡ .
ϕ
guv
(zu , zv )
(57)
ϕ
guv
(i, zv )
Proof. The scheme of the proof is similar to Lemma A.6. The unary
components of g and ḡ are equal. If we show inequality (44), the
implication q ∈ Sḡ ⇒ q ∈ Sg will follow as in Lemma A.6. For our
ḡ , inequality (44) reduces to
By adding (56) and (57) we obtain
≤
and
hence (zu , zv ) is active. Therefore δ(z) and dual point ϕ satisfy
complementarity slackness and hence they are primal-dual optimal
ϕ
and g∅
= Eg (z) = 0.
+
∆+
uv (qu (i), qv (j)) ≤ ∆uv (i, j)
Lemma 5.4 (Correctness of Algorithm 2). If (∀v ∈ V) Ov (ϕ) ∩
Yv = ∅ holds for an arc consistent dual vector ϕ, then ϕ is optimal.
(54)
and due to idempotency of q the left hand side is identically zero.
Therefore inequality (44) is verified.
Direction ⇒. Assume q ∈ Sg . By Theorem A.5, there exist dual
multipliers ϕ satisfying inequalities (48). Consider
Proof. Corollary from Theorem A.8.
Fast Message Passing
(55)
Theorem 8.2 (Fast message passing). Message passing for an edge
term ḡuv (27) can be reduced to that for fuv in time O(|Yu | + |Yv |).
Let us show that component-wise inequalities (48) hold for ḡ ϕ .
Clearly they hold for unary components and for pairwise components
+
0
where ∆+
uv (i, j) = 0. Let uv ∈ E and ∆uv (i, j) > 0. Let i = qu (i)
and j 0 = qu (j). It must be that i0 6= i and j 0 6= j . Let us denote
ϕ
ϕ
ϕ
ϕ
a = guv
(i0 , j 0 ), b = guv
(i0 , j), c = guv
(i, j 0 ) and d = guv
(i, j).
ϕ
Let d¯ := guv
(i, j) − ∆+
(i,
j)
=
d
−
(a
+
d
−
b
−
c)
= b + c − a.
uv
From (48) we have that a ≤ b, c, d. It follows that 2a ≤ b + c or a ≤
Proof. The components of the reduced problem ḡ (27) can be expressed directly in components of f as in Table 6. Passing a message
on edge uv amounts to calculating ϕvu (j) := mini∈Xu a(i) +
Xu
ḡuv (i, j) for some vector a ∈ R . For j ∈
/ Yv , substituting pairwise
terms of ḡ , it expands as
ϕvu (j) := mini∈Xu a(i) + ∆uv (i) .
(58)
ḡ ϕ = (g − ∆+ )ϕ = g ϕ − ∆+ .
16
i∈
/ Yu
i ∈ Yu
i∈Y
/ u , j ∈Y
/ v
i∈Y
/ u , j∈Yv
i∈Yu , j ∈Y
/ v
i∈Yu , j∈Yv
ḡu (i) = 0;
ḡu (i) = fu (i) − fu (yu );
ḡuv (i, j) = 0; ∆uv (i) := ∆vu (j) := 0;
0
0
∆vu (j) := mini0 ∈Y
/ u fuv (i , j) − fuv (i , yv ) ,
ḡuv (i, j) = ∆vu (j);
0
0
∆uv (i) := minj 0 ∈Y
/ v fuv (i, j ) − fuv (yu , j ) ,
ḡuv (i, j) = ∆uv (i);
ḡuv (i, j) = min fuv (i, j) − fuv (yu , yv ),
∆vu (j) + ∆uv (i) .
TABLE 6: Components of the Reduced Verification Problem
Since the message is equal for all j ∈
/ Yv , it is sufficient to represent
it by ϕvu (yv ) (recall that yv ∈
/ Yv ). For j ∈ Yv , substituting pairwise
terms of ḡ and denoting c = fuv (yu , yv ),
ϕvu (j) := min mini∈Y
(59)
/ u a(i) + ∆vu (j),
mini∈Yu a(i) + min fuv (i, j) − c, ∆uv (i) + ∆vu (j)
= min mini∈Y
(60a)
/ u a(i) + ∆uv (j),
mini∈Yu a(i)+fuv (i, j) − c,
(60b)
mini∈Yu a(i) + ∆uv (i) + ∆vu (j)
(60c)
Adding ∆uv (i) inside (60a) (it is zero for i ∈
/ Yu ) and grouping (60a)
and (60c) together, we obtain for j ∈ Yv , ϕvu (j) =
min min a(i)+fuv (i, j) − c, ϕuv (yv )+∆vu (j) .
(61)
i∈Yu
Expression (60b) is a message passing for f , but the minimum is only
over Yu and the result is needed only for j ∈ Yv . This message can be
computed in time O(|Yu |+|Yv |) using the same algorithms [15, 28, 3]
(see also non-uniform min-convolution in [52]). Evaluating (61) takes
additional O(|Yv |) time and minimum in (58) takes O(|Yu |) time
assuming that components of a(i) are equal for j ∈
/ Yv (because it is
already true for ḡ and ϕ).
17
A PPENDIX B
R ESULTS P ER I NSTANCE
Instance
Algorithm
Time
needed
overall (s)
Time for
initial
solution (s)
1CKK
Our-CPLEX
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-CPLEX
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-CPLEX
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-CPLEX
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-CPLEX
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-CPLEX
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-CPLEX
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-CPLEX
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-CPLEX
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-CPLEX
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-CPLEX
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-CPLEX
Our-TRWS
2757.62
5.76
5670.00
825.00
2502.69
47.57
4070.00
6.03
5520.00
723.00
2388.46
51.33
2628.72
6.88
7494.00
2112.00
1067.46
66.73
9677.42
10.00
1736.00
2429.00
9776.60
54.21
36222.90
9.14
1008.00
1288.00
11419.60
55.26
1381.60
4.67
3728.00
540.00
1552.95
40.12
3628.90
5.83
5138.00
928.00
4618.78
41.87
7458.75
6.25
4961.00
1054.00
4473.58
61.92
5721.95
6.49
7228.00
1193.00
2163.98
44.07
12085.40
7.74
7516.00
1997.00
6137.07
93.87
3247.11
7.44
6709.00
1291.00
1776.23
44.71
1599.67
5.18
ProteinFolding
1177.62
5
5.00
3
0.00
0
0.00
0
2493.65
1
30.19
2
992.15
7
4.70
4
0.00
0
0.00
0
2198.04
3
21.60
3
416.74
5
5.50
4
0.00
0
0.00
0
910.90
4
46.77
5
5476.81
5
8.60
3
0.00
0
0.00
0
9771.18
1
42.90
2
6998.66
5
7.90
3
0.00
0
0.00
0
11409.90
2
39.60
2
765.84
4
3.91
4
0.00
0
0.00
0
1552.88
1
28.12
2
2546.68
5
5.20
3
0.00
0
0.00
0
4618.76
1
33.03
3
1996.55
5
5.30
4
0.00
0
0.00
0
4440.51
1
44.42
2
1946.20
6
4.80
6
0.00
0
0.00
0
2161.07
1
35.46
2
3007.95
6
6.50
3
0.00
0
0.00
0
6128.14
1
46.42
2
915.86
7
5.90
4
0.00
0
0.00
0
1598.07
2
25.52
2
55.01
28
2.92
18
1CM1
1SY9
2BBN
2BCX
2BE6
2F3Y
2FOT
2HQW
2O60
3BXL
pdb1b25
18
#iterations
#iterations
Algorithm 1,2 TRWS
†
1000+15
†
†
†
288+185
†
1000+65
†
†
†
242+358
†
1000+15
†
†
†
400+174
†
1000+10
†
†
†
242+146
†
1000+55
†
†
†
252+194
†
1000+60
†
†
†
363+230
†
1000+10
†
†
†
321+164
†
1000+25
†
†
†
398+222
†
1000+160
†
†
†
382+121
†
1000+55
†
†
†
352+369
†
1000+60
†
†
†
227+216
†
530+150
Logarithmic
percentage
partial
optimality
Percentage
excluded
labels
14.24%
13.83%
0.00%
0.00%
0.00%
0.00%
8.38%
8.07%
0.00%
0.00%
0.00%
0.00%
25.34%
28.06%
0.00%
0.00%
0.00%
0.00%
2.12%
2.64%
0.00%
0.00%
0.00%
0.00%
4.81%
4.39%
0.00%
0.00%
0.00%
0.00%
9.14%
8.96%
0.00%
0.00%
0.00%
0.00%
6.22%
8.39%
0.00%
0.00%
0.00%
0.00%
4.10%
4.01%
0.00%
0.00%
0.00%
0.00%
10.30%
8.33%
0.00%
0.00%
0.00%
0.00%
4.22%
4.94%
0.00%
0.00%
0.00%
0.00%
4.97%
4.66%
0.00%
0.00%
0.00%
0.00%
76.76%
83.00%
27.04%
26.53%
0.00%
0.00%
0.00%
0.00%
34.28%
29.98%
0.00%
0.00%
0.00%
0.00%
51.30%
57.98%
0.00%
0.11%
0.00%
0.00%
8.58%
14.17%
0.00%
0.00%
0.00%
0.00%
15.66%
14.21%
0.00%
0.00%
0.00%
0.00%
17.68%
15.12%
0.05%
0.00%
0.00%
0.00%
10.66%
13.74%
0.00%
0.05%
0.00%
0.00%
11.64%
11.01%
0.09%
0.07%
0.00%
0.00%
17.30%
18.08%
0.00%
0.00%
0.00%
0.00%
12.81%
15.55%
0.00%
0.00%
0.00%
0.00%
17.18%
12.35%
0.00%
0.00%
0.00%
0.00%
84.05%
87.84%
Instance
pdb1d2e
pdb1fmj
pdb1i24
pdb1iqc
pdb1jmx
pdb1kgn
pdb1kwh
pdb1m3y
pdb1qks
colseg-cow3
colseg-cow4
colseg-garden4
Algorithm
Time
needed
overall (s)
Time for
initial
solution (s)
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-CPLEX
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-CPLEX
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-CPLEX
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-CPLEX
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-CPLEX
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-CPLEX
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-CPLEX
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
Our-CPLEX
Our-TRWS
MQPBO-10
MQPBO
[43]-CPLEX
[43]-TRWS
27.00
2.00
324.64
119.71
154.76
1.67
12.00
0.00
483.55
83.82
99.33
1.05
6.00
0.00
77.30
16.67
0.06
3.00
0.00
5.66
0.82
111.58
0.74
8.00
0.00
229.09
36.03
142.20
0.67
8.00
0.00
121.71
20.02
196.03
1.37
9.00
0.00
161.57
53.49
105.77
0.46
5.00
0.00
51.33
9.15
73.60
0.79
8.00
0.00
120.60
28.19
138.12
0.30
9.00
0.00
96.77
27.99
Our-TRWS
Kovtun[24]
MQPBO-10
MQPBO
[43]-TRWS
Our-TRWS
Kovtun[24]
MQPBO-10
MQPBO
[43]-TRWS
Our-TRWS
Kovtun[24]
MQPBO-10
MQPBO
[43]-TRWS
66.30
1.00
206.00
24.00
7530.72
91.26
2.00
46.00
40.00
7395.03
0.49
0.00
14.00
1.00
33.68
0.00
0.00
72.11
27.62
25.44
1.13
0.00
0.00
34.59
6.16
12.35
0.60
0.00
0.00
16.97
3.10
0.02
0.00
0.00
5.66
0.82
18.51
0.40
0.00
0.00
24.36
4.06
15.52
0.29
0.00
0.00
16.21
3.59
17.97
0.76
0.00
0.00
24.37
6.45
9.63
0.27
0.00
0.00
12.89
2.43
18.58
0.65
0.00
0.00
25.18
4.82
15.19
0.12
0.00
0.00
15.82
3.24
color-seg
48.10
0.00
0.00
0.00
690.55
48.31
0.00
0.00
0.00
742.58
0.15
0.00
0.00
0.00
6.75
19
0
0
14
14
5
7
0
0
25
47
7
14
0
0
11
11
2
0
0
0
0
5
8
0
0
20
28
9
10
0
0
19
24
10
9
0
0
12
17
10
8
0
0
9
8
3
3
0
0
14
12
8
4
0
0
12
15
†
†
†
443+1238
†
420+75
†
†
†
190+2775
†
540+135
†
†
†
186+677
60+5
†
†
†
115+0
†
200+35
†
†
†
169+2058
†
200+75
†
†
†
188+1098
†
400+170
†
†
†
268+1824
†
440+50
†
†
†
208+401
†
340+10
†
†
†
200+1135
†
80+20
†
†
†
161+1154
Logarithmic
percentage
partial
optimality
0.00%
0.00%
18.84%
18.92%
97.30%
96.97%
0.00%
0.00%
55.53%
55.69%
92.58%
83.18%
0.00%
0.00%
15.94%
16.18%
99.73%
0.00%
0.00%
100.00%
100.00%
99.10%
96.39%
0.00%
0.00%
35.32%
40.50%
97.24%
93.46%
0.00%
0.00%
35.86%
35.26%
89.22%
88.92%
0.00%
0.00%
39.42%
13.20%
79.09%
76.53%
0.00%
0.00%
25.54%
25.43%
98.54%
97.53%
0.00%
0.00%
31.05%
31.08%
98.30%
98.57%
0.00%
0.00%
28.18%
30.63%
6
0
0
0
14
13
0
0
0
10
5
0
0
0
5
1000+140
†
†
†
826+6056
1000+310
†
†
†
848+6349
70+20
†
†
†
167+488
99.96%
0.00%
0.00%
0.00%
99.95%
99.92%
0.00%
0.00%
0.00%
99.80%
99.91%
0.00%
0.00%
0.00%
99.89%
#iterations
#iterations
Algorithm 1,2 TRWS
Percentage
excluded
labels
2.53%
1.99%
22.32%
22.34%
97.98%
98.25%
4.61%
2.74%
58.94%
58.98%
94.90%
87.09%
2.92%
2.04%
18.83%
18.91%
99.94%
2.85%
3.43%
100.00%
100.00%
99.63%
97.10%
6.06%
4.90%
41.15%
45.56%
98.69%
95.83%
3.76%
3.73%
39.98%
39.12%
93.23%
93.16%
3.24%
2.27%
39.67%
13.36%
85.64%
83.26%
2.99%
3.43%
31.15%
31.13%
99.47%
99.08%
6.38%
5.72%
27.97%
27.98%
98.93%
99.38%
5.09%
3.68%
26.37%
28.46%
99.97%
99.89%
43.55%
32.06%
99.95%
99.93%
99.90%
0.56%
0.37%
99.80%
99.94%
94.96%
4.27%
0.21%
99.89%
Instance
Algorithm
Time
needed
overall (s)
clownfish-small
Our-TRWS
Kovtun[24]
MQPBO-10
MQPBO
[43]-TRWS
Our-TRWS
Kovtun[24]
MQPBO-10
MQPBO
[43]-TRWS
Our-TRWS
Kovtun[24]
MQPBO-10
MQPBO
[43]-TRWS
Our-TRWS
Kovtun[24]
MQPBO-10
MQPBO
[43]-TRWS
Our-TRWS
Kovtun[24]
MQPBO-10
MQPBO
[43]-TRWS
Our-TRWS
Kovtun[24]
MQPBO-10
MQPBO
[43]-TRWS
Our-TRWS
Kovtun[24]
MQPBO-10
MQPBO
[43]-TRWS
Our-TRWS
Kovtun[24]
MQPBO-10
MQPBO
[43]-TRWS
Our-TRWS
Kovtun[24]
MQPBO-10
MQPBO
[43]-TRWS
1.72
1.00
536.00
41.00
151.98
1.87
1.00
577.00
33.00
677.08
0.57
0.00
37.00
3.00
31.28
1.28
1.00
607.00
31.00
13.75
2.48
1.00
510.00
19.00
846.27
1.21
0.00
193.00
13.00
15.67
18.77
1.00
591.00
16.00
799.08
0.79
0.00
7.00
1.00
46.20
1.35
1.00
528.00
39.00
311.54
Our-TRWS
MQPBO-10
MQPBO
[43]-TRWS
Our-TRWS
MQPBO-10
MQPBO
[43]-TRWS
286.40
1087.00
90.00
12726.45
320.00
646.00
97.00
14360.45
Our-TRWS
[43]-TRWS
Our-TRWS
[43]-TRWS
Our-TRWS
[43]-TRWS
231.97
3837.51
19.75
9277.99
108.73
14737.47
crops-small
fourcolors
lake-small
palm-small
penguin-small
pfau-small
snail
strawberry-glass-2-small
family-gm
pano-gm
ted-gm
tsu-gm
ven-gm
Time for
initial
solution (s)
Logarithmic
percentage
partial
optimality
Percentage
excluded
labels
80+10
†
†
†
223+610
120+5
†
†
†
260+3578
20+5
†
†
†
34+238
50+5
†
†
†
95+-95
160+10
†
†
†
291+4582
90+5
†
†
†
152+-152
950+470
†
†
†
654+10857
50+5
†
†
†
83+332
80+5
†
†
†
259+1721
>99.99%
0.00%
0.00%
0.00%
99.97%
100.00%
0.00%
0.00%
0.00%
99.00%
99.96%
0.00%
0.00%
0.00%
99.92%
100.00%
0.00%
0.00%
0.00%
100.00%
>99.99%
0.00%
0.00%
0.00%
98.20%
100.00%
0.00%
0.00%
0.00%
100.00%
89.43%
0.00%
0.00%
0.00%
10.43%
99.99%
0.00%
0.00%
0.00%
99.98%
100.00%
0.00%
0.00%
0.00%
99.31%
>99.99%
74.11%
15.83%
4.67%
99.97%
100.00%
64.70%
14.32%
0.71%
99.00%
99.97%
69.52%
0.00%
0.00%
99.92%
100.00%
74.87%
15.31%
6.65%
100.00%
>99.99%
68.65%
0.48%
0.00%
98.20%
100.00%
91.99%
1.42%
1.03%
100.00%
93.41%
5.59%
0.70%
0.00%
10.43%
99.99%
97.77%
77.91%
58.35%
99.98%
100.00%
54.99%
2.78%
0.00%
99.31%
1000+1265
†
†
1015+22483
1000+1105
†
†
911+11193
4.75%
0.00%
0.00%
4.41%
67.73%
0.00%
0.00%
27.55%
4.80%
4.41%
4.34%
4.41%
79.17%
28.06%
40.37%
27.55%
1000+715
689+10383
670+75
377+17421
1000+40
993+16592
67.27%
38.13%
99.91%
0.39%
0.01%
0.00%
72.05%
38.13%
99.94%
0.39%
0.02%
0.00%
#iterations
#iterations
Algorithm 1,2 TRWS
color-seg-n4
0.68
3
0.00
0
0.00
0
0.00
0
30.01
6
1.02
2
0.00
0
0.00
0
0.00
0
34.88
40
0.08
2
0.00
0
0.00
0
0.00
0
2.60
8
0.43
2
0.00
0
0.00
0
0.00
0
13.75
0
1.37
3
0.00
0
0.00
0
0.00
0
39.97
19
0.54
2
0.00
0
0.00
0
0.00
0
15.67
0
7.22
48
0.00
0
0.00
0
0.00
0
79.34
44
0.23
2
0.00
0
0.00
0
0.00
0
6.47
5
0.60
2
0.00
0
0.00
0
0.00
0
31.00
11
mrf-photomontage
93.08
77
0.00
0
0.00
0
1291.11
50
112.17
59
0.00
0
0.00
0
1871.14
33
mrf-stereo
72.67
119
436.30
28
14.67
10
267.55
54
94.44
9
1451.83
55
TABLE 7: Detailed experimental evaluation for Algorithm 1 utilising CPLEX [17] as a subsolver, denoted as Our-CPLEX, Algorithm 2
utilising TRW-S [20] as a subsolver, denoted as Our-TRWS, their counterparts from [43] denoted by [43]-CPLEX and [43]-TRWS and
MQPBO [19] run for one iteration with predefined label order, denoted by MQPBO, and run 10 iterations in 10 random label orders, denoted
by MQPBO-10.
20
| 8cs.DS
|
APPROXIMATING GEODESICS VIA RANDOM POINTS
arXiv:1711.06952v1 [math.PR] 19 Nov 2017
ERIK DAVIS AND SUNDER SETHURAMAN
Abstract. Given
a ‘cost’ functional F on paths γ in a domain D ⊂ Rd , in the
R
form F (γ) = 01 f (γ(t), γ̇(t))dt, it is of interest to approximate its minimum
cost and geodesic paths. Let X1 , . . . , Xn be points drawn independently from
D according to a distribution with a density. Form a random geometric graph
on the points where Xi and Xj are connected when 0 < |Xi − Xj | < , and
the length scale = n vanishes at a suitable rate.
For a general class of functionals F , associated to Finsler and other distances on D, using a probabilistic form of Gamma convergence, we show that
the minimum costs and geodesic paths, with respect to types of approximating
discrete ‘cost’ functionals, built from the random geometric graph, converge
almost surely in various senses to those corresponding to the continuum cost
F , as the number of sample points diverges. In particular, the geodesic path
convergence shown appears to be among the first results of its kind.
1. Introduction
Understanding the ‘shortest’ or geodesic paths between points in a medium is
an intrinsic concern in diverse applied problems, from ‘optimal routing’ in networks
and disordered materials to ‘identifying manifold structure in large data sets’, as
well as in studies of probabilistic Zd -percolation models, since the seminal paper
of [5] (cf. recent survey [4]). See also [17], [18], [19], [20], [21], [22] which consider
percolation in Rd continuum settings.
There are sometimes abstract formulas for the geodesics, from the calculus of
variations, or other differential equation approaches. For instance, with respect to
a patch of a Riemannian manifold (M, g), with M ⊂ Rd and tensor field g(·), it is
known that the distance function U (·) = d(x, ·), for fixed x, is a viscosity solution
of the Eikonal equation k∇U (y)kg(y)−1 = 1 for y 6= x, with boundary condition
p
U (x) = 0. Here, kvkA = hv, Avi, where h·, ·i is the standard innerproduct on
Rd . Then, a geodesic γ connecting x and z may be recovered from U by solving a
‘descent’ equation, γ̇(t) = −η(t)g −1 (γ(t))∇U (γ(t)), where η(t) is a scalar function
controlling the speed.
On the other hand, computing numerically the distances and geodesics may be a
complicated issue. One of the standard approaches is the ‘fast marching method’ to
approximate the distance U , by solving the Eikonal equation on a regular grid of n
points. This method has been extended in a variety of ways, including with respect
to triangulated domains, as well as irregular samples {x1 , . . . , xn } of an Euclidean
submanifold (cf. [30], [24]). See also [28] in the above contexts for a review.
2010 Mathematics Subject Classification. 60D05, 58E10, 62-07, 49J55, 49J45, 53C22, 05C82.
Key words and phrases. geodesic, shortest path, distance, consistency, random geometric
graph, Gamma convergence, scaling limit, Finsler .
1
2
ERIK DAVIS AND SUNDER SETHURAMAN
Alternatively, variants of Dijkstra’s or ‘heat flow’ methods, on graphs approximating the space are sometimes used. In Dijkstra’s algorithm, distances and shortest paths are found by successively computing optimal routes to nearest-neighbor
edges. In ‘heat flow’ methods, geodesic distances can be found in terms of the small
time asymptotics of a heat kernel on the space. For instance, see [10], [11], [12],
[13], [33].
Another idea has been to collect a random sample Xn of n points from a manifold embedded in Rd , put a network structure on these points, say in terms of a
-random geometric or k-nearest neighbor graph, and then approximate the ‘continuum’ geodesics lengths by lengths of ‘discrete’ geodesic paths found in this network.
Presumably, under assumptions on how the points are sampled and how the random graphs are formed, as the number of points diverge, these ‘discrete’ distances
should converge almost surely to the ‘continuum’ shortest path lengths. Such a
statistical consistency result is fundamental in ‘manifold learning’ [6]. For instance,
the popular ISOMAP procedure [32], [7] is based on these notions to elicit manifold
structure in data sets.
More specifically, let D be a subset of Rd corresponding to a patch of the manifold, and consider a ‘kernel’ f (x, v) : D × Rd → [0, ∞). Define the f -cost of a
R1
path γ(t) : [0, 1] → D from γ(0) = a to γ(1) = b as F (γ) = 0 f (γ(t), γ̇(t))dt. The
f -distance from a to b is then the infimum of such costs over paths γ. For example,
if f (x, v) = |v|p , the f -distance is |b − a|p , the pth power of the Euclidean distance.
With respect to a class of functions f and samples drawn from a distribution on
the D with density ρ, papers [7], [29], and [3] address, among other results, how
= n and k = kn should decrease and increase respectively so that various concentration type bounds between types of discrete and continuum optimal distances
hold with high probability, leading to consistent estimates.
For instance, in [29], for n -random graphs and smooth ρ, certain density dependent estimators of continuum distances were considered, where f (x, v) = h(ρ(x))|v|
and h(y) is decreasing, smooth, constant for |y| small, and bounded away from 0.
This work extends [7], which considered f (x, v) = |v| and uniformly distributed
samples. On the other hand, in [3], among other results, on kn -nearest-neighbor
graphs, continuum distances, where f (x, v) = h(ρ(x))|v| and h is increasing, Lipschitz, and bounded away from 0, were approximated (see also [15]).
In these contexts, the purpose of this article is twofold. First, we identify a
general class of f -distances for which different associated discrete distances, formed
from random n -random geometric graphs on a domain D ⊂ Rd , converge almost
surely to them. Second, we describe when the associated discrete geodesic paths
converge almost surely, in uniform and Hausdorff norms, to continuum f -distance
geodesic paths, a type of consistency which appears to be among the first contributions of this kind. The main results are Theorems 2.1, 2.2, 2.4, 2.7, 2.8, and
Corollary 2.3. See Section 2 and Subsection 2.3 for precise statements and related
remarks.
We consider the following three different discrete costs. The first, d1 , optimizes
on paths γ, starting and ending at a and b respectively, linearly interpolated between
points in Xn ∪ {a, b}, where consecutive points are within n of each other, and the
time to traverse each link is the same. The second, d2 , optimizes with respect
to ‘quasinormal’ interpolations between the points, using however the f -geodesic
paths. The third, d3 , does not interpolate at all, and optimizes a ‘Riemann sum’
APPROXIMATING GEODESICS VIA RANDOM POINTS
3
Pm−1
1
cost m
i=0 f (vi , m(vi+1 − vi )) where m is the number of edges in the discrete
path {v0 , . . . , vm } ⊂ Xn ∪ {a, b}. We note, discrete distances d2 and d3 , in the
setting f (x, v) = |v| were introduced in [7], and density dependent versions were
used in the results in [3] and [29]. The discrete distance d1 , although natural, seems
not well considered in the literature.
The conditions we impose on f include p-homogeneity in v for p ≥ 1, convexity
and an ellipticity condition with respect to v, and a smoothness assumption away
from v 6= 0. Such conditions include a large class of kernels f associated to Finsler
spaces, as well as those kernels considered in [29] and [7], with respect to n -random
graphs. The domain D ⊂ Rd is assumed to be bounded and convex. Also, we
assume that the rate of decrease of n is such that the graph on Xn is connected
for all large n.
While a main contribution of the article is to provide a general setting in which
the ‘discrete to continuum’ convergences hold, we remark our proof method is quite
different from that in the literature, where specific features of f , such as f (x, v) =
|v| in [7], are important in estimation of distances, not easily generalized. We
give a probabilistic form of ‘Gamma convergence’ to derive the almost sure limits,
which may be of interest itself. This method involves showing ‘liminf’, ‘limsup’
and ‘compactness’ elements, as in the analysis context, but here on appropriate
probability 1 sets. Part of the output of the technique, beyond giving convergence
of the distances, is that it yields convergence of the minimizing discrete paths to
continuum geodesics in various senses.
The f -costs share different properties depending on if p = 1 or p > 1, and also
when d ≥ 2 or d = 1. For instance, the f -cost is invariant to reparametrization of
the path exactly when p = 1. Also, when p > 1, the form of the f -cost may be seen
to be coercive on the modulus of γ, not the case when p = 1. In fact, the p = 1
case is the most troublesome, and more assumptions on f and n are required in
Theorems 2.7 and 2.8 to deal with the ‘linear’ path cost d1 and ‘Riemann’ cost d3 ,
which are ‘rougher’ than the ‘quasinormal’ cost d2 .
At the same time, in d = 1, in contrast to d ≥ 2, all paths lie in the interval
[a, b] ⊂ R. When also p = 1, the problem is somewhat degenerate: By invariance
to reparametrization, the costs d1 and d2 turn out to be nonrandom and to reduce
Rb
to the integral a f (s, 1)ds. Also, the cost d3 is a Riemann sum which converges to
this integral.
Finally, we comment on a difference in viewpoint with respect to results in
continuum percolation. The ‘Riemann sum’ cost considered here seems related to
but is different than the cost optimized in the works [18], [19]. There, for p > 1,
one optimizes the cost of a path {w0 , . . . , wm }, along random points, from the
Pm−1
origin 0 to nx, for x ∈ Rd , given by i=0 |wi+1 − wi |p , and infers a scaled distance
d(x) = c(d, p)|x|, in law of large numbers scale n, where the proportionality constant
c(d, p) is not explicit. In contrast, however, in this article, given already an integral
f -distance, the viewpoint is to optimize costs of paths of length order 1 (not n as
in [18], [19]), where the length scale between points is being scaled of order n ,
and then to recover the f -distance in the limit. We note also another difference:
When f (x, v) = |v|p , as remarked above, the f -distance from the origin to x is |x|p ,
instead of ∼ |x| as in the continuum percolation studies.
In Section 2, the setting, assumptions and results are given with respect to three
types of discrete costs. In Section 3, proofs of Theorems 2.1 and 2.2 and Corollary
4
ERIK DAVIS AND SUNDER SETHURAMAN
(a) F -minimizing path, with level sets of w in- (b) n -graph, on n = 400 uniform points with
n = (1/400)0.3
dicated.
Figure 1. Continuum geodesic and n -graph for f (x, v) = w(x)|v| on the
domain D = [−1, 1]×[−1, 1], where w(x) = 1+8 exp(−2(x1 −1/2)2 +xy+2y 2 ),
and a = (−0.8, −0.8), b = (0.8, 0.8)
2.3 on the ‘interpolating’ costs are given. In Section 4, proofs of Theorems 2.4, 2.7,
and 2.8, with respect to ‘Riemann’ costs, and ‘interpolating’ costs when p = 1, are
given. In Section 5, some technical results, used in the course of the main proofs,
are collected.
2. Setting and Results
We now introduce the setting of the problem, and ‘standing assumptions’, which
hold throughout the article.
For d ≥ 1, we will be working on a subset D ⊂ Rd ,
which is the closure of an open, bounded, convex domain.
(2.1)
Therefore, D is a Lipschitz domain (cf. Corollary 9.1.2 in [2], Section 1.1.8 in [23]).
Consider points a, b ∈ D and let Ω(a, b) denote the space of Lipschitz paths
γ : [0, 1] → D with γ(0) = a and γ(1) = b. Given f : D × Rd → [0, ∞), we define
the cost F : Ω(a, b) → [0, ∞) by
Z 1
F (γ) =
f (γ(t), γ̇(t)) dt,
0
and associated optimal cost
df (a, b) =
inf
F (γ).
(2.2)
γ∈Ω(a,b)
We will make the following assumptions on the integrand f :
(A0) f is continuous on D × Rd , and C 1 on D × (Rd \ {0}),
(A1) f (x, v) is convex in v,
(A2) there exists p ≥ 1 such that f (x, v) is p-homogenous in v,
f (x, λv) = λp f (x, v) for λ > 0,
(2.3)
APPROXIMATING GEODESICS VIA RANDOM POINTS
5
(A3) there exist constants m1 , m2 > 0 such that
m1 |v|p ≤ f (x, v) ≤ m2 |v|p for all x ∈ D.
(2.4)
We remark, when p > 1 and p-homogenity (A2) holds, that f may be extended
to a C 1 function on D × Rd .
Part of the reasoning for the assumptions (A0)-(A3) is that they include, for
p ≥ 1, the familiar kernel f (x, v) = |v|p , for which, when p = 1, F (γ) is the
arclength of the path γ and df (a, b) is the length of the line segment from a to b.
Also, under these assumptions on f , it is known that the infimum in (2.2) is
attained at a path in Ω(a, b), perhaps nonuniquely (see Proposition 5.2 of the appendix). In addition, we remark, when p = 1, under additional differentiability
assumptions, df represents a Finsler distance (cf. [25], [31] and references therein).
When p = 1, the cost has an interesting scaling property: By 1-homogeneity of
f , the cost F is invariant under smooth reparameterization of paths. That is, given
a path γ ∈ Ω(a, b) and smooth, increasing s : [0, 1] → [0, 1], with s(0) = 0 and
s(1) = 1, one has F (γ̃) = F (γ) where γ̃(t) = γ(s(t)).
This property allows to deduce, when p = 1, that df satisfies the triangle property (not guaranteed when p > 1): Let γ1 be a path from u to w, and γ2 be a path
from w to z. Write
Z 1
Z 1
f (γ1 (t), γ̇1 (t))dt +
f (γ2 (t), γ̇2 (t))dt
0
0
Z
1/2
=
Z
f (γ1 (2s), 2γ̇1 (2s))ds +
0
Z
=
1/2
f (γ2 (2s), 2γ̇2 (2s))ds
0
1
f (γ3 (t), γ̇3 (t))dt,
(2.5)
0
where γ3 is a path from u to z, following γ1 (2·) and γ2 (2·) on time intervals [0, 1/2]
and [1/2, 1] respectively. Optimizing over γ1 , γ2 and γ3 gives df (u, w) + df (w, z) ≥
df (u, z).
We now construct a random geometric graph on D through which approximations of df and its geodesics will be made. Let {Xi , X2 , . . .} ⊂ D be a sequence
of independent points, identically distributed according to a distribution ν with
probability density ρ. For each n ∈ N, let Xn = {X1 , . . . , Xn } and fix a length
scale n > 0. With respect to a realization {Xi }, we define a graph Gn (a, b), on
the vertex set Xn ∪ {a, b}, by connecting an edge between u, v in Xn ∪ {a, b} iff
0 < |u − v| < n , where | · | refers to the Euclidean distance in Rd .
For u, v ∈ Xn ∪ {a, b}, we say that a finite sequence (v0 , v1 , . . . , vm ) of vertices is
a path with m-steps from u to v in Gn (a, b) if v0 = u, vm = v, and there is an edge
from vi to vi+1 for 0 ≤ i < m. Let Vn (a, b) denote the set of paths from a to b in
Gn (a, b).
We will assume a certain decay rate on n , namely that limn↑∞ n = 0 and
lim sup
n→∞
(log n)1/d 1
= 0.
n1/d n
(2.6)
Under this type of decay rate, almost surely, for all large n and a, b ∈ D, points a, b
will be connected by a path in the graph Gn (a, b), in other words, Vn (a, b) will be
nonempty. Indeed, under this rate, the degree of a point in the graph will diverge
to infinity. See Proposition 5.1 in the appendix, and remarks in Section 2.3.
6
ERIK DAVIS AND SUNDER SETHURAMAN
We will also assume that the underlying probability density ρ is uniformly
bounded, that is, there exists a constant c > 0 such that
c ≤ ρ(x) ≤ c−1
for all x ∈ D.
(2.7)
See Figure 1, parts (a) and (b), which depict a geodesic path with respect to a
cost F , and an n -random graph.
‘Standing assumptions’. To summarize, the assumptions, dimension d ≥ 1, (2.1)
on D, items (A0)-(A3) on f when p ≥ 1, decay rate (2.6) on n , and density bound
(2.7) on ρ, denoted as the ‘standing assumptions’, will hold throughout the article.
In the next two Subsections, we present results on approximation of df (a, b) and
its geodesics with respect to two types of schemes, where approximating costs are
built (1) in terms of ‘interpolations’ of points in Vn (a, b) and also (2) in terms of
‘Riemann sums’.
2.1. Interpolating costs. We introduce two types of discrete costs based on ‘linear’ and ‘quasinormal’ paths.
Linear interpolations. With respect to a realization {Xi }, for u, v ∈ D, let
lu,v : [0, 1] → D denote the constant-speed linear path from a to b, given by
lu,v (t) = (1 − t)u + tv.
Consider now v = (v0 , v1 , . . . , vm ) ∈ Vn (a, b). We define lv ∈ Ω(a, b) to be the
concatenation of the linear segments {lvi−1 ,vi }m
i=1 , where each segment is traversed
in the same time 1/m. More precisely, for i/m ≤ t ≤ (i + 1)/m, define
lv (t) = lvi ,vi+1 (mt − i),
and note that the resulting piecewise linear path is in Ω(a, b).
Define now a subset Ωln (a, b) of Ω(a, b) by
Ωln (a, b) = {lv |v ∈ Vn (a, b)} ,
and define the (random) discrete cost Ln : Ωln (a, b) → [0, ∞] by
Ln (γ) = F (γ) for γ ∈ Ωln (a, b).
In other words, Ln is the restriction of F to Ωln (a, b), noting the p-homogenity of
f , taking form
m−1
X Z (i+1)/m
Ln (lv ) =
f (lvi ,vi+1 (mt − i), m(vi+1 − vi ))dt
i=0
= mp−1
i/m
m−1
XZ 1
i=0
f (lvi ,vi+1 (t), vi+1 − vi )dt.
(2.8)
0
Quasinormal interpolations. Define now a different discrete cost which may
nonlinearly interpolate among points in paths of Vn (a, b). We say that a Lipschitz
path γ is quasinormal with respect to f if there exists a c > 0 such that
f (γ(t), γ̇(t)) = c for a.e. t ∈ [0, 1].
It is known, under the ‘standard assumptions’ on f (see Proposition 5.2) that, for
u, v ∈ D, there exists a quasinormal path γ : [0, 1] → D, with γ(0) = u, γ(1) = v,
R1
which is optimal, df (u, v) = 0 f (γ(t), γ̇(t)) dt. For what follows, when we refer
APPROXIMATING GEODESICS VIA RANDOM POINTS
7
to a ‘quasinormal’ path connecting u and v, we mean such a fixed optimal path
denoted by γu,v .
Given a path v = (v0 , . . . , vm ) ∈ Vn (a, b), let γv ∈ Ω(a, b) denote the concatenation of {γvi−1 ,vi }m
i=1 , where each segment uses the same time 1/m. More precisely,
for i/m ≤ t ≤ (i + 1)/m, define
γv (t) = γvi ,vi+1 (mt − i).
As with piecewise linear functions, define the subset Ωγn (a, b) of Ω(a, b) by
Ωγn (a, b) = {γv |v ∈ Vn (a, b)} .
Let Gn : Ωγn (a, b) → R denote the restriction of F to Ωγn (a, b).
Then, with respect to a path γ = γv ∈ Ωγn (a, b), by the p-homogenity of f , we
evaluate that
Z 1
Gn (γ) =
f (γ(t), γ̇(t)) dt
(2.9)
0
=
m Z
X
i=1
i/m
f (γvi−1 ,vi (mt − i), mγ̇vi−1 ,vi (mt − i)) dt
(i−1)/m
= mp−1
m Z
X
i=1
1
f (γvi−1 ,vi (t), γ̇vi−1 ,vi (t)) dt = mp−1
0
m
X
df (vi−1 , vi ).
i=1
Further, by p-homogeneity of f and optimality of {γvi ,vi+1 }m
i=1 , the segments of
γ = γv are also optimal, in the sense that
Z (i+1)/m
Z 1
f (γ(t), γ̇(t)) dt = mp−1
f (γvi ,vi+1 (t), γ̇vi ,vi+1 (t))dt
i/m
0
= inf mp−1
γ
e
Z
0
1
f (e
γ (t), γ
e˙ (t)) dt = inf
γ
b
Z
(i+1)/m
i/m
f (b
γ (t), γ
b˙ (t))dt, (2.10)
where the infima are over Lipschitz paths γ
e : [0, 1] → D and γ
b : [i/m, (i+1)/m] → D
with γ
e(0) = vi , γ
e(1) = vi+1 , γ
b(i/m) = vi and γ
b((i + 1)/m) = vi+1 .
Relations between Gn and Ln . At this point, we remark there are kernels f for
which Gn = Ln , namely those such that linear segments are in fact quasinormal
geodesics. An example is f (x, v) = |v|. Identifying these kernels is a question
with a long history, going back to Hilbert, whose 4th problem paraphrased asks for
which geometries are the geodesics straight lines (cf. surveys [25], [26]). Hamel’s
criterion, namely ∂xi ∂vj f = ∂xj ∂vi f for 1 ≤ i, j ≤ d, is a well-known solution to
this question (see [14], [25], [26] and references therein).
We also note, as mentioned in the introduction, that the case d = p = 1 is ‘degenerate’ in that min Gn and min Ln are not random. Indeed, let γv ∈ arg min Gn and
suppose v = (v0 , . . . , vm ) ∈ Vn (a, b). We observe that γv must be nondecreasing, as
otherwise, one could build a smaller cost path, from parts of γv using invariance to
reparametrization, violating optimality of γv . In particular, γ̇v ≥ 0 and vi < vi+1
for 0 ≤ i ≤ m − 1. Then,
Z b
m−1
m−1
X Z (i+1)/m
X Z vi+1
Gn (γv ) =
f (γv (t), γ̇v (t))dt =
f (s, 1)ds =
f (s, 1)ds,
i=0
i/m
i=0
vi
a
using the 1-homogeneity of f and changing variables. The same argument yields
Rb
that min Ln = a f (s, 1)ds. We do not consider this ‘degenerate’ case further.
8
ERIK DAVIS AND SUNDER SETHURAMAN
The first result is for linearly interpolated paths.
Theorem 2.1. Suppose that p > 1. With respect to realizations {Xi } in a probability 1 set, the following holds. The minimum values of the costs Ln converge to
the minimum of F ,
lim
min
n→∞ γ∈Ωln (a,b)
Ln (γ) =
min F (γ).
γ∈Ω(a,b)
Moreover, consider a sequence of optimal paths γn ∈ arg min Ln . Any subsequence of {γn } has a further subsequence that converges uniformly to a limit path
γ ∈ arg min F ,
lim sup |γnk (t) − γ(t)| = 0.
k→∞ 0≤t≤1
In addition, if γ is the unique minimizer of F , then the whole sequence γn converges uniformly to γ.
The case d ≥ 2 and p = 1 requires further development, and is addressed with a
few more assumptions in Theorem 2.8.
We now address quasinormal interpolations.
Theorem 2.2. Suppose that either (1) p > 1 or (2) d ≥ 2 and p = 1. Then,
with respect to realizations {Xi } in a probability 1 set, the following holds. The
minimum values of the energies Gn converge to the minimum of F ,
lim
min
n→∞ γ∈Ωγ
n (a,b)
Gn (γ) =
min F (γ).
γ∈Ω(a,b)
Moreover, consider a sequence of optimal paths γn ∈ arg min Gn . Any subsequence of {γn } has a further subsequence that converges uniformly to a limit path
γ ∈ arg min F ,
lim sup |γnk (t) − γ(t)| = 0.
k→∞ 0≤t≤1
In addition, if γ is the unique minimzer of F , then the whole sequence γn converges uniformly to γ.
We remark, when d ≥ 2 and p = 1, that there is a certain ambiguity in the
results of Theorem 2.2, due to the invariance of F under reparametrization of
paths. In this case, there is no unique
R 1minimizer of F . Consider for example the
case where f (x, v) = |v| and F (γ) = 0 |γ̇(t)|dt. Any minimizer of this functional
is a parameterization of a line, but of course such minimizers are not unique.
One way to address this is to formulate a certain Hausdorff convergence with
respect to images of the paths. Given γ ∈ Ω(a, b), we denote the image of γ by
Sγ = {γ(t) | 0 ≤ t ≤ 1} .
Consider the Hausdorff metric dhaus , defined on compact subsets A, B of D by
dhaus (A, B) = max{sup inf d(x, y), sup inf d(x, y)}.
x∈A y∈B
y∈B x∈A
Corollary 2.3. Suppose that either (1) d ≥ 2 and p = 1 or (2) p > 1. Consider paths {γv(n) }, for all large n either in form γv(n) ∈ arg min Gn , or γv(n) ∈
arg min Ln .
Then, with respect to realizations {Xi } in a probability 1 set, any subsequence
of {v(n) } has a further subsequence which converges in the Hausdorff sense to Sγ ,
where γ ∈ arg min F is an optimal path.
APPROXIMATING GEODESICS VIA RANDOM POINTS
9
Figure 2. H400 -minimizing discrete path in the setting of Figure 1, linearly
interpolated for visual clarity.
Moreover, if F has a unique (up to reparametrization) minimizer γ, then the
whole sequence converges,
lim dhaus (v(n) , Sγ ) = 0.
n→∞
2.2. Riemann sum costs and p = 1-linear interpolating costs. We first introduce a cost which requires knowledge of f only on discrete points and, as a
consequence, more ‘applicable’. At the end of the subsection, we return to linear
interpolated costs when p = 1.
Define Hn : Vn (a, b) → R, for v = (v0 , v1 , . . . , vm ), by
m
1 X
Hn (v) =
f (vi , m(vi+1 − vi )).
m i=1
(2.11)
The functional Hn is, in a sense, a ‘Riemann sum’ approximation to Ln and Gn ,
and therefore its behavior, and the behavior of its minimizing paths, should be
similar to that of Ln and Gn . See Figure 2 for an example of an optimal Hn path.
We make this intuition rigorous by establishing variants of Theorems 2.1 and 2.2
with respect to the cost Hn . Given the ‘rougher’ nature of Hn , however, additional
assumptions on f and n , beyond those in the ‘standing assumptions’, will be helpful
in this regard. As in the previous Subsection, our results differ between the two
cases p = 1 and p > 1.
Define the following smoothness condition:
(Lip) There exists a c such that for all x, y ∈ D and v ∈ Rd we have
|f (x, v) − f (y, v)| ≤ c|x − y||v|p .
We note when f satisfies the homogeneity condition (2.3), and ∇x f (x, v) is uniformly bounded on D × {y : |y| = 1}, that (Lip) holds.
10
ERIK DAVIS AND SUNDER SETHURAMAN
We now consider the behavior of Hn when p > 1. The analogue to Theorem 2.1
and Corollary 2.3 in this setting is the following.
Theorem 2.4. Suppose p > 1, and that f in addition satisfies (Lip). With respect
to realizations {Xi } in a probability 1 set, the minimum values of the energies Hn
converge to the minimum of F ,
lim
min
n→∞ v∈Vn (a,b)
Hn (v) =
min F (γ).
γ∈Ω(a,b)
Further, consider a sequence of optimal discrete paths w(n) ∈ arg min Hn , and
their linear interpolations {lw(n) }. Then, for any subsequence of {lw(n) } and correspondingly of {w(n) }, there is a further subsequence of the linear paths which
converges uniformly to a limit path γ ∈ arg min F , and of the discrete paths in the
Hausdorff sense to Sγ .
If F has a unique minimizer γ, the whole sequence of linear paths converges
uniformly to γ, and the whole sequence of discrete paths converges in the Hausdorff
sense to Sγ .
We will need to impose further assumptions on the integrand f to state results
in the case p = 1. See below for examples of f satisfying these conditions, and also
Subsection 2.3 for further comments.
(Hilb) We say that f satisfies the ‘Hilbert condition’ if, for each x,
Z 1
inf
f (x, γ̇(t))dt = f (x, b − a),
γ∈Ω(a,b)
0
that is, straight lines are geodesics for the kernel f (x, ·).
(TrIneq) We say f satisfies the ‘triangle inequality’ if, for each x,
f (x, v − w) ≤ f (x, v − u) + f (x, u − w)
for all u, v, w ∈ Rd .
(Pythag) Let α > 1. Consider points u, v, w where |uw|, |vw|, |uv| < η for an η < 1.
Suppose there is a constant c such that, for 0 < r < 1,
– dist(w, line(u, v)) ≥ r, and
– |uv| ≤ cr1/α .
Then, we say f satisfies the ‘Pythagoras α-condition’ if there is a constant
C = C(α, f, c) such that, for all x,
f (x, w − u) + f (x, v − w) ≥ f (x, v − u) + Crα .
Here, line(u, v) is the line segment between u and v.
Here, in the statement of (Hilb), the kernel function, for fixed x, is only a function
of v. The following lemma is a case of the Hamel’s criterion discussed in the previous
Subsection.
Lemma 2.5. Given the ‘standing assumptions’, suppose also, for fixed x ∈ D,
that v 7→ f (x, v) is C 2 on Rd \ {0} with positive definite Hessian. Then, (Hilb) is
satisfied.
2
Proof. Fix an x0 ∈ D. There
R 1 is a quasinormal minimizer γ ∈ C where both
c = f (x0 , γ̇(t)) = inf γ∈Ω(a,b) 0 f (x0 , γ̇(t))dt and
Z 1
2
2
c = f (x0 , γ̇(t)) = inf
f 2 (x0 , γ̇(t))dt
γ∈Ω(a,b)
0
APPROXIMATING GEODESICS VIA RANDOM POINTS
11
for a.e. 0 ≤ t ≤ 1 (cf. Prop. 5.25 in [8]). Let g(v) = f 2 (x0 , v). Then, γ satisfies the
d
Euler-Lagrange equation dt
∇v g(γ̇(t)) = ∇x g(γ̇(t)) = 0. In other words, H γ̈(t) = 0,
where H denotes the Hessian of g. By assumption, H is positive definite. Hence,
γ̈(t) ≡ 0, and so γ is a parametrization of a straight line.
An example of a class of kernels f satisfying the ‘standing assumptions’ and the
additional conditions above is given in the following result. Recall h·, ·i denotes the
Euclidean inner product on Rd .
Lemma 2.6. Let x 7→ M = M (x) be a C 1 , strictly elliptic matrix-valued function
on D. The kernel f (x, v) = hv, M (x)vi1/2 satisfies the ‘standing assumptions’, and
also (Lip), (Hilb), (TrIneq) and (Pythag) for all α > 1.
Proof. The kernel clearly satisfies the ‘standing assumptions’ and (Lip). Next, for
fixed x, the map v 7→ f (x, v) satisfies the conditions of Lemma 2.5, and so satisfies
(Hilb). Also, v 7→ f (x, v) trivially satisfies (TrIneq).
We show (Pythag) in the case f (x, v) = h(x)|v|, that is M (x) = h2 (x)Id, as the
notation is easier and all the ideas carry over to the more general case. Consider a
right triangle joining u, w, z where z is on the line through u, v (cf. Figure 3).
Figure 3. Geometric argument used in proof of Lemma 2.6 with respect to (Pythag).
If z is not on the line segment connecting u and v, then either (a) |uw| ≥
|uv| and |wv| ≥ r ≥ rα or (b) |wv| ≥ |uv| and |uw| ≥ r ≥ rα . In either case,
h(x)[|uw| + |wv|] ≥ h(x)|uv| + m1 rα and (Pythag) is satisfied.
Suppose now z is on the line segment connecting u and v. In the triangle, uw
is the hypotenuse, and wz and uz are the legs, such that |uw|2 = |wz|2 + |uz|2 .
Hence, as |uw| < 1, all the lengths are less than 1. We are given that |wz| ≥ r and
|uz| ≤ |uv| ≤ cr1/α . Then, as r ≤ |wz| < 1 and 2 < min{2α, α + 1/α}, we have
|uw|2
= |uz|2 + |wz|2
≥
|uz|2 + (1/9)r2α + 2(1/3)rα+1/α
|uz|2 + (9 max{c, 1}2 )−1 r2α + 2(3 max{c, 1})−1 |uz|rα
2
= |uz| + (3 max{c, 1})−1 rα ,
≥
and so |uw| ≥ |uz| + (3 max{c, 1})−1 rα .
12
ERIK DAVIS AND SUNDER SETHURAMAN
A similar inequality, |wv| ≥ |vz| + (3 max{c, 1})−1 rα , holds with the same argument. Hence, h(x)|uw| + h(x)|wv| ≥ h(x)|uv| + C(m1 , c)rα .
We will also need to limit the decay properties of n for the next result; see
Subsection 2.3 for comments on this limitation. Namely, we will suppose that n is
in form n = n−δ where
δ > max{[(2 − α2 )η + d]−1 , [(α(d − 1) + 1]−1 },
(2.12)
√
for an 0 < η < 1 and 1 < α < 2.
We note that condition (2.6), when n is in form n = n−δ , yields that δ < 1/d.
However, when d ≥ 2, we have max{[(2 − α2 )η + d]−1 , [(α(d − 1) + 1]−1 } < 1/d,
and so (2.12), in conjuction with (2.6), limits δ to an interval.
Theorem 2.7. Suppose d ≥ 2 and p = 1, and that f also satisfies (Lip) and (Hilb).
With respect to realizations {Xn }n≥1 in a probability 1 set, the minimum values of
the cost Hn converge to the minimum of F ,
lim
min
n→∞ v∈Vn (a,b)
Hn (v) =
min F (γ).
γ∈Ω(a,b)
Moreover,
√ suppose now that f in addition satisfies (TrIneq) and (Pythag) for an
1 < α < 2, and that n satisfies (2.12).
Consider a sequence of optimal discrete paths v(n) ∈ arg min Hn , and their linear
interpolations {lv(n) }. Then, for any subsequence of {lv(n) } and so of {v(n) }, there
is a further subsequence of the linear paths which converges uniformly to a limit
path γ ∈ arg min F , and of the discrete paths in the Hausdorff sense to Sγ .
If F has a unique (up to reparametrization) minimizer γ, the whole sequence of
discrete paths converges, limn→∞ dhaus (v(n) , Sγ ) = 0.
See Figure 2 for an example of an Hn -cost geodesic path.
As noted in the introduction, when d = p = 1, Hn (v) is a certain Riemann
sum. Let w ∈ arg min Hn , and observe by optimality that w = (w0 , . . . , wm ) ∈
Vn (a, b) satisfies wi < wi+1 for 0 ≤ i ≤ m − 1. Hence, by 1-homogenity of f ,
Pm−1
Hn (w) = i=0 f (wi , 1)|wi+1 − wi |, and Hn (w) strongly approximates the integral
Rb
f (s, 1)ds, given that the partition length max |wi+1 − wi | ≤ n → 0. For this
a
reason, this case is not included in the above theorem.
Linear interpolating costs when p = 1. Having now introduced (Lip), (Hilb),
(TrIneq) and (Pythag), we address the case d ≥ 2 and p = 1 with respect to
the cost Ln .
Theorem 2.8. Suppose d ≥ 2 and p = 1, and that f also satisfies (Lip) and (Hilb).
With respect to realizations {Xn }n≥1 in a probability 1 set, the minimum values of
the cost Ln converge to the minimum of F ,
lim
min
n→∞ γ∈Ωln (a,b)
Ln (γ) =
min F (γ).
γ∈Ω(a,b)
Moreover,
√ suppose now that f in addition satisfies (TrIneq) and (Pythag) for an
1 < α < 2, and that n satisfies (2.12).
Consider a sequence of optimal paths lv(n) ∈ arg min Ln . Then, for any subsequence of {lv(n) } and so of the discrete paths {v(n) }, there is a further subsequence
of the linear paths which converges uniformly to a limit path γ ∈ arg min F , and of
the discrete paths in the Hausdorff sense to Sγ .
APPROXIMATING GEODESICS VIA RANDOM POINTS
13
If F has a unique (up to reparametrization) minimizer γ, the whole sequence of
discrete paths converges, limn→∞ dhaus (v(n) , Sγ ) = 0.
2.3. Remarks. We make several comments about the assumptions and related
issues.
1. Domain. The requirements that D should be closed and connected are needed
for the quasinormal path results in [8] and [16] to hold. Also, the proof of Proposition 5.1, on the maximum distance to a nearest neighbor vertex, requires that
the domain boundary should be Lipschitz, true for convex domains. The convexity of the domain also ensures that all the linearly interpolated paths are within
the domain, and allows comparison with quasinormal ones, which by definition are
constrained in the domain, as in the proof of the ‘limsup’ inequality, Lemma 3.5. In
addition, a bound on the domain allows the Arzela-Ascoli equicontinuity criterion
to be applied in the compactness property, Lemma 3.7.
2. Ellipticity of ρ. The bound on ρ is useful to compare ν to the uniform distribution in the nearest-neighbor map result, Proposition 5.1, as well as in bounding
the number of points in certain sets in Lemma 4.8. We note, as our approximating
costs, Ln , Gn , Hn , do not involve density estimators, our results do not depend on
the specifics of ρ, unlike for ‘density based distances’ discussed in [29].
3. Decay of n (2.6). Intuitively, the rate n cannot vanish too quickly, as then
the graph may be disconnected with respect to a postive set of realizations {Xi }.
However, the estimate in (2.6) ensures that the graph Gn (a, b) is connected for all
large n almost surely–see Proposition 5.1. This is a version of the ‘δ-sampling’
condition in [7], and is related to connectivity estimates in continuum percolation
[27]. Moreover, we note, the prescribed rate yields in fact that any vertex Xi will
have degree tending to infinity as n grows, as long as ρ is elliptic: One calculates
that the mean number of points in the n ball around Xi is of order ndn which
grows faster than log(n).
4. Assumptions (A0)-(A3) on f . These are somewhat standard assumptions to
treat parametric variational integrals such as F (cf. [8] and [16]), which include the
basic case f (x, v) = |v|p .
5. Assumption on p. The assumption p ≥ 1 is useful to show existence of
quasinormal paths in Proposition 5.2, and compactness of minimizers. The case
p < 1 is more problematic in this sense and not discussed here.
6. Extra assumptions in Theorems 2.7 and 2.8. The main difficulty is in showing
compactness of optimal Hn and Ln paths when p = 1. With respect to Theorem
2.4, when p > 1, the form of the cost allows a Holder’s inequality argument to
deduce equicontinuity of the paths, from which compactness follows using AscoliArzela’s theorem. However, there is no such coercivity when p = 1. Yet, with the
additional assumptions, one can approximate a geodesic locally by straight lines.
Several geometric estimates on the number of points in small windows around these
straight lines are needed to ensure accuracy of the approximation, for which the
upperbound on n in (2.12) is useful.
7. Unique minimizers of F . Given that our results achieve their strongest form
when arg min F consists of a unique minimizing path, perhaps up to reparametrization, we comment on this possibility. Under suitable smoothness conditions on the
14
ERIK DAVIS AND SUNDER SETHURAMAN
integrand f , uniqueness criteria for ordinary differential equations allow to deduce from the Euler-Lagrange equations, d/dt∇v f (γ(t), γ̇(t)) = ∇x f (γ(t), γ̇(t)),
that there is a unique geodesic between points a, b sufficiently close together (cf.
Proposition 5.25 in [8] and Chapter 5 in [9]). On the other hand, for general a, b,
‘nonuniqueness’ may hold depending on the structure of f . For instance, one may
construct an f , satisfying the ‘standing assumptions’, with several F -minimizing
paths, by penalizing portions of D so as to induce ‘forks’.
8. k-nearest neighbor graphs. It is not clear if our approximation results, say
Theorem 2.7, hold with respect to the k-nearest-neighbor graph with k bounded–
that is the graph formed by attaching edges from a vertex to the nearest k points.
For instance, when {Xi } is arranged along a fine regular grid, f (x, v) ≡ |v|, d = 2,
and k = 4, the optimal Hn route of moving from the origin to (1, 1) is on a ‘staircase’
path
√ with length ∼ 2, no matter how refined the grid is, yet the Euclidean distance
is 2. In this respect, the random geometric graph setting of Theorem 2.7 allows
enough choices among nearby points, as long as ρ is elliptic, for the optimal path
to approximate the straight line from (0, 0) to (1, 1). It would be of interest to
investigate the extent to which our results extend to k-nearest neighbor graphs.
3. Proof of Theorems 2.1, 2.2, and Corollary 2.3
As mentioned in the introduction, the proof of Theorems 2.1 and 2.2 relies on a
probabilistic ‘Gamma Convergence’ argument. After establishing some basic notation and results on quasinormal minimizers, we present three main proof elements,
‘liminf inequality’, ‘limsup inequality’ and ‘compactness’, in the following Subsections. Proofs of Theorems 2.1 and 2.2, and Corollary 2.3 are the end of the
Subsection.
3.1. Preliminaries. Define a ‘nearest-neighbor’ map Tn : D → Xn where, for
x ∈ D, Tn (x) is the point of Xn closest to x with respect to the Euclidean distance.
In the event of a tie, we adopt the convention that Tn (x) is that nearest neighbor in
Xn with the smallest subscript. Since Xn is random, we note Tn and the distortion
kTn − Idk∞ = sup |Tn (x) − x| = kTn − Idk∞ = sup min |Xi − y|
y∈D 1≤i≤n
x∈D
are also random. In Proposition 5.1 of the appendix, we show for a, b ∈ D that,
almost surely,
the graph Gn (a, b) is connected for all a, b ∈ D and all large n.
(3.13)
Moreover, it is shown there that exists a constant C such that almost surely,
lim sup
n→∞
kTn − Idk∞ n1/d
≤ C.
(log n)1/d
(3.14)
Throughout, we will be working with realizations where both (3.13) and (3.14) are
satisfied. Let
A1 be the probability 1 event that (3.13) and (3.14) hold.
We observe, when the decay rate (2.6) on n holds, on the set of realizations A1 ,
we have limn→∞ kTn − Idk∞ /n = 0.
To rule out certain degenerate configurations of points, in d ≥ 2, let
A2 be the event that Xk 6∈ SγXi ,Xj ∀ distinct i, j, k ∈ N.
APPROXIMATING GEODESICS VIA RANDOM POINTS
15
Since the {Xi } come from a continuous distribution, and the image of the Lipschitz
path γXi ,Xj in D ⊂ Rd , when d ≥ 2, is of lower dimension, A2 has probability 1.
Recall the definitions of quasinormal and linear paths γu,v and lu,v .
Proposition 3.1. Let m1 and m2 be the constants in (2.4). Then for u, v ∈ D,
m1 |u − v|p ≤ df (u, v) ≤ m2 |u − v|p .
(3.15)
Further, the path γu,v satisfies, for 0 ≤ s, t ≤ 1, that
|γu,v (s) − γu,v (t)| ≤ m2 /m1 )1/p |u − v||s − t|,
(3.16)
(3.17)
and
sup |γu,v (t) − lu,v (t)| ≤
m2 /m1
1/p
+ 1 |u − v|.
0≤t≤1
Proof. For a Lipschitz path γ between u and v, we have m1 |γ̇(t)|p ≤ f (γ(t), γ̇(t)) ≤
R1
m2 |γ̇(t)|p by (2.4). Also, inf γ∈Ω(a,b) 0 |γ̇(t)|p dt = |b − a|p by a standard calculus of
variations argument (see also Proposition 5.2). So, by taking infimum over γ, we
obtain (3.15).
Suppose now γ = γu,v is quasinormal, so that f (γ(t), γ̇(t)) = c for some constant
R1
c and a.e. t. Integrating, and noting (3.15), gives c = 0 f (γ, γ̇) dt = df (u, v) ≤
m2 |u − v|p . On the other hand, by (2.4), m1 |γ̇(t)|p ≤ f (γ(t), γ̇(t)) = c. Hence,
|γ̇(t)| ≤ (m2 /m1 )1/p |u − v| and (3.16) follows.
Finally, to establish
(3.17), suppose that there is a t such that |γu,v (t)−lu,v (t)| >
(m2 /m1 )1/p +1 |u−v|. Then, considering that |lu,v (t)−v| ≤ |u−v|, an application
of the triangle inequality gives |γu,v (t)−v| > (m2 /m1 )1/p |u−v|. However, by (3.16),
|γu,v (t) − v| = |γu,v (t) − γu,v (1)| ≤ (m2 /m1 )1/p |u − v||t − 1| ≤ (m2 /m1 )1/p |u − v|,
a contradiction. Thus, inequality (3.17) holds.
3.2. Liminf Inequality. A first step in getting some control over the limit cost F
in terms of the discrete costs is the following bound.
Lemma 3.2 (Liminf Inequality). Consider γ ∈ Ω(a, b), and suppose we have a
sequence of paths γn ∈ Ω(a, b) such that
Z 1
lim sup |γn (t) − γ(t)| = 0
and
sup
|γ̇n |p dt < ∞.
n→∞ 0≤t≤1
n
0
Then, F (γ) ≤ lim inf n→∞ F (γn ).
Proof. A sufficient condition for this inequality, a ‘lower semicontinuity’ property
of F , to hold is that f (x, v) be jointly continuous and convex in v. See Theorem
3.5 (and the subsequent Remark 2) of [8] for more discussion on this matter.
3.3. Limsup Inequality. To make effective use of the liminf inequality, we need
to identify a sufficiently rich set of sequences for which a reverse inequality holds.
To this end, we develop certain approximations of Lipschitz paths by piecewise
linear or piecewise quasinormal paths.
The following result gives a method for recovering an element of Vn (a, b) from a
suitable element of Ω(a, b).
Proposition 3.3. Suppose, for constants c, C, that γ ∈ Ω(a, b) satisfies
c≤
|γ(s) − γ(t)|
≤ C,
|s − t|
(3.18)
16
ERIK DAVIS AND SUNDER SETHURAMAN
for all 0 ≤ s < t ≤ 1. Let N = N (n) = dK/n e, where K = C + 1 say, and define
v0 = a, vN = b, and vi = Tn γ(i/N ) for 0 < i < N .
Then, with respect to realizations {Xi } in the probability 1 set A1 , we have
v = (v0 , . . . , vN ) ∈ Vn (a, b) for all sufficiently large n.
Proof. To show that v ∈ Vn (a, b), it is sufficient to verify that consecutive vertices
vi−1 and vi are connected by an edge in Gn (a, b), or in other words
0 < |vi − vi−1 | < n ,
for i = 1, . . . , N.
We first show that |vi − vi−1 | < n . Note that |γ(i/N ) − γ((i − 1)/N )| ≤ C/N ≤
(C/K)n and C/K < 1. For 1 < i < N , we have
|vi − vi−1 | = |Tn γ(i/N ) − Tn γ((i − 1)/N )|
≤ (C/K)n + 2kTn − Idk∞ .
Similarly, for segments incident to an endpoint a or b, we have
max(|v1 − v0 |, |vN − vN −1 |) ≤ (C/K)n + kTn − Idk∞ .
In either case, assumption (2.6) on the decay of n implies that, for realizations
{Xi } in A1 , we have |vi − vi−1 | < n for all 1 ≤ i ≤ N and sufficiently large n.
Now, we show that 0 < |vi − vi−1 |. By the Lipschitz lower bound on γ, we have
|γ(i/N ) − γ((i − 1)/N )| ≥ c/N > (c/(K + 1))n
for 1 ≤ i ≤ N . By a triangle inequality argument, the distance between vi and
vi−1 is bounded below by (c/(K + 1))n − 2kTn − Idk∞ , which on the set A1 , as
n satisfies (2.6) and therefore vanishes slower than kTn − Idk∞ , is positive for all
large n.
We now establish some approximation properties obtained by interpolating paths
between points in v = (v0 , . . . , vN ).
Proposition 3.4. Fix γ ∈ Ω(a, b) satisfying (3.18), and a realization {Xi } in
the probability 1 set A1 . Let γn = γv and ln = lv , where N = N (n) and v =
(v0 , . . . , vN ) are defined as in Proposition 3.3. Then, we obtain
lim sup |γn (t) − γ(t)| = 0, and lim sup |ln (t) − γ(t)| = 0.
n→∞ 0≤t≤1
n→∞ 0≤t≤1
In addition,
sup kln0 k∞ < ∞, and
n
ln0 (t) → γ 0 (t) for a.e. t ∈ [0, 1].
(3.19)
Proof. We first argue that limn→∞ sup0≤t≤1 |ln (t)−γ(t)| = 0. Let ui = γ(i/N ), and
let ˜ln = lu(n) ∈ Ω(a, b) be the piecewise linear interpolation of u(n) = (u0 , . . . , uN ).
As γ is Lipschitz and limn→∞ N (n) = ∞, we have limn→∞ sup0≤t≤1 |˜ln (t) − γ(t)| =
0, and also limn→∞ ˜ln0 (t) = γ 0 (t) for a.e. t ∈ [0, 1]. By construction, ln (i/N ) = vi =
Tn γ(i/N ) and ˜ln (i/n) = ui = γ(i/N ) so that
max |ln (i/N ) − ˜ln (i/N )| ≤ kId − Tn k∞ .
0≤i≤N
Then, as ln and ˜ln are piecewise linear, it follows that sup0≤t≤1 |ln (t)−˜ln (t)| ≤ kId−
Tn k∞ and, as kId − Tn k∞ vanishes on A1 , that limn→∞ sup0≤t≤1 |ln (t) − γ(t)| = 0.
For i/N < t < (i + 1)/N , we have
ln0 (t) = N (vi+1 − vi ).
APPROXIMATING GEODESICS VIA RANDOM POINTS
17
As |vi+1 − vi | ≤ n (Proposition 3.3), it follows that |ln0 (t)| ≤ N n ≤ K + n . Hence,
supn kln0 k∞ < ∞.
Likewise, ˜ln0 (t) = N (ui+1 − ui ), and so
|l0 (t) − ˜l0 (t)| ≤ N (|vi+1 − ui+1 | + |vi − ui |) ≤ 2N kTn − Idk∞ .
n
n
For realizations in the probability 1 set A1 , since N = dK/n e and n satisfies (2.6)
and therefore vanishes slower than kTn −Idk∞ , we have limn→∞ N kTn −Idk∞ = 0.
Hence, ln0 (t) → γ 0 (t) for a.e. t ∈ [0, 1].
Now, considering the bound (3.17), it follows that
sup |γn (t) − ln (t)| ≤ max (C|v0 − v1 |, . . . , C|vN −1 − vN |) ≤ Cn ,
0≤t≤1
and hence kγn − γk∞ → 0.
With the above work in place, we proceed to the main result of this subsection.
Lemma 3.5 (Limsup Inequality). Let γ ∈ Ω(a, b) satisfy inequality (3.18). Then,
with respect to realizations {Xi } in the probability 1 set A1 , we may find a sequence
of paths {γn } taken either in form for all large n as (1) γn ∈ Ωln (a, b) or (2)
γn ∈ Ωγn (a, b) such that limn→∞ sup0≤t≤1 |γn (t) − γ(t)| = 0 and
F (γ) ≥ lim sup F (γn ).
(3.20)
n→∞
We remark that the sequence {γn } in the last lemma is called the ‘recovery
sequence’ since the liminf inequality in Lemma 3.2 and the limsup inequality in
Lemma 3.5 together imply the limit, limn F (γn ) = F (γ).
Proof. Let N = dK/n e, where K = C + 1 say is a constant greater than C in
(3.18). Define v0 = a, vN = b, and vi = Tn γ(i/N ) for 0 < i < N . Then, by
Proposition 3.3, v = v(n) = (v0 , . . . , vN ) ∈ Vn (a, b).
We now consider paths in case (1). By Proposition 3.4, the interpolated paths
ln = lv ∈ Ωln (a, b) converge uniformly to γ. Consider the bound
Z 1
|F (ln ) − F (γ)| ≤
|f (ln (t), ln0 (t)) − f (γ(t), γ 0 (t))| dt.
0
By Proposition 3.4, ln0 converges almost everywhere to γ 0 , and supn kln0 k∞ < ∞.
Hence (ln (t), ln0 (t)) → (γ(t), γ 0 (t)) for almost every t. Also, kγ̇k∞ < C by (3.18).
Since, by (2.4), f (x, v) ≤ m2 |v|p , an application of the bounded convergence theorem yields limn→∞ |F (ln )−F (γ)| = 0. Here, {ln } is the desired ‘recovery’ sequence.
We now consider case (2). Let γn = γv ∈ Ωγ (a, b). By Proposition 3.4, it follows
that limn→∞ sup0≤t≤1 |γn (t) − γ(t)| = 0. To show (3.20) for this sequence, write
Z 1
F (γn ) =
f (γn (t), γ˙n (t)) dt
0
=
N Z
X
i=1
≤
f (γn (t), γ˙n (t)) dt
(i−1)/N
N Z
X
i=1
i/N
i/N
f (ln (t), l˙n (t)) dt = F (ln ),
(i−1)/N
as γn on the time interval [i/N, (i+1)/N ] corresponds to the minimum cost, geodesic
path moving from vi to vi+1 (cf. (2.10)), and ln is a possibly more expensive path.
18
ERIK DAVIS AND SUNDER SETHURAMAN
But, by case (1), lim sup F (γn ) ≤ lim sup F (ln ) ≤ F (γ).
3.4. Compactness. In this Subsection, we consider circumstances under which a
sequence of paths {γn }, in the context of Theorems 2.1 and 2.2, has a limit point
with respect to uniform convergence. Here, the arguments when p = 1 differ from
those when p > 1.
R1
In particular, consider paths γn where 0 f (γn (t), γ˙n (t)) dt is uniformly bounded.
One has m1 |v|p ≤ f (x, v) and it follows that {γn } is bounded in the W 1,p Sobolev
space. When p > 1, this is sufficient to derive a suitable compactness result. But,
when p = 1, this is no longer the case.
However, when p = 1, our general outlook is that it is enough to establish
a compactness result for sequences of optimal paths, on which certain eccentric
possibilities are ruled out.
We begin by considering such compactness when p = 1, when the paths lie in
Ωγn (a, b). The setting p > 1 is discussed afterwards.
Proposition 3.6. Suppose d ≥ 2 and that p = 1. Then, with respect to realizations
{Xi } in the probability 1 set A1 ∩ A2 , for all large n, if γ ∈ arg min Gn and 0 ≤
s, t ≤ 1, we have that
|γ(s) − γ(t)| ≤ (4m2 /m21 )Gn (γ)|s − t|.
Proof. The path γ ∈ Ωγn (a, b) is a piecewise quasinormal path of the form γ = γv
where v = (v0 , v1 , . . . , vm ) ∈ Vn (a, b). We now try to relate m, the number of
segments in the path, to Gn (γ), the path energy. Recall the formula (2.9).
Let Bi denote the (open) Euclidean ball of radius n /2 around vi . We claim
that |Bi ∩ {v0 , . . . , vm }| ≤ 2. To see this, suppose that there are at least 3 points
of {v0 , . . . , vm } in Bi . Let vj and vl denote the points in Bi with the smallest and
largest index, respectively. By minimality of γ, vj 6= vl . Let vk denote a third point
in Bi .
As vk , vl ∈ Bi , we have |vk − vl | < n , and so these points are connected in the
graph. Applying the triangle inequality for df , valid when p = 1 (cf. (2.5)), and
noting on the event A2 that vk 6∈ Sγvj ,vl , we have
df (vj , vl ) < df (vj , vk ) + df (vk , vl ) ≤ df (vj , vj+1 ) . . . + df (vl−1 , vl ).
Thus, the path γ̃ = γw , where w = (v0 , . . . , vj , vl , . . . , vm ), satisfies Gn (γ̃) < Gn (γ).
This contradicts the optimality of γ, and therefore |Bi ∩ {v0 , . . . , vm }| ≤ 2.
We may thus cover the vertices of γ with balls {Bi }m
i=1 , centered on the vertices
{vi }m
i=1 , and each of these balls contains at most two vertices. It follows that there
is a subcover by s ≥ m/2 balls, {B10 , . . . , Bs0 }, no two of them containing a common
point in v.
A lower bound for Gn (γ) is found by considering that part of the Gn -integral
contributed to by the portion of the path γ in Bi0 . Each such portion, if it does not
terminate in Bi0 , must visit both the center of Bi0 and the boundary ∂Bi0 , and hence
has Euclidean length at least n /2. Summing over these portions, we obtain
s
X n
mn
≤
≤ L,
4
2
i=1
where L =
R1
0
|γ̇(t)| dt is the Euclidean arclength of γ.
APPROXIMATING GEODESICS VIA RANDOM POINTS
19
By (2.4), it follows that
m1
mn
≤ m1 L ≤
4
Z
t
f (γ(t), γ̇(t))dt = Gn (γ).
(3.21)
0
To get a Lipschitz bound for γ, recall the bound (3.16). Then, γvi−1 ,vi (m ·
−i) is Lipschitz with constant (m2 /m1 )m|vi−1 − vi |. It follows that γ, being the
concatenation of these segments, satisfies
|γ(s) − γ(t)| =
q−1
X
γ(wr ) − γ(wr+1 )
r=0
≤
(m2 /m1 )m max (|v0 − v1 |, . . . , |vm−1 − vm |)
q−1
X
|wr − wr+1 |
r=0
≤
(m2 /m1 )mn |s − t|,
m−1
where s = w0 < · · · wq = t, {wr }q−1
r=1 ⊂ {j/m}j=1 so that |wr − wr+1 | ≤ 1/m for
0 ≤ r ≤ q − 1. Then, with (3.21), we obtain
|γ(t) − γ(s)| ≤ (m2 /m1 )mn |s − t| ≤ (4/m1 )(m2 /m1 )Gn (γ)|s − t|,
finishing the proof.
We now prove our compactness property.
Lemma 3.7 (Compactness Property).
(I). Suppose for all large n that either γn ∈ arg min Ln or γn ∈ arg min Gn .
Then, for realizations {Xi } in the probability 1 set A1 , we have supn F (γn ) < ∞.
(II). Consider now the following cases:
(i) Suppose paths γn ∈ arg min Gn for all large n.
(ii) Suppose p > 1, and paths γn ∈ Ω(a, b) for all large n such that supn F (γn ) < ∞.
Then, in case (i) when p > 1, and in case (ii), with respect to realizations {Xi }
in the probability 1 set A1 , we have {γn } is relatively compact for the topology of
uniform convergence. For case (i) when d ≥ 2 and p = 1, the same conclusion
holds with respect to realizations {Xi } in the probability 1 set A1 ∩ A2 .
Proof. We first prove the bound supn F (γn ) < ∞ in part (I). Choose a γ̃ ∈ Ω(a, b),
where (3.18) holds, and F (γ̃) < ∞. By Lemma 3.5, there is a sequence {γ̃n } of
either piecewise linear or quasinormal paths such that lim supn→∞ F (γ̃n ) ≤ F (γ̃).
Hence, by minimality of {γn }, with respect to paths in either Ωln or Ωγn , we have
sup F (γn ) ≤ sup F (γ̃n ) < ∞.
n
(3.22)
n
We now argue the claims for cases (i) and (ii). In both cases, as D is bounded,
the paths γn : [0, 1] → D are uniformly bounded. To invoke the Arzela-Ascoli
theorem, we must show that {γn } is an equicontinuous family.
In case (i), when d ≥ 2 and p = 1, by Lemma 3.6 on realizations in A1 ∩ A2 , we
have |γn (s)−γn (t)| ≤ CGn (γn )|s−t|, with C independent of n. As Gn (γn ) = F (γn ),
combining with (3.22), it follows that {γn } is equicontinuous.
If p > 1, with respect to realizations in A1 , (3.22) implies that, if case (i) holds
for the sequence, then case (ii) holds.
Without loss of generality, then, we focus our attention now on case (ii). Recall,
by (2.4), that m1 |v|p ≤ f (x, v). Let q be the conjugate of p, that is 1/p + 1/q = 1.
20
ERIK DAVIS AND SUNDER SETHURAMAN
Then, for 0 ≤ s < t ≤ 1,
Z
|γn (s) − γn (t)| ≤
t
|γn0 (r)| dr
s
≤ |t − s|
1/q
1
Z
|γn0 (t)|p
1/p
dt
(3.23)
0
1/p
≤ (|t − s|1/q /m1 )
Z
1
f (γn (t), γn0 (t)) dt
1/p
0
1/p
= (|t − s|1/q /m1 )F (γn )1/p
Combining (3.23) and the assumption in case (ii) that supn F (γn ) < ∞, we have
|γn (s) − γn (t)| ≤ C|t − s|1/q for a constant C independent of n, and hence {γn } is
equicontinuous.
3.5. Proof of Theorems 2.1 and 2.2. With the preceding ‘Gamma convergence’
ingredients in place, the proofs of Theorem 2.1 and 2.2 are similar, and will be given
together.
Proofs of Theorems 2.1 and 2.2. Fix a realization {Xi } in the probability 1
set A1 . Let {γn } be a sequence of paths such that, for all large n, we have either
γn ∈ arg min Ln or γn ∈ arg min Gn . Supposing that {γn } has a subsequential limit
limk→∞ γnk = γ, with respect to the topology of uniform convergence, we now
argue that γ ∈ arg min F .
By the ‘liminf’ Lemma 3.2, we have
F (γ) ≤ lim inf F (γnk ).
k→∞
(3.24)
Let γ ∗ ∈ arg min F . Then, by inequality (5.49) of Proposition 5.2, there exist
constants c1 , c2 such that c1 |s − t| ≤ |γ ∗ (s) − γ ∗ (t)| ≤ c2 |s − t| for 0 ≤ s, t ≤ 1.
Hence, by the ‘limsup’ Lemma 3.5, there exists a sequence {γn∗ }, of either piecewise
linear or quasinormal paths, converging uniformly to γ ∗ and lim supn→∞ F (γn∗ ) ≤
F (γ ∗ ). Recall that F (γ) = Ln (γ) or F (γ) = Gn (γ) when γ is piecewise linear or
quasinormal respectively. Combining with (3.24) and minimality of γnk , we have
F (γ) ≤ lim inf F (γnk ) ≤ lim sup F (γn∗k ) ≤ F (γ ∗ ) = min F,
k→∞
(3.25)
k→∞
and so γ ∈ arg min F .
In the case the paths {γn } are piecewise linear, since F (γn ) = min Ln , it follows
from (3.25) that
lim min Lnk = min F.
k→∞
Similarly, when {γn } are piecewise quasinormal, limk→∞ min Gnk = min F .
Therefore, we have shown that, if a subsequential limit of {γn } exists, it is an
optimal continuum path γ ∈ arg min F .
Consider now Theorem 2.1, where p > 1 and γn ∈ arg min Ln , and part (1) of
Theorem 2.2 where p > 1 and γn ∈ arg min Gn . By the ‘compactness’ Lemma 3.7,
supn F (γn ) < ∞ and a subsequential limit exists.
Consider now part (2) of Theorem 2.2 where d ≥ 2, p = 1 and γn ∈ arg min Gn .
Suppose that the realization {Xi } belongs also to the probability 1 set A2 . Then,
subsequential limits follow again from the ‘compactness’ Lemma 3.7.
APPROXIMATING GEODESICS VIA RANDOM POINTS
21
Now, consider any subsequence {nk } of N. Then, by the work above, applied to the sequence {nk }, there is a further subsequence {nkj }, and a γ ∈
arg min F , with γnkj → γ uniformly, in the settings of Theorems 2.1 and 2.2.
Moreover, limj→∞ min Lnkj = min F when the paths {γnkj } are piecewise linear,
and limj→∞ min Gnkj = min F when the paths are piecewise quasinormal.
Since this argument is valid for any subsequence {nk } of N, we recover that
min Ln → min F or min Gn → min F when respectively the paths are piecewise
linear or quasinormal. Finally, if F has a unique minimizer γ, by considering
subsequences again, the whole sequence {γn } must converges uniformly to γ.
3.6. Proof of Corollary 2.3. Corollary 2.3 is a statement about Hausdorff convergence. In order to adapt the results of Theorems 2.1 and 2.2 to this end, we
make the following observation.
Proposition 3.8. Fix a realization {Xi } in A1 , and consider a sequence of paths
{γn } such that γn for all large n is either in the form (1) γn = γv(n) or (2) γn =
lv(n) , where v(n) ∈ Vn (a, b). Suppose that γn converges uniformly to a limit γ ∈
Ω(a, b). Then,
lim dhaus (v(n) , Sγ ) = 0.
n→∞
(n)
(n)
(n)
Proof. Write v(n) = (v0 , . . . , vm(n) ). Since γn → γ uniformly and vi
it follows that
(n)
lim max inf |vi − x| = 0.
i
),
= γn ( m(n)
n→∞ 1≤i≤m(n) x∈Sγ
(3.26)
On the other hand, consider t with i/m(n) ≤ t < (i + 1)/m(n). In case (1),
(n)
|γ(t) − vi | ≤ |γ(t) − γ(i/m(n))| + kγn − γk∞ ≤ C/m(n) + kγn − γk∞ ,
where C is the Lipschitz constant of γ. In case (2), using linearity of the path,
(n)
(n)
(n)
|γ(t) − vi | ≤ |vi+1 − vi | + kγn − γk∞ ≤ n + kγn − γk∞ .
Pm(n) (n)
(n)
Since v(n) is a path in Vn (a, b), one may bound n m(n) ≥ i=0 |vi − vi+1 | ≥
|a − b|, and so m(n) ≥ |a − b|/n diverges. Hence, in both cases,
lim sup
min
n→∞ x∈Sγ 1≤i≤m(n)
(n)
|x − vi | = 0.
Combining (3.26) and (3.27), it follows that limn→∞ dhaus (v(n) , Sγ ) = 0.
(3.27)
We now proceed to prove Corollary 2.3.
Proof of Corollary 2.3. We give the argument for the case of piecewise linear
optimizers, as the the argument is exactly the same for piecewise quasinormal paths,
using Theorem 2.2 instead of Theorem 2.1 below.
Suppose ln = lv(n) ∈ arg min Ln is a sequence of paths where v(n) ∈ Vn (a, b).
By Theorem 2.1, with respect to a probability 1 set of realizations {Xi }, any subsequence of {ln } has a further subsequence {lnk } which converges uniformly to a
γ ∈ arg min F . By Proposition 3.8, it follows that limk→∞ dhaus (v(nk ) , Sγ ) = 0.
Suppose now that F has a unique (up to reparametrization) minimizer γ. Note
that Sγ is invariant under reparametrization of γ. Then, we conclude that all limit
points of {v (n) } correspond to Sγ , and hence the whole sequence v (n) converges to
Sγ , limn→∞ dhaus (v (n) , Sγ ) = 0.
22
ERIK DAVIS AND SUNDER SETHURAMAN
4. Proof of Theorems 2.4, 2.7 and 2.8
The proofs of Theorems 2.4, 2.7 and 2.8 all make use of Theorems 2.1 and 2.2 in
comparing the costs Hn and Ln to Gn . When p = 1, as with respect to Theorem 2.2,
the arguments in Theorems 2.7 and 2.8 are more involved, especially with respect
to the minimal cost Hn -path convergence, where several geometric estimates are
used to show a compactness principle.
We begin with the following useful fact.
Proposition 4.1. Suppose U, W, C : X → R are functions such that |U (x) −
W (x)| ≤ C(x) for all x ∈ X. If U (x1 ) = min U and W (x2 ) = min W then
−C(x1 ) ≤ min U − min W ≤ C(x2 ).
Proof. For any y ∈ X we have min U = U (x1 ) ≤ U (y) ≤ W (y) + C(y). Taking
y = x2 gives min U ≤ min W + C(x2 ). The other inequality follows similarly.
4.1. Proof of Theorem 2.4. Suppose v = (v0 , v1 , . . . , vm ) ∈ Vn (a, b). Then (Lip)
implies, for 0 ≤ i ≤ m − 1, that
Z 1
f (vi + t(vi+1 − vi ), vi+1 − vi ) dt − f (vi , vi+1 − vi ) ≤ c|vi+1 − vi |p+1 . (4.28)
0
Since vi , vi+1 are neighbors in the n -graph, |vi+1 − vi | ≤ n . Thus, from the
homogeneity (2.3) and bounds (2.4) of f , rescaling (4.28) gives
Z i+1
m
i
1
f (vi + m(s − )(vi+1 − vi ), m(vi+1 − vi )) ds − f (vi , m(vi+1 − vi ))
i
m
m
m
cn
≤
|vi+1 − vi |p mp .
m
Recall formulas (2.8) and (2.11). Summing over i gives the following estimate
relating Ln and Hn :
|Ln (lv ) − Hn (v)| ≤
m−1
cn X
|vi+1 − vi |p mp .
m i=0
(4.29)
Applying (2.4), the right-side of (4.29) can be bounded above in terms of both
Ln (lv ) and Hn (v). Hence, with c0 = cm−1
1 ,
|Ln (lv ) − Hn (v)| ≤ c0 n min(Ln (lv ), Hn (v)).
Suppose lv(n) ∈ arg min Ln and w
of (4.30) and Proposition 4.1 is
−c0 n min Ln
(n)
(4.30)
∈ arg min Hn . An immediate consequence
≤
−c0 n min(Ln (lv(n) ), Hn (v(n) ))
≤
min Ln − min Hn
≤
c0 n min(Ln (lw(n) ), Hn (w(n) )) ≤ c0 n min Hn .
By Theorem 2.1, we have limn→∞ min Ln = min F for almost all realizations
{Xi } (those in A1 as the proof shows). Then, min Hn ≤ (1 + c0 n ) min Ln and so
lim sup min Hn ≤ min F a.s. In particular, as min F < ∞, we have supn min Hn <
∞ a.s.
On the other hand, min Hn ≥ min Ln − c0 n min Hn a.s. As supn min Hn < ∞,
we observe that lim inf min Hn ≥ min F a.s. Hence, min Hn → min F a.s. This
finishes one part of Theorem 2.4.
APPROXIMATING GEODESICS VIA RANDOM POINTS
23
To address the others, consider lw(n) , the piecewise linear interpolation of w(n) .
By (4.30), we have
Ln (lw(n) ) ≤ (1 + c0 n )Hn (w(n) ) = (1 + c0 n ) min Hn .
Moreover, noting the optimality of lv(n) and w(n) gives
Ln (lv(n) ) ≤ Ln (lw(n) ) ≤ (c0 n + 1)Hn (v(n) ).
(4.31)
Another application of (4.30) yields
Hn (v(n) ) ≤ c0 n Ln (lv(n) ) + Ln (lv(n) ).
Hence, the left-side of (4.31) is bounded as
min Ln = Ln (lv(n) ) ≤
=
Ln (lw(n) ) ≤ (c0 n + 1)2 Ln (v(n) )
(c0 n + 1)2 min Ln .
(4.32)
Hence, as min Ln → min F a.s., we have
min F = lim Ln (lv(n) ) = lim Ln (lw(n) ) a.s.
n→∞
n→∞
(4.33)
We also observe, as a consequence, that supn Ln (lw(n) ) < ∞ a.s.
Given that p > 1, by the ‘compactness’ Lemma 3.7, with respect to realizations
{Xi } in the probability 1 set A1 , any subsequence of {lw(n) } has a further uniformly
convergent subsequence to a limit γ̃ ∈ Ω(a, b). By the ‘liminf’ Lemma 3.2, F (γ̃) ≤
lim inf n→∞ Ln (lw(n) ) a.s. Finally, by (4.33), it follows that F (γ̃) = min F and
so γ̃ ∈ arg min F . Consequently, if F has a unique minimizer γ, then the whole
sequence {lw(n) } converges uniformly almost surely to it.
The proofs of statements about Hausdorff convergence follow the same arguments
as given for Corollary 2.3, and are omitted.
4.2. Proof of Theorems 2.7 and 2.8. We prove Theorems 2.7 and 2.8 in two
parts.
Proof of Theorems 2.7 and 2.8. First, we prove in Proposition 4.3 that the
minimal costs of Hn and Ln converge to min F , making use of comparisions with
quasinormal paths, for which we have control in Theorem 2.2.
Second, in Proposition 4.9 in Subsection 4.2.2, we show that the minimizing paths
converge in the various senses desired. A main tool in this proof is a compactness
property (Proposition 4.4), for minimal Hn and Ln -paths when p = 1, shown in
Subsection 4.2.1.
To supply the proofs of the desired propositions, we now obtain an useful estimate
between the cost of a quasinormal path and a linear one.
Proposition 4.2. Suppose d ≥ 2, p = 1, and that f also satisfies (Lip) and (Hilb).
For a, b ∈ D such that |b − a| ≤ 1, there is a constant c1 such that
df (a, b) − f (a, b − a) ≤ c1 |b − a|2 .
R1
In particular, as df (a, b) = 0 f (γ(t), γ̇(t))dt for the quasinormal path γ = γa,b
connecting a and b, we have
Z 1
f (γ(t), γ̇(t))dt − f (a, b − a) ≤ c1 |b − a|2 .
0
24
ERIK DAVIS AND SUNDER SETHURAMAN
Proof. By (2.4), for a Lipschitz path β from a to b, we have
Z 1
Z 1
Z 1
f (β(t), β̇(t))dt ≤ m2
|β̇(t)|dt.
|β̇(t)|dt ≤
m1 |
0
0
0
R1
Optimizing over β, we recover that m1 |b − a| ≤ 0 f (γ(t), γ̇(t))dt ≤ m2 |b − a|. By
R1
(2.4) again, we have that the arclength of γ satisfies 0 |γ̇(t)|dt ≤ (m2 /m1 )|b − a|.
In particular, the path γ is constrained in the Euclidean ball B around a of radius
(m2 /m1 )|b − a|. Note also that the minimizing Euclidean path γ
e, with constant
speed |b − a| on the straight line from a to b in times 0 ≤ t ≤ 1, is also constrained
in this ball.
Now, for a Lipschitz path β, constrained in the ball B, expand
Z 1
Z 1
Z 1
f (β(t), β̇(t))dt =
f (a, β̇(t))dt +
f (β(t), β̇(t)) − f (a, β̇(t))dt.
0
0
0
As the paths are in B, by (Lip), with respect to a Lipschitz constant C,
|f (β(t), β̇(t)) − f (a, β̇(t))| ≤ C|β(t) − a||β̇(t)|
≤ C(m2 /m1 )|b − a||β̇(t)|.
Therefore, with respect to Lipschitz paths β constrained in B,
Z 1
Z 1
Z
f (β(t), β̇(t))dt −
f (a, β̇(t))dt ≤ C(m2 /m1 )|b − a|
0
0
1
|β̇(t)|dt.
0
Note, by condition (Hilb) that, for the cost with respect to f (a, ·), straight lines
are geodesics, and in particular γ
e(t) = (1 − t)a + tb is optimal. Hence, the minimal F -cost, with respect to f (a, ·), of moving from a to b, given invariance to
parametrization when p = 1, is f (a, b − a).
Then, by Proposition 4.1, applied to the two functionals of β on the left-hand
sides, we obtain
Z 1
hZ 1
i
df (a, b) − f (a, b − a) ≤ C(m2 /m1 )|b − a| max
|γ̇(t)|dt,
|γ
e˙ (t)|dt ,
0
≤
0
C(m2 /m1 )2 |b − a|2 ,
noting the arclength bounds of γ = γa,b and γ
e above.
Proposition 4.3. Suppose d ≥ 2, p = 1, and that f also satisfies (Lip) and (Hilb).
With respect to realizations {Xi } in the probability 1 set A1 ∩ A2 , the minimum
values of Hn and Ln converge to the minimum of F ,
lim
min
n→∞ v∈Vn (a,b)
Hn (v) = lim
min
n→∞ γ∈Ωln (a,b)
Ln (γ) =
min F (γ).
γ∈Ω(a,b)
Proof. Consider the energies Gn and Hn in (2.9) and (2.11). For γ = γv , the
piecewise quasinormal path through the vertices v = (v0 , v1 , . . . , vm ) ∈ Vn (a, b), we
have, noting p = 1, that
m−1
XZ 1
Gn (γ) =
f (γi (t), γ̇i (t)) dt,
i=0
0
where γi = γvi ,vi+1 is a quasinormal path from vi to vi+1 .
APPROXIMATING GEODESICS VIA RANDOM POINTS
25
An application of Proposition 4.2, noting that |vi+1 − vi | ≤ n , gives
Z 1
f (γi (t), γ˙i (t)) dt − f (vi , vi+1 − vi ) ≤ c1 n |vi+1 − vi |.
0
Summing this over i gives
|Gn (γ) − Hn (v)| ≤ c1 n
≤
m−1
X
i=0
c1 m−1
1 n
|vi+1 − vi |
(4.34)
min(Gn (γ), Hn (v)),
where the last inequality follows from applying (2.4) to both Gn and Hn .
Recall the energy Ln in (2.8). Similarly, and more directly, using (Lip), we have
for a linear path l = lv ∈ Ωln (a, b) through vertices v = (v0 , . . . , vm ) ∈ Vn (a, b) that
Z 1
f (li (t), l˙i (t)) dt − f (vi , vi+1 − vi ) ≤ c1 n |vi+1 − vi |,
0
where li = lvi ,vi+1 is the linear path from vi to vi+1 with slope vi+1 − vi . Summing
over i, using (2.4), we obtain
|Ln (l) − Hn (v)| ≤ c1 n
m−1
X
|vi+1 − vi | ≤ c1 m−1
1 n min(Ln (l), Hn (v)).
(4.35)
i=0
We now reprise some of the argument for Theorem 2.4. A consequence of (4.34)
and Proposition 4.1 is
−1
−c1 m−1
1 n min Gn ≤ min Gn − min Hn ≤ c1 m1 n min Hn .
Hence, supn min Hn < supn (1 + c1 m−1
1 n ) min Gn . By Theorem 2.2, as seen in its
proof, for realizations in the probablility 1 set A1 ∩ A2 , we have limn→∞ min Gn =
min F , which is finite. Then, we conclude that also limn→∞ min Hn = min F a.s.
Now, we can repeat this same argument with Ln and (4.35) in place of Gn and
(4.34), using now min Hn → min F a.s., to conclude that also min Ln converges to
min F a.s.
4.2.1. Compactness Property. When d ≥ 2 and p = 1, analogous to Lemma 3.6, we
formulate now a compactness property for minimal paths w(n) ∈ arg min Hn and
lv(n) ∈ arg min Ln .
It will be useful to consider a partition of D by a regular grid. Let z ∈ Zd
Qd
and let
√n,z be the intersection of the box i=1 [zi τn , (zi + 1)τn ) with D, where
τn = n / d. We will refer to these sets as ‘boxes’, with the understanding that the
boundary of D results in some of these being irregularly shaped. Regardless, each
n,z has diameter at most n , and so points of {Xi }ni=1 in n,z are all connected
in the random geometric graph.
Proposition 4.4. Consider the assumptions in the second parts of Theorems 2.7
and 2.8. Suppose w(n) ∈ arg min Hn , and consider the piecewise linear interpolations ln = lw(n) . Then, with respect to a realizations {Xi }i≥1 in a probability 1
subset of A1 ∩A2 , the sequence {ln } is relatively compact for the topology of uniform
convergence.
Suppose now lv(n) ∈ arg min Ln . Then, the same conclusion holds for the optimal
linear interpolations {lv(n) }.
26
ERIK DAVIS AND SUNDER SETHURAMAN
Proof. We show that the sequence {ln } is equicontinuous for almost all realizations
in A1 . As the paths belong to a bounded set D, the proposition would then follow
from the Arzela-Ascoli criterion.
Partition D by boxes {n,z }z∈Zd . By Lemma 4.5 below, the number of boxes
visited by w(n) and v(n) is a.s. bounded by C/n a.s., for all large n, where
C = C(f, d). By Lemma 4.8 below, the number of vertices in w(n) and v(n) in
a box is a.s. bounded by a constant K = K(d, ρ, α) for all large n. Thus, the
maximum number kn of points in w(n) and v(n) is a.s. bounded,
kn ≤ KC/n .
(n)
(n)
(n)
(n)
Since |wi+1 − wi |, |vi+1 − vi | ≤ n , we obtain
(n)
(n)
supi kn |vi+1 − vi | ≤ KC a.s. for all large n.
(n)
(n)
supi kn |wi+1 − wi | ≤ KC and
This implies a.s. that the piecewise linear paths ln and lv(n) are Lipschitz, with
respect to the fixed constant KC, for all large n, and so in particular equicontinuous.
(n)
(n)
Indeed, for ln = lw(n) , where say w(n) = (w0 , . . . , wkn ), consider the part of the
(n)
path connecting wi
(n)
and wi+1 from times i/kn to (i + 1)/kn , namely ln (t) =
(n)
(n)
(n)
(n)
wi (i + 1 − kn t) + wi+1 (kn t − i). Then, we have |l˙n (t)| = kn |wi+1 − wi | ≤ kn n ≤
KC. The same argument holds for the paths lv(n) .
We now show the lemmas used in the proof Proposition 4.4. We first bound the
number of boxes visited by an optimal path.
Lemma 4.5. Suppose d ≥ 2, p = 1, and that f also satisfies (Lip) and (Hilb).
Suppose w ∈ arg min Hn and lv ∈ arg min Ln are optimal paths. Then, for realizations {Xi } in the probability 1 set A1 ∩ A2 , for all large n, the number of distinct
boxes {n,z }z∈Zd visited by w and v is bounded by C/n , where C = C(d, f ).
d
Proof. Any visit
√ of the path w or v to 2 + 1 distinct boxes has an Euclidean length
of at least n / d, since not all 2d + 1 boxes can be adjacent. Recalling (2.4), and
the formulas (2.8)√and (2.11), such a visitation therefore has a Hn cost or Ln cost
of at least m1 n / d. So, we may √
bound the number of boxes visited by w or v by
C 0 Hn (w)/n , where C 0 = (2d + 1) d/m1 depends on the dimension and f , but not
on the path w or v. Recalling Proposition 4.3, we have with respect to realizations
in A1 ∩ A2 that lim min Hn = lim min Ln = min F < ∞. The lemma then follows
with say C = 2C 0 min F .
The next result shows that optimal paths w ∈ arg min Hn and lv ∈ arg min Ln
cannot have ‘long necks’, and gives a bound on the number of points nearby an
edge in the graph.
Lemma 4.6. Suppose d ≥ 2 and p = 1. Fix a realization {Xi } in the probability
1 set A1 . Suppose w ∈ arg min Hn is an optimal path. If i < j is such that
|wi − wj | < n , then
wk ∈ B(wi , Cn ), for i ≤ k ≤ j,
(4.36)
where C = 2(m2 /m1 ).
Further, let Θn = supi,j:|wi −wj |<n |i − j|, and suppose n = n−δ , where δ >
1/(β + d) and β > 0. Then, with respect to realizations in a probability 1 subset of
A1 , for all large n, we have
Θn ≤ −β
(4.37)
n .
APPROXIMATING GEODESICS VIA RANDOM POINTS
27
Suppose now lv ∈ arg min Ln . The same conclusions (4.36) and (4.37) hold with
v in place of w.
Proof. We first show (4.36). If one of the points {wk }jk=i is more than an Euclidean
distance 2(m2 /m1 )n away from wi , then, recalling (2.4), we have
j−1
X
f (wk , wk+1 − wk ) ≥ m1
k=i
j−1
X
|wk+1 − wk |
k=i
≥ 2m2 n ≥ 2m2 |wj − wi | ≥ 2f (wi , wj − wi ).
But, this implies that the path connecting wi and wj in one step would be less
costly, with respect to Hn , than w. Since w was taken to be minimal, all points
{wk }jk=i therefore must belong to B(wi , 2(m2 /m1 )n ).
Suppose now lv ∈ arg min Ln and recall the form of Ln when p = 1 in (2.8).
Similarly, if one of the points {vk }jk=i is away from vi by 2(m2 /m1 )n , we have
j−1 Z 1
j−1
X
X
f (lvk ,vk+1 (t), vk+1 − vk )dt ≥ m1
|vk+1 − vk |
k=i
0
k=i
1
Z
≥ 2m2 n ≥ 2m2 |vj − vi | ≥ 2
f (lvi ,vj (t), vj − vi )dt,
0
also a contradiction of minimality of lv .
We now consider (4.37). The proof here is a count bound with respect to w.
The argument with respect to v is exactly the same with v in place of w.
First, |j − i| is bounded by the number Ni,n of points {Xi }ni=1 , distinct from wi ,
in the ball B(wi , Cn ). Then, Ni,n is Binomial(n − 1, p) where p = ν(B(Xi , Cn )).
For k ≥ 1, we have
n−1 k
(np)k
.
P(Ni,n ≥ k) ≤
p ≤
k!
k
Recalling that ν = ρdx and ρ is bounded, we have p ≤ kρk∞ Vol(B(0, 1))dn , and so
P (Ni,n ≥ k) ≤ (C 0 ndn )k /k! for some constant C 0 .
Let Nn = max{Ni,n }ni=1 . Then, a union bound gives that
n
exp{k(log n + d log n + log C 0 )}.
P (Nn ≥ k) ≤
k!
√
Taking k = d−β
2πe−k k k+1/2 , yields that
n e, and noting k! ≥
n
0
P (Nn ≥ −β
exp{−β
(4.38)
n (log n + (β + d) log n + log C + 1)}.
n )≤ √
2π
If n is in the form n = n−δ , then the right hand side of (4.38) is summable when
(β + d)δ > 1.
Hence, by Borel-Cantelli lemma, for realizations in the intersection of a probability 1 set and A1 say, we have Θn ≤ maxi Ni,n ≤ Nn ≤ −β
n for all large n, and
(4.37) follows.
We now give a lower bound on the cost of certain ‘long necks’, that is the cost
of an optimal Hn -path w of moving away from two close by vertices.
Lemma 4.7. Suppose d ≥ 2, p = 1 and that f also satisfies (T rIneq), and
(P ythag) with α > 1. Fix a realization {Xi } in the probability 1 set A1 . Suppose
w ∈ arg min Hn is an optimal path, and let i < j be indices such that |wi −wj | < n .
28
ERIK DAVIS AND SUNDER SETHURAMAN
Let ` denote the straight line segment from wi to wj . Consider a neighborhood
A = ∪x∈` B(x, r) of `, with r = α
n.
Then, if there is a point wk 6∈ A for some i < k < j, there is a constant
C = C(α, f ) such that
j−1
X
f (wi , wq+1 − wq ) ≥ f (wi , wj − wi ) + Crα .
(4.39)
q=i
Suppose now lv ∈ arg min Ln . Then, (4.39) holds with v in place of w.
Proof. . The argument for v is the same as for w, which we now present. Suppose
a point wk is at least an Euclidean distance r from `. By the (TrIneq) condition,
Pj−1
q=i f (wi , wq+1 − wq ) ≥ f (wi , wk − wi ) + f (wi , wj − wk ).
By Lemma 4.6, as |wi − wj | < n , we have |wk wi | = |wk − wi | ≤ 2(m2 /m1 )n ,
which is strictly less than an η < 1 for all large n. We also conclude |wj wk |, |wi wj | <
η < 1 for all large n. In addition, 2(m2 /m1 )r1/α = 2(m2 /m1 )n ≥ |wi wj |. Thus,
by (Pythag) with x = wi , we obtain f (wi , wk − wi ) + f (wi , wj − wk ) ≥ f (wi , wj −
wi ) + Crα , where C = C(α, f ).
Hence, (4.39) follows by combining the inequalities.
We now bound the number of points of an optimal path in a box, the main
estimate used in the proof of Proposition 4.4. The argument is in two steps. In the
first step, using a rough count on the number of vertices of the path within a given
box, we may approximate the contribution to Hn and Ln from the vertices in the
box in terms of a ‘localized’ cost. Then, we use (Pythag), applied to the ‘localized’
cost, to deduce that the optimal path in the box is trapped in a ‘small’ set in the
box. The second step then is to show that such ‘small’ sets contain only a constant
number of points in {Xi }.
Lemma 4.8. Consider the assumptions of the second parts of Theorems 2.7 and
2.8. Suppose w ∈ arg min Hn is an optimal path. Then, with respect to realizations
{Xi } in a probability 1 subset of A1 , for all large n, there is a constant K, such
that |{wj }nj=1 ∩ n,z | ≤ K for all z ∈ Zd .
Suppose now lv ∈ arg min Ln . Then, the same statement above holds with v in
place of w.
Proof. We will give the main argument for w and indicate modifications with respect to v. Consider a box := n,z . Boxes with at most one point trivially
satisfy the claim in the lemma if say K ≥ 2. Suppose now that there are at least
two points in the box .
Step 1. Let wi and wj be the first and last points of w in the box, that is, with the
smallest and largest indices respectively. By Lemma 4.6, as |wi − wj | < n , we have
Pj−1
wk ∈ B(wi , C 0 n ) for i ≤ k ≤ j. Hence, by (Lip), we have | k=i f (wk , wk+1 −wk )−
Pj−1
0 2
−1
k=i f (wi , wk+1 −wk )| ≤ C n |j −i|. Now, also by Lemma 4.6, when δ > (β +d)
for β > 0, we have |j − i| ≤ n−β . Hence, the following estimate, with respect to a
‘localized’ energy, where x = wi is fixed, is obtained:
j−1
X
k=i
f (wk , wk+1 − wk ) −
j−1
X
k=i
f (wi , wk+1 − wk ) ≤ C 0 2−β
.
n
(4.40)
APPROXIMATING GEODESICS VIA RANDOM POINTS
29
Similarly, when v is considered, following the same reasoning using Lemma 4.6
and (Lip), we may obtain (4.40) with v in place of w, and moreover
j−1
j−1 Z 1
X
X
f (lvk ,vk+1 (t), vk+1 − vk )dt −
f (vk , vk+1 − vk ) ≤ C 0 2n |j − i| ≤ C 0 2−β .
k=i
0
k=i
Hence, combining these two estimates, we obtain that
j−1 Z 1
j−1
X
X
f (lvk ,vk+1 (t), vk+1 − vk )dt −
f (vi , vk+1 − vk ) ≤ 2C 0 2−β
.
n
k=i
0
(4.41)
k=i
Returning to the path w, by Lemma 4.7, noting (4.40), any path (wi , . . . , wj )
exiting A, the r = α
n -neighborhood of the line segment from wi to wj , is costlier,
with respect to Hn , than the Hn -cost f (wi , wj − wi ) of a straight path connecting
wi to wj in a single hop, as follows:
j−1
X
0 2−β
f (wq , wq+1 − wq ) − f (wi , wj − wi ) ≥ Crα − C 0 2−β
= C2α
. (4.42)
n
n − C n
q=i
Let us now consider v. Since |vi − vj | ≤ Cn by Lemma 4.6, using (Lip), we
have that
Z 1
f (vi , vj − vi ) −
f (lvi ,vj (t), vj − vi )dt ≤ C 0 2n .
(4.43)
0
Following the same reasoning given with respect to w, we may obtain (4.42) with
v in place of w. Then, noting (4.41), a path (vi , . . . , vj ), exiting the r = α
nneighborhood of the line segment from vi to vj , has Ln cost more than the one step
Hn cost f (vi , vj − vi ) by the amount Crα − 2C 0 2−β
. By (4.43), this Hn cost differs
n
R1
from the one step Ln cost 0 f (lvi ,vj (t), vj − vi )dt of moving from vi to vj by C 0 2n .
Therefore, the cost savings of moving in one step, in considering w or v which
2−β
),
= O(2α
exit the r-neighborhood, is bounded below by Crα − 3C 0 2−β
n − n
n
2
which is positive, for all large n, when α < 2 − β.√ This is the case when we fix
β = (2 − α2 )η > 0, for an 0 < η < 1, since 1 < α < 2.
Hence, with this choice of β, such exiting paths are not optimal, and all the points
{wi , . . . , wj } or {vi , . . . , vj } in the box must belong to the r = α
n -neighborhood of
the line segment connecting the ith and jth points.
We note, given the value of β, to use Lemma 4.6 above, the exponent δ should
satisfy δ > [(2 − α2 )η + d]−1 , afforded by our assumptions.
Step 2.. We now focus on w as the following counting argument is the same
with respect to v. We will count the points in the small set A. The cardinality |{wk }jk=i | = |j − i| is bounded by |Xn ∩ A| = 2 + Nn,z , where Nn,z is the
Binomial(n − 2, ν(A)) count of the number of points in Xn distinct from wi , wj in
the set A. Note, as A is nearly a cylinder with length n and radius α
n , and ρ is
α(d−1)+1
bounded, we have that ν(A) ≤ C(ρ)n
. Then,
P (Nn,z ≥ K) ≤ (nK /K!)ν(A)K ≤ C(ρ)K nK nKαd+K(1−α) .
Hence, by a union of events bound, as the number of boxes intersecting D is bounded
by C 0 −d
n , we have
P (∃z ∈ Zd such that Nn,z ≥ K) ≤ C 0 C(ρ)K nK nKαd+K(1−α)−d .
(4.44)
30
ERIK DAVIS AND SUNDER SETHURAMAN
Suppose n is of the form n = n−δ for 0 < δ < 1/d. If d < K < ∞ and
δ > (K + 1)/[K(α(d − 1) + 1) − d],
(4.45)
the display (4.44) is summable in n. In particular, when δ > [α(d − 1) + 1]−1 , part
of our assumptions, a large but fixed K can be chosen so that (4.45) holds.
Hence, by Borel-Cantelli lemma, on the intersection of a probability 1 set and
A1 say, we recover the claim for all large n that the path visits at most K points
between the first and last visit to a visited box.
4.2.2. Convergence of Optimal Paths. We now consider the behavior of the optimal
paths, in analogy to Theorem 2.2, for the energy Hn .
Proposition 4.9. Consider the assumptions for the second parts of Theorems 2.7
and 2.8. Consider a discrete path w(n) ∈ arg min Hn and its linear interpolation,
lw(n) . Then, with respect to realizations in a probability 1 subset of A1 ∩ A2 , for
any subsequence of {lw(n) }, and correspondingly of {w(n) }, there is a further subsequence of the linear paths which converges uniformly to a limit path γ ∈ arg min F ,
and of the discrete paths in the Hausdorff sense to Sγ .
If F has a unique (up to reparametrization) minimizer γ, then the whole sequence
{w(n) } converges, limn→∞ dhaus (w(n) , Sγ ) = 0.
Consider now a path lv(n) ∈ arg min Ln . The same conclusions holds for {v(n) }
in place of {w(n) }.
Proof. Consider first w(n) ∈ arg min Hn . By the compactness criterion, Proposition
4.4, almost surely, any subsequence of the paths {lw(n) } has a further subsequence
{lw(nk ) } converging uniformly to a limit γ. By the ‘liminf’ Lemma 3.2, F (γ) ≤
lim inf k→∞ F (lw(nk ) ).
The same argument and conclusion holds with lv(n) ∈ arg min Ln and v(n) in
place of lw(n) and w(n) .
We now show that γ ∈ arg min F . With respect to optimal Ln paths, as
F (lv(n) ) = min Ln , and min Ln → min F a.s. by Proposition 4.3, we obtain
F (γ) ≤ min F , and so the desired conclusion.
For Hn optimal paths, we recall an argument in the proof of Theorem 2.4. Using
only the ‘standing assumptions’ (allowing p = 1) and (Lip), we derived (4.30),
namely, for u ∈ Vn (a, b), that |Ln (lu ) − Hn (u)| ≤ cm−1
1 n min(Ln (lu ), Hn (u))
where c is the constant in (Lip). Then, as a consequence of Proposition 4.1, we saw
2
in (4.32) that min Ln ≤ Ln (lw(n) ) ≤ (cm−1
1 n + 1) min Ln . Since, by Proposition
4.3, min Ln → min F a.s., we conclude that γ ∈ arg min F .
Finally, we remark that the Hausdorff convergences are argued as in the proof
of Corollary 2.3.
5. Appendix
Here we collect some results which we had previously assumed.
5.1. Nearest-Neighbor Rate.
Proposition 5.1. Let {Xi } be i.i.d. samples from a probability measure ν =
ρ(x) dx on a Lipschitz domain D, and let
Rn = sup min |Xi − y|.
y∈D 1≤i≤n
APPROXIMATING GEODESICS VIA RANDOM POINTS
31
Suppose ρ(x) is uniformly bounded below by a positive constant. Then, there exists
a constant C, independent of n, such that, for almost all realizations {Xi },
lim sup
n→∞
Rn n1/d
≤ C.
(log n)1/d
In particular, when n satisfies (2.6), for a, b ∈ D, almost surely for all large n,
there is a path in Vn (a, b) connecting a, b via the graph Gn .
Proof. We first address the claim with respect to Rn . Let B(y, r) be the Euclidean
ball of radius r centered at y ∈ D. Since D is Lipschitz, there is a constant c such
that m(B(y, r) ∩ D) ≥ cm(B(y, r)) for all small r > 0, where m denotes Lebesgue
measure (cf. the discussion about cone conditions in Section 4.11 of [1]). It follows
that there is a constant c such that m(B(y, r) ∩ D)/m(D) ≥ crd for all y ∈ D and
all small r > 0. Since ν has density ρ bounded below by a positive constant, there
exists a constant c such that ν(B(y, r)) ≥ crd for all y ∈ D and 0 < r < r0 , where
r0 is a sufficiently small constant. Therefore, recalling Xn = {X1 , X2 , . . . , Xn }, we
have
P(|B(y, r) ∩ Xn | = 0) = (1 − ν(B(y, r)))n
(5.46)
d
≤ (1 − crd )n ≤ e−cnr .
Let {y1 , . . . , yk } ⊂ D be a collection of points so that supx∈X min1≤i≤k |x − yi | ≤ r.
We may take the number of points k to satisfy k ≤ c/rd for some constant c
independent of r, say, by choosing {yi } to be a regular grid, with grid length ∼ r.
Let Ei denote the event that |B(yi , r) ∩ Xn | = 0, and consider the event {Rn >
2r} that there exists a y ∈ D with min1≤i≤n |Xi − y| > 2r. Then, by a triangle
inequality argument, we have {Rn > 2r} ⊂ ∪ki=1 Ei . Hence, together with (5.46),
we have
k
X
d
c
P(Rn > 2r) ≤
P(Ei ) ≤ d e−cnr .
(5.47)
r
i=1
Let rd = (3 log n)/(cn). Then, (5.47) gives a summable term,
c2
(3 log n)1/d
≤
.
P Rn > 2
3n2
n1/d
By Borel-Cantelli lemma, Rn ≤ 2(3 log n)1/d /n1/d for all large n.
We now show that Gn (a, b) is connected when n satisfies (2.6). Let v1 , v2 be any
vertices in Xn ∪ {a, b}, and consider the line `(t) = v1 (1 − t) + v2 (t) between them
for t ∈ [0, 1]. By convexity of D, the path ` is contained in D. Consider points on
the path v1 = `(0), `(Rn ), `(2Rn ), . . . , `(kRn ), `(1) = v1 , where k = b|v2 − v1 |/Rn c
so that |1 − kRn | ≤ Rn . Each point y = `(jRn ) is within Euclidean distance Rn
of a point uj ∈ Xn by the ‘Rn -limit’ a.s. for all large n. By construction, for
1 ≤ j ≤ k − 1, the Euclidean distance between uj and uj+1 is less than sum of the
distances, from uj to `(jRn ), from `(jRn ) to `((j + 1)Rn ), and from `((j + 1)Rn ) to
uj+1 , which is bounded by 3Rn . Similarly, the endpoints v1 , v2 are within Euclidean
distance 2Rn of u1 and uk respectively. Since, by (2.6), 3Rn /n < 1 for all large n,
the path along vertices {v1 , u1 , . . . , uk , v2 } belongs to Vn (a, b) and so v1 and v2 are
connected in Gn (a, b) a.s. for all large n.
32
ERIK DAVIS AND SUNDER SETHURAMAN
5.2. Existence of Quasinormal Minimizers. We discuss a ‘conservation law’
for F -minimizing paths, and existence of F -minimizing Lipschitz paths, following
the treatment in [8].
R1
Proposition 5.2. Consider the integral functional F (γ) = 0 f (γ, γ̇) dt, where
f satisfies (A0)-(A3). Then, F attains a minimum on the set Ω(a, b) of Lipschitz paths from a to b. In other words, there exists a γ ∗ ∈ Ω(a, b) with F (γ ∗ ) =
inf γ∈Ω(a,b) F (γ).
In the case that p = 1, there exists a γ ∈ arg min F and constants c, c1 , c2 such
that
f (γ(t), γ̇(t)) = c a.e. t ∈ [0, 1],
(5.48)
and
c1 ≤ |γ̇(t)| ≤ c2
a.e. t ∈ [0, 1].
(5.49)
If p > 1, then there are constants c, c1 , c2 such that (5.48) and (5.49) hold for any
γ ∈ arg min F .
Proof. We first give an argument in the case where p > 1. Note, by assumption, the
integrand f is continuous and C 1 on D × (Rd \ {0}), convex and p-homogenous in
the second argument, and satisfies (2.4). As p > 1, f may be extended continuously
to a C 1 function on D × Rd .
When the domain of F is extended to all Sobolev paths γ ∈ W 1,p ([0, 1]; D) with
γ(0) = a, γ(1) = b, the existence of a minimizer follows from Remark 2 of Section
3.2 in [8]. In particular, the continuity and convexity assumptions (A0) and (A1)
imply that F is lower-semicontinuous with respect to weak convergence of Sobolev
functions. The existence of a minimizer then follows from a standard compactness
argument.
Let γ denote such a Sobolev minimizer. Consider now an ‘inner variation’
ω(t, ) = γ(ξ(t, )) of γ, where ξ is C 1 on [0, 1]×(−0 , 0 ) for some 0 > 0 and ξ(·, ) is
a C 1 diffeomorphism of the interval [0, 1] to itself, with γ(t, 0) = t. It may be shown
(see the discussion on pages 19-21, Proposition 1.16 and Remark 3 in Section 1.1 of
d
F (ω(·, ))|=0 over the class of inner variations,
[8]) that the optimality condition d
together with Euler’s identity for homogenous functions, v · ∇v f (x, v) = pf (x, v),
together imply that
(p − 1)f (γ(t), γ̇(t)) = c a.e. t,
(5.50)
for some constant c.
Finally, by assumption (A3) on f , it follows that c > 0 and there exist constants
c1 , c2 > 0 with c1 ≤ |γ̇(t)| ≤ c2 for almost every t. In particular, γ ∈ Ω(a, b), and
the proposition is proved for p > 1.
The argument for the p = 1 case is complicated by a lack of compactness with
respect to weak convergence in the Sobolev space W 1,1 ([0, 1]; D), as well as difficulty in establishing an analogue of (5.50). By a more involved argument, relating
R1
optimizers of F to optimizers of the quadratic functional Q(γ) := 0 f 2 (γ, γ̇) dt, the
existence of a Lipschitz path γ ∈ arg min F satisfying (5.48) is established in Theorem 1 of [16] (see also Theorem 5.22 of [8] which gives an alternative argument).
From this and assumption (A3), inequality (5.49) follows.
Acknowledgement. This work was partially supported by ARO W911NF-14-10179.
APPROXIMATING GEODESICS VIA RANDOM POINTS
33
References
[1] Adams, R. and Fournier, J. (2003). Sobolev Spaces Academic Press
[2] Agronowich, M.S. (2010). Sobolev Spaces, their Generalizations, and Elliptic Problems in
Smooth and Lipschitz Domains. Springer Monographs in Mathematics
[3] Alamgir, M., von Luxburg, U. (2012). Shortest path distance in random k-nearest neighbor
graphs. in Proc. 29th Int. Conf. on Machine Learning; arXiv:1206.6381
[4] Auffinger, A., Damron, M., Hanson, J. (2016). 50 years of first passage percolation.
arXiv:1511.03262
[5] Beardwood, J., Halton, J.H., Hammersley, J.M. (1959). The shortest path through many
points. Math. Proc. Cambridge Philosophical Soc. 55 299–327.
[6] Bern, M. et. al. (1999). Emerging Challenges in Computational Topology. NSF report.
arXiv:cs/9909001
[7] Bernstein, M., De Silva, V., Langford, J. and Tenenbaum, J. (2000). Graph approximations to geodesics on embedded manifolds. Technical Report, Department of Psychology,
Stanford University
[8] Buttazzo, G., Giaquinta, M. and Hildebrandt, S. (1998). One-dimensional variational
problems: an introduction. Oxford University Press, Oxford
[9] Burago, D., Burago, Y., and Ivanov, S. (2001). A Course in Metric Geometry. American
Mathematical Society, Providence, R.I.
[10] Cabello, S., Jejcic, M. (2014). Shortest paths in intersection graphs of unit disks.
arXiv:1402.4855
[11] Crane, K., Weischedel, C., Wardetzky, M. (2013). Geodesics in heat: A new approach
to computing distance based on heat flow. ACM Trans. Graph. 28 106
[12] Giesen, J. and Wagner, U. (2004). Shape dimension and intrinsic metric from samples of
manifolds. Discrete & Computational Geometry 32 245–267.
[13] von Deylen, S.W., Glickenstein, D., Wardetzky, M. (2016). Distortion estimates for
barycentric coordinates on Riemannian simplices. arXiv:1610.01168
[14] Gelfand, I.M., Smirnov, M.M. (1994). Lagrangians satisying Crofton formulas, Radon
transforms, and nonlocal differentials. Adv. in Math. 109 188–227.
[15] Hashimoto, T.B, Sun, Y., and Jaakkola, T.S. (2015). Metric recovery from directed unweighted graphs. in Proc. 18th Int. Conference on AI and Stat (AISTATS) San Diego, CA.
JMLR: W&CP 38.
[16] Hildebrandt, S. (2016). Minimizers of one-dimensional parametric variational integrals. St.
Petersburg Math. J. 27 569–576.
[17] Hirsch, C., Neuhäuser, D., Gloaguen, C., Schmidt, V. (2015). First passage percolation
on random geometric graphs and an application to shortest-path trees. Adv. Appl. Probab.
47 328–354.
[18] Howard, D., Newman, C.M. (1997). Euclidean models of first-passage percolation. Probab.
Theory Relat. Fields 108 153–170.
[19] Howard, D., Newman, C.M. (2001). Geodesics and spanning trees for Euclidean first-passage
percolation. Ann. Probab. 29 577–623.
[20] Hwang, S.J., Damelin, S.B., Hero, A.O. (2016). Shortest path through random points.
Ann. Appl. Probab. 26 2791–2823.
[21] LaGatta, T., Wehr, J. (2010). A shape theorem for Riemannian first-passage percolation.
J. Math. Phys. 51 053502
[22] LaGatta, T., Wehr, J. (2014). Geodesics of random Riemannian metrics. Commun. Math.
Phys. 327 181–241.
[23] Maz’ya, V. (2011). Sobolev Spaces with Application to Elliptic Partial Differential Equations.
2nd Ed. Springer. Grundlehren der mathematischen Wissenschaften 342
[24] Mémoli, F. and Sapiro, G. (2005). Distance functions and geodesics on submanifolds of Rd
and Point Clouds. SIAM Journal on Applied Mathematics 65 1227–1260.
[25] Alvarez Paiva, J.C. (2006). Some problems on Finsler Geometry. in Handbook of Differential
Geometry Vol. II Ed. Dillen, F.J.E., Verstraelen, L.C.A., Elsevier, Chapter 1, 1–33.
[26] Papadopoulos, A. (2013). On Hilbert’s fourth problem. in Handbook of Hilbert geometry
Ed. Papadopoulos, A. and Troyanov, M., European Mathematical Society. arXiv:1312.3172
[27] Penrose, M. (2003). Random Geometric Graphs. Oxford University Press, Oxford.
34
ERIK DAVIS AND SUNDER SETHURAMAN
[28] Peyré, G., Péchaud, M., Keriven, R., Cohen, L., and others. (2010). Geodesic methods
in computer vision and graphics. Foundations and Trends in Computer Graphics and Vision
5 197–397.
[29] Sajama, Orlitzky, A. (2005). Estimating and computing density based distance metrics. in
Proc. 22nd Int. Cont. on Machine Learning Bonn, Germany.
[30] Sethian, J. (1999). Fast marching methods. SIAM Review 41 199–235.
[31] Tamassy, L. (2008). Relation between metric spaces and Finsler spaces. Differential Geometry and its Applications 26 483–494.
[32] Tenenbaum, J.B., de Silva, V., Langford, J.C. (2000). A global geometric framework for
nonlinear dimensionality reduction. Science. 290 2319–2323.
[33] Yu, H., Zhang, J.J., Jiao, Z. (2014). Geodesics on point clouds. Mathematical Problems in
Engineering 2014 860136
Department of Mathematics, University of Arizona, Tucson, AZ 85721
e-mail: [email protected]
Department of Mathematics, University of Arizona, Tucson, AZ 85721
e-mail: [email protected]
| 10math.ST
|
An analysis of the SPARSEVA estimate for the finite sample data case ?
Huong Ha a , James S. Welsh a , Cristian R. Rojas b , Bo Wahlberg b
arXiv:1703.09351v1 [math.ST] 27 Mar 2017
a School
b Department
of Electrical Engineering and Computer Science, The University of Newcastle, Australia
of Automatic Control and ACCESS, School of Electrical Engineering, KTH Royal Institute of Technology, SE-100 44
Stockholm, Sweden
Abstract
In this paper, we develop an upper bound for the SPARSEVA (SPARSe Estimation based on a VAlidation criterion) estimation error in
a general scheme, i.e., when the cost function is strongly convex and the regularized norm is decomposable for a pair of subspaces. We
show how this general bound can be applied to a sparse regression problem to obtain an upper bound for the traditional SPARSEVA
problem. Numerical results are used to illustrate the effectiveness of the suggested bound.
Key words: SPARSEVA estimate; upper bound; finite sample data.
1
Introduction
Regularization is a well known technique for estimating
model parameters from measured input-output data. Its applications are in any fields that are related to constructing
mathematical models from observed data, such as system
identification, machine learning and econometrics. The idea
of the regularization technique is to solve a convex optimization problem constructed from a cost function and a
weighted regularizer (regularized M-estimators). There are
various types of regularizers that have been suggested so far,
such as the l1 [18], l2 [19] and nuclear norms [5] [6].
During the last few decades, in the system identification
community, regularization has been utilised extensively [14].
Most of this work has focused on analysing the asymptotic
properties of an estimator, i.e., when the length of the data
goes to infinity. The purpose of this type of analysis is to
evaluate the performance of the estimation method to determine if the estimate is acceptable. However, in practice,
the data sample size for any estimation problem is always
finite, hence, it is difficult to judge the performance of the
estimated parameters based on asymptotic properties, especially when the data length is short.
Recently, a number of authors have published research ([1],
? The material in this paper was not presented at any conference.
Email addresses: [email protected] (Huong Ha),
[email protected] (James S. Welsh),
[email protected] (Cristian R. Rojas),
[email protected] (Bo Wahlberg).
Preprint submitted to Automatica
[4], [12]) aimed at analysing estimation error properties of
the regularized M-estimators when the sample size of the
data is finite. Specifically, they develop upper bounds on the
estimation error for high dimensional problems, i.e., when
the number of parameters is comparable to or larger than the
sample size of the data. Most of these activities are from the
statistics and machine learning communities. Among these
works, the paper [12] provides a very elegant and interesting framework for establishing consistency and convergence
rates of estimates obtained from a regularized procedure under high dimensional scaling. It determines a general upper bound for regularized M-estimators and then shows how
it can be used to derive bounds for some specific scenarios. However, this framework is only applicable for the case
when the number of parameters is comparable, or larger,
than the sample size of the data whereas in a typical system
identification problem, the number of parameters is generally smaller than the sample size of the data.
Here in this paper we utilize the framework suggested in
[12] to develop an upper bound for the estimation error of
the M-estimators used in a system identification problem.
Here, the M-estimator problems are implemented using the
SPARSEVA (SPARSe Estimation based on a VAlidation criterion) framework [15], [16]. Our aim is to derive an upper
bound for the estimation error of the general SPARSEVA
estimate. We then apply this bound to a sparse linear regression problem to obtain an upper bound for the traditional
SPARSEVA problem. In addition, we also provide numerical simulation results to illustrate the suggested bound of
the SPARSEVA estimation error.
2 April 2018
The paper is organized as follows. Section 2 formulates
the problem. Section 3 provides definitions and properties
required for the later analysis. The general bound for the
SPARSEVA estimation error is then developed in Section 4.
In Section 5, we apply the general bound to the special case
when the model is cast in a linear regression framework.
Section 6 illustrates the developed bound by numerical simulation. Finally, Section 7 provides conclusions.
2
based the chosen validation criterion. For example, as suggested in [15] [16], εN can be chosen as 2n/N (Akaike Information Criterion (AIC)), nlog(N)/N (Bayesian Information Criterion (BIC)); or as suggested in [8], n/N (Prediction
Error Criterion).
For the traditional regularization method described in (1),
[12] recently developed an upper bound on the estimation
error between the estimate θ̂λN and the unknown parameter
vector θ∗ . This bound is a function of some constants related
to the nature of the data, the regularization parameter λN ,
the cost function L and the data length N. The beauty of
this bound is that it quantifies the relationship between the
estimation error and the finite data length N. Through this
relationship, it is easy to confirm most of the properties of the
estimate θ̂λN in the asymptotic scenario, i.e. N → ∞, which
were developed in the literature some time ago ([9], [11]).
Problem Formulation
Let Z1N = {Z1 , ..., ZN } denote N identically distributed observations with marginal distribution P. L : Rn × Z N →
R denotes a convex and differentiable cost function. Let
θ∗ ∈ argminθ∈Rn L (θ) be a minimizer of the population risk
L (θ) = EZ N [L (θ; Z1N )].
1
Inspired by [12], our goal is to derive a similar bound for the
SPARSEVA estimate θ̂εN , i.e, we want to know how much
the SPARSEVA estimate θ̂εN differs from the true parameter
θ∗ when the data sample size N is finite. Note that the notation and techniques used in this paper are similar to [12];
however, in [12], the convex optimization problem is posed
in the traditional regularization framework (1), while in this
paper, the optimization problem is based on the SPARSEVA
regularization (2).
The task here is to estimate the unknown parameter θ∗ from
the data Z1N . A well known approach to this problem is to
use a regularization technique, i.e., to solve the following
convex optimization problem,
θ̂λN ∈ arg min {L (θ; Z1N ) + λN R (θ)},
(1)
θ∈Rn
where λN > 0 is a user-defined regularization parameter and
R : Rn → R+ is a norm.
3
A difficulty when estimating the parameter θ∗ using the
above regularization technique is that one needs to find
the regularization parameter λN . The traditional method to
choose λN is to use cross validation, i.e., to estimate the parameter θ∗ with different values of λN , then select the value
of λN that provides the best fit to the validation data. This
cross validation method is quite time consuming and very
dependent on the data. Here we are specifically interested
in the SPARSEVA (SPARSe Estimation based on a VAlidation criterion) framework, suggested in [15] and [16], which
provides automatic tuning of the regularization parameters.
Utilizing the SPARSEVA framework, an estimate of θ∗ can
be computed using the following convex optimization problem:
In this section, we provide descriptions of some definitions and properties of the norm R (θ) and the cost function
L (θ; Z1N ), needed to establish an upper bound on the estimation error. Note that we only provide a brief summary such
that the research described in this paper can be understood.
Readers can find a more detailed discussion in [12].
3.1
s.t.
(2)
L (θ; Z1N ) ≤ L (θ̂NR ; Z1N )(1 + εN ),
The norm R is said to be decomposable with respect to
where εN > 0 is the regularization parameter and θ̂NR is the
“non-regularized” estimate obtained from minimizing the
cost function L (θ; Z1N ), i.e.
θ̂NR ∈ arg min
θ∈Rn
L (θ; Z1N ).
Decomposability of a Norm
Let us consider a pair of linear subspaces M ⊆ M of Rn .
The orthogonal complement of the space M is then defined
as,
⊥
M = {v ∈ Rn |hu, vi = 0 for all u ∈ M },
where h·, ·i is the inner product that maps Rn × Rn → R.
θ̂εN ∈ arg min R (θ)
θ∈Rn
Definitions and Properties of the Norm R (θ) and the
Cost Function L (θ)
⊥
(M , M ) if
R (θ + γ) = R (θ) + R (γ)
(3)
(4)
⊥
for all θ ∈ M and γ ∈ M .
There are many combinations of norms and vector spaces
that satisfy this property (cf. [12]). An example is the l1 norm
The advantage of the SPARSEVA framework is that there
are some natural choices of the regularization parameter εN
2
to the pair (R , k · k) is defined as,
and the sparse vector space defined (5). For any subset S ⊆
{1, 2, . . . , n} with cardinality s, define the model subspace
M as,
M (S) = {θ ∈ Rn | θ j = 0 for all j 6∈ S}.
Ψ(M ) =
u∈M \{0}
(5)
⊥
M (S) = {γ ∈ Rn |γ j = 0 for all j ∈ S}.
3.5
As shown in [12], the l1 -norm is decomposable with respect
to the pair (M (S), M ⊥ (S)).
(9)
ΠM (u) = arg min ku − vk2 .
hu, vi
= sup hu, vi,
u∈Rn \{0} R (u)
R (u)≤1
sup
In the sequel, to simplify the notation, we will write uM to
denote ΠM (u).
(6)
4
A twice differentiable function L (θ) : Rn → R is strongly
convex on Rn when there exists an m > 0 such that its Hessian
52 L (θ) satisfies,
Theorem 4.1 Assume R is a norm and is decomposable
⊥
with respect to the subspace pair (M , M ) and the cost
function L (θ) is differentiable and strongly convex with curvature κL . Consider the SPARSEVA problem in (2), then the
following properties hold:
(7)
for all θ ∈ Rn [3]. This is equivalent to the statement that
the minimum eigenvalue of 52 L (θ) is not smaller than m
for all θ ∈ Rn .
An interesting consequence of the strong convexity property
in (7) is that for all θ, ∆ ∈ Rn , we have,
m
2
L (θ + ∆) ≥ L (θ) + 5L (θ)T ∆ + k∆k22 .
Analysis of the Regularization Technique using the
SPARSEVA
In this section, we apply the properties described in Section
3 to derive an upper bound on the error between the SPARSEVA estimate θ̂εN and the unknown parameter θ∗ . This
upper bound is described in the following theorem.
Strong Convexity
52 L (θ) mI,
(10)
v∈M
where sup is the supremum operator.
Based on the above definition, one can easily see that the
dual of the l1 norm, with respect to the Euclidean inner
product, is the l∞ norm [12].
i. When εN > 0, there exists a Lagrange multiplier, λN =
λεN , such that (1) and (2) have the same solution.
ii. Any optimal solution θ̂εN 6= 0 of the SPARSEVA problem
(2) satisfies the following inequalities:
• If εN is chosen such that
(8)
λεN ≤ 1/R ∗ (∇L (θ∗ )),
The inequality in (8) has a geometric interpretation in that
the graph of the function L (θ) has a positive curvature at
any θ ∈ Rn . The term m/2 for the largest m satisfying (7) is
typically known as the curvature of L (θ).
3.4
.
Projection Operator
For a given inner product h·, ·i, the dual of the norm R is
defined by,
3.3
kuk
The projection of a vector u onto a space M , with respect
to the Euclidean norm, is defined by the following,
Dual Norm
R ∗ (v) =
R (u)
This quantity measures how well the norm R is compatible
with the error norm k.k over the subspace M . As shown in
[12], when M is Rs , the regularized norm R is the l1 norm,
and the error norm is the l2 norm,
√ then the subspace compatibility constant is Ψ(M ) = s. Notice also that Ψ(M )
is finite, due to the equivalence of finite dimensional norms.
Now if we define M (S) = M (S), then the orthogonal complement M (S), with respect to the Euclidean inner product,
can be computed as follows,
3.2
sup
then
4
4
Ψ2 (M ) +
R (θ∗M ⊥ ).
κL λεN
κ2L λ2εN
(11)
• If εN is chosen such that
kθ̂εN − θ∗ k22 ≤
Subspace Compatibility Constant
For a given norm R and an error norm k · k, the subspace
compatibility constant of a subspace M ∈ Rn with respect
λεN > 1/R ∗ (∇L (θ∗ )),
3
with θ̂NR = (ΦN ΦTN )−1 ΦN YN and εN > 0 being the userdefined regularization parameter. Now εN can be chosen as
either 2n/N or log(N)n/N as suggested in [15]; or n/N as
suggested in [8].
then
kθ̂εN − θ∗ k22 ≤
2
{R ∗ (∇L (θ∗ ))}2 Ψ2 (M )
κ2L
⊥
8
+ 2 {R ∗ (∇L (θ∗ ))}2 Ψ2 (M )
κL
4
+
R (θ∗M ⊥ ).
κL λεN
(12)
Remark 5.1 Note that the sparse regression problem is very
common in system identification and is often used to obtain
a low order linear model by regularization.
5.1
Proof. See the Appendix (Section A.2).
Consider the convex optimization problem in (15), the Hessian matrix of the cost function L (θ) is computed as,
Remark 4.1 Note that Theorem 4.1 is intended to provide
an upper bound on the estimation error for the general
SPARSEVA problem (2). At this stage, it is hard to evaluate,
or quantify, the value on the right hand side of the inequalities (11) and (12) as they still contain the term λεN and other
abstract terms. However, in the later sections of this paper,
from this general upper bound, we will provide bounds on
the estimation errors for some specific scenarios.
52 L (θ) =
⊥
(M , M ), there is one bound for the estimation error.
Hence, in the usual sense, to apply Theorem 4.1 for any
5
⊥
1
ΦN ΦTN .
N
In order to prove that L (θ) is strongly convex, we need to
prove,
1
ΦN ΦTN 2κL I.
(16)
∃κL > 0 s.t.
N
We see that the requirement in (16) coincides with the requirement of persistent excitation of the input signal in a
system identification problem. If an experiment is welldesigned, then the input signal u(t) needs to be persistently
exciting of order n, i.e., the matrix ΦN ΦTN is a positive definite matrix. This means that the condition in (16) is always
satisfied for any linear regression problem derived from a
well posed system identification problem. Hence, the cost
function L (θ) of the optimization problem in (15) is generally strongly convex. This means that for any choice of the
regression matrix ΦN that satisfies the persistent excitation
condition, there exists a positive curvature κL of the cost
function L (θ).
Remark 4.2 The bound in Theorem 4.1 is actually a family of bounds. For each choice of the pair of subspaces
specific scenario, the goal is to choose M and M
obtain an optimal rate of the bound.
An Analysis on the Strong Convexity Property and the
Curvature of the l2 norm Cost Function
to
An Upper Bound for Sparse Regression
In this section, we illustrate how to apply Theorem 4.1 to
derive an upper bound of the error between the SPARSEVA
estimate θ̂εN and the true parameter θ∗ for a linear regression
model.
where θ∗ ∈ Rn is the unknown parameter that is required to
be estimated; e ∈ RN is the disturbance noise; ΦN ∈ Rn×N
is the regression matrix and YN ∈ RN is the output vector.
We now need to consider whether there exists a “global” curvature κ for any regression matrix ΦN . Consider ΦN ∈ Rn×N
to be any random matrix in which each column ΦN, j is independently sampled from a Normal distribution of zero mean
and covariance matrix Σ ∈ Rn×n (Σ-Gaussian ensemble), i.e.,
ΦN, j ∼ N (0, Σ), ∀ j = 1, .., N. The question we then need
to address is whether there exists a constant value κ, which
only depends on Σ, that gives,
Using the SPARSEVA framework in (2) with R chosen as
the l1 norm and the cost function L (θ) chosen as,
1
ΦN ΦTN 2κI,
N
Consider the following linear regression model,
YN = ΦTN θ∗ + e,
L (θ) =
1
kYN − ΦTN θk22 ,
2N
(13)
for any random matrix ΦN sampled from a Σ-Gaussian ensemble. Notice that, in the case N n, κ can be chosen as
0.5 times the smallest eigenvalue of N −1 ΦN ΦTN . To answer
this question, we investigate the distribution of the smallest
eigenvalue of N −1 ΦN ΦTN .
(14)
then an estimate of θ∗ in (13) can be found by solving the
following problem,
The distribution of the smallest eigenvalue of N −1 ΦN ΦTN has
been discussed frequently in the literature, especially from
the principal component analysis (PCA) community. It is
θ̂εN ∈ arg min kθk1
θ∈Rn
s.t.
(17)
(15)
L (θ) ≤ L (θ̂NR )(1 + εN ),
4
well known that when N n and the matrix ΦN being ΣGaussian ensemble, there exists a distribution of the smallest
eigenvalue of the covariance matrix N −1 ΦN ΦTN whose parameters only depend on the eigenvalues of Σ, N and n [10].
A closed form for this distribution can be found in [10].
5.3
In this paper, we will use the following notation:
• f (x|0, σ2 ) = (2πσ2 )−1/2 exp(−x2 /2σ2 ) denotes the probability density function (pdf) of the Normal distribution
N (0, σ2 ).
• χ2β (N) denotes the value that P(X < χ2β (N)) = 1 − β,
where X is Chi square distributed with N degrees of
freedom.
The existence of the distribution of the smallest eigenvalue of
N −1 ΦN ΦTN , which we denote P(x|Σ, N, n), means that given
a probability 1 − α , 0 ≤ α ≤ 1, there exists a value wmin
such that N −1 ΦN ΦTN wmin I, for any random Σ-Gaussian
ensemble matrix ΦN . Then κ in (17) can be expressed as
(1/2)wmin . In this paper, we will denote by κα lower bound
on the global curvature in (17) with probability 1 − α , 0 ≤
α ≤ 1.
5.4
Developing the Upper Bound
We now present three propositions that assist in the development of the upper bound on the estimation error kθ̂εN − θ∗ k2
for the optimization problem in (15).
Related to the computation of the universal constant κ corresponding to the distribution N (0, Σ), note that, in reality, it is
very difficult to compute the exact distribution P(x|Σ, N, n)
using the formula in [10]. Here we suggest to use an empirical method to compute the distribution P(x|Σ, N, n) due
to inexact knowledge of the formula for the distribution.
The idea is to generate a large number of random matrices ΦN , compute the smallest eigenvalue of N −1 ΦN ΦTN , and
then build a histogram of these values, which is an approximation of P(x|Σ, N, n). Then one can compute the value
of wmin to ensure the inequality N −1 ΦN ΦTN wmin I occurs
with probability 1 − α. Finally, compute κα using the formula κα = (1/2)wmin .
5.2
Notation
Proposition 5.1 Consider the optimization problem in (15),
and denote by λεN the corresponding Lagrange multiplier
of its constraint. Then, if θ̂εN 6= 0, λεN can be computed as
1
.
k∇L (θ̂εN )k∞
(19)
Proof. See the Appendix (Section A.3).
λεN =
Assumptions
Proposition 5.2 Suppose Assumptions 5.1 and 5.2 hold.
Then, with probability 1 − nβ (0 ≤ β ≤ 1), we have
For the linear regression in (13), the following assumptions
are made:
Assumption 5.1 The column ΦN, j , j = 1, ..., N of the regressor matrix ΦN are mutually independent, and they are distributed as ΦN, j ∼ N (0, Σ), where Σ ∈ Rn×n is a constant,
symmetric, positive definite matrix.
ΦN e
N
P
≤ t ΦN
∞
"
N 2t 2
≥ 1 − 2 exp − 2
2σe smax χ2β (N)
#!n
,
where smax is the maximum element on the diagonal of the
matrix Σ. In particular,
Assumption 5.2 The noise vector e ∈ RN is Gaussian with
i.i.d. N (0, σ2e ) entries.
q
∗
∇L (θ )
Assumption 5.3 The true parameter θ∗ is “weakly” sparse,
i.e. θ∗ ∈ Bq (Rq ), where,
≤
2σ2e smax χ2β (N) ln(2/β)
∞
N
,
(20)
with probability at least 1 − 2nβ.
(
Bq (Rq ) :=
θ ∈ Rp
p
∑ |θi |q ≤ Rq
)
,
(18)
i=1
Proof. See the Appendix (Section A.4).
with q ∈ [0, 1] being a constant.
Proposition 5.3 Suppose Assumptions 5.1 and 5.2 hold,
then with probability at least 1 − nβ (0 ≤ β ≤ 1), we have
For Assumption 5.3, note that when q = 0, under the convention that 00 = 0, the set in (18) corresponds to an exact
sparsity set, where all the elements belonging to the set have
at most R0 non-zero entries. Generally, for q ∈ (0, 1], the set
Bq (Rq ) forces the ordered absolute values of θ∗ to decay
with a certain rate.
P
∇L (θ̂εN )
(
≥
5
≤t e
∞
N 2t 2
1 − 2 exp − 2
2σe smax χ2β (N − n)(1 + εN )
!)n
Using the definition of the subspace compatibility constant
described in Section 3, we have,
where smax is the maximum element on the diagonal of the
matrix Σ. In particular,
q
∇L (θ̂εN )
≤
2σ2e smax χ2β (N − n)(1 + εN ) ln(2/β)
N
∞
Ψ2 (M (Sη )) = |Sη | = nη ,
Ψ2 (M ⊥ (Sη )) = |Sηc | = n − nη .
,
(21)
where |S| denotes the cardinality of S.
with probability at least 1 − 2nβ.
Now, for Theorem 4.1 to generate an upper bound for
the problem (15), we need to establish an upper bound
on kθ∗M ⊥ (S ) k1 . Based on the definition of the subspace
Proof. See the Appendix (Section A.5).
(25)
η
M ⊥ (Sη ), we have,
From the three propositions above, we can now state the following theorem that gives an upper bound on the estimation
error in the case of weakly sparse estimates.
kθ∗M ⊥ (Sη ) k1 = kθ∗[nη +1:n] k1 ,
Theorem 5.1 Suppose Assumptions 5.1, 5.2 and 5.3 hold,
when N is large, then with probability (1 − α)(1 − 4nβ)
(0 ≤ α ≤ 1, 0 ≤ β ≤ 1), if θ̂εN 6= 0 we have the following
inequality
where θ∗[nη +1:n] denotes the vector formed from the n − nη
smallest (in magnitude) entries of θ∗ . Define κα as a lower
bound on the global curvature of the regression matrix, i.e.
half the smallest eigenvalue of ΦTN ΦN , with probability 1 −
α, 0 ≤ α ≤ 1. Substituting (19), (20), (21), (25) and (26) into
the bound in Theorem 4.1, then with nη being any integer
between 1 and n, we have the following bounds:
kθ̂εN − θ∗ k22 ≤ max(a1 , a2 ),
(22)
where
a1 =
a2 =
• If εN is chosen such that
8nη σ2e smax χ2β (N − n)(1 + εN ) ln(2/β)
+
κ2α N 2
q
32σ2e smax χ2β (N − n)(1 + εN ) ln(2/β)
κα N
2
(16n − 12nη )σe smax χ2β (N) ln(2/β)
λεN ≤ 1/R ∗ (∇L (θ∗ )) = 1/k∇L (θ∗ )k∞ ,
then, with probability at least (1 − α)(1 − 2nβ),
kθ∗[nη +1:n] k1 ,
kθ̂εN − θ∗ k22
8nη σ2e smax χ2β (N − n)(1 + εN ) ln(2/β)
≤
κ2α N 2
q
32σ2e smax χ2β (N − n)(1 + εN ) ln(2/β)
+
kθ∗[nη +1:n] k1 .
κα N
κ2α N 2
q
+
(26)
32σ2e smax χ2β (N − n)(1 + εN ) ln(2/β)
κα N
kθ∗[nη +1:n] k1 ,
where κα is a lower bound on the curvature of the regression matrix (i.e., half the smallest eigenvalue of N −1 ΦTN ΦN )
with probability 1 − α, nη is any integer between 1 and n,
θ∗[nη +1:n] is the vector formed from the n − nη smallest (in
magnitude) entries of θ∗ , and smax is the maximum element
on the diagonal of the matrix Σ.
• If εN is chosen such that
λεN > 1/R ∗ (∇L (θ∗ )) = 1/k∇L (θ∗ )k∞ ,
then, with probability at least (1 − α)(1 − 4nβ),
kθ̂εN − θ∗ k22
(16n − 12nη )σ2e smax χ2β (N) ln(2/β)
≤
κ2α N 2
q
32σ2e smax χ2β (N − n)(1 + εN ) ln(2/β)
+
kθ∗[nη +1:n] k1 .
κα N
Proof. For an integer nη ∈ {1, . . . , n}, define Sη as the set
of the indices of the nη largest (in magnitude) entries of θ∗ ,
and its complementary set Sηc as
Sηc = {1, 2, ..., n} \ Sη ;
(23)
with the corresponding subspaces M (Sη ) and M ⊥ (Sη ) as,
M (Sη ) = {θ ∈ Rn | θ j = 0 ∀ j 6∈ Sη },
M ⊥ (Sη ) = {γ ∈ Rn | γ j = 0 ∀ j ∈ Sη }.
Therefore, for nη being any integer between 1 and n, with
probability at least (1 − α)(1 − 4nβ), we have
(24)
kθ̂εN − θ∗ k22 ≤ max(a1 , a2 ),
6
(27)
6
where
a1 =
8nη σ2e smax χ2β (N − n)(1 + εN ) ln(2/β)
+
κ2α N 2
q
32σ2e smax χ2β (N − n)(1 + εN ) ln(2/β)
κα N
In this section, numerical examples are presented to illustrate
the bound kθ̂εN − θ∗ k22 as stated in Theorem 5.1. In Section
6.1, we consider the case when the input is Gaussian white
noise whilst in Section 6.2, the input is a correlated signal
with zero mean.
kθ∗[nη +1:n] k1 ,
6.1
a2 =
(16n − 12nη )σ2e smax χ2β (N) ln(2/β)
+
32σ2e smax χ2β (N − n)(1 + εN ) ln(2/β)
κα N
kθ∗[nη +1:n] k1 .
Remark 5.2 The bound in Theorem 5.1 is also a family of
bounds, one for each value of nη .
The FIR model structure is used here in order to construct the
SPARSEVA problem (15). The number of parameters n of
the FIR model is set to be 35. The regularization parameter
εN is chosen as n/N [8].
Remark 5.3 Note that the developed bound in Theorem 5.1
depends on the true parameter θ∗ , which is unknown but
constant. Using a similar proof as in Proposition 2.3 of [7],
we can derive under Assumption 5.3 an upper bound for the
term kθ∗[nη +1:n] k1 . Specifically, we have,
kθ∗[nη +1:n] k1 =
=
n
∑
|θ∗[i] |
∑
|θ∗[i] |1−q |θ∗[i] |q
i=nη +1
n
We then compute the upper bound of kθ̂εN − θ∗ k2 using
(22) with different values of nη , i.e. nη = [10, 15, 25]. The
probability parameters α and β are chosen to be 0.02 and
0.001 respectively. With this setting, the probability of the
upper bound being correct is (1 − α)(1 − 4nβ) = 0.84. This
upper bound will be compared with kθ̂εN − θ∗ k2 . Note that
we plot both the upper bound and the true estimation errors
on a logarithmic scale.
i=nη +1
≤ |θ∗nη |1−q
n
∑
Gaussian White Noise Input
In this section, a random discrete time system with a random model order between 1 and 10 is generated using the
command drss from Matlab. The system has poles with
magnitude less than 0.9. Gaussian white noise is added
to the system output to give different levels of SNR, e.g.
30dB, 20dB and 10dB. For each noise level, 50 different input excitation signals (Gaussian white noise with
variance 1) and output noise realizations are generated.
For each set of input and output data, the system parameters are estimated using a different sample size, i.e.,
N = [450, 1000, 5000, 10000, 50000, 100000].
κ2α N 2
q
Numerical Evaluation
|θ∗[i] |q
i=nη +1
Plots of the estimation error versus the data length N with
different noise levels are displayed in Figures 1 to 3. In Figures 1 to 3, the red lines are the true estimation errors from 50
estimates using the SPARSEVA framework. The magenta,
blue and cyan lines are the upper bounds developed in Theorem 5.1, which correspond to nη = [10, 15, 25], respectively.
We can see that the plots confirm the bound developed in
Theorem 5.1 for all noise levels. When N becomes large,
the estimation error and the corresponding upper bound become smaller. When N goes to infinity, the estimation error
will tend to 0. Note that the bounds are slightly different for
the chosen values of nη , however, not significantly. As can
be seen, the bounds are relatively insensitive to the choice
of nη .
(1−q)/q n
1 nη
∗ q
|θ
|
≤
∑ [i]
∑ |θ∗[i] |q
nη i=1
i=nη +1
(1−q)/q n
1 n
≤
∑ |θ∗[i] |q
∑ |θ∗[i] |q
nη i=1
i=1
1 1/q−1
∗ q
≤
kθ∗ k1−q
q kθ kq
nη
≤ (nη )1−1/q kθ∗ kq
≤ (nη )1−1/q (Rq )1/q .
Hence, we can always place an upper bound on the term
kθ∗[nη +1:n] k1 by a known constant which depends on the nature of the true parameter θ∗ .
In addition, we plot another graph, shown in Fig. 4, to compare the proposed upper bound and the true estimation ern
rors corresponding to different value of εN , i.e.
(PEC),
N
2n
log(N)n
(AIC) and
(BIC). The blue lines are the upN
N
per bounds developed in Theorem 5.1, which correspond to
Therefore, from Theorem 5.1, we can see that the estimation
error kθ̂εN − θ∗ k22 = O p (N −1/2 ) [16]. This confirms the result in [16], that in the asymptotic case, when εN > 0, the
SPARSEVA estimate θ̂εN converges to the true parameter θ∗ .
7
-0.5
-0.5
n =10
Developed bounds, n =25
2
2
n =15
-1
True estimation errors (BIC)
True estimation errors (AIC)
True estimation errors (PEC)
-1
2
n =25
2
True estimation errors
-1.5
^ 0 k2 )
log(k "
N
^ 0 k2 )
log(k "
N
-1.5
-2
-2.5
-2.5
-3
-3.5
0
-2
-3
1
2
3
4
5
N
6
7
8
9
-3.5
0
10
#10
1
2
3
4
4
Fig. 1. Plot of the estimation error for a SNR=30dB and a Gaussian
white input signal.
5
N
6
7
8
9
10
#10
4
Fig. 4. Plot of the proposed bound and the true estimation errors
corresponding to different choices of εN , for a SNR=30dB and a
white Gaussian input signal (magenta BIC, green AIC and red
PEC).
0
n =10
choices of εN .
2
n =15
-0.5
2
n =25
^ 0 k2 )
log(k "
N
2
True estimation errors
-1
6.2
In this section, a random discrete time system with a random
model order between 1 and 10 is generated using the command drss from Matlab. The system has poles with magnitude less than 0.9. White noise is added to the system output
with different levels of SNR, e.g. 30dB, 20dB and 10dB. For
each noise level, 50 different input excitation signals and output noise realizations are generated. For each set of input and
output data, the system parameters are estimated using different sample sizes, i.e. N = [450, 1000, 5000, 10000, 50000].
-1.5
-2
-2.5
-3
0
1
2
3
4
5
N
6
7
8
9
10
#10
Coloured Noise Input
4
Fig. 2. Plot of the estimation error for a SNR=20dB and a Gaussian
white input signal.
Here, the input signal is generated by filtering a zero mean
Gaussian white noise with unit variance through the filter,
0.5
n =10
2
Fu (q) =
n =15
0
2
n =25
2
True estimation errors
^ 0 k2 )
log(k "
N
-0.5
Due to this filtering, the covariance matrix of the regression
matrix distribution will not be of a diagonal form. In addition, to ensure that the columns of the regression matrix are
mutually independent, when constructing the regression matrix, we only take the output y(tk ) every nTs seconds, where
Ts is the sampling time.
-1
-1.5
-2
-2.5
0
0.9798
.
1 − 0.2q−1
1
2
3
4
5
N
6
7
8
9
This set-up allows us to evaluate the tightness of the upper
bound in Theorem (5.1). Note that this is a completely different scenario to that in Section 6.1.
10
#10
4
Fig. 3. Plot of the estimation error for a SNR=10dB and a Gaussian
white input signal.
The FIR model structure is used here in order to construct
the linear regression for the SPARSEVA problem (15). The
number of parameters n of the FIR model is set to be 35.
The regularization parameter, εN , is chosen as n/N [8].
nη = 25, with the three different values of εN . The magenta
(BIC), green (AIC) and red (PEC) lines are the true estimation errors from 50 estimates (for each value of εN ) using
the SPARSEVA framework. We can see that the plot again
confirms the validity of the proposed upper bound for all
We then compute the upper bound of kθ̂εN − θ∗ k2 using
(22) with different values of nη , i.e. nη = [10, 15, 25]. The
8
probability parameters α and β are chosen to be 0.02 and
0.001 respectively. With this setting, the probability of the
upper bound being correct is (1 − α)(1 − 4nβ) = 0.84. This
upper bound will be compared with kθ̂εN − θ∗ k2 . Note that
here we plot both the upper bound and the true estimation
errors in logarithmic scale.
1
n =10
2
2
^ 0 k2 )
log(k "
N
0
-1
-1.5
-2
-2.5
0
1
2
3
N
4
5
#10
4
Fig. 7. Plot of the estimation error for SNR=10dB for a coloured
input signal.
7
Conclusion
The paper provides an upper bound on the SPARSEVA estimation error in the general case, for any choice of strongly
convex cost function and decomposable norm. We also evaluate the bound for a specific scenario, i.e., a sparse regression estimate problem. Numerical results confirm the validity of the developed bound for different input signals with
different output noise levels for different choices of the regularization parameters.
n =10
2
n =15
2
n =25
2
^ 0 k2 )
log(k "
N
True estimation errors
-0.5
0
-1
2
n =25
Plots of the upper bound as stated in Theorem 5.1 and the
true estimation error kθ̂εN − θ∗ k2 are displayed in Figures 5
to 7. In Figures 5 to 7, the red lines are the true estimation
errors from 50 estimates using the SPARSEVA framework.
The magenta, blue and cyan lines are the upper bounds developed in Theorem 5.1, which correspond to nη = [10, 15, 25]
respectively. We can see that the plots confirmed the bound
developed in Theorem 5.1 for all noise levels. When N becomes large, the estimation error and the corresponding upper bound become smaller. When N goes to infinity, the estimation error will tend to 0. Note that the bounds are slightly
different for the chosen values of nη , however, not significantly. As can be seen, the bounds are relatively insensitive
to the choice of nη .
-0.5
n =15
0.5
True estimation errors
-1.5
-2
References
-2.5
[1]
-3
-3.5
0
1
2
3
4
N
5
#10
[2]
4
Fig. 5. Plot of the estimation error for SNR=30dB for a coloured
input signal.
[3]
[4]
0.5
n =10
2
n =15
0
[5]
2
n =25
2
^ 0 k2 )
log(k "
N
-0.5
True estimation errors
-1
[6]
-1.5
-2
[7]
-2.5
-3
0
1
2
3
N
4
[8]
5
#10
4
Fig. 6. Plot of the estimation error for SNR=20dB for a coloured
input signal.
9
P.J. Bickel, Y. Ritov, and A.B. Tsybakov. Simultaneous analysis of lasso and Dantzig selector. Annals of
Statistics, 37:1705–1732, 2009.
J.M. Borwein and Q.J. Zhu. A variational approach to
lagrange multipliers. Journal of Optimization Theory
and Applications, 171(3):727–756, 2016.
S. Boyd and L. Vandenberghe. Convex Optimization.
Cambridge University Press, 2004.
E. Candes and T. Tao. The Dantzig selector: Statistical
estimation when p is much larger than n. Annals of
Statistics, 35:2313–2351, 2007.
M. Fazel, H. Hindi, and S. Boyd. A rank minimization
heuristic with application to minimum order system
approximation. Proceedings 2001 American Control
Conference, 2001.
M. Fazel, H. Hindi, and S. Boyd. Log-det heuristic for
matrix rank minimization with applications to hankel
and euclidean distance matrices. Proceedings 2003
American Control Conference, 2003.
S. Foucart and H. Rauhut. A Mathematical Introduction
to Compressive Sensing. Birkhäuser Basel, 2013.
H. Ha, J.S Welsh, N. Blomberg, C.R. Rojas, and
B. Wahlberg. Reweighted nuclear norm regularization:
A sparseva approach. Proceedings of the 17th IFAC
Symposium on System Identification, 48(28):1172 –
1177, 2015.
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
A
Proof. See the supplementary material of [12].
J. Huang, J.L. Horowitz, and S. Ma. Asymptotic properties of bridge estimators in sparse high-dimensional
regression models. Annals of Statistics, 36(2):587–613,
2008.
A.T. James. Distributions of matrix variates and latent
roots derived from normal samples. The Annals of
Mathematical Statistics, 35(2):475–501, 1964.
K. Knight and W. Fu. Asymptotics for Lasso-Type Estimators. Annals of Statistics, 28(5):1356–1378, 2000.
S.N. Negahban, P. Ravikumar, M.J. Wainwright, and
B. Yu. A Unified Framework for High-Dimensional
Analysis of M-Estimators with Decomposable Regularizers. Statistical Science, 27(4):538–557, 2012.
M. Osborne, B. Presnell, and B. Turlach. On the
LASSO and its dual. Journal of Computational and
Graphical Statistics, 9(2):319–337, 2000.
G. Pillonetto, F. Dinuzzo, T. Chen, G. De Nicolao, and
L. Ljung. Kernel methods in system identification,
machine learning and function estimation: A survey.
Automatica, 50(3):657–682, 2014.
C.R. Rojas and H. Hjalmarsson. Sparse estimation
based on a validation criterion. 2011 50th IEEE Conference on Decision and Control and European Control Conference (CDC-ECC), pages 2825–2830, 2011.
C.R. Rojas, R. Tóth, and H. Hjalmarsson. Sparse estimation of polynomial and rational dynamical models.
IEEE Transactions on Automatic Control, 59:2962–
2977, 2014.
T. Söderström and P. Stoica. System Identification.
Prentice Hall, 1989.
R. Tibshirani. Regression Shrinkage and Selection Via
the Lasso. Journal of the Royal Statistical Society,
Series B, 58:267–288, 1996.
A.N Tikhonov and V.Y Arsenin. Solutions of Ill-Posed
Problems. Winston & Sons, 1977.
R. Vershynin. Introduction to the non-asymptotic analysis of random matrices. In Y.C. Eldar and G. Kutyniok, editors, Compressed Sensing: Theory and Applications. Cambridge University Press, 2012.
We now quote the following lemma from [2] with modification to fit with the notation used in the SPARSEVA problem (2). This lemma helps us to find some important properties related to the SPARSEVA estimate θ̂λN . Based on these
properties, in the next section we can derive the upper bound
on the estimation error. Note that for notational simplicity,
we will denote L (θ; Z1N ) as L (θ).
Lemma A.2 Consider the convex optimization problem in
(2). Then the pair (θ̂εN , λεN ), with θ̂εN 6= 0, has the property
that θ̂εN is the solution of the problem (2) and λεN is the
Lagrange multiplier if and only if all of the following hold:
(1) λεN ∈ R+ ;
(2) the function R (θ)+λ{L (θ)− L (θ̂NR )(1+εN )} attains
its minimum over Rn at θ̂εN ; and
(3) L (θ̂εN ) − L (θ̂NR )(1 + εN ) = 0.
Proof. See the proof of Theorem 2.2 in [2]. The third condition in the cited theorem is a complementary slackness condition, which reduces to condition (3) here if θ̂εN 6= 0 (cf.
[16, Lemma II.2]).
A.2
Background knowledge
Now that we have confirmed the existence of the Lagrange
multiplier λεN , consider the function F (∆) defined as follows,
First, we cite a lemma directly from [12], to enable the proof
of Theorem 4.1 to be constructed.
Lemma A.1 For any norm R that is decomposable with
F (∆) = R (θ∗ + ∆) − R (θ∗ ) + λεN {L (θ∗ + ∆) − L (θ∗ )}.
⊥
respect to (M , M ); and any vectors θ, ∆, we have
R (θ + ∆) − R (θ) ≥ R (∆
M
Recall that ∆
⊥
M
⊥
⊥
Proof of Theorem 4.1
First we need to prove that there exists a Lagrange multiplier for the SPARSEVA problem (2). We can assume without loss of generality that L (θ̂NR ) 6= 0, since otherwise we
can take λεN = 0. According to [2], the Lagrange multiplier
for a convex optimization problem with constraint exists
when the Slater condition is satisfied. Specifically, for the
SPARSEVA problem (2), the Lagrange multiplier λεN exists
when there exists a θ1 such that L (θ1 ) < L (θ̂NR )(1 + εN ).
If εN > 0, and L (θ̂NR ) 6= 0, there always exists a parameter vector θ1 such that L (θ1 ) < L (θ̂NR )(1 + εN ) (just take
θ1 = θ̂NR ). Therefore, there exists a Lagrange multiplier λεN
for the SPARSEVA problem.
Appendix
A.1
(A.2)
Using the strong convexity condition of L (θ),
) − R (∆M ) − 2R (θM ⊥ ),
(A.1)
L (θ∗ + ∆) − L (θ∗ ) ≥ h∇L (θ∗ ), ∆i + κL k∆k22 .
(A.3)
Applying the Cauchy-Schwartz inequality to R and its dual
norm R ∗ , we have,
is the Euclidean projection of ∆ onto
M (see Section 3.5), and similarly for the other terms in
|h∇L (θ∗ ), ∆i| ≤ R ∗ (∇L (θ∗ ))R (∆).
(A.1).
10
(A.4)
By the definition of subspace compatibility,
Next, combining the inequality (A.4) and the triangle inequality, i.e., R (∆) ≤ R (∆M ) + R (∆ ⊥ ), we have,
M
R (∆ˆ εN ,M ) ≤ Ψ(M )k∆ˆ εN ,M k2 .
|h∇L (θ∗ ), ∆i| ≤ R ∗ (∇L (θ∗ ))R (∆)
≤ R ∗ (∇L (θ∗ ))(R (∆M ) + R (∆
M
⊥
)),
Now since 0 ∈ M , we have,
therefore,
∗
∗
∗
h∇L (θ ), ∆i ≥ −R (∇L (θ ))(R (∆M ) + R (∆
k∆ˆ ε ,M k2 = kΠM (∆ˆ εN ) − ΠM (0)k2
N
≤ k∆ˆ εN − 0k2 = k∆ˆ εN k2 .
)).
(A.5)
Now, combining (A.2), (A.3), (A.5), and Lemma A.1,
M
⊥
Therefore,
R (∆ˆ εN ,M ) ≤ Ψ(M )k∆ˆ εN k2 .
F (∆) = R (θ∗ + ∆) − R (θ∗ ) + λεN {L (θ∗ + ∆) − L (θ∗ )}
≥ R (∆ ⊥ ) − R (∆M ) − 2R (θ∗M ⊥ )
M
n
+ λεN − R ∗ (∇L (θ∗ ))(R (∆M ) + R (∆
∗
M
∗
≥ {1 − λεN R (∇L (θ ))}R (∆
∗
M
∗
⊥
⊥
Substituting this into (A.9) gives,
)) + κL k∆k22
o
0 ≥ κL λεN k∆ˆ εN k22 − {1 + λεN R ∗ (∇L (θ∗ ))}Ψ(M )k∆ˆ εN k2
− 2R (θ∗M ⊥ ). (A.11)
)
− {1 + λεN R (∇L (θ ))}R (∆M ) + κL λεN k∆k22
− 2R (θ∗M ⊥ ).
(A.10)
Note that for a quadratic polynomial f (x) = ax2 + bx + c,
with a > 0, if there exists x ∈ R+ that makes f (x) ≤ 0, then
such x must satisfy
(A.6)
Notice that, when θ̂εN is the estimate of the SPARSEVA
problem, then property 2 in Lemma A.2 states that the function R (θ) + λ{L (θ) − L (θ̂NR )(1 + εN )} attains its minimum
over Rn at θ̂εN , which means,
√
−b + b2 − 4ac
.
x≤
2a
Since (A + B)2 ≤ 2A2 + 2B2 for all A, B ∈ R,
∀θ ∈ R n , R (θ̂εN ) + λεN (L (θ̂εN ) − L (θ̂NR )(1 + εN ))
b2
b2 − 4ac
b2 − 2ac
x ≤2
+
=
.
4a2
4a2
a2
≤ R (θ) + λεN (L (θ) − L (θ̂NR )(1 + εN )).
2
(A.12)
Hence,
Applying this inequality to (A.11), we have,
∀θ ∈ R n , R (θ̂εN ) − R (θ) + λεN {L (θ̂εN ) − L (θ)} ≤ 0,
1
{1 + λεN R ∗ (∇L (θ∗ ))}2 Ψ2 (M )
κ2L λ2εN
4
+
R (θ∗M ⊥ )
κL λεN
4
4
≤ 2 2 Ψ2 (M ) +
R (θ∗M ⊥ ). (A.13)
κ
κL λεN
L λεN
k∆ˆ εN k22 ≤
or, taking θ = θ∗ and defining ∆ˆ εN := θ̂εN − θ∗ ,
F (∆ˆ εN ) ≤ 0.
(A.7)
Combining (A.6) with (A.7), we then have,
0 ≥ κL λεN k∆ˆ εN k22 + {1 − λεN R ∗ (∇L (θ∗ ))}R (∆ˆ
∗
∗
εN ,M
∗
⊥
)
Case 2: λεN > 1/R ∗ (∇L (θ∗ ))
− {1 + λεN R (∇L (θ ))}R (∆ˆ ε ,M ) − 2R (θM ⊥ ).
N
(A.8)
Using a similar analysis as in Case 1,
Now we consider two cases.
R (∆ˆ
εN ,M
Case 1: λεN ≤ 1/R ∗ (∇L (θ∗ ))
⊥
⊥
) ≤ Ψ(M )k∆ˆ εN k2 .
(A.14)
Substituting (A.14) and (A.10) into (A.8), we obtain,
From (A.8), we have,
⊥
0 ≥ κL λεN k∆ˆ εN k22 + {1 − λεN R ∗ (∇L (θ∗ ))}Ψ(M )k∆ˆ εN k2
− {1 + λεN R ∗ (∇L (θ∗ ))}Ψ(M )k∆ˆ εN k2 − 2R (θ∗M ⊥ ).
(A.15)
0 ≥ κL λεN k∆ˆ εN k22 − {1 + λεN R ∗ (∇L (θ∗ ))}R (∆ˆ ε ,M )
N
− 2R (θ∗M ⊥ ).
(A.9)
11
Now using the inequality (A.12), yields,
g(θ, λ) = kθk1 + λ(L (θ) − L (θ̂NR )(1 + εN ))
λ
= kθk1 +
(kYN − ΦTN θk22 − kYN − ΦTN θ̂NR k22 (1 + εN )).
2N
(A.20)
⊥
1 1
∗
∗
2
ˆ
− R (∇L (θ )) Ψ(M )
k∆εN k2 ≤ 2
λεN
κL
2
1
∗
∗
−
+ R (∇L (θ )) Ψ(M )
λεN
4
+
R (θ∗M ⊥ ).
κL λεN
(A.16)
Applying the inequality (A + B)2 ≤ 2A2 + 2B2 to the first
term in (A.16) gives,
The subdifferential of g(θ, λ) can be computed as
∂g(θ, λ)
λ
= v − ΦN (YN − ΦTN θ),
∂θ
N
where v = (v1 , . . . , vm )T is of the form
2
⊥
1
∗
∗
− R (∇L (θ )) Ψ2 (M )
λεN
2
2
1
+ 2
+ R ∗ (∇L (θ∗ )) Ψ2 (M )
κL λεN
4
R (θ∗M ⊥ ).
+
κL λεN
2
k∆ˆ εN k22 ≤ 2
κL
vi
v
i
vi
2
1
∗
∗
− R (∇L (θ )) ≤ {R ∗ (∇L (θ∗ ))}2 .
λεN
2
1
∗
∗
+ R (∇L (θ )) ≤ 4{R ∗ (∇L (θ∗ ))}2 .
λεN
∂g(θ, λ)
∂θ
θ=θ̂εN ,λ=λεN
λεN
=−
ΦN (YN − ΦTN θ̂εN ) + vθ̂ε ,
N
N
λεN =
2
{R ∗ (∇L (θ∗ ))}2 Ψ2 (M )
κ2L
⊥
8
+ 2 {R ∗ (∇L (θ∗ ))}2 Ψ2 (M )
κL
4
+
R (θ∗M ⊥ ).
κL λεN
k∆ˆ εN k22 ≤
N
.
kΦN (YN − ΦTN θ̂εN )k∞
Note that this proof is similar to the one in [13], where an expression was derived for the Lagrange multiplier in the traditional l1 norm regularization problem (the LASSO). Here
we have derived the Lagrange multiplier for the SPARSEVA
problem as given in (15).
Proof of Proposition 5.1
Let us rewrite the SPARSEVA problem (15),
A.4
θ̂εN ∈ arg min kθk1
s.t.
(A.23)
1
Since ∇L (θ̂εN ) = ΦN (YN −ΦTN θ̂εN ), we can also write λεN
N
as
1
λεN =
.
k∇L (θ̂εN )k∞
θ∈Rn
(A.22)
for some v of the form in (A.22). Note that when θ̂εN 6= 0,
kvk∞ = 1, which means that
(A.18)
Therefore, combining (A.17) and (A.18),
A.3
if θi > 0
if θi < 0
if θi = 0.
0=
(A.17)
We also have,
=1
= −1
∈ [−1, 1]
Using property 2 of Lemma A.2, when θ̂εN is a solution of the
SPARSEVA problem (15) and λεN is a Lagrange multiplier,
we have,
Note that 0 < 1/λεN < R ∗ (∇L (θ∗ )), therefore,
(A.21)
Proof of Proposition 5.2
(A.19)
L (θ) − L (θ̂NR )(1 + εN ) ≤ 0,
For the linear regression (13) and the choice of L (θ) in (14),
in the Lagrangian form (1) using Lemma A.2. The Lagrangian of the optimization problem (A.19) is,
∇L (θ∗ ) =
12
1
1
ΦN (YN − ΦTN θ∗ ) = ΦN e.
N
N
Denote R j as the jth row of the matrix ΦN , then ∇L (θ∗ ) can
be computed as,
The expression in parentheses on the right hand side is
monotonically decreasing in R j RTj , so using (A.26) gives
R1
R1 e
1 R2 e
1
1 R2
∇L (θ∗ ) = ΦN e = . e = .
N
N ..
N ..
Rn
Rn e
P
ΦN e
P
N
≤
n
Z
N
ΦN ≥ (1 − β)n
≤
q
2σ2e smax χ2β (N) ln(2/β)
N
∞
with probability at least 1 − 2nβ.
A.5
Proof of Proposition 5.3
When θ̂εN is the solution of the problem in (15), we have,
(A.25)
∇L (θ̂εN ) =
1
ΦN (YN − ΦTN θ̂εN ).
N
(A.29)
Denote eεN = YN − ΦTN θ̂εN , and R j as the jth row of the
matrix ΦN , then (A.29) becomes,
(A.26)
∇L (θ̂εN ) =
1
ΦN eεN
N
R1
R1 eεN
1
1
R2
R2 eεN
= . eεN = . .
N ..
N ..
Rn
Rn eεN
From Assumption 5.1, and using the same argument as in
Proposition 5.2, we have that each element of R j is distributed as N (0, Σ( j, j)).
Now consider the variable Z = N −1 eTεN RTj . Since each element of R j is N (0, Σ( j, j)),
σ2e
T
P
≤ t ΦN = ∏
f x 0, 2 R j R j dx .
N
∞
−t
j=1
(A.28)
This expression can be bounded from below using the standard result that P(|N (0, σ2 )| > t) ≤ 2 exp(−t 2 /2σ2 ) [20,
Eq. (5.5)], to obtain
"
#!
n
ΦN e
N 2t 2
P
≤ t ΦN ≥ ∏ 1 − 2 exp − 2
.
N ∞
2σe R j RTj
j=1
q
2σ2e smax χ2β (N) ln(2/β)
∇L (θ∗ )
Note that from (A.24), for any t > 0, we have,
Zt
1
σ2e
T
P
R j e ≤ t ΦN =
f x 0, 2 R j R j dx,
N
N
−t
(A.27)
where f (x|0, N −2 σ2e R j RTj ) denotes the pdf of the Normal
distribution N (0, N −2 σ2e R j RTj ). This gives,
ΦN e
N
q
2σ2e smax χ2β (N) ln(2/β)/N gives
∞
with probability 1 − β.
,
with probability at least 1 − nβ, or equivalently,
Hence, the variance of the distribution of the variable
N −1 R j e, is,
σ2e
σ2e
T
R
R
≤
Σ( j, j)χ2β (N),
j
j
N2
N2
#!n
≥ 1 − nβ
Now in order to derive a bound for ∇L (θ∗ ), we first derive an
upper bound for the variance N −2 σ2e R j RTj of the distribution
in (A.24). Since each element of R j ∼ N (0, Σ( j, j)), we
have,
R j RTj ∼ Σ( j, j)χ2 (N).
Hence, with probability 1 − β, 0 ≤ β ≤ 1, we have,
≤ Σ( j,
≤ t ΦN
∞
"
N 2t 2
≥ 1 − 2 exp − 2
2σe smax χ2β (N)
In particular, taking t =
Now consider the variable Z = N −1 R j e. Using Assumption
5.2 on the disturbance noise e, e ∼ N (0, σ2e ) and the property
related to the sum of independent normal random variables,
we then have,
σ2
Z|e ∼ N 0, e2 R j RTj .
(A.24)
N
j)χ2β (N).
which holds with probability 1 at least 1 − nβ, where smax
denotes the maximum diagonal element of matrix Σ.
Note that R j contains the jth element of each column ΦN, j
of the matrix ΦN , and following the Assumption 5.1, ΦN, j ∼
N (0, Σ), each element of R j ∼ N (0, Σ( j, j)).
R j RTj
ΦN e
N
t
Z∼N
1
Σ( j, j) T
0,
e
e
.
ε
N 2 εN N
This bound follows because the events A j that (A.26) holds
are not necessarily independent, but their joint probability can be
bounded like P(A1 ∩· · ·∩An ) = 1−P(AC1 ∪· · ·∪ACn ) ≥ 1−P(AC1 )−
· · · − P(ACn ) = 1 − nβ.
13
with probability at least 1 − nβ, or, equivalently,
Note that,
eTεN eεN = (YN − ΦTN θ̂εN )T (YN − ΦTN θ̂εN ) = 2N L (θ̂εN )
q
∇L (θ̂εN )
= 2N L (θ̂NR )(1 + εN ).
(A.30)
From Section 4.4 in [17], we have,
L (θ̂NR )|ΦN ∼
σ2e 2
χ (N − n),
2N
σ2e 2
χ (N − n),
2N β
with probability 1 − β, 0 ≤ β ≤ 1. Combining this inequality
with (A.30) gives, with probability 1 − β,
Σ2 ( j, j)eTεN eεN
σ2
≤ e2 Σ( j, j)χ2β (N − n)(1 + εN ).
2
N
N
Hence,
R j eεN
≤t e
P
N
Z t
σ2e 2
≥
f x 0, Σ( j, j) 2 χβ (N − n)(1 + εN ) dx, (A.31)
N
−t
with probability 1 − β. This means,
ΦN eεN
P
(A.32)
≤t e
N
∞
n Z t
σ2
=∏
f x 0, e2 Σ( j, j)χ2β (N − n)(1 + εN ) dx
N
−t
j=1
(
!)n
N 2t 2
≥ 1 − 2 exp − 2
2σe smax χ2β (N − n)(1 + εN )
with probability at least 1−nβ, following the same reasoning
as in the proof of Proposition 5.2. Therefore,
P ∇L (θ̂εN ) ≤ t e
∞
!)n
(
N 2t 2
≥ 1 − 2 exp − 2
2σe smax χ2β (N − n)(1 + εN )
with probability at least 1 − nβ.
Taking t =
q
2σ2e smax χ2β (N − n)(1 + εN ) ln(2/β)/N gives
P ∇L (θ̂εN )
≤
q
2σ2e smax χ2β (N − n)(1 + εN ) ln(2/β)
∞
2σ2e smax χ2β (N − n)(1 + εN ) ln(2/β)
∞
with probability at least 1 − 2nβ.
which gives,
L (θ̂NR ) ≤
≤
N
e
≥ (1 − β)n ≥ 1 − nβ,
14
N
,
| 10math.ST
|
1
Low-Complexity Sub-band Digital Predistortion for
Spurious Emission Suppression in Noncontiguous
Spectrum Access
arXiv:1607.02249v2 [cs.IT] 19 Aug 2016
Mahmoud Abdelaziz, Student Member, IEEE, Lauri Anttila, Member, IEEE,
Chance Tarver, Student Member, IEEE, Kaipeng Li, Student Member, IEEE, Joseph R. Cavallaro, Fellow, IEEE,
Mikko Valkama, Senior Member, IEEE
Abstract—Noncontiguous transmission schemes combined with
high power-efficiency requirements pose big challenges for radio
transmitter and power amplifier (PA) design and implementation.
Due to the nonlinear nature of the PA, severe unwanted emissions
can occur, which can potentially interfere with neighboring
channel signals or even desensitize the own receiver in frequency division duplexing (FDD) transceivers. In this article, to
suppress such unwanted emissions, a low-complexity sub-band
DPD solution, specifically tailored for spectrally noncontiguous
transmission schemes in low-cost devices, is proposed. The
proposed technique aims at mitigating only the selected spurious intermodulation distortion components at the PA output,
hence allowing for substantially reduced processing complexity
compared to classical linearization solutions. Furthermore, novel
decorrelation based parameter learning solutions are also proposed and formulated, which offer reduced computing complexity
in parameter estimation as well as the ability to track timevarying features adaptively. Comprehensive simulation and RF
measurement results are provided, using a commercial LTEAdvanced mobile PA, to evaluate and validate the effectiveness
of the proposed solution in real world scenarios. The obtained
results demonstrate that highly efficient spurious component
suppression can be obtained using the proposed solutions.
Index Terms—Adaptive filters, carrier aggregation, digital
predistortion, frequency division duplexing, nonlinear distortion,
power amplifier, software defined radio, spectrally agile radio,
spurious emission, 3GPP LTE-Advanced.
I. I NTRODUCTION
Spectrum scarcity and ever-increasing data rate requirements are the two main motivating factors behind introducing
carrier aggregation (CA) in modern wireless communication systems, such as 3GPP LTE-Advanced [1], [2]. In CA
transmission, multiple component carriers (CCs) at different
Mahmoud Abdelaziz, Lauri Anttila, and Mikko Valkama are with the
Department of Electronics and Communications Engineering, Tampere University of Technology, Tampere, Finland.
Chance Tarver, Kaipeng Li, and Joseph R. Cavallaro are with the Department of Electrical and Computer Engineering, Rice University, Houston, TX.
This work was supported by the Finnish Funding Agency for Technology
and Innovation (Tekes) under the project “Future Small-Cell Networks using
Reconfigurable Antennas (FUNERA)”, and by the Linz Center of Mechatronics (LCM) in the framework of the Austrian COMET-K2 programme. The
work was also funded by the Academy of Finland under the projects 288670
“Massive MIMO: Advanced Antennas, Systems and Signal Processing at mmWaves”, 284694 “Fundamentals of Ultra Dense 5G Networks with Application
to Machine Type Communication”, and 301820 “Competitive Funding to
Strengthen University Research Profiles”. This work was also supported in
part by the US National Science Foundation under grants ECCS-1408370,
CNS-1265332, and ECCS-1232274.
RF frequencies are adopted simultaneously, either within the
same RF band (intraband CA), or at different RF bands
(interband CA) [3]. While interband CA naturally leads to a
non-contiguous transmit spectrum, this can also happen in the
intraband CA case when the adopted component carriers are
located in a non-contiguous manner. Adopting non-contiguous
CA provides lots of flexibility in RF spectrum use, but
also poses substantial challenges in the transmitter design,
especially in lower-cost devices such as LTE-Advanced user
equipment (UE). Moreover, multicarrier type modulations with
very large peak-to-average power ratio (PAPR) are used in
most state-of-the-art wireless communication systems. When
combined with noncontiguous transmission schemes and high
power-efficiency requirements, controlling the transmitter unwanted emissions due to power amplifier (PA) nonlinearity
becomes a true challenge [4]–[6].
In [3], a comparison of different transmitter architectures
for CA based transmission schemes was presented, showing
that it is more power efficient to combine the CCs before
the PA, and thus use only a single PA, instead of having a
separate PA for each CC. This is also technically feasible,
especially in the intraband CA case. The power savings can
be quite significant since the overall transmitter efficiency is
dominated by the PA [3], and is particularly critical in mobile
terminals and other lower-cost devices, such as small-cell basestations, with limited computing and cooling capabilities [7],
[8]. However, the power efficiency gained from using a single
PA for all CCs comes at the expense of more severe unwanted
emissions, stemming from the PA nonlinear characteristics,
that can occur at adjacent channels or bands, or even at own
receiver band in FDD transceiver case [4]–[6], [9].
The levels of different unwanted emissions are, in general,
regulated by the standardization bodies and ITU-R [10].
It has been recently demonstrated that in the context of
LTE-Advanced mobile transmitter with noncontiguous CA or
multicluster type transmission, the PA nonlinearities lead to
spurious emissions that can seriously violate the given spectrum and spurious emission limits if not properly controlled
[2], [4]–[6]. Furthermore, in FDD devices, in addition to
violating the general spurious emission limits, the generated
spurious components can also overlap with the device’s own
receive band, causing own receiver desensitization [4], [9],
[11], [12]. One obvious solution to decrease the levels of
unwanted emissions is to back off the transmit power from
2
its saturation region. In 3GPP UE terminology, this is known
as Maximum Power Reduction (MPR), and MPR values up to
16 dB are allowed in some use cases for mobile terminals [2].
However, this approach will end up yielding a significantly
lower PA efficiency as well as a substantial reduction in the
network coverage. There is thus a clear need for alternative PA
linearization solutions which do not have such drastic adverse
effects.
Digital predistortion (DPD) is, in general, one of the most
effective solutions for mitigating transmitter nonlinear distortion. There have been substantial research efforts in the past
20 years in developing ever more efficient and elaborate DPD
techniques, mostly for high-end macro base-station type of
devices. The conventional DPD approaches seek to linearize
the full composite transmit signal, and we thus refer to such
solutions as full-band DPD in this article. There have also been
a handful of recent works on efficient concurrent linearization
techniques for multi-band transmitters that employ only a
single PA [13]–[15]. These works assume that the component
carriers are separated by a large distance such that the spurious
emissions are filtered out by the transmit RF filter, and hence
linearization of only the main carriers is pursued.
Complementary to such methods, the scope of this paper is
to introduce a low-complexity DPD solution for suppressing
the spurious emissions in non-contiguous transmission cases,
while not concentrating specifically on the linearization of the
main component carriers. Hereafter, we will refer to such
linearization solutions as sub-band DPD. This approach is
motivated by the following two factors. First, the emission
limits in the spurious region are generally stricter than in the
spectral regrowth region around the component carriers, and
are thus more easily violated. This has been recognized also
in 3GPP recently, in the context of intraband noncontiguous
carrier aggregation [2], [4]–[6]. Moreover, even in interband
carrier aggregation scenarios in FDD devices, some of the
spurious components can be hitting the own RX band causing
own receiver desensitization [11]. Second, by concentrating
the linearization efforts to the most critical spurious emissions
only, the processing and instrumentation requirements can
be significantly relaxed, thus potentially facilitating the DPD
processing also in mobile terminals and other lower-cost
devices. This applies not only to the DPD main path processing
complexity but also to the feedback receiver instrumentation
complexity which can also be substantially reduced.
There have been some recent studies in the literature that
consider the mitigation of the spurious emissions explicitly. In
[16]–[18], such processing was added to complement a concurrent linearization system while assuming a frequency-flat PA
response within the spurious sub-band. In [16], [17], the DPD
parameter estimation was done offline, based on extracting the
quasi-memoryless PA parameters using large-signal network
analyzer (LSNA) measurements, and covering up to fifth-order
processing. In [18], a memoryless least-squares fit between the
observed intermodulation distortion (IMD) at the considered
sub-band and certain basis functions was performed. The basis
functions were generated using a wideband composite carrier
baseband equivalent signal, followed by sub-band filtering,
implying a very high sample rate and processing complex-
ity, especially for widely spaced carriers. The estimated and
regenerated IMD was then applied at the PA input, oppositely
phased, such that it is canceled at the PA output. In [19],
[20], on the other hand, third-order nonlinearities at the IM3
sub-bands were specifically targeted, through explicit, lowrate, behavioral modeling of the baseband equivalent IM3
sub-band emissions. Furthermore, the parameter estimation
of the DPD was based on a closed-loop feedback with a
decorrelation-based learning rule, which was shown to have
better linearization performance compared to the third-order
inverse solution in [16] in terms of the spurious emission
suppression. Moreover, an FPGA implementation of this thirdorder decorrelation-based sub-band DPD was presented in
[21], demonstrating fast and reliable performance under real
time constraints. A recent overview article [22] highlighted the
main principles and advantages of such low complexity subband DPD solutions, while not concentrating on the details of
the DPD processing or parameter estimation and adaptation
algorithms at technical level. In [23], on the other hand, a
flexible full-band DPD solution was also proposed by the
authors which can optimize the DPD coefficients to minimize
the nonlinear distortion at a particular frequency or sub-band in
the out-of-band or spurious regions. However, like other fullband DPD techniques, it requires very high sampling rates in
the transmitter and feedback receiver when the carrier spacing
between the CCs increases.
In this article, we extend the elementary third-order IM3
sub-band DPD solution, proposed by the authors in [19],
[20], in two ways. First, the third-order IM3 sub-band DPD
is extended to incorporate higher-order processing based on
explicit modeling of the higher-order spurious components at
the IM3 sub-band. This will enhance the IM3 spurious emission suppression considerably. Furthermore, we also extend the
sub-band DPD solution to include higher-order sub-bands, i.e.,
IM5, IM7, etc., thus offering more flexibility and linearization
capabilities beyond the basic IM3 sub-band. All the proposed
solutions are derived for wideband nonlinear PAs with memory. Furthermore, we also formulate novel decorrelation-based
parameter estimation methods, covering both sample-adaptive
and block-adaptive learning rules, to efficiently identify the
needed DPD parameters with low complexity. The proposed
learning solutions are also shown to offer better performance
than the earlier proposed third-order or fifth-order inverse
based methods. We also provide comprehensive simulation and
RF measurement results, using a commercial LTE-Advanced
mobile PA, to evaluate and validate the effectiveness of the
proposed solutions in real world scenarios.
The rest of this article is organized as follows: Section II presents the mathematical modeling of the different
considered spurious components at different sub-bands, produced by a nonlinear PA with memory. Stemming from this
modeling, also the corresponding core processing principles
of the proposed DPD solutions at different sub-bands are
formulated. Then, Section III presents the proposed DPD
parameter learning solutions, covering both sample-adaptive
and block-adaptive decorrelation solutions at different subbands. Section IV addresses then different implementation
alternatives of the proposed sub-band DPD concept, and also
3
6ϑΚΤΦ 1ΤΦΓΤ 0ΘΠΝΚΠΓΧΤΚς[
%ΧΤΤΚΓΤ ∃9
%ΧΤΤΚΓΤ ∃9
(ΚΗςϑ 1ΤΦΓΤ 0ΘΠΝΚΠΓΧΤΚς[
+/ 4
+/ 4
+/ 4
5ΓΞΓΠςϑ 1ΤΦΓΤ 0ΘΠΝΚΠΓΧΤΚς[
5ΡΩΤ .ΚΟΚς
5ΡΩΤ .ΚΟΚς
+/
+/
+/
%%
%ΧΤΤΚΓΤ ∃9
&f
&f
%%
/∗∴
&f
+/
%ΧΤΤΚΓΤ ∃9
&f
+/
f
+/
/∗∴
&f
&f
&f
Fig. 1. Illustration of different intermodulation distortion components created by a nonlinear PA when excited with a non-contiguous CA signal with two
component carriers. Here, nonlinear distortion components up to order 7 are shown.
analyzes the computing complexity of the sub-band DPD in
terms of the number of floating point operations together
with some hardware complexity aspects. Finally, Sections V
and VI report comprehensive simulation and RF measurement
results, evidencing excellent spurious emission suppression in
different realistic scenarios.
II. S PURIOUS C OMPONENT M ODELING AND P ROPOSED
S UB - BAND DPD P ROCESSING
In this manuscript, we assume a practical case of noncontiguous carrier aggregation with two component carriers.
When such noncontiguous dual-carrier signal is applied at
the PA input, the PA nonlinearity leads to intermodulation
distortion at different sub-bands, as shown in Fig. 1. Assuming
a CC separation of ∆f , in addition to the spectral regrowth
around the main carriers, intermodulation between the two
CCs yields strong IMD at integer multiples of ∆f from the
main CCs. In this article, we refer to the intermodulation (IM)
sub-bands located at ±∆f from the main CCs as the IM3
sub-bands. Similarly, the IM5 sub-bands are located at ±2∆f
from the main CCs, and so on. In general, each IM sub-band
includes nonlinear distortion components of different orders as
shown in Fig. 1. For example, in the case of a seventh-order
PA nonlinearity, the IM3 sub-band contains third, fifth and
seventh-order nonlinearities, while the IM5 sub-band contains
fifth and seventh-order components.
In this section, we start with the fundamental modeling of
the nonlinear distortion at the IM3 sub-band as a concrete
example. Stemming from that modeling, we then formulate the
basic processing of the proposed IM3 sub-band DPD. After
that, the modeling is extended to cover the nonlinear distortion
components at the higher-order IM sub-bands, followed by
the corresponding higher-order sub-band DPD processing.
The actual parameter estimation and learning algorithms for
the proposed sub-band DPD structures are then presented in
details in Section III. In all the modeling and developments, we
adopt the widely-used wideband Parallel Hammerstein (PH)
PA model [24] to describe the fundamental nonlinear behavior
of the PA, as it has been shown to model accurately the
measured nonlinear behavior of different classes of true PAs.
A. Spurious Component Modeling at IM3 Sub-bands
The modeling is carried out at composite baseband equivalent level, where the two component carriers are assumed to be
separated by 2fIF . The composite baseband equivalent input
and output signals of the P th order Parallel Hammerstein PA
model, with monomial nonlinearities and FIR branch filters,
respectively, read
f
j2π fIF n
x(n) = x1 (n)e
y(n) =
P
X
s
f
−j2π fIF n
+ x2 (n)e
s
fp,n ? |x(n)|p−1 x(n),
,
(1)
(2)
p=1
p odd
where x1 (n) and x2 (n) are the baseband component carrier
signals, fp,n denotes the PH branch filter impulse responses
of order p, and ? is the convolution operator. Intermodulation
between the two component carriers leads to the appearance
of IMD components at ±3fIF , ±5fIF , etc., as shown in Fig.
1 for the corresponding RF spectrum.
As a concrete example, let us analyze the IMD at ±3fIF .
Direct substitution of (1) in (2) allows extracting the baseband
equivalent distortion terms at the IM3 sub-bands, which,
through straight-forward manipulations, yields
yIM 3± (n) =
P
X
f3±,p,n ? u3±,p (n).
(3)
p=3
p odd
Here, f3±,p,n denote the baseband equivalent impulse responses corresponding to the wideband PH PA model filters
fp,n , evaluated at the IM3± sub-bands, formally defined as
3f
∓j2π fIF n
s
fp,n },
f3±,p,n = LP F {e
(4)
with LP F {.} denoting an ideal low pass filtering operation
with a passband width P times the bandwidth of the wider
CC. Furthermore, u3+,p (n) and u3−,p (n) in (3) are the corresponding pth order static nonlinear (SNL) basis functions,
related to the nonlinear distortion at the IM3+ and IM3- subbands, respectively. Assuming then an 11th order PA model
(i.e., P = 11), as a concrete high-order example, the IM3+
sub-band basis functions read
4
2# 1ΩςΡΩς 9Κςϑ +/
f IF
Ω
f
+/
+/
Ζ
Ω
+/
fRF %%
%%
+ 3
/1&
&#%
+/
+/
&2&
+/
f
f IF
2#
3 f IF
f
5Ω∆ ∃ΧΠΦ
&2& (ΚΝςΓΤ
&52
f
&#%
Ξ(,
Ζ
f RF
#&%
f
(10)
p=3
p odd
Ζ
Ω
Q
X
j2π 3ffIF
n
s
x̃(n) = x(n) +
α3+,p,n ? u3+,p (n) e
3 f IF
.2(
Fig. 2. The proposed sub-band DPD processing principle focusing on the
IM3+ sub-band. Thick lines are used to indicate complex I/Q processing. For
presentation simplicity, TX filtering between the feedback coupler and the
antenna is not shown.
˜ variables to indicate
Here, and in the continuation, we use (.)
sub-band DPD processing and the corresponding predistorted
signals. This DPD processing principle is illustrated in Fig.
2 at a conceptual level, where the TX/RX duplexer filters
are omitted for simplicity of the presentation since they do
not directly impact the sub-band DPD processing or learning.
Similar convention is followed also in other figures. Different
implementation alternatives as well as the parameter learning
and feedback receiver aspects are addressed in more details in
Sections III and IV.
C. Generalization to Higher-Order IM Sub-bands
u3+,3 (n) = x∗2 (n)x21 (n)
(5)
u3+,5 (n) = u3+,3 (n) × (2|x1 (n)|2 + 3|x2 (n)|2 )
(6)
u3+,7 (n) = u3+,3 (n) × (3|x1 (n)|4 + 6|x2 (n)|4
+ 12|x1 (n)|2 |x2 (n)|2 )
(7)
u3+,9 (n) = u3+,3 (n) × (4|x1 (n)|6 + 10|x2 (n)|6
+ 30|x1 (n)|4 |x2 (n)|2 + 40|x1 (n)|2 |x2 (n)|4 )
(8)
u3+,11 (n) = u3+,3 (n) × (5|x1 (n)|8 + 15|x2 (n)|8
+ 60|x1 (n)|6 |x2 (n)|2 + 100|x1 (n)|2 |x2 (n)|6
+ 150|x1 (n)|4 |x2 (n)|4 ).
(9)
The corresponding basis functions for the IM3- sub-band, i.e.,
u3−,p (n), can be obtained by simply interchanging x1 (n)
and x2 (n) in the above expressions. Next, these behavioral
modeling results are utilized to formulate the proposed IM3
sub-band DPD concept, specifically tailored to suppress the
distortion at the IM3 sub-band.
The direct substitution of (1) in (2) leads to the appearance
of spurious intermodulation terms also at higher-order IM subbands, in addition to the previously considered IM3 sub-bands,
as illustrated already in Fig. 1. These higher-order sub-band
emissions can also be harmful, since they can violate the
emission limits, cause own receiver desensitization, or both.
Thus, developing a sub-band DPD solution that can tackle also
the distortion at these higher-order IM sub-bands with feasible
complexity is important. Similar to the above IM3+ sub-band
developments, we next extract the IMD terms at the higherorder IM sub-bands. The baseband equivalent IMD terms at
the IM5+, IM7+, IM9+, and IM11+ sub-bands, as concrete
examples, can be extracted using (1) and (2), interpreted at
proper sub-bands, yielding
yIM 5+ (n) =
f5+,p,n ? u5+,p (n)
(11)
f7+,p,n ? u7+,p (n)
(12)
f9+,p,n ? u9+,p (n)
(13)
f11+,p,n ? u11+,p (n).
(14)
p=5
p odd
yIM 7+ (n) =
P
X
p=7
p odd
B. Proposed IM3 Sub-band DPD Principle
For presentation purposes, we focus here on suppressing
the spurious emissions at the IM3+ sub-band, while the corresponding emissions at the IM3- sub-band can also be easily
mitigated using a very similar structure with minor changes
as will be elaborated later in Section II-C. The key idea of the
sub-band DPD concept is to inject a proper additional lowpower cancellation signal, with structural similarity to (3) and
located at +3fIF , into the PA input, such that the level of
the IM3+ term at the PA output is reduced. Now, stemming
from the IMD structure in (3), an appropriate digital injection
signal can be obtained by adopting the IM3+ basis functions
u3+,p (n) in (5)-(9), combined with proper filtering using a
bank of sub-band DPD filters α3+,p,n . Incorporating such subband DPD processing with polynomial order Q, the composite
baseband equivalent PA input signal reads
P
X
yIM 9+ (n) =
P
X
p=9
p odd
yIM 11+ (n) =
P
X
p=11
p odd
In the above models, f5+,p,n , f7+,p,n , f9+,p,n , and f11+,p,n
denote the baseband equivalent impulse responses corresponding to the wideband PA model filters fp,n in (2), evaluated at
the IM5+, IM7+, IM9+, and IM11+ sub-bands, respectively.
These are formally obtained similar to (4) but replacing 3fIF
with either 5fIF , 7fIF , 9fIF , or 11fIF , respectively. The pth
order basis functions at the these higher-order IM sub-bands,
denoted by u5+,p (n), u7+,p (n), u9+,p (n), and u11+,p (n), and
assuming again an 11th order PH PA model, read
5
∗ 2
u5+,5 (n) = (x2 (n) ) x1 (n)
3
(15)
2
2
u5+,7 (n) = u5+,5 (n)(4|x2 (n)| + 3|x1 (n)| )
(16)
u5+,9 (n) = u5+,5 (n)(10|x2 (n)|4 + 6|x1 (n)|4
+ 20|x1 (n)|2 |x2 (n)|2 )
(17)
u5+,11 (n) = u5+,5 (n)(20|x2 (n)|6 + 10|x1 (n)|6
+ 75|x2 (n)|4 |x1 (n)|2 + 60|x2 (n)|2 |x1 (n)|4 )
u7+,7 (n) = (x2 (n)∗ )3 x1 (n)4
(18)
(19)
2
2
u7+,9 (n) = u7+,7 (n)(5|x2 (n)| + 4|x1 (n)| )
(20)
u7+,11 (n) = u7+,7 (n)(15|x2 (n)|4 + 10|x1 (n)|4
+ 30|x2 (n)|2 |x1 (n)|2 )
(21)
u9+,9 (n) = (x2 (n)∗ )4 x1 (n)5
(22)
2
2
u9+,11 (n) = u9+,9 (n)(6|x2 (n)| + 5|x1 (n| )
∗ 5
6
u11+,11 (n) = (x2 (n) ) x1 (n) .
(23)
(24)
The corresponding baseband equivalent IMD terms at the negative IM sub-bands can be obtained by simply interchanging
x1 (n) and x2 (n) in (15)-(24).
Then, stemming from the distortion structure in (11)-(14),
and adopting similar ideology as in the previous IM3 subband DPD case, a natural injection signal for suppressing the
spur components at higher-order IM sub-bands can be obtained
by properly filtering and combining the above higher-order
sub-band basis functions. The sub-band specific filters for the
pth order basis functions are denoted with α5+,p,n , α7+,p,n ,
α9+,p,n , and α11+,p,n . Incorporating such DPD processing
with DPD polynomial order Q, and aggregating at the same
time parallel sub-band DPDs simultaneously for the IM3+,
IM5+, and IM7+ sub-bands, as a concrete example, the
composite baseband equivalent PA input signal now reads
Q
X
j2π 3ffIF
n
s
x̃(n) = x(n) +
α3+,p,n ? u3+,p (n) e
X
j2π 5ffIF
n
s
+
α5+,p,n ? u5+,p (n) e
p=5
p odd
Q
X
j2π 7ffIF
n
s
+
α7+,p,n ? u7+,p (n) e
In this subsection, we derive three analytical reference
solutions for calculating the IM3 sub-band DPD coefficients.
The considered approaches are the third-order inverse solution, also adopted in [16], the IM3 power minimization
solution, and the analytical IM3 decorrelation-based solution.
To keep the analytical developments simple and tractable,
we consider a simplified case with a third-order memoryless
IM3+ sub-band DPD and a third-order memoryless PA. The
actual sample-adaptive and block-adaptive decorrelation based
learning solutions, devised later in this section, are formulated
for the general cases of higher-order processing with memory.
Now, starting with the dual carrier signal in (1), and limiting
the study to the simplified case of a third-order memoryless
IM3+ sub-band DPD and a third-order memoryless PA, the
basic signal models are given by (26)-(30).
y(n) = f1 x(n) + f3 |x(n)|2 x(n),
x∗2 (n)x21 (n),
u(n) =
yIM 3+ (n) = f3 u(n),
(26)
(27)
(28)
3f
j2π fIF n
s
,
x̃(n) = x(n) + αu(n)e
+ f3 |α|2 α|x1 (n)|4 |x2 (n)|2 u(n).
p=3
p odd
Q
A. Analytical Reference Solutions
(29)
ỹIM 3+ (n) = (f3 + f1 α)u(n) + 2f3 α(|x1 (n)|2 + |x2 (n)|2 )u(n)
analytical reference solutions, taking the third-order IM3 subband DPD as a simple and tractable example, in order to
demonstrate that minimizing the IM3 power is essentially
identical to decorrelating the IM3 observation against the
corresponding sub-band DPD basis functions. Then, both
sample-adaptive and block-adaptive decorrelation based learning rules are devised, covering the general cases of higherorder processing with memory at IM3 and higher-order IM
sub-bands.
(25)
p=7
p odd
In general, the achievable suppression of the distortion at
different IM sub-bands depends directly on the selection and
optimization of the different filters α3+,p,n , α5+,p,n , α7+,p,n ,
and so on. This is addressed in detail in the next section.
Here, f1 and f3 are the memoryless PA model parameters, α
denotes the memoryless sub-band DPD parameter to be optimized, while yIM 3+ (n) and ỹIM 3+ (n) refer to the baseband
equivalent PA output at the positive IM3 sub-band without and
with sub-band DPD, respectively.
Now, as (30) clearly shows, the IM3 sub-band distortion at
the PA output, with DPD adopted, depends directly on, and
can thus be controlled by, the DPD coefficient α. In the wellknown third-order inverse solution, the DPD parameter α is
selected such that the third-order term in (30) is canceled, i.e.,
f3 + f1 α = 0. The corresponding solution, denoted by αinv ,
thus reads
αinv = −f3 /f1 .
III. S UB - BAND DPD PARAMETER L EARNING U SING THE
D ECORRELATION P RINCIPLE
In this section, based on the previous spurious component
modeling and the proposed sub-band DPD principle, we
formulate computing feasible and highly efficient estimation
algorithms for learning and optimizing the sub-band DPD filter
coefficients such that the spurious emissions are minimized
at the considered IM sub-bands. We start by deriving some
(30)
(31)
However, this will not remove all the distortion, because
higher-order terms will be created due to the predistortion,
as shown in (30).
A more elaborate method for selecting the DPD parameter
α is the one that minimizes the power of the total IM3
sub-band signal, ỹIM 3+ (n), referred to as the minimum IM3
power or minimum mean-squared error (MMSE) solution in
the continuation. For notational convenience, we define the
6
so-called error signal as e(n) = ỹIM 3+ (n), as with ideal
predistortion the IM3 sub-band signal would be zero, and thus
the optimization means minimizing the power of this error
signal. The detailed derivation of the optimum DPD parameter
which minimizes the mean squared error IE[|e(n)|2 ] is given
in Appendix A, yielding
B. Sample-Adaptive Decorrelation-based Learning
In this subsection, we provide the actual sample-adaptive
learning rules for both IM3 and higher-order IM sub-band
DPD structures with memory, whose basic operating principles
were described in Section II. To facilitate the learning, we
assume a feedback or observation receiver measuring the particular IM sub-band whose sub-band DPD filter coefficients are
currently under learning. Notice that opposed to classical wideαM M SE =
∗
∗
2
band DPD principles, where a wideband observation receiver
− f1 f3 IE42 + 2|f3 | (IE62 + IE44 )
∗ , is typically needed, a more narrowband receiver is sufficient
∗
2
2
[|f1 | IE42 + 4R(f1 f3 )(IE62 + IE44 ) + 4|f3 | (IE46 + 2IE64 + IE82 )]
(32) here since only the particular IM sub-band is observed. As both
the IM3 and the higher-order IM sub-band DPDs are based
where IEij refers to the products of the CC signals’ higher- on multiple strongly correlated basis functions, given in (5)order moments of the form IE[|x1 |i ]IE[|x2 |j ], while IE[.] de- (9) and (15)-(24), respectively, we start by introducing a basis
notes the statistical expectation operator.
function orthogonalization procedure, after which the actual
As will be shown with concrete examples in Section V, proposed adaptive decorrelation algorithms are described.
this MMSE solution provides better linearization performance
1) Basis Function Orthogonalization: When a nonlinearcompared to the third-order inverse solution. However, as ity order higher than the IM sub-band order is considered,
shown in (32), the analytical MMSE solution requires the multiple basis functions are adopted in the sub-band DPD
knowledge of various higher-order moments of the CC signals. processing, as described in Section II. Taking as an example
Furthermore, the above solution is valid only in the case case the IM3+ sub-band, the SNL basis functions u
3+,p (n),
of a memoryless third-order nonlinear system, beyond which p = 1, 3, ..., Q, given in (5)-(9), are highly correlated. This
obtaining an analytical expression for the DPD coefficients can negatively impact the convergence and stability of the
becomes overly tedious. Thus, to relax these constraints, an adaptive decorrelation-based learning. Therefore, the SNL
alternative solution, based on minimizing the correlation be- basis functions are first orthogonalized [25], which yields a
tween the IM3 sub-band observation and the third-order basis new set of DPD basis functions s
3+,p (n), p = 1, 3, ..., Q,
function u(n) = x∗2 (n)x21 (n) was initially discussed in [19], written formally at sample level as
and will be extended to higher nonlinearity orders and higherorder IM sub-bands in this paper. The analytical reference
s3+ (n) = Wu3+ (n),
(35)
solution for the decorrelation-based sub-band DPD structure
is obtained by setting the correlation between the error signal,
where
e(n), and the third-order basis function, u(n) = x∗2 (n)x21 (n),
to zero, i.e., IE[e(n)u∗ (n)] = 0. Then, through straightu3+ (n) = [u3+,3 (n) u3+,5 (n) ... u3+,Q (n)]T ,
(36)
forward algebraic manipulations, the decorrelation-based DPD
T
coefficient, denoted by αo , can be shown to read
s3+ (n) = [s3+,3 (n) s3+,5 (n) ... s3+,Q (n)] ,
(37)
1
w5,3
w7,3
W=
.
.
.
αo =
−f3
I 60
+
f1 + 2f3 ( E
E
I 40
E
I 04
)
E
I 02
.
(33)
Assuming then, for simplicity, that the CC baseband equivalents are complex Gaussians, (33) simplifies to
αo =
−f3
,
2
2
f1 + 2f3 (3σx,1
+ 2σx,2
)
(34)
2
2
= IE[|x1 |2 ] and σx,2
= IE[|x2 |]2 are the variances
where σx,1
of the two CCs [19].
A clear advantage of the decorrelation-based approach
compared to the earlier MMSE solution lies in the simple
and straightforward adaptive filtering based practical computing solutions, sketched initially in [19] for simple thirdorder processing at IM3, which do not require any prior
knowledge about signal moments or any other parameters.
Furthermore, opposed to the MMSE solution, the decorrelation
based approach can be easily extended to include higher-order
nonlinearities and memory effects, at both the IM3 sub-band
as well as the other higher-order IM sub-bands, as will be
described in details in the next subsection.
wQ,3
0
w5,5
w7,5
..
.
wQ,5
0
0
w7,7
..
.
wQ,7
0
0
...
..
.
...
0
0
0
..
.
.
(38)
wQ,Q
The lower triangular matrix W can be obtained, e.g., through
Gram-Schmidt orthogonalization, QR or singular value decomposition, or using a lower complexity iterative orthogonalization algorithm [26].
Similarly, the SNL basis functions in (15)-(24) corresponding to the higher-order IM sub-bands are also orthogonalized,
to obtain new sets of orthogonal basis functions s5+ (n),
s7+ (n), etc.
2) Adaptive Learning for IM3+ Sub-band: We now present
the actual sample-adaptive decorrelation based learning algorithm for the IM3+ sub-band DPD coefficients. For notational
convenience, we introduce the following vectors
α3+,l (n) = [α3+,3,l (n) α3+,5,l (n) ... α3+,Q,l (n)]T ,
ᾱ3+ (n) = [α3+,0 (n)
T
α3+,1 (n)
T
T T
... α3+,N (n) ] ,
s̄3+ (n) = [s3+ (n)T s3+ (n − 1)T ... s3+ (n − N )T ]T ,
(39)
(40)
(41)
7
x1
x∃
x
Upsampling
and IF
upconversion
x2
S/P
Adaptive Higher Order IM3+ Sub-band DPD
x1
x1
I1
Q1
x2
SNL3
u3
s3
f RF , 3 fIF
Third Order IM3 Decorrelator
Z-1
Z-1
I
Z-1
Upsampling
and 3xIF
upconversion
1*
N*
Q
Z-D
f RF ∗ 3 f IF
PA
I/Q
LO
f
0*
f RF , f IF f RF ∗ f IF
D/A
LPF
D/A
LPF
f RF ∗ 3 f IF
yIM 3 ∗
Adaptive Algorithm
I2
x2 Q
2
f
x1
x2
x1
x2
SNL5
u5
s5
DSP
Fifth Order IM3 Decorrelator
Baseband
SNLQ
uQ
sQ
LPF
A/D
f
Qth Order IM3 Decorrelator
Fig. 3. Detailed block diagram of the sample-adaptive decorrelation-based IM3+ sub-band DPD with higher-order nonlinearities and memory.
where now α3+,p,l (n) denotes the lth adaptive filter coefficient
of the pth order orthogonalized IM3+ basis function s3+,p (n)
at time index n, and N denotes the adaptive filter memory
depth. Furthermore, the vectors ᾱ3+ (n) and s̄3+ (n) incorporate all the coefficients and basis function samples up to
polynomial order Q. Adopting this notation, the instantaneous
sample of the composite baseband equivalent PA input signal
x̃(n) now reads
objective function J(ᾱ3+ ) = IE[x̃3+ (n)e∗3+ (n)], where the
target is to search for the DPD coefficients ᾱ3+ that minimize
the ensemble correlation between e3+ (n) and the baseband
equivalent of the IM3+ injection signal x̃3+ (n) in (43). The
overall processing flow is graphically illustrated in Fig. 3.
3) Adaptive Learning for Higher-Order IM Sub-bands: We
now extend the decorrelation based learning to higher-order
IM sub-bands. First, adopting vector-based notations again,
we introduce the following vectors for notational convenience
3f
j2π fIF n
s
,
x̃(n) = x(n) + x̃3+ (n)e
(42)
where the instantaneous sample of the baseband equivalent
IM3+ injection signal x̃3+ (n) reads
x̃3+ (n) = ᾱ3+ (n)H s̄3+ (n).
(43)
α5+,l (n) = [α5+,5,l (n) α5+,7,l (n) ... α5+,Q,l (n)]T ,
(46)
T T
... α5+,N (n) ] ,
(47)
ᾱ7+ (n) = [α7+,0 (n)T α7+,1 (n)T ... α7+,N (n)T ]T ,
(48)
α7+,l (n) = [α7+,7,l (n) α7+,9,l (n) ... α7+,Q,l (n)] ,
ᾱ5+ (n) = [α5+,0 (n)
T
α5+,1 (n)
T
T
s5+ (n) = [s5+,5 (n) s5+,7 (n) ... s5+,Q (n)] ,
Then, in order to adaptively update the filter coefficients
ᾱ3+ (n), the IM3+ sub-band is observed with the feedback
receiver, and the coefficient is updated as
µ
ᾱ3+ (n + 1) = ᾱ3+ (n) −
s̄3+ (n)e∗3+ (n),
||s̄3+ (n)||2 + C
(44)
where e3+ (n) = ỹIM 3+ (n) denotes the baseband equivalent
observation of the PA output at the IM3+ sub-band with the
current DPD coefficients. The scaling factor ||s̄3+ (n)||2 + C
normalizing the learning step-size µ in (44) is philosophically
similar to that of the Normalized Least Mean Square (NLMS)
algorithm, in effect making the learning characteristics more
robust against the input data dynamics.
The proposed coefficient update in (44) is seeking to decorrelate the IM3+ observation and the adopted orthogonalized
basis functions. This type of learning algorithm can also
be interpreted as a stochastic Newton root search in the
(45)
T
(49)
s7+ (n) = [s7+,7 (n) s7+,9 (n) ... s7+,Q (n)]T ,
T
s̄5+ (n) = [s5+ (n)
T
s̄7+ (n) = [s7+ (n)
s5+ (n − 1)
T
s7+ (n − 1)
T
(50)
T T
(51)
T T
(52)
... s5+ (n − N ) ] ,
... s7+ (n − N ) ] .
Then, similar to x̃3+ (n) in (43), the baseband equivalents of
the IM5+ and IM7+ sub-band DPD injection signals, denoted
by x̃5+ (n) and x̃7+ (n), respectively, read
x̃5+ (n) = ᾱ5+ (n)H s̄5+ (n), x̃7+ (n) = ᾱ7+ (n)H s̄7+ (n).
(53)
Thus, the composite baseband equivalent PA input signal when
the IM3+, IM5+, and IM7+ sub-band DPDs are all included,
reads
5f
3f
j2π fIF n
s
x̃(n) = x(n) + x̃3+ (n)e
+ x̃5+ (n)e
j2π fIF n
s
7f
+ x̃7+ (n)e
j2π fIF n
s
.
(54)
8
x1
M Samples
L Samples
É M is the estimation block size in samples
É L is the DPD update interval in samples
Estimation
Block m
Estimation
Block m+1
Update DPD
Paramters
Update DPD
Paramters
Upsampling and
IF upconversion
x2
x1
Fig. 4. Block-adaptive DPD learning concept. The DPD parameters estimated
in the current estimation block m are applied in the sub-band DPD main path
processing during the next block.
x2
Then, similar to the IM3+ coefficient update in (44), the
higher-order IM5+ and IM7+ coefficient updates read
x
IM3 + DPD
Upsampling and
+3xIF upconversion
IM3 - DPD
Upsampling and
-3xIF upconversion
IM5 + DPD
Upsampling and
+5xIF upconversion
IM5 - DPD
Upsampling and
-5xIF upconversion
IMP + DPD
Upsampling and
+PxIF upconversion
IMP - DPD
Upsampling and
-PxIF upconversion
x
x
”RF”
I/Q
MOD
DAC
Error Signal
PA
Feedback
Receiver
Sub-band Decorrelating DPD
µ
ᾱ5+ (n + 1) = ᾱ5+ (n) −
s̄5+ (n)e∗5+ (n),
||s̄5+ (n)||2 + C
µ
s̄7+ (n)e∗7+ (n),
ᾱ7+ (n + 1) = ᾱ7+ (n) −
||s̄7+ (n)||2 + C
(55)
(a) Sub-band DPD Architecture I: Digital IF injection
(56)
where e5+ (n) = ỹIM 5+ (n) and e7+ (n) = ỹIM 7+ (n) denote
the baseband equivalents of the PA output at the IM5+ and
IM7+ sub-bands, with the corresponding sub-band DPDs included adopting the current coefficients, ᾱ5+ (n) and ᾱ7+ (n),
respectively. From the learning perspective, observing a single
IM sub-band at a time is the most obvious alternative, which
means that the learning for multiple sub-band DPDs happens
one at a time. Furthermore, extending the decorrelation-based
learning for the negative IM sub-bands is straight-forward.
This can be obtained by interchanging x1 (n) and x2 (n) in
the SNL basis functions expressions, and observing the PA
output at the corresponding negative IM sub-bands.
x1
x2
x1
x2
Upsampling and
IF upconversion
x
x
DAC 1
IM3 + DPD
DAC 2
Analog +3xIF
upconversion
IM3 - DPD
DAC 3
Analog -3xIF
upconversion
IM5 + DPD
DAC 4
Analog +5xIF
upconversion
IM5 - DPD
DAC 5
Analog -5xIF
upconversion
IMP + DPD
DAC
M-1
Analog +PxIF
upconversion
IMP - DPD
DAC
M
Analog -PxIF
upconversion
”RF”
I/Q
MOD
PA
Error Signal
Feedback
Receiver
Sub-band Decorrelating DPD
C. Block-Adaptive Decorrelation-based Learning
As Fig. 3 illustrates, the previous sample-adaptive
decorrelation-based learning concept is in principle a closedloop feedback system with nonlinear adaptive processing
inside the loop. During the DPD learning phase, and under
the potential hardware processing and latency constraints,
the DPD parameter convergence and consequently the DPD
linearization performance can be affected, especially if the
learning loop delay becomes large. Stemming from this, an
alternative and new block-adaptive decorrelation-based learning solution is developed next.
The proposed block-adaptive learning rule implies defining
two distinct blocks in the processing, as illustrated in Fig. 4.
A single update cycle of the learning algorithm will utilize
M samples whereas the DPD parameter update interval is L
samples, with M ≤ L. Thus, by proper choice of M and
L, arbitrarily long loop delays can in principle be tolerated,
facilitating stable operation under various hardware and software processing latency constraints. Notice that proper timing
synchronization between the observation receiver output and
the basis functions is, in general, needed, which can be
accomplished prior to executing the actual coefficient learning
procedure.
Now, assuming an estimation block size of M samples and
DPD filter memory depth of N per each of the IM3+ orthogonalized basis functions, the following vectors and matrices,
(b) Sub-band DPD Architecture II: Analog IF injection
Fig. 5. Overall sub-band DPD architecture with multiple IM sub-bands
included. Thick lines indicate complex I/Q processing.
which stack all the samples and the corresponding DPD filter
coefficients within block m, can be defined:
s3+ (nm ) = [s3+,3 (nm ) s3+,5 (nm ) ... s3+,Q (nm )]T ,
T
T T
s̄3+ (nm ) = [s3+ (nm ) ... s3+ (nm − N ) ] ,
S3+ (m) = [s̄3+ (nm ) ... s̄3+ (nm + M − 1)],
α3+,l (m) = [α3+,3,l (m) α3+,5,l (m) ... α3+,Q,l (m)]T ,
ᾱ3+ (m) = [α3+,0 (m)T α3+,1 (m)T ... α3+,N (m)T ]T .
(57)
(58)
(59)
(60)
(61)
Here, the index of the first sample of block m is denoted
by nm . The block-adaptive IM3+ sub-band DPD coefficient
update then reads
e3+ (m) = [ỹIM 3+ (nm ) ... ỹIM 3+ (nm + M − 1)]T , (62)
µ
ᾱ3+ (m + 1) = ᾱ3+ (m) −
S3+ (m)e∗3+ (m),
||S3+ (m)||2 + C
(63)
where e∗3+ (m) refers to the element-wise conjugated error
signal vector, while S3+ (m) denotes the filter input data
matrix, all within the processing block m. The obtained new
9
TABLE I
C OMPARISON OF RUNNING COMPLEXITIES OF NINTH ORDER SUB - BAND AND FULL - BAND DPD S . T WO 1 MH Z CC S WITH 20 MH Z CARRIER SPACING IS
ASSUMED . DPD SAMPLE RATE IS 189 MSPS FOR THE FULL - BAND AND 9 MSPS FOR THE SUB - BAND DPD.
Basis Function Generation FLOPs
DPD Filtering FLOPs
Total Number of FLOPs
GFLOPS (N1 = 1, N2 = 3)
IM3+/- Sub-band
37
32(N1 +1) - 2
32(N1 +1) + 35
0.891
Sub-band DPD
IM5+/- Sub-band
IM7+/- Sub-band
40
45
24(N1 +1) - 2
16(N1 +1) - 2
24(N1 +1) + 38
16(N1 +1) + 43
0.774
0.675
DPD coefficients ᾱ3+ (m + 1) are then applied on the next
block of L samples as illustrated in Fig. 4. While the above
presentation describes the block-adaptive learning only at the
IM3+ sub-band, extending the principle to higher-order IM
sub-bands is straightforward, and thus not explicitly shown.
IV. I MPLEMENTATION A SPECTS AND C OMPLEXITY
A NALYSIS
One of the main advantages of the proposed decorrelationbased sub-band DPD technique is its reduced complexity
compared to the classical full-band DPD processing, especially
in scenarios where the CCs are widely spaced and thus very
high speed ADCs and DACs are required in the classical
solutions. In this section, we first address some implementation aspects of the proposed sub-band DPD concept. We
then provide a thorough comparison of the computing and
hardware complexity perspectives between the full-band and
the proposed sub-band DPDs. Finally, some system power
efficiency considerations are also presented, with particular
focus on low-cost mobile devices.
A. Sub-band DPD Implementation Alternatives
Fig. 5 shows two alternative architectures of the overall
sub-band DPD processing with multiple IM sub-bands. The
first architecture shown in Fig. 5(a) adds an upsampling and
digital IF upconversion block after each sub-band DPD stage
in order to digitally place the generated injection signals at
proper intermediate frequencies. After adding the sub-band
DPD outputs in the digital domain, a single wideband DAC per
I and Q branch is used after which the signal is upconverted
to RF and amplified by the PA.
The second architecture in Fig. 5(b) adds the outputs of
the sub-band DPDs in the analog domain, implying that each
sub-band DPD is followed by a narrow-band DAC per I and
Q branch, together with an analog complex IF upconversion.
A common I/Q RF modulator is then used for all sub-bands
prior to the PA module.
Both of these two architectures have their own advantages
and disadvantages. In particular, if the carrier spacing between
the component carriers is not very large, architecture I is
likely to be more suitable. On the other hand, when the
carrier spacing starts to increase, using a single wideband
DAC may not be efficient from cost and power consumption
perspectives, so in this case architecture II is likely to be more
attractive. However, some extra processing may be required
in this architecture in order to achieve proper synchronization
between the DACs. This forms an interesting topic for future
research.
IM9+/- Sub-band
48
8(N1 +1) - 2
8(N1 +1) + 46
0.558
Full-band DPD
Full-band
11
40(N2 +1) - 2
40(N2 +1) + 9
31.941
A common advantage of both architectures is that each subband DPD block can be switched on or off according to the
prevailing emission levels and limits at the considered subbands. Such flexibility is not available in the classical fullband DPD solutions since, by design, a full-band predistorter
always tries to linearize the full composite transmit band. With
the sub-band DPD concept, the linearization can be flexibly
tailored and optimized to those frequencies which are the most
critical from the emission limits perspective.
B. Sub-band versus Full-band DPD Running Complexity
In general, the computational complexity of any DPD can
be classified into three main parts [24]: identification complexity, adaptation complexity and running complexity. The
identification part is basically the estimation complexity of
the DPD parameters, while the adaptation complexity includes
the required processing in order to adapt to new operating
conditions or device aging. Finally, the running complexity,
which is the most critical especially for mobile-type devices,
involves the number of computations done per second while
the DPD is operating. In this subsection, we will focus on the
DPD running complexity in details, while the DPD parameter
identification and feedback receiver complexity perspectives
are shortly discussed in section IV-C. For a quantitative
comparison of the running complexities, we shall use the
number of floating point operations (FLOPs) per sample, the
number of DPD coefficients, and the required sample rate in
the predistortion path as the main quantitative metrics.
In general, the running complexity is divided into two
main parts: the first is the basis function generation, and
the second is the actual predistortion filtering using the basis
functions [24]. The number of FLOPs required to perform
these two operations is shown in Table I for ninth-order
sub-band and full-band DPDs, respectively, where N1 and
N2 are the corresponding memory depths per adopted basis
function. The full-band DPD architecture that we use also in
our comparative performance simulations in Section V, and
which is also widely applied otherwise, is based on the PH
architecture with ninth order nonlinearity, while the sub-band
DPD is based on architecture II shown in Fig. 5(b). Frequency
selectivity of the nonlinear PA is another important factor to be
considered when comparing the two DPD architectures. This
implies that when memory effects of the PA are considered,
substantially longer filters are needed for the full-band DPD
compared to the sub-band DPD for a certain performance
requirement. In the complexity analysis, in the sub-band DPD
case, we thus assume a memory depth of 1 per basis function,
10
while a memory depth of 3 is assumed for the full-band PH
DPD per basis function.
A substantial reduction in the needed sample rates can be
achieved when adopting the sub-band DPD. As a concrete
example, if we consider a challenging LTE-A UL scenario of
two 1 MHz CCs with 20 MHz carrier spacing, the required
sample rate by the full-band DPD is 9×(20+1) = 189M SP S,
while for the sub-band DPD it becomes 9 × 1 = 9M SP S.
Consequently, as shown in Table I, a huge reduction in the
number of FLOPs per second (FLOPS) can be achieved using
the sub-band DPD. Furthermore, the processing complexity
of the sub-band DPD solution is clearly feasible for modern
mobile device processing platforms, in terms of the GFLOPS
count, while that of the classical fullband DPD is clearly
infeasible. However, in some scenarios, when the carrier spacing between the CCs is decreased and/or the CC bandwidth
increases, the benefit from using the sub-band DPD approach
is reduced since the sample rates of the sub-band and fullband DPD will become more comparable. Notice also that,
e.g., in interband CA cases, one can adopt a concurrent 2DDPD [13], [14] to linearize the main carriers, which can then
be easily complemented with the sub-band DPD processing
to protect the own receiver if TX filtering does not offer
sufficient isolation. Such a scenario could take place, e.g., in
uplink Band 1 (1920-1980 MHz) + Band 3 (1710-1785 MHz)
interband CA where it is technologically feasible to adopt a
single multiband/multimode PA module for amplification. In
such scenario, the IM3 sub-band is then directly at the own
RX frequencies of Band 1 (2110-2170 MHz).
C. Feedback Receiver and Parameter Estimation Complexity
In addition to the complexity reduction in the DPD main
processing path, the complexity of the feedback or observation
receiver, used for DPD parameter estimation and adaptation, is
also greatly reduced. In order to estimate, e.g., the parameters
of the IM3+ sub-band DPD, we only need to observe the
IM3+ sub-band at the PA output, instead of observing the fullband (including all the IM sub-bands) which is the case in the
full-band DPD. This reduces the cost, complexity and power
consumption in the feedback path allowing use of simpler instrumentation, in particular the ADC. Moreover, only a single
observation receiver is required, even if linearizing multiple
sub-bands, since the parameter learning and corresponding
observation of the PA output can be done sequentially, in
a per sub-band manner. Finally, in terms of the parameter
estimation algorithmic complexity, the proposed decorrelation
based solutions are extremely simple, when compared to any
classical full-band DPD related method, such as least-squares
based parameter fitting and indirect learning architecture [8]
that are commonly adopted.
D. Overall Transmitter Power Efficiency Perspectives
When a DPD is adopted, a less linear but more efficient
PA, that can operate near its saturation region, can generally
be used. However, the overall power efficiency of the device is
only improved if the extra power consumed by the DPD stage
is less than the power savings due to increased PA efficiency.
Here, we address this aspect from a mobile device perspective.
We consider a practical scenario where the transmit power at
the output of the mobile PA is +26 dBm (i.e., 400 mW), stemming from 3GPP LTE-Advanced requirements [2], [5], and
assuming that the Tx duplexer filter and connector insertion
losses are 3 dB. Then, good examples of practical PA power
efficiency figures when operating in highly nonlinear or linear
modes are around 35% and 20% (or even less), respectively.
This means that the power consumed by the highly nonlinear
PA is roughly 1150 mwatt, while the corresponding linear PA
consumes roughly 2000 mwatt. In other words, adopting a
highly power-efficient nonlinear PA saves 850 mwatt of power
in this particular example. Then, in order to suppress the
spurious emissions at the PA output, in the nonlinear PA case,
we adopt the sub-band DPD solution.
In [27], a state of the art 28 nm implementation of the
Qualcomm Hexagon DSP capable of supporting up to 4.8
GFLOPS at 1.2 GHz is reported, which is more than enough
for carrying out the needed sub-band DPD processing for
the IM3 and IM5 sub-bands in our example. In the fullband DPD case, on the other hand, which requires 31.9
GFLOPS as shown in Table I, this is clearly insufficient. The
power consumption of the DSP platform in [27] is shown
to be approximately 100 mwatt, when running at 840 MHz
(i.e., 3.36 GFLOPS assuming 4 FLOPS per cycle), which is
again sufficient for linearizing the IM3 and IM5 sub-bands
in our example. Thus, adopting the nonlinear PA already
saves 850 milliwatt, when it comes to the PA interface, as
explained above, while when complemented with the sub-band
DPD processing for enhanced linearity, only 100 milliwatt of
additional power is consumed. Thus, the overall power budget
and power-efficiency are clearly in favor of using a highly nonlinear PA, complemented with the sub-band DPD structure,
even in a mobile device with non-contiguous uplink carrier
aggregation. Furthermore, if the sub-band DPD processing
is implemented using a dedicated hardware solution (e.g., a
digital ASIC), an even more power-efficient DPD stage can
most likely be realized.
V. S IMULATION R ESULTS
In this section, a quantitative performance analysis of the
proposed sub-band DPD solution is presented using Matlab simulations with practical models for mobile-like PAs
designed for low-cost devices. In general, we quantify the
suppression of intermodulation power at the IM3 and IM5
sub-bands through the power ratios relative to the component
carrier wanted signal power as shown in Fig. 1, which are
defined as
Pwanted
Pwanted
, IM 5R = 10 log10
.
IM 3R = 10 log10
PIM 3
PIM 5
(64)
The inband transmit waveform purity is measured through the
error vector magnitude (EVM), which is defined as
s
Perror
EV M% =
× 100%.
(65)
Pref
11
TABLE II
C OMPARISON OF IM3+ SUPPRESSION AND EVM OF THIRD - ORDER
INVERSE , MMSE AND DECORRELATION - BASED ANALYTICAL SUB - BAND
DPD SOLUTIONS AT +23 D B M PA OUTPUT POWER
No DPD
Third-order Inverse
Analytical Decorrelation-based
Analytical MMSE
Positive IM3R [dBc]
28.8342
41.9230
51.8248
51.8468
EVM [%]
1.9711
1.8912
1.8914
1.8917
20
10
0
-10
-20
Here, Perror is the power of the error signal, while Pref is the
reference power of the ideal symbol constellation. The error
signal is defined as the difference between the ideal symbol
values and the corresponding synchronized and equalized
samples at the PA output, both normalized to have identical
linear gains.
-30
-40
-50
-20
-10
0
10
20
A. Comparison of Analytical IM3 Sub-band DPD Solutions
In Section III-A, analytical reference expressions for the
third-order inverse, minimum MSE, and decorrelation-based
sub-band DPD solutions were presented. In this subsection,
we shortly evaluate and compare the performance of these
analytical solutions, assuming a third-order memoryless PA
and known parameters, and focus only on the positive IM3
sub-band for simplicity. The memoryless PA model has been
identified using a true mobile PA, with transmit power of
+23dBm. The main objective is to compare the performance
of the three analytical solutions in terms of linearization
performance, and thereon to verify that the decorrelation based
solution is essentially identifcal to the minimum MSE solution.
The signal used in this performance evaluation is composed
of two 1MHz LTE-A UL SC-FDMA CCs with QPSK data
modulation, and the CC spacing is 10MHz. The obtained
results shown in Table II demonstrate that the decorrelationbased solution is giving almost the same performance as
the minimum MSE solution, and also that both of them
are substantially better than the classical third-order inverse
solution [16] in terms of the IMD suppression at the considered
sub-band. It can also be seen from Table II that the EVM is
essentially not affected by the sub-band DPD processing.
B. P th -order Inverse vs. Decorrelation-based IM3 Sub-band
DPD Solutions
Next we evaluate and compare the performance of 5th-order
inverse and decorrelation-based IM3 sub-band DPD solutions.
The 5th-order inverse reference solution for IM3+ sub-band
is derived in Appendix B. In these simulations, the transmit
waveform is again composed of two 1 MHz LTE-A UL SCFDMA component carriers with QPSK data modulation, and
the CC spacing is 10 MHz. The PA model, in turn, is a
memoryless 5th-order model whose parameters have been
identified using a true mobile PA transmitting at +23 dBm.
The 5th-order inverse DPD is using known parameters, while
the decorrelation-based one is adopting the proposed sampleadaptive learning described in detail in Section IV.
The PA output spectra with different solutions are illustrated
in Fig. 6. It can clearly be seen that the decorrelation-based
Fig. 6. Baseband equivalent PA output spectrum at +23 dBm with two 1 MHz
LTE-A UL carriers and 10 MHz CC separation using a fifth-order memoryless
PA. P th -order inverse and decorrelation-based IM3+ sub-band DPD solutions
are compared, with both third-order and fifth-order nonlinear processing.
sub-band DPD substantially outperforms the P th -order inverse
based sub-band DPD, in both third and fifth-order cases,
despite the fact that the P th -order inverse solutions are using
known parameters. The reason is that the inverse solutions
cancel only the third and fifth order terms, as described in Appendix B, but do not suppress the other induced higher-order
terms, which have structural similarity and correlation with
the third and fifth-order basis functions. On the other hand, the
proposed decorrelation-based solution takes this explicitly into
account, and thus achieves clearly better spurious emission
suppression. The figure also illustrates that the performance
of the fifth-order decorrelation-based sub-band DPD is clearly
better than that of the third-order one.
In the remaining parts of this section, we shall focus
on more detailed performance evaluations of the proposed
decorrelation-based sub-band DPD solution, incorporating
memory both in the PA and in the predistortion processing.
From now on, we refer to the decorrelation-based sub-band
DPD simply as ’sub-band DPD’, to simplify the presentation.
C. Performance of Proposed Higher-Order Sub-band DPD at
IM3, IM5 and IM7
In this subsection, we evaluate the performance of the
proposed sub-band DPD in the more realistic case of having
memory in the PA. The PA model is a 9th-order parallel
Hammerstein model, with four memory taps per branch, and
the parameters have been identified using measurements with
a true mobile PA transmitting at +24 dBm. The sub-band
DPD structure contains now also memory, with two taps
(N = 1) per basis function. Block-adaptive learning principle
is adopted, with 200 blocks each containing 1000 samples.
The transmit waveform is otherwise identical to earlier cases,
but the CC separation is now 12 MHz.
Fig. 7 shows the effectiveness of the proposed higherorder sub-band DPDs, here processing the IM3+ sub-band,
12
TABLE III
C OMPARISON OF QUANTITATIVE RUNNING COMPLEXITY AND LINEARIZATION PERFORMANCE OF FULL - BAND VERSUS SUB - BAND DPD. T WO 1 MH Z
LTE-A UL CARRIERS WITH QPSK DATA MODULATION AND 20 MH Z CARRIER SPACING ARE USED .
No DPD
Full-Band ILA DPD
IM3+ Sub-Band DPD
IM5+ Sub-Band DPD
DPD
Coeffs
N/A
20
8
6
Running Complexity
Fs [MSPS]
GFLOPS
N/A
N/A
189
31.941
9
0.891
9
0.774
EVM [%]
1.2527
0.1058
1.2489
1.2529
Transmitter Performance
Positive IM3R [dBc]
Positive IM5R [dBc]
35.5744
56.4272
61.9428
63.9842
68.3291
N/A
N/A
71.9823
20
20
10
10
0
0
-10
-10
-20
-20
-30
-30
-40
-40
-50
-50
-60
-60
-70
-60
-40
-20
0
20
40
60
Fig. 7. Baseband equivalent PA output spectrum at +24 dBm with two 1
MHz LTE-A UL carriers and 12 MHz CC separation using a ninth-order PA
with memory. Different orders of the IM3+ sub-band DPD are compared, with
memory depth equal to 1 per DPD SNL basis function.
-10
-15
-20
-25
-30
-35
-40
-45
-50
-55
-60
8
10
12
14
16
18
20
22
Fig. 8. IM3 spurious emissions vs. TX power using different orders of the
IM3+ sub-band DPD processing, with memory depth equal to 1 per DPD
SNL basis function. A ninth-order PA with memory is used, and 2 dB Tx
filter insertion loss is assumed.
compared to the basic third-order solution presented earlier in
[20]. Up to 40 dB suppression for the IM3 spurious emissions
is shown when adopting ninth-order processing, something
-70
-60
-40
-20
0
Output Power [dBm]
+20
+19
+20
+20
20
40
60
Fig. 9. Baseband equivalent of two 1 MHz LTE-A UL carriers with 12 MHz
separation using a ninth order PH PA extracted from a real mobile PA at +24
dBm. Ninth order negative IM3, IM5, and IM7 sub-band DPD solutions are
shown, with memory depth equal to 1 per DPD SNL basis function.
that has not been reported before in any prior works using
a practical PA model with memory effects. Fig. 8 shows then
how the IM3 spurious emission level changes with varying
the TX power and using different sub-band DPD orders. The
IM3 spurious emissions are clearly below the general spurious
emission limit (-30 dBm/MHz) even at very high TX powers
up to +23 dBm, when using a seventh-order, or higher, subband DPD.
Another main contribution in this paper is the extension of
the IM3 sub-band DPD solution to include also higher-order
IM sub-bands. Fig. 9 shows the performance of the negative
IM5 and IM7 sub-band DPDs, in addition to the negative
IM3 sub-band DPD, when ninth-order nonlinear processing
is adopted in the sub-band DPDs. Up to 40 dB, 35 dB, and
18 dB of suppression is achieved at the negative IM3, IM5,
and IM7 sub-bands, respectively. This shows the effectiveness
of the proposed solutions in processing and suppressing also
the spurious emissions at higher IM sub-bands.
D. Full-Band versus Sub-band DPD Complexity and Performance Analysis
In this subsection, we compare the performance of our
proposed sub-band DPD technique against that of a classical
full-band DPD adopting parallel Hammerstein based wideband
linearization and indirect learning architecture (ILA). The
13
(a)
Host-processor based processing:
· Generate baseband I/Q samples.
· Transfer block of samples to VST for I/Q
modulation and RF upconversion.
· Extract received block of samples from VST after
RF downconversion and I/Q demodulation.
· Estimate sub-band DPD parameters using blockadaptive estimation processing and transmit
predistorted signal to VST.
NI Vector Signal Transceiver
VST RF Out
Mobile PA
Attenuator
VST RF In
(b)
Fig. 10. Hardware setup used in the RF measurements for testing and evaluating the proposed sub-band DPD.
evaluation considers the linearization performance, complexity, and transmitter efficiency utilizing our earlier complexity
analysis results reported in Section IV. The ninth-order fullband PH ILA DPD uses 100k samples for parameter learning,
per ILA iteration, with a total of 2 ILA iterations, and the
number of memory taps per PH branch is 4. The ninth-order
block-adpative sub-band DPD, in turn, uses also a total of 200k
samples with block size M = 1000, and adopts two memory
taps per DPD SNL basis function.
The results are collected in Table III, which shows that
in addition to the significantly lower complexity measured
by the number of GFLOPS, the sub-band DPD achieves
better linearization performance in terms of spurious IMD
suppression at the considered IM3 and IM5 sub-bands. On
the other hand, the full-band DPD outperforms the sub-band
DPD in terms of the inband distortion mitigation (i.e., EVM).
This is expected, since the full-band DPD linearizes the whole
transmit band, including the main CCs and the IMD spurious
emissions. However, the EVM with the sub-band DPD is only
around 1.25% which is by far sufficient for modulations at
least up to 64-QAM. Additionally, the full-band DPD based
on the ILA structure requires an additional 1 to 2 dB back
off to guarantee stable operation, which is not required in
the sub-band DPD. Thus the transmitter becomes more power
efficient when using the sub-band DPD as shown in Table III.
For fairness, it is to be acknowledged that a full-band DPD
can typically enhance the EVM and ACLR, while the subband DPD concept is specifically targeting only the spurious
emissions.
VI. RF M EASUREMENT R ESULTS
In order to further demonstrate the operation of the proposed sub-band DPD solution, we next report results of
comprehensive RF measurements using a commercial LTEAdvanced mobile terminal PA together with a vector signal
transceiver (VST), which is implementing the RF modulation
and demodulation. The actual sub-band DPD processing and
parameter learning algorithms are running on a host processor.
The ACPM-5002-TR1 mobile power amplifier used in our
measurements is designed for LTE-A UL band 25 (18501915 MHz), with 29 dB gain. The National Instruments (NI)
PXIe-5645R VST includes both a vector signal generator
(VSG), and a vector signal analyzer (VSA) with 80 MHz
instantaneous bandwidth. In these experiments, the digital
baseband waveform is divided into 50 blocks of size M = 10k
samples which are first generated locally on the host processor,
and then transferred to the VSG to perform RF I/Q modulation
at the desired power level at the PA input. The VST RF output
is then connected to the input port of the external power
amplifier, whose output port is connected to the VST RF input
through a 40 dB attenuator, implementing the observation
receiver as illustrated also in Fig. 10. The VSA performs RF
I/Q demodulation to bring the signal back to baseband. The
baseband I/Q observation block is then filtered to select the
IM3 sub-band which is used for block-based sub-band DPD
learning, after proper alignment with the locally generated
basis functions, as explained in Section III-C. The sub-band
DPD block size M used in these experiments is 10k, and the
DPD memory depth N is 1.
In general, two different intra-band LTE-A CA RF measurement examples are demonstrated in this subsection. The first
experiment demonstrates the violation of the spurious emission
limit due to the inband emission of the IM3+ spur, thus not
being attenuated by the TX filter. The second experiment
demonstrates an own RX desensitization example, in FDD
transceiver context, where the IM3+ spur is located at the own
RX band (1930-1995 MHz) and is not sufficiently attenuated
by the duplexer TX filter. Notice that in principle, when
tackling specifically the own RX desensitization problem with
the proposed sub-band DPD solution, the main receiver of
14
A. Spurious Emission Limit Violation
20
10
0
Tx Band
-10
-20
Spur Limit
-30
-40
-50
1810
1830
1850
1870
1890
1910
Fig. 11. An LTE-A band 25 RF measurement example at the UE PA output
showing the gain from using an IM3+ sub-band DPD. IM3 spur reduction
with third, fifth, seventh, and ninth-order sub-band DPDs are demonstrated,
using a real commercial mobile PA operating at +25 dBm. An LTE-A UL
CA signal with two 3 MHz CCs and 20 MHz carrier spacing is used.
20
10
0
Tx Band
-10
Rx Band
-20
-30
-40
-50
1835
1855
1875
1895
1915
1935
Fig. 12. An LTE-A band 25 RF measurement example at the UE PA output
showing the gain from using a ninth-order IM3+ sub-band DPD when the
IM3+ is falling at own RX band. An LTE-A UL CA signal with two 5 MHz
CCs and 40 MHz carrier spacing is used with a real commercial mobile PA
operating at +25 dBm.
the FDD device could potentially be used as the observation
receiver (i.e., learning the sub-band DPD coefficients without
an extra auxiliary/observation receiver). However, such an
approach would indeed be applicable only in the own RX
desensitization case, while when mitigating other harmful
emissions, an auxiliary observation receiver would anyway
have to be adopted. Thus, in these RF measurements, we adopt
the auxiliary receiver based approach, while consider further
developments of the main receiver based parameter learning
as an important topic for our future work.
Fig. 11 shows the measured power spectral density at the
PA output using the IM3 sub-band DPD with different orders.
The adopted waveform is an intra-band CA LTE-A UL signal
with 3 MHz per CC and 20 MHz carrier spacing. The power
level at the PA output in this example is +25 dBm. The
intermodulation distortion at the IM3+ sub-band is emitted
inband, since the total TX band covers 1850-1915 MHz, and is
clearly violating the spurious emission limit (-30 dBm/MHz)
when no DPD is used. When using the sub-band DPD, the
spurious emission level is well below the emission limit, given
that at least 5th-order processing is deployed. In general, more
than 20 dB of measured spurious emission suppression is
achieved with the ninth order IM3 sub-band DPD, thus giving
more than 10 dB additional gain compared to the basic thirdorder sub-band DPD as shown in Fig. 11. Notice that there is
no need for predistorting the IM3- sub-band in this example,
since those emissions will be filtered out by the TX filter.
B. Own Receiver Desensitization
Fig. 12 illustrates another LTE-A intra-band CA UL band
25 example, now with 5 MHz CC bandwidths and 40 MHz
carrier spacing. The duplexing distance at LTE-A band 25 is
only 80 MHz, thus the IM3+ spur at 1935 MHz in this example
is falling on the own RX band, and therefore potentially
desensitizing the own receiver. The power level at the PA
output in this example is +25 dBm.
Then, by adopting a ninth-order IM3+ sub-band DPD, more
than 20 dB of spurious emission suppression is achieved, as
can be seen in Fig. 12. Again, two taps (N = 1) per basis
function are used, and block-adaptive parameter learning is
deployed. Fig. 13 shows the real-time convergence of the
sub-band DPD coefficients for the third, fifth, seventh, and
ninth-order basis functions, respectively. It can be seen that
the coefficients converge in a stable manner in a real RF
environment due to the orthogonalization of the SNL basis
functions, as explained in Section III-B.
In general, assuming a UE duplexer TX filter with 65 dB
attenuation at the own RX band (e.g., ACMD-6125 Band
25 LTE-A UE Duplexer), the integrated power of the IM3+
spur at the RX band without DPD will be approximately -73
dBm/5MHz. This is 25 dB above the effective RX noise floor
when assuming 9 dB UE RX noise figure (NF) [28]. This
would thus cause significant own receiver desensitization that
could lead to a complete blocking of the desired RX signal.
On the other hand, when the proposed IM3+ sub-band DPD
is deployed, the integrated power of the IM3+ spur at the RX
band will be approximately -95 dBm/5MHz, which is only 3
dB above the effective RX noise floor, as shown also in Fig. 14.
Though the residual spur is still slightly above the effective RX
noise floor, the sensitivity degradation is substantially relaxed,
despite operating at a maximum PA output power of +25
dBm. We elaborate on this further in Fig. 14, showing the
integrated power of the IM3+ spur at the RX LNA input while
changing the PA output power level. With a ninth order subband DPD, we can transmit up to +18 dBm PA output power
with an effectively perfectly linear TX, in terms of IM3+ spur
15
15
10
5
0
1
2
3
4
5
Fig. 13. Convergence of the ninth-order sub-band DPD coefficients, with
memory depth N = 1 per basis function, when the positive IM3 sub-band
is considered and block-adaptive learning is deployed. An RF measurement
example with two 5 MHz LTE-A UL CCs and 40 MHz carrier spacing is
used with a real commercial mobile PA operating at +25 dBm.
methods were also formulated, allowing efficient estimation
and tracking with low computational complexity. All algorithm
derivations and modeling were carried out in the general case
of having memory in the PA as well as in the sub-band
DPD processing. Different nonlinear distortion and processing
orders, beyond classical third-order cases, were also reported.
The proposed technique can find application in suppressing
inband spurs which would violate the spurious emission limit,
suppressing out-of-band spurs falling, e.g., on the own receiver
band, or in protecting primary user transmissions in cognitive radio systems. A quantitative complexity analysis was
presented, comparing the proposed solution to conventional
full-band DPD solutions available in the literature. The performance was evaluated in a comprehensive manner, showing
excellent linearization performance despite the considerably
reduced complexity compared to classical full-band solutions.
Finally, extensive RF measurement results using a commercial LTE-Advanced mobile power amplifier were reported,
evidencing up to 22 dB suppression of the most problematic
third-order spurious emissions.
A PPENDIX A: A NALYTICAL MMSE S OLUTION FOR
T HIRD -O RDER IM3 S UB - BAND DPD
Here, we derive the analytical minimum mean-square error
solution shown in (32), which is used as a reference solution
in simulations in Section V-A. We first define the so-called
error signal as e(n) = ỹIM 3+ (n), since with ideal predistortion the IM3+ sub-band signal would be zero, and thus the
optimization means minimizing the power of this error signal.
From (30), the error signal reads
e(n) = (f3 + f1 α)u(n) + 2f3 α(|x1 (n)|2 + |x2 (n)|2 )u(n)
+ f3 |α|2 α|x1 (n)|4 |x2 (n)|2 u(n),
Fig. 14. Integrated IM3+ sub-band power at own RX LNA input over 5 MHz
vs. PA output power using ninth order IM3+ sub-band DPD processing, with
memory depth equal to 1 per DPD SNL basis function. A commercial mobile
PA operating at LTE band 25 is used in the RF measurements, and 65 dB
duplexer TX filter attenuation [?] is assumed at the RX band.
level, compared to +10 dBm PA output power without DPD.
Additionally, with DPD, the integrated IM3+ power is less
than the effective RX noise floor up to +23 dBm PA output
power, while without DPD it is already 20 dB above the noise
floor at the same power level.
VII. C ONCLUSIONS
In this article, a novel low-complexity sub-band digital
predistortion (DPD) solution was proposed for suppressing
unwanted spurious emissions in non-contiguous spectrum access. Novel decorrelation based adaptive parameter learning
(66)
where f1 and f3 are the third-order memoryless PA model
parameters, α denotes the DPD coefficient to be optimized,
and x1 (n) and x2 (n) are the baseband equivalents of the two
component carriers. The statistical expectation IE[|e(n)|2 ] =
IE[e(n)e∗ (n)], assuming that the component carrier signals
x1 (n) and x2 (n) are statistically independent, and ignoring
some high-order vanishingly small terms, then reads
IE[e(n)e∗ (n)] = IE42 |f3 |2 + |f1 |2 |α|2 + αf1 f3∗ + α∗ f1∗ f3
+ (2IE62 + 2IE44 ) |f3 |2 (α + α∗ ) + |α|2 (f1 f3∗ + f1∗ f3 )
+ (4IE46 + 8IE64 + 4IE82 )|f3 |2 |α|2 ,
(67)
where IEij is used as a shorthand notation for IE[|x1 |i ]IE[|x2 |j ].
Now, differentiating (67) with respect to α, yields
∂
IE[e(n)e∗ (n)] = IE42 α∗ |f1 |2 + f1 f3∗
∂α
+ (2IE62 + 2IE44 ) |f3 |2 + α∗ (f1 f3∗ + f1∗ f3 )
+ (4IE46 + 8IE64 + 4IE82 )|f3 |2 α∗ .
(68)
Then, setting (68) to zero and solving for α yields the optimal
MMSE DPD parameter, given by
16
α3,inv = −
αM M SE =
∗
− f1 f3∗ IE42 + 2|f3 |2 (IE62 + IE44 )
,
[|f1 |2 IE42 + 4R(f1 f3∗ )(IE62 + IE44 ) + 4|f3 |2 (IE46 + 2IE64 + IE82 )]∗
(69)
where R(.) denotes the real-part operator. This concludes the
proof.
A PPENDIX B: A NALYTICAL F IFTH -O RDER I NVERSE
BASED IM3 S UB - BAND DPD S OLUTION
Here, we derive the 5th order inverse solution for the IM3+
sub-band DPD, to be used as a reference solution in the
simulations in Section V-B. The output of a memoryless 5th
order polynomial PA model is
y(n) = f1 x(n) + f3 |x(n)|2 x(n) + f5 |x(n)|4 x(n),
(70)
where f1 , f3 and f5 are the polynomial coefficients, and x(n)
is the composite baseband equivalent input signal, as given in
(1). Through direct substitution of (1) in (70), the baseband
equivalent IM3+ distortion term, located at three times the IF
frequency, can be extracted and reads
yIM 3+ (n) = f3 (x∗2 (n)x21 (n))
+ 2f5 |x1 (n)|2 (x∗2 (n)x21 (n))
+ 3f5 |x2 (n)|2 (x∗2 (n)x21 (n)).
(71)
Stemming from the signal structure in (71), the sub-band
DPD injection signal is composed of three basis functions of the form x∗2 (n)x21 (n), |x1 (n)|2 x∗2 (n)x21 (n) and
|x2 (n)|2 x∗2 (n)x21 (n). In case of fifth-order inverse based subband DPD, these basis functions are multiplied by proper
coefficients such that all the distortion terms at the IM3+ subband at the PA output, up to order five, are cancelled. Thus,
incorporating such DPD processing, yet with arbitrary coefficients, the composite baseband equivalent PA input signal
reads
f
j2π fIF n
x̃(n) = x1 (n)e
s
f
−j2π fIF n
+ x2 (n)e
3f
j2π fIF
s
+ α3,inv (x∗2 (n)x21 (n))e
s
n
3f
+ α51,inv |x1 (n)|2 (x∗2 (n)x21 (n))e
j2π fIF n
s
+ α52,inv |x2 (n)|2 (x∗2 (n)x21 (n))e
j2π fIF n
s
,
(72)
3f
where the subscript inv in the coefficients is emphasizing the
P th -order inverse based solution. Substituting (72) in (70),
and extracting the third and fifth order IM3+ terms, yields
ỹ IM 3+ (n) = (f1 α3,inv + f3 )x∗2 (n)x21 (n)
+ (f1 α51,inv + 2f3 α3,inv + 2f5 )|x1 (n)|2 (x∗2 (n)x21 (n))
(73)
+ (f1 α52,inv + 2f3 α3,inv + 3f5 )|x2 (n)|2 (x∗2 (n)x21 (n)).
From (73), we now can easily obtain the fifth-order inverse
coefficients that null the third and fifth-order terms, yielding
α51,inv = 2
f3
,
f1
f5
f5
f32
f2
− 2 , α52,inv = 2 32 − 3 .
2
f1
f1
f1
f1
This concludes the derivation.
R EFERENCES
[1] S. Parkvall, A. Furuskar, and E. Dahlman, “Evolution of LTE toward
IMT-advanced,” IEEE Commun. Mag., vol. 49, no. 2, pp. 84–91,
February 2011.
[2] LTE Evolved Universal Terrestrial Radio Access (E-UTRA) User Equipment (UE) radio transmission and reception, 3GPP TS 36.101 V12.4.0
(Release 12), June 2014.
[3] S.A. Bassam, W. Chen, M. Helaoui, and F.M. Ghannouchi, “Transmitter
architecture for CA: Carrier Aggregation in LTE-Advanced systems,”
IEEE Microw. Mag., vol. 14, no. 5, pp. 78–86, July 2013.
[4] C. Park, L. Sundström, A. Wallen, and A. Khayrallah, “Carrier
aggregation for LTE-advanced: Design challenges of terminals,” IEEE
Commun. Mag., vol. 51, no. 12, pp. 76–84, Dec. 2013.
[5] R4-121205, “Way forward for non-contiguous intraband transmitter
aspects.”, 3GPP Tech. Rep., Feb. 2013.
[6] R4-124353, “Non-contiguous intraband unwanted emission.”, 3GPP
Tech. Rep., Feb. 2013.
[7] H. Gandhi, D. Greenstreet, and J. Quintal, “Digital Radio Front-End
strategies provide game-changing benefits for small cell base stations,”
Tech. Rep., Texas Instruments, May 2013.
[8] L. Guan and A. Zhu, “Green communications: Digital predistortion for
wideband RF power amplifiers,” IEEE Microw. Mag., vol. 15, no. 7, pp.
84–89, Dec. 2014.
[9] R4-123797, “UE reference sensitivity requirements with two UL
carriers.,” 3GPP Tech. Rep., Feb. 2013.
[10] International Telecommunication Union Radio Communication Sector,
Recommendation ITU-R SM.329-12 Unwanted emissions in the spurious
domain..
[11] Chao Yu, Wenhui Cao, Yan Guo, and A. Zhu, “Digital Compensation
for Transmitter Leakage in Non-Contiguous Carrier Aggregation Applications With FPGA Implementation,” IEEE Trans. Microw. Theory
Techn., vol. 63, no. 12, pp. 4306–4318, Dec 2015.
[12] A. Kiayani, M. Abdelaziz, L. Anttila, V. Lehtinen, and M. Valkama,
“Digital Mitigation of Transmitter-Induced Receiver Desensitization in
Carrier Aggregation FDD Transceivers,” IEEE Trans. Microw. Theory
Techn., vol. 63, no. 11, pp. 3608–3623, Nov 2015.
[13] S.A. Bassam, F.M. Ghannouchi, and M. Helaoui, “2-D Digital Predistortion (2-D-DPD) architecture for concurrent dual-band transmitters,”
IEEE Trans. Microw. Theory Techn., vol. 59, no. 10, pp. 2547–2553,
Oct. 2011.
[14] P. Roblin, C. Quindroit, N. Naraharisetti, S. Gheitanchi, and M. Fitton,
“Concurrent linearization,” IEEE Microw. Mag., vol. 14, no. 7, pp. 75–
91, Nov. 2013.
[15] Y. Liu, J. Yan, and P. Asbeck, “Concurrent dual-band digital predistortion with a single feedback loop,” IEEE Trans. Microw. Theory Techn.,
vol. 63, no. 5, pp. 1556–1568, May 2015.
[16] P. Roblin, S. K. Myoung, D. Chaillot, Y. G. Kim, A. Fathimulla,
J. Strahler, and S. Bibyk, “Frequency-selective predistortion linearization
of RF power amplifiers,” IEEE Trans. Microw. Theory Techn., vol. 56,
no. 1, pp. 65–76, Jan. 2008.
[17] J. Kim, P. Roblin, D. Chaillot, and Z. Xie, “A generalized architecture
for the frequency-selective digital predistortion linearization technique,”
IEEE Trans. Microw. Theory Techn., vol. 61, no. 1, pp. 596–605, Jan.
2013.
[18] S.A. Bassam, M. Helaoui, and F.M. Ghannouchi, “Channel-selective
multi-cell digital predistorter for multi-carrier transmitters,” IEEE Trans.
Microw. Theory Techn., vol. 60, no. 8, pp. 2344–2352, Aug. 2012.
[19] M. Abdelaziz, L. Anttila, A. Mohammadi, F. Ghannouchi, and
M. Valkama, “Reduced-complexity power amplifier linearization for
carrier aggregation mobile transceivers,” in IEEE International Conference on Acoustics, Speech, and Signal Processing, pp. 3908–3912, May
2014.
[20] M. Abdelaziz, L. Anttila, J.R. Cavallaro, S. Bhattacharyya, A. Mohammadi, F.M. Ghannouchi, M. Juntti, and M. Valkama, “Low-complexity
digital predistortion for reducing power amplifier spurious emissions
in spectrally-agile flexible radio,” in 9th International Conference on
Cognitive Radio Oriented Wireless Networks, pp. 323–328, June 2014.
17
[21] M. Abdelaziz, C. Tarver, K. Li, L. Anttila, R. Martinez, M. Valkama,
J. R. Cavallaro, “Sub-band digital predistortion for noncontiguous
transmissions: Algorithm development and real-time prototype implementation,” in 49th Asilomar Conference on Signals, Systems and
Computers, pp. 1180–1186, Nov. 2015.
[22] M. Abdelaziz, Z. Fu, L. Anttila, A. M. Wyglinski and M. Valkama,
“Digital predistortion for mitigating spurious emissions in spectrally
agile radios,” IEEE Commun. Mag., vol. 54, no. 3, pp. 60–69, March
2016.
[23] Z. Fu, L. Anttila, M. Abdelaziz, M. Valkama, and A. M. Wyglinski,
“Frequency-Selective Digital Predistortion for Unwanted Emission Reduction,” IEEE Trans. Commun., vol. 63, no. 1, pp. 254–267, Jan. 2015.
[24] A.S. Tehrani, H. Cao, S. Afsardoost, T. Eriksson, M. Isaksson, and
C. Fager, “A comparative analysis of the complexity/accuracy tradeoff
in power amplifier behavioral models,” IEEE Trans. Microw. Theory
Techn., vol. 58, no. 6, pp. 1510–1520, June 2010.
[25] H. Qian, S. Yao, H. Huang, and W. Feng, “A low-complexity digital
predistortion algorithm for power amplier linearization,” IEEE Trans.
Broadcast., vol. 60, no. 4, pp. 670–678, Dec. 2014.
[26] W. Hoffmann, “Iterative algorithms for Gram Schmidt orthogonalization,” Computing, vol. 41, no. 4, pp. 335–348, 1989.
[27] M. Saint-Laurent et al., “A 28 nm DSP powered by an on-chip LDO
for high-performance and energy-efficient mobile applications,” IEEE
J. Solid-State Circuits, vol. 50, no. 1, pp. 81–91, Jan 2015.
[28] LTE Evolved Universal Terrestrial Radio Access (E-UTRA) RF system
scenarios, 3GPP TR 36.942 V10.2.0 (Release 10), May 2011.
Mahmoud Abdelaziz (S’13) received the B.Sc.
degree (with honors) and M.Sc. degree in Electronics
and Electrical Communications Engineering from
Cairo University, Egypt, in 2006 and 2011. He is
currently pursuing the Doctoral degree at Tampere
University of Technology, Finland where he works
as a researcher with the Department of Electronics
and Communications. From 2007 to 2012 he has
been working as a communication systems and
signal processing engineer at Newport Media, as
well as other companies in the wireless industry.
His research interests include statistical and adaptive signal processing in
flexible radio transceivers, wideband digital pre-distortion, and cognitive radio
systems.
Lauri Anttila (S’06, M’11) received the M.Sc.
degree and D.Sc. (Tech) degree (with honors) in
electrical engineering from Tampere University of
Technology (TUT), Tampere, Finland, in 2004 and
2011. Currently, he is a Senior Research Fellow
at the Department of Electronics and Communications Engineering at TUT. His research interests
are in signal processing for wireless communications, in particular radio implementation challenges
in 5G cellular radio and full-duplex radio, flexible
duplexing techniques, and transmitter and receiver
linearization. He has co-authored over 60 peer reviewed articles in these areas,
as well as two book chapters.
Chance Tarver received a B.S. degree in electrical
engineering from Louisiana Tech University, Ruston,
LA, in 2014, and his M.S. degree in electrical and
computer engineering in 2016 from Rice University,
Houston, TX. He is currently a Ph.D. student in the
Department of Electrical and Computer Engineering
at Rice University. His research interests include
software defined radio and signal processing for
wireless communications.
Kaipeng Li (S’14) received his B.S. degree in
physics from Nanjing University, Nanjing, China,
in 2013, and his M.S. degree in electrical and
computer engineering from Rice University in 2015.
He is currently a Ph.D. candidate in the Department
of Electrical and Computer Engineering at Rice
University, Houston, Texas. His research interests
include digital signal processing, parallel computing
on GPGPU and multicore CPU, software-defined
radios and massive MIMO systems.
Joseph R. Cavallaro (S’78, M’82, SM’05, F’15)
received the B.S. degree from the University of
Pennsylvania, Philadelphia, PA, in 1981, the M.S.
degree from Princeton University, Princeton, NJ, in
1982, and the Ph.D. degree from Cornell University,
Ithaca, NY, in 1988, all in electrical engineering.
From 1981 to 1983, he was with AT&T Bell
Laboratories, Holmdel, NJ. In 1988, he joined the
faculty of Rice University, Houston, TX, where he
is currently a Professor of electrical and computer
engineering. His research interests include computer
arithmetic, and DSP, GPU, FPGA, and VLSI architectures for applications
in wireless communications. During the 1996-1997 academic year, he served
at the National Science Foundation as Director of the Prototyping Tools and
Methodology Program. He was a Nokia Foundation Fellow and a Visiting
Professor at the University of Oulu, Finland in 2005 and continues his
affiliation there as an Adjunct Professor. He is currently the Director of the
Center for Multimedia Communication at Rice University. He is a Fellow of
the IEEE and a Member of the IEEE SPS TC on Design and Implementation
of Signal Processing Systems and the Chair-Elect of the IEEE CAS TC on
Circuits and Systems for Communications. He is currently an Associate Editor
of the IEEE Transactions on Signal Processing, the IEEE Signal Processing
Letters, and the Journal of Signal Processing Systems. He was Co-chair of
the 2004 Signal Processing for Communications Symposium at the IEEE
Global Communications Conference and General/Program Co-chair of the
2003, 2004, and 2011 IEEE International Conference on Application-Specific
Systems, Architectures and Processors (ASAP), General/Program Co-chair
for the 2012, 2014 ACM/IEEE GLSVLSI, Finance Chair for the 2013 IEEE
GlobalSIP conference, and TPC Co-Chair of the 2016 IEEE SiPS workshop.
He was a member of the IEEE CAS Society Board of Governors during 2014.
Mikko Valkama (S’00, M’01, SM’15) was born
in Pirkkala, Finland, on November 27, 1975. He
received the M.Sc. and Ph.D. Degrees (both with
honors) in electrical engineering (EE) from Tampere
University of Technology (TUT), Finland, in 2000
and 2001, respectively. In 2002, he received the Best
Ph.D. Thesis -award by the Finnish Academy of
Science and Letters for his dissertation entitled ”Advanced I/Q signal processing for wideband receivers:
Models and algorithms”. In 2003, he was working
as a visiting researcher with the Communications
Systems and Signal Processing Institute at SDSU, San Diego, CA. Currently,
he is a Full Professor and Department Vice-Head at the Department of
Electronics and Communications Engineering at TUT, Finland. His general
research interests include communications signal processing, estimation and
detection techniques, signal processing algorithms for software defined flexible radios, cognitive radio, full-duplex radio, radio localization, 5G mobile
cellular radio, digital transmission techniques such as different variants of
multicarrier modulation methods and OFDM, and radio resource management
for ad-hoc and mobile networks.
| 3cs.SY
|
Characteristic Matrices and Trellis Reduction for
Tail-Biting Convolutional Codes
Masato Tajima, Senior Member, IEEE
arXiv:1705.03982v2 [cs.IT] 23 May 2017
Abstract
Basic properties of a characteristic matrix for a tail-biting convolutional code are investigated. A tail-biting convolutional code
can be regarded as a linear block code. Since the corresponding scalar generator matrix Gtb has a kind of cyclic structure, an
associated characteristic matrix also has a cyclic structure, from which basic properties of a characteristic matrix are obtained.
Next, using the derived results, we discuss the possibility of trellis reduction for a given tail-biting convolutional code. There
are cases where we can find a scalar generator matrix G′ equivalent to Gtb based on a characteristic matrix. In this case, if
the polynomial generator matrix corresponding to G′ has been reduced, or can be reduced by using appropriate transformations,
then trellis reduction for the original tail-biting convolutional code is realized. In many cases, the polynomial generator matrix
corresponding to G′ has a monomial factor in some column and is reduced by dividing the column by the factor. Note that
this transformation corresponds to cyclically shifting the associated code subsequence (a tail-biting path is regarded as a code
sequence) to the left. Thus if we allow partial cyclic shifts of a tail-biting path, then trellis reduction is accomplished.
Index Terms
Tail-biting convolutional codes, tail-biting trellis, characteristic matrix, cyclic shift, trellis reduction.
M. Tajima was with the graduate School of Science and Engineering, University of Toyama, 3190 Gofuku, Toyama 930-8555, Japan (e-mail:
[email protected]).
Manuscript received April 19, 2005; revised August 26, 2015. This paper was presented in part at the IEICE Technical Committee IT Conference in March
2017.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
1
Characteristic Matrices and Trellis Reduction for
Tail-Biting Convolutional Codes
I. I NTRODUCTION
From the 1980s to 1990s, trellis representations of linear block codes were studied with a great interest [2], [7], [8], [14],
[17], [18], [19]. Subsequently, tail-biting trellises of linear block codes have received much attention. Given a linear block
code, there exists a unique minimal conventional trellis. This trellis simultaneously minimizes all measures of trellis complexity.
However, tail-biting trellises do not have such a property. That is, minimality of tail-biting trellises depends on the measure
being used [13]. In general, the complexity of a tail-biting trellis may be much lower than that of the minimal conventional
trellis. There have been many contributions to the subject, including [3], [4], [9], [10], [11], [13], [15], [20], [22], [32]. The
works [3], [13] had a strong influence on the subsequent studies. A remarkable progress has been made by Koetter and Vardy
in their paper [13]. They showed that for a k-dimensional linear block code of length n with full support, there exists a list of n
characteristic generators (i.e., a characteristic matrix [13]) from which all minimal tail-biting minimal trellises can be obtained.
A different method of producing tail-biting trellises was proposed by Nori and Shankar [20]. They used the Bahl-CockeJelinek-Raviv (BCJR) construction [2]. These works were further investigated by Gluesing-Luerssen and Weaver [9], [10]. In
particular, noting that a characteristic matrix for a given code is not necessarily unique, they have refined and generalized the
previous works. More recent works [4], [11] provide further research on the subject.
On the other hand, tail-biting convolutional codes were proposed by Ma and Wolf in 1986 [16] (tail-biting representations
of block codes were introduced by Solomon and van Tilborg [24]). Tail-biting (abbreviated TB) is a technique by which a
convolutional code can be used to construct a block code without any loss of rate. In connection with the subject, there have
been also many works, including [1], [16], [23], [25], [28], [33]. Since a TB convolutional code is identified with a linear
block code, the results on TB trellises for linear block codes can be used. In particular, we can think of a characteristic matrix
of a given TB convolutional code. In this paper, we first investigate a characteristic matrix for a TB convolutional code. And
then, based on the derived results, we discuss the possibility of trellis reduction for a given TB convolutional code. An outline
of the rest of the paper is as follows:
In Section II, we review the basic notions needed for this paper.
In Section III, we investigate the basic properties of a characteristic matrix for a TB convolutional code. When a TB
convolutional code with generator matrix G(D) is regarded as a linear block code C, a (scalar) generator matrix (denoted by
Gtb ) for C is constructed using the coefficients which appear in the polynomial expansion of G(D). We see that Gtb has a
kind of cyclic structure. Then it is shown that the (characteristic) span list associated with a characteristic matrix for C consists
of some basic spans and their right cyclic shifts, from which basic properties of a characteristic matrix are derived.
In Section IV, we deal with transformations of G(D) and discuss the relationship between these transformations and the
corresponding scalar generator matrices Gtb ’s. We see that dividing a column of G(D) by a monomial factor corresponds to
cyclically shifting a column subsequence of Gtb to the left, whereas multiplying a column of G(D) by a monomial corresponds
to cyclically shifting a column subsequence of Gtb to the right. These properties are essentially used for trellis reduction to
be discussed in Section V.
In Section V, we discuss the possibility of trellis reduction for a given TB convolutional code (we identify the code with
an (n, k) block code C). As is stated above, we can think of a characteristic matrix for C. Consider the case where some k
characteristic generators, which consist of some basic generators and their right cyclic shifts, can generate the same code C.
We see that these characteristic generators form a (scalar) generator matrix associated with a (polynomial) generator matrix of
another convolutional code. In this case, if the constraint length of the obtained generator matrix is smaller than that of the
original one, then trellis reduction is realized. Even if this kind of reduction is not possible, there are cases where a newly
obtained generator matrix contains a monomial factor in some column. Then there is a possibility that the generator matrix
is reduced by sweeping the monomial factor out of the column. Note that this operation corresponds to cyclically shifting
the corresponding code subsequence to the left. In this way, trellis reduction can be accomplished. We also present a trellis
reduction method for high rate codes which uses a reciprocal dual encoder. We remark that the (trellis) section length is an
important parameter and the proposed method is restricted to TB convolutional codes with short to moderate section length.
We give an upper bound for the section length by evaluating the span lengths of characteristic generators.
Finally, conclusions are provided in Section VI.
II. P RELIMINARIES
We begin with the basic notions needed in this paper, where the underlying field is assumed to be F = GF(2). Let C be an
△
(n, k) linear block code, where the set of indices for a codeword in C is denoted by I = {0, 1, · · · , n − 1}. Then a codeword
x ∈ C is expressed as x = (x0 , x1 , · · · , xn−1 ). I is also regarded as the time axis for TB trellises for C. Since TB trellises
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
2
for C are considered in this paper, it is convenient to identify I with Zn , the ring of integers modulo n. Hence, when dealing
with TB trellises, all index arithmetic will be implicitly performed modulo n [13].
The notion of span is fundamental in trellis theory. Given a codeword x ∈ C, a span of x, denoted by [x], is a semiopen
interval (a, b] ∈ I such that the corresponding closed interval [a, b] contains all the nonzero positions of x [13]. Due to the
cyclic structure of the time axis I, we adopt the following interpretation of intervals [9], [10], [13]. For a, b ∈ I, we define
△
{a, a + 1, · · · , b},
if a ≤ b
[a, b] =
(1)
{a, a + 1, · · · , n − 1, 0, 1, · · · , b},
if a > b
△
and (a, b] = [a, b]\{a}. We call the intervals (a, b] and [a, b] conventional if a ≤ b and circular otherwise.
In connection with the construction of minimal TB trellises for C, Koetter and Vardy [13] introduced the notion of
characteristic generator for C. Denote by σj (·) a cyclic shift to the left by j positions [13]. Similarly, denote by ρj (·) a
△
cyclic shift to the right by j positions. Let Xj∗ be a basis in minimal-span form [17] for the code Cj = σj (C). A characteristic
generator for C is a pair consisting of a codeword x = (x0 , x1 , · · · , xn−1 ) ∈ C and a span [x] = (a, b] such that xa , xb are
nonzero. The set of all the characteristic generators for C is given by
△
∗
).
X = X0∗ ∪ ρ1 (X1∗ ) ∪ · · · ∪ ρn−1 (Xn−1
(2)
Here we have an understanding that if x∗ ∈ Xj∗ , then [ρj (x∗ )] = (a + j, b + j], where [x∗ ] = (a, b].
Assume that C has a full support. Then a characteristic matrix for C is the n × n matrix having the elements of X as its
rows. The above definition implies that when we refer to a characteristic matrix, the associated spans are taken into account.
Here note that a basis in minimal-span form is not necessarily unique. Hence, X may not be uniquely determined. On the
other hand, the set of spans (denoted by T ) accompanied by X is, up to ordering, uniquely determined by the code C [9], [10],
[13]. T is called the characteristic span list of C (an element∈ T is called a characteristic span of C) [9], [10]. In order to
clarify this fact, Gluesing-Luerssen and Weaver introduced the notion of characteristic pair (X, T ) of C [9, Definition III.8],
where X is a generating set of C and T represents the associated spans. In this paper, we basically follow the definition of
Gluesing-Luerssen and Weaver, but in order to emphasize the fact that a characteristic matrix inherently assumes the associated
spans, we leave the term characteristic matrix for the definition. Thus we define as follows (cf. [9, Definition III.8]).
Definition 2.1: Let C be an (n, k) linear block code with support I. A characteristic matrix for C with (characteristic) span
list T is defined to be a pair (X, T ), where
x1
x2
(3)
X =
···
xn
T = {(al , bl ] : l = 1, 2, · · · , n}
(4)
have the properties:
1) {x1 , · · · , xn } generates C.
2) (al , bl ] is a span of xl , l = 1, · · · , n.
3) a1 , · · · , an are distinct and b1 , · · · , bn are distinct.
4) For all j ∈ I, there exist exactly n − k row indices, l1 , · · · , ln−k , such that j ∈ (ali , bli ] for i = 1, · · · , n − k.
Remark: Property 3) is derived from [13, Lemma 5.7] and the related remarks. Also, Property 4) is derived from the proof
of [13, Theorem 5.10].
In the following, when there is no danger of confusion, we shall use the terms characteristic matrix X and characteristic
matrix X with span list T interchangeably.
III. C HARACTERISTIC M ATRICES
FOR A
TAIL -B ITING C ONVOLUTIONAL C ODE
Let G(D) be a polynomial generator matrix of size k0 × n0 . Denote by H(D) a corresponding polynomial check matrix.
Both G(D) and H(D) are assumed to be canonical [18]. Consider a standard trellis of N sections for a convolutional code
defined by G(D). Here max(L, M ) + 1 ≤ N is assumed, where L and M are the memory lengths of G(D) and H(D),
respectively. The TB condition is a restriction that the encoder starts and ends in the same state. That is, only those paths in
the trellis that start and end in the same state are admissible. We call such paths TB paths. Let C be the set of all TB paths.
In the following, we call C a TB convolutional code of section length N defined by G(D) (cf. Fig.1 in Section V). When
there is no danger of confusion, we will omit the phrase of section length N . C can be regarded as a linear block code B tb
of length n = n0 N . To simplify the notations, B tb is identified with C and is denoted simply by C. Let
G(D) = G0 + G1 D + · · · + GL DL
(5)
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
3
be the polynomial expansion of G(D), where Gi (0 ≤ i ≤ L) are k0 × n0 matrices. Then the scalar generator matrix for
C (= B tb ) is given by
G0 G1 ... GL−1 GL
G0 ... ... GL−1 GL
...
...
...
... ...
△
G
G
...
...
G
0
1
L
tb
(6)
GN =
G
G
G
...
G
0
1
L−1
L
GL−1 GL
G0 ... ...
... ... ...
... G1
G1 G2 ... GL
G0
with size k × n = k0 N × n0 N [12]. Hence, we can say that a TB convolutional code C is generated by Gtb
N . In the following,
we call Gtb
the
tail-biting
generator
matrix
(abbreviated
TBGM)
associated
with
a
TB
convolutional
code
C
defined by G(D),
N
or simply the TBGM associated with G(D).
A. Computation of Characteristic Matrices
Koetter and Vardy [13] have given an algorithm which can compute a characteristic matrix for a linear block code. Consider
tb
tb
tb
a TB convolutional code C generated by Gtb
N . Note that σn0 (GN ) is equivalent to GN . That is, GN has a periodic structure
of period n0 . Using this property, a characteristic matrix X for C can be computed efficiently.
△
∗
Let Cj = σj (C). Cj is the code generated by σj (Gtb
N ). Let Xj be a basis in minimal-span form for the code Cj . Then a
characteristic matrix X for C is defined as follows [13]:
X
△
=
X0∗ ∪ ρ1 (X1∗ ) ∪ · · · ∪ ρn0 −1 (Xn∗0 −1 )
∗
)
∪ρn0 (Xn∗0 ) ∪ ρn0 +1 (Xn∗0 +1 ) ∪ · · · ∪ ρ2n0 −1 (X2n
0 −1
···
∗
∗
∗
∪ρ(N −1)n0 (X(N
−1)n0 ) ∪ ρ(N −1)n0 +1 (X(N −1)n0 +1 ) ∪ · · · ∪ ρN n0 −1 (XN n0 −1 ).
(7)
tb
Since σn0 (Gtb
N ) is equivalent to GN , we have
ρn0 (Xn∗0 ) =
ρn0 +1 (Xn∗0 +1 ) =
∗
ρ2n0 −1 (X2n
) =
0 −1
where
ρn0 (X0∗ )
ρn0 (ρ1 (X1∗ )) = ρn0 (X̃1∗ )
···
ρn0 (ρn0 −1 (Xn∗0 −1 )) = ρn0 (X̃n∗0 −1 ),
△
X̃1∗ = ρ1 (X1∗ )
∗△
X̃2 = ρ2 (X2∗ )
···
∗
△
X̃n0 −1 = ρn0 −1 (Xn∗0 −1 ).
Similarly, we have
∗
)
ρ2n0 (X2n
0
=
ρ2n0 (X0∗ )
∗
ρ2n0 +1 (X2n
)
0 +1
=
ρ2n0 (ρ1 (X1∗ )) = ρ2n0 (X̃1∗ )
∗
ρ3n0 −1 (X3n
)
0 −1
···
= ρ2n0 (ρn0 −1 (Xn∗0 −1 )) = ρ2n0 (X̃n∗0 −1 ).
In general, for i = 1, · · · , N − 1, we have
∗
)
ρin0 (Xin
0
∗
ρin0 +1 (Xin
)
0 +1
∗
ρ(i+1)n0 −1 (X(i+1)n
)
0 −1
=
ρin0 (X0∗ )
= ρin0 (X̃1∗ )
···
=
ρin0 (X̃n∗0 −1 ).
(8)
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
4
Hence,
X
= X0∗ ∪ ρ1 (X1∗ ) ∪ · · · ∪ ρn0 −1 (Xn∗0 −1 )
∪ρn0 (X0∗ ) ∪ ρn0 (X̃1∗ ) ∪ · · · ∪ ρn0 (X̃n∗0 −1 )
···
∪ρ(N −1)n0 (X0∗ ) ∪ ρ(N −1)n0 (X̃1∗ ) ∪ · · · ∪ ρ(N −1)n0 (X̃n∗0 −1 )
is obtained. Thus we have shown the following.
Proposition 3.1: A characteristic matrix X for a TB convolutional code C generated by Gtb
N is given by
X
=
X0∗ ∪ ρ1 (X1∗ ) ∪ · · · ∪ ρn0 −1 (Xn∗0 −1 )
∪ρn0 (X0∗ ) ∪ ρn0 (X̃1∗ ) ∪ · · · ∪ ρn0 (X̃n∗0 −1 )
···
∪ρ(N −1)n0 (X0∗ ) ∪ ρ(N −1)n0 (X̃1∗ ) ∪ · · · ∪ ρ(N −1)n0 (X̃n∗0 −1 ).
(9)
Corollary 3.1: If the relation
X0∗ ∪ ρ1 (X1∗ ) ∪ · · · ∪ ρn0 −1 (Xn∗0 −1 ) ⊆ X0∗ ∪ ρn0 (X0∗ )
(10)
holds, then a characteristic matrix X is given by
X = X0∗ ∪ ρn0 (X0∗ ) ∪ ρ2n0 (X0∗ ) ∪ · · · ∪ ρ(N −1)n0 (X0∗ ).
(11)
Proof: From the assumption, we have
ρn0 (X0∗ ) ∪ ρn0 (X̃1∗ ) ∪ · · · ∪ ρn0 (X̃n∗0 −1 ) ⊆ ρn0 (X0∗ ) ∪ ρ2n0 (X0∗ ).
Similarly, we have
⊆ ρ2n0 (X0∗ ) ∪ ρ3n0 (X0∗ )
···
ρ2n0 (X0∗ ) ∪ ρ2n0 (X̃1∗ ) ∪ · · · ∪ ρ2n0 (X̃n∗0 −1 )
ρ(N −1)n0 (X0∗ ) ∪ ρ(N −1)n0 (X̃1∗ ) ∪ · · · ∪ ρ(N −1)n0 (X̃n∗0 −1 )
⊆
=
ρ(N −1)n0 (X0∗ ) ∪ ρN n0 (X0∗ )
ρ(N −1)n0 (X0∗ ) ∪ X0∗ .
Then, from Proposition 3.1, it follows that
X = X0∗ ∪ ρn0 (X0∗ ) ∪ ρ2n0 (X0∗ ) ∪ · · · ∪ ρ(N −1)n0 (X0∗ ).
We remark that in many practical applications, a characteristic matrix for a TB convolutional code is obtained based on the
above corollary.
Example 1: Consider the TB convolutional code of section length N = 3 defined by
G(D)
=
=
△
=
(1 + D, D, 1 + D)
(1, 0, 1) + (1, 1, 1)D
(12)
G0 + G1 D.
(13)
The associated TBGM is given by
Gtb
3
=
=
In this case, we have
G0 G1
0 G0
G1 0
1 0 1
0 0 0
1 1 1
1 0
X0∗ = 0 0
0 1
1 0
X̃1∗ = 0 0
0 1
0
G1
G0
1
1
0
1 1
0 1
0 0
0
1
1
1
0
0
1
1
0
1 1
0 1
1 0
0
1
0
0
1
1
1
0
0
1
1
0
1 1
0 1
1 0
0
1
0
0
1
1
0 0
1 1 .
0 1
0
(0, 5]
1 (3, 8]
0
(1, 7]
0
(2, 0]
1 (3, 8]
0
(1, 7]
(14)
(15)
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
1
X̃2∗ = 0
0
0
0
1
5
1 1
0 1
0 0
1
0
1
1
1
0
By applying ρ3 and ρ6 to these matrices, a characteristic matrix X
1 0 1 1 1 1
0 1 0 0 1 0
1 0 1 1 1 1
0 0 0 1 0 1
X=
0 1 0 0 1 0
0 0 0 1 0 1
1 1 1 0 0 0
0 1 0 0 1 0
1 1 1 0 0 0
0 0
1 1
0 1
0
(2, 0]
1 (3, 8]
0
(4, 1].
is obtained as follows:
(0, 5]
0 0 0
0 1 0
(1, 7]
0 0 0
(2, 0]
1 1 1
(3, 8]
0 1 0
(4, 1]
1 1 1
(5, 3]
1 0 1
(6, 2]
0 1 0 (7, 4]
(8, 6].
1 0 1
(16)
(17)
Note that the spans (0, 5], (1, 7], (2, 0], (3, 8], (4, 1] are connected with X0∗ ∪ ρ1 (X1∗ ) ∪ ρ2 (X2∗ ), whereas the spans
(0, 5], (1, 7], (3, 8], (4, 1], (6, 2] are connected with X0∗ ∪ ρ3 (X0∗ ). Hence,
X0∗ ∪ ρ1 (X1∗ ) ∪ ρ2 (X2∗ ) * X0∗ ∪ ρ3 (X0∗ ).
We see that a characteristic matrix X cannot be obtained simply by applying ρ3 and ρ6 to X0∗ .
B. Structure of the Characteristic Span List
Let (X, T ), where
x1
x2
X =
···
xn
(a1 , b1 ]
(a2 , b2 ]
,
T =
···
(an , bn ]
be a characteristic matrix for C with span list T , then (σ1 (X), σ1 (T )) is a characteristic matrix for σ1 (C) with span list
σ1 (T ) [9, Remark III.9 (b)], where
σ1 (x1 )
△ σ1 (x2 )
(18)
σ1 (X) =
···
σ1 (xn )
(a1 − 1, b1 − 1]
△ (a2 − 1, b2 − 1]
.
(19)
σ1 (T ) =
···
(an − 1, bn − 1]
Using repeatedly this relation, we see that (σj (X), σj (T )) is a characteristic matrix for σj (C) with span list σj (T ). Consider
tb
tb
a TB convolutional code C generated by Gtb
N and set j to n0 . Since σn0 (GN ) is equivalent to GN , σn0 (C) = C holds. Thus
we have the following.
Lemma 3.1: Let C be a TB convolutional code generated by Gtb
N . If (X, T ) is a characteristic matrix for C with span list
T , then (σn0 (X), σn0 (T )) is also a characteristic matrix for C with span list σn0 (T ). Let
T = {(al , bl ] : l = 1, 2, · · · , n}.
(20)
σn0 (T ) = {(al − n0 , bl − n0 ] : l = 1, 2, · · · , n}.
(21)
Then σn0 (T ) is given by
Since the characteristic span list is uniquely determined, T and σn0 (T ) coincide up to ordering.
Proposition 3.2: The characteristic span list T of a TB convolutional code C generated by Gtb
N consists of the set of basic
spans
(0, b0 ]
△
(1, b1 ]
(22)
T0 =
···
(n0 − 1, bn0 −1 ]
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
6
and ρin0 (T0 ) (i = 1, 2, · · · , N − 1).
Proof: Suppose that the spans in T are sorted such that
T = {(l, bl ] : l = 0, 1, · · · , n − 1}.
Then we have
σn0 (T ) = {(l − n0 , bl − n0 ] : l = 0, 1, · · · , n − 1}.
Here take notice of the following set of spans in T :
(n0 , bn0 ]
(n0 + 1, bn0 +1 ]
···
(2n0 − 1, b2n0 −1 ].
In σn0 (T ), it is transformed to
(0, bn0 − n0 ]
(1, bn0 +1 − n0 ]
···
(n0 − 1, b2n0 −1 − n0 ].
Since T and σn0 (T ) coincide up to ordering,
(0, bn0 − n0 ] = (0, b0 ]
(1, bn0 +1 − n0 ] = (1, b1 ]
·
··
(n0 − 1, b2n0 −1 − n0 ] = (n0 − 1, bn0 −1 ]
holds. Hence, we have
b n = b 0 + n0
0
bn0 +1 = b1 + n0
···
b2n0 −1 = bn0 −1 + n0 .
Similarly, the set of spans
(2n0 , b2n0 ]
(2n0 + 1, b2n0 +1 ]
···
(3n0 − 1, b3n0 −1 ]
is transformed to
Then for the same reason,
holds. Hence, we have
(n0 , b2n0 − n0 ] = (n0 , bn0 ]
(n0 + 1, b2n0 +1 − n0 ] = (n0 + 1, bn0 +1 ]
···
(2n0 − 1, b3n0 −1 − n0 ] = (2n0 − 1, b2n0 −1 ]
Continuing the same argument, we have
for i = 1, · · · , N − 1.
(n0 , b2n0 − n0 ]
(n0 + 1, b2n0 +1 − n0 ]
·
··
(2n0 − 1, b3n0 −1 − n0 ].
b2n0 = bn0 + n0 = b0 + 2n0
b2n0 +1 = bn0 +1 + n0 = b1 + 2n0
···
b3n0 −1 = b2n0 −1 + n0 = bn0 −1 + 2n0 .
bin0 = b0 + in0
bin0 +1 = b1 + in0
···
b(i+1)n0 −1 = bn0 −1 + in0
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
7
Example 2: Consider the TB convolutional code of section length N = 3 defined by the rate R = 2/3 encoder
1+D D 1+D
G(D) =
D
1
1
1 0 1
1 1 1
=
+
D
0 1 1
1 0 0
△
=
G0 + G1 D.
Using the associated TBGM, i.e.,
G0
0
Gtb
3 =
G1
a charactreristic matrix X is computed as follows:
1 0
0 1
0 0
0 0
X=
0 0
0 0
1 1
1 0
0 0
0
1
1
0
0
0
0
0
1
1
1
0
1
0
0
0
0
0
1
0
0
0
1
0
0
0
0
(23)
(24)
0
G1 ,
G0
G1
G0
0
0
0
1
0
1
1
0
0
0
0
0
0
1
1
0
1
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
1
0
1
1
We see that the characteristic span list T consists of the set of basic spans
(0, 4]
T0 =
(1, 3]
(2, 5]
(0,
(1,
(2,
(3,
(4,
(5,
(6,
(7,
(8,
4]
3]
5]
7]
6]
8]
1]
0]
2].
(25)
(26)
and its right cyclic shifts by 3 and 6 positions.
C. Counting Characteristic Matrices
Recall the definition of a characteristic matrix X for a given code C, i.e.,
△
∗
X = X0∗ ∪ ρ1 (X1∗ ) ∪ · · · ∪ ρn−1 (Xn−1
),
where Xj∗ is a basis in minimal-span form for the code Cj = σj (C). Note that Xj∗ is not necessarily unique. Hence, X is not
uniquely determined [9]. With respect to this subject, Weaver [32] discussed the relationship between the characteristic span
list of C and the number of characteristic matrices for C.
Let T = {(al , bl ] : l = 1, 2, · · · , n} be the characteristic span list of C. Define the set Θl as follows [32]:
△
Θl = {r : (ar , br ] ( (al , bl ]}.
(27)
|Θl | represents the number of spans (in T ) included in a specified span (al , bl ]. Weaver [32] proved the following.
Lemma 3.2 (Weaver [32]): Let (al , bl ] be a characteristic span of C. Then there exist 2|Θl | characteristic generators for C
having this span.
This fact is derived from the next observation:
Let (ar , br ] ( (al , bl ] and consider two characteristic generators xr and xl with spans (ar , br ] and (al , bl ], respectively. Then
xl + xr is also a characteristic generator with span (al , bl ].
Consider a TB convolutional code C generated by Gtb
N . We have already shown that the characteristic span list T of C
consists of the set of basic spans
(0, b0 ]
(1, b1 ]
T0 =
···
(n0 − 1, bn0 −1 ]
and ρin0 (T0 ) (i = 1, 2, · · · , N − 1). Hence, it suffices to consider the spans in T0 for the purpose of counting the number of
characteristic matrices. Define Θi (1 ≤ i ≤ n0 ) as follows:
△
Θ1 = {r : (ar , br ] ( (0, b0 ]}
△
Θ2 = {r : (ar , br ] ( (1, b1 ]}
(28)
···
△
Θn0 = {r : (ar , br ] ( (n0 − 1, bn0 −1 ]}.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
Also, let
8
△
θi = |Θi | (i = 1, 2, · · · , n0 )
△
θ = θ1 + θ2 + · · · + θn0 .
(29)
(30)
Then we have the following:
θ
• There exist 2 1 characteristic generators having span (0, b0 ].
θ
• There exist 2 2 characteristic generators having span (1, b1 ].
···
θ
• There exist 2 n0 characteristic generators having span (n0 − 1, bn0 −1 ].
As a result, the degree of freedom related to the spans in T0 is given by
2θ1 × 2θ2 × · · · × 2θn0
= 2θ1 +θ2 +···+θn0
= 2θ .
(31)
Since this degree of freedom is common to other (N − 1) blocks of spans in T , the overall degree of freedom related to T
becomes
(2θ )N = 2θN .
(32)
Thus we have shown the following.
Proposition 3.3: Let C be a TB convolutional code generated by Gtb
N . Let θi (1 ≤ i ≤ n0 ) and θ be as above. Then there
exist 2θN characteristic matrices for C.
Example 2 (Continued): Take notice of the first three rows of the characteristic matrix X. We have
Θ1 = {2}
Θ2 = φ
Θ3 = φ.
Hence, θ = 1 + 0 + 0 = 1 and there exist 21×3 = 8 characteristic matrices.
D. Span Lengths of Characteristic Generators
Let [x] = (a, b] be a span of a codeword x. Then the span length of x is defined by |[a, b]|, i.e., the number of elements in
the closed interval [a, b]. When a span alone is referred to without specifying the accompanied codeword, we use the term the
span length of a span (a, b]. Let T be the characteristic span list of a TB convolutional code C generated by Gtb
N . Suppose
that the spans in T are sorted such that
T = {(l, bl ] : l = 0, 1, · · · , n − 1}.
Then by [13, Theorem 5.10],
|(0, b0 ]| + |(1, b1 ]| + · · · + |(n − 1, bn−1 ]| = n(n − k)
(33)
holds. Due to the structure of T (see Proposition 3.2), the left-hand side of the above equality becomes
N (|(0, b0 ]| + |(1, b1 ]| + · · · + |(n0 − 1, bn0 −1 ]|)
= N ((ℓ1 − 1) + (ℓ2 − 1) + · · · + (ℓn0 − 1))
= N ((ℓ1 + ℓ2 + · · · + ℓn0 ) − n0 )
= N (ℓ − n0 ),
where
△
ℓ1 = |[0, b0 ]|
△
ℓ2 = |[1, b1 ]|
···
△
ℓn0 = |[n0 − 1, bn0 −1 ]|
△
ℓ = ℓ1 + ℓ2 + · · · + ℓn0 .
In the derivation, we also used the relation
ℓi = |[i − 1, bi−1 ]| = |(i − 1, bi−1 ]| + 1 (i = 1, 2, · · · , n0 ).
(34)
(35)
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
9
Replacing n and k by n0 N and k0 N , respectively, the above equality reduces to
N (ℓ − n0 ) =
ℓ − n0 =
ℓ
=
n0 N (n0 N − k0 N )
n0 ((n0 − k0 )N )
n0 ((n0 − k0 )N + 1).
Thus we have shown the following.
Proposition 3.4: Let T be the characteristic span list of a TB convolutional code C generated by Gtb
N . Denote by T0 the set
of basic spans in T . Then the sum ℓ of span lengths of spans in T0 is given by
ℓ = n0 ((n0 − k0 )N + 1).
Example 2 (Continued): We have
(36)
ℓ1 = |[0, 4]| = 5
ℓ2 = |[1, 3]| = 3
ℓ3 = |[2, 5]| = 4
ℓ = ℓ1 + ℓ2 + ℓ3 = 5 + 3 + 4 = 12.
Also, we have
n0 ((n0 − k0 )N + 1) = 3((3 − 2) × 3 + 1) = 12.
IV. T RANSFORMATIONS
OF
G(D) AND
THE
C ORRESPONDING TBGM’ S
In this section, we discuss the relationship between transformations of a generator matrix G(D) and the corresponding
TBGM’s (Gtb
N ’s). We consider the following transformations of G(D):
a) Dividing the jth column by Dp .
b) Multiplying the jth column by Dq .
c) Adding the ith row multiplied by Dq to the jth row.
d) Implicit transformations.
In the next section, we will see that these transformations play an essential role in trellis reduction for TB convolutional
codes.
A. Dividing a Column of G(D) by Dp
Suppose that the jth column of G(D) has a monomial factor Dp (1 ≤ p ≤ L). We
that j = 1 and p = 1. Hence, G(D) has the form
′
g1,1 (D)D g1,2 (D) . . . g1,n0 (D)
′
g2,1
(D)D g2,2 (D) . . . g2,n0 (D)
G(D) =
...
...
...
...
gk′ 0 ,1 (D)D gk0 ,2 (D) . . . gk0 ,n0 (D)
can assume without loss of generality
.
(37)
Let
G(D)
(1)
(1)
(0)
(0)
(0)
g1,1 g1,2
g1,1 g1,2 . . . g1,n0
(1)
(0)
(1)
(0)
(0)
g2,2 . . . g2,n0 g2,1 g2,2
g
= 2,1
+
...
...
... ...
... ...
(1)
(1)
(0)
(0)
(0)
gk0 ,1 gk0 ,2
gk0 ,1 gk0 ,2 . . . gk0 ,n0
(2)
(2)
(2)
g1,1 g1,2 . . . g1,n0
(2)
(2)
(2)
g2,2 . . . g2,n0 2
g
+ 2,1
D + ···
...
... ...
...
(2)
(2)
(2)
gk0 ,1 gk0 ,2 . . . gk0 ,n0
(L)
(L)
(L)
g1,1 g1,2 . . . g1,n0
(L)
(L)
(L)
g2,2 . . . g2,n0 L
g
+ 2,1
D
... ...
...
...
(L)
(L)
(L)
gk0 ,1 gk0 ,2 . . . gk0 ,n0
△
= G0 + G1 D + · · · + GL−1 DL−1 + GL DL
(1)
. . . g1,n0
(1)
. . . g2,n0
...
...
(1)
. . . gk0 ,n0
D
(38)
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
10
be the polynomial expansion of G(D). Comparing the (i, 1) (1 ≤ i ≤ k0 ) entries, we have
(L)
(2)
(1)
(0)
′
g1,1
(D)D = g1,1 + g1,1 D + g1,1 D2 + · · · + g1,1 DL
′
(0)
(1)
(2)
(L)
g2,1 (D)D = g2,1 + g2,1 D + g2,1 D2 + · · · + g2,1 DL
···
′
(0)
(1)
(2)
(L)
gk0 ,1 (D)D = gk0 ,1 + gk0 ,1 D + gk0 ,1 D2 + · · · + gk0 ,1 DL .
By these equations, we have
(0)
g1,1 = 0
(0)
g2,1 = 0
·
··
(0)
gk0 ,1 = 0
(1)
(2)
(L)
′
g1,1
(D) = g1,1 + g1,1 D + · · · + g1,1 DL−1 + 0 · DL
′
(1)
(2)
(L)
g2,1 (D) = g2,1 + g2,1 D + · · · + g2,1 DL−1 + 0 · DL
···
′
(1)
(2)
(L)
gk0 ,1 (D) = gk0 ,1 + gk0 ,1 D + · · · + gk0 ,1 DL−1 + 0 · DL .
(39)
(40)
Dividing the first column of G(D) by D, let the resulting matrix be G′ (D). Then G′ (D) has the polynomial expansion:
′
g1,1 (D) g1,2 (D) . . . g1,n0 (D)
′
g2,1
(D) g2,2 (D) . . . g2,n0 (D)
(41)
G′ (D) =
...
...
...
...
′
gk0 ,1 (D) gk0 ,2 (D) . . . gk0 ,n0 (D)
(2)
(1)
(1)
(1)
(0)
(0)
g1,1 g1,2 . . . g1,n0
g1,1 g1,2 . . . g1,n0
(2)
(1)
(1)
(1)
(0)
(0)
g2,2 . . . g2,n0 g2,1 g2,2 . . . g2,n0
g
= 2,1
D
+
... ...
...
... ...
... ...
...
(2)
(1)
(1)
(1)
(0)
(0)
gk0 ,1 gk0 ,2 . . . gk0 ,n0
gk ,1 gk ,2 . . . gk0 ,n0
0 (3) 0 (2)
(2)
g1,1 g1,2 . . . g1,n0
(3)
(2)
(2)
g
g2,2 . . . g2,n0 2
+ 2,1
D + ···
...
... ...
...
(3)
(2)
(2)
gk0 ,1 gk0 ,2 . . . gk0 ,n0
(L)
(L)
(L)
(L−1)
(L−1)
0 g1,2 . . . g1,n0
g1,1 g1,2
. . . g1,n0
(L)
(L)
(L)
(L−1)
(L−1)
g
g2,2
. . . g2,n0 L−1 0 g2,2 . . . g2,n0 L
+
+ 2,1
D
D
... ... ...
...
...
...
...
...
(L)
(L)
(L)
(L−1)
(L−1)
0 gk0 ,2 . . . gk0 ,n0
gk0 ,1 gk0 ,2
. . . gk0 ,n0
△
=
G′0 + G′1 D + · · · + G′L−1 DL−1 + G′L DL .
Consider the TBGM associated with G′ (D), denoted by G′tb
N , where
′
′
G0 G1 ... G′L−1 G′L
G′0 ... ... G′L−1 G′L
... ...
...
... ...
′
′
′
G0 G1 ... ... GL
′tb
GN = ′
.
G′0 G′1 ... G′L−1
L
G
′
′
G′
G0 ... ...
L−1 GL
′
... ... ...
... G1
′
′
′
G′0
G1 G2 ... GL
(42)
(43)
′tb
Note that both Gtb
N and GN can be regarded as matrices having N blocks of n0 columns. Then in view of the entries of
′
Gi (0 ≤ i ≤ L) and the relation
(0)
g1,1 = 0
(0)
g2,1 = 0
···
(0)
gk0 ,1 = 0,
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
11
tb
G′tb
N is obtained from GN by cyclically shifting the first column of each block to the left by n0 positions. Thus we have the
following.
Proposition 4.1: Regard Gtb
N as a matrix having N blocks of n0 columns. Suppose that the jth column of G(D) has a
monomial factor Dp (1 ≤ p ≤ L). Then dividing the jth column of G(D) by Dp is equivalent to cyclically shifting the jth
column of each block of Gtb
N to the left by pn0 positions.
Let C be a TB convolutional code of section length N defined by G(D). Note that each codeword in C consists of N
blocks of n0 components. Here let us cyclically shift the jth component of each block to the left by pn0 positions. Denote
tb
by C ′ the set of resulting (modified) codewords. We have already shown that G′tb
N is obtained from GN by cyclically shifting
′
′tb
the jth column of each block to the left by pn0 positions. Hence, C is generated by GN . In words, C ′ is represented as a
TB convolutional code defined by G′ (D).
B. Multiplying a Column of G(D) by Dq
Consider multiplication of the jth column of G(D) by Dq , where q + L + 1 ≤ N . In the following, we assume without
loss of generality that j = 1Cq = 1 (L + 2 ≤ N ). Hence, the resulting matrix G′ (D) has the form
g1,1 (D)D g1,2 (D) . . . g1,n0 (D)
g2,1 (D)D g2,2 (D) . . . g2,n0 (D)
(44)
G′ (D) =
...
...
...
...
gk0 ,1 (D)D gk0 ,2 (D) . . . gk0 ,n0 (D)
′
g1,1 (D) g1,2 (D) . . . g1,n0 (D)
′
g2,2 (D) . . . g2,n0 (D)
△ g2,1 (D)
.
(45)
=
...
...
...
...
′
gk0 ,1 (D) gk0 ,2 (D) . . . gk0 ,n0 (D)
Then we have
′
g1,1
(D)
=
g1,1 (D)D
=
(g1,1 + g1,1 D + · · · + g1,1 DL )D
(0)
(0)
(1)
(1)
(L)
(L−1)
= g1,1 D + g1,1 D2 + · · · + g1,1
···
gk′ 0 ,1 (D)
(L)
DL + g1,1 DL+1
=
gk0 ,1 (D)D
=
(gk0 ,1 + gk0 ,1 D + · · · + gk0 ,1 DL )D
=
gk0 ,1 D + gk0 ,1 D2 + · · · + gk0 ,1 DL + gk0 ,1 DL+1 .
(0)
(0)
(1)
(L)
(1)
(L−1)
(L)
Accordingly, the polynomial expansion of G′ (D) becomes
(0)
(1)
(1)
(0)
(0)
g1,1 g1,2 . . . g1,n0
0
g1,2 . . . g1,n0
(0)
(1)
(1)
(0)
(0)
0
g2,2 . . . g2,n0 g2,1 g2,2 . . . g2,n0
G′ (D) =
+
... ... ...
... ...
...
... ...
(0)
(1)
(1)
(0)
(0)
gk0 ,1 gk0 ,2 . . . gk0 ,n0
0 gk0 ,2 . . . gk0 ,n0
(1)
(2)
(2)
g1,1 g1,2 . . . g1,n0
(1)
(2)
(2)
g
g2,2 . . . g2,n0 2
+ 2,1
D + ···
...
... ...
...
(1)
(2)
(2)
gk0 ,1 gk0 ,2 . . . gk0 ,n0
(L)
(L−1)
(L)
(L)
0 ...
g1,1
g1,2 . . . g1,n0
g1,1
(L)
(L−1)
(L)
(L)
0 ...
g2,2 . . . g2,n0 L g2,1
g
+ 2,1
D +
... ... ...
...
... ...
...
(L)
(L−1)
(L)
(L)
gk0 ,1 0 . . .
gk0 ,1
gk0 ,2 . . . gk0 ,n0
△
=
(46)
G′0 + G′1 D + · · · + G′L DL + G′L+1 DL+1 .
(47)
D
0
0
...
0
L+1
D
(48)
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
12
Consider the TBGM associated with G′ (D), denoted by G′tb
N , where
′
′
′
′
′
G0 G1
... GL−1 GL GL+1
′
′
′
G′0
...
... GL−1 GL GL+1
...
...
...
...
...
...
′
′
′
′
G0 G1
...
... GL GL+1
′
′
G′0 G′1
...
... GL .
G′tb
N = GL+1
G′ G′
G′0 G′1 ... G′L−1
L+1
′L
′
′
G
G′0 ... ...
L−1 GL GL+1
...
...
...
...
...
G′1
G′1 G′2
...
G′L G′L+1
G′0
(49)
′tb
′
Note that Gtb
N and GN consist of N blocks of n0 columns as above. In view of the entries of Gi (0 ≤ i ≤ L + 1), we see
′tb
tb
that GN is obtained from GN by cyclically shifting the first column of each block to the right by n0 positions. Thus we have
the following.
Proposition 4.2: Regard Gtb
N as a matrix having N blocks of n0 columns. Suppose that q + L + 1 ≤ N . Then multiplying
the jth column of G(D) by Dq is equivalent to cyclically shifting the jth column of each block of Gtb
N to the right by qn0
positions.
Remark: In order for G′tb
N to be defined, the condition q + L + 1 ≤ N is required.
Let C be a TB convolutional code of section length N with generator matrix G(D). Let C ′ be as in the previous section.
In this case, however, the jth component of each block is cyclically shifted to the right by qn0 positions. We have shown
′
tb
that G′tb
N is obtained from GN by cyclically shifting the jth column of each block to the right by qn0 positions. Hence, C is
′tb
′
′
generated by GN . In words, C is represented as a TB convolutional code defined by G (D).
C. gj (D) ← gj (D) + Dq gi (D)
Consider addition of the ith row gi (D) multiplied by Dq to the jth row gj (D), denoted by gj (D) ← gj (D) + Dq gi (D),
where q + L + 1 ≤ N . In the following, we assume without loss of generality that i = 1 and j = 2. Let the first row of G(D)
be
g1 (D) = (g1,1 (D), g1,2 (D), · · · , g1,n0 (D)).
Also, let
g1 (D)
=
(0)
(0)
(L−1)
+(g1,1
△
=
(1)
(0)
(1)
(1)
(g1,1 , g1,2 , · · · , g1,n0 ) + (g1,1 , g1,2 , · · · , g1,n0 )D + · · ·
(L−1)
, g1,2
(L−1)
(L)
(L)
(L)
, · · · , g1,n0 )DL−1 + (g1,1 , g1,2 , · · · , g1,n0 )DL
g0 + g1 D + · · · + gL−1 DL−1 + gL DL
be the polynomial expansion, where the size of gi (0 ≤ i ≤ L) is 1 × n0 . Then the polynomial expansion of g1 (D)Dq becomes
g1 (D)Dq = g0 Dq + g1 Dq+1 + · · · + gL−1 DL+q−1 + gL DL+q .
Note that the first row of Gtb
N is expressed as
g0 , g1 , · · · , gL−1 , gL , 0, · · · , 0 .
|
{z
}
N
Hence, its right cyclic shift by qn0 positions, i.e.,
q
L+1
z }| { z
}|
{
0, · · · , 0, g0 , g1 , · · · , gL−1 , gL , 0, · · · , 0
|
{z
}
N
q
coincides with the (qk0 + 1)th row of Gtb
N . That is, g2 (D) ← g2 (D) + D g1 (D)
tb
row to the second row within the matrix GN . Note that this is an elementary row
corresponds to addition of the (qk0 + 1)th
operation. Thus we have the following.
Proposition 4.3: Suppose that q + L + 1 ≤ N . Consider the operation gj (D) ← gj (D) + Dq gi (D). Let the resulting matrix
′tb
tb
be G′ (D) and the associated TBGM be G′tb
N . Then GN is equivalent to GN .
Taking into consideration Proposition 4.3, let us introduce a useful notion. Let C and C ′ be TB convolutional codes of
section length N defined by G(D) and G′ (D), respectively. Denote by L and L′ the memory lengths of G(D) and G′ (D),
′tb
′
respectively, where max(L, L′ ) + 1 ≤ N . Let Gtb
N and GN be the TBGM’s associated with C and C , respectively. We see
tb
′tb
′
that if GN and GN are equivalent, then C = C . All of this leads to the following definition.
′tb
′
Definition 4.1: When Gtb
N and GN are equivalent, we say that G(D) and G (D) are “TB-equivalent”.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
13
Thus we have the following.
Proposition 4.4: If G(D) and G′ (D) are TB-equivalent, then a TB convolutional code defined by G(D) is represented as
a TB convolutional code defined by G′ (D), and vice versa.
Proof: A direct consequence of the definition of TB-equivalent.
V. T RELLIS R EDUCTION
FOR
TB C ONVOLUTIONAL C ODES
In this section, we will show that for a TB convolutional code of short to moderate section length, the associated TB trellis
can be reduced. We begin with an example.
A. An Example of Trellis Reduction
(00)
t=0
00
t=1
00
11
t=2
00
11
t=3
00
11
t=4
00
11
t=5
11
(01)
10
10
01
11
11
11
00
01
10
01
11
00
01
10
10
01
11
00
01
(11)
10
01
00
(10)
10
01
00
01
10
01
10
10
Fig. 1. The tail-biting convolutional code defined by G(D) = (1 + D + D 2 , 1 + D 2 ) (N = 5).
Consider the TB convolutional code C defined by G(D) = (1 + D + D2 , 1 + D2 ), where the section length N is set to 5.
The corresponding TB trellis is shown in Fig.1, where the paths which start and end in the same state are TB paths (i.e., valid
codewords). Then C is the set of all TB paths. Since G(D) has the polynomial expansion
G(D)
=
(1, 1) + (1, 0)D + (1, 1)D2
(50)
△
G0 + G1 D + G2 D2 ,
(51)
=
the TBGM associated with C is given by
Gtb
5
=
=
Based on Gtb
5 , a characteristic matrix X for C is
1 1
0 1
0 0
0 0
0 0
X=
1 0
1 1
0 0
1 0
0 1
G0
0
0
G2
G1
1
0
0
1
1
G1
G0
0
0
G2
1 1
0 1
0 0
1 0
0 1
G2
G1
G0
0
0
0
1
0
0
1
0
0
G2 0
G1 G2
G0 G1
0 G0
1 1 0 0 0 0
1 0 1 1 0 0
1 1 1 0 1 1
.
0 0 1 1 1 0
0 0 0 0 1 1
(52)
(53)
computed as follows:
1
0
1
0
0
0
0
1
1
0
0
1
1
1
0
0
0
0
1
0
1
0
1
0
1
0
0
0
0
1
1
0
0
1
1
1
0
0
0
0
0
1
1
0
1
0
1
0
0
0
0
0
1
0
0
1
1
1
0
0
0
0
0
1
1
0
1
0
1
0
0
0
0
0
1
0
0
1
1
1
(0,
(1,
(2,
(3,
(4,
(5,
(6,
(7,
(8,
(9,
5]
6]
7]
8]
9]
0]
1]
2]
3]
4].
(54)
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
14
Choosing 5 rows from X, let
′
G =
0
0
1
0
0
1
0
0
0
1
0
0
0
1
0
1 0
1 0
0 0
0 0
0 1
0
1
1
0
0
0 0
0 1
1 0
1 0
0 0
1
0
0
0
0
0
0
0
1
1
(1, 6]
(3, 8]
(5, 0]
(7, 2]
(9, 4].
(55)
We see that the rows of G′ are linearly independent and thus generate C, i.e., G′ is equivalent to Gtb
5 .
Here note that G′ consists of the first row and its right cyclic shifts by i × 2 (1 ≤ i ≤ 4) positions. Accordingly, G′ can be
regarded as the TBGM associated with
G′ (D) =
=
(D3 , 1 + D)
(0, 1) + (0, 1)D + (1, 0)D3 .
(56)
(57)
Hence, C is equally represented as a TB convolutional code defined by G′ (D) = (D3 , 1 + D). We remark that the constraint
length of G′ (D) is ν ′ = 3 and is greater than that of G(D).
On the other hand, observe that the first column of G′ (D) has a factor D2 . Then dividing the first column by D2 , we have
G′ (D) = (D3 , 1 + D) → G′′ (D) = (D, 1 + D).
(58)
Note that this transformation corresponds to cyclically shifting the first component of each branch (of a TB path) to the left
by two branches (cf. Proposition 4.1). By this transformation, the original TB convolutional code is represented using a trellis
associated with G′′ (D) as well. The trellis for G′′ (D) = (D, 1 + D) is shown in Fig.2. For example, take notice of the TB
path in Fig.1 which starts and ends in state (00):
w = 11 01 10 01 11.
Cyclically shifting the first component of each branch to the left by two branches, it becomes
wm = 11 01 10 11 01.
We see that the modified path is represented as a path which starts and ends in state (1) in Fig.2.
This example shows that there are cases where a given TB convolutional code is represented using a reduced trellis with
less state complexity, if we allow partial cyclic shifts of a TB path.
(0)
(1)
t=0
00
t=1
00
t=2
00
t=3
00
t=4
00
01
01
01
01
01
11
11
11
11
11
10
10
10
10
t=5
10
Fig. 2. The tail-biting convolutional code defined by G′′ (D) = (D, 1 + D) (N = 5).
B. Trellis Reduction for TB Convolutional Codes
The argument in the previous section, though it was presented in terms of a specific example, is entirely general. Then
the method can be directly extended to a general case. Let G(D) be as in Section III. Denote by ν the constraint length of
G(D). Consider a TB convolutional code C of section length N defined by G(D). The trellis reduction procedure becomes
as follows.
Procedure for trellis reduction:
i) Compute a characteristic matrix X for C based on the TBGM Gtb
N , where X consists of n0 rows and their right cyclic
shifts by integer multiple of n0 .
ii) Choosing k rows from X, form G′ , where G′ has the properties:
1) The rows of G′ are linearly independent and thus generate C.
2) G′ consists of k0 rows and their right cyclic shifts by integer multiple of n0 .
iii) (Direct reduction) G′ is regarded as the TBGM associated with another generator matrix G′ (D). Let ν ′ be the constraint
length of G′ (D). If ν ′ < ν, then trellis reduction for C is realized.
iv) (Indirect reduction) Even if ν ′ ≥ ν, there are cases where G′ (D) has a monomial factor Dp in some (jth) column. Then
there is a possibility that ν ′ is reduced by dividing the jth column of G′ (D) by Dp (the resulting matrix is denoted by
G′′ (D)). Let ν ′′ be the constraint length of G′′ (D). If ν ′′ < ν, then the original TB trellis can be reduced. That is, by
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
15
cyclically shifting the jth component of each branch of a TB path (∈ C) to the left by p branches, the set of modified
paths are equally represented as a TB convolutional code defined by G′′ (D) (this is justified by Proposition 4.1). Thus
trellis reduction is accomplished.
v) X is not necessarily unique [9]. Hence, if necessary, try i) ∼ iv) using another characteristic matrix X ′ for C.
Remark: For row rate codes, it is rather easy to find G′ which is equivalent to Gtb
N . Also, row rate codes make it easy to
determine whether G′ (D) can be reduced or not.
As is stated above, there are some restrictions on the selection of X and G′ . We have the following.
Proposition 5.1: The number of characteristic matrices X’s in i) is given by 2θ , where θ is defined in Section III-C. For a
fixed X, the number of G′ ’s which satisfy the condition 2) in ii) is given by n0 Ck0 .
Proof: G′ is a candidate for a TBGM associated with an encoder. Hence, the above is a consequence of the structure of
TBGM.
Example 3: Consider the TB convolutional code C of section lengh N = 6 defined by
G(D) = (1 + D + D2 + D3 , 1 + D2 + D3 ).
Using the associated TBGM, i.e.,
Gtb
6 ,
X=
(59)
a characteristic matrix X for C is computed as follows:
1 0 0 0 1 0 0 1 0 0 0 0
(0, 7]
0 1 1 0 0 1 1 0 0 0 0 0
(1, 6]
0 0 1 0 0 0 1 0 0 1 0 0
(2, 9]
0 0 0 1 1 0 0 1 1 0 0 0
(3, 8]
0 0 0 0 1 0 0 0 1 0 0 1
(4, 11]
0 0 0 0 0 1 1 0 0 1 1 0
(5, 10]
0 1 0 0 0 0 1 0 0 0 1 0
(6, 1]
1 0 0 0 0 0 0 1 1 0 0 1
(7, 0]
1 0 0 1 0 0 0 0 1 0 0 0
(8, 3]
0 1 1 0 0 0 0 0 0 1 1 0
(9, 2]
0 0 1 0 0 1 0 0 0 0 1 0 (10, 5]
1 0 0 1 1 0 0 0 0 0 0 1
(11, 4].
Choosing 6 rows from X, define G′ as
1
0
0
G′ =
0
1
0
0
0
0
1
0
0
0
1
0
0
0
1
0
0
0
0
1
0
1
0
1
0
0
0
0
0
0
0
0
1
0
1
0
1
0
0
1
0
0
0
0
0
0
0
1
0
1
0
0
1
0
0
0
0
0
0
0
1
0
1
0
0
1
0
0
0
(0, 7]
(2, 9]
(4, 11]
(6, 1]
(8, 3]
(10, 5].
(60)
(61)
′
We see that G′ is equivalent to Gtb
6 . Also, we see that G is the TBGM associated with
G′ (D) = (1 + D2 , D3 ).
(62)
Note that the constraint length ν ′ = 3 of G′ (D) is not reduced compared to that of G(D). On the other hand, observe that
the second column of G′ (D) has a factor D. Then dividing the column by D, we have
G′ (D) = (1 + D2 , D3 ) → G′′ (D) = (1 + D2 , D2 ).
(63)
This transformation corresponds to cyclically shifting the second component of each branch of a TB path to the left by one
branch (cf. Proposition 4.1). As a result, the modified paths are represented using the trellis for G′′ (D). Thus trellis reduction
for C is accomplished.
Remark: As is stated above, X is not necessarily unique. For example, if a characteristic matrix
1 1 1 0 1 1 1 1 0 0 0 0
(0, 7]
0 1 1 0 0 1 1 0 0 0 0 0 (1, 6]
0 0 1 1 1 0 1 1 1 1 0 0 (2, 9]
0 0 0 1 1 0 0 1 1 0 0 0 (3, 8]
0 0 0 0 1 1 1 0 1 1 1 1 (4, 11]
0 0 0 0 0 1 1 0 0 1 1 0 (5, 10]
X′ =
(64)
1 1 0 0 0 0 1 1 1 0 1 1 (6, 1]
1 0 0 0 0 0 0 1 1 0 0 1 (7, 0]
1 1 1 1 0 0 0 0 1 1 1 0 (8, 3]
0 1 1 0 0 0 0 0 0 1 1 0 (9, 2]
1 0 1 1 1 1 0 0 0 0 1 1 (10, 5]
1 0 0 1 1 0 0 0 0 0 0 1
(11, 4]
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
16
is used, then trellis reduction cannot be realized using the above procedure.
Using appropriate characteristic matrices, the above reduction method can also be applied to the following cases:
(1) R = 1/2, ν = 4, N = 6 :
G(D)
= (1 + D + D4 , 1 + D2 + D3 + D4 )
→ G′ (D) = (D2 , 1 + D)
→ G′′ (D) = (D, 1 + D).
(2) R = 1/2, ν = 5, N = 10 :
G(D)
=
(1 + D + D2 + D3 + D4 + D5 , 1 + D3 + D5 )
→ G′ (D) = (D4 + D5 , 1 + D + D4 )
→ G′′ (D) = (D3 + D4 , 1 + D + D4 ).
(3) R = 1/2, ν = 6, N = 8 :
G(D)
= (1 + D + D4 + D5 + D6 , 1 + D2 + D3 + D4 + D6 )
→ G′ (D) = (D + D2 + D3 , 1 + D2 + D3 ).
G(D)
=
(4) R = 1/2, ν = 6, N = 8 :
(1 + D + D2 + D3 + D6 , 1 + D2 + D3 + D5 + D6 )
→ G′ (D) = (1 + D2 + D3 , D2 + D3 + D4 )
→ G′′ (D) = (1 + D2 + D3 , D + D2 + D3 ).
(5) R = 1/3, ν = 3, N = 5 :
G(D)
= (1 + D + D2 + D3 , 1 + D + D3 , 1 + D2 + D3 )
→ G′ (D) = (D + D2 , D3 , 1)
→ G′′ (D) = (1 + D, D, 1).
(6) R = 2/3, ν = 4, N = 6 :
G(D)
D + D2 1 + D 1 + D + D2
=
1
1 + D2
1 + D2
2
D 1+D
D
′
→ G (D) =
D
0
1 + D2
D 1+D
D
→ G′′ (D) =
.
1
0
1 + D2
C. Trellis Reduction Using a Reciprocal Dual Encoder
For high rate codes, G′ (D) may not have a monomial factor in any columns. Then it is not easily determined whether
G′ (D) can be reduced or not. In such cases, it is useful to consider a reciprocal dual encoder H̃ ′ (D) associated with G′ (D).
A reciprocal dual encoder [21] is defined as follows: Let G(D) be as in Section III. Also, let H(D) be a corresponding check
matrix with size (n0 − k0 ) × n0 . A reciprocal dual encoder H̃(D) is obtained by substituting D−1 for D in H(D) and by
⊥
multiplying the ith (1 ≤ i ≤ n0 − k0 ) row of the resulting matrix by Dνi , where νi⊥ is the degree of the ith row of H(D).
Definition 5.1 (McEliece and Lin [18]): Let Gscalar be a scalar generator matrix for a terminated convolutional code defined
by G(D) [18], [21]. Gscalar is given by
G0 G1 ... GL
G0 ... GL−1 GL
... ...
... ...
.
G0 G1 ... GL
Gscalar =
(65)
G
... GL−1 GL
0
... ...
... ...
G0 G1 ... GL
The (L + 1)k0 × n0 matrix
GL
△ GL−1
Ĝ =
··· ,
G0
(66)
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
17
which repeatedly appears as a vertical slice in Gscalar except initial and final transient sections, is called the matrix module.
Then the trellis module T0 for the trellis associated with Gscalar corresponds to Ĝ. If Gscalar is in minimal-span form, then T0
is minimal. The state complexity profile of T0 is an n0 -tuple consisting of the dimensions of state spaces Vi (0 ≤ i ≤ n0 − 1)
of T0 .
The meaning of obtaining a reciprocal dual encoder is based on the following result [26], [30], [31].
Proposition 5.2 (Tang and Lin [30]): Consider a minimal trellis module of G(D) and that of an associated reciprocal dual
encoder H̃(D). Then their state complex profiles are identical.
Hence, in order to determine whether G′ (D) is reduced or not, we can compute a reciprocal dual encoder H̃ ′ (D) associated
with G′ (D). In connection with an encoder G(D) and an associated reciprocal dual encoder H̃(D), we have the following.
tb
Proposition 5.3: Let Gtb
N be the TBGM associated with G(D). Then a check matrix corresponding to GN is obtained as
tb
the TBGM (denoted by H̃N
) associated with a reciprocal dual encoder H̃(D).
Proof: Let the polynomial expansion of H(D) be
H(D) = H0 + H1 D + · · · + HM−1 DM−1 + HM DM ,
(67)
where M is the memory length of H(D) and Hi (0 ≤ i ≤ M ) are (n0 − k0 ) × n0 matrices. It is known (e.g., [28]) that a
check matrix corresponding to Gtb
N is given by
H0
HM ... H2 H1
H1
H0
... ... H2
...
H
...
HM ...
1
HM−1 ... ... H0
HM
tb
H =
(68)
H0
HM HM−1 ... H1
HM ... ...
H1 ...
... HM−1
...
... H0
HM HM−1 ... H1 H0
with size (n0 − k0 )N × n0 N . On the other hand, let the polynomial expansion of H̃(D) be
H̃(D) = H̃0 + H̃1 D + · · · + H̃M−1 DM−1 + H̃M DM .
tb
H̃N
)
Then the TBGM associated with H̃(D) (denoted by
H̃0 H̃1
H̃0
tb
H̃N =
H̃M
H̃M−1 H̃M
...
...
H̃1 H̃2
(69)
is defined by
... H̃M−1
...
...
...
...
...
H̃0
H̃1
H̃0
...
...
H̃M
H̃M−1 H̃M
H̃M
... ...
... ... H̃M
H̃1 ... H̃M−1
H̃0 ... ...
... H̃1
H̃0
(70)
with size (n0 − k0 )N × n0 N .
Here take notice of the ith (1 ≤ i ≤ n0 − k0 ) row of
tb
H̃0 , H̃1 , · · · , H̃M−1 , H̃M , 0, · · · , 0 ⊂ H̃N
.
We see that the row is identical to the ith row of
Hνi⊥ , · · · , H1 , H0 , 0, · · · , 0, HM , · · · , Hνi⊥ +1 ⊂ H tb .
Similarly, the ith (1 ≤ i ≤ n0 − k0 ) row of
tb
0, H̃0 , H̃1 , · · · , H̃M−1 , H̃M , 0, · · · , 0 ⊂ H̃N
is identical to the ith row of
Hνi⊥ +1 , · · · , H1 , H0 , 0, · · · , 0, HM , · · · , Hνi⊥ +2 ⊂ H tb .
tb
tb
Due to the cyclic structures of H tb and H̃N
, similar correspondences hold successively. Hence, H̃N
is given as a row permutation
tb
of H .
A procedure for computing H̃ ′ (D) is obtained based on the above proposition.
Procedure for computing H̃ ′ (D):
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
18
tb
i) Compute a characteristic matrix Y for the dual code C ⊥ based on H̃N
, where Y consists of n0 rows and their right
cyclic shifts by integer multiple of n0 .
ii) Choosing (n − k) rows from Y , form H̃ ′ , where H̃ ′ has the properties:
1) The rows of H̃ ′ are linearly independent and thus generate C ⊥ .
2) H̃ ′ consists of (n0 − k0 ) rows and their right cyclic shifts by integer multiple of n0 .
iii) Let H̃ ′ (D) be the polynomial matrix whose TBGM is H̃ ′ .
tb
′
′
iv) Note that G′ and H̃ ′ are equivalent to Gtb
N and H̃N , respectively. Hence, H̃ is a check matrix corresponding to G . Then
′
′
it follows from Proposition 5.3 that H̃ (D) is a reciprocal dual encoder associated with G (D).
v) Y is not necessarily unique. Hence, if necessary, try i) ∼ iv) using another characteristic matrix Y ′ for C ⊥ .
The following is an example where trellis reduction is realized using a reciprocal dual encoder.
Example 4: Consider the rate R = 2/3 TB convolutional code C of section length N = 5 with generator matrix
1+D D
1
G(D) =
(71)
D2
1 1 + D + D2
1 0 1
1 1 0
0 0 0
=
+
D+
D2
0 1 1
0 0 1
1 0 1
△
=
G0 + G1 D + G2 D2 .
Based on the associated TBGM, i.e.,
Gtb
5
=
a characteristic matrix X for C is computed as
1 0 1
0 1 1
0 0 1
0 0 0
0 0 0
0 0 0
0 0 0
X =
0 0 0
0 0 0
0 0 0
1 0 1
1 0 0
1 1 0
0 0 1
0 0 1
G0
0
0
G2
G1
G1
G0
0
0
G2
(72)
G2
G1
G0
0
0
0
G2
G1
G0
0
0
0
G2
G1
G0
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
1
1
1
1
1
0
0
0
,
(73)
follows:
1
0
0
1
0
0
0
0
0
0
0
0
0
1
1
1
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
1
1
1
1
1
0
0
0
0
0
0
0
1
0
0
1
1
1
0
0
1
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
1
1
1
1
1
0
0
0
0
0
0
0
0
0
0
1
1
1
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
1
1
1
0
0
1
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
1
0
0
0
0
0
0
0
0
1
0
0
1
1
1
1
1
.
(74)
The span list for X is given by
T
=
{(0, 4], (1, 8], (2, 6], (3, 7], (4, 11], (5, 9], (6, 10], (7, 14],
(8, 12], (9, 13], (10, 2], (11, 0], (12, 1], (13, 5], (14, 3]}.
(75)
Choosing 10 rows from X, let
G′ =
1
0
0
0
0
0
0
1
1
0
0
0
0
0
0
0
0
0
1
0
1
1
0
0
0
0
0
0
0
1
1
0
1
0
0
0
0
0
0
1
1
0
0
0
0
0
0
0
0
0
0
1
1
1
0
0
0
0
0
0
0 0
1 0
1 1
0 0
1 0
0 0
0 0
0 0
0 0
0 0
0
0
0
1
1
1
0
0
0
0
0
0
0
1
1
0
1
0
0
0
0 0
0 0
0 0
0 0
1 0
0 1
0 1
0 1
0 0
0 0
0
0
0
0
0
1
1
0
1
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
1
1
1
.
(76)
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
19
The span list for G′ is given by
S
= {(0, 4], (2, 6], (3, 7], (5, 9], (6, 10],
(8, 12], (9, 13], (11, 0], (12, 1], (14, 3]}.
(77)
′
We see that the rows of G′ are linearly independent and thus generate C, i.e., G′ is equivalent to Gtb
5 . Also, note that G is
the TBGM associated with
1+D D
1
′
G (D) =
(78)
D2
0 1+D
1 0 1
1 1 0
0 0 0
=
+
D+
D2
0 0 1
0 0 1
1 0 0
△
=
G′0 + G′1 D + G′2 D2 .
(79)
Hence, the original TB convolutional code is equally represented as a TB convolutional code defined by G′ (D).
Observe that the constraint length of G′ (D) is ν ′ = 3 and is equal to that of G(D). Also, notice that the second column of
′
G (D) has a factor D. However, ν ′ is not reduced by dividing the column by D. In general, it is difficult to tell a possibility
of reduction of G′ (D) just by looking at its entries. So, we will compute a reciprocal dual encoder H̃ ′ (D) associated with
G′ (D).
We begin with a reciprocal dual encoder H̃(D) associated with G(D). H̃(D) is given by
H̃(D) = (1 + D + D2 + D3 , 1 + D + D3 , 1 + D2 + D3 ).
Based on H̃5tb , a characteristic matrix Y
1
0
0
0
1
0
1
Y =
1
0
1
0
1
1
0
1
for C ⊥ is computed as follows:
1
1
0
0
0
0
1
0
1
0
0
0
1
0
0
1
0
1
0
0
0
1
0
0
1
1
0
0
0
0
1
0
1
1
0
0
0
1
0
1
1
0
1
0
1
1
0
0
1
1
0
0
0
0
1
0
1
0
0
0
0
0
0
1
0
1
0
0
0
1
0
0
1
1
0
1
0
1
1
0
1
1
0
0
0
1
0
1
1
0
0
0
0
1
0
0
1
1
0
0
0
0
1
0
1
1
1
0
0
0
0
1
0
1
0
0
0
1
0
0
1
1
0
1
0
1
1
0
1
1
0
0
0
1
0
1
0
1
0
0
0
1
0
0
1
1
0
0
0
0
1
0
0
1
1
0
0
0
0
1
0
1
0
0
0
0
1
0
1
1
0
1
0
1
1
0
1
1
0
0
0
0
0
1
0
1
0
0
0
1
0
0
1
1
0
0
0
0
1
0
0
1
1
0
0
0
0
1
0
1
The span list for Y is given by
T̂
(80)
.
(81)
= {(0, 11], (1, 12], (2, 10], (3, 14], (4, 0], (5, 13], (6, 2], (7, 3],
(8, 1], (9, 5], (10, 6], (11, 4], (12, 8], (13, 9], (14, 7]}.
Note that if the span list for X is T =
1, · · · , 15} [10], [13].
Next, choosing 5 rows from Y , let
0
0
H̃ ′ =
0
1
1
(82)
{(al , bl ], l = 1, · · · , 15}, then the span list for Y is given by T̂ = {(bl , al ], l =
0
0
1
0
0
1
0
0
0
0
1 0
0 0
0 0
0 1
1 0
0
1
0
0
0
1
1
0
0
0
0 0
0 0
0 1
0 0
1 0
0
1
1
0
0
1
0
0
0
0
0 0
0 0
0 1
1 1
0 0
0
1
0
0
0
0
0
0
0
1
.
(83)
The span list for H̃ ′ is given by
Ŝ = {(2, 10], (5, 13], (8, 1], (11, 4], (14, 7]}.
(84)
We see that H̃ ′ is equivalent to H̃5tb . Thus H̃ ′ is a scalar check matrix corresponding to G′ . Also, note that H̃ ′ is the
TBGM associated with H̃ ′ (D) = (D + D2 , D3 , 1). We already know that G′ is the TBGM associated with G′ (D). Hence, by
Proposition 5.3, a reciprocal dual encoder associated with G′ (D) is given by
H̃ ′ (D) = (D + D2 , D3 , 1).
(85)
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
20
Observe that H̃ ′ (D) = (D + D2 , D3 , 1) has a factor D in the first column and a factor D2 in the second column. Then
sweeping these factors out of the corresponding columns, the constraint length of H̃ ′ (D) is reduced to one. This fact implies
that the constraint length of G′ (D) can also be reduced.
In the following, we will show that reduction of G′ (D) is actually realized. For the purpose, a check matrix corresponding
to G′ (D), i.e.,
H ′ (D) = (D−1 + D−2 , D−3 , 1) × D3 = (D + D2 , 1, D3 )
(86)
is used.
Let G(D) and H(D) be a generator matrix and a corresponding check matrix for a convolutional code, respectively. In the
following, this relation is denoted by G(D) ⇔ H(D). It is shown [27] that G(D) and H(D) can be reduced simultaneously,
if reduction is possible, where the relation ⇔ is retained in the whole reduction process. We apply the method to our case
under consideration.
Step 1: For G′ (D), add the first row multiplied by D to the second row. By Proposition 4.3, this is a TB-equivalent
transformation. As a result, we have
1+D D 1
G′′ (D) =
⇔ H ′ (D) = (D + D2 , 1, D3 ).
D
D2 1
Step 2: Divide the second column of G′′ (D) by D, while divide the first and third columns of H ′ (D) by D. Then we have
1+D 1 1
G′′′ (D) =
⇔ H ′′′ (D) = (1 + D, 1, D2 ).
D
D 1
Step 3: Multiply the third column of G′′′ (D) by D, while divide the third column of H ′′′ (D) by D. Then we have
1+D 1 D
(4)
G (D) =
⇔ H (4) (D) = (1 + D, 1, D).
D
D D
1+D 1 D
(4)
Step 4: Note that G (D) =
is not basic [5]. Using an invariant-factor decomposition [5] of G(4) (D),
D
D D
an equivalent basic matrix
1+D 1 D
G(5) (D) =
(87)
1
1 1
is obtained. Note that G(4) (D) and G(5) (D) are TB-equivalent (cf. Proposition 4.4).
In the above reduction process for G′ (D), except for TB-equivalent transformations, the second column is divided by D,
whereas the third column is multiplied by D. Accordingly, for each TB path, let us cyclically shift the second component of
each branch to the left by one branch and cyclically shift the third component of each branch to the right by one branch. Then
the modified TB paths are represented as a TB convolutional code defined by G(5) (D) (see Propositions 4.1 and 4.2). The
trellis for G(5) (D) is shown in Fig.3. For example, take an information sequence
u = 01 00 01 00 00
and the corresponding TB path
w = uGtb
5 = 011 001 110 001 101.
By cyclically shifting the second component of each branch to the left by one branch, and by cyclically shifting the third
component of each branch to the right by one branch, we have
w m = 001 011 101 000 111.
We see that w m is a TB path which starts and ends in state (0) in Fig.3.
(0)
(1)
t=0 111
000
001
110
t=1 111
000
001
110
t=2 111
000
001
110
t=3 111
000
001
110
t=4 111
000
001
110
010
010
010
010
010
101
101
101
101
101
011
100
011
100
011
100
011
100
011
100
t=5
Fig. 3. The tail-biting convolutional code defined by G(5) (D) (N = 5).
Remark: We remark that in the above argument, it is assumed that G′ is equivalent to Gtb
5 (i.e., the equivalence has been
checked beforehand). In general, however, k = k0 N is relatively large for high rate codes. Hence, it is preferable that the
′
tb
equivalence of G′ and Gtb
5 is derived without checking it beforehand. Actually, the equivalence of G and G5 is derived from
′
tb
the equivalence of H̃ and H̃5 using the result of Gluesing-Luerssen and Weaver [10, Theorem IV.3] (see Appendix A).
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
21
D. Relation Between Trellis Reduction and Section Length
In the proposed trellis reduction method, the section length N is an important parameter. Actually, the method is effective for
TB convolutional codes of short to moderate section length. This is because the span lengths of characteristic generators increase
as N grows (see Section III-D). We have already shown that a TB trellis with generator matrix G(D) = (1 + D + D2 , 1 + D2 )
can be reduced for the case of N = 5. Consider the same trellis. This time, however, N is set to 6. Then Gtb
6 is given by
G0 G1 G2 0
0
0
0 G0 G1 G2 0
0
0
0 G0 G1 G2 0
Gtb
=
(88)
6
0
0
0 G0 G1 G2
G2 0
0
0 G0 G1
G1 G2 0
0
0 G0
1 1 1 0 1 1 0 0 0 0 0 0
(0, 5]
0 0 1 1 1 0 1 1 0 0 0 0 (2, 7]
0 0 0 0 1 1 1 0 1 1 0 0 (4, 9]
=
(89)
0 0 0 0 0 0 1 1 1 0 1 1 (6, 11]
1 1 0 0 0 0 0 0 1 1 1 0 (8, 1]
1 0 1 1 0 0 0 0 0 0 1 1
(10, 3].
Note that to each generator in Gtb
6 , its span is assigned in the natural
are the same, i.e., 6. A characteristic matrix is computed as follows:
1 1 1 0 1 1 0 0 0
0 1 0 1 1 1 1 1 1
0 0 1 1 1 0 1 1 0
0 0 0 1 0 1 1 1 1
0 0 0 0 1 1 1 0 1
1 0 0 0 0 1 0 1 1
X =
0 0 0 0 0 0 1 1 1
1 1 1 0 0 0 0 1 0
1 1 0 0 0 0 0 0 1
1 1 1 1 1 0 0 0 0
1 0 1 1 0 0 0 0 0
0 1 1 1 1 1 1 0 0
Thus the set of basic spans is given by
T0 =
manner. Observe that the span lengths of these spans
0
0
0
1
1
1
0
1
1
1
0
0
0
0
0
1
0
1
1
1
1
0
1
0
0
0
0
0
0
1
1
1
0
1
1
1
(0, 5]
(1, 8]
(2, 7]
(3, 10]
(4, 9]
(5, 0]
(6, 11]
(7, 2]
(8, 1]
(9, 4]
(10, 3]
(11, 6].
(0, 5]
(1, 8].
(90)
(91)
With respect to G′ , there are two cases. When the first row of X is used as a basic generator of G′ , G′ is identical to Gtb
6 .
When the second row of X is used as a basic generator of G′ , the span lengths of rows of G′ are 8 and are greater than
6. These facts mean that in either case, trellis reduction is not realized using the proposed method. On the other hand, this
example implies that the upper bound for N can be estimated by comparing the span lengths of generators in G′ with those
of generators in Gtb
N.
Let X be a characteristic matrix for a TB convolutional code of section length N . We already know that the associated span
list T consists of the set of basic spans
(0, b0 ]
(1, b1 ]
T0 =
···
(n0 − 1, bn0 −1 ]
and ρin0 (T0 ) (i = 1, 2, · · · , N − 1). Also, the sum of span lengths of spans in T0 is given by
ℓ = n0 ((n0 − k0 )N + 1).
In the proposed method, G′ consists of k generators in X. From a span viewpoint, this corresponds to choosing k0 spans from
T0 . Accordingly, the sum of span lengths of these k0 spans, denoted by ℓ′ , is approximated by
ℓ′
;
(k0 /n0 ) × n0 ((n0 − k0 )N + 1)
=
k0 ((n0 − k0 )N + 1).
(92)
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
22
On the other hand, consider Gtb
N , where to each generator, its span is assigned in the natural manner. Then the span list
consists of the set of basic spans T̂0 and ρin0 (T̂0 ) (i = 1, 2, · · · , N − 1). We evaluate the sum of span lengths of spans in T̂0 ,
denoted by ℓ̂. Let νi be the degree of the ith row of G(D). Here take notice of the first block of k0 rows in Gtb
N , i.e.,
(G0 , G1 , · · · , GL−1 , GL , 0, · · · , 0) .
The span length of the ith (1 ≤ i ≤ k0 ) row is approximated by n0 (νi + 1). Hence, we have
ℓ̂
; n0 (ν1 + 1) + n0 (ν2 + 1) + · · · + n0 (νk0 + 1)
= n0 (ν1 + ν2 + · · · + νk0 ) + n0 k0
= n0 (ν + k0 ),
(93)
△
where ν = ν1 + ν2 + · · · + νk0 is the constraint length of G(D). Since trellis reduction is realized in the case where G′ consists
of generators with short span length, we can take the inequality
ℓ′ ≤ ℓ̂
as a criterion for trellis reduction. That is, we can estimate the upper bound for N using the inequality
(♯) k0 ((n0 − k0 )N + 1) ≤ n0 (ν + k0 ).
(94)
For several concrete cases, we will show the condition (♯).
(1) R = 1/2:
′
ℓ ;N +1
ℓ̂ ; 2ν + 2
(♯) N ≤ 2ν + 1.
(2) R = 1/3:
ℓ′ ; 2N + 1
ℓ̂ ; 3ν + 3
(♯) N ≤ ⌊(3/2)ν⌋ + 1.
(3) R = 2/3:
ℓ′ ; 2N + 2
ℓ̂ ; 3ν + 6
(♯) N ≤ ⌊(3/2)ν⌋ + 2.
We observe that the TB convolutional codes presented in Section V-B all satisfy the condition (♯). Also, the rate R = 2/3
TB convolutional code discussed in the previous section satisfies the condition (♯).
VI. C ONCLUSION
In this paper, we have derived several basic properties of a characteristic matrix for a TB convolutional code. We have
shown that the characteristic span list consists of some basic spans and their right cyclic shifts. Using the derived results, we
have shown that a trellis associated with a given TB convolutional code can be reduced in some cases. As candidates for trellis
reduction, we have taken the generator matrices from the tables in [12, Chapter 8] in principle. For example, the rate R = 1/2
encoders in Section V-B were chosen from [12, TABLE 8.1]. On the other hand, good TB convolutional encoders have been
obtained [12], [25]. Here, for a given rate R = k0 /n0 , the optimal encoder of memory length L produces the largest minimum
distance d for each section length N . We have applied the proposed reduction method to some of such encoders (see [12,
TABLE 8.19]). As a result, for example, we have obtained G = (6, 7) (ν = 2, N = 5) from G = (50, 64) (ν = 3, N = 5),
where the octal notation for generator matrices is used. Similarly, we have obtained G = (54, 60) (ν = 3, N = 6) from
G = (46, 60) (ν = 4, N = 6). Note that both G = (6, 7) (ν = 2, N = 5) and G = (54, 60) (ν = 3, N = 6) are listed in the
same table.
Finally, we remark that the proposed trellis reduction method depends on the choice of a characteristic matrix for a given
convolutional code. Though the number of characteristic matrices to be examined is rather restricted (cf. Proposition 5.1), the
method is not fully constructive. Also, a detailed condition that trellis reduction is realized has to be clarified.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
23
A PPENDIX A
P ROOF
OF THE EQUIVALENCE OF
G′
AND
Gtb
5
We first prove the following.
Proposition A.1: Let G(D) be as in Section III. Consider a TB convolutional code C generated by Gtb
N and the corresponding
dual code C ⊥ . Let X be a characteristic matrix for C with span list T . Also, let Y be a characteristic matrix for C ⊥ with
span list T̂ . Let X̃ and Ỹ be submatrices of X and Y , respectively, where X̃ consists of k rows in X, whereas Ỹ consists of
(n − k) rows in Y . Denote by S and Ŝ the span lists for X̃ and Ỹ , respectively. Here assume the following:
i) X̃ consists of k0 rows and their right cyclic shifts by integer multiple of n0 .
ii) Each span in S does not include any spans in T except itself.
iii) The rows of Ỹ are linearly independent and thus generate C ⊥ .
iv) Ỹ consists of (n0 − k0 ) rows and their right cyclic shifts by integer multiple of n0 .
v) Each span in Ŝ does not include any spans in T̂ except itself.
vi) S and Ŝ satisfy (b, a] ∈ Ŝ ↔ (a, b] ∈
/ S (that is, Ŝ consists of the spans in T̂ whose reverse is not in S).
Then the rows of X̃ are linearly independent, thus generate C, i.e., X̃ is equivalent to Gtb
N.
Remark: When X̃ consists of generators in X with short span length, it is probable that the condition ii) holds. Similarly,
when Ỹ consists of generators in Y with short span length, it is probable that the condition v) holds.
Proof: From ii), it follows that X̃ is common to all the characteristic matrices for C. Similarly, from v), it follows that Ỹ
is common to all the characteristic matrices for C ⊥ . Also, by vi), (X̃, Ỹ ) is a dual selection of (X, Y ) [10, Definition IV.2].
As a result [10, Theorem IV.3], we have
1) rank X̃ = k ↔ rank Ỹ = n − k
2) Let rank X̃ = k. Then the KV trellises [10] based on (X̃, S) and (Ỹ , Ŝ) are dual to each other.
By iii), rank Ỹ = n − k. Hence, by 1), rank X̃ = k.
⊥
Let us go back to Example 4. In this example, the code C generated by Gtb
generated by H̃5tb are
5 and the dual code C
⊥
considered. X is a characteristic matrix for C, whereas Y is a characteristic matrix for C . Note that neither X nor Y are
unique. These are observed from the relation of inclusion in the associated span lists T and T̂ , where
T
= {(0, 4], (1, 8], (2, 6], (3, 7], (4, 11], (5, 9], (6, 10], (7, 14],
T̂
= {(0, 11], (1, 12], (2, 10], (3, 14], (4, 0], (5, 13], (6, 2], (7, 3],
(8, 1], (9, 5], (10, 6], (11, 4], (12, 8], (13, 9], (14, 7]}.
(8, 12], (9, 13], (10, 2], (11, 0], (12, 1], (13, 5], (14, 3]}
Next, take notice of the matrices G′ and H̃ ′ , which are submatrices of X and Y , respectively. The corresponding span lists
are given by
S
= {(0, 4], (2, 6], (3, 7], (5, 9], (6, 10],
Ŝ
(8, 12], (9, 13], (11, 0], (12, 1], (14, 3]}
= {(2, 10], (5, 13], (8, 1], (11, 4], (14, 7]}.
Here note the following:
• Each span in S does not include any spans in T except itself.
• Each span in Ŝ does not include any spans in T̂ except itself.
Moreover, Ŝ consists of the spans in T̂ whose reverse is not in S. Actually, by reversing the spans in Ŝ, we have
(10, 2], (13, 5], (1, 8], (4, 11], (7, 14].
We see that these spans are not in S.
All these facts show that the conditions in Proposition A.1 are satisfied, when X̃ and Ỹ are replaced by G′ and H̃ ′ ,
respectively. Hence, the equivalence of G′ and Gtb
5 is derived.
Remark: [10, Theorem IV.3] holds only for a pair (X, Y ), where X is a characteristic matrix for C and Y is the corresponding
dual one for C ⊥ (see [10]). On the other hand, the pair (X, Y ) computed above may not be in the duality relation. However,
G′ is common to all the characteristic matrices for C, and H̃ ′ is common to all the characteristic matrices for C ⊥ as well.
Hence, the theorem can be applied to our case.
ACKNOWLEDGMENT
The author would like to thank Prof. Heide Gluesing-Luerssen for valuable comments on the duality of Koetter-Vardy (KV)
trellises.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
24
R EFERENCES
[1] J. B. Anderson and S. M. Hladik, “An optimal circular Viterbi decoder for the bounbed distance criterion,” IEEE Trans. Commun., vol. 50, no. 11,
pp. 1736–1742, Nov. 2002.
[2] L. R. Bahl, J. Cocke, F. Jelinek, and J. Raviv, “Optimal decoding of linear codes for minimizing symbol error rate,” IEEE Trans. Inform. Theory,
vol. IT-20, no. 2, pp. 284–287, March 1974.
[3] A. R. Calderbank, G. D. Forney. Jr., and A. Vardy, “Minimal tail-biting trellises: The Golay code and more,” IEEE Trans. Inform. Theory, vol. 45, no. 5,
pp. 1435–1455, July 1999.
[4] D. Conti and N. Boston, “On the algebraic structure of linear tail-biting trellises,” IEEE Trans. Inform. Theory, vol. 61, no. 5, pp. 2283–2299, May
2015.
[5] G. D. Forney, Jr., “Convolutional codes I: Algebraic structure,” IEEE Trans. Inform. Theory, vol. IT-16, no. 6, pp. 720–738, Nov. 1970.
, “Structural analysis of convolutional codes via dual codes,” IEEE Trans. Inform. Theory, vol. IT-19, no. 4, pp. 512–518, July 1973.
[6]
[7]
, “Coset codes–Part II: Binary lattices and related codes,” IEEE Trans. Inform. Theory, vol. 34, no. 5, pp. 1152–1187 (Appendix A), Sept. 1988.
[8]
, “Dimension/length profiles and trellis complexity of linear block codes,” IEEE Trans. Inform. Theory, vol. 40, no. 6, pp. 1741–1752, Nov.
1994.
[9] H. Gluesing-Luerssen and E. A. Weaver, “Linear tail-biting trellises: Characteristic generators and the BCJR-construction,” IEEE Trans. Inform. Theory,
vol. 57, no. 2, pp. 738–751, Feb. 2011.
, “Characteristic generators and dualization for tail-biting trellises,” IEEE Trans. Inform. Theory, vol. 57, no. 11, pp. 7418–7430, Nov. 2011.
[10]
[11] H. Gluesing-Luerssen and G. D. Forney, Jr., “Local irreducibility of tail-biting trellises,” IEEE Trans. Inform. Theory, vol. 59, no. 10, pp. 6597–6610,
Oct. 2013.
[12] R. Johannesson and K. S. Zigangirov, Fundamentals of Convolutional Coding. New York: IEEE Press, 1999.
[13] R. Koetter and A. Vardy, “The structure of tail-biting trellises: Minimality and basic principles,” IEEE Trans. Inform. Theory, vol. 49, no. 9, pp. 2081–2105,
Sep. 2003.
[14] F. R. Kschischang and V. Sorokine, “On the trellis structure of block codes,” IEEE Trans. Inform. Theory, vol. 41, no. 6, pp. 1924–1937, Nov. 1995.
[15] S. Lin and R. Y. Shao, “General structure and construction of tail biting trellises for linear block codes,” in Proc. ISIT2000, p. 117, June 2000.
[16] H. H. Ma and J. K. Wolf, “On tail biting convolutional codes,” IEEE Trans. Commun., vol. COM-34, no. 2, pp. 104–111, Feb. 1986.
[17] R. J. McEliece, “On the BCJR trellis for linear block codes,” IEEE Trans. Inform. Theory, vol. 42, no. 4, pp. 1072–1092, July 1996.
[18] R. J. McEliece and W. Lin, “The trellis complexity of convolutional codes,” IEEE Trans. Inform. Theory, vol. 42, no. 6, pp. 1855–1864, Nov. 1996.
[19] D. J. Muder, “Minimal trellises for block codes,” IEEE Trans. Inform. Theory, vol. 34, no. 5, pp. 1049–1053, Sept. 1988.
[20] A. V. Nori and P. Shankar, “Unifying views of tail-biting trellis constructions for linear block codes,” IEEE Trans. Inform. Theory, vol. 52, no. 10,
pp. 4431–4443, Oct. 2006.
[21] S. Riedel, “Symbol-by-symbol MAP decoding algorithm for high-rate convolutional codes that use reciprocal dual codes,” IEEE J. Select. Areas Commun.,
vol. 16, no. 2, pp. 175–185, Feb. 1998.
[22] Y. Shany and Y. Be’ery, “Linear tail-biting trellises, the square-root bound, and applications for Reed-Muller codes,” IEEE Trans. Inform. Theory, vol. 46,
no. 4, pp. 1514–1523, July 2000.
[23] R. Y. Shao, S. Lin, and M. P. C. Fossorier, “Two decoding algorithms for tailbiting codes,” IEEE Trans. Commun., vol. 51, no. 10, pp. 1658–1665, Oct.
2003.
[24] G. Solomon and H. C. A. van Tilborg, “A connection between block and convolutional codes,” SIAM J. Appl. Math., vol. 37, no. 2, pp. 358–369, 1979.
[25] P. Ståhl, J. B. Anderson, and R. Johannesson, “Optimal and near-optimal encoders for short and moderate-length tail-biting trellises,” IEEE
Trans. Inform. Theory, vol. 45, no. 7, pp. 2562–2571, Nov. 1999.
[26] M. Tajima, K. Okino, and T. Miyagoshi, “Trellis complexity analysis of convolutional codes using reciprocal dual codes,” in Proc. SITA2004 (Japanese
Ed.), pp. 91–94, Dec. 2004.
[27]
, “Simultaneous code/error-trellis reduction for convolutional codes using shifted code/error-subsequences,” in Proc. ISIT2011, pp. 2409–2413,
Aug. 2011.
[28] M. Tajima and K. Okino, “Error-trellises for tailbiting convolutional codes,” in Proc. ISITA2012, pp. 648–652, Oct. 2012.
[29] M. Tajima, “Trellis reduction for tail-biting convolutional codes using characteristic matrices and cyclically shifted code-subsequences,” IEICE Technical
Report (Japanese Ed.), IT2016-108, pp. 61–66, March 2017.
[30] H.-H. Tang and M.-C. Lin, “On (n, n-1) convolutional codes with low trellis complexity,” IEEE Trans. Commun., vol. 50, no. 1, pp. 37–47, Jan. 2002.
[31] H.-H. Tang, M.-C. Lin, and B. F. U. Filho, “Minimal trellis modules and equivalent convolutional codes,” IEEE Trans. Inform. Theory, vol. 52, no. 8,
pp. 3738–3746, Aug. 2006.
[32] E. A. Weaver, “Minimality and duality of tail-biting trellises for linear codes,” Ph.D. dissertation, submitted to the University of Kentucky, Lexington,
Kentucky, USA, April 2012.
[33] C. Weiss, C. Bettstetter, and S. Riedel, “Code construction and decoding of parallel concatenated tail-biting codes,” IEEE Trans. Inform. Theory, vol. 47,
no. 1, pp. 366–386, Jan. 2001.
Masato Tajima (M’86–SM’13) was born in Toyama, Japan, on August 13, 1949. He received the B.E., M.E., and Dr. of Eng. degrees all in electrical
engineering from Waseda University, Tokyo, Japan, in 1972, 1974, and 1979, respectively. He joined the Electronics Equipment Laboratory of Toshiba R&D
Center in 1979, where he engaged in research and development of channel coding techniques with applications to satellite communication systems. From 1993
to 2006, he was with the Department of Intellectual Information Systems Engineering, Toyama University, first as an Associate Professor, next as a Professor.
From 2006 to 2015, he was with the Graduate School of Science and Engineering, University of Toyama, as a Professor. He is currently a Professor Emeritus
at University of Toyama. His research interests are in coding theory and its applications.
| 7cs.IT
|
arXiv:1705.00985v1 [cs.DS] 2 May 2017
Determinant-Preserving Sparsification of SDDM Matrices with
Applications to Counting and Sampling Spanning Trees
David Durfee∗
John Peebles†
Richard Peng‡
Anup B. Rao§
May 3, 2017
Abstract
We show variants of spectral sparsification routines can preserve the total spanning tree
counts of graphs, which by Kirchhoff’s matrix-tree theorem, is equivalent to determinant of
a graph Laplacian minor, or equivalently, of any SDDM matrix. Our analyses utilizes this
combinatorial connection to bridge between statistical leverage scores / effective resistances
and the analysis of random graphs by [Janson, Combinatorics, Probability and Computing ‘94]. This leads to a routine that in quadratic time, sparsifies a graph down to about
n1.5 edges in ways that preserve both the determinant and the distribution of spanning trees
(provided the sparsified graph is viewed as a random object). Extending this algorithm to
work with Schur complements and approximate Choleksy factorizations leads to algorithms
for counting and sampling spanning trees which are nearly optimal for dense graphs.
We give an algorithm that computes a (1 ± δ) approximation to the determinant of any
SDDM matrix with constant probability in about n2 δ −2 time. This is the first routine for
graphs that outperforms general-purpose routines for computing determinants of arbitrary
matrices. We also give an algorithm that generates in about n2 δ −2 time a spanning tree of
a weighted undirected graph from a distribution with total variation distance of δ from the
w -uniform distribution .
1
Introduction
The determinant of a matrix is a fundamental quantity in numerical algorithms due to its connection to the rank of the matrix and its interpretation as the volume of the ellipsoid corresponding
of the matrix. For graph Laplacians, which are at the core of spectral graph theory and spectral algorithms, the matrix-tree theorem gives that the determinant of the minor obtained by
removing one row and the corresponding column equals to the total weight of all the spanning
trees in the graph [Kir47] . Formally on a weighted graph G with n vertices we have:
det LG
1:n−1,1:n−1 = TG ,
where LG is the graph Laplacian of G and and TG is the total weight of all the spanning
trees of G. As the all-ones vector is in the null space of LG , we need to drop its last row
∗
Georgia Institute of Technology. email:[email protected]
Massachusetts Institute of Technology. email:[email protected]
‡
Georgia Institute of Technology. email:[email protected]
§
Georgia Institute of Technology. email:[email protected]
†
1
and column and work with LG
1:n−1,1:n−1 , which is precisely the definition of SDDM matrices
in numerical analysis [ST14]. The study of random spanning trees builds directly upon this
connection between tree counts and determinants, and also plays an important role in graph
theory [GRV09, AGM+ 10, FHHP11].
While there has been much progress in the development of faster spectral algorithms, the
estimation of determinants encapsulates many shortcomings of existing techniques. Many of the
nearly linear time algorithms rely on sparsification procedures that remove edges from a graph
while provably preserving the Laplacian matrix as an operator, and in turn, crucial algorithmic
quantities such as cut sizes, Rayleigh quotients, and eigenvalues. The determinant of a matrix
on the other hand is the product of all of its eigenvalues. As a result, a worst case guarantee
of 1 ± (ǫ/n) per eigenvalue is needed to obtain a good overall approximation, and this in turn
leads to additional factors of n in the number of edges needed in the sparse approximate.
Due to this amplification of error by a factor of n, previous works on numerically approximating determinants without dense-matrix multiplications [BDKZ15a, HAB14a, HMS15] usually
focus on the log-determinant, and (under a nearly-linear running time) give errors of additive
ǫn in the log determinant estimate, or a multiplicative error of exp(ǫn) for the determinant.
The lack of a sparsification procedure also led to the running time of random spanning tree
sampling algorithms to be limited by the sizes of the dense graphs generated in intermediate
steps [KM09, MST15, DKP+ 16].
In this paper, we show that a slight variant of spectral sparsification preserves determinant
approximations to a much higher accuracy than applying the guarantees to individual edges.
Specifically, we show that sampling ω(n1.5 ) edges from a distribution given by leverage scores,
or weight times effective resistances, produces a sparser graph whose determinant approximates
that of the original graph. Furthermore, by treating the sparsifier itself as a random object,
we can show that the spanning tree distribution produced by sampling a random tree from a
random sparsifier is close to the spanning tree distribution in the original graph in total variation
distance. Combining extensions of these algorithms with sparsification based algorithms for
graph Laplacians then leads to quadratic time algorithms for counting and sampling random
spanning trees, which are nearly optimal for dense graphs with m = Θ(n2 ).
This determinant-preserving sparsification phenomenon is surprising in several aspects: because we can also show—both experimentally and mathematically—that on the complete graph,
about n1.5 edges are necessary to preserve the determinant, this is one of the first graph sparsification phenomenons that requires the number of edges to be between >> n. The proof of
correctness of this procedure also hinges upon combinatorial arguments based on the matrix-tree
theorem in ways motivated by a result for Janson for complete graphs [Jan94], instead of the
more common matrix-concentration bound based proofs [SS11, Tro12, CP15, Coh16]. Furthermore, this algorithm appears far more delicate than spectral sparsification: it requires global
control on the number of samples, high quality estimates of resistances (which is the running
time bottleneck in Theorem 4.1 below), and only holds with constant probability. Nonetheless, the use of this procedure into our determinant estimation and spanning tree generation
algorithms still demonstrates that it can serve as a useful algorithmic tool.
2
1.1
Our Results
def P
We will use G = (V, E, w ) to denote weighted multigraphs, and d u = e:e∋u w e to denote the
weighted degree of vertex u. The weight of a spanning tree in a weighed undirected multigraph
is:
Y
def
w (T ) =
w e.
e∈T
def
We will use TG to denote the total weight of trees, TG =
result can be described by the following theorem:
P
T ∈T
w (T ). Our key sparsification
Theorem 1.1. Given any graph G and any parameter δ, we can compute in O(n2 δ−2 ) time a
graph H with O(n1.5 δ−2 ) edges such that with constant probability we have
(1 − δ) TG ≤ TH ≤ (1 + δ) TG .
This implies that graphs can be sparsified in a manner that preserves the determinant, albeit
to a density that is not nearly-linear in n.
We show how to make our sparsification routine to errors in estimating leverage scores, and
how our scheme can be adapted to implicitly sparsify dense objects that we do not have explicit
access to. In particular, we utilize tools such as rejection sampling and high quality effective
resistance estimation via projections to extend this routine to give determinant-preserving sparsification algorithms for Schur complements, which are intermediate states of Gaussian elimination
on graphs, using ideas from the sparsification of random walk polynomials.
We use these extensions of our routine to obtain a variety of algorithms built around our
graph sparsifiers. Our two main algorithmic applications are as follows. We achieve the first
algorithm for estimating the determinant of an SDDM matrix that is faster than general purpose algorithms for the matrix determinant problem. Since the determinant of an SDDM m
corresponds to the determinant of a graph Laplacian with one row/column removed.
e n2 δ−2
Theorem 1.2. Given an SDDM matrix M , there is a routine DetApprox which in O
time outputs D such that D = (1 ± δ) det(M ) with high probability
A crucial thing to note which distinguishes the above guarantee from most other similar
results is that we give a multiplicative approximation of the det(M ). This is much stronger than
giving a multiplicative approximation of log det(M ), which is what other work typically tries to
achieve.
The sparsifiers we construct will also approximately preserve the spanning tree distribution,
which we leverage to yield a faster algorithm for sampling random spanning trees. Our new
algorithm improves upon the current fastest algorithm for general weighted graphs when one
wishes to achieve constant—or slightly sub-constant—total variation distance.
Theorem 1.3. Given an undirected,
weighted graph G = (V, E, w ), there is a routine ApproxTree
2
−2
e
outputs a random spanning tree from a distribution that has
which in expected time O n δ
total variation distance ≤ δ from the w -uniform distribution on G.
3
1.2
1.2.1
Prior Work
Graph Sparsification
In the most general sense, a graph sparsification procedure is a method for taking a potentially
dense graph and returning a sparse graph called a sparsifier that approximately still has many of
the same properties of the original graph. It was introduced in [EGIN97] for preserving properties
related to minimum spanning trees, edge connectivity, and related problems. [BK96] defined the
notion of cut sparsification in which one produces a graph whose cut sizes approximate those
in the original graph. [ST11] defined the more general notion of spectral sparsification which
requires that the two graphs’ Laplacian matrices approximate each other as quadratic forms.1
2 ) edges from the original graph, yielding
e
In particular, this spectral sparsification samples O(n/ǫ
2
e
a graph with O(n/ǫ
) whose quadratic forms—and hence, eigenvalues—approximate each other
within a factor of (1 ± ǫ). This implies that their determinants approximate each other within
(1 ± ǫ)n . This is not useful from the perspective of preserving the determinant: since one would
need to samples Ω(n3 ) edges to get a constant factor approximation, one could instead exactly
compute the determinant or sample spanning trees using exact algorithms with this runtime.
All of the above results on sparsification are for undirected graphs. Recently, [CKP+ 17]
has defined a useful notion of sparsification for directed graphs along with a nearly linear time
algorithm for constructing sparsifiers under this notion of sparsification.
1.2.2
Determinant Estimation
Exactly calculating the the determinant of an arbitrary matrix is known to be equivalent to
matrix multiplication [BS83]. For approximately computing the log of the determinant, [IL11]
uses the identity log(det(A)) = tr(log(B)) + tr(log(B −1 A)) to do this whenever one can find a
matrix B such that the tr(log(B)) = log(det(B)) and tr(log(B −1 A)) = log(det(B −1 A) can both
be quickly approximated.2
For the special case of approximating the log determinant of an SDD matrix, [HAB14b]
applies this same identity recursively where the B matrices are a sequence of ultrasparsifiers
that are inspired by the recursive preconditioning framework of [ST14]. They obtain a running
time of O(m(n−1 ǫ−2 + ǫ−1 )polylog(nκ/ǫ)) for estimating the log determinant to additive error
ǫ.
[BDKZ15b] estimates the log determinant of arbitrary positive definite matrices, but has
runtime that depends linearly on the condition number of the matrix.
In contrast, our work is the first we know of that gives a multiplicative approximation of
the determinant itself, rather than its log. Despite achieving a much stronger approximation
guarantee, our algorithm has essentially the same runtime as that of [HAB14b] when the graph
is dense. Note also that if one wishes to conduct an “apples to apples” comparison by setting
their value of ǫ small enough in order to match our approximation guarantee, their algorithm
would only achieve a runtime bound of O(mnδ−2 polylog(nκ/ǫ)), which is never better than our
runtime and can be as bad as a factor of n worse.3
1
If two graphs Laplacian matrices approximate each other as quadratic forms then their cut sizes also approximate each other.
2
Specifically, they take B as the diagonal of A and prove sufficient conditions for when the log determinant of
−1
B A can be quickly approximated with this choice of B.
3
This simplification of their runtime is using the substitution ǫ = δ/n which gives roughly (1±δ) multiplicative
4
1.2.3
Sampling Spanning Trees
Previous works on sampling random spanning trees are a combination of two ideas: that they
could be generated using random walks, and that they could be mapped from a random integer
via Kirchoff’s matrix tree theorem. The former leads to running times of the form O(nm)
[Bro89, Ald90], while the latter approach[Gue83, Kul90, CMN96, HX16] led to routines that
run in O(nω ) time, where ω ≈ 2.373 is the matrix multiplication exponent [Wil12].
These approaches have been combined in algorithms by Kelner and Madry [KM09] and
Madry, Straszak and Tarnawski [MST15]. These algorithms are based on simulating the walk
more efficiently on parts of the graphs, and combining this with graph decompositions to handle
the more expensive portions of the walks globally. Due to the connection with random-walk
based spanning tree sampling algorithms, these routines often have inherent dependencies on
the edge weights. Furthermore, on dense graphs their running times are still worse than the
matrix-multiplication time routines.
The previous best running time
for generating a random spanning tree from a weighted
e n5/3 m1/3 log2 (1/δ) achieved by [DKP+ 16]. It works by combining a recursive
graph was O
procedure similar to those used in the more recent O(nω ) time algorithms [HX16]
with spectral
5/3
13
2
e
sparsification ideas, achieving a runtime of O(n m ). When m = Θ n , the algorithm in
e n7/3 time to produce a tree from a distribution that is o(1) away from the
[DKP+ 16] takes O
w -uniform distribution, which is slower by nearly a n1/3 factor than the algorithm given in this
paper.
Our algorithm can be viewed as a natural extension of the sparsification0-based approach
from [DKP+ 16]: instead of preserving the probability of a single edge being chosen in a random
spanning tree, we instead aim to preserve the entire distribution over spanning trees, with the
sparsifier itself also considered as a random variable. This allow us to significantly reduce the
sizes of intermediate graphs, but at the cost of a higher total variation distance in the spanning
tree distributions. This characterization of a random spanning tree is not present in any of
the previous works, and we believe it is an interesting direction to combine our sparsification
procedure with the other algorithms.
1.3
Organization
Section 2 will introduce the necessary notation and some of the previously known fundamental
results regarding the mathematical objects that we work with throughout the paper. Section 3
will give a high-level sketch of our primary results and concentration bounds for total tree
weight under specific sampling schemes. Section 4 leverages these concentration bounds to give
a quadratic time sparsification procedure (down to Ω(n1.5 ) edges) for general graphs. Section 5
uses random walk connections to extend our sparsification procedure to the Schur complement
of a graph. Section 6 utilizes the previous routines to achieve a quadratic time algorithm for
computing the determinant of SDDM matrices. Section 7 combines our results and modifies
previously known routines to give a quadratic time algorithm for sampling random spanning
trees with low total variation distance. Section 8 extends our concentration bounds to random
samplings where an arbitrary tree is fixed, and is necessary for the error accounting of our
error in estimating the determinant for their algorithm. This simplification is also assuming δ ≤ 1, which is the
only regime we analyze our algorithm in and thus the only regime in which we can compare the two.
5
random spanning tree sampling algorithm. Section 9 proves the total variation distance bounds
given for our random sampling tree algorithm.
2
Background
2.1
Graphs, Matrices, and Random Spanning Trees
The goal of generating a random spanning tree is to pick tree T with probability proportional
to its weight, which we formalize in the following definition.
Definition 2.1 (w -uniform distribution on trees). Let P rG
T (·) be a probability distribution on
TG such that
Πe∈T0 w e
P rG
.
T (T = T0 ) =
TG
We refer to P rG
T (·) as the w -uniform distribution on the trees of G.
When the graph G is unweighted, this corresponds to the uniform distribution on TG .
We refer to P r G
T (·) as the w -uniform distribution on TG . When the graph G is unweighted,
this corresponds to the uniform distribution on TG . Furthermore, as we will manipulate the
probability of a particular tree being chosen extensively, we will denote such probabilities with
P rG (Tb), aka:
def
P rG Tb = P rG T = Tb .
T
The Laplacian of a graph G = (V, E, w ) is an n × n matrix specified by:
(
du
if u = v
def
Luv =
−w uv if u 6= v
We will write LG when we wish to indicate which graph G that the Laplacian corresponds to
and L when the context is clear. When the graph has multi-edges, we define w uv as the sum of
weights of all the edges e that go between vertices u, v. Laplacians are natural objects to consider
when dealing with random spanning trees due to the matrix tree theorem, which states that
the determinant of L with any row/column corresponding to some vertex removed is the total
weight of spanning trees. We denote this removal of a vertex u as L−u . As the index of vertex
removed does not affect the result, we will usually work with L−n . Furthermore, we will use
det (M ) to denote the determinant of a matrix. As we will work mostly with graph Laplacians,
it is also useful for us to define the ‘positive determinant’ det+ , where we remove the last row
and column. Using this notation, the matrix tree theorem can be stated as:
G
.
TG = det(LG
−n ) = det+ L
We measure the distance between two probability distributions by total variation distance.
Definition 2.2. Given two probability distributions p and q on the same index set Ω, the total
variation distance between p and q is given by
X
def 1
|p(x) − q(x)| .
dT V (p, q) =
2
x∈Ω
6
Let G = (V, E, w ) be a graph and e ∈ E an edge. We write G/e to denote the graph
obtained by contracting the edge e, i.e., identifying the two endpoints of e and deleting any self
loops formed in the resulting graph. We write G\e to denote the graph obtained by deleting the
edge e from G. We extend these definitions to G/F and G\F for F ⊆ E to refer to the graph
obtained by contracting all the edges in F and deleting all the edges in F , respectively.
Also, for a subset of vertices V1 , we use G[V1 ] to denote the graph induced on the vertex of
V1 . letting G(V1 ) be the edges associated with L[V1 ,V1 ] in the Schur complement.
2.2
Effective Resistances and Leverage Scores
The matrix tree theorem also gives connections to another important algebraic quantity: the
def
effective resistance between two vertices. This quantity is formally given as Ref f (u, v) =
⊺
χuv L−1 χuv where χuv is the indicator vector with 1 at u, −1 at v, and 0 everywhere else. Via
the adjugate matrix, it can be shown that the effective resistance of an edge is precisely the ratio
of the number of spanning trees in G/e over the number in G:
Ref f (u, v) =
TG/e
.
TG
As w e · TG/e is the total weight of all trees in G that contain edge e, the fraction4 of spanning
trees that contain e = uv is given by w e Ref f (u, v). This quantity is called the statistical leverage
score of an edge, and we denote it by τ e . It is fundamental component of many randomized
algorithms for sampling / sparsifying graphs and matrices [SS11, Vis12, Tro12].
The fact that τ e is the fraction of trees containing e also gives one way of deriving the sum
of these quantities:
Fact 2.3. (Foster’s Theorem) On any graph G we have
X
τ e = n − 1.
e
The resistance Ref f (u, v), and in turn the statistical leverage scores τ e can be estimated using
linear system solves and random projections [SS11]. For simplicity, we follow the abstraction
utilized by Madry, Straszak, and Tarnawski [MST15], except we also allow the intermediate
linear system solves to utilize a sparsifier instead of the original graph.
Lemma 2.4. (Theorem 2.1. of [MST15])
Let G = (V, E) be a graph with m edges. For every ǫ > 0 we can find in Õ(min{mǫ−2 , m +
−2 log m )
nǫ−4 }) time an embedding of the effective resistance metric into ℜO(ǫ
such that with high
e ef f (u, v) of any effective resistance satisfying
probability allows one to compute an estimate R
∀u, v ∈ V
e ef f (u, v) ≤ Ref f (u, v) ≤ (1 + ǫ) R
e ef f (u, v) .
(1 − ǫ) R
Specifically, each vertex u in this embedding is associated with an (explicitly stored) z u ∈
−2 log m )
e ef f (u, v) is given by:
ℜO(ǫ
, and for any pair of vertices, the estimate R
e ef f (u, v) = kz u − z v k2 ,
R
2
which takes O(ǫ−2 log m) time to compute once we have the embedding.
4
provided one thinks of an edge with weight w as representing w parallel edges, or equivalently, counts spanning
trees with multiplicity according to their weight
7
2.3
Schur Complements
For our applications, we will utilize our determinant-preserving sparsification algorithms in
recursions based on Schur complements. A partition of the vertices, which we will denote
using
V = V1 ⊔ V2 ,
partitions the corresponding graph Laplacian into blocks which we will denote using indices in
the subscripts:
L[V1 ,V1 ] L[V1 ,V2 ]
.
L=
L[V2 ,V1 ] L[V2 ,V2 ]
The Schur complement of G, or L, onto V1 is then:
−1
def
G
G
LG
Sc (G, V1 ) = Sc LG , V1 = LG
[V2 ,V1 ] ,
[V1 ,V1 ] − L[V1 ,V2 ] L[V2 ,V2 ]
and we will use Sc (G, V1 ) and Sc LG , V1 interchangeably. We further note that we will
always consider V1 to be the vertex set we Schur complement onto, and V2 to be the vertex set
we eliminate, except for instances in which we need to consider both Sc (G, V1 ) and Sc (G, V2 ).
Schur complements behave nicely with respect to determinants determinants, which suggests
the general structure of the recursion we will use for estimating the determinant.
Fact 2.5. For any matrix M where M [V2 ,V2 ] is invertible,
det (M −n ) = det M [V2 ,V2 ] · det+ (Sc (M , V1 )).
This relationship also suggests that there should exist a bijection between spanning tree
distribution in G and the product distribution given by sampling spanning trees independently
from Sc (L, V1 ) and the graph Laplacian formed by adding one row/column to L[V2 ,V2 ] .
Finally, our algorithms for approximating Schur complements rely on the fact that they
preserve certain marginal probabilities. The algorithms of [CDN89, CMN96, HX16, DKP+ 16]
also use variants of some of these facts, which are closely related to the preservation of the
spanning tree distribution on Sc (L, V1 ). (See Section 7 for details.)
Fact 2.6. Let V1 be a subset of vertices of a graph G, then for any vertices u, v ∈ V1 , we have:
Sc(G,V1 )
RG
ef f (u, v) = Ref f
(u, v) .
Theorem 2.7 (Burton and Premantle [BP93]). For any set of edges F ⊆ E in a graph G =
(V, E, w ), the probability F is contained in a w -uniform random spanning tree is
P rG
T (F ⊆ T ) = det(M (L,F ) ),
where M (L,F ) is a |F |×|F | matrix whose (e, f )’th entry, for e, f ∈ F, is given by
By a standard property of Schur complements (see [HJ12]), we have
L−1 [V1 , V1 ] = Sc (G, V1 )† .
p
w (e)w (f )χTe L† χf .
Here (L† )[V1 , V1 ] is the minor of L† with row and column indices in V1 . This immediately implies
that when F is incident only on vertices in V1 , we have M (L,F ) = M (Sc(G,V1 ),F ). Putting these
together, we have
8
Fact 2.8. Given a partition of the vertices V = V1 ⊔ V2 . For any set of edges F contained in
G[V1 ], we have
Sc(G,V1 )
(F ⊆ T ).
P rG
T (F ⊆ T ) = P r T
3
Sketch of the Results
The starting point for us is the paper by Janson [Jan94] which gives (among other things) the
limiting distribution of the number of spanning trees in the Gn,m model of random graphs. Our
concentration result for the number of spanning trees in the sparsified graph is inspired by this
paper, and our algorithmic use of this sparsification routine is motivated by sparsification based
algorithms for matrices related to graphs [PS14, CCL+ 15, KLP+ 16]. The key result we will
prove is a concentration bound on the number of spanning trees when the graph is sparsified by
sampling edges with probability approximately proportional to effective resistance.
3.1
Concentration Bound
Let G be a weighted graph with n vertices and m edges, and H be a random subgraph obtained
by choosing a subset of edges of size s uniformly randomly. The probability of a subset of edges,
which could either be a single tree, or the union of several trees, being kept in H can be bounded
precisely. Since we will eventually choose s > n1.5 , we will treat the quantity n3 /s2 as negligible.
The probability of H containing a fixed tree was shown by Janson to be:
Lemma 3.1. If m ≥
s2
n,
then for any tree T , the probability of it being included in H is
2
3
n
n
(s)n−1
n−1
=p
· exp − − O
.
PrH [T ∈ H] =
(m)n−1
2s
s2
where (a)b denotes the product a · (a − 1) · · · (a − (b − 1)).
By linearity of expectation, the expected total weight of spanning trees in H is:
2
3
n
n
EH [TH ] = TG · pn−1 · exp − − O
.
(1)
2s
s2
hP
i
As in [Jan94], the second moment, EH TH2 = EH
w
(T
)w
(T
)P
r
(T
,
T
∈
H)
, can
1
2
1 2
(T1 ,T2 )
be written as a sum over all pairs of trees (T1 , T2 ) . Due to symmetry, the probability of a
particular pair of trees T1 , T2 both being subgraphs of H depends only on the size of their
intersection. The following bound is shown in Appendix A.
Lemma 3.2. Let G be a graph with n vertices and m edges, and H be a uniformly random
2
subset of s > 10n edges chosen from G, where m ≥ sn . Then for any two spanning trees T1 and
T2 of G with |T1 ∩ T2 | = k, we have:
PrH [T1 , T2 ∈ H] ≤ p
2n−2
2n2
exp −
s
where p = s/m.
9
k
1
2n
,
1+
p
s
The crux of the bound on the second moment in Janson’s proof is getting a handle on the
number of tree pairs (T1 , T2 ) with |T1 ∩ T2 | = k in the complete graph where all edges are
symmetric. An alternate way to obtain a bound on the number of spanning trees can also be
obtained using leverage scores, which describe the fraction of spanning trees that utilize a single
edge. A well known fact about random spanning tree distributions [BP93] is that the edges are
negatively correlated:
Fact 3.3 (Negative Correlation). Suppose F is subset of edges in a graph G, then
G
P rG
T (F ⊆ T ) ≤ Πe∈F P r T (e ∈ T ) .
An easy consequence of Fact 3.3 is
Lemma 3.4. For any subset of edges F we have that the total weight of all spanning trees
containing F is given by
X
Y
τ e.
w (T ) ≤ TG
e∈F
T is a spanning tree of G
F ⊆T
The combinatorial view of all edges being interchangable in the complete graph can therefore be replaced with an algebraic view in terms of the leverage scores. Specifically, invoking
n
gives the following lemma
Lemma 3.4 in the case where all edges have leverage score at most m
which is proven in Appendix A.
Lemma 3.5. In a graph G where all edges have leverage scores at most
X
T1 ,T2
|T1 ∩T2 |=k
w (T1 ) · w (T2 ) ≤
TG2
1
·
k!
n2
m
n
m,
we have
k
With Lemma 3.5, we can finally prove the following bound on the second moment which
gives our concentration result.
Lemma 3.6. Let G be a graph on n vertices and m edges such that all edges have statistical
2
n
leverage scores ≤ m
. For a random subset of s > 10n edges, H, where m ≥ sn we have:
3
3
2
n
n
n
2
=
E
[T
]
exp
O
.
EH TH2 ≤ TG2 p2n−2 exp − + O
H H
s
s2
s2
Proof. By definition of the second moment, we have:
X
w (T1 ) · w (T2 ) · PrH [T1 ∪ T2 ⊆ H] .
EH TH2 =
T1 ,T2
Re-writing the above sum in terms of the size of the intersection k, and invoking Lemma 3.2
gives:
EH TH2
≤
n−1
X
k=0
X
T1 ,T2
|T1 ∩T2 |=k
w (T1 ) · w (T2 ) · p
10
2n−2
2n2
exp −
s
k
1
2n
.
1+
p
s
Note that the trailing term only depends on k and can be pulled outside the summation of
T1 , T2 , so we then use Lemma 3.5 to bound this by:
EH TH2
≤
n−1
X
k=0
TG2
1
·
k!
n2
m
k
·p
2n−2
2n2
exp −
s
k
1
2n
.
1+
p
s
Which upon pulling out the terms that are independent of k, and substituting in p = s/m gives:
EH TH2
≤
TG2
·p
2n−2
2n2
· exp −
s
n−1
k
X 1 n2
2n
·
.
·
1+
k!
s
s
k=0
From the Taylor expansion of exp(·), we have:
2
n
2n
2n2
· exp
1+
EH TH2 ≤ TG2 · p2n−2 · exp −
s
s
s
2
3
n
n
= TG2 · p2n−2 · exp −
.
· exp O
s
s2
This bound implies that once we set s2 > n3 , the variance becomes less than the square of the
expectation. It forms the basis of our key concentration results, which we show in Section 4, and
also leads to Theorem 1.1. In particular, we demonstrate that this sampling scheme extends to
importance sampling, where edges are picked with probabilities proportional to (approximations
of) of their leverage scores.
A somewhat surprising aspect of this concentration result is that there is a difference between
models Gn,m and the Erdos-Renyi model Gn,p when the quantity of interest is the number of
spanning trees. In particular, the number of spanning trees of a graph G ∼ Gn,m is approximately
normally distributed when m = ω n1.5 , whereas it has approximate log-normal distribution
when G ∼ Gn,p and p < 1.
An immediate consequence of this is that we can now approximate det+ (LG ) by computing
det+ (LH ). It also becomes natural to consider speedups of random spanning tree sampling
algorithms that generate a spanning tree from a sparsifier. Note however that we cannot hope
to preserve the distribution over all spanning trees via a single sparsifier, as some of the edges
are no longer present.
To account for this change in support, we instead consider the randomness used in generating
the sparsifier as also part of the randomness needed to produce spanning trees. In Section 9.1,
we show that just bounds on the variance of TH suffices for a bound on the TV distances of the
trees.
Lemma 3.7. Suppose H is a distribution over rescaled subgraphs of G such that for some
parameter some 0 < δ < 1 we have
EH∼H TH2
≤ 1 + δ,
EH∼H [TH ]2
11
and for any tree T̂ and any graph from the distribution that contain it, H we have:
h
i−1 E ′
H ∼H [TH ′ ]
,
w H Tb = w G Tb · PrH ′ ∼H Tb ⊆ H ′
·
TG
then the distribution given by P rG (T ), p, and the distribution induced by EH∼H P rH (T ) , p̃
satisfies
√
dT V (p, p̃) ≤ δ.
Note that uniform sampling meets the property about w H (T ) because of linearity of expectation. We can also check that the importance sampling based routine that we will discuss in Section 4.2 also meets this criteria. Combining this with the running time bounds from Theorem 1.1,
as well as the Õ(m1/3 n5/3 ) time random spanning tree sampling algorithm from [DKP+ 16] then
leads to a faster algorithm.
Corollary 3.8. For any graph G on n vertices and any δ > 0, there is an algorithm that
generates a tree from a distribution whose total variation is at most δ from the random tree
13
distribution of G in time Õ(n 6 =2.1666... δ−2/3 + n2 δ−2 ).
3.2
Integration Into Recursive Algorithms
As a one-step invocation of our concentration bound leads to speedups over previous routines,
we investigate tighter integrations of the sparsification routine into algorithms. In particular, the
sparsified Schur complement algorithms [KLP+ 16] provide a natural place to substitute spectral
sparsifiers with determinant-preserving ones. In particular, the identity of
det+ (L) = det (L[V2 ,V2 ] ) · det+ (Sc (L, V1 )).
where det+ is the determinant of the matrix minor, suggests that we can approximate det (L−n )
by approximating det (L[V2 ,V2 ] ) and det+ (Sc (L, V1 )) instead. Both of these subproblems are
smaller by a constant factor, and we also have |V1 | + |V2 | = n. So this leads to a recursive
scheme where the total number of vertices involved at all layers is O(n log n). This type of
recursion underlies both our determinant estimation and spanning tree sampling algorithms.
The main difficulty remaining for the determinant estimation algorithm is then sparsifying
Sc (G, V1 ) while preserving its determinant. For this, we note that some V1 are significantly
easier than others: in particular, when V2 = V \ V1 is an independent set, the Schur complement
of each of the vertices in V2 can be computed independently. Furthermore, it is well understood
how to sample these complements, which are weighted cliques, by a distribution that exceeds
their true leverage scores.
Lemma 3.9. There is a procedure that takes a graph G with n vertices, a parameter δ, and
produces in Õ(n2 δ−1 ) time a subset of vertices V1 with |V1 | = Θ(n), along with a graph H V1 such
that
TSc(G,V1 ) exp (−δ) ≤ EH V1 [TH V1 ] ≤ TSc(G,V1 ) exp (δ) ,
and
EH V1 TH2 V1
EH V1 [TH V1 ]2
12
≤ exp (δ) .
Lemma 2.4 holds w.h.p., and we condition on this event. In our algorithmic applications we
will be able to add the polynomially small failure probability of Lemma 2.4 to the error bounds.
The bound on variance implies that the number of spanning trees is concentrated close to its
expectation, TSc(G,V1 ) , and that a random spanning tree drawn from the generated graph H V1
is —over the randomness of the sparsification procedure—close in total variation distance to a
random spanning tree of the true Schur complement.
As a result, we can design schemes that:
1. Finds an O(1)-DD subset V2 , and set V1 ← V \ V2 .
2. Produce a determinant-preserving sparsifier H V1 for Sc (G, V1 ).
3. Recurse on both L[V2 ,V2 ] and H V1 .
However, in this case, the accumulation of error is too rapid for yielding a good approximation
of determinants. Instead, it becomes necesary to track the accumulation of variance during all
recursive calls. Formally, the cost of sparsifying so that the variance is at most δ is about n2 δ−1 ,
where δ is the size of the problem. This means that for a problem on Gi of size βi n for 0 ≤ βi ≤ 1,
we can afford an error of βi δ when working with it, since:
1. The sum of βi on any layer is at most 2,
5
so the sum of variance per layer is O(δ).
2. The cost of each sparsification step is now βi n2 δ−1 , which sums to about n2 δ−1 per layer.
Our random spanning tree sampling algorithm in Section 7 is similarly based on this careful
accounting of variance. We first modify the recursive Schur complement algorithm introduced
by Coulburn et al. [CDN89] to give a simpler algorithm that only braches two ways at each
step in Section 7.1, leading to a high level scheme fairly similar to the recursive determinant
algorithm. Despite these similarities, the accumulation of errors becomes far more involved
here due to the choice of trees in earlier recursive calls affecting the graph in later steps. More
specifically, the recursive structure of our determinant algorithm can be considered analogous to
a breadth-first-search, which allows us to consider all subgraphs at each layer to be independent.
In contrast, the recursive structure of our random spanning tree algorithm, which we show in
Section 7.2 is more analogous to a depth-first traversal of the tree, where the output solution of
one subproblem will affect the input of all subsequent subproblems.
These dependency issues will be the key difficulty in considering variance across levels. The
total variation distance tracks the discrepancy over all trees of G between their probability
of being returned by the overall recursive algorithm, and their probability in the w -uniform
distribution. Accounting for this over all trees leads us to bounding variances in the probabilities
of individual trees being picked. As this is, in turn, is equivalent to the weight of the tree divided
by the determinant of the graph, the inverse of the probability of a tree being picked can play
a simliar role to the determinant in the determinant sparsification algorithm described above.
However, tracking this value requires analyzing extending our concentration bounds to the case
where an arbitrary tree is fixed in the graph and we sample from the remaining edges. We
study this Section 8, prove bounds analogous to the concentration bounds from Section 4, and
incorporate the guarantees back into the recursive algorithm in Section 7.2.
5
each recursive call may introduce one new vertex
13
4
Determinant Preserving Sparsification
In this section we will ultimately prove Theorem 1.1, our primary result regarding determinantpreserving sparsification. However, most of this section will be devoted to proving the following
general determinant-preserving sparsification routine that also forms the core of subsequent
algorithms:
Theorem 4.1. Given an undirected, weighted graph G = (V, E, w ), an error threshold ǫ > 0,
parameter ρ along with routines:
P
1. SampleEdgeG () that samples an edge e from a probability distribution p ( e p e = 1),
as well as returning the corresponding value of p e . Here p e must satisfy:
τe
≤ ρ · pe
n−1
where τ e is the true leverage score of e in G.
2. ApproxLeverageG (u, v, ǫ) that returns the leverage score of an edge u, v in G to an error
of ǫ. Specifically, given an edge e, it returns a value τee such that:
(1 − ǫ) τ e ≤ τee ≤ (1 + ǫ) τ e .
There is a routine DetSparsify(G, s, ǫ) that computes a graph H with s edges such that its tree
count, TH , satisfies:
3
n
,
EH [TH ] = TG 1 ± O
s2
and:
EH TH2
EH [TH ]2
≤ exp
ǫ2 n 2
+O
s
n3
s2
Furthermore, the expected running time is bounded by:
1. O(s · ρ) calls to SampleEdgeG (e) and ApproxLeverage(e) with constant error,
2. O(s) calls to ApproxLeverage(e) with ǫ error.
We establish guarantees for this algorithm using the following steps:
1. Showing that the concentration bounds as sketched in Section 3 holds for approximate
leverage scores in Section 4.1.
2. Show via taking the limit of probabilistic processes that the analog of this process works
for sampling a general graph where edges can have varying leverage scores. This proof is
in Section 4.2.
3. Show via rejection sampling that (high error) one sided bounds on statistical leverage
scores, such as those that suffice for spectral sparsification, can also be to do the initial
round of sampling instead of two-sided approximations of leverage scores. This, as well as
pseudocode and guarantees of the overall algorithm are given in Section 4.3.
14
4.1
Concentration Bound with Approximately Uniform Leverage Scores
Similar to the simplified proof as outlined in Section 3, our proofs relied on uniformly sampling
2
s edges from a multi-graph with m ≥ sn edges, such that all edges have leverage score within
multiplicative 1 ± ǫ of n−1
s , aka. approximately uniform. The bound that we prove is an analog
of Lemma 3.6
2
Lemma 4.2. Given a weighted multi-graph G such that m ≥ sn , s ≥ n, and all edges e ∈ E
≤ τ e ≤ (1+ǫ)(n−1)
, with 0 ≤ ǫ < 1, then
have (1−ǫ)(n−1)
m
m
3
2 2
EH TH2
n
n ǫ
+O
2 ≤ exp
s
s2
EH [TH ]
Similar to the proof of Lemma 3.6 in Section 3, we can utilize the bounds on the probability
of k edges being chosen using Lemma 3.2. The only assumption that changed was the bounds
on τ e , which does not affect EH [TH ]2 . The only term that changes is our upper bound the total
weight of trees that contain some subset of k edges that was the produce of k leverage scores. At
a glance, this product can change by a factor of up to (1 + ǫ)k , which when substituted naively
into the proof of Lemma 3.2 directly would yield an additional term of
2
n ǫ
,
exp
s
and in turn necessitating ǫ < n−1/2 for a sample count of s ≈ n1.5 .
However, note that this is the worst case distortion over a subset F . The upper bound that
we use, Lemma 3.5 sums over these bounds over all subsets, and over all edges we still have
X
e ∈ Gτ e = n − 1.
Incorporating this allows us to show a tighter bound that depends on ǫ2 .
m
Similar to the proof of Lemma 3.5, we can regroup the summation
over
all
subsets
of
k
P
Q
E(G), and bound the fraction of trees containing each subset F via T :F ⊆T w (T ) ≤ TG e∈F τ e
via Lemma 3.4.
Y
X
X
τ 2e
w (T1 ) · w (T2 ) ≤
TG2
F ⊆E
|F |=k
T1 ,T2
|T1 ∩T2 |=k
e∈F
P
The proof will heavily utilize the fact that e∈E τ e = n − 1. We bound this in first two
steps: first treat it as a symmetric product over τ 2e , and bound the total as a function of
X
τ 2e ,
e
P
then we bound this sum using the fact that e τ e = n − 1.
The first step utilizes the concavity of the product function, and bound the total by the sum:
P
Lemma 4.3. For any set of non-negative values x 1 . . . x m with i x i ≤ z, we have
X Y
m
z k
.
xi ≤
m
k
F ⊆[1...m] i∈F
|F |=k
15
z
Proof. We claim that this sum is maximized when x i = m
for all e.
Consider fixing all variables other than some x i and x j , which we assume to be x 1 ≤ x 2
without loss of generality as the function is symmetric on all variables:
X
Y
F ⊆[1...m] i∈F
|F |=k
X
x i = x 1x 2
Y
F ⊆[3...m] i∈F
|F |=k−2
X
x i
+ (x 1 + x 2 ) ·
Y
F ⊆[3...m] i∈F
|F |=k−1
x i
+
X
Y
x i.
F ⊆[3...m] i∈F
|F |=k
Then if x 1 < x 2 , locally changing their values to x 1 + ǫ and x 2 − ǫ keeps the second term
the same. While the first term becomes
(x 1 + ǫ) (x 2 − ǫ) = x 1 x 2 + ǫ (x 2 − x 1 ) − ǫ2 ,
which is greater than x 1 x 2 when 0 < ǫ < (x 2 − x 1 ).
This shows that the overall summation is maximized when all x i are equal, aka
z
,
m
xi =
which upon substitution gives the result.
The second step is in fact the k = 1 case of Lemma 3.5.
Lemma 4.4. For any set of values y e such that
X
y = n − 1,
e
and
(1 + ǫ) n
(1 − ǫ) n
≤ ye ≤
,
m
m
we have
X
e
y 2e ≤
(1 + ǫ2 )(n − 1)2
.
m
Proof. Note that for any a ≤ b, and any ǫ, we have
(a − ǫ)2 + (b + ǫ)2 = a2 + b2 + 2ǫ2 + 2ǫ (b − a) ,
and this transformation must increase the sum for ǫ > 0. This means the sum is maximized
and the other half are (1+ǫ)(n−1)
. This then gives
when half of the leverage scores are (1−ǫ)(n−1)
m
m
X
e∈E
y 2e
m
≤
2
(1 + ǫ)(n − 1)
m
2
m
+
2
(1 − ǫ)(n − 1)
m
16
2
=
(1 + ǫ2 )(n − 1)2
.
m
Proof. (of Lemma 4.2)
We first derive an analog of Lemma 3.5 for bounding the total weights of pairs of trees
containing subsets of size k, where we again start with the bounds
X
T1 ,T2
|T1 ∩T2 |=k
X
w (T1 ) · w (T2 ) ≤
X
F ⊆E T1 ,T2
|F |=k F ⊆T1 ∩T2
w (T1 ) · w (T2 ) =
X
F ⊆E
|F |=k
X
T :F ⊆T
2
w (T )
Applying Lemma 3.4 to the inner term of the summation then gives
X
X
Y
w (T1 ) · w (T2 ) ≤
TG2 ·
τ 2e
F ⊆E
|F |=k
T1 ,T2
|T1 ∩T2 |=k
The bounds on τ e and
P
e τe
e∈F
= n − 1 gives, via Lemma 4.4
X
e
τ 2e ≤
(1 + ǫ2 )(n − 1)2
.
m
Substituting this into Lemma 4.3 with x i = τ 2e then gives
X Y
τ 2e
F ⊆E e∈F
|F |=k
k
k
k
mk (1 + ǫ2 )n2
1 (1 + ǫ2 )n2
m
(1 + ǫ2 )n2
≤
=
.
≤
m2
k!
m2
k!
m
k
which implies our analog of Lemma 3.5
X
T1 ,T2
|T1 ∩T2 |=k
w (T1 ) · w (T2 ) ≤
TG2
1
·
k!
(1 + ǫ2 )n2
m
k
.
We can then duplicate the proof of Lemma 3.6. Similar to that proof, we can regroup the
summation by k = |T1 ∩ T2 | and invoking Lemma 3.2 to get:
EH TH2
≤
n−1
X
k=0
X
T1 ,T2
|T1 ∩T2 |=k
w (T1 ) · w (T2 ) · p
2n−2
2n2
exp −
s
k
1
2n
.
1+
p
s
where p = s/m. When incorporated with our analog of Lemma 3.5 gives:
EH TH2
≤
n−1
X
k=0
p
2n−2
k
k
1
(1 + ǫ2 )n2
2n
2 1
· TG
1+
p
s
k!
m
n−1
k
X 1 (1 + ǫ2 )n2
2n
2n2
2 2n−2
.
·
·
1+
= TG p
· exp −
s
k!
s
s
2n2
exp −
s
k=0
17
P
zk
k k!
≤ exp(z) then leaves us with:
3
2
2
n
n 2 ǫ2
n
2
2n−2
+O
EH TH ≤ TG · p
· exp − +
s
s
s2
Substituting in the Taylor expansion of
and finishes the proof.
4.2
Generalization to Graphs with Arbitrary Leverage Score Distributions
2
The first condition of m ≥ sn will be easily achieved by splitting each edge a sufficient number
of times, which does not need to be done explicitly in the sparsification algorithm. Furthermore,
from the definition of statistical leverage score splitting an edge into k copies will give each
copy a kth fraction of the edge’s leverage score. Careful splitting can then ensure the second
condition, but will require ǫ-approximate leverage score estimates on the edges. The simple
approach would compute this for all edges, then split each edge according to this estimate and
draw from the resulting edge set. Instead, we only utilize this algorithm as a proof technique,
and give a sampling scheme that’s equivalent to this algorithm’s limiting behavior as m → ∞.
Pseudocode of this routine is in Algorithm 1.
Algorithm 1: IdealSparsify(G, τe, s): Sample s (multi) edges of G to produce H such
that TG ≈ TH .
Input: Graph G, approximate leverage scores τe, sample count s
1 Initialize H as the empty graph, H ← ∅;
2 for i = 1 . . . s do
3
Pick edge e with probability proportional to τee ;
4
Add e to H with new weight:
5
Output H
w e (n − 1)
exp
τee s
n2
2 (n − 1) s
.
Note that this sampling scheme is with replacement: the probability of a ‘collision’ as the
number of copies tend to ∞ is sufficiently small that it can be covered by the proof as well.
The guarantee that we will show for Algorithm 1 is:
Lemma 4.5. For any graph G and any set of approximate leverage scores τe such that
(1 − ǫ) τ e ≤ τee ≤ (1 + ǫ) τ e
for all edges e. The graph H = IdealSparsify(G, τe, s) satisfies:
3
n
TG ≤ EH [TH ] ≤ TG ,
1−O
s2
and
EH TH2
2 2
ǫ n
n3
≤ exp O
+ 2
.
s
s
EH [TH ]2
18
Our proof strategy is simple: claim that this algorithm is statistically close to simulating
splitting each edge into a very large number of copies. Note that these proofs are purely for
showing the convergence of statistical processes, so all that’s needed is for the numbers that
arise in this proof (in particular, m) to be finite.
We first show that G and τe can be perturbed to become rational numbers.
(G)
(G)
Lemma 4.6. For any graph G and any set of τe such that (1 − ǫ)τ e ≤ τee ≤ (1 + ǫ)τ e for
all edges e for some constant ǫ > 0, and any perturbation threshold δ, we can find graph G′ with
all edge weights rationals, and τe′ with all entries rational numbers such that:
1. TG ≤ TG′ ≤ (1 + δ)TG , and
(G′ )
2. (1 − 2ǫ)τ e
(G′ )
≤ τe′e ≤ (1 + 2ǫ)τ e
for all edges e.
Proof. This is a direct consequence of the rational numbers being everywhere dense, and that
perturbing edge weights by a factor of 1±α perturbs leverage scores by a factor of up to 1±O(α),
and total weights of trees by a factor of (1 ± α)n−1 .
Having all leverage scores as integers means that we can do an exact splitting by setting m,
the total number of split edges, to a multiple of the common denominator of all the τe′e values
times n − 1. Specifically, an edge with approximate leverage score τe′e becomes
τe′e ·
copies, each with weight
m
n−1
w e (n − 1)
,
τe′e m
and ‘true’ leverage score
τ e (n − 1)
.
τe′e m
In particular, since
(1 − 2ǫ) ≤
τe
≤ (1 + 2ǫ) ,
τee
this splitted graph satisfies the condition of Lemma 4.2. This then enables us to obtain the
guarantees of Lemma 4.5 by once again letting m tend to ∞.
Proof. (of Lemma 4.5) We first show that Algorithm 1 works for the graph with rational weights
and approximate leverage scores as generated by Lemma 4.6.
The condition established above means that we can apply Lemma 4.2 to the output of picking
s random edges among these m split copies. This graph H ′ satisfies
2
3
s n−1
n
n
exp − − O
EH ′ [TH ′ ] = TG′
,
m
2s
s2
and
EH ′ TH2 ′
EH ′ [TH ′ ]2
≤ exp
n 2 ǫ2
+O
s
19
n3
s2
.
The ratio of the second moment is not affected by rescaling, so the graph
m
n2
H ′′ ←
exp
s
2s (n − 1)
meets the requirements on both the expectation and variances. Furthermore, the rescaled weight
of an single edge being picked is:
n2
w e (n − 1) m
n2
w e (n − 1)
· exp
exp
=
,
s
2s (n − 1)
2s (n − 1)
τe′e m
τe′e s
which is exactly what Algorithm 1 assigns.
It remains to resolve the discrepancy between sampling with and without replacement: the
probability of the same edge being picked twice in two different steps is at most 1/m, so the
total probability of a duplicate sample is bounded by s2 /m. We then give a finite bound on the
size of m for which this probability becomes negligible in our routine. The rescaling factor of a
single edge is (very crudely) bounded by
1
n2
(n − 1)
,
≤ exp n3
exp
′
2s (n − 1)
mine τe′e
τee s
which means that any of the H ′′ returned must satisfy
n
1
4
TH ′′ ≤ exp n
T G′ ,
mine τe′e
which is finite. As a result, as m → ∞, the difference that this causes to both the first and
second moments become negligible.
The result for H ← IdealSparsify(G, τe, s) then follows from the infinitesimal perturbation
made to G, as the rational numbers are dense everywhere.
4.3
Incorporating Crude Edge Sampler Using Rejection Sampling
Under Lemma 4.5 we assumed access to ǫ-approximate leverage scores, which could be computed
with m calls to our assumed subroutine ApproxLeverageG , where m here is the number of
edges of G. However, we roughly associate ApproxLeverageG with Lemma 2.4 that requires
Õ(ǫ−2 ) time per call (and we deal with the w.h.p. aspect in the proof of Theorem 1.1), and
to achieve our desired sparsification of O(n1.5 ) edges, we will need ǫ = n−1/4 for the necessary
concentration bounds. Instead, we will show that we can use rejection sampling to take s edges
drawn from approximate leverage scores using a cruder distribution p e , which will only require
application of ApproxLeverageG with error ǫ for an expected O(s) number of edges.
Rejection sampling is a known technique that allows us to sample from some distribution f
by instead sampling from a distribution g that approximates f and accept the sample with a
specific probability based on the probability of drawing that sample from f and g.
More specifically, suppose we are given two probability distributions f and g over the same
state space X, such that for all x ∈ X we have Cg(x) ≥ f (x) for some constant C. Then we
f (x)
.
can draw from f by instead drawing x ∼ g, and accepting the draw with probability Cg(x)
20
This procedure only requires a lower bound on g with respect to f , but in order to accept a
draw with constant probability, there need to be weaker upper bound guarantees. Our guarantees
on τee will fulfill these requirements, and the rejection sampling will accept a constant fraction
of the draws. By splitting into a sufficient number of edges, we ensure that drawing the same
multi-edge from any split edge will occur with at most constant probability.
Specifically, each sample is drawn via. the following steps:
1. Draw a sample according the distribution g, e.
2. Evaluate the values of f (e) and g(e).
3. Keep the sample with probability f (e)/g(e).
As the running time of ApproxLeverageG (e, ǫ) will ultimately depend on the value of ǫ apply this algorithmic framework, we also need to perform rejection sampling twice, once with
constant error, and once with leverage scores extracted from the true approximate distribution.
Pseudocode of this routine is shown in Algorithm 2.
Algorithm 2: DetSparsify(G, s, SampleEdgeG ()), ρ, ApproxLeverageG (u, v, ǫ)):
Sample s (multi) edges of G to produce H such that TG ≈ TH .
Input: Graph G.
Sample count s, leverage score approximation error 0 < ǫ < 1/2,
P
SampleEdgeG () that samples an edge e from a probability distribution p ( e p e = 1),
and returning the corresponding value of p e .
ρ that bounds the under-sampling rate of SampleEdgeG ().
ApproxLeverageG (u, v, ǫ) that returns the approximate leverage score of an edge u, v
in G to an error of ǫ.
1 Initialize H as the empty graph, H ← ∅;
2 while H has fewer than s edges do
3
e, p e ← SampleEdgeG ().
2
ApproxLeverageG (u, v, 0.1)
4
Let p ′e ← n−1
5
Reject e with probability 1 − p ′e /(4ρ · p e ).
1
ApproxLeverageG (u, v, ǫ)
6
Let p ′′e ← n−1
7
Reject e with probability 1 − p ′′e /p ′e .
8
Add e to H with new weight
we
exp
p ′′e s
9
n2
2(n − 1)s
.
Output H
We first show that this routine will in fact sample edges according to ǫ-approximate leverage
scores, as was assumed in IdealSparsify
Lemma 4.7. The edges are being sampled with probability proportional to τe(G,ǫ) , the leverage
score estimates given by ApproxLeverageG (·, ǫ).
21
Note that this algorithm does not, at any time, have access to the full distribution τe(G,ǫ) .
Proof. Our proof will assume the known guarantees of rejection sampling, which is to say that
the following are true:
1. Given distributions p and p ′ , sampling an edge e from p and accepting with probability
p ′e /(4ρ · p e ) is equivalent to drawing an edge from p ′ as long as p ′e /(4ρ · p e ) ∈ [0, 1] for all
e.
2. Given distributions p ′ and p ′′ , sampling an edge e from p ′ and accepting with probability
p ′′e /p ′e is equivalent to drawing an edge from p ′′ as long as p ′′e /p ′e ∈ [0, 1] for all e.
As a result, we only need to check that p ′e /(4ρp e ) and p ′′e /p ′e are at most 1.
The guarantees of SampleEdgeG () gives
τe
≤ ρp e .
n−1
As p ′e was generated with error 1.1, we have
p ′e ≤
2.2τ e
≤ 2.2ρp e ,
(n − 1)
so p ′e /(4ρp e ) ≤ 1. To show p ′′e /p ′e ≤ 1, once again the guarantees of SampleEdgeG () gives:
p ′′e ≤ (1 + ǫ)
τe
τe
≤ 2 · 0.9
≤ p ′e .
n−1
n−1
It remains to show that this rejection sampling process still makes sufficiently progress, yet
also does not call ApproxLeverageG (e, ǫ) (the more accurate leverage score estimator) too
many times.
Lemma 4.8. At each step, the probability of DetSparsify calling ApproxLeverageG (e, ǫ)
1
is at most ρ1 , while the probability of it adding an edge to H is at least 8ρ
.
P
Proof. The proof utilizes the fact e τ e = n − 1 (Fact 2.3) extensively.
If the edge e is picked, ApproxLeverageG (e, ǫ) is called with probability
2.2τ e
p ′e
≤
4ρ · p e
4ρ · p e · (n − 1)
Summing over this over all edge e by the probability of picking them gives:
P
X
2.2 e τ e
2.2τ e
1
pe
=
≤ .
4ρ · p e · (n − 1)
4ρ · (n − 1)
ρ
e
On the other hand, the probability of picking edge e, and not rejecting it is:
pe ·
p ′′
τe(G,ǫ)
p ′e
· e′ =
.
4ρ · p e p e
4ρ(n − 1)
22
where this follows by cancellation and how we set p ′′e in our algorithm. Summing over all edges
then gives the probability of not rejecting an edge to be
X
e
P
X (1 − ǫ)τ e
(1 − ǫ) e τ e
τe(G,ǫ)
1
≥
=
≥
4ρ(n − 1)
4ρ(n
−
1)
4ρ(n
−
1)
8ρ
e
Proof. (of Theorem 4.1) Lemma 4.7 implies that edges are sampled in DetSparsify with probability proportional to ǫ-approximate leverage scores guaranteed by ApproxLeverageG (·, ǫ).
Therefore, we can apply Lemma 4.5 to achieve the desired expectation and concentration bounds.
Finally, Lemma 4.8 implies that we expect to sample at most O(s · ρ) edges, each of which require a call to SampleEdgeG (e) and ApproxLeverageG with constant error. It additionally
implies that we expect to make O(s) calls to ApproxLeverageG with ǫ error.
Directly invoking this theorem leads to the sparsification algorithm.
Proof. (of Theorem 1.1) Consider invoking Theorem 4.1 with parameters
s ← O n1.5 δ−2 ,
ǫ ← n−1/4 .
This gives:
ǫ2 n 2 n 3
,
≤ δ,
s s2
which then implies
1 − O δ2
and
TG ≤ EH [TH ] ≤ 1 + O δ2
TG ,
EH TH2 ≤ 1 + O δ2 EH [TH ]2 .
The second condition is equivalent to VarH [TH ] ≤ δ2 EH [TH ], which by Chebyshev inequality
gives that with constant probability we have
(1 − O (δ)) TG ≤ TH ≤ (1 + O (δ)) TG .
Combining this with the bounds on EH [TH ], and adjusting constants gives the overall bound.
Constructing the probability distribution p for sampling edges only requires computing constant approximate leverage scores for all edges, and then sampling proportionally for each edge,
giving a constant value for ρ. By Lemma 2.4, this requires Õ(m) time. The running time then
is dominated by the O(s) calls made to the effective resistance oracle with error ǫ = n−1/4 .
Invoking Lemma 2.4 gives that this cost is bounded by
O nǫ−4 + sǫ−2 = O n2 δ−2 .
Furthermore, because Lemma 2.4 holds w.h.p. we can absorb the probability of failure into
our constant probability bound
23
Another immediate consequence of our sparsification routine in Theorem 4.1, along with
bounds on total variation distance that we prove in Section 9, is that we can give a faster
spanning tree sampling algorithm for dense graphs by plugging the sparsified graph into previous
algorithms for generating random spanning trees.
Proof. (of Corollary 3.8) As in the proof of Theorem 1.1, we invoke Theorem 4.1 with parameters
s ← O n1.5 δ−2 ,
ǫ ← n−1/4 .
giving
EH TH2
EH [TH ]2
≤ 1 + δ2 .
Applying Lemma 3.7, which is proven in Section 9.1, we then have that drawing a tree from H
according to the w -uniform distribution gives a total variation distance of δ from drawing a tree
according to the w -uniform distribution of G. The running time of drawing H is dominated by
the O(s) calls made to the effective resistance oracle with error ǫ = n−1/4 . Invoking Lemma 2.4
gives that this cost is bounded by
O nǫ−4 + sǫ−2 = O n2 δ−2 .
Furthermore, because Lemma 2.4 holds w.h.p. we can absorb the probability of failure into
our total variation distance bound (where we implicitly assume that δ is at most polynomially
small).
We then use the Õ(m1/3 n5/3 ) time algorithm in [DKP+ 16] with m = O(n1.5 δ−2 ) to draw a
tree from H. This then achieves our desired running time and total variation distance bound.
5
Implicit Sparsification of the Schur Complement
Note that the determinant sparsification routine in Theorem 4.1 only requires an oracle that
samples edges by an approximate distribution to resistance, as well as access to approximate
leverage scores on the graph. This suggests that a variety of naturally dense objects, such as
random walk matrices [CCL+ 15, JKPS17] and Schur complements [KLP+ 16, DKP+ 16] can also
be sparsified in ways that preserve the determinant (of the minor with one vertex removed)
or the spanning tree distributions. The latter objects, Schur complements, have already been
shown to lead to speedups in random spanning tree generation algorithms recently [DKP+ 16].
Furthermore the fact that Schur complements preserve effective resistances exactly (2.6)
means that we can directly invoke the effective resistances data structure as constructed in
Lemma 2.4 to produce effective resistance estimates on any of its Schur complements. As a result,
the main focus of this section is an efficient way of producing samples from a distribution that
approximates drawing a multi-edge from the Schur complement with probabilities proportional
to its leverage score. Here we follow the template introduced in [KLP+ 16] of only eliminating
(1 + α)-diagonally-dominant subsets of vertices, as it in turn allows the use of walk sampling
based implicit sparsification similar to those in [CCL+ 15, JKPS17].
24
(1 + α)-diagonally-dominant subsets have been used in Schur complement based linear system solvers to facilitate the convergence of iterative methods in the L[V2 ,V2 ] block [KLP+ 16].
Formally, the condition that we require is:
Definition 5.1. In a weighted graph G = (V, E, w ), a subset of vertices V2 ⊆ V is (1 + α)diagonally-dominant, or (1 + α)-DD if for every u ∈ V2 with weighted degree d u we have:
X
1
1 X
w uv ≥
w uv .
du =
1+α
1 + α v∼u
v∼u,v∈V
/ 2
It was shown in [KLP+ 16] that large sets of such vertices can be found by trimming a
uniformly random sample.
Lemma 5.2. (Lemma 3.5. of [KLP+ 16] instantiated on graphs)
There is a routine AlmostIndependent(G, α) that for a graph G with n vertices, and a parameter α ≥ 0, returns in O(m) expected time a subset V2 with |V2 | ≥ n/(8(1 + α)) such that
LG,[V2 ,V2 ] is (1 + α)-DD.
Given such a subset V2 , we then proceed to sample edges in Sc (G, V1 ) via the following
simple random walk sampling algorithm:
1. Pick a random edge in G.
2. Extend both of its endpoints in random walks until they first reach somewhere in V1 .
Incorporating this scheme into the determinant preserving sparsification schemes then leads
these guarantees:
Theorem 5.3. Conditioned on Lemma 2.4 holding, there is a procedure SchurSparse that
takes a graph G, and an 1.1-DD subset of vertices V2 , returns a graph H V1 in Õ(n2 δ−1 ) expected
time such that the distribution over H V1 satisfies:
TSc(G,V1 ) exp (−δ) ≤ EH V1 [TH V1 ] ≤ TSc(G,V1 ) exp (δ) ,
and
EH V1 TH2 V1
EH V1 [TH V1 ]2
≤ exp (δ) .
Furthermore, the number of edges of H V1 can be set to anywhere between O(n1.5 δ−1 ) and
O(n2 δ−1 ) without affecting the final bound.
We let this subset of vertices produced to be V2 , and let its complement be V1 . Our key idea
is to view Sc (G, V1 ) as a multi-graph where each multi-edge corresponds to a walk in G that
starts and ends in V1 , but has all intermediate vertices in V2 . Specifically a length k walk
u0 , u1 , . . . uk ,
with u0 , uk ∈ V1 and ui ∈ V2 for all 0 < i < k, corresponds to a multi-edge between u0 and uk
in Sc (G, V1 ) with weight given by
Q
G
0≤i<k w ui ui+1
Sc(G,V1 ) def
.
(2)
w u0 ,u1 ,...uk = Q
G
0<i<k d ui
25
We check formally that this multi-graph defined on V1 is exactly the same as Sc (G, V1 ) via the
Taylor expansion of L−1
[V2 ,V2 ] based Jacobi iteration.
Lemma 5.4. Given a graph G and a partition of its vertices into V1 and V2 , the graph GV1
formed by all the multi-edges corresponding to walks starting and ending at V1 , but stays entirely
within V2 with weights given by Equation 2 is exactly Sc (G, V1 ).
Proof. Consider the Schur complement:
Sc (G, V1 ) = L[V1 ,V1 ] − L[V2 ,V1 ] L†[V2 ,V2 ] L[V1 ,V2 ] .
If there are no edges leaving V2 , then the result holds trivially. Otherwise, L[V2 ,V2 ] is a strictly
diagonally dominant matrix, and is therefore full rank. We can write it as
L[V2 ,V2 ] = D − A
where D is the diagonal of L[V2 ,V2 ] and A is the negation of the off-diagonal entries, and then
expand L−1
[V2 ,V2 ] via the Jacobi series:
−1
−1
−1/2
−1/2
−1/2
I
−
D
AD
D −1/2
=
(D
−
A)
=
D
L−1
[V2 ,V2 ]
"∞
#
∞
k
X
X
k
D −1/2 =
= D −1/2
D −1 A D −1 . (3)
D −1/2 AD −1/2
k=0
k=0
Note that this series converges because the strict diagonal dominance of L[V2 ,V2 ] implies (AD −1 )k
tends to zero as k → ∞. Substituting this in place of L−1
[V2 ,V2 ] gives:
Sc (G, V1 ) = L[V1 ,V1 ] −
∞
X
k=0
k
L[V1 ,V2 ] D −1 A D −1 L[V2 ,V1 ] .
As all the off-diagonal entries in L are non-positive, we can replace L[V1 ,V2 ] with −L[V1 ,V2 ] to
make all the terms in the trailing summation positive. As these are the only ways to form new
off-diagonal entries, the identity based on matrix multiplication of
Q
G
h
X
k −1
i
0≤i<k w ui ui+1
−1
−L[V1 ,V2 ] D A D
−L[V2 ,V1 ]
=
Q
G
u0 ,uk
0<i<k d ui
u1 ...uk−1
gives the required identity.
This characterization of Sc (G, V1 ), coupled with the (1 + α)-diagonal-dominance of V2 ,
allows us to sample the multi-edges in Sc (G, V1 ) in the same way as the (short) random walk
sparsification algorithms from [CCL+ 15, JKPS17].
Lemma 5.5. Given any graph G = (V, E, w ), an (1 + α)-DD subset V2 , and access to 2approximations of statistical leverage scores on G, τeG , SampleEdgeSchur returns edges in G
according to the distribution p e in O(α) expected time per sample. Furthermore, the distribution
that it samples edges in Sc (G, V1 ) from, p, satisfies
Sc(G,V )
τ u0 ,...uk1
.
n−1
for every edge in Sc (G, V1 ) corresponding to the walk u0 , . . . uk .
O (1) · p u0 ,...uk ≥
26
Algorithm 3: SampleEdgeSchur(G = (V, E, w ), V1 ): samples an edge from Sc (G, V1 )
Input: Graph G, vertices V1 to complement onto, and (implicit) access to a
2-approximation of the leverage scores of G, τeG .
Output: A multi-edge e in Sc (G, V1 ) corresponding to a walk u0 , u1 , . . . uk , and the
probability of it being picked in this distribution p u0 ,u1 ,...uk
eG
1 Sample an edge e from G randomly with probability drawn from τ
e;
2 Perform two independent random walks from the endpoints of e until they both reach
some vertex in V1 , let the walk be u0 . . . uk ;
3 Output edge u0 uk (corresponding to the path u0 , u1 , . . . uk ) with
w u0 ...uk ←
Q
0≤i<k
Q
p u0 ...uk ← P
wG
ui ui+1
0<i<k
1
eG
e′
e′ τ
dG
ui
X
0≤i<k
,
(same as Equation 2)
τeG
ui ui+1 ·
Y wG
uj uj+1
0≤j<i
d uj+1
·
Y
i+1≤j<k
wG
uj uj+1
d uj
.
The guarantees of this procedure are analogous to the random walk sampling sparsification
scheme from [CCL+ 15, JKPS17], with the main difference being the terminating condition for
the walks leads to the removal of an overhead related to the number of steps in the walk. The
modification of the initial step to picking the initial edge from G by resistance is necessary to
get ρ to a constant, as the about n1.5 samples limits the amount of overhead that we can have
per sample.
Proof. We first verify that p is indeed a probability on the multi-edges of Sc (G, V1 ), partitioned
by the walks that they correspond to in G, or formally
X
p u0 ,u1 ...uk = 1.
u0 ,u1 ,...uk :
u0 ,uk ∈V1 ,
ui ∈V2 ∀1≤i<k
To obtain this equality, note that for any random walk starting at vertex i, the total probabilities
of walks starting at i and ending in V1 is upper bounded by 1. Algebraically this becomes:
Y w ui ui+1
X
= 1,
d
u
i
u ,u ,...u
1
2
k
0≤i<k
so applying this to both terms of each edge e gives that the total probability mass over any
G
starting edge is PτeeeG , and in turn the total.
e′
τ e′
For the running time, since V2 is (1 + α)-almost independent, each step of the walk takes
expected time O(α). Also, the value of p u0 ,u1 ,...uk can be computed in O(k) time by computing
prefix/suffix products of the transition probabilities along the path (instead of evaluating each
summand in O(k) time for a total of O(k 2 )).
Finally, we need to bound the approximation of p compared
P G to the true leverage scores τ .
ee is within a constant factor of n.
As τeG
is
a
2-approximation
of
the
true
leverage
scores,
e
eτ
27
So it suffices to show
O (1) ·
X
0≤i<k
τeG
ui ui+1
Y wG
uj uj+1
0≤j<i
d uj+1
·
wG
uj uj+1
Y
i+1≤j<k
d uj
≥ RSc(G,V1 ) (u0 , uk ) · w u0 ,u1 ,...uk .
ef f
Here we invoke the equivalence of effective resistances in G and Sc (G, V1 ) given by Fact 2.6 in
the reverse direction. Then by Rayleigh’s monotonicity principle, we have
Sc(G,V1 )
Ref f
(u0 , uk ) = RG
ef f (u0 , uk ) ≤
X 2e
τG
ui ui+1
0≤i<k
w ui ui+1
,
which when substituted into the expression for w u0 ,u1 ,...uk from Equation 2 gives
G
G
G
X
Y
Y
X 2e
w uj uj+1
w uj uj+1
τ ui ui+1
w u0 ,u1 ,...uk =
.
·
2e
τG
ui ui+1
w ui ui+1
d uj+1
d uj
0≤i<k
0≤j<i
0≤i<k
i+1≤j<k
This sampling procedure can be immediately combined with Theorem 4.1 to give algorithms
for generating approximate Schur complements. Pseudocode of this routine is in Algorithm 4.
Algorithm 4: SchurSparse(G, V1 , δ)
Input: Graph G, 1.1-DD subset of vertices V2 and error parameter δ
Output: Sparse Schur complement of Sc (G, V1 )
1 Set ǫ ← 0.1;
2 Set s ← n2 δ −1 ;
3 Build leverage score data structure on G with errors 0.1 (via Lemma 2.4);
4 Let H V1 ←
DetSparsify(Sc (G, V1 ) , s, SampleEdgeSchur(G, V1 ), LeverageApproxG , ǫ);
5 Output H V1 ;
Proof. (Of Theorem 5.3) Note that the choices of ǫ and s must ensure that
n 2 ǫ2
=δ
s
n3
≤δ
s2
This is then equivalent to s ≥ n1.5 δ−1 and ǫs2 = n2 δ−1 . This further implies that ǫ ≥ n1/4 .
Our ǫ and s in SchurSparse meet these conditions (and the ones specifically chosen in the
algorithm will also be necessary for one of our applications). The guarantees then follow from
putting the quality of the sampler from Lemma 5.5 into the requirements of the determinant
preserving sampling procedure from Theorem 4.1. Additionally, Lemma 5.5 only requires access
to 2-approximate leverage scores, which can be computed by Lemma 2.4 in Õ(m) time. Furthermore, Lemma 5.5 gives that our ρ value is constant, and our assumption in Theorem 5.3 that we
28
are given an 1.1-DD subset V2 implies that our expected O(s · ρ) calls to SampleEdgeSchur
will require O(1) time. The only other overheads are the computation and invocations of the
various copies of approximate resistance data structures. Since m ≤ n2 and ǫ ≥ n1/4 , Lemma 2.4
gives that this cost is bounded by Õ(m + n2 + ǫs2 ) = Õ(n2 δ−1 ).
6
Approximate Determinant of SDDM Matrices
In this section, we provide an algorithm for computing an approximate determinant of SDDM
matrices, which are minors of graph Laplacians formed by removing one row/column.
Theorem 1.1 allows us to sparsify a dense graph while still approximately preserving the
determinant of the graph minor. If there were some existing algorithm for computing the determinant that had good dependence on sparsity, we could achieve an improved runtime for determinant computation by simply invoking such an algorithm on a minor of the sparsified graph.6
Unfortunately, current determinant computation algorithms (that achieve high-accuracy) are
only dependent on n, so simply reducing the edge count does not directly improve the runtime
for determinant computation. Instead the algorithm we give will utilize Fact 2.5
det+ (L) = det L[V2 ,V2 ] · det+ (Sc (L, V1 )).
(where we recall that det+ is the determinant of the matrix minor) to recursively split the matrix.
Specifically, we partition the vertex set based upon the routine AlmostIndependent from
Lemma 5.2, then compute Schur complements according to SchurSparse in Theorem 5.3. Our
algorithm will take as input a Laplacian matrix. However, this recursion naturally produces two
matrices, the second of which is a Laplacian and the first of which is a submatrix of a Laplacian.
Therefore, we need to convert L[V2 ,V2 ] into a Laplacian. We do this by adding one vertex with
appropriate edge weights such that each row and column sums to 0. Pseudocode of this routine
is in Algorithm 5, and we call it with the parameters LV2 ← AddRowColumn(L[V2 ,V2 ] ).
Algorithm 5: AddRowColumn(M ) : complete M into a graph Laplacian by adding
one more row/column
Input: SDDM Matrix M
Output: Laplacian matrix L with one extra row / column than M
1 Let n be the dimension of M ;
2 for i = 1 to n do
3
Sum non-zero entries of row i, call s i ;
4
Set L(n + 1, i), L(i, n + 1) ← −s i ;
Pn
5 Let L(n + 1, n + 1) ←
i=1 s i ;
6 Output L;
The procedure AddRowColumn outputs a Laplacian LV2 such that L[V2 ,V2 ] can be obtained
if one removes this added row/column. This immediately gives det+ (LV2 ) = det(L[V2 ,V2 ] ) by
definition, and we can now give our determinant computation algorithm of the minor of a graph
Laplacian.
6
To get with high probability one could use standard boosting tricks involving taking the median of several
estimates of the determinant obtained in this fashion.
29
Algorithm 6: DetApprox(L, δ, n) : Compute det+ (L) with error parameter δ
Input: Laplacian matrix L, top level error threshold δ, and top level graph size n
Output: Approximate det+ (L)
1 if this is the top-level invocation of this function in the recursion tree then
2
δ′ ← Θ(δ2 / log3 n)
3 else
4
δ′ ← δ
if L is 2 × 2 then
return the weight on the (unique) edge in the graph
5
6
1
)
{Via Lemma 5.2}
V2 ← AlmostIndependent(L, 10
V1 ← V \ V2 ;
LV1 ← SchurSparse(L, V1 , δ′ );
{|V1 | /n is the value of β in Lemma 6.1.}
LV2 ← AddRowColumn(L[V2 ,V2 ] );
Output DetApprox(LV1 , δ′ |V1 | /n, n) · DetApprox(LV2 , δ′ |V2 | /n, n);
7
8
9
10
11
Our analysis of this recursive routine consists of bounding the distortions incurred at each
level of the recursion tree. This in turn uses the fact that the number of vertices across all calls
within a level and the total “amount” of δ across all calls within a level both remain unchanged
from one level to the next. This can be summarized by the following Lemma which bounds the
error accumulated within one level of recursion in our algorithm.
Lemma
6.1. Suppose we are given some small δ ≥ 0 and non-negative β1 , ..., βk such that
Pk
i=1 βi = O(1), along with Laplacian matrices L(1), . . . , L(k) and each having a corresponding
vertex partition V1 (i), V2 (i), where
#
"
L (i)[V1 (i),V1 (i)] L (i)[V1 (i),V2 (i)]
.
L(i) =
L (i)[V2 (i),V1 (i)] L (i)[V2 (i),V2 (i)]
Let LV1 (i) denote the result of running SchurSparse to remove the V2 (i) block in each of these
matrices:7
def
LV1 (i) = SchurSparse (L(i), V1 (i), βi δ) .
Then conditioning upon a with high probability event8 in each of these calls to SchurSparse,
for any p we have with probability at least 1 − p:
k
Y
i=1
7
k
p Y
δ/p
det L[V2 (i),V2 (i)] (i) · det+ LV1 (i) .
det+ (L (i)) = 1 ± O
i=1
This Lemma only applies when the matrices are fixed with respect to the randomness used in the invocations of
SchurSparse mentioned in the Lemma. In other words, it only applies when the result of running SchurSparse
on each of these L(i) matrices is independent of the result of running it on the other matrices. This is why the
Lemma only immediately bounds error within a level of the recursion—where this independence holds—rather
than for the entire algorithm.
8
namely, the event that all the leverage score estimation calls to Lemma 2.4 from SchurSparse succeed
30
Here the βi corresponds to the |V1 | /n and |V2 | /n values that δ is multiplied against in each
call parameter to SchurSparse. An example of the main steps in this determinant approximation algorithm, as well as the graphs corresponding to applying Lemma 6.1 to one of the layers
is in Figure 1.
G on n vertices with Laplacian L
SchurSparse(LG , V1 , βδ)
AddRowColumn(LG
[V2 ,V2 ] )
L(1)
L(2)
SchurSparse(L(1), V1 (1), β1 δ)
LV1 (1) , |V1 (1)| = β1 n
SchurSparse(L(2), V1 (2), β2 δ)
LV1 (2) , |V1 (2)| = β2 n
LV2 (1)
LV2 (2)
Figure 1: Two layers of the call Structure of the determinant approximation algorithm DetApprox (algorithm 6), with the transition from the first to the second layer labeled as in
Lemma 6.1.
Applying Lemma 6.1 to all the layers of the recursion tree gives the overall guarantees.
Proof of Theorem 1.2.
Running Time: Let the number of vertices and edges in the current graph corresponding
to L be n and m respectively. Calling AlmostIndependent takes expected time O(m) and
guarantees
n
n
≤ |V2 | ≤ ,
16
8
which means the total recursion terminates in O(log n) steps.
For the running time, note that as there are at most O(n) recursive calls, the total number
of vertices per level of the recursion is O(n). The running time on each level are also dominated
by the calls to SchurSparse, which comes out to
n
2
Õ |V1 (i)| ′
= Õ |V1 (i)| nδ−2 ,
δ |V1 (i)|
and once again sums to Õ(n2 δ−2 ). We note that this running time can also be obtained from more
standard analyses of recursive algorithms, specifically applying guess-and-check to a running
time recurrence of the form of:
T (n, δ) = T (θn, θδ) + T ((1 − θ) n + 1, (1 − θ) δ) + Õ(n2 δ−1 ).
31
Correctness. As shown in the running time analysis, our recursion tree has depth at most
O(log n), and there are at most O(n) total vertices at any given level. We associate each level
of the recursion in our algorithm with the list of matrices which are given as input to the calls
making up that level of recursion. For any level in our recursion, consider the product of det+
applied to each of these matrices. We refer to this quantity for level j as qj . Notice that q0 is
the determinant we wish to compute and q# levels−1 is what our algorithm actually outputs. As
δ
such, it suffices to prove that for any j, qj = (1 ± # levels
)qj−1 with probability of failure at most
3
1
′
2
10·# levels . However, by the fact that we set δ = Θ(δ / log n) in the top level of recursion with
sufficiently small constants, this immediately follows from Lemma 6.1.
A minor technical issue is that Lemma 6.1 only gives guarantees conditioned on a WHP
event. However, we only need to invoke this Lemma a logarithmic number of times, so we can
absorb this polynomially small failure probability into the our total failure probability without
issue.
Standard boosting techniques—such as running O(log n) independent instances and taking
the medians of the estimates— give our desired with high probability statement.
It remains to bound the variances per level of the recursion.
Proof. (Of Lemma 6.1) As a result of Fact 2.5
k
Y
det+ (L(i)) =
k
Y
i=1
i=1
det L (i)[V2 (i),V2 (i)] det+ (Sc (L (i) , V1 (i))) .
Consequently, it suffices to show that with probability at least 1 − p
k
Y
i=1
k
p Y
det+ (Sc (L (i) , V1 (i))) = 1 ± O
δ/p
det+ LV1 (i) .
i=1
Recall that LV1 (i) denotes the random variable that is the approximate Schur complement generated through the call to SchurSparse(L(i), V1 (i), βi δ).
Using the fact that our calls to SchurSparse are independent along with the assumption
P
of ki=1 βi = O(1), we can apply the guarantees of Theorem 5.3 to obtain
#
" k
k
i
h
Y
Y
V1 (1)
ELV1 (i) det+ LV1 (i)
=
ELV1 (1) ...LV1 (k)
det+ L
i=1
i=1
= (1 ± O (δ))
and
k
Y
det+ (Sc (L(i), V1 (i))) ,
i=1
2
2
V1 (i)
V1 (i)
ELV1 (1) ...LV1 (k)
k ELV1 (i) det+ L
i=1 det+ L
Y
=
i2
h
2
Qk
V (i)
V1 (i)
i=1 ELV1 (i) det+ L 1
L
ELV1 (1) ...LV1 (k)
det
+
i=1
Qk
≤
32
k
Y
i=1
exp (O (βi δ)) ≤ exp (O (δ)).
By assumption δ is small, so we can approximate exp (O(δ)) with 1+O(δ), which with bound
above gives
#
#
" k
" k
2
Y
Y
V1 (i)
V1 (i)
det+ L
≤ O (δ) ELV1 (1) ...LV1 (k)
,
det+ L
VarLV1 (1) ...LV1 (k)
i=1
Then applying the approximation on E
VarLV1 (1) ...LV1 (k)
" k
Y
i=1
hQ
det+ LV1 (i)
i=1
k
i=1 det+ (SchurSparse(L(i), V1 (i), βi δ))
#
≤ O (δ)
k
Y
i=1
i
!2
det+ (Sc (L(i), V1 (i)))
gives
.
At which point we can apply Chebyshev’s inequality to obtain our desired result.
7
Random Spanning Tree Sampling
In this section we will give an algorithm for generating a random spanning tree from a weighted
graph, that uses SchurSparse as a subroutine, and ultimately prove Theorem 1.3.
In order to do so, we will first give an O(nω ) time recursive algorithm using Schur complement
that exactly generates a random tree from the w -uniform distribution. The given algorithm is
inspired by the one introduced in [CDN89], and its variants utilized in [CMN96, HX16, DKP+ 16].
However, we will (out of necessity for our further extensions) reduce the number of branches in
the recursion to two, by giving an efficient algorithmic implementation of a bijective mapping
between spanning trees in G and spanning trees in Sc (G, V2 ) when V1 , the set of vertices
removed, is an independent set. We note that this also yields an alternative algorithm for
generating random spanning trees from the w -uniform distribution in O(nω ) time.
The runtime of this recursion will then be achieved similar to our determinant algorithm.
We reduce δ proportional to the decrease in the number of vertices for every successive recursive
call in exactly the same was as the determinant approximation algorithm from Section 6. As has
been previously stated and which is proven in Section 9.1, drawing a random spanning tree from
e 2 −1
a graph after
√ running our sparsification routine which takes O(n δ ), will have total variation
distance δ from the w -uniform distribution.
Similar to our analysis of the determinant algorithm, we cannot directly apply this bound
to each tree because the lower levels of the recursion will contribute far too much error when
δ is not decreasing at a proportional rate to the total variation distance. Thus we will again
need to give better bounds on the variance across each level, allowing stronger bounds on the
contribution to total variation distance of the entire level.
This accounting for total variance is more difficult here due to the stronger dependence
between the recursive calls. Specifically, the input to the graph on V2 depends on the set of
edges chosen in the first recursive call on V1 , specifically Sc (G, V1 ), or a sparsified version of it.
Accounting for this dependency will require proving additional concentration bounds shown
in Section 8, which we specifically achieve by sampling s = O(n2 δ−1 ) edges in each call to
SchurSparse. While this might seem contradictory to the notion of “sampling”, we instead
consider this to be sampling from the graph in which all the edges generated from the Schur
complement are kept separate and could be far more than n2 edges.
33
7.1
Exact O(nω ) Time Recursive Algorithm
We start by showing an algorithm that samples trees from the exact w -uniform distribution
via the computation of Schur complements. Its pseudocode is in Algorithm 7, and it forms the
basis of our approximate algorithm: the faster routine in Section 7.2 is essentially the same as
inserting sparsification steps between recursive calls.
Algorithm 7: ExactTree(G) : Take a graph and output a tree randomly from the
w -uniform distribution
Input: Graph G
Output: A tree randomly generated from the w -uniform distribution of G
1 If there is only one edge e in G, return G ;
2 Partition V evenly into V1 and V2 ;
3 T1 = ExactTree(Sc(G, V1 ));
4 for each e ∈ T1 do
w e (G)
, G ← G/e, T ← T ∪ e ;
5
with probability w e (Sc(G,V
1 ))
6
7
8
9
Delete the remaining edges, i.e., G ← G \ E(V1 );
T2 = ExactTree(Sc(G, V2 ));
T ← T ∪ ProlongateTree(G, V1 ⊔ V2 , T2 );
Output T ;
The procedure ProlongateTree is invoked when V1 = V \ V2 maps a tree T2 from the
Schur complement Sc (G, V2 ) to a tree back in G. It crucially uses the property that V1 is an
independent set, and its pseudocode is given in Algorithm 8.
Lemma 7.1. The procedure ExactTree(G) will generate a random tree of G from the w uniform distribution in O(nω ) time.
The algorithm we give is similar to the divide and conquer approaches of [CDN89, CMN96,
HX16, DKP+ 16]. The two main facts used by these approaches can be summarized as follows:
1. Schur complements preserves the leverage score of original edges, and
2. The operation of taking Schur complements, and the operation of deleting or contracting
an edge are associative.
We too will make use of these two facts. But unlike all previous approaches, at every stage
we need to recurse on only two sub-problems. All previous approaches have a branching factor
of at least four.
We can do this by exploiting the structure of the Schur complement when one eliminates an
independent set of vertices. We formalize this in Lemma 7.5.
Before we can prove the lemma, we need to state an important property of Schur complements that follows from Fact 2.8. Recall the notation from Section 2 that for a weighted graph
G = (V, E, w ), P rG
T (·) denotes the probability of · over trees T picked from the w -uniform
distribution on spanning trees of G.
34
Algorithm 8: ProlongateTree(G, V1 ⊔ V2 , T2 ): prolongating a tree on Sc (G, V2 ) to a
tree on G.
Input: A graph G, a splitting of vertices V1 ⊔ V2 such that V1 is an independent set, tree
T2 of Sc (G, V2 ).
Output: A tree in G
1 T ← ∅;
2 for each e = xy ∈ T2 do
3
Create distribution λe , set λe (∅) = w e (G);
4
for each v ∈ V1 such that (v, x), (v, y) ∈ E(G) do
5
Set λe (v) = w (v,x) (G)w (v,y) (G)d v (G)−1 ;
6
7
8
9
10
11
12
13
14
Randomly assign f (e) to {∅ ∪ V1 } with probability proportional to λ;
for each v ∈ V1 do
for each e = (x, y) ∈ T2 such that (v, x), (v, y) ∈ E(G) do
if f (e) 6= v then
Contract x and y ;
for each contracted vertex X in the neighborhood of v do
Connect X to v with edge (v, u) ∈ G with probability proportional to wG ((v, u)) ;
T ← T ∪ (v, u);
Output T ;
Lemma 7.2. Let G be a graph with a partition of vertices V = V1 ⊔ V2 . Then for any set of
edges F contained in G[V1 ], the induced subgraph on V1 , we have:
Sc(G,V1 )
P rG
T (T ∩ E (G [V1 ]) = F ) = P r T
(T ∩ E (G [V1 ]) = F ) ,
where the edges in Sc (G, V1 ) are treated as the sum of G[V1 ] and Gsc [V1 ], the new edges added
to the Schur complement.
Sc(G,V )
1
(T ∩ E (G [V1 ]) =
Proof. If F contains a cycle, then P r G
T (T ∩ E (G [V1 ]) = F ) = 0 = P r T
F ). Therefore, we will assume F does not contain any cycle, and we will prove by induction on
the size of F . If |F | > |V1 | − 1, then F will have to contain a cycle. When |F | = |V1 | − 1, then F
will have to be the edge set of a tree in Sc(G, V1 ). Then by Fact 2.8, the corollary holds. Now
suppose that the corollary holds for all F with |F | = |V1 | − 1 − k. Now consider some F with
|F | = |V1 | − 1 − (k + 1). We know
X
G
′
P rG
P rG
T (F ⊆ T ) = P r T (F = (T ∩ E (G [V1 ]))) +
T F = (T ∩ E (G [V1 ])) .
F ′ ⊃F
Since |F ′ | > |F |, by assumption
X
X
Sc(G,V1 )
′
P rG
F ′ = (T ∩ E (G [V1 ])) ,
P rT
T F = (T ∩ E (G [V1 ])) =
F ′ ⊃F
F ′ ⊃F
Sc(G,V1 )
then by Fact 2.8 we have P rG
T (F ⊆ T ) = P r T
(F ⊆ T ), which implies
Sc(G,V1 )
P rG
T (F = (T ∩ E (G [V1 ]))) = P r T
35
(F = (T ∩ E (G [V1 ]))) .
The tracking of edges from various layers of the Schur complement leads to another layer of
overhead in recursive algorithms. They can be circumvented by merging the edges, generating a
random spanning tree, and the ‘unsplit’ the edge by random spanning. The following is a direct
consequence of the definition of w (T ):
b be a multi-graph, and G be the simple graph formed by summing the weights
Lemma 7.3. Let G
of overlapping edges. Then the procedure of:
1. Sampling a random spanning tree from G, T .
b eb with probability
2. For each edge e ∈ T , assign it to an original edge from G,
b
w eb G
.
w e (G)
b the original multi-graph.
Produces a w -uniform spanning tree from G,
This then leads to the following proto-algorithm:
1. Partition the vertices (roughly evenly) into
V = V1 ⊔ V2 .
2. Generate a w -uniform tree of Sc (G, V1 ), and create F1 = T ∩ E(G[V1 ]) by re-sampling
edges in G[V1 ] using Lemma 7.3. By Lemma 7.2, this subset is precisely the intersection
of a random spanning tree with G[V1 ].
3. This means we have ‘decided’ on all edges in G[V1 ]. So we can proceed by contracting all
the edges of F1 , and deleting all the edges corresponding to E(G[V1 ])\F . Let the resulting
graph be G′ and let V1′ be the remaining vertices in V1 after this contraction.
4. Observe that V1′ is an independent set, and its complement is V2 . We can use another
recursive call to generate a w -uniform tree in Sc(G′ , V2 ). Then we utilize the fact that V1′
is an independent set to lift this to a tree in G′ efficiently via Lemma 7.5.
Our key idea for reducing the number of recursive calls of the algorithm, that when V1 (from
the partition of vertices V = V1 ⊔ V2 ) is an independent set, we can directly lift a tree from
Sc(G, V2 ) to a tree in G. This will require viewing GSc [V2 ] as a sum of cliques, one per vertex
of V1 , plus the original edges in G[V2 ].
Fact 7.4. Given a graph G and a vertex v, the graph Sc(G, V \v) is the induced graph G[V \{v}]
plus a weighted complete graph K(v) on the neighbors of v. This graph K(v) is formed by adding
one edge xy for every pair of x and y incident to v with weight
w (v,x) w (v,y)
,
degv
def
where d v =
P
x w (v,x)
is the weighted degree of v in G.
36
Lemma 7.5. Let G be a graph on n vertices and V1 an independent set. If T is drawn from
the w -uniform distribution of Sc(G, V2 ), then in O(n2 ) time ProlongateTree(G, V1 ⊔ V2 , T2 )
returns a tree from the w -uniform distribution of G.
Proof. The running time of ProlongateTree is O(n2 ) as T2 has ≤ n − 1 edges and |V1 | ≤ n.
Now we will show the correctness. Let V1 = {v1 , ..., vk }. We will represent Sc(G, V2 ) as a
multi-graph arising by Schur complementing out the vertices in V1 one by one and keeping the
new edges created in the process separate from each other as a multi-graph. We represent this
multi-graph as
Sc(G, V2 ) = G [V2 ] + K (v1 ) + ... + K (vk ) ,
where G[V2 ] is the induced subgraph on V2 and K(vi ) is the weighted complete graph on the
neighbors of vi . Then
• By the unsplitting procedure from Lemma 7.3, the function f maps T2 to a tree in the
multi-graph G[V2 ] + K(v1 ) + ... + K(vk ), and
• the rest of the sampling steps maps this tree to one in G.
We will now prove correctness by induction on the size of the independent set V1 . The case
of |V1 | = 0 follows from Sc (G, V2 ) = G. If |V1 | = 1, i.e, V1 = {v} for some vertex v, then
Sc(G, V2 ) is G[V2 ] + K(v). Given a tree T2 of Sc(G, V2 ), the creation of f will first map T2 to
a tree in the multigraph G[V2 ] + K(v) by randomly deciding for each edge e ∈ T to be in G(V1 )
or K(v) depending on it’s weight. If we let T ′ (V2 ) = T ′ ∩ G[V2 ], then by Lemma 7.2,
G[V2 ]+K(v)
′
P rG
T ∩ E(G [V2 ]) = T ′ (V2 ) .
T T ∩ E(G [V2 ]) = T (V2 ) = P r T
Therefore, we can contract all the edges of T ′ (V2 ) ∩ G [V2 ] and delete all other edges of G [V2 ].
This results in a multi-graph star with v at the center. Now, ProlongateTree does the
following to decide on the remaining edges. For every multi-edge of the star graph obtained by
contracting or deleting edges in G[V2 ], we choose exactly one edge, randomly according to its
weight. This process generates a random tree of multi-graph star.
Now we assume that the lemma holds for all V1′ with |V1′ | < k. Let V1 = {v1 , ..., vk }. The
key thing to note is that when V1 is an independent set, we can write
Sc (G, V2 ) = G [V2 ] + K (v1 ) + . . . + K (vk ) ,
and
Sc (G, V2 ∪ vk ) = G [V2 ∪ vk ] + K (v1 ) + . . . + K (vk−1 ) .
Therefore, by the same reasoning as above, we can take a random tree T ′ of the multi-graph
G[V2 ] + K(v1 ) + ... + K(vk ) and map it to a tree on G[V2 ∪ vk ] + K(v1 ) + ... + K(vk−1 ) =
Sc (G, V2 ∪ vk ) by our procedure ProlongateTree. We then apply our inductive hypothesis
on the set V1 \ {vk } to map Sc(G, V2 ∪ vk ) to a tree of G by ProlongateTree, which implies
the lemma.
We also remark that the running time of ProlongateTree can be reduced to O(m log n)
using dynamic trees, which can be abstracted as a data structure supporting operations on rooted
forests [ST85, AHLT05]. We omit the details here as this does not bottleneck the running time.
With this procedure fixed, we can now show the overall guarantees of the exact algorithm.
37
Proof. of Lemma 7.1 Correctness follows immediately from Lemmas 7.2 and 7.5. The running
time of ProlongateTree is O(n2 ) and contracting or deleting all edges contained in G[V1 ]
takes O(m) time. Note that in this new contracted graph, the vertex set containing V1 is an
independent set. Furthermore, computing the Schur complement takes O(nω ) time, giving the
running time recurrence
T (n) = 2T (n/2) + O (nω ) = O (nω ) .
7.2
Fast Random Spanning Tree Sampling using Determinant Sparsification
of Schur complement
Next, we note that the most expensive operation from the exact sampling algorithm from Section 7.1 was the Schur complement procedure. Accordingly, we will substitute in our sparse
Schur complement procedure to speed up the running time.
However, this will add some complication in applying Line 5 of ExactTree. To address
this, we need the observation that the SchurSparse procedure can be extended to distinguish
edges from the original graph, and the Schur complement in the multi-graph that it produces.
Lemma 7.6. The procedure SchurSparse(G, V1 , δ) given in Algorithm 4 can be modified to
record whether an edge in its output, H V1 is a rescaled copy of an edge from the original induced
subgraph on V1 , G[V1 ], or one of the new edges generated from the Schur complement, GSC (V1 ).
Proof. The edges for H V1 are generated by the random walks via SampleEdgeSchur(G, V1 ),
whose pseudocode is given in Algorithm 3. Each of these produces a walk between two vertices
in V1 , and such a walk belongs to G[V1 ] if it is length 1, and GSC (V1 ) otherwise.
We can now give our algorithm for generating random spanning trees and prove the guarantees that lead to the main result from Theorem 1.3.
Note that the splitting on Line 7 is mapping T1 first back to a tree on a the sparsified multigraph of Sc (G, V1 ): where the rescaled edges that originated from G[V1 ] are tracked separately
from the edges that arise from new edges involving random walks that go through vertices in
V2 .
The desired runtime will follow equivalently to the analysis of the determinant algorithm in
Section 6 as we are decreasing δ proportionally to the number of vertices. It remains to bound
the distortion to the spanning tree distribution caused by the calls to SchurSparse.
Bounds on this distortion will not follow equivalently to that of the determinant algorithm,
which also substitutes SchurSparse for exact Schur complements, due to the dependencies in
our recursive structure. In particular, while the calls to SchurSparse are independent, the
graphs that they are called upon depend on the randomness in Line 6 and ProlongateTree,
which more specifically, are simply the resulting edge contractions/deletions in previously visited vertex partitions within the recursion. Each subgraph SchurSparse is called upon is
additionally dependent on the vertex partitioning from AlmostIndependent.
The key idea to our proof will then be a layer-by-layer analysis of distortion incurred by
SchurSparse at each layer to the probability of sampling a fixed tree. By considering an
alternate procedure where we consider exactly sampling a random spanning tree after some
layer, along with the fact that our consideration is restricted to a fixed tree, this will allow
38
Algorithm 9: ApproxTree(G, δ, n) Take a graph and output a tree randomly from a
distribution δ-close to the w -uniform distribution
Input: Graph G, error parameter δ, and initial number of vertices n
Output: A tree randomly generated from a distribution δ-close to the w -uniform
distribution of G
1
{Via Lemma 5.2}
1 V2 ← AlmostIndependent(G, 10 );
2 H1 ← SchurSparse(G, V1 , δ · |V1 |/n)), while tracking whether the edge is from G[V1 ] via
the modifications from Lemma 7.6 ;
3 T1 = ApproxTree(H1 , δ, n);
4 G′ ← G ;
5 for each e ∈ T1 do
6
if RAND[0, 1] ≤ w ori
e (G1 )/w e (G1 ) then
{w ori
(G
1 ) is calculated using the weights tracked from Line 2);
e
7
G′ ← G′ /{e} ;
8
T ← T ∪ {e};
9
10
11
12
13
Delete all edges between (remaining) vertices in V1 in G′ , G′ ← G′ \ E(G′ [V1 ]) ;
H2 ← SchurSparse(G′ , V2 , δ · |V2 |/n) ;
T2 = ApproxTree(H2 , δ, n);
T ← T ∪ ProlongateTree(G, V1 ⊔ V2 , T2 ) ;
Output T ;
us to separate the randomness incurred by calls to SchurSparse from the other sources of
randomness mentioned above. Accordingly, we will provide the following definition.
Definition 7.7. For any L ≥ 0, the level-L truncated algorithm is the algorithm given by
modifying ApproxTree(G, δ, n) so that all computations of sparsified Schur complements are
replaced by exact calls to Schur complements (aka. Sc (G, V1 ) or Sc (G′ , V2 ))) after level l.
The tree distribution T (L) is defined as the output of the level-L truncated algorithm.
Note that in particular, T (0) is the tree distribution produced by ExactTree(G), or the
w -uniform distribution; while T (O(log n)) is the distribution outputted by ApproxTree(G, δ).
The primary motivation of this definition is that we can separate the randomness between
(l)
T and T (l+1) by only the calls to SchurSparse at level l + 1, which will ultimately give the
following lemma that we prove at the end of this section
Lemma 7.8. For an invocation of ApproxTree on a graph G with variance bound δ, for any
layer L > 0, we have
√
dT V T (L−1) , T (L) ≤ O( δ).
To begin, we consider the differences between T (0) and T (1) and the probability of sampling
a fixed tree Tb on a recursive call on G. The most crucial observation is that the two recursive
calls to ApproxTree(G1 , δ, n) and ApproxTree(G2 , δ, n) can be viewed as independent:
Claim 7.9. For a call to ApproxTree(G, δ, n) (Algorithm 9) to return Tb, there is only one
possible choice of G′ as generated via Lines 4 to 9.
39
Proof. Note that the edges removed from Line 7 are precisely the edges in T with both endpoints
contained in V1 , E(T [V1 ]). For a fixed Tb, this set is unique, so G′ is unique as well.
This allows us to analyze a truncated algorithm by splitting the probabilities into those that
occur at level l or above. Specifically, at the first level, this can be viewed as pairs of graphs
Sc (G, V1 ) and Sc (G, V2 ) along with the ‘intended’ trees from them:
Definition 7.10. We define the level-one probabilities of returning a pair of trees T1 and T2
that belong a pair of graphs G1 , G2 ,
p(≤1) (G, G1 , G2 ) , (T1 , T2 ) , Tb .
as the product of:
1. The probability (from running AlmostIndependent) that G is partitioned into V1 ⊔ V2
so that Sc (G, V1 ) = G1 and Sc (G′ , V2 ) = G2 , where G′ is G with the edges T ∩ G[V1 ]
contracted and all other edges in G[V1 ] are deleted.
2. The probability that T1 is mapped to Tb[V1 ] in Line 6.
3. The probability that T2 is mapped to Tb/Tb[V1 ] by the call to ProlongateTree on Line 12.
This definition then allows us to formalize the splitting of probabilities above and below level
1. More importantly, we note that if we instead call SchurSparse to generate G1 and G2 , this
will not affect the level-one probability because (1) both the calls to AlmostIndependent and
ProlongateTree do not depend on G1 and G2 , and (2) we can consider T1 to be drawn from
the multi-graph of G1 where we track which edges are from the original graph and which were
generated by the Schur complement.
Consequently, the only difference between the distributions T (0) and T (1) will be the distortion of drawing T1 and T2 from G1 and G2 vs the sparsified version of G1 and G2 . This handling
of sparsifiers of the Schur complements is further simplified with by the following observation:
Claim 7.11. The output of SchurSparse(G, V ′ , δ) is identical to the output of
IdealSparsify Sc G, V ′ , τe, n2 δ−1 ,
for some set of 1.1-approximate statistical leverage scores of Sc (G, V ′ ), τe.
This can be seen by revisiting the Schur complement sparsification and rejection sampling algorithms from Section 5 and 4.3 which show that this statement also extends to the approximate
Schur complements produced on lines 2 and 10 of Algorithm 9.
This means we can let H1 and H2 denote the distribution produced by IdealSparsify on
G1 and G2 respectively.
~ T~ )≤1 such that for any tree
Lemma 7.12. There exists a collection of graphs and tree pairs (G,
b
T , with the probabilities given above in Definition 7.10 we have:
X
(0)
P rT
Tb =
p(≤1) (G, G1 , G2 ) , (T1 , T2 ) , Tb · P rG1 (T1 ) · P rG2 (T2 ) .
((G,G1 ,G2 ),(T1 ,T2 ))∈(G,T )(≤1)
40
and
P rT
(1)
Tb =
X
((G,G1 ,G2 ),(T1 ,T2 ))∈(G,T )(≤1)
p(≤1) (G, G1 , G2 ) , (T1 , T2 ) , Tb
· EH1 ∈H1 P r H1 (T1 ) · EH2 ∈H2 P r G2 (T2 ) .
We can then in turn extend this via induction to multiple levels. It is important to note that
in comparing the distributions T (L−1) and T (L) for L ≥ 1 both will make calls to IdealSparsify
through level L. We will then need to additionally consider the possible graphs generated by
sparsification through level L, then restrict to the corresponding exact graphs at level L + 1.
Definition 7.13. We will use G~ (≤L) , T~ (L) to denote a sequence of graphs on levels up to L − 1,
plus the peripheral exact Schur complements on level L, along with the spanning trees generated
on these peripheral graphs.
~ T~ )(≤L) to denote
As these graphs and trees can exist on different vertex sets, we will use (G,
~ ≤L
the set of graph/tree pairs that are on the same set of vertices. For a sequence of graphs G
and a sequence of trees on their peripherals, T~ L , we will use
~ (≤L) , T~ (L) , Tb
p(≤L) G
to denote the product of the probabilities of the level-by-level vertex split and resulting trees
mapping back correctly as defined in Definition 7.10, times the probabilities that the subsequent
graphs are generated as sparsifiers of the ones above
~ (L) to denote just the peripheral graphs, and H(
~ G
~ (L) ) to denote
Furthermore, we will use G
~ (L) to denote one particular
the product distribution over sparsifiers of these graphs, and H
sequence of such sparsifiers on this level. We can also define the probabilities of trees being
picked in a vector-wise sense:
~ (L)
~ (L)
~ (L) ~ (L) def Y
~ (L) ~ (L) def Y
(L)
(L)
.
P r Hj
=
,
P rH
T
T~j
P r Gj
=
P rG
T
T~j
j
j
Applying Lemma 7.12 inductively then allows us to extend this to multiple levels.
~ T~ )(≤L) such that for any
Corollary 7.14. There exists a collection of graphs and tree pairs (G,
b
tree T we have:
X
(L−1)
~ (≤L) , T~ (L) , Tb · P r G~ (L) T~ (L) ,
p(≤L) G
P rT
Tb =
(G~ (≤L) ,T~ (L) )∈(G,T )(≤L)
and
P rT
(L)
Tb =
X
(G~ (≤L) ,T~ (L) )∈(G,T )(≤L)
i
h
~ (≤L) , T~ (L) , Tb · E ~ (L) ~ (L) P rH~ (L) T~ (L) .
p(≤L) G
H ∼H(G )
This reduces our necessary proof of bounding the total variation distance between T (L−1)
and T (L) to examining the difference between
i
h
~ (L) ~ (L)
~ (L) ~ (L)
.
T
and
EH~ (L) ∼H~ (G(L) ) P rH
P rG
T
41
~ (L)
Recalling the definition of P rH
expectation is
~ (L)
P r Hj
(T~ (L) ): we have that the inverse of each probability in the
(L) −1
=
T~j
TH~ (L)
j
,
~ (L)
(L)
H
~
j
w
Tj
~ (L) . However, it is critical to note that
and we have concentration bounds for the total trees in H
j
(L)
~
~ (L) for some j.
this probability is 0 (and cannot be inverted) when Tj is not contained in H
j
This necessitates extending our concentration bounds to random graphs where we condition
upon a certain tree remaining in the graph. This will be done in the following Lemma, proven
2
in Section 8, and we recall that we set s such that δ = O( ns ) in SchurSparse.
Lemma 7.15. Let G be a graph on n vertices and m edges, τe be an 1.1-approximate estimates of
2
leverage scores, s be a sample count such that s ≥ 4n2 and m ≥ sn . Let H denote the distribution
over the outputs of IdealSparsify(G, τe, s), and for a any fixed spanning Tb, let H|T denote the
distribution formed by conditioning on the graph containing Tb. Then we have:
2
−1
h
i−1
−1
n
H|Tb b
b
P rG Tb
,
PrH∼H T ⊆ H
· EH|Tb ∼H| P r
T
= 1±O
b
s
T
and
2
−2
h
i−2
−1
n
P rG Tb
.
· VarH|Tb ∼H| P rH|Tb Tb
PrH∼H Tb ⊆ H
≤O
b
s
T
Due to the independence of each call to IdealSparsify, we can apply these concentration
bounds across the product
Y
~ (L)
~ (L) ~ (L)
(L)
T~j
T
=
P r Hj
P rH
j
and use the fact that δ decreases proportionally to vertex size in our algorithm:
~ (l) , with associated sparsifier distribuCorollary 7.16. For any sequence of peripheral graphs G
~ (L)
tion HS , and any sequence of trees T~ (L) as defined in Definition 7.13 such that P r G (T~ (L) ) > 0,
we have
i−1
h
~ (L) ~ (L) −1
~ (L) ~ (L)
H
H
Pr
T
>0
· E ~ (L) ~ (L)
T
PrH~ (L) ∼H~ (G(L) ) P r
~ (L) ~ (L)
H ∼H(G ) P r H
(T )>0
~ (L) ~ (L) −1
= (1 ± δ) P rG
T
,
and
h
PrH~ (L) ∼H~ (G(L) ) P r
~ (L)
H
i−2
T~ (L) > 0
· E ~ (L)
H
~ (L)
~ (G(L) ) P r H
∼H
42
(T~ (L) )>0
Pr
~ (L)
H
−2
(L)
~
T
~ (L)
≤ (1 + δ) P rG
−2
.
T~ (L)
Proof. The independence of the calls to IdealSparsify, and the definition of
~ (L)
~ (L)
~ (L) ~ (L) def Y
~ (L) ~ (L) def Y
(L)
(L)
.
,
P rH
T
=
P r Hj
T~j
P rG
T
=
P r Gj
T~j
j
j
Applying Lemma 7.15 to each call of IdealSparsify, where s was set such that δ/n =
gives that the total error bounded by
X V G(L)
,
exp
n
n2
s
gives
j
and the bound then follows form the total size of each level of the recursion being O(n̄).
It then remains to use concentration bounds on the inverse of the desired probability to
bound the total variation distance, which can be done by the following lemma which can be
viewed as an extension of Lemma 3.7, and is also proven in Section 9.
Lemma 7.17. Let U be a distribution over a universe of elements, u, each associated with
random variable Pu such that
Eu∼U [E [Pu ]] = 1,
and for each Pu we have
1. Pu ≥ 0, and
2. Pr [Pu > 0]−1 · Ep∼Pu |p>0 p−1 = 1 ± δ, and
3. Pr [Pu > 0]−2 Ep∼Pu|p>0 p−2 ≤ 1 + δ,
then
Eu∼U [|1 − E [Pu ]|] ≤ O
√
δ .
To utilize this lemma, we observe that the values
~ (≤L) , T~ (L) , Tb · P r G~ (L) T~ (L)
p(≤L) G
~ (≤L) , T~ (L) , Tb, while the distribution H(G
~ (L) ), once
forms a probability distribution over tuples G
rescaled, can play the role of Pu . Decoupling the total variation distance per tree into the
~ (≤L) , T~ (L) then allows us to bound the overall total variation
corresponding terms on pairs of G
distance between T (L−1) and T (L) .
Proof of Lemma 7.8. By the definition of total variation distance
X
(L−1)
(L)
P rT
Tb − P rT
Tb .
dT V T (L−1) , T (L) =
Tb
43
By Corollary 7.14 and triangle inequality we can then upper bound this probability by
X
dT V T (L−1) , T (L) ≤
Tb
X
(G~ (≤L) ,T~ (L) )∈(G,T )(≤L)
~ (≤L) , T~ (L) , Tb
p(≤L) G
~ (L)
· P rG
i
h
~ (L) ~ (L)
.
T
T~ (L) − EH~ (L) ∼H~ (G(L) ) P rH
~ (≤L) , T~ (L) , Tb) is crucially the same for each, and the inner term in the
The scalar p(≤L) (G
summation is equivalent to
i
h
~ (L) ~ (L)
~ (≤L) , T~ (L) , Tb · E ~ (L) ~ (L) P r H
~ (≤L) , T~ (L) , Tb · P rG~ (L) T~ (L) − p(≤L) G
T
p(≤L) G
H ∼H(G )
~ (L) , T~ (L) , Tb)
Our goal is to use Lemma 7.17 where U here is the distribution over tuples (G
with density equaling:
~ (≤L) , T~ (L) , Tb · P rG~ (L) T~ (L) ,
p(≤L) G
~ (L) ), with the same density, and
and Pu is the distribution over the corresponding value of H(G
values equaling to:
~ (L) ~ (L)
~ (L) ~ (L) −1
.
P rH
T
P rG
T
Note that the fact that each T~ L maps back to some tree T̂ imply that U is a distribution, as well
as Eu∼U [E [Pu ]] = 1. A rescaled version of Corollary 7.16 then gives the required conditions for
Lemma 7.17, which in turn gives the overall bound.
Proof of Theorem 1.3. The running time follows the same way as the analysis of the determinant
estimation algorithm in the Proof of Theorem 1.2 at the end of Section 6.
For correctness, the total variation distance bound is implied by appropriately setting δ, and
then invoking the per-layer bound from Lemma 7.8. Note that factors of log n are absorbed by
e notation.
the O
Finally, note that for simplicity our analysis of total variation distance does not account for
the failure probability of Lemma 2.4. To account for these, we can simply use the fact that
only O(n log n) calls to SchurSparse are made. Hence, the probability of any call failing is
polynomially small, which can be absorbed into the total variation distance.
8
Conditional Concentration Bounds
In this section, we extend our concentration bounds to conditioning on a certain tree being in the
sampled graph, specifically with the goal of proving Lemma 7.15. By edge splitting arguments
similar to those in Section 4.2, it suffices to analyze the case where all edges have about the
same leverage score.
Lemma 8.1. Let G be a graph on n vertices and m edges such that all edges have statistical
s2
2
leverage scores τ e ≤ 2n
m , and s be a sample count such that s ≥ 4n and m ≥ n . Let H
be a subgraph containing s edges picked at random without replacement, and let H denote this
44
distribution over subgraphs on s edges. Furthermore for any fixed spanning tree, Tb, let H|T
denote the distribution induced by those in H that contain Tb, and use H|Tb to denote such a
graph, then
2
−1
h
i−1
−1
n
H|Tb b
b
P rG Tb
,
PrH∼H T ⊆ H
· EH|Tb ∼H| P r
T
= 1±O
b
s
T
and
2
−1
−2
n
b
PrH∼H
T
· VarH|Tb ∼H| P r
≤O
P rG Tb
.
b
s
T
Note that the ‘uniform leverage score’ requirement here is not as strict as the analysis from
Lemma 4.2. This is because we’re eventually aiming for a bound of s ≈ n2 samples. This also
means that constant factor leverage score approximations suffices for this routine.
The starting point of this proof is the observation that because we’re doing uniform sampling,
the only term in
w H|Tb Tb
w G Tb
=
P rH|Tb Tb =
TH|Tb
TH|Tb
h
Tb ⊆ H
i−2
H|Tb
that is dependent on H|Tb is TH|Tb . The proof will then follow by showing concentration of this
variable which will be done similarly to the concentration of TH that was done in Section 3
and 4.
The primary difficulty of extending the proof will come from the fact that trees will have
different probabilities of being in the sampled graph depending on how many edges they share
with Tb. Much of this will be dealt with by the assumption that s ≥ 4n2 , which makes the
exponential terms in the probabilities associated with a tree being in a sampled graph negligible.
Additionally, this assumption implies that for any fixed tree Tb the expected number of edges
it shares with a random tree is close to 0. As a result, trees that intersect with Tb will have
negligible contributions, and our analysis can follow similarly to that in Section 3 and 4.
We further note that due to the larger sample count of s ≥ 4n2 , the concentration bounds
in this section will also hold, and would in fact be slightly simpler to prove, if the edges were
sampled independently with probability s/m. We keep our assumption of sampling s edges
globally without replacement though in order to avoid changing our algorithm, and the analysis
will not require much additional work.
The section will be organized as follows: In Section 8.1 we give upper and lower bounds on
the expectation of TH|Tb . In Section 8.2 we give an upper bound on the variance of TH|Tb . In
Section 8.3 we combine the bounds from the previous two sections to prove Lemma 8.1.
8.1
Upper and Lower Bounds on Conditional Expectation
i
h
In order to prove upper and lower bounds on EH|Tb TH|Tb , we will first give several helpful
h
i
definitions, corollaries, and lemmas to assist in the proof. Our examination of EH|Tb TH|Tb will
require approximations of PrH|Tb T ⊆ H|Tb , and, as we are now fixing n − 1 edges and drawing
s−n+1
s − n + 1 edges from the remaining m − n + 1 edges, each edge will now have probability m−n+1
of being in the sampled graph. We will denote this probability with
def s − n + 1
.
pb =
m−n+1
45
It will often be easier to exchange pb for
def
p =
s
,
m
the probability of a single edge being picked without the conditioning on Tb. The errors of doing
so is governed by:
s−n
s−n+1
s
n
p=
≤
= p̂ ≤
= p.
(4)
1−
s
m
m−n+1
m
We remark that these errors turn out to be acceptable even when pb is raised to the O(n) power.
Furthermore, our assumption of s ≥ 4n2 implies that we expect a randomly chosen tree not
to intersect with Tb. This will often implicitly show up in the form of the geometric series below,
for which a bound is immediately implied by our assumption.
Lemma 8.2. If s ≥ 4n2 , then
k
∞
X
2n2
k=1
s
=O
n2
s
.
The change in our sampling procedure will alter the formulation of PrH|Tb T ⊆ H|Tb , so we
i
h
first want to write EH|Tb TH|Tb in terms of values that we are familiar with while only losing small
errors. Additionally, many of the exponential terms in the previous analysis will immediately
be absorbed into approximation error by our assumption that s ≥ 4n2 .
2
Lemma 8.3. Let G be a graph on n vertices and m edges and s a value such that m ≥ sn , Fix
some tree Tb ∈ G. For a random subset of s ≥ 4n2 edges containing Tb, H|Tb ⊇ Tb, we have
h
i
EH|Tb TH|Tb =
1−O
n2
s
n−1
X
X
pn−1−k
k=0
w (T ),
T : |T ∩Tb|=k
where p = s/m is the probability of each edge being picked in the sample.
Proof. Given that all edges of Tb are in H|Tb , the remaining s − n + 1 edges are chosen uniformly from all m − n + 1 edges not in Tb. Accordingly, for any tree T ∈ G, the probability
PrH|Tb T ⊆ H|Tb is obtained by dividing the number of subsets of s − n + 1 edges that contain
all edges in T \ Tb, against the number of subsets of s − n + 1 edges from m − n + 1:
PrH|Tb T ⊆ H|Tb
m − n + 1 − T \ Tb
(s − n + 1)|T \Tb|
m−n+1
.
/
=
=
(m − n + 1)|T \Tb|
s−n+1
s − n + 1 − T \ Tb
Following the proof Lemma 3.1, this reduces to
b
b
T \T
PrH|Tb T ⊆ H|Tb = pb|T \T | exp −
2s
46
2
−O
3
n
,
s2
which we can further reduce using the assumption of s ≥ 4n2 to:
2
n
b
PrH|Tb T ⊆ H|Tb = 1 − O
pb|T \T | ,
s
and in turn obtain via linearity of expectation:
2 X
i
h
n
b
w (T )b
p|T \T | .
EH|Tb TH|Tb = 1 − O
s
T
We then subdivide the summation based on the amount of edges in the intersection of T and
Tb and move our pb term inside the summation
h
i
EH|Tb TH|Tb =
1−O
n2
s
n−1
X
k=0
pbn−1−k
X
w (T ).
T : T ∩Tb=k
Finally, we can use Equation 4 to replace pb by p because
2n2
n n
≥ 1−
1 ≥ 1−
s
s
where n2 s < 0.1.
We will also require a strong lower bound of the expectation. The following lemma shows
that most of the trees do not intersect with Tb. Restricting ourh consideration
to such trees will
i
be much easier to work in obtaining the lower bound on EH|Tb TH|Tb .
Lemma 8.4. Let G be a graph on n vertices and m edges such that m ≥ 4n2 and all edges have
b
statistical leverage scores ≤ 2n
m . For any tree T ∈ G.
X
T : |T ∩Tb|=0
w (T ) ≥
1−O
n2
s
TG .
Proof. By definition, we can classify the trees by their intersection with Tb:
TG =
n−1
X
X
w (T ).
k=0 T : |T ∩Tb|=k
Consider each inner summation and further separating into each possible forest of Tb with k
edges gives:
X X
X
X X
w (T ) ≤
w (T ).
w (T ) =
T
:
F
⊆T
T
b
b
b
F ⊆T
T : |T ∩T |=k
F ⊆T
b
|F |=k F =T ∩T
47
|F |=k
Invoking Lemma 3.4 on the inner summation and the fact that there are
with k edges, gives an upper bound of
X
T : |T ∩Tb|=k
w (T ) ≤
k
2 k
n−1
2n
2n
≤ TG
.
TG
m
m
k
n−1
k
forests of Tb
We will utilize this upper bound for all k > 0 and achieve a lower bound from rearranging
our initial summation
!
n−1
n−1
X 2n2 k
X
X
X
w (T ) ≥ TG 1 −
w (T ) = TG −
.
m
k=1
k=1
b
b
T : |T ∩T |=k
T : |T ∩T |=0
Applying the assumption of m ≥ 4n2 and Lemma 8.2 gives our desired result.
With the necessary tools in place, we will now give upper and lower bounds on the expectation
in terms of TG pn−1 , which we note is also a close approximation of EH [TH ] by our assumption
that s ≥ 4n2 .
Lemma 8.5. Let G be a graph on n vertices and m edges such that all edges have statistical
s2
b
leverage scores ≤ 2n
m , and let s be such that m ≥ n . Fix some tree T ∈ G. For a random subset
2
of s ≥ 4n edges that contain Tb, H|Tb ⊆ Tb we have:
2
h
i
n
TG pn−1 .
EH|Tb TH|Tb = 1 ± O
s
Proof. We will first prove the upper bound. From Lemma 8.3 we have
h
i n−1
X
EH|Tb TH|Tb ≤
pn−1−k
k=0
X
w (T ) ,
T : |T ∩Tb|=k
while a proof similar to Lemma 8.4 gives
X
T : |T ∩Tb|=k
w (T ) ≤ TG
2n2
m
Moving pn−1 outside the summation and substituting
h
i
EH|Tb TH|Tb ≤ TG p
n−1
n−1
X
k=0
k
s
m
.
for p gives
2n2
s
k
,
and applying Corollary 8.2 to upper bound the summation gives
2
i
h
n
TG pn−1 .
EH|Tb TH|Tb ≤ 1 + O
s
48
For the lower bound, we again first using Lemma 8.3 and then restrict to trees that do not
intersect Tb using Lemma 8.4. Formally we have:
2 n−1
h
i
X
n
EH|Tb TH|Tb = 1 − O
pn−1−k
s
k=0
≥
8.2
1−O
w (T )
T : |T ∩Tb|=k
2
n
s
X
p
n−1
X
T : |T ∩Tb|=0
w (T ) ≥
1−O
n2
s
pn−1 TG .
Upper Bound on Conditional Variance
i
h
2
The bound on variance is by upper bounding EH|Tb TH| b in a way similar to Lemma 3.6. Once
T
again, the assumption of s > 4n2 means the situation is simpler because the exponential term
is negligible.
As with the proof of Lemma 3.6, we will often separate summations of pairs of trees based
upon the number of edges in their intersection, then frequently invoke Lemma 3.4. However
there will be more moving pieces in each summation due to intersections with Tb, so Lemma 8.7
proven later in this section, which is analogous to Lemma 3.5, will be much more involved.
Lemma 8.6. Let G be a graph on n vertices and m edges such that all edges have statistical
s2
b
leverage scores ≤ 2n
m , and s a sample count such that m ≥ n . For some tree T ∈ G, let H|Tb
denote a random subset of s edges such that Tb ⊆ H|Tb , then:
i
h
2
2
EH|Tb TH|
n
b
T
.
h
i2 ≤ 1 + O
s
EH|Tb TH|Tb
Proof. By analogous reasoning to the proof in Lemma 8.3, for any pair of trees T1 , T2 ∈ G we
have
PrH|Tb T1 , T2 ⊆ H|Tb
m − n + 1 − (T ∪ T ) \ Tb
(s − n + 1)|(T1 ∪T2 )\Tb|
1
2
m−n+1
/
=
=
.
s−n+1
(m − n + 1)|(T1 ∪T2 )\Tb|
s − n + 1 − (T ∪ T ) \ Tb
1
2
s−k
≤
As a consequence of Equation 4, specifically the bound m−k
the upper bound
b
PrH|Tb T1 , T2 ⊆ H|Tb ≤ p|(T1 ∪T2 )\T | ,
s
m
when k ≥ 0, we can obtain
and in turn summing over all pairs of trees:
i
h
X
b
2
w (T1 ) w (T2 ) p|(T1 ∪T2 )\T | .
≤
EH|Tb TH|
b
T
T1 ,T2
49
We note that |(T1 ∪ T2 ) \ Tb| = |T1 \ Tb| + |T2 \ Tb| − |(T1 ∩ T2 ) \ Tb|. Furthermore, |T1 \ Tb| =
n − 1 − |T1 ∩ Tb, so we separate the summation as per usual by each possible size of |T1 ∩ Tb| and
|T2 ∩ Tb|, and bring the terms outside of the summation that only depend on these values.
i
h
X
X
b
2n−2
2
−k1 −k2
≤
p
w (T1 ) w (T2 ) p−(T1 ∩T2 )\T .
EH|Tb TH|
p
b
T
T1 ,T2
k1 ,k2
|T1 ∩Tb|=k1
|T2 ∩Tb|=k2
In order to deal with the inner most summation we will need to again separate based on the size
of |(T1 ∪ T2 ) \ Tb|, and we further note that |(T1 ∩ T2 ) \ Tb| = |(T1 \ Tb) ∩ (T2 \ Tb)|:
n−1
i
h
X
X
2n−2
2
−k1 −k2
≤
p
EH|Tb TH|
p
p−k
b
T
k1 ,k2
X
w (T1 ) w (T2 ) .
T1 ,T2
k=0
|T1 ∩Tb|=k1
|T2 ∩Tb|=k2
|(T1 \Tb)∩(T2 \Tb)|=k
The last term is bounded in Lemma 8.7, which is stated and proven immediately after this.
Incorporating the resulting bound, and grouping the terms by the summations over k1 , k2 , and
k respectively gives:
2k+k1 +k2
m
n
n
2n
p
p
TG2
≤p
EH|Tb
m
k
k1
k2
k=0
k1 ,k2
k2
2k !
n−1
n−1
n−1
k1
X
X
X
n
2n
m
2n
n
2n
.
p−k2
p−k
= TG2 p2n−2
p−k1
m
m
m
k2
k
k1
h
2
TH|
b
T
i
2n−2
X
−k1 −k2
n−1
X
−k
k2 =0
k1 =0
k=0
for p in each summation and use the very crude upper bound ab ≤ ab :
!
n−1
n−1
n−1
i
h
X 2n2 k2
X 2n2 k1
X 2n2 k
2 2n−2
2
.
EH|Tb TH| b ≤ TG p
T
s
s
s
We then plug in
s
m
k2 =0
k1 =0
k=0
Lemma 8.2 then upper bounds each summation by 1 + O(n2 /s), giving
2
i
h
n
2
EH|Tb TH| b ≤ 1 + O
TG2 p2n−2 .
T
s
It remains to prove the following bound on the number of of pairs of trees with a certain
intersection size with Tb, and each other. The following Lemma is a generalization to Lemma 3.5,
and is proven analogously using the negative correlation of edges in spanning trees from Fact 3.3
and Lemma 3.4.
50
Lemma 8.7. Let G be graph with m edges and n vertices such that every edges has leverage
b
score ≤ 2n
m . For any tree T ∈ G and any integers k, k1 , k2 ∈ [0, n − 1],
X
T1 ,T2
|T1 ∩Tb|=k1
|T2 ∩Tb|=k2
2k+k1 +k2
m
n
n
2n
TG2 .
w (T1 ) w (T2 ) ≤
m
k
k1
k2
|(T1 \Tb)∩(T2 \Tb)|=k
Proof. We will first separate the summation over all possible forests F of size k that could be
the intersection of T1 \ Tb and T2 \ Tb:
X
X
X
w (T1 ) w (T2 ) =
w (T1 ) w (T2 ) .
F ⊆E
|F |=k
T1 ,T2
|T1 ∩Tb|=k1
|T2 ∩Tb|=k2
|(T1 \Tb)∩(T2 \Tb)|=k
T1 ,T2
|T1 ∩Tb|=k1
|T2 ∩Tb|=k2
F =(T1 \Tb)∩(T2 \Tb)
We first consider the inner summation, and will relax the requirement to only needing
F ⊆ (T1 \ Tb) ∩ (T2 \ Tb),
which we note is equivalent to F ⊆ (T1 \ Tb) and F ⊆ (T2 \ Tb). This then allows us to separate
the summation again for a particular F into terms involving just T1 and T2 :
X
T1 ,T2
|T1 ∩Tb|=k1
|T2 ∩Tb|=k2
F =(T1 \Tb)∩(T2 \Tb)
w (T1 ) w (T2 ) ≤
T
1:
X
|T1 ∩Tb|=k1
F ⊆(T1 \Tb)
w (T1 )
T
2:
X
|T2 ∩Tb|=k2
F ⊆(T2 \Tb)
w (T2 )
.
We further examine the first term in the product, and the second will follow equivalently.
Once again, we will split the summation by all possible forests Fb of Tb with size k1 that T1 \ Tb
could intersect in, and further relax to them only having to contain Fb.
X
X
X
w (T1 ) .
w (T1 ) ≤
T
1
T1 : |T1 ∩Tb|=k1
Fb⊆Tb
|Fb|=k1 Fb⊆(T1 ∩Tb)
F ⊆(T1 \Tb)
F ⊆(T1 \Tb)
Since T1 ∩ Tb and T1 \ Tb are disjoint, we can restrict to Fb that are disjoint from F , as well as
relaxing to requiring (Fb ∪ F ) ⊆ T1 (instead of Fb ⊆ (T1 ∩ Tb) and F ⊆ (T1 \ Tb)):
X
X
X
w (T ) .
w (T1 ) ≤
b
b
b
b
T1 : |T1 ∩T |=k1
F ⊆T
(F ∪F )⊆T
|Fb|=k1
F ⊆(T1 \Tb)
(Fb∩F )=∅
51
The assumption of Fb and F being disjoint means their union must have exactly k + k1 edges.
We can
then apply Lemma 3.4 to the inner summation and use the fact that there are at most
n−1
sets Fb to achieve the upper bound
k1
X
k+k1
n
2n
w (T1 ) ≤
TG .
k1
m
X
k+k2
n
2n
TG ,
w (T2 ) ≤
k2
m
T1 : |T1 ∩Tb|=k1
F ⊆(T1 \Tb)
Similarly, we can also obtain
T2 : |T2 ∩Tb|=k2
F ⊆(T2 \Tb)
m
k
which, along with the fact that there are
8.3
edge sets F of size k, gives our desired bound.
Concentration of Inverse Probabilities
We now complete a proof of Lemma 8.1 using the concentration results on the number of trees
in a sampled graph, conditioned upon a certain tree being contained in the graph.
Proof of Lemma 8.1. The definition of
−1
TH|Tb
P rH|Tb Tb
=
w (Tb)
and Lemma 3.1 give
h
PrH∼H Tb ⊆ H
i−1
·EH|Tb ∼H|
b
T
i
h
2
3 E
−1 1 n−1
T
H|
∼H
H|
b
b
|
n
n
T
T
Tb
P rH|Tb Tb
=
.
exp
+O
p
2s
s2
w Tb
Our condition of s ≥ 4n2 allows us to bound the term hexp(ni2 /(2s) + O(n3 /s2 )) by (1 +
O(n2 /s)), and incorporating our approximation of EH|Tb ∼H| TH|Tb from Lemma 8.5 gives
b
T
h
PrH∼H Tb ⊆ H
i−1
· EH|Tb ∼H|
b
T
2
−1
n
TG
H|Tb b
Pr
T
= 1±O
· ,
s
w Tb
−1
and the definition of P rG Tb
implies the bounds on expectation.
For the variance bound, we use the identity
−1
−2
VarH| ∼H
P rH|Tb Tb
= EH| ∼H
P rH|Tb Tb
− EH| ∼H
b
T
|b
T
which by the definition
b
T
|b
T
b
T
−1
TH|Tb
=
P rH|Tb Tb
w (Tb)
52
|b
T
−1 2
P rH|Tb Tb
,
reduces to
VarH|Tb ∼H|
b
T
i
h
i2
h
2
2
−1 EH| b ∼H| TH|
−
E
T
H|
∼H
H|
b
b
|b
T 2 p2n−2
n
T
T
T
b
b
T
H|Tb b
T
T
Pr
,
T
=
≤O
· G
2
s
w (Tb)2
w Tb
where the last inequality is from incorporating Lemmas 8.5 and 8.6. Applying Lemma 3.1, and
once again using the condition of s ≥ 4n2 to bound
2
2
n
n
n3
exp
+ O( 2 ≤ 1 + O
≤ O (1)
2s
s
s
gives:
2
−1
TG2
n
H|Tb b
PrH∼H
· VarH|Tb ∼H| P r
T
≤O
·
,
b
s
T
w (Tb)2
−1
and the variance bound follows from the definition of P r G Tb
.
9
h
Tb ⊆ H
i−2
Bounding Total Variation Distance
In this section we will first bound the total variation distance between drawing a tree from the
w -uniform distribution of G, and uniformly sampling s edges, H, from G, then drawing a tree
from the w -uniform distribution of H. The first bound will only be based on a concentration
for the number of trees in H, and will give the Õ(n13/6 ) time algorithm for sampling spanning
trees from Corollary 3.8.
Next we will give a more general bound on the total variation distance between two distributions based on concentration of inverse probabilities. The resulting Lemma 7.17 is used
for proving the bound on total variation distance in the recursive algorithm given in Section 7.
However, as this bound requires a higher sample count of about n2 , the direct derivation of TV
distances from concentration bounds is still necessary for uses of the Õ(n1.5 ) edge sparsifier in
Corollary 3.8.
9.1
Simple Total Variation Distance Bound from Concentration Bounds
We give here a proof of total variation distance being bounded based on the concentration of
spanning trees in the sampled graph.
Proof. (of Lemma 3.7) Substituting the definition of p and p̃ into the definition of total variation
distance gives:
i
h
X
dT V (p, p̃) =
P rG Tb − EH∼H P rH Tb
.
Tb
Substituting in the conditions of:
w H (T )
,
(by definition of P rH (Tb))
P rH Tb =
TH
h
i−1 E ′
H ∼H [TH ′ ]
w H Tb = w G Tb · PrH ′ ∼H Tb ⊆ H ′
·
,
TG
53
(by given condition)
Using the fact that
h
i
h
i
EH∼H 1 Tb ⊆ H = PrH ′ ∼H Tb ⊆ H ′ ,
we can distribute the first term into:
h
i−1
X
′
G
H
EH∼H 1 Tb ⊆ H · PrH ′ ∼H Tb ⊆ H
dT V (p, p̃) =
· P r Tb − P r
Tb
,
T̂
which by the condition on w H (Tb) simplifies to:
X
dT V (p, p̃) =
T̂
EH∼H 1 Tb ⊆ H ·
w H Tb
EH ′ ∼H [TH ′ ]
As 1(Tb ⊆ H) = 1 iff P rH (Tb) > 0, this further simplifies into
dT V (p, p̃) =
X
T̂
h
i
h
i
PrH ′ ∼H′ Tb ⊆ H ′
which by triangle inequality gives:
dT V (p, p̃) =
X
T̂
EH∼H|T
PrH ′ ∼H′ Tb ⊆ H ′ · EH∼H|T
− P r H Tb .
w H Tb
− P rH Tb ,
w H Tb
− P rH Tb ,
EH ′ ∼H [TH ′ ]
EH ′ ∼H [TH ′ ]
at which point we can rearrange the summation to obtain:
w H Tb
X
X
1
1
= EH
.
−
dT V (p, p̃) ≤ EH
w H Tb ·
P rH Tb −
EH ′ [TH ′ ]
TH
EH ′ [TH ′ ]
Tb⊆H
Tb⊆H
which by definition of TH simplifies to:
dT V (p, p̃) ≤ EH
TH
1−
EH ′ [TH ′ ]
.
By the Cauchy-Schwarz inequality, which for distributions can be instantiated as EX [f (X)] ≤
p
EX [f (X)2 ] for any random variable X and function f (X), we then get:
v
v
"
"
u
2 # u
2 #
u
u
√
T
T
H
H
dT V (p, p̃) ≤ tEH
= tE H
− 1 = δ.
1−
EH ′ [TH ′ ]
EH ′ [TH ′ ]
54
9.2
Total Variation Distance Bound from Inverse Probability Concentration
We give here our proof of Lemma 7.17, that is a more general bound on total variation distance
based upon concentration results of the inverse probabilities.
Lemma 9.1. Let X be a random variable such that X > 0 over its entire support, and given
some δ ≥ 0, such that E [X] = (1 ± δ)µ and Var [X] ≤ δµ2 , then
h
i
√
1
Pr |X −1 − µ−1 | > 4k δµ−1 ≤ 2
k
if 1 < k < δ−1/2 /4
Proof. Chebyshev’s inequality gives
h
√ i
1
Pr |X − (1 ± δ)µ| > k δµ ≤ 2 .
k
Furthermore, if we assume X such that
√
|X − (1 ± δ)µ| ≤ k δµ
which reduces to
√
√
1 − 2k δ µ ≤ X ≤ 1 + 2k δ µ.
Inverting and reversing the inequalities gives
µ−1
µ−1
√ ≤ X −1 ≤
√ .
1 + 2k δ
1 − 2k δ
Using the fact that
we can then conclude,
1
1+ǫ
= 1−
ǫ
1+ǫ
≤ 1 − ǫ for ǫ > 0, and
1
1+ǫ
= 1+
ǫ
1−ǫ
≤ 1 + 2ǫ for ǫ ≤ 1/2,
√
√
1 − 4k δ µ−1 ≤ X −1 ≤ 1 + 4k δ µ−1 ,
which implies
Pr
h
X
and proves the lemma.
−1
−µ
−1
√
> 4k δµ
−1
i
h
√ i
≤ Pr |X − (1 ± δ) µ| > k δµ
This bound does not allow us to bound EX [|X − µ] because when X close to 0, the value
of X −1 can be arbitrarily large, while this bound only bounds the probability of such events
by O(δ−1 ). We handle this by treating the case of X small separately, and account for the
total probability of such cases via summations over I and xb. First we show that once these
distributions are truncated to avoid the small X case, its variance is bounded.
Lemma 9.2. Let Y be a random variable
such
µY > 0 we have 0 < Y ≤
δ,
−1
that for parameters
−1
−1
≤ δµ−2
2µY over its entire support, and that E Y
= (1 ± δ)µY , Var Y
Y , then
√
δ µY .
E [|Y − µY |] ≤ O
55
Proof. Since |Y − µY | ≤ µY , we can decompose this expected value into buckets of 2 via:
log(δ−1/2 /4)
X
E [|Y − µY |] ≤
i=0
h
i √
√
PrY |Y − µY | ≥ 2i δµY · 2i δµY ,
where the last term is from the guarantee of Y ≤ 1. Lemma 9.1 gives that each of the intermediate probability terms is bounded by O(2−2i ), while the last one is bounded by 1δ , so this gives
a total of
log(δ−1/2 )
X
√
√
2i δµY O 2−2i ≤ δµY
E [|Y − µ|] ≤
i=0
We can now complete the proof via an argument similar to the proof of Lemma 3.7 in
Section 9.1. The only additional step is the definition of BAD u , which represents the portion of
the random variable Pu with high deviation.
Proof of Lemma 7.17. For each u, we define a scaling factor corresponding to the probability
that Pu is non-zero:
def
pu+ = Prp∼Pu [p > 0] .
By triangle inequality, we have for each Pu
|1 − E [Pu ]| ≤ pu+ · Ep∼Pu |p>0
p−1
u+ − p .
We will handle the case where p is close and far from p−1
u+ separately. This requires defining
the portion of Pu with non-zero values, but large variance as
1 −1
def
−1
BAD u = p ∈ supp (Pu ) : pu+ − p > pu+ .
2
Lemma 9.1 gives that for each u,
Prp∼Pu |p>0 [p ∈ BAD u ] ≤ O
√
δ ,
which with the outer distribution and factoring the value of p−1
u+ gives gives:
√
Eu∼U pu+ · Ep∼Pu|p>0 1 (p ∈ BAD u ) · p−1
δ ,
≤O
u+
√
Eu∼U pu+ · Ep∼Pu|p>0 1 (p ∈
/ BAD u ) · p−1
δ .
≥1−O
u+
(5)
(6)
We then define the ‘fixed’ distributions Peu with the same distribution over p as Pu , but
whose values are set to p−1
u+ whenever p ∈ BADu . Lemma 9.2 then gives:
√
−1
δp
−
p
≤
O
Ep∼Peu|p>0 p−1
u+ ,
u+
56
or taken over the support of U , and written with indicator variables:
√
/ BAD u ) · p−1
−
p
≤
O
Eu∼U pu+ · Ep∼Pu |p>0 1 (p ∈
δ .
u+
Combining this with the lower bound on the mass of p−1
u+ on the complements of the bad sets
−1
from Equation 6 via the triangle inequality p ≥ p u+ − |p −1
u+ − p| gives:
√
δ ,
/ BAD u ) · p] ≥ 1 − O
Eu∼U pu+ · Ep∼Pu |p>0 [1 (p ∈
or upon taking complement again:
Eu∼U pu+ · Ep∼Pu |p>0
√
δ ,
[1 (p ∈ BAD u ) · p] ≤ O
which together with Equation 5 and the non-negativity of p−1
u+ and p gives
√
−
p
≤
O
Eu∼U pu+ · Ep∼Pu |p>0 1 (p ∈ BAD u ) · p−1
δ .
u+
Combining these two summations, and invoking the triangle inequality at the start then gives
the bound.
References
[AGM+ 10] Arash Asadpour, Michel X. Goemans, Aleksander Madry, Shayan Oveis Gharan,
and Amin Saberi. An o(log n/ log log n)-approximation algorithm for the asymmetric traveling salesman problem. In Proceedings of the Twenty-first Annual ACMSIAM Symposium on Discrete Algorithms, SODA ’10, pages 379–389, Philadelphia,
PA, USA, 2010. Society for Industrial and Applied Mathematics.
[AHLT05]
Stephen Alstrup, Jacob Holm, Kristian De Lichtenberg, and Mikkel Thorup. Maintaining information in fully dynamic trees with top trees. Acm Transactions on
Algorithms (talg), 1(2):243–264, 2005.
[Ald90]
David Aldous. The random walk construction of uniform spanning trees and uniform
labelled trees. In SIAM Journal on Discrete Mathematics, pages 450–465, 1990.
[BDKZ15a] Christos Boutsidis, Petros Drineas, Prabhanjan Kambadur, and Anastasios Zouzias.
A randomized algorithm for approximating the log determinant of a symmetric positive definite matrix. CoRR, abs/1503.00374, 2015. Available at:
http://arxiv.org/abs/1503.00374.
[BDKZ15b] Christos Boutsidis, Petros Drineas, Prabhanjan Kambadur, and Anastasios Zouzias.
A randomized algorithm for approximating the log determinant of a symmetric
positive definite matrix. CoRR, abs/1503.00374, 2015.
[BK96]
András A. Benczúr and David R. Karger. Approximating s-t minimum cuts in Õ(n2)
time. In Proceedings of the Twenty-eighth Annual ACM Symposium on Theory of
Computing, STOC ’96, pages 47–55, New York, NY, USA, 1996. ACM.
57
[BP93]
Robert Burton and Robin Pemantle. Local characteristics, entropy and limit theorems for spanning trees and domino tilings via transfer-impedances. The Annals of
Probability, pages 1329–1371, 1993.
[Bro89]
Andrei Broder. Generating random spanning trees. In Proceedings of the 30th
annual Symposium on Foundations of Computer Science, FOCS 1989, pages 442–
447, 1989.
[BS83]
Walter Baur and Volker Strassen. The complexity of partial derivatives. Theoretical
Computer Science, 22(3):317 – 330, 1983.
[CCL+ 15]
Dehua Cheng, Yu Cheng, Yan Liu, Richard Peng, and Shang-Hua Teng. Efficient
sampling for Gaussian graphical models via spectral sparsification. Proceedings
of The 28th Conference on Learning Theory, pages 364–390, 2015. Available at
http://jmlr.org/proceedings/papers/v40/Cheng15.pdf.
[CDN89]
Charles J Colbourn, Robert PJ Day, and Louis D Nel. Unranking and ranking
spanning trees of a graph. Journal of Algorithms, 10(2):271–286, 1989.
[CKP+ 17]
Michael B. Cohen, Jonathan A. Kelner, John Peebles, Richard Peng, Anup Rao,
Aaron Sidford, and Adrian Vladu. Almost-linear-time algorithms for markov chains
and new spectral primitives for directed graphs. Accepted to STOC 2017. Preprint
available at https://arxiv.org/abs/1611.00755., 2017.
[CMN96]
Charles J Colbourn, Wendy J Myrvold, and Eugene Neufeld. Two algorithms for
unranking arborescences. Journal of Algorithms, 20(2):268–281, 1996.
[Coh16]
Michael B Cohen. Nearly tight oblivious subspace embeddings by trace inequalities.
In Proceedings of the Twenty-Seventh Annual ACM-SIAM Symposium on Discrete
Algorithms, pages 278–287. SIAM, 2016.
[CP15]
Michael B. Cohen and Richard Peng. ℓp row sampling by Lewis weights. In Proceedings of the Forty-Seventh Annual ACM on Symposium on Theory of Computing, STOC ’15, pages 183–192, New York, NY, USA, 2015. ACM. Available at
http://arxiv.org/abs/1412.0588.
[DKP+ 16]
David Durfee, Rasmus Kyng, John Peebles, Anup B. Rao, and Sushant Sachdeva.
Sampling random spanning trees faster than matrix multiplication. CoRR,
abs/1611.07451, 2016.
[EGIN97]
David Eppstein, Zvi Galil, Giuseppe F. Italiano, and Amnon Nissenzweig. Sparsification—a technique for speeding up dynamic graph algorithms. J. ACM,
44(5):669–696, September 1997.
[FHHP11]
Wai Shing Fung, Ramesh Hariharan, Nicholas JA Harvey, and Debmalya Panigrahi. A general framework for graph sparsification. In Proceedings of the fortythird annual ACM symposium on Theory of computing, pages 71–80. ACM, 2011.
https://arxiv.org/abs/1004.4080.
58
[GRV09]
Navin Goyal, Luis Rademacher, and Santosh Vempala. Expanders via random
spanning trees. In Proceedings of the Twentieth Annual ACM-SIAM Symposium
on Discrete Algorithms, SODA ’09, pages 576–585, Philadelphia, PA, USA, 2009.
Society for Industrial and Applied Mathematics.
[Gue83]
Alain Guenoche. Random spanning tree. Journal of Algorithms, 4(3):214–220, 1983.
[HAB14a]
Timothy Hunter, Ahmed El Alaoui, and Alexandre M. Bayen. Computing the
log-determinant of symmetric, diagonally dominant matrices in near-linear time.
CoRR, abs/1408.1693, 2014. Available at: http://arxiv.org/abs/1408.1693.
[HAB14b]
Timothy Hunter, Ahmed El Alaoui, and Alexandre M. Bayen. Computing the
log-determinant of symmetric, diagonally dominant matrices in near-linear time.
CoRR, abs/1408.1693, 2014.
[HJ12]
Roger A Horn and Charles R Johnson. Matrix analysis. Cambridge university press,
2012.
[HMS15]
Insu Han, Dmitry Malioutov, and Jinwoo Shin. Large-scale log-determinant computation through stochastic chebyshev expansions. In ICML, pages 908–917, 2015.
Available at: https://arxiv.org/abs/1606.00942.
[HX16]
Nicholas J. A. Harvey and Keyulu Xu. Generating random spanning trees via fast
matrix multiplication. In LATIN 2016: Theoretical Informatics, volume 9644, pages
522–535, 2016.
[IL11]
Ilse C. F. Ipsen and Dean J. Lee. Determinant approximations, 2011.
[Jan94]
Svante Janson. The numbers of spanning trees, hamilton cycles and perfect matchings in a random graph. Combinatorics, Probability and Computing, 3(01):97–126,
1994.
[JKPS17]
Gorav Jindal, Pavel Kolev, Richard Peng, and Saurabh Sawlani. Density independent algorithms for sparsifying k-step random walks. CoRR, abs/1702.06110,
2017.
[Kir47]
Gustav Kirchhoff. Ü ber die auflösung der gliechungen, auf welche man bei der untersuchung der linearen vertheilung galvanischer ströme geführt wird. In Poggendorgs
Ann. Phys. Chem., pages 497–508, 1847.
[KLP+ 16]
Rasmus Kyng, Yin Tat Lee, Richard Peng, Sushant Sachdeva, and Daniel A Spielman. Sparsified cholesky and multigrid solvers for connection laplacians. In Proceedings of the 48th Annual ACM SIGACT Symposium on Theory of Computing,
pages 842–850. ACM, 2016. Available at http://arxiv.org/abs/1512.01892.
[KM09]
Jonathan Kelner and Aleksander Madry. Faster generation of random spanning trees.
In Proceedings of the 50th annual Symposium on Foundations of Computer Science, FOCS 2009, pages 13–21, 2009.
Available at
https://arxiv.org/abs/0908.1448.
59
[Kul90]
Vidyadhar G. Kulkarni. Generating random combinatorial objects. Journal of
Algorithms, 11(2):185–207, 1990.
[MST15]
Aleksander Madry, Damian Straszak, and Jakub Tarnawski. Fast generation of
random spanning trees and the effective resistance metric. In Proceedings of the
Twenty-Sixth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2015,
pages 2019–2036, 2015. Available at http://arxiv.org/pdf/1501.00267v1.pdf.
[PS14]
Richard Peng and Daniel A. Spielman. An efficient parallel solver for SDD linear
systems. In Proceedings of the 46th Annual ACM Symposium on Theory of Computing, STOC ’14, pages 333–342, New York, NY, USA, 2014. ACM. Available at
http://arxiv.org/abs/1311.3286.
[SS11]
Daniel A. Spielman and Nikhil Srivastava. Graph sparsification by effective resistances. SIAM Journal on Computing, 40(6):1913–1926, 2011.
[ST85]
Daniel Dominic Sleator and Robert Endre Tarjan. Self-adjusting binary search
trees. Journal of the ACM (JACM), 32(3):652–686, 1985.
[ST11]
Daniel A. Spielman and Shang-Hua Teng. Spectral sparsification of graphs. SIAM
J. Comput., 40(4):981–1025, July 2011.
[ST14]
Daniel A. Spielman and Shang-Hua Teng. Nearly linear time algorithms for preconditioning and solving symmetric, diagonally dominant linear systems. SIAM
Journal on Matrix Analysis and Applications, 35(3):835–885, 2014. Available at
http://arxiv.org/abs/cs/0607105.
[Tro12]
Joel A. Tropp.
User-friendly tail bounds for sums of random matrices.
Found. Comput. Math., 12(4):389–434, August 2012.
Available at
http://arxiv.org/abs/1004.4389.
[Vis12]
N. K. Vishnoi. Lx = b laplacian solvers and their algorithmic applications, 2012.
[Wil12]
Virginia Vassilevska Williams.
Multiplying matrices faster than
coppersmith-winograd.
In
Proceedings of the Forty-fourth Annual
ACM
Symposium
on
Theory
of
Computing,
STOC
’12,
pages 887–898, New York, NY, USA, 2012. ACM.
Available at:
https://pdfs.semanticscholar.org/3030/fa2aecda339d593b86a260bfab9988b42df7.pdf.
A
Deferred Proofs
We now provide detailed proofs of the combinatorial facts about random subsets of edges that
are discussed briefly in Section 3.
Proof. (of Lemma 3.1)
60
This probability is obtained by dividing the number of subsets of s edges that
the
contain
b
n − 1 edges in T , against the number of subsets of s edges from m, which using ab = (a)
,
gives:
(b)b
(m − n + 1)s−n+1 (s)s
m−n+1
m
/
=
,
s−n+1
s
(m)s (s − n + 1)s−n+1
(7)
and the two terms can be simplified by the rule (a)b /(a − k)b−k = (a)k .
Furthermore,
(a)b = a
b
1
1−
a
b−1
··· 1 −
a
= a exp
We then use the Taylor expansion of ln(1 − x) = −
= ab exp −
Pb−1
i=1 i
−
a
Pb−1
2
i=1 i
−
2a2
b
b−1
X
i=1
P∞
xi
i=1 i
i
ln 1 −
a
!
to obtain
!
2
3
3
b
b
i
b
i=1
− .... = a exp − − O
3
3a
2a
a2
Pb−1
(s)n−1
Substituting into (m)
gives
n−1
3
3
3
2
2
n
n
n
n2
n
n
n−1
n−1
−O
+O
=p
exp − − O
p
exp − +
2s 2m
s2
m2
2s
s2
3
2
n2
is absorbed by O ns2 because m ≥ sn was assumed.
where 2m
Proof. (Of Lemma 3.2)
As before, we have
|T1 ∪ T2 |2
−O
PrH [T1 , T2 ∈ H] = p|T1 ∪T2 | exp −
2s
n3
s2
!
Invoking the identity:
|T1 ∪ T2 | = 2n − 2 − |T1 ∩ T2 |
gives
PrH [T1 , T2 ∈ H] = p
Using the algebraic identity
2n−2 −k
p
3
n
(2n − 2 − k)2
−O
.
exp −
2s
s2
(2n − 2 − k)2 ≥ 4n2 + 4nk
and dropping the trailing (negative) lower order term gives:
4n2 4nk
2n−2
−k
+
,
PrH [T1 , T2 ∈ H] ≤ p
· p exp −
2s
2s
2
upon which we can pull out the 4n
2s term in the exponential to get a term that only depends k.
k
−k
Grouping the p term together with the exp( 2n
s ) term, and using the fact that exp(t) ≤ 1 + 2t
when t ≤ 0.1 then gives the result.
61
Proof. (of Lemma 3.5) We first separate the summation in terms of all possible forests F of size
k that any pair of trees could intersect in
X
T1 ,T2
|T1 ∩T2 |=k
X
w (T1 ) · w (T2 ) =
X
F ⊆E T1 ,T2
|F |=k F =T1 ∩T2
w (T1 ) · w (T2 )
We then consider the inner summation, the number of pairs of trees T1 , T2 with T1 ∩ T2 = F
for some particular set F of size k. This is upper bounded by the square of the number of trees
containing F :
X
T1 ,T2
F =T1 ∩T2
w (T1 ) · w (T2 ) ≤
X
T1 ,T2
F ⊆T1 ∩T2
w (T1 ) · w (T2 ) =
X
T :F ⊆T
2
w (T )
This allow us to directly incorporate the bounds from Lemma 3.4, and in turn the assumption
n
of τ e ≤ m
to obtain the bound:
X
T1 ,T2
F =T1 ∩T2
w (T1 ) · w (T2 ) ≤
TG
n k 2
m
.
Furthermore, the number of possible subsets of F is bounded by
k
even more crudely by mk! . Incorporating this then gives:
X
T1 ,T2
|T1 ∩T2 |=k
m
k ,
which can be bounded
2
2 k
mk
n k
n
2 1
w (T1 ) · w (T2 ) ≤
= TG ·
· TG
.
k!
m
k! m
62
| 8cs.DS
|
Computing the L1 Geodesic Diameter and Center
of a Polygonal Domain∗
Sang Won Bae1 , Matias Korman2 , Joseph S.B. Mitchell3 , Yoshio
Okamoto4 , Valentin Polishchuk5 , and Haitao Wang6
arXiv:1512.07160v2 [cs.CG] 9 Nov 2016
1
2
3
4
5
6
Kyonggi University, Suwon, South Korea. [email protected]
Tohoku University, Sendai, Japan. [email protected]
Stony Brook University, New York, USA. [email protected]
The University of Electro-Communications, Tokyo, Japan. [email protected]
Linköping University, Linköping, Sweden. [email protected]
Utah State University, Utah, USA. [email protected]
Abstract
For a polygonal domain with h holes and a total of n vertices, we present algorithms that compute
the L1 geodesic diameter in O(n2 + h4 ) time and the L1 geodesic center in O((n4 + n2 h4 )α(n))
time, respectively, where α(·) denotes the inverse Ackermann function. No algorithms were
known for these problems before. For the Euclidean counterpart, the best algorithms compute
the geodesic diameter in O(n7.73 ) or O(n7 (h + log n)) time, and compute the geodesic center in
O(n11 log n) time. Therefore, our algorithms are significantly faster than the algorithms for the
Euclidean problems. Our algorithms are based on several interesting observations on L1 shortest
paths in polygonal domains.
Keywords and phrases geodesic diameter, geodesic center, shortest paths, polygonal domains,
L1 metric
1
Introduction
A polygonal domain P is a closed and connected polygonal region in the plane R2 , with h ≥ 0
holes (i.e., simple polygons). Let n be the total number of vertices of P. Regarding the
boundary of P as obstacles, we consider shortest obstacle-avoiding paths lying in P between
any two points p, q ∈ P. Their geodesic distance d(p, q) is the length of a shortest path
between p and q in P. The geodesic diameter (or simply diameter) of P is the maximum
geodesic distance over all pairs of points p, q ∈ P, i.e., maxp∈P maxq∈P d(p, q). Closely
related to the diameter is the min-max quantity minp∈P maxq∈P d(p, q), in which a point p∗
that minimizes maxq∈P d(p∗ , q) is called a geodesic center (or simply center) of P. Each of
the above quantities is called Euclidean or L1 depending on which of the Euclidean or L1
metric is adopted to measure the length of paths.
For simple polygons (i.e., h = 0), the Euclidean geodesic diameter and center have been
studied since the 1980s [3, 8, 23]. For the diameter, Chazelle [8] gave the first O(n2 )-time
algorithm, followed by an O(n log n)-time algorithm by Suri [23]. Finally, Hershberger and
Suri [15] gave a linear-time algorithm for computing the diameter. For the center, after an
O(n4 log n)-time algorithm by Asano and Toussaint [3], Pollack, Sharir, and Rote [21] gave
an O(n log n) time algorithm for computing the geodesic center. Recently, Ahn et al. [1]
solved the problem in O(n) time.
∗
A preliminary version of this paper appeared in the Proceedings of the 33rd International Symposium
on Theoretical Aspects of Computer Science (STACS 2016).
© S.W. Bae, M. Korman, J.S.B. Mitchell, Y. Okamoto, V. Polishchuk, and H. Wang;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
2
The L1 Geodesic Diameter and Center
For the general case (i.e., h > 0), the problems are more difficult. The Euclidean diameter
problem was solved in O(n7.73 ) or O(n7 (h + log n)) time [4]. The Euclidean center problem
was first solved in O(n12+ ) time for any > 0 [5] and then an improved O(n11 log n) time
algorithm was given in [24].
For the L1 versions, the geodesic diameter and center of simple polygons can be computed
in linear time [6, 22], but we are unaware of any previous algorithms for polygonal domains.
In this paper, we present the first algorithms that compute the geodesic diameter and center
of a polygonal domain P (as defined above) in O(n2 + h4 ) and O((n4 + n2 h4 )α(n)) time,
respectively, where α(·) is the inverse Ackermann function. Comparing with the algorithms
for the same problems under the Euclidean metric, our algorithms are much more efficient,
especially when h is significantly smaller than n.
As discussed in [4], a main difficulty of polygonal domains seemingly arises from the fact
that there can be several topologically different shortest paths between two points, which is
not the case for simple polygons. Bae, Korman, and Okamoto [4] observed that the Euclidean
diameter can be realized by two interior points of a polygonal domain, in which case the two
points have at least five distinct shortest paths. This difficulty makes their algorithm suffer
a fairly large running time. Similar issues also arise in the L1 metric, where a diameter may
also be realized by two interior points (this can be seen by extending the examples in [4]).
We take a different approach from [4]. We first construct an O(n2 )-sized cell decomposition
of P such that the L1 geodesic distance function restricted to any pair of two cells can be
explicitly described in O(1) complexity. Consequently, the L1 diameter and center can be
obtained by exploring these cell-restricted pieces of the geodesic distance. This leads to
simple algorithms that compute the diameter in O(n4 ) time and the center in O(n6 α(n))
time. With the help of an “extended corridor structure” of P [9, 10, 11], we reduce the
O(n2 ) complexity of our decomposition to another “coarser” decomposition of O(n + h2 )
complexity; with another crucial observation (Lemma 7), one may compute the diameter
in O(n3 + h4 ) time by using our techniques for the above O(n4 ) time algorithm. One of
our main contributions is an additional series of observations (Lemmas 9 to 18) that allow
us to further reduce the running time to O(n2 + h4 ). These observations along with the
decomposition may have other applications as well. The idea for computing the center is
similar.
We are motivated to study the L1 versions of the diameter and center problems in
polygonal (even non-rectilinear) domains for several reasons. First, the L1 metric is natural
and well studied in optimization and routing problems, as it models actual costs in rectilinear
road networks and certain robotics/VLSI applications. Indeed, the L1 diameter and center
problems in the simpler setting of simply connected domains have been studied [6, 22].
Second, the L1 metric approximates the Euclidean metric. Further, improved understanding
of algorithmic results in one metric can assist in understanding in other metrics; e.g., the
continuous Dijkstra methods for L1 shortest paths of [18, 19] directly led to improved results
for Euclidean shortest paths.
1.1
Preliminaries
For any subset A ⊂ R2 , denote by ∂A the boundary of A. Denote by pq the line segment
with endpoints p and q. The L1 length of pq is defined to be |xp − xq | + |yp − yq |, where
xp and yp are the x- and y-coordinates of p, respectively, and xq and yq are the x- and
y-coordinates of q, respectively. For any polygonal path π ∈ R2 , let |π| be the L1 length of
π, which is the sum of the L1 lengths of all segments of π. In the following, a path always
refers to a polygonal path. A path is xy-monotone (or monotone for short) if every vertical
S.W. Bae, M. Korman, J.S.B. Mitchell, Y. Okamoto, V. Polishchuk, and H. Wang
or horizontal line intersects it in at most one connected component. Following is a basic
observation on the L1 length of paths in R2 , which will be used in our discussion.
I Fact 1. For any monotone path π between two points p, q ∈ R2 , |π| = |pq| holds.
We view the boundary ∂P of our polygonal domain P as a series of obstacles so that no
path in P is allowed to cross ∂P. Throughout the paper, unless otherwise stated, a shortest
path always refers to an L1 shortest path and the distance/length of a path (e.g., d(p, q))
always refers to its L1 distance/length. The diameter/center always refers to the L1 geodesic
diameter/center. For simplicity of discussion, we make a general position assumption that
no two vertices of P have the same x- or y-coordinate.
The following will also be exploited as a basic fact in further discussion.
I Fact 2 ([13, 14]). In any simple polygon P , there is a unique Euclidean shortest path π
between any two points in P . The path π is also an L1 shortest path in P .
The rest of the paper is organized as follows. In Section 2, we introduce our cell
decomposition of P and exploit it to have preliminary algorithms for computing the diameter
and center of P. The algorithms will be improved later in Section 4, based on the extended
corridor structure and new observations discussed in Section 3. One may consider the
preliminary algorithms in Section 2 relatively straightforward, but we present them for the
following reasons. First, they provide an overview on the problem structure. Second, they
will help the reader to understand the more sophisticated algorithms given in Section 4.
Third, some parts of them will also be needed in the algorithms in Section 4.
2
The Cell Decomposition and Preliminary Algorithms
In this section, we introduce our cell decomposition D of P and exploit it to have preliminary
algorithms that compute the diameter and center of P.
We first build the horizontal trapezoidal map by extending a horizontal line from each
vertex of P until each end of the line hits ∂P. Next, we compute the vertical trapezoidal
map by extending a vertical line from each vertex of P and each of the ends of the above
extended lines. We then overlay the two trapezoidal maps, resulting in a cell decomposition
D of P (e.g., see Fig. 1). The above extended horizontal or vertical line segments are called
the diagonals of D. Note that D has O(n) diagonals and O(n2 ) cells. Each cell σ of D is
bounded by two to four diagonals and at most one edge of P, and thus appears as a trapezoid
or a triangle; let Vσ be the set of vertices of D that are incident to σ (note that |Vσ | ≤ 4).
By an abuse of notation, we let D also denote the set of all the cells of the decomposition.
Each cell of D is an intersection between a trapezoid of the horizontal trapezoidal map
and another one of the vertical trapezoidal map. Two cells of D are aligned if they are
contained in the same trapezoid of the horizontal or vertical trapezoidal map, and unaligned
otherwise. Lemma 1 is crucial for computing both the diameter and the center of P.
I Lemma 1. Let σ, σ 0 be any two cells of D. For any point s ∈ σ and any point t ∈ σ 0 , if σ
and σ 0 are aligned, then d(s, t) = |st|; otherwise, there exists an L1 shortest path between s
and t that passes through two vertices v ∈ Vσ and v 0 ∈ Vσ0 (e.g., see Fig. 2).
Proof. If two cells σ, σ 0 ∈ D are aligned, then they are contained in a trapezoid τ of the
vertical or horizontal trapezoidal map. Since τ is convex, any two points in τ can be joined
by a straight segment, so we have d(s, t) = |st| for any s ∈ σ and t ∈ σ.
Now, suppose that σ and σ 0 are unaligned (e.g., see Fig. 1). Let π be any shortest path
between s and t. We first observe that π intersects one horizontal diagonal e1 of D and one
3
4
The L1 Geodesic Diameter and Center
t
t
v0
s
s
e1
e2
e1
v
e2
Figure 1 The cell decomposition D of P, and a
Figure 2 Illustrating Lemma 1: a shortest path
shortest path from s to t.
through vertices v ∈ Vσ and v 0 ∈ Vσ0 .
vertical diagonal e2 , both of which bound σ (e.g., see Fig. 1, where e1 and e2 are highlighted
with red color): otherwise, σ and σ 0 must be aligned. Since e1 and e2 are bounding σ, the
intersection v = e1 ∩ e2 is a vertex of σ (e.g., see Fig. 2). Let p1 be the first intersection of π
with e1 while we go along π from s to t. Similarly, define p2 to be the first intersection of π
with e2 .
Since e1 is horizontal and e2 is vertical, the union of the two line segments p1 v ∪ vp2
is a shortest path between p1 and p2 . We replace the portion of π between p1 and p2 by
p1 v ∪ vp2 to have another s-t path π 0 . Since p1 v ∪ vp2 is monotone, its length is equal to
|p1 p2 | by Fact 1. This implies that π 0 is a shortest path between s and t and passes through
the vertex v.
Symmetrically, the above argument can be applied to the other side, the destination t
and the cell σ 0 , which implies that π 0 can be modified to a s-t shortest path π 00 that passes
through v and simultaneously a vertex v 0 of σ 0 . The lemma thus follows.
J
2.1
Computing the Geodesic Diameter
In this section, we present an O(n4 ) time algorithm for computing the diameter of P.
The general idea is to consider every pair of cells of D separately. For each pair of
such cells σ, σ 0 ∈ D, we compute the maximum geodesic distance between σ and σ 0 , that is,
maxs∈σ,t∈σ0 d(s, t), called the (σ, σ 0 )-constrained diameter. Since D is a decomposition of P,
the diameter of P is equal to the maximum value of the constrained diameters over all pairs
of cells of D. We handle two cases depending on whether σ and σ 0 are aligned.
If σ and σ 0 are aligned, by Lemma 1, for any s ∈ σ and t ∈ σ 0 , we have d(s, t) = |st|,
i.e, the L1 distance of st. Since the L1 distance function is convex, the (σ, σ 0 )-constrained
diameter is always realized by some pair (v, v 0 ) of two vertices with v ∈ Vσ and v 0 ∈ Vσ0 . We
are thus done by checking at most 16 pairs of vertices, in O(1) time.
In the following, we assume that σ and σ 0 are unaligned. Consider any point s ∈ σ
and any point t ∈ σ 0 . For any vertex v ∈ Vσ and any vertex v 0 ∈ Vσ0 , consider the path
from s to t obtained by concatenating sv, a shortest path from v to v 0 , and v 0 t, and let
dvv0 (s, t) be its length. Lemma 1 ensures that d(s, t) = minv∈Vσ ,v0 ∈Vσ0 dvv0 (s, t). Since
dvv0 (s, t) = |sv| + |v 0 t| + d(v, v 0 ) and d(v, v 0 ) is constant over all (s, t) ∈ σ × σ 0 , the function
dvv0 is linear on σ × σ 0 . Thus, it is easy to compute the (σ, σ 0 )-constrained diameter once we
know the value of d(v, v 0 ) for every pair (v, v 0 ) of vertices.
S.W. Bae, M. Korman, J.S.B. Mitchell, Y. Okamoto, V. Polishchuk, and H. Wang
I Lemma 2. For any two cells σ, σ 0 ∈ D, the (σ, σ 0 )-constrained diameter can be computed
in constant time, provided that d(v, v 0 ) for every pair (v, v 0 ) with v ∈ Vσ and v 0 ∈ Vσ0 has
been computed.
Proof. The case where σ and σ 0 are aligned is easy as discussed above. We thus assume
they are unaligned.
Assume that we know the value of d(v, v 0 ) for every pair (v, v 0 ) with v ∈ Vσ and v 0 ∈ Vσ0 .
Recall that d(s, t) = minv∈Vσ ,v0 ∈Vσ dvv0 (s, t) and dvv0 (s, t) = |sv| + |v 0 t| + d(v, v 0 ). Further,
note that |Vσ | ≤ 4, |Vσ0 | ≤ 4.
Since each dvv0 is a linear function on its domain σ × σ 0 , its graph appears as a hyperplane
in a 5-dimensional space. Thus, the geodesic distance function d restricted on σ × σ 0
corresponds to the lower envelope of those hyperplanes. Since there are only a constant
number of pairs (v, v 0 ), the function d can also be explicitly constructed in O(1) time. Finally,
we find the highest point on the graph of d by traversing all of its faces.
J
For each vertex v of D, a straightforward method can compute d(v, v 0 ) for all other vertices
v 0 of D in O(n2 log n) time, by first computing the shortest path map SPM (v) [18, 19] in
O(n log n) time and then computing d(v, v 0 ) for all v 0 ∈ D in O(n2 log n) time. We instead
give a faster sweeping algorithm in Lemma 3 by making use of the property that all vertices
on every diagonal of D are sorted.
I Lemma 3. For each vertex v of D, we can evaluate d(v, v 0 ) for all vertices v 0 of D in
O(n2 ) time.
Proof. Our algorithm attains its efficiency by using the property that all vertices on each
diagonal of D are sorted. Specifically, suppose that D is represented by a standard data
structure, e.g., the doubly connected edge list. Then, by traversing each diagonal (either
vertical or horizontal), we can obtain a vertically or horizontally sorted list of vertices on
that diagonal.
We first compute the shortest path map SPM (v) in O(n log n) time [18, 19]. We then
apply a standard sweeping technique, say, we sweep SPM (v) by a vertical line from left to
right. The events are when the sweep line hits vertices of SPM (v), obstacle vertices of P,
or vertical diagonals of D. Note that each vertex of D is either on a vertical diagonal or an
obstacle vertex. We use the standard technique to handle the events of vertices of SPM (v)
and P, and each such event costs O(log n) time. For each event of a vertical diagonal of D,
we simply do a linear search on the sweeping status to find the cells of SPM (v) that contain
the cell vertices of D on the diagonal. Each such event takes O(n) time since each diagonal
of D has O(n) vertices. Note that the total number of events is O(n). Hence, the running
time of the sweeping algorithm is O(n2 ).
J
Thus, after O(n4 )-time preprocessing, for any two cells σ, σ 0 ∈ D, the (σ, σ 0 )-constrained
diameter can be computed in O(1) time by Lemma 2. Since D has O(n2 ) cells, it suffices to
handle at most O(n4 ) pairs of cells, resulting in O(n4 ) candidates for the diameter, and the
maximum is the diameter. Hence, we obtain the following result.
I Theorem 4. The L1 geodesic diameter of P can be computed in O(n4 ) time.
2.2
Computing the Geodesic Center
We now present an algorithm that computes an L1 center of P. The observation in Lemma 1
plays an important role in our algorithm.
5
6
The L1 Geodesic Diameter and Center
For any point q ∈ P, we define R(q) to be the maximum geodesic distance between q
and any point in P, i.e., R(q) := maxp∈P d(p, q). A center q ∗ of P is defined to be a point
with R(q ∗ ) = minq∈P R(q). Our approach is again based on the decomposition D: for each
cell σ ∈ D, we want to find a point q ∈ σ that minimizes the maximum geodesic distance
d(p, q) over all p ∈ P. We call such a point q ∈ σ a σ-constrained center. Thus, if q 0 is a
σ-constrained center, then we have R(q 0 ) = minq∈σ R(q). Clearly, the center q ∗ of P must
be a σ-constrained center for some σ ∈ D. Our algorithm thus finds a σ-constrained center
for every σ ∈ D, which at last results in O(n2 ) candidates for a center of P.
Consider any cell σ ∈ D. To compute a σ-constrained center, we investigate the function
R restricted to σ and exploit Lemma 1 again. To utilize Lemma 1, for any point q ∈ σ, we
define Rσ0 (q) := maxp∈σ0 d(p, q) for any σ 0 ∈ D. For any q ∈ σ, R(q) = maxσ0 ∈D Rσ0 (q),
that is, R is the upper envelope of all the Rσ0 on the domain σ. Our algorithm explicitly
computes the functions Rσ0 for all σ 0 ∈ D and computes the upper envelope U of the graphs
of the Rσ0 . Then, a σ-constrained center corresponds to a lowest point on U.
We observe the following for the function Rσ0 .
I Lemma 5. The function Rσ0 is piecewise linear on σ and has O(1) complexity.
Proof. Recall that Rσ0 (q) = maxp∈σ0 d(p, q) for q ∈ σ. In this proof, we regard d to be
restricted on σ × σ 0 ⊂ R4 and use a coordinate system of R4 by introducing 4 axes, x, y, u,
and w with p = (x, y) ∈ σ and q = (u, w) ∈ σ 0 . Thus, we may write Rσ0 (q) = Rσ0 (x, y) =
max(u,w)∈σ0 d(x, y, u, w).
The graph of function d consists of O(1) linear patches as shown in the proof of Lemma 2.
Once we fully identify the geodesic distance function d on σ × σ 0 , we consider its graph
S := {z = d(x, y, u, w)} for all (x, y, u, w) ∈ σ × σ 0 , which is a hypersurface in a 5-dimensional
space (σ × σ 0 ) × R ⊂ R5 with an additional axis z. We then project the graph S onto the
(x, y, z)-space. More precisely, the projection of S is the set {(x, y, z) | (x, y, u, w, z) ∈ S}.
Thus, for any (x, y) ∈ σ, Rσ0 (x, y) is determined by the highest point in the intersection
of the projection with the z-axis parallel line through point (x, y, 0). This implies that the
function Rσ0 simply corresponds to the upper envelope (in the z-coordinate) of the projection
of S. Since S consists of O(1) linear patches, so does the upper envelope of its projection,
which concludes the proof.
J
Now, we are ready to describe how to compute a σ-constrained center. We first handle
every cell σ 0 ∈ D to compute the graph of Rσ0 and thus gather its linear patches. Let Γ
be the family of those linear patches for all σ 0 ∈ D. We then compute the upper envelope
of Γ and find a lowest point on the upper envelope, which corresponds to a σ-constrained
center. Since |Γ| = O(n2 ) by Lemma 5, the upper envelope can be computed in O(n4 α(n))
time by executing the algorithm by Edelsbrunner et al. [12], where α(·) denotes the inverse
Ackermann function. The following theorem summarizes our algorithm.
I Theorem 6. An L1 geodesic center of P can be computed in O(n6 α(n)) time.
Proof. As a preprocessing, we compute all of the vertex-to-vertex geodesic distances d(v, v 0 )
for all pairs of vertices of D in O(n4 ) time. We show that for a fixed σ ∈ D, a σ-constrained
center can be computed in O(n4 α(n)) time. As discussed in the proof of Lemma 2, for any
σ 0 ∈ D, the geodesic distance function d restricted on σ × σ 0 , along with its graph D over
σ × σ 0 , can be specified in O(1) time. By Lemma 5, from D we can describe the function
Rσ0 in O(1) time. The last task is to compute the upper envelope of all Rσ0 in O(n4 α(n))
time, as discussed above, by executing the algorithm by Edelsbrunner et al. [12].
J
S.W. Bae, M. Korman, J.S.B. Mitchell, Y. Okamoto, V. Polishchuk, and H. Wang
e
b
e
b
d
d
Pi
c
Pj
Pj
x
Pi
πK
y
f
c
a
f
a
(a)
Figure 3 A triangulation T of P and the 3regular graph obtained from the dual graph of
T whose nodes and edges are depicted by black
dots and red solid curves. Each junction triangle
corresponds to a node and removing all junction
triangles results in three corridors, in this figure,
each of which corresponds to an edge of the graph.
3
(b)
Figure 4 Hourglasses HK in corridors K. The
dashed segments be and af are diagonals of junction
triangles in T . (a) HK is open. Five bays can be seen.
A bay with gate cd is shown as the shaded region. (b)
HK is closed. There are three bays and a canal, and the
shaded region depicts the canal with two gates dx and
cy.
Exploiting the Extended Corridor Structure
In this section, we briefly review the extended corridor structure of P and present new
observations, which will be crucial for our improved algorithms in Section 4. The corridor
structure has been used for solving shortest path problems [9, 16, 17]. Later some new
concepts such as “bays,” “canals,” and the “ocean” were introduced [10, 11], referred to as
the “extended corridor structure.”
3.1
The Extended Corridor Structure
Let T denote an arbitrary triangulation of P (e.g., see Figure 3). We can obtain T in
O(n log n) time or O(n + h log1+ h) time for any > 0 [7]. Let G denote the dual graph of
T , i.e., each node of G(P) corresponds to a triangle in T (P) and each edge connects two
nodes of G(P) corresponding to two triangles sharing a diagonal of T (P). Based on G, one
can obtain a planar 3-regular graph, possibly with loops and multi-edges, by repeatedly
removing all degree-one nodes and then contracting all degree-two nodes. The resulting
3-regular graph has O(h) faces, nodes, and edges [17]. Each node of the graph corresponds
to a triangle in T , called a junction triangle. The removal of all junction triangles from P
results in O(h) components, called corridors, each of which corresponds to an edge of the
graph. See Figure 3. Refer to [17] for more details.
Next we briefly review the concepts of bays, canals, and the ocean. Refer to [10, 11] for
more details.
Let P1 , . . . , Ph be the h holes of P and P0 be the outer polygon of P. For simplicity, a
hole may also refer to the unbounded region outside P0 hereafter. The boundary ∂K of a
corridor K consists of two diagonals of T and two paths along the boundary of holes Pi
and Pj , respectively (it is possible that Pi and Pj are the same hole, in which case one may
consider Pi and Pj as the above two paths respectively). Let a, b ∈ Pi and e, f ∈ Pj be the
endpoints of the two paths, respectively, such that be and f a are diagonals of T , each of
which bounds a junction triangle. See Figure 4. Let πab (resp., πef ) denote the Euclidean
shortest path from a to b (resp., e to f ) inside K. The region HK bounded by πab , πef , be,
and f a is called an hourglass, which is either open if πab ∩ πef = ∅, or closed, otherwise.
If HK is open, then both πab and πef are convex chains and are called the sides of HK ;
otherwise, HK consists of two “funnels” and a path πK = πab ∩ πef joining the two apices
7
8
The L1 Geodesic Diameter and Center
of the two funnels, called the corridor path of K. The two funnel apices (e.g., x and y in
Figure 4(b)) connected by πK are called the corridor path terminals. Note that each funnel
comprises two convex chains.
We consider the region of K minus the interior of HK , which consists of a number of
simple polygons facing (i.e., sharing an edge with) one or both of Pi and Pj . We call each
of these simple polygons a bay if it is facing a single hole, or a canal if facing both holes.
Each bay is bounded by a portion of the boundary of a hole and a segment cd between two
obstacle vertices c, d that are consecutive along a side of HK . We call the segment cd the
gate of the bay. (See Figure 4(a).) On the other hand, there exists a unique canal for each
corridor K only when HK is closed and the two holes Pi and Pj both bound the canal. The
canal in K in this case completely contains the corridor path πK . A canal has two gates xd
and yc that are two segments facing the two funnels, respectively, where x, y are the corridor
path terminals and d, c are vertices of the funnels. (See Figure 4(b).) Note that each bay or
canal is a simple polygon.
Let M ⊆ P be the union of all junction triangles, open hourglasses, and funnels. We call
M the ocean. Its boundary ∂M consists of O(h) convex vertices and O(h) reflex chains each
of which is a side of an open hourglass or of a funnel. Note that P \ M consists of all bays
and canals of P.
For convenience of discussion, define each bay/canal in such a way that they do not
contain their gates and hence their gates are contained in M; therefore, each point of P
is either in a bay/canal or in M, but not in both. After the triangulation T is obtained,
computing the ocean, all bays and canals can be done in O(n) time [10, 11, 17].
Roughly speaking, the reason we partition P into the ocean, bays, and canals is to
facilitate evaluating the distance d(s, t) for any two points s and t in P. For example, if
both s and t are in M, then we can use a similar method as in Section 2 to evaluate d(s, t).
However, the challenging case happens when one of s and t is in M and the other is in a bay
or canal.
The following lemma is one of our key observations for our improved algorithms in
Section 4. It essentially tells that for any point s ∈ P and any bay or canal A, the farthest
point of s in A is achieved on the boundary ∂A, which is similar in spirit to the simple
polygon case.
I Lemma 7. Let s ∈ P be any point and A be a bay or canal of P. Then, for any t ∈ A, there
exists t0 ∈ ∂A such that d(s, t) ≤ d(s, t0 ). Equivalently, maxt∈A d(s, t) = maxt∈∂A d(s, t).
Proof. Recall that the gates of A are not contained in A but in the ocean M. Let Ā be the
closure of A, that is, Ā consists of A and its gates. For any p, q ∈ Ā, let dA (p, q) be the L1
geodesic distance in Ā. Since Ā is a simple polygon, Fact 2 implies that there is a unique
Euclidean shortest path π2 (p, q) in Ā between any p, q ∈ Ā, and dA (p, q) = |π2 (p, q)|. In
general, we have dA (p, q) ≥ d(p, q).
Depending on whether A is a bay or a canal, our proof will consider two cases. We first
prove a basic property as follows.
A Basic Property
Consider any point s0 ∈ P and any point t0 ∈ A. Then, we claim that there exists a shortest
s0 -t0 path π with the following property (*):
(*) π crosses each gate of A at most once and each component of π ∩ Ā is the unique
Euclidean shortest path π2 (p, q) for some points p, q ∈ Ā.
S.W. Bae, M. Korman, J.S.B. Mitchell, Y. Okamoto, V. Polishchuk, and H. Wang
A
u1
g
s
B u1
t
u1
t
u2
t
g
A
(a)
t′
(b)
γ∗
y
By
u2
B u2
(c)
Figure 5 Illustration to the proof of Lemma 7 when A is a bay. (a) When s ∈ Ā and (b)(c) when
s∈
/ Ā.
Consider any shortest s0 -t0 path π 0 . If π 0 crosses a gate g of A at least twice, then let p ∈ g
and q ∈ g be the first and last points on g we encounter when walking along π 0 from s0 to
t0 . We can replace the portion of π 0 between p and q with the line segment pq by Fact 1 to
obtain another shortest path that crosses g at most once. One can repeat this procedure for
all gates of A to have a shortest path π 00 crossing each gate of A at most once. Then, we take
a connected component of π 00 ∩ Ā, which is an L1 shortest path between its two endpoints p, q
inside Ā. This implies that d(p, q) = dA (p, q) = |π2 (p, q)|, so we can replace the component
by π2 (p, q). Repeat this for all components of π 00 ∩ Ā to obtain another shortest s0 -t0 path π
with the desired property (*).
The Bay Case
To prove the lemma, we first prove the case where A is a bay. Then, A has a unique gate g.
Recall that the gate g is not contained in A. Depending on whether s is in Ā, there are two
cases.
Suppose that s ∈ Ā. Let t ∈ A be any point in A. By our claim, there exists a shortest s-t
path π in P with property (*). Since both s and t lie in Ā, π does not cross g, and is thus
contained in Ā. Moreover, by property (*), we have π = π2 (s, t). Hence, d(s, t) = dA (s, t).
If t lies on ∂A, then the lemma trivially holds. Suppose that t lies in the interior of A.
Then, we can extend the last segment of π2 (s, t) until it hits a point t0 on the boundary
∂A. See Figure 5(a). Again since Ā is a simple polygon, the extended path is indeed
π2 (s, t0 ). By the above argument, d(s, t0 ) = dA (s, t0 ) = |π2 (s, t0 )|, which is strictly larger
than |π2 (s, t)| = d(s, t).
Suppose that s ∈
/ Ā. Then, any shortest path from s to any point t ∈ A must cross the
gate g. This implies that d(s, t) = minx∈g {d(s, x) + dA (x, t)} for any t ∈ A. We show that
there exists t0 ∈ ∂A such that for any point x ∈ g, it holds that dA (x, t) ≤ dA (x, t0 ), which
implies that d(s, t) = minx∈g {d(s, x) + dA (x, t)} ≤ minx∈g {d(s, x) + dA (x, t0 )} = d(s, t0 ).
For the purpose, we consider the union of π2 (x, t) for all x ∈ g. The union forms a funnel
Fg (t) plus the Euclidean shortest path π2 (u, t) from the apex u of Fg (t) to t. If u 6= t,
then we extend the last segment of π2 (u, t) to a point t0 on the boundary ∂A, similarly to
the previous case so that dA (u, t) ≤ dA (u, t0 ) and thus dA (x, t) ≤ dA (x, t0 ) for any x ∈ g.
Otherwise, if u = t, then let u1 and u2 be the two vertices of Fg (t) adjacent to the apex
t. See Figure 5(b).
Observe that the segment u1 u2 separates t and the gate g, and hence path π2 (x, t)
for any x ∈ g crosses u1 u2 . We now claim that there exists a ray γ ∗ from t such
that as t moves along γ ∗ , dA (t, y) for any fixed y ∈ u1 u2 is nondecreasing. If the
claim is true, then we select t0 = γ ∗ ∩ ∂A, so the lemma follows since it holds that
dA (t, x) = miny∈ u1 u2 {dA (t, y) + dA (y, x)} ≤ miny∈ u1 u2 {dA (t0 , y) + dA (y, x)} = dA (t0 , x)
9
10
The L1 Geodesic Diameter and Center
d
g
Ag
x
x′
g′
c
y
Ag′
y′
Figure 6 Illustration to the proof of Lemma 7 when A is a canal with two gates g = dx and g 0 = cy.
for any x ∈ g. Next, we prove the claim.
For each y ∈ u1 u2 , let By be the L1 disk centered at y with radius dA (y, t) = |yt|.
See Figure 5(c). Since t lies on the boundary of By , as we move t along a ray γ in
some direction outwards By , |yt| is not decreasing, that is, for any p ∈ γ, |yp| ≥ |yt|.
This also implies that dA (y, p) ≥ |yp| ≥ |yt| = dA (y, t). Let hy be the set of all such
rays γ that as t moves along γ, |yt| is not decreasing. Our goal is thus to show that
T
6 ∅, and pick γ ∗ as any ray in the intersection. For the purpose, we consider
y∈u1 u2 hy =
the four quadrants—left-upper, right-upper, left-lower, and right-lower—centered at t.
Then, since the By are all L1 disks, the set hy only depends on which quadrant y belongs
to; more precisely, for any y in a common quadrant, the set hy stays constant. For
example, for any y ∈ u1 u2 lying in the right-upper quadrant, then hy is commonly the
set of all rays from t in direction between 135◦ and 315◦ , inclusively, since t lies on the
bottom-left edge of By in this case. Thus, the directions of all rays in hy span an angle of
exactly 180◦ . Moreover, u1 u2 is a line segment and thus intersects only three quadrants.
T
Therefore, y∈u1 u2 hy is equal to the intersection of at most three different sets of rays,
whose directions span an angle of 180◦ . (Figure 5(c) illustrates an example scene when
T
u1 u2 intersects three quadrants centered at t.) This implies that y∈u1 u2 hy 6= ∅.
The Canal Case
Above, we have proved the lemma for the bay case where A is a bay. Next, we turn to the
canal case: suppose that A is a canal.
Let g = dx and g 0 = cy be the two gates of A, where x and y are the two corridor path
terminals (e.g., see Figure 6). We extend g from x into the interior of A, in the direction
opposite to d. Note that due to the definition of canals, this extension always goes into the
interior of A (refer to [10] for detailed discussion). Let x0 be the first point on ∂A hit by
the extension. The line segment xx0 partitions A into two simple polygons, and the one
containing d is denoted by Ag . We consider dx0 as an edge of Ag , but for convenience of
discussion, we assume that Ag does not contain the segment dx0 . Define Ag0 analogously for
the other gate g 0 = cy. If t ∈ Ag , then we can view Ag as a “bay” with gate dx0 , and apply
the identical argument as done in the bay case, concluding that for any t ∈ Ag there exists
t0 ∈ ∂Ag such that d(s, t) ≤ d(s, t0 ). If t0 ∈ ∂A, then we are done. Otherwise, if t0 ∈ xx0 ,
then s ∈ Ag according to our analysis on the bay case. In this case, we move t0 along xx0 to
x or x0 , since max{d(s, x), d(s, x0 )} ≥ d(s, t0 ) ≥ d(s, t), we are done. The case of t ∈ Ag0 is
analogous.
S.W. Bae, M. Korman, J.S.B. Mitchell, Y. Okamoto, V. Polishchuk, and H. Wang
b Observe that for any p ∈ g,
b := A \ (Ag ∪ Ag0 ). From now on, we suppose t ∈ A.
Let A
π2 (p, t) passes through the corridor path terminal x since t ∈
/ Ag [10]; symmetrically, for
any p0 ∈ g 0 , π2 (p0 , t) passes through y. Consider any L1 shortest s-t path π in P with
property (*). We classify π into one of the following three types: (a) π lies inside A, that
is, π = π2 (s, t), (b) when walking along π from s to t, the last gate crossed by π is g, or (c)
is g 0 . Note that π falls into one of the three cases. In case (a), indeed we have s ∈ A and
d(s, t) = dA (s, t). In case (b), π consists of a shortest path from s to x and π2 (x, t), and thus
d(s, t) = d(s, x) + dA (x, t). Symmetrically, in case (c), we have d(s, t) = d(s, y) + dA (y, t).
Depending on whether s ∈ Ā or not, we handle two possibilities. In the following, we
assume property (*) when we discuss any shortest s-t path.
Suppose that s ∈
/ Ā. Then, any shortest s-t path in P must cross a gate of A. This
means that there is no shortest s-t path of type (a), and we have d(s, t) = min{d(s, x) +
b into three regions Rx , Ry , and
dA (x, t), d(s, y) + dA (y, t)}. Consider a decomposition of A
b
b | d(s, x) +
B such that Rx = {p ∈ A | d(s, x) + dA (x, p) < d(s, y) + dA (y, p)}, Ry = {p ∈ A
b
dA (x, p) > d(s, y) + dA (y, p)}, and B = {p ∈ A | d(s, x) + dA (x, p) = d(s, y) + dA (y, p)}.
b of two
This decomposition is clearly the geodesic Voronoi diagram in simple polygon A
sites {x, y} with additive weights. See Aronov [2] and Papadopoulou and Lee [20]. Also,
we have that d(s, t) = d(s, x) + dA (x, t) for any t ∈ Rx ; d(s, t) = d(s, y) + dA (y, t) for any
t ∈ Ry . The region B is called the bisector between x and y. By the property of Voronoi
b Let
diagrams [2, 20], B = ∂Rx ∩ ∂Ry and B is a path connecting two points on ∂ A.
b0 ∈ B be the intersection π2 (x, y) ∩ B. Then, dA (x, b0 ) = dA (y, b0 ) ≤ dA (x, b) = dA (y, b)
for any b ∈ B, and moreover if we move b along B in one direction from b0 , dA (x, b) is
nondecreasing. Thus, maxb∈B dA (x, b) is attained when b is an endpoint of B.
When t ∈ ∂A, the lemma is trivial. If t ∈ B, then we let t0 be the endpoint of B in direction
away from b0 . Then, by the property of the bisector B, we have dA (x, t) ≤ dA (x, t0 )
and dA (y, t) ≤ dA (y, t0 ), and hence d(s, t) ≤ d(s, t0 ). If t0 lies on the boundary ∂A, we
b \ ∂A, say on xx0 . In this case, we
conclude the lemma; otherwise, t0 may lie on ∂ A
apply the analysis of the bay case where s ∈
/ Ag to find a point t00 on ∂A such that
d(s, t0 ) ≤ d(s, t00 ).
If t lies in the interior of Rx , then we extend the last segment of π2 (x, t) until it hits
b or
a point t0 on ∂Rx . Then, we have that d(s, t0 ) ≥ d(s, t). Note that t0 lies on ∂ A
00
on B; in any case, we apply the above argument so that we can find a point t on ∂A
with d(s, t00 ) ≥ d(s, t0 ) ≥ d(s, t). The case where t lies in the interior of Ry is handled
analogously.
b of
Finally, suppose that s ∈ Ā. We again consider the L1 geodesic Voronoi diagram in A
three sites {s, x, y} with additive weights 0, d(s, x), d(s, y), respectively. As done above,
we observe that the maximum value maxt∈Ab d(s, t) is attained when t ∈ ∂A or t is a
Voronoi vertex. The former case is analyzed above. Here, we prove that the latter case
cannot happen. Note that there are three shortest paths of different types between s
and the Voronoi vertex while there are exactly two shortest paths of types (b) and (c) to
any point on the bisector between x and y. In the following, we show that the bisector
between x and y cannot appear in the Voronoi diagram, which implies that the Voronoi
diagram has no vertex.
Suppose to the contrary that the bisector between x and y appears as a nonempty
Voronoi edge of the Voronoi diagram, and that t is a point on it. That is, d(s, t) =
d(s, x) + dA (x, t) = d(s, y) + dA (y, t) > dA (s, t). Let π and π 0 be two shortest s-t paths
of type (b) and (c), respectively. Thus, π passes through x and π 0 passes through y to
b By property (*), π crosses g 0 first and then g when walking along π from s
reach t ∈ A.
11
12
The L1 Geodesic Diameter and Center
p g
x
g
q p
g′
t
s
p
x
p′ g ′
y
z
t
′
L
s
Figure 7 Illustration to the proof of Lemma 7
Figure 8 Illustrating two shortest s-t paths π (red
when A is a canal.
dotted) and π 0 (blue solid), intersecting at z, in a canal
with two gates g and g 0 .
to t, and π 0 crosses g first and then g 0 . Let p ∈ g and p0 ∈ g 0 be the last point of π 0 ∩ g
and π ∩ g 0 , respectively, when we walk along π and π 0 from s to t.
We claim that π and π 0 intersect each other in a point other than s and t. Indeed, consider
the loop L formed by the subpath of π between s and x, the segment xp, and the subpath
of π 0 between s and p. See Figure 7. Also, let U be a disk centered at p with arbitrarily
small radius and q ∈ π 0 ∩ ∂U be the point on π 0 not lying in A. If the loop L does not
separate q and t, then the subpath of π from x to t must intersect π 0 in a point other
than t (see Fig. 8), and thus the claim follows; otherwise, the subpath of π 0 from q to t
must cross L at some point other than s and t, and thus the claim also follows.
0
Let z ∈ π ∩ π 0 \ {s, t} (see Fig. 8), and πzt and πzt
be the subpath of π and π 0 , respectively,
0
from z to t. By the property of shortest paths, we have |πzt | = |πzt
| = d(z, t). Hence,
0
00
replacing πzt by πzt in π results in another shortest s-t path π . If π 00 lies inside Ā,
then we have d(s, t) = |π 00 | ≥ dA (s, t). Otherwise, π 00 crosses g 0 twice at p0 and y, and
thus replacing the subpath of π 00 from p0 to y by p0 y results in another shortest path
inside Ā. In either way, there is another L1 shortest s-t path of type (a), and hence
d(s, t) = dA (s, t), a contradiction to the assumption that t lies on the bisector between x
and y.
This finishes the proof of the lemma.
3.2
J
Shortest Paths in the Ocean M
We now discuss shortest paths in the ocean M. Recall that corridor paths are contained in
canals, but their terminals are on ∂M. By using the corridor paths and M, finding an L1
or Euclidean shortest path between two points s and t in M can be reduced to the convex
case since ∂M consists of O(h) convex chains. For example, suppose both s and t are in
M. Then, there must be a shortest s-t path π that lies in the union of M and all corridor
paths [9, 11, 17].
Consider any two points s and t in M. A shortest s-t path π(s, t) in P is a shortest
path in M that possibly contains some corridor paths. Intuitively, one may view corridor
paths as “shortcuts” among the components of the space M. As in [17], since ∂M consists
of O(h) convex vertices and O(h) reflex chains, the complementary region P 0 \ M (where
P 0 refers to the union of P and all its holes) can be partitioned into a set B of O(h) convex
objects with a total of O(n) vertices (e.g., by extending an angle-bisecting segment inward
from each convex vertex [17]). If we view the objects in B as obstacles, then π is a shortest
path avoiding all obstacles of B but possibly containing some corridor paths. Note that our
S.W. Bae, M. Korman, J.S.B. Mitchell, Y. Okamoto, V. Polishchuk, and H. Wang
σ
core
Figure 9 Illustrating the core of a convex
Figure 10 Illustrating the core-based cell decomposiobstacle: the (red) squared vertices are corridor tion DM : the (red) squared vertices are core vertices and
path terminals.
the green cell σ is a boundary cell.
algorithms can work on P and M directly without using B; but for ease of exposition, we
will discuss our algorithm with the help of B.
Each convex obstacle Q of B has at most four extreme vertices: the topmost, bottommost,
leftmost, and rightmost vertices, and there may be some corridor path terminals on the
boundary of Q. We connect the extreme vertices and the corridor path terminals on ∂Q
consecutively by line segments to obtain another polygon, denoted by core(Q) and called the
core of Q (see Figure 9). Let Pcore denote the complement of the union of all cores core(Q)
for all Q ∈ B and corridor paths in P. Note that the number of vertices of Pcore is O(h) and
M ⊆ Pcore . For s, t ∈ Pcore , let dcore (s, t) be the geodesic distance between s and t in Pcore .
The core structure leads to a more efficient way to find an L1 shortest path between two
points in P. Chen and Wang [9] proved that an L1 shortest path between s, t ∈ M in Pcore
can be locally modified to an L1 shortest path in P without increasing its L1 length.
I Lemma 8 ([9]). For any two points s and t in M, d(s, t) = dcore (s, t) holds.
Hence, to compute d(s, t) between two points s and t in M, it is sufficient to consider
only the cores and the corridor paths, that is, Pcore . We thus reduce the problem size from
O(n) to O(h). Let SPMcore (s) be a shortest path map for any source point s ∈ M. Then,
SPMcore (s) has O(h) complexity and can be computed in O(h log h) time [9].
3.3
Decomposition of the Ocean M
We introduce a core-based cell decomposition DM of the ocean M (see Figure 10) in order
to fully exploit the advantage of the core structure in designing algorithms computing the L1
geodesic diameter and center. For any Q ∈ B, the vertices of core(Q) are called core vertices.
The construction of DM is analogous to that of the previous cell decomposition D for
P. We first extend a horizontal line only from each core vertex until it hits ∂M to have a
horizontal diagonal, and then extend a vertical line from each core vertex and each endpoint
of the above horizontal diagonal. The resulting cell decomposition induced by the above
diagonals is DM . Hence, DM is constructed in M with respect to core vertices. Note that
DM consists of O(h2 ) cells and can be built in O(n log n + h2 ) time by a typical plane sweep
algorithm. We call a cell σ of DM a boundary cell if ∂σ ∩ ∂M =
6 ∅. For any boundary cell σ,
the portion ∂σ ∩ ∂M appears as a convex chain of Q ∈ B by our construction of its core
and DM ; since ∂σ ∩ ∂M may contain multiple vertices of M, the complexity of σ may not
be constant. Any non-boundary cell of DM is a rectangle bounded by four diagonals. Each
13
14
The L1 Geodesic Diameter and Center
vertex of DM is either an endpoint of its diagonal or an intersection of two diagonals; thus,
the number of vertices of DM is O(h2 ).
Below we prove an analogue of Lemma 1 for the decomposition DM of M. Let Vσ be the
set of vertices of DM incident to σ. Note that |Vσ | ≤ 4. We define the alignedness relation
between two cells of DM analogously to that for D. We then observe an analogy to Lemma 1.
I Lemma 9. Let σ, σ 0 be any two cells of DM . If they are aligned, then d(s, t) = |st| for
any s ∈ σ and t ∈ σ 0 ; otherwise, there exists a shortest s-t path in P containing two vertices
v ∈ Vσ and v 0 ∈ Vσ0 with d(s, t) = |sv| + d(v, v 0 ) + |v 0 t|.
Proof. We first discuss the case where σ and σ 0 are aligned. In this case, they are bounded
by two consecutive parallel diagonals of DM , and let S ⊂ M be the region in between the
two diagonals. Since S consists of two monotone concave chains and the two diagonals by
our construction of DM , it is not difficult to see that any s ∈ σ and t ∈ σ 0 can be joined by a
monotone path π inside S. This implies that |π| = |st| = d(s, t) by Fact 1.
Next, we consider the unaligned case. Suppose that σ and σ 0 are unaligned. By Lemma 8,
there exists a shortest s-t path π in P such that π lies inside the union M and all the corridor
paths of P. Our proof for this case is analogous to that of Lemma 1. Since σ and σ 0 are
unaligned, there are two possibilities when we walk along π from s to t: either we meet a
horizontal diagonal e1 and a vertical diagonal e2 of DM that bound σ, or enter a corridor
path via its terminal x. In the former case, we can apply the same argument as done in the
proof of Lemma 1 to show that π can be modified to pass through a vertex v ∈ Vσ with
v = e1 ∩ e2 without increasing the length of the resulting path. In the latter case, observe by
our construction of DM that x is also a vertex of DM and there is a diagonal extended from
x. If x ∈ Vσ , we are done since d(s, x) = |sx| as discussed above (any cell σ is aligned with
itself). Otherwise, there is a unique cell σ 00 6= σ ∈ DM with x ∈ Vσ00 that is aligned with σ,
and there is a common diagonal e bounding σ and σ 00 . In this case, since π passes through x,
it indeed intersects two diagonals, which means that this is the former case.
J
4
Improved Algorithms
In this section, we further explore the geometric structures and give more observations about
our decomposition. These results, together with our results in Section 3, help us to give
improved algorithms that compute the diameter and center, using a similar algorithmic
framework as in Section 2.
4.1
The Cell-to-Cell Geodesic Distance Functions
Recall that our preliminary algorithms in Section 2 rely on the nice behavior of the cell-to-cell
geodesic distance function: specifically, d restricted to σ × σ 0 for any two cells σ, σ 0 ∈ D is
the lower envelope of O(1) linear functions. We now have two different cell decompositions,
D of P and DM of M. Here, we observe analogues of Lemmas 1 and 9 for any two cells in
D ∪ DM , by extending the alignedness relation between cells in D and DM , as follows.
Consider the geodesic distance function d restricted to σ × σ 0 for any two cells σ, σ 0 ∈
D ∪ DM . We call a cell σ ∈ D ∪ DM oceanic if σ ⊂ M, or coastal, otherwise. If both
σ, σ 0 ∈ D ∪ DM are coastal, then σ, σ 0 ∈ D and the case is well understood as discussed in
Section 2. Otherwise, there are two cases: the ocean-to-ocean case where both σ and σ 0 are
oceanic, and the coast-to-ocean case where only one of them is oceanic. We discuss the two
cases below.
S.W. Bae, M. Korman, J.S.B. Mitchell, Y. Okamoto, V. Polishchuk, and H. Wang
Ocean-to-ocean For the ocean-to-ocean case, we extend the alignedness relation for all
oceanic cells in D ∪ DM . To this end, when both σ and σ 0 are in D or DM , the alignedness
has already been defined. For any two oceanic cells σ ∈ D and σ 0 ∈ DM , we define their
alignedness relation in the following way. If σ is contained in a cell σ 00 ∈ DM that is aligned
with σ 0 , then we say that σ and σ 0 are aligned. However, σ may not be contained in a cell of
DM because the endpoints of horizontal diagonals of DM that are on bay/canal gates are
not vertices of D and those endpoints create vertical diagonals in DM that are not in D. To
resolve this issue, we augment D by adding the vertical diagonals of DM to D. Specifically, for
each vertical diagonal l of DM , if no diagonal in D contains l, then we add l to D and extend
l vertically until it hits the boundary of P. In this way, we add O(h) vertical diagonals to D,
and the size of D is still O(n2 ). Further, all results we obtained before are still applicable to
the new D. With a little abuse of notation, we still use D to denote the new version of D.
Now, for any two oceanic cells σ ∈ D and σ 0 ∈ DM , there must be a unique cell σ 00 ∈ DM
that contains σ, and σ and σ 0 are defined to be aligned if and only if σ 00 and σ 0 are aligned.
Lemmas 1 and 9 are naturally extended as follows, along with this extended alignedness
relation.
I Lemma 10. Let σ, σ 0 ∈ D ∪ DM be two oceanic cells. For any s ∈ σ and t ∈ σ 0 , it holds
that d(s, t) = |st| if σ and σ 0 are aligned; otherwise, there exists a shortest s-t path that
passes through a vertex v ∈ Vσ and a vertex v 0 ∈ Vσ0 .
Proof. If both of σ and σ 0 belong to D or DM , Lemmas 1 and 9 are applied. Suppose that
σ ∈ D and σ 0 ∈ DM . If they are aligned, then σ is contained in σ 00 ∈ DM that is aligned
with σ 0 by definition. Hence, we have d(s, t) = |st| by Lemma 9.
J
Coast-to-ocean We then turn to the coast-to-ocean case. We now focus on a bay or canal
A. Since A has gates, we need to somehow incorporate the influence of its gates into the
decomposition D. To this end, we add O(1) additional diagonals into DM as follows: extend
a horizontal line from each endpoint of each gate of A until it hits ∂M, and then extend
a vertical line from each endpoint of each gate of A and each endpoint of the horizontal
A
diagonals that are added above. Let DM
denote the resulting decomposition of M. Note
A
that there are some cells of DM each of which is partitioned into O(1) cells of DM
but the
A
2
combinatorial complexity of DM is still O(h ). For any gate g of A, let Cg ⊂ P be the
cross-shaped region of points in P that can be joined with a point on g by a vertical or
horizontal line segment inside P. Since the endpoints of g are also obstacle vertices, the
A
boundary of Cg is formed by four diagonals of D. Hence, any cell in D or DM
is either
A
completely contained in Cg or interior-disjoint from Cg . A cell of D or DM in the former
case is said to be g-aligned.
A
In the following, we let σ ∈ D be any coastal cell that intersects A and σ 0 ∈ DM
be any
0
oceanic cell. Depending on whether σ and σ are g-aligned for a gate g of A, there are three
cases: (1) both cells are g-aligned; (2) σ 0 is not g-aligned; (3) σ 0 is g-aligned but σ is not.
Lemma 11 handles the first case. Lemma 12 deals with a special case for the latter two cases.
Lemma 13 is for the second case. Lemma 15 is for the third case and Lemma 14 is for proving
Lemma 15. The proof of Lemma 16 summarizes the entire algorithm for all three cases.
I Lemma 11. Suppose that σ and σ 0 are both g-aligned for a gate g of A. Then, for any
s ∈ σ and t ∈ σ 0 , we have d(s, t) = |st|.
Proof. It suffices to observe that s ∈ σ and t ∈ σ 0 in Cg can be joined by an L-shaped
rectilinear path, whose length is equal to the L1 distance between them by Fact 1.
J
15
16
The L1 Geodesic Diameter and Center
Consider any path π in P from s ∈ σ to t ∈ σ 0 , and assume π is directed from s to t. For
a gate g of A, we call π g-through if g is the last gate of A crossed by π. The path π is a
shortest g-through path if its L1 length is the smallest among all g-through paths from s to t.
Suppose π is a shortest path from s to t in P. Since σ may intersect M, if s ∈ σ is not in A,
then π may avoid A (i.e., π does not intersect A). If A is a bay, then either π avoids A or π
is a shortest g-through path for the only gate g of A; otherwise (i.e., A is a canal), either π
avoids A or π is a shortest g-through or g 0 -through path for the two gates g and g 0 of A. We
have the following lemma, which is self-evident.
I Lemma 12. Suppose that for any gate g of A, at least one of σ and σ 0 is not g-aligned.
For any s ∈ σ and t ∈ σ 0 , if there exists a shortest s-t path that avoids A, then a shortest s-t
path passes through a vertex v ∈ Vσ and another vertex v 0 ∈ Vσ0 .
We then focus on shortest g-through paths according to the g-alignedness of σ and σ 0 .
I Lemma 13. Suppose σ 0 is not g-aligned for a gate g of A and there are no shortest s-t
paths that avoid A. Then, for any s ∈ σ and t ∈ σ 0 , there exists a shortest g-through s-t path
containing a vertex v ∈ Vσ and a vertex v 0 ∈ Vσ0 .
Proof. If A is a bay, since there are no shortest s-t paths that avoid A, s ∈ σ must be
contained in A, and thus there must exist g-through paths from s ∈ σ and t ∈ σ 0 . If A is a
canal, although σ may be crossed by the other gate of A, there also exist g-through paths
from s ∈ σ and t ∈ σ 0 . More specifically, if s ∈ A, then there are g-through paths from s to
any t ∈ σ 0 ; otherwise there are also g-through paths from s to any t ∈ σ 0 that cross both
gates of A.
Let π be any shortest g-through path between s ∈ σ and t ∈ σ 0 . Since π is g-through
and σ 0 is not g-aligned, π crosses a horizontal and a vertical diagonals of D that define Cg ,
and escape Cg to reach t in σ 0 . This implies that π intersects a horizontal and a vertical
diagonals defining σ and thus can be modified to pass through a vertex v ∈ Vσ of σ as done
in the proof of Lemma 1. At the opposite end t, since σ 0 ∩ Cg = ∅, we can apply the above
argument symmetrically to modify π to pass through a vertex v 0 ∈ Vσ0 of σ 0 . Thus, the
lemma follows.
J
A
The remaining case is when σ 0 ∈ DM
is g-aligned but σ ∈ D is not. Recall σ is coastal
0
and intersects A, and σ is oceanic (implying σ 0 does not intersect A).
I Lemma 14. Let g be a gate of A, and suppose that σ is not g-aligned. Then, there exists
a unique vertex vg ∈ Vσ ∩ A such that for any s ∈ σ and x ∈ g, the concatenation of segment
svg and any L1 shortest path from vg to x inside A ∪ σ results in an L1 shortest path from s
to x in A ∪ σ.
Proof. Let P := A ∪ σ. Since σ is not g-aligned, σ does not intersect the gate g. Therefore,
if A is a bay, σ must be contained in A and thus P = A; if A is a canal, σ may intersect the
other gate of A while the union P forms a simple polygon. Thus, P is a simple polygon, and
we apply Fact 2 to P . Let π2 (s, t) be the unique Euclidean shortest path between s, t ∈ P in
P . Consider the union H of π2 (s, x) for all s ∈ σ and all points x ∈ g, and suppose π2 (s, x)
is directed from s to x. Then, H forms an hourglass. We distinguish two possibilities: either
H is open or closed.
Assume that H is open. See Figure 11(a). Then, there exist s0 ∈ σ and x0 ∈ g such that
s0 x0 ⊂ P , and thus π2 (s0 , x0 ) = s0 x0 . Without loss of generality, we assume that s0 lies to the
right of and above x0 so that π2 (s0 , x0 ) is left-downwards. We then observe that the first
S.W. Bae, M. Korman, J.S.B. Mitchell, Y. Okamoto, V. Polishchuk, and H. Wang
A
vg
s′
σ
F
A
x′
σ
vg
u
g
H
g
(a)
(b)
Figure 11 Illustration to the proof of Lemma 14. (a) When H is open or (b) closed.
segment su of π(s, x) for any s ∈ σ and any x ∈ g is also left-downwards since σ is a cell of
D and is not g-aligned. This implies that the shortest path π2 (s, x) contained in H crosses
the same pair of a vertical and a horizontal diagonals that define σ; more precisely, it crosses
the left vertical and the lower horizontal diagonals of D. Letting vg be the vertex defined
by the two diagonals, we apply the same argument as in the proof of Lemma 1 to modify
π2 (s, x) to pass through vg .
Now, assume that H is closed. See Figure 11(b). Then, H has two funnels and let F be
the one that contains σ. Let u be the apex of F , that is every Euclidean shortest path in H
passes through u. Note that u is an obstacle vertex, and thus u ∈ Vσ00 for a cell σ 00 ∈ D that
is not aligned with σ. Without loss of generality, we assume that σ lies to the right of and
above σ 0 . We observe that the Euclidean shortest path π2 (s, u) for any s ∈ σ is monotone
since u is the apex of F . Since u lies to the left of and below any s ∈ σ and π2 (s, u) is
monotone, we can modify the path to pass through the bottom-left vertex vg ∈ Vσ , as in the
previous case.
J
From now on, let vg be the vertex as described in Lemma 14 (vg can be found efficiently,
as shown in the proof of Lemma 16). Consider the union of the Euclidean shortest paths
inside A from vg to all points x ∈ g. Since A is a simple polygon, the union forms a funnel
Fg (vg ) with base g, plus the Euclidean shortest path from vg to the apex of Fg (vg ). Recall
Fact 2 that any Euclidean shortest path inside a simple polygon is also an L1 shortest path.
Let Wg (vg ) be the set of horizontally and vertically extreme points in each convex chain of
Fg (vg ), that is, Wg (vg ) gathers the leftmost, rightmost, uppermost, and lowermost points in
each chain of Fg (vg ). Note that |Wg (vg )| ≤ 8 and Wg (vg ) includes the endpoints of g and
the apex of Fg (vg ). We then observe the following lemma.
I Lemma 15. Suppose that σ 0 is g-aligned but σ is not. Then, for any s ∈ σ and t ∈ σ 0 , there
exists a shortest g-through s-t path that passes through vg and some w ∈ Wg (vg ). Moreover,
the length of such a path is |svg | + d(vg , w) + |wt|.
Proof. Since A is a simple polygon, any Euclidean shortest path in A is also an L1 shortest
path by Fact 2. Thus, the L1 length of a shortest path from vg to any point x in the
funnel Fg (vg ) is equal to the L1 length of the unique Euclidean shortest path in A, which is
contained in Fg (vg ).
By Lemma 14 and the assumption that σ 0 is g-aligned, among the paths from s to t that
cross the gate g, there exists an L1 shortest g-through s-t path π consisting of three portions:
svg , the unique Euclidean shortest path from vg to a vertex u on a convex chain of Fg (vg ),
and ut. Let w ∈ Wg (vg ) be the last one among Wg (vg ) that we encounter during the walk
from s to t along π. Consider the segment wt, which may cross ∂Fg (vg ). If wt ∩ ∂Fg (vg ) = ∅,
17
18
The L1 Geodesic Diameter and Center
then we are done by replacing the subpath of π from u to t by wt. Otherwise, wt crosses
∂Fg (vg ) at two points p, q ∈ ∂Fg (vg ). Since Wg (vg ) includes all extreme points of each chain
of Fg (vg ), there is no w0 ∈ Wg (vg ) on the subchain of Fg (vg ) between p and q. Hence, we
can replace the subpath of π from w to t by a monotone path from w to t, which consists
of wp, the convex path from p to q along ∂Fg (vg ), and qt, and the L1 length of the above
monotone path is equal to |wt| by Fact 1. Consequently, the resulting path is also an L1
shortest path with the desired property.
J
For any cell σ ∈ D ∪ DM , let nσ be the combinatorial complexity of σ. If σ is a boundary
cell of DM , then nσ may not be bounded by a constant; otherwise, σ is a trapezoid or a
triangle, and thus nσ ≤ 4. The geodesic distance function d defined on σ × σ 0 for any two
cells σ, σ 0 ∈ D ∪ DM can be explicitly computed in O(nσ nσ0 ) time after some preprocessing,
as shown in Lemma 16.
I Lemma 16. Let σ be any cell of D or DM . After O(n)-time preprocessing, the function d
on σ × σ 0 for any cell σ 0 ∈ D ∪ DM can be explicitly computed in O(nσ nσ0 ) time, provided
that d(v, v 0 ) has been computed for any v ∈ Vσ and any v 0 ∈ Vσ0 . Moreover, d on σ × σ 0 is
the lower envelope of O(1) linear functions.
Proof. If both σ and σ 0 are oceanic, then Lemma 10 implies that for any (s, t) ∈ σ × σ 0 ,
d(s, t) = |st| if they are aligned, or d(s, t) = minv∈Vσ ,v0 ∈Vσ0 dvv0 (s, t), where dvv0 (s, t) =
|sv| + d(v, v 0 ) + |v 0 t|. On the other hand, if σ and σ 0 are coastal, then both are cells of D
and Lemma 1 implies the same conclusion. Since |Vσ | ≤ 4 and |Vσ0 | ≤ 4 in either case, the
geodesic distance d on σ × σ is the lower envelope of at most 16 linear functions. Hence,
provided that the values of d(v, v 0 ) for all pairs (v, v 0 ) are known, the envelope can be
computed in time proportional to the complexity of the domain σ × σ 0 , which is O(nσ nσ0 ).
From now on, suppose that σ is coastal and σ 0 is oceanic. Then, σ is a cell of D and
intersects some bay or canal A. If σ 0 is also a cell of D, then Lemma 1 implies the lemma, as
discussed in Section 2; thus, we assume σ 0 is a cell of DM .
A
As above, we add diagonals extended from each endpoint of each gate of A to obtain DM
,
and specify all g-aligned cells for each gate g of A in O(n) time. In the following, let σ 0 be
A
an oceanic cell of D or of DM
. Note that a cell of DM can be partitioned into O(1) cells of
A
DM . We have two cases depending on whether A is a bay or a canal.
First, suppose that A is a bay; let g be the unique gate of A. In this case, any L1 shortest
path is g-through, provided that it intersects A, since g is unique. There are two subcases
depending on whether σ is g-aligned or not.
If σ is g-aligned, then by Lemmas 11, 12, and 13, we have d(s, t) = |st| if σ 0 is g-aligned,
or d(s, t) = minv∈Vσ ,v0 ∈Vσ0 dvv0 (s, t), otherwise, where dvv0 (s, t) = |sv| + d(v, v 0 ) + |v 0 t|.
Thus, the lemma follows by an identical argument as above.
Suppose that σ is not g-aligned. Then, σ ⊂ A since A has a unique gate g. In this
case, we need to find the vertex vg ∈ Vσ . For the purpose, we compute at most four
Euclidean shortest path maps SPMA (v) inside A for all v ∈ Vσ in O(n) time [13]. By
Fact 2, SPMA (v) is also an L1 shortest path map in A. We then specify the L1 geodesic
distance from v to all points on g, which results in a piecewise linear function fv on g.
For each v ∈ Vσ , we test whether it holds that fv (x) + |vv 0 | ≤ fv0 (x) for all x ∈ g and all
v 0 ∈ Vσ . By Lemma 14, there exists a vertex in Vσ for which the above test is passed,
and such a vertex is vg . Since each shortest path map SPMA (v) is of O(n) complexity,
all the above effort to find vg is bounded by O(n). Next, we compute the funnel Fg (vg )
and the extreme vertices Wg (vg ) as done above by exploring SPMA (vg ) in O(n) time.
S.W. Bae, M. Korman, J.S.B. Mitchell, Y. Okamoto, V. Polishchuk, and H. Wang
If σ 0 is not g-aligned, we apply Lemma 13 to obtain d(s, t) = minv∈Vσ ,v0 ∈Vσ0 dvv0 (s, t).
Thus, d is the lower envelope of at most 16 linear functions over σ × σ 0 . Otherwise,
if σ 0 is g-aligned, then we have d(s, t) = minw∈Wg (vg ) dvg w (s, t) by Lemma 15. Since
|Wg (vg )| ≤ 8, d is the lower envelope of a constant number of linear functions.
Thus, in any case, we conclude the bay case.
Now, suppose that A is a canal. Then, A has two gates g and g 0 , and σ falls into one of
the three case: (i) σ is both g-aligned and g 0 -aligned, (ii) σ is neither g-aligned nor g 0 -aligned,
or (iii) σ is g- or g 0 -aligned but not both. As a preprocessing, if σ is not g-aligned, then
we compute vg , Fg (vg ), and Wg (vg ) as done in the bay case; analogously, if not g 0 -aligned,
compute vg0 , Fg0 (vg0 ), and Wg0 (vg0 ). Note that any shortest path in P is either g-through
or g 0 -through, provided that it intersects A. Thus, d(s, t) chooses the minimum among
a shortest g-through path, a shortest g 0 -through path, and a shortest path avoiding A if
possible. We consider each of the three cases of σ.
1. Suppose that σ is both g-aligned and g 0 -aligned. In this case, if σ 0 is either g-aligned or
g 0 -aligned, then we have d(s, t) = |st| by Lemma 11. Otherwise, if σ 0 is neither g-aligned
nor g 0 -aligned, then we apply Lemmas 12 and 13 to have d(s, t) = minv∈Vσ ,v0 ∈Vσ0 dvv0 (s, t).
Hence, the lemma follows.
2. Suppose that σ is neither g-aligned nor g 0 -aligned. If σ 0 is both g-aligned and g 0 -aligned,
then by Lemma 15 the length of a shortest g-through path is equal to minw∈Wg (vg ) dvg w (s, t)
while the length of a shortest g 0 -through path is equal to minw∈Wg0 (vg0 ) dvg0 w (s, t). The
geodesic distance d(s, t) is the minimum of the above two quantities, and thus the lower
envelope of O(1) linear function on σ × σ 0 .
If σ 0 is g-aligned but not g 0 -aligned, then by Lemmas 13 and 15, we have
d(s, t) = min{
min
w∈Wg (vg )
dvg w (s, t),
min
v∈Vσ ,v 0 ∈Vσ0
dvv0 (s, t)}.
The case where σ 0 is g 0 -aligned but not g-aligned is analogous.
If σ 0 is neither g-aligned nor g 0 -aligned, then d(s, t) = minv∈Vσ ,v0 ∈Vσ0 dvv0 (s, t) by Lemma 13.
3. Suppose that σ is g 0 -aligned but not g-aligned. The other case where it is g-aligned but
not g 0 -aligned can be handled symmetrically. If σ 0 is g 0 -aligned, then we have d(s, t) = |st|
by Lemma 11. If σ 0 is neither g-aligned nor g 0 -aligned, then, by Lemmas 12 and 13,
d(s, t) = minv∈Vσ ,v0 ∈Vσ0 dvv0 (s, t).
The remaining case is when σ 0 is g-aligned but not g 0 -aligned. In this case, the length
of a shortest g-through path is equal to minw∈Wg (vg ) dvg w (s, t) by Lemma 15 for gate
g while the length of a shortest g 0 -through path is equal to minv∈Vσ ,v0 ∈Vσ0 dvv0 (s, t) by
Lemmas 12 and 13. Thus, the geodesic distance d(s, t) is the smaller of the two quantities.
Consequently, we have verified every case of (σ, σ 0 ).
As the last step of the proof, observe that it is sufficient to handle separately all the cells
A
σ 0 ∈ DM
whose union forms the original cell of DM , since any cell of DM can be decomposed
A
into O(1) cells of DM
.
J
4.2
Computing the Geodesic Diameter and Center
Lemma 7 assures that we can ignore coastal cells that are completely contained in the interior
of a bay or canal, in order to find a farthest point from any s ∈ P. This suggests a combined
set Df of cells from the two different decompositions D and DM : Let Df be the set of all
cells σ such that either σ belongs to DM or σ ∈ D is a coastal cell with ∂σ ∩ ∂P =
6 ∅. Note
that Df consists of O(h2 ) oceanic cells from DM and O(n) coastal cells from D. Since the
19
20
The L1 Geodesic Diameter and Center
boundary ∂A of any bay or canal A is covered by the cells of Df , Lemma 7 implies the
following lemma.
I Lemma 17. For any point s ∈ P, maxt∈P d(s, t) = maxσ0 ∈Df maxt∈σ0 d(s, t).
We apply the same approach as in Section 2 but we use Df instead of D.
To compute the L1 geodesic diameter, we compute the (σ, σ 0 )-constrained diameter for
each pair of cells σ, σ 0 ∈ Df . Suppose we know the value of d(v, v 0 ) for any v ∈ Vσ and any
v ∈ Vσ0 over all σ, σ 0 ∈ Df . Our algorithm handles each pair (σ, σ 0 ) of cells in Df according
to their types by applying Lemma 16. The following lemma computes d(v, v 0 ) for all cell
vertices v and v 0 of Df .
I Lemma 18. In O(n2 + h4 ) time, one can compute the geodesic distances d(v, v 0 ) between
every v ∈ Vσ and v 0 ∈ Vσ0 for all pairs of two cells σ, σ 0 ∈ Df .
Proof. Let VM be the set of vertices Vσ for all oceanic cells σ ∈ DM , and Vc be the set of
vertices Vσ for all coastal cells σ ∈ Df . Note that |VM | = O(h2 ) and |Vc | = O(n). We handle
the pairs (v, v 0 ) of vertices separately in three cases: (i) when v, v 0 ∈ VM , (ii) when v ∈ Vc
and v 0 ∈ VM , and (iii) when v, v 0 ∈ Vc .
Let v ∈ VM be any vertex. We compute the shortest path map SPMcore (v) in the core
domain Pcore as discussed in Section 3. Recall that SPMcore (v) is of O(h) complexity and can
be computed in O(n + h log h) time (after P is triangulated) [9, 11]. For any point p ∈ M, the
geodesic distance dcore (v, p) can be determined in constant time after locating the region of
SPMcore (v) that contains p. By Lemma 8, we have d(v, p) = dcore (v, p). Computing d(v, v 0 )
for all v 0 ∈ VM can be done in O(h2 ) time by running the plane sweep algorithm of Lemma 3
on SPMcore (v). Thus, for each v ∈ VM we spend O(n + h2 ) time. Since |VM | = O(h2 ), we
can compute d(v, v 0 ) for all pairs of vertices v, v 0 ∈ VM in time O(nh2 + h4 ).
Case (ii) can also be handled in a similar fashion. Let v ∈ Vc . We also show that
computing d(v, v 0 ) for all v 0 ∈ VM can be done in O(n + h2 ) time. If v lies in the ocean
M, then we can apply the same argument as in case (i). Thus, we assume v ∈
/ M. For
the purpose, we consider v as a point hole (i.e., a hole or an obstacle consisting of only
one point) into the polygonal domain P to obtain a new domain Pv , and compute the
corresponding corridor structures of Pv , which can be done in O(n + h log1+ h) time (or
O(n) time after a triangulation of P is given), as discussed in Section 3.1. Let Mv denote the
ocean corresponding to the new polygonal domain Pv . Since v lies in a bay or canal of P, M
is a subset of Mv by the definition of bays, canals and the ocean. Thus, we have VM ⊂ Mv .
0
We then compute the core structure of Pv and the shortest path map SPMcore
(v) in Mv in
0
O(n + h log h) time [9, 11]. Analogously, the complexity of SPMcore
(v) is bounded by O(h).
0
Finally, perform the plane sweep algorithm on SPMcore
(v) as in case (i) to get the values of
0
0
2
d(v, v ) for all v ∈ VM in O(h ) time.
What remains is case (iii) where v, v 0 ∈ Vc . Fix v ∈ Vc . The vertices in Vc either lie on
∂P or in its interior. In this case, we assume that we have a triangulation of P as discussed
in Section 3. Recall that we can compute the shortest path map SPM (v) in O(n + h log h)
time [9, 11]. Since SPM (v) stores d(v, v 0 ) for all obstacle vertices v 0 of P, computing d(v, v 0 )
for all v 0 ∈ Vc ∩ ∂P can be done in the same time bound by adding Vc ∩ ∂P into P as
obstacle vertices. Thus, the case where v, v 0 ∈ Vc and one of them lies in ∂P can be handled
in O(n2 + nh log h) time.
In the following, we focus on how to compute d(v, v 0 ) for all v, v 0 ∈ Vc \ ∂P. Let e1 , . . . , en
be the edges of P in an arbitrary order. We modify the original polygonal domain P to
obtain the rectified polygonal domain Prect as follows. For each i = 1, . . . , n, we define Vi to
S.W. Bae, M. Korman, J.S.B. Mitchell, Y. Okamoto, V. Polishchuk, and H. Wang
v △v
g
g
(a)
(b)
Figure 12 Illustration to how to construct Prect . (a) A bay A with gate g and the decomposition D
inside A. The dark gray region depicts the hole of P bounding A, coastal cells of D intersecting ∂A are
shaded by light gray color, and black dots are vertices in Vc \ ∂P. One of those in Vc is labeled v and its
triangle 4v is highlighted. (b) The rectified polygonal domain Prect obtained by expanding the hole into
4v for all v ∈ Vc . The boundary of Prect is depicted by solid segments. Here, the edges of A are ordered
in the counter-clockwise order.
be the set of all vertices v ∈ (Vc \ ∂P) \ (V1 ∪ · · · ∪ Vi−1 ) such that v ∈ Vσ for some coastal cell
σ ∈ Df with ∂σ ∩ ei =
6 ∅. For each v ∈ Vi , we shoot two rays from v, vertical and horizontal,
towards ei until each hits ei . Let 4v be the triangle formed by v and the two points on ei
hit by the rays. Since v is a vertex of a cell of D facing ei , by the construction of D, the two
rays must hit ei and thus the triangle 4v is well defined. We then expand each hole of P
into the triangles 4v for all v ∈ Vc \ ∂P. Let Prect be the resulting polygonal domain; that
is, every triangle 4v is regarded as an obstacle in Prect . We also add all those in Vc \ ∂P
into Prect as obstacle vertices. See Figure 12. Observe that Prect is a subset of P as subsets
of R2 and all those in Vc \ ∂P lie on the boundary of Prect as its obstacle vertices. For any
two points s, t ∈ Prect , let drect (s, t) be the L1 geodesic distance between s and t in Prect .
We then claim the following:
For any s, t ∈ Prect , it holds that d(s, t) = drect (s, t).
Suppose that the claim is true. The construction of Prect can be done in O(n2 ) time. Then,
for any v ∈ Vc \ ∂P, we compute the shortest path map SPMrect (v) in the rectified domain
Prect and obtain d(v, v 0 ) for all other v 0 ∈ Vc \ ∂P. Since Prect has h holes and O(n) vertices
by our construction of Prect , this task can be done in O(n + h log h) time [9, 11]. At last,
case (iii) can be processed in total O(n2 + nh log h) time.
We now prove the claim, as follows.
Proof of the claim. For any v ∈ Vc \ ∂P, the triangle 4v is well defined. We call a triangle
4v maximal if there is no other 4v0 with 4v ⊂ 4v0 . Note that any two maximal triangles
are interior-disjoint but may share a portion of their sides. Indeed, P \ Prect is the union of
all maximal triangles 4v . Pick any connected component C of P \ Prect . The set C is either
a maximal triangle 4v itself or the union of two maximal triangles that share a portion of
their sides by our construction of Prect and of D. In either case, observe that the portion
∂C \ ∂P is a monotone path.
Consider any s, t ∈ Prect and any shortest s-t path π in P, that is, d(s, t) = |π|. If π lies
inside Prect , then we are done since the L1 length of any s-t path inside Prect is at least
d(s, t). Otherwise, π may cross a number of connected components of P \Prect . Pick any such
connected component C that is crossed by π. Let p and q be the first and the last points on
∂C we encounter when walking along π from s to t. Since ∂C \ ∂P is monotone as observed
21
22
The L1 Geodesic Diameter and Center
above, the path πpq along the boundary of C is also monotone. By Fact 1, |πpq | = |pq| and
thus we can replace the subpath of π between p and q by πpq without increasing the L1
length. The resulting path thus has length equal to |π| and avoids the interior of C. We
repeat the above procedure for all such connected components C crossed by π. At last, the
final path π 0 has length equal to |π| and avoids the interior of P \ Prect . That is, π 0 is a
s-t path in Prect with |π 0 | = |π| = d(s, t). Since drect (s, t) ≥ d(s, t) in general, π 0 is an L1
shortest s-t path in Prect , and hence drect (s, t) = d(s, t).
This proves the above claim.
Consequently, the total time complexity is bounded by
O(nh2 + h4 ) + O(n2 + nh log h) = O(n2 + nh2 + h4 ) = O(n2 + h4 ).
J
The lemma thus follows.
Our algorithms for computing the diameter and center are summarized in the proof of
the following theorem.
I Theorem 19. The L1 geodesic diameter and center of P can be computed in O(n2 + h4 )
and O((n4 + n2 h4 )α(n)) time, respectively.
Proof. We first discuss the diameter algorithm, whose correctness follows directly from
Lemma 17.
After the execution of the procedure of Lemma 18 as a preprocessing, our algorithm
considers three cases for two cells σ, σ 0 ∈ Df : (i) both are oceanic, (ii) both are coastal, or
(iii) σ is coastal and σ 0 is oceanic. In either case, we apply Lemma 16.
P
For case (i), we have O(h2 ) oceanic cells and the total complexity is σ∈DM nσ = O(n+h2 ).
Thus, the total time for case (i) is bounded by
X X
X
O(nσ nσ0 ) =
O(nσ (n + h2 )) = O((n + h2 )2 ) = O(n2 + h4 ).
σ∈DM σ 0 ∈DM
σ∈DM
For case (ii), we have O(n) coastal cells in DM and their total complexity is O(n) since
they are all trapezoidal. Thus, the total time for case (ii) is bounded by O(n2 ).
For case (iii), we fix a coastal cell σ ∈ Df and iterate over all oceanic cells σ 0 ∈ DM ,
after an O(n)-time preprocessing, as done in the proof of Lemma 16. For each σ, we take
P
O(n + h2 ) time since σ∈DM nσ = O(n + h2 ). Thus, the total time for case (iii) is bounded
by O(n2 + nh2 ) = O((n + h2 )2 ) = O(n2 + h4 ).
Next, we discuss our algorithm for computing a geodesic center of P. We consider O(n2 )
cells σ ∈ D and compute all the σ-constrained centers. As a preprocessing, we spend O(n4 )
time to compute the geodesic distances d(v, v 0 ) for all pairs of vertices of D by Lemma 3. Fix
a cell σ ∈ D. For all σ 0 ∈ Df , we compute the geodesic distance function d restricted to σ × σ 0
by applying Lemma 16. As in Section 2, compute the graph of Rσ0 (q) = maxp∈σ0 d(p, q)
by projecting the graph of d over σ × σ 0 , and take the upper envelope of the graphs of Rσ0
for all σ 0 ∈ Df . By Lemma 16, we have an analogue of Lemma 5 and thus a σ-constrained
center can be computed in O(m2 α(m)) time, where m denotes the total complexity of all
Rσ0 . Lemma 16 implies that m = O(n + h2 ).
P
P
For the time complexity, note that σ∈DM nσ = O(n + h2 ) and σ∈Df \DM nσ = O(n).
Since any cell in D is either a triangle or a trapezoid, its complexity is O(1). Thus, for each
σ ∈ D, by Lemma 16, computing a σ-constrained center takes O((n+h2 )2 α(n)) time, after an
O(n4 )-time preprocessing (Lemma 3). Iterating over all σ ∈ D takes O(n2 (n + h2 )2 α(n)) =
O((n4 + n2 h4 )α(n)) time.
J
S.W. Bae, M. Korman, J.S.B. Mitchell, Y. Okamoto, V. Polishchuk, and H. Wang
5
Conclusions
We gave efficient algorithms for computing the L1 geodesic diameter and center of a polygonal
domain. In particular, we exploited the extended corridor structure to make the running
times depend on the number of holes in the domain (which may be much smaller than the
number of vertices). It would be interesting to find further improvements to the algorithms
in hopes of reducing the worst-case running times; it would also be interesting to prove
non-trivial lower bounds on the time complexities of the problems.
Acknowledgements. Work by S.W. Bae was supported by Basic Science Research Program
through the National Research Foundation of Korea (NRF) funded by the Ministry of
Science, ICT & Future Planning (2013R1A1A1A05006927), and by the Ministry of Education
(2015R1D1A1A01057220). M. Korman is partially supported by JSPS/MEXT Grant-in-Aid
for Scientific Research Grant Numbers 12H00855 and 15H02665. J. Mitchell acknowledges
support from the US-Israel Binational Science Foundation (grant 2010074) and the National
Science Foundation (CCF-1018388, CCF-1526406). Y. Okamoto is partially supported by JST,
CREST, Foundation of Innovative Algorithms for Big Data and JSPS/MEXT Grant-in-Aid for
Scientific Research Grant Numbers JP24106005, JP24700008, JP24220003, and JP15K00009.
V. Polishchuk is supported in part by Grant 2014-03476 from the Sweden’s innovation agency
VINNOVA and the project UTM-OK from the Swedish Transport Administration Trafikverket.
H. Wang was supported in part by the National Science Foundation (CCF-1317143).
References
1
2
3
4
5
6
7
8
9
10
H.-K. Ahn, L. Barba, P. Bose, J.-L. De Carufel, M. Korman, and E. Oh. A linear-time
algorithm for the geodesic center of a simple polygon. In Proc. of the 31st Symposium on
Computational Geometry, pages 209–223, 2015.
B. Aronov. On the geodesic Voronoi diagram of point sites in a simple polygon. Algorithmica, 4(1–4):109–140, 1989.
T. Asano and G. Toussaint. Computing the geodesic center of a simple polygon. Technical
Report SOCS-85.32, McGill University, Montreal, Canada, 1985.
S.W. Bae, M. Korman, and Y. Okamoto. The geodesic diameter of polygonal domains.
Discrete and Computational Geometry, 50:306–329, 2013.
S.W. Bae, M. Korman, and Y. Okamoto. Computing the geodesic centers of a polygonal domain. In Proc. of the 26th Canadian Conference on Computational Geometry,
2014. Journal version to appear in Computational Geometry: Theory and Applications,
http://dx.doi.org/10.1016/j.comgeo.2015.10.009.
S.W. Bae, M. Korman, Y. Okamoto, and H. Wang. Computing the L1 geodesic diameter
and center of a simple polygon in linear time. Computational Geometry: Theory and
Applications, 48:495–505, 2015.
R. Bar-Yehuda and B. Chazelle. Triangulating disjoint Jordan chains. International Journal
of Computational Geometry and Applications, 4(4):475–481, 1994.
B. Chazelle. A theorem on polygon cutting with applications. In Proc. of the 23rd Annual
Symposium on Foundations of Computer Science, pages 339–349, 1982.
D.Z. Chen and H. Wang. A nearly optimal algorithm for finding L1 shortest paths among
polygonal obstacles in the plane. In Proc. of the 19th European Symposium on Algorithms,
pages 481–492, 2011.
D.Z. Chen and H. Wang. Computing the visibility polygon of an island in a polygonal
domain. In Proc. of the 39th International Colloquium on Automata, Languages and Programming, pages 218–229, 2012. Journal version published online in Algorithmica, 2015.
23
24
The L1 Geodesic Diameter and Center
11
12
13
14
15
16
17
18
19
20
21
22
23
24
D.Z. Chen and H. Wang. L1 shortest path queries among polygonal obstacles in the plane.
In Proc. of the 30th Symposium on Theoretical Aspects of Computer Science, pages 293–304,
2013.
H. Edelsbrunner, L.J. Guibas, and M. Sharir. The upper envelope of piecewise linear
functions: Algorithms and applications. Discrete and Computational Geometry, 4:311–336,
1989.
L.J. Guibas, J. Hershberger, D. Leven, M. Sharir, and R.E. Tarjan. Linear-time algorithms
for visibility and shortest path problems inside triangulated simple polygons. Algorithmica,
2(1-4):209–233, 1987.
J. Hershberger and J. Snoeyink. Computing minimum length paths of a given homotopy
class. Computational Geometry: Theory and Applications, 4(2):63–97, 1994.
J. Hershberger and S. Suri. Matrix searching with the shortest-path metric. SIAM Journal
on Computing, 26(6):1612–1634, 1997.
R. Inkulu and S. Kapoor. Planar rectilinear shortest path computation using corridors.
Computational Geometry: Theory and Applications, 42(9):873–884, 2009.
S. Kapoor, S.N. Maheshwari, and J.S.B. Mitchell. An efficient algorithm for Euclidean
shortest paths among polygonal obstacles in the plane. Discrete and Computational Geometry, 18(4):377–383, 1997.
J.S.B. Mitchell. An optimal algorithm for shortest rectilinear paths among obstacles. In
the 1st Canadian Conference on Computational Geometry, 1989.
J.S.B. Mitchell. L1 shortest paths among polygonal obstacles in the plane. Algorithmica,
8(1):55–88, 1992.
E. Papadopoulou and D.T. Lee. A new approach for the geodesic Voronoi diagram of points
in a simple polygon and other restricted polygonal domains. Algorithmica, 20(4):319–352,
1998.
R. Pollack, M. Sharir, and G. Rote. Computing the geodesic center of a simple polygon.
Discrete and Computational Geometry, 4(1):611–626, 1989.
S. Schuierer. Computing the L1 -diameter and center of a simple rectilinear polygon. In
Proc. of the International Conference on Computing and Information, pages 214–229, 1994.
S. Suri. Computing geodesic furthest neighbors in simple polygons. Journal of Computer
and System Sciences, 39:220–235, 1989.
H. Wang. On the geodesic centers of polygonal domains. In Proc. of the 24th European
Symposium on Algorithms, pages 77:1–77:17, 2016.
| 8cs.DS
|
MMSE PRECODER FOR MASSIVE MIMO USING 1-BIT QUANTIZATION
Ovais Bin Usman, Hela Jedda, Amine Mezghani and Josef A. Nossek
arXiv:1706.08717v1 [cs.IT] 27 Jun 2017
Institute for Circuit Theory and Signal Processing
Technische Universität München, 80290 Munich, Germany
Email:{hela.jedda, josef.a.nossek}@tum.de
ABSTRACT
We propose a novel linear minimum-mean-squared-error
(MMSE) precoder design for a downlink (DL) massive
multiple-input-multiple-output (MIMO) scenario. For economical and computational efficiency reasons low resolution
1-bit digital-to-analog (DAC) and analog-to-digital (ADC)
converters are used. This comes at the cost of performance
gain that can be recovered by the large number of antennas
deployed at the base station (BS) and an appropiate precoder design to mitigate the distortions due to the coarse
quantization. The proposed precoder takes the quantization
non-linearities into account and is split into a digital precoder
and an analog precoder. We formulate the two-stage precoding problem such that the MSE of the users is minimized
under the 1-bit constraint. In the simulations, we compare the
new optimized precoding scheme with previously proposed
linear precoders in terms of uncoded bit error ratio (BER).
Index Terms— Massive MIMO, Precoding, 1-bit quantization, Transmit signal processing
1. INTRODUCTION
The massive MIMO system, or named large-scale antenna
system has been seen as a promising technology for the next
generation wireless communication systems [1, 2]. The huge
increase in the number of antennas at BS can improve spectral efficiency (SE), energy efficiency (EE) and reliability.
The BS with large number of antennas, say 100 antennas
or more, simultaneously serves a much smaller number of
single-antenna users. With the knowledge of CSI at the BS
(CSIT), this large spatial DoF of massive MIMO systems
can be exploited to significantly increase the spatial multiplexing/diversity gain using MU-MIMO precoding [3, 4].
The linear precoders, such as MF, ZF [5] and the regularized
zero-forcing (RZF) scheme [3] are shown to be near-optimal.
Thus, it is more practical to use low-complexity linear precoding techniques in massive MIMO systems. Therefore, we
mainly focus on linear precoding techniques in this work. The
price to pay for massive MIMO systems is increased complexity of the hardware (number of radio frequency (RF) and
ADC/DAC chains) and the signal processing and resulting
increased energy consumption at the transmitter [5]. Several approaches are considered in the literature to decrease
the power consumption such as spatial modulation [6], load
modulation [7], the use of parasitic antennas [8] and the use
of low-cost transceivers [9].
One attractive solution to overcome the issues of high
complexity and high energy consumption associated with
massive MIMO, is the use of very low resolution ADCs
and DACs. The power consumption of the ADC and the
DAC, one of the most power-hungry devices, can be reduced
exponentially by decreasing the resolution [10] and 1-bit
quantization can drastically simplify other RF-components,
e.g., amplifiers and mixers. Therefore, we design an MMSE
linear precoder in a DL massive MIMO scenario where the
resolution of the DACs and ADCs is restricted to 1 bit. This
precoder design aims at mitigating the distortions due to
the coarse quantization in addition to inter-user interference
(IUI). A similar work has been presented in [11], where the
authors optimize first the quantizer’s levels and then give a
closed-form expression of an MMSE precoder that takes into
account the quantizer non-linearities. However, only quantization at the transmitter was considered. In this contribution,
we do not optimize the quantizer. The quantizer in our work
has constant levels. But we introduce a second precoding
stage in the analog domain after the quantizer to minimize
the distortions due to 1-bit DAC/ADC in i.i.d. complex Gaussian channels. The proposed two-stage precoder is designed
based on iterative methods. We assume perfect CSIT and
study how the new precoder scheme is improving the BER
compared to the precoder introduced in [11].
This paper is organized as follows: in Section 2 the system
model is presented. In Section 3 some derivations related to
the 1-bit quantization are introduced. In Section 4 we formulate our optimization problem and show the derivations and
the corresponding solution. In Sections 5 and 6 we interpret
the simulation results and summarize this work.
Notation: Bold letters indicate vectors and matrices, nonbold letters express scalars. The operators (.)∗ , (.)T , (.)H and
E [.] stand for complex conjugation, the transposition, Hermitian transposition and the expectation, respectively. The n× n
identity matrix is denoted by In while the zeros (ones) matrix
with n rows and m columns is defined as 0n,m (1n,m ). We
For a complex-valued signal vector x we get
Q(x) = Q (xR + j xI ) = Q (xR ) + j Q (xI ) .
Fig. 1. System Model
define (•)R = ℜ{•}, (•)I = ℑ{•} and Q(x) = sign(xR ) +
j sign(xI ). Additionally, diag(A) denotes a diagonal matrix
containing only the diagonal elements of A. σα and ραβ denote the standard deviation of α and the correlation coefficient between α and β, respectively. For a circular distributed
Gaussian complex-valued signal α we have σαR = σαI .
2. SYSTEM MODEL
We consider a massive MIMO downlink scenario as depicted
in Fig. 1. The BS with N antennas serves M single-antenna
users, where N ≫ M . The signal vector s ∈ OM contains the data symbols for each of the M users, where O
represents the set of QPSK
constellation. We assume that
s ∼ ON 0M , σs2 IM . In this system 1-bit quantization at
the transmitter Qt as well as at the receiver Qr is deployed.
Therefore, in order to mitigate the IUI, we make use of a
two-stage precoder consisting of the digital precoder P and
the analog precoder D. The use of the 1-bit quantizer at
the transmitter Qt delivers a signal yQ that belongs to the
set of {±1 ± j }. It means that the magnitude of the entry yQ,n , n = 1, ..., N , is constant and its phase belongs to
5π 7π
{ π4 , 3π
4 , 4 , 4 }. As a result, all the antennas end up getting the same power. To recover the information loss of the
power allocation due to Qt , we employ an analog precoder
D of real-valued diagonal structure. So, we end up with
D ∈ RN ×N , P ∈ C N ×M and
yQD = D Qt (Ps), where
T
P = p1 . . . pN . The received decoded signal vector ŝ ∈ C M×1 of the M single-antenna users reads as ŝ =
Qr (HyQD + η), where H ∈ C M×N is the channel matrix
with i.i.d. complex-valued entries of zero mean and unit variance and η ∼ CN (0M , Cη = IM ) is the AWG noise vector.
3. STATISTICAL THEORY OF 1-BIT
QUANTIZATION
To design a precoder which takes into account the effects
of Qt and Qr , we need to know some statistical properties of quantization especially the auto and cross-correlation
properties for a Gaussian input signal. Since quantization is
non-linear, it has strong effects on the statistical properties
of the signal. The statistical properties of hard limiters dealing with real-valued Gaussian-distributed signals are derived
in [12]. These derivations are applied to complex-valued
Gaussian-distributed signals and introduced in this section.
(1)
The covariance matrix between an unquantized circular distributed complex-valued signal x of covariance matrix Cx and
its 1-bit quantized signal xQ = Q (x) is given by
r
4
−1
KCx , where K = diag (Cx ) 2 .
CxQ x =
(2)
π
The covariance matrix of the 1-bit quantized circular distributed complex-valued signals xQ is given by
CxQ =
4
(arcsin (Kℜ {Cx } K) + j arcsin (Kℑ {Cx } K)) .
π
(3)
Note that the diagonal entries of CxQ are the squared norm of
quantized signals, which lead to the following result
diag (CxQ ) = 2I.
(4)
These four equations are the basis for solving the optimization
problem presented in the next section.
4. OPTIMIZATION PROBLEM
The optimization problem is formulated as follows
h
i
2
{PMMSE-Q , DMMSE-Q } = arg min E kŝ − sk2
P,D
s.t. E
h
2
kyQD k2
i
(5)
≤ Etx and D ∈ RN ×N is diagonal.
4.1. Objective function
We aim at minimizing the MSE between the desired signals
s and the received signals ŝ given that the power of the transmitted signal yQD is limited by the available transmit power
Etx . We end up with the following expression for MSE
MSE = σs2 tr (IM)+tr E ŝŝH −E sŝH −E ŝsH . (6)
We have to find the three expectation terms in the above for
which we make use of the covariance and cross correlation
matrices. We have already mentioned that the input signal
covariance matrix Cs = σs2 IM . The covariance matrix of the
precoder’s output y is given by
(7)
Cy = E yyH = σs2 PPH .
Tofind a linear expression for the covariance matrix CyQ =
H
E yQ yQ
, we make use of (3), (4) and the approximation
arcsin (x) ≃ x, for x 6= 1. So, we get
Cy Q =
4
K2 PPH K2 + cIN ,
π
(8)
with
4.2. Constraint
1
H −2
K2 = diag PP
and c =
− 1 . The covariance matrix CyQD of the transmitted signal yQD is given by
π
2
CyQD = DCyQ D.
(9)
The received signal covariance matrix reads as
Cx = HDCyQ DHH + Cη .
(10)
If we look at our MSE expression
in (6), one of the terms
H
which we need
to
find
is
tr
E
ŝŝ
. Since the structure of
Cŝ = E ŝŝH is very similar to CyQ , we end up with
tr E ŝŝH = 2M
(11)
i 4σ
h
s
E ŝsH =
K1 HDK2 P
π
4σs H
E sŝH =
P K2 DHH K1 ,
π
− 12
K1 = diag (Cx )
(12)
(13)
.
Finally, putting the expressions of (11), (12) and (13) in (6),
we end up with the following closed-form expression for
MSE
MSE = σs2 M + 2M
4σs
tr K1 HDK2 P + PH K2 DHH K1 .
−
π
(14)
The MSE expression in (14) contains two unknown variables
P and D. Intuitively, D should be a function of P, since it reallocates the power to the transmit signal originally intended
by P which gets lost due to Qt . To this end, we define a new
matrix P′ = K2 P. P′ is a row-normalized version of P,
such that each row of P′ has unit norm. Note that the MSE
expression contains the product K2 P and PH K2 . K1 also
contains these products. Thus, the MSE expression found so
far is in P′ rather than P. The purpose of D is to remove
this row-normalization of P introduced by K2 . Therefore, an
obvious choice for D is
H
D = K−1
2 = diag PP
Using (8) and the fact that D is a diagonal matrix, we can
simplify the constraint to
(16)
2 tr D2 ≤ Etx .
4.3. Final Optimization Problem
Using (14), (15) and (16), we can finally write our optimization problem as
12
.
(15)
In other words, the optimization with respect to D is reformulated as a one with respect to the norm of each row of P.
4σs
tr K1 HP + PH HH K1
π
(17)
1
Etx
and D = diag PPH 2 .
s.t. tr PPH ≤
2
min σs2 M + 2M −
P
We still
more
terms in the MSE, i.e.
need
to find two
tr E ŝsH and tr E sŝH before we can proceed
to
H
H H
. We
solve for P. Note that tr E ŝs
= tr E sŝ
use (2) to calculate the above mentioned expectations, which
can be expressed as follows
with
The constraint can be simply expressed as
i
h
2
E yQD 2 = tr DCyQ D ≤ Etx .
4.4. Solving the Optimization Problem
Note that our objective function in (17) is non-linear in P because of K1 . Furthermore, the solution set has to satisfy the
constraint in (17). Thus, we resort to the gradient projection
algorithm to solve our optimization problem [13]. The steps
for this algorithm can be found in Table 1. The needed derivative of the MSE with respect to P is given by
∂MSE (P)
2
4
= − σs HT K1 − HT K31 diag (H∗ P∗ ) H∗ P∗
∂P
π
π
2c
− diag HT diag H∗ P∗ K31 H∗ P∗
π
2
− HT K31 diag PT HT H∗ P∗
π
∗ ∗
2c
T
3 T T
−
diag H diag K1 P H H P .
π
5. SIMULATION RESULTS
In this section, we compare our proposed precoder with different precoding schemes in terms of the uncoded BER. All
the simulation results are averaged over 200 channel realizations. The used modulation scheme is QPSK, where σs2 = 2.
N = 20 antennas at the BS serve M = 4 users with Nb =
1000 transmit symbols per channel use. The tolerable error ǫ
and the iteration step µ of the gradient projection algorithm
are set to 10−6 and 0.05, respectively.
In Fig. 2 the uncoded BER is simulated as function of the
available transmit power Etx . ”WF, no Quant.” refers to the
linear Wiener filter precoder while no quantization is applied
in the system model. ”WF, D=I” is the linear Wiener filter
precoder that does not take the quantization into account and
BER performance sensitivity to inaccuracy in D implementation is studied and plotted in Fig. 3. It can be seen that even
with 10% error in D, the BER performance does not degrade
much as compared to the ideal case.
Table 1. Gradient Projection Algorithm
Iteration step µ,Tolerable error ǫ
InitializationP(0) = HH , n = 0
Step 1: If tr P(0) PH
> E2tx ,
r(0)
0.5Etx P(0) ;
let P(0) =
tr P(0) PH
(0)
∗
∂MSE(P(n) )
Step 2: P(n+1) = P(n) − µ
;
∂P
> E2tx ,
If tr P(n+1) PH
r(n+1)
0.5Etx
P(n+1) ;
let P(n+1) =
H
Uncoded BER
10
QP-GP, D with 5% error
QP-GP, D with 10% error
QP-GP
QP-GP, D=I
-2
tr P(n+1) P(n+1)
Step 3: If MSE(n+1) − MSE(n) ≤ ǫ
→ terminate the algorithm.
Otherwise, let n = n + 1 and return to Step 2.
10 -3
8
10
12
14
16
18
20
Transmit Power/dB
Fig. 3. Sensitivity analysis with respect to D
equal power allocation is performed [14]. The transmit power
constraint is still satisfied by appropriate scaling. ”QP-GP”
denotes our proposed precoder design: Quantized Precoder
with Gradient Projection method. ”QP-GP, D=I” refers to the
proposed precoder design when the power allocation is equal
for all transmit antenns. So, no additional analog processing
D is required. ”QWP” designates the Quantized Wiener filter
Precoder introduced in [11]. It can be seen from the results
that ignoring the distortions due to the 1-bit quantization in
WF leads to the worst case scenario. When taking them into
account in QP-GP, D=I a significant improvement in the uncoded BER can be achieved. This performance improvement
can be further increased when unequal power allocation at the
transmit antennas is deployed, as shown in the case of QP-GP
and QWF. The proposed precoder design QP-GP outperforms
the other designs. This iterative design converges to the same
solution for different initial values.
The analog real-valued diagonal precoder D can be built
within the power amplifiers at each antenna. Fig. 4 shows
the distribution of the normalized diagonal coefficients of D.
We observe that the deviation of these coefficients among the
different antennas and the channel realizations with respect
to the mean value (at max 6dB) is quite small. Therefore,
the requirements in terms of the dynamic range of the power
amplifier are still reasonable.
600
500
400
300
200
100
0
0.5
1
1.5
2
dnn/sqrt(E tx/2N)
Uncoded BER
10 0
Fig. 4. Distribution of the diagonal D coefficients over 200 channel
realizations: Etx = 10dB
WF, no Quant.
WF, D=I
QP-GP, D=I
QWP
QP-GP
10 -1
6. CONCLUSION
10 -2
10 -3
10 -4
-5
0
5
10
15
20
E tx[dB]
Fig. 2. BER comparison between different precoding schemes
In general, the analog processing exhibits higher complexity and lower accuracy as compared to the digital counterpart due to hardware implementations and imperfections
(aging, temperature,...). The analog precoder D offers less
complexity due to its positive real-valued diagonal structure,
and since it has to be updated only every coherence time. The
We present a new MMSE precoder design to mitigate the IUI
in a DL massive MIMO scenario assuming perfect CSIT. The
proposed precoder design takes into account the signal distortions due to the 1-bit quantization at the transmitter and at
the receiver. The precoder is split into a digital precoder that
separates the users in the direction and a real-valued diagonal analog precoder for the power allocation at each antenna.
Our precoding method shows better performance in terms of
the uncoded BER compared to the precoder designed in [11].
The analog precoder involved in the proposed scheme is updated every coherence time thus reducing the implementation
complexity. Furthermore, the BER performance is insensitive
to imperfections in the analog precoder implementation.
7. REFERENCES
[1] T.L. Marzetta, “Noncooperative cellular wireless with
unlimited numbers of base station antennas,” Wireless
Communications, IEEE Transactions on, vol. 9, no. 11,
pp. 3590–3600, November 2010.
[2] E. Bjornson, M. Kountouris, and M. Debbah, “Massive mimo and small cells: Improving energy efficiency
by optimal soft-cell coordination,” in Telecommunications (ICT), 2013 20th International Conference on,
May 2013, pp. 1–5.
[3] C.B. Peel, B.M. Hochwald, and A.L. Swindlehurst, “A
vector-perturbation technique for near-capacity multiantenna multiuser communication-part i: channel inversion and regularization,” Communications, IEEE Transactions on, vol. 53, no. 1, pp. 195–202, Jan 2005.
[4] A.B. Gershman, N.D. Sidiropoulos, S. ShahbazPanahi,
M. Bengtsson, and B. Ottersten, “Convex optimizationbased beamforming,” Signal Processing Magazine,
IEEE, vol. 27, no. 3, pp. 62–75, May 2010.
[5] F. Rusek, D. Persson, Buon Kiong Lau, E.G. Larsson,
T.L. Marzetta, O. Edfors, and F. Tufvesson, “Scaling
up mimo: Opportunities and challenges with very large
arrays,” Signal Processing Magazine, IEEE, vol. 30, no.
1, pp. 40–60, Jan 2013.
[6] M. Di Renzo, H. Haas, A. Ghrayeb, S. Sugiura, and
L. Hanzo, “Spatial modulation for generalized mimo:
Challenges, opportunities, and implementation,” Proceedings of the IEEE, vol. 102, no. 1, pp. 56–103, Jan
2014.
[7] R.R. Muller, M.A. Sedaghat, and G. Fischer, “Load
modulated massive mimo,” in Signal and Information
Processing (GlobalSIP), 2014 IEEE Global Conference
on, Dec 2014, pp. 622–626.
[8] A. Kalis, A. Kanatas, and C. Papadias, Parasitic Antenna Arrays for Wireless MIMO Systems, Springer,
2014.
[9] E. Bjornson, M. Matthaiou, and M. Debbah, “Massive MIMO with Non-Ideal Arbitrary Arrays: Hardware
scaling laws and circuit-aware design,” Wireless Communications, IEEE Transactions on, vol. PP, no. 99, pp.
1–1, 2015.
[10] C. Svensson, S. Andersson, and P. Bogner, “On the
power consumption of analog to digital converters,” in
Norchip Conference, 2006. 24th, Nov 2006, pp. 49–52.
[11] A. Mezghani, R. Ghiat, and J.A. Nossek, “Transmit
processing with low resolution d/a-converters,” in Electronics, Circuits, and Systems, 2009. ICECS 2009. 16th
IEEE International Conference on, Dec 2009, pp. 683–
686.
[12] R. Price, “A useful theorem for nonlinear devices having
gaussian inputs,” Information Theory, IRE Transactions
on, vol. 4, no. 2, pp. 69–72, June 1958.
[13] D. P. Bertsekas and J. N. Tsitsiklis, Parallel and Distributed Computation:Numerical Methods, PrenticeHall, 1989.
[14] M. Joham, Optimization of Linear and Nonlinear Transmit Signal Processing, Ph.D. thesis, Lehrstuhl für Netzwerktheorie und Signalverarbeitung, Technische Universität München, 2004.
| 7cs.IT
|
On using Tra er Driver for External Dynami
Pro ess Observation ⋆ (extended abstra t)
arXiv:cs/0701106v1 [cs.PL] 17 Jan 2007
Pierre Deransart
Inria
Ro quen ourt, BP 105, 78153 Le Chesnay Cedex, Fran e
Pierre.Deransartinria.fr
Abstra t
One is interested here in the observation of dynami
pro esses starting from the
tra es whi h they leave or those that one makes them produ e. It is
onsidered
here that it should be possible to make several observations simultaneously, using
a large variety of independently developed analyzers.
For this purpose, we introdu e the original notion of full tra e to
apture
the idea that a pro ess
an be instrumented in su h a way that it may broad ast
all information whi h
ould ever be requested by any kind of observer. Ea h
analyzer
an then nd in the full tra e the data elements whi h it needs. This
approa h uses what has been
alled a "tra er driver" whi h
ompletes the tra er
and drives it to answer the requests of the analyzers.
A tra er driver allows to restri t the ow of information and makes this
approa h tra table. On the other side, the potential size of a full tra e seems to
make the idea of full tra e unrealisti .
In this work we explore the
onsequen es of this notion in term of potential
e ien y, by analyzing the respe tive workloads between the (full) tra er and
many dierent analyzers, all being likely run in true parallel environments.
To illustrate this study, we use the example of the observation of the resolution of
onstraints systems (proof-tree, sear h-tree and propagation) using so-
phisti ated visualization tools, as developed in the proje t OADymPPaC (20012004).
The pro esses
approa h
1
onsidered here are
omputer programs, but we believe the
an be extended to many other kinds of pro esses.
Introdu tion
One is interested here in the observation of dynami
pro esses starting from the
tra es whi h they leave or those that one makes them produ e. It is
onsidered
here that it should be possible to make several observations simultaneously, using
a large variety of independently developed analyzers.
When one wants to observe a pro ess, the pra ti e is to instrument it for
ea h type of observation whi h one wants to make on it. One thus implements a
⋆
This work has been partly supported by OADymPPaC [1℄, a Fren h RNTL proje t.
new "ad ho " tra er for ea h analyzer, or one adapts and
one. Su h a work
approa h whi h
ompletes an existing
an be largely avoided if one adopts from the start a general
onsists to instrument it su h that it produ es what we
all a
"full tra e". This unique tra e will be useful then for all the later observations
whi h one
an plan to make. Ea h analyzer
an then nd in the full tra e the
data elements whi h it needs. This approa h uses what was
driver" whi h
alled a "tra er
ompletes the tra er and drives it to answer the requests of the
analyzers.
This approa h is parti ularly tempting in pra ti e as the full tra e never needs
to be
ompletely expressed (the ex hanges of information remain limited) and
the work of the tra er implementation and of its driver is made on e for all. The
evaluation in terms of feasibility and performan e remains however problemati .
In fa t, this approa h allows to redu e the size of the tra e emitted to a
useful bare minimum and thus to speed up the whole pro ess. In
it allows to
onsider very large full tra es. This has also a
the size of the tra e. Beyond a
ompensation,
ost whi h grows with
ertain size, the produ tion
ost of the tra e is
likely to be ome prohibitory. It is pre isely the question in whi h one is interested
here. Until where
an one go in this approa h on a pra ti al level without slowing
down a pro ess ex essively and how? To get a more pre ise idea, one must take
into a
ount not only the time of tra e produ tion, but also how the tra e will
be used.
In [2℄ and [3℄ the notion of tra er driver is presented and experimented is the
ontext of nite domain
onstraint resolution [4℄. But the question of the nature
of the tra e whose emission is
ontrolled by the tra er driver is not dire tly ta k-
led. In this work we explore the
onsequen es of this notion in term of potential
e ien y, by analyzing the respe tive workloads between the (full) tra er and
many dierent analyzers, all being likely run in a true parallel environment.
In this work, we introdu e the notion of full tra e to
pro ess
whi h
apture the idea that a
an be instrumented in su h a way that it may broad ast all information
ould ever be requested by any kind of observer. We then analyze the
nature of the work of the tra er and the driver, and the distribution of the
fun tions between the tra er and its driver on the one hand, and the analyzers
on the other hand. This allows us to better estimate how powerful, useful and
e ient the
on ept of full tra e
ar hite ture of all involved
an be, provided it is a
ompanied by the right
omponents.
To illustrate this study, one will take the example of the observation of the
resolution of
onstraints systems (proof-tree, sear h-tree and propagation) us-
ing sophisti ated tools for visualization, a
ording to the method developed in
the proje ts DiSCiPl (1997-2000) [5℄ then OADymPPaC (2001-2004) [1℄. This
eld is of parti ular interest be ause the tra es in lude the representations of
ompli ated and potentially bulky obje ts, and the
omputations (evolution of
the domain of the variables) are at the same time logi al and sto hasti . The
onstraints systems, be ause of the
to true
omplex systems.
omplexity of their resolution, are very
lose
In this extended abstra t we present su
essively the
on epts of full tra e
and its in remental ompressed version, then the question of their semanti s.
We analyze nally the problem of the distribution of work between a driven
tra er and external analyzers whi h work only with the useful tra e ow whi h
is provided to them by their requests.
2
Full Tra e
We introdu e here the
pro ess
on ept of
full tra e.
By de nition the full tra e of a
ontains all one may like to know about it during its exe ution (this
in ludes likely a des ription of the pro ess itself ) . A pro ess is
at a given moment
t
by a state
St .
hara terized
It does not enter in the framework of this
arti le to de ne what is exa tly su h a state. It will be supposed only that it
an be des ribed by a nite set of parameters
for its value at moment
A
t.
The
pn
for the
nth
parameter, and
pn,t
on ept of "moment" will be spe i ed hereafter.
urrent state will be denoted by the list of the values of its parameters. It is
also assumed that the transformation of a state into another is made by "steps"
hara terized by an a tion
a. The set of a
tions performed at moment
t is labelled
at .
So de ned, the
on ept of full tra e seems not to have any appli ation. In
pra ti e, there are only approximations; but the important thing here, is to admit
that, whatever is the level of details with whi h one wishes to observe a pro ess,
there is always a threshold whi h makes it possible to de ne su h a tra e. One
an
onsider that in the
ase of a program, it a ts of a more or less thorough
instrumentation whi h produ es a tra e, in other words a program augmented
with a tra er.
De nition 1 (Virtual Full Tra e).
sequen e of tra e events of the form
. An virtual full tra e is an unbounded
et : (t, at , St+1 )
omprising the following
elements:
et : unique identi er of the event.
t: hrono. Time of the tra e. It only
varies by unit values and is always
in reasing. To distinguish from the time of the observed pro esses or of the
analyzers whi h may not be monotonous
ompared to the
hrono.
St = p1,t ..., pn,t : parameters at hrono t. In a tra e event the parameters
are alled attributes and St the full urrent state. The parameters may
St is the last
urs. The parameters (or attributes)
des ribe obje ts or a tions performed to rea h the new state.
observed state before the event
et
orrespond to the new rea hed state
o
St+1 .
at : a tion an identi er of the set of a
state
St
to the state
tions
hara terizing the step from the
St+1 .
Any tra e ee tively produ ed by a pro ess
an be regarded as a partial full
tra e. In pra ti e, one "sees" only partial tra es whi h start after the moment
0 where the pro ess observed is in a presumably initial state
S0 .
We will limit
ourselves in this summary to only a single example of tra e: the tra e of the
proof-trees in Prolog systems based on standard Prolog [6℄. This tra e is far
from being a full tra e (many useful information, even easily available at any
moment are not represented there). Here is an extra t:
1
2
3
3
4
4
5
....
1
2
3
3
3
3
3
Call:
Call:
Call:
Exit:
Call:
Exit:
Call:
'$ all$'(ben h(2))
ben h(2)
2>0
2>0
_182 is 2-1
1 is 2-1
ben h(1)
This is a Byrd's tra e [7℄, adopted in the majority of the Prolog systems. The
moment "0"
event
orresponds to the laun hing of the resolution of a goal. Ea h tra e
orresponds to a stage in its resolution. An event
ontains the following
information: two attributes whi h are indi ation of depths (the rst, the depth
in the sear h-tree and the se ond in the proof-tree), a "port" whi h
orresponds
to the a tion whi h made it possible to rea h this stage. Thus the port
orresponds to a
goals to solve, and
total of 5 in the
all
all to the indi ated subgoal and the installation of the sub-
exit
to the su
ess of the subgoal. There are other ports (a
ase of GNU-Prolog) not detailed here. The rst " all" de nes
the laun hing of the rst goals (if there is only one of them). If the resolution
terminates, the tra e is nite, unbounded otherwise. The last attribute gives
the subgoal to solve. This tra e does not
hrono
omprise identi ers nor
hrono; the
orresponds to the sequential order of emission of the tra e events and
the resulting
hrono plays the role of an identi er.
It is interesting to note that the obje tive of su h a tra e is to display all the
steps of evolution of a proof-tree until obtaining all the possible proofs (the port
fail
orresponds to a failure and
redo
to a nondeterministi
goal where a new
resolution will be tried), and to some extend to des ribe also the sear h-tree.
However, the (partial) proof-tree is never expli itly des ribed (not more than
the sear h-tree). This tra e thus does not provide the parameters of interest dire tly (the partial proof-tree). This brings us to the following observation. The
parameters of the full state are not given expli itly in this tra e, but only some
attributes (port and goal) whi h possibly make it possible to nd it. We will
re onsider this point later. These attributes thus give only an in remental information whi h makes it possible to obtain the new proof-tree after a resolution
step. We will say that the tra e is "in remental". This leads us to de ne the
following parti ular tra es whi h we de ne here without more details.
De nition 2 (Dis ontinuous, Ee tive, In remental Full Tra e).
A tra e is dis ontinuous if it is a (full) tra e whi h ontains su
events whose variation of the
.
essive
hrono may be higher than a unit. There are "holes"
in it, either be ause the emission is dis ontinuous, or be ause the observing proess "listens" only o
asionally to it. Noti e that the dis ontinuity
on erns only
the tra e emission and not its extra tion; in other words moment 0
orresponds
to the a tivation of the tra er.
A tra e is
ee tive if it is a (full) tra
starting from the knowledge of
(St , At )
e of the form
one
an dedu e
et : (t, At ) and
(at , St+1 ). At
su h that,
denotes a
set of attributes. The ee tive tra e is the tra e emitted by a tra er, that whi h
is a tually visible. The virtual full tra e is a parti ular
where the attributes
parti ular
At
are the a tion label
at
ase of ee tive tra e
and the parameters
St .
Another
ase is the full in remental tra e.
A tra e is
in remental if the attributes are su
h that only the
urrent state are noted. It has the form
et : (t, Deltat +)
hanges ae t-
Deltat +
ontains the des ription of the a tions whi h modify the values of the parameters
ing the
where
of the moment t. To remain a full tra e, this tra e must satisfy the following
dition: starting from the knowledge of
(St , Deltat +)
one
an dedu e
on-
(at , St+1 ).
In this extended summary, when the distin tion is not absolutely ne essary,
one will not distinguish between virtual and ee tive tra es, and one will speak
indierently about parameters or attributes.
Pra ti ally all the tra es are in remental (thus it uses attributes), as the
pre eding example illustrates it well, be ause the emission of a full state in ea h
tra e event would be obviously prohibitive. In fa t the size whi h would take the
events would be mu h too high and the tra e would be extremely redundant.
The
ondition imposes simply that one
an retrieve the full tra e starting from
the transmitted attributes and from the pre eding full state. In pra ti e the
observed pro esses will produ e partial tra es only. In this
ase the retrieval of
the full state is impossible. If one wants to nd a full tra e, or at least a more
omplete one, it will be ne essary "to ask" the tra er to provide at least one full
state. If the observing pro ess needs to take into a
that
an be su ient to enable it to maintain a
ount only a partial state,
onsistent partial state. On the
other hand if it needs, at a given moment, to know a full state, or at least a
more
full
omplete state, it will have to ask the tra er to provide him at least a
urrent state or a part of it.
Pra ti ally all the tra es are dis ontinuous, even if often ea h part of them
(ne essarily nite)
an be regarded as a nite single tra e. This also raises the
problem of the knowledge of the initial state
S0
in whi h the observed pro ess
was at the moment of the initial tra e event ( hrono equals to 0) and thus of the
ommuni ation to the analyzers of the full initial state
S0 ,
before any event of
tra e.
These two reasons justify that one is interested in the manner of obtaining
su h a state. The urrent value of a parameter may exist just as it is in the pro ess
and requires only a small
omputation to extra t it. On the other hand it may
not exist and should require a partial re-exe ution of the pro ess (this
is used in the analyzers of the CHIP environment of Cosyte
apa ity
[5℄ or CLPGUI
[1℄). This obliges to stop the exe ution of the observed pro ess (at least when
that is possible), and to give the ability to the observing pro ess to stop and
resume the observed pro ess. This leads to the idea sket hed in the introdu tion
of syn hronization primitives between pro esses, a kind "image freezing" whi h
makes it possible to
omplete the information at any moment, a
need. This also results in the need for a fun tion "a
ess to the
ording to the
urrent state".
It is important to note that with an ee tive or in remental tra e, keeping
a full
urrent state is then in
harge of the observing pro ess and not of the
observed pro ess. This assumes however that the observed pro ess
least re overy points in whi h full
ontains at
urrent state will be maintained and a
essible,
so as to allow the observing pro ess to resume the tra e and to be able to restart
from a full
urrent state. It is important to note that with a full ee tive tra e,
dierent from the full virtual tra e, to keep a full
pro ess will be an additional
urrent state in the observing
harge of this pro ess itself, be ause the tra er of
the observed pro ess then does not have any more obligation to
al ulate the
requested parameters expli itly (only the requested attributes are
This supposes however that the observed pro ess
points in whi h a full
omputed).
ontains at least re overy
urrent state will be preserved and a
essible. This may
allow the observing pro ess to resume the tra e and to be able to restart from a
full
urrent state. This aspe t is not treated here.
3
Observational Semanti s
One is interested in this se tion in the semanti s of the full tra e,
alled here
"observational semanti s".
The tra e does not explain anything. It is only a
olle tion of fa ts. The
question arises however to understand the tra e. For that one needs two levels of
semanti s. The rst level
orresponds to the des ription of the a tions and ob-
je ts appearing in the parameters and attributes of the tra es, i.e. a semanti s of
the observed data. The se ond level
orresponds to a kind of tra e explanation,
i.e. a semanti s des ribing how the values of the parameters at the moment
are derived from the values of the parameters of the moment
a tions
whi h a
at
t−1
t
and how the
are sele ted. Clearly this means that one has a model of the pro ess
urately des ribes the evolution of the tra e between two events. The
form of this semanti s is the subje t of another work in progress.
To read a tra e, one needs only the rst level of semanti s. Only the relations
between the parameters in the same state or attributes in the emitted tra e
events must be known. In the example of the Prolog tra e, the properties of the
obje t "proof-tree" must be known to understand the relation between the depth
of the tree and the tree itself. The semanti s of the tra e (full and/or in remental)
thus
ontains the des ription of the relations between the parameters whi h
relates the semanti s and the produ ed attributes. But to understand the tra e
omprehensively, it is ne essary to go further in the knowledge of the pro ess
itself.
What we
all here "observational semanti s" (OS) is the semanti s of the
tra er (the rst semanti s, although being part of the OS,
an be seen like a
semanti s of the tra e). It uses the semanti s of the obje ts and a tions of the
full tra e and
an be seen as an abstra t semanti s in the sense of Cousot [8℄. It
l ∈ dom(Σ) ∧ choice-point(Σ(l))
N ← l, S ← Σ(l)
ba k to
Fig. 1. A rule of the abstra
t model of Prolog Proof-Tree displayed in [13℄
is a kind of natural semanti s [9℄ whi h
an be expressed by a nite set of rules
of the form
a : Condition(S) → S ′ = a(S)
su h that any event of tra e et : (t, at , St+1 ) an be obtained starting from
the state St and the a tion at by appli ation of the rule a of the OS. The (set
of ) a tions a performed at hrono t is denoted at . This semanti s an take the
form of a stru tural operational semanti s [10℄ or of an evolving algebra [11℄
and being more or less re ned, "big-step" or "small-step" [12℄. It is in fa t very
tempting to have a
omplete semanti s for the tra er in order to allow a
lear
implementation of it.
Coming ba k to the example of Prolog, a tra er augmented with the tra e
of
onstraints resolution (proof-tree, sear h-tree, labeling-tree and propagation),
alled Codeine, has been implemented in this manner on GNU-Prolog [13℄. An
abstra t model has been de ned [14℄, whi h has been then implemented on several solvers [1℄. This made it possible to meet two prin ipal obje tives: portability
of the analyzers whose input data are based on the full tra e, and robustness of
the tra ers whose implementation is guided and improved by a good methodologi al approa h [15℄ based on a pre-spe i ed rigorous semanti s.
It is useful however to note that a
would be a
omplete semanti s of the tra e, whi h
omplete formalization of the observed pro ess, is almost impossible
in pra ti e, be ause of the degree of re nement that would imply. For example,
an attribute in luded in many tra es is the CPU time
onsumed by the pro ess
sin e the "beginning" of the tra e emission. To formalize its variations, in fa t
issued from the host system, would amount introdu ing into the observational
semanti s a model of the system in whi h the pro ess is exe uted.
Finally an OS should not be
onfused with a
whi h would make it possible to re all the
the sole initial state and its rules. The
omplete operational semanti s
ourse of the pro ess starting from
urrent full state is not su ient to
know whi h rule may apply. For example, the model developed for nite domain
onstraints resolution des ribed in [13℄
as at least one rule
ontains a set of operational rules su h
an be applied at ea h state, but whose
always pre ise enough to de ide how it
onditions are not
an be applied. So is the
ba k to
rule,
depi ted in the Figure 1.
It
ontains a
the new
ondition
l ∈ dom(Σ)
whi h must be satis ed, meaning that
urrent node belongs to the sear h-tree, but nothing says whi h node
must be sele ted. The sole knowledge of the full
urrent state (whi h in ludes the
urrent sear h-tree) is not su ient. The knowledge of the
urrent tra e event
is ne essary to know whi h is this new node and thus to know how this rule is
in fa t instantiated.
4
Tra er Driver and Evaluation
The idea to use a tra er driver was proposed by M. Du assé and J. Noyé sin e
the origin of Prolog [16℄ in the
ontext of logi
programming, then regularly de-
veloped and tested in various environments [17℄. The originality of this approa h
onsists in proposing that the full tra e should not be
ommuni ated in its in-
tegrity to an analyzer, but that this one re eives only the part of the tra e whi h
relates to it. Instead of broad asting "urbi et orbi" a full tra e that any analyzer
would have the task to lter, ltering is performed at the sour e level, i.e. by
the observed pro ess whi h behaves then like a server of tra es. The analyzer
is a
lient whi h restri ts himself to indi ate to the observed pro ess the tra e
whi h it needs indeed. The
pro esses then brings to
information: the
omplete separation of the observed and observing
onsider an ar hite ture server/ lient with ex hanges of
lient indi ates to the server the tra e it needs and the server
provides him only what it requires. The ltering of the tra e is no longer
arried
out by the analyzer, but by the observed pro ess. It is the task of the tra er
driver to perform all requested ltering and to dispat h the tra es requested by
the
lients. This ar hite ture as well as the ex hanges between the pro esses have
been des ribed in parti ular in [1℄ and [2℄. We will not go more into the details in
this extended abstra t. In addition, only the aspe ts related to the possibilities
of modulating the tra es emitted a
be
ording to the needs of the analyzers will
onsidered; the possibility previously mentioned to syn hronize the pro esses
will not be
onsidered here.
This approa h allows to redu e the size of the tra e emitted to a useful bare
minimum and thus to speed up the whole pro ess. In
to
onsider very large full tra es. This has also a
of the tra e. Beyond a
ompensation, it allows
ost whi h grows with the size
ertain size, the produ tion
ost of the tra e is likely to
be ome prohibitory. It is pre isely the question in whi h one is interested. Until
where
an one go into this approa h on a pra ti al level without slowing down
a pro ess ex essively and how? To get a more pre ise idea, one must take into
a
ount not only the time of tra e produ tion, but also how the tra e will be
used.
One ould think at rst sight that the fa t, for the observed pro ess, to have to
produ e a full tra e makes this approa h unrealisti . The server is indeed slowed
down by the simple fa t of having to
ompute a great number of parameters
whi h will perhaps never be used. This approa h would be thus very penalizing
for the pro ess instrumented with an expensive tra er whenever only a weak
portion of the full tra e would be used. On another side, and pre isely in this
ase, a
onsiderable e onomy is realized be ause of ltering at the sour e, and
the transmission of a limited tra e whose
osts of
oding and diusion are then
extremely redu ed. The idea is that in pra ti e the fa t of emitting only one
small part of the tra e, ltered at the sour e,
ompensates for the over
ost of
work mainly related to the updates of all the parameters of the full tra e in the
analyzed pro ess.
On the other hand, if many analyzers are a tivated simultaneously and use
at a given moment a tra e equivalent to a full tra e, the problem to produ e or
not the full tra e does not arise any more (it must in any
ase be produ ed), and
the question about the interest of using a tra er driver is worth to be posed. We
show that even in this
ase the driver
an save time.
Additionally another type of e onomy must be
onsidered whi h is
omple-
mentary to the previous one: the use of an in remental tra e as an ee tive tra e
instead of the virtual full tra e. In this
ase there is no loss of information (the
equivalent of the full tra e is still transmitted), but there is a redu tion of the
amount of data transmitted (a kind of data
ompression is used in order to
limit the size of the emitted data ow). In this later
ase the tra er performs a
kind of ompression and the analyzer a kind of de ompression and both task
must be taken into a
ount in the analysis of the workloads. It will be shown
that the performan es of the tra er and the analyzers
an be in uen ed more by
this kind of load rather than by the size of the full tra e.
In order to be more pre ise, it is ne essary to analyze the repartition of work
within the various pro esses.
The times to take into a
ount for a detailed performan es analysis are times
on erning the pro ess itself, its tra er and its pilot on the one hand, and times
on erning one analyzer on the other hand (one must suppose here that, if there
are several analyzers, they are running with true parallelism):
Pro ess
Tra er and Driver
--------------- ------------------------------------|
| |
|
T_prog + T_ ore + T_ ond + T_extra t + T_en ode-and- om
Analyser
-----------------------------------------|
|
T_filter + T_de ode + T_rebuild + T_exe
From the side of the pro ess, tra er and driver:
Tprog :
time devoted to the exe ution of the not instrumented program (or
instrumented but with dea tivated tra er).
Tcore: additional exe
ution time of the pro ess on e instrumented to produ e
the full tra e and with a tivated tra er. It is a time devoted to the
onstru -
tion of the elements ne essary to a likely later extra tion of the parameters of
the
urrent state. This time is largely in uen ed by the size of the full tra e.
At this stage all
approa h, one
omputations must be performed be ause, in the presented
onsiders that it must be possible to produ e the full
state at any moment (in
urrent
ase in parti ular of dis ontinuous tra e). This time
is related to the form of the full tra e only and does not depend on what
will be emitted. If all the parameters of the full tra e are already part of the
pro ess, this time is just null.
Tcond :
time of
he king of the
onditions de ning the tra es to be emitted
for ea h analyzer ( ltering). This time is null if there is no ltering (emission
of the full tra e).
Textract : omputing time of the parameters requested during or after ltering.
Tencode−and−com : time for formatting the tra e (en oding), possible ompression and emission.
Time spe i
Textract
and
to the driver
Tencode−and−com
Tcond .
orresponds to
Other times, namely,
Tcore ,
an be regarded as times related to the tra er.
It should be noti ed here that, in this approa h, the driver a ts only on
the
hoi e of the parameters and the attributes to be
tra e, i.e. on the
ontained in the ee tive
ommuni ated information. It does not have the possibility of
in uen ing the form of the attributes, for example the degree of ompression
of the information. In fa t, this ompression is
oded here in the form of the
attributes. The nature of the attributes (in remental information or not) is part
of the tra er and it
annot be modi ed or adapted through the tra er driver.
On the other side any additional
ompression algorithm used to redu e the size
of the information ow belongs to the stage of en oding. Generalizing this
idea, one
ould also
onsider the possibility to put into the tra e more abstra t
attributes, adapted to some more spe i
a redu ed size. The
use, su h that the whole tra e has
orresponding attribute
omputation time would thus be
related to the extra tion stage. This generalization is not studied here.
From the side of the analyzer:
Tf ilter :
time of ltering by an analyzer. This time is null if ltering is per-
formed at the sour e (the tra e events sent to a parti ular analyzer
an
indeed be tagged at the sour e). In fa t by pre aution but mainly be ause of
la k of implemented tra er driver, many external analyzers will lter again
the tra e. However it is not ne essary to
Tdecode:
ir umvent, as the time of
well. It must be
not be
onsider this
ase here.
time of de oding of the re eived tra e. This time is impossible to
oding and
ompared with the
ompletely eliminated. If
used, it is be ause it is
Tencode−and−com as
Tencode−and−com. It an-
ommuni ation
encode
part of
ompression/de ompression algorithms are
onsidered that, even if both times are
umulated,
one will save a substantial amount of time over the transmission.
Trebuild :
time of rebuilding the full tra e starting from the ee tive tra e
( omputation of the urrent parameters starting from the emitted attributes).
This time must be
tributes. It is
ompared with
Textract ,
equilibrates most favorably with the
Texec :
the
omputation time of the at-
onsidered that su h a time, even if
umulated with
Textract ,
ost of tra e emission.
exe ution time of the proper fun tions of the analyzer. With very
sophisti ated analyzers (as those used for data analysis for example), this
time
an be ome so important that it makes negligible the one
orresponding
to the tra e produ tion.
Example: Codeine, des ribed in [13℄, implements the generation of a full tra e
for the analysis of
onstraints resolution, as an extension of GNU-Prolog. The
urrent state
St
ontains (among other attributes): proof-tree, sear h-tree,
straints state and variables with their domains. For a
the full tra e see [1℄ ( alled in this proje t "generi
not generate a full tra e (as full as it
on-
omplete des ription of
tra e"). Even if Codeine does
ould be possible), the Codeine tra e is
onsiderably ri her than the simple Byrd's tra e of GNU-Prolog whi h is stri tly
ontained. Only an in remental tra e is generated, and only the proof-tree and
the
onstraints and variables states
urrent state of the pro ess (the
an be rebuilt later starting just from the
ost of extra tion is redu ed to the
proper data management realized by the tra er). To obtain the
ost of the
urrent sear h-
tree the pro ess would have to be re-exe uted partially, therefore it would be
ne essary to freeze its
nently a
urrent exe ution. The
ost of management of a perma-
essible sear h-tree at any moment would be
of its size. On the other hand an analyzer
learly intra table be ause
an, using the produ ed in remental
tra e, maintain all these obje ts permanently (obtaining
useful for it). Codeine also
St
or those parameters
ontains a tra er driver su h that the de nitions of
the tra es to be emitted (spe i ation of the emitted tra e) are stored in a data
le whi h must be provided before the pro ess starts.
Times are distributed as follows:
From the side of the pro ess, tra er and driver:
Tprog : exe
part of
Tcore:
ution time of GNU-Prolog with "swit hes" of the tra er (a small
Tcore
of negligible duration in general).
time of
onstru tion of the parameters of the full Codeine tra e ( ol-
le ting of all data useful to extra t the full tra e); a kind of GNU-Prolog
plug-in.
Tcond : time of ltering the full tra e to sele
Textract : omputing time of the attributes
t all the requested tra es.
orresponding to the parameters
requested during ltering.
Tencode−and−com :
omputing time of the attributes
orresponding to the re-
quested parameters, en oding (XML format or Prolog term) of the emitted
tra e, and emission time of the in remental tra e.
From the side of the analyzers: experiments in the framework of the OADymPPaC proje t with sophisti ated analyzers (intensive visualization of graphs,
visual data analyses) revealed the following
Tf ilter
and
Tdecode :
osts.
both times are intri ate in a synta ti
(XML) of the full tra e. Re- ltering, a part of whi h
But the driven tra er approa h
analysis module
ould have been avoided.
ould not be taken into a
ount in the
analyzers built during this proje t.
Trebuild :
time of
onstru tion of the parameters of the full tra e (variables
domains, a tive
onstraints set, sear h-tree ...). This time grows (non esti-
mated fa tor) a
ording to the size of the data, sometimes related to the size
of the tra e, with a
onsiderable slow down of the analyzer during the read-
ing of the tra e . The low speed of an analyzer may
ause, in the
ase of
analysis of the tra e on the y, a strong slow down of the observed pro ess.
Texec :
time of
onstru tion of the obje ts to be visualized (graphs, data
tables). These times
an grow exponentially a
The e ien y of the used algorithms is
ording to the size of the data.
ru ial here. This
an also
ause a
slow down of the pro ess, and pleads in favor of a preliminary treatment
of information before transmission (for example, sele tion of distinguished
nodes to put in the tra e to redu e the size of a drawn graphs, or to
olle t
groups of variables as a unique attribute to redu e the number of lines in a
matrix).
It was been shown experimentally in [3℄ that the behavior of the Codeine
tra er with the full tra e above
ompares favorably with the behavior of GNU-
Prolog with the Byrd's tra e only. Furthermore, the ltering realized by the
tra er driver does not prejudi es the performan es.
We give here a theoreti al justi ation to this result, showing that this approa h, already justi ed experimentally,
an also be justi ed theoreti ally.
In [2℄ L. Langevine observes that ltering the tra es all together is more
ee tive than to lter them one after the other. This
omes from the fa t that
running several automata together may be more e ient than running one only.
Indeed the essen e of ltering relates to simpli ed
sele t rst the tra e events
ner additional ltering will be thus
of events mu h more redu ed. One
a tra e whose language
to
onditions whose role is to
ontaining the ports requested by an analyzer. A
arried out thereafter but on a number
an admit that this rst ltering relates to
orresponds to a regular language. It is then possible
onsider that ea h lter is itself a regular expression whose re ognition on
the full tra e
Filtering
an be done using a non deterministi
nite-state automaton.
orresponds then to the re ognition task by a union of as many nite-
state automata than there are a tive analyzers requesting a tra e. However the
resulting automaton, on e optimized,
an be mu h more e ient than the most
e ient of the automata asso iated with a single analyzer (the union of the
automata
an be more e ient in terms of
omputation steps than only one of
them [18℄). As these operations of ltering are extremely frequent, be ause they
apply to all the tra e events, the speed up
an be
onsiderable.
We are now in position to analyze the respe tive workloads.
First observe that respe tive times of both sides may be
onsidered as
u-
mulative or not, depending whether the respe tive pro ess are run sequentially
or with true parallelism. In the later
essors (a situation whi h
an be
ase, if all pro ess are run on dierent pro-
onsidered with su h an approa h), only the
slowest pro ess must be taken into a
ount to evaluate the exe ution time of the
whole system.
Tprog
and
Tcore
These times
on one side,
Texec
on the other side.
orrespond to times spe i
to the tra er and the analyzers (the
slowest analyzer has the main in uen e on the exe ution time). These times
Tcore
Tf ilter
are in ompressible and
Tcond
on one side and
depends only on the size of the virtual tra e.
on the other side.
At least one of these times must be null, and, if the ltering is at the sour e,
the performan e
an be improved. In any
whatever is the size of the full tra e,
Textract
on one side and
These times
Trebuild
ases this time appears negligible,
ompared with the other times.
on the other side.
orrespond respe tively to the
omputation of the attributes
from the parameters and re ipro ally. If one
an redu e the extra tion time
thanks to the ltering ( onsidering the low number of sele ted tra e events
and attributes, and the fa t that the most of the parameters
work is already in luded in
Tcore ),
the time of rebuilding
omputation
an be very im-
portant and will probably be signi antly greater than the extra tion time.
Tencodea ndc om
on one side and
These times
Tdecode
on the other side.
an be important, but the interest lies in the fa t that the
pro ts realized on the
ompensates the
ommuni ation time (redu ed emitted volume) largely
oding/de oding times.
To summarize, the tra er and the slowest analyzer are the main fa tors in uen ing the whole performan e and they
sides
an in rease the workload of both
onsiderably. However the use of a tra er driver and of te hniques of tra e
ompression make it possible to
ompensate partially, but sometimes in very
ee tive manner, the over- osts related to the use of a full tra e.
5
Con lusion
We introdu ed the
on ept of full tra e in order to take into a
possibilities of analysis of a dynami
ount the multiple
pro ess. The analyzed pro ess is instru-
mented with, in addition to its tra er, a tra e driver, and the analyzers have
the possibility of addressing orders to the driver. An ar hite ture
then
lient/server is
onsidered to des ribe the intera tions between the pro esses analyzers and
the analyzed pro ess. This enabled us to ta kle the problem of the evaluation of
this approa h in terms of e ien y.
We tried to appre iate how the introdu tion of a tra er driver
the global e ien y of the analysis of a dynami
ould improve
pro ess using several analyzers
observing the pro ess in a simultaneous way. We initially observed that a full
tra e
ould be parti ularly expensive to produ e, but that part of this
ould be transferred, without loss of
apa ity of analysis (the full tra e
retrieved by the analyzers), on the analyzers. The most unfavorable
of e ien y
ost
an be
ase in term
orresponds to the situation where the equivalent of a full tra e must
be built, extra ted and emitted. In this
ase, the full tra e is in fa t a union of
all the redu ed tra es requested by several analyzers. We showed that, even in
this
ase, a ltering realized by a tra er driver was able to bring an important
bene e. Our observations, during the proje ts DiSCiPl [5℄ and OADymPPaC
[1℄, using very sophisti ated analyzers as powerful tools for visualizations as well
([19℄ [20℄), also showed that the limits of performan e
analyzers that tra er, even with a full tra e.
This study opens nally on a series of questions:
ame often more of the
How deep is it possible to implement a very broad full tra e?
ourse the limits of the approa h are obvious: the
Of
on ept of full tra e is
meaningful only with regards to a family of possible analyses, well known
and sele ted in advan e. Nothing guarantees a priori that an additional instrumentation of the observed pro ess will never be ne essary. But beyond
this aspe t, the interesting question relates to the feasibility of the implementation of a very ne grained full tra e. On one side indeed, one will be
able to
ompensate for
ertain produ tion
osts of su h large tra e, by emit-
ting an ee tive tra e of limited size; but on the other side, the analyzers
(whose use
an be temporary or ex eptional) will be loaded with most of
time of development/re-building of the full tra e (ne essary, even if it uses
only a part of it) and will o
asionally slow down the observed pro ess.
Whi h intera tion language to use and whi h dialogue between
the driver and the analyzers? This question has been little ta kled
here and mainly remains out of the s ope of this arti le. The tenden y however is with the use of a language like XML. It is the way
hosen by the
OADymPPaC proje t, and the possibility of redu ing the ow of tra e to its
bare minimum en ourages this. Nevertheless our experiments showed that
the
data
ommuni ation needed to be optimized by
ompression with spe i
tra e
ombining usual methods of
ompression like using in remental at-
tributes, but also by introdu ing more abstra t attributes in the tra e. This
leads to the idea that the dialogue between the involved pro esses, limited
here to the
hoi e of the tra e events and the attributes in the tra e, and
apa ities of syn hronization, must be extended su h that it allows also to
in uen e the design of the attributes themselves.
Finally a
ru ial question related to the
omprehension of the tra e.
to understand a tra e, or how to des ribe its semanti s ?
How
If the
semanti s of the tra e (and of the tra er), or at least a large part of it, is
given a priori (be ause one has a model for the observed pro ess), then the
omprehension of the tra e as well as the implementation of the tra er are
largely fa ilitated. In the opposite
or arti ial
ase - and it is the
ase for many natural
omplex pro esses - one has only vast tra es to study. Even in
the elds of programming languages where semanti s seems better
ontrolled
a priori, one tries to analyze a program behavior by trying to understand its
tra es. Thus one sees stinging the usefulness of general te hniques based on
data mining [21℄ or on Web mining [22℄ for su h purposes. However one must
re ognize that any full tra e will probably always in lude portions es aping
any kind of des ription based on formal semanti s.
Referen es
1. Deransart, P., & al: Outils d'Analyse Dynamique Pour la Programmation par
Contraintes (OADymPPaC). Te hni al report, Inria Ro quen ourt and É ole des
Mines de Nantes and Insa de Rennes and Université d'Orléans and Cosyte and
ILOG (2004) Projet RNTL. http:// ontraintes.inria.fr/OADymPPaC .
2. Langevine, L., Du assé, M.: A Tra er Driver for Hybrid Exe ution Analyses. In
Press, A., ed.: Pro eedings of the 6th Automated Debugging Symposium. (2005)
3. Langevine, L., Du assé, M.: A Tra er Driver for Versatile Dynami Analyses of
Constraint Logi Programs. In Serebrenik, A., Muñoz-Hernandez, S., eds.: Pro eedings of the 15th Workshop on Logi -based Methods in Programming Environments
(WLPE'05), a pre- onferen e workshop of ICLP'05, Sitges, Spain (2005) Computer
Resear h Repository s.SE/0508105.
4. Langevine, L., Du assé, M.: A Tra er Driver for CLP: Debugging, Monitoring and
Visualization of an Exe ution from a Single Tra er. In Demoen, B., Lifs hitz, V.,
eds.: Pro . of the 20th Int. Conf. on Logi Programming, (ICLP'04). Number 3132
in LNCS, Saint-Malo, Fran e (2004) 462463
5. Deransart, P., Hermenegildo, M., Maªuszy«ski, J., eds.: Analysis and Visualisation
Tools for Constraint Programming. Number 1870 in LNCS. Springer Verlag (2000)
6. GNU-Prolog: A CLP(FD) System Based on Standard Prolog (ISO) developed by
D. Diaz (2003) http://gprolog.sour eforge.net/ Distributed under the GNU li ense.
7. Byrd, L.: Understanding the ontrol ow of Prolog programs. Logi Programming
Workshop (1980)
8. Cousot, P.: Abstra t Interpretation. Te hnique et S ien e Informatiques 19 (1994)
19
9. Kahn, G.: Natural Semanti s. In Brandenburg, F.J., Vidal-Naquet, G., Wirsing,
M., eds.: Pro eedings of STACS, Springer (1987) also INRIA, RR 416, Natural
Semanti s on the Computer.
10. Plotkin, G.: A Stru tural Approa h to Operational Semanti s. TR DAIMI FN 19,
Computer S ien e Department, Aarhus University, Aarhus, Denmark (1981)
11. Gurevit h, Y.: Evolving algebras, a tutorial introdu tion. Bulletin of the European
Asso iation for Theoreti al Computer S ien e 43 (1991) 264284
12. Lu as, S.: Observational Semanti s and Dynami Analysis of Computational Proess. RR 00/02, E ole Polyte hnique, Laboratoire d'Informatique, LIX, Palaiseau
(Fran e) (2000)
13. Langevine, L., Du assé, M., Deransart, P.: A Propagation Tra er for GNU-Prolog:
from Formal De nition to E ient Implementation. In Palamidessi, C., ed.: Proeedings of the 19th International Conferen e on Logi Programming, ICLP'03,
Mumbai, India (2003)
14. Langevine, L., Deransart, P., Du assé, M.: A Generi Tra e S hema for the
Portability of CP(FD) Debugging Tools. In Apt, K., Fages, F., Rossi, F., P.,
S., Van za, J., eds.: Re ent Advan es in Constraints. Number 3010 in Le ture
Notes in Arti ial Intelligen e. Springer Verlag (2004) Sele ted papers of the Joint
ERCIM/CoLogNET International Workshop on Constraint Solving and Constraint
Logi Programming.
15. Du assé, M., Langevine, L., Deransart, P.: Rigorous design of tra ers: an experiment for onstraint logi programming. In Ronsse, M., Boss here, K.D., eds.:
Pro eedings of the 5th International Workshop on Automated and Algorithmi
Debugging, AADEBUG'03, Ghent, Belgium (2003) Computer Resear h Repository s.SE/0310042.
16. Du assé, M., Noyé, J.: Logi Programming Environments: Dynami Program Analysis and Debugging. The Journal of Logi Programming 19/20 (1994) 351384
17. Du assé, M.: Tra e S hemata for (Multi-language) Dynami Analysis. In Choi,
J.D., Ryder, B., Zeller, A., eds.: Dagstuhl Seminar 03491, "Understanding Program
Dynami s", Saint-Malo, Fran e, Dagstuhl, Germany (2003)
18. Hop roft, J.E., Ullman, J.D., eds.: Introdu tion to Automata Theory, Languages,
and Computation. Computer S ien e. Addison-Wesley (1979)
19. Baudel, T., & al: Visual CP, Referen e Manual (2004) Manufa tured and distributed by Ilog, http://www2.ilog. om/preview/Dis overy/ .
20. Ghoniem, M., Jussien, N., Fekete, J.D.: VISEXP: Visualizing Constraint Solver
Dynami s Using Explanations. In Barr, V., Markov, Z., eds.: Pro . of the 17th Int.
of the Florida Arti ial Intelligen e Resear h So iety Conferen e (FLAIRS'04),
AAAI Press (2004)
21. Denmat, T., Du assé, M., Ridoux, O.: Data Mining and Cross-Che king of Exe ution Tra es. A reinterpretation of Jones, Harrold and Satsko test Information
Visualization. RI 1743, IRISA, Rennes (Fran e) (2005)
22. Zaidman, A., Calders, T., Demeyer, S., Paredaens, J.: Applying Webmining Te hniques to Exe ution Tra es to Support the Program Comprehension Pro ess. In
So iety, I.C., ed.: Pro eedings of the 9th European Conferen e on Software Maintenan e and Rengineering (CSMR 2005), Man hester, UK (2005) 134142
| 2cs.AI
|
Sample Efficient Policy Search for Optimal Stopping Domains
arXiv:1702.06238v2 [cs.AI] 24 May 2017
Karan Goel
Carnegie Mellon University
[email protected]
Christoph Dann
Carnegie Mellon University
[email protected]
Abstract
Optimal stopping problems consider the question
of deciding when to stop an observation-generating
process in order to maximize a return. We examine
the problem of simultaneously learning and planning in such domains, when data is collected directly from the environment. We propose GFSE,
a simple and flexible model-free policy search
method that reuses data for sample efficiency by
leveraging problem structure. We bound the sample
complexity of our approach to guarantee uniform
convergence of policy value estimates, tightening
existing PAC bounds to achieve logarithmic dependence on horizon length for our setting. We also
examine the benefit of our method against prevalent model-based and model-free approaches on 3
domains taken from diverse fields.
Introduction
Sequential decision making and learning in unknown environments, commonly modeled as reinforcement learning
(RL), is a key aspect of artificial intelligence. An important subclass of RL is optimal stopping processes, where an
agent decides at each step whether to continue or terminate a
stochastic process and the reward upon termination is a function of the observations seen so far. Many common problems
in Computer Science and Operations Research can be modeled within this setting, including the secretary problem [Ferguson, 1989], house selling [Glower et al., 1998; Lippman
and McCall, 1976], American options trading [Jacka, 1991;
Mordecki, 2002], product pricing [Feng and Gallego, 1995]
and asset replacement [Jiang and Powell, 2015], as well as
problems in artificial intelligence like mission monitoring
robots [Best et al., 2015], metareasoning about the value of
additional computation [Zilberstein, 1995] and automatically
deciding when to purchase an airline ticket [Etzioni et al.,
2003]. Often the stopping process dynamics are unknown in
advance and so finding a good stopping policy (when to halt)
requires learning from experience in the environment. As real
experience can incur real losses, we desire algorithms that
can quickly (with minimal samples) learn good policies that
achieve high reward for these problems.
Emma Brunskill
Stanford University
[email protected]
Interestingly, most prior work on optimal stopping has focused on the planning problem: how to compute near-optimal
policies given access to the dynamics and reward of the
stochastic stopping process [Peskir and Shiryaev, 2006]. Optimal stopping problems can also be framed as a partially observable Markov decision process (POMDP), and there also
exists work on learning a good policy for acting in POMDPs,
that bounds the number of samples required to identify the
near optimal policy out of a class of policies [Kearns et al.,
1999; Ng and Jordan, 2000]. However, such work either (i)
makes the strong assumption that the algorithm has access
to a generative model (ability to simulate from any state) of
the stochastic process, which makes this work more suited to
improving the efficiency of planning using simulations of the
domain, or (ii) can use trajectories directly collected from the
environment, but incurs exponential horizon dependence.
In this paper, we consider how to quickly learn a nearoptimal policy in a stochastic optimal stopping process with
unknown dynamics, given an input class of policies. We assume there is a fixed maximum length horizon for acting,
and then make a simple but powerful observation: for stopping problems with process-dependent rewards, the outcomes
of a full length trajectory (that is, a trajectory in which the
policy only halts after the entire horizon) provide an estimated return for halting after one step, after two steps, and
so on, till the horizon. In this way, a single full-length trajectory yields a sample return for any stopping policy. Based
on this, we propose an algorithm that first acts by stopping
only after the full length horizon for a number of trajectories, and then performs policy search over an input policy class, where the full length trajectories are used to provide estimates of the expected return of each policy considered in the policy class. The policy in the set with the
highest expected performance is selected for future use. We
provide sample complexity bounds on the number of full
length trajectories sufficient to identify a near optimal policy within the input policy class. Our results are similar
to more general results for POMDPs [Kearns et al., 1999;
Ng and Jordan, 2000], but due to the structure of optimal stopping we achieve two key benefits: our bounds’ dependence on
the horizon is only logarithmic instead of linear (with a generative model) and exponential (without), and our results apply
to learning in stochastic stopping processes, with no generative model required. Simulation results on student tutoring,
ticket purchase, and asset replacement show our approach
significantly improves over state-of-the-art approaches.
Problem Formulation
We consider the standard stochastic discrete-time optimal
stopping process setting. As in Tsitsiklis and Van Roy [1999],
we assume there is a stochastic process P that generates observations o1 , . . . , ot (they may be vectors). There are two
actions: halt or continue the process. The reward model is
a known, deterministic function of the sequence of observations and the choice of whether to continue or halt. While
there do exist domains where the reward model can be a
nondeterministic function of the observations and the actions
(such as a medical procedure that reveals the patient’s true
condition after a sequence of waiting), most common optimal stopping problems fall within the framework considered
here, including the secretary problem (the quality of each secretary is directly observed), house selling (the price for the
house from each bidder is known), asset replacement (published guides on the worth of an asset, plus knowledge of the
cost of buying a new one), etc. We focus on the episodic setting where there is a fixed maximum time horizon for each
process. The finite horizon value of a policy Vπ = EP [r|π]
is the expected return from following π over a horizon of H
steps, where the expectation is taken over the stochastic process dynamics P. Note the policy may choose to halt before
H steps. The goal is to maximize return across episodes.
We focus here on direct policy search methods (see e.g.
[Sutton et al., 1999]). More precisely, we assume as input a
parameterized policy class Π = {f (θ)|θ ∈ Θ} where Θ is the
set of policy parameters. Direct policy search does not require
building a model of the domain, and has been very successful
in a variety of reinforcement learning (RL) contexts [Deisenroth and Rasmussen, 2011; Levine and Abbeel, 2014].
Sample Efficient Policy Search
We are particularly interested in domains where evaluation
of a policy incurs real cost in the environment, such as stock
market options selling. In such settings we wish to find sample efficient methods for doing policy search, that can minimize the number of poor outcomes in the real world. The
challenge is that we do not know the stochastic dynamics
P and so it is not possible to, in advance of acting, perform
policy search to identify a good policy. Instead we can only
obtain information about the domain dynamics by executing
policies in the real world. We seek to efficiently leverage such
experience to quickly make good decisions.
We now present a simple approach, GFSE (Gather Full,
Search and Execute) (Algorithm 1), to do sample efficient
policy search. GFSE collects a set of full-length (horizon H)
trajectories, uses these to evaluate the performance of any policy in the input policy class Π, identifies a good policy, and
then executes the resulting policy on all future episodes.
The key insight is in the first step, gathering the data to be
used to evaluate the performance of any policy in the policy
class Π. Monte Carlo estimation can be used to estimate the
expected return of a policy by running it many times. However, this scales poorly with the cardinality of the policy class.
Algorithm 1: Gather Full, Search and Execute (GFSE)
Input: policy class Π, search method S, , δ
n ← Use Theorem 1 with , δ
Γ ← Gather n full trajectories from environment
π∗ ← Identify policy using S(Π, Γ) \\evaluation uses Γ
Execute π∗
Building a dynamics model from a set of data is more efficient, as a model can be used to simulate the performance of
any policy, but this requires us to make certain assumptions
about the domain (for ex. the Markov property) Which can
lead to biased estimates. Alternatively, importance sampling
can be used to do off-policy evaluation [Precup, 2000] but
unfortunately such estimates tend to be very high variance.
However, a simple but powerful observation is that a fullhorizon (H-step) trajectory can be used to yield a sample return for all optimal stopping policies in Π. Given a full length
trajectory (o1 , . . . , oH ), the performance of a particular policy π can be simulated by providing (o1 , o2 , . . . ) to the target
policy until it halts at some time step t ≤ H. Therefore we
can take the subsequence of observations (o1 , o2 , . . . , ot ) and
use it to directly compute the return that would have been observed for executing π on this trajectory. A single full-horizon
trajectory will provide just one sample of the return of any
policy. But a set of n full-horizon trajectories can be used to
provide n sample returns for a given policy π, thereby providing an empirical estimate of V π . We can do this off policy
evaluation of V π for any policy in the class Π.
Prior work has shown that given access to a generative
model of the domain, policy search can be done in an efficient way by using common random numbers to evaluate
policies that act differently in an episode [Kearns et al., 1999;
Ng and Jordan, 2000]. In our setting, a full-horizon trajectory is essentially equivalent to having access to a generative
model that can produce a single return for any policy. However, access to a full length trajectory can be obtained by running in the environment, whereas generic generative models
typically require ”teleporation”: the ability to simulate what
would happen next under a particular action given an arbitrary prior history, which is hard unless in a planning scenario
in which one already has knowedge of the dynamics process.
Our results require weaker assumptions than prior results that
use stronger generative models to obtain similar sample efficiency, while also achieving better sample efficiency than
approaches with access to similar generative models.
We will shortly provide a sufficient condition on the number of full length trajectories n, to guarantee that we can
evaluate any policy sufficiently accurately to enable policy
search to identify a near-optimal policy (within the input policy class). Of course, empirically, we will often wish to select
a smaller n: our simulation experiments will demonstrate that
often a small n still enables us to identify a good policy.
Theoretical Analysis
We now provide bounds on the sample complexity of GFSE:
the number of full length trajectories required to obtain near
accurate estimates of all policies in a policy class. This is suf-
ficient to identify the optimal (or near-optimal) policy in the
policy class with the highest expected return.
First, we note that the optimal stopping problems we consider in this paper can be viewed as a particular instance of
a POMDP. Briefly, there is some hidden state space, with a
dynamics model that determines how the current state transitions to a new state stochastically, given the continue action.
The observation is a function of the hidden state, and the reward is also a function of the hidden state and action.
Our main result is that, given a policy class Π, the sample complexity scales logarithmically with the horizon. We
make no assumption of access to a generative model. This
is a significant improvement over prior sample complexity results for policy search for generic POMDPs and large
MDPs [Kearns et al., 1999; Ng and Jordan, 2000] which required access to a generative model of the environment and
had a sample complexity that scaled linearly with the horizon. These results can be thought of as bounding the computation/simulation time required during planning, when one
has access to a generative model that can be used to sample
an outcome (reward, observation) given any prior history and
action. In contrast, our results apply during learning, where
the agent has no generative model of the domain, but must instead explore to observe different outcomes. Without a generative model of the domain, sample complexity results for
policy search in generic POMDPs when learning scale exponentially with the horizon [Kearns et al., 1999].
Optimal stopping trajectories are related to the trajectory
trees of Kearns et al. which were used to evaluate the returns
of different POMDP policies. For a POMDP with 2 actions,
each trajectory tree is a complete binary tree (of depth H)
rooted at a start state. Nodes in the tree are labeled with a
state and observation, and a path from the root to any node
in the tree denotes a series of actions taken by a policy. A
trajectory tree can be used to evaluate any policy in Π, since
every action sequence is part of the tree. However, while for
generic POMDPs the size of a trajectory tree is exponential in
the horizon, for optimal stopping problems the tree size is linear in the horizon (Figure 1). This allows us to obtain significantly tighter dependence on H than for generic POMDPs.
Our analysis closely follows the prior sample complexity
results of [Kearns et al., 1999]. Kearns et al. proceeded by
first considering a bound on the VC-dimension of Π when
viewed as a set of real-valued mappings from histories to returns, as a function of the VC-dimension of Π when viewed
as mappings of histories to actions. Then they use this result
to bound the sample complexity needed to get near-accurate
estimates of the returns of all policies in the policy class.
We will follow a similar procedure to bound the sample
complexity when Π contains a potentially infinite number of
deterministic policies.1 Let d = VC(Π) be the VC-dimension
of our policy class. This is well-defined, since each optimalstopping policy maps trajectories to 2 actions (binary labeling). Let VCr (Π) be the VC-dimension of Π when viewed as
a set of real-valued mappings from full trajectories to returns
1
Similar to Kearns et al. (1999) our results extend to finite Π and
infinite, stochastic Π, as well as the discounted infinite-horizon case
(using an -approximation to Vπ with horizon H ).
Figure 1: The structure of a full trajectory for horizon H.
Each node represents an observation while arrows represent
one of the two available actions.
and assume Vπ is bounded by Vmax . From [Vapnik and Kotz,
1982], we know that VCr (Π) can be computed as VC(H),
where H = {I(π, ∆, ·)|π ∈ Π, ∆ ∈ [−Vmax , Vmax ]} with
I(π, ∆, x) = 1 if π(x) ≥ ∆, and 0 otherwise (π(x) is the
return for full trajectory x under π).
Lemma 1. Let Π be a set of deterministic optimal-stopping
policies with VC-dimension d when viewed as a set of maps
from trajectories to actions. Then, when viewed as a set of
maps from the space of all full trajectories to [−Vmax , Vmax ],
Π has dimension bounded by,
VCr (Π) = O(d log H)
Proof. Our proof proceeds similarly to Lemma A.1 in Kearns
et al.. The crucial difference is that our policies operate on
a full-trajectory structure that contains H nodes (Figure 1),
rather than Kearns et al.’s trajectory trees with 2H+1 nodes.
In our setting at each point the agent only gets to consider
whether to halt or continue, and if the halt action is chosen,
the trajectory terminates. This implies that in contrast to standard expectimax trees where the size of the tree depends on
the action space as an exponential of the horizon, |A|H , in our
setting the dependence induced by the actions is only linear in
H. Thus Π can produce a much smaller set of behaviors, and
our dependence on H is logarithmic, rather than polynomial.
More formally, by Sauer’s lemma, k trajectories can be lad
beled in atmost ( ek
d ) ways by Π. First note that n full trajectories contain at most nH distinct trajectories across them
(one per node; refer to Figure 1 for the structure of full trajectories). Each action labeling of these k = nH trajectories
by Π, corresponds to selecting n paths (1 path per full trajectory), where each path starts at the first observation and ends
at a terminal node. The number of possible selections by Π is
d
thus atmost ( enH
d ) . Each path can be viewed as mapping a
full trajectory to a return, and a selection therefore maps the
n full trajectories to n real-valued returns.
There are nH terminal nodes across the n full trajectories.
Thus there are at most nH distinct real-valued returns on the
n full trajectories under Π. If we set the indicator threshold
∆ to equal each of these nH returns in turn, there would be
d
atmost ( enH
d ) distinct binary labelings of the full trajectories for each such ∆. Thus, the set of indicator functions that
d
define VCr (Π) can generate atmost nH( enH
d ) distinct labelings on n full trajectories. To shatter the n full trajectories,
d
n
we set nH( enH
d ) ≥ 2 , and the result follows.
We now proceed similarly to Theorem 3.2 in Kearns et al.
Theorem 1. Let Π be a potentially infinite set of deterministic
optimal stopping policies and let d be the VC-dimension of Π.
Let Γ be n full trajectories collected from the environment,
and let Vˆπ be the value estimates for π ∈ Π using Γ. Let the
return Vπ be bounded by Vmax for any trajectory. If
2
!
1
Vmax
d log H + log
n=O
δ
then with probability at least 1 − δ, |Vπ − V̂π | ≤ holds
simultaneously for all π ∈ Π.
Proof. Let X be the space of full trajectories. Every policy π ∈ Π is a bounded real-valued map π : X →
[−Vmax , Vmax ]. Let x1 , . . . , xn ∼ P be i.i.d. full trajectories generated by the environment dynamics. Using a result of [Vapnik and Kotz, 1982],Pwe have with
n
− n1 i=1 π(xi ) ≤
probability
1 − δ, supπ∈Π EP [π(x)]
q
VCr (Π) log VC n(Π) +log δ1
r
O Vmax
. Substitute EP [π(x)] =
n
P
n
Vπ , n1 i=1 π(xi ) = V̂π , VCr (Π) = O(d log H) in the inequality and upper-bound by to get the result.
In practice it may be impossible for us to evaluate every
policy in Π, and then select the one with the best estimated
mean. In such cases, we can use a different search method (S
in Algorithm 1) to find a local optima in Π, while using our
bound to ensure that policy values are estimated accurately.
Lastly, we discuss [Tsitsiklis and Van Roy, 1999], who
estimate Q values for finite-horizon Markov optimal stopping problems using a linear combination of basis functions,
and then use that to find a threshold policy. They outline a
procedure to tune the basis function weights that asymptotically guarantees their policy value’s convergence to the best
basis-function approximation. Under their assumptions, if we
construct a policy class using basis functions, we inherit the
useful convergence results relying on their search procedure,
along with retaining our finite sample complexity results.
Experiments
We now demonstrate the setting we consider is sufficiently
general to capture several problems of interest and that our
approach, GFSE, can improve performance in optimal stopping problems over some state-of-the-art baselines.
Ticket Purchase
Many purchasing problems can be posed as an optimal stopping process where the return from stopping is simply the advertised cost. We consider deciding when to purchase an airline ticket for a later trip date in order to minimize cost. The
opaque way in which prices are set, and competitive pricing
makes this domain difficult to model. Prior work [Etzioni et
al., 2003; Groves and Gini, 2015] has focused on identifying
features to create sophisticated models that make good purchase decisions. Surprisingly, it can be hard to improve on an
earliest purchase baseline that buys after the first observation.
We use data from Groves and Gini (2015) who collected
real pricing data for a fixed set of routes over a period of 2
years, querying travel sites regularly to collect price information. Each route has several departure dates distributed over
the 2 year period. For a price observation sequence of length
Table 1: Mean expenditure of deploying different policies on
the test set for ticket purchase. Earliest purchase buys immediately, latest purchase waits till the departure date.
Method
NYC-MSP
MSP-NYC
SEA-IAD
Ours (Πsimple )
Ours (Πcomplex )
Earliest purchase
Latest purchase
$355
$351
$380
$631
$374
$344
$383
$647
$565
$560
$578
$938
Best possible price
$307
$306
$513
T , a customer could commence his ticket search at any point
in the sequence (e.g. some customer starts 60 days before departure while another only a week before). Thus, we consider
all such commencement points separately to get T distinct
full trajectories (similar to [Groves and Gini, 2015]).
We construct a parameterized policy class (Πsimple ) based
on Ripper’s decision rules in [Etzioni et al., 2003]: WAIT if
(curr price > θ0 AND days to depart > θ1 ) else BUY,
where BUY corresponds to halting. We also constructed a
more complex class (Πcomplex ) with 6 parameters, that learns
different price thresholds depending on how far the departure
date is. We consider nonstop flights on 3 routes, NYC-MSP,
MSP-NYC and SEA-IAD, training/testing each separately.
Our method, GFSE collects full length trajectories during the first 200 days (∼1000 trajectories) and uses them to
construct a single stopping policy. It performs a simple policy search by sampling and evaluating 500 policies randomly
from the policy space. It then uses the best identified policy
to simulate ticket purchasing decisions for departure dates occurring during the remaining part of the 2 years (∼2000 trajectories). We restrict the data to departure dates that contain
at-least 30 price observations.2
Results on the test sets are shown in Table 1. Our policy search method succeeds in finding a policy that leads to
non-trivial improvement over the difficult earliest purchase
baseline. Our improvements are in line with prior approaches
specifically designed for this particular domain.3 These results highlight how our setting can capture important purchasing tasks and how our approach, even with a simple policy
search, can find policies with significantly better performance
than competitive domain-specific baselines.
Tutoring and Asset Replacement
We now consider 2 simulated domains and compare GFSE
to several approaches for learning to act quickly in these
domains. Unless specified, all results are averaged over 20
rounds and error bars indicate 95% confidence intervals.
Baselines. One natural idea is to proceed as GFSE, but use
the gathered data to build parametric domain models that can
be used to estimate the performance of potential policies. We
2
We found that shorter trajectories were collected close to the
departure date, where prices fluctuate more and for which our illustrative policy classes are inadequate. In such cases, our method
adopted a risk-averse earliest purchase policy.
3
Unfortunately, the authors were unable to provide us with the
train/test split used in [Groves and Gini, 2015].
call these ”model-based” approaches. A second idea is to consider the initial set of collected data as a budget of free exploration, and instead use this budget to do Monte Carlo onpolicy evaluation of a set of policies.
Of course, doing all exploration, as we do in GFSE, is
not always optimal. We also consider a state-of-the-art approach for quickly identifying the global optima of a function where the function is initially unknown and each function
evaluation is expensive, Bayesian Optimization (BO). Multiple papers have shown BO can be used to speed online policy
search for reinforcement learning tasks [Wilson et al., 2014;
Deisenroth and Rasmussen, 2011]. Given the policy class, BO
selects a policy to evaluate at each step, and maintains estimates over the expected value of every policy. We use Yelp’s
MOE for BO [Yelp, 2016] with a Gaussian kernel and the
popular expected improvement heuristic for picking policies
from Π. The hyper-parameters for BO are picked by a separate optimization to find maximum-likelihood estimates.
Simulated Student Learning. We first consider a simulated
student tutor domain. A number of tutoring systems use mastery teaching, in which a student is provided with practice
examples until they are estimated to have mastered the material. This is an optimal stopping problem because at each time
step, after observing whether a student got the activity correct
or not, the tutor can decide whether to halt or continue providing the student with additional practice. On halting, the
student is given the next problem in the sequence; the objective is to maximize the score on this ‘posttest’, while giving
as few problems as possible overall. As is popular in the literature, we model student learning using the Bayesian Knowledge Tracing (BKT) model [Corbett and Anderson, 1995].
BKT is a 2-state Hidden Markov Model (HMM) with the
state capturing whether the student has mastered the skill or
not. Within the HMM, 4 probabilities – pi (prior mastery),
pt (transition to mastery), pg (guess) and ps (slip) describe
the model. To simulate student data, we fix BKT parameters4
pi = 0.18, pt = 0.2, pg = 0.2, ps = 0.1 and generate student
trajectories using this BKT model for H = 20 problems.
For GFSE, we consider two policy classes, both of which
halt when the probability of the student’s next response (according to the model in use) being correct crosses a threshold. Thus, we halt if Pr(ot = correct|o1 , . . . , ot−1 ) > θ0
where θ0 is some threshold. In fact, policies of this kind are
widely used in commercial tutoring systems [Koedinger et
al., 2013]. If we use the BKT model to implement this policy class, it is parameterized by (pi , pt , pg , ps , θ0 ). Π then
contains all possible instantiations of these parameters for
our model-free approach to search over. We also consider
a policy class based on another popular educational data
mining model of student learning: Additive Factors Model
(AFM) [Draney et al., 1995]. AFM is a logistic regression
model used to predict the probability that a student will get
the next problem correct given their past responses. Thus,
Pr(ot = correct|o1 , . . . , ot−1 ) = 1+e−(β11 +β2 nc ) where nc
is the number of correct past attempts.
We first note that GFSE is significantly more effective than
4
Our results hold for other instantiations of these parameters as
well. See [Ritter et al., 2009] for other reasonable parameter settings.
Figure 2: Comparison of best policy found on varying budget
under (a) matched model setting; (b) model mismatch setting.
taking the same budget of exploration, and using it to evaluate
each policy in an on-policy manner using Monte Carlo (MC)
estimation. More precisely, we sample k = 100 policies from
the BKT policy class, and fix a budget of B ∈ {100, 1000}
trajectories. GFSE uses B trajectories to evaluates all k policies while MC runs every policy on B
k trajectories (e.g. 1 trajectory/policy for B = 100) and selects the one with the highest mean performance. Averaging results across 20 separate
runs, we found that GFSE identifies a much better policy;
MC chose poor policies because it is mislead by the potential
performance of a policy due to the limited data.
We also explored the performance of building a model of
the domain, both in the setting when the model matches the
true domain (here, a BKT model), and a model-mismatch
case, where the policy class is based on a student AFM model
(which does not match the BKT process dynamics). We use
maximum likelihood estimation to fit the assumed model’s
parameters given the collected data and then separately optimize over the threshold parameters θ0 . We compare, on varying the budget B: (a) GFSE; (b) model-based; (c) BO. All
results are averaged over 50 trials.
The results of this experiment are shown in Figure 2. Our
approach does well in both settings, quickly finding a near optimal policy. As one would expect, the model-based approach
does well under the matched model setting, making full use
of the knowledge of the underlying process dynamics. However, on fitting the mismatched AFM model, the model-based
approach suffers. As has been noted by prior work [Mandel et
al., 2014] model-fitting procedures focus on maximizing the
likelihood of the observed data rather than trying to directly
identify a policy that is expected to perform well. BO can find
a good policy, but takes more samples to do so.
Since BO is an online approach whereas GFSE uses a fixed
budget of exploration, we also compare the averaged cumulative performance of BO to variants of GFSE in Figure 3. This
mimics a scenario where we care about online performance
on every individual trajectory, rather than having access to a
fixed budget before deploying a policy. For our method, we
can choose to collect more or less full trajectories before finding the best policy. Interestingly, if we use 5 trajectories as the
initial budget to collect full length trajectories, GFSE meets
or exceeds BO performance in this setting in both the matched
and mismatched model cases, within 10 trajectories.
BO suffers from the highly stochastic returns of policies in
this setting. For more efficient data reuse, we also consider a
variant of BO (BO-REuse) where we evaluate each proposed
Figure 3: Average cumulative performance on the simulated
student domain. (a) matched model; (b) model-mismatch.
GFSE-k collects k initial full-length trajectories before identifying and then executing the best policy.
Figure 4: Cumulative performance of augmented methods on
(a) matched model; (b) model-mismatch.
policy online and also using previously collected trajectories,
yielding a more robust estimate of the policy’s performance.
Similarly, for GFSE we deploy a policy using an initial budget of trajectories, and then use its on-policy trajectory (in addition to earlier trajectories) to rerun policy search and identify another policy for the next time step (GFSE-RE). Figure 4 shows this improved both methods (especially in the
mismatch case), with our approaches still performing best.
Asset Replacement. Another natural problem that falls into
our optimal stopping problem is when to replace a depreciating asset (such as a car, machine, etc). For simulation, we use
a model described in [Jiang and Powell, 2015]. Variants of
this model are widely used in that field [Feldstein and Rothschild, 1974; Rust, 1987]. In the model, observations are d
dimensional vectors of the form (X, Y1 , . . . , Yd−1 ). Each asset starts at a fixed valuation X = Xmax which depreciates
stochastically5 while emitting observations Y1 , . . . , Yd−1 at
every time step. The reward function used incorporates the
cost of replacement (which increases over time), the utility
derived from the asset and a penalty if the asset becomes
worthless before replacement. We use d = 3 for experiments.
We construct a logistic threshold policy class; replacing the
asset if 1+e−(β11+β2 ·depr) > β3 where depr is the total depreciation from Xmax seen so far (normalized to lie in [0, 1]).
In addition to the approaches seen before, we also include
baseline policies that choose to (i) replace the asset immediately; (ii) never replace. Lastly, we include the optimal value
(known only in hindsight) for reference.
The results are shown in Fig 5. Surprisingly, our method
5
Details of the model can be found in [Jiang and Powell, 2015].
Figure 5: Results from asset replacement.
outperforms competing methods by a considerable margin.
It appears that our chosen policy class is tricky to optimize
over: most policies in the space perform poorly. For 500 random policies chosen from this space, the mean cost is around
240 with a 95% confidence interval of only 16. However, the
domain itself is not very noisy, with robust value estimation
requiring less than 5 trajectories (see Figure 5). This enables
our method to consistently find a good policy even with a
low budget: one that corresponds to replacing the asset when
depreciation is around 50%. BO improves slowly; either sampling bad policies due to the sparse nature of the space, or disbelieving the estimate of a good policy due to the bad policies
surrounding it. Manually adjusting the BO hyperparameters
to account for this did not improve performance significantly.
Discussion and Conclusion
GFSE performed well, outperforming state-of-the-art algorithms and common baselines, in a variety of simulations of
important domains. While we randomly searched over policies in relatively simple policy classes for illustration, more
sophisticated search methods and policy classes could be employed, without effecting the theoretical guarantees we derived. Another extension is in using shorter trajectories that
terminate before the horizon for policy evaluation (similar
to how full trajectories are used). This is useful in a scenario where we get trajectories on-policy using the best policy found by GFSE. We can then rerun policy search with all
trajectories (full length or short) collected so far. Our policy
value estimates will be biased in this case, since only a policy that halts earlier than a shorter trajectory can use it for
evaluation. Values for policies that halt later may be overestimated (higher variance of estimation due to fewer trajectories), biasing us to pick them. If the number of evaluations
per policy exceeds the number in Theorem 1, our estimates
would remain within of the true values (with high probability), which would minimize the effect of this bias. As we saw
from Figure 4, this (GFSE-RE) works well empirically.
To summarize, we introduced a method for learning to act
in optimal stopping problems, which reuses full length trajectories to perform policy search. Our theoretical analysis and
empirical simulations demonstrate that this simple observation can lead to benefits in sample complexity and practice.
Acknowledgments
We appreciate the financial support of a NSF BigData award
#1546510, a Google research award and a Yahoo gift.
References
[Best et al., 2015] Graeme Best, Wolfram Martens, and
Robert Fitch. A spatiotemporal optimal stopping problem for mission monitoring with stationary viewpoints. In
Robotics: Science and Systems, 2015.
[Corbett and Anderson, 1995] Albert T. Corbett and John R.
Anderson. Knowledge tracing: Modelling the acquisition
of procedural knowledge. User Model. User-Adapt. Interact., 4(4):253–278, 1995.
[Deisenroth and Rasmussen, 2011] Marc Deisenroth and
Carl E Rasmussen. Pilco: A model-based and dataefficient approach to policy search. In ICML, pages
465–472, 2011.
[Draney et al., 1995] Karen L Draney, Peter Pirolli, and
Mark Wilson. A measurement model for a complex cognitive skill. Cognitively diagnostic assessment, pages 103–
125, 1995.
[Etzioni et al., 2003] Oren Etzioni, Rattapoom Tuchinda,
Craig A. Knoblock, and Alexander Yates. To buy or not to
buy: mining airfare data to minimize ticket purchase price.
In KDD, pages 119–128, 2003.
[Feldstein and Rothschild, 1974] Martin S Feldstein and
Michael Rothschild. Towards an economic theory of
replacement investment. Econometrica: Journal of the
Econometric Society, pages 393–423, 1974.
[Feng and Gallego, 1995] Youyi Feng and Guillermo Gallego. Optimal starting times for end-of-season sales and
optimal stopping times for promotional fares. Management Science, 41(8):1371–1391, 1995.
[Ferguson, 1989] Thomas S Ferguson. Who solved the secretary problem? Statistical science, pages 282–289, 1989.
[Glower et al., 1998] Michel Glower, Donald R Haurin, and
Patric H Hendershott. Selling time and selling price: The
influence of seller motivation. Real estate economics,
26(4):719–740, 1998.
[Groves and Gini, 2015] William Groves and Maria L. Gini.
On optimizing airline ticket purchase timing. ACM TIST,
7(1):3:1–3:28, 2015.
[Jacka, 1991] SD 1 Jacka. Optimal stopping and the american put. Mathematical Finance, 1(2):1–14, 1991.
[Jiang and Powell, 2015] Daniel R. Jiang and Warren B.
Powell. An approximate dynamic programming algorithm for monotone value functions. Operations Research,
63(6):1489–1511, 2015.
[Kearns et al., 1999] Michael J. Kearns, Yishay Mansour,
and Andrew Y. Ng. Approximate planning in large pomdps
via reusable trajectories. In NIPS, pages 1001–1007, 1999.
[Koedinger et al., 2013] Kenneth R Koedinger, Emma Brunskill, Ryan SJd Baker, Elizabeth A McLaughlin, and John
Stamper. New potentials for data-driven intelligent tutoring system development and optimization. AI Magazine,
34(3):27–41, 2013.
[Levine and Abbeel, 2014] Sergey Levine and Pieter
Abbeel. Learning neural network policies with guided
policy search under unknown dynamics. In NIPS, pages
1071–1079, 2014.
[Lippman and McCall, 1976] Steven A Lippman and John
McCall. The economics of job search: A survey. Economic inquiry, 14(2):155–189, 1976.
[Mandel et al., 2014] Travis Mandel, Yun-En Liu, Sergey
Levine, Emma Brunskill, and Zoran Popovic. Offline policy evaluation across representations with applications to
educational games. In AAMAS, pages 1077–1084. International Foundation for Autonomous Agents and Multiagent
Systems, 2014.
[Mordecki, 2002] Ernesto Mordecki. Optimal stopping and
perpetual options for lévy processes. Finance and Stochastics, 6(4):473–493, 2002.
[Ng and Jordan, 2000] Andrew Y. Ng and Michael I. Jordan.
PEGASUS: A policy search method for large mdps and
pomdps. In UAI, 2000.
[Peskir and Shiryaev, 2006] Goran Peskir and Albert
Shiryaev. Optimal stopping and free-boundary problems.
Springer, 2006.
[Precup, 2000] Doina Precup. Eligibility traces for offpolicy policy evaluation. Computer Science Department
Faculty Publication Series, page 80, 2000.
[Ritter et al., 2009] Steven Ritter, Thomas K. Harris, Tristan
Nixon, Daniel Dickison, R. Charles Murray, and Brendon
Towle. Reducing the knowledge tracing space. In Educational Data Mining, pages 151–160, 2009.
[Rust, 1987] John Rust. Optimal replacement of gmc bus engines: An empirical model of harold zurcher. Econometrica: Journal of the Econometric Society, pages 999–1033,
1987.
[Sutton et al., 1999] Richard S. Sutton, David A.
McAllester, Satinder P. Singh, and Yishay Mansour.
Policy gradient methods for reinforcement learning with
function approximation. In NIPS, pages 1057–1063,
1999.
[Tsitsiklis and Van Roy, 1999] John N Tsitsiklis and Benjamin Van Roy. Optimal stopping of markov processes:
Hilbert space theory, approximation algorithms, and an application to pricing high-dimensional financial derivatives.
IEEE Transactions on Automatic Control, 44(10):1840–
1851, 1999.
[Vapnik and Kotz, 1982] Vladimir Naumovich Vapnik and
Samuel Kotz. Estimation of dependences based on empirical data, volume 40. Springer-Verlag New York, 1982.
[Wilson et al., 2014] Aaron Wilson, Alan Fern, and Prasad
Tadepalli. Using trajectory data to improve bayesian optimization for reinforcement learning. Journal of Machine
Learning Research, 15(1):253–282, 2014.
[Yelp, 2016] Yelp. Metric optimization engine. https://
github.com/Yelp/MOE, 2016.
[Zilberstein, 1995] Shlomo Zilberstein. Operational rationality through compilation of anytime algorithms. AI Magazine, 16(2):79, 1995.
| 2cs.AI
|
Exact Learning of Lightweight Description Logic Ontologies
Exact Learning of Lightweight Description Logic Ontologies
Boris Konev
[email protected]
Department of Computer Science
University of Liverpool, United Kingdom
arXiv:1709.07314v1 [cs.LG] 20 Sep 2017
Carsten Lutz
[email protected]
Department of Computer Science
University of Bremen, Germany
Ana Ozaki
[email protected]
Department of Computer Science
University of Liverpool, United Kingdom
Frank Wolter
[email protected]
Department of Computer Science
University of Liverpool, United Kingdom
Editor:
Abstract
We study the problem of learning description logic (DL) ontologies in Angluin et al.’s framework of exact learning via queries. We admit membership queries (“is a given subsumption
entailed by the target ontology?”) and equivalence queries (“is a given ontology equivalent
to the target ontology?”). We present three main results: (1) ontologies formulated in
(two relevant versions of) the description logic DL-Lite can be learned with polynomially
many queries of polynomial size; (2) this is not the case for ontologies formulated in the
description logic EL, even when only acyclic ontologies are admitted; and (3) ontologies
formulated in a fragment of EL related to the web ontology language OWL 2 RL can be
learned in polynomial time. We also show that neither membership nor equivalence queries
alone are sufficient in cases (1) and (3).
Keywords: Exact Learning, Description Logic, Complexity
1. Introduction
In many subfields of artificial intelligence, ontologies are used to provide a common vocabulary for the application domain of interest and to give a meaning to the terms in the vocabulary, and to describe the relations between them. Description logics (DLs) are a prominent
family of ontology languages with a long history that goes back to Brachman’s famous
knowledge representation system KL-ONE in the early 1980s (Brachman and Schmolze,
1985). Today, there are several widely used families of DLs that differ in expressive power,
computational complexity, and intended application. The most important ones are the ALC
family which aims at high expressive power, the EL family (Baader et al., 2005) which aims
to provide scalable reasoning, and the DL-Lite family (Calvanese et al., 2007; Artale et al.,
2009) which is tailored specifically towards applications in data access. In 2004, the World
Wide Web Committee (W3C) has standardised a DL of the ALC family as an ontology
1
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
language for the web, called OWL. The standard was updated to OWL 2 in 2009, and since
then comprises a family of five languages including the OWL 2 profiles OWL 2 EL, OWL 2
QL, and OWL 2 RL. While OWL 2 EL is based on EL and OWL 2 QL on DL-Lite, OWL 2
RL is closely related to the fragment of EL that is obtained by allowing only concept names
on the right-hand side of concept inclusions. In this paper we study DLs from the EL and
DL-Lite families.
Designing an ontology for an application domain is a subtle, error-prone, and time consuming task. From its beginnings, DL research was driven by the aim to provide various
forms of support for ontology engineers, assisting them in the design of high-quality ontologies; examples include the ubiquitous task of ontology classification (Baader et al., 2017),
reasoning support for debugging ontologies (Wang et al., 2005; Schlobach et al., 2007), support for modular ontology design (Stuckenschmidt et al., 2009), and checking the completeness of the modelling in a systematic way (Baader et al., 2007). The same aim is pursued
by the field of ontology learning, where the goal is to use machine learning techniques for
various ontology engineering tasks such as to identify the relevant vocabulary of the application domain (Cimiano et al., 2010; Wong et al., 2012), to learn an initial version of the
ontology that is then refined manually (Borchmann and Distel, 2011; Ma and Distel, 2013;
Jiménez-Ruiz et al., 2015), and to learn concept expressions as building blocks of an ontology (Lehmann and Hitzler, 2010); see the recent collection (Lehmann and Völker, 2014)
and the related work section at the end of the paper for details.
In this paper we concentrate on learning the full logical structure of a description logic
ontology. Our starting point is the observation that building a high-quality ontology relies
on the successful communication between an ontology engineer and a domain expert because
the former is typically not sufficiently familiar with the domain and the latter is rarely an
expert in ontology engineering. We study the foundations of this communication process
in terms of a simple communication model and analyse, within this model, the complexity
of constructing a correct and complete domain ontology. Our model rests on the following
assumptions:
1. The domain expert has perfect knowledge of the domain, but is not able to formalise
or communicate the target ontology O to be constructed.
2. The domain expert is able to communicate the vocabulary (predicate symbols, which
in the case of DLs take the form of concept and role names) of O and shares it with
the ontology engineer. The ontology engineer knows nothing else about the domain.
3. The ontology engineer can pose queries to the domain expert which the domain expert
answers truthfully. The main queries posed by the ontology engineer are of the form
“Is the concept inclusion C ⊑ D entailed by O?”
4. In addition, the ontology engineer needs a way to find out whether the ontology H
that has been constructed so far, called the hypothesis ontology, is complete. If not,
he requests an example illustrating the incompleteness. The engineer can thus ask:
“Is the ontology H complete? If not, then return a concept inclusion C ⊑ D
entailed by O but not by H.”
2
Exact Learning of Lightweight Description Logic Ontologies
We are then interested in whether the target ontology O can be constructed with only polynomially many queries of polynomial size (polynomial query learnability) or, even better,
with overall polynomial time (polynomial time learnability). In both cases, the polynomial is
in the size of the ontology to be constructed plus the size of the counterexamples returned
by the domain expert. Without taking into account the latter, one can never expect to
achieve polynomial time learnability because the domain expert could provide unnecessarily large counterexamples. Note that polynomial time learnability implies polynomial query
learnability, but that the converse is false because polynomial query learnability allows the
ontology engineer to run computationally costly procedures between posing queries.
The above model is an instance of Angluin et al.’s framework of exact learning via
queries (Angluin, 1987b). In this context, the queries mentioned in Point 3 above are called
membership queries. The queries in Point 4 are a form of equivalence queries. In Angluin’s
framework, however, such queries are slightly more general:
“Is the hypothesis ontology H equivalent to the target ontology O? If not,
then return a concept inclusion C ⊑ D entailed by O but not by H (a positive
counterexample) or vice versa (a negative counterexample).”
In our upper bounds (that is, polynomial learnability results), we admit only queries of the
more restricted form in Point 4 above: the learning algorithm is designed in a way so that
the hypothesis ontology H is a consequence of the target ontology O at all times, and thus
the only meaningful equivalence query is a query of the form “Is H already complete?”.
Our lower bounds (results saying that polynomial learnability is impossible), in contrast,
apply to unrestricted equivalence queries, that is, they do not assume that the hypothesis
is implied by the target. In this way, we achieve maximum generality.
Within the setup outlined above, we study the following description logics:
(a) DL-Lite∃R , which is a member of the DL-Lite family that admits role inclusions and
allows nested existential quantification on the right-hand side of concept inclusions;
(b) the extension DL-Lite∃R,horn of DL-Lite∃R with conjunction on the left-hand side of
concept inclusions;
(c) the basic member EL of the EL family;
(d) the fragment ELlhs of EL where only concept names (but no compound concept expressions) are admitted on the right-hand side of concept inclusions.
We remark that DL-Lite∃R is closely related to OWL 2 QL, which is based on the fragment
of DL-Lite∃R that does not allow nested existential quantification on the right-hand side of
concept inclusions. In this more restricted case, though, polynomial learnability is uninteresting. In fact, the number of concept inclusions formulated in a fixed finite vocabulary Σ
is bounded polynomially in the size of Σ instead of being infinite as in the description logics
studied in this paper; consequently, TBoxes are trivially learnable in polynomial time, even
when only membership queries (but no equivalence queries) are available or vice versa. The
extension DL-Lite∃R,horn of DL-Lite∃R is not part of the OWL 2 QL standard, but admitting
conjunctions on the left-hand side of concept inclusions is a useful and widely considered
3
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
Not polynomial query learnable
EL
DL-Lite∃R DL-Lite∃R,horn
Polynomial query learnable
ELrhs
Polynomial time learnable
ELlhs
Figure 1: Summary of main results
extension of basic DL-Lite dialects, see for example (Artale et al., 2009). ELlhs is a significant part of the OWL 2 RL language and can be viewed as a natural fragment of Datalog.
An even better approximation of OWL 2 RL would be the extension of ELlhs with inverse
roles, but polynomial learnability in that language remains an open problem. And finally,
unrestricted EL can be viewed as a logical core of the OWL 2 EL language.
After introducing preliminaries in Section 2, we study exact learning of DL-Lite∃R ontologies in Section 3, establishing polynomial query learnability. We strengthen this result
to DL-Lite∃R,horn in Section 4, using a significantly more subtle algorithm. It remains open
whether DL-Lite∃R and DL-Lite∃R,horn admit polynomial time learnability. Our algorithms
do not yield such a stronger result since they use subsumption checks to analyse counterexamples provided by the oracle and to integrate them into the current hypothesis ontology,
and subsumption is NP-complete in these DLs (Kikot et al., 2011). In Section 5, we show
that ELlhs ontologies are learnable in polynomial time, a result that extends the known
polynomial time learnability of propositional Horn formulas (Angluin et al., 1992), which
correspond to EL ontologies without existential restrictions. In fact, our algorithms take
inspiration from learning algorithms for propositional Horn formulas and combine the underlying ideas with modern concepts from DL such as canonical models, simulations, and
products. The algorithm for ELlhs also uses subsumption checks, which in this case does not
get in the way of polynomial time learnability since subsumption in ELlhs can be decided
in polynomial time.
In Section 6, we then establish that EL ontologies are not polynomial query learnable.
Note that the fragment ELrhs of EL, which is symmetric to ELlhs and only admits concept
names on the left-hand side of concept inclusions is a fragment of DL-Lite∃R . Together,
our upper bounds for DL-Lite∃R and ELlhs thus establish that failure of polynomial query
learnability of EL ontologies is caused by the interaction between existential restrictions
on the left- and right-hand sides of concept inclusions. Interestingly, our result already
applies to acyclic EL TBoxes, which disallow recursive definitions of concepts and are of
a rather restricted syntactic form. However, the result does rely on concept inclusions as
counterexamples that are of a form not allowed in acyclic TBoxes. We also show that
ontologies formulated in DL-Lite∃R,horn and in ELlhs are neither polynomial query learnable
with membership queries alone nor with equivalence queries alone; corresponding results for
propositional Horn formulas can be found in (Frazier and Pitt, 1993; Angluin et al., 1992;
Angluin, 1987a), see also (Arias and Balcázar, 2011). Figure 1 summarises the main results
obtained in this paper.
In Section 7 we provide an extensive discussion of related work on the exact learning of
logical formulas and theories, and we close the paper with a discussion of open problems.
A small number of proofs are deferred to an appendix.
4
Exact Learning of Lightweight Description Logic Ontologies
2. Preliminaries
We introduce the description logics studied in this paper, then consider a representation
of concept expressions in terms of labelled trees and show how important semantic notions
such as subsumption between concept expressions can be characterised by homomorphisms
between the corresponding trees. This also involves introducing canonical models, which
are an important tool throughout the paper. Finally, we formally introduce the framework
of exact learning.
2.1 Description Logics
Let NC be a countably infinite set of concept names (denoted by upper case letters A, B, etc)
and let NR be a countably infinite set of role names disjoint from NC (denoted by lower case
letters r, s, etc). Concept and role names can be regarded as unary and binary predicates,
respectively. In description logic, constructors are used to define compound concept and
role expressions from concept and role names. In this paper, the only role constructor is the
inverse role constructor: for r ∈ NR , the expression r − is the inverse role of r. Semantically,
r − represents the converse of the binary relation r. A role expression is a role name or an
inverse role. We set r − := s if r = s− for a role name s. For brevity, we will typically
speak of roles rather than of role expressions. The concept constructors used in this paper
are ⊤ (everything), ⊓ (conjunction), and ∃r.C (qualified existential restriction). Formally,
concept expressions C are defined according to the following syntactic rule:
C, D
:=
⊤
| A | C ⊓D
| ∃r.C
where A is a concept name and r is a role. For example, ∃hasChild.⊤ ⊓ ∃gender.Male denotes
the class of individuals who have a child and whose gender is male.
Terminological knowledge is captured by finite sets of inclusions between concept expressions or roles. Specifically,
• a concept inclusion (CI) takes the form C ⊑ D, where C and D are concept expressions, and
• a role inclusion (RI) takes the form r ⊑ s, where r and s are roles.
An ontology or TBox is a finite set of CIs and RIs.1 We use C ≡ D as an abbreviations
for the two CIs C ⊑ D and D ⊑ C and likewise for r ≡ s; we speak of concept equivalences
(CEs) and role equivalences (REs), respectively.
1. In the description logic literature, CIs of the form introduced here are often called E LI CIs to distinguish
them from CIs that use concept expressions formulated in other description logics. The TBoxes are called
E LIH TBoxes (TBoxes that consist of E LI CIs and RIs).
5
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
Example 1 Consider the following TBox:
Prof ⊑ ∃supervisor of.Student ⊓ ∃conduct research.⊤
Graduate ≡ ∃has degree.⊤
GraduateStudent ≡ Student ⊓ Graduate
−
GraduateStudent ⊑ ∃supervisor of .Prof
(1)
(2)
(3)
(4)
supervisor of ⊑ advisor of
(5)
CS Graduate ≡ ∃has degree.CS Degree
(6)
The CI in Line 1 states that every professor supervises students and conducts research.
Notice that we do not specify the specific area of research, hence we use an unqualified
existential restriction of the form ∃r.⊤. The CE in Line 2 defines a graduate as anyone
who has a degree. The CE in Line 3 defines a graduate student as a student who is a
graduate. The CI in Line 4 states that graduate students are supervised by professors.
Notice that we use the inverse role of supervisor of here. Line 5 shows an RI which states
that every supervisor is an advisor. The CE in the last line defines a computer science
graduate as someone with a degree in computer science.
A signature is a set of concept and role names and we use ΣT to denote the signature of
the TBox T , that is, the set of concept and role names that occur in it. The size |C| of
a concept expression C is the length of the string that represents C, where concept names
and
P role names are considered to be of length one. The size |T | of a TBox T is defined as
C⊑D∈T |C| + |D|.
The semantics of concept expressions and TBoxes is defined as follows (Baader et al.,
2017). An interpretation I = (∆I , ·I ) is given by a non-empty set ∆I (the domain of I)
and a mapping ·I that maps every concept name A to a subset AI of ∆I and every role
name r to a subset r I of ∆I × ∆I . The interpretation r I of an inverse role r = s− is
given by r I = {(d, d′ ) | (d′ , d) ∈ sI } and the interpretation C I of a concept expression C is
defined inductively by
⊤I
(C1 ⊓ C2 )I
(∃r.C)I
= ∆I
= C1I ∩ C2I
= {d ∈ ∆I | there exists d′ ∈ C I with (d, d′ ) ∈ r I }.
An interpretation I satisfies a concept expression C if C I is not empty. It satisfies the CI
C ⊑ D if C I ⊆ D I , written as I |= C ⊑ D. Similarly, I satisfies RI r ⊑ s if r I ⊆ sI ,
written as I |= r ⊑ s. I is a model of a TBox T if it satisfies all CIs and RIs in T . A
TBox T entails a CI or RI α, in symbols T |= α, if α is satisfied in every model of T .
Concept expressions C and D are equivalent w.r.t. T , written T |= C ≡ D, if T |= C ⊑ D
and T |= D ⊑ C; equivalence of roles r and s is defined accordingly, written T |= s ≡ r.
TBoxes T and T ′ are logically equivalent, in symbols T ≡ T ′ , if T |= α for all α ∈ T ′ and
vice versa.
6
Exact Learning of Lightweight Description Logic Ontologies
CS Deg
d1
d5
has deg
Stud Grad
d4
Grad Stud
d3
con res
d0
CS Grad
Grad
d2
Prof
has deg
sup of
adv of
Figure 2: Illustration to Example 2.
Example 2 Consider the TBox T from Example 1 and the interpretation I that is illustrated in Figure 2 and defined by setting ∆I = {d0 , . . . , d5 } and
Prof I
StudentI
Graduate StudentI
GraduateI
CS GraduateI
=
=
=
=
=
{d2 },
{d4 },
{d4 },
{d4 , d0 },
{d0 },
conduct researchI
supervisor of I
advisor of I
has degreeI
CS DegreeI
=
=
=
=
=
{(d2 , d3 )}
{(d2 , d4 )},
{(d2 , d4 )},
{(d4 , d5 ), (d0 , d1 )},
{d1 }.
It is easy to see that I is a model of T . Moreover, I 6|= Graduate ⊑ CS Graduate as
GraduateI = {d4 , d0 } but CS GraduateI = {d0 }, thus T 6|= Graduate ⊑ CS Graduate. It can
be shown that T |= CS Graduate ⊑ Graduate.
It is ExpTime-complete to decide, given a TBox T and a concept inclusion C ⊑ D, whether
T |= C ⊑ D (Baader et al., 2008); this reasoning problem is known as subsumption. Because
of this high complexity, the profiles of OWL 2 are based on syntactically more restricted
description logics in which subsumption is less complex. We next introduce a few relevant
such logics. A basic concept is a concept name or a concept expression of the form ∃r.⊤,
where r is a role. For example, ∃hasChild− .⊤ is a basic concept, but ∃hasChild− .Graduate is
not.
DL-Lite∃R . A DL-Lite∃R CI takes the form
B⊑C
where B is a basic concept and C is a concept expression. A DL-Lite∃R inclusion is a
DL-Lite∃R CI or an RI. A DL-Lite∃R TBox is a finite set of DL-Lite∃R inclusions.
Example 3 Lines (1), (4), and (5) of Example 1 are DL-Lite∃R inclusions and Line (2)
abbreviates the two DL-Lite∃R CIs Graduate ⊑ has degree.⊤ and has degree.⊤ ⊑ Graduate.
Lines (3) and (6) do not fall within DL-Lite∃R .
DL-Lite∃R,horn . In the extension DL-Lite∃R,horn of DL-Lite∃R , CIs take the form
B1 ⊓ · · · ⊓ Bn ⊑ C
where B1 , . . . , Bn are basic concepts and C is a concept expression. A DL-Lite∃R,horn TBox
T is a finite set of DL-Lite∃R,horn CIs and RIs. Both DL-Lite∃R and DL-Lite∃R,horn have been
investigated in detail in (Artale et al., 2009).
7
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
Example 4 As Lines (1), (2), (4) and (5) from Example 1 fall within DL-Lite∃R , they
also fall within DL-Lite∃R,horn . Line (3) falls within DL-Lite∃R,horn . Line (6) is not in
DL-Lite∃R,horn .
EL. An EL concept expression is a concept expression that does not use inverse roles. An
EL concept inclusion is a CI of the form
C⊑D
where C and D are EL concept expressions. An EL TBox is a finite set of EL CIs. Thus, EL
does neither admit role inclusions nor inverse roles. In contrast to DL-Lite∃R,horn , however,
it allows existential restrictions ∃r.C with C 6= ⊤ on the left-hand side of CIs.
Example 5 Inclusions (1), (2), (3) and (6) from Example 1 are EL inclusions. Inclusion
(5) is not an EL inclusion.
Subsumption is NP-complete in DL-Lite∃R and in DL-Lite∃R,horn , see (Kikot et al., 2011)
for the lower bound and (Calvanese et al., 2007; Artale et al., 2009) for the upper bound.
Subsumption in EL is in PTime (Baader et al., 2005) and this is still true if RIs that do
not use inverse roles are admitted in the TBox. Given a TBox T and an RI r ⊑ s, deciding
whether T |= r ⊑ s is possible in PTime in all description logics considered in this paper.
In fact, T |= r ⊑ s if, and only if, there exists a sequence r0 , . . . , rn of roles such that r = r0 ,
−
s = rn , and for every i < n either ri ⊑ ri+1 ∈ T or ri− ⊑ ri+1
∈ T . Our learning algorithms
will carry out various subsumption checks as a subprocedure, as detailed later on.
2.2 Tree representation of concept expressions
To achieve an elegant and succinct exposition of our learning algorithms, it will be convenient to represent concept expressions C as a finite directed tree TC whose nodes are
labelled with sets of concept names and whose edges are labelled with roles, and to describe
manipulations of concept expressions in terms of manipulations of the corresponding tree
such as merging nodes, replacing subgraphs, modifying node and edge labels, etc. We generally use ρC to denote the root node of the tree TC . In detail, TC is defined as follows. For
C = ⊤, the tree TC has a single node d with label l(d) = ∅; if C = A, where A is a concept
name, then TC has a single node d with l(d) = {A}; if C = ∃r.D, then TC is obtained from
TD by adding a new root d0 and an edge from d0 to the root d of TD with label l(d0 , d) = r
(we then call d an r-successor of d0 ); if C = D1 ⊓ D2 , then TC is obtained by identifying
the roots of TD1 and TD2 .
Example 6 For C = Student ⊓ ∃has degree.∃has degree− .Graduate Student, TC has three
nodes, e0 , e1 , e2 , where e0 is the root ρC of TC , e1 is a successor of e0 and e2 is a successor
of e1 , the labelling of the nodes is given by l(e0 ) = {Student}, l(e1 ) = ∅, and l(e2 ) =
{Graduate Student}, and the labelling of the edges is given by l(e0 , e1 ) = has degree and
l(e1 , e2 ) = has degree− , see Figure 3 (left).
8
Exact Learning of Lightweight Description Logic Ontologies
Grad Stud
e2
has deg−
Grad
e5
e4
e1
con res
has deg
e0
Stud
e3
Prof
sup of
Figure 3: Illustration to Examples 6 (left) and 7 (right).
Conversely, every labelled finite directed tree T of the described form gives rise to a concept
expression CT in the following way: if T has a single node d labelled by {A1 , . . . , An }, then
CT = A1 ⊓ · · · ⊓ An (we treat ⊤ as the empty conjunction here, so if l(d) = ∅ then CT = ⊤).
Inductively, let d be the root of T labelled with l(d) = {A1 , . . . , An }, let d1 , . . . , dm be
the successors of d, and let l(d, d1 ) = r1 , . . . , l(d, dm ) = rm . Assume Cd1 , . . . , Cdm are the
concept expressions corresponding to the subtrees of T with roots d1 , . . . , dm , respectively.
Then CT = A1 ⊓ · · · ⊓ An ⊓ ∃r1 .Cd1 ⊓ · · · ⊓ ∃rm .Cdm .
Example 7 Let T be the tree with root e3 labelled by {Prof} and successors e4 , e5 labelled by
∅ and {Graduate}, respectively, and with edge labelling given by l(e3 , e4 ) = conduct research
and l(e3 , e5 ) = supervisor of. Then CT = Prof⊓∃conduct research.⊤⊓∃supervisor of.Graduate;
see Figure 3 (right).
In what follows, we will not always distinguish explicitly between C and its tree representation TC which allows us to speak, for example, about the nodes and subtrees of a
concept expression.
One important use of the tree representation of concept expressions is that both the
truth relation ‘d ∈ C I ’ and the entailment ‘T |= C ⊑ D’ can be characterised in terms of
homomorphisms between labelled trees and interpretations. A mapping h from a tree TC
corresponding to a concept expression C to an interpretation I is a homomorphism if A ∈
l(d) implies h(d) ∈ AI for every concept name A and r = l(d, d′ ) implies (h(d), h(d′ )) ∈ r I
for all role names r. The following characterisation of the truth relation d ∈ C I by means
of homomorphisms is well-known.
Lemma 8 Let I be an interpretation, d ∈ ∆I , and C a concept expression. Then d ∈ C I
if, and only if, there is a homomorphism from TC to I mapping ρC to d.
The proof is by a straightforward induction on the structure of C, see for example Baader et al.
(1999) for details.
Example 9 Consider the interpretation I from Example 2 and the tree representations of
the concept expressions given in Figure 3. It can be seen that functions g and h defined as
g(e0 ) = d4 , g(e1 ) = d5 , g(e2 ) = d4 and h(e3 ) = d2 , h(e4 ) = d3 , h(e5 ) = d4 are homomorphisms and so, by Lemma 8, d4 ∈ (Student ⊓ ∃has degree.∃has degree− .Graduate Student)I
and d2 ∈ (Prof ⊓ ∃conduct research.⊤ ⊓ ∃supervisor of.Graduate)I .
9
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
It is also standard to characterise the subsumption relation ∅ |= C ⊑ D (that is, subsumption relative to the empty TBox) by means of homomorphisms between the tree representations TD and TC . A homomorphism h from labelled tree T1 to labelled tree T2 is a mapping
from the nodes of T1 to the nodes of T2 such that A ∈ l(d) implies A ∈ l(h(d)) for every
concept name A and r = l(d, d′ ) implies r = l(h(d), h(d′ )) for every role r.
Lemma 10 Let C and D be concept expressions. Then ∅ |= C ⊑ D if, and only if, there
is a homomorphism from TD to TC that maps ρD to ρC .
The ‘if’ direction is essentially a consequence of Lemma 8 and the fact that the composition
of two homomorphisms is again a homomorphism. For the ‘only if’ direction, one can
consider TC as an interpretation I and apply Lemma 8. We again refer to Baader et al.
(1999) for details.
Next, we characterise subsumption in the presence of TBoxes in terms of homomorphisms. To achieve this, we make use of the canonical model IC0 ,T of a concept expression C0 and a TBox T . If T = ∅, then we want IC0 ,T to be TC0 viewed as a tree-shaped
interpretation which we denote by IC0 rather than by IC0 ,T . More precisely, the domain of
IC0 is the set of nodes of TC0 and
d ∈ AIC0
iff
A ∈ l(d), for all d ∈ ∆IC0 and concept names A
(d, d′ ) ∈ r IC0
iff
r = l(d, d′ ), for all d, d′ ∈ ∆IC0 and roles names r
We call the root ρC0 of TC0 the root of IC0 . If T 6= ∅, then IC0 ,T is obtained by extending
IC0 so that the CIs in T are satisfied. For example, if T = {A ⊑ ∃r.B} and C0 = A,
then IC0 is a single node ρC0 with AIC0 = {ρC0 } and X IC0 = ∅ for all concept and role
names X distinct from A. To define IC0 ,T we add a node d to ∆IC0 and set B IC0 ,T = {d}
and r IC0 ,T = {(ρC0 , d)}. In general, IC0 ,T is defined as the limit of a sequence I0 , I1 , . . . of
interpretations, where I0 = IC0 . For the inductive definition of the sequence, assume that
In has been defined. Then obtain In+1 by applying one of the following rules once:
1. if C ⊑ D ∈ T and d ∈ C In but d 6∈ D In , then take the interpretation ID and add it
to In by identifying its root ρC with d. In more detail, assume that ∆In ∩ ∆IC = {d}
and d = ρC and define In+1 by setting, for all concept names A and role names r:
∆In+1 = ∆In ∪ ∆IC ,
AIn+1 = AIn ∪ AIC ,
r In+1 = r In ∪ r IC ;
2. if r ⊑ s ∈ T and (d, d′ ) ∈ r In but (d, d′ ) 6∈ sIn , then define In+1 as In except that
sIn+1 := sIn ∪ {(d, d′ )} if s is a role name; otherwise there is a role name s0 with
In+1
s = s−
= sI0 n ∪ {(d′ , d)}.
0 and we define In+1 as In except that s0
We assume that rule application is fair, that is, if a rule is applicable in a certain place,
then it will indeed eventually be applied there. If for some n > 0 no rule is applicable then
we set In+1 = In . We obtain IC0 ,T by setting for all concept names A and role names r:
∆IC0 ,T =
[
∆ In ,
AIC0 ,T =
[
n≥0
n≥0
10
AIn ,
r IC0 ,T =
[
n≥0
r In .
Exact Learning of Lightweight Description Logic Ontologies
I0
I1
Grad
con res
sup of
Grad
sup of
adv of
I4
Stud
con res
Stud
con res
sup of
Prof
Prof
Grad
sup of
adv of
con res
sup of
adv of
Prof
I3
I2
Grad
has deg
Stud
Grad
sup of
adv of
sup of
adv of
Prof
con res
sup of
adv of
Prof
Figure 4: Canonical model construction for Example 11.
Note that the interpretation IC0 ,T obtained in the limit is tree-shaped and might be infinite.2
The following example illustrates the definition of IC0 ,T .
Example 11 Consider the following TBox T :
Prof ⊑ ∃supervisor of.Student
(7)
Prof ⊑ ∃conduct research.⊤
(8)
Graduate ⊑ ∃has degree.⊤
(9)
∃has degree.⊤ ⊑ Graduate
(10)
supervisor of ⊑ advisor of
(11)
and the concept expression
C0 = Prof ⊓ ∃conduct research.⊤ ⊓ ∃supervisor of.Graduate.
Figure 4 illustrates the steps of the canonical model construction with I0 being IC0 and I4
being the canonical model IC0 ,T .
The following lemma provides the announced characterisation of subsumption in the presence of TBoxes.
Lemma 12 Let T be a TBox and C a concept expression. Then IC,T is a model of T and
the following conditions are equivalent, for every concept expression D:
1. T |= C ⊑ D;
2. ρC ∈ D IC,T ;
3. there is a homomorphism from TD to IC,T that maps ρD to ρC .
2. The exact shape of IC0 ,T depends on the order of rule applications. However, all possible resulting
interpretations IC0 ,T are homomorphically equivalent and, as a consequence, the order of rule application
is not important for our purposes.
11
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
The proof is completely standard (see, for example, (Krötzsch, 2012)), we only give a highlevel overview. Using the construction of IC,T , it is not hard to show that IC,T is a model
of T and that ρC ∈ C IC,T . This implies ‘1 ⇒ 2’ and ‘2 ⇒ 3’ follows from Lemma 8. For ‘3
⇒ 1’, one can show that for any model I of T and any d ∈ C I , there is a homomorphism
from IC,T to I that maps ρC to d. In fact, one constructs a homomorphism to I from each
of the interpretations I0 , I1 , . . . built during the construction of IC,T , which is not hard
by analysing the rules applied during that construction. The homomorphism built for each
In+1 extends that for In and thus we can take the unions of all those homomorphisms to
obtain a homomorphism from IC,T to I. It remains to compose homomorphisms and apply
Lemma 8.
We are only going to use canonical models and Lemma 12 in the context of DL-Lite∃R and
DL-Lite∃R,horn . Next, we identify a more subtle property of canonical models of DL-Lite∃R
and DL-Lite∃R,horn TBoxes that we need later on. Very roughly speaking, it states a form
of locality which is due to the fact that existential restrictions on the left-hand side of CIs
in DL-Lite∃R,horn are unqualified. Assume that d ∈ (∃r.D)IC,T for the canonical model IC,T
of a concept expression C and TBox T . Assume d ∈ ∆IC . We know that there exists
a homomorphism h from T∃r.D to IC,T mapping ρ∃r.D to d. Then either h maps some
elements of TD into ∆IC or it maps the whole tree TD into ∆IC,T \ ∆IC . We are interested
in the latter case. The following lemma states that if T is a DL-Lite∃R TBox, then there
is a basic concept B with d ∈ B IC such that T |= B ⊑ ∃r.D. Thus, the question whether
d ∈ (∃r.D)IC,T only depends on the concept names A with d ∈ AIC and the roles r with
d ∈ (∃r.⊤)IC . If T is a DL-Lite∃R,horn TBox, it might not be sufficient to take a single
basic concept but at least a set of basic concepts suffices (corresponding to the fact that
DL-Lite∃R,horn admits conjunctions on the left-hand side of CIs). This observation does not
hold for EL TBoxes and is ultimately the reason for the fact that one cannot polynomially
learn EL TBoxes. The following example illustrates this observation.
Example 13 Consider the EL TBox T = {∃r.A ⊑ A, A ⊑ ∃r.B} and let C = ∃r.∃r.A.
Then ρC ∈ (∃r.B)IC,T since T |= C ⊑ ∃r.B. We therefore find a homomorphism h from
T∃r.B to IC,T mapping ρ∃r.B to ρC . This homomorphism maps TB (which has a single
node only) into ∆IC \ ∆IC,T . The only basic concept B with ρC ∈ B IC is ∃r.⊤ but clearly
T 6|= ∃r.⊤ ⊑ ∃r.B and so the observation we sketched above does not hold for EL.
We present this result in a more formal way. Let T1 and T2 be trees with labelling functions
l1 and l2 , respectively. We call T1 a subtree of T2 if the following conditions hold: T1 ⊆ T2 ,
l1 is the restriction of l2 to T1 , and if d ∈ T1 and d′ is a successor of d in T2 , then d′
is a successor of d in T1 as well. The one-neigbourhood NIC (d) of d ∈ ∆IC is the set of
concept names A with d ∈ AIC and basic concepts ∃r.⊤ such that there exists d′ ∈ ∆IC
with (d, d′ ) ∈ r IC .
Lemma 14 Let T be a DL-Lite∃R,horn TBox, D = ∃r.D ′ and assume h : TD → IC,T is
′
such that h(ρD ) = d ∈ ∆IC and the image of the subtree TD
d of TD under h is included in
I
I
C
C,T
\ ∆ . Then there exists I ⊆ NIC (d) such that T |= E∈I E ⊑ D. Moreover, if T is
∆
a DL-Lite∃R TBox, then there exists such a set I ⊆ NIC (d) with a single concept.
12
Exact Learning of Lightweight Description Logic Ontologies
Proof (sketch) This property of canonical models for DL-Lite∃R,horn has been proved implicitly in many papers, for example (Artale et al., 2009). We give a sketch. Let N be the
conjunction of all E ∈ NIC (d) and assume T 6|= N ⊑ D. Consider the canonical model
IN,T . By definition, the one-neighbourhoods of ρN in IN,T and d in IC,T coincide. Now
observe that the canonical model of any concept expression C0 and TBox T is obtained
from IC0 by hooking tree-shaped interpretations Id with root d to every d in IC0 . As
in DL-Lite∃R,horn the concept expressions on the left-hand side of CIs are basic concepts,
the interpretations Id only depend on the one-neighbourhood NIC0 (d) of d in IC0 . Thus,
the tree-shaped interpretations hooked to ρN in IN,T and to d in IC,T coincide and the
homomorphism h given in Lemma 14 provides a homomorphism h : TD → IN such that
h(ρD ) = ρN . By Lemma 12, T |= N ⊑ D. We have derived a contradiction. For DL-Lite∃R
one only requires a single member of NIC (d) since the left-hand side of CIs in DL-Lite∃R
consists of a single basic concept only.
❏
We close the introduction of description logics with some comments about the choice of our
languages. In the DL literature it is not uncommon to consider the weaker variant DL-LiteR
of DL-Lite∃R in which only basic concepts are admitted on the right-hand side of CIs, but
compound concepts are not. This is often without loss of generality since every DL-Lite∃R
TBox can be expressed in DL-LiteR by using additional role names; in this way, standard
DL-Lite∃R reasoning tasks such as subsumption and conjunctive query answering can be
reduced in polynomial time to the corresponding tasks for DL-LiteR . Such a reduction is
not possible in the framework of exact learning that we are concerned with in this paper.
In fact, in contrast to DL-Lite∃R TBoxes, TBoxes in DL-LiteR are trivially polynomial time
learnable using either membership queries only or equivalence queries only as there are only
polynomially many CIs and RIs over a given signature.
2.3 Exact Learning
We introduce the relevant notation for exact learning. A learning framework F is a triple
(X, L, µ), where X is a set of examples (also called domain or instance space), L is a set of
concepts,3 and µ is a mapping from L to 2X . We say that x ∈ X is a positive example for
l ∈ L if x ∈ µ(l) and a negative example for l if x 6∈ µ(l).
We give a formal definition of polynomial query and time learnability within a learning
framework. Let F = (X, L, µ) be a learning framework. We are interested in the exact
identification of a target concept representation l ∈ L by posing queries to oracles. Let
MEMF,l be the oracle that takes as input some x ∈ X and returns ‘yes’ if x ∈ µ(l) and
‘no’ otherwise. A membership query is a call to the oracle MEMF,l . Similarly, for every
l ∈ L, we denote by EQF,l the oracle that takes as input a hypothesis concept representation
h ∈ L and returns ‘yes’ if µ(h) = µ(l) and a counterexample x ∈ µ(h) ⊕ µ(l) otherwise,
where ⊕ denotes the symmetric set difference. There is no assumption regarding which
counterexample in µ(h) ⊕ µ(l) is chosen by the oracle. An equivalence query is a call to the
oracle EQF,l .
A learning algorithm for F is a deterministic algorithm that takes no input, is allowed
to make queries to MEMF,l and EQF,l (without knowing what the target l to be learned is),
3. The similarity of this name to ‘concept expression’ is accidental and should not be taken to mean that
these two notions are closely related. Both is standard terminology in the respective area.
13
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
and that eventually halts and outputs some h ∈ L with µ(h) = µ(l). We say that F is exact
learnable if there is a learning algorithm for F and that F is polynomial query learnable
if it is exact learnable by an algorithm A such that at every step the sum of the sizes of
the inputs to membership and equivalence queries made by A up to that step is bounded
by a polynomial p(|l|, |x|), where l is the target and x ∈ X is the largest counterexample
seen so far (Arias, 2004). Finally, F is polynomial time learnable if it is exact learnable
by an algorithm A such that at every step (we count each call to an oracle as one step of
computation) of computation the time used by A up to that step is bounded by a polynomial
p(|l|, |x|), where l ∈ L is the target and x ∈ X is the largest counterexample seen so far.
Clearly, a learning framework F that is polynomial time learnable is also polynomial query
learnable.
The aim of this paper is to study learnability of description logic TBoxes. In this context,
each DL L gives rise to a learning framework (X, L, µ), as follows: L is the set of all TBoxes
formulated in L, X is the set of all CIs and RIs formulated in L, and µ(T ) = {α ∈ X |
T |= α} for every T ∈ L. Observe that µ(T ) = µ(T ′ ) iff T ≡ T ′ , for all TBoxes T and
T ′ . We say that L TBoxes are polynomial query learnable if the learning framework defined
by L is polynomial query learnable, and likewise for polynomial time learnability. What
does not show up directly in this representation is our assumption that the signature of the
target TBox is known to the learner. Note that this is a standard assumption. For example,
when learning propositional Horn formulas, it is common to assume that the variables in
the target formula are known to the learner.
3. Learning DL-Lite∃R TBoxes
We prove that DL-Lite∃R TBoxes are polynomial query learnable. If inverse roles are disallowed in CIs and RIs of the target TBox then our algorithm runs in polynomial time and
thus shows that TBoxes in this restricted language are polynomial time learnable. Without
this restriction, polynomial time learnability remains open.
To simplify the presentation, we make two minor assumptions about the target TBox T .
We will show later how these assumptions can be overcomed. First, we assume that T does
not entail non-trivial role equivalences, that is, there do not exist distinct roles r and s such
that T |= r ≡ s. This allows us to avoid dealing with classes of equivalent roles, simplifying
notation. The second requirement is a bit more subtle. A concept inclusion is in reduced
form if it is between basic concepts or its left-hand side is a concept name. A TBox T is in
named form if all CIs in it are in reduced form and it contains a concept name Ar such that
Ar ≡ ∃r.⊤ ∈ T , for each role r. We assume that the target TBox is in named form and
that all CIs considered by the learner are in reduced form. In particular, counterexamples
returned by the oracle are immediately converted into this form.
Example 15 Although the TBox T from Example 11 does not entail role equivalences and
all its CIs are in reduced form, it is not in named form. To fix this, we introduce concept
names Asupervisor of , Aconduct research and Aadvisor of and extend T with the following equivalences:
Asupervisor
Aconduct
of
research
≡ ∃supervisor of.⊤
(12)
≡ ∃conduct research.⊤
(13)
14
Exact Learning of Lightweight Description Logic Ontologies
Algorithm 1 Naı̈ve learning algorithm for DL-Lite∃R
Input: A DL-Lite∃R TBox T in named form given to the oracle; ΣT given to the learner.
Output: TBox H, computed by the learner, such that T ≡ H.
1: Compute Hbasic = {r ⊑ s | T |= r ⊑ s} ∪ {B1 ⊑ B2 | T |= B1 ⊑ B2 , B1 , B2 basic}
2: Set Hadd = ∅
3: while Hbasic ∪ Hadd 6≡ T do
4:
Let A ⊑ C be the returned positive counterexample for T relative to Hbasic ∪ Hadd
5:
if there is A ⊑ C ′ ∈ Hadd then
6:
Replace A ⊑ C ′ by A ⊑ C ⊓ C ′ in Hadd
7:
else
8:
Add A ⊑ C to Hadd
9:
end if
10: end while
11: return H = Hbasic ∪ Hadd
Aadvisor of
≡ ∃advisor of.⊤.
(14)
Notice that Graduate acts as a name for ∃has degree.⊤ so no new definition is needed for
the role has degree. The TBox T ′ = T ∪ {(12), (13), (14)} is in named form.
To develop the learning algorithm it is instructive to start with a naı̈ve version that
does not always terminate but which can be refined to obtain the desired algorithm. This
version is presented as Algorithm 1. Given the signature ΣT of the target TBox T , the
learner starts with computing the set Hbasic by posing to the oracle the membership query
‘T |= r ⊑ s?’ for all r, s ∈ ΣT and ‘T |= B1 ⊑ B2 ?’ for all basic concept B1 , B2 over ΣT .
Observe that T |= Hbasic . Then it enters the main while loop. Note that the condition
‘Hbasic ∪ Hadd 6≡ T ?’ in Line 3 is implemented using an equivalence query to the oracle,
and that A ⊑ C in Line 4 refers to the counterexample returned by the oracle in the case
that equivalence does not hold. The counterexample must be positive since we maintain
the invariant T |= Hbasic ∪ Hadd throughout the run of the algorithm. If there is no CI of
the form A ⊑ C ′ in Hadd then A ⊑ C is added to Hadd , otherwise A ⊑ C ⊓ C ′ is (Lines 6
and 8). The algorithm terminates when Hbasic ∪ Hadd ≡ T , implying that the target TBox
has been learned.
Example 16 For the TBox T ′ from Example 15, Algorithm 1 first computes Hbasic which
coincides with T ′ except that Prof ⊑ ∃supervisor of.Student is not included since the concept
∃supervisor of.Student is not basic. In the main loop the only counterexamples to Hbasic ∪
Hadd ≡ T ′ are (up to logical equivalence modulo Hbasic ) the CIs
Prof ⊑ ∃supervisor of.Student,
Prof ⊑ ∃advisor of.Student.
If the oracle returns the first CI in the first iteration, the algorithm terminates immediately
having learned T ′ . Otherwise the oracle first returns the second CI and then returns the
first CI in the second iteration. The algorithm terminates with
Hadd = {Prof ⊑ ∃supervisor of.Student ⊓ ∃advisor of.Student}
15
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
which is equivalent to T ′ .
We now consider five examples on which this naı̈ve algorithm fails to terminate after polynomially many steps (or at all), each example motivating a different modification step that
is added to Algorithm 1 after Lines 4 and 5. The final, corrected algorithm is given as
Algorithm 2 below. Each modification step takes as input a counterexample A ⊑ C against
the equivalence Hbasic ∪ Hadd ≡ T and modifies it by posing membership queries to the
oracle to obtain a CI A′ ⊑ C ′ which is still a counterexample and has additional desired
properties. CIs satisfying all five additional properties will be called T -essential. The five
modification steps are of three different types:
1. two saturations steps: the underlying tree of TC is left unchanged but the labelling is
modified by adding concept names to node labels or replacing roles in edge labels;
2. two merging steps: nodes in the tree TC are merged, resulting in a tree with fewer
nodes;
3. a decomposition step: TC is replaced by a subtree or a subtree is removed from TC ,
and the concept name A on the left-hand side might be replaced.
The saturation and merging steps do not change the left-hand side A of the CI A ⊑ C and
result in a logically stronger CI A ⊑ C ′ in the sense that ∅ |= C ′ ⊑ C. In contrast, the
decomposition step can be regarded as a reset operation in which also the left-hand side
can change and which is logically not related to A ⊑ C. We start with an example which
motivates the first saturation step.
Example 17 Let
T = {A ⊑ ∃r.A} ∪ Tnf ,
where Tnf = {Ar ≡ ∃r.⊤} ensures that T is in named form. First, Algorithm 1 computes
Hbasic . Afterwards the oracle can provide for the n-th equivalence query in the while loop the
positive counterexample A ⊑ ∃r n+1 .⊤, for any n ≥ 1 (here we set inductively ∃r m+1 .⊤ =
∃r.∃r m .⊤ and ∃r 1 .⊤ = ∃r.⊤). Thus, the algorithm does not terminate.
Informally, the problem for the learner in Example 17 is that the concepts ∃r n .⊤ used in
the counterexamples A ⊑ ∃r n .⊤ get larger and larger, but still none of the counterexamples
implies A ⊑ ∃r.A. We address this problem by saturating TC with implied concept names.
For the following discussion, recall that we do not distinguish between the concept expression
C and its tree representation TC . For example, if we say that C ′ is obtained from C by
adding a concept name B to the label of node d in C, then this stands for: C ′ is the concept
expression corresponding to the tree obtained from TC by adding B to the label of d in TC .
Definition 18 (Concept saturation for T ) Let A ⊑ C be a CI with T |= A ⊑ C. A
CI A ⊑ C ′ is obtained from A ⊑ C by concept saturation for T if T |= A ⊑ C ′ and C ′
is obtained from C by adding a concept name to the label of some node of C. We say that
A ⊑ C is concept saturated for T if there is no A ⊑ C ′ with C 6= C ′ that can be obtained
from A ⊑ C by concept saturation.
16
Exact Learning of Lightweight Description Logic Ontologies
Observe that the learner can compute a concept saturated A ⊑ C ′ from a counterexample
A ⊑ C by posing polynomially many membership queries to the oracle: it simply asks
for any node d in TC and concept name E ∈ ΣT whether T |= A ⊑ C E,d , where C E,d is
obtained from C by adding E to the label of d. If the answer is positive, it replaces C by
C E,d and proceeds.
Example 19 (Example 17 continued) The CIs A ⊑ ∃r n .⊤ are not concept saturated
for T . For example, for n = 2, the concept saturation of A ⊑ ∃r 2 .⊤ is A ⊑ A ⊓ Ar ⊓
∃r.(A ⊓ Ar ⊓ ∃r.(Ar ⊓ A)). Now observe that if the CI A ⊑ C returned by the oracle to
the first equivalence query is transformed by the learner into a concept saturated CI (after
Line 4), then the TBox T = {A ⊑ ∃r.A} ∪ Tnf is learned in one step: the only possible
counterexamples returned by the oracle to the equivalence query Hbasic ≡ T are of the form
A ⊑ C1 ⊓ ∃r.C2 for some concepts C1 and C2 . Concept saturation results in a concept of
the form C1′ ⊓ ∃r.(A ⊓ C2′ ) and {A ⊑ C1′ ⊓ ∃r.(A ⊓ C2′ )} |= A ⊑ ∃r.A.
The following example motivates the second saturation step. Here and in the subsequent
examples we do not transform the TBoxes into named form as this does not effect the
argument and simplifies presentation.
Example 20 Consider for n ≥ 1 the TBoxes
Tn = {A ⊑ ∃e1 .∃e2 . . . . ∃en .⊤} ∪ {ei ⊑ ri , ei ⊑ si | 1 ≤ i ≤ n}.
For M ⊆ {1, . . . , n}, set CM = ∃t1 .∃t2 . . . . ∃tn .⊤, where ti = ri if i ∈ M and ti = si if
i∈
/ M . Then for the first 2n equivalence queries in the while loop the oracle can provide a
positive counterexample A ⊑ CM by always choosing a fresh set M ⊆ {1, . . . , n}.
Intuitively, the problem for the learner in Example 20 is that there are exponentially many
logically incomparable CIs that are entailed by Tn but do not entail A ⊑ ∃e1 .∃e2 . . . . ∃en .⊤.
A step towards resolving this problem is to replace the roles ri and si by the roles ei in the
counterexamples A ⊑ CM .
Definition 21 (Role saturation for T ) Let A ⊑ C be a CI with T |= A ⊑ C. A CI
A ⊑ C ′ is obtained from A ⊑ C by role saturation for T if T |= A ⊑ C ′ and C ′ is obtained
from C by replacing in some edge label a role r by a role s with T |= s ⊑ r. We say that
A ⊑ C is role saturated for T if there is no A ⊑ C ′ with C 6= C ′ that can be obtained from
A ⊑ C by role saturation.
Similarly to concept saturation, the learner can compute a role saturated A ⊑ C ′ from a
counterexample A ⊑ C by posing polynomially many membership queries. Observe that
in Example 20 the role saturation of any A ⊑ CM is A ⊑ ∃e1 .∃e2 . . . . ∃en .⊤. Thus, if the
counterexample A ⊑ CM returned by the first equivalence query is transformed into a role
saturated CI, then the algorithm terminates after one step. We now introduce and motivate
our two merging rules.
Example 22 Consider the TBox
T = {A ⊑ ∃r.⊤ ⊓ ∃s.⊤ ⊓ ∃e.B}.
17
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
e
r−
r
s−
s
B
h
e
..
.
s
r
r−
r
Figure 5: Tree representation of C{1,3} and homomorphism to ∃r.⊤ ⊓ ∃s.⊤ ⊓ ∃e.B.
−
−
and fix an n ≥ 1. For M ⊆ {1, . . . , n}, set CM = ∃t1 .∃t−
1 .∃t2 .∃t2 . . . . ∃tn .∃tn .∃e.⊤, where
/ M . Figure 5 (left) illustrates the concept expression C{1,3} ,
ti = r if i ∈ M and ti = s if i ∈
assuming n = 3. By Lemma 10, T |= A ⊑ CM since there is a homomorphism h from TCM
to the labelled tree that corresponds to ∃r.⊤ ⊓ ∃s.⊤ ⊓ ∃e.B, as shown in Figure 5. Thus, the
oracle can provide for the first 2n equivalence queries a positive counterexample A ⊑ CM
by always choosing a fresh set M ⊆ {1, . . . , n}.
The problem for the learner in Example 22 is similar to that in Example 20: there are
exponentially many logically incomparable CIs that are entailed by T but do not entail
A ⊑ ∃r.⊤ ⊓ ∃s.⊤ ⊓ ∃e.B. A step towards solving this problem is to merge the predecessor
and successor nodes of a node if the edge labels are inverse to each other and the resulting
CI is still implied by the TBox.
Definition 23 (Parent/Child Merging for T ) A concept C ′ is obtained from a concept C by parent/child merging if C ′ is obtained from C by choosing nodes d, d′ , d′′ such
that d is an r-successor of d′ , and d′′ is an r − -successor of d, for some role r, and then
removing d′′ , setting l(d′ ) = l(d′ ) ∪ l(d′′ ), and making every s-successor e of d′′ in C an
s-successor of d′ , for any role s.
Let A ⊑ C be a CI with T |= A ⊑ C. A CI A ⊑ C ′ is obtained from A ⊑ C by
parent/child merging if T |= A ⊑ C ′ and C ′ is obtained from C by parent/child merging.
We say that A ⊑ C is parent/child merged for T if there is no A ⊑ C ′ with C 6= C ′ that
can be obtained from A ⊑ C by parent/child merging.
Note that when C ′ is obtained from C by parent/child merging with d′ and d′′ as in Definition 23, then ∅ |= C ′ ⊑ C. To show this, one can use Lemma 10 and the natural
homomorphism h from TC to TC ′ , that is, the identity except that h(d′′ ) = d′ .
Similarly to the saturation operations, the learner can compute a parent/child merged
A ⊑ C ′ by posing polynomially many membership queries. In Example 22 the parent/child
merging of any A ⊑ CM with ∅ =
6 M 6= {1, . . . , n} is A ⊑ ∃r.⊤ ⊓ ∃s.⊤ ⊓ ∃e.⊤, as illustrated
in Figure 6: in the first step the nodes d0 and d2 are merged, two additional merging steps
give ∃r.⊤ ⊓ ∃s.⊤ ⊓ ∃e.⊤. The following example motivates the second merging operation.
Example 24 Define concept expressions Ci by induction as follows:
C1 = ∃r.⊤ ⊓ ∃s.⊤,
Ci+1 = C1 ⊓ ∃e.Ci
18
Exact Learning of Lightweight Description Logic Ontologies
e
r−
r
e
s−
s
r−
r
d2
rd−
1
r
···
s−
s
d1
r
r
s
e
d0 = d2
d0
Figure 6: Parent/Child Merging of C{1,3} for T
s
r
C3
s
e
e
r
r
{1,2}
e
s
r
s
r
{1,3}
C3
e
C3
r
s
e
e
r
{1,3}
Figure 7: Tree representation of the concept expressions C3 , C3
{1,2}
and C3
.
and let
Tn = {A ⊑ ∃e.Cn }.
/ M . Also, let
For M ⊆ {1, . . . , n}, set C1M = ∃r.⊤ if 1 ∈ M and C1M = ∃s.⊤ if 1 ∈
M = ∃r.⊤ ⊓ ∃e.C M if i + 1 ∈ M and C M = ∃s.⊤ ⊓ ∃e.C M if i + 1 ∈
Ci+1
/
M , 1 ≤ i < n.
i
i+1
i
M
Figure 7 illustrates concept expressions of the form Cn and Cn .
As an answer to the first 2n equivalence queries the oracle can compute a positive counterexample A ⊑ ∃e.CnM by always choosing a fresh set M ⊆ {1, . . . , n}.
To deal with this example we introduce a modification step that identifies siblings in C
rather than a parent and a child.
Definition 25 (Sibling Merging for T ) A concept C ′ is obtained from a concept C by
sibling merging if C ′ is obtained from C by choosing nodes d, d′ , d′′ such that d′ and d′′ are
r-successors of d, for some role r, and then removing d′′ , setting l(d′ ) = l(d′ ) ∪ l(d′′ ), and
making every s-successor e of d′′ in TC an s-successor of d′ , for any role s.
Let A ⊑ C be a CI with T |= A ⊑ C. A CI A ⊑ C ′ is obtained from A ⊑ C by sibling
merging if T |= A ⊑ C ′ and C ′ is obtained from C by sibling merging. We say that A ⊑ C
is sibling merged for T if there is no A ⊑ C ′ with C 6= C ′ that can be obtained from A ⊑ C
by sibling merging.
It can be verified that when C ′ is obtained from C by sibling merging, then ∅ |= C ′ ⊑ C.
In Example 24 the counterexamples A ⊑ CnM are actually sibling merged for Tn . Thus,
producing a sibling merged A ⊑ C ′ directly from the counterexamples returned by the
oracle does not overcome the problem illustrated by the example. Instead, we apply sibling
19
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
merging after Line 5 of the algorithm: instead of adding A ⊑ C ⊓ C ′ to Hadd , the learner
computes a sibling merged A ⊑ D from this CI and adds it to Hadd . For Example 24, this is
illustrated in Figure 8. Clearly, after at most n + 1 counterexamples, the learner has added
A ⊑ ∃e.Cn , as required.
{1,3}
C5
r
r
s
e
e
{1,2}
C5
r
s
r
e
r
s
r
s
e
e
e
r
{1,3}
Figure 8: Sibling Merging C3
{1,2}
and C3
for T .
Finally, we need a decomposition rule. The following variant of Example 17 illustrates
that the four modification steps introduced so far do not yet lead to a polynomial learning
algorithm even if they are applied both after Line 4 and after Line 5 in Algorithm 1.
Example 26 Let
T = {A ⊑ B, B ⊑ ∃r.B}.
The oracle can provide for the n-th equivalence query the positive counterexample A ⊑ CB,n ,
where CB,n = A ⊓ DB,n and, inductively, DB,0 = B and DB,n+1 = B ⊓ ∃r.DB,n , for any
n ≥ 0. The algorithm does not terminate even with the four modification steps introduced
above applied after Lines 4 and 5: the CIs A ⊑ CB,n are concept and role saturated and
they are parent/child and sibling merged.
The problem illustrated in Example 26 is that so far the learning algorithm attempts to
learn T without ever considering to add to Hadd a CI whose left-hand side is B (rather than
A). To deal with this problem we introduce a ‘reset step’ that, in contrast to the previous
modification steps, can lead to a different left-hand side and also to a CI that does not
imply the original CI given T , as in all previous modification steps.
Definition 27 (Decomposed CI for T ) Let A ⊑ C be a CI with T |= A ⊑ C. We say
that A ⊑ C is decomposed for T if for every non-root node d in C, every concept name
A′ ∈ l(d), and every r-successor d′ of d in C, we have T 6|= A′ ⊑ ∃r.C ′ where C ′ corresponds
to the subtree of C rooted at d′ .
In contrast to the previous four modification steps, the membership queries used by the
learner to obtain a decomposed CI do not only depend on T but also on the hypothesis
Hadd ∪ Hbasic computed up to that point: starting from CI A ⊑ C, the learner takes a
non-root node d in C, a concept name A′ ∈ l(d), and an r-successor d′ of d in C, and then
checks using a membership query whether T |= A′ ⊑ ∃r.C ′ , where C ′ is the subtree rooted
at d′ in C. If the check succeeds, A ⊑ C is replaced by
(a) A′ ⊑ ∃r.C ′ if Hbasic ∪ Hadd 6|= A′ ⊑ ∃r.C ′ ; and otherwise by
−
′
(b) A ⊑ C|−
d′ ↓ , where C|d′ ↓ is obtained from C by removing the subtree rooted in d
from C.
20
Exact Learning of Lightweight Description Logic Ontologies
B
B
d′
r
r
B
d′
d
d
B
r
d
r
A, B
A, B
′
Figure 9: Illustration of decomposition of CIs. Here C|−
d′ ↓ = A ⊓ B ⊓ ∃r.B and C = B.
Algorithm 2 The learning algorithm for DL-Lite∃R
Input: A DL-Lite∃R TBox T in named form given to the oracle; ΣT given to the learner.
Output: TBox H, computed by the learner, such that T ≡ H.
1: Compute Hbasic = {r ⊑ s | T |= r ⊑ s} ∪ {B1 ⊑ B2 | T |= B1 ⊑ B2 , B1 , B2 basic}
2: Set Hadd = ∅
3: while Hbasic ∪ Hadd 6≡ T do
4:
Let A ⊑ C be the returned positive counterexample for T relative to Hbasic ∪ Hadd
5:
Find a T -essential CI A′ ⊑ C ′ such that Hbasic ∪ Hadd 6|= A′ ⊑ C ′
6:
if there is A′ ⊑ C ′′ ∈ Hadd then
7:
Find T -essential CI A′ ⊑ C ∗ such that ∅ |= C ∗ ⊑ C ′′ ⊓ C ′
8:
Replace A′ ⊑ C ′′ by A′ ⊑ C ∗ in Hadd
9:
else
10:
Add A′ ⊑ C ′ to Hadd
11:
end if
12: end while
13: return H = Hbasic ∪ Hadd
−
′
′
′
′
Note that {A ⊑ C|−
d′ ↓ , A ⊑ ∃r.C } |= A ⊑ C. Thus, one of the CIs A ⊑ C|d′ ↓ and A ⊑ ∃r.C
is not entailed by Hbasic ∪ Hadd , and this is the CI that replaces the original CI.
In Example 26, assume that the oracle returns A ⊑ C with C = A ⊓ B ⊓ ∃r.(B ⊓ ∃r.B)
as the first counterexample. The tree TC corresponding to C is shown on the left-hand side
of Figure 9. This CI is not decomposed for T : the label of node d contains B, the concept
C ′ rooted in d′ in C is B and T |= B ⊑ ∃r.B. Since H ∪ Hadd 6|= B ⊑ ∃r.B, Case (a) applies
and A ⊑ B is replaced by B ⊑ ∃r.B.
This finishes the description of the modification steps. It turns out that they cure all
problems with the initial version of the algorithm and enable polynomial query learnability.
Definition 28 A DL-Lite∃R CI is T -essential if it is concept saturated, role saturated, parent/child merged, sibling merged, and decomposed for T .
After Lines 4 and 5 of Algorithm 1, we need to make the CI currently considered T -essential,
by exhaustively applying the modification steps described above in all possible orders. The
resulting refined version of the learning algorithm is shown as Algorithm 2. We next analyse
the properties of this algorithm.
21
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
Polynomial Query Bound on the Algorithm
If Algorithm 2 terminates, then it obviously has found a TBox Hbasic ∪ Hadd that is logically
equivalent to T . It thus remains to show that the algorithm terminates after polynomially
many polynomial size queries. Observe that Hadd contains at most one CI A ⊑ C for each
concept name A. At each step in the while loop, either some A′ ⊑ C ′ is added to Hadd such
that no CI with A′ on the left-hand side existed in Hadd before (Line 10) or an existing CI
A′ ⊑ C ′′ in Hadd is replaced by a fresh CI A′ ⊑ C ∗ with ∅ |= C ∗ ⊑ C ′′ .
We start with showing that Lines 5 and 7 can be implemented with polynomially many
membership queries. The next lemma addresses Line 5.
Lemma 29 Given a positive counterexample A ⊑ C for T relative to Hbasic ∪Hadd , one can
construct a T -essential counterexample A′ ⊑ C ′ using only polynomially many polynomial
size membership queries in |C| + |T |.
Proof Let A ⊑ C be a positive counterexample for T relative to Hbasic ∪ Hadd and assume
the five modification steps introduced above are applied exhaustively by posing membership
queries to the oracle. Observe that the number of applications of modifications steps is
bounded polynomially in |C| × |T |. To show this, let nC be the number of nodes in TC .
Then nC ′′ = nC ′ if A′′ ⊑ C ′′ is obtained from A′ ⊑ C ′ by a concept or role saturation step
and nC ′′ < nC ′ if A′′ ⊑ C ′′ is obtained from A′ ⊑ C ′ by a merging or decomposition step.
Thus, the number of applications of merging and decomposition steps is bounded by nC
and the number of applications of concept and role saturated steps is bounded by |ΣT | × nC
and |ΣT | × n2C , respectively. Thus, after at most nC + |ΣT | × nC + |ΣT | × n2C steps no
modification step is applicable and the final CI is T -essential. We verify that it is also a
positive counterexample for T relative to Hbasic ∪ Hadd . It suffices to show that the CI
resulting from each single modification step is entailed by T , but not by Hbasic ∪ Hadd . The
former has been shown when we introduced the modification steps. Regarding the latter, in
the first four modification steps we have Hbasic |= C ′ ⊑ C if A ⊑ C is replaced by A ⊑ C ′ .
Hence Hbasic ∪ Hadd 6|= A ⊑ C ′ . For the decomposition step, we have already argued, after
Definition 27, that the added CI is not entailed by Hbasic ∪ Hadd .
❏
The following lemma addresses Line 7.
Lemma 30 Assume that A ⊑ C1 and A ⊑ C2 are T -essential. Then one can construct a
T -essential A ⊑ C such that ∅ |= C ⊑ C1 ⊓ C2 using polynomially many polynomial size
membership queries in |C1 | + |C2 |.
Proof We start with A ⊑ C1 ⊓ C2 . Using the fact that C1 and C2 are both T -essential,
one can show that this CI is (i) concept saturated for T , (ii) role saturated for T , (iii)
parent/child merged for T , and (iv) decomposed for T . Assume, for example, that A ⊑
C1 ⊓ C2 is not concept saturated. Then one can add a new concept name A′ to the label
l(d) for some node d in TC1 ⊓C2 and T |= A ⊑ C ′ for the resulting concept C ′ . Clearly d is
a node in TC1 or in TC2 . Assume without loss of generality that d is in TC1 and let C1′ be
the concept obtained from C1 by adding A′ to l(d). Then T |= A ⊑ C1′ since T |= A ⊑ C ′
which contradicts the assumption that A ⊑ C1 is concept saturated. The remaining three
modification steps are considered similarly. We now exhaustively apply the modification
22
Exact Learning of Lightweight Description Logic Ontologies
step ‘Sibling merging for T ’ and use the resulting CI as the desired A ⊑ C. Similarly to the
argument above one can show that a CI with properties (i)–(iv) still has those properties
after applying sibling merging. Thus, A ⊑ C is T -essential. We have argued in the proof of
Lemma 29 already that the number of applications of a sibling merging step to a CI of the
form A ⊑ D is bounded by the number of nodes in TD . Thus, the number of modification
steps is bounded polynomially in |C1 | + |C2 |.
❏
To analyse the algorithm further we first prove a polynomial upper bound on the size
of T -essential CIs. To this end, we require the notion of an isomorphic embedding and an
auxiliary lemma. A homomorphism h : TC → I is an isomorphic embedding for T if it
is injective, A ∈ l(d) if h(d) ∈ AI for all concept names A, and for r = l(d, d′ ) it holds
that T |= r ⊑ s for all (h(d), h(d′ )) ∈ sI . The following lemma shows that for T -essential
CIs A ⊑ C and any D that interpolates between A and C (meaning that T |= A ⊑ D
and T |= D ⊑ C) the homomorphism h from TC to ID,T that witnesses ρD ∈ C ID,T (see
Lemma 12) is an isomorphic embedding.
Lemma 31 Assume the A ⊑ C is T -essential, T |= A ⊑ D and T |= D ⊑ C. Then any
homomorphism h : TC → ID,T that maps ρC to ρD is an isomorphic embedding.
Proof Assume first that h is not injective. Then there is a parent/child or sibling merging
C ′ of C and a homomorphism f : TC ′ → ID,T such that h = f ◦ g for the natural homomorphism g : TC → TC ′ . By Lemma 12, T |= D ⊑ C ′ . Thus, T |= A ⊑ C ′ and we have
derived a contradiction to the assumption that C is parent/child and sibling merged.
Now let T ′ be the following labelled tree: the nodes in T ′ are the same as in T , A ∈ l′ (d)
iff h(d) ∈ AID,T , and for any two nodes d, d′ with d′ a successor of d: l′ (d, d′ ) = r for the
unique role r with T |= r ⊑ s for all s with (d, d′ ) ∈ sID,T . Let C ′ be the concept expression
that corresponds to T ′ . Then ρD ∈ C ′ ID,T and so, by Lemma 12, T |= D ⊑ C ′ . Thus
T |= A ⊑ C ′ . But then A ⊑ C ′ can be obtained from A ⊑ C by concept and role saturation
steps. As A ⊑ C is concept and role saturated already, C = C ′ and so h is an isomorphic
embedding.
❏
We are now able to prove that T -essential CIs are of polynomial size in T . Let nC denote
the number of nodes in the tree representation TC of C and let
AT = {A} ∪ {B | A ⊑ B ∈ Hbasic } ∪ {D | T |= A ⊑ B, B ⊑ D ∈ T }.
Lemma 32 If A ⊑ C is T -essential, then nC ≤
X
D∈AT
nD .
d
Proof Assume A ⊑ C is T -essential. Let D0 := D∈AT D and ID0 ,T be the canonical
model of D0 and T . By Lemma 12, there is a homomorphism h : TC → ID0 ,T mapping ρC
to ρD0 . By Lemma 31, h is an isomorphic embedding. Using that A ⊑ C is decomposed for
T we now show that h maps TC into the restriction of ID0 to ∆ID0 from which the lemma
follows since h is injective.
For a proof by contradiction, assume that there exists d′ in TC with h(d′ ) 6∈ ∆ID0 . As
h(ρC ) ∈ ∆ID0 , we may assume that all d′ on the path from ρC to d′ are mapped to ∆ID0 .
23
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
In particular, the parent d of d′ in TC is mapped into ∆ID0 . Let l(d, d′ ) = r. Observe that
the whole subtree rooted in d′ must be mapped into ∆ID0 ,T \ ∆ID0 since otherwise h would
not be injective.
Let C ′ = ∃r.C ′′ , where C ′′ corresponds to the subtree rooted in d′ in C. By Lemma 14
there exists a basic concept B such that h(d) ∈ B ID0 ,T and T |= B ⊑ C ′ . As T is in
named form there exists a concept name E with T |= E ≡ B. Thus h(d) ∈ E ID0 ,T and
T |= E ⊑ C ′ . As h is an isomorphic embedding, E ∈ l(d). We make a case distinction:
• h(d) 6= ρD0 . Then A ⊑ C is not decomposed for T since C contains an edge (d, d′ )
such that E is in the node label of d, l(d, d′ ) = r, and T |= E ⊑ ∃r.C ′′ . We have
derived a contradiction.
• h(d) = ρD0 . As h(d′ ) ∈ ∆ID0 ,T \ ∆ID0 , by construction of ID0 ,T , there exists a CI
ID
B0 ⊑ D ∈ T with B0 a basic concept such that ρD0 ∈ B0 0 and h(d′ ) is in the copy
of the tree-shaped interpretation ID which was attached to ρD0 in the construction of
ID0 ,T . But since T |= A ⊑ B0 we have D ∈ AT and so ρD0 ∈ D ID0 . But then, by the
construction of ID0 ,T , no fresh ID was attached to ρD0 because D is already satisfied
in ID0 and we have derived a contradiction.
❏
,T
We are now in the position to prove that the learning algorithm terminates after posing a
polynomial number of queries.
Lemma 33 For every concept name A, the number of replacements of a CI A ⊑ C in Hadd
by a CI of the form A ⊑ C ′ is bounded polynomially in |T |.
Proof All CIs A ⊑ C ever added to Hadd are T -essential. We show that when A ⊑ C is
replaced with A ⊑ C ′ , then the number of nodes in the tree representation of C ′ is strictly
larger than the number of nodes in the
P tree representation of C. By Lemma 32, the number
of replacements is thus bounded by D∈AT nD , which is polynomial in |T |.
Note that when A ⊑ C is replaced with A ⊑ C ′ , then ∅ |= C ′ ⊑ C and ∅ 6|= C ⊑ C ′ . It
thus suffices to establish the following.
Claim. If A ⊑ C and A ⊑ C ′ are T -essential, ∅ |= C ′ ⊑ C and ∅ 6|= C ⊑ C ′ , then TC is
obtained from TC ′ by removing at least one subtree.
We prove the claim. Since ∅ |= C ′ ⊑ C, by Lemma 12 there is a homomorphism h from TC
to the canonical model IC ′ that maps ρC to ρC ′ . Then h is also a homomorphism into the
canonical model IC ′ ,T and thus, by Lemma 31, h is an isomorphic embedding into IC ′ ,T .
Then, trivially, h is also an isomorphic embedding into IC ′ which means that TC is obtained
from TC ′ by removing subtrees. Since ∅ 6|= C ⊑ C ′ , at least one subtree must in fact have
been removed.
❏
We have obtained the following main result of this section.
Theorem 34 DL-Lite∃R TBoxes are polynomial query learnable using membership and equivalence queries. Moreover, DL-Lite∃R TBoxes without inverse roles can be learned in polynomial time using membership and equivalence queries.
24
Exact Learning of Lightweight Description Logic Ontologies
Proof Recall that our algorithm requires the target TBox to be in named form. We
first show Theorem 34 under that assumption and then argue that the assumption can be
dropped.
In each iteration of Algorithm 2, either a CI is added to Hadd or a CI is replaced in
Hadd . Since the number of times the former happens is bounded by |ΣT | and (by Lemma 33)
the number of times the latter happens is polynomial in |T |, the number of iterations of
Algorithm 2 is polynomial in |T |. For polynomial query learnability of DL-Lite∃R TBoxes,
it remains to show that in each iteration Algorithm 2 makes only polynomially many polynomial size queries in |T | and the size of the largest counterexample seen so far. We start
with equivalence queries, made only in Line 3. We have already argued that the number
of iterations is polynomial in |T | and thus so is the number of equivalence queries made.
Regarding their size, we observe that there are at most |ΣT |2 CIs in Hbasic and at most
|ΣT | CIs in Hadd , that the size of CIs in Hbasic is constant and by Lemma 32 the size of CIs
in Hadd is polynomial in |T |. Membership queries are made only in Lines 5 and 7 for which
it suffices to invoke Lemmas 29 and 30.
Now for the “moreover” part of Theorem 34. Observe that since each (membership or
equivalence) query counts as one step of computation, the only potentially costly step of Algorithm 2 is the implementation of the decomposition step in Line 5, which relies on making
subsumption checks of the form Hbasic ∪ Hadd |= A ⊑ C. As discussed in Section 2, deciding
subsumption in DL-Lite∃R is NP-complete while in EL with role inclusions subsumption is
in PTime. As DL-Lite∃R without inverse roles is a fragment of EL with role inclusions, we
obtain polynomial time learnability for TBoxes in this case.
To drop the requirement that the target TBox is in named form, we show that any
polynomial (query or time) learning algorithm for TBoxes in named form can be transformed
into the same kind of algorithm for unrestricted target TBoxes. In fact, the learner can
use at most O(|ΣT |2 ) membership queries “Does T entail r ⊑ s?” to compute for every
role r the class [r]T of roles s with T |= s ≡ r and choose a representative rT for this
class. Then whenever some s ∈ [r]T is used in any counterexample returned by the oracle,
it gets replaced with rT . Likewise, whenever T does not have a name for some ∃r.⊤, the
algorithm still uses the concept name Ar in its internal representations (although they are
no longer included in the signature ΣT of the target TBox) and replaces ∃r.⊤ with Ar in the
counterexamples returned by the oracle. It also replaces each Ar with ∃r.⊤ in membership
queries to the oracle and in the hypothesis used for posing equivalence queries.
❏
4. Learning DL-Lite∃R,horn TBoxes
We study exact learnability of TBoxes in DL-Lite∃R,horn , the extension of DL-Lite∃R that
admits conjunctions of basic concepts on the left-hand side of CIs. This language is a
generalisation of both DL-Lite∃R and propositional Horn logic. In fact, the algorithm we
present combines the classical algorithms for propositional Horn logic (Angluin et al., 1992;
Frazier and Pitt, 1993) with the algorithm for DL-Lite∃R presented in Section 3. The resulting algorithm is quite subtle and indeed this is the reason why we treated the DL-Lite∃R
case separately in Section 3.
To simplify the presentation, we make the same assumptions as in Section 3 about
the target TBox T with signature ΣT . In particular, we assume that T is in named form,
25
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
Algorithm 3 The learning algorithm for DL-Lite∃R,horn TBoxes
Input: A DL-Lite∃R,horn TBox T in named form given to the oracle; ΣT given to the learner.
Output: TBox H, computed by the learner, such that T ≡ H.
1: Compute Hbasic = {r ⊑ s | T |= r ⊑ s} ∪ {B1 ⊑ B2 | T |= B1 ⊑ B2 , B1 , B2 basic}
2: Set Hadd to be the empty list and H = Hbasic ∪ Hadd
3: while H 6≡ T do
4:
Let γ be the returned positive counterexample for T and H
5:
Find a T -essential γ ′ with H 6|= γ ′ and |{∃r.F | ∃r.F ∈ rhs(γ ′ )}| ≤ 1
6:
Left saturate γ ′ for H
7:
if there is A ∈ NC such that T |= lhs(γ ′ ) ⊑ A and H 6|= lhs(γ ′ ) ⊑ A then
8:
H :=CN-Refine(H, lhs(γ ′ ) ⊑ A)
9:
else
10:
H :=∃-Refine(H, γ ′ )
11:
end if
12:
Set H = Hbasic ∪ Hadd
13: end while
14: return H
suitably generalised to DL-Lite∃R,horn : there are no distinct roles r and s such that T |= r ≡ s,
for each role r the TBox T contains an equivalence Ar ≡ ∃r.⊤ and all CIs of T are either CIs
between basic concepts or contain no concept expressions of the form ∃r.⊤ on the left-hand
side, for any role r. Denote by lhs(α) the set of concept names that occur as conjuncts
on the left-hand side of a CI α and denote by rhs(α) the set of concept expressions that
occur as top-level conjuncts on the right-hand side of α (that is, they are not nested inside
restrictions). We often do not distinguish between the set lhs(α) and the conjunction over
all its concept expressions, and similarly for rhs(α). For example, if α1 = C1 ⊑ D1 and
α2 = C2 ⊑ D2 then lhs(α1 ) ⊑ rhs(α2 ) stands for C1 ⊑ D2 . Also, if lhs(α1 ) = {A1 , A2 , A3 }
and lhs(α2 ) = {A2 , A3 , A4 } then lhs(α1 ) ∩ lhs(α2 ) ⊑ D stands for A2 ⊓ A3 ⊑ D.
The algorithm for learning DL-Lite∃R,horn TBoxes is shown as Algorithm 3. Like Algorithm 2, Algorithm 3 first determines the set Hbasic that contains all CIs B1 ⊑ B2 with
B1 , B2 basic concepts such that T |= B1 ⊑ B2 and all RIs r ⊑ s such that T |= r ⊑ s.
The hypothesis H is the union of Hbasic and Hadd . In contrast to Algorithm 2, Hadd is an
ordered list of CIs rather than a set. We write αi to denote the CI α at position i in the
list Hadd . In the learning algorithm, working with an ordered list of CIs allows the learner
to pick the first αi in Hadd with a certain property and merge it with a new CI, a technique
we adopt from (Angluin et al., 1992; Frazier and Pitt, 1993). As in Algorithm 2, T |= H is
a loop invariant, thus, γ is necessarily positive. The algorithm terminates when H ≡ T .
Algorithm 3 uses membership queries to compute a T -essential counterexample γ such
that rhs(γ) contains at most one concept expression of the form ∃r.F (Line 5) and which
is ‘Left saturated for H’ (Line 6); here, a CI is left saturated for H if its left-hand side
contains all subsuming concept names w.r.t. H (Definition 35 below) and T -essential if it
satisfies the conditions for T -essential CIs from Section 3, appropriately modified for CIs
with conjunctions of concept names on the left-hand side (Definition 37 below). Then, the
algorithm checks whether there is a concept name A such that lhs(γ) ⊑ A is a positive
26
Exact Learning of Lightweight Description Logic Ontologies
Algorithm 4 Function CN-Refine(H, γ)
1: if there is A ∈ NC and αi ∈ Hadd such that T |= lhs(αi ) ∩ lhs(γ) ⊑ A,
2:
and H 6|= lhs(αi ) ∩ lhs(γ) ⊑ A then
3:
Concept saturate γ ′ = lhs(αi ) ∩ lhs(γ) ⊑ A for T
4:
Replace the first such αi in Hadd by γ ′
5: else
6:
Concept saturate γ for T
7:
Append γ to the list Hadd
8: end if
9: return H
Algorithm 5 Function ∃-Refine(H, γ)
1: if there is C ∈ rhs(γ) of the form ∃r.D and αi ∈ Hadd such that T |= lhs(αi )∩lhs(γ) ⊑ C
2:
and H 6|= lhs(αi ) ∩ lhs(γ) ⊑ C then
3:
if T |= lhs(αi ) ∩ lhs(γ) ⊑ C ⊓ rhs(αi ) then
4:
Find a T -essential lhs(αi ) ∩ lhs(γ) ⊑ D ∗ with ∅ |= D ∗ ⊑ C ⊓ rhs(αi )
5:
Replace the first such αi in Hadd by lhs(αi ) ∩ lhs(γ) ⊑ D ∗
6:
else
7:
Concept saturate γ ′ = lhs(αi ) ∩ lhs(γ) ⊑ C for T
8:
Replace the first such αi in Hadd by γ ′
9:
end if
10: else
11:
Append γ to the list Hadd
12: end if
13: return H
counterexample. If so, then it calls Function CN-Refine (Algorithm 4) and updates the
hypothesis either by refining some αi in Hadd or by appending a new CI to Hadd . The
number of replacements of any given αi in Hadd in CN-Refine is bounded by |ΣT | since
whenever αi is replaced in CN-Refine(H, γ), then lhs(αi ) ∩ lhs(γ) ( lhs(αi ).4
If there is no concept name A such that lhs(γ) ⊑ A is a positive counterexample then
Algorithm 3 calls Function ∃-Refine (Algorithm 5). In this case one considers the existential
restrictions that occur on top-level on the right-hand side of γ. Note that ∃-Refine can be
viewed as a variation of the body of the while loop in Algorithm 2 in which one considers
sets of concept names on the left-hand side of CIs rather than a single concept name. Recall
that in Algorithm 2, the new CI γ and a CI α in Hadd are merged if they have the same
concept name on the left-hand side. In contrast, now they are merged if the intersection
of their left-sides is still subsumed by some existential restriction C from rhs(γ) (Lines 1
and 2). There are two cases: if the intersection is also subsumed by rhs(α) (checked in
Line 3), then in the next line a T -essential counterexample is computed and the first such
αi is replaced by the new CI. Otherwise it follows that lhs(αi ) ∩ lhs(γ) ( lhs(αi ) and the
4. This is a consequence of the fact that Hadd only contains concept saturated CIs (defined essentially as
in the previous section, see Definition 37 below): lhs(αi ) ∩ lhs(γ) = lhs(αi ) and T |= lhs(αi ) ∩ lhs(γ) ⊑ A
implies A ∈ rhs(αi ) by concept saturatedness, thus contradicting H 6|= lhs(αi ) ∩ lhs(γ) ⊑ A.
27
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
first such αi is replaced by the CI computed in Line 7. Note that the latter can happen at
most |ΣT | times for each CI in Hadd (and the former can happen at most |T | times for each
CI in Hadd , see Lemma 43 below). If no CI can be refined with γ then ∃-Refine appends γ
to Hadd .
We now define the step ‘left-saturate γ for H’ used in Line 6 of Algorithm 3. Observe
that this step is meaningless for DL-Lite∃R .
Definition 35 (Left saturation for H) A CI γ ′ is obtained from a CI γ by left saturation for H if rhs(γ ′ ) = rhs(γ) and lhs(γ ′ ) = {A ∈ ΣT | H |= lhs(γ) ⊑ A}. A CI γ is left
saturated for H if it coincides with its left saturation for H.
One can clearly left saturate any CI γ for H by checking whether H |= lhs(γ) |= A for
every A ∈ ΣT . The following example shows that Line 6 is necessary for Algorithm 3
to be polynomial. A similar step is also necessary in Frazier et al.’s algorithm learning
propositional Horn logic from entailments (Frazier and Pitt, 1993).
Example 36 Assume Line 6 of Algorithm 3 is omitted. Let for n ≥ 2,
Tn = {E1 ⊓ · · · ⊓ En ⊑ A} ∪ {Ai ⊑ Ei , Bi ⊑ Ei | 1 ≤ i ≤ n}.
d
/ M . Then
For M ⊆ {1, . . . , n}, set CM = i≤n Ci , where Ci = Ai if i ∈ M and Ci = Bi if i ∈
n
the oracle can provide for the first 2 equivalence queries in the while loop of Algorithm 3
a positive counterexample CM ⊑ A by always choosing a fresh set M ⊆ {1, . . . , n}.
For the refinements on the right-hand side, we extend to DL-Lite∃R,horn the notion of T essential CIs introduced in the previous section:
1. (Concept saturation for T ) A CI γ ′ is obtained from a CI γ by concept saturation
for T if lhs(γ) = lhs(γ ′ ), rhs(γ ′ ) is obtained from rhs(γ) by adding a concept name
to the label of some node of rhs(γ), and T |= γ ′ . A CI γ is concept saturated for T
if T |= γ and there is no γ ′ with γ 6= γ ′ that can be obtained from γ by concept
saturation for T .
2. (Role saturation for T ) A CI γ ′ is obtained from γ by role saturation for T if lhs(γ) =
lhs(γ ′ ), rhs(γ ′ ) is obtained from rhs(γ) by replacing in some edge label a role r by a
role s with T |= s ⊑ r, and T |= γ ′ . A CI γ is role saturated for T if T |= γ and there
is no γ ′ with γ 6= γ ′ that can be obtained from γ by role saturation for T .
3. (Parent/child merged for T ) A CI γ ′ is obtained from a CI γ by parent/child merging
for T if lhs(γ) = lhs(γ ′ ), rhs(γ ′ ) is obtained from rhs(γ ′ ) by parent/child merging (as
in Definition 23), and T |= γ ′ . A CI γ is parent/child merged for T if T |= γ and
there is no γ ′ with γ 6= γ ′ that can be obtained from γ by parent/child merging for T .
4. (Sibling merged for T ) A CI γ ′ is obtained from a CI γ by sibling merging for T
if lhs(γ) = lhs(γ ′ ), rhs(γ ′ ) is obtained from rhs(γ ′ ) by sibling merging (as in Definition 25), and T |= γ ′ . A CI γ is sibling merged for T if T |= γ and there is no γ ′ with
γ 6= γ ′ that can be obtained from γ by sibling merging for T .
28
Exact Learning of Lightweight Description Logic Ontologies
5. (Decomposed CI for T ) A CI γ is decomposed for T if T |= γ and for every nonroot node d in rhs(γ), every role r, and every r-successor d′ of d in rhs(γ) we have
T 6|= l(d) ⊑ ∃r.C ′ , where C ′ corresponds to the subtree of rhs(γ) rooted at d′ .
Definition 37 A DL-Lite∃R,horn CI is T -essential if it is concept saturated, role saturated,
parent/child merged, sibling merged, and decomposed for T .
The saturation, merging and decomposition steps defined above are straightforward generalisations of Definitions 18 to 27 to CIs with conjunctions on the left-hand side. One can
easily generalise the arguments from DL-Lite∃R to show that for any CI γ with H 6|= γ one
can compute a T -essential γ ′ with H 6|= γ ′ using polynomially many membership queries
and entailment checks relative to H. For the analysis of the learning algorithm it is crucial
that all CIs in the ordered list Hadd are T -essential at all times, which we prove next.
Lemma 38 At any point in the execution of Algorithm 3, all CIs in Hadd are T -essential.
Proof If a CI γ is of the form A1 ⊓ · · · ⊓ An ⊑ A with A a concept name, then the set
rhs(γ ′ ) of the concept saturation γ ′ of γ for T contains concept names only. Thus, γ ′ is
T -essential. It follows that the CIs added to Hadd in Lines 4 and 7 of CN-Refine are T essential. Also, it is easy to see that if γ is T -essential and C ∈ rhs(γ) then the concept
saturation of lhs(γ) ⊑ C for T is T -essential as well. Thus, the CI γ ′ in Line 8 of ∃-Refine
is T -essential.
❏
Polynomial Query Bound on the Algorithm
As in the previous section it is immediate that upon termination the algorithm has found a
TBox H = Hbasic ∪ Hadd that is logically equivalent to the target TBox T . It thus remains
to show that it issues only polynomially many queries of polynomial size. We first discuss
how Lines 5 and 6 of Algorithm 3 and Line 4 of ∃-Refine can be implemented. The next
lemma addresses Lines 5 and 6 of Algorithm 3.
Lemma 39 Given a positive counterexample γ for T relative to H, one can construct with
polynomially many polynomial size membership queries in |γ| and |T |, a counterexample γ ′
that is left saturated for H, T -essential and such that |{∃r.F | ∃r.F ∈ rhs(γ ′ )}| ≤ 1.
The proof of Lemma 39 is a straightforward extension of the proof of Lemma 29 and
uses the observation that a left-saturated γ ′ for H can be computed from γ by adding all
concept names A ∈ ΣT with H |= lhs(γ) ⊑ A to lhs(γ). This lemma also requires that
|{∃r.F | ∃r.F ∈ rhs(γ ′ )}| ≤ 1. If there is A ∈ rhs(γ ′ ) such that H 6|= lhs(γ ′ ) ⊑ A then we
can simply drop all conjuncts of the form ∃r.F from rhs(γ ′ ). Otherwise, we can satisfy the
condition by simply choosing a conjunct ∃r.F ∈ rhs(γ ′ ) such that H 6|= lhs(γ ′ ) ⊑ ∃r.F and
then apply ‘Concept saturation for T ’ to lhs(γ ′ ) ⊑ ∃r.F . The resulting γ ′ is left saturated
for H, T -essential and has at most one conjunct of the form ∃r.F in rhs(γ ′ ).
The following lemma addresses Line 4 of ∃-Refine.
Lemma 40 Assume that α and γ are T -essential and there is C ∈ rhs(γ) such that T |=
lhs(α) ∩ lhs(γ) ⊑ rhs(α) ⊓ C. Then one can construct, with polynomially many polynomial
29
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
size membership queries in |rhs(α)| and |C|, a T -essential lhs(α) ∩ lhs(γ) ⊑ D ∗ such that
∅ |= D ∗ ⊑ rhs(α) ⊓ C.
Proof Assume T |= lhs(α) ∩ lhs(γ) ⊑ rhs(α) ⊓ C. Then, similar to Lemma 30, one can
show that the only property of T -essential CIs that can fail is being sibling merged for T
and that after applying the step ‘Sibling merging for T ’ to lhs(α) ∩ lhs(γ) ⊑ rhs(α) ⊓ C the
resulting CI is T -essential, as required.
❏
We also have to show that the number of CIs in Hadd is bounded polynomially in |T |
and for each position of Hadd the number of replacements is bounded polynomially in |T |.
These properties follow from the following lemma.
Lemma 41 Let Hadd be a ordered list of CIs computed at some point of an execution of
Algorithm 3. Then
(i) the length of Hadd is bounded by the number of CIs in T and
(ii) The number of replacements of an existing CI α ∈ Hadd is bounded polynomially in
|T |.
The rest of the section is devoted to proving Lemma 41. We first show Point (ii) of Lemma 41
and start by generalising Lemma 32 on the size of T -essentials CIs. For any conjunction C
of concept names we set
CT
= {D | T |= C ⊑ A1 ⊓ · · · ⊓ Ak and A1 ⊓ · · · ⊓ Ak ⊑ D ∈ T } ∪
{B | T |= C ⊑ B, B basic concept over ΣT }
Recall that for any concept expression C we denote by nC the number of nodes in the tree
TC corresponding to C.
Lemma 42 If α is T -essential, then nrhs(α) ≤
P
D∈lhs(α)T
nD .
Proof Thedproof is almost the same as the proof of Lemma 32.
Let D0 := D∈lhs(α)T D and let ID0 ,T be the canonical model
can prove in almost the same way as in the proof of Lemma 32
h from Trhs(α) into ID0 ,T mapping ρrhs(α) to ρD0 ,T is an injective
Lemma 14 for DL-Lite∃R,horn instead of DL-Lite∃R ).
Assume α is T -essential.
of D0 and T . Now one
that the homomorphism
mapping into ID0 (using
❏
We are now in the position to prove Point (ii) of Lemma 41.
Lemma 43 The number of replacements of an existing CI α ∈ Hadd is bounded polynomially in |T |.
30
Exact Learning of Lightweight Description Logic Ontologies
Proof A CI α ∈ Hadd can be replaced in Line 4 of CN-Refine or in Lines 5 or 8 of ∃-Refine.
If α is replaced by α′ in Line 4 of CN-Refine or in Line 8 of ∃-Refine then lhs(α′ ) ( lhs(α), so
the number of replacements is bounded by |ΣT |. If α is replaced by α′ in Line 5 of ∃-Refine,
then either lhs(α′ ) ( lhs(α) or lhs(α′ ) = lhs(α). For the latter case one can show as in the
proof of Lemma 33 for DL-Lite∃R , the following
Claim. If A1 ⊓ · · · ⊓ An ⊑ C and A1 ⊓ · · · ⊓ An ⊑ C ′ are T -essential, and ∅ |= C ′ ⊑ C, then
TC is obtained from TC ′ by removing subtrees.
Thus, each time α ∈ Hadd is replaced in Line 5 of ∃-Refine without decreasing the
number of concept names in lhs(α), the number nrhs(α) of nodes in the tree representation
of rhs(α) strictly increases. By Lemma 42, nrhs(α) is bounded polynomially in |T | and the
lemma follows.
❏
We now come to the proof of Point (i) of Lemma 41. To formulate an upper bound on the
length of Hadd in terms of T it is convenient to assume that the right-hand side of every CI
in T is primitive, that is, either a concept name or a concept expression of the form ∃r.D.
This assumption is w.l.o.g. since one can equivalently transform every CI C ⊑ D1 ⊓ D2 into
two CIs C ⊑ D1 and C ⊑ D2 . We call such a TBox rhs-primitive. Note that CIs in H may
still have multiple concepts on the right-hand side.
A concept C is called concept saturated for T if T |= C ⊑ C ′ whenever C ′ results from
C by adding a new concept name A′ to the label of some node in TC . Denote by C sat the
(unique) concept obtained from C by adding concept names to the node labels of TC until
it is concept saturated for T . The following definition enables us to link the CIs in Hadd to
the CIs in T .
Definition 44 Let T be rhs-primitive. We say that a CI α has target β ∈ T if
1. lhs(β) ⊆ lhs(α) and
2. there exists D ∈ rhs(α) \ lhs(α) such that ∅ |= rhs(β)sat ⊑ D.
We aim to show that Algorithm 3 maintains the invariant that
(iii) every α ∈ Hadd has some target β ∈ T and
(iv) every β ∈ T is the target of at most one α ∈ Hadd .
Then Point (i) of Lemma 41 clearly follows.
Example 45 To illustrate Definition 44, suppose that
T = {A1 ⊓ A4 ⊑ A2 , A2 ⊑ ∃r.A3 , A3 ⊑ A4 , Ar ≡ ∃r.⊤}
is the target TBox. T is rhs-primitive. To simplify notation, we use βi to denote the i-th
CI occurring in T above. Assume Hbasic = {β3 , β4 } and Hadd = ∅. Let α1 = A1 ⊓ A3 ⊑ A2 .
Then there is no βi ∈ T such that α1 has target βi . However, by applying left saturation
for H to α1 we obtain α′1 = A1 ⊓ A3 ⊓ A4 ⊑ A2 and since lhs(β1 ) ⊆ lhs(α′1 ) and A2 6∈ lhs(α′1 ),
α′1 has target β1 . For α2 = A1 ⊓ A4 ⊑ ∃r.A3 , there is no βi ∈ T such that α2 has
target βi . But α2 is not T -essential and making it T -essential results in α′2 = A1 ⊓ A4 ⊑
31
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
Ar ⊓ A1 ⊓ A2 ⊓ A4 ⊓ ∃r.(A3 ⊓ A4 ) which again has target β1 . Finally, let α3 = A2 ⊑ ∃r.A4 .
As lhs(β2 ) ⊆ lhs(α3 ) and ∅ |= Ar ⊓ ∃r.(A3 ⊓ A4 ) ⊑ ∃r.A4 , α3 has target β2 . Note that α3
is not T -essential, but the result of making it T -essential also has target β2 .
Point (iii) is a consequence of the following lemma.
Lemma 46 Let T be rhs-primitive and let γ be a T -essential CI such that ∅ 6|= γ. Then γ
has some target β ∈ T .
Proof Assume γ is T -essential and ∅ 6|= γ. Assume for a proof by contradiction that γ has
no target in T . We first show the following
Claim 1. If γ has no target in T and T |= lhs(γ) ⊑ A then A ∈ lhs(γ), for all A ∈ NC .
For the proof of Claim 1, consider the canonical model Ilhs(γ),T of lhs(γ) and T . Recall
that ρlhs(γ),T denotes the root of Ilhs(γ),T . By Lemma 12, ρlhs(γ),T ∈ D Ilhs(γ),T iff T |=
lhs(γ) ⊑ D, for any concept D. Thus, it suffices to prove that ρlhs(γ),T ∈ AIlhs(γ),T implies
A ∈ lhs(γ), for all concept names A. The proof is by induction over the sequence I0 , . . .
used to construct Ilhs(γ),T , where I0 = Ilhs(γ) . For Ilhs(γ) this is the case by definition.
Now suppose the claim holds for In and ρlhs(γ),T ∈ AIn+1 \ AIn . Then there either exist
concept names A1 , . . . , Ak with A1 ⊓ · · · ⊓ Ak ⊑ A ∈ T and ρlhs(γ),T ∈ (A1 ⊓ · · · ⊓ Ak )In
or there exists ∃r.⊤ with ∃r.⊤ ⊑ A ∈ T and ρlhs(γ),T ∈ (∃r.⊤)In . In the first case, we
have {A1 , . . . , Ak } ⊆ lhs(γ) by induction hypothesis and so A ∈ lhs(γ) because otherwise
A1 ⊓ · · · ⊓ Ak ⊑ A would be a target of γ. In the second case there must be an Im with
m < n such that there are E1 ⊓ · · · ⊓ Ek ⊑ ∃s.D ∈ T and s ⊑ r ∈ T with ρlhs(γ),T ∈
(E1 ⊓ · · · ⊓ Ek )Im (the case s = r is similar and omitted). It follows that A ∈ lhs(γ) because
otherwise E1 ⊓ · · · ⊓ Ek ⊑ ∃s.D would be a target of γ since, by induction hypothesis,
{E1 , . . . , Ek } ⊆ lhs(γ) and A ∈ (∃s.D)sat . This finishes the proof of Claim 1.
By Claim 1, as ∅ 6|= γ, there is a conjunct of thed
form ∃r.F in rhs(γ). Let (lhs(α))T be as
above and ID0 ,T be the canonical model of D0 = D∈(lhs(α))T D and T . As ∃r.F ∈ rhs(γ)
and γ is T -essential one can show in the same way as in the proof of Lemma 32 that there
is an injective homomorphism from the labelled tree T∃r.F corresponding to ∃r.F into the
restriction of ID0 ,T to ∆ID0 mapping the root of T∃r.F to the root ρD0 ,T of ID0 ,T . Thus, by
definition of (lhs(α))T , there is β ∈ T such that T |= lhs(α) ⊑ lhs(β) and ∅ |= rhs(β)sat ⊑
∃r.F . By Lemma 12, ρlhs(γ),T ∈ AIlhs(γ),T , for all A ∈ lhs(β). Hence, by Claim 1 and
again Lemma 12, lhs(β) ⊆ lhs(γ). We have shown that γ has target β and so derived a
contradiction.
❏
Point (iii) is a direct consequence of Lemma 46 and the fact that all CIs in Hadd are
T -essential (Lemma 38). To prove Point (iv), we first establish the following intermediate
Lemmas 47 and 48.
Lemma 47 Let T be rhs-primitive and let H, γ be inputs to CN-Refine. Let αi ∈ Hadd ,
β ∈ T , and concept name A 6∈ lhs(γ) satisfy the following conditions: (a) lhs(β) ⊆ lhs(γ);
(b) T |= lhs(β) ⊑ A; (c) lhs(β) ⊆ lhs(αi ). Then there is some j ≤ i such that αj is replaced
in Line 4 of CN-Refine.
32
Exact Learning of Lightweight Description Logic Ontologies
Proof Assume H, γ, αi , β, and A satisfy the conditions of the lemma. If CN-Refine replaces
some αj with j < i then we are done. Suppose this does not happen. Then we need to
show that αi is replaced. By Conditions (a), (b), and (c), T |= lhs(γ) ∩ lhs(αi ) ⊑ A. As γ is
left saturated for H, A 6∈ lhs(γ) implies that H 6|= lhs(γ) ⊑ A. So H 6|= lhs(γ) ∩ lhs(αi ) ⊑ A.
Then, the condition in Lines 1 and 2 of CN-Refine is satisfied and αi is replaced.
❏
Lemma 48 Let T be rhs-primitive and let H, γ be inputs to ∃-Refine. If γ has target β ∈ T
and αi ∈ Hadd satisfies lhs(β) ⊆ lhs(αi ), then there is some j ≤ i such that αj is replaced
in Line 5 or 8 of ∃-Refine.
Proof Let H, γ, β, and αi satisfy the conditions of the lemma. If ∃-Refine replaces some
αj with j < i then we are done. Suppose this does not happen. We need to show that αi
is replaced. We first show that there is a concept C of the form ∃r.F in rhs(γ) such that
lhs(γ) ⊑ C has target β. Note that if Algorithm 3 calls ∃-Refine then there is no concept
name A such that T |= lhs(γ) ⊑ A and H 6|= lhs(γ) ⊑ A (Line 10). As γ is left saturated
for H and T -essential, this implies NC ∩ rhs(γ) ⊆ lhs(γ). But then any C ∈ rhs(γ) \ lhs(γ)
with ∅ |= rhs(β)sat ⊑ C is compound. As γ has target β it follows that lhs(γ) ⊑ C has
target β for some C of the form ∃r.F in rhs(γ). By Line 5 of Algorithm 3, there is only
one such conjunct C in rhs(γ). From ∅ |= rhs(β)sat ⊑ C we obtain T |= lhs(β) ⊑ C.
Since lhs(β) ⊆ lhs(γ) ∩ lhs(αi ), we have that T |= lhs(γ) ∩ lhs(αi ) ⊑ C. As γ is a positive
counterexample, H 6|= γ. From NC ∩ rhs(γ) ⊆ lhs(γ) we thus obtain H 6|= lhs(γ) ⊑ C, and
so, H 6|= lhs(αi ) ∩ lhs(γ) ⊑ C. Hence, the condition in Lines 1 and 2 of ∃-Refine is satisfied
and αi is replaced (in Line 5 or 8).
❏
Point (iv) above is now a direct consequence of the following lemma.
Lemma 49 At any point in the execution of Algorithm 3, if αj ∈ Hadd has target β ∈ T
then lhs(β) 6⊆ lhs(αi ), for all i < j.
Proof The proof is by induction on the number k of iterations. For k = 1 the lemma is
vacuously true. Assume it holds for k = n, n ≥ 1. Now the algorithm modifies Hadd in
response to receiving a positive counterexample in iteration k = n + 1. We make a case
distinction:
Case 1. Algorithm 3 calls CN-Refine: Let H, γ be the inputs to CN-Refine. Assume first
that the condition in Lines 1 and 2 is not satisfied. Then CN-Refine appends the result
of concept saturating γ for T to Hadd . Call this CI γ ′ . Suppose that the lemma fails
to hold. This can only happen if γ ′ has a target β ∈ T and there is αi ∈ Hadd such
that lhs(β) ⊆ lhs(αi ). Then, since lhs(γ ′ ) = lhs(γ), we have that lhs(β) ⊆ lhs(γ) and,
since rhs(γ ′ ) ⊆ NC , there is a concept name A 6∈ lhs(γ) such that ∅ |= rhs(β)sat ⊑ A. So
T |= lhs(β) ⊑ A. Then Lemma 47 applies to H, γ, αi , β and A which contradicts the
assumption that CN-Refine did not replace any αj ∈ Hadd , j ≤ i.
Now assume that the condition in Lines 1 and 2 is satisfied. Suppose that the lemma
fails to hold. This can only happen if there are αi , αj ∈ Hadd with i < j such that either
(a) αi is replaced by α′i , lhs(β) ⊆ lhs(α′i ) and αj has target β; or (b) αj is replaced by α′j ,
α′j has target β and lhs(β) ⊆ lhs(αi ). In case (a), from lhs(γ) ∩ lhs(αi ) = lhs(α′i ), we obtain
33
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
lhs(β) ⊆ lhs(γ) ∩ lhs(αi ). Thus, lhs(β) ⊆ lhs(αi ). This contradicts the induction hypothesis.
Now assume case (b). Since α′j has target β, we obtain:
1. lhs(β) ⊆ lhs(α′j ); and
2. as rhs(α′j ) ⊆ NC , there is A ∈ NC with A ∈ rhs(α′j ) \ lhs(α′j ) and ∅ |= rhs(β)sat ⊑ A.
Since lhs(γ) ∩ lhs(αj ) = lhs(α′j ), it follows from Point 1 that lhs(β) ⊆ lhs(αj ) and lhs(β) ⊆
lhs(γ). From ∅ |= rhs(β)sat ⊑ A we obtain T |= lhs(β) ⊑ A. If A ∈ rhs(α′j ) \ lhs(α′j ) then
either A ∈ rhs(αj ) \ lhs(αj ) or A 6∈ lhs(γ). So either αj has target β or A 6∈ lhs(γ). αj does
not have target β as this would contradict the induction hypothesis. Thus, A 6∈ lhs(γ) and
the conditions of Lemma 47 are satisfied by H, γ, αi , β, and A. Thus, some αi′ with i′ ≤ i
is replaced which contradicts the assumption that αj is replaced.
Case 2. Algorithm 3 calls ∃-Refine: Let H, γ be the inputs to ∃-Refine. Assume first that
the condition in Lines 1 and 2 is not satisfied. Then ∃-Refine appends γ to Hadd . Suppose
the lemma fails to hold. This can only happen if γ has a target β ∈ T and there is αi ∈ Hadd
such that lhs(β) ⊆ lhs(αi ). By Lemma 48, this contradicts the assumption that ∃-Refine
did not replace any αj ∈ Hadd , j ≤ i.
Assume now that the condition in Lines 1 and 2 is satisfied. Suppose that the lemma
fails to hold. This can only happen if there are αi , αj ∈ Hadd with i < j such that either (a)
αi is replaced by α′i , lhs(β) ⊆ lhs(α′i ) and αj has target β; or (b) αj is replaced by α′j , α′j has
target β and lhs(β) ⊆ lhs(αi ). For case (a) we argue as above: from lhs(γ)∩lhs(αi ) = lhs(α′i ),
we obtain lhs(β) ⊆ lhs(γ) ∩ lhs(αi ). Thus, lhs(β) ⊆ lhs(αi ), which contradicts the induction
hypothesis. Now assume case (b). As α′j has target β, we obtain the following:
1. lhs(β) ⊆ lhs(α′j ); and
2. there is D ∈ rhs(α′j ) \ lhs(α′j ) and ∅ |= rhs(β)sat ⊑ D.
Since lhs(γ) ∩ lhs(αj ) = lhs(α′j ), it follows from Point 1 that lhs(β) ⊆ lhs(αj ) and lhs(β) ⊆
lhs(γ). Recall that if Algorithm 3 calls ∃-Refine then there is no A ∈ NC such that T |=
lhs(γ) ⊑ A and H 6|= lhs(γ) ⊑ A. So NC ∩ rhs(γ) ⊆ lhs(γ) (by left saturation of γ for H).
Assume D ∈ NC . Since D ∈ rhs(α′j ) \ lhs(α′j ) (Point 2), it follows that D 6∈ lhs(αj ). As
lhs(α′j ) ⊆ lhs(αj ), we have that D ∈ rhs(αj ). So D ∈ rhs(αj ) \ lhs(αj ). This means that
αj has target β, which contradicts the induction hypothesis. Otherwise, D is of the form
∃r.F . Then, either D ∈ rhs(γ) or there is D ′ ∈ rhs(αj ) such that ∅ |= D ⊑ D ′ . In the latter
case, D ′ ∈ rhs(αj ) \ lhs(αj ) and ∅ |= rhs(β)sat ⊑ D ′ , so αj has target β, which contradicts
the induction hypothesis. In the former case, γ has target β. Then H, γ, and αi satisfy
the conditions of Lemma 48. Thus, some αi′ with i′ ≤ i is replaced which contradicts the
assumption that αj is replaced.
❏
We have proved the main result of this section.
Theorem 50 DL-Lite∃R,horn TBoxes are polynomial query learnable using membership and
equivalence queries. Moreover, DL-Lite∃R,horn TBoxes without inverse roles can be learned
in polynomial time using membership and equivalence queries.
34
Exact Learning of Lightweight Description Logic Ontologies
Proof Polynomial query learnability of DL-Lite∃R,horn TBoxes follows from Lemma 41 and
the analysis of the number of membership queries in Lemmas 39 and 40, see the proof
of Theorem 34. For the second part observe that the only potentially costly steps are
entailment checks of the form H |= α, where H is a DL-Lite∃R,horn TBox and α a DL-Lite∃R,horn
CI, both without inverse roles. Then both H and α are in EL with role inclusions for which
entailment is known to be in PTime (Baader et al., 2005).
❏
5. Learning ELlhs TBoxes
We study polynomial learnability of TBoxes in the restriction ELlhs of EL in which only
concept names are allowed on the right-hand side of CIs. We assume that CIs used in
membership queries and in equivalence queries and those returned as counterexamples are
also of this restricted form and show that under this assumption ELlhs TBoxes can be learned
in polynomial time. As in the previous section, our learning algorithm is an extension
of the polynomial time algorithm for learning propositional Horn theories presented by
(Angluin et al., 1992; Arias and Balcázar, 2011).
There is a certain similarity between the learning algorithm of this section and the
DL-Lite∃R,horn learning algorithm introduced in Section 4. In both cases the left-hand side
of inclusions can contain complex concept expressions, which, unless addressed, might lead
to several counterexamples with unnecessarily strong left-hand sides targeting the same
inclusion in the target TBox. In Algorithm 3 storing multiple such counterexamples in Hadd
is prevented by taking the intersection of the set of conjuncts of the left-hand sides. To deal
with the more complex left-hand sides of inclusions in ELlhs , a more sophisticated way of
‘taking the intersection’ of concept expressions is required. To define it, we identify concept
expressions with tree-shaped interpretations and then take their product. Products have
also been employed in the construction of least common subsumers (Baader et al., 1999).
In detail, we say
interpretation I is a ditree interpretation if the directed graph
S that an
I
with E = r∈NR r is a directed tree and r I ∩ sI = ∅ for all distinct r, s ∈ NR . We
denote the root of a ditree interpretation I with ρI . The interpretation IC corresponding
to an EL concept expression C is a ditree interpretation with root ρC . Conversely, every
ditree interpretation I can be viewed as an EL concept expression CI in the same way as
any labelled tree T with edge labels that are role names (rather than arbitrary roles) can
be seen as an EL concept expression.
(∆I , E)
An interpretation I is a T -countermodel for a given ELlhs TBox T if I 6|= T . Notice
that for any ELlhs inclusion C ⊑ A with T |= C ⊑ A and ∅ 6|= C ⊑ A the interpretation IC
is a T -countermodel. Indeed, by construction of IC , we have ρC ∈ C IC and, as ∅ 6|= C ⊑ A,
we have ρC ∈
/ AIC . So IC 6|= C ⊑ A and, as T |= C ⊑ A, we have IC 6|= T . Conversely,
given a T -countermodel I, a learning algorithm can construct in polynomial time in |ΣT |
all inclusions of the form CI ⊑ A, where A is a concept name, such that T |= CI ⊑ A by
posing membership queries to the oracle. Thus a learning algorithm can use inclusions and
T -countermodels interchangeably. We prefer working with interpretations as we can then
use the notion of products to define the ‘intersection of concept expressions’ and the results
of Section 2 linking homomorphisms with entailment in a direct way.
35
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
A
B
d1
Ae
d2
(d2 , e1 )
=
×
d0
A
(d1 , e1 )
1
e0
(d2 , e0 )
(d1 , e0 )
(d0 , e1 )
(d0 , e0 )
Figure 10: Illustration to Example 52.
The product of two interpretations I and J is the interpretation I × J with
∆I×J
= ∆I × ∆J
AI×J
= {(d, e) | d ∈ AI , e ∈ AJ }
r I×J
= {((d, e), (d′ , e′ )) | (d, d′ ) ∈ r I , (e, e′ ) ∈ r J }
Products preserve the membership in EL concept expressions (Lutz et al., 2011):
Lemma 51 For all interpretations I and J , all d ∈ ∆I and e ∈ ∆J , and for all EL concept
expressions C the following holds: d ∈ C I and e ∈ C J if, and only if, (d, e) ∈ C I×J .
One can easily show that the product of ditree interpretations is a disjoint union of ditree
interpretations. If I and J are ditree interpretations, we denote by I ×ρ J the maximal
ditree interpretation that is a subinterpretation of I × J and contains (ρI , ρJ ).
Example 52 Figure 10 depicts the product of the ditree interpretations I with root d0 and
J with root e0 . The ditree interpretation I ×ρ J has root (d0 , e0 ) and does not contain the
nodes (d2 , e0 ), (d1 , e0 ) and (d0 , e1 ) from I × J .
Observe that the product IC ×ID of concept expressions C = A1 ⊓· · ·⊓An and D = B1 ⊓· · ·⊓
Bm , where A1 , . . . , An and B1 , . . . , Bm are concept names, coincides with the interpretation
IE , where E is the conjunction of all concept names in {A1 , . . . , An } ∩ {B1 , . . . , Bm }. Thus,
products can be seen as a generalisation of taking the intersection of the concept names
from the left-hand side of DL-Lite∃R,horn concept inclusions used in Section 4.
We will now describe a class of T -countermodels that are in a sense minimal and central
to our learning algorithm. Let T be the ELlhs TBox to be learned, and assume that its
signature ΣT is known to the learner. For a ditree interpretation I, we use I|−
ρ to denote
the interpretation obtained from I by removing the root ρI of I. For any d ∈ ∆I \ {ρI }, we
use I|−
d↓ to denote I with the subtree rooted at d removed. A T -countermodel is essential
if the following conditions are satisfied:
1. I|−
ρ |= T ;
I
2. I|−
d↓ |= T for all d ∈ ∆ \ {ρI }.
Intuitively, Condition 1 states that I contradicts T only at the root, that is, the only
reason for why I does not satisfy T is that for at least one CI C ⊑ A ∈ T , we have that
ρI ∈ C I and ρI ∈
/ AI . Condition 2 is a minimality condition which states that for any such
36
Exact Learning of Lightweight Description Logic Ontologies
Algorithm 6 The learning algorithm for ELlhs TBoxes
Input: ELlhs TBox T given to the oracle; ΣT given to the learner.
Output: TBox H, computed by the learner, such that T ≡ H.
1: Set I to the empty list (of ditree interpretations)
2: Set H = ∅
3: while H 6≡ T do
4:
Let C ⊑ A be the returned positive counterexample for T relative to H
5:
Find an essential T -countermodel I with I |= H
6:
if there is a J ∈ I such that J 6→ρ (I ×ρ J ) and I ×ρ J 6|= T then
7:
Let J be the first such element of I
8:
Find an essential T -countermodel J ′ ⊆ I ×ρ J
9:
Replace J in I with J ′
10:
else
11:
Append I to I
12:
end if
13:
Construct H = {CI ⊑ A | I ∈ I, A a concept name in ΣT , T |= CI ⊑ A}
14: end while
15: return H
C ⊑ A ∈ T , ρI is no longer in C I if we remove any node from I. Example 61 at the end of
this section shows that working with essential T -countermodels is needed for our learning
algorithm to be in polynomial time.
The algorithm for learning ELlhs TBoxes is given as Algorithm 6. It maintains an ordered
list I of ditree interpretations that intuitively represents the TBox H constructed in Line 13.
In Line 6 we write I →ρ J if there is a homomorphism from a ditree interpretation I to a
ditree interpretation J mapping ρI to ρJ . I 6→ρ J denotes that no such homomorphism
exists. By Lemma 10, I →ρ J iff ∅ |= CJ ⊑ CI which can be checked in polynomial time
in the size of I and J . In Line 8, we write J ′ ⊆ I ×ρ J as shorthand for the condition that
J ′ is a subinterpretation of I ×ρ J that is obtained from I ×ρ J by removing subtrees.
Note that the assumption in Line 4 that a positive counterexample is returned is justified
by the construction of H in Lines 2 and 13, which ensures that, at all times, T |= H.
We now provide additional details on how to realise lines 5, 8 and 13. Line 13 is the
easiest: we simply use membership queries ‘T |= CI ⊑ A?’ with I ∈ I and A ∈ ΣT to
find all CIs CI ⊑ A entailed by T . We will later show that the length of I is bounded
polynomially in |T | and that each interpretation in I is replaced only polynomially many
times, therefore polynomially many membership queries suffice. Lines 5 and 8 are addressed
by Lemmas 53 and 54 below.
Lemma 53 Given a positive counterexample C ⊑ A for T relative to H, one can construct
an essential T -countermodel I with I |= H using only polynomially many membership
queries in |T | + |C|.
37
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
Proof Let C ⊑ A be a positive counterexample for T relative to H. Let IC be the ditree
interpretation of C. First observe that IC 6|= T : since H 6|= C ⊑ A, we know that A does
not occur as a top-level conjunct in C. Consequently, ρC ∈ C IC \ AIC and thus IC 6|= T .
We construct an essential T -countermodel I with I |= H by applying the following rules
to I := IC .
1. Saturate I by exhaustively applying the CIs from H as rules: if D ⊑ B ∈ H and
d ∈ D I , then add d to B I .
2. Replace I by a minimal subtree of I refuting T to address Condition 1 of essential
T -countermodels. To describe how this can be achieved using membership queries
denote for d ∈ ∆I by I|d the ditree interpretation obtained from I by taking the
subtree of I rooted in d. Now check using membership queries for any d ∈ ∆I \ {ρI }
and concept name B whether T |= CI|d ⊑ B. Then replace I by any I|d such that
there exists a B with T |= CI|d ⊑ B and d 6∈ B I|d but there does not exist a d′ in
∆I|d and a B ′ with T |= CI|d′ ⊑ B ′ and d′ 6∈ B I|d′ . If no such d and B exist, then I
is not replaced.
3. Exhaustively remove subtrees from I until Condition 2 of essential T -countermodels
−
is also satisfied: if I|−
d↓ 6|= T , then replace I by I|d↓ . This can again be achieved using
the membership queries T |= CI|− ⊑ B for B a concept name.
d↓
Now we show that the interpretation J constructed above has the required properties. First
observe that J |= H: clearly, the interpretation I constructed in Step 1 is a model of H. As
taking subtrees and removing subtrees from I preserves being a model of H, we conclude
that J |= H. Next we show that J 6|= T : the interpretation I constructed in Step 1 is not
a model of T . In fact, we can use CI ⊑ A as a positive counterexample for T relative to H
instead of C ⊑ A. Observe that ∅ |= CI ⊑ C, and thus T |= C ⊑ A implies T |= CI ⊑ A.
On the other hand, ρI ∈ B I implies H |= C ⊑ B for all concept names B. Consequently
and since H 6|= C ⊑ A, we have ρI ∈
/ AI . Thus I 6|= T . By construction, Steps 2 and
3 preserve the condition that I is not a model of T and so J 6|= T . It remains to argue
that J satisfies Conditions 1 and 2 for essential T -countermodels for H. But Condition 1
is ensured by Step 2 and Condition 2 is ensured by Step 3, respectively.
❏
Lemma 54 Given essential T -countermodels I and J with I ×ρ J 6|= T , one can construct
an essential T -countermodel J ′ ⊆ I ×ρ J using only polynomially many membership queries
in |T | + |I| + |J |.
Proof Let I and J be essential T -countermodels with I ×ρ J 6|= T . Obtain the interpretation J ′ from I ×ρ J by exhaustively applying Rule 3 from the proof of Lemma 53.
As argued above, applying Rule 3 can be implemented using membership queries and J ′
is a T -countermodel. Thus, it remains to argue that it satisfies Conditions 1 and 2 for
T -essential countermodels. For Condition 1, we have to show that J ′ |−
ρ |= T . We know
−
−
−
−
that I|ρ |= T and J |ρ |= T . Thus, by Lemma 51, I|ρ × J |ρ |= T . Now J ′ |−
ρ is obtained
− by removing subtrees and removing subtrees preserves being a model of an
from I|−
×
J
|
ρ
ρ
38
Exact Learning of Lightweight Description Logic Ontologies
′ −
ELlhs TBox. Thus, J ′ |−
ρ |= T . For Condition 2, we have to show that J |d↓ |= T for all
′
d ∈ ∆J \ {ρJ ′ }. But if this is not the case, then the subtree rooted at d would have been
removed during the construction of J ′ from I ×ρ J using Rule 3.
❏
If Algorithm 6 terminates, then it obviously returns a TBox H that is equivalent to the
target TBox T . It thus remains to prove that the algorithm terminates after polynomially
many steps, which is a consequence of the following lemma.
Lemma 55 Let I be a list computed at some point of an execution of Algorithm 6. Then
(i) the length of I is bounded by the number of CIs in T and (ii) each interpretation in each
position of I is replaced only |T | + |T |2 often with a new interpretation.
The rest of this section is devoted to proving Lemma 55. For easy reference, assume
that at each point of the execution of the algorithm, I has the form I0 , . . . , Ik for some
k ≥ 0. To establish Point (i) of Lemma 55, we closely follow (Angluin et al., 1992) and
show that
(iii) for every Ii , there is a Di ⊑ Ai ∈ T with Ii 6|= Di ⊑ Ai and
(iv) if i 6= j, then Di ⊑ Ai and Dj ⊑ Aj are not identical.
In fact, Point (iii) is immediate since whenever a new Ii is added to I in the algorithm, then
Ii is a T -countermodel. To prove Point (iv), we first establish the intermediate Lemma 56
below. For a ditree interpretation I and a CI C ⊑ A, we write I |=ρ C ⊑ A if ρI ∈
/ C I or
ρI ∈ AI ; that is, the CI C ⊑ A is satisfied at the root of I, but not necessarily at other
points in I. It is easy to see that if some interpretation I is a T -countermodel, then there
is C ⊑ A ∈ T such that I 6|=ρ C ⊑ A.
The following lemma shows under which conditions Algorithm 6 replaces an interpretation in the list I.
Lemma 56 If the interpretation I constructed in Line 5 of Algorithm 6 satisfies I 6|=ρ C ⊑
A ∈ T and ρIj ∈ C Ij for some j, then J = Ii is replaced with J ′ in Line 9 for some i ≤ j.
Proof Assume that the interpretation I constructed in Line 5 of Algorithm 6 satisfies
I 6|=ρ C ⊑ A ∈ T and that there is some j with ρIj ∈ C Ij . If there is some i < j such that
Ii 6→ρ (I ×ρ Ii ) and I ×ρ Ii 6|= T , then J = Ii′ will be replaced with J ′ in Line 9 for some
i′ ≤ i and we are done. Thus assume that there is no such i. We aim to show that J = Ij
is replaced with J ′ in Line 9. To this end, it suffices to prove that Ij 6→ρ (I ×ρ Ij ) and
I ×ρ Ij 6|= T . The latter is a consequence of I 6|=ρ C ⊑ A and ρIj ∈ C Ij .
Assume to the contrary of what we have to show that Ij →ρ (I ×ρ Ij ). We establish a
contradiction against I |= H (which holds by construction of I in the algorithm) by showing
that
1. I 6|=ρ CIj ⊑ A and
2. CIj ⊑ A ∈ H.
39
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
For Point 1, Ij →ρ (I ×ρ Ij ) and ρIj ∈ (CIj )Ij imply ρI×ρ Ij ∈ (CIj )I×ρ Ij , which gives
ρI ∈ (CIj )I , by Lemma 51. It remains to observe that I 6|=ρ C ⊑ A implies ρI ∈
/ AI .
In view of the construction of H in the algorithm, Point 2 can be established by showing
that T |= CIj ⊑ A. Since C ⊑ A ∈ T , it suffices to prove that ∅ |= CIj ⊑ C. This, however,
is an immediate consequence of the fact that ρIj ∈ C Ij and the definition of CIj .
❏
Now, Point (iv) above is a consequence of the following.
Lemma 57 At any time of the algorithm execution, the following condition holds: if Ii 6|=ρ
C ⊑ A ∈ T and j < i, then ρIj ∈
/ C Ij .
Proof We prove the invariant formulated in Lemma 57 by induction on the number of
iterations of the while loop. Clearly, the invariant is satisfied before the loop is entered. We
now consider the two places where I is modified, that is, Line 9 and Line 11, starting with
the latter.
In Line 11, I is appended to I. Assume that I |6 =ρ C ⊑ A ∈ T . We have to show that,
before I was added to I, there was no Ii ∈ I with ρIi ∈ C Ii . This, however, is immediate
by Lemma 56.
Now assume that J was replaced in Line 9 with J ′ . We have to show two properties:
/ C Ij .
1. If J ′ = Ii 6|=ρ C ⊑ A ∈ T and j < i, then ρIj ∈
Assume to the contrary that ρIj ∈ C Ij . Since J ′ is obtained from I × J by removing
subtrees (see Lemma 54), J ′ 6|=ρ C ⊑ A implies I × J 6|=ρ C ⊑ A. Consequently,
I 6|=ρ C ⊑ A or J 6|=ρ C ⊑ A. The former and ρIj ∈ C Ij yields i ≤ j by Lemma 56,
in contradiction to j < i. In the latter case, since Ii = J before the replacement of
J with J ′ , we have a contradiction against the induction hypothesis.
/ C Ij .
2. If J ′ = Ij and Ii 6|=ρ C ⊑ A ∈ T with i > j, then ρIj ∈
Assume to the contrary that ρIj ∈ C Ij . Since J ′ is obtained from I ×ρ J by removing
subtrees, we then have ρI×ρ J ∈ C I×ρ J , thus ρJ ∈ C J . Since Ij = J before the
replacement of J with J ′ , we have a contradiction against the induction hypothesis.
❏
We now turn towards proving Point (ii) of Lemma 55. It is a consequence of Lemma 59
below.
Lemma 58 If I is an essential T -countermodel, then |∆I | ≤ |T |.
Proof Let I be an essential T -countermodel. Then I 6|= T , but I|−
ρ |= T . It follows that
I
I
there is a C ⊑ A ∈ T such that ρI ∈ C \ A . By Lemma 8, there is a homomorphism h
from IC to I mapping ρIC to ρI . We show that |∆I | ≤ |C|, from which |∆I | ≤ |T | follows.
It suffices to show that h is surjective. Assume that this is not the case and let d ∈ ∆I
be outside the range of h. Then h is a homomorphism from IC to J := I|−
d↓ . Therefore,
J
J
ρ ∈ C by Lemma 8, which implies J 6|= C ⊑ A. But J 6|= C ⊑ A contradicts the
assumption that I is an essential T -countermodel as it violates Condition 2 of being an
essential T -countermodel.
❏
40
Exact Learning of Lightweight Description Logic Ontologies
Lemma 59 Let I0 , . . . , In be a list of interpretations such that Ii+1 replaces Ii in Line 9
for all i < n. Then n ≤ |T | + |T |2 .
Proof Let I and J be ditree interpretations. We set I ≤ρ J if ρI ∈ AI implies ρJ ∈ AJ
for all concept names A. We first show that for every i < n either
(a) Ii 6≤ρ Ii+1 or
(b) Ii+1 →ρ Ii via a surjective homomorphism.
For a proof by contradiction assume that there is i < n such that neither (a) nor (b) holds.
Since Ii+1 is obtained from some I ×ρ Ii by removing subtrees and (I ×ρ Ii ) →ρ Ii we
obtain that Ii+1 →ρ Ii . Since Ii+1 is an essential T -countermodel, there is a C ⊑ A ∈ T
such that Ii+1 6|=ρ C ⊑ A. Let J be the subinterpretation of Ii determined by the range of
the homomorphism h from Ii+1 to Ii mapping ρIi+1 to ρIi . By Lemma 8, ρJ ∈ C J and so,
since ρJ 6∈ AJ because (a) does not hold, J 6|=ρ C ⊑ A. Ii is an essential T -countermodel
and so J = I. But then h is surjective and we have derived a contradiction.
In addition to the property stated above, we also have for all i < n:
(c) Ii+1 ≤ρ Ii and
(d) Ii 6→ρ Ii+1 .
It follows that for any i < n with Ii ≤ρ Ii+1 either |∆Ii | < |∆Ii+1 | or |AIi | < |AIi+1 | for
some concept name A. By Lemma 58 we have |∆Ii | ≤ |T | for all i ≤ n. Hence k − j ≤ |T |2
for any subsequence Ij . . . , Ik of I0 , . . . , In with Ii ≤ρ Ii+1 for all j ≤ i < k. It follows that
n ≤ |T | + |T |2 .
❏
We have thus established the main result of this section. Note that we obtain a polynomial
time learning algorithm since checking T |= α is in polynomial times for EL TBoxes T and
EL CIs α (as discussed in Section 2).
Theorem 60 ELlhs TBoxes are polynomial time learnable using membership and equivalence queries.
The following example shows that Algorithm 6 does not terminate in polynomial time if in
Line 5 it does not transform the given counterexample into an essential T -countermodel.
Example 61 Assume that Line 5 of Algorithm 6 does not modify the counterexample
C ⊑ A given in Line 4 if the second condition for essential T -countermodels (IC |−
d↓ |= T
−
I
for all d ∈ ∆ C \ {ρIC }) is satisfied but the first condition (IC |ρ |= T ) does not hold. Then
for the target TBox T = {∃r.A ⊑ A} the oracle can return the infinite sequence of positive
counterexamples ∃r n .A ⊑ A, with n a prime number. In fact, Algorithm 6 would simply
construct the list I of interpretations I∃rn .A , n a prime number, and would not terminate.
To show this observe that Algorithm 6 would never replace a CI in the list I by another CI
since I∃rn .A ×ρ I∃rn+m .A = I∃rn .⊤ and I∃rn .⊤ |= T .
41
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
Now assume that Line 5 of Algorithm 6 does not modify the counterexample C ⊑ A
given in Line 4 if the first condition for essential T -countermodels is satisfied but the second
condition does not hold. Let T be a TBox containing ∃r.A ⊑ A and some CIs containing the
concept names B1 and B2 , say, for simplicity, B1 ⊑ B1 and B2 ⊑ B2 . Let ϕ1 = ∃r.(B1 ⊓ B2 )
and ϕn+1 = ∃r.(ϕn ⊓ B1 ⊓ B2 ). Then the oracle can return n positive counterexamples
∃r.A ⊓ Ci ⊑ A, where the tree TCi corresponding to Ci is the result of identifying the
i-th node of the tree Tϕi corresponding to ϕi with the root of the tree corresponding to
∃r.(B1 ⊓ ϕn ) ⊓ ∃r.(B2 ⊓ ϕn ). Note that the product of IC1 , . . . , ICn is an interpretation with
O(2n ) elements. Then, at the n-th iteration, Algorithm 6 computes an interpretation of
exponential size in n.
6. Limits of Polynomial Learnability
The main result of this section is that EL TBoxes are not polynomial query learnable using
membership and equivalence queries. We also show that DL-Lite∃R TBoxes are not polynomial query learnable using membership or equivalence queries alone. The latter result also
holds for ELlhs TBoxes. In this case, however, it follows already from the fact that propositional Horn logic is not polynomial query learnable from entailments using membership or
equivalence queries alone (Frazier and Pitt, 1993; Angluin et al., 1992; Angluin, 1987a).
We start by proving the non-polynomial query learnability result for EL TBoxes. On our
way, we also prove non-polynomial query learnability of DL-Lite∃R TBoxes using membership
queries only. Our proof shows that even acyclic EL TBoxes are not polynomial query
learnable and, in fact, heavily relies on the additional properties of acyclic TBoxes. Recall
that an EL TBox is called acyclic if it satisfies the following conditions (Baader et al., 2017;
Konev et al., 2012):
• all CIs and CEs are of the form A ⊑ C or A ≡ C, where A is a concept name;
• no concept name occurs more than once on the left-hand side of a CI;
• there are no cyclic definitions: there is no sequence α0 , . . . , αn of CIs such that the
concept name on the left-hand side of α0 occurs in αn and the concept name on the
left-hand side of αi+1 occurs in the right-hand side of αi for all i < n.
Our non-polynomial query learnability proof is inspired by Angluin’s lower bound for
the following abstract learning problem (Angluin, 1987b): a learner aims to identify one
of N distinct sets L1 , . . . , LN which have the property that there exists a set L∩ for which
Li ∩ Lj = L∩ , for any i 6= j. It is assumed that L∩ is not a valid argument to an equivalence
query. The learner can pose membership queries “x ∈ L?” and equivalence queries “H =
L?”. Then in the worst case it takes at least N − 1 membership and equivalence queries to
exactly identify a hypothesis Li from L1 , . . . , LN . The proof proceeds as follows. At every
stage of computation, the oracle (which here should be viewed as an adversary) maintains
a set of hypotheses S, which the learner is not able to distinguish based on the answers
given so far. Initially, S = {L1 , . . . , LN }. When the learner asks a membership query x,
the oracle returns ’Yes’ if x ∈ L∩ and ’No’ otherwise. In the latter case, the (unique) Li
such that x ∈ Li is removed from S. When the learner asks an equivalence query H, the
42
Exact Learning of Lightweight Description Logic Ontologies
oracle returns ‘No’ and a counterexample x ∈ L∩ ⊕ H (the symmetric difference of L∩ and
H). This always exists as L∩ is not a valid query. If the counterexample x is not a member
of L∩ , (at most one) Li ∈ S such that x ∈ Li is eliminated from S. In the worst case, the
learner has to reduce the cardinality of S to one to exactly identify a hypothesis, which
takes N − 1 queries.
Similarly to the method outlined above, in our proof we maintain a set of acyclic EL
TBoxes whose members the learning algorithm is not able to distinguish based on the
answers obtained so far. For didactic purposes, we first present a set of acyclic TBoxes
SN = {T1 , . . . , TN }, where N is superpolynomial in the size of every TBox Ti , for which
the oracle can respond to membership queries in the way described above but which is
polynomial time learnable when equivalence queries are also allowed. We then show how
the TBoxes can be modified to obtain a family of acyclic TBoxes that is not polynomial
query learnable using membership and equivalence queries.
To present the TBoxes in SN , fix two role names r and s. We use the following abbreviation. For any sequence σ = σ 1 σ 2 . . . σ n ∈ {r, s}n , the expression ∃σ.C stands for
∃σ 1 .∃σ 2 . . . ∃σ n .C. Then for every such sequence σ, of which there are N = 2n many,
consider the acyclic EL TBox Tσ defined as
Tσ = {A ⊑ ∃σ.M ⊓ X0 } ∪ T0 with
T0 = {Xi ⊑ ∃r.Xi+1 ⊓ ∃s.Xi+1 | 0 ≤ i < n} .
Observe that the canonical model IX0 ,T0 of X0 and T0 consists of a full binary tree whose
edges are labelled with the role names r and s and with X0 at the root ρX0 , X1 at level 1,
and so on. In the canonical model IA,Tσ of A and Tσ , the root is labelled by A and X0 and,
in addition to the binary tree, there is a path given by the sequence σ whose endpoint is
marked by the concept name M .
One can use Angluin’s strategy to show that TBoxes from the set SN of all such TBoxes
Tσ cannot be learned using polynomially many polynomial size membership queries only:
notice that for no sequence σ′ 6= σ of length n, we have Tσ |= A ⊑ ∃σ′ .M . Thus a
membership query of the form A ⊑ ∃σ.M eliminates at most one TBox from the set of
TBoxes that the learner cannot distinguish. This observation can be generalised to arbitrary
membership queries C ⊑ D in EL; however, we instead observe that the TBoxes Tσ are
formulated in DL-Lite∃R and prove a stronger result. The proof, given in the appendix, uses
the canonical model construction introduced in Section 2.
Lemma 62 For every DL-Lite∃R CI B ⊑ D over the signature of Tσ ,
• either Tσ |= B ⊑ D for every Tσ ∈ SN
• or there is at most one Tσ ∈ SN such that Tσ |= B ⊑ D.
The argument outlined above immediately gives us the following side result.
Theorem 63 DL-Lite∃R TBoxes (even without inverse roles) are not polynomial query
learnable using only membership queries.
43
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
We return now to our proof that EL TBoxes are not polynomial query learnable using
both membership and equivalence queries. Notice that the set of TBoxes SN is not suitable
as a single equivalence query is sufficient to learn any TBox from SN in two steps: given the
equivalence query {A ⊑ X0 } ∪ T0 , the oracle has no other option but to reveal the target
TBox Tσ as A ⊑ ∃σ.M can be found ‘inside’ every counterexample.
Our strategy to rule out equivalence queries with the ‘intersection TBox’ is to modify
T1 , . . . , TN in such a way that although a TBox T∩ axiomatising the intersection over the
set of consequences of each Ti , i ≤ N , exists, its size is superpolynomial and so it cannot
be used as an equivalence query by a polynomial query learning algorithm.
For every n > 0 and every n-tuple L = (σ1 , . . . , σn ), where every σi is a role sequence
of length n as above, we define an acyclic EL TBox TL as the union of T0 and the following
CIs and CEs:5
A1 ⊑ ∃σ1 .M ⊓ X0
An ⊑ ∃σn .M ⊓ X0
...
B1 ⊑ ∃σ1 .M ⊓ X0
Bn ⊑ ∃σn .M ⊓ X0
A ≡ X0 ⊓ ∃σ1 .M ⊓ · · · ⊓ ∃σn .M.
Observe that every TL contains the TBoxes Tσi , 1 ≤ i ≤ n, discussed above with A replaced
by any
d of the three concept names A, Ai , Bi . In addition, every TL entails, among other
CIs, ni=1 Ci ⊑ A, where every Ci is either Ai or Bi . There are 2n different such CIs, which
indicates that every representation of the ‘intersection TBox’ requires superpolynomially
many axioms. It follows from Lemma 67 below that this is indeed the case.
Let Ln be a set of n-tuples such that for 1 ≤ i ≤ n and every L, L′ ∈ Ln with L =
(σ1 , . . . , σn ), L′ = (σ1′ , . . . , σn′ ), if σi = σj′ then L = L′ and i = j. Then for any sequence
σ of length n there exists at most one L ∈ Ln and at most one i ≤ n such that TL |= Ai ⊑
∃σ.M and TL |= Bi ⊑ ∃σ.M . We can choose Ln such that there are N = ⌊2n /n⌋ different
tuples in Ln . Notice that the size of each TL with L ∈ Ln is polynomial in n and so N is
superpolynomial in the size of each TL with L ∈ Ln . Let the set of TBoxes that the learner
cannot distinguish initially be SL = {TL | L ∈ Ln }. We use Σn to denote the signature
of TL .
For the proof of non-polynomial query learnability, we show that the oracle has a strategy
to answer both membership and equivalence queries without eliminating too many TBoxes
from SL . We start with the former.
Unlike the DL-Lite∃R case presented above, membership query can eliminate more than
one TBox from SL . Consider, for example, two TBoxes TL and TL′ , where {L, L′ } ⊆ Ln
with L = (σ1 , . . . , σn ) and L′ = (σ1′ , . . . , σn′ ). Then the CI
X0 ⊓ ∃σ1 .M ⊓ ∃σ1′ .M ⊓ A2 ⊓ · · · ⊓ An ⊑ A
is entailed by both TL and TL′ but not by any other TL′′ with L′′ ∈ Ln . We prove, however,
that the number of TBoxes eliminated from SL by a single membership query can be linearly
bounded by the size of the query.
5. In fact, to prove non-polynomial query learnability, it suffices to consider ∃σ1 .M ⊓ · · · ⊓ ∃σn .M ⊑ A in
place of the concept equivalence; however, CIs of this form are not allowed in acyclic TBoxes. CIs with a
complex left-hand side or concept equivalences are essential for non-polynomial query learnability as any
acyclic TBox containing expressions of the form A ⊑ C only is a DL-Lite∃R TBox and thus polynomially
learnable with membership and equivalence queries (Section 3).
44
Exact Learning of Lightweight Description Logic Ontologies
Lemma 64 For all EL CIs C ⊑ D over Σn :
• either TL |= C ⊑ D for every L ∈ Ln
• or the number of L ∈ Ln such that TL |= C ⊑ D does not exceed |C|.
The proof of Lemma 64 is technical and is deferred to the appendix. To illustrate our proof
method here we consider a particular case that deals with membership queries of the form
C ⊑ ∃σ.M and is used in the proof of the general case. Both proofs rely on the following
lemma from (Konev et al., 2012) that characterises CIs entailed by acyclic EL TBoxes.
Lemma 65 (Konev et al.) Let T be andacyclic EL d
TBox, r a role name and D an EL
concept expression. Suppose that T |= 1≤i≤n Ai ⊓ 1≤j≤m ∃rj .Cj ⊑ D, where Ai are
concept names for 1 ≤ i ≤ n, Cj are EL concept expressions for 1 ≤ j ≤ m, and m, n ≥ 0.
Then the following holds:
• if D is a concept name such that T does not contain any CE D ≡ C for any concept
expression C, then there exists Ai , 1 ≤ i ≤ n, such that T |= Ai ⊑ D;
• if D is of the form ∃r.D ′ then either (i) there exists Ai , 1 ≤ i ≤ n, such that T |=
Ai ⊑ ∃r.D ′ or (ii) there exists rj , 1 ≤ j ≤ m, such that rj = r and T |= Cj ⊑ D ′ .
The following lemma considers membership queries of the form C ⊑ ∃σ.M .
Lemma 66 For any 0 ≤ m ≤ n, any sequence of role names σ = σ 1 . . . σ m ∈ {r, s}m , and
any EL concept expression C over Σn :
• either TL |= C ⊑ ∃σ.M for every TL with L ∈ Ln ;
• or there is at most one TL such that TL |= C ⊑ ∃σ.M .
Proof The lemma follows from the following claim.
Claim. Let L = (σ1 , . . . , σn ) ∈ Ln be such that TL |= C ⊑ ∃σ.M . Then either (1) there
exists i ≤ n such that σ = σi and C is of the form A ⊓ C ′ , Ai ⊓ C ′ or Bi ⊓ C ′ , for some EL
concept expression C ′ ; or (2) we have ∅ |= C ⊑ ∃σ.M .
Proof of Claim. We prove the claim by induction on m. If m = 0, by Lemma 65, the concept
expression C is of the form Z ⊓ C ′ , for some concept name Z and concept expression C ′
such that TL |= Z ⊑ M . As TL |= Z ⊑ M does not hold for any concept name Z distinct
from M , we obtain Z = M . Thus, ∅ |= C ⊑ M and Point (2) follows.
Let m > 0. By Lemma 65 we have one of the following two cases:
• C is of the form X ⊓C ′ , for some concept name X and concept expression C ′ such that
TL |= X ⊑ ∃σ.M . But then there exists i ≤ n such that σ = σi and X ∈ {A, Ai , Bi }
and Point (1) follows.
• C is of the form ∃σ 1 .C ′ ⊓ C ′′ , for some concept expressions C ′ and C ′′ , and TL |= C ′ ⊑
∃σ 2 . · · · ∃σ m .M . Notice that the length of the sequence σ 2 . . . σ n is strictly less than n.
Thus, by induction hypothesis, ∅ |= C ′ ⊑ ∃σ 2 . · · · ∃σ m .M . But then ∅ |= C ⊑ ∃σ.M
and Point (2) follows.
45
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
This finishes the proof of the claim. To see that the claim entails the lemma observe that
at most one L ∈ Ln can satisfy Point (1). Point (2) entails that TL |= C ⊑ ∃σ.M for every
TL with L ∈ Ln .
❏
We now show how the oracle can answer equivalence queries, aiming to show that for any
polynomial size equivalence query H, the oracle can return a counterexample C ⊑ D such
that either (i) H |= C ⊑ D and TL |= C ⊑ D for at most one L ∈ Ln or (ii) H 6|= C ⊑ D
and TL |= C ⊑ D for every L ∈ Ln . Thus, such a counterexample eliminates at most one
TL from the set SL of TBoxes that the learner cannot distinguish. In addition, however, we
have to take extra care of the size of counterexamples as the learning algorithm is allowed
to formulate queries polynomial not only in the size of the target TBox but also in the
size of the counterexamples returned by the oracle. For instance, if the hypothesis TBox H
contains a CI C ⊑ D which is not entailed by any TL , one cannot simply return C ⊑ D as a
counterexample since the learner will be able to ‘pump up’ its capacity by asking a sequence
of equivalence queries Hi = {Ci ⊑ Di } such that the size of Ci+1 ⊑ Di+1 is twice the size
of Ci ⊑ Di . Then at every stage in a run of the learning algorithm, the query size will be
polynomial in the size of the input and the size of the largest counterexample received so
far, but exponential size queries will become available to the learner. The following lemma
addresses this issue.
Lemma 67 For any n > 1 and any EL TBox H in Σn with |H| < 2n , there exists an EL
CI C ⊑ D over Σn such that the size of C ⊑ D does not exceed 6n and
• if H |= C ⊑ D, then TL |= C ⊑ D for at most one L ∈ Ln ;
• if H 6|= C ⊑ D, then TL |= C ⊑ D for every L ∈ Ln .
Proof We define an exponentially large TBox T∩ and use it to prove that one can select
the required EL CI C ⊑ D in such a way that either H |= C ⊑ D and T∩ 6|= C ⊑ D, or vice
versa.
n
d To define T∩ , denote for any sequence b = b1 . . . bn ∈ {0, 1} by Cb the conjunction
i≤n Ci , where Ci = Ai if bi = 1 and Ci = Bi if bi = 0. Then we define
T∩ = T0 ∪ {Cb ⊑ A ⊓ X0 | b ∈ {0, 1}n }.
Consider the following cases for H and T∩ .
1. Suppose H 6|= T∩ . Then there exists a CI C ⊑ D ∈ T∩ such that H 6|= C ⊑ D. Clearly,
C ⊑ D is entailed by every TL , for L ∈ Ln , and the size of C ⊑ D does not exceed
6n. Thus C ⊑ D is as required.
2. Suppose there exist b ∈ {0, 1}n and a concept expression of the form ∃t.D ′ such that
H |= Cb ⊑ ∃t.D ′ and T0 6|= Cb ⊑ ∃t.D ′ . It can be seen (Lemma 73 in the appendix),
that there exists a sequence of role names t1 , . . . , tl ∈ {r, s}l with 0 ≤ l ≤ n + 1 and
Y ∈ {⊤} ∪ NC such that ∅ |= ∃t.D ′ ⊑ ∃t1 . · · · ∃tl .Y . Thus, H |= Cb ⊑ ∃t1 . · · · ∃tl .Y
and T0 6|= X0 ⊑ ∃t1 . · · · ∃tl .Y . We show that the inclusion Cb ⊑ ∃t1 . · · · ∃tl .Y is as
required. Clearly, the size of Cb ⊑ ∃t1 . · · · ∃tl .Y does not exceed 6n. It remains to
prove that TL |= Cb ⊑ ∃t1 · · · ∃tl .Y for at most one L ∈ Ln .
46
Exact Learning of Lightweight Description Logic Ontologies
Suppose there exists L ∈ Ln such that TL |= Cb ⊑ ∃t1 . · · · ∃tl .Y . By Lemma 65,
there exists Aj or Bj such that TL |= Aj ⊑ ∃t1 . · · · ∃tl .Y or TL |= Bj ⊑ ∃t1 . · · · ∃tl .Y ,
respectively. As T0 6|= X0 ⊑ ∃t1 . · · · ∃tl .Y it is easy to see that l = n, t1 t2 . . . tn = σj ,
and Y = M follow. As TL′ 6|= Cb ⊑ ∃σj .M for any L′ ∈ Ln such that L′ 6= L, it
follows that TL |= Cb ⊑ ∃t1 · · · ∃tl .Y for at most one L ∈ Ln
3. Finally, suppose that neither Case 1 nor 2 above apply. Then H |= T∩ and for
every b ∈ {0, 1}n and every EL concept expression over Σn of the form ∃t.D ′ : if
H |= Cb ⊑ ∃t.D ′ then T0 |= X0 ⊑ ∃t.D ′ . We show that unless there exists a CI C ⊑ D
satisfying the conditions of the lemma, H contains at least 2n different CIs (and thus
derive a contradiction).
Fix some b = b1 . . . bn ∈ {0, 1}n . From H |= T∩ we obtain H |= Cb ⊑ A. Then there
must exist at least one CI C ⊑ A ⊓ D ∈ H such that H |= Cb ⊑ C and ∅ 6|= C ⊑ A.
Let C = Z1 ⊓ · · · ⊓ Zm ⊓ ∃t1 .C1′ ⊓ · · · ⊓ ∃tl .Cl′ , where Z1 ,. . . , Zm are different concept
names. As H |= Cb ⊑ ∃tj .Cj′ we have T0 |= X0 ⊑ ∃tj .Cj′ , for j = 1, . . . l. As H |= T∩
we have H |= X0 ⊑ ∃tj .Cj′ , for j = 1, . . . l. So H |= Z1 ⊓ · · · ⊓ Zm ⊓ X0 ⊑ A.
• Suppose there exists i such that there is no Zj ∈ {Ai , Bi }. Then we have TL 6|=
Z1 ⊓· · ·⊓Zm ⊓X0 ⊑ A, for any L ∈ Ln . Notice that Z1 ⊓· · ·⊓Zm contains at most
all concepts names in Σn , except Ai , Bi . Thus, the size of Z1 ⊓ · · · ⊓ Zm ⊓ X0 ⊑ A
does not exceed 6n, and Z1 ⊓ · · · ⊓ Zm ⊓ X0 ⊑ A is as required.
• Assume that Z0 ⊓ · · · ⊓ Zm ⊓ X0 contains a conjunct Bi such that bi 6= 0. Then
H |= Cb ⊑ Bi and there is no L ∈ Ln such that TL |= Cb ⊑ Bi . The size of
Cb ⊑ Bi does not exceed 6n, so Cb ⊑ Bi is as required.
• Assume that Z0 ⊓ · · · ⊓ Zm ⊓ X0 contains a conjunct Ai such that bi 6= 1. Then
H |= Cb ⊑ Ai and there is no L ∈ Ln such that TL |= Cb ⊑ Ai . The size of
Cb ⊑ Ai does not exceed 6n, so Cb ⊑ Ai is as required.
• If none of the above applies, then Z1 ⊓ · · · ⊓ Zm ⊓ X0 contains exactly the Ai with
bi = 1 and exactly the Bi with bi = 0.
This argument applies to arbitrary b ∈ {0, 1}n . Thus, if there exists no CI C ⊑ D
satisfying the conditions of the lemma then, by the final case, H contains at least 2n
CIs.
❏
Now we have all the ingredients to prove that EL TBoxes are not polynomial query learnable
using membership and equivalence queries.
Theorem 68 EL TBoxes are not polynomial query learnable using membership and equivalence queries.
Proof Assume that TBoxes are polynomial query learnable. Then there exists a learning
algorithm whose query complexity (the sum of the sizes of the inputs to membership and
equivalence queries made by the algorithm up to a computation step) is bounded at any
stage by a polynomial p(n, m). Choose n such that ⌊2n /n⌋ > (p(n, 6n))2 and let SL =
{TL | L ∈ Ln }. We follow Angluin’s strategy of letting the oracle remove TBoxes from SL
47
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
in such a way that the learner cannot distinguish between any of the remaining TBoxes.
Given a membership query C ⊑ D, if TL |= C ⊑ D for every L ∈ Ln , then the answer is
‘yes’; otherwise the answer is ‘no’ and all TL with TL |= C ⊑ D are removed from SL (by
Lemma 64, there are at most |C| such TBoxes). Given an equivalence query H, the answer
is ‘no’, a counterexample C ⊑ D guaranteed by Lemma 67 is produced, and (at most one)
TL such that TL |= C ⊑ D is removed from SL .
As all counterexamples produced are smaller than 6n, the overall query complexity
of the algorithm is bounded by p(n, 6n). Hence, the learner asks no more than p(n, 6n)
queries and the size of every query does not exceed p(n, 6n). By Lemmas 64 and 67, at
most (p(n, 6n))2 TBoxes are removed from SL during the run of the algorithm. But then,
the algorithm cannot distinguish between any remaining TBoxes and we have derived a
contradiction.
❏
We conclude this section by showing that DL-Lite∃R TBoxes cannot be learned using
polynomially many polynomial size equivalence queries only. We use the following result on
non-polynomial query learnability of monotone DNF formulas, that is, DNF formulas that
do not use negation, using equivalence queries due to Angluin (1990). Here, equivalence
queries take a hypothesis ψ in the form of a monotone DNF formula and return as a
counterexample either a truth assignment that satisfies ψ but not the target formula φ or
vice versa. Let M (n, t, s) denote the set of all monotone DNF formulas whose variables are
x1 , . . . , xn , that have exactly t conjunctions, and where each conjunction contains exactly
s variables.
Theorem 69 (Angluin) For any polynomial q(·) there exist constants t0 and s0 and a
strategy 6 for the oracle O to answer equivalence queries posed by a learning algorithm in
such a way that for sufficiently large n any learning algorithm that asks at most q(n) equivalence queries, each bounded in size by q(n), cannot exactly identify elements of M (n, t0 , s0 ).
To employ Theorem 69, we associate with every monotone DNF formula
φ=
t
_
(xi1 ∧ · · · ∧ xisi ),
i=1
where {xi1 , . . . , xisi } ⊆ {x1 , . . . , xn }, a DL-Lite∃R TBox Tφ as follows. With each conjunct
xi1 ∧ · · · ∧ xisi we associate a concept expression
Ci := ∃ρi1 .∃ρi2 . . . . ∃ρin .⊤,
where ρij = r if xj occurs in xi1 ∧ · · · ∧ xisi and ρij = r̄ otherwise (r and r̄ are role names).
Let A be a concept name and set
d
Tφ = {A ⊑ ti=1 Ci , r̄ ⊑ r}.
6. The existence of this strategy is a direct consequence of Theorem 8 of (Angluin, 1990), which states
that the class of DNF formulae has the approximate fingerprint property, and the proof of Theorem 1
of (Angluin, 1990), where such a strategy is explicitly constructed for any class having approximate
fingerprints.
48
Exact Learning of Lightweight Description Logic Ontologies
For example, for n = 4 and φ = (x1 ∧ x4 ) ∨ x2 we have
Tφ = {A ⊑ ∃r.∃r̄.∃r̄.∃r.⊤, A ⊑ ∃r̄.∃r.∃r̄.∃r̄.⊤, r̄ ⊑ r}.
We say that a TBox T has a DNF-representation for n if it is obtained by the translation
of a monotone DNF-formula with n variables; that is, if T is of the following form, for some
Γ ⊆ {r, r̄}n :
l
{A ⊑
∃ρ1 .∃ρ2 . . . . ∃ρn .⊤, r̄ ⊑ r}.
ρ1 ···ρn ∈Γ
A truth assignment I (for the variables x1 . . . , xn ) also corresponds to a concept expression
CI := ∃ρi1 .∃ρi2 . . . . ∃ρin .⊤,
where ρij = r if I makes xj true and ρij = r̄ otherwise. Then
I |= φ if, and only if, Tφ |= A ⊑ CI
holds for all truth assignments I.
Note that r̄ represents that a variable is false and r that a variable is true. Thus, the
RI r̄ ⊑ r captures the monotonicity of the DNF formulas considered. For any fixed values
n, s and t, we set
T (n, t, s) = {Tφ | φ ∈ M (n, t, s)}.
Note that the TBoxes in T (n, t, s) are exactly those TBoxes that have a DNF-representation
for n and satisfy additionally the conditions that the DNF represented by Tφ has exactly t
conjunctions each conjunction of which has exactly s variables.
We describe now the strategy for the oracle O′ to answer equivalence queries so that
no learning algorithm is able to exactly identify members of T (n, t, s) based on the answers
to polynomially many equivalence queries of polynomial size. If the TBox in the equivalence query is ‘obviously’ not within the class T (n, t, s), then we will explicitly produce
a counterexample that the oracle can return. If, on the other hand, the TBox H from
the equivalence query is ‘similar’ to TBoxes that have a DNF-representation for n, then
we approximate H by a TBox H′ that has a DNF-representation for n and return the
counterexample A ⊑ CI corresponding to the truth assignment I that the oracle O from
Theorem 69 would return when given ψ.
In detail the strategy is as follows. Assume q is the given polynomial in Theorem 69
and that t0 , s0 and the strategy of the oracle O are chosen so that for sufficiently large n
no learning algorithm for DNF formulas that asks at most q(n) equivalence queries, each
bounded in size by q(n), can distinguish all members of M (n, t0 , s0 ). Choose a sufficiently
large n. Let H be an equivalence TBox query issued by a learning algorithm. Then O′ does
the following:
1. If H entails some A ⊑ ∃ρ1 .∃ρ2 . . . . ∃ρn+1 .⊤ with ρi ∈ {r, r̄} for 1 ≤ i ≤ n + 1, then
return this CI as a negative counterexample;
2. If H entails some ∃ρ1 .⊤ ⊑ ∃ρ2 .⊤ such that {ρ1 , ρ2 } ⊆ {r, r̄, r − , r̄ − } and {r̄ ⊑ r} 6|=
∃ρ1 .⊤ ⊑ ∃ρ2 .⊤, then return this CI as a negative counterexample;
49
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
3. If H |= ∃ρ1 .⊤ ⊑ ∃ρ2 .∃ρ3 .⊤ such that {ρ1 , ρ2 , ρ3 } ⊆ {r, r̄}, then return this CI as a
negative counterexample;
4. If there exists no ρ1 , . . . , ρn ∈ {r, r̄}n such that H |= A ⊑ ∃ρ1 . · · · ∃ρn .⊤ then return
A ⊑ |∃r ·{z
· · ∃r} .⊤ as a positive counterexample.
n
5. Suppose now that none of the above applies. We say that a sequence ρ1 , . . . , ρn ∈
{r, r̄}n is r-minimal for H if H |= A ⊑ ∃ρ1 . · · · ∃ρn .⊤ and whenever ρi = r, for
1 ≤ i ≤ n, we have H 6|= ∃ρ1 . · · · ∃ρi−1 .∃r̄.∃ρi+1 . · · · ∃ρn .⊤. We obtain a TBox H′ with
a DNF representation by setting
l
∃ρ1 . · · · ∃ρn .⊤, r̄ ⊑ r}.
H′ = {A ⊑
ρ1 ,...,ρn is
r-minimal for H
Observe that for any sequence ρ1 , . . . , ρn ∈ {r, r̄}n we have H |= A ⊑ ∃ρ1 . · · · ∃ρn .⊤ if,
and only if, H′ |= A ⊑ ∃ρ1 . · · · ∃ρn .⊤. We convert H′ into its corresponding monotone
DNF formula φH′ by reversing the translation from monotone DNF formulas into
DL-Lite∃R TBoxes of the above form in the obvious way. Note that the size of φH′
is linear in the size of H′ . Given φH′ the oracle O returns a (positive or negative)
counterexample (a truth assignment) I. Then return the counterexample in the form
of the CI A ⊑ CI .
Observe that the answers given in Points 1 to 3 are correct in the sense that if an inclusion
α is returned as a negative example then T 6|= α for any T ∈ T (n, t, s). Point 4 is trivially
correct, since any monotone DNF is satisfied by the truth assignment that makes every
variable true. We analyse the size of the TBox H′ computed in Point 5.
Lemma 70 Assume that Points 1 to 4 do not apply to H. Then the number of sequences
ρ1 , . . . , ρn ∈ {r, r̄}n which are r-minimal for H is bounded by |H|.
Proof We first show that if ρ1 , . . . , ρn ∈ {r, r̄}n is r-minimal for H, then there exists a CI
A ⊑ C ∈ H such that
(∗) there are concept expressions C0 , . . . , Cn with C0 = C and ∃ρi+1 .Ci+1 a top-level
conjunct of Ci , for all i < n.
For the proof we require the canonical model IA,H of A and H (Lemma 12). Denote
the root of IA,H by ρA . Let ρ1 , . . . , ρn ∈ {r, r̄}n be r-minimal for H. Then there are
I
d0 , . . . , dn ∈ ∆IA,H with d0 = ρA such that (di , di+1 ) ∈ ρi A,H for all i < n. By the
canonical model construction and the assumption that Points 2 and 3 do not hold, there
either exists di ∈ AIA,H or there is a CI A ⊑ C ∈ H such that (∗) holds. We show that
the first condition does not hold. Assume for a prove by contradiction that di ∈ AIA,H . By
Lemma 12, H |= A ⊑ ∃ρ1 · · · ∃ρi .A. But then H |= A ⊑ ∃(ρ1 · · · ρi )n .⊤ for all n > 0 which
contradicts the assumption that Point 1 does not apply to H.
It follows that the number of distinct r-minimal sequences is bounded by the number of
distinct sequences C0 , . . . , Cn with A ⊑ C0 ∈ H and ∃ρi+1 .Ci+1 a top-level conjunct of Ci
for all i < n. Thus, the number of distinct r-minimal sequences is bounded by |H|.
❏
50
Exact Learning of Lightweight Description Logic Ontologies
It follows from Lemma 70 that the size of the TBox H′ computed in Point 5 is bounded by
4n|H| + 2.
Theorem 71 DL-Lite∃R TBoxes (even without inverse roles) are not polynomial query
learnable using only equivalence queries.
Proof Suppose that the query complexity of a learning algorithm A for DL-Lite∃R TBoxes
in Σ = {A, r, r̄} is bounded at every stage of computation by a polynomial p(x, y), where
x is the size of the target TBox, and y is the maximal size of a counterexample returned
by the oracle up to the current stage of computation. Let q(n) = (p(n2 , 4n + 6))2 , and let
constants t0 and s0 be as guaranteed by Lemma 69. We claim that, for sufficiently large n,
A cannot distinguish some Tφ and Tψ for φ, ψ ∈ M (n, t0 , s0 ).
Assuming that n > 11 (the maximal size of counterexamples given under Point 2 and
3), the largest counterexample returned by our strategy described above is of the form
A ⊑ ∃ρ1 . · · · ∃ρn+1 .⊤, so for sufficiently large n the maximal size of any counterexample in
any run of A is bounded by 4n+6 = 4(n+1)+2. Similarly, the size of every potential target
TBox Tφ ∈ T (n, t0 , s0 ) does not exceed t0 · (4n + 2) and, as t0 is a constant, for sufficiently
large n it is bounded by n2 . Thus, for sufficiently large n the total query complexity of
A on any input from T (n, t0 , s0 ) is bounded by p(n2 , 4n + 6). Obviously, the size of each
query is bounded by the query complexity of the learning algorithm. So, the size of a DNF
equivalence query forwarded to the strategy O guaranteed by Lemma 69 is bounded by
4n × p(n2 , 4n + 6) + 2 ≤ q(n), and there will be at most q(n) queries forwarded. But then
O can return answers such that some φ and ψ from M (n, t0 , s0 ) cannot be distinguished.
It remains to observe that A cannot distinguish Tφ and Tψ .
❏
7. Related Work
Some related work has already been discussed in the introduction to this paper. Here we
discuss in more detail related work from ontology learning in general and exact learning of
ontologies in particular. We start with the former.
Ontology Learning. Research in ontology learning has a rich history that we cannot discuss here in full detail. The collection (Lehmann and Völker, 2014) and surveys (Cimiano et al.,
2010; Wong et al., 2012) provide an excellent introduction to the state of the art in this field.
The techniques applied in ontology learning range from information extraction and text mining to interactive learning and inductive logic programming (ILP). Of particular relevance
for this paper are the approaches to learning logical expressions (rather than subsumption
hierarchies between concept names). For example, the work in (Lehmann and Haase, 2009;
Lehmann and Hitzler, 2010; Bühmann et al., 2014) applies techniques from ILP to learn description logic concept expressions. ILP is applied as well in (Lisi, 2011) for learning logical
rules for ontologies. The learning of fuzzy DLs has been considered in (Lisi and Straccia,
2015). Other machine learning methods which have been applied to learn ontology axioms
include Association Rule Mining (ARM) (Völker and Niepert, 2011; Fleischhacker et al.,
2012; Völker et al., 2015) and Formal Concept Analysis (FCA) (Rudolph, 2004; Baader et al.,
2007; Distel, 2011; Borchmann, 2014; Ganter et al., 2016). Recently, learnability of lightweight
51
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
DL TBoxes from finite sets of interpretations has been investigated in (Klarman and Britz,
2015).
Exact Learning of Description Logic Concept Expressions. Rather than aiming
to learn a TBox here one is interested in learning a target concept expression C∗ . This was
first studied in (Cohen and Hirsh, 1994a,b; Frazier and Pitt, 1996). The standard learning
protocol is as follows:
• a membership query asks whether a concept expression C is subsumed by the target
concept expression C∗ (in symbols, ∅ |= C ⊑ C∗ ?);
• an equivalence query asks whether a concept expression C is equivalent to the target
concept expression C∗ (in symbols, ∅ |= C ≡ C∗ ?). If C and C∗ are not equivalent
then the oracle gives a counterexample, that is, a concept expression C ′ such that
either ∅ |= C ′ ⊑ C∗ and ∅ 6|= C ′ ⊑ C or ∅ 6|= C ′ ⊑ C∗ and ∅ |= C ′ ⊑ C.
(Cohen and Hirsh, 1994a,b; Frazier and Pitt, 1996) consider concept expressions in (variations of) the now largely historic description logic Classic (Borgida et al., 1989; Patel-Schneider et al.,
1991; Borgida and Patel-Schneider, 1994). The expressive power of Classic and its variants
is incomparable to the expressive power of modern lightweight description logics. Classic only shares conjunction and unqualified existential restrictions of the form ∃r.⊤ with
the DLs considered in this paper. It additionally admits value restrictions ∀r.C whose
interpretation is given as
(∀r.C)I = {d ∈ ∆I |d′ ∈ C I for all d′ with (d, d′ ) ∈ r I }
and unqualified number restrictions (≤ n r) and (≥ n r) interpreted as
(≤ nr)I
= {d ∈ ∆I | |{d′ | (d, d′ ) ∈ r I }| ≤ n}
(≥ nr)I
= {d ∈ ∆I | |{d′ | (d, d′ ) ∈ r I }| ≥ n}
as well as various constructors using individual names. For example, if a1 , . . . , an are names
for individual objects, then ONE-OF(a1 , . . . , an ) is a Classic concept denoting the set
{aI1 , . . . , aIn }, where aIi denotes the individual with name ai in interpretation I. It is proved
in (Cohen and Hirsh, 1994a,b; Frazier and Pitt, 1996) that in many fragments of Classic
concept expressions cannot be learned polynomially using only membership or equivalence
queries but that they can be learned in polynomial time using both. Exact learning of
concept expressions in modern lightweight description logics has not yet been investigated.
Exact Learning of TBoxes using Concept Inclusions as Queries. First results on
exact learning of description logic TBoxes using concept inclusions as queries were presented in (Konev et al., 2013, 2014), of which this paper is an extension. In contrast
to (Konev et al., 2013, 2014), we make the distinction between polynomial time and polynomial query learnability which enables us to formulate and prove results on a more fine
grained level. TBoxes in DL-Lite∃R,horn , for which we prove polynomial query learnability,
were not considered in (Konev et al., 2013, 2014). The current paper is also closely related
to the PhD thesis of the third author (Ozaki, 2016). In addition to the results presented
here, it is shown there that even in the extension of ELlhs with role inclusions, TBoxes can
be learned in polynomial time. The learning algorithm is a non-trivial extension of the
algorithm presented here for ELlhs TBoxes.
52
Exact Learning of Lightweight Description Logic Ontologies
Exact Learning of TBoxes using Certain Answers. In recent years, data access mediated by ontologies has become one of the most important applications of DLs, see (Poggi et al.,
2008; Bienvenu et al., 2014; Kontchakov and Zakharyaschev, 2014; Bienvenu and Ortiz, 2015)
and references therein. The idea is to use a TBox to specify semantics and background
knowledge for the data and use it for deriving more complete answers to queries over the
data. In this context, the data is stored in an ABox consisting of a finite set of assertions of
the form A(a) or r(a, b), where A is a concept names, r a role name, and a, b are individual
names. Given a query q(~x) (typically a conjunctive query), a TBox T , and an ABox A, a
tuple of individual names ~a from A and of the same length as ~x is called a certain answer
to q(~x) over A w.r.t. T , in symbols T , A |= q(~a), if every model I of T and A satisfies q(~a).
Motivated by this setup, (Konev et al., 2016; Ozaki, 2016) study polynomial learnability
of TBoxes using membership queries that ask whether a tuple of individuals names is a
certain answer to a query over an ABox w.r.t. the target TBox. This is a natural alternative to learning using concept inclusions since domain experts are often more familiar with
querying data in a particular domain than with the logical notion of subsumption between
concept expressions. In detail, the learning protocol is as follows:
• a membership query takes the form (A, q(~a)) and asks whether the tuple ~a of individual
names is a certain answer to the query q(~x) over the ABox A w.r.t. the target TBox T ;
• an equivalence query asks whether a TBox H is equivalent to the target TBox T . If
T and H are not equivalent then a counterexample of the form (A, q(~a)) is given such
that T , A |= q(~a) and H, A 6|= q(~a) (a positive counterexample) or T , A 6|= q(~a) and
H, A |= q(~a) (a negative counterexample).
In the learning protocol above we have not yet specified the class of queries from which the
q(~x) are drawn and which strongly influences the classes of TBoxes that can be learned. In
the context of data access using TBoxes the two most popular classes of queries are:
• conjunctive queries (CQs), that is, existentially quantified conjunctions of atoms; and
• instance queries (IQs), which take the form C(x) or r(x, y) with C a concept expression
from the DL under consideration and r a role name.
In (Konev et al., 2016; Ozaki, 2016), exact learning of TBoxes in the languages EL, ELlhs
and DL-Lite∃R is studied for both IQs and CQs in queries. The positive learnability results
are proved by polynomial reductions to the learnability results presented in this paper
and (Ozaki, 2016). The basic link between learning using concept inclusions as queries and
learning by certain answers is as follows: if T is a TBox and C, D are concept expressions in
any of the DLs discussed above then one can regard the labelled tree TC corresponding to C
as an ABox AC with root ρC and it holds that T |= C ⊑ D if, and only if, T , AC |= D(ρC ).
The converse direction (obtaining a concept expression from an ABox) is more involved
since ABoxes are not tree-shaped and an additional unfolding step is needed to compute
a corresponding concept expression. Using this link it is proved in (Konev et al., 2016;
Ozaki, 2016) that DL-Lite∃R and ELlhs TBoxes with role inclusions can be learned with
polynomially many queries using certain answers to IQs. It is also proved that EL is still
not learnable with polynomially many queries using certain answers with neither IQs nor
CQs as the query language and that DL-Lite∃R TBoxes cannot be learned with polynomially
many queries using certain answers with CQs as the query language.
53
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
Exact Learning in (other) Fragments of FO Horn. We discuss results on exact
learning of finite sets of FO Horn clauses or fragments of this logic, where a FO Horn
clause is a universally quantified clause with at most one positive literal (Page Jr, 1993;
Arimura, 1997; Reddy and Tadepalli, 1998; Arias and Khardon, 2002; Arias et al., 2007;
Selman and Fern, 2011). Depending on what is used as membership queries and as counterexamples to equivalence queries, one can distinguish between exact learning FO Horn
clauses using interpretations and using entailments. As learning using entailments is closer
to our approach we focus on that setting. The exact learning protocol is then as follows:
• a membership query asks whether an FO Horn clause is entailed by the target set T
of FO Horn clauses;
• an equivalence query asks whether a set H of FO Horn clauses is equivalent to the
target set T . If H and T are not equivalent then a counterexample is given, that is,
an FO Horn clause entailed by T but not by H (a positive counterexample) or vice
versa.
Considering how terms (with function symbols allowed) can appear in an FO Horn clause,
two main restrictions have been studied in the literature:
1. Range restricted clauses: when the set of terms in the positive literal (if existent) is a
subset of the terms in the negative literals and their subterms; and
2. Constrained clauses: when the set of terms and subterms in the positive literal (if
existent) is a superset of the terms in the negative literals.
For example, the FO Horn clause ∀x(¬P (f (x)) ∨ P (x)) is range restricted but not constrained and the FO Horn clause ∀x(¬P (x) ∨ P (f (x))) is constrained but not range restricted, where P is a predicate symbol and f a function symbol. In (Reddy and Tadepalli,
1998) and (Arimura, 1997), it is shown that under certain acyclicity conditions FO Horn
with range restricted clauses and, respectively, constrained clauses are polynomial time
learnable from entailments if the arity of predicates is bounded by a constant. A learning algorithm for a fragment of FO Horn (called closed FO Horn) that subsumes the two
languages defined above is presented in (Arias and Khardon, 2002). The algorithm is polynomial in the number of clauses, terms and predicates and the size of the counterexamples,
but exponential not only in the arity of predicates but also in the number of variables per
clause. In fact, it is an open question whether there exists a learning algorithm for closed
FO Horn that is polynomial in the number of variables per clause.
We relate the learnability results for FO Horn to the learnability results for lightweight
description logics presented in this paper. Observe that most DLs (and in particular all
DLs investigated in this paper) can be translated into FO (Baader et al., 2003). For example, a translation of the ELlhs CI ∃r.A ⊑ B is ∀x∀y(¬r(x, y) ∨ ¬A(y) ∨ B(x)) and a
translation of the DL-Lite∃R CI A ⊑ ∃r.A is ∀x(A(x) → ∃y.(r(x, y) ∧ A(y))). Under this
translation, every ELlhs TBox can be regarded as a set of range restricted FO Horn clauses,
where the arity of predicates is bounded by 2. In contrast, since in DL-Lite∃R existential
quantifiers can be nested in the right side of CIs, DL-Lite∃R CIs cannot be translated into
FO Horn clauses. We can now summarise the relationship between our learnability results
54
Exact Learning of Lightweight Description Logic Ontologies
for ELlhs , DL-Lite∃R and DL-Lite∃R,horn and the results on exact learnability of FO Horn from
entailments as follows: since the arity of DL predicates is at most 2 and since no function
symbols are admitted in DLs, none of the DLs considered in this paper can express the
fragments of FO Horn discussed above. On the other hand, we do not impose an acyclicity condition on the TBoxes (in contrast to (Reddy and Tadepalli, 1998; Arimura, 1997))
and our algorithms are polynomial in the number of variables permitted in any clause (in
contrast to (Arias and Khardon, 2002)). Thus, the results discussed above for FO Horn
do not translate into polynomial learning algorithms for ELlhs and are not applicable to
DL-Lite∃R nor DL-Lite∃R,horn . Our results thus cover new fragments of FO that have not yet
been considered for exact learning. This is not surprising, given the fact that the fragments
of FO considered previously were not motivated by applications in ontology learning.
Also related to exact learning of Horn FO is recent work on exact learning of schema
mappings in data exchange (ten Cate et al., 2012). Schema mappings are tuples (S, T, M )
where S is a source schema (a finite set of predicates), T is a target schema (a finite set
of predicates), and M is a finite set of sentences of the form ∀~x(ϕ(~x) → ∃~yψ(~x, ~y )) where
ϕ(~x) and ψ(~x, ~y ) are conjunctions of atoms over S and T , respectively (Fagin et al., 2005).
(S, T, M ) is a GAV schema mapping if ~y is empty and ψ(~x, ~y ) is an atom. In (ten Cate et al.,
2012), the authors study exact learnability of GAV schema mappings from data examples
(I, J) consisting of a database I over the source schema S and a database J over the
target schema T . Such a data example satisfies M if I ∪ J |= M . The authors present
both polynomial query learnability results for protocols using membership and equivalence
queries and non-polynomial query learnability results if either only membership or only
equivalence queries are allowed. The results presented in (ten Cate et al., 2012) are not
applicable to the setting considered in this paper since the learning protocol uses data
examples instead of entailments.
8. Conclusion
We have presented the first study of learnability of DL ontologies in Angluin et al’s framework of exact learning, obtaining both positive and negative results. Several research questions remain to be explored. One immediate question is whether acyclic EL TBoxes can
be learned in polynomial time using queries and counterexamples of the form A ≡ C and
A ⊑ C only. Note that our non-polynomial query learnability result for acyclic EL TBoxes
relies heavily on counterexamples that are not of this form. Another immediate question is
whether the extension of ELlhs with inverse roles (which is a better approximation of OWL2
RL than ELlhs itself) can still be learned in polynomial time, or at least with polynomially
many queries of polynomial size. Other interesting research directions are non-polynomial
time learning algorithms for EL TBoxes and the admission of different types of membership
queries and counterexamples in the learning protocol. For example, one could replace CIs
as counterexamples with interpretations.
Acknowledgements Lutz was supported by the DFG project Prob-DL (LU1417/1-1).
Konev and Wolter were supported by the EPSRC project EP/H043594/1. Ozaki was supported by the Science without Borders scholarship programme.
55
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
References
Dana Angluin. Learning propositional Horn sentences with hints. Technical report, Yale
University, 1987a.
Dana Angluin. Queries and concept learning. Machine Learning, 2(4):319–342, 1987b.
Dana Angluin. Negative results for equivalence queries. Machine Learning, 5:121–150, 1990.
Dana Angluin, Michael Frazier, and Leonard Pitt. Learning conjunctions of Horn clauses.
Machine Learning, 9:147–164, 1992.
Marta Arias. Exact learning of first-order expressions from queries. PhD thesis, Tufts
University, 2004.
Marta Arias and José L. Balcázar. Construction and learnability of canonical Horn formulas.
Machine Learning, 85(3):273–297, 2011.
Marta Arias and Roni Khardon. Learning closed Horn expressions. Information and Computation, 178(1):214–240, 2002.
Marta Arias, Roni Khardon, and Jérôme Maloberti. Learning Horn expressions with
LOGAN-H. Journal of Machine Learning Research, 8:549–587, 2007.
Hiroki Arimura. Learning acyclic first-order horn sentences from entailment. In International Workshop on Algorithmic Learning Theory, pages 432–445, 1997.
Alessandro Artale, Diego Calvanese, Roman Kontchakov, and Michael Zakharyaschev. The
DL-Lite family and relations. Journal of Artificial Intelligence Research (JAIR), 36:1–69,
2009.
Franz Baader, Ralf Küsters, and Ralf Molitor. Computing least common subsumers in
description logics with existential restrictions. In International Joint Conference on Artificial Intelligence (IJCAI), pages 96–103, 1999.
Franz Baader, Diego Calvanese, Deborah L. McGuinness, Daniele Nardi, and Peter F.
Patel-Schneider, editors. The Description Logic Handbook: Theory, Implementation, and
Applications. Cambridge University Press, New York, NY, USA, 2003. ISBN 0-52178176-0.
Franz Baader, Sebastian Brandt, and Carsten Lutz. Pushing the EL envelope. In International Joint Conference on Artificial Intelligence (IJCAI), pages 364–369, 2005.
Franz Baader, Bernhard Ganter, Baris Sertkaya, and Ulrike Sattler. Completing description
logic knowledge bases using formal concept analysis. In International Joint Conference
on Artificial Intelligence (IJCAI), pages 230–235, 2007.
Franz Baader, Carsten Lutz, and Sebastian Brandt. Pushing the EL envelope further. In
Proceedings of the Fourth OWLED Workshop on OWL: Experiences and Directions, 2008.
56
Exact Learning of Lightweight Description Logic Ontologies
Franz Baader, Ian Horrocks, Carsten Lutz, and Ulrike Sattler. An Introduction to Description Logic. Cambridge University Press, 2017.
Meghyn Bienvenu and Magdalena Ortiz. Ontology-mediated query answering with datatractable description logics. In Reasoning Web. Semantic Technologies for Advanced
Query Answering - 11th International Summer School, pages 218–307, 2015.
Meghyn Bienvenu, Balder ten Cate, Carsten Lutz, and Frank Wolter. Ontology-based data
access: A study through disjunctive datalog, csp, and MMSNP. ACM Trans. Database
Syst., 39(4):33:1–33:44, 2014.
Daniel Borchmann and Felix Distel. Mining of EL-GCIs. In The 11th IEEE International
Conference on Data Mining Workshops, Vancouver, Canada, 2011.
Dipl-Math Daniel Borchmann. Learning terminological knowledge with high confidence from
erroneous data. PhD thesis, Higher School of Economics, 2014.
Alexander Borgida and Peter F Patel-Schneider. A semantics and complete algorithm for
subsumption in the classic description logic. Journal of Artificial Intelligence Research,
1:277–308, 1994.
Alexander Borgida, Ronald J. Brachman, Deborah L. McGuinness, and Lori Alperin
Resnick. CLASSIC: A structural data model for objects. In Proceedings of the 1989
ACM SIGMOD International Conference on Management of Data, pages 58–67, 1989.
Ronald J. Brachman and James G. Schmolze. An overview of the KL-ONE knowledge
representation system. Cognitive Science, 9(2):171–216, 1985.
Lorenz Bühmann, Daniel Fleischhacker, Jens Lehmann, André Melo, and Johanna Völker.
Inductive lexical learning of class expressions. In Knowledge Engineering and Knowledge
Management - 19th International Conference (EKAW), pages 42–53, 2014.
Diego Calvanese, Giuseppe De Giacomo, Domenico Lembo, Maurizio Lenzerini, and Riccardo Rosati. Tractable reasoning and efficient query answering in description logics: The
DL-Lite family. Journal of Automated reasoning, 39(3):385–429, 2007.
Philipp Cimiano, Johanna Völker, and Paul Buitelaar. Ontology construction. In Handbook of Natural Language Processing, Second Edition., pages 577–604. Chapman and
Hall/CRC, 2010.
William W Cohen and Haym Hirsh. The learnability of description logics with equality
constraints. Machine Learning, 17(2-3):169–199, 1994a.
William W. Cohen and Haym Hirsh. Learning the CLASSIC description logic: Theoretical and experimental results. In Principles of Knowledge Representation and Reasoning
(KR), pages 121–133, 1994b.
Felix Distel. Learning description logic knowledge bases from data using methods from formal
concept analysis. PhD thesis, Dresden University of Technology, 2011.
57
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
Ronald Fagin, Phokion G Kolaitis, Renée J Miller, and Lucian Popa. Data exchange:
semantics and query answering. Theoretical Computer Science, 336(1):89–124, 2005.
Daniel Fleischhacker, Johanna Völker, and Heiner Stuckenschmidt. Mining RDF data for
property axioms. In On the Move to Meaningful Internet Systems: OTM 2012, pages
718–735. Springer, 2012.
Michael Frazier and Leonard Pitt. Learning from entailment: An application to propositional horn sentences. In International Conference on Machine Learning (ICML), pages
120–127, 1993.
Michael Frazier and Leonard Pitt. Classic learning. Machine Learning, 25(2-3):151–193,
1996.
Bernhard Ganter, Sergei A Obiedkov, Sebastian Rudolph, and Gerd Stumme. Conceptual
exploration. Springer, 2016.
Ernesto Jiménez-Ruiz, Evgeny Kharlamov, Dmitriy Zheleznyakov, Ian Horrocks, Christoph
Pinkel, Martin G Skjæveland, Evgenij Thorstensen, and Jose Mora. BootOX: practical
mapping of RDBs to OWL 2. In International Semantic Web Conference, (ISWC), pages
113–132. 2015.
Stanislav Kikot, Roman Kontchakov, and Michael Zakharyaschev. On (in)tractability of
OBDA with OWL 2 QL. In Proceedings of the 24th International Workshop on Description Logics (DL 2011), 2011.
Szymon Klarman and Katarina Britz. Ontology learning from interpretations in lightweight
description logics. In Inductive Logic Programming, 2015.
Boris Konev, Michel Ludwig, Dirk Walther, and Frank Wolter. The logical difference for
the lightweight description logic EL. Journal of Artificial Intelligence Research (JAIR),
44:633–708, 2012.
Boris Konev, Carsten Lutz, and Frank Wolter. Exact learning of Tboxes in EL and DL-Lite.
In Informal Proceedings of the 26th International Workshop on Description Logics, pages
341–352, 2013.
Boris Konev, Carsten Lutz, Ana Ozaki, and Frank Wolter. Exact learning of lightweight
description logic ontologies. In Principles of Knowledge Representation and Reasoning
(KR), 2014.
Boris Konev, Ana Ozaki, and Frank Wolter. A model for learning description logic ontologies
based on exact learning. In Conference on Artificial Intelligence (AAAI), pages 1008–
1015, 2016.
Roman Kontchakov and Michael Zakharyaschev. An introduction to description logics
and query rewriting. In Reasoning Web. Semantic Technologies for Advanced Query
Answering - 10th International Summer School, pages 195–244, 2014.
58
Exact Learning of Lightweight Description Logic Ontologies
Markus Krötzsch. OWL 2 profiles: An introduction to lightweight ontology languages. In
Reasoning Web. Semantic Technologies for Advanced Query Answering - 8th International Summer School, pages 112–183, 2012.
Jens Lehmann and Christoph Haase. Ideal downward refinement in the\ mathcal {EL}
description logic. In International Conference on Inductive Logic Programming (ILP),
pages 73–87, 2009.
Jens Lehmann and Pascal Hitzler. Concept learning in description logics using refinement
operators. Machine Learning, 78(1-2):203–250, 2010.
Jens Lehmann and Johanna Völker. Perspectives on Ontology Learning, volume 18. IOS
Press, 2014.
Francesca A. Lisi. Al-quin: An onto-relational learning system for semantic web mining.
International Journal on Semantic Web and Information Systems, 7(3):1–22, 2011.
Francesca A. Lisi and Umberto Straccia. Learning in description logics with fuzzy concrete
domains. Fundamenta Informaticae, 140(3-4):373–391, 2015.
Carsten Lutz, Robert Piro, and Frank Wolter. Description logic TBoxes: Model-theoretic
characterizations and rewritability. In International Joint Conference on Artificial Intelligence (IJCAI), pages 983–988, 2011.
Yue Ma and Felix Distel. Learning formal definitions for snomed ct from text. In Artificial
Intelligence in Medicine Conference (AIME), pages 73–77, 2013.
Ana Ozaki. Exact Learning of Description Logic Ontologies. PhD thesis, University of
Liverpool, 2016.
Charles David Page Jr. Anti-unification in constraint logics: foundations and applications
to learnability in first-order logic, to speed-up learning, and to deduction. PhD thesis,
University of Illinois at Urbana-Champaign, 1993.
Peter F. Patel-Schneider, Deborah L. McGuinness, and Alexander Borgida. The CLASSIC knowledge representation system: Guiding principles and implementation rationale.
SIGART Bulletin, 2(3):108–113, 1991.
Antonella Poggi, Domenico Lembo, Diego Calvanese, Giuseppe De Giacomo, Maurizio Lenzerini, and Riccardo Rosati. Linking data to ontologies. J. Data Semantics, 10:133–173,
2008.
Chandra Reddy and Prasad Tadepalli. Learning first-order acyclic Horn programs from
entailment. Inductive Logic Programming, pages 23–37, 1998.
Sebastian Rudolph. Exploring relational structures via FLE. In International Conference
on Conceptual Structures, ICCS, pages 196–212, 2004.
Stefan Schlobach, Zhisheng Huang, Ronald Cornet, and Frank Van Harmelen. Debugging
incoherent terminologies. Journal of Automated Reasoning, 39(3):317–349, 2007.
59
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
Joseph Selman and Alan Fern. Learning first-order definite theories via object-based queries.
In European Conference on Machine Learning and Principles and Practice of Knowledge
Discovery in Databases (ECML/PKDD), pages 159–174, 2011.
Heiner Stuckenschmidt, Christine Parent, and Stefano Spaccapietra, editors. Modular Ontologies: Concepts, Theories and Techniques for Knowledge Modularization, volume 5445
of Lecture Notes in Computer Science. Springer, 2009.
Balder ten Cate, Vı́ctor Dalmau, and Phokion G. Kolaitis. Learning schema mappings. In
International Conference on Database Theory (ICDT), pages 182–195, 2012.
Johanna Völker and Mathias Niepert. Statistical schema induction. In The Semantic Web:
Research and Applications, pages 124–138. Springer, 2011.
Johanna Völker, Daniel Fleischhacker, and Heiner Stuckenschmidt. Automatic acquisition
of class disjointness. Journal of Web Semantics, 35:124–139, 2015.
Hai Wang, Matthew Horridge, Alan L. Rector, Nick Drummond, and Julian Seidenberg.
Debugging OWL-DL ontologies: A heuristic approach. In 4th International Semantic
Web Conference ISWC, pages 745–757, 2005.
Wilson Wong, Wei Liu, and Mohammed Bennamoun. Ontology learning from text: A look
back and into the future. ACM Computing Surveys, 44(4):20:1–20:36, 2012.
60
Exact Learning of Lightweight Description Logic Ontologies
Appendix A. Proofs for Section 6
We supply proofs for Lemma 62 and Lemma 64. In addition, we prove a claim used in the
proof of Lemma 67. We start by giving the proof of Lemma 62.
Lemma 62 For every DL-Lite∃R CI B ⊑ D over the signature of Tσ ,
• either Tσ |= B ⊑ D for every Tσ ∈ SN
• or there is at most one Tσ ∈ SN such that Tσ |= B ⊑ D.
Proof Assume the CI B ⊑ D is given. If B 6= A or M does not occur in D, then the claim
can be readily checked. Thus, we assume that B = A and M occurs in D. Assume there
exists σ0 such that Tσ0 |= A ⊑ D (if no such σ0 exists, we are done). For any σ, let IA,Tσ
be the canonical model of A and Tσ (Lemma 12). Apply the following restricted form of
parent/child merging exhaustively to the concept expression D:
• if there are nodes d, d1 , d2 ∈ TD with l(d1 , d) = σ and l(d, d2 ) = σ − for some σ ∈ {r, s},
then replace D by the resulting concept expression after d1 and d2 are merged in D.
Let D ′ be the resulting concept expression. Recall from Lemma 12 that Tσ |= A ⊑ D iff
there is a homomorphism from TD to IA,Tσ mapping ρD to ρA . Using the fact that IA,Tσ is
a ditree interpretation, one can readily check that any homomorphism h from TD to IA,Tσ
mapping ρD to ρA factors through TD′ and that D ′ is an EL concept expression. Thus, if
there is an additional σ′ 6= σ0 such that Tσ′ |= A ⊑ D, then there are two homomorphisms
hσ0 and hσ′ with the same domain TD′ into IA,Tσ0 and IA,Tσ′ and mapping the root of TD′
to the roots of IA,Tσ0 and IA,Tσ′ , respectively. Since M occurs in D ′ and D ′ is an EL concept
expression we find a sequence D0 , . . . , Dm with D0 = D ′ and Dm = M such that ∃si+1 .Di+1
is a top-level conjunct of Di for si ∈ {r, s} and all i < m. But then s1 · · · sm = σ0 and
s1 · · · sm = σ′ and we have derived a contradiction to the assumption that σ0 and σ′ are
distinct.
❏
To prove Lemma 64 we require the following observation.
Lemma 72 For any acyclic EL TBox T , any CI A ⊑ C ∈ T and any concept expression
of the form ∃t.D we have T |= A ⊑ ∃t.D if, and only if, T |= C ⊑ ∃t.D.
We are now ready to prove Lemma 64.
Lemma 64 For every EL CI C ⊑ D over Σn :
• either TL |= C ⊑ D for every L ∈ Ln
• or the number of L ∈ Ln such that TL |= C ⊑ D does not exceed |C|.
Proof We prove the lemma by induction on the structure of D. We assume throughout
the proof that there exists some L0 ∈ Ln such that TL0 |= C ⊑ D.
Base case: D is a concept name. We make the following case distinction.
• D ∈ {Xi , Ai , Bi | 1 ≤ i ≤ n} or D = M . By Lemma 65, C is of the form Z ⊓ C ′ ,
for some concept name Z, and TL0 |= Z ⊑ D. But then Z = D and it follows that
TL |= C ⊑ D for every L ∈ Ln .
61
Boris Konev, Carsten Lutz, Ana Ozaki and Frank Wolter
• D = X0 . By Lemma 65, C is of the form Z ⊓ C ′ , for some concept name Z, and
TL0 |= Z ⊑ X0 . This is the case if either Z = X0 , or Z ∈ {A, A1 , B1 , . . . , An , Bn }. In
either case, TL |= C ⊑ X0 for every L ∈ Ln .
• D = A. If C is of the form A ⊓ C ′ or for all i such that 1 ≤ i ≤ n, Ai or Bi is a
conjunct of C, then TL |= C ⊑ A for every L ∈ Ln . Assume now that C is not of this
form. Then for some j such that 1 ≤ j ≤ n, C is neither of the form A ⊓ C ′ nor of
the form Aj ⊓ C ′ nor of the form Bj ⊓ C ′ . Let L = (σ1 , . . . , σn ) ∈ Ln be such that
TL |= C ⊑ A. Notice that TL |= C ⊑ A, for L = (σ1 , . . . , σn ) ∈ Ln , if, and only if,
TL |= C ⊑ X0 ⊓∃σ1 .M ⊓· · ·⊓∃σn .M . By the claim in the proof of Lemma 66, for such
a TL we must have ∅ |= C ⊑ ∃σj .M . Clearly, the number of L = (σ1 , . . . , σn ) ∈ Ln
with ∅ |= C ⊑ ∃σj .M does not exceed |C|.
Thus, either TL |= C ⊑ A for every L ∈ Ln or the number of L ∈ Ln such that
TL |= C ⊑ A does not exceed |C|.
Induction step. If D = D1 ⊓ D2 , then TL |= C ⊑ D if, and only if, T |= C ⊑ Di , i = 1, 2.
By induction hypothesis, for i = 1, 2 either TL |= C ⊑ Di for every L ∈ Ln or there exist
at most |C| different L ∈ Ln such that TL |= C ⊑ Di . Thus either TL |= C ⊑ D for every
L ∈ Ln or the number of L ∈ Ln such that TL |= C ⊑ D also does not exceed |C|.
Now assume that D = ∃t.D ′ . Suppose that TL |= C ⊑ D for some L ∈ Ln . Then, by
Lemma 65, either there exists a conjunct Z of C, Z a concept name, such that TL |= Z ⊑
∃t.D ′ or there exists a conjunct ∃t.C ′ of C with TL |= C ′ ⊑ D ′ . We analyse for every
conjunct of C of the form Z or ∃t.C ′ the number of L ∈ Ln such that TL |= Z ⊑ ∃t.D ′ or
TL |= ∃t.C ′ ⊑ ∃t.D ′ , respectively.
(i) Let Z be a conjunct of C such that Z is a concept name and TL |= Z ⊑ ∃t.D ′ . Notice
that Z 6= M as there is no L ∈ Ln such that TL |= M ⊑ ∃t.D ′ . We consider the
remaining cases.
– Z = Xi , for some i ≥ 0. It is easy to see that for L, L′ ∈ Ln we have TL |= Xi ⊑
∃t.D ′ if, and only if TL′ |= Xi ⊑ ∃t.D ′ . Thus, TL |= Z ⊑ ∃t.D ′ for every L ∈ Ln .
– Z ∈ {Ai , Bi | 1 ≤ i ≤ n}. By Lemma 72, TL |= Z ⊑ ∃t.D ′ if, and only
if, TL |= X0 ⊓ ∃σi .M ⊑ ∃t.D ′ . By Lemma 65, either TL |= X0 ⊑ ∃t.D ′ or
TL |= ∃σi .M ⊑ ∃t.D ′ . If TL |= X0 ⊑ ∃t.D ′ then, as above, for TL |= C ⊑ ∃t.D ′
every L ∈ Ln . Suppose that ∃t.D ′ is such that TL 6|= X0 ⊑ ∃t.D ′ and TL |=
∃σi .M ⊑ ∃t.D ′ . By inductive applications of Lemma 65, this is only possible if
∃t.D ′ = ∃σi .M . Thus, there is exactly one L ∈ Ln (namely, L = L0 ) such that
TL |= Z ⊑ ∃σi .M .
– Z = A. Suppose that for some L = (σ1 , . . . , σn ) ∈ Ln we have TL |= A ⊑ ∃t.D ′ .
Equivalently, TL |= X0 ⊓ ∃σ1 .M ⊓ . . . σn .M ⊑ ∃t.D ′ . By Lemma 65, either
TL |= X0 ⊑ ∃t.D ′ or TL |= ∃σi .M ⊑ ∃t.D ′ for some i with 1 ≤ i ≤ n. Thus, as
above, unless TL |= X0 ⊑ ∃t.D ′ we have ∃t.D ′ is ∃σi .M . But then L = L0 .
(ii) Let ∃t.C ′ be a conjunct of C with TL |= C ′ ⊑ D ′ . The induction hypothesis implies
that the number of L ∈ Ln such that TL |= C ′ ⊑ D ′ does not exceed |C ′ |.
62
Exact Learning of Lightweight Description Logic Ontologies
To summarise, either TL |= C ⊑ ∃t.D ′ for every L ∈ Ln or for every conjunct C0 of C of
the form Z or ∃t.C ′ , the number of L ∈ Ln such that TL |= C0 ⊑ ∃t.D ′ does not exceed
|C0 |. Hence the number of L ∈ Ln such that TL |= C ⊑ ∃t.D ′ does not exceed |C|.
❏
The next result is used in the proof of Lemma 67.
Lemma 73 For any 0 ≤ i ≤ n and concept expression D over Σn , if T0 6|= Xi ⊑ D then
there exists a sequence of role names t1 , . . . tl ∈ {r, s}l such that ∅ |= D ⊑ ∃t1 . · · · ∃tl .Y and
T0 6|= Xi ⊑ ∃t1 . · · · ∃tl .Y , where Y is either ⊤ or a concept name and 0 ≤ l ≤ n − i + 1.
Proof We prove the lemma by induction on i from i = n to i = 0. If i = n, then
T0 6|= Xi ⊑ D if either ∅ |= D ⊑ ∃t.⊤, for some role name t, or ∅ |= D ⊑ Y , for some
concept name Y 6= Xi .
Suppose that the lemma is proved for 0 < j ≤ n and let i = j − 1. We proceed by
induction on the structure of D. If D is a concept name, we are done as T0 |= Xi ⊑ Z does
not hold for any concept name Z 6= Xi . If D is of the form ∃t.D ′ , where t ∈ {r, s}, then
T0 6|= Xi+1 ⊑ D ′ , and so, by induction hypothesis, there exists a sequence of role names
t1 , . . . , tl , with l ≤ n − i, such that T0 6|= Xi+1 ⊑ ∃t1 . · · · ∃tl .Y and ∅ |= D ′ ⊑ ∃t1 . · · · ∃tl .Y .
But then, by Lemma 72 and Lemma 65, T0 6|= Xi ⊑ ∃t.∃t1 . · · · ∃tl .Y and ∅ |= ∃t.D ′ ⊑
∃t.∃t1 . · · · ∃tl .Y . If D is of the form D = D1 ⊓ D2 , there there exists Di , i = 1, 2, such that
T0 6|= Xi ⊑ Di and the lemma holds by induction hypothesis.
❏
63
| 2cs.AI
|
FLIC: Fast Linear Iterative Clustering with Active Search
Jiaxing Zhao1
Bo Ren1∗
arXiv:1612.01810v2 [cs.CV] 21 Nov 2017
1
Qibin Hou1
Nankai University
Abstract
In this paper, we reconsider the clustering problem for image over-segmentation from a new perspective. We propose a
novel search algorithm named “active search” which explicitly considers neighboring continuity. Based on this search
method, we design a back-and-forth traversal strategy and
a “joint” assignment and update step to speed up the algorithm. Compared to earlier works, such as Simple Linear Iterative Clustering (SLIC) and its follow-ups, who use fixed
search regions and perform the assignment and the update
step separately, our novel scheme reduces the number of iterations required for convergence, and also improves the boundary sensitivity of the over-segmentation results. Extensive
evaluations on the Berkeley segmentation benchmark verify
that our method outperforms competing methods under various evaluation metrics. In particular, lowest time cost is reported among existing methods (approximately 30 fps for a
481 × 321 image on a single CPU core). To facilitate the
development of over-segmentation, the code will be publicly
available.
1
Introduction
Superpixels, generated by image over-segmentation, take
the place of pixels to become the fundamental units in
various computer vision tasks, including image segmentation (Cheng et al. 2016), image classification (Wang et al.
2013b), 3D reconstruction (Hoiem, Efros, and Hebert 2005),
object tracking (Wang et al. 2011), etc. Such a technique
can greatly reduce computational complexity, avoid undersegmentation, and reduce the influence caused by noise.
Therefore, how to generate superpixels with high efficiency
plays an important role in many vision and image processing
applications.
Generating superpixels has been an important research
issue, and a group of classical methods have been developed, including FH (Felzenszwalb and Huttenlocher 2004),
Mean Shift (Comaniciu and Meer 2002), Watershed (Vincent and Soille 1991), etc. The lack of compactness and
the irregularity of superpixels restrict their applications, especially when contrast is poor or shadows are present. To
solve the above-mentioned problems, Shi and Malik proposed Normalized Cuts (NC) (Shi and Malik 2000) that generated compact superpixels. However, this method does not
∗
Bo Ren is the corresponding author: [email protected].
Ming-Ming Cheng1
2
Paul L. Rosin2
Cardiff University
adhere to image boundaries very well, and the complexity is high. GraphCut (Boykov, Veksler, and Zabih 2001;
Veksler, Boykov, and Mehrani 2010) regarded the segmentation problem as an energy optimization process. It solved
the compactness problem by using min-cut/max-flow algorithms (Boykov and Kolmogorov 2004; Kolmogorov and
Zabin 2004), but their parameters are hard to control. Turbopixel (Levinshtein et al. 2009) is another method that is
proposed to solve the compactness problem. However, the
inefficiency of the underlying level-set method (Osher and
Sethian 1988) restricts its applications. Bergh et al. (2012)
proposed an energy-driven algorithm SEEDS whose results
adhered to the boundaries well, but unfortunately it suffers
from irregularity and the number of superpixels is uncertain. ERS (Liu et al. 2011), although it performs well on
the Berkeley segmentation benchmark, has a high computational cost that limits its practical use.
Achanta et al. (2012) proposed a linear clustering based
algorithm SLIC. It generates superpixels based on Lloyd’s
algorithm (Lloyd 1982) (also known as Voronoi iteration or
k-means). In the assignment step of SLIC, as a key point to
speed up the algorithm, each pixel p is associated with those
cluster seeds whose search regions overlap its location. Such
a strategy is also adopted by most subsequent works based
on SLIC. SLIC is widely used in various applications (Wang
et al. 2011) because of its high efficiency and good performance. Inspired by SLIC, Wang et al. (2013a) implemented
an algorithm SSS that considered the structural information within images. It uses the geodesic distance (Peyré et
al. 2010) computed by the geometric flows instead of the
simple Euclidean distance. However, efficiency is poor because of the bottleneck caused by the high computational
cost of measuring geodesic distances. Very recently, Liu et
al. proposed Manifold SLIC (2016) that generated contentsensitive superpixels by computing Centroidal Voronoi Tessellation (CVT) (Du, Faber, and Gunzburger 1999) in a special feature space. Such an advanced technique makes it
much faster than SSS but still slower than SLIC owing to
the cost of its mapping, splitting and merging processes.
From the aforementioned descriptions, we see that abovementioned methods improve the results by either using more
complicated distance measurements or providing more suitable transformations of the feature space. However, the assignment and update steps within these methods are per-
formed separately, leading to low convergence rate.
In this paper, we consider the over-segmentation problem
from a new perspective. Each pixel in our algorithm is allowed to actively search which superpixel it should belong
to, according to its neighboring pixels as shown in Figure 1.
In the meantime, the seeds of the superpixels can be adaptively changed during this process, which allows our assignment and update steps to be performed jointly. This property
enables our approach to converge rapidly. To sum up, our
main advantages are:
• Our algorithm features good awareness of neighboringpixel continuity, and produces results with good boundary
sensitivity regardless of image complexity and contrast.
• Our algorithm performs the assignment step and the update step in a joint manner, and has a high convergence
rate as well as the lowest time cost among all superpixel
segmentation approaches. Experiments show that our approach is able to converge in two scan loops, with better
performance measured under a variety of evaluation metrics on the Berkeley segmentation benchmark.
2
Preliminaries
Before introducing our approach that allows adaptive search
regions and joint assignment and update steps, we first
briefly recap the standard previous scheme with fixed search
regions and separate steps. A typical one is the SLIC algorithm who improves Lloyd’s algorithm, reducing the time
complexity from O(KN ) to O(N ), where K is the number
of the superpixels and N is the number of pixels.
Let {Ii }N
i=1 be a color image, where Ii represents the corresponding variable of each pixel. Given a set of evenly distributed seeds {Sk }K
k=1 , SLIC simplifies the Lloyd’s algorithm to get the Centroidal Voronoi Tessellation (CVT) (Du,
Faber, and Gunzburger 1999) that will be introduced in Section 3.4. In the assignment step, each pixel Ii is associated
with those cluster seeds whose search regions overlap its
location as shown in Figure 1(a). The area of a search
p region can be denoted by 2T × 2T , where T =
N/K.
Specifically, SLIC considers Ii to lie in a five dimensional
space that contains a three dimensional CIELAB color space
(li , ai , bi ) and a two dimensional spatial space (xi , yi ). SLIC
measures the distance between two points using a weighted
Euclidean distance, which can be computed by
s
2
ds ∗ m
2
D(Ii , Ij ) = dc +
,
(1)
Ns
p
s
s
p
p
p
s
s: seed
p: pixel
(a) SLIC
s
s: seed
p: pixel
(b) Active search
Figure 1: (a) The search method used in SLIC. Each seed
only searches a limited region to reduce computation complexity. (b) Our proposed active search. Each pixel is able to
decide its own label by searching its surroundings.
obtains the over-segmentation results by iteratively performing the assignment and update steps.
The follow-up works of SLIC also use a similar procedure as SLIC. They improve the performance of SLIC using better distance measures or more suitable transformation function between color space and spatial space. However, in these algorithms, each search region is fixed in the
assignment step of a single loop, and the relationship among
neighboring pixels is largely ignored when allocating pixels
to superpixels. Separately performing the assignment step
and the update step also leads to a delayed feedback of pixel
label change.
3
The Proposed Approach
Since superpixels normally serve as the first step of other vision related applications, how to generate superpixels with
good boundaries and very fast speed is a crucial problem. Here, unlike previous algorithms (Achanta et al. 2012;
Liu et al. 2016), we consider this problem from a new aspect,
in which only the surrounding pixels are considered for determining the label of the current pixel. Each pixel actively
selects which superpixel it should belong to in a back-andforth order to provide better estimation of over-segmentation
regions. Moreover, the assignment step and the update step
are performed jointly. Very few iterations are required for
our approach to reach convergence. An overview of our algorithm can be found in Alg. 1.
3.1
Problem Setup
(3)
Given the desired number of superpixels K and an input image I = {Ii }N
i=1 , where N is the number of pixels, our goal
is to produce a series of disjoint small regions (or superpixels). Following most previous works (Achanta et al. 2012),
the original RGB color space is transformed to the CIELAB
color space (which has been proven useful). Thus, each pixel
Ii in an image I can be represented in a five dimensional
space,
Ii = (li , ai , bi , xi , yi ).
(4)
In the update step, SLIC recomputes the center of each superpixel and moves the seeds to these new centers. Then it
We first divide the original image into a regular grid
p containing K elements {Gk }K
with
step
length
υ
=
N/K as
k=1
where m is a variable that controls the weight of the spatial
term, and Ns = T . Variables ds and dc are respectively the
spatial and color distances, which can be expressed as
q
ds = (xi − xj )2 + (yi − yj )2 ,
(2)
and
dc =
q
(li − lj )2 + (ai − aj )2 + (bi − bj )2 .
Algorithm 1 FLIC
Require: Image I with N pixels, the desired number of superpixels K, the maximal iteration numbers itrmax and
the spatial distance weight m.
Ensure: K superpixels
K
Divide the original image
pinto regular grids {Gk }k=1
with step length υ = N/K.
Initialize labels {Lk }K
k=1 for pixels according to their
locations.
Move each seed to the lowest gradient position in 3×3
neighborhoods.
Initialize seeds {Sk }K
k=1 .
Regard pixels sharing the same label as a superpixel ζ.
Initialize distance d(i) = ∞ for each pixel and itr = 0.
while itr < itrmax do
for each superpixel ζk do
Use back-and-forth scan to traverse superpixel ζk
to get the pixels processing sequence (Section
3.3).
for each pixel Ii in the sequence do
Set d(i) = D(Ii , SLi ) by Eqn. 1
for Ij in the four-neighborhood of Ii do
if Lj 6= Li then
Compute D = D(Ii , SLj ).(Eqn. 1)
if D < d(i) then
d(i) = D; Li = Lj .
end if
end if
end for
if Li is changed to Lj then
Use Eqn. 10 to update ζLi ;
Use Eqn. 11 to update ζLj ;
Update the bounding box of ζLj .(Section
3.4)
end if
end for
end for
itr++;
end while
in (Achanta et al. 2012), and the initial label for each pixel
Ii is assigned as:
Li = k, if Ii ∈ Gk .
(5)
We initialize the seed Sk in Gk as the centroid. Therefore,
Sk can also be defined in the same five dimensional space
Sk = {lk , ak , bk , xk , yk }.
3.2
(6)
Label Decision
In most natural images adjacent pixels tend to share the same
labels, i.e. neighboring pixels have natural continuity. Thus,
we propose an active search method that is able to leverage as much of this a priori information as possible. In our
method, unlike most previous works (Achanta et al. 2012;
Liu et al. 2016), the label of the current pixel is only determined by its neighbors. We will compute the distances
between the current pixel and the seeds of its four or eight
adjacent pixels. Figure 1 provides a more intuitive illustration. Specifically, for a pixel Ii , our assignment principle is
Li = argmin D(Ii , SLj ), Ij ∈ Ai ,
(7)
Lj
where Ai consists of Ii and its four neighboring pixels, SLj
is Ij ’s corresponding superpixel seed. We use Eqn. 1 to measure the distance D(Ii , SLj ).
Since each pixel can only be assigned to a superpixel containing at least one of its neighbors, the local pixel continuity has a stronger effect in the proposed strategy, allowing
each pixel to actively assign itself to one of its surrounding closely connected superpixel regions. The advantages
of such a strategy are obvious. First, the nearby assignment
principle can avoid the occurrence of too many isolated regions, indirectly preserving the desired number of superpixels. Second, such an assignment operation is not limited by
a fixed range in space, resulting in better boundary adherence despite some irregular shapes of those superpixels with
very complicated content. Furthermore, in the assignment
process, the superpixel centers are also self-adaptively modified, leading to faster convergence. Detailed demonstration
and analysis can be found in Section 4.3. It is worth mentioning that the neighbors of the internal pixels in a superpixels normally share the same labels, so it is unnecessary to
process them any more. This fact allows us to process each
superpixel extremely quickly.
3.3
Traversal Order
The traversal order plays a very important role in our approach in that an appropriate scanning order may lead to
a visually better segmentation. As demonstrated in Section
3.2, the label of each pixel only depends on the seeds of its
surrounding pixels. This indicates that, in a superpixel, the
label of the current pixel is directly or indirectly related to
those pixels that have already been dealt with. To better take
advantage of this avalanche effect, we adopt a back-andforth traversal order as in PatchMatch (Barnes et al. 2009),
in which the pixels that are processed later will benefit from
the previously processed pixels. Figure 2 makes this process
clear. In the forward pass, the label decision of each pixel
considers the information from the top surrounding pixels of
the superpixel, and similarly, the backward pass will provide
the information from the bottom surrounding pixels of the
superpixel. With such a scanning order, all the surrounding
information can be taken into consideration, yielding better
segments.
Considering that an arbitrary superpixel might have an irregular shape instead of a simple rectangle or square, we actually use a simplified strategy to traverse the whole superpixel. For any superpixel, we first find a minimum bounding
box within which all its pixels are enclosed, as shown in Figure 2. We then perform the scanning process for all the pixels
in the corresponding minimum bounding box and only deal
with those pixels that are within the superpixel.
3.4
Joint assignment and update step
A common phenomenon in existing methods, such as SLIC
(Achanta et al. 2012), is that the assignment step and the up-
(a) A region and its bounding box
(b) Forward scan order
(c) Backward scan order
(d) Bounding box after updating
Figure 2: Illustration of our scanning order for each superpixel. We use gray regions enclosed by blue lines to represent superpixels and use red dashed rectangles to denote their corresponding bounding boxes. As shown in (b) and (c), we first scan the
bounding box from left to right and top to bottom and then in the opposite direction. The shape of each superpixel might change,
and so we update the bounding box if this occurs, as in (d), and leave it unchanged if there are no changes to the superpixel
shape.
date step are performed separately, leading to delayed feedback from pixel label changes to superpixel seeds. An obvious problem of such a strategy is that many (normally
more than five) iterations are required which becomes the
bottleneck of fast convergence. In our approach, based on
the assignment principle Eqn. 7, we design a “joint” assignment and update strategy which operate these two steps at
a finer granularity. The approximately joint step is able to
adjust the superpixel seed center position on the fly, drastically reducing the number of iterations needed for convergence. Since most clustering-based superpixel methods use
the Centroidal Voronoi Tessellation (CVT), we will briefly
introduce the CVT first and then describe our method.
Let S = {Sk }K
k=1 be the set of seeds in the image, where
K is the expected number of superpixels. The Voronoi cell
V(Sk ) of a seed Sk is denoted by:
VSk = {Ii ∈ I | d(Ii , Sk ) ≤ d(Ii , Sj ), ∀Sj ∈ S},
(8)
where d(Ii , Sk ) is an arbitrary distance measure from pixel
Ii to the seed Sk . The Voronoi Diagram VD (S) is defined by
VD (S) = {VSk 6= φ | ∀Sk ∈ S}.
(9)
A CVT is then defined as a Voronoi Diagram whose generator point of each Voronoi cell is also its center of mass. As
mentioned above, the traditional CVT is usually obtained by
heuristic algorithms, such as Lloyd’s algorithm, iteratively
performing updates after each assignment step until convergence is reached.
In our approach, on account of our novel label decision
strategy as shown in Eqn. 7, we are able to jointly perform
the update step and the assignment step instead of separately.
More specifically, after pixel Ii is processed, if its label is
changed to, for instance, Lj , we immediately update the current seed SLi using the following equation:
SLi =
SLi ∗ |ζLi | − Ii
,
|ζLi | − 1
update SLj using the following equation
SLj =
where |ζLi | is the number of pixels in superpixel ζLi , and
(11)
The bounding box of ζLj is also updated thereafter.
It is noteworthy to mention that the above updates only
contain very simple arithmetic operations and hence can be
performed very efficiently. Such an immediate update will
help later pixels make a better choice during assignment,
leading to better convergence. Figure 6 shows the convergence speed of our approach.
4
Experiments
Our method is implemented in C++ and runs on a PC
with an Intel Core i7-4790K CPU with 4.0GHz, 32GB
RAM, and 64 bit operating system. We compare our
method with many previous and current state-of-the-art
works, including FH (Felzenszwalb and Huttenlocher 2004),
SLIC (Achanta et al. 2012), Manifold SLIC (Liu et al. 2016),
SEEDS (Van den Bergh et al. 2012), and ERS (Liu et al.
2011) on the BSDS500 benchmark, using the evaluation
methods proposed in (Arbelaez et al. 2011; Stutz, Hermans,
and Leibe 2014). Note that the source codes used in evaluation of the above works may be of different versions, and
we find this leads to performance difference from the original reports when a different implementation of evaluation
code is applied. To give a fair comparison, we uniformly
use publicly available source code (Arbelaez et al. 2011;
Stutz, Hermans, and Leibe 2014) for all the methods. As
in previous researches in the literature (Liu et al. 2016;
Wang et al. 2013a), we evaluate all algorithms on 200 randomly selected images of resolution 481 × 321 from the
Berkeley dataset.
4.1
(10)
SLj ∗ ζLj + Ii
.
ζL j + 1
Parameters
In our approach, three parameters need to be set. The first
one is the number of superpixels K. One of the common advantages of clustering-based algorithms is that the expected
Ours
ERS
SEEDS
FH
SLIC
M-SLIC
0.8
0.75
0.7
0.65
0.6
100
200
300
400
500
Undersegment error
Boundary recall
0.9
0.85
0.15
0.1
0.05
600
Number of superpixels
0.2
(a) BR curves
0
0.2
0.4
Time(seconds)
0.6
0.8
(b) UE trade-off
1
0.9
0.95
0.85
Ours
ERS
SEEDS
FH
SLIC
M-SLIC
0.9
0.85
0.8
0
0.2
0.4
Time(seconds)
0.6
(c) ASA trade-off
0.8
Boundary Recall
Ours
ERS
SEEDS
FH
SLIC
M-SLIC
Achievable segmentation accuracy
0.25
1
0.95
Ours
ERS
SEEDS
FH
SLIC
M-SLIC
0.8
0.75
0.7
0
0.2
0.4
Time(seconds)
0.6
0.8
(d) BR trade-off
Figure 3: Comparisons between existing state-of-the-art methods and our approach (FLIC) on the BSDS500 benchmark. In (b)(d), K is fixed to 200 to demonstrate our best trade-off between performance and efficiency between competing methods. As
can be seen, our strategy significantly outperforms methods that have similar time cost in boundary recall. At least competitive
results are also achieved compared to slower methods (e.g. the state-of-the-art method ERS (Liu et al. 2011)) in all the evaluation
metrics, but at an order of magnitude faster speed.
0.85
itr = 4 (Normal scan-line)
itr = 1 (Forth-and-back)
itr = 2 (Forth-and-back)
0.8
0.75
Time(seconds)
Boundary recall
0.9
0.7
100
100
200
300
400
BR Time BR Time BR Time BR Time
4-N 78.6 34 85.9 35 89.1 36 91.8 38
8-N 80.5 54 87.4 56 90.5 59 92.7 61
0.08
1
0.95
0.06
0.04
0.02
itr = 3 (Forth-and-back)
200
300
400
500
Number of superpixels
(a) BR
600
0
100
200
300
400
500
Number of superpixels
600
(b) Time Cost
Figure 4: Part of sensitivity analysis under the standard evaluation metrics and time cost.
number of superpixels can be directly obtained by setting
the clustering parameter K. The second one is the spatial
distance weight m. Parameter m has a large effect on the
smoothness and compactness of superpixels. We shall show
that our performance will increase as m decreases. However,
a small m can also lead to the irregularity of superpixels. To
achieve a good trade-off between compactness and performance, in the following experiments, we set m = 5 as default. The last parameter is the number of iterations itr. Here
we set itr = 2 in default to get the balance between time
cost and performance. What should be stressed out here is
that to compare with other methods in a fair way, for each
method we optimize its parameters to maximize the recall
value computed on the BSDS500 benchmark.
4.2
Table 1: Boundary recall and time cost comparisons between
4-neighborhood and 8-neighborhood with different superpixel counts: 100, 200, 300, 400.
Comparison with Existing Methods
Our approach outperforms previous methods that have similar computational efficiency , and achieve at least comparable results compared to slower algorithms with an order of
magnitude faster speed. Details are discussed below.
Boundary Recall (BR). Boundary recall is a measurement which denotes the adherence to the boundaries. It computes what fraction of the ground truth edges falls within εpixel length from at least one superpixel boundary. The BR
(Achanta et al. 2012) can be computed by
P
p∈ξG Π(minq∈ξS kp − qk < ε)
,
(12)
BRG (S) =
|ξG |
where ξS and ξG respectively denote the union set of superpixel boundaries and the union set of ground truth boundaries. The indicator function Π checks if the nearest pixel
is within ε distance. Here we follow (Achanta et al. 2012;
Liu et al. 2016) and set ε = 2 in our experiment. The boundary recall curves of different methods are plotted in Figure
3(a). One can easily observe that our FLIC method outperforms all other methods.
Undersegment Error (UE). The undersegment error reflects the extent that superpixels do not exactly overlap the
ground truth segmentation. Similar to BR, UE can also reflect the boundary adherence. The difference is that UE uses
segmentation regions instead of boundaries in the measurement. Mathematically, the UE (Neubert and Protzel 2012)
can be computed by
P
P
G∈%G (
S:S∩G6=φ min(Sin , Sout ))
, (13)
U EG (S) =
N
where %S is the union set of superpixels, %G is the union set
of the segments of the ground truth, Sin denotes the overlapping of the superpixel S and the ground truth segment
G, and Sout denotes the rest of the superpixel S. As shown
in Figure 3(b), our results are nearly the same as those of
the best approach ERS(Liu et al. 2011) and run significantly
faster.
Achievable Segmentation Accuracy (ASA). ASA gives
the highest accuracy achievable for object segmentation that
utilizes superpixels as units. Similar to UE, ASA utilizes
segments instead of the boundaries, which can be computed
by (Liu et al. 2011)
P
maxi |Sk ∩ Gi |
ASAG (S) = k P
,
(14)
i Gi
0.9
Ours
SLIC
0.85
Boundary recall
Boundary recall
0.9
0.8
0.75
0.7
0
5
10
15
Spatial distance weight m
0.85
Ours
SLIC
0.8
0.75
0.7
0
5
(a)
Iteration
10
15
(b)
4.3
Figure 5: (a) The BR-m curves, where m is the spatial distance weight in Eqn. 1. Our overall performance is far better
than SLIC for all the tested m. (b) The BR-Iteration curves.
Our method converges within 2 iterations, which is much
faster than SLIC.
Boundary recall
0.9
Jointly
Separately
0.88
0.86
0.84
0.82
0.8
0
5
Iteration
10
method, our approach adheres to boundaries very well and
runs twice as fast. Compared to ERS method, our resulting
superpixels are much more regular and the mean execution
time of our approach is 20 times shorter.
All the aforementioned facts and Figure 3 reflect that our
approach achieves an excellent compromise among adherence, compactness, and time cost.
15
Figure 6: Comparison of the convergence rate between joint
and separate assignment and update steps.
where Sk represents the superpixel and Gi represents the
ground truth segment. A better superpixel segmentation will
have a larger ASA value. As shown in Figure 3(c), compared to the ERS (Liu et al. 2011), the performance of our
approach is competitive and our method achieves the best
trade-off between the performance and time cost.
Time Cost (TC). Similar to SLIC, our method also
achieves an O(N ) time complexity. We know that computation efficiency is one of the most important points for using superpixels as elementary units. Many approaches are
limited by their speeds, such as SSS (Wang et al. 2013a)
and ERS (Liu et al. 2011). As shown in Figure 3, the average time cost of FLIC with two iterations when processing
an image is 0.035s, while the time costs of ERS, Manifold
SLIC, SLIC, and FH are 0.625, 0.281s, 0.072s, and 0.047s,
respectively. It is obvious that FLIC has the lowest time cost
among all methods and it runs nearly 20 times faster than
ERS with comparable result quality.
Visual Results and Analysis. In Figure 7, we show several superpixel segmentation results using different algorithms. As can be seen, our approach is more sensitive to
image boundaries, especially when there is poor contrast between the foreground and background. Compared to SLIC
Algorithm Analysis
Efficacy of the back-and-forth Traverse Order. As shown
in Figure 2, we adopt a back-and-forth traverse order to scan
the whole region enclosed by a bounding box for each superpixel. Actually, a couple of forward scans can also perform
very well for our method. We provide a comparison between
two strategies: using pure forward scan order for four iterations versus using the proposed back-and-forth scan order
twice (which is also four iterations). Figure 4 shows quantitative comparisons between these two strategies. The blue
line represents the results using normal forward scan order
while the red line stands for the results using our method. It
can be seen that the red curve significantly outperforms the
blue one and achieves competitive time cost compared to the
blue curve. This fact reflects that our back-and-forth scan order considers more information about the regions outside the
bounding box, leading to more reliable boundaries.
The Role of the Spatial Distance Weight. As shown in
Figure 5(a), unlike SLIC (Achanta et al. 2012), the BR curve
with respect to the spatial distance weight m is monotonically decreasing in our approach. The reason for this phenomenon is that in our method local region continuity is
mostly ensured by the active search algorithm, and color
boundaries are less well preserved for larger m. On the other
hand, small m will result in less regular superpixels, so we
choose m = 5 for our comparison with previous works. It
is noteworthy to mention that superpixels are normally considered as the first step of most vision tasks and these vision tasks often favor those superpixel methods with good
boundaries. Therefore, users can select a reasonable value
for m according to their specific conditions. In any case, our
overall performance is significantly better for all m values.
Convergence Rate. FLIC significantly accelerates the
evolution so that we only need a few iterations before convergence. We compare the performance curves with different iterations on the Berkeley benchmark. It can be easily
found in Figure 5(b) that our algorithm quickly converges
within only two iterations and more iterations only bring
marginal benefits to the results. Numerically, the boundary
recall of the superpixels with only one iteration is 0.835
when K is set to 200. The value after two iterations is 0.859
and after three iterations it is 0.860 when generating the
same number of superpixels. The undersegment error values are 0.115, 0.108, and 0.107, respectively. The achievable
segmentation accuracy values are 0.941, 0.945, and 0.946,
respectively. As can be seen in Figure 5(b), our algorithm not
only converges much faster than SLIC (which requires ten
iterations to converge), but also obtains better performance.
The Role of the Joint Assignment and Update. Our algorithm jointly performs the assignment and update steps.
In Figure 6, we show the convergence rates of both our
(a) Ours
(b) FH
(c) SLIC
(d) M-SLIC
(e) SEEDS
(f) ERS
Figure 7: Visual comparison of superpixel segmentation results using different existing algorithm with 100 superpixels and
m = 10. Our approach adheres to boundaries very well and at the same time produces compact superpixels.
Figure 8: Images segmented by our proposed approach with m = 20 and the number of superpixels set to 1000, 400, and 200,
respectively. The resulting superpixels adhere to region boundaries very well.
Figure 9: Images segmented by our proposed approach with m = 10, 20, and 30, respectively. When m tends to be a smaller
value, the superpixels adhere well to boundaries. When m becomes larger, the superpixels become more compact.
joint approach and that of separately performing assignment
and update steps. One can observe that our joint approach
converges very quickly and only two iterations are needed,
while the separate approach needs another two iterations to
reach the same BR value. This phenomenon demonstrates
that our joint approach is efficient and without any negative
effect on our final results.
Effect on the Size of Neighborhoods. As mentioned in
Section 3, in our implementation, the label of the current
pixel relies on its four neighborhood pixels. Actually, using
eight neighborhood pixels is also reasonable as more neighbors will definitely provide more useful information. In Table 1, we briefly compare the results for these two cases. A
natural observation is that using larger neighborhoods leads
to an increase in performance but at the cost of reducing running speed. With regard to real applications, users can select
either case to suit their own preferences.
Qualitative Results. In Figure 8 we show some segmentation results produced by our approach with m = 20 and
the number of superpixels set to 1000, 400, and 200, respectively. It is seen that, over the range of K value, the edges of
the resulting superpixels are always very close to the boundaries. This phenomenon is especially obvious in the first image and the third image. We also show some segmentation
results with different values of m in Figure 9. When m tends
to smaller values, for example 10, the shapes of the resulting
superpixels become less regular. When m is larger, for example 30, the resulting superpixels become more compact.
5
Conclusions
In this paper we present a novel algorithm using active
search, which is able to improve the performance and significantly reduce the time cost for using superpixels to oversegment an image. Taking advantage of local continuity, our
algorithm provides results with good boundary sensitivity
regardless of image contrast and complexity. Moreover, it is
able to converge in only two iterations, achieving the lowest time cost compared to previous methods while obtain-
ing performance comparable to the state-of-the-art method
ERS with 1/20th of its running time. We have used various
evaluation metrics on the Berkeley segmentation benchmark
dataset to demonstrate the high efficiency and high performance of our approach.
Acknowledgments
This research was sponsored by NSFC (61620106008,
61572264), CAST (YESS20150117), Huawei Innovation
Research Program (HIRP), and IBM Global SUR award.
References
Achanta, R.; Shaji, A.; Smith, K.; Lucchi, A.; Fua, P.; and
Süsstrunk, S. 2012. SLIC superpixels compared to state-of-theart superpixel methods. IEEE Trans. PAMI 34(11):2274–2282.
Arbelaez, P.; Maire, M.; Fowlkes, C.; and Malik, J. 2011. Contour
detection and hierarchical image segmentation. IEEE Trans. PAMI
33(5):898–916.
Barnes, C.; Shechtman, E.; Finkelstein, A.; and Goldman, D. 2009.
Patchmatch: a randomized correspondence algorithm for structural
image editing. ACM Transactions on Graphics-TOG 28(3):24.
Boykov, Y., and Kolmogorov, V. 2004. An experimental comparison of min-cut/max-flow algorithms for energy minimization in
vision. IEEE Trans. PAMI 26(9):1124–1137.
Boykov, Y.; Veksler, O.; and Zabih, R. 2001. Fast approximate energy minimization via graph cuts. IEEE Trans. PAMI 23(11):1222–
1239.
Cheng, M.-M.; Liu, Y.; Hou, Q.; Bian, J.; Torr, P.; Hu, S.-M.; and
Tu, Z. 2016. HFS: Hierarchical feature selection for efficient image
segmentation. In ECCV, 867–882. Springer.
Comaniciu, D., and Meer, P. 2002. Mean shift: A robust approach
toward feature space analysis. IEEE Trans. PAMI 24(5):603–619.
Du, Q.; Faber, V.; and Gunzburger, M. 1999. Centroidal Voronoi
tessellations: applications and algorithms. SIAM review 41(4):637–
676.
Felzenszwalb, P. F., and Huttenlocher, D. P. 2004. Efficient graphbased image segmentation. International Journal of Computer Vision 59(2):167–181.
Hoiem, D.; Efros, A. A.; and Hebert, M. 2005. Automatic photo
pop-up. ACM transactions on graphics (TOG) 24(3):577–584.
Kolmogorov, V., and Zabin, R. 2004. What energy functions can
be minimized via graph cuts? IEEE Trans. PAMI 26(2):147–159.
Levinshtein, A.; Stere, A.; Kutulakos, K. N.; Fleet, D. J.; Dickinson, S. J.; and Siddiqi, K. 2009. Turbopixels: Fast superpixels
using geometric flows. IEEE Trans. PAMI (12):2290–2297.
Liu, M.-Y.; Tuzel, O.; Ramalingam, S.; and Chellappa, R. 2011.
Entropy rate superpixel segmentation. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, 2097–2104.
IEEE.
Liu, Y.-J.; Yu, C.-C.; Yu, M.-J.; and He, Y. 2016. Manifold SLIC:
A fast method to compute content-sensitive superpixels. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, 651–659.
Lloyd, S. 1982. Least squares quantization in PCM. IEEE Transactions on Information Theory 28(2):129–137.
Neubert, P., and Protzel, P. 2012. Superpixel benchmark and comparison. In Proc. Forum Bildverarbeitung, 1–12.
Osher, S., and Sethian, J. A. 1988. Fronts propagating with
curvature-dependent speed: algorithms based on Hamilton-Jacobi
formulations. Journal of computational physics 79(1):12–49.
Peyré, G.; Péchaud, M.; Keriven, R.; and Cohen, L. D. 2010.
Geodesic methods in computer vision and graphics. Foundations
and Trends R in Computer Graphics and Vision 5(3–4):197–397.
Shi, J., and Malik, J. 2000. Normalized cuts and image segmentation. IEEE Trans. PAMI 22(8):888–905.
Stutz, D.; Hermans, A.; and Leibe, B. 2014. Superpixel segmentation using depth information. RWTH Aachen University, Aachen,
Germany.
Van den Bergh, M.; Boix, X.; Roig, G.; de Capitani, B.; and
Van Gool, L. 2012. Seeds: Superpixels extracted via energy-driven
sampling. In European conference on computer vision, 13–26.
Springer.
Veksler, O.; Boykov, Y.; and Mehrani, P. 2010. Superpixels and
supervoxels in an energy optimization framework. In European
conference on Computer vision, 211–224. Springer.
Vincent, L., and Soille, P. 1991. Watersheds in digital spaces: an
efficient algorithm based on immersion simulations. IEEE Trans.
PAMI 13(6):583–598.
Wang, S.; Lu, H.; Yang, F.; and Yang, M.-H. 2011. Superpixel
tracking. In 2011 International Conference on Computer Vision,
1323–1330. IEEE.
Wang, P.; Zeng, G.; Gan, R.; Wang, J.; and Zha, H. 2013a.
Structure-sensitive superpixels via geodesic distance. International
journal of computer vision 103(1):1–21.
Wang, Z.; Feng, J.; Yan, S.; and Xi, H. 2013b. Image classification
via object-aware holistic superpixel selection. IEEE Transactions
on Image Processing 22(11):4341–4352.
| 1cs.CV
|
An asymptotically optimal indirect approach to continuous-time system
identification
arXiv:1803.08449v1 [cs.SY] 22 Mar 2018
Rodrigo A. González, Cristian R. Rojas and James S. Welsh
Abstract— The indirect approach to continuous-time system
identification consists in estimating continuous-time models by
first determining an appropriate discrete-time model. For a
zero-order hold sampling mechanism, this approach usually
leads to a transfer function estimate with relative degree 1,
independent of the relative degree of the strictly proper real system. In this paper, a refinement of these methods is developed.
Inspired by indirect PEM, we propose a method that enforces
a fixed relative degree in the continuous-time transfer function
estimate, and show that the resulting estimator is consistent
and asymptotically efficient. Extensive numerical simulations
are put forward to show the performance of this estimator
when contrasted with other indirect and direct methods for
continuous-time system identification.
Index Terms— System identification; Continuous-time systems; Parameter estimation; Sampled data.
I. I NTRODUCTION
System identification deals with the problem of estimating
adequate models of dynamical systems from input-output
data. The methods developed over the years in this field have
seen applications in many areas of science and engineering,
and comprehensive literature has been written on the subject
[1]–[3].
When postulating a mathematical model for describing a
dynamical system based on sampled data, one must decide
between obtaining a discrete-time (DT), or a continuoustime (CT) model. In a predominantly digital era, DT system
identification has been studied thoroughly (see, e.g., [1],
[2]). Nevertheless, interest in CT models still persists due
to its advantages over discrete-time. For example, greybox modelling [4], which is commonly based on physical
principles and conservation laws, is naturally suited for
continuous time, as the parameters can usually be better
interpreted in this domain. Also, CT models are known to
have more intuitive dynamics, and they do not depend on a
sampling period.
In CT system identification there are two main approaches,
namely the direct and indirect approaches. For direct CT
system identification, a CT model is obtained directly from
the sampled data. The main difficulty present in the direct
methods is the handling of derivatives, as they are not immediately available from discrete data points without amplifying
This work was supported by the Swedish Research Council under contract
number 2016-06079 (NewLEADS).
R.A. González and C.R. Rojas are with the Department of Automatic
Control and ACCESS Linnaeus Centre, KTH Royal Institute of Technology,
10044 Stockholm, Sweden (e-mails: [email protected], [email protected]).
James S. Welsh is with the School of Electrical Engineering
and Computer Science, University of Newcastle, Australia (e-mail:
[email protected]).
noise [5]. To effectively deal with this issue, many well
known methods have been proposed [6], with success in
real applications [7]. On the other hand, indirect methods
for CT modelling first determine a suitable DT model via
DT system identification methods like the Prediction Error
Methods (PEM) or Maximum Likelihood (ML) [2], and then
transform this model into a CT equivalent model. Evidence
has been shown regarding the advantages of direct over
indirect CT model identification [8], although with a precise
initialisation of PEM, the approaches seem comparable for
certain sampling periods [9].
Even though the indirect approach seems easy to implement, as there is much theory and literature concerning DT
system identification, there are reasons why this approach is
not always recommended. First, it may suffer from numerical
inaccuracies at fast sampling, and requires a precise initialisation. In addition, it is not possible to select the desired
numerator order of the CT model, as the estimated DT model
will generally lead to a CT model with relative degree 1 in
the case of sampling by a zero-order hold mechanism. Hence,
an unnecessarily complex model structure is indirectly being
estimated, which leads to a loss in accuracy according to the
parsimony principle [1].
In this paper, we introduce a method that optimally
imposes a desired relative degree in the indirect approach
to continuous-time system identification. Based on Indirect
PEM [10], we prove that the proposed estimator is a consistent and asymptotically efficient estimator of the system’s
true parameter vector. Extensive numerical simulations show
that the new method imposes the correct relative degree,
while improving the statistical properties of the transfer
function estimate, and achieves a performance that compares favourably against both standard direct and indirect
approaches.
The remainder of this paper is organised as follows. In
Section II the problem is formulated. Section III provides an
introduction to the indirect approach for CT system identification. In Section IV we derive an estimator that optimally
enforces the desired relative degree for the indirect approach,
and determine its properties. Section V illustrates the method
with extensive numerical examples. Finally, conclusions are
drawn in Section VI.
II. P ROBLEM FORMULATION
Consider a linear time-invariant, causal, stable, single input
single output, CT system
y(t) = G0 (ρ)u(t)
βn−r ρn−r + βn−r−1 ρn−r−1 + · · · + β1 ρ + β0
=
u(t),
ρn + αn−1 ρn−1 + · · · + α1 ρ + α0
(1)
where ρ is the Heavyside operator, i.e., ρg(t) = dg(t)/dt,
and r is the relative degree of the system. In this paper, we
denote θc0 := [βn−r . . . β0 αn−1 . . . α0 ]> as the true CT
system parameter vector.
Suppose that the input-output signals are sampled with
period h and the resulting output is contaminated by an
additive zero-mean white noise sequence {e(kh)}k∈N of
variance σ 2 . That is,
ym (kh) = y(kh) + e(kh),
k ∈ N.
(2)
The goal of CT system identification is to obtain a
CT transfer function estimate for G0 (ρ), given N discrete
input-output data measurements {u(kh), ym (kh)}N
k=1 and
knowledge about the physical characteristics of the system,
or the intersample behaviour. In this paper, we assume that
the input is a piecewise constant signal between samples (i.e.,
zero-order hold behaviour [11]).
For obtaining a model of G0 (ρ), a simple way to proceed
is to identify the zero-order hold equivalent model given
the input and output data measurements by using standard
PEM in the DT domain, and then return to the CT domain
via zero-order hold equivalences. Although this procedure
has good statistical properties, it does not impose relative
degree constraints in the CT domain. Our goal is to optimally
impose this constraint, which should lead to a statistically
improved estimate of G(s).
III. T HE INDIRECT APPROACH TO CONTINUOUS - TIME
SYSTEM IDENTIFICATION
One approach to identifying a CT system is to first
estimate the DT model given the input and output data
samples, and then translate this model into continuous time.
This is called the indirect approach, since it relies on DT
system identification theory instead of obtaining immediately
a CT model using CT system identification methods.
Much literature has been written regarding the first step
of the indirect approach [1], [2]. The theoretically optimal
solution is to apply the maximum likelihood method. This
method is known to give consistent and asymptotically
efficient estimates under very general conditions. Under the
assumption that the additive white noise is Gaussian, the
ML method is equivalent to PEM, which is one of the
most celebrated parametric methods, and is available in the
MATLAB System Identification Toolbox [12].
If a CT model for (1) is required, we should propose a
DT model structure of the form
bn−1 z n−1 + bn−2 z n−2 + · · · + b1 z + b0
.
H(z) =
z n + an−1 z n−1 + · · · + a1 z + a0
(3)
If we define θd = [bn−1 . . . b0 an−1 . . . a0 ]> and denote ŷ
as the model’s output, then the ML estimate is
θ̂d = arg min
θd
N
1 X
kym (kh) − ŷ(kh, θd )k2 .
σ2
k=1
The next step is to transform this DT transfer function into
an adequate CT model. This can be done in several ways.
For example, the well-known Tustin transformation can be
applied on the DT transfer function estimate by letting
sh + 2
,
z=
2 − sh
as reported in [13]. If it is assumed that the CT input signal
is piecewise constant, the most natural mapping (used in e.g.
[9], [14]), is the zero-order hold sampling equivalence
G0 (s)
−1
−1
, (4)
H0 (z) = (1 − z )Z L
s
t=kh
where Z and L denote the Z and Laplace transforms,
respectively. This mapping is known to be a troublesome
part of the indirect approach, as it can be ill conditioned and
its uniqueness depends on a correct choice of the sampling
period.
In both of these mappings, the resulting CT model can
have numerator parameters that exceed the desired numerator
orders. This is generally the case when the relative degree
of G0 (s) is greater than 1. This problem contributes to poor
accuracy and high standard deviations at high frequencies.
One very simple way of treating this issue is setting to zero
the numerator coefficients which should be zero, but this is
not the best way of taking care of the information [9].
IV. O PTIMAL ENFORCEMENT OF RELATIVE DEGREE
In this section we develop an indirect-approach estimator
for the CT parameter vector θc0 that renders a CT transfer
function estimate Ĝ(s) of a desired relative degree r. For
this matter, we first focus on the PEM estimator of the zeroorder hold equivalent model of G0 (s).
For simplicity, we assume that the correct model order
has been found. A model with structure (3) is obtained by
PEM, and the covariance matrix of θ̂d is also estimated. We
know that the CT zero-order hold equivalent of this estimated
model is in general given by
Ĝ(s) =
β̂n−1 sn−1 + β̂n−2 sn−2 + · · · + β̂1 s + β̂0
.
sn + α̂n−1 sn−1 + · · · + α̂1 s + α̂0
Define θ̂c = [β̂n−1 . . . β̂0 α̂n−1 . . . α̂0 ]> , and denote by
θd0 the true DT parameter vector. The parameters in θ̂c are
related to θ̂d by the zero-order hold equivalence equations
that can be derived by using (4) and comparing coefficients.
This relation is a nonlinear mapping f : θ̂c → f (θ̂c ) =
θ̂d , which is differentiable almost everywhere. Hence, the
following asymptotic relationship is valid for the covariance
matrices of θ̂d and θ̂c :
Σθ̂d = E{(θ̂d − θd0 )(θ̂d − θd0 )> }
≈ E{J(θ̂c − θc0 )(θ̂c − θc0 )> J > }
= JΣθ̂c J > ,
(5)
where J is the Jacobian matrix of f evaluated at the naive
estimation of θc0 , that is, throwing away the high order
coefficients of the numerator of Ĝ which should be zero1 .
Now, we propose to find an appropriate projection of θ̂c
into a proper subspace of the parameter space that yields
the desired relative degree. This subspace is simply the one
formed by all vectors with first r − 1 elements set to zero.
Hence, we decide to study the following problem:
1
(θ̂c − θ)
θ̃c = arg min (θ̂c − θ)> Σ−1
θ̂c
θ 2
Ir−1 0 θ = 0,
s.t.
(6)
(7)
where Ir−1 is the identity matrix of dimension r − 1, 0 is the
null matrix of appropriate dimensions, and Σ−1
= J > Σ−1
J.
θ̂c
θ̂d
The optimisation problem in this context can be interpreted
as an application of the Indirect PEM [10].
By Lagrange multiplier theory, the optimization problem
in (6) is equivalent to calculating, for a suitable λ,
1
(θ̂c − θ) + λ> Ir−1 0 θ.
θ̃c = arg min (θ̂c − θ)> Σ−1
θ̂
c
θ 2
Partitioning Σθ̂c appropriately (dropping the subindex for
simplicity), and differentiating with respect to θ we obtain
Σ11 Σ12 Ir−1
θ̃c = θ̂c −
λ,
(8)
Σ21 Σ22
0
Imposing (7) we obtain λ = Σ−1
11 Ir−1 0 θ̂c . If we denote
by C the Cholesky factorization matrix of Σθ̂c [15] (i.e., a
lower triangular matrix with positive diagonal entries such
that Σθ̂c = CC > ) we can write (8) as
Ir−1 −1
θ̃c = θ̂c − Σ
Σ11 Ir−1 0 θ̂c
0
C
0
Ir−1
−1
Ir−1 0 θ̂c
= θ̂c − 11
C11
C21 C22
0
>
0
0
= C r−1
C −1 θ̂c .
(9)
0
I2n−r+1
That is, the estimator (9) can be seen as an L2 best approximation to the PEM CT estimate θ̂c that imposes the desired
relative degree.
A. Properties
We briefly present the most important properties of estimator (9) in the following theorems.
Theorem 4.1: Consider the system described by (1) and
(2), where {e(kh)}N
k=1 is a Gaussian white noise sequence.
Assume that the sampling frequency 2π/h is larger than
twice the largest imaginary part of the s-domain poles and
there is no delay in the real system2 . Then, the estimator (9)
is a consistent and asymptotically efficient estimator of the
real vector parameter θc0 , provided the DT model set (with
the chosen relative degree) contains the real system.
1 Note that the standard PEM estimate could have also been used for this
matter, and the asymptotic relation still holds.
2 These conditions can be relaxed, as long as the sampling frequency is
such that the z → s transformation is well defined.
Proof: Under the Gaussian noise assumption, the DT
PEM estimate can be interpreted as the ML estimate. Under
the proposed sampling frequency, the z → s transformation
is unique for sufficiently large N [16]. Hence, by the
invariance principle of ML estimators [17], the CT equivalent
of the system’s parameters is also an ML estimate.
To prove the theorem, we only require that the assumptions
in [10] are satisfied in this scenario, and then directly apply
the results obtained in the cited contribution. First, note that
the model structure given by this procedure contains the
models with the desired relative degree (and the contention
is proper if r > 1), with a linear mapping between parameter
vectors given by the matrix
T := 0r−1 I2n−r+1 .
Furthermore, provided that the DT model set contains the
real system, both structures give parameter identifiability.
Also note that Σθ̂c , obtained via (5), is a consistent estimate of the covariance matrix of θ̂c . Hence, the results
in
Namely,
estimation errors3
√ the normalized
√ [10] follow.
0
0
N (θ̂c − θc )/σ and N (θ̃c − θc )/σ are asymptotically
normally distributed with zero means and their asymptotic
covariance matrices satisfy the relation
[AsCov(θ̃c − θc0 )]−1 = T [AsCov(θ̂c − θc0 )]−1 T > .
(10)
Moreover, following the √
steps in [10, Section 3], the improved PEM estimate is N −consistent and has the same
asymptotic distribution as θ̂c , thereby proving its asymptotic
efficiency.
Remark 4.1: Note that by (9) and (10), the asymptotic
covariances can be shown to satisfy the following properties:
• AsCov(θ̃c − θc0 , θ̂c − θ̃c ) = 0,
• AsCov(θ̃c − θc0 ) = AsCov(θ̂c − θc0 ) − AsCov(θ̂c − θ̃c ).
Both of these claims follow by applying properties 10.5 and
10.6 from [18, Chapter 10] to this context. These properties
imply that for sufficiently large N , the proposed estimator
can only decrease the covariance of the estimated parameters
compared to standard PEM. The asymptotic covariances satisfy a Pythagorean relation, as the PEM estimate is projected
orthogonally on to the proper subspace where θc0 lies.
Next, we establish that imposing a larger relative degree
improves the accuracy of the estimates, provided that the
highest relative degree model structure contains the real
system.
Theorem 4.2: Given a plant G0 (s) of order n and relative
degree r > 1, consider CT candidate models of relative
degree r1 and r2 and their improved PEM parameter vector
estimates θ̃cr1 and θ̃cr2 respectively. If r1 < r2 ≤ r, then
their asymptotic covariance matrices satisfy AsCov(θ̃cr2 )
AsCov(θ̃cr1 ).
Proof: The proof follows by applying [19, Theorem 2].
Details are omitted due to length restrictions.
3 For simplicity, we assume that the vector θ 0 has the appropriate
c
dimension, where zeros have been considered in the first terms if necessary.
Remark 4.2: The relative degree of the CT system is
not always known. In some practical applications, physical
knowledge about the system can give intuition. In addition,
statistical measures can be used such as the coefficient of
determination [7], or the Young Information Criterion [20].
differences, and the correct relative degree has been imposed
for this estimator. We have used the command d2c of
MATLAB in both cases. SRIVC has been implemented with
the CONTSID toolbox for MATLAB [6, Chapter 9] with
default initialisation, and has been set to estimate the model
V. M ONTE C ARLO SIMULATION STUDIES
β1 s + β0
.
s4 + α3 s3 + α2 s2 + α1 s + α0
A. Effect of the number of data points and sampling rate
In this study, we have designed the input as a pseudorandom binary sequence (PRBS) of amplitude switches between
0 and 2. For the first input sequence, the number of stages
of the shift register is n = 10 and the data length of the
shortest interval is p = 7. Hence, a sequence of N = 7161
data points has been obtained. The noise is a zero-mean
white Gaussian noise signal, where the variance has been
set such that the signal-to-noise ratio (SNR in dB) between
the noiseless output sequence and the noise equals 10 dB.
Three Monte Carlo studies have been performed with
M = 500 simulations of different noise realisations each
one, for h = 0.01, 0.05, 0.1. The results are shown in Table
I.
We will now study the performance of the proposed
estimator under a series of experiments.
The system considered is the Rao-Garnier system [8],
which is a linear fourth-order non-minimum phase system
with complex poles that has been tested in many publications
(see e.g. [7], [21], [22]) of CT system identification:
−6400s + 1600
.
s4 + 5s3 + 408s2 + 416s + 1600
This system is interesting since it has two damped oscillatory
modes at 2 and 20 rad/sec with damping of 0.1 and 0.25
respectively, and has a non-minimum phase zero at s = 0.25.
It is known that this is a particularly difficult system to
estimate by PEM/ML, since these methods may converge
to a local minimum if they are not well initialised [21].
Three methods have been compared: PEM, PEM with
relative degree enforcement (labeled PEMrd (9)), and the
simplified refined IV method for continuous-time systems
(SRIVC), which is one of the most successful direct methods
available, and has been suggested for general use in one
of the most recent surveys on CT system identification [7].
Each method has been tested in M different Monte Carlo
simulations, and they have been evaluated according to the
average normalized square error of the system estimate
G(s) =
M
1 X kĜi − G0 k22
MSE Ĝ =
,
M i=1
kG0 k22
Ĝ(s) =
TABLE I
M ONTE C ARLO SIMULATION RESULTS FOR PRBS INPUT OF TOTAL
LENGTH
h
0.01
0.05
0.1
Method
PEM
PEMrd
SRIVC
PEM
PEMrd
SRIVC
PEM
PEMrd
SRIVC
N = 7161.
MSE Ĝ
1.113 · 10−4
0.732 · 10−4
0.733 · 10−4
1.996 · 10−3
1.406 · 10−3
1.397 · 10−3
3.275 · 10−3
1.914 · 10−3
1.922 · 10−3
MSE θ̂
6.757 · 10−5
4.283 · 10−5
4.292 · 10−5
5.423 · 10−4
3.933 · 10−4
3.925 · 10−4
7.937 · 10−4
4.797 · 10−4
4.749 · 10−4
Fit
98.9742
99.1219
99.1217
98.9645
99.1141
99.1146
98.9436
99.0884
99.0870
the average normalized square error of the parameter vectors
M
1 X kθ̂ci − θc0 k22
MSE θ̂ =
,
M i=1 kθc0 k22
and the fit measure
Fit =
M
kŷ i − yk2
100 X
1−
,
M i=1
ky − ȳk2
where y is the noise-free output sequence (the simulated data
without the additive measurement noise), ŷ i is the simulated
output sequence of the i-th estimated model, and ȳ is the
average value of {y(kh)}N
k=1 .
We have run PEM using the standard MATLAB System
Identification Toolbox [12] with the oe command, and
assumed that the correct order of the system is known. The
search algorithm has been initialised with the estimate given
by the Null Space Fitting method4 [23]. We based PEMrd
on the PEM estimate previously obtained. The required
Jacobian matrix has been numerically calculated via finite
4 PEM initialised with the estimate from SRIVC (as in [9]) has also been
tested with similar results.
In order to analyse the performance of each estimator
under less data, we have set the number of stages to n = 9
and the data length of the shortest interval to p = 3, resulting
in a input of N = 1533 data points. With the same SNR as
the test above, the results for 500 Monte Carlo simulations
for each sampling period can be found in Table II.
TABLE II
M ONTE C ARLO SIMULATION RESULTS FOR PRBS INPUT OF TOTAL
LENGTH N = 1533.
h
0.01
0.05
0.1
Method
PEM
PEMrd
SRIVC
PEM
PEMrd
SRIVC
PEM
PEMrd
SRIVC
MSE Ĝ
5.882 · 10−4
4.025 · 10−4
4.017 · 10−4
7.431 · 10−4
4.316 · 10−4
4.319 · 10−4
1.959 · 10−2
1.077 · 10−2
9.992 · 10−3
MSE θ̂
3.651 · 10−4
2.057 · 10−4
2.060 · 10−4
4.539 · 10−4
2.788 · 10−4
2.789 · 10−4
2.915 · 10−3
2.184 · 10−3
3.877 · 10−3
Fit
97.7791
98.0705
98.0719
97.7892
98.1172
98.1167
97.7416
98.0339
97.9754
Both Tables I and II show that the refined PEM estimator
statistically improves the estimates given by PEM, and is
A different input signal has also been tested. We have
taken 2000 data measurements of a multisine input given
by the sum of sine waves of angular frequencies ω =
0.5, 1, 5, 8, 10, 12, 15, 20, 25, 30. The standard deviation of
the additive noise has been set to 0.1, and the median of
500 Monte Carlo simulations of the normalized model error,
normalized parameter error, and fit have been obtained. The
results are shown in Table III.
Bode Diagram
0
-50
-100
-150
-200
0
-180
-360
-540
-720
10-2
TABLE III
Bode Diagrams
50
Magnitude (dB)
B. Multisine input
simulations under the setup in Section V-A, for h = 0.05,
N = 7161.
It is clear by Fig. 1 that the improvement over PEM is
mainly at high frequencies. This is intuitive, since the relative
degree determines the asymptotic slope of the Bode diagram
of magnitude. The proposed method enforces the true asymptotic slope, leading to an important gain in accuracy in both
magnitude and phase.
Phase (deg)
a very competitive method against SRIVC, even for high
frequency sampling. Note that under less data points, PEMrd
still outperforms PEM for every sampling period, which
indicates that the asymptotic properties studied in Section
IV can be observed in practical finite data cases as well.
Remark 5.1: In Tables I and II we have discarded cases
where PEM has delivered estimates with one pole in the negative real axis. Fortunately this scenario is very uncommon
(9 cases seen in 3000 simulations). A similar phenomenon
was observed in Table II for h = 0.1 in SRIVC, where 2
estimates gave negative fit values. These simulations were
not considered either.
10-1
M ONTE C ARLO SIMULATION RESULTS FOR MULTISINE WAVE INPUT OF
TOTAL LENGTH
h
0.01
0.02
Method
PEM
PEMrd
SRIVC
PEM
PEMrd
SRIVC
100
101
Frequency (rad/s)
102
103
104
N = 2000.
kĜ−G0 k2
kG0 k2
8.799 · 10−5
kθ̂−θ0 k2
kθ0 k2
7.269 · 10−5
1.352 · 10−5
1.439 · 10−1
1.422 · 10−4
2.294 · 10−5
2.141 · 10−2
4.435 · 10−6
6.205 · 10−2
8.368 · 10−5
1.013 · 10−5
4.599 · 10−3
Fit
99.4319
99.8173
80.8446
99.2684
99.7325
94.3144
For this input and sampling method, SRIVC performs
poorly, while PEM and PEMrd normally reach the global optimum. Even though the fit is very near the optimal, PEMrd
is consistently better than standard PEM at all metrics.
C. Mean and covariance of the estimated parameters
As established in Theorem 4.1, the improved PEM estimate should reduce (at least asymptotically) the covariance of
the parameter vector. To test this, we have obtained the mean
and standard deviation of each parameter given by the Monte
Carlo study under the setup of Section V-A with h = 0.05
and N = 7161. The results are shown in Table IV. It can
be observed that the mean values are similar in all methods
(except standard PEM, which does not estimate the correct
model structure), and PEMrd provides the lowest standard
deviation for every parameter.
D. Direct comparison with standard PEM
In this subsection we analyse the improvements of the
novel method over the standard PEM estimates.
To show the impact of selecting and enforcing the correct
relative degree, we have focused on the Bode Diagrams of the
resulting models for PEM and PEMrd. In Figure 1 we have
plotted the frequency response estimates of 100 Monte Carlo
Fig. 1. Bode Diagram of 100 estimates by PEM (green), 100 estimates by
PEMrd (blue), and the real system (red).
Also, direct comparison plots between PEM and PEMrd
are shown in Figure 2. These plots compare the normalized
model and parameter error, and the fit of PEM and PEMrd
for each Monte Carlo experiment (500 in total). PEMrd
outperforms standard PEM in most Monte Carlo simulations,
specially in the fit comparison, where 496 out of 500 experiments have lead to an increase in fit under PEM refinement.
E. Random systems
To obtain the average performance of each estimator, we
have tested them on a data set created with 500 random
systems of order 3 and relative degree 2, by using the rss
command in MATLAB. The slowest pole has been set to
have real part less than -0.1. The input was a unit variance
Gaussian white noise, and the additive white noise was also
Gaussian, of standard deviation equal to 5% of the maximum
value of the noiseless output. The sampling period has been
chosen as 10 times faster than the fastest pole or zero of the
real system.
We have computed the median of the metrics used above,
as in Section V-B, for the 500 random systems. We also have
counted failures of the estimators, which are the cases when
the estimates produced a negative fit, when the algorithm
crashed, or when it was not possible to correctly initialise
the PEM estimate having tried reducing by a factor of 2 the
sampling rate for initialisation estimation.
The results can be seen in Table V. Although all estimators
report failures, the PEMrd again shows promising results in
TABLE IV
E STIMATED PARAMETER VALUE MEANS AND STANDARD DEVIATIONS FOR EACH METHOD , CONSIDERING h = 0.05.
Method
PEM
PEMrd
SRIVC
10 -3 Model error norms
8
b2
0
0.301
11.414
0
0
0
0
b3
-6400
-6403.01
147.13
-6397.25
122.39
-6399.19
132.05
4
PEMrd
4
b4
1600
1601.33
47.85
1599.79
42.39
1600.49
44.21
a1
5
5.006
0.399
4.994
0.315
5.014
0.338
a2
408
408.16
7.98
407.88
7.11
407.99
7.75
a3
416
416.25
9.27
415.84
8.33
416.61
8.98
a4
1600
1600.49
33.29
1599.27
28.59
1599.71
31.1
random systems have been put forward with promising
results. These have shown that the refinement to standard
PEM leads to an important improvement in all the statistical
metrics studied, and its performance is comparable, if not
better, to SRIVC for all sampling periods in this study
provided PEM is initialised correctly.
10 -3 Parameter error norms
5
6
PEMrd
b1
0
-0.044
0.963
0
0
0
0
Parameter
True value
Mean
Std. Dev
Mean
Std. Dev.
Mean
Std. Dev.
3
2
R EFERENCES
2
1
0
0
0
2
4
6
8
0
10 -3
PEM
1
2
3
4
Fit
100
5
10 -3
PEM
PEMrd
99.5
99
98.5
98.5
99
99.5
100
PEM
Fig. 2. Direct comparison plots between PEM and PEMrd for 500 Monte
Carlo simulations. The green dots correspond to Monte Carlo simulations
where PEMrd outperforms PEM. Red dots represent the opposite, and the
dashed blue line is the separatrix.
TABLE V
M ONTE C ARLO SIMULATION RESULTS FOR RANDOM SYSTEMS .
Method
PEM
PEMrd
SRIVC
kĜ−G0 k2
kG0 k2
1.449 · 10−4
kθ̂−θ0 k2
kθ0 k2
3.328 · 10−3
1.046 · 10−4
1.111 · 10−4
1.872 · 10−3
2.126 · 10−3
Fit
N◦ Failures
98.871
98.997
98.9723
12
16
30
all metrics. As pointed out in [9] and seen in these tests,
initialisation aspects are in fact a major issue concerning the
reliability of these algorithms.
VI. C ONCLUSIONS
We have proposed a refinement to the standard PEM estimator for indirect continuous-time system identification that
achieves an asymptotically optimal desired relative degree
enforcement. An explicit expression for this estimator has
been found, and its statistical properties have been analysed.
Extensive simulations using both standard benchmarks and
[1] T. Söderström, P. Stoica, System Identification, Prentice-Hall, 1988.
[2] L. Ljung, System Identification: Theory for the User, 2nd edition,
Prentice-Hall, 1999.
[3] R. Pintelon, J. Schoukens, System Identification: A Frequency Domain
Approach, John Wiley & Sons, 2012.
[4] T. P. Bohlin, Practical Grey-box Process Identification: Theory and
Applications, Springer, 2006.
[5] G. Rao, H. Unbehauen, Identification of continuous-time systems, IEE
Proceedings-Control theory and applications 153 (2) (2006) 185–220.
[6] H. Garnier, L. Wang, Identification of Continuous-time Models from
Sampled Data, Springer, 2008.
[7] H. Garnier, Direct continuous-time approaches to system identification. Overview and benefits for practical applications, European
Journal of control 24 (2015) 50–62.
[8] G. Rao, H. Garnier, Numerical illustrations of the relevance of direct continuous-time model identification, IFAC Proceedings Volumes
35 (1) (2002) 133–138.
[9] L. Ljung, Experiments with identification of continuous time models,
IFAC Proceedings Volumes 42 (10) (2009) 1175–1180.
[10] T. Söderström, P. Stoica, B. Friedlander, An Indirect Prediction Error
Method for System Identification, Automatica 27 (1) (1991) 183–188.
[11] K. J. Åström, B. Wittenmark, Computer Controlled Systems: Theory
and Design, Prentice-Hall, 1984.
[12] L. Ljung, System Identification Toolbox 7: Getting Started Guide.
[13] H. Unbehauen, G. Rao, Continuous-time Approaches to System Identification - A Survey, Automatica 26 (1) (1990) 23–35.
[14] N. K. Sinha, Identification of continuous-time systems from samples
of input-output data: An introduction, Sadhana 25 (2) (2000) 75–83.
[15] R. A. Horn, C. R. Johnson, Matrix Analysis, 2nd Edition, Cambridge
University Press, 2012.
[16] I. Kollar, G. Franklin, R. Pintelon, On the Equivalence of z-domain
and s-domain Models in System Identification, in: Instrumentation and
Measurement Technology Conference, IMTC-96, Vol. 1, IEEE, 1996,
pp. 14–19.
[17] P. W. Zehna, et al., Invariance of maximum likelihood estimators,
Annals of Mathematical Statistics 37 (3) (1966) 744.
[18] C. Gourieroux, A. Monfort, Statistics and econometric models, Vol. 1,
Cambridge University Press, 1995.
[19] R. A. González, P. E. Valenzuela, C. R. Rojas, R. A. Rojas, Optimal
enforcement of causality in non-parametric transfer function estimation, IEEE Control Systems Letters 1 (2) (2017) 268–273.
[20] P. Young, Recursive estimation, forecasting and adaptive control,
Control and Dynamic systems 30 (Part 3) (1989) 119–165.
[21] L. Ljung, Initialisation aspects for subspace and Output-Error identification methods, in: European Control Conference (ECC), 2003, 2003,
pp. 773–778.
[22] J. S. Welsh, J. C. Agüero, M. Alamir, Continuous-time System
Identification using Indirect Inference, IFAC Proceedings Volumes
42 (10) (2009) 1169–1174.
[23] M. Galrinho, C. R. Rojas, H. Hjalmarsson, The weighted null-space
fitting method, arXiv preprint arXiv:1708.03946.
| 3cs.SY
|
1
When data mining meets optimization: A case
study on the quadratic assignment problem
arXiv:1708.05214v2 [cs.DM] 7 Oct 2017
Yangming Zhou, Jin-Kao Hao∗ , and Béatrice Duval
Abstract—This paper presents a hybrid approach called frequent pattern based search that combines data mining and optimization.
The proposed method uses a data mining procedure to mine frequent patterns from a set of high-quality solutions collected from
previous search, and the mined frequent patterns are then employed to build starting solutions that are improved by an optimization
procedure. After presenting the general approach and its composing ingredients, we illustrate its application to solve the well-known
and challenging quadratic assignment problem. Computational results on the 21 hardest benchmark instances show that the proposed
approach competes favorably with state-of-the-art algorithms both in terms of solution quality and computing time.
Index Terms—Frequent pattern mining, heuristic design, learning-based optimization, quadratic assignment
✦
1
I NTRODUCTION
R
ECENT years, hybridization of data mining techniques
with metaheuristics has received increasing attention
from the optimization community. A number of attempts
have been made to use data mining techniques to enhance
the performance of metaheuristic optimization for solving
hard combinatorial problems [6], [35], [45], [46], [48], [49].
Data mining involves discovering useful rules and hidden patterns from data [47]. With the help of data mining
techniques, heuristic search algorithms can hopefully make
their search strategies more informed and thus improve
their search performances. For instance, candidate solutions
visited by a search algorithm can be collected and exploited
by supervised or unsupervised learning techniques to extract useful hidden patterns and rules, which can be used
to guide future search decisions of the search algorithm.
More generally, data mining techniques could help to build
the starting solution or initial population, choose the right
operators, set the suitable parameters or make automatic
algorithm configuration [21], [40].
In this work, we introduce a hybrid approach called
frequent pattern based search (FPBS) that combines data
mining techniques and optimization methods for solving
combinatorial search problems. Basically, FPBS employs a
data mining procedure to mine useful patterns that frequently occur in high-quality (or elite) solutions collected
from previous search and then uses the mined patterns to
construct new starting solutions that are further improved
by an optimization method. To update the set of elite solutions used for pattern mining, FPBS calls for a dedicated
procedure to manage newly discovered elite solutions. The
key intuition behind the proposed approach is that a frequent pattern extracted from high-quality solutions identi-
fies a particularly promising region in the search space that
is worthy of an intensive examination by the optimization
procedure. By using multiple mined patterns combined with
an effective optimization procedure, FPBS is expected to
achieve a suitable balance between search exploration and
exploitation that is critical for a high performance of the
search algorithm [19].
To verify the interest of the proposed FPBS approach, we
consider the well-known and highly challenging quadratic
assignment problem (QAP) as a case study. Besides its popularity as one of the most studied NP-hard combinatorial
optimization problem, QAP is also a relevant representative
of many permutation problems. To apply the general FPBS
approach to solve QAP, we specify the underlying patterns
and frequent patterns, the frequent pattern mining algorithm and the dedicated optimization procedure. We then
assess the resulting algorithm on the set of the hardest QAP
benchmark instances from the QAPLIB. Our experimental
results show that the proposed algorithm is highly competitive compared to state-of-the-art algorithms both in terms
of solution quality and computing time.
The rest of this paper is organized as follows. In the next
section, we introduce the concept of frequent pattern mining
and provide a literature review on hybridization of association rule mining with metaheuristic optimization. In Section
3, we present the proposed frequent pattern based search
approach. Section 4 shows the application of the general
FPBS approach to solve the quadratic assignment problem.
Section 5 is dedicated to an experimental analysis of key
components of the proposed approach. Finally, conclusions
and further work are discussed in Section 6.
•
2
Y. Zhou, J.K. Hao (corresponding author) and B. Duval are with the
Department of Computer Science, LERIA, Université d’Angers, 2
Boulevard Lavoisier, 49045 Angers, France, J.K. Hao is also affiliated with
the Institut Universitaire de France, 1 rue Descartes, 75231 Paris, France
(E-mails:
[email protected];
[email protected];
[email protected]).
F REQUENT PATTERN MINING AND HEURISTIC
SEARCH : STATE OF THE ART
In this section, we briefly recall the concept of frequent
pattern mining and review related literature on combination
of search methods with association rule mining (or frequent
2
pattern mining) techniques for solving combinatorial optimization problems.
2.1 Frequent pattern mining
Frequent pattern mining was originally introduced for market basket analysis in the form of association rules mining [3], with the purpose of identifying associations between different items that customers place in their “shopping baskets”. Also, the concept of frequent itemsets was
first introduced for mining transaction database. Let D =
{T1 , T2 , . . . , TN } be a transaction database defined over a
set of items I = {x1 , x2 , x3 , . . . , xM }. A frequent itemset
typically refers to a set of items that often appear together
in a transactional dataset [18], e.g., milk and bread, which
are frequently bought together in grocery stores by many
customers. Furthermore, a l-itemset (or l-pattern) p ⊆ I ,
which consists of l items from I , is frequent if p occurs in
the transaction database D no lower than θ times, where θ
is a user-defined minimum support threshold.
In the original model of frequent pattern mining [3], the
problem of finding association rules has also been proposed.
Association rules are closely related to frequent patterns
in the sense that association rules can be considered as a
“second-stage” output, which are derived from the mined
frequent patterns. Given two itemsets U ⊆ I and V ⊆ I ,
the rule U ⇒ V is considered to be an association rule at
minimum support θ and minimum confidence η , when the
following two conditions are satisfied simultaneously: the
set U ∪ V is a frequent pattern for the minimum support θ
and the ratio of the support of U ∪ V over U is at least η .
2.2 Association rule (or frequent pattern) mining for
heuristic search
We now review existing studies on hybridization between
data mining and heuristic search, starting by the issue of
representation of frequent patterns.
2.2.1 Representation of frequent patterns
The mined knowledge or information can be represented in
the form of frequent patterns or association rules. Frequent
patterns are patterns that occur frequently in the given data
[18]. There are various types of patterns, such as itemsets,
subsequences, and substructures.
To apply frequent pattern mining for solving combinatorial optimization problems, one of the main challenges
is to define a suitable pattern for the problem under consideration. In the following, we introduce the definition
of frequent patterns (in terms of frequent items) for two
categories of representative optimization problems.
•
Frequent pattern for subset selection problems. Subset selection is basically to determine a subset of
specific elements among a set of given elements
while optimizing an objective function defined over
the selected elements or unselected elements. Subset
selection is encountered in many situations and includes, for instance, knapsack problems, clique problems, diversity problems, maximum stable problems
and critical node problems. A candidate solution to a
subset selection problem is usually represented by a
•
set of selected elements. Since it is natural to treat
each element as an item, frequent pattern mining
techniques can be directly applied to subset selection
problems. In this setting, a transaction corresponds
to a solution of the subset selection problem and
a pattern can be conveniently defined as a subset
of elements that frequently appear in some specific
(e.g., high-quality) solutions. Thus, given a database
D (i.e., a set of visited solutions), a frequent pattern p
mined from D with support θ corresponds to a set of
elements that occur at least θ times in the database.
Frequent pattern for permutation problems. Permutation problems cover another large range of important combinatorial problems. Many classic NP-hard
problems, such as the traveling salesman problem,
the quadratic assignment problem, graph labeling
problems and flow shop scheduling problems, are
typical permutation examples. Even if the solutions
of this category of problems are represented as permutations, the practical meaning of a permutation
depends on the problem under consideration. To
apply frequent pattern mining techniques to such
a problem, a transformation between a permutation
and an itemset is necessary. For instance, in the context of a vehicle routing problem, a transformation
was proposed to map a permutation (a sequence of
visited cities) into a set of pairs between any two
consecutive elements (cities) in the permutation [16].
This transformation emphasizes the order between
two consecutive elements. In Section 4.2.3, we introduce a new transformation for QAP whose main
idea is to decompose a permutation into a set of
element-position pairs, which focuses on not only
the order between elements but also the relation
between an element and its location. In these settings,
a frequent pattern can be considered as a set of pairs
that frequently appear in some specific solutions.
2.2.2 Mining and heuristics
In this section, we present a brief literature review on hybridizing association rule mining (or frequent pattern mining) techniques with metaheuristics. The reviewed studies
are summarized in Table 1.
GRASP (Greedy randomized adaptive search procedure)
was the first metaheuristic to be hybridized with data mining techniques (i.e., association rule mining) [33], [34]. Denoted as DM-GRASP, this approach was originally designed
to solve the set packing problem. DM-GRASP organizes its
search process into two sequential phases, and incorporates
an association rule mining procedure at the second phase.
Specifically, high-quality solutions found in the first phase
(GRASP) were stored in an elite set, and then a data mining
procedure is invoked to extract some patterns from the elite
set. At the second phase, a new solution is constructed based
on a mined pattern instead of using the greedy randomized
construction procedure of GRASP. This approach has been
applied to solve several problems including the maximum
diversity problem [37], the server replication for reliable
multicast problem [38], and the p-median problem [28]. A
survey on some significant applications of DM-GRASP can
3
TABLE 1
Main work on hybridizing association rules mining techniques with metaheuristics for solving combinatorial optimization problems.
Algorithm
metaheuristic⋆
optimization problem
DM-GRASP [33], [34]
DM-GRASP [37]
DM-GRASP [38]
GADMLS [36]
DM-GRASP [28]
HDMNS [32]
DM-GRASP-PR [7]
MDM-GRASP [27]
DM-HH [25]
VALS [44]
GAAR [31]
GAAR [30]
MDM-GRASP/VND [16]
DM-ILS [26]
GRASP
GRASP
GRASP
GA+LS
GRASP
NS
GRASP+PR
GRASP
PR+LS
LS
GA
GA
GRASP+VND
ILS
set packing problem
maximum diversity problem
server replication for reliable multicast problem
single-vehicle routing problem
p-median problem
p-median problem
2-path network design problem
server replication for reliable multicast problem
p-median problem
set partitioning problem
constrain satisfaction problem
weighted constrain satisfaction problem
one-commodity pickup-and-delivery traveling salesman problem
set covering with pairs problem
⋆
year
2004,2006
2005
2006
2006
2011
2012
2013
2014
2014
2015
2015
2015
2016
2016
Greedy randomized adaptive search procedure (GRASP), genetic algorithm (GA), local search (LS), neighborhood search (NS),
path-relinking (PR), variable neighborhood descent (VND), and iterated local search (ILS).
be found in [39]. An interesting extension of DM-GRASP is
to execute the data mining procedure multiple times instead
of only once [16], [26], [27]. Compared to DM-GRASP where
the data mining call occurs once at the midway of the
whole search process, the multi-mining version performs
the mining task when the elite set stagnates. The same idea
has been explored recently by hybridizing data mining and
GRASP enhanced with path-relinking (PR) [7] or variable
neighborhood descent (VND) [16].
In addition to GRASP, data mining has also been hybridized with other metaheuristics like evolutionary algorithms. To improve the performance of an evolutionary
algorithm applied to an oil collecting vehicle routing problem, a hybrid algorithm (GADMLS) combining genetic algorithm, local search and data mining was proposed in [36].
Another hybrid approach (GAAR) that uses a data mining
module to guide an evolutionary algorithm was presented
in [30], [31] to solve the constraint satisfaction problem
(CSP). Besides the standard components of a genetic algorithm, a data mining module is added to find association
rules (between variables and values) from an archive of best
individuals found in the previous generations.
Apart from GRASP and evolutionary algorithms, it has
been shown that other heuristics can also benefit from the
incorporation of a data mining procedure. For example,
a data mining approach was applied to extract variable
associations from previously solved instances for identifying promising pairs of flipping variables in a large neighborhood search method for the set partitioning problem,
thus reducing the search space of local search algorithms
for the set partitioning problem [44]. Another example is
the hybridization of neighborhood search with data mining techniques for solving the p-median problem [32]. A
data mining procedure was also integrated into a multistart hybrid heuristic for the p-median problem [25], which
combines elements of different traditional metaheuristics
(e.g., local search) and uses path-relinking as a memorybased intensification mechanism. Finally, the widely-used
iterated local search method was recently hybridized with
data mining for solving the set covering with pairs problem
[26].
3
F REQUENT PATTERN BASED SEARCH
In this section, we propose frequent pattern based search
(FPBS) for solving combinatorial search problems. This
general-purpose search approach tightly integrates frequent
pattern mining with an optimization procedure. Below, we
first show the general scheme of the proposed FPBS approach, and then present its composing ingredients.
3.1 General scheme
The FPBS approach uses relevant frequent patterns extracted from high-quality solutions to build promising starting solutions which are further improved by an optimization procedure. The improved solutions are in turn used
to help mine additional patterns. By iterating the mining
procedure and the optimization procedure, FPBS is expected
to examine the search space effectively and efficiently. As a
case study, we illustrate in Section 4 its application to solve
the well-known quadratic assignment problem.
From a perspective of system architecture, FPBS maintains an archive of high-quality solutions (called elite set)
for the purpose of pattern mining and includes five critical
operating components: an initialization procedure (Section
3.2), an optimization search procedure (Section 3.4), a data
mining procedure (Section 3.3), a frequent pattern based
solution construction procedure (Section 3.5) and an elite
solution management procedure (Section 3.6).
The general framework of the proposed FPBS approach
is presented in Algorithm 1. FPBS starts from a set of
high-quality solutions that are obtained by the initialization
procedure (line 3). From these high-quality solutions, a data
mining procedure is invoked to mine a number of frequent
patterns (line 5). A new solution is then constructed based
on a mined pattern and further improved by the optimization procedure (lines 7-10). The improved solution is
finally inserted to the elite set according to the elite solution
management policy (line 15). The process is repeated until a
stopping condition (e.g., a time limit) is satisfied. In addition
to its invocation just after the initialization procedure, the
data mining procedure is also called each time the elite set
is judged to be stagnating (lines 17-18), i.e., it has not been
updated during a predefined number of iterations.
4
Algorithm 1: General framework of Frequent Pattern
Based Search
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
Input: Problem instance I , elite set size k and number of
patterns to be mined m
Output: The best solution S ∗ found
begin
// I is supposed to be a minimization problem
ES ← EliteSetInitialize(); /* ES is the set of elite
solution */
S ∗ ← arg min{f (Si ) : i = 1, 2, . . . , k}; /* S ∗ is the
best solution found so far */
P ← F requentP atternM ine(ES, m);
while a stopping condition is not reached do
p ← P atternSelection(P);
// construct a new solution based mined frequent
pattern S ← P atternBasedConstruct(p);
// improve the constructed solution
S ′ ← Optimize(S);
// update the best solution found so far
if f (S ′ ) < f (S ∗ ) then
S∗ ← S′;
// update the elite set
ES ← EliteSetU pdate(ES, S ′ );
// restart the mining procedure when the elite set
stagnates
if elite set stagnates then
P ← F requentP atternM ine(ES, m);
3.2 Elite set initialization
FPBS starts its search with an elite set (ES ) composed of k
distinct high-quality solutions. To build such an elite set, we
first generate, by any means (e.g., with a random or greedy
construction method), an initial solution that is improved by
an optimization procedure (see Section 3.4). The improved
solution is then inserted into the elite set according to the
elite set management strategy (see Section 3.6). We repeat
this process until an elite set of k different solutions is built.
Note that similar ideas have been successfully applied to
build a high-quality initial population for memetic algorithms [10], [24], [49].
3.3 Frequent pattern mining procedure
In our approach, the frequent pattern mining procedure
is used to discover some patterns that frequently occurs
in high-quality solutions stored in the elite set. Beside the
itemsets (see Section 2.2.1), mined patterns can also be subsequences or substructures [2]. A subsequence is an order
of items, e.g., buying a mobile phone first, then a power
bank, and finally a memory card. If a subsequence occurs
frequently in a transaction database, then it is a frequent
sequential pattern. A substructure can refer to different
structural forms, such as subgraphs, subtrees, or sublattices,
which may be combined with itemsets or subsequences. If
a substructure occurs frequently in a graph database, it is
called a frequent structural pattern.
To handle a wide diversity of data types, numerous
mining tasks and algorithms have been proposed in the
literature [2], [15], [17]. A simple summary of frequent
pattern mining algorithms is provided in Table 2. For a
specific application, the patterns of high-quality solutions
collected in the elite set can be expressed as itemsets,
subsequences, or substructures. Once the form of patterns
is determined, a suitable mining algorithm can be selected
accordingly. For our case study on QAP presented in Section
4, a pattern corresponds to a set of element-location pairs.
Thus frequent patterns can be conveniently represented as
frequent itemsets (i.e., a set of identical element-location assignments). Consequently, we adopt the FPmax* algorithm
(see Section 4.2.3 for more details) to mine only the maximal
frequent items. Detailed reviews on frequent pattern mining
algorithms can be found in [2], [17].
3.4 Optimization procedure
For the purpose of solution improvement (to built the initial
elite set and to improve each new solution built from a
mined pattern), any optimization procedure dedicated to
the given problem can be applied in principle. On the
other hand, since the optimization component ensures the
key role of search intensification, it is desirable to call for
a powerful search algorithm. Basically, the optimization
procedure can be considered to be a black box optimizer
that is called to improve the input solution. In practice,
the optimization procedure can be based on local search,
population-based search or even hybrid memetic search. In
any case, the search procedure must be carefully designed
with respect to the problem under consideration and should
ideally be effective both in terms of search capacity and
time efficiency. As we show in Section 4, for the quadratic
assignment problem considered in this work, we will adopt
the powerful breakout local search procedure [9] as our
optimization procedure.
3.5 Solution construction based on mined pattern
Once a set of frequent patterns P is extracted from the elite
set, new solutions are constructed based on these mined
patterns. For this purpose, we first select a mined frequent
pattern by the tournament selection strategy as follows. Let
λ be the size of the tournament pool. We randomly choose
λ (1 6 λ 6 |P|) individuals with replacement from the
mined pattern set P , and then pick the best one (i.e., with
the largest size), where λ is a parameter. The computational
complexity of this selection strategy is O(|P|). The advantage of the tournament selection strategy is that the selection
pressure can be easily adjusted by changing the size of the
tournament pool λ. The larger the tournament pool is, the
smaller the chance for shorter patterns to be selected.
Since frequent patterns usually correspond to a set of
common elements shared by the high-quality solutions
examined by the mining procedure, each mined pattern
directly defines a partial solution. To obtain a whole solution, we can apply a greedy or random procedure to
complete the partial solution. The way to build such a
solution shares similarity with the general backbone-based
crossover procedure [8], [49]. However, compared to the
notion of backbones that are typically shared by two parent
solutions, our frequent patterns are naturally shared by two
or more high-quality solutions. In this sense, backbones can
be considered as a special case of more general frequent
patterns.
5
TABLE 2
A simple summary of frequent pattern mining algorithms [2].
algorithms
tasks
patterns
apriori-based
pattern-growth
frequent itemset mining
sequential pattern mining
structural pattern mining
itemsets
subsequences
substructures
e.g., Apriori
e.g., GSP, SPADE
e.g., AGM, FSG
e.g., FP-growth, FPmax
e.g., PrefixSpan
e.g., gSpan, FFSM
Finally, it is possible to construct a new solution for each
mined pattern instead of using a long pattern selected by the
tournament selection strategy, as explained above. Also, an
elite solution can be selected to guide the construction of a
new solution. In particular, the way to use frequent patterns
should be determined according to the studied problem.
3.6 Elite set management
As explained above, each new solution constructed using
a mined frequent pattern is improved by the optimization
procedure. Then, we decide whether the improved solution
should be inserted into the elite set ES .
There are a number of updating strategies in the literature [41] that can be applied within the FPBS approach.
For example, the classic quality-based replacement strategy
simply inserts the solution into ES to replace the worst
solution if it is better than the worst solution in ES .
In addition, more elaborated updating strategies consider
other criteria than the quality of solutions. For example, the
quality-and-distance updating strategy not only considers
the quality of the solution, but also its distance to other
solutions in the population [24], [29], [49]. A suitable elite
set management strategy can be determined according to
the practical problem.
4
where a and b are the flow and distance matrices respectively, and π ∈ Ω is a solution and π(i) represents the location chosen for facility i. The optimization objective is to find
a permutation π ∗ in Ω such that the sum of the products of
the flow and distance matrices, i.e., f (π ∗ ) 6 f (π), ∀π ∈ Ω,
is minimized.
In addition to the facility location application, QAP can
be used to formulate many other real-world problems [12],
[13], [23] such as electrical circuit wiring/routing, transportation engineering, parallel and distributed computing,
image processing and analysis of chemical reactions for
organic compounds. Moreover, a number of classic NPhard problems, such as the traveling salesman problem, the
maximum clique problem, the bin packing problem and the
graph partitioning problem, can also be recast as QAPs [23].
Due to its practical and theoretical significance, QAP has
attracted much research effort since its first formulation [1],
[9], [10], [20], [43]. In fact, QAP is among the most studied
and the most competitive combinatorial optimization problems. Since exact algorithms are unpractical for instances
of size larger than 36 [5], a large number of heuristic
methods have been proposed for QAP, which can provide
near-optimal solutions in reasonable computation times.
Detailed reviews of heuristic and metaheuristic algorithms
developed till 2007 for QAP are available in [12], [23]. An
updated review of more recent studies on QAP can be found
in [9], [10].
FPBS APPLIED TO THE QUADRATIC ASSIGN 4.2 FPBS for QAP
MENT PROBLEM
In this section, we present a case study of applying the general FPBS approach to the well-known quadratic assignment
problem (QAP) and show its competitiveness compared to
state-of-the-art QAP algorithms.
4.1 Quadratic assignment problem
The quadratic assignment problem is a well-known NPhard combinatorial optimization problem. It was originally
introduced by Koopmans and Beckman [22] in 1957 to
model the locations of indivisible economic activities such
as capital equipment. QAP aims to determine a minimal
cost assignment of n facilities to n locations, given a flow
aij from facility i to facility j for all i, j ∈ {1, . . . , n}
and a distance buv between locations u and v for all
u, v ∈ {1, . . . , n}. Let Ω denotes the set of all possible
permutations π : {1, . . . , n} → {1, . . . , n}, then QAP can
mathematically be formulated as follows.
min f (π) =
π∈Ω
n
n X
X
i=1 j=1
aij bπ(i)π(j)
(1)
Algorithm 2 shows the FPBS algorithm for QAP (denoted
as FPBS-QAP), which is an instantiation of the general
scheme of Algorithm 1. Since FPBS-QAP inherits the main
components of FPBS, hereafter we only present the specific
features related to QAP: solution representation and evaluation, optimization procedure, frequent pattern mining for
QAP, solution construction using QAP patterns, and elite set
update strategy.
4.2.1 Solution representation, neighborhood and evaluation
Given a QAP instance with n facilities (or locations), a candidate solution is naturally represented by a permutation
π of {1, 2, . . . , n}, such that π(i) is the location assigned to
facility i. The search space Ω is thus composed of all possible
n! permutations. For any solution π ∈ Ω, its quality is given
by Eq. (1).
To examine the search space, we adopt an iterated local
search algorithm called BLS (see Section 4.2.2). For this
purpose, we introduce the neighborhood used by BLS. Give
a solution, i.e., a permutation π , its neighborhood N (π)
is defined as the set of all possible permutations that can
6
Algorithm 2: The FPBS algorithm for QAP.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
Input: Instance G, elite set size k, the number of mined
patterns m, time limit tmax and the maximum
number of iterations without updating
max no update
Output: The best solution π ∗ found so far
begin
ES ← EliteSetInitialize();
π ∗ ← arg min{f (πi ) : i = 1, 2, . . . , k};
P ← F requentP atternM ine(ES, m);
no update ← 0;
t ← 0;
while t < tmax do
pi ← P atternSelection(P);
// build a new solution based selected pattern
π ← P atternBasedConstruct(pi);
// improve the constructed solution
π ′ ← BreakoutLocalSearch(π);
// update the best solution found so far
if f (π ′ ) < f (π ∗ ) then
π∗ ← π′;
// update the elite set
if EliteSetU pdate(ES, π ′ ) = T rue then
no update ← 0;
else
no update ← no update + 1;
// restart the mining procedure when the elite set
is steady
if no update > max no update then
P ← F requentP atternM ine(ES, m);
no update ← 0;
be obtained by exchanging the values of any two different
positions π(u) and π(v) in π , i.e., N (π) = {π ′ | π ′ (u) =
π(v), π ′ (v) = π(u), u 6= v and π ′ (i) = π(i), ∀i 6= u, v},
which has a size of n(n − 1)/2.
As indicated in [42], given a permutation π and its
objective value f (π), the objective value of any neighnoring
permutation π ′ can be effectively calculated according to an
incremental evaluation technique.
4.2.2 Breakout local search
To ensure an effective examination of the search space, we
adopt the breakout local search (BLS) algorithm [9], which
is a state-of-the-art algorithms for QAP currently available
in the literature.
BLS follows the iterated local search scheme and iteratively alternates between a descent search phase (to find
local optima) and a dedicated perturbation phase (to discover new promising regions). BLS starts from an initial
random permutation, and then improves the initial solution
to a local optimum by the best improvement descent search
with the above exchange-based neighborhood. Upon the
discovery of a local optimum, BLS triggers a perturbation
mechanism. The perturbation mechanism adaptively selects
a tabu-based perturbation (called directed perturbation) or a
random perturbation (called undirected perturbation). BLS
also determines the number of perturbation steps (called
perturbation strength) in an adaptive way [9].
The tabu-based perturbation and random perturbation
provide two complementary means for search diversifica-
tion. The former applies a selection rule that favors neighboring solutions that minimize the objective degradation,
under the constraint that the neighboring solutions have
not been visited during the last γ iterations (where γ is the
pre-defined tabu tenure), while the latter performs moves
selected uniformly at random. To keep a suitable balance
between an intensified search and a diversified search, BLS
alternates probabilistically between these two perturbations.
The probability to select a particular perturbation is determined dynamically according to the current number of
times to visit local optima without improvement on the best
solution found. The probability of applying the tabu-based
perturbation over the random perturbation is empirically
limited to be at least as Q. The perturbation strength L
is determined based on a simple reactive strategy. One
increases L by one if the search returns to the immediate
previous local optimum, and otherwise resets L to a given
initial value L0 . Once the type and the strength L of the
perturbation are determined, the selected perturbation with
strength L is applied to the current solution. The resulting
solution is then used as the starting solution of the next
round of the descent search procedure (see [9] for more
details).
4.2.3 Mining frequent patterns for QAP
The quadratic assignment problem is a typical permutation
problem whose solutions are naturally represented by permutations. For QAP, we define a frequent pattern to be a
set of identical location-facility assignments shared by highquality solutions, and represent a frequent pattern by an
itemset. To apply a frequent itemset mining algorithm, we
need to transform a permutation into a set of items. A transformation is recently proposed in [16]. The transformation
works as follows. For each pair of elements (π(i) and π(j))
of a given permutation π , an arc (π(i), π(j)) is generated,
which maps a permutation π to a set S ′ of |π| − 1 arcs. For
example, considering a permutation π = {5, 4, 7, 2, 1, 6, 3}
represented by a sequence of these elements, π can be
mapped as S ′ = {(5, 4), (4, 7), (7, 2), (2, 1), (1, 6), (6, 3)}. By
this transformation, a permutation is divided into a set of
arcs, which conserves the order of elements. However, this
transformation loses the information between the elements
and their locations. In practice, we can not identify the
true location of an element when only a part of pairs are
available.
To overcome this difficulty, we propose a new transformation. Our transformation decomposes a permutation into
a set of ordered pairs, each pair being formed by an element
(facility) i and its position (location) π(i). Specifically, let π
be a candidate solution of QAP, for each element (facility) i,
a corresponding element-position pair (i, π(i)) is generated,
which transforms permutation π into a set of n elementposition pairs. For example, given π = {5, 4, 7, 2, 1, 6, 3},
{(1, 5), (2, 4), (3, 7), (4, 2), (5, 1), (6, 6), (7, 3)} is the corresponding set of element-position pairs.
Once a permutation is transformed into a set of pairs,
we will treat each pair (x, y) as an item z = x ∗ n + y ,
where n is the length of the permutation. Now, the task of
mining frequent patterns from multiple permutations can be
conveniently transformed into the task of mining frequent
itemsets. The main drawback of mining all frequent itemsets
7
is that if there is a large frequent itemset, then almost all
subsets of the itemset might be examined. However, it is
usually sufficient to find only the maximal frequent itemsets
(a maximal frequent itemset is that it has no superset that is
frequent). Thus mining frequent itemsets can be reduced to
mine only the maximal frequent itemsets. For this purpose,
we adopt the popular FPmax*1 algorithm [14].
Algorithm 3: Pseudo code of the FPmax* Algorithm
[14]
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
Input: T : a FP-tree
M : the MFI-tree for T.base.
Output: Updated M
begin
if T only contains a single path P then
insert P into M ;
else
for each i in T .header do
set Y = T.base ∪ {i};
if T.array is not N ull then
T ail={frequent items for i in T.array};
4.2.4 Solution construction based on mined pattern
Algorithm 4 describes the main steps to construct new
solutions based on the mined frequent patterns. Initially,
a pattern is first selected from the set of mined frequent
patterns P according to the tournament selection strategy
(line 2). Then, we re-map this chosen pattern into a partial
solution π (line 3). If the length of the partial solution |π|
is less than a given threshold (i.e., β ∗ n), we use an elite
solution to guide the construction (lines 4-6). Specifically,
we first select an elite solution π 0 (called guiding solution)
from the elite set (line 5), and then we complete π based
on the guiding solution π 0 (i.e., directly copy the elements
of all unassigned positions of π 0 to π if the elements have
not been assigned in π ). Finally, if π is still an incomplete
solution, we randomly assign the remaining elements to the
unassigned positions until a full solution is obtained (line
7).
Algorithm 4: Solution construction based on mined
frequent patterns
else
T ail={frequent items in i’s conditional
pattern base};
sort T ail in decreasing order of the items’
counts;
if subset checking(Y ∪ T ail, M ) = F alse then
construct Y’s conditional FP-tree TY and
its array AY ;
initialize Y’s conditional MFI-tree MY ;
call FPmax*(TY , MY );
merge MY with M ;
1
2
3
4
5
6
7
Algorithm 3 shows the pseudo code of the FPmax* algorithm. A description of the basic concepts used in FPmax*
such as the frequent pattern tree (FP-tree), the FP-growth
method, and the maximal frequent itemset tree (MFI-tree)
can be found in [14]. Here we only provide a brief presentation of its general procedure. In the initial call, a FP-tree T
is constructed from the first scan of the database, together
with an initial empty MFI-tree. During the recursion, if there
is only one single path in the FP-tree T , this single path
together with T.base is a MFI of the dataset. The MFI is
then inserted into M (line 3). Otherwise, for each item i
in the header table, we set Y = T.base ∪ {i} and prepare
for the recursive call FPmax*(TY , MY ). The items in the
header table are processed in increasing order of frequency,
so that maximal frequent itemsets will be found before
any of their frequent subsets (line 11). Lines 7-10 use the
array technique, and line 12 invokes the subset checking()
function to check if Y together with all frequent items in
Y ’s conditional pattern base is a subset of any existing
MFI in M , thus we perform superset pruning. If function
subset checking() returns F alse, FPmax* will be called
recursively, with (TY , MY ) (lines 13-16). For a detailed
description of the FPmax* algorithm, please refer to [14].
Input: Instance G, a set of m mined patterns P and an
elite set ES of size k
Output: A new solution π
begin
p ← P atternSelection(P); /∗ select a mined pattern
∗/
π ← re-map(p); /∗ generate a partial solution based
on selected pattern ∗/
if |π| < β ∗ n then
π 0 ← GuidedSolutionSelect(ES); /∗ select a
guiding solution ∗/
π ← GuidedComplete(π, π 0 ); /∗ complete based
on guiding solution ∗/
π ← RandomComplete(π); /∗ complete at random ∗/
4.2.5 Elite set update strategy
Once a new improved solution π ′ is obtained by the BLS
algorithm, we decide whether π ′ should be inserted into the
elite set ES . In our case, we adopt the following qualitybased strategy that π ′ is inserted into ES if two conditions
are satisfied simultaneously: (i) π ′ is different from any
solution in ES and (ii) π ′ is not worse than any solution in
ES , i.e., f (π ′ ) 6 f (π w ), where π w ← arg maxπ∈ES {f (π)}
is the worst solution in the ES .
4.3 Computational studies of FPBS for QAP
Our computational studies aim to evaluate the efficiency of
the FPBS-QAP algorithm. For this purpose, we first perform
a detailed performance comparisons between FPBS-QAP
and two state-of-the-art algorithms, i.e., BLS [9] and BMA
[10], whose source codes are available to us. Furthermore,
we compare FPBS-QAP with four additional algorithms that
are published very recently since 2015.
4.3.1 Benchmark instances
1. The source code of the FPmax* algorithm is publicly available at
http://fimi.ua.ac.be/src/
The experimental evaluations of QAP algorithms are usually performed on 135 popular benchmark instances from
8
QAPLIB2 , with n ranging from 12 to 150. These instances
can be classified into four categories:
•
•
•
•
Type I. 114 real-life instances are obtained from
practical QAP applications;
Type II. 5 unstructured, randomly generated instances whose distance and flow matrices are randomly generated based on a uniform distribution;
Type III. 5 real-like-life instances are generated instances that are similar to the real-life QAP instances;
Type IV. 11 instances with grid-based distances
in which the distances are the Manhattan distance
between points on a grid.
Like [9], [10], we do not consider the 114 instances from
Type I because they are very easy for modern QAP algorithms (i.e., their optimal solutions can be found easily with
a short time, often less than one second). Our experiments
focus on the remaining 21 hard instances from Type II,
Type III and Type IV. Notice that for these 21 challenging
instances, no single algorithm can attain the best-known results for all the instances. Indeed, even the best performing
algorithm misses at least two best-known results.
4.3.2 Experimental settings
The proposed FPBS-QAP algorithm3 was implemented in
the C++ programming language and complied with gcc
4.1.2 and flag ‘-O3’. All the experiments were carried out
on a computer equipped with an Intel E5-2670 processor
with 2.5 GHz and 2 GB RAM operating under the Linux
system. Without using any compiler flag, running the wellknown DIMACS machine benchmark procedure dfmax.c4
on our machine requires 0.19, 1.17 and 4.54 seconds to solve
the benchmark graphs r300.5, r400.5 and r500.5 respectively.
Our computational results were obtained by running the
FPBS-QAP algorithm with the parameter settings provided
in Table 3. To identify an appropriate value for a given
parameter, we compare the performance of the algorithm
with different parameter values, while fixing other parameter values. An example to select a appropriate m (frequent
pattern set size) value is provided in Section 5.3.
TABLE 3
Parameter settings of FPBS-QAP algorithm.
Parameter
description
value
tmax
k
max no update
θ
m
λ
β
max iter
time limit (hours)
elite set size
number of times without updating
minimum support
frequent pattern set size
tournament pool size
length threshold
number of iterations for BLS⋆
0.5 or 2.0
15
15
2
11
3
0.75
10000
⋆
Other parameters of BLS adopt the default values provided in [9].
Given its stochastic nature, the proposed FPBS-QAP
algorithm was independently ran 10 times on each test
instance, which is a standard practice for solving QAP [1],
[9], [10], [43]. Our assessment is based on the percentage
2. https://www.opt.math.tugraz.at/qaplib/
3. The source code of our FPBS-QAP algorithm will be made available at http://www.info.univ-angers.fr/∼hao/fpbs.html
4. dfmax: ftp://dimacs.rutgers.edu/pub/dsj/clique
deviation (PD) metrics that are widely used in the literature.
The PD metrics measures the percentage deviation from the
best-known value (BKV). For example, the best percentage
deviation (BAD), the average percentage deviation (APD)
and the worst percentage deviation (WPD), are respectively
calculated according to:
XP D = 100 ∗
X − BKV
[%]
BKV
(2)
where X ∈ {B, A, W } corresponds to the best objective
value, average objective value and worst objective value
achieved by an algorithm. The smaller the XPD value, the
better the evaluated algorithm.
4.3.3 Comparison of FPBS-QAP with BLS and BMA
To demonstrate the effectiveness of the FPBS-QAP algorithm, we first show a detailed comparison of FPBS-QAP
with two main reference algorithms: BLS (breakout local
search) [9] and BMA (population-based memetic algorithm)
[10]. This experiment was based on two motivations. First,
BLS and BMA are among the best performing QAP algorithms in the literature. Second, the source codes of BLS
and BMA are available to us, making it possible to make
a fair comparision (using the same computing platform
and stopping conditions). Third, both FPBS-QAP and BMA
use BLS as their underlying optimization procedure, this
comparison allows us to assess the added value of the data
mining component of FPBS-QAP. For this experiment, we
run FPBS-QAP and the two reference algorithms under two
stopping conditions, i.e., a limit of tmax = 30 minutes (0.5
hour) and a limit of tmax = 120 minutes (2 hours). This
allows us to study the behavior of the compared algorithms
under short and long conditions.
The comparative performances of FPBS-QAP with BLS
and BMA under tmax = 30 minutes and tmax = 120
minutes are presented in Table 4 and Table 5, respectively.
In these two tables, we report the BPD, APD, WPD values
of each algorithm, and the average time in minutes (T (m))
to achieve the results. At the last row of each table, we also
indicate the average value of each indicator. The smaller the
value, the better the performance of an algorithm.
From Table 4, we observe that FPBS-QAP achieves the
best performance compared to the algorithms BLS and BMA
under tmax = 30 minutes. First, FPBS-QAP achieves all
best-known values except two cases while BLS and BMA
fail to find the best-known values for five instances. Second,
FPBS-QAP is able to reach the best-known values of the two
largest instances, i.e., tai150b and tho150 within the given
computing time. BLS fails to find the best-known values
of tai150b and tho150 within the time limit of 30 minutes
(it can find these values only under a very long time limit
of tmax = 10 hours). BMA performs better than BLS by
attaining the best-known value of tai150b, but still fails on
tho150. The BPD value of FPBS-QAP for the 21 benchmark
instances is only 0.043%, which is smaller than 0.059% of
BLS, and 0.051% of BMA respectively. Similar observations
can also be found for the APD and WPD indicators. It is
worth noting that FPBS-QAP needs less time to achieve
these (better) results than BLS, while it consumes nearly the
same computing time as BMA.
9
TABLE 4
Performance comparison of the proposed FPBS-QAP algorithm with BLS and BMA on 21 hard instances under tmax = 30 minutes. The number
of times of reaching the best-known value over 10 runs is indicated in parentheses.
BLS
BMA
Instance
BKV
BPD
APD
WPD
T (m)
tai40a
tai50a
tai60a
tai80a
tai100a
tai50b
tai60b
tai80b
tai100b
tai150b
sko72
sko81
sko90
sko100a
sko100b
sko100c
sko100d
sko100e
sko100f
wil100
tho150
3139370
4938796
7205962
13499184
21052466
458821517
608215054
818415043
1185996137
498896643
66256
90998
115534
152002
153890
147862
149576
149150
149036
273038
8133398
0.000(1)
0.000(1)
0.273(0)
0.474(0)
0.467(0)
0.000(10)
0.000(10)
0.000(10)
0.000(9)
0.014(0)
0.000(10)
0.000(9)
0.000(4)
0.000(4)
0.000(8)
0.000(7)
0.000(3)
0.000(5)
0.000(5)
0.000(6)
0.013(0)
0.067
0.181
0.371
0.571
0.566
0.000
0.000
0.000
0.005
0.219
0.000
0.001
0.023
0.006
0.001
0.001
0.004
0.002
0.006
0.001
0.091
0.074
0.364
0.442
0.647
0.630
0.000
0.000
0.000
0.045
0.390
0.000
0.011
0.095
0.018
0.004
0.004
0.009
0.005
0.032
0.003
0.128
5.5
14.5
17.7
13.5
16.4
0.2
0.6
3.3
10.3
14.5
4.4
8.0
9.5
12.4
4.6
10.2
13.5
13.7
11.5
11.4
11.9
0.059
0.101
0.138
9.8
avg.
APD
WPD
T (m)
0.000(1)
0.053(0)
0.165(0)
0.468(0)
0.389(0)
0.000(10)
0.000(10)
0.000(10)
0.000(6)
0.000(1)
0.000(9)
0.000(10)
0.000(9)
0.000(8)
0.000(10)
0.000(10)
0.000(9)
0.000(10)
0.000(6)
0.000(10)
0.003(0)
0.067
0.216
0.285
0.553
0.510
0.000
0.000
0.000
0.044
0.137
0.006
0.000
0.004
0.003
0.000
0.000
0.007
0.000
0.005
0.000
0.021
0.074
0.317
0.381
0.621
0.623
0.000
0.000
0.000
0.143
0.316
0.063
0.000
0.038
0.016
0.000
0.000
0.066
0.000
0.021
0.000
0.067
7.0
12.8
17.6
14.5
19.0
0.2
0.3
1.8
2.4
21.5
0.7
3.5
5.9
5.2
8.5
6.9
6.6
6.5
6.3
6.4
19.9
0.051
0.095
0.139
8.2
When a long time limit of tmax = 120 minutes is
allowed, our FPBS-QAP algorithm is able to achieve even
better results. As we see from Table 5, the best-known values
are obtained with a higher success rate compared to the
results under tmax = 30 minutes in Table 4. The average
BPD value of FPBS-QAP is 0.028%, which is the smallest
compared to 0.037% of BMA, and 0.038% of BLS. FPBSQAP also achieves the smallest average APD value and
average WPD value. As to the computing times, FPBS-QAP
requires on average 22.0 minutes to reach its best solution,
which is the shortest time among the compared algorithms
(32.2 minutes for BLS, and 23.1 minutes for BMA).
In summary, our FPBS-QAP algorithm competes favorably with the two best-performing QAP algorithms (i.e., BLS
and BMA) in terms of both solution quality and computing
time. The computational results demonstrate the effectiveness of the proposed FPBS-QAP algorithm, and further
shows the usefulness of using frequent patterns mined from
high-quality solutions to guide the search for an effective
exploration of the solution space.
4.3.4 Comparison with four other state-of-the-art algorithms
We now extend our experimental study by comparing FPBSQAP with four other very recent state-of-the-art QAP algorithms in the literature.
•
•
FPBS-QAP
BPD
Parallel hybrid algorithm (PHA) [43] (2015) used
the MPI libraries and was implemented on a highperformance cluster with 46 nodes, a total RAM of
736 GB and a total disk capacity of 6.5 TB configured
in a high-performance RAID. Each node includes 2
CPUs (4 cores per CPU) and 16 GB of RAM.
Two-stage memory powered great deluge algorithm
(TMSGD) [1] (2015) was implemented on a personal
computer with 2.1 GHz and 8 GB RAM. The algorithm was stopped when the number of fitness
evaluations reaches 20000 ∗ n (n is the instance size).
•
•
BPD
APD
WPD
T (m)
0.000(1)
0.000(1)
0.165(0)
0.430(0)
0.311(0)
0.000(10)
0.000(10)
0.000(10)
0.000(6)
0.000(1)
0.000(10)
0.000(10)
0.000(6)
0.000(10)
0.000(10)
0.000(10)
0.000(10)
0.000(7)
0.000(7)
0.000(10)
0.000(1)
0.067
0.279
0.377
0.516
0.402
0.000
0.000
0.000
0.040
0.191
0.000
0.000
0.015
0.000
0.000
0.000
0.000
0.001
0.003
0.000
0.051
0.074
0.415
0.469
0.614
0.553
0.000
0.000
0.000
0.100
0.321
0.000
0.000
0.038
0.000
0.000
0.000
0.000
0.004
0.021
0.000
0.123
6.4
17.3
8.7
18.2
13.4
0.1
0.5
1.2
4.0
20.5
1.4
2.7
4.2
7.7
7.7
8.7
9.0
10.3
4.4
9.5
24.2
0.043
0.092
0.130
8.6
Parallel multi-start hyper-heuristic algorithm (MSH)
[11] (2016) was implemented on the same high performance cluster as the above PHA algorithm.
Breakout local search using OpenMP (BLS-OpenMP)
[4] (2017) was implemented on OpenMP (i.e., an API
for shared-memory parallel computations that runs
on multi-core computers.) and was executed on a
personal computer with an Intel Core i7-6700 CPU
3.4 GHZ with 4 cores and 16 GB RAM. It is possible
to execute 8 logical processors on this computer.
One notices that three of these four recent QAP algorithms are implemented and run on parallel machines.
Moreover, their results have been obtained on different computing platforms, with different stopping conditions. The
comparison shown in this section is provided mainly for indicative purposes. Still, the comparison provides interesting
indications on the performance of the proposed algorithm
relative to these state-of-the-art algorithms. Moreover, the
availability of our FPBS-QAP code makes it possible for
researchers to make fair comparisons with FPBS-QAP (see
footnote 3).
Table 6 presents the comparative results between the
proposed FPBS-QAP algorithm and the four reference algorithms. Like [1], [4], [11], we adopt the APD indicator
(defined in Section 4.3.2) for this comparative study and
indicate the running time (T (m)) as an additional indicator,
which should be interpreted with cautions for the reasons
raised above. For completeness, we also include the results
of BLS and BMA from Table 6. In the last row of the table,
we again indicate the average value of each indicator.
From Table 6, we observe that our FPBS-QAP algorithm
achieves a highly competitive performance compared to
these state-of-the-art algorithms. The average APD value
of FPBS-QAP is 0.061%, which is only slightly worse than
0.058% of the parallel PHA algorithm and better than all
remaining reference algorithms. Moreover, even if PHA was
10
TABLE 5
Performance comparison of the proposed FPBS-QAP algorithm with BLS and BMA on 21 hard instances under tmax = 120 minutes. The number
of times of reaching the best-known value over 10 runs is indicated in parentheses.
BLS
BMA
Instance
BKV
BPD
APD
WPD
T (m)
tai40a
tai50a
tai60a
tai80a
tai100a
tai50b
tai60b
tai80b
tai100b
tai150b
sko72
sko81
sko90
sko100a
sko100b
sko100c
sko100d
sko100e
sko100f
wil100
tho150
3139370
4938796
7205962
13499184
21052466
458821517
608215054
818415043
1185996137
498896643
66256
90998
115534
152002
153890
147862
149576
149150
149036
273038
8133398
0.000(7)
0.000(1)
0.036(0)
0.416(0)
0.335(0)
0.000(10)
0.000(10)
0.000(10)
0.000(10)
0.001(0)
0.000(10)
0.000(10)
0.000(10)
0.000(10)
0.000(10)
0.000(10)
0.000(6)
0.000(10)
0.000(7)
0.000(9)
0.009(0)
0.022
0.100
0.233
0.502
0.460
0.000
0.000
0.000
0.000
0.040
0.000
0.000
0.000
0.000
0.000
0.000
0.001
0.000
0.002
0.000
0.056
0.074
0.245
0.331
0.587
0.560
0.000
0.000
0.000
0.000
0.183
0.000
0.000
0.000
0.000
0.000
0.000
0.005
0.000
0.005
0.002
0.116
0.038
0.067
0.100
avg.
FPBS-QAP
BPD
APD
WPD
T (m)
BPD
APD
WPD
T (m)
40.7
47.0
73.9
58.0
58.4
0.2
0.3
3.9
6.1
42.7
3.0
10.3
19.6
51.0
21.6
22.4
38.5
44.2
40.2
28.9
64.2
0.000(5)
0.000(3)
0.165(0)
0.332(0)
0.223(0)
0.000(10)
0.000(10)
0.000(10)
0.000(6)
0.055(0)
0.000(10)
0.000(10)
0.000(9)
0.000(8)
0.000(10)
0.000(10)
0.000(10)
0.000(10)
0.000(8)
0.000(10)
0.000(1)
0.037
0.098
0.221
0.428
0.370
0.000
0.000
0.000
0.040
0.213
0.000
0.000
0.004
0.003
0.000
0.000
0.000
0.000
0.001
0.000
0.036
0.074
0.291
0.352
0.505
0.511
0.000
0.000
0.000
0.100
0.414
0.000
0.000
0.038
0.016
0.000
0.000
0.000
0.000
0.005
0.000
0.065
28.9
38.8
34.7
69.9
59.9
0.1
0.4
2.0
8.1
56.0
0.7
3.2
3.6
28.6
11.0
7.1
12.6
5.3
23.7
6.9
83.9
0.000(7)
0.000(2)
0.000(1)
0.313(0)
0.280(0)
0.000(10)
0.000(10)
0.000(10)
0.000(10)
0.000(5)
0.000(10)
0.000(10)
0.000(9)
0.000(10)
0.000(10)
0.000(10)
0.000(10)
0.000(10)
0.000(6)
0.000(10)
0.000(3)
0.022
0.107
0.216
0.451
0.378
0.000
0.000
0.000
0.000
0.099
0.000
0.000
0.004
0.000
0.000
0.000
0.000
0.000
0.002
0.000
0.008
0.074
0.231
0.300
0.618
0.466
0.000
0.000
0.000
0.000
0.313
0.000
0.000
0.038
0.000
0.000
0.000
0.000
0.000
0.005
0.000
0.064
52.5
67.8
60.0
55.2
36.1
0.2
0.4
1.4
2.9
46.4
4.8
3.3
2.4
8.5
5.8
8.7
16.2
12.2
4.0
16.4
57.4
32.2
0.037
0.069
0.113
23.1
0.028
0.061
0.100
22.0
TABLE 6
Comparative performance between the FPBS-QAP algorithm and state-of-the-art algorithms on hard instances in terms of the APD value.
Computational time are given in minutes for indicative purposes.
BLS⋆
PHA◦
BMA⋆
TMSGD
MSH◦
BLS-OpenMP◦
FPBS-QAP
Instance
BKV
APD
T (m)
APD
T (m)
APD
T (m)
APD
T (m)
APD
T (m)
APD
T (m)
APD
T (m)
tai40a
tai50a
tai60a
tai80a
tai100a
tai50b
tai60b
tai80b
tai100b
tai150b
sko72
sko81
sko90
sko100a
sko100b
sko100c
sko100d
sko100e
sko100f
wil100
tho150
3139370
4938796
7205962
13499184
21052466
458821517
608215054
818415043
1185996137
498896643
66256
90998
115534
152002
153890
147862
149576
149150
149036
273038
8133398
0.022
0.100
0.233
0.502
0.460
0.000
0.000
0.000
0.000
0.040
0.000
0.000
0.000
0.000
0.000
0.000
0.001
0.000
0.002
0.000
0.056
40.7
47.0
73.9
58.0
58.4
0.2
0.3
3.9
6.1
42.7
3.0
10.3
19.6
51.0
21.6
22.4
38.5
44.2
40.2
28.9
64.2
0.000
0.000
0.000
0.644
0.537
0.000
0.000
0.000
0.000
0.026
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.009
10.6
12.7
19.6
40.0
71.9
5.8
9.5
27.7
42.5
177.4
33.6
39.9
40.5
41.7
42.3
42.2
41.9
42.5
42.0
42.0
177.4
0.037
0.098
0.221
0.428
0.370
0.000
0.000
0.000
0.040
0.213
0.000
0.000
0.004
0.003
0.000
0.000
0.000
0.000
0.001
0.000
0.036
28.9
38.8
34.7
69.9
59.9
0.1
0.4
2.0
8.1
56.0
0.7
3.2
3.6
28.6
11.0
7.1
12.6
5.3
23.7
6.9
83.9
0.261
0.276
0.448
0.832
0.874
0.005
0.000
0.025
0.028
0.051
0.007
0.019
0.031
0.029
0.015
0.013
0.017
0.016
0.013
0.008
0.039
27.8
41.1
78.9
111.3
138.3
10.2
33.6
0.0
72.6
258.0
38.0
57.1
93.8
153.2
164.3
154.5
148.9
146.1
153.4
155.1
512.8
0.261
0.165
0.270
0.530
0.338
0.000
0.000
0.000
0.000
*
0.000
0.000
0.000
0.003
0.004
0.003
0.004
0.000
0.000
*
*
30.0
37.5
45.0
60.0
75.0
3.0
3.2
4.0
5.0
*
3.6
4.1
4.5
75.0
75.0
75.0
75.0
75.0
75.0
*
*
0.000
0.000
0.000
0.504
0.617
0.000
0.000
0.000
0.000
*
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
*
*
32.2
68.2
107.9
236.0
448.5
0.7
18.6
218.1
160.8
*
1.8
2.4
3.3
29.8
8.5
4.3
12.9
4.3
17.1
*
*
0.022
0.107
0.216
0.451
0.378
0.000
0.000
0.000
0.000
0.099
0.000
0.000
0.004
0.000
0.000
0.000
0.000
0.000
0.002
0.000
0.008
52.5
67.8
60.0
55.2
36.1
0.2
0.4
1.4
2.9
46.4
4.8
3.3
2.4
8.5
5.8
8.7
16.2
12.2
4.0
16.4
57.4
0.067
32.2
0.058
47.8
0.069
23.1
0.143
121.4
0.088
40.3
0.062
76.4
0.061
22.0
avg.
⋆
The results of BLS and BMA were obtained by running the programs on our computer with tmax = 120 minutes. These results are slightly different from
the results reported in [9], [10].
◦
PHA, MSH and BLS-OpenMP are parallel algorithms that were run on high-performance platforms under various stopping conditions.
run on a parallel high-performance computing platform, its
average computing time (≥ 177.4 minutes) is almost three
times of the time required by FPBS-QAP (≤ 67.8 minutes)
to obtain very similar results. Note that the results of three
instances, including the two hardest and largest instances
tai150b and tho150, are not reported for BLS-OpenMP. The
APD value of BLS-OpenMP is computed for the remaining
18 instances. Importantly, our algorithm requires the least
time to achieve the best results, and its average time is only
22.0 minutes. These observations show that our FPBS-QAP
algorithm is highly competitive compared to the state-of-
the-art algorithms in terms of solution quality and computing time.
5
A NALYSIS AND DISCUSSION
In this section, we perform additional experiments to gain
understandings of the proposed FPBS algorithm including
the rationale behind the solution construction based on frequent patterns, the effectiveness of the solution construction
based on mined frequent patterns, and the impact of the
number of the largest patterns m on the performance of the
proposed algorithm.
11
5.1 Rationale behind the solution construction based
on mined patterns
To explain the rationale behind the solution construction
based on mined frequent patterns, we analyze the structural
similarity between high-quality solutions in the elite set, and
the length distribution of the frequent patterns mined from
the elite set. Given two high-quality solutions π s and π t , we
define their similarity as follows.
5.2 Effectiveness of the solution construction based
on frequent pattern
The frequent pattern based solution construction method
is a good alternative to the general crossover operator
0.8
0.8
0.6
0.6
0.4
0.4
0.2
max_sim
avg_sim
min_sim
0
0.2
max_len
avg_len
min_len
0
tho150
wil100
sko100f
sko100e
sko100d
sko100c
sko100b
sko100a
sko90
sko81
sko72
tai150b
tai100b
tai80b
tai60b
tai50b
tai100a
tai80a
tai60a
tai50a
tai40a
|p|
(4)
n
where the length of a pattern is the proportion of the
number of identical elements over the total number of
elements. A larger pattern length indicates thus more shared
elements. The solution similarity defined in Def. (3) can be
considered as a special case of the pattern length defined in
Def. (4). Specifically, when the support value of a mined
pattern equals 2, the pattern is simplified as the set of
common elements shared by only two solutions. This is
further confirmed according to the results reported in Figure
1, where the curve of the maximum solution similarity (left
sub-figure) is exactly the same as the curve of the maximum
length (right sub-figure).
In this experiment, we solved each benchmark instance
with a time limit of tmax = 30 minutes. To analyze the solution similarity of high-quality solutions stored in the elite set
according to Eq. (3), we calculate the length distribution of
a set of frequent patterns mined from the elite set according
to Eq. (4). The results of the similarity between high-quality
solutions and the length distribution of the mined frequent
patterns are presented in Figure 1.
In Figure 1, we report the maximum value, average
value, and minimum value of the solution similarity and
the pattern length, respectively. We can clearly observe that
there is a high similarity between the high-quality solutions.
Specifically, for all instances, the maximum solution similarity is larger than 0.9. Also, the average solution similarities
between any two high-quality solutions are larger than 0.5
except for sko72, for which the average solution similarity
is about 0.4. A more significant observation can be derived
based on the lengths of the mined patterns showed in the
right sub-figure. The high structural similarities between
the high-quality solutions provide the rationale behind our
solution construction based on mined patterns.
len(p) =
pattern length
1
tho150
wil100
sko100f
sko100e
sko100d
sko100c
sko100b
sko100a
sko90
sko81
sko72
tai150b
tai100b
tai80b
tai60b
tai50b
tai100a
tai80a
tai60a
tai50a
tai40a
|π s ∩ π t |
sim(π s , π t ) =
(3)
n
where π s ∩ π t is the set of common elements shared by
s
π and π t . The larger the similarity between two solutions,
the more common elements they share.
As we mentioned above, a mined frequent pattern represents a set of identical elements shared by two or more
solutions under a given minimum support θ. A frequent
pattern can be directly converted to a partial solution, thus
we define the length of a pattern p as follows.
solution similarity
1
Fig. 1. Solution similarity between high-quality solutions (left sub-figure)
and length distribution of the mined patterns (right sub-figure).
in evolutionary algorithms and memetic algorithms. To
demonstrate the effectiveness of the solution construction
using frequent patterns, we compare this approach with
the general crossover operator within the framework of
our FPBS-QAP algorithm. In this experiment, we compared
FPBS-QAP with its alternative version FPBS-QAP0 where
the frequent pattern based solution construction of FPBSQAP was replaced by the standard uniform crossover operator used in [10]. We ran both algorithms on each benchmark
instance 10 times with a time limit of tmax = 30 minutes.
The comparative results between FPBS-QAP and FPBSQAP0 are summarized in Table 7.
Table 7 indicates that FPBS-QAP performs better than
FPBS-QAP0 . FPBS-QAP is able to achieve a better or the
same BPD value on all instances except tai60a. For tai60a,
the BPD value of FPBS-QAP is 0.165%, which is only
marginally worse than 0.164% achieved by FPBS-QAP0 . The
average BPD value of FPBS-QAP is also better than that
of FPBS-QAP0 , i.e., 0.043% < 0.056%. Finally, FPBS-QAP
achieves better results in terms of the average APD value
and the average WPD value. To achieve these results, the average run time of FPBS-QAP0 is slightly shorter than that of
FPBS-QAP (i.e., 7.9 < 8.6 minutes). This can be explained by
the fact that FPBS-QAP needs to execute a frequent pattern
mining procedure during the search. These observations
confirms the interest of the solution construction method
using mined frequent patterns.
5.3 Impact of the number of the largest frequent patterns m
The number of the longest frequent patterns m (m > 1)
influences the diversity of the new solutions constructed by
the solution construction method using mined frequent patterns. To investigate the impact of this parameter, we varied
the values of m within a reasonable range and compared
their performances. The box and whisker plots showed in
Figure 2 are obtained by considering ten different values
m ∈ {1, 3, . . . , 21}. The experiments were conducted on
four representative instances selected from different families
(tai100a, tai150b, sko100f and tho150). For each m value
12
TABLE 7
Comparisons between FPBS-QAP0 and FPBS-QAP on hard instances under the time limit of tmax = 30 minutes. The success rate of reaching
the best-known value over 10 runs is indicated in parentheses.
FPBS-QAP0 ⋆
BKV
BPD
APD
WPD
T (m)
tai40a
tai50a
tai60a
tai80a
tai100a
tai50b
tai60b
tai80b
tai100b
tai150b
sko72
sko81
sko90
sko100a
sko100b
sko100c
sko100d
sko100e
sko100f
wil100
tho150
3139370
4938796
7205962
13499184
21052466
458821517
608215054
818415043
1185996137
498896643
66256
90998
115534
152002
153890
147862
149576
149150
149036
273038
8133398
0.000(2)
0.241(0)
0.164(0)
0.446(0)
0.316(0)
0.000(10)
0.000(10)
0.000(10)
0.000(8)
0.000(1)
0.000(9)
0.000(10)
0.000(9)
0.000(9)
0.000(8)
0.000(10)
0.000(10)
0.000(10)
0.000(7)
0.000(9)
0.002(0)
0.059
0.318
0.334
0.533
0.466
0.000
0.000
0.000
0.018
0.204
0.006
0.000
0.004
0.002
0.001
0.000
0.000
0.000
0.002
0.000
0.022
0.074
0.392
0.486
0.622
0.615
0.000
0.000
0.000
0.100
0.358
0.063
0.000
0.038
0.016
0.004
0.000
0.000
0.000
0.005
0.002
0.080
7.4
11.8
13.1
15.7
16.8
0.1
0.3
1.5
3.1
20.2
2.2
5.1
4.1
7.0
6.0
6.0
9.5
7.6
3.6
6.0
19.1
0.056
0.094
0.136
7.9
avg.
⋆
5
7
9
11
13
15
17
19
21
PD to the best known value(%)
PD to the best known value(%)
tai100a
0.6
0.55
0.5
0.45
0.4
0.35
0.3
0.25
0.2
3
tai150b
0.6
0.5
0.4
0.3
0.2
0.1
0
1
3
5
7
sko100f
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
1
3
5
7
9
11
13
number of patterns
9
11
13
15
17
19
21
15
17
19
21
number of patterns
15
17
19
21
PD to the best known value(%)
PD to the best known value(%)
number of patterns
0.08
tho150
0.15
0.1
0.05
0
1
3
5
7
9
11
13
number of patterns
Fig. 2. Impact of the number of the largest frequent patterns m. Box and
whisker plots corresponding to 10 different values of m ∈ {1, 3, . . . , 21}
in terms of the percentage deviation (PD) to the best-known value.
In Figure 2, X-axis indicates the different values for the
number of the largest frequent patterns m and Y-axis shows
the performance (i.e., the percentage deviation to the bestknown value). We observe that the performance of the FPBSQAP algorithm strongly depends on the m value except for
sko100f. FPBS-QAP achieves a good performance when the
number of the largest pattern m is fixed to 11. This justifies
the default value for m shown in Table 3.
To tune parameters β and max no update, we used the
same method and chosen β = 0.75, max no update = 15
as their default values. It is possible that FPBS-QAP improves its performance when its parameters are tuned for
each specific problem instance.
6
BPD
APD
WPD
T (m)
0.000(1)
0.000(1)
0.165(0)
0.430(0)
0.311(0)
0.000(10)
0.000(10)
0.000(10)
0.000(6)
0.000(1)
0.000(10)
0.000(10)
0.000(6)
0.000(10)
0.000(10)
0.000(10)
0.000(10)
0.000(7)
0.000(7)
0.000(10)
0.000(1)
0.067
0.279
0.377
0.516
0.402
0.000
0.000
0.000
0.040
0.191
0.000
0.000
0.015
0.000
0.000
0.000
0.000
0.001
0.003
0.000
0.051
0.074
0.415
0.469
0.614
0.553
0.000
0.000
0.000
0.100
0.321
0.000
0.000
0.038
0.000
0.000
0.000
0.000
0.004
0.021
0.000
0.123
6.4
17.3
8.7
18.2
13.4
0.1
0.5
1.2
4.0
20.5
1.4
2.7
4.2
7.7
7.7
8.7
9.0
10.3
4.4
9.5
24.2
0.043
0.092
0.130
8.6
FPBS-QAP0 can also be considered as an alternative version of BMA [10] by removing the mutation procedure.
and each instance, we ran the algorithm 10 times with the
stopping condition of tmax = 30 minutes.
1
FPBS-QAP
Instance
C ONCLUSIONS AND FURTHER WORK
In this paper, we proposed a general-purpose optimization
approach called frequent pattern based search (FPBS). The
proposed approach relies on a data mining procedure to
mine frequent patterns from high-quality solutions collected
during the search. The mined patterns are then used to
create new starting solutions for further improvements. By
iterating the pattern mining phase and the optimization
phase, FPBS is designed to ensure an effective exploration
of the combinatorial search space.
The viability of the proposed approach was verified
on the well-known quadratic assignment problem. Extensive computational results on popular QAPLIB benchmarks
showed that FPBS performs remarkably well compared to
very recent and state-of-the-art algorithms both in terms of
solution quality and computational efficiency. Specifically,
our approach is able to find the best-known objective values
for all the benchmark instances except tai80a and tai100a
within a time limit of 0.5 hour or 2 hours. To the best
of our knowledge, very few QAP algorithms can achieve
such a performance. Furthermore, we performed additional
experiments to investigate three key issues of the proposed
FPBS algorithm.
As future work, three directions can be followed. First,
this study focused on exploring maximal frequent itemsets.
However, there are other interesting patterns available in the
area of pattern mining. It is worth studying alternative patterns like sequential patterns and graph patterns. Second,
FPBS is a general-purpose approach, it would be interesting to investigate its interest to solve other optimization
problems, particularly other permutation problems (e.g.,
linear ordering problem and traveling salesman problem)
and subset selection problems (e.g., diversity or dispersion
problems and critical node problems).
R EFERENCES
[1] A. Acan and A. Ünveren. A great deluge and tabu search hybrid
with two-stage memory support for quadratic assignment problem.
Applied Soft Computing, 36:185–203, 2015.
13
[2] C.C. Aggarwal, M.A. Bhuiyan, and M.A. Hasan. Frequent pattern
mining algorithms: A survey. In Frequent Pattern Mining, pages 19–
64. Springer, 2014.
[3] R. Agrawal, T. Imieliński, and A. Swami. Mining association rules
between sets of items in large databases. In Proceedings of the
1993 ACM SIGMOD International Conference on Management of Data,
pages 207–216, New York, USA, 1993.
[4] Y. Aksan, T. Dokeroglu, and A. Cosar. A stagnation-aware cooperative parallel breakout local search algorithm for the quadratic
assignment problem. Computers & Industrial Engineering, 103:105–
115, 2017.
[5] K. Anstreicher, N. Brixius, J.P. Goux, and J. Linderoth. Solving large
quadratic assignment problems on computational grids. Mathematical Programming, 91(3):563–588, 2002.
[6] S. Asta and E. Özcan. A tensor-based selection hyper-heuristic
for cross-domain heuristic search. Information Sciences, 299:412–432,
2015.
[7] H. Barbalho, I. Rosseti, S.L. Martins, and A. Plastino. A hybrid
data mining GRASP with path-relinking. Computers & Operations
Research, 40(12):3159–3173, 2013.
[8] U. Benlic and J.K. Hao A multilevel memetic approach for
improving graph k-partitions. IEEE Transactions on Evolutionary
Computation, 15(5):624–472, 2011.
[9] U. Benlic and J.K. Hao. Breakout local search for the quadratic
assignment problem.
Applied Mathematics and Computation,
219(9):4800–4815, 2013.
[10] U. Benlic and J.K. Hao. Memetic search for the quadratic assignment problem. Expert Systems with Applications, 42(1):584–595, 2015.
[11] T. Dokeroglu and A. Cosar. A novel multistart hyper-heuristic
algorithm on the grid for the quadratic assignment problem. Engineering Applications of Artificial Intelligence, 52:10–25, 2016.
[12] Z. Drezner, P.M. Hahn, and É.D. Taillard. Recent advances for the
quadratic assignment problem with special emphasis on instances
that are difficult for metaheuristic methods. Annals of Operations
Research, 139(1):65–94, 2005.
[13] E. Duman and I. Or. The quadratic assignment problem in the
context of the printed circuit board assembly process. Computers &
Operations Research, 34(1):163–179, 2007.
[14] G. Grahne and J. Zhu. Efficiently using prefix-trees in mining
frequent itemsets. In Proceedings of the ICDM 2003 Workshop on
Frequent Itemset Mining Implementations, 19 December 2003.
[15] G. Grahne and J. Zhu. Fast Algorithms for Frequent Itemset
Mining Using FP-Trees. IEEE Transactions on Knowledge and Data
Engineering, 17(10):1347–1362, 2005.
[16] M. Guerine, I. Rosseti, and A. Plastino. Extending the hybridization of metaheuristics with data mining: Dealing with sequences.
Intelligent Data Analysis, 20(5):1133–1156, 2016.
[17] J. Han, H. Cheng, D. Xin, and X. Yan. Frequent pattern mining:
current status and future directions. Data Mining and Knowledge
Discovery, 15(1):55–86, 2007.
[18] J. Han, J. Pei, and M. Kamber. Data Mining: Concepts and Techniques.
Elsevier, 2011.
[19] J.K. Hao. Memetic Algorithms in Discrete Optimization. In F. Neri,
C. Cotta, P. Moscato (Eds.) Handbook of Memetic Algorithms.
Studies in Computational Intelligence 379, pages 73-94, 2012.
[20] T. James, C. Rego, and F. Glover. Multistart tabu search and
diversification strategies for the quadratic assignment problem.
IEEE Transactions on Systems, Man, and Cybernetics - Part A: Systems
and Humans, 39(3):579–596, May 2009.
[21] L. Jourdan, C. Dhaenens, and E.G. Talbi. Using datamining techniques to help metaheuristics: A short survey, pages 57–69. Springer
Berlin Heidelberg, Berlin, Heidelberg, 2006.
[22] T.C. Koopmans and M. Beckmann. Assignment problems and
the location of economic activities. Econometrica: Journal of the
Econometric Society, pages 53–76, 1957.
[23] E.M. Loiola, N.M.M. de Abreu, P.O. Boaventura-Netto, P. Hahn,
and T. Querido. A survey for the quadratic assignment problem.
European Journal of Operational Research, 176(2):657–690, 2007.
[24] Z. Lü and J.K. Hao. A memetic algorithm for graph coloring.
European Journal of Operational Research, 203(1):241–250, 2010.
[25] D. Martins, G.M. Vianna, I. Rosseti, S.L. Martins, and A. Plastino.
Making a state-of-the-art heuristic faster with data mining. Annals
of Operations Research, pages 1–22, 2014.
[26] S.L. Martins, I. Rosseti, and A. Plastino. Data mining in stochastic
local search, pages 1–49. Springer, 2016.
[27] A. Plastino, H. Barbalho, L.F.M. Santos, R. Fuchshuber, and
S.L. Martins. Adaptive and multi-mining versions of the DMGRASP hybrid metaheuristic. Journal of Heuristics, 20(1):39–74, 2014.
[28] A. Plastino, R. Fuchshuber, S.L. Martins, A.A. Freitas, and S. Salhi.
A hybrid data mining metaheuristic for the p-median problem.
Statistical Analysis and Data Mining, 4(3):313–335, 2011.
[29] D.C. Porumbel, J.K. Hao, and P. Kuntz. An Evolutionary Approach with Diversity Guarantee and Well-Informed Grouping Recombination for Graph Coloring. Computers & Operations Research,
37(10):1822–1832, 2010.
[30] M. Raschip, C. Croitoru, and K. Stoffel. Guiding evolutionary
search with association rules for solving weighted CSPs. In Proceedings of the 2015 Annual Conference on Genetic and Evolutionary
Computation, pages 481–488. ACM, 2015.
[31] M. Raschip, C. Croitoru, and K. Stoffel. Using association rules
to guide evolutionary search in solving constraint satisfaction. In
Proceedings of the 2015 IEEE Congress on Evolutionary Computation
(CEC), pages 744–750. IEEE, 2015.
[32] D.S. Reddy, A. Govardhan, and S. Sarma. Hybridization of
neighbourhood search metaheuristic with data mining technique
to solve p-median problem. International Journal of Computational
Engineering Research, 2(7):159–163, 2012.
[33] M.H. Ribeiro, V.A. Trindade, A. Plastino, and S.L. Martins. Hybridization of GRASP metaheuristics with data mining techniques.
In Proceedings of First International Workshop on Hybrid Metaheuristics,
Valencia, Spain, August 22-23,, pages 69–78, 2004.
[34] M.H. Ribeiro, A. Plastino, and S.L. Martins. Hybridization of
GRASP metaheuristic with data mining techniques. Journal of
Mathematical Modelling and Algorithms, 5(1):23–41, 2006.
[35] M. Samorani and M. Laguna. Data-mining-driven neighborhood
search. INFORMS Journal on Computing, 24(2):210–227, 2012.
[36] H.G. Santos, L.S. Ochi, E.H. Marinho, and L.M. Drummond.
Combining an evolutionary algorithm with data mining to solve a
single-vehicle routing problem. Neurocomputing, 70(1):70–77, 2006.
[37] L.F. Santos, M.H. Ribeiro, A. Plastino, and S.L. Martins. A hybrid
GRASP with data mining for the maximum diversity problem, pages
116–127. Springer, Berlin, Heidelberg, 2005.
[38] L.F. Santos, R. Milagres, C.V. Albuquerque, S. Martins, and A. Plastino. NGL01-4: A hybrid GRASP with data mining for efficient
server replication for reliable multicast. In Proceedings of the 2006
Global Telecommunications Conference, Sans Francisco, USA, pages 1–
6. IEEE, 2006.
[39] L.F. Santos, S.L. Martins, and A. Plastino. Applications of the DMGRASP heuristic: a survey. International Transactions in Operational
Research, 15(4):387–416, 2008.
[40] T. Stützle and M. López-Ibáñez. Automatic (offline) configuration
of algorithms. Proceedings of the Companion Publication of the 2015
Annual Conference on Genetic and Evolutionary Computation, pages
681–702.
[41] K. Sörensen and M. Sevaux. MA mid PM: memetic algorithms
with population management. Computers & Operations Research,
33(5):1214–1225, 2006.
[42] É. Taillard. Robust taboo search for the quadratic assignment
problem. Parallel Computing, 17(4-5):443–455, 1991.
[43] U. Tosun. On the performance of parallel hybrid algorithms for
the solution of the quadratic assignment problem. Engineering
Applications of Artificial Intelligence, 39:267–278, 2015.
[44] S. Umetani. Exploiting variable associations to configure efficient
local search in large-scale set partitioning problems. In Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence,
AAAI’15, pages 1226–1232. AAAI Press, 2015.
[45] T. Wauters, K. Verbeeck, P. De Causmaecker and G.V. Berghe.
Boosting metaheuristic search using reinforcement learning, Hybrid
Metaheuristics, Springer Berlin Heidelberg, 433-452, 2013.
[46] T. Wauters, K. Verbeeck, P. De Causmaecker and G.V. Berghe. A
learning-based optimization approach to multi-project scheduling,
Journal of Scheduling, Springer, 18(1):61–74, 2015.
[47] I.H. Witten and E. Frank. Data Mining: Practical Machine Learning
Tools and Techniques, Second Edition (Morgan Kaufmann Series in
Data Management Systems) Morgan Kaufmann Publishers Inc., San
Francisco, CA, USA, 2005.
[48] Y. Zhou, J.K. Hao, and B. Duval. Reinforcement learning based
local search for grouping problems: A case study on graph coloring.
Expert Systems with Applications, 64:412–422, 2016.
[49] Y. Zhou, J.K. Hao, and B. Duval. Opposition-based memetic
search for the maximum diversity problem. IEEE Transactions on
Evolutionary Computation, 21(5):731–745, 2017.
| 8cs.DS
|
Divergence, Entropy, Information
An Opinionated Introduction to Information Theory
Phil Chodrow 1
1
arXiv:1708.07459v1 [cs.IT] 24 Aug 2017
MIT Operations Research Center
[email protected]
https://philchodrow.github.io/
August 25, 2017
Information theory is a mathematical theory of learning with deep connections with topics as diverse as artificial intelligence, statistical physics,
and biological evolution. Many primers on the topic paint a broad picture
with relatively little mathematical sophistication, while many others develop
specific application areas in detail. In contrast, these informal notes aim to
outline some elements of the information-theoretic “way of thinking,” by cutting a rapid and interesting path through some of the theory’s foundational
concepts and theorems. They are aimed at practicing systems scientists who
are interested in exploring potential connections between information theory
and their own fields. The main mathematical prerequisite for the notes is
comfort with elementary probability, including sample spaces, conditioning,
and expectations.
We take the Kullback-Leibler divergence as our foundational concept, and
then proceed to develop the entropy and mutual information. We discuss
some of the main foundational results, including the Chernoff bounds as a
characterization of the divergence; Gibbs’ Theorem; and the Data Processing
Inequality. A recurring theme is that the definitions of information theory
support natural theorems that sound “obvious” when translated into English.
More pithily, “information theory makes common sense precise.” Since the
focus of the notes is not primarily on technical details, proofs are provided
only where the relevant techniques are illustrative of broader themes. Otherwise, proofs and intriguing tangents are referenced in liberally-sprinkled
footnotes. The notes close with a highly nonexhaustive list of references to
resources and other perspectives on the field.
1
Contents
1
Why Information Theory?
3
2
Why Not Start with Entropy?
3
3
Introducing the Divergence
4
4
Entropy
8
5
Conditional Entropy
10
6
Information Three Ways
12
7
Why Information Shrinks
14
8
Some Further Reading
16
2
1 Why Information Theory?
Briefly, information theory is a mathematical theory of learning with rich
connections to physics, statistics, and biology. Information-theoretic methods quantify complexity and predictability in systems, and make precise
how observing one feature of a system assists in predicting other features.
Information-theoretic thinking helps to structure algorithms; describe natural processes; and draw surprising connections between seemingly disparate
fields.
Formally, information theory is a subfield of probability, the mathematical study of uncertainty and randomness. What is distinctive about information theory is its emphasis on properties of probability distributions
that are independent of how those distributions are represented. Because
of this representation-independence, information-theoretic quantities often
have claim to be the most “fundamental” properties of a system or problem,
governing its complexity and learnability.
In its original formulation Shannon (1948), information theory is a theory
of communication: specifically, the transmission of a signal of some given
complexity over an unreliable channel, such as a telephone line corrupted by
a certain amount of white noise. In these notes, we will emphasize a slightly
different role for information theory, as a theory of learning more generally.
This emphasis is consistent with the original formulation, since the communication problem may be viewed as the problem of the message receiver
learning the intent of the sender based on the potentially corrupted transmission. However, the emphasis on learning allows us to more easily glimpse
some of the rich connections of information theory to other disciplines. Special consideration in these notes will be given to statistical motivations for
information-theoretic concepts. Theoretical statistics is the design of mathematical methods for learning from data; information-theoretic considerations
determine when such learning is possible and to what degree. We will close
with a connection to physics; some connections to biology are cited in the
references.
2 Why Not Start with Entropy?
Entropy is easily the information-theoretic concept with the widest popular
currency, and many expositions of that theory take entropy as their starting
point. We, however, will choose a different point of departure for these notes,
and derive entropy along the way. Our point of choice is the Kullback-Leibler
(KL) divergence between two distributions, also called in some contexts the
relative entropy, relative information, or free energy.1 Why start with the
1
For the remainder of these notes I’ll stick with “divergence” – though there are many other interesting
objects called “divergences” in mathematics, we won’t be discussing any of them here, so no confusion
3
divergence? Well, there’s a simple reason – while we’ll focus on discrete
random variables here, we’d like to develop a theory that, wherever possible,
applies to continuous random variables as well. The divergence is well-defined
for both discrete random variables and continuous ones; that is, if p and q
are two distributions satisfying certain regularity properties, then d(p, q) is
a uniquely determined, nonnegative real number whether p and q are discrete or continuous. In contrast, the natural definition of entropy (so called
differential entropy) for continuous random variables has two bad behaviors.
First, it can be negative, which is undesirable for a measure of uncertainty.
Second, and arguably worse, the differential entropy is not even uniquely defined. There are multiple ways to describe the same continuous distribution
– for example, the following three distributions are the same:
1. “The Gaussian distribution with mean 0 and variance 1.”
2. “The Gaussian distribution with mean 0 and standard deviation 1.”
3. “The Gaussian distribution with mean 0 and second moment equal to
1.”
Technically, the act of switching from one of these descriptions to another
can be viewed as a smooth change of coordinates2 in the space of distribution
parameters. For example, we move from the first description to the second
by changing coordinates from (µ, σ 2 ) to (µ, σ), which we can do by applying
√
the function f (x, y) = (x, y). Regrettably, the differential entropy is not
invariant under such coordinate changes – change the way you describe the
distribution, and the differential entropy changes as well. This is undesirable
– the foundations of our theory should be independent of the contingencies
of how we describe the distributions under study. The divergence passes this
test in both discrete and continuous cases; the differential entropy does not.3
Since we can define the entropy in terms of the divergence in the discrete
case, we’ll start with the divergence and derive the entropy along the way.
3 Introducing the Divergence
It is often said that the divergence d(p, q) between distributions p and q
measures how “surprised” you are if you think the state of the world is q
but then measure it to be p. However, this idea of surprise isn’t typically
explained or made precise. To motivate the KL divergence, we’ll start from a
somewhat unusual beginning – the Chernoff bounds – that makes exact the
role that the divergence plays in governing how surprised you ought to be.
should arise.
Technically, the idea of a smooth coordinate change is captured by diffeomorphisms: invertible functions on coordinate space whose inverses are also smooth.
3
It
is
possible
to
define
alternative
notions
of
entropy
that
attempt
to
skirt
these
issues;
however,
they
have
their
own
difficulties.
https://en.wikipedia.org/wiki/Limiting_density_of_discrete_points
2
4
Let’s begin with a simple running example. You are drawing from an
(infinite) deck of standard playing cards, with four card suits {♠, ♣, ♦, ♥}
and thirteen card values {1, . . . , 13}. We’ll view the sets of possible values as alphabets: X = {♠, ♣, ♦, ♥} is the alphabet of possible suits, and
Y = {1, . . . , 13} the alphabet of possible values. We’ll let X and Y be the
corresponding random variables, so for each realization, X ∈ X and Y ∈ Y.
Suppose that I have a prior belief that the distribution of suits in the
deck is uniform. My belief about the suits can be summarized by a vector
q = ( 14 , 14 , 14 , 14 ). It’s extremely convenient to view q as a single point in the
probability simplex P X of all valid probability distributions over X .
Definition 1 (Probability Simplex). For any finite alphabet X with |X| = m,
the probability simplex P X is the set
(
)
X
X
m
P , q∈R |
qi = 1, qi ≥ 0 ∀i .
(1)
i
Remark. It’s helpful to remember that P X is an P
m − 1-dimensional space;
the “missing” dimension is due to the constraint i qi = 1. When m = 3,
P X is an equilateral triangle; when m = 4 a tetrahedron, and so on.
If q is your belief, you would naturally expect that, if you drew enough
cards, the observed distribution of suits would be “close” to q, and that if
you could draw infinitely many cards, the distribution would indeed converge
to q. Let’s make this precise: define p̂n to be the distribution of suits you
observe after pulling n cards. It’s important to remember that p̂ is a random
vector, which changes in each realization. But it would be reasonable to
expect that p̂ → q as n → ∞, and indeed this is true almost surely (with
probability 1) according to the Strong Law of Large Numbers, if q is in fact
the true distribution of cards in the deck.
But what happens if you keep drawing cards and the observed distribution
p̂n is much different than your belief q? Then you would justifiably be surprised, and your “level of surprise” can be quantified by the probability of
observing p̂n if the true distribution were q, which I’ll denote P(p̂n ; q). We’d
expect P(p̂n ; q) to become small when n grows large. Indeed, there is a quite
strong result here – P(p̂n ; q) decays exponentially in n, with a very special
exponent.
Definition 2 (Kullback-Leibler Divergence). For p, q ∈ P X , the KullbackLeibler (KL) divergence of q from p is
d(p, q) ,
X
p(x) log
x∈X
p(x)
,
q(x)
(2)
where we are using the conventions that log ∞ = ∞, log 0 = −∞, 0/0 = 0,
and 0 × ∞ = 0.
5
Theorem 1 (Chernoff Bounds). Suppose that the card suits are truly distributed according to p 6= q. Then,
e−nd(p,q)
≤ P(p̂n ; q) ≤ e−nd(p,q) .
(n + 1)m
(3)
So, the probability of observing p̂n when you thought the distribution was
q decays exponentially, with the exponent given by the divergence of your
belief q from the true distribution p. Another way to say this: ignoring
non-exponential factors,
1
∼ d(p, q) ,
(4)
− log P(p̂n ; q) =
n
that is, d(p, q) is the minus the log of your average surprise per card drawn.
The Chernoff bounds thus provide firm mathematical content to the idea
that the divergence measures surprise.
Just to make this concrete, suppose that I start with the belief that the
deck is uniform over suits; that is, my belief is q = ( 14 , 14 , 14 , 14 ) over the
alphabet {♠, ♣, ♦, ♥}. Unbeknownst to me, you have removed all the black
cards from the deck, which therefore has true distribution p = 0, 0, 21 , 12 .
I draw 100 cards and record the suits. How surprised am I by the suit
distribution I observe? The divergence between my belief and the true deck
distribution is d(p, q) ≈ 0.69, so Theorem 1 states that the dominating factor
in the probability of my observing an empirical distribution close to p in 100
draws is e−0.69×100 ≈ 10−30 . I am quite surprised indeed!
We’ll now state one of the most important properties of the divergence.
Theorem 2 (Gibbs’ Inequality). For all p, q ∈ P X , it holds that d(p, q) ≥ 0,
with equality iff p = q.
In words, you can never have “negative surprise,” and you are only unsurprised if you what you observed is exactly what you expected.
Proof. There are lots of ways to prove Gibbs’ Inequality; here’s one with
Lagrange multipliers. Fix p ∈ P X . We’d like to show that the problem
min d(p, q)
q∈P X
(5)
has value 0 and that this value is achieved at the unique point q = p. We
need two gradients: the gradient P
of d(p, q) with respect to q and the gradient
of the implicit constraint g(q) , x∈X q(x) = 1. The former is
#
"
X
p(x)
(6)
p(x) log
∇q d(p, q) = ∇q
q(x)
x∈X
#
"
X
p(x) log q(x)
(7)
= −∇q
x∈X
= −p ⊘ q,
6
(8)
where p ⊘ q is the elementwise quotient of vectors, and where we recall the
convention 0/0 = 0. On the other hand,
∇q g(q) = 1,
(9)
the vector whose entries are all unity. The method of Lagrange multipliers
states that we should seek λ ∈ R such that
∇q d(p, q) = λ∇q (gq ),
(10)
− p ⊘ q = λ1,
(11)
or
from which it’s easy to see that the only solution is q = p and λ = −1.
It’s a quick check that the corresponding solution value is d(p, p) = 0, which
completes the proof.
Remark. Theorem 2 is the primary sense in which d behaves “like a distance”
on the simplex P. On the other hand, d is unlike a distance in that it is not
symmetric and does not satisfy the triangle inequality.4
Let’s close out this section by noting one of the many connections between the divergence and classical statistics. Maximum likelihood estimation
is a foundational method of modern statistical practice; tools from linear
regression to neural networks may be viewed as likelihood maximizers. The
divergence allows a particularly elegant formulation of maximum likelihood
estimation: likelihood maximization is the same as divergence minimization. Let θ be some statistical parameter, which may be multidimensional; for example, in the context of normal distributions, we may have
θ = (µ, σ 2 ); in the context of regression, θ may be the regression coefficients
β. Let pθX be the probability distribution over X with parameters θ. Let
{x1 , . . . , xn } be a sequence of i.i.d. observations of X. Maximum likelihood
estimation encourages us to find the parameter θ such that
θ ∗ = argmax
θ
n
Y
pθX (xi ) ,
(12)
i=1
i.e. the parameter value that makes the data most probable. To express this
in terms of the divergence, we need just one more piece of notation: let p̂X be
the empirical distribution of observations of X. Then, it’s a slightly involved
algebraic exercise to show that the maximum likelihood estimation problem
can also be written
θ ∗ = argmin d(p̂X , pθX ).
(13)
θ
4
In fact, d is related to a “proper” distance metric on P X , which is usually called the Fisher Information Metric and is the fundamental object of study in the field of information geometry
(Amari and Nagaoka, 2007).
7
This is rather nice – maximum likelihood estimation consists in making the
parameterized distribution pθX as close as possible to the observed data distribution p̂X in the sense of the divergence.5
4 Entropy
After having put it off for a while, let’s define the Shannon entropy. If
we think about the divergence as a (metaphorical) distance, the entropy
measures how close a distribution is to the uniform.
Definition 3 (Shannon Entropy). The Shannon entropy of p ∈ P X is
X
H(p) , −
p(x) log p(x) .
(14)
x∈X
When convenient, we will also use the notation H(X) to refer to the entropy
of a random variable X distributed according to p.
Theorem 3. The Shannon entropy is related to the divergence according to
the formula
H(p) = log m − d(p, u) ,
(15)
where m = |X | is the size of the alphabet X and where u is the uniform
distribution on X .6
Remark. This formula makes it easy to remember the entropy of the uniform
distribution – it’s just log m, where m is the number of possible choices. If
we are playing a game in which I draw a card from the infinite deck, the suit
of the card is uniform, and the entropy of the suit distribution is therefore
H(X) = log 4 = 2 log 2.
Remark. In words, d(p, u) is your surprise if you thought the suit distribution
was uniform and then found it was in fact p. If you are relatively unsurprised,
then p is very close to uniform. Indeed, Gibbs’ inequality (Theorem 2) immediately implies that H(p) assumes its largest value of log m exactly when
p = u.
5
This result is another hint at the beautiful geometry of the divergence: the operation of minimizing a
distance-like measure is often called “projection.” Maximum likelihood estimation thus consists in a
kind of statistical projection.
6
Here is as good a place as any to note that for discrete random variables, the divergence can also be
defined in terms of the entropy. Technically, d is the Bregman divergence induced by the Shannon
entropy, and can be characterized by the equation
d(p, q) = − [H(p) − H(q) − h∇q H, p − qi] .
(16)
Intuitively, d(p, q) is minus the approximation loss associated with estimating the difference in entropy
between p and q using a first-order Taylor expansion centered at q. This somewhat artificial-seeming
construction turns out to lead in some very interesting directions in statistics and machine learning.
8
Remark. Theorem 3 provides one useful insight into why the Shannon entropy
does not generalize naturally to continuous distributions – whereas equation
(15) involves the uniform distribution on X , but there can be no analogous
formula for continuous random variables on R because there is no uniform
distribution on R.
A Bayesian Interpretation of Entropy
The construction of the entropy in terms of the divergence is fairly natural –
we use the divergence to measure how close p is to the uniform distribution,
flip the sign so that high entropy distributions are more uniform, and add
a constant term to make the entropy nonnegative. This formulation of the
entropy turns out to have another interesting characterization in the context
of Bayesian prediction. In Bayesian prediction, I will pull a single card from
the deck. Before I do, I ask you to provide a distribution p over the alphabet
{♠, ♣, ♦, ♥} representing your prediction about the suit of the card I pulled.
As examples, you can choose p = (1, 0, 0, 0) if you are certain that the suite
will be ♠, or p = ( 41 , 41 , 14 , 14 ) to express maximal ignorance. After you guess,
I pull the card, obtaining a sample x ∈ X , and reward you based on the
quality of your prediction relative to the outcome x. I do this based on a
loss function f (p, x); after your guess I give you f (p, x) dollars, say. If we
assume that my aim is to encourage you to (a) report your true beliefs about
the deck and (b) reward you based only on what happened (i.e. not on what
could have happened), then there is only one appropriate loss function f ,
which turns out to be closely related to the entropy. More formally,
Definition 4. A loss function f is proper if, for any alphabet Y and random
variable Y on Y,
pX|Y =y = argmin E[f (p, x)|Y = y].
(17)
p∈P X
Remark. In this definition, it’s useful to think of Y as some kind of “side
information” or “additional data.” For example, Y could be my telling you
that the card I pulled is a red card, which could influence your predictive
distribution. When f is proper, you have an incentive to factor that into
your predictive distribution. While it may feel that “of course” you should
factor this in, not all loss functions encourage you to do so. For example, if
f is constant, then you have no incentive to use Y at all, since each guess
is just as good as any other. A proper loss function guarantees that you
can maximize your payout (minimize your loss) by completely accounting
for all available data when forming your prediction, which should therefore
be pX|Y =y . Thus, a proper loss function ensures that the Bayesian prediction
game is “honest”.
Definition 5. A loss function f is local if f (p, x) = ψ(p, p(x)) for some
function ψ.
9
Remark. The function f is local iff f can be written as a function only of
my prediction p and how much probabilistic weight I put on the event that
actually occurred – not events that “could have happened” but didn’t. Thus,
a proper loss funciton ensures that the Bayesian prediction game is “fair.”
Somewhat amazingly, the log loss function given by f (p, x) = − log p(x) is
the only loss function that is both proper and local (both honest and fair),
up to an affine transformation.
Theorem 4 (Uniqueness of the Log-Loss). Let f be a local and proper reward
function. Then, f (p, x) = A log p(x) + B for some constants A < 0 and
B ∈ R.
Without loss of generality, we’ll take A = −1 and B = 0. The entropy in
this context occurs as the expected log-loss when you know the distribution
of suits in the deck. If you know, say, that the proportions in the deck are
p = ( 41 , 14 , 14 , 14 ) and need to formulate your predictive distribution, Theorem
4 implies that your best guess is just p, since you have no additional side
information. Then....
Definition 6 (Entropy, Bayesian Characterization). The (Shannon) entropy of p is your minimal expected loss when playing the Bayesian prediction
game in which the true distribution of suits is p.
Remark. To see that this definition is consistent with the one we saw before,
we can simply compute the expectation:
E[f (p, X)] = E[− log p(X)]
X
=−
p(x) log p(x),
(18)
(19)
x∈X
which matches Definition 3. The second inequality follows from the fact that,
if you are playing optimally, p is both the true distribution of X and your
best predictive distribution.
5 Conditional Entropy
The true magic of probability theory is conditional probabilities, which formalize the idea of learning: P(A|B) represents my best belief about A given
what I know about B. While the Shannon entropy itself is quite interesting,
information theory really starts becoming a useful framework for thinking
probabilistically when we formulate the conditional entropy, which encodes
the idea of learning as a process of uncertainty reduction.
In this section and the next, we’ll need to keep track of multiple random
variables and distributions. To fix notation, we’ll let pX ∈ P X be the distribution of a discrete random variable X on alphabet X , pY ∈ P Y the distribution of a discrete random variable Y on alphabet Y, and pX,Y ∈ P X ×Y their
10
joint distribution on alphabet X × Y. Additionally, we’ll denote the product
distribution of marginals as pX ⊗ pY ∈ P X ×Y ; that is, (pX ⊗ pY )(x, y) =
pX (x)pY (y).
Definition 7 (Conditional Entropy). The conditional entropy of X given
Y is
X
H(X|Y ) ,
pX,Y (x, y) log pX|Y (x|y) .
(20)
x,y∈X ×Y
Remark. It might seem as though H(X|Y ) ought to be defined as
X
H̃(X|Y ) =
pX|Y (x|y) log pX|Y (x|y) ,
(21)
x,y∈X ×Y
which looks more symmetrical. However, a quick think makes clear that this
definition isn’t appropriate, because it doesn’t include any information about
the distribution of Y . If Y is concentrated around some very informative (or
uninformative) values, then H̃ won’t notice that some values of Y are more
valuable than others.
In the framework of our Bayesian interpretation of the entropy above, the
conditional entropy is your expected reward in the guessing game assuming
you receive some additional side information. For example, consider playing
the suit-guessing game in the infinite deck of cards. Recall that the suit
distribution is uniform, with entropy H(X) = H(u) = 2 log 2. Suppose now
that you get side information – when I draw the card from the deck, before
I ask you to guess the suit, I tell you the color (black or red). Since for each
color there are just two possible suits, the entropy decreases. Formally, if
X is the suit and Y the color, it’s easy to compute that H(X|Y ) = log 2 –
knowing the color reduces your uncertainty by half.
The conditional entropy is somewhat more difficult to express in terms of
the divergence, but it does have a useful relationship with the (unconditional)
entropy.
Theorem 5. The conditional entropy is related to the unconditional entropy
as
H(X|Y ) = H(X, Y ) − H(Y ),
(22)
where H(X, Y ) is the entropy of the distribution pX,Y .
Remark. This theorem is easy to remember, because it looks like what you
get by recalling the definition of the conditional probability and taking logs:
pX|Y (x|y) =
pX,Y (x, y)
.
pY (y)
(23)
Indeed, take logs and compute the expectations over X and Y to prove the
theorem directly. Another way to remember this theorem is to just say it
11
out: the uncertainty you have about X given that you’ve been told Y is
equal to the uncertainty you had about both X and Y , less the uncertainty
that was resolved when you learned Y .
From this theorem, it’s a quick use of Gibbs’ inequality to show:
Theorem 6 (Side Information Reduces Uncertainty).
H(X|Y ) ≤ H(X).
(24)
That is, knowing Y can never make you more uncertain about X, only
less. This makes sense – after all, if Y is not actually informative about X,
you can just ignore it.
Theorem 6 implies that H(X) − H(X|Y ) ≥ 0. This difference quantifies
how much Y reduces uncertainty about X; if H(X) − H(X|Y ) = 0, for
example, then H(X|Y ) = H(X) and it is natural to say that Y “carries no
information” about X. We encode the idea of information as uncertainty
reduction in the next section.
6 Information Three Ways
Thus far, we’ve seen two concepts – divergence and entropy – that play fundamentals role in information theory. But neither of them exactly resemble
an idea of “information,” so how does the theory earn its name? Our brief
note at the end of the last section suggests that we think about information as a relationship between two variables X and Y , in which knowing
Y decreases our uncertainty (entropy) about X. As it turns out, the idea
of information information that falls out of this motivation is a remarkably
useful one, and can be formulated in many interesting and different ways.
Let’s start by naming this difference:
Definition 8 (Mutual Information). The mutual information I(X, Y ) in
Y about X is
I(X, Y ) , H(X) − H(X|Y ).
(25)
The mutual information is just the uncertainty reduction associated with
knowing Y . In the context of the Bayesian guessing game, I(X, Y ) is the
“value” of being told the suit color, compared to having to play the game
without that information. From our calculations above, in the suit-guessing
game, I(X, Y ) = H(X) − H(X|Y ) = 2 log 2 − log 2 = log 2.
Let’s now express mutual information in two other ways. Remarkably,
these follow directly via simple algebra, but each identity gives a new way to
think about the meaning of the mutual information.
Theorem 7. The mutual information may also be written as:
I(X, Y ) = d(pX,Y , pX ⊗ pY )
= EY [d(pX|Y , pX )]
12
(26)
(27)
Let’s start by unpacking equation (26). I(X, Y ) is the divergence between
the actual joint distribution pX,Y and the product distribution pX ⊗ pY . Importantly, the latter is what the distribution of X and Y would be, were they
independent random variables with the same marginals. This plus Gibb’s inequality implies:
Corrolary 1. Random variables X and Y are independent if and only if
I(X, Y ) = 0.
So, I(X, Y ) is something like a super-charged correlation coefficient – it
measures the degree of statistical correlation between X and Y , but it is
stronger than the correlation coefficient in two ways. First, I(X, Y ) detects
all kinds of statistical relationships, not just linear ones. Second, while the
correlation coefficient can vanish for dependent variables, this never happens
for the mutual information – zero mutual information implies dependence,
period. As a quick illustration, it’s not hard to see that if X is the suit color
and Z is the numerical value of the card pulled, then I(X, Z) = 0. Intuitively, if we were playing the suit-guessing game and I offered to tell you the
card’s face-value, you would be rightly annoyed – that’s an unhelpful (“uninformative”) offer, because the face-values and suit colors are independent.7
Equation (26) has another useful consequence. Since that formulation is
symmetric in X and Y ,
Corrolary 2. The mutual information is symmetric:
I(X, Y ) = I(Y, X) .
(28)
Finally, as we noted briefly at the end of the previous section, the following
is a direct consequence of Equation (26) and Gibbs’ inequality.
Corrolary 3. The mutual information is nonnegative:
I(X, Y ) ≥ 0.
(29)
Now let’s unpack equation (27). One way to read this is as quantifying
the danger of ignoring available information: d(pX|Y =y , pX ) is how surprised
you would be if you ignored the information Y and instead kept using pX
as your belief. If I told you that the deck contained only red
cards, but you
chose to ignore this and continue guessing u = 14 , 14 , 14 , 14 as your guess,
you would be surprised to keep seeing red cards turn up draw after draw.
Formulation (27) expresses the mutual information as the expected surprise
7
So, why don’t we just dispose of correlation coefficients and use I(X, Y ) instead? Well, correlation
coefficients can be estimated from data relatively simply and are fairly robust to error. In contrast,
I(X, Y ) requires that we have reasonably good estimates of the joint distribution pX,Y , which is not
usually available. Furthermore, it can be hard to distinguish I(X, Y ) = 10−6 from I(X, Y ) = 0, and
statistical tests of significance that would solve this problem are much more complex than those for
correlation coefficients.
13
you would experience by ignoring your available side information Y , with
the expectation taken over all the possible values the side information could
assume. While this formulation may seem much more opaque than (26), it
turns out to be remarkably useful when thinking geometrically, as it expresses
the mutual information as the average “distance” between the marginal pX
and the conditionals pX|Y . Pursuing this thought turns out to express the
mutual information as something like the “moment of inertia” for the joint
distribution pX,Y .
7 Why Information Shrinks
The famous 2nd Law of Thermodynamics states that, in a closed system,
entropy increases. The physicists’ concept of entropy is closely related to but
slightly different from the information theorist’s concept, and we therefore
won’t make a direct attack on the 2nd Law in these notes. However, there
is a close analog of the 2nd Law that gives much of the flavor and can be
formulated in information theoretic terms. Whereas the 2nd Law states
that entropy grows, the Data Processing Inequality states that information
shrinks.
Theorem 8 (Data Processing Inequality). Let X and Y be random variables,
and let Z = g(Y ), where g is some function g : Y → Z. Then,
I(X, Z) ≤ I(X, Y ).
(30)
This is not the most general possible form of the Data Processing Inequality, but it has the right flavor. The meaning of this theorem is both “obvious”
and striking in its generality.8 Intuitively, if you are using Y to predict X,
then any processing you do to Y can only reduce your predictive power. Data
processing can enable tractable computations; reduce the impact of noise in
your observations; and improve your visualizations. The one thing it can’t do
is create information out of thin air. No amount of processing is a substitute
for having enough of the data you really want.
We’ll pursue the proof the Data Processing Inequality, as the steps are
quite enlightening. First, we need the conditional mutual information:
Definition 9 (Conditional Mutual Information). The conditional mutual
information of X and Y given Z is
X
I(X, Y |Z) =
pZ (z)d(pX,Y |Z=z , pX|Z=z ⊗ pY |Z=z ).
(31)
z∈Z
8
My first thought when seeing this was: “g can be ANY function? Really?” I then spent half an hour
fruitlessly attempting to produce a counter-example.
14
The divergence in the summand
P is naturally written I(X, Y |Z = z), in
which case we have I(X, Y |Z) = z∈Z pZ (z)I(X, Y |Z = z), which has the
form of an expectation of mutual informations conditioned on specific values
of Z. The conditional mutual information is naturally understood as the
value of knowing Y for the prediction of X, given that you also already
know Z. Somewhat surprisingly, both of the cases I(X, Y |Z) > I(X, Y ) and
I(X, Y |Z) < I(X, Y ) may hold; that is, knowing Z can either increase or
decrease the value of knowing Y in the context of predicting X.
Theorem 9 (Chain Rule of Mutual Information). We have
I(X, (Y, Z)) = I(X, Z) + I(X, Y |Z).
(32)
Remark. The notation I(X, (Y, Z)) refers to the (regular) mutual information
between X and the random variable (Y, Z), which we can regard as a single
random variable on the alphabet Y × Z.
Proof. We can compute directly, dividing up sums and remembering relations
like pX,Y,Z (x, y, z) = pX,Y |Z (x, y|z)pZ (z). Omitting some of the more tedious
algebra,
I(X, (Y, Z)) = d(pX,Y,Z , pX ⊗ pY,Z )
X
pX,Z (x, z)
+
=
pX,Y |Z (x, z) log
pX (x)pZ (z)
x,z∈X ×Z
X
pZ (z)pX,Y |Z (x, y|z) log
x,y,z∈X ×Y×Z
pX,Y |Z (x, y|z)
pY |Z (y|z)pX|Z (x|z)
= I(X, Z) + I(X, Y |Z) ,
as was to be shown.
As always, the Chain Rule has a nice interpretation if you think about
estimating X by first learning Z, and then Y . At the end of this process,
you know both Y and Z, and therefore have information I(X, (Y, Z)). This
total information splits into two pieces: the information you gained when
you learned Z, and the information you gained when you learned Y after
already knowing Z.
We are now ready to prove the Data Processing Inequality.9
Proof. Since Z = g(Y ), that is, is a function of Y alone, we have that
Z ⊥ X|Y , that is, given Y , Z and X are independent.10 So, I(X, Z|Y ) = 0.
9
10
Proof borrowed from http://www.cs.cmu.edu/~ aarti/Class/10704/lec2-dataprocess.pdf
In fact, Z ⊥ X|Y is often taken as the hypothesis of the Data Processing inequality rather than
Z = g(Y ), as it is somewhat weaker and sufficient to prove the result.
15
On the other hand, using the chain rule in two ways,
I(X, (Y, Z)) = I(X, Z) + I(X, Y |Z)
= I(X, Y ) + I(X, Z|Y ).
Since I(X, Z|Y ) = 0 by our argument above, we obtain I(X, Y ) = I(X, Z) +
I(X, Y |Z). Since I(X, Y |Z) is nonnegative by Gibbs’ inequality, we conclude
that I(X, Z) ≤ I(X, Y ), as was to be shown.
The Data Processing Inequality states that, in the absence of additional
information sources, processing leaves you with less information than you
started. The 2nd Law of Thermodynamics states that, in the absence of
additional energy sources, the system dynamics leave you with less order
than you started. These formulations suggest a natural parallel between the
concepts of information and order, and therefore a natural parallel between
the two theorems. We’ll close out this note with an extremely simplistic-yetsuggestive way to think about this.
Let X0 and Y0 each be random variables reflecting the possible locations
and momenta of two particles at time t = 0. We’ll assume (a) that the
particles don’t interact, but that (b) the experimenter has placed the two
particles very close to each other with similar momenta. Thus, the initial
configuration of the system is highly ordered, reflected by I(X0 , Y0 ) > 0. If
we knew Y0 , then we’d also significantly reduce our uncertainty about X0 .
How does this system evolve over time? We’re assuming no interactions, so
each of the particles evolve separately according to some dynamics, which
we can write X1 = gx (X0 ) and Y1 = gy (Y0 ). Using the data processing
inequality twice, we have
I(X1 , Y1 ) ≤ I(X0 , Y1 ) ≤ I(X0 , Y0 ).
(33)
Thus, the dynamics tend to reduce information. Of course, we can complicate
this picture in various ways, by considering particle interactions or external
potentials, either of which require a more sophisticated analysis. The full
2nd Law, which is beyond the scope of these notes, is most appropriate for
considering these cases.
8 Some Further Reading
This introduction has been far from exhaustive, and I heartily encourage
those interested to explore these topics in more detail. The below is a short
list of some of the resources I have found most intriguing and useful, in
addition to those cited in the introduction.
16
Information Theory “in General”
1. Shannon’s original work (Shannon, 1948) – in the words of one of my
professors, “the most important masters’ thesis of the 20th century.”
2. Shannon’s entertaining information-theoretic study of written English
(Shannon, 1951).
3. The text of Cover and Thomas (1991) is the standard modern overview
of the field for both theorists and practitioners.
4. Colah’s blog post “Visual Information Theory” at http://colah.github.io/posts/2015-09-Vi
is both entertaining and extremely helpful for getting basic intuition
around the relationship between entropy and communication.
Information Theory, Statistics, and Machine Learning
1. An excellent and entertaining introduction to these topics is the alreadymentioned MacKay (2003).
2. Those who want to further explore will likely enjoy Csiszar and Shields
(2004), but I would suggest doing this one after MacKay.
3. Readers interested in pursuing the Bayesian development of entropy
much more deeply may enjoy Bernardo and Smith (2008), which provides an extremely thorough development of decision theory with a
strong information-theoretic perspective.
4. The notes for the course “Information Processing and Learning” at
Carnegie-Mellon’s famous Machine Learning department are excellent
and accessible; find them at http://www.cs.cmu.edu/~aarti/Class/10704/lecs.html
Information Theory, Physics, and Biology
1. Marc Harper has a number of very fun papers in which he views biological evolutionary dynamics as learning processes through the framework
of information theory; a few are (Harper, 2009, 2010).
2. John Baez and his student Blake Pollard wrote a very nice and easyreading review article of the role of information concepts in biological
and chemical systems (Baez and Pollard, 2016).
3. More generally, John Baez’s blog is a treasure-trove of interesting vignettes and insights on the role that information plays in the physical
and biological worlds: https://johncarlosbaez.wordpress.com/category/information-and
For a more thoroughly worked-out connection between information dissipation and the Second Law of Thermodynamics, see this one: https://johncarlosbaez.wordp
17
References
Amari, S.-I. and Nagaoka, H. (2007). Methods of Information Geometry.
American Mathematical Society.
Baez, J. and Pollard, B. (2016). Relative entropy in biological systems.
Entropy, 18(2):46.
Bernardo, J. M. and Smith, A. F. (2008). Bayesian Theory. John Wiley and
Sons, New York.
Cover, T. M. and Thomas, J. A. (1991). Elements of Information Theory.
John Wiley and Sons, New York.
Csiszar, I. and Shields, P. C. (2004). Information Theory and Statistics: A
Tutorial. Foundations and Trends in Communications and Information
Theory, 1(4):417–528.
Harper, M. (2009). Information geometry and evolutionary game theory,
arXiv: 0911.1383. pages 1–13.
Harper, M. (2010). The replicator equation as an inference dynamic, arXiv:
0911.1763. pages 1–10.
MacKay, D. J. C. (2003). Information Theory, Inference, and Learning Algorithms. Cambridge Univeristy Press, 4th edition.
Shannon, C. E. (1948). A mathematical theory of communication. The Bell
System Technical Journal, 27:379–423.
Shannon, C. E. (1951). Prediction and entropy of printed English. Bell
System Technical Journal, 30(1):50–64.
18
| 7cs.IT
|
On the topological dimension of the Gromov boundaries of
some hyperbolic Out(FN )-graphs
Mladen Bestvina, Camille Horbez and Richard D. Wade
arXiv:1610.02115v1 [math.GR] 7 Oct 2016
October 10, 2016
Abstract
We give upper bounds, linear in the rank, to the topological dimensions of the
Gromov boundaries of the intersection graph, the free factor graph and the cyclic
splitting graph of a finitely generated free group.
Contents
1 Hyperbolic Out(FN )-graphs and their boundaries
1.1 The free factor graph . . . . . . . . . . . . . . . . .
1.2 The intersection graph . . . . . . . . . . . . . . . .
1.3 The cyclic splitting graph . . . . . . . . . . . . . .
1.4 More on mixing Z-averse trees . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5
5
6
6
7
2 Train-tracks, indices and stratifications
9
2.1 Train-tracks and carried trees . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Indices and stratifications . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3 Closedness of P (τ ) in its stratum
4 The
4.1
4.2
4.3
15
cells P (τ ) have dimension at most 0.
20
Existence of folding sequences of train-tracks directed by a boundary point 21
Openness of the set of points carried by a specialization or a fold . . . . . 27
Control on the diameter: getting finer and finer decompositions . . . . . . 29
5 End of the proof of the main theorem
30
A Why equivalence classes of vertices and specializations?
31
Introduction
The curve graph C(Σ) of an orientable hyperbolic surface of finite type Σ is an essential
tool in the study of the mapping class group of Σ. It was proved to be Gromov hyperbolic
1
by Masur–Minsky [22], and its Gromov boundary was identified with the space of ending
laminations on Σ by Klarreich [17]. A striking application of the curve graph and
its geometry at infinity is the recent proof by Bestvina–Bromberg–Fujiwara [2] that
Mod(Σ) has finite asymptotic dimension, which implies in turn that it satisfies the
integral Novikov conjecture. A crucial ingredient in the proof is the finite asymptotic
dimension of the curve graph: this was first proved by Bell–Fujiwara [4] using Masur–
Minsky’s tight geodesics, and recently recovered (with a bound linear in the genus of Σ
and the number of punctures) by Bestvina–Bromberg [1], via finite capacity dimension of
the Gromov boundary. The latter approach builds upon work of Gabai [10], who bounded
the topological dimension of ∂∞ C(Σ) by building covers in terms of train-tracks on the
surface; bounding the capacity dimension requires getting more metric control on such
covers.
The importance of the curve graph in the study of mapping class groups led people to
look for hyperbolic graphs with Out(FN )-actions. In the present paper, we will be mainly
interested in three of them: the free factor graph F FN , the cyclic splitting graph F ZN
and the intersection graph IN . The Gromov boundary of each graph is homeomorphic
to a quotient of a subspace of the boundary ∂CVN of outer space [5, 12, 14, 8]. As ∂CVN
has dimension equal to 3N − 5 [3, 11] and the quotient maps are cell-like, this bounds
the cohomological dimension of each of these Gromov boundaries. A priori this does not
imply finiteness of their topological dimensions. This is the goal of the present paper.
Main Theorem. The boundary ∂∞ IN has topological dimension at most 2N − 3.
The boundary ∂∞ F FN has topological dimension at most 2N − 2.
The boundary ∂∞ F ZN has topological dimension at most 3N − 5.
We do not know whether equality holds in any of these cases (see the open questions
at the end of the introduction). Following Gabai’s work, our proof relies on constructing
a decomposition of each Gromov boundary in terms of a notion of train-tracks. We hope
that, by getting further control on the covers we construct, this approach may pave
the way towards a proof of finite asymptotic dimension of F FN , F ZN or IN (via finite
capacity dimension of their Gromov boundaries).
The rest of the introduction is devoted to explaining the strategy of our proof. Although we treat the cases of ∂∞ IN , ∂∞ F FN and ∂∞ F ZN all at once in the paper, we
will mainly focus on the free factor graph in this introduction for simplicity, and only
say a word about how the proof works for F ZN (the proof for IN is similar).
Bounding the topological dimension. To establish our main theorem, we will use
the following two topological facts [9, Lemma 1.5.2 and Proposition 1.5.3]. Let X be a
separable metric space.
1. If X can be written as a finite union X = X0 ∪ X1 ∪ · · · ∪ Xk where each Xi is
0-dimensional, then dim(X) ≤ k.
Zero-dimensionality of each Xi will be proved by appealing to the following fact.
2. If there exists a countable cover of Xi by closed 0-dimensional subsets, then
dim(Xi ) = 0.
For example, one can recover the fact that dim(Rn ) ≤ n from the above two facts,
by using the decomposition of Rn where Xi is the set of points having exactly i rational
2
coordinates. Zero-dimensionality of each Xi can be proved using the second point, by
decomposing each Xi into countably many closed subsets Xij , two points being in the
same set precisely when they have the same rational coordinates (one easily checks that
these sets Xij are 0-dimensional by finding arbitrarily small boxes around each point of
Xij with boundary outside of Xij ).
Notice that a decomposition (hereafter, a stratification) of X into finitely many
subsets Xi as in the first point can be provided by a map X → {0, . . . , k} (hereafter,
an index map). Showing that each Xi is zero-dimensional amounts to proving that
every point in Xi has clopen neighborhoods (within Xi ) of arbitrary small diameter
(equivalently, every point in Xi has arbitrary small open neighborhoods in Xi with
empty boundary). Actually, thanks to the second point above, it is enough to write each
stratum Xi as a countable union of closed subsets (called hereafter a cell decomposition)
and prove 0-dimensionality of each cell of the decomposition.
Stratification and cell decomposition of ∂∞ F FN . The boundary ∂∞ F FN is a
separable metric space when equipped with a visual metric. Points in the boundary are
represented by FN -trees, and a first reasonable attempt could be to define a stratification
of ∂∞ F FN by using an index map similar to the one introduced by Gaboriau–Levitt in
[11], that roughly counts orbits of branch points and of directions at these points in
the trees. Although we make use of these features, our definition of the stratification is
slightly different; it is based on a notion of train-tracks.
A train-track τ consists of a triple (S τ , ∼V , ∼D ), where S τ is a free FN -action on a
simplicial tree, and ∼V (respectively ∼D ) is an equivalence relation on the set of vertices
(respectively directions) in S τ . We then say that a tree T ∈ cv N is carried by τ if there is
an FN -equivariant map f : S τ → T (called a carrying map) that is in a sense compatible
with the train-track structure. The typical situation is when f identifies two vertices in
S τ if and only if they are ∼V -equivalent, and identifies the germs of two directions based
at equivalent vertices in S τ if and only if the directions are ∼D -equivalent. For technical
reasons, our general definition of a carrying map is slightly weaker; in particular it is a
bit more flexible about possible images of vertices in S τ at which there are only three
equivalence classes of directions. A point ξ ∈ ∂∞ F FN – which is an equivalence class of
trees that all admit alignment-preserving bijections to one another – is then carried by
τ if some (equivalently, any) representative is carried.
A train-track τ as above has an index i(τ ) of at most 2N − 2: this is a combinatorial
datum which mainly counts orbits of equivalence classes of vertices and directions in
τ . The index i(ξ) of a point ξ ∈ ∂∞ F FN is then defined as the maximal index of a
train-track that carries ξ. A train-track determines a cell P (τ ) in ∂∞ F FN , defined as
the set of points ξ ∈ ∂∞ F FN such that τ carries ξ and i(τ ) = i(ξ).
We define a stratification of ∂∞ F FN by letting Xi be the collection of points of index
i. Each Xi is covered by the countable collection of the cells P (τ ), where τ varies over all
train-tracks of index i. In view of the topological facts recalled above, our main theorem
follows for ∂∞ F FN from the following points:
1. (Proposition 3.1) The boundary of P (τ ) in ∂∞ F FN is contained in a union of cells
of strictly greater index. This implies that P (τ ) is closed in Xi(τ ) .
2. (Proposition 4.1) Each cell P (τ ) has dimension at most 0.
We say that P (τ ) has dimension ‘at most’ 0, and not ‘equal to’ 0, as we do not
exclude the possibility that P (τ ) is empty.
3
For the first point, we show that if (ξn )n∈N ∈ P (τ )N converges to ξ ∈ ∂P (τ ), then
the carrying maps fn from S τ to representatives of ξn converge to a map f from S τ to
a representative T of ξ. However, the limiting map f may no longer be a carrying map:
for example, inequivalent vertices in S τ may have the same image in the limit, and edges
in S τ may be collapsed to a point by f . We then collapse all edges in S τ that map to
a point under f and get an induced map f ′ from the collapse S ′ to T . This determines
a new train-track structure τ ′ on S ′ . A combinatorial argument enables us to count the
number of directions ‘lost’ when passing from τ to τ ′ , and show that i(τ ′ ) > i(τ ) unless
T is carried by τ .
Our proof of the second point relies on a cell decomposition process similar to Gabai’s
[10], who used splitting sequences of train-tracks on surfaces to get finer and finer covers
of the space of ending laminations. In our context, starting from a cell P (τ ), we will
construct finer and finer decompositions of P (τ ) into clopen subsets by means of folding
sequences of train-tracks. Eventually the subsets in the decomposition have small enough
diameter.
More precisely, starting from a train-track τ one can ‘resolve’ an illegal turn in τ
by folding it. There are several possibilities for the folded track (see the figures in
Section 4), so this operation yields a subdivision of P (τ ) into various P (τ j ). A first
crucial fact we prove is that these P (τ j ) are all open in P (τ ). Now, if ξ ∈ P (τ ) and
ε > 0, then by folding the train-track τ for sufficiently long we can reach a train-track τ ′
with diam(P (τ ′ )) < ε. Here, hyperbolicity of F FN is crucial: our folding sequence will
determine an (unparameterized) quasi-geodesic going to infinity (towards ξ) in F FN .
The set P (τi ) defined by the train-track obtained at time i of the process is contained in
the set of endpoints of geodesic rays in F FN starting at the simplicial tree associated to
τ0 and passing at a bounded distance from τi . From the definition of the visual metric on
the boundary, this implies that the diameter of P (τi ) converges to 0 as we move along the
folding path. All in all, we find a train-track τ ′ such that P (τ ′ ) is an open neighborhood
of ξ in P (τ ), and diam(P (τ ′ )) < ε. The boundary of P (τ ′ ) in P (τ ) is empty (it contains
points of strictly greater index), so we get dim(P (τ )) = 0, as required.
A word on ∂∞ F ZN . Some points in ∂∞ F ZN are represented by trees in which some
free factor system of FN is elliptic. This leads us to work with train-tracks that are
allowed to have bigger vertex stabilizers, and the index of a train-track now also takes
into account the complexity of this elliptic free factor system. A priori, our definition of
index in this setting only yields a quadratic bound in N to the topological dimension of
∂∞ F ZN , but we then get the linear bound from cohomological dimension. We also have
to deal with the fact that a point ξ ∈ ∂∞ F ZN is an equivalence class of trees that may
not admit alignment-preserving bijections between different representatives; this leads
us to analysing preferred (mixing) representatives of these classes more closely.
Organization of the paper. In Section 1, we review the definitions of F FN , IN
and F ZN , and the descriptions of their Gromov boundaries. We also prove a few facts
concerning mixing representatives of points in ∂∞ F ZN , which are used to tackle the last
difficulty mentioned in the above paragraph. Train-tracks and indices are then defined
in Section 2, and the stratifications of the Gromov boundaries are given. We prove in
Section 3 that each cell P (τ ) is closed in its stratum, by showing that ∂P (τ ) is made of
points with strictly higher index. We introduce folding moves in Section 4, and use them
to prove that each cell P (τ ) has dimension at most 0. The proof of our main theorem
4
is then completed in Section 5. The paper has an appendix in which we illustrate the
motivations behind some technical requirements that appear in our definitions of traintracks and carrying maps.
Some open questions. As mentioned earlier, we hope that the cell decompositions
of the boundaries defined in the present paper may provide a tool to tackle the question
of finiteness of the asymptotic dimension of the various graphs, following the blueprint
of [1].
Gabai uses the cell decomposition of the ending lamination space EL to show that
it is highly connected for sufficiently complicated surfaces [10] (connectivity had been
established earlier for most surfaces by Leininger–Schleimer [18]). It is unknown whether
the boundaries of IN , F FN or F ZN are connected. The question of local connectivity
of these boundaries (or the related question of local connectivity of the boundary of
Culler–Vogtmann’s outer space) are also open to our knowledge.
Finally, we would like to address the question of finding lower bounds on the dimensions. Gabai shows that the topological dimension of EL is bounded from below by
3g + p − 4 (where g is the genus of the surface and p is the number of punctures). Since
the ending lamination space of a once-punctured surface sits as a subspace of ∂∞ F F2g
and ∂∞ F Z2g , this gives a lower bound on the dimension of ∂∞ F F2g and ∂∞ F Z2g . Improving the gap between the upper and lower bounds, as well as finding a lower bound
for ∂∞ IN , is an interesting problem.
Acknowledgments. We would like to thank Patrick Reynolds for conversations we
had related to the present project, and Vera Tonić for pointing out to us the reference
[9] relating topological and cohomological dimensions. The present work was supported
by the National Science Foundation under Grants No. DMS-1440140 while the authors
were in residence at the Mathematical Sciences Research Institute in Berkeley, California,
during the Fall 2016 semester. The first author was supported by the NSF under Grant
No. DMS-1607236.
1
Hyperbolic Out(FN )-graphs and their boundaries
We review the definitions of three hyperbolic Out(FN )-graphs (the free factor graph,
the intersection graph and the Z-splitting graph) and the descriptions of their Gromov
boundaries. The only novelties are some facts concerning mixing Z-averse trees in the
final subsection.
1.1
The free factor graph
The free factor graph F FN is the graph (equipped with the simplicial metric) whose
vertices are the conjugacy classes of proper free factors of FN . Two conjugacy classes
[A] and [B] of free factors are joined by an edge if they have representatives A, B such
that A ( B or B ( A. Its hyperbolicity was proved in [4].
To describe its Gromov boundary, we first recall that unprojectivized outer space cvN
is the space of all FN -equivariant isometry classes of minimal, free, simplicial, isometric
FN -actions on simplicial metric trees. Its closure cv N (for the equivariant Gromov–
Hausdorff topology) was identified in [6, 4] with the space of all minimal very small
FN -trees, i.e. FN -actions on R-trees in which all stabilizers of nondegenerate arcs are
5
cyclic (possibly trivial) and root-closed, and tripod stabilizers are trivial. There exists a
coarsely Out(FN )-equivariant map π : cvN → F FN , which sends a tree T to a free factor
that is elliptic in a tree T obtained from T by collapsing some edges to points.
An FN -tree T ∈ ∂cvN is arational if no proper free factor of FN is elliptic in T , or acts
with dense orbits on its minimal subtree. Two arational trees T and T ′ are equivalent
(T ∼ T ′ ) if there exists an FN -equivariant alignment-preserving bijection from T to T ′ .
We denote by AT N the subspace of ∂cvN made of arational trees. Arational trees were
introduced by Reynolds in [25], where he proved that every arational FN -tree is either
free, or dual to an arational measured lamination on a once-holed surface.
Theorem 1.1. (Bestvina–Reynolds [5], Hamenstädt [12]) There is a homeomorphism
∂π : AT N /∼ → ∂∞ F FN ,
which extends the map π continuously to the boundary. By this we mean that for all
sequences (Sn )n∈N ∈ (cvN )N converging to a tree S∞ ∈ AT N (for the topology on cv N ),
the sequence (π(Sn ))n∈N converges to ∂π(S∞ ) (for the topology on F FN ∪ ∂∞ F FN ).
1.2
The intersection graph
A conjugacy class α of FN is geometric if it is either part of a free basis of FN , or
else corresponds to the boundary curve of a once-holed surface with fundamental group
identified with FN . The intersection graph IN (with Mann’s definition [21], a variation
on Kapovich–Lustig’s [16]) is the bipartite graph whose vertices are the simplicial FN trees in ∂cvN together with the set of geometric conjugacy classes of FN . A tree T is
joined by an edge to a conjugacy class α whenever α is elliptic in T . Its hyperbolicity
was proved in [21]. The intersection graph is also quasi-isometric to Dowdall–Taylor’s
co-surface graph [8, Section 4]. We denote by FAT N ⊆ AT N the space of free arational
trees in cv N . Again, there is a coarsely Out(FN )-equivariant map π : cvN → IN , which
sends a tree T ∈ cvN to a tree T ∈ ∂cvN obtained by collapsing some of the edges of T
to points.
Theorem 1.2. (Dowdall–Taylor [8]) There is a homeomorphism
∂π : FAT N /∼ → ∂∞ IN ,
which extends the map π continuously to the boundary.
Continuity of the extension is understood in the same way as in the statement of
Theorem 1.1 above.
1.3
The cyclic splitting graph
A cyclic splitting of FN is a simplicial, minimal FN -tree in which all edge stabilizers are
cyclic (possibly trivial). The Z-splitting graph F ZN is the graph whose vertices are the
FN -equivariant homeomorphism classes of Z-splittings of FN . Two splittings are joined
by an edge if they have a common refinement. Its hyperbolicity was proved by Mann
[20].
We recall that two trees T, T ′ ∈ cv N are compatible if there exists a tree Tb ∈ cv N that
admits alignment-preserving FN -equivariant maps onto both T and T ′ . A tree T ∈ cv N
is Z-averse if it is not compatible with any tree T ′ ∈ cv N which is compatible with a
6
Z-splitting of FN . We denote by XN the subspace of cv N made of Z-averse trees. Two
Z-averse trees are equivalent if they are both compatible with a common tree; although
not obvious, this was shown in [14] to be an equivalence relation on XN , which we denote
by ≈.
There is an Out(FN )-equivariant map π : cvN → F ZN , given by forgetting the
metric.
Theorem 1.3. (Hamenstädt [12], Horbez [14]) There is a homeomorphism
∂π : XN /≈ → ∂∞ F ZN ,
which extends the map π continuously to the boundary.
Each ≈-equivalence class in XN has preferred representatives that are mixing. We
recall that a tree T ∈ cv N is mixing if for all segments I, J ⊆ T , there exists a finite set
{g1 , . . . , gk } ⊆ FN such that J is contained in the union of finitely many translates gi I.
Theorem 1.4. (Horbez [14]) Every ≈-class in XN contains a mixing tree, and any two
mixing trees in the same ≈-class admit FN -equivariant alignment-preserving bijections
between each other. Any Z-averse tree admits an FN -equivariant alignment-preserving
map onto every mixing tree in its ≈-class.
The space of arational trees AT N is contained in the space of Z-averse trees XN ,
and the equivalence relation ∼ we have defined on AT N is the restriction to AT N of
the equivalence relation ≈ on XN (all arational trees are mixing [25]). The inclusion
AT N ⊆ XN then induces a subspace inclusion ∂∞ F FN ⊆ ∂∞ F ZN .
1.4
More on mixing Z-averse trees
In this section, we establish a few more facts concerning mixing Z-averse trees and their
possible point stabilizers, building on the work in [14]. If we were only concerned with
the free factor graph, we would not need these results. The reader may decide to skim
through the section and avoid the technicalities in the proofs in a first reading.
Lemma 1.5. Let T ∈ cv N be mixing and Z-averse, and let A ⊆ FN be a proper free
factor. Then the A-action on its minimal subtree TA ⊆ T is discrete (possibly TA is
reduced to a point).
In particular, if no proper free factor is elliptic in T , then T is arational.
Proof. Assume towards a contradiction that TA is not simplicial. Since T has trivial
arc stabilizers, the Levitt decomposition of TA [19] has trivial arc stabilizers (it may be
reduced to a point in the case where T has dense orbits). Let B ⊆ A be a free factor of A
(hence of FN ) which is a vertex group of this decomposition. The B-minimal subtree TB
of T (which is also the B-minimal subtree of TA ) has dense orbits, so by [24, Lemma 3.10]
the family {gTB |g ∈ FN } is a transverse family in T . As T is mixing, it is a transverse
covering. In addition, by [25, Corollary 6.4] the stabilizer of TB is equal to B. But by
[14, Proposition 4.23], the stabilizer of a subtree in a transverse covering of a mixing
Z-averse tree cannot be a free factor (in fact, it cannot be elliptic in a Z-splitting of
FN ), so we get a contradiction.
A collection A of subgroups of FN is a free factor system if it coincides with the
set of nontrivial point stabilizers in some simplicial FN -tree with trivial arc stabilizers.
7
There is a natural order on the collection of free factor systems, by saying that a free
factor system A is contained in a free factor system A′ whenever every factor in A is
contained in one of the factors in A′ .
Proposition 1.6. Let T ∈ cv N be mixing and Z-averse. Then either T is dual to an
arational measured lamination on a closed hyperbolic surface with finitely many points
removed, or else the collection of point stabilizers in T is a free factor system.
Proof. First assume that there does not exist any free splitting of FN in which all point
stabilizers of T are elliptic. Then, with the terminology of [13, Section 5], the tree T is
of surface type (by the same argument as in [13, Lemma 5.8]). Since T is Z-averse, the
skeleton of the dynamical decomposition of T is reduced to a point, in other words T is
dual to an arational measured lamination on a surface.
Assume now that there exists a free splitting S of FN in which all point stabilizers of
T are elliptic. Let A be the smallest free factor system such that every point stabilizer
of T is contained within some factor in A. By Lemma 1.5, each factor A in A acts
discretely on its minimal subtree TA . In addition TA has trivial arc stabilizers because
T has trivial arc stabilizers. So either A is elliptic in T , or else TA is a free splitting of
A. The second situation cannot occur, as otherwise the point stabilizers in TA would
form a free factor system of A, contradicting the minimality of A. Hence A coincides
with the collection of point stabilizers of T .
We will also establish a characterization of the mixing representatives in a given
equivalence class of Z-averse trees. Before that, we start with the following lemma.
Lemma 1.7. Let T, T ′ ∈ cv N . Assume that T has dense orbits, and that there exists an
FN -equivariant alignment-preserving map p : T → T ′ .
Then either p is a bijection, or else there exists g ∈ FN that is elliptic in T ′ but not in
T . In the latter case, some point stabilizer in T ′ is non-cyclic.
Proof. Notice that surjectivity of p follows from the minimality of T ′ . The collection
of all subtrees of the form p−1 ({x}) with x ∈ T ′ is a transverse family in T . If p is
not a bijection, then one of the subtrees Y in this family is nondegenerate, and [25,
Proposition 7.6] implies that its stabilizer A = Stab(p(Y )) is nontrivial, and that the
A-action on Y is not discrete. This implies that the minimal A-invariant subtree of Y is
not reduced to a point, and A contains an element that acts hyperbolically on T , while
it fixes a point in T ′ . In addition A is not cyclic, so the last assertion of the lemma
holds.
Proposition 1.8. A Z-averse tree T is mixing if and only if for all T ′ ≈ T , every
element of FN that is elliptic in T ′ is also elliptic in T .
If T is dual to an arational lamination on a surface, then all trees in the ≈-class of T
are mixing.
Proof. If T is mixing, then all trees T ′ in its ≈-class admit an FN -equivariant alignmentpreserving map onto T , so every element elliptic in T ′ is also elliptic in T . If T is not
mixing, then it admits an FN -equivariant alignment-preserving map p onto a mixing tree
T in the same ≈-class, and p is not a bijection. By Lemma 1.7, there exists an element
of FN that is elliptic in T but not in T . The second assertion in Proposition 1.8 follows
from the last assertion of Lemma 1.7 because if T is dual to an arational lamination on
a surface, then T is mixing and all point stabilizers in T are cyclic.
8
Corollary 1.9. Let T be a Z-averse tree, and let A be a maximal free factor system
elliptic in T . Let T be a mixing representative of the ≈-class of T .
Then A is a maximal elliptic free factor system in T if and only if T is mixing.
Proof. If T is mixing, then T and T have the same point stabilizers, so the conclusion
is obvious. If T is not mixing, then Proposition 1.8 implies that T is not dual to a
lamination on a surface, and that there exists an element g ∈ FN not contained in A
that is elliptic in T . Proposition 1.6 shows that the collection of elliptic subgroups in T
is a free factor system of FN , and this free factor system strictly contains A.
2
Train-tracks, indices and stratifications
2.1
Train-tracks and carried trees
Definition 2.1 (Train-tracks). A train-track τ is the data of
• a minimal, simplicial FN -tree S τ with trivial edge stabilizers,
• an equivalence relation ∼τV on the set V (S τ ) of vertices of S τ , such that if v ∼τV
v ′ , then gv ∼τV gv ′ for all g ∈ FN , and such that no two adjacent vertices are
equivalent,
• for each ∼τV -class X, an equivalence relation ∼τD,X on the set D(X) of directions
at the vertices in X, such that if two directions d, d′ ∈ D(X) are equivalent, then
for all g ∈ FN , the directions gd, gd′ ∈ D(gX) are also equivalent.
Equivalence classes of directions at X are called gates at X.
We denote by A(τ ) the free factor system made of all point stabilizers in S τ .
Remark 2.2. Including an equivalence relation on the vertex set of S τ in the definition
may look surprising to the reader, as this is not standard in train-track theory for free
groups. Roughly speaking, the equivalence classes of vertices correspond to branch points
in trees carried by a track; this is explained in more detail in the appendix of the paper.
We will usually also impose that the train-tracks we work with satisfy some additional
assumptions. Let τ be a train-track. A pair (d, d′ ) of directions based at a common vertex
of S τ is called a turn, and is said to be legal if d and d′ are inequivalent. A subtree A ⊂ S τ
crosses the turn (d, d′ ) if the intersection of A with both directions d and d′ is non-empty.
We say that A is legal in τ if each turn crossed by A is legal in τ .
Definition 2.3 (Admissible train-tracks). A train-track τ is admissible if for every
vertex v ∈ S τ , there exist three pairwise inequivalent directions d1 , d2 , d3 at v such that
for all i ∈ {1, 2, 3}, there exists an element gi ∈ FN acting hyperbolically on S τ whose
axis in S τ is legal and crosses the turn (di , di+1 ) (mod 3).
In particular, there are at least three gates at every equivalence class of vertices in
admissible train-tracks. We will call a triple (g1 , g2 , g3 ) as in Definition 2.3 a tripod of
legal elements at v, and we say that their axes form a tripod of legal axes at v.
Edges in simplicial trees are given an affine structure, which enables us to consider
maps from simplicial trees to R-trees that are linear on edges. If f : S → T is linear then
there is a unique metric on S for which f is isometric when restricted to every edge. We
say that this is the metric on S determined by the linear map f .
9
Train-tracks naturally arise from morphisms between trees. Suppose that f : S → T
is an FN -equivariant map from a simplicial FN -tree S to an FN -tree T which is linear on
edges and does not collapse any edge of S to a point. There is an induced equivalence
relation ∼V on the set of vertices of S given by saying that two vertices are equivalent
if they have the same f -image in T . Let w be a point in T and let X := f −1 (w) ∩ V (S)
be its associated equivalence class in V (S). As soon as X 6= ∅, equivariance of f implies
that its (setwise) stabilizer Stab(X) in S is equal to the stabilizer of the point w in T .
Let D(X) be the set of directions in S based at points in X. Since f does not collapse
any edge, it induces an equivalence relation ∼D,X on D(X), where two directions in
D(X) are equivalent if they have germs that map into the same direction at w in T . The
set of Stab(X)-orbits of equivalence classes in D(X) then maps injectively into the set
of Stab(w)-orbits of directions at w in T . We call this collection of equivalence classes
τf the train-track structure on S induced by f . This will be the typical example where
a tree T is carried by the train-track τf . Our general definition of carrying is slightly
more technical, as it involves a bit more flexibility with respect to the definition of f at
exceptional classes of vertices, defined as follows.
Definition 2.4 (Exceptional classes of vertices). Let τ be an admissible train-track.
An equivalence class X of vertices of S τ is called exceptional if there are exactly 3 gates
at X.
Definition 2.5 (Specialization). Let τ and τ ′ be two train-tracks, and let [v0 ] ∈ S τ
be an exceptional equivalence class of vertices. We say that τ ′ is a specialization of τ at
[v0 ] if
′
• Sτ = Sτ ,
′
• there exists a vertex v1 ∈ S τ , not in the same orbit as v0 , such that ∼τV is the
′
coarsest FN -invariant equivalence relation finer than ∼τV and such that v1 ∼τV v0 ,
′
• if d, d′ are two directions not based at any vertex in the orbit of [v0 ], then d ∼τ d′
if and only if d ∼τ d′ ,
• every direction at a vertex in [v0 ] is equivalent to some direction at a vertex which
is τ -equivalent to v1 .
Definition 2.6 (Carrying ). We say that an FN -tree T is carried by a train-track τ
if there is an FN -equivariant map f : S τ → T , which is linear on edges and does not
collapse any edge to a point, such that the train-track structure τf induced by f is obtained
from τ by a finite (possibly trivial) sequence of specializations.
In this situation, we call the map f a carrying map (with respect to τ ).
Remark 2.7. The motivation for introducing specializations and allowing them in the
definition of carrying is again explained in the appendix; specializations will appear
naturally later in the paper when we start performing folds on tracks.
Lemma 2.8. The collection of all train-tracks τ for which there exists a tree T ∈ cv N
carried by τ is countable.
Proof. There are countably many minimal, simplicial FN -trees with trivial edge stabilizers. In addition, if τ carries a tree T ∈ cv N , then the stabilizer of every equivalence
class X of vertices in τ is a finitely generated subgroup of FN (because every point
10
stabilizer in a very small FN -tree is finitely generated by [11]). The set X/Stab(X) is
finite because there are finitely many FN -orbits of vertices in S τ , and if v, gv are two
vertices in X, then g ∈ Stab(X). The stabilizer of every gate [d] is at most cyclic, and
again [d]/Stab([d]) is finite. The equivalence relation on vertices is recovered by taking
a finite set A1 , . . . , Ak of representatives of FN -orbits of equivalence classes, and in each
Ai taking a finite set Vi of vertices such that Stab(Ai ).Vi = Ai . Each equivalence class
is of the form gStab(Ai ).Vi , for some g ∈ FN . The equivalence relation on edges is
recovered by taking a finite set B1 , . . . , Bl of representatives of FN -orbits of gates in τ ,
and for each Bi taking a finite set Ei of representatives of FN -orbits of oriented edges
determining the directions in Bi . Then each gate in τ is of the form gStab(Bi ).Ei for
some g ∈ FN . Hence τ is determined by the simplicial tree S τ and the finite family
({Stab(Ai )}, {Vi }, {Stab(Bi )}, {Ei }), which gives a countable number of possible traintracks.
We will also need the following observation.
Lemma 2.9. Let τ be an admissible train-track, and let [a, b] ⊆ S τ be a legal segment.
Then there exists an element g ∈ FN acting hyperbolically on S τ , whose axis in S τ is
legal and contains [a, b].
Proof. Since τ is admissible, there exist elements g, h ∈ FN acting hyperbolically in S τ ,
whose axes Ag , Ah are legal and pass through a, b respectively, only meet [a, b] at one
extremity, and such that the subtree Y := Ag ∪ [a, b] ∪ Ah is legal. Standard properties of
group actions on trees [7] imply that all turns in the axis of gh are contained in translates
of Y . So the axis of gh is legal, and it contains [a, b].
The following two lemmas state that every tree T ∈ cv N is carried by an admissible
train-track, and in addition the carrying map f : S τ → T is completely determined by
the train-track structure. Given a free factor system A, we say that a tree T ∈ cv N is
an A-tree if all subgroups in A are elliptic in T .
Lemma 2.10. Let A be a free factor system, and let T ∈ cv N be an A-tree with trivial
arc stabilizers. Then there exists an admissible train-track τ such that T is carried by τ
and A(τ ) = A.
Proof. Without loss of generality, we can assume that A is a maximal free factor system
such that T is an A-tree. We let A/FN := {[A1 ], . . . , [Ak ]}. If A =
6 ∅, then (up to
replacing the subgroups Ai by appropriate conjugates) we can choose for S the universal
cover of the graph of groups depicted on Figure 1. There is a unique FN -equivariant map
f : S → T that is linear on edges: indeed, every vertex v of S has nontrivial stabilizer
Gv , and must be sent by f to the unique point in T fixed by Gv . The elements ai and
the subgroups A2 , . . . , Ak cannot fix the same point as A1 in T by maximality of A, so
f does not collapse any edge, and therefore τf is well-defined. Admissibility of τf can
be checked by taking advantage of the fact that all vertex groups are infinite: one can
construct the required legal elements by taking appropriate products of elliptic elements
in T .
We now assume that A = ∅. Let {a1 , . . . , aN } be a free basis of FN such that
A = ha1 , a2 i acts freely with discrete orbits on T , and no ai is elliptic in T . Such a basis
exists as the reducing factors of T (if there are any) form a bounded subset of F FN , see
[5, Corollary 5.3]. Let TA be the minimal A-invariant subtree for the A-action on T . Let
ΓA = TA /A. Then ΓA is a finite graph in cv2 with one or two vertices. We extend ΓA
11
a1
A2
A1
al
Ak
Figure 1: The tree S in the proof of Lemma 2.10, in the case where A 6= ∅.
to a marked FN –graph Γ with the same number of vertices, by attaching loops labelled
by the elements a3 , . . . , aN at a fixed vertex of ΓA . Let S be the universal cover of Γ.
The isometric embedding from TA into T extends to a map f : S → T , and f does
not collapse any edges because no element ai is elliptic in T . The induced train-track
τf is admissible as each vertex of TA has valence at least 3 and f is an FN -equivariant
isometric embedding when restricted to TA .
Lemma 2.11. Let T ∈ cv N , and let τ be an admissible train-track such that T is carried
by τ . Then there exists a unique carrying map f : S τ → T for the train-track structure.
Furthermore, the carrying map f varies continuously on the set of trees that are carried
by τ (in the equivariant Gromov–Hausdorff topology).
Proof. To prove uniqueness, it is enough to show that the f -image of any vertex v ∈ S τ
is completely determined by the train-track structure. Let (g1 , g2 , g3 ) be a tripod of legal
elements at v, and let e1 , e2 , e3 be the three edges at v taken by their axes. Then f is
isometric when restricted to e1 ∪ e2 ∪ e3 , and also when restricted to the axes of g1 , g2 , g3 .
This implies that the intersection of the axes of g1 , g2 , g3 in T is reduced to a point, and
f must send v to that point. Continuity also follows from the above argument.
In order to define what it means for an equivalence class of Z-averse trees to be
carried by a train-track, we will need the following lemma.
Lemma 2.12. Let T, T ′ ∈ cv N , and let τ be a train-track.
If there is an FN -equivariant alignment-preserving bijection from T to T ′ , then T is
carried by τ if and only if T ′ is carried by τ .
Proof. Let f : S τ → T be a carrying map, and let θ : T → T ′ be an FN -equivariant
alignment-preserving bijection. Let f ′ : S τ → T ′ be the unique linear map that coincides
with θ ◦ f on the vertices of S τ . We claim that f ′ is a carrying map. Indeed, since θ is
a bijection, two vertices in S τ have the same f ′ -image if and only if they have the same
f -image. In addition, since θ preserves alignment, the germs of two directions in S τ are
identified by f ′ if and only if they are identified by f .
We recall that XN ⊆ cv N denotes the subspace made of Z-averse trees.
Definition 2.13 (Carrying equivalence classes of Z-averse trees). An equivalence class ξ ∈ XN /≈ is carried by a train-track τ if some (equivalently, any) mixing
representative of ξ is carried by τ .
12
2.2
Indices and stratifications
We now define stratifications of the Gromov boundaries of IN , F FN and F ZN by means
of an index function taking finitely many values. We will define both the index of a
train-track and the index of a tree T ∈ cv N , and explain how the two are related, before
defining the index of a boundary point (which is an equivalence class of trees).
2.2.1
Geometric index of a tree in cv N
The following definition is reminiscent of Gaboriau–Levitt’s index for trees in cv N [11],
with a slight difference in the constants we use.
Definition 2.14 (Geometric index of a tree T ∈ cv N ). Let T ∈ cv N be a tree with
trivial arc stabilizers. The geometric index of T is defined as
X
igeom (T ) :=
(αv + 3rank(Stab(v)) − 3),
v∈V (T )/FN
where V (T ) denotes the set of branch points in T , and αv denotes the number of FN orbits of directions at v.
Lemma 2.15. For all T ∈ cv N with trivial arc stabilizers, we have igeom (T ) ≤ 3N − 3.
If T is arational, then igeom (T ) ≤ 2N − 2.
If T is free and arational, then igeom (T ) ≤ 2N − 3.
Proof. Let
iGL (T ) :=
X
(αv + 2rank(Stab(v)) − 2).
v∈V (T )/FN
We have igeom (T ) ≤ 23 iGL (T ) for all T ∈ cv N , and Gaboriau–Levitt proved in [11,
Theorem III.2] that iGL (T ) ≤ 2N − 2, so igeom (T ) ≤ 3N − 3. In addition we have
X
igeom (T ) = iGL (T ) +
(rank(Stab(v)) − 1).
(1)
v∈V (T )/FN
Assume now that T is arational. If the FN -action on T is free, since there is at least
one orbit of branch points in T , we get from Equation (1) that igeom (T ) < iGL (T ) ≤
2N − 2. Otherwise, by [25], the FN -action on T is dual to a measured lamination on a
once-holed surface, so all point stabilizers in T are either trivial or cyclic, and we get
igeom (T ) ≤ iGL (T ) ≤ 2N − 2.
2.2.2
Index of a train-track and carrying index of a tree
The height h(A) of a free factor system A is defined as the maximal length k of a proper
chain of free factor systems ∅ = A0 ( A1 ( · · · ( Ak = A. We recall that the free
factor system A(τ ) associated to a train-track τ is the free factor system consisting of
the vertex stabilizers in the associated tree S τ .
Definition 2.16 (Index of a train-track ). The geometric index of a train-track τ is
defined as
k
X
(αi + 3ri − 3),
igeom (τ ) :=
i=1
13
where the sum is taken over a finite set {X1 , . . . , Xk } of representatives of the FN -orbits
of equivalence classes in V (S τ ), αi denotes the number of Stab(Xi )-orbits of gates at the
vertices in Xi , and ri denotes the rank of Stab(Xi ).
The index of τ is defined to be the pair i(τ ) := (h(A(τ )), igeom (τ )).
Indices of train-tracks will be ordered lexicographically.
Remark 2.17. If τ is admissible, then every equivalence class X of vertices in τ has a
nonnegative contribution to the geometric index of τ . The contribution of a class X to
the geometric index of τ is zero if and only if Stab(X) is trivial, and there are exactly
three gates at X; this is precisely when X is exceptional.
Lemma 2.18. Let T ∈ cv N be a tree with trivial arc stabilizers, and let τ be an admissible
train-track that carries T . Then igeom (τ ) ≤ igeom (T ).
Proof. Let f : S τ → T be the unique carrying map. Since τ is admissible, vertices in S τ
are mapped to branch points in T . As f is a carrying map, distinct equivalence classes
of nonexceptional vertices in S τ are mapped to distinct branch points in T . If X is
an equivalence class in S τ mapping to a branch point v ∈ T then Stab(X) = Stab(v).
Furthermore, one checks that two gates based at X in distinct Stab(X)-orbits are mapped
under f to directions at v in distinct Stab(v)-orbits. Since exceptional vertices do not
contribute to the geometric index, it follows that igeom (τ ) ≤ igeom (T ).
Notice that the inequality from Lemma 2.18 might be strict if some branch direction
in the tree T is not “visible” in the track. In the following definition, instead of directly
counting branch points and branch directions in T , we will count the maximal number
of such directions that are visible from a train-track.
Definition 2.19 (Carrying index of a tree T ∈ cv N ). Let T ∈ cv N be a tree with
trivial arc stabilizers. We define the carrying index of T , denoted by i(T ), as the maximal
index of an admissible train-track τ that carries T .
An ideal carrier for T is an admissible train-track τ that carries T such that i(T ) = i(τ ).
Remark 2.20. In view of Lemma 2.10, if τ is an ideal carrier of a tree T , then A(τ ) is a
maximal free factor system that is elliptic in T .
Lemma 2.21. The index of a tree T ∈ cv N with trivial arc stabilizers can only take
boundedly many values (with a bound only depending on N ).
The index of an arational tree is comprised between (0, 0) and (0, 2N − 2).
The index of a free arational tree is comprised between (0, 0) and (0, 2N − 3).
Proof. The first assertion is a consequence of Proposition 2.15, together with the fact
that there is a bound (only depending on N ) on the height of a free factor system of
FN . The other assertions follow from Proposition 2.15 because no nontrivial free factor
is elliptic in an arational tree, so the height h(A(τ )) for a train-track τ that carries an
arational tree is 0.
2.2.3
Index of a boundary point and stratifications
We now define the index of an equivalence class ξ of Z-averse trees. We recall that by
definition a train-track τ carries ξ if and only if τ carries the mixing representatives of
ξ.
14
Definition 2.22 (Index of a boundary point). The index i(ξ) of an equivalence class
ξ ∈ XN /≈ is defined as the maximal index of an admissible train-track that carries ξ
(equivalently i(ξ) is the carrying index of the mixing representatives of ξ).
Given an admissible train-track τ , we define the cell P (τ ) as the subspace of XN /≈ made
of all classes ξ that are carried by τ and such that i(ξ) = i(τ ).
For all i, we define the stratum Xi ⊆ XN /≈ as the set of all points ξ ∈ XN /≈ such that
i(ξ) = i.
The Gromov boundary ∂∞ F ZN can be written as the union of all strata Xi , where
i varies over the finite set of all possible indices for mixing Z-averse trees.
If a tree T ∈ cv N is mixing and Z-averse, and if no proper free factor is elliptic in
T , then Lemma 1.5 says that T is arational. Therefore, the boundary ∂∞ F FN coincides
with the subspace of XN /≈ which is the union of all strata Xi with i comprised between
(0, 0) and (0, 2N − 2).
′
′
Finally, the boundary ∂∞ IN can be written as the union X0′ ∪ · · · ∪ X2N
−3 , where Xi
is the subspace of X(0,i) made of equivalence classes of free actions.
In view of the topological facts recalled in the introduction, we are left showing that
each cell P (τ ) is closed in its stratum Xi(τ ) , and that dim(P (τ )) ≤ 0. This will be the
contents of Sections 3 and 4, respectively. We give a complete overview of the proof of
our main theorem in Section 5.
3
Closedness of P (τ ) in its stratum
In general, the property of a tree being carried by a train-track is not a closed condition.
However, cells determined by train-tracks are closed in their own strata. This is the goal
of the present section. As each boundary we study is metrizable, throughout the paper
we will use sequential arguments to work with the topology where this is appropriate.
Proposition 3.1. Let τ be an admissible train-track. Then all points in ∂P (τ ) =
P (τ ) \ P (τ ) ⊆ XN /≈ have index strictly greater than i(τ ).
In particular P (τ ) is closed in Xi(τ ) .
Our proof of Proposition 3.1 is based on Lemma 3.2 and Proposition 3.4 below.
Lemma 3.2. Let τ be an admissible train-track, and let (Tn )n∈N be a sequence of mixing
Z-averse trees carried by τ . Assume that the trees Tn converge to a Z-averse tree T ,
and denote by ξ the ≈-class of T .
Then either T is mixing and A(τ ) is a maximal free factor system that is elliptic in T ,
or else i(ξ) > i(τ ).
Proof. The free factor system A(τ ) is elliptic in all trees Tn , and therefore it is also elliptic
in T . If A(τ ) is not a maximal free factor system elliptic in T , then as T collapses to
any mixing representative T ∈ ξ the free factor system A(τ ) is not maximally elliptic
in T . Remark 2.20 then implies that i(ξ) > i(τ ). If T is not mixing (but possibly A(τ )
is maximally elliptic in T ), then Corollary 1.9 implies that again A(τ ) is not maximally
elliptic in any mixing representative of ξ. Hence i(ξ) > i(τ ) in this case, also.
To complete the proof of Proposition 3.1, we are thus left understanding the case
where the limiting tree T is mixing, and A(τ ) is a maximal elliptic free factor system
in T : this will be done in Proposition 3.4 below. The idea is that the carrying maps
fn : S τ → Tn will always converge to an FN -equivariant map f : S τ → T ; in general
15
the limiting map f can fail to be a carrying map (it may even collapse some edges to
points), and in this case we will prove in Proposition 3.4 that there is a jump in index.
We start by proving the existence of the limiting map f .
Lemma 3.3. Let τ be an admissible train-track, and let (Tn )n∈N be a sequence of trees
in cv N with trivial arc stabilizers, converging to a tree T ∈ cv N . Assume that all trees
Tn are carried by τ , and for all n ∈ N, let fn : S τ → Tn be the carrying map.
Then the maps fn converge (in the equivariant Gromov–Hausdorff topology) to a map
f : Sτ → T .
Proof. Let v be a vertex in S τ . Since τ is admissible, there exist three inequivalent
edges e1 , e2 , e3 that form a legal tripod at v, and two elements g, h ∈ FN that both act
hyperbolically in S τ , whose axes are legal, and such that the axis of g (resp. h) crosses
the turn (e1 , e3 ) (resp. (e2 , e3 )). Then the axes of g and h in S τ intersect in a compact
non-degenerate segment with initial point v, and up to replacing g and h by their inverses,
we can assume that g and h both translate along this segment in the direction going out
of v. Since all maps fn are carrying maps, the elements g and h are hyperbolic in all
trees Tn , and their axes in Tn intersect in a compact nondegenerate segment, on which
they translate in the same direction. In the limiting tree T the intersection ATg ∩ ATh of
the axes (or fixed sets) of g and h is still a compact segment, and if it is non-degenerate
the elements g and h are hyperbolic in T and still translate in the same direction along
the intersection. We then define f (v) to be the initial point of ATg ∩ ATh . We repeat this
process over each orbit of vertices to obtain an FN -equivariant map from the vertices of
S τ to T , and extend the map linearly over edges. Distances between intersections of axes
(as well as their initial points) are determined by the Gromov–Hausdorff topology [23],
so it follows that for any two vertices v and v ′ , the distance dTn (fn (v), fn (v ′ )) converges
to dT (f (v), f (v ′ )) as n goes to infinity. This implies that the sequence of maps (fn )
converges to f .
Proposition 3.4. Let τ be an admissible train-track, and let (Tn )n∈N be a sequence of
mixing Z-averse trees carried by τ . Assume that (Tn ) converges to a mixing Z-averse
tree T , and that A(τ ) is a maximal free factor system elliptic in T . For all n ∈ N, let
fn : S τ → Tn be the carrying map, and let f : S τ → T be the limit of the maps fn .
Let S ′ be the tree obtained from S τ by collapsing all edges whose f -image is reduced to
a point, let f ′ : S ′ → T be the induced map, and let τ ′ := τf ′ be the train-track on S ′
induced by f ′ .
Then τ ′ is admissible. Either igeom (τ ′ ) > igeom (τ ), or else S ′ = S τ and τ ′ is obtained
from τ by a finite (possibly trivial) sequence of specializations (hence τ carries T ).
Before proving Proposition 3.4, we first complete the proof of Proposition 3.1 from
the above facts.
Proof of Proposition 3.1. Let ξ ∈ ∂P (τ ). If ξ is carried by τ , then i(ξ) ≥ i(τ ), and
this inequality is strict because otherwise τ would be an ideal carrier of ξ, contradicting
ξ ∈ ∂P (τ ).
We now assume that ξ is not carried by τ . Let (ξn )n∈N ∈ P (τ )N be a sequence
converging to ξ, and for all n ∈ N, let Tn be a mixing representative of the equivalence
class ξn . Since cv N is projectively compact, up to a subsequence we can assume that
(Tn )n∈N converges to a tree T . Since the boundary map ∂π : XN → ∂∞ F ZN is closed,
the tree T is Z-averse, and in the ≈-class ξ. Using Lemma 3.2, the proof reduces to the
case where T is mixing, and A(τ ) is a maximal free factor system elliptic in T . Since
16
ξ is not carried by τ , Proposition 3.4 then shows that T is carried by a train-track τ ′
satisfying igeom (τ ′ ) > igeom (τ ). This implies in turn that i(ξ) > i(τ ).
The rest of the section is devoted to the proof of Proposition 3.4.
Proof of Proposition 3.4.
1. The track τ ′ is admissible. Before proving that τ ′ is admissible, we first observe
that all hyperbolic elements in S τ are still hyperbolic in S ′ . Indeed, in view of Proposition 1.6 and of our assumption that A(τ ) is a maximal elliptic free factor system in
T , if g ∈ FN \ {e} is elliptic in T but is not contained in a free factor from A(τ ), then
the conjugacy class of g is given by (some power of) a boundary curve. Hence g is not
contained in any proper free factor of FN relative to A(τ ). Therefore the axis of g in
S τ crosses all orbits of edges, and so cannot be collapsed to a point by the collapse map
π : Sτ → S′.
We now prove that τ ′ is admissible. Let v ∈ V (S ′ ), and let X ⊆ S τ be the πpreimage of v. We first observe that X is a bounded subtree of S τ : indeed, otherwise,
we would find two oriented edges e, e′ in X in the same FN -orbit (say e′ = ge) and
pointing in the same direction; this would imply that g is hyperbolic in S τ but not in
S ′ , a contradiction. Let Y ⊆ X be a maximal legal subtree of X. Using the fact that
τ is admissible, we can find three pairwise inequivalent edges e1 , e2 , e3 lying outside of
X based at extremal vertices v1 , v2 , v3 of Y (these vertices are not necessarily distinct),
such that the subtree Y ∪ e1 ∪ e2 ∪ e3 is legal in τ . In particular, no edge ei is collapsed
to a point by π. Since τ is admissible, Lemma 2.9 gives three elements g1 , g2 , g3 , which
act hyperbolically in S τ , whose axes are legal, and such that the axis of gi in S τ crosses
the segment ei ∪ [vi , vi+1 ] ∪ ei+1 (mod 3). For all n ∈ N, the map fn preserves alignment
when restricted to each of these axes, so in the limit f also preserves alignment when
restricted to these axes. This implies that g1 , g2 and g3 are legal in S ′ . In addition, their
axes form a legal tripod at v, so τ ′ is admissible.
2. Controlling the index of τ ′ . We now prove that igeom (τ ′ ) > igeom (τ ) unless
τ carries T . Given a vertex v ∈ V (S ′ ), we will first establish that the contribution of the
equivalence class [v] to the index of τ ′ is no less than the sum of the contributions of its
π-preimages in V (S τ ) to the index of S τ .
Let x := f ′ (v). Let X be the set of vertices in S τ that are mapped to x under f . As
equivalent vertices in τ are mapped to the same point in T under f , the set X is a union
of equivalence classes of vertices in τ . As Stab(x) = Stab(X), if two vertices in X are in
the same FN -orbit, then they are in the same Stab(x)-orbit. Hence we may pick a finite
set E1 , . . . , Ek of representatives of the Stab(x)-orbits of equivalence classes of vertices
in X. Suppose that the images of these equivalence classes correspond to l orbits of
points in Tn (where l can be chosen to be independent of n by passing to an appropriate
subsequence). After reordering (and possibly passing to a further subsequence) we may
assume E1 , . . . , El are mapped to distinct Stab(x)-orbits of points in Tn for all n ∈ N,
and El+1 , . . . , Ek are exceptional classes in τ (notice that some of the classes Ei with
i ∈ {1, . . . , l} might be exceptional as well).
Suppose that each equivalence class Ei has αi gates in τ and has a stabilizer of
rank ri . As exceptional classes do not contribute to the geometric index, the amount
E1 , . . . , Ek contribute to the geometric index of τ is
l
X
i=1
αi + 3
l
X
ri − 3l.
i=1
17
(2)
Let E = [v] be the equivalence class corresponding to the image of X in τ ′ . Then E
contributes α′ + 3r ′ − 3 to the index of τ ′ , where α′ isPthe number of orbits
P of gates
at E and r ′ is the rank of Stab(x). If we define α = li=1 αi and r = li=1 ri , then
the difference between the index contribution of E in τ ′ and the index contribution of
E1 , . . . , Ek in τ is
α′ − α + 3(r ′ − r) + 3(l − 1).
(3)
Our goal is to control the number of Stab(x)-orbits of gates lost when passing from τ
to τ ′ . For all n ∈ N, we let Yn be the subtree of Tn spanned by the points in fn (X). Since
there are finitely many Stab(x)-orbits of equivalence classes of vertices in X, the tree Yn
is obtained from the Stab(x)-minimal invariant subtree of Tn by attaching finitely many
orbits of finite trees (if Stab(x) = {e}, then Yn is a finite subtree of Tn ). Recall that we
assumed that A(τ ) is a maximal elliptic free factor system in T . Using Proposition 1.6,
this implies that Stab(x) is either cyclic or contained in A(τ ) (it may be trivial). In the
second case A(τ ) is elliptic in Tn . Hence the quotient Gn = Yn /Stab(x) is a finite graph
of rank 0 or 1. There are l marked points in Gn corresponding to the images of E1 , . . . , El .
Claims:
1. If e is an oriented edge in S τ which is collapsed under π, and such that f ′ (π(e)) = x,
then the gate corresponding to e in τ is mapped under fn to a direction in Yn at
a point in fn (X).
2. If e and e′ are oriented edges in S τ that determine inequivalent directions d, d′ in τ ,
based at vertices in Stab(x).{E1 , . . . , El }, and if π(e) and π(e′ ) are nondegenerate
and equivalent in τ ′ , then for all sufficiently large n ∈ N, one of the directions d, d′
is mapped under fn to a direction in Yn at a point in fn (X).
Proof of claims: For the first claim, as e is collapsed its endpoints lie in X, so that
fn (e) is an arc between two points of fn (X) in Yn . For the second claim, if π(e) and
π(e′ ) are equivalent, then f (e) ∩ f (e′ ) is nondegenerate. Then for all sufficiently large
n ∈ N, the intersection of fn (e) with fn (e′ ) contains a nondegenerate segment. As e
and e′ are not equivalent in τ , they are based at distinct equivalence classes of vertices
in Stab(x).{E1 , . . . , El }. Let us call these [v1 ] and [v2 ]. We have fn (v1 ) 6= fn (v2 ), and
the arc connecting fn (v1 ) with fn (v2 ) is contained in the tree Yn , and is covered by the
union of the two arcs corresponding to fn (e) (starting at fn (v1 )) and fn (e′ ) (starting at
fn (v2 )). One can check that the initial direction of either fn (e) or fn (e′ ) must then be
contained in Yn .
Since T has trivial arc stabilizers, no two oriented edges in the same orbit are equivalent in τ ′ , and therefore there are only finitely many orbits of pairs of equivalent directions in τ ′ . Therefore, we can choose n ∈ N large enough so that the conclusion of
the second claim holds for all pairs of inequivalent directions in S τ based at vertices in
Stab(x).{E1 , . . . , El }, whose π-images are (nondegenerate and) equivalent in S ′ .
Denote by Gnout (resp. Gnin ) the set of gates based at vertices in Stab(x).{E1 , . . . , El }
which are mapped outside of Yn (resp. inside Yn ) by fn . Claim 1 implies that there
is a map Ψ from the set Enout of edges in Gnout , to the set of gates at E in τ ′ . Claim 2
then implies that any two edges in Enout in distinct gates have distinct Ψ-images, so
|Gnout /Stab(x)| ≤ α′ . Furthermore, our definition of E1 , . . . , El implies that the natural
map (induced by fn ) from Gnin /Stab(x) to the set Dir(Gn ) of directions based at the
18
marked points in the quotient graph Gn , is injective, so |Gnin /Stab(x)| ≤ |Dir(Gn )|. By
summing the above two inequalities, it follows that α − α′ is bounded above by the
number of directions based at the marked points in the quotient graph Gn .
We will now distinguish three cases. Notice that up to passing to a subsequence, we
can assume that one of them occurs.
Case 1: The stabilizer Stab(x) is either trivial, or fixes a point xn in all trees Tn ,
and for all n ∈ N, there is a vertex vn ∈ V (S τ ) such that fn (vn ) = xn (this happens in
particular if Stab(x) is contained in A(τ )).
In this case, we have r = r ′ , and the graph Gn has rank 0. The l marked points of Gn
include the leaves of Gn : indeed, every leaf of Gn is either the projection to Gn of a point
in fn (X), or else it is the projection to Gn of the unique point y ∈ Tn with stabilizer
equal to Stab(x). In the latter case, we again have y ∈ fn (X) by assumption. We will
apply the following fact to the graph Gn .
Fact: If G is a finite connected graph of rank r with l marked points containing all
leaves of G, then there are at most 2(l + r − 1) directions at these marked points.
Proof of the fact: An Euler characteristic argument shows that in any finite connected graph G with v vertices, there are exactly 2(v + r − 1) directions at the vertices.
Viewing the vertex set as a set of marked points, removing any non-leaf vertex from this
set loses at least two directions, from which the fact follows.
This fact applied to the quotient graph Gn shows that
α′ − α ≥ −2(l − 1).
(4)
From (3) we see that the new equivalence class contributes at least l − 1 more to the
geometric index than the sum of the contributions of the previous equivalence classes to
the index of τ . This shows that the index increases as soon as l ≥ 2, so we are left with
the case where l = 1.
In this remaining situation, the graph Gn is a single point, so all but possibly one
(say E1 ) of the equivalence classes E1 , . . . , Ek are exceptional. By Claim 1 no edges corresponding to directions at X are collapsed under π. Furthermore, by Claim 2 distinct
equivalence classes of gates based at vertices in the orbit of E1 are mapped to distinct
gates under π. It follows that either there are more gates at the equivalence class X when
passing from τ to τ ′ (so the index increases), or the gates corresponding to directions
at E1 are mapped bijectively and each exceptional class in E2 , . . . , Ek corresponds to a
specialization of τ (otherwise we would see extra gates). Repeating this argument across
all equivalence classes in τ ′ , we find that either the geometric index of τ ′ is greater than
that of τ , or τ ′ is obtained from τ by applying a finite number of specializations. In this
latter case, this implies that T is carried by τ .
Case 2: The group Stab(x) is not elliptic in Tn (in particular Stab(x) is cyclic).
In this case, we have r ′ = 1 and r = 0. The graph Gn is then a circle, with finitely many
finite trees attached, whose leaves are the projections of points in fn (X). In particular
Gn has rank r ′ − r = 1, and all its leaves are marked. By applying the above fact, we
thus get that
α′ − α ≥ −2(l + r ′ − r − 1).
(5)
19
Since r ′ − r = 1, we get from (3) that the new equivalence class contributes at least l
more to the geometric index than the sum of the contributions of the previous equivalence classes to the index of τ , so i(τ ′ ) > i(τ ).
Case 3: The stabilizer Stab(x) is cyclic, fixes a point xn in all trees Tn , but no vertex
in S τ is mapped to xn under fn .
In this case, we have r = 0 and r ′ = 1. The graph Gn has rank 0. The image of xn in
the quotient graph Gn might not be marked, however the set of marked points in Gn
includes all other leaves. We will use the following variation on the above fact (which is
easily proved by first including the missing leaf to the set of marked points).
Fact: Suppose G is a finite connected graph of rank r with a set of l marked points containing all leaves except possibly one. Then there are at most 2(l + r − 1) + 1 directions
at these marked points.
This fact, applied to the graph Gn , shows that
α′ − α ≥ −2(l − 1) − 1.
(6)
Since r ′ − r = 1, we get from (3) that the new equivalence class contributes at least
l + 1 more to the geometric index than the sum of the contributions of the previous
equivalence classes to the index of τ , and again we are done.
4
The cells P (τ ) have dimension at most 0.
The goal of this section is to prove the following fact.
Proposition 4.1. Let τ be an admissible train-track. Then P (τ ) has dimension at most
0.
The strategy of our proof of Proposition 4.1 is the following. Given any point ξ ∈
P (τ ), our goal is to construct arbitrarily small open neighborhoods of ξ with empty
boundary in P (τ ). This will be done using folding sequences of train-tracks. In Section 2,
we defined the notion of a specialization, which gives a new train-track τ ′ from a traintrack τ . In Section 4.1, we will introduce other operations called folding moves. These
will enable us to define a new train-track τ ′ from a train-track τ by folding at an illegal
turn. We will then make the following definition.
Definition 4.2 (Folding sequence of train-tracks). A folding sequence of traintracks is an infinite sequence (τi )i∈N of admissible train-tracks such that for all i ∈ N,
the train-track τi+1 is obtained from τi by applying a folding move followed by a finite
(possibly trivial) sequence of specializations.
Given ξ ∈ ∂∞ F ZN , we say that the folding sequence of train-tracks (τi )i∈N is directed
by ξ if ξ ∈ P (τi ) for all i ∈ N.
We will first prove in Section 4.1 that folding sequences of train-tracks exist.
Lemma 4.3. Let τ be an admissible train-track, and let ξ ∈ P (τ ).
Then there exists a folding sequence of train-tracks (τi )i∈N directed by ξ, such that τ0 is
obtained from τ by a finite (possibly trivial) sequence of specializations.
20
We will then prove in Section 4.2 that all sets P (τi ) in a folding sequence of traintracks are open in P (τ0 ), by showing the following two facts.
Lemma 4.4. Let τ be an admissible train-track, and let τ ′ be a specialization of τ . Then
τ ′ is admissible, and P (τ ′ ) is an open subset of P (τ ).
Lemma 4.5. Let τ be an admissible train-track, and let τ ′ be a train-track obtained from
τ by folding an illegal turn. Then τ ′ is admissible, and P (τ ′ ) is an open subset of P (τ ).
In other words, the sets P (τi ) are open neighborhoods of ξ in P (τ ), and they are
closed in P (τ ) (because their boundaries are made of trees with higher index in view
of Proposition 3.1). So to complete the proof of Proposition 4.1, we are left showing
that P (τi ) can be made arbitrary small. This is proved in Section 4.3 in the form of the
following proposition.
Lemma 4.6. Let τ be an admissible train-track, let ξ ∈ P (τ ), and let (τi )i∈N be a folding
sequence of train-tracks directed by ξ.
Then the diameter of P (τi ) converges to 0.
We now sum up the proof of Proposition 4.1 from the above four lemmas.
Proof of Proposition 4.1. Assume that P (τ ) 6= ∅, and let ξ ∈ P (τ ). Let ε > 0. Let
(τi )i∈N be a folding sequence of train-tracks directed by ξ provided by Lemma 4.3, where
τ0 is obtained from τ by a finite (possibly trivial) sequence of specializations. Lemma 4.6
shows that we can find k ∈ N such that P (τk ) contains ξ and has diameter at most
ε. An iterative application of Lemmas 4.4 and 4.5 then ensure that P (τk ) is an open
neighborhood of ξ in P (τ ). Furthermore, the boundary of P (τk ) in P (τ ) is empty (it is
made of trees of higher index by Proposition 3.1). As each point in P (τ ) has arbitrarily
small open neighborhoods with empty boundary, dim(P (τ )) = 0.
4.1
4.1.1
Existence of folding sequences of train-tracks directed by a boundary point
More on specializations
The notion of specialization was given in Definition 2.5.
Lemma 4.7. Let τ be an admissible train-track, and let τ ′ be a specialization of τ .
Then τ ′ is admissible, and i(τ ′ ) = i(τ ).
Proof. Admissibility is clear, as the underlying tree S τ and the collection of legal turns
at each vertex of S τ are unchanged after a specialization. To see that i(τ ′ ) = i(τ ),
notice first that the exceptional vertices in S τ contribute 0 to the geometric index of τ .
Stabilizers do not increase by definition of a specialization (otherwise two inequivalent
vertices in τ that are not in the orbit of v0 would become equivalent in τ ′ ), and no new
gate is created. So igeom (τ ′ ) = igeom (τ ), and since stabilizers of equivalence classes of
vertices are the same in τ and in τ ′ , we have i(τ ′ ) = i(τ ).
4.1.2
Folding moves
We now introduce three types of folding moves and discuss some of their properties.
21
τ
τ′
v1
v2
singular fold
e1
e2
v
Figure 2: A singular fold. The two black vertices v1 and v2 are equivalent in τ , and the
colors give the gates at the equivalence class of vertices they define.
Definition 4.8 (Singular fold, see Figure 2). Let τ be a train-track. Let e1 = [v, v1 ]
and e2 = [v, v2 ] be two edges in S τ that are based at a common vertex v, determine
equivalent directions at v, and such that v1 ∼τ v2 .
A train-track τ ′ is obtained from τ by a singular fold at {e1 , e2 } if
′
• there is an FN -equivariant map g : S τ → S τ that consists in equivariantly identifying e1 with e2 ,
′
• for all vertices v, v ′ ∈ V (S τ ), we have g(v ′ ) ∼τ g(v) if and only if v ′ ∼τ v,
• for all directions d, d′ in S τ based at equivalent vertices of S τ , we have g(d′ ) ∼τ
g(d) if and only if d′ ∼τ d.
′
Remark 4.9. Note that if d1 (resp. d2 ) is the direction in S τ at v1 (resp. v2 ) pointing
towards v, one can perform a singular fold at {e1 , e2 } if d1 and d2 (as well as v1 and v2 )
are equivalent: indeed, the first condition in the definition implies that g(d1 ) = g(d2 ),
and the last condition then implies that d1 ∼τ d2 .
Definition 4.10. Given an admissible train-track τ , we denote by P(τ ) the subspace of
∂cv N made of trees T with dense orbits such that τ is an ideal carrier of T .
Lemma 4.11. Let τ be an admissible train-track, and let e1 = [v, v1 ] and e2 = [v, v2 ] be
two edges that form an illegal turn in τ , such that v1 ∼τ v2 , and such that the directions
at v1 and v2 pointing towards v are τ -equivalent. Let τ ′ be the train-track obtained from
τ by applying a singular fold at {e1 , e2 }.
Then τ ′ is admissible, and P(τ ′ ) = P(τ ).
′
Proof. Let g : S τ → S τ be the folding map given in Definition 4.8. The numbers and
the stabilizers of equivalence classes of vertices and gates are unchanged under g, so that
′
i(τ ′ ) = i(τ ). To check admissibility of τ ′ , let v ′ be a vertex in S τ . Then there exists
22
τ
τ′
v1
v2
e1
e2
partial fold
v′
v
Figure 3: A partial fold.
v ∈ V (S τ ) such that v ′ = g(v), and the g-image of any tripod of legal axes at v is a
tripod of legal axes at v ′ (because g sends legal turns to legal turns).
We finally show that P(τ ) = P(τ ′ ). Let T ∈ P(τ ), with carrying map f : S τ → T .
Since the extremities of e1 and e2 are equivalent, these two edges are mapped to the
same segment in T . It follows that the carrying map f factors through the fold g to
′
attain a map f ′ : S τ → T , and one can check that the train-track induced by f ′ is
obtained from τ ′ by a finite (possibly trivial) sequence of specializations (using the same
sequence of specializations as when passing from τ to τf ). So T is carried by τ ′ , and
equality of the indices of τ and τ ′ shows that τ ′ is also an ideal carrier of T . It follows
that P(τ ) ⊆ P(τ ′ ). Conversely, if τ ′ is an ideal carrier of T with carrying map f ′ then
the composition f = f ′ ◦ g is such that τf is obtained from τ by a finite (possibly trivial)
sequence of specializations, and it follows that P(τ ′ ) ⊆ P(τ ).
Definition 4.12 (Partial fold, see Figure 3). Let τ be a train-track, and let {e1 , e2 }
be an illegal turn at a vertex v ∈ S τ . A train-track τ ′ is obtained from τ by a partial
fold at {e1 , e2 } if
′
• there is a map g : S τ → S τ that consists in equivariantly identifying a proper
initial segment [v, v1′ ] of e1 with a proper initial segment [v, v2′ ] of e2 , so that the
′
′
vertex v ′ = g(v1′ ) = g(v2′ ) is trivalent in S τ (all vertices of S τ that are not in the
FN -orbit of v ′ have a unique g-preimage in S τ ),
′
• for all v, w ∈ V (S τ ), we have g(v) ∼τV g(w) if and only if v ∼τV w,
′
• for all directions d, d′ based at vertices of S τ , we have g(d′ ) ∼τ g(d) if and only if
d′ ∼τ d,
• the vertex v ′ is not equivalent to any other vertex in τ ′ , and all directions at v ′ in
′
S τ are pairwise inequivalent.
Remark 4.13. The last condition in the definition implies that the new vertex v ′ is
exceptional in τ ′ .
23
Lemma 4.14. Let τ be an admissible train-track, and let τ ′ be a train-track obtained
from τ by a partial fold.
Then τ ′ is admissible, and i(τ ′ ) = i(τ ).
Proof. The condition from the definition of admissibility is easy to check at any vertex
′
in S τ not in the orbit of v ′ (with the notations from Definition 4.12). Let d0 , d1 , d2 be
the three directions at v ′ , containing the g-images of v = v0 , v1 , v2 , respectively. Let
i ∈ {0, 1, 2} (considered modulo 3). To construct a legal element whose axis crosses the
turn {di , di+1 }, we argue as in the proof of Lemma 2.9: one first finds two elements h, h′
that are legal in both τ and τ ′ such that
′
• the axes Ah , Ah′ in S τ go through g(vi ) and g(vi+1 ) respectively,
• the axes Ah and Ah′ do not go through v ′ ,
• the subtree Y spanned by Ah and Ah′ (which contains the turn {di , di+1 }) is legal.
The element hh′ is then a legal element that crosses the turn {di , di+1 }, as required.
To see that i(τ ′ ) = i(τ ), we first note that A(τ ′ ) = A(τ ) because the stabilizer of any
equivalence class corresponding to a new vertex is trivial, and the stabilizers of the other
equivalence classes of vertices are unchanged when passing from τ to τ ′ . In addition,
under the folding process the number of nonexceptional equivalence classes of vertices,
along with the ranks of their stabilizers and the number of associated gates, remains
the same, and the new exceptional trivalent vertex contributes 0 to the index. Hence
igeom (τ ′ ) = igeom (τ ), and therefore i(τ ′ ) = i(τ ) also.
Lemma 4.15. Let τ be an admissible train-track, and let T ∈ P(τ ) with carrying map
f : S τ → T . Assume that there exist two edges e1 , e2 in S τ that form an illegal turn,
such that |f (e1 ) ∩ f (e2 )| is smaller than both |f (e1 )| and |f (e2 )|. Let τ ′ be the train-track
obtained from τ by applying a partial fold at {e1 , e2 }.
Then T ∈ P(τ ′ ).
Proof. Recall that S τ is equipped with the metric induced by the map f : S τ → T , so
that any edge e of S τ has length |f (e)|. Let S ′ be the simplicial tree obtained from S τ
by equivariantly identifying a proper initial segment [v, v1′ ] of e1 of length |f (e1 ) ∩ f (e2 )|
with a proper initial segment [v, v2′ ] of e2 of the same length. Let g : S τ → S ′ be the
folding map. We first claim that the vertex v ′ := g(v1′ ) = g(v2′ ) is trivalent in S ′ . Indeed,
otherwise, the vertex v ′ would have infinite stabilizer, so A(τ ) would not be a maximal
′
free factor system elliptic in T , a contradiction. We thus have S ′ = S τ . The map f
′
factors through a map f ′ : S τ → T . Since i(τ ′ ) = i(τ ), it is enough to prove that the
train-track τf ′ is obtained from τ ′ by a finite sequence of specializations.
By definition, the train-track τf is obtained from τ by a finite sequence of specializations. Let (τf )′ be the partial fold of τf at {e1 , e2 }, and notice that (τf )′ is obtained
from τ ′ by the same sequence of specializations as when passing from τ to τf .
If the f ′ -image of the new trivalent vertex v ′ is not equal to the f -image of any other
vertex in S τ , then τf ′ = (τf )′ and we are done. Otherwise, we will show that τf ′ is a
specialization of (τf )′ , from which the lemma will follow. Indeed, if f ′ (v ′ ) is equal to the
f -image of another vertex v0 ∈ S τ , we first observe that we can always find such a v0
which is not in the orbit of v ′ . This is because if v ′ is only identified with vertices in
its own orbit, then the equivalence class of v ′ in τf ′ has nontrivial stabilizer in τf ′ , and
contributes positively to the geometric index of τf ′ . Then i(τf ′ ) > i(τ ), contradicting
24
τ
τ′
vd
v2
v2′
g(vd )
g(v1 ) = g(v2′ )
full fold
v1
v
Figure 4: A full fold. The new direction at g(v1 ) = g(v2′ ) has been identified with the
g-image of the direction d0 at v1 with initial edge [v1 , vd ] (in general, the direction d0
may also be based at a vertex v1′ 6= v1 as long as v1 and v1′ are τ -equivalent).
the fact that τ is an ideal carrier. So we can assume that v0 and v ′ are not in the same
orbit. Next, we observe that all germs of directions at v ′ are identified by f ′ with germs
of directions at vertices that are τ -equivalent to v0 , as otherwise we would be creating a
new gate when passing from τ to τf ′ , again increasing the index. So τf ′ is a specialization
of (τf )′ .
Definition 4.16 (Full fold, see Figure 4). Let τ be a train-track, and let {e1 , e2 } be
an illegal turn at a vertex v ∈ S τ . Denote by v1 , v2 the other extremities of e1 , e2 . Let
d0 be a direction in τ based at a vertex τ -equivalent to v1 . A train-track τ ′ is obtained
from τ by a full fold of e1 into e2 with special gate [d0 ] if
′
• there is an FN -equivariant map g : S τ → S τ that consists in equivariantly identifying [v, v1 ] with a proper initial segment [v, v2′ ] ( [v, v2 ],
′
• for all vertices v, v ′ ∈ V (S τ ), we have g(v ′ ) ∼τ g(v) if and only if v ′ ∼τ v,
′
• for all directions d, d′ based at vertices of S τ , we have g(d′ ) ∼τ g(d) if and only if
d′ ∼τ d,
′
• if d is the direction in S τ based at g(v1 ) = g(v2′ ) and pointing towards g(v2 ), then
d ∼ g(d0 ).
The number of equivalence classes of vertices remains unchanged under a full fold,
′
and the vertex stabilizers in S τ are the same as the vertex stabilizers in S τ . The fold
creates no new gates, so the index remains unchanged. If τ is admissible then τ ′ is also
admissible. Hence:
Lemma 4.17. Let τ be an admissible train-track, and let τ ′ be a train-track obtained
from τ by a full fold.
Then τ ′ is admissible and i(τ ′ ) = i(τ ).
25
Lemma 4.18. Let τ be an admissible train-track, and let T ∈ P(τ ), with carrying map
f : S τ → T . Let e1 = [v, v1 ] and e2 = [v, v2 ] be two edges in S τ that form an illegal turn.
Assume that f (e1 ) ( f (e2 ). Also assume that there is a direction d0 (whose initial edge
we denote by [v1′ , vd ]) at a vertex v1′ ∼τ v1 such that
dT (f (vd ), f (v2 )) < dT (f (vd ), f (v1′ )) + dT (f (v1′ ), f (v2 )).
(7)
Let τ ′ be the train-track obtained from τ by fully folding e1 into e2 , with special gate [d0 ].
Then T ∈ P(τ ′ ).
′
Proof. Since f (e1 ) ( f (e2 ), the map f factors through the fold g : S τ → S τ to reach
′
an FN -equivariant map f ′ : S τ → T , and Equation (7) ensures that f ′ identifies germs
of [g(v1 ), g(v2 )] and of g(d0 ). So τf ′ is obtained from τ ′ by the same finite sequence of
specializations as when passing from τ to τf . We know in addition that i(τ ′ ) = i(τ )
(Lemma 4.17). This implies that T ∈ P(τ ′ ).
Corollary 4.19. Let τ be an admissible train-track, and let T ∈ P(τ ), with carrying
map f : S τ → T . Let e1 = [v, v1 ] and e2 = [v, v2 ] be two edges in S τ that form an illegal
turn. Assume that f (e1 ) ( f (e2 ). Also assume that T is not carried by any specialization
of τ .
Then there exists a train-track τ ′ obtained from τ by fully folding e1 into e2 , such that
T ∈ P(τ ′ ).
Proof. Let S ′ be the underlying simplicial tree of any full fold of τ . Since f (e1 ) ( f (e2 ),
′
the map f factors through an FN -equivariant map f ′ : S τ → T . Equality in indices
shows the existence of a direction d0 in τ for which Equation (7) from Lemma 4.18 holds,
as otherwise the fold would create a new gate, and T would be carried by a train-track
with higher index. The direction d0 is based at a vertex v1′ satisfying f (v1′ ) = f (v1 ).
Since T is not carried by any specialization of τ , this implies that v1′ is τ -equivalent to
v1 . Therefore, Lemma 4.18 implies that T ∈ P(τ ′ ), where τ ′ is the full fold of τ with
special gate [d0 ].
4.1.3
Folding sequences: proof of Lemma 4.3
Given two train-tracks τ and τ ′ , we say that τ ′ is obtained from τ by applying a folding
move if τ ′ is obtained from τ by applying either a singular fold, a partial fold or a full
fold. We then define a folding sequence of train-tracks as in Definition 4.2: recall that
this is a sequence (τi )i∈N of train-tracks such that τi+1 is obtained from τi by applying a
folding move, followed by a finite (possibly trivial) sequence of specializations. The goal
of the present section is to prove Lemma 4.3. We will actually prove a slightly stronger
version of it, given in Lemma 4.22 below. The main fact we will use in the proof is the
following.
Lemma 4.20. Let τ be an admissible train-track, and let T ∈ P(τ ).
Then there exists a train-track τ ′ 6= τ obtained either by a specialization, or by applying
a folding move, such that T ∈ P(τ ′ ).
Proof. Assume that T is not carried by any train-track τ ′ 6= τ obtained from τ by
performing a specialization. Let f : S τ → T be the carrying map, and let e1 = [v, v1 ]
and e2 = [v, v2 ] be two edges that form an illegal turn in τ : this exists because T is not
simplicial by definition of P(τ ). If f (v1 ) = f (v2 ), then v1 ∼τ v2 , as otherwise T would
26
be carried by a specialization of τ . In this case f identifies germs of the directions at
v1 and v2 pointing towards v, so Lemma 4.11 implies that T ∈ P(τ ′ ), where τ ′ is the
singular fold of τ . If f (v1 ) 6= f (v2 ), and if |f (e1 ) ∩ f (e2 )| is smaller than both |f (e1 )| and
|f (e2 )|, then Lemma 4.15 implies that T is carried by the partial fold of τ at {e1 , e2 }.
Finally, if f (e1 ) ( f (e2 ) (or vice versa), then Corollary 4.19 implies that T is carried by
a full fold of τ .
We also make the following observation.
Lemma 4.21. Let τ be an admissible train-track, and let τ ′ be a train-track obtained
from τ by applying either a specialization or a folding move.
Then P(τ ′ ) ⊆ P(τ ) (and hence P (τ ′ ) ⊆ P (τ )).
Proof. If τ ′ is obtained from τ by applying a specialization, then the conclusion follows
from the definition of being carried, together with the fact that i(τ ′ ) = i(τ ) (Lemma 4.7).
If τ ′ is obtained from τ by applying a folding move, then the conclusion follows from the
fact that i(τ ′ ) = i(τ ) (Lemmas 4.11, 4.14 and 4.17), together with the observation that
′
′
if g : S τ → S τ is the fold map, and f : S τ → T is a carrying map, then f ◦ g : S τ → T
is also a carrying map.
We are now in position to prove the following stronger version of Lemma 4.3.
Lemma 4.22. Let ξ ∈ ∂∞ F ZN . Let τ0 , . . . , τk be a finite sequence of admissible traintracks such that ξ ∈ P (τk ), and for each i ∈ {0, . . . , k − 1}, the train-track τi+1 is
obtained from τi by applying a folding move followed by a finite (possibly trivial) sequence
of specializations.
Then there exists a folding sequence (τi′ )i∈N of train-tracks directed by ξ, such that τi = τi′
for all i ∈ {0, . . . , k − 1}, and τk′ is obtained from τk by a finite (possibly trivial) sequence
of specializations.
Proof. Lemma 4.21 shows that ξ ∈ P (τi ) for all i ∈ {1, . . . , k}. The conclusion is then
obtained by iteratively applying Lemma 4.20 to some mixing representative of the class
ξ (starting with the train-track τk ), and noticing that we can only perform finitely many
specializations in a row.
4.2
4.2.1
Openness of the set of points carried by a specialization or a fold
From trees to equivalence classes of trees
We first reduce the proofs of Lemmas 4.4 and 4.5 to their analogous versions for trees in
cv N .
Lemma 4.23. Let τ and τ ′ be admissible train-tracks.
If P(τ ′ ) is an open subset of P(τ ) in ∂cv N , then P (τ ′ ) is an open subset of P (τ ) in
XN /≈.
Proof. The conclusion is obvious if P(τ ′ ) is empty (in this case P (τ ′ ) is also empty), so
we assume otherwise. Since P(τ ′ ) ⊆ P(τ ), we have i(τ ′ ) = i(τ ). Let ξ ∈ P (τ ′ ), and let
(ξn )n∈N ∈ P (τ )N be a sequence converging to ξ. We wish to prove that ξn ∈ P (τ ′ ) for
all sufficiently large n ∈ N (we may use sequential arguments as XN /≈ is a separable
metric space).
For all n ∈ N, let Tn be a mixing representative of ξn . Since cv N is projectively
compact, up to passing to a subsequence, we can assume that (Tn )n∈N converges to a
27
tree T ∈ cv N . Closedness of the boundary map ∂π : XN → ∂∞ F ZN shows that T ∈ XN ,
and T is a representative of the class ξ. Since A(τ ) is elliptic in each tree Tn , it is
also elliptic in T . If T is not mixing, then Corollary 1.9 implies that A(τ ) is not the
largest free factor system elliptic in the mixing representatives of ξ, contradicting that
i(ξ) = i(τ ). So T is mixing, and the fact that ξ ∈ P (τ ′ ) thus implies that T ∈ P(τ ′ ).
Openness of P(τ ′ ) in P(τ ) then shows that Tn ∈ P(τ ′ ) for all sufficiently large n ∈ N.
Since Tn is a mixing representative of ξn , this precisely means that ξn ∈ P (τ ′ ) for all
sufficiently large n ∈ N, as desired.
4.2.2
Specializations
In this section, we will prove Lemma 4.4. More precisely, we will prove its analogous
version for trees in cv N , from which Lemma 4.4 follows thanks to Lemma 4.23.
Lemma 4.24. Let τ be an admissible train-track, and let τ ′ be a specialization of τ .
Then P(τ ′ ) is an open subset of P(τ ).
Proof. By Lemma 4.21, we have P(τ ′ ) ⊆ P(τ ). Let T ∈ P(τ ′ ), and let (Tn )n∈N ∈ P(τ )N
be a sequence of trees that converges to T . We aim to show that Tn is carried by τ ′ for all
sufficiently large n ∈ N, which will imply that Tn ∈ P(τ ′ ) by equality of the indices. Let
[v0 ] be the class of vertices in S τ at which the specialization occurs, and let v1 ∈ V (S τ )
be a vertex identified with v0 in τ ′ .
Let e10 , e20 , e30 be three edges adjacent to vertices in the class [v0 ], determining distinct
gates, and let e11 , e21 , e31 be edges based at vertices τ -equivalent to v1 that are identified
with e10 , e20 , e30 in τ ′ . We aim to show that for all sufficiently large n ∈ N, the vertices v0
and v1 are identified by the carrying map fn : S τ → Tn , and so are initial segments of
ei0 and ei1 for all i ∈ {1, 2, 3}.
For all i ∈ {1, 2, 3}, the intersection f (ei0 ) ∩ f (ei1 ) (where f : S τ → T is the carrying
map) is nondegenerate. Since the carrying map varies continuously with the carried tree
(Lemma 2.11), for all sufficiently large n ∈ N, the image fn (ei1 ) has nondegenerate intersection with fn (ei0 ). Since the three directions determined by e10 , e20 , e30 are inequivalent,
their images fn (e10 ), fn (e20 ), fn (e30 ) form a tripod at fn (v0 ). If fn (v1 ) 6= fn (v0 ), then at
least two of the images fn (e11 ), fn (e21 ), fn (e31 ) would then have a nondegenerate intersection (containing [fn (v0 ), fn (v1 )]), which is a contradiction because these three edges are
inequivalent. So fn (v1 ) = fn (v0 ), and since fn (ei1 ) has nondegenerate intersection with
fn (ei0 ), it follows that Tn is carried by τ ′ .
4.2.3
Folds
Lemma 4.5 follows from the following proposition together with Lemma 4.23.
Proposition 4.25. Let τ be an admissible train-track. Let τ ′ be a train-track obtained
from τ by folding an illegal turn.
Then P(τ ′ ) is open in P(τ ).
Proof. By Lemma 4.21, we have P(τ ′ ) ⊆ P(τ ). If τ ′ is obtained from τ by applying
a singular fold, then Lemma 4.11 ensures that P(τ ′ ) = P(τ ), so the conclusion holds.
Denoting by f : S τ → T the carrying map, we can thus assume that f (v1 ) 6= f (v2 )
(using the notations of the previous sections). Let p := f (e1 ) ∩ f (e2 ). The endpoints of
p, f (e1 ), and f (e2 ) are branch points in T (because τ is admissible), so the lengths of
these segments are determined by a finite set of translation lengths in T .
28
If τ ′ is obtained from τ by a partial fold, then |p| is strictly less than both |f (e1 )| and
|f (e2 )|. Using the fact that the carrying map f varies continuously with the carried tree
(Lemma 2.11), we see that this property remains true for all trees in a neighborhood of
T , and Lemma 4.15 implies that all trees in this neighborhood are carried by τ ′ .
Assume now that τ ′ is obtained from τ by a full fold, with e1 fully folded into e2 .
Let vd be the endpoint of an edge corresponding to a direction d based at a vertex v1′
equivalent to v1 , in the special gate. Then |f (e1 )| < |f (e2 )|, and dT (f (vd ), f (v2 )) <
dT (f (vd ), f (v1′ )) + dT (f (v1′ ), f (v2 )). These are open conditions in cv N . Therefore,
Lemma 4.19 gives the existence of an open neighborhood U of T in P(τ ) such that
all trees in U are carried by the same full fold.
4.3
Control on the diameter: getting finer and finer decompositions
The goal of the present section is to prove Lemma 4.6. The key lemma is the following.
Lemma 4.26. Let ξ ∈ ∂∞ F ZN , and let (τi )i∈N be a folding sequence of train-tracks
directed by ξ. Then (S τi ) converges to ξ for the topology on F ZN ∪ ∂∞ F ZN .
Proof. Let T be a mixing representative of ξ. We can find a sequence (Sn )n∈N of simplicial metric FN -trees with trivial edge stabilizers such that
• for all n ∈ N, the simplicial tree S τn is obtained from Sn by forgetting the metric,
• for all n ∈ N, the unique carrying map fn : Sn → T is isometric on edges.
In addition, for all i < j, there are natural morphisms fij : Si → Sj , such that fik =
fjk ◦ fij for all i < j < k. The sequence (Sn )n∈N converges to a tree S∞ : indeed, for
all g ∈ FN , the sequence (||g||Sn )n∈N is non-increasing, so it converges. In addition, the
tree S∞ is not reduced to a point, because the legal structure on S0 induced by the
morphisms f0j will stabilize as j goes to +∞, and a legal element (which exists because
τ0 is admissible, and every legal turn for the train-track τ0 is also legal for the train-track
structure on S0 induced by the morphisms f0j ) cannot become elliptic in the limit. By
taking a limit of the maps f0,j , we get an FN -equivariant map f0∞ : S0 → S∞ , which is
isometric when restricted to every edge of S0 .
We will show that S∞ is Z-averse and ≈-equivalent to T , which implies the convergence of S τi to ξ (for the topology on F ZN ∪ ∂∞ F ZN ) by the continuity statement for
the boundary map ∂π given in Theorem 1.3.
We first claim that the tree S∞ is not simplicial. Indeed, assume towards a contra′ be the simplicial tree obtained from S
diction that it is, and let S∞
∞ by adding the
f0∞ -images of all vertices of S0 to subdivide the edges of S∞ . Subdivide the edges of S0
′ . Since T has trivial arc stabilizers,
so that every edge of S0 is mapped to an edge of S∞
′ never identifies two edges in the same orbit, so there
the folding process from S0 to S∞
is a lower bound on the difference between the volume of Si+1 /FN and the volume of
Si /FN . Therefore, the folding process has to stop, contradicting the fact that the folding
sequence is infinite. This shows that S∞ is not simplicial.
We now assume towards a contradiction that S∞ is not equivalent to T . There
exists a 1-Lipschitz FN -equivariant map f from S∞ to the metric completion of T ,
obtained by taking a limit of the maps fnm as m goes to infinity, see the construction
from [15, Theorem 4.3]. Therefore, if S∞ has dense orbits, then there is a 1-Lipschitz
FN -equivariant alignment-preserving map from S∞ to T by [15, Proposition 5.7], which
implies that S∞ is Z-averse and equivalent to T by [14, Theorem 4.1]. Therefore, there
29
remains to prove that S∞ has dense orbits. Assume towards a contradiction that it does
not. Since S∞ is a limit of free and simplicial FN -trees that all admit 1-Lipschitz maps
onto it, it has trivial arc stabilizers. Since S∞ is not simplicial, there is a free factor A
acting with dense orbits on its minimal subtree SA in S∞ . By [25, Lemma 3.10], the
translates in T of f (SA ) (which cannot be reduced to a point) form a transverse family
in T : this is a contradiction because the stabilizer of a subtree in a transverse family
in a mixing Z-averse tree cannot be a free factor (in fact, it cannot be elliptic in any
Z-splitting of FN by [14, Proposition 4.23]).
Proof of Lemma 4.6. Let ξ ′ ∈ P (τi ), and let T ′ be a mixing representative of ξ ′ . By
Lemma 4.22, there exists a folding sequence of train-tracks (τj′ )j∈N directed by ξ ′ , such
that τj′ = τj for all j < i, and τi′ is obtained from τi by a finite (possibly trivial) sequence
′
of specializations. All trees S τi then lie on the image in F ZN of an optimal liberal folding
′
path, which is a quasi-geodesic by [20], and (S τi )i∈N converges to ξ ′ by Lemma 4.26. This
shows that any quasi-geodesic ray in F ZN from S τ0 to a point in P (τi ) passes within
bounded distance of S τi . As S τi converges to ξ ∈ ∂∞ F ZN (Lemma 4.26), it follows from
the definition of a visual distance that the diameter of P (τi ) converges to 0.
5
End of the proof of the main theorem
We now sum up the arguments from the previous sections to complete the proof of our
main theorem.
Proof of the main theorem. We start with the case of F ZN . The Gromov boundary
∂∞ F ZN is a separable metric space (equipped with a visual metric). It can be written
as the union of all strata Xi (made of boundary points of index i), where i ∈ N2 can
only take finitely many values. Each stratum Xi is the union of the sets P (τ ), where
τ varies over the collection of all train-tracks of index i, and the collection of all traintracks τ for which P (τ ) is nonempty is countable by Lemma 2.8. By Proposition 3.1,
each cell P (τ ) is closed in its stratum Xi , and by Proposition 4.1 it has dimension at
most 0. So each Xi is a countable union of closed 0-dimensional subsets, so Xi is 0dimensional by the countable union theorem [9, Lemma 1.5.2]. Since ∂∞ F ZN is a union
of finitely many subsets of dimension 0, the union theorem [9, Proposition 1.5.3] implies
that ∂∞ F ZN has finite topological dimension (bounded by the number of strata, minus
1). In particular, the topological dimension of ∂∞ F ZN is equal to its cohomological
dimension, see the discussion in [9, pp. 94-95]. This gives the desired bound because
the cohomological dimension of ∂∞ F ZN is bounded by 3N − 5 by [14, Corollary 7.3]
(this is proved by using the existence of a cell-like map from a subset of ∂CVN – whose
topological dimension is equal to 3N − 5 by [11], and applying [26]).
It was shown in Section 2.2.3 that the Gromov boundary ∂∞ F FN is equal to the
union of the strata Xi , with i comprised between (0, 0) and (0, 2N − 2). Therefore, the
above argument directly shows that the topological dimension of ∂∞ F FN is at most
2N − 2 (without appealing to cohomological dimension).
Finally, the Gromov boundary ∂∞ IN is equal to a union of strata Xi′ , with i ∈ Z+
comprised between 0 and 2N − 3, where each stratum Xi′ is a subspace of the stratum
X(0,i) from above. Given a train-track τ of index (0, i), we let P ′ (τ ) := P (τ ) ∩ Xi′ . Being
a subspace of P (τ ), the set P ′ (τ ) has dimension at most 0. In addition P ′ (τ ) is closed
in its stratum Xi′ , because its boundary in ∂∞ IN is made of points of higher index by
30
v
v
Step 1
v′
v
v
Step 2
Step 3
Figure 5: The following three steps may repeat infinitely often along a folding sequence,
showing that the homeomorphism type of the underlying simplicial tree might never
stabilize.
Proposition 3.1. The same argument as above thus shows that the topological dimension
of ∂∞ IN is at most 2N − 3.
A
Why equivalence classes of vertices and specializations?
In this appendix, we would like to illustrate the reason why we needed to introduce an
equivalence relation on the vertex set of S τ in the definition of a train-track (Definition 2.1), and allow for specializations in the definition of carrying (Definition 2.6).
Why equivalence classes of vertices in the definition of a train-track? In
the world of train-tracks on surfaces, when one defines a train-track splitting sequence
towards an arational foliation, the homeomorphism type of the complement of the traintrack eventually stabilizes, after which each singularity in the limiting foliation is determined by a complementary region of a track, and the visible prongs determine the index
of the singular point.
On the contrary, in a folding sequence of train-tracks towards an arational tree T as
defined in the present paper, the number of vertices in the preimage of a branch point in
T , as well as the number of directions at these vertices, may never stabilize, as illustrated
by the following situation (see Figure 5).
At some point on the folding sequence, one may have to perform a singular fold as
depicted on the right-hand side of Figure 5 (Step 3). If we had not declared the two
identified vertices equivalent before computing the index, such an operation would have
resulted in a drop in index. If such a situation could only occur finitely many times along
the folding sequence, we could have defined a ‘stable’ index along the folding sequence,
but this might not be true in general. Indeed, later on along the folding sequence, a new
trivalent vertex v can be created due to a partial fold (Step 1 on Figure 5), and this
exceptional vertex v may then be declared equivalent to another vertex v ′ in the track
by applying a specialization: this results in the possibility of performing a new singular
fold later on, identifying v and v ′ . One could then hope that eventually, all singular folds
in the folding sequence involve an exceptional vertex, which would not cause trouble as
far as index is concerned; but even this may fail to be true in general. Indeed, it might
31
happen that later on in the process, a full fold involving v creates a fourth direction at v
(this is the red direction in Step 2 of Figure 5), which is not equivalent to any direction
at v (but is instead equivalent to a direction at another vertex v ′ in the same class as v).
Introducing an equivalence class on the set of vertices in the definition of a train-track
ensures that the index of our train-tracks remains constant along a folding path, and
prevents overcounting the number of branch points in the limiting tree when counting
branch points in the train-track.
Why introducing specializations in the definition of carrying? We would finally
like to give a word of motivation for the necessity to introduce specializations in the
definition of carrying, instead of just saying that a train-track τ carries a tree T if
τ = τf for some FN -equivariant map f : S τ → T . As already observed, in a folding
sequence of train-tracks directed by T , partial folds create new trivalent vertices, and
it can happen that the newly created vertex v gets mapped to the same point in T as
another vertex v ′ ∈ S τ . We could have tried in this case to perform the partial fold and
the specialization at the same time, in other words declare that there are several distinct
partial folds of τ , including one (denoted by τ1 ) where v is not declared equivalent with
any other vertex, and a second (denoted by τ2 ) where v is declared equivalent to v ′ (and
there would be infinitely many other partial folds where v is identified with any possible
vertex from S τ ). With this approach, the difficulty comes when proving openness of the
set P(τ1 ) within P(τ ). Indeed, for a certain point T ∈ P(τ ), the carrying map may
′
′
not identify the new trivalent vertex v in S τ with any other vertex in S τ (hence T is
carried by τ1 ), while for nearby trees Tn in P(τ ) (for the Gromov–Hausdorff topology),
the carrying map identifies v with a vertex vn going further and further away in S τ as
n goes to infinity (hence no tree Tn is carried by τ1 ). So without the extra flexibility in
the definition of carrying, this would lead to P(τ1 ) not being open in P(τ ). This justifies
our definition of carrying. We have a single way of performing a partial fold of a turn,
and the specialization is performed later on along the folding sequence if needed.
References
[1] M. Bestvina and K. Bromberg, On the asymptotic dimension of the curve complex,
arXiv:1508.04832v1 (2015).
[2] M. Bestvina, K. Bromberg, and K. Fujiwara, Constructing group actions on quasitrees and applications to mapping class groups, Publ. math. IHÉS 122 (2015), no. 1,
1–64.
[3] M. Bestvina and M. Feighn, Outer limits, preprint (1994).
, Hyperbolicity of the complex of free factors, Adv. Maths. 256 (2014), 104–
[4]
155.
[5] M. Bestvina and P. Reynolds, The boundary of the complex of free factors, Duke
Math. J. 164 (2015), no. 11, 2213–2251.
[6] M.M. Cohen and M. Lustig, Very small group actions on R-trees and Dehn twist
automorphisms, Topology 34 (1995), no. 3, 575–617.
[7] M. Culler and J.W. Morgan, Group actions on R-trees, Proc. London Math. Soc.
55 (1987), no. 3, 571–604.
32
[8] S. Dowdall and S.J. Taylor, The co-surface graph and the geometry of hyperbolic
free group extensions, arXiv:1601.00101v1 (2016).
[9] R. Engelking, Dimension theory, North Holland Publishing Company, 1978.
[10] D. Gabai, Almost filling laminations and the connectivity of ending lamination
space, Geom. Topol. 13 (2009), no. 2, 1017–1041.
[11] D. Gaboriau and G. Levitt, The rank of actions on R-trees, Ann. Scient. Ec. Norm.
Sup. 28 (1995), no. 5, 549–570.
[12] U. Hamenstädt, The boundary of the free factor graph and the free splitting graph,
arXiv:1211.1630v5 (2014).
[13] C. Horbez, The boundary of the outer space of a free product, to appear in Israel J.
Math.
[14]
, Hyperbolic graphs for free products, and the Gromov boundary of the graph
of cyclic splittings, to appear in J. Topol.
[15]
, Spectral rigidity for primitive elements of FN , J. Group Theory 19 (2016),
no. 1, 55–123.
[16] I. Kapovich and M. Lustig, Geometric intersection number and analogues of the
curve complex for free groups, Geom. Topol. 13 (2009), no. 3, 1805–1833.
[17] E. Klarreich, The Boundary at Infinity of the Curve Complex and the Relative
Teichmüller Space, preprint (1999).
[18] C.J. Leininger and S. Schleimer, Connectivity of the space of ending laminations,
Duke Math. J. 150 (2009), no. 3, 533–575.
[19] G. Levitt, Graphs of actions on R-trees, Comment. Math. Helv. 69 (1994), no. 1,
28–38.
[20] B. Mann, Hyperbolicity of the cyclic splitting graph, Geom. Dedic. 173 (2014), no. 1,
271–280.
[21]
, Some hyperbolic Out(FN )-graphs and nonunique ergodicity of very small
FN -trees, Ph.D. thesis, University of Utah, 2014.
[22] H.A. Masur and Y.N. Minsky, Geometry of the complex of curves I : Hyperbolicity,
Invent. math. 138 (1999), no. 1, 103–149.
[23] F. Paulin, The Gromov topology on R-trees, Topology Appl. 32 (1989), no. 3, 197–
221.
[24] P. Reynolds, On indecomposable trees in the boundary of outer space, Geom. Dedic.
153 (2010), no. 1, 59–71.
[25]
, Reducing systems for very small trees, arXiv:1211.3378v1 (2012).
[26] L.R. Rubin and P.J. Schapiro, Cell-like maps onto non-compact spaces of finite
cohomological dimension, Topol. Appl. 27 (1987), 221–244.
33
Mladen Bestvina, Department of Mathematics, University of Utah, 155 South
1400 East, JWB 233, Salt Lake City, Utah 84112-0090, United States
e-mail:[email protected]
Camille Horbez, Laboratoire de Mathématiques d’Orsay, Univ. Paris-Sud, CNRS,
Université Paris-Saclay, 91405 Orsay, France.
e-mail:[email protected]
Richard D. Wade, Department of Mathematics, The University of British Columbia,
1984 Mathematics Road, Vancouver, B.C., Canada V6T 1Z2
email:[email protected]
34
| 4math.GR
|
Accepted in the IEEE Transactions on Cognitive and Developmental Systems (TCDS), 2017
Seamless Integration and Coordination of Cognitive Skills in Humanoid Robots: A Deep
Learning Approach
Jungsik Hwang and Jun Tani
Abstract
This study investigates how adequate coordination among the different cognitive processes of a humanoid robot can
be developed through end-to-end learning of direct perception of visuomotor stream. We propose a deep dynamic
neural network model built on a dynamic vision network, a motor generation network, and a higher-level network. The
proposed model was designed to process and to integrate direct perception of dynamic visuomotor patterns in a
hierarchical model characterized by different spatial and temporal constraints imposed on each level. We conducted
synthetic robotic experiments in which a robot learned to read human's intention through observing the gestures and
then to generate the corresponding goal-directed actions. Results verify that the proposed model is able to learn the
tutored skills and to generalize them to novel situations. The model showed synergic coordination of perception, action
and decision making, and it integrated and coordinated a set of cognitive skills including visual perception, intention
reading, attention switching, working memory, action preparation and execution in a seamless manner. Analysis
reveals that coherent internal representations emerged at each level of the hierarchy. Higher-level representation
reflecting actional intention developed by means of continuous integration of the lower-level visuo-proprioceptive
stream.
Index Terms—Neurorobotics, deep learning, sensorimotor learning
This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIP) [No. 2014R1A2A2A01005491].
Jungsik Hwang is with the school of Electrical Engineering in Korea Advanced Institute of Science and Technology, Daejeon, South Korea (e-mail:
[email protected]). Jun Tani is a corresponding author and he is a professor at the school of Electrical Engineering in Korea Advanced Institute of Science
and Technology, Daejeon, South Korea and an adjunct professor at Okinawa Institute of Science and Technology, Okinawa, Japan. (e-mail:
[email protected]).
I. INTRODUCTION
It would be desirable if a robot could learn to generate complex goal-directed behaviors from its own sensorimotor
experience, as human beings do. One challenge in reaching this goal is that such complex behaviors require an agent to
coordinate multiple cognitive processes. For instance, imagine a robot conducting an object manipulation task with a
human partner, such as reaching for and grasping an object. The human partner indicates a target objects located on the
workspace through a gesture. Then, the robot observes the workspace, finds the indicated object by combining the
perceived gesture information as well as the perceived object’s properties, switches its attention to the object, prepares
an action and executes it. Even this simple task is complex, involving diverse cognitive skills such as visual perception,
intention reading, working memory, decision-making, action preparation and execution. It is essential to link these
skills with synergy by developing spatio-temporal coordination among them. Furthermore, these skills ideally arise
from the robot’s experience (of reaching for and grasping objects, for example), rather than from hand-engineered
features reflecting a human engineer’s understanding of what any given task may require [1-9].
In this study, we employed a deep learning approach to build a robotic system which can directly and autonomously
learn from its own visuomotor experience. Deep learning is a fast-growing field in machine learning and artificial
intelligence with remarkable advances, such as text recognition, speech recognition, image recognition and many
others (See [10-12] for recent reviews of deep learning). One of the most important characteristics of deep learning is
that deep networks can autonomously extract task-related features in high-dimensional data, such as images and action
sequences, without the necessity of hand-engineered feature extraction methods [1, 10, 12]. So, deep learning provides
an important tool for robotics, because through deep learning a robot can learn directly from its huge-dimensional
sensorimotor data acquired through dynamic interaction with the environment [1]. A few recent studies [13-16] have
demonstrated the plausibility of deep learning in the field of robotics. However, several challenges in adapting deep
learning schemes to robotics remain. For example, a robotic system must process spatio-temporally dynamic patterns,
whereas deep learning schemes have generally been designed to process static patterns. In addition, robotic tasks
typically incorporate multiple sensory modalities, such as vision, proprioception and audition, while most deep
learning applications attend to single modalities, e.g. visual facial recognition. Also, Sigaud and Droniou [1] have
pointed out that it is still unclear how higher-level representations can be built by stacking several networks.
In this paper, we propose a dynamic deep neural network model called the Visuo-Motor Deep Dynamic Neural
Network (VMDNN) which can learn to generate goal-directed behaviors by coordinating multiple cognitive processes
including visual perception, intention reading1, attention switching, memorization and retrieval with working memory,
action preparation and generation. The model was designed to process and to integrate dynamic visuomotor patterns
directly perceived through the robot’s interaction with its environment. The VMDNN is composed of three different
types of subnetwork: the Multiple Spatio-Temporal scales Neural Network (MSTNN) [17], the Multiple Timescales
Recurrent Neural Network (MTRNN) [18] and the PFC (Prefrontal Cortex) subnetworks. The MSTNN has
demonstrated an ability to recognize dynamic visual scenes [17], and the MTRNN to learn compositional actions [18].
In the VMDNN model, these two subnetworks are tightly coupled via the PFC subnetwork, enabling the system to
process dynamic visuomotor patterns simultaneously. In other words, the problems of perceiving visual patterns and
generating motor patterns were regarded as inseparable problem and therefore the visual perception and the motor
generation were performed simultaneously in a single VMDNN model in the current study. This approach is based on
the previous studies which emphasized the importance of perception-action coupling in robotic manipulation [13] as
well as in developmental robotics [40]. Each part of the architecture is imposed with different spatial and temporal
constraints to enable the hierarchical computation of visuomotor processing in the lower-level and its abstraction in the
higher-level. We conducted a set of synthetic robotics experiments to examine the proposed model and also to gain
insight into mechanisms involved in learning goal-directed actions in biological systems. Here, it is worth noting that
artificial neural networks are meant to model the essential features of the nervous system, not its detailed
implementation [19-21].
In our experiments, a humanoid robot learned goal-directed actions from huge-dimensional visuomotor data
acquired from repeated tutoring during which the robot’s actions were guided by the experimenter. We investigated
how a set of cognitive skills can be integrated and coordinated in a seamless manner to generate goal-directed
sequential behaviors of the robot. Particularly, we focused on identifying the human gestures, reading intention
underlying them, and generating corresponding sequential behaviors of the humanoid robot in a reaching-and-grasping
task. In our experiment, the robot was trained to recognize human gestures and to grasp the target object indicated by
the gestures. This task thus required a set of cognitive skills such as visual perception, intention reading, working
memory, action preparation and execution. Recognizing gestures is more challenging than recognizing static images
1
Here, the term of “reading” or “categorizing” is used rather than “recognizing” because recognition may involve reconstruction of patterns but categorization in
the current study does not.
since both spatial and temporal information in the gesture need to be identified. Moreover, reading intention of others
by observing their behavior has been considered as one of the core abilities required for social cognition [5, 22-24]. In
addition, reaching and grasping are fundamental skills that have significant influences on the development of
perceptual and cognitive abilities [5]. The reaching-and-grasping task has been extensively studied in child
development [25, 26] as well as in robotics [19, 20, 27-29]. In a robotic context, they require robust perception and
action systems as well as simultaneous coordination of a set of cognitive skills, making hand-designing features
demanding and time-consuming [14]. Moreover, the visuomotor task in our experiment was not explicitly segmented
into the different sub-tasks such as gesture classification and action generation by the experimenter. Therefore, the task
requires the robot to adapt to different task phases autonomously by coordinating a set of cognitive skills in a seamless
manner throughout the task. In addition, this task requires the robot to have working memory capability to keep the
contextual information such that the robot could compare the categorized human intention with perceived object
properties. This implies that simply mapping perception to action cannot perform the task successfully since it lacks
contextual dynamics. We expected that the synergic coordination of the above mentioned cognitive skills would arise
through end-to-end learning of the tightly coupled structure in which the multiple subnetworks densely interact.
In the learning stage, the robot learned a task in the supervised end-to-end manner. In the testing stage, we examined
the model’s learning and generalization capabilities. Furthermore, the robot was examined under a visual occlusion
experimental paradigm in which the visual input to the model was unexpectedly and completely occluded. This was to
verify whether the proposed model was equipped with a sort of memory capability for maintaining task-related
information. In addition, we analyzed the neuronal activation in order to clarify the internal dynamics and
representations emerging in the model during different phases of operation.
The remaining part of the paper is organized as follows. In Section II, we review several previous studies employing
deep learning schemes in the robotic context. In Section III, we introduce the proposed model in detail. Section IV and
V are devoted to the experimental settings and the results respectively. Several key aspects and the implications of the
proposed model are discussed in Section VI. Finally, we conclude our paper and indicate current and future research
directions in Section VII.
II. RELATED WORKS
Due to the remarkable success of deep learning in various fields, recent studies have attempted to employ deep
learning in the field of robotics (See [1] for a recent review). For instance, Di Nuovo, et al. [15] employed a deep neural
network architecture in order to study number cognition in a humanoid robot. The robot was trained to classify numbers
by learning the association between an auditory signal and corresponding finger counting activity. They found that
their model was quicker and more accurate when both modalities were associated. Similarly, Droniou, et al. [2]
introduced a deep network architecture which could learn from different sensory modalities, including vision, audition
and proprioception. During experiments, their model was trained to classify handwritten digits, and they demonstrated
that learning across multiple modalities significantly improved classification performance. Yu and Lee [30] employed
a deep learning approach on reading human intention. A supervised MTRNN model was employed in their experiments,
and they showed that their model could successfully recognize human intention through observing a set of motions.
Lenz, et al. [14] proposed a two-stage cascaded detection system to detect robotic grasps in an RGB-D view of a scene
and conducted experiments on different robotic platforms. They found that their deep learning method performed
significantly better than one with well-designed hand-engineered features. Pinto and Gupta [31] also addressed the
problem of detecting robotic grasps, adopting a convolutional neural network (CNN) model to predict grasp location
and angle.
Although these studies demonstrate the utility of deep learning in a robotic context, they most focused on robotic
perception and robots were not directly controlled by deep learning schemes. A few recent studies have attempted to
utilize deep learning to control a robot. Wahlström, et al. [16], addressed the pixels-to-torques problem by introducing
a reinforcement learning algorithm that enabled their agent to learn control policy from pixel information. Deep
auto-encoders and a multi-layer feedforward neural network were employed in their model, and a 1-link robotic
pendulum was used as a testing platform. Their model was able to learn closed-loop policies in continuous state-action
spaces directly from pixel information. Noda, et al. [32] introduced a deep auto-encoder-based computational
framework designed to integrate sensorimotor data. A humanoid robot was used in their object manipulation
experiments, and they showed that their model was able to form higher-level multimodal representations by learning
sensorimotor information including joint angles, RGB images and audio data. Levine, et al. [13] proposed a deep neural
network model which learned a control policy that linked raw image percepts to motor torques of the robot. They
showed that the robot was able to conduct various object manipulation tasks by learning perception and control
together in an end-to-end manner. Park and Tani [33] investigated how a robot could infer the underlying intention of
human gestures and generate corresponding behaviors of a humanoid robot. In their work, an MTRNN model was
employed and the robot could successfully achieve the task by extracting the compositional semantic rules latent in the
various combinations of human gestures.
Several problems confront ongoing deep-learning research in robotics applications, as represented by limitations in
existing studies such as relatively simple testing platform [16, 33], separate processing of individual modalities [32],
and inability to handle temporal information [13]. In the current study, we aim to address these challenges with a deep
dynamic neural network model for a humanoid robot which can process and integrate spatio-temporal dynamic
visuomotor patterns in an end-to-end manner.
III. THE DEEP NEURAL NETWORK MODEL
In this section, we describe the Visuo-Motor Deep Dynamic Neural Network (VMDNN) in detail. The proposed
model was designed to process and to integrate direct perception of visuomotor patterns in a hierarchical structure
characterized by different spatio-temporal constraints imposed on each part of the hierarchy. It has several distinctive
characteristics. First, the model can perform low-level visuomotor processing without hand-engineered feature
extraction methods by means of deep learning schemes. Second, the model processes dynamic visuomotor patterns in a
hierarchical structure essential to cortical computation [15, 34]. Third, perception and action are tightly intertwined
within the system, enabling the model to form multimodal representations across sensory modalities. The VMDNN
model consists of three types of subnetworks: (1) MSTNN subnetworks for processing dynamic visual images, (2)
MTRNN subnetworks for controlling the robot’s action and attention and (3) a prefrontal cortex (PFC) subnetwork
located on top of these two subnetworks which dynamically integrates them (Fig. 1).
Fig. 1. The VMDNN model consists of three types of subnetworks: MSTNN subnetworks (VI, VF and VS) for dynamic visual
image processing (left), MTRNN subnetworks (MO, MF and MS) for action generation (right) and the PFC subnetwork for
integration of perception and action (top)
A. MSTNN Subnetwork
In our study, we employed Multiple Spatio-Temporal Scales Neural Network (MSTNN) to process dynamic visual
images perceived by a robot conducting a visuomotor task. The MSTNN is an extended Convolutional Neural
Network (CNN) [35] employing leaky integrator neural units with different time constants [17, 36]. Although
conventional CNN models have been shown the ability to process spatial data such as static images, they lack the
ability to process spatio-temporally dynamic patterns. To successfully conduct a visuomotor task, the robot needs to
extract both spatial and temporal features latent in the sequential observations. Unlike conventional CNN models that
utilize spatial constraints only, the MSTNN model has been shown that it can process both spatial and temporal patterns
by imposing multiple spatio-temporal scales constraints on local neural activity [17, 36]. Consequently, the MSTNN
can extract the task-related visual features latent in the dynamic visual images while the robot was tutored for the task
iteratively.
The MSTNN subnetwork was composed of three layers imposed with different spatio-temporal constraints: the
MSTNN-input (VI) layer containing a current visual scene, the MSTNN-fast (VF) layer with shorter-distance
connectivity and smaller time constants, and the MSTNN-slow (VS) layer with longer-distant connectivity and larger
time constants. Each MSTNN layer organizes into a specific set of feature maps retaining spatial information of visual
input. They are connected successively from VI to VS.
B. MTRNN Subnetwork
In the current study, we employed Multiple Timescales Recurrent Neural Network (MTRNN) for generating the
robot’s behavior and controlling the attention of the robot. The MTRNN is a hierarchical neural network model
consisting of a multiple continuous time recurrent neural networks with leaky integrator neurons [18]. MTRNN has
been shown superior performance in modeling robot’s sequential action by utilizing its temporal hierarchy. To be more
specific, the lower level in the MTRNN has a smaller time constant showing fast dynamics whereas the higher level has
a bigger time constant exhibiting slow dynamics. Due to this temporal hierarchy, an MTRNN can learn compositional
action sequences as a meaningful functional hierarchy emerges within the system [18, 34, 37]. Consequently, the entire
behavior of the robot including reaching and grasping as well as visual attention control can be decomposed into a set
of primitives for their flexible recombination adapting to various situations [18].
In our model, the MTRNN subnetwork is composed of three layers characterized by different temporal constraints:
the MTRNN-slow (MS) showing slow dynamics with the larger time constant, the MTRNN-fast (MF) showing fast
dynamics with smaller and the MTRNN-output (MO) layer with the smallest time constant. Neurons in the MS and MF
layers are asymmetrically connected to each other and to themselves. The MO layer is composed of groups of softmax
neurons indicating the sparse representation of the model’s output. The MO layer receives inputs from the MF layer and
generates behavior outputs as well as attention control signals.
C. PFC Subnetwork
At the top, the PFC (Prefrontal Cortex) layer tightly couples the MSTNN and MTRNN subnetworks for achieving
tight association between the visual perception and the motor generation. The PFC layer is a recurrent neural network
consisting of a set of leaky-integrator neurons equipped with recurrent loops in order to process abstract sequential
information. The PFC layer receives inputs from the VS layer in the MSTNN subnetwork as well as from the MS layer
in the MTRNN subnetwork, meaning that both abstracted visual information (V S) and proprioceptive information (MS)
are integrated in the PFC layer. The PFC layer also has a forward connection to the M S layer to control the robot’s
behavior and attention.
The PFC layer can be characterized by several key features. First, neurons in the PFC layer are assigned the largest
time constant. As a result, the PFC subnetwork exhibits the slowest-scale dynamics and this enables the PFC
subnetwork to carry more information about a situation [38]. Second, neurons in the PFC layer are equipped with
recurrent connections which are essential to handle dynamic sequential data [1, 9, 39]. Third, perception and action are
coupled via the PFC layer which integrates two monomodal subnetworks (MSTNN and MTRNN) and builds
higher-level multimodal representations from abstracted visuomotor information in each pathway. Lungarella and
Metta [40] argued that perception and action are not separated but tightly coupled, with this coupling is gradually
getting refined during developmental process.
D. Problem Formulation
Fig. 1 illustrates the structure of the VMDNN model. The input to the model (𝐼 𝑡 ) is an observation of the world at
time step t which can be obtained from the robot’s camera at the beginning to the end of the task. The observation is a
pixel-level image represented as a matrix H × W, where H is a height and W is a width of the image. Robot’s behavior
outputs as well as attention control signals are generated at the output layer MO at each time step t. Let 𝑦 𝑡 =
[𝑦1𝑡 , 𝑦2𝑡 , … , 𝑦𝑛𝑡 ] denotes the output of the model at the time step t where n is the number of neurons at the output layer
M O.
In forward dynamics computation, the problem is defined as to the behavior output and the attention signal (𝑦 𝑡 ) at
time step t given a visual observation 𝐼 𝑡 and the model parameters θ such as kernels, weights, and biases. In the training
phase, the problem is defined as to optimize the model’s parameters θ in order to minimize the error E at the output
layer MO represented by the Kullback-Leibler divergence between the teaching signal 𝑦̅ 𝑡 and the network’s output 𝑦 𝑡 .
The training data was visuomotor sequences obtained from the repeated tutoring prior to the training phase. The
detailed description about the forward dynamics and the training phase are described in the following sections.
E. Forward Dynamics for Action Generation
The internal states of all neural units were initialized with neutral values at the onset of the action generation mode.
Then, a pixel image 𝐼 𝑡 in grayscale obtained from the robot’s camera was given to the vision input layer (VI) and neural
unit’s internal states and activations were successively computed in every layer from the input layer (VI) to the output
layer (MO). The neuron activation at the MO layer was transformed to analog values using the softmax activation to
control the robot’s joints and attention. A detailed description of the computational procedure follows.
txy
At each time step t, the internal state ui
txy
and the dynamic activation vi
of the neural unit located at the (x, y)
position in the ith feature map in the MSTNN layers (i ∈ VF ˅ VS) are computed according to the following formulas:
𝑡𝑥𝑦
𝑢𝑖
= (1 −
1 (𝑡−1)𝑥𝑦 1
)𝑢
+ [ ∑ (𝑘𝑖𝑗 ∗ 𝑣𝑗𝑡 ) + 𝑏𝑖 ]
𝑥𝑦
𝜏𝑖 𝑖
𝜏𝑖
𝑗∈𝑉𝑗
𝑡𝑥𝑦
v𝑖
2 𝑡𝑥𝑦
= 1.7159 × tanh( 𝑢𝑖 )
3
(1)
(2)
τ is the time constant, Vj is the feature maps in the previous layer (if i ∈ VF, then Vj = VI and if i ∈ VS, then Vj = VF),
* is the convolution operator, kij is the kernel connecting jth feature map in Vj with the ith feature map in the current
layer, and b is the bias. Please note that the hyperbolic tangent recommended in [41] was used as an activation function
to enhance convergence.
From the PFC layer to the MO layer, the internal state uti and the dynamic activation yit of the ith neuron in the PFC
and MTRNN layers (i ∈ PFC ˅ MS ˅ MF ˅ MO) are determined by the following equations:
𝑢𝑖𝑡
= (1 −
1
)𝑢𝑡−1
𝜏𝑖 𝑖
+
1
[∑𝑗∈𝑉𝑆 𝑘𝑖𝑗 ∗ 𝑣𝑗𝑡 + ∑𝑘∈𝑀𝑆 ˅𝑃𝐹𝐶 𝑤𝑖𝑘 𝑦𝑘𝑡−1 + 𝑏𝑖 ]
𝜏𝑖
1
[∑𝑗∈𝑃𝐹𝐶 𝑤𝑖𝑗 𝑦𝑗𝑡 + ∑𝑘∈𝑀𝐹 ˅𝑀𝑆 𝑤𝑖𝑘 𝑦𝑘𝑡−1 + 𝑏𝑖 ]
𝜏𝑖
1
[∑𝑗∈𝑀𝑆 ˅𝑀𝐹 𝑤𝑖𝑗 𝑦𝑗𝑡−1 + 𝑏𝑖 ]
𝜏𝑖
1
[∑𝑗∈𝑀𝐹 𝑤𝑖𝑗 𝑦𝑗𝑡 + 𝑏𝑖 ]
𝜏
{
y𝑖𝑡
=
𝑖
2
1.7159 × tanh ( 𝑢𝑖𝑡 )
3
exp(𝑢𝑖𝑡 )
𝑡
{∑𝑗∈𝑀𝑂 exp(𝑢𝑗 )
𝑖𝑓 𝑖 ∈ 𝑃𝐹𝐶
𝑖𝑓 𝑖 ∈ 𝑀𝑆
𝑖𝑓 𝑖 ∈ 𝑀𝐹
(3)
𝑖𝑓 𝑖 ∈ 𝑀𝑂
𝑖𝑓 𝑖 ∈ 𝑃𝐹𝐶˅𝑀𝑆 ˅𝑀𝐹
𝑖𝑓 𝑖 ∈ 𝑀𝑂
(4)
wij are the weight from the jth neural unit to the ith neural unit.
In sum, the image obtained from the robot’s camera was sent to the input layer (VI) of the model at each time step of
the action generation mode. Then, the internal states and activations were successively computed from the input layer
(VI) to the output layer (MO). The robot was operated based on the output of the model including the joint position
values. After the execution of an action, the image acquired from the robot’s camera was sent to the input layer (VI) of
the model again in the next time step. In this sense, an image given to the model can be considered as a visual feedback
since it reflects the effect of the robot’s action in the preceding time step.
F. Training Phase
The model was trained in a supervised end-to-end learning fashion. The training data consisted of raw visuomotor
sequences obtained from repeated tutoring during which the robot was manually operated by the experimenter. From
the beginning to the end of tutoring, a visual image perceived from the robot’s camera (visual observation) was jointly
collected with the encoder values of the robot’s joint positions as well as the level of grasping and the level of foveation
at each time step. The model was trained to abstract and associate visual perception with proprioceptive information
using these visuomotor sequence patterns. Backpropagation through time (BPTT) [42] was employed in learning
values of parameters, such as the kernels, weights and biases of the model.
Prior to end-to-end training, learnable parameters in the visual pathway (from VI to PFC) were initialized by means
of pre-training. Studies have demonstrated that pre-training is an efficient method for initializing network parameters
[10, 13, 43]. The pre-training method in our study is similar to that of [13] in which the visual part of the model was
pre-trained prior to the end-to-end learning phase of the experiment. During pre-training in our study, the softmax
output layer was connected to the PFC layer and the connections between the MTRNN subnetworks and the PFC layer,
as well as the recurrent connections within the PFC layer were removed. In this condition, the system operates as an
MSTNN model, and it was trained as a typical classifier using BPTT as described in [36]. Then, the values of the
parameters in the visual pathway acquired during pre-training were used as initial values for the parameters in the same
pathways during the end-to-end training phases of the experiment.
After the pre-training, end-to-end training was conducted from the VI layer to the MO layer. During end-to-end
training, the model’s entire learnable parameters (k, w and b) were updated to minimize the error E at the MO layer
represented by Kullback-Leibler divergence between the teaching signal y̅it and the network’s output yit .
𝐸 = ∑ ∑ 𝑦̅𝑖𝑡 𝑙𝑜𝑔
𝑡 𝑖∈𝑀𝑂
𝑦̅𝑖𝑡
𝑦𝑖𝑡
(5)
A stochastic gradient descent method was applied during end-to-end training and the entire learnable parameters
were updated when each training data was presented as follows.
𝜕𝐸
𝑤𝑖𝑗 (𝑛 + 1) = 𝑤𝑖𝑗 (𝑛) − 𝜂 (
+ 0.0005𝑤𝑖𝑗 (𝑛))
𝜕𝑤𝑖𝑗
(6)
𝜕𝐸
𝑘𝑖𝑗 (𝑛 + 1) = 𝑘𝑖𝑗 (𝑛) − 𝜂 (
+ 0.0005𝑘𝑖𝑗 (𝑛))
𝜕𝑘𝒊𝒋
(7)
𝜕𝐸
b𝑖 (𝑛 + 1) = b𝑖 (𝑛) − 𝜂 ( )
𝜕𝑏𝑖
(8)
n is an index of the learning step and η is the learning rate. The weight decay method was used to prevent overfitting
[35] with the weight decay rate of 0.0005.
IV. EXPERIMENT SETTINGS
A. Robotic Platform
iCub [44] is a child-like humanoid robot consisting of 53 degrees of freedom (DOFs) distributed in the body. We
used a simulation of the iCub [45] in our experiments. The iCub simulator accurately models the actual robot’s physical
interaction with the environment, making it an adequate research platform for studying developmental robotics [15, 45].
Simulation of the iCub is shown in (Fig. 2). In our simulation environment, a screen was located in front of the robot to
display human gestures, the task table was placed between the screen and the robot and the two objects were placed on
the task table.
At each time step, the interfacing program captured the image perceived from the robot’s camera, preprocessed the
captured image, sent it to the VMDNN model, received the output of the VMDNN model and then operated the robot
based on the output of the model. To be more specific, the pixel image obtained from the robot’s camera embedded in
its left eye was passed to the vision input layer (VI) of the model at each time step. The obtained image was
preprocessed by resizing to 64 (w) × 48 (h), converting to grayscale, and normalizing to -1 to 1. The softmax values of
the output layer (MO) were converted into the values of the corresponding joint positions (7 DOFs in the right arm, 2
DOFs in the neck) as well as into grasping and attention control signals. Then, the interfacing program operated the
robot based on the VMDNN model’s outputs using the motor controller provided in the iCub software package [45].
Then, the interfacing program captured an image from the robot’s camera and sent it back to the VMDNN model in the
next time step.
Fig. 2. The experimental settings using an iCub simulator. (a) Four types of human gestures (b) Two objects consisting of a tall and
a long object were placed in the task space, and the robot grasped the target object indicated by a human gesture (c) Object
locations on a task space.
Regarding the output for the robot’s behavior, we used the robot’s right arm consisting of 7 DOFs (shoulder’s pitch,
roll, yaw, elbow, wrist’s pronosupination, pitch, and yaw). In addition, the network output the level of extension or
flexion of finger joints in order to control grasping similar to [20]. The level varied from 1 (not grasping) to 10 (fully
grasping) and it was used to control the 9 DOFs of the right hand. In our experiments, the VMDNN model also
controlled visual attention which is essential to generate adequate robot’s behavior [37]. Jeong, et al. [37] demonstrated
that MTRNN could seamlessly coordinate visual attention and motor behaviors. In their work, the MTRNN model
outputted the category of the object to be attended to, and the external visual guiding system localized the position of
the specified object based on the network’s output and the retina image so that the robot’s camera could fixate the
object. In contrast, no external module was employed in our study, and the VMDNN model itself controlled robot’s
visual attention. More specifically, two attention control mechanisms were employed in our study: attention shift and
foveation. First, the robot located the object of attention at the center of the visual scene by orienting its head (2 DOFs
– neck’s pitch and yaw). Second, the robot controlled the resolution of the visual scene given to the network by
controlling the level of foveation from 1 (minimum foveation) to 10 (maximum foveation). For instance, the level of
foveation increased while the robot’s hand was closely approaching to the attended object. Consequently, images
containing the target object and the robot’s hand were given to the network with higher resolution so that the model
could more clearly perceive the object’s properties including orientation and the location of the hand.
B. Network Configuration
The VMDNN model was composed of 7 layers: the VI, VF, VS layers in the MSTNN subnetwork, the PFC layer, and
the MS, MF, MO layers in the MTRNN subnetwork. The structure of the VMDNN model used in this study was found
empirically in our preliminary experiments [46]. Note that the structure of the VMDNN model including the number of
layers in each subnetwork can be extended depending on the complexity of the task since the ‘deeper’ structure can
enhance learning of complex functions in visuomotor patterns [11]. For instance, more number of layers in the MSTNN
subnetwork can be employed to process more complex visual images [17]. Similarly, more complex robot’s behavior
can be learned by employing more number of layers in the MTRNN subnetwork as reported in [33].
Each MSTNN layer consisted of a set of feature maps retaining the spatial information of the visual input. The
number and size of feature maps varied between layers. The vision input layer (VI) had a single feature map containing
the current visual scene (64 (w) × 48 (h)) obtained from the left eye of the robot. The VF layer consisted of 4 feature
maps with sized 15 × 11 and the VS layer consisted of 8 feature maps with sized 5 × 3. The size of kernels in the VF and
VS layers were set to 8 × 8 and 7 × 7 respectively. The sampling factors denoting the amount of shift of the kernel in a
convolution operation were set to 4 and 2 for the VF and VS respectively. The PFC layer was composed of 20 neurons
and the kernel size was 5 × 3 with the sampling factor set to 1. The numbers of neurons employed in the MTRNN layers
(MS, MF and MO) were 30, 50 and 110 respectively. The 110 neurons in the MO layer were comprised of 11 groups of
softmax neurons representing the 11 categories of the model’s output: the joint position values of the robot’s neck (2),
the joint position values of the right arm (7), the level of grasping (1) and the level of foveation (1). The softmax output
values in each group at the MO layer were inversely transformed to analog values that directly set the joint angles of the
robots, the level of grasping and the level of foveation.
Regarding the time scale properties, we compared two different types of visual pathway: CNN and MSTNN. In the
CNN condition, the time constants of both the VF and the VS layers were set to 1, resulting in no temporal hierarchy. On
the other hand, in the MSTNN condition, the time constants of the VF and the VS layers were set to 1 and 15
respectively, resulting in a temporal hierarchy in the visual pathway. We also compared two different temporal scales
(fast and slow) in the PFC layer. The time constant was set to 1 in the fast PFC condition whereas it was set to 150 in the
slow PFC condition. In sum, there were 4 different network conditions examined in our experiments. Throughout the
experiments, the time constants of the MS, MF and MO layers were fixed to 70, 2 and 1 respectively. The proper values
for the time constant at each level of the model were found heuristically in our preliminary study [46].
Prior to end-to-end learning, learnable parameters were initialized with the values acquired from the pre-training
stage to enhance learning capability. During the pre-training stage, the model was trained to grasp an object without
human gestures. Also, the visual pathway was additionally pre-trained to classify the four types of human gestures as
described in Section III. During the end-to-end learning, the network was trained for 13,000 epochs with a learning rate
of 0.01.
C. Task
The objective of the task was to grasp the target object indicated by the human gesture displayed on the screen at the
beginning of the task. The overall task flow was as follows. At the beginning of the task, the robot was set to the home
position, orienting its head toward the screen in front of the robot and stretching both arms sideways. From the home
position, the robot observed a human gesture being displayed on the screen (Fig. 2). After observing the gesture, the
robot oriented its head to the task table on which two objects consisting of a long and a tall object were placed. While
observing the task space, the robot figured out the target object indicated by the human gesture and oriented its head to
the target object. Then, the robot reached out its right arm and grasped the target object. For example, when the human
gesture indicated the tall object, the robot had to pick out the tall object from among the two available along with
information about its orientation. Similarly, when the human gesture indicated the right side, the robot had to figure out
the type and orientation of the object on the right side (see the supplementary video). This task, therefore, inherently
required the robot to have working memory capability to maintain the human gesture information throughout the task
phases and dynamically combine it with the perceived object properties.
The visuomotor sequence for each training trial consisted of images perceived from the robot’s camera (visual
observation), the values of robot’s joint positions and the values of the grasping and foveation signals and they were
collected simultaneously from the beginning to the end of tutoring. Consequently, the images in the visuomotor
sequences included the ones that were perceived while observing the human gestures on the screen as well as the ones
perceived while completing the behavioral task. Similarly, the values of the robot’s joint positions in the visuomotor
sequences included the ones that were recorded while observing the human gestures as well as the ones recorded while
acting in the task space after the observation. The robot was trained with 200 trials consisting of varying object
configurations and human gestures. Two objects consisting of one tall and one long object were presented in each trial
and the location and the orientation of the two objects were controlled so that the way of presenting the two objects did
not bias the robot’s behavior toward certain reaching and grasping behaviors and away from others. Regarding human
gestures, a target object was specified by one of four different human gestures indicating the location (either left or
right) or the type (either tall or long) of the target object (Fig. 2 (a)). A 40-frame video clip of the human gesture was
displayed on the screen in front of the robot. We collected several gesture trials from 7 human subjects and
pseudo-randomly selected from amongst them so that each type of gesture appeared the same number of times in the
training dataset. Throughout the experiments, two different types of box-shape objects were used: a tall object with a
size of 2.8cm × 5cm × 10cm and a long object with a size of 2.8cm × 10cm × 5cm. Objects were placed with 5 different
orientations (-45°, -22.5°, 0°, 22.5°, 45°) at 10 positions symmetrically distributed on the XY-plane of the task space
(Fig. 2 (c)).
During the testing stage, we evaluated the model’s performance with respect to the learned trials (TR) as well as with
a set of novel situations in order to examine the model’s generalization capability. First, we tested the model with 60
trials in which two objects were randomly located (OBJ) to examine whether the robot was able to generalize reaching
and grasping skills to unlearned object positions and orientations. We also examined the model’s generalization
capability with respect to human gestures, using 200 training trials with the gestures of a novel subject (SUB). Then, we
examined the model’s generalization capability with 60 cases of the most novel situation, in which randomly located
objects were indicated by gestures of the novel subject (OBJ × SUB). It should be noted that main focus was on the
model’s ability to coordinate cognitive skills such as gesture recognition and working memory rather than solely on
gesture recognition, since the latter is more concerned with typical perceptual classification tasks. In addition, we
evaluated the model under a visual occlusion experimental paradigm in which the vision input to the network was
completely occluded. The main focus of this evaluation was to verify whether the network was equipped with a sort of
internal memory, enabling it to show robust behavior even the visual information was unexpectedly and completely
occluded. From the training trials (TR) and the testing trials (OBJ × SUB), we selected 80 and 60 trials respectively.
During the visual occlusion experiment, vision input to the network was occluded at the onset of observing the task
space (t = 51), at the onset of attending to the target object (t = 55), at the onset of observing the target object (t = 61), at
the onset of reaching (t = 66), and during reaching (t = 86). Prior to the testing, the learnable parameters were initialized
to the ones obtained from the training.
V. RESULTS
A. Generalization Performances
Table 1 shows the success rate of each network condition. Each trial was evaluated as “successful” if the robot
grasped and lifted an object and “failure” otherwise. In general, the MSTNN vision with the slow PFC condition
showed the better performance than did other conditions. In this condition, the model successfully learned the training
trials (TR = 98.5%) and was able to generalize learned skills to different testing conditions. This model was able to
generalize reaching and grasping skills when the objects were randomly located (OBJ = 85.0%), when the gestures of
the novel subject were displayed (SUB = 96.5%), and when the randomly located objects were specified by the novel
subject (OBJ × SUB = 83.3%). It is worth noting that the model demonstrated relatively low success rates when in the
fast PFC conditions, especially CNN vision with the fast PFC condition. The task required the robot to maintain human
gesture information displayed at the beginning of the trial throughout all task phases. In the end, the model equipped
with the temporal hierarchy performed significantly better than did those without it.
TABLE 1
THE SUCCESS RATE OF FOUR NETWORK CONDITIONS
Network Conditions
Type of
PFC
Vision
Scales
Layer
Testing Conditions
TR
OBJ
SUB
OBJ × SUB
CNN
Fast
47.0%
41.7%
44.0%
41.7%
CNN
Slow
87.5%
85.0%
87.5%
80.0%
MSTNN
Fast
92.0%
60.0%
86.0%
56.7%
MSTNN
Slow
98.5%
85.0%
96.5%
83.3%
TABLE 2
THE PERCENTAGE OF TASK FAILURE CAUSED BY CONFUSION
Network Conditions
Type of
PFC Scales
Vision Layer
Testing Conditions
TR
OBJ × SUB
CNN
Fast
59.4%
71.4%
CNN
Slow
44.0%
41.7%
MSTNN
Fast
68.8%
65.4%
MSTNN
Slow
0.0%
20.0%
We analyzed failure cases in training trials (TR) and in the most novel trials (OBJ × SUB). Particularly, we focused
on task failure caused by apparent confusion between gesture-indicated and other objects. In these cases, the robot
simply grasped the incorrect object. Table 2 shows the percentage of the error caused by confusion in the four network
conditions. The model showed the least confusion errors in the MSTNN with slow PFC condition. Moreover, the error
caused by confusion was more pronounced in fast PFC than in slow PFC conditions. This result implies that the model
was able to maintain correct and stable representations by means of slow dynamics in the PFC layer.
Fig. 3. The success rate of the four different network conditions in the visual occlusion experiment: (a) testing on learned human
gestures and object positions and orientations (TR) and (b) testing on unlearned human gestures and object positions and
orientations (SUB × POS).
Fig. 3 illustrates the success rate for each network condition with respect to (a) training trials (TR) and (b) testing
trials (OBJ × SUB) with different occlusion timings. When vision input to the model was occluded, performance of the
four different network conditions differed significantly. As expected, the model’s performance in all conditions
generally degraded when vision input was occluded during earlier task phases. Especially, the model showed relatively
worse performance in the CNN vision with the fast PFC condition than other conditions. Therefore, it can be inferred
that memory capability achieved by the slow dynamics subnetwork at the higher level plays an important role when
vision input is occluded. That is, by means of memory, the robot was able to maintain information about human
gestures as well as information about the target objects including position, type and orientation throughout the all
phases of task execution. This result shows the importance of the internal contextual dynamics of the proposed model
and highlights a difference between the proposed model and the previous study [13] which was prone to occlusion due
to the lack of capability of keeping memory.
B. Development of Internal Representation
In order to reveal the model’s self-organized internal representation, we analyzed neural activation during training
trials using the t-Distributed Stochastic Neighbor Embedding (t-SNE) dimensionality reduction algorithm [47]. During
t-SNE analysis, the parameters for the initial dimensions and the perplexity were set to 10 and 30 respectively. Fig. 4
depicts the internal representation emerging at each layer during the three different task phases. Each point indicates a
single training trial and distances between those points represent relative similarity between trials. The color and the
shape of each point denote the type of human gesture and the type of the object respectively. The number next to each
point indicates the object’s position.
Fig. 4. Internal representations emerging during three different task phases at each layer in the MSTNN with the slow PFC. Each
point represents a single trial and the distance between points represents the similarity between trials. The shape and the color of
each point indicate the type of the target object and the type of human gesture, respectively. The number next to each point signifies
the object’s position. In some cases, numbers are omitted due to overlap. Please note that the mode of grasping is different
depending on the type of object. We focused on the relationships between patterns, so we did not plot the axes varying between
plots.
In MSTNN layers (VF and VS), sequential visual images were abstracted in both temporal and spatial dimensions
through hierarchical processing. For instance, where the robot begins observation of the task space, five clusters can be
observed in the VF layer (b). These clusters correspond to the five possible pair locations employed during training.
Objects appearing in the same position appeared in the same cluster regardless of the type of object and type of human
gesture. In the VS layer, those representations were further separated, with the type of target object differentiated within
the clusters reflecting the relative locations of the two objects (e). Similarly, in the mid-reaching phase, the VF layer (c)
and the VS layer (f) encoded both the type and the location of the target object, but the distinction with respect to the
object’s type was less clear in the VF layer.
Transitions of representations from those reflecting types of human gesture to those reflecting specific target objects
organized in the PFC layer. After observing the gesture (g), four clusters reflecting the type of gesture appeared,
suggesting that human gestures were successfully recognized. Then, those internal representations started to develop
progressively, and smaller clusters indicating specific target objects emerged in the mid-reaching phase regardless of
the presented human gesture (i). Our interpretation is that the higher-level of the model read the human’s intention and
translated it into robot’s own intention to reach and to grasp the specified object. To be more specific, lower-level
visual images containing human gestures were processed hierarchically by MSTNN subnetworks and the abstracted
representation underlying one of the four human’s intentions appeared in the higher-level (the PFC layer). Then, the
robot simultaneously incorporated the perceived object’s properties including location, type and orientation, and it
formed its own intention for reaching and grasping the target object. In the end, this result suggests that the robot’s
intention was not explicitly mapped but arose dynamically from perceived information.
In the MS layer, four clusters representing each type of gesture appeared at the end of human gesture observation (j)
and at the onset of target object observation (k). This implies that higher-level proprioception was calibrated based on
the perceived human gestures before the robot exhibited different behaviors depending on the target object. At
mid-reach (l), the MS layer encoded both the type and the location of the target object, but differences with respect to
the object’s location were less clear than those appearing in the PFC layer. In the MF layer, the internal representations
developed similarly, but less clearly than those in the MS layer. For instance, the MF layer encoded both the type and the
location of the object during mid-reach (o), but representations were less clear than those of the PFC (i) and MS (l).
Fig. 5. The development of internal representations in PFC, MS and MF for exemplar cases. 8 trials consisting of 4 different
configurations of the target object indicated by 2 different types of human gesture were examined using the t-SNE algorithm. Line
colors indicate the type of human gesture, markers indicate the task phase and an arrow indicates the direction of the time step. At
the end of the trajectories, the configuration of each target object is specified.
We further analyzed the development of internal representations for exemplar cases in the PFC, M S and MF layers
using the t-SNE algorithm (Fig. 5). A total number of 8 representative training trials consisting of 4 different target
objects indicated by 2 different human gestures were compared. In the t-SNE analysis, the number of the initial
dimensions and the perplexity were set to 10 and 50 respectively. Fig. 5 illustrates that the proposed model dynamically
integrated perception and computed motor plans. The PFC layer developed higher-level task-related representations
encompassing both visual and motor information whenever available. For instance, the PFC layer identified the type of
gesture even before the end of human gesture presentation. While the robot was observing the gestures, the
representations at the PFC layer already started to develop differently according to the type of gesture. When the robot
started observing the task space, representations developed differently depending on the object’s features, such as
location and type. Analysis shows that the MS layer was calibrated based on the perceived human gesture. For example,
when the robot started observing the task space, representations in the M S layer began to differentiate themselves
depending on the perceived human gesture. The implication here is that higher-level proprioception was calibrated
based on perceived visual information. The representations of the MF layer developed similarly to those of the MS layer,
showing very similar development of representations for the same target object, but they were more closely related the
robot’s current action. In sum, the MS layer mainly encoded higher-level motor actions and mediated between
higher-level cognition (PFC) and the lower-level motor actions (MF).
Fig. 6. The development of the first three principal components (PCs) at each layer in the two opposite cases: (a) the CNN with the
fast PFC condition and (b) the MSTNN with the slow PFC condition. The numbers at the bottom (horizontal lines) indicate the time
step from the beginning of the task to the end of the task, and the colors denote component score values. Legends are omitted since
as our focus is on similarities in the development of neuronal activation with different occlusion timings.
In order to clarify differences in the internal dynamics of the model when vision input was occluded, we compared
two different timings in the two diametrically opposed cases: not occluded and occluded at the onset of reaching, in the
CNN with the fast PFC condition and in the MSTNN with the slow PFC condition. The development of the first three
principal components (PCs) of each layer for the same target object is depicted from the beginning to the end of the task
in Fig. 6. In the CNN with fast PFC condition, the robot failed to grasp the target object when visual input was
completely occluded at the onset of reaching. However, it was able to grasp the target object when the visual input was
not occluded. Analysis clearly reveals that PCs in the PFC, MS and MF layers developed differently in both cases,
implying that the model cannot form consistent representations. In the MSTNN with the slow PFC condition, the robot
was able to grasp the target object in both cases. Analysis reveals that the development of PCs of the PFC, MS and MF
layers in occluded and non-occluded cases were similar. Especially, when vision input was occluded at the onset of
reaching, the activation of MS and MF developed similarly with those in the non-occluded case, implying the
development of the coherent proprioceptive representations regardless of occlusion conditions. In sum, the PFC layer
as well as MTRNN layers developed consistent representations even when visual information was lost at the initiation
of a reaching action, resulting in successful task performance.
VI. DISCUSSION
Throughout the experiments, we verified several key aspects of the proposed model. In this section, we discuss them
in detail.
A. Self-organized Coordinated Dynamic Structure
The proposed model developed a coordinated dynamic mechanism in the whole network, enabling the robot to learn
goal-directed behaviors through the seamless coordination of cognitive skills. In terms of downward causation,
spatio-temporal constraints imposed on the each level of the hierarchy and end-to-end learning performed on the tightly
coupled structure are the two key factors in achieving the coordinated dynamic mechanism.
First, multiple scales of spatio-temporal constraint enabled the model to dynamically compose conceptual
representations in the higher level by assembling perceptual features in the lower level, suggesting the emergence of
different cognitive functionalities in its hierarchy. Sporns [48] argues that cognitive functions develop in human brains
through anatomical spatio-temporal constraints including connectivity and timescales among local regions. Similarly,
different spatio-temporal constraints imposed on different parts of the proposed model lead to the development of
different cognitive functionalities at each level of the hierarchy. For example, the PFC layer mainly encoded
higher-level task-related information whereas the MTRNN layers encoded information related to the robot’s current
action. Specifically, lower-level proprioception layer (MF) developed the representation closely reflecting the robot’s
current action whereas higher-level proprioception layer (MS) played the role of mediating between higher-level
cognition (PFC) and lower-level proprioception (MF). This result is analogous to findings in [49] that higher-level
task-related information was encoded in the PFC and the lower-level arm movements were encoded in the primary
motor cortex of monkeys.
Second, end-to-end learning performed on a tightly coupled structure enabled the model to form coordinated
representations without explicitly encapsulating each perceptual modality, i.e. perception, action, and decision making.
On the present model, the coupling of perception and action is achieved by connecting two different deep networks
through the PFC layer. This coupling enabled the model to generate motor plans dynamically in response to visual
perceptual flow. This finding is in line with the previous studies [13, 40] in which the importance of coupling
perception and action was emphasized. Analysis of neural activation shows that this dynamic motor plan generation
involved continuous integration of the visual perceptual flow, rather than being directly mapped from visual perception.
This dynamic transformation of visual information into behavioral information was also observed in the experiments
with macaque monkey’s brains [50].
Multimodal representation helps with distinguishing object type and orientation [3] and competency for multimodal
information integration is considered essential for embodied cognition [2, 4, 51, 52]. Ultimately, the present model was
able to develop multimodal representations by abstracting and associating visual perception with proprioceptive
information in different pathways. In other words, higher level representations were based not only on visual
information about and object and corresponding human gesture, but also on information about movement trajectories
resulting in the successful grasp of the target object.
B. Memory and Pre-planning Capability
Results indicate that the proposed model is capable of developing and employing working memory. We found that
the robot was able to maintain task-related information in higher levels throughout the task phases, and dynamically
combining it with object percepts. For instance, the robot maintained human intention categorized at the beginning of
the task and combined it with the object percept, so that it could reach and grasp the target object. Furthermore, the
proposed model performed robustly even when visual input was completely and unexpectedly occluded. This memory
capability was achieved by the temporal hierarchy of the model as well as by the recurrent connections of the PFC layer.
Particularly, when the time constants of the higher-level were larger than those of the lower-level, the model showed
the most robust performance in the various circumstances including the experiments with the novel object
configurations as well as the ones with unexpected visual occlusion. Although the suitable values of time constant at
each level might differ depending on the task, the finding of our study suggests that the progressively larger time
constants from the low-level to the higher-level of the architecture played an important role to form the coordinated
dynamic structure including a memory capability. This finding is also in consistency with the other previous studies [9,
17-18, 33-34, 36-37, 46] that have shown the importance of a similar temporal hierarchy in the multiple timescales
neural network such as MSTNN and MTRNN. This internal contextual dynamics of the proposed model highlights a
key difference between the proposed model and the previous study [13] which lacked capability for keeping memory.
Furthermore, the model was able to pre-plan an action prior to action execution. More specifically, neuronal
activation in the higher-level proprioception layer (MS) was calibrated based on the perceived visual images prior to
action execution. This pre-planning capability played a particularly important role in the visual occlusion experiment,
enabling the robot to reach and grasp the target object even without monitoring the target object and the hand during
reaching. This result is analogous to findings in [53] which reported that F5 neurons in the brain of the macaque
monkey encoded grip-specific information even when no movement was intended.
VII. CONCLUSION
The current study introduced the Visuo-Motor Deep Dynamic Neural Network (VMDNN) model which can learn to
read human intention and generate corresponding behaviors in robots by coordinating multiple cognitive processes
including visual recognition, attention switching, memorizing and retrieving with working memory, action preparation
and generation in a seamless manner. The simulation study on the model using the iCub simulator revealed that the
robot could categorize human intention through observing gestures, preserve the contextual information, and execute
corresponding goal-directed actions. The analysis further showed that synergic coordination among these cognitive
processes can be developed when end-to-end learning of tutored experience is performed on the whole network,
allowing dense interaction between subnetworks. In conclusion, the aforementioned cognitive mechanism can be
developed by means of downward causation in terms of spatio-temporal scale differentiation among local subnetworks,
topological connectivity among them, and the way of interacting through sensory-motor coupling.
There are several research directions suggested by this study. First, experiments incorporating other cognitive skills
as well as other sensory modalities need to be conducted for a better understanding of the mechanisms of learning
goal-directed actions in human and other biological neural systems. Second, the scalability of the proposed model will
need to be examined in experiments with a real robot and with a larger variety of objects. The complexity of the task
that the model can perform seems proportional to the size of the learnable parameters including the number of layers in
each subnetwork and the number of neurons in each layer. Third, the robot’s sensorimotor experience in our study was
acquired from tutoring by the experimenter. In some cases, tutoring may be impossible, so it would be worth
investigating how such experience can be obtained autonomously.
REFERENCES
[1] O. Sigaud and A. Droniou, "Towards Deep Developmental Learning," IEEE Transactions on Cognitive and Developmental
Systems, vol. 8, pp. 99-114, 2016.
[2] A. Droniou, S. Ivaldi, and O. Sigaud, "Deep unsupervised network for multimodal perception, representation and
classification," Robotics and Autonomous Systems, vol. 71, pp. 83-98, 2015.
[3] F. Guerin, N. Kruger, and D. Kraft, "A Survey of the Ontogeny of Tool Use: From Sensorimotor Experience to Planning,"
IEEE Transactions on Autonomous Mental Development, vol. 5, pp. 18-45, 2013.
[4] M. Lungarella, G. Metta, R. Pfeifer, and G. Sandini, "Developmental robotics: a survey," Connection Science, vol. 15, pp.
151-190, 2003.
[5] A. Cangelosi, M. Schlesinger, and L. B. Smith, Developmental robotics: From babies to robots: MIT Press, 2015.
[6] L. Berthouze and T. Ziemke, "Epigenetic robotics—modelling cognitive development in robotic systems," Connection
Science, vol. 15, pp. 147-150, 2003.
[7] M. Asada, K. F. MacDorman, H. Ishiguro, and Y. Kuniyoshi, "Cognitive developmental robotics as a new paradigm for the
design of humanoid robots," Robotics and Autonomous Systems, vol. 37, pp. 185-193, 2001.
[8] T. Taniguchi, T. Nagai, T. Nakamura, N. Iwahashi, T. Ogata, and H. Asoh, "Symbol emergence in robotics: a survey,"
Advanced Robotics, vol. 30, pp. 706-728, 2016.
[9] J. Tani, "Self-Organization and Compositionality in Cognitive Brains: A Neurorobotics Study," Proceedings of the IEEE, vol. 102,
pp. 586-605, 2014.
[10] Y. Bengio, A. Courville, and P. Vincent, "Representation Learning: A Review and New Perspectives," Pattern Analysis and
Machine Intelligence, IEEE Transactions on, vol. 35, pp. 1798-1828, 2013.
[11] Y. LeCun, Y. Bengio, and G. Hinton, "Deep learning," Nature, vol. 521, pp. 436-444, 2015.
[12] J. Schmidhuber, "Deep learning in neural networks: An overview," Neural Networks, vol. 61, pp. 85-117, 2015.
[13] S. Levine, C. Finn, T. Darrell, and P. Abbeel, "End-to-end training of deep visuomotor policies," arXiv preprint
arXiv:1504.00702, 2015.
[14] I. Lenz, H. Lee, and A. Saxena, "Deep learning for detecting robotic grasps," The International Journal of Robotics Research,
vol. 34, pp. 705-724, 2015.
[15] A. Di Nuovo, V. M. De La Cruz, and A. Cangelosi, "A Deep Learning Neural Network for Number Cognition: A bi-cultural
study with the iCub," in 2015 Joint IEEE International Conference on Development and Learning and Epigenetic Robotics
(ICDL-EpiRob), 2015, pp. 320-325.
[16] N. Wahlström, T. B. Schön, and M. P. Deisenroth, "From pixels to torques: Policy learning with deep dynamical models,"
arXiv preprint arXiv:1502.02251, 2015.
[17] M. Jung, J. Hwang, and J. Tani, "Self-Organization of Spatio-Temporal Hierarchy via Learning of Dynamic Visual Image
Patterns on Action Sequences," PLoS ONE, 2015.
[18] Y. Yamashita and J. Tani, "Emergence of functional hierarchy in a multiple timescale neural network model: a humanoid robot
experiment," PLoS Computational Biology, vol. 4, p. e1000220, 2008.
[19] P. Savastano and S. Nolfi, "Incremental Learning in a 14 DOF Simulated iCub Robot: Modeling Infant Reach/Grasp
Development," in Biomimetic and Biohybrid Systems: First International Conference, Living Machines 2012, Barcelona,
Spain, July 9-12, 2012. Proceedings, T. J. Prescott, N. F. Lepora, A. Mura, and P. F. M. J. Verschure, Eds., ed Berlin,
Heidelberg: Springer Berlin Heidelberg, 2012, pp. 250-261.
[20] P. Savastano and S. Nolfi, "A Robotic Model of Reaching and Grasping Development," IEEE Transactions on Autonomous
Mental Development, vol. 5, pp. 326-336, 2013.
[21] J. L. McClelland, M. M. Botvinick, D. C. Noelle, D. C. Plaut, T. T. Rogers, M. S. Seidenberg, et al., "Letting structure emerge:
connectionist and dynamical systems approaches to cognition," Trends in Cognitive Sciences, vol. 14, pp. 348-356, 2010.
[22] P. F. Dominey and F. Warneken, "The basis of shared intentions in human and robot cognition," New Ideas in Psychology, vol. 29,
pp. 260-274, 12// 2011.
[23] M. Tomasello, M. Carpenter, J. Call, T. Behne, and H. Moll, "Understanding and sharing intentions: the origins of cultural
cognition," Behav Brain Sci, vol. 28, pp. 675-91; discussion 691-735, Oct 2005.
[24] D. Vernon, S. Thill, and T. Ziemke, "The Role of Intention in Cognitive Robotics," in Toward Robotic Socially Believable
Behaving Systems - Volume I : Modeling Emotions, A. Esposito and C. L. Jain, Eds., ed Cham: Springer International
Publishing, 2016, pp. 15-27.
[25] M. E. McCarty, R. K. Clifton, D. H. Ashmead, P. Lee, and N. Goubet, "How infants use vision for grasping objects," Child
Dev, vol. 72, pp. 973-87, Jul-Aug 2001.
[26] R. E. Lasky, "The effect of visual feedback of the hand on the reaching and retrieval behavior of young infants," Child Dev, vol. 48, pp.
112-7, Mar 1977.
[27] E. Oztop, N. S. Bradley, and M. A. Arbib, "Infant grasp learning: a computational model," Exp Brain Res, vol. 158, pp.
480-503, Oct 2004.
[28] E. Ugur, E. Şahin, and E. Oztop, "Self-discovery of motor primitives and learning grasp affordances," in 2012 IEEE/RSJ
International Conference on Intelligent Robots and Systems, 2012, pp. 3260-3267.
[29] E. Ugur, Y. Nagai, E. Sahin, and E. Oztop, "Staged Development of Robot Skills: Behavior Formation, Affordance Learning
and Imitation with Motionese," IEEE Transactions on Autonomous Mental Development, vol. 7, pp. 119-139, 2015.
[30] Z. Yu and M. Lee, "Human motion based intent recognition using a deep dynamic neural model," Robotics and Autonomous
Systems, vol. 71, pp. 134-149, 9// 2015.
[31] L. Pinto and A. Gupta, "Supersizing Self-supervision: Learning to Grasp from 50K Tries and 700 Robot Hours," arXiv preprint
arXiv:1509.06825, 2015.
[32] K. Noda, H. Arie, Y. Suga, and T. Ogata, "Multimodal integration learning of robot behavior using deep neural networks,"
Robotics and Autonomous Systems, vol. 62, pp. 721-736, 2014.
[33] G. Park and J. Tani, "Development of Compositional and Contextual Communication of Robots by Using the Multiple
Timescales Dynamic Neural Network," presented at the The 5th Joint IEEE International Conference on Development and
Learning and on Epigenetic Robotics, Rhode Island, USA, 2015.
[34] R. Nishimoto and J. Tani, "Development of hierarchical structures for actions and motor imagery: a constructivist view from
synthetic neuro-robotics study," Psychol Res, vol. 73, pp. 545-58, Jul 2009.
[35] A. Krizhevsky, I. Sutskever, and G. E. Hinton, "Imagenet classification with deep convolutional neural networks," in
Advances in neural information processing systems, 2012, pp. 1097-1105.
[36] M. Jung, J. Hwang, and J. Tani, "Multiple spatio-temporal scales neural network for contextual visual recognition of human
actions," in 2014 Joint IEEE International Conferences on Development and Learning and Epigenetic Robotics
(ICDL-Epirob), 2014, pp. 235-241.
[37] S. Jeong, H. Arie, M. Lee, and J. Tani, "Neuro-robotics study on integrative learning of proactive visual attention and motor
behaviors," Cognitive Neurodynamics, vol. 6, pp. 43-59, 2012.
[38] H. Celikkanat, E. Sahin, and S. Kalkan, "Recurrent Slow Feature Analysis for Developing Object Permanence in Robots,"
presented at the IROS 2013 Workshop on Neuroscience and Robotics, Tokyo, Japan, 2013.
[39] P. F. Dominey, M. Hoen, J. M. Blanc, and T. Lelekov-Boissard, "Neurological basis of language and sequential cognition:
evidence from simulation, aphasia, and ERP studies," Brain Lang, vol. 86, pp. 207-25, Aug 2003.
[40] M. Lungarella and G. Metta, "Beyond Gazing, Pointing, and Reaching: A Survey of Developmental Robotics," vol. 101, C. G.
Prince, L. Berthouze, H. Kozima, D. Bullock, G. Stojanov, and C. Balkenius, Eds., ed: Lund University Cognitive Studies,
2003, pp. 81-89.
[41] Y. A. LeCun, L. Bottou, G. B. Orr, and K.-R. Müller, "Efficient backprop," in Neural Networks: Tricks of the Trade. vol. 7700,
G. Montavon, G. B. Orr, and K.-R. Müller, Eds., ed: Springer Berlin Heidelberg, 2012, pp. 9-48.
[42] D. E. Rumelhart, J. L. McClelland, and P. R. Group, Parallel distributed processing vol. 1: MIT press, 1986.
[43] Y. Bengio, P. Lamblin, D. Popovici, and H. Larochelle, "Greedy layer-wise training of deep networks," Advances in neural
information processing systems, vol. 19, p. 153, 2007.
[44] N. G. Tsagarakis, G. Metta, G. Sandini, D. Vernon, R. Beira, F. Becchi, et al., "iCub: the design and realization of an open humanoid
platform for cognitive and neuroscience research," Advanced Robotics, vol. 21, pp. 1151-1175, 2007.
[45] V. Tikhanoff, A. Cangelosi, P. Fitzpatrick, G. Metta, L. Natale, and F. Nori, "An open-source simulator for cognitive robotics
research: the prototype of the iCub humanoid robot simulator," presented at the Proceedings of the 8th Workshop on
Performance Metrics for Intelligent Systems, Gaithersburg, Maryland, 2008.
[46] J. Hwang, M. Jung, N. Madapana, J. Kim, M. Choi, and J. Tani, "Achieving "synergy" in cognitive behavior of humanoids via
deep learning of dynamic visuo-motor-attentional coordination," in 2015 IEEE-RAS 15th International Conference on
Humanoid Robots (Humanoids), 2015, pp. 817-824.
[47] L. Van der Maaten and G. Hinton, "Visualizing data using t-SNE," Journal of Machine Learning Research, vol. 9, p. 85, 2008.
[48] O. Sporns, Networks of the Brain. Cambridge, MA: MIT press, 2011.
[49] H. Mushiake, N. Saito, K. Sakamoto, Y. Itoyama, and J. Tanji, "Activity in the lateral prefrontal cortex reflects multiple steps
of future events in action plans," Neuron, vol. 50, pp. 631-41, May 18 2006.
[50] C. E. Vargas-Irwin, L. Franquemont, M. J. Black, and J. P. Donoghue, "Linking Objects to Actions: Encoding of Target Object
and Grasping Strategy in Primate Ventral Premotor Cortex," The Journal of Neuroscience, vol. 35, pp. 10888-10897, 2015.
[51] K. Meyer and A. Damasio, "Convergence and divergence in a neural architecture for recognition and memory," Trends
Neurosci, vol. 32, pp. 376-82, Jul 2009.
[52] O. L. Georgeon, J. B. Marshall, and R. Manzotti, "ECA: An enactivist cognitive architecture based on sensorimotor
modeling," Biologically Inspired Cognitive Architectures, vol. 6, pp. 46-57, 2013.
[53] J. Carpaneto, M. A. Umilta, L. Fogassi, A. Murata, V. Gallese, S. Micera, et al., "Decoding the activity of grasping neurons
recorded from the ventral premotor area F5 of the macaque monkey," Neuroscience, vol. 188, pp. 80-94, Aug 11 2011.
| 2cs.AI
|
MINIMUM DISTANCE FUNCTIONS OF COMPLETE INTERSECTIONS
arXiv:1601.07604v4 [math.AC] 24 Oct 2017
JOSÉ MARTÍNEZ-BERNAL, YURIKO PITONES, AND RAFAEL H. VILLARREAL
Abstract. We study the minimum distance function of a complete intersection graded ideal in
a polynomial ring with coefficients in a field. For graded ideals of dimension one, whose initial
ideal is a complete intersection, we use the footprint function to give a sharp lower bound for
the minimum distance function. Then we show some applications to coding theory.
1. Introduction
Let S = K[t1 , . . . , ts ] = ⊕∞
d=0 Sd be a polynomial ring over a field K with the standard grading
and let I 6= (0) be a graded ideal of S. The degree or multiplicity of S/I is denoted by deg(S/I).
Fix a graded monomial order ≺ on S and let in≺ (I) be the initial ideal of I.
The footprint of S/I or Gröbner éscalier of I, denoted ∆≺ (I), is the set of all monomials of S
not in the ideal in≺ (I) [28, p. 13, p. 133]. This notion occurs in other branches of mathematics
under different names; see [20, p. 6] for a list of alternative names.
Given an integer d ≥ 1, let M≺,d be the set of all zero-divisors of S/in≺ (I) of degree d that
are in ∆≺ (I), and let F≺,d be the set of all zero-divisors of S/I that are not zero and are a
K-linear combination of monomials in ∆≺ (I) of degree d.
The footprint function of I, denoted fpI , is the function fpI : N+ → Z given by
deg(S/I) − max{deg(S/(in≺ (I), ta )) | ta ∈ M≺,d } if M≺,d 6= ∅,
fpI (d) :=
deg(S/I)
if M≺,d = ∅,
and the minimum distance function of I, denoted δI , is the function δI : N+ → Z given by
deg(S/I) − max{deg(S/(I, f ))| f ∈ F≺,d } if F≺,d 6= ∅,
δI (d) :=
deg(S/I)
if F≺,d = ∅.
These two functions were introduced and studied in [24]. Notice that δI is independent of the
monomial order ≺ (see Lemma 3.9). To compute δI is a difficult problem but to compute fpI is
much easier.
We come to the main result of this paper which gives an explicit lower bound for δI and a
formula for fpI for a family of complete intersection graded ideals:
Theorem 3.14 If the initial ideal in≺ (I) of I is a complete intersection of height s−1 generated
by tα2 , . . . , tαs , with di = deg(tαi ) and 1 ≤ di ≤ di+1 for i ≥ 2, then
s
P
(di − 1) − 1,
(dk+2 − ℓ)dk+3 · · · ds if d ≤
i=2
δI (d) ≥ fpI (d) =
s
P
if d ≥
(di − 1) ,
1
i=2
where 0 ≤ k ≤ s − 2 and ℓ are integers such that d =
Pk+1
i=2
(di − 1) + ℓ and 1 ≤ ℓ ≤ dk+2 − 1.
2010 Mathematics Subject Classification. Primary 13P25; Secondary 94B60, 11T71.
The first and third author were supported by SNI. The second author was supported by CONACyT.
1
2
JOSÉ MARTÍNEZ-BERNAL, YURIKO PITONES, AND RAFAEL H. VILLARREAL
An important case of this theorem, from the viewpoint of applications, is when I is the
vanishing ideal of a finite set of projective points over a finite field (see the discussion below
about the connection of fpI and δI with coding theory). If I is a complete intersection monomial
ideal of dimension 1, then δI (d) = fpI (d) for d ≥ 1 (see Proposition 3.11), but this case is only
of theoretical interest because, by Proposition 2.9, a monomial ideal is a vanishing ideal only in
particular cases.
Let I ⊂ S be a graded ideal such that L = in≺ (I) is a complete intersection of dimension 1.
We give a formula for the degree of S/(L, ta ) when ta is in M≺,d , that is, ta is not in L and is
a zero-divisor of S/L. By an easy classification of the complete intersection property of L (see
Lemma 3.1) there are basically two cases to consider. One of them is Lemma 3.4, and the other
is the following:
Lemma 3.3 If L = in≺ (I) is generated by td22 , . . . , tds s and ta = ta11 · · · tas s is in M≺,d , then
deg(S/(L, ta )) = d2 · · · ds − (d2 − a2 ) · · · (ds − as ).
To show our main result we use the formula for the degree of the ring S/(L, ta ), and then
use Proposition 3.13 to bound the degrees uniformly. The proof of the main result takes place
in an abstract algebraic setting with no reference to vanishing ideals or finite fields.
The formulas for the degree are useful in the following setting. If I = I(X) is the vanishing
ideal of a finite set X of projective points, and in≺ (I) is generated by td22 , . . . , tds s , then Lemma 3.3
can be used to give upper bounds for the number of zeros in X of homogeneous polynomials of
S. In fact, if f ∈ F≺,d and in≺ (f ) = ta11 · · · tas s , then in≺ (f ) is in M≺,d , and by Corollary 3.8
one has:
|VX (f )| ≤ d2 · · · ds − (d2 − a2 ) · · · (ds − as ),
where VX (f ) is the set of zeros or variety of f in X. This upper bound depends on the exponent
of the leading term of f . A more complex upper bound is obtained when the initial ideal of I(X)
is as in Lemma 3.1(ii). In this case one uses the formula for the degree given in Lemma 3.4.
The interest in studying fpI and δI comes from algebraic coding theory. Indeed, if I = I(X) is
the vanishing ideal of a finite subset X of a projective space Ps−1 over a finite field K = Fq , then
the minimum distance δX (d) of the corresponding projective Reed-Muller-type code is equal to
δI(X) (d), and fpI(X) (d) is a lower bound for δX (d) for d ≥ 1 (see Theorem 2.6 and Lemma 3.10).
Therefore, one has the formula:
(2.2)
δI(X) (d) = deg(S/I(X)) − max{|VX (f )| : f 6≡ 0; f ∈ Sd },
where f 6≡ 0 means that f is not the zero function on X. Our abstract study of the minimum
distance and footprint functions provides fresh techniques to study δX (d).
It is well-known that the degree of S/I(X) is equal to |X| [19, Lecture 13]. Hence, using
Eq. (2.2) and our main result, we get the following uniform upper bound for the number of zeros
of all polynomials f ∈ Sd that do not vanish at all points of X.
Corollary 4.2 If the initial ideal in≺ (I(X)) is a complete intersection generated by tα2 , . . . , tαs ,
with di = deg(tαi ) and 1 ≤ di ≤ di+1 for i ≥ 2, then
(4.1)
|VX (f )| ≤ |X| − (dk+2 − ℓ) dk+3 · · · ds ,
for any f ∈ Sd that does not vanish at all point of X, where 0 ≤ k ≤ s − 2 and ℓ are integers
P
such that d = k+1
i=2 (di − 1) + ℓ and 1 ≤ ℓ ≤ dk+2 − 1.
This result gives a tool for finding good uniform upper bounds for the number of zeros in X
of polynomials over finite fields. This is a problem of fundamental interest in algebraic coding
COMPLETE INTERSECTIONS
3
theory [31] and algebraic geometry [30]. We leave as an open question whether this uniform
bound is optimal, that is, whether the equality is attained for some polynomial f .
Tohǎneanu and Van Tuyl conjectured [33, Conjecture 4.9] that if the vanishing ideal I(X) is
a complete intersection generated by polynomials of degrees d2 , . . . , ds and di ≤ di+1 for all i,
then δX (1) ≥ (d2 − 1)d3 . . . ds . By Corollary 4.2 this conjecture is true if in≺ (I(X)) is a complete
intersection. We leave as another open question whether Corollary 4.2 is true if we only assume
that I(X) is a complete intersection (cf. Proposition 3.12).
To illustrate the use of Corollary 4.2 in a concrete situation, consider the lexicographical order
on S with t1 ≺ · · · ≺ ts and a projective torus over a finite field Fq with q 6= 2:
T = {[(x1 , . . . , xs )] ∈ Ps−1 | xi ∈ F∗q for i = 1, . . . , s},
where F∗q = Fq \ {0}. As I(T) is generated by the Gröbner basis {tiq−1 − t1q−1 }si=2 , its initial ideal
is a complete intersection generated by t2q−1 , . . . , tsq−1 . Therefore, noticing that deg(S/I(T)) is
equal to (q − 1)s−1 and setting di = q − 1 for i = 2, . . . , s in Eq. (4.1), we obtain that any
homogeneous polynomial f of degree d, not vanishing at all points of T, has at most
(q − 1)s−1 − (q − 1)s−(k+2) (q − 1 − ℓ)
zeros in T if d ≤ (q − 2)(s − 1) − 1, and k and ℓ are the unique integers such that k ≥ 0,
1 ≤ ℓ ≤ q − 2 and d = k(q − 2) + ℓ. This uniform bound was given in [29, Theorem 3.5] and it
is seen that this bound is in fact optimal by constructing an appropriate polynomial f .
If fpI (d) = δI (d) for d ≥ 1, we say that I is a Geil–Carvalho ideal. For vanishing ideals
over finite fields, this notion is essentially another way of saying that the bound of Eq. (4.1) is
optimal. The first interesting family of ideals where equality holds is due to Geil [13, Theorem 2].
His result essentially shows that fpI (d) = δI (d) for d ≥ 1 when ≺ is a graded lexicographical
order and I is the homogenization of the vanishing ideal of the affine space As−1 over a finite
field K = Fq . Recently, Carvalho [5, Proposition 2.3] extended this result by replacing As−1
by a cartesian product of subsets of Fq . In this case the underlying Reed-Muller-type code is
called an affine cartesian code and an explicit formula for the minimum distance was first given
in [14, 21]. In a very recent paper, Bishnoi, Clark, Potukuchi, and Schmitt give another proof
of this formula [3, Theorem 5.2] using a result of Alon and Füredi [1, Theorem 5] (see also [4]).
As the two most relevant applications of our main result to algebraic coding theory, we recover
the formula for the minimum distance of an affine cartesian code given in [21, Theorem 3.8] and
[14, Proposition 5], and the fact that the homogenization of the corresponding vanishing ideal
is a Geil–Carvalho ideal [5] (see Corollary 4.4).
Then we present an extension of a result of Alon and Füredi [1, Theorem 1]—in terms of the
regularity of a vanishing ideal—about coverings of the cube {0, 1}n by affine hyperplanes, that
can be applied to any finite subset of a projective space whose vanishing ideal has a complete
intersection initial ideal (see Corollary 4.5 and Example 4.6).
Finally, using Macaulay2 [17], we exemplify how some of our results can be used in practice,
and show that the vanishing ideal of P2 over F2 is not Geil–Carvalho by computing all possible
initial ideals (see Example 4.8).
In Section 2 we introduce projective Reed-Muller-type codes and present some of the results
and terminology that will be needed in the paper. For all unexplained terminology and additional
information, we refer to [36] (for deeper advances on the knowledge of the degree), [8] (for the
theory of Gröbner bases), [2, 11, 32] (for commutative algebra and Hilbert functions), and [23, 34]
(for the theory of error-correcting codes and linear codes).
4
JOSÉ MARTÍNEZ-BERNAL, YURIKO PITONES, AND RAFAEL H. VILLARREAL
2. Preliminaries
In this section, we present some of the results that will be needed throughout the paper and
introduce some more notation. All results of this section are well-known.
Let S = K[t1 , . . . , ts ] = ⊕∞
d=0 Sd be a graded polynomial ring over a field K with the standard
grading and let (0) 6= I ⊂ S be a graded ideal. The Hilbert function of S/I is:
HI (d) := dimK (Sd /Id ),
d = 0, 1, 2, . . .
where Id = I ∩ Sd . By the dimension of I we mean the Krull dimension of S/I.
The degree or multiplicity of S/I is the positive integer
(
(k − 1)! lim HI (d)/dk−1 if k ≥ 1,
d→∞
deg(S/I) :=
dimK (S/I)
if k = 0,
and the regularity of the Hilbert function of S/I, or simply the regularity of S/I, denoted
reg(S/I), is the least integer r ≥ 0 such that HI (d) is equal to hI (d) for d ≥ r, where hI is the
Hilbert polynomial of S/I.
Let ≺ be a monomial order on S and let (0) 6= I ⊂ S be an ideal. The leading monomial of
f is denoted by in≺ (f ) and the initial ideal of I is denoted by in≺ (I). A monomial ta is called
a standard monomial of S/I, with respect to ≺, if ta is not in the ideal in≺ (I). A polynomial f
is called standard if f 6= 0 and f is a K-linear combination of standard monomials. The set of
standard monomials, denoted ∆≺ (I), is called the footprint of S/I. If I is graded, then HI (d)
is the number of standard monomials of degree d.
Lemma 2.1. Let ≺ be a monomial order, let I ⊂ S be an ideal, and let f be a polynomial of S
of positive degree. If in≺ (f ) is regular on S/in≺ (I), then f is regular on S/I.
Proof. Let g be a polynomial of S such that gf ∈ I. It suffices to show that g ∈ I. Pick a
Gröbner basis
P g1 , . . . , gr of I. Then, by the division algorithm [8, Theorem 3, p. 63], we can
write g = ri=1 ai gi + h, where h = 0 or h is a standard polynomial of S/I. We need only show
that h = 0. If h 6= 0, then hf is in I and in≺ (h)in≺ (f ) is in in≺ (I). Therefore in≺ (h) is in
in≺ (I), a contradiction.
Remark 2.2. Given an integer d ≥ 1, there is a map in≺ : F≺,d → M≺,d given by f 7→ in≺ (f ).
This follows from Lemma 2.1. If I is a monomial ideal, then M≺,d ⊂ F≺,d .
Projective Reed-Muller-type codes. Let K = Fq be a finite field with q elements, let Ps−1
be a projective space over K, and let X be a subset of Ps−1 . As usual, points of Ps−1 are denoted
by [α], where 0 6= α ∈ K s . In this paragraph all results are valid if we assume that K is any field
and X is a finite subset of Ps−1 , instead of assuming that K is finite. However, the interesting
case for coding theory is when K is finite.
The vanishing ideal of X, denoted I(X), is the ideal of S generated by the homogeneous
polynomials that vanish at all points of X. In this case the Hilbert function of S/I(X) is denoted
by HX (d). Let P1 , . . . , Pm be a set of representatives for the points of X with m = |X|. Fix a
degree d ≥ 1. For each i there is fi ∈ Sd such that fi (Pi ) 6= 0. Indeed suppose Pi = [(a1 , . . . , as )],
there is at least one k in {1, . . . , s} such that ak 6= 0. Setting fi (t1 , . . . , ts ) = tdk one has that
fi ∈ Sd and fi (Pi ) 6= 0. There is a K-linear map:
f (Pm )
f (P1 )
,...,
.
(2.1)
evd : Sd = K[t1 , . . . , ts ]d → K |X| ,
f 7→
f1 (P1 )
fm (Pm )
COMPLETE INTERSECTIONS
5
The map evd is called an evaluation map. The image of Sd under evd , denoted by CX (d), is
called a projective Reed-Muller-type code of degree d over X [10]. It is also called an evaluation
code associated to X [15]. This type of codes have been studied using commutative algebra
methods and especially Hilbert functions, see [9, 16, 26, 31] and the references therein.
Definition 2.3. A linear code is a linear subspace of K m for some m. The basic parameters of
the linear code CX (d) are its length |X|, dimension dimK CX (d), and minimum distance
δX (d) := min{kvk : 0 6= v ∈ CX (d)},
where kvk is the number of non-zero entries of v.
Lemma 2.4. [24, Lemma 2.13] (a) The map evd is well-defined, i.e., it is independent of the
set of representatives that we choose for the points of X. (b) The basic parameters of the ReedMuller-type code CX (d) are independent of f1 , . . . , fm .
The following summarizes the well-known relation between projective Reed-Muller-type codes
and the theory of Hilbert functions. Notice that items (i) and (iv) follow directly from Eq. (2.1)
and item (iii), respectively.
Proposition 2.5. The following hold.
(i)
(ii)
(iii)
(iv)
HX (d) = dimK CX (d) for d ≥ 1.
[19, Lecture 13] deg(S/I(X)) = |X|.
(Singleton bound) 1 ≤ δX (d) ≤ |X| − HX (d) + 1 for d ≥ 1.
δX (d) = 1 for d ≥ reg(S/I(X)).
The next result gives an algebraic formulation of the minimum distance of a projective ReedMuller-type code in terms of the degree and the structure of the underlying vanishing ideal.
Theorem 2.6. [24, Theorem 4.7] If |X| ≥ 2, then δX (d) = δI(X) (d) ≥ 1 for d ≥ 1.
This result gives an algorithm, that can be implemented in CoCoA [7], Macaulay2 [17], or
Singular [18], to compute δX (d) for small values of q and s, where q is the cardinality of Fq and
s is the number of variables of S (see the procedure of Example 4.6). Using SAGE [27] one can
also compute δX (d) by finding a generator matrix of CX (d).
As a direct consequence of Theorem 2.6 one has:
(2.2)
δI(X) (d) = deg(S/I(X)) − max{|VX (f )| : f 6≡ 0; f ∈ Sd },
where VX (f ) is the zero set of f in X and f 6≡ 0 means that f does not vanish at all points of X.
The next lemma follows using the division algorithm [8] (cf. [12, Problem 1-17]).
Lemma 2.7. Let X be a finite subset of Ps−1 , let [α] be a point in X, with α = (α1 , . . . , αs ) and
αk 6= 0 for some k, and let I[α] be the vanishing ideal of [α]. Then I[α] is a prime ideal,
I[α] = ({αk ti − αi tk | k 6= i ∈ {1, . . . , s}), deg(S/I[α] ) = 1,
T
ht(I[α] ) = s − 1, and I(X) = [β]∈X I[β] is the primary decomposition of I(X).
Remark 2.8. If X is a finite set of projective points, then S/I(X) is a Cohen–Macaulay reduced
graded ring of dimension 1. This is very well-known, and it follows directly from Lemma 2.7.
In particular, the regularity of the Hilbert function of S/I(X) is the Castelnuovo–Mumford
regularity of S/I(X).
An ideal I ⊂ S is called unmixed if all its associated primes have the same height. The next
result classifies monomial vanishing ideals of finite sets in a projective space.
6
JOSÉ MARTÍNEZ-BERNAL, YURIKO PITONES, AND RAFAEL H. VILLARREAL
Proposition 2.9. Let X be a finite subset of Ps−1 . The following are equivalent:
(a) I(X) is a monomial ideal.
(b) I(X) = ∩m
i=1 pi , where the pi ’s are generated by s − 1 variables.
(c) X ⊂ {[e1 ], . . . , [es ]}, where ei is the i-th unit vector.
Proof. (a) ⇒ (b): By Remark 2.8, I(X) is a radical Cohen–Macaulay graded ideal of dimension
1. Hence, I(X) is an unmixed square-free monomial ideal of height s − 1. Therefore, I(X) is
equal to ∩m
i=1 pi , where the pi ’s are face ideals (i.e., ideals generated by variables) of height s − 1.
(b) ⇒ (c): Let X be the Zariski closure of X. As X is finite, one has X = X = V (I(X)) =
(pi ). Thus it suffices to notice that V (pi ) = {[eij ]} for some ij .
(c) ⇒ (a): This follows from Lemma 2.7.
∪m
i=1 V
3. Complete intersections
Let S = K[t1 , . . . , ts ] = ⊕∞
d=0 Sd be a polynomial ring over a field K with the standard grading
and s ≥ 2. An ideal I ⊂ S is called a complete intersection if there exist g1 , . . . , gr in S such
that I = (g1 , . . . , gr ), where r is the height of I.
In what follows by a monomial order ≺ we mean a graded monomial order in the sense that
≺ is defined first by total degree [8].
Lemma 3.1. Let L ⊂ S be an ideal generated by monomials. If dim(S/L) = 1, then L is a
complete intersection if and only if, up to permutation of the variables t1 , . . . , ts , we can write
(i) L = (td22 , . . . , tds s ) with 1 ≤ di ≤ di+1 for i ≥ 2, or
dp
c cp+1 dp+2
(ii) L = (td12 , . . . , tp−1
, tpp tp+1
, tp+2 , . . . , tds s ) for some p ≥ 1 such that 1 ≤ cp ≤ cp+1 and
1 ≤ di ≤ di+1 for 2 ≤ i ≤ s − 1, where dp+1 = cp + cp+1 .
Proof. ⇒) Let tα1 , . . . , tαs−1 be the minimal set of generators of L consisting of monomials.
These monomials form a regular sequence. Hence tαi and tαj have no common variables for
i 6= j. Then, either all variables occur in tα1 , . . . , tαs−1 and we are in case (ii), up to permutation
of the variables t1 , . . . , ts , or there is one variable that is not in any of the tαi ’s and we are in
case (i), up to permutation of the variables t1 , . . . , ts .
⇐) In both cases L is an ideal of height s − 1 generated by s − 1 elements, that is, L is a
complete intersection.
Proposition 3.2. [35, Propositions 3.1.33 and 5.1.11] Let A = R1 /I1 , B = R2 /I2 be two
standard graded algebras over a field K, where R1 = K[x], R2 = K[y] are polynomial rings in
disjoint sets of variables and Ii is an ideal of Ri . If R = K[x, y] and I = I1 R + I2 R, then
(R1 /I1 ) ⊗K (R2 /I2 ) ≃ R/I and F (A ⊗K B, x) = F (A, x)F (B, x),
where F (A, x) and F (B, x) are the Hilbert series of A and B, respectively.
Lemma 3.3. Let L be the ideal of S generated by td22 , . . . , tds s . If ta = ta11 tar r · · · tas s , r ≥ 2,
ar ≥ 1, and ai ≤ di − 1 for i ≥ r, then
deg(S/(L, ta )) = deg(S/(L, tar r · · · tas s )) = d2 · · · ds − d2 · · · dr−1 (dr − ar ) · · · (ds − as ),
where ai = 0 if 2 ≤ i < r.
COMPLETE INTERSECTIONS
7
Proof. In what follows we will use the fact that Hilbert functions and Hilbert series are additive
on short exact sequences [12, Chapter 2, Proposition 7]. If a1 ≥ 1, then taking Hilbert functions
in the exact sequence
a1
t
1
S/(L, ta ) −→ S/(L, ta11 ) −→ 0,
0 −→ S/(L, tar r · · · tas s )[−a1 ] −→
and noticing that dim(S/(L, ta11 )) = 0, the first equality follows. Thus we may assume that ta
has the form ta = tar r · · · tas s and ai = 0 for i < r.
We proceed by induction on s ≥ 2. Assume s = 2. Then r = 2, ta = ta22 , (L, ta ) = (ta22 ), and
the degree of S/(L, ta ) is a2 , as required. Assume s ≥ 3. If ai = 0 for i > r, then (L, ta ) = (L, tar r )
is a complete intersection and the required formula follows from [32, Corollary 3.3]. Thus we
may assume that ai ≥ 1 for some i > r. There is an exact sequence
(3.1)
tar
a
d
d
r
r+1
r+1
r−1
0 −→ S/(td22 , . . . , tr−1
· · · tas s )[−ar ] −→
, . . . , tds s , tr+1
, trdr −ar , tr+1
d
d
r−1
r+1
S/(L, ta ) −→ S/(td22 , . . . , tr−1
, tar r , tr+1
, . . . , tds s ) −→ 0.
Notice that the ring on the right is a complete intersection and the ring on the left is isomorphic
to the tensor product
(3.2)
d
d
a
r−1
r+1
r+1
K[t2 , . . . , tr ]/(td22 , . . . , tr−1
, trdr −ar ) ⊗K K[t1 , tr+1 , . . . , ts ]/(tr+1
, . . . , tds s , tr+1
· · · tas s ).
Hence, taking Hilbert series in Eq. (3.1), and applying [32, Corollary 3.3], the theorem of
Hilbert–Serre [32, p. 58], and Proposition 3.2, we can write the Hilbert series of S/(L, ta ) as
F (S/(L, ta ), x) =
xar (1 − xd2 ) · · · (1 − xdr−1 )(1 − xdr −ar ) g(x)
+
(1 − x)r−1
(1 − x)
(1 − xd2 ) · · · (1 − xdr−1 )(1 − xar )(1 − xdr+1 ) · · · (1 − xds )
,
(1 − x)s
where g(x)/(1 − x) is the Hilbert series of the second ring in the tensor product of Eq. (3.2) and
g(1) is its degree. By induction hypothesis
g(1) = dr+1 · · · ds − (dr+1 − ar+1 ) · · · (ds − as ).
Therefore, writing F (S/(L, ta ), x) = h(x)/(1−x) with h(x) ∈ Z[x] and h(1) > 0, and recalling
that h(1) is the degree of S/(L, ta ), we get
deg(S/(L, ta )) = h(1) = d2 · · · dr−1 (dr − ar )g(1) + d2 · · · dr−1 ar dr+1 · · · ds
= d2 · · · ds − (d2 − a2 ) · · · (ds − as ).
✷
d
c
c
d
p
p+1
p+2
, tpp tp+1
Lemma 3.4. (A) Let L be the ideal of S generated by td12 , . . . , tp−1
, tp+2
, . . . , tds s , where
a1
a
a
p ≥ 1, 1 ≤ cp ≤ cp+1 and di ≥ 1 for all i. If t = t1 · · · ts s is not in L, dp+1 = cp + cp+1 , and
ai ≥ 1 for some i, then the degree of S/(L, ta ) is equal to
(i)
d2 · · · ds − (cp+1 − ap+1 )
p−1
Y
(ii.1) d2 · · · ds − (cp − ap )
(di+1 − ai )
s
Y
(di − ai ) if ap < cp , ap+1 ≥ cp+1 ;
i=p+2
i=1
(ii.2) d2 · · · ds − (dp+1 − ap − ap+1 )
(di − ai ) if ap ≥ cp ;
i=p+2
i=1
p−1
Y
s
Y
(di+1 − ai )
p−1
Y
(di+1 − ai )
i=1
s
Y
(di − ai ) if ap < cp , ap+1 < cp+1 .
i=p+2
8
JOSÉ MARTÍNEZ-BERNAL, YURIKO PITONES, AND RAFAEL H. VILLARREAL
(B) Let I be a graded ideal such that L = in≺ (I). If 0 ≤ k ≤ s − 2, ℓ are integers such that
P
d = k+1
i=2 (di − 1) + ℓ and 1 ≤ ℓ ≤ dk+2 − 1, then fpI (d) ≤ (dk+2 − ℓ)dk+3 · · · ds .
a
Proof. (A) Case (i): Assume ap ≥ cp . If ai = 0 for i 6= p, then ta = tpp , and by the first equality
of Lemma 3.3, and using [32, Corollary 3.3], we get
d
a
d
d
c
d
c
c
p
p+2
p+1
, tpp , tp+2
deg(S/(L, ta )) = deg(S/(td12 , . . . , tp−1
, . . . , tds s , tpp tp+1
))
p
p+2
= deg(S/(td12 , . . . , tp−1
, tpp , tp+2
, . . . , tds s )) = d2 · · · dp cp dp+2 · · · ds
= d2 · · · dp (dp+1 − cp+1 )dp+2 · · · ds = d2 · · · ds − cp+1 d2 · · · dp dp+2 · · · ds ,
as required. We may now assume that ai ≥ 1 for some i 6= p. As ta ∈
/ L and ap ≥ cp , one has
ai < di+1 for i = 1, . . . , p − 1, ap+1 < cp+1 , and ai < di for i = p + 2, . . . , s. Therefore from the
exact sequence
c
d
d
a −cp ap+1
tp+1
a
p
p+1
p+2
p−1
0 −→ S/(td12 , . . . , tp−1
, tp+1
, tp+2
, . . . , tds s , ta11 · · · tp−1
tp p
d
c
d
c
tpp
· · · tas s )[−cp ] −→
p
p+2
, tpp , tp+2
S/(L, ta ) −→ S/(td12 , . . . , tp−1
, . . . , tds s ) −→ 0,
and using Lemma 3.3 and [32, Corollary 3.3], the required equality follows.
a
Case (ii): Assume ap < cp . If ai = 0 for i 6= p, then ta = tpp and 0 = ap+1 < cp+1 . Hence, by
[32, Corollary 3.3], we get
a
d
d
p
p+2
deg(S/(L, ta )) = deg(S/(td12 , . . . , tp−1
, tpp , tp+2
, . . . , tds s ))
= d2 · · · dp ap dp+2 · · · ds
= d2 · · · ds − (dp+1 − ap )d2 · · · dp dp+2 · · · ds ,
as required. We may now assume that ai ≥ 1 for some i 6= p. Consider the exact sequence
d
c
d
a
a
c
tpp
p
p+1
p+2
p−1
p+1
0 −→ S/(td12 , . . . , tp−1
, tp+1
, tp+2
, . . . , tds s , t1a1 · · · tp−1
tp+1
· · · tas s )[−cp ] −→
(3.3)
d
d
c
p
p+2
S/(L, ta ) −→ S/(td12 , . . . , tp−1
, tpp , tp+2
, . . . , tds s , ta11 · · · tas s ) −→ 0.
Subcase (ii.1): Assume ap+1 ≥ cp+1 . As ta ∈
/ L, in our situation, one has ai < di+1 for
i = 1, . . . , p − 1, ap < cp , and ai < di for i = p + 2, . . . , s. If ai = 0 for i 6= p + 1, then taking
Hilbert series in Eq. (3.3), and noticing that the ring on the right has dimension 0, we get
deg(S/(L, ta )) = d2 · · · dp cp+1 dp+2 · · · ds
= d2 · · · ds − cp d2 · · · dp dp+2 · · · ds ,
as required. Thus we may now assume that ai ≥ 1 for some i 6= p + 1. Taking Hilbert series in
Eq. (3.3), and using [32, Corollary 3.3], we obtain
deg(S/(L, ta )) = d2 · · · dp cp+1 dp+2 · · · ds +
d
c
d
p
p+2
, tpp , tp+2
deg(S/(td12 , . . . , tp−1
, . . . , tds s , ta11 · · · tas s )).
Therefore, using Lemma 3.3, the required equality follows.
Subcase (ii.2): Assume ap+1 < cp+1 . If ai = 0 for i 6= p + 1, taking Hilbert series in Eq. (3.3),
and noticing that the ring on the right has dimension 0, by Lemma 3.3, we get
deg(S/(L, ta )) = d2 · · · dp cp+1 dp+2 · · · ds − d2 · · · dp (cp+1 − ap+1 )dp+2 · · · ds
= d2 · · · dp ap+1 dp+2 · · · ds
= d2 · · · ds − (dp+1 − ap+1 )d2 · · · dp dp+2 · · · ds ,
COMPLETE INTERSECTIONS
9
as required. Thus we may now assume that ai ≥ 1 for some i 6= p + 1. Taking Hilbert series in
Eq. (3.3), and applying Lemma 3.3 to the ends of Eq. (3.3), the required equality follows.
(B) It suffices to find a monomial tb in M≺,d such that
(3.4)
deg(S/(in≺ (I), tb )) = (dk+2 − ℓ)dk+3 · · · ds .
There are five cases to consider:
dk+1 −1 ℓ
dp −1 cp cp+1 −1 dp+2 −1
td12 −1 · · · tp−1
tp tp+1 tp+2 · · · tk+1
tk+2
dp −1 cp cp+1 −1 ℓ
d2 −1
· · · tp−1 tp tp+1 tp+2
t1
d
−1
b
d
−1
2
t =
· · · tkk+1 tℓk+1
t1
dp −1 cp ℓ−cp
td12 −1 · · · tp−1
tp tp+1
d2 −1
dp −1 ℓ
t1
· · · tp−1 tp
if
if
if
if
if
k
k
k
k
k
≥ p + 1,
= p,
≤ p − 2,
= p − 1 and ℓ ≥ cp ,
= p − 1 and ℓ < cp .
In each case, by the formulas for the degree of part (A), we get the equality of Eq. (3.4).
If f ∈ S, the quotient ideal of I with respect to f is given by (I : f ) = {h ∈ S| hf ∈ I}.
Lemma 3.5. [24, Lemma 4.1] Let I ⊂ S be an unmixed graded ideal and let ≺ be a monomial
order. If f ∈ S is homogeneous and (I : f ) 6= I, then
deg(S/(I, f )) ≤ deg(S/(in≺ (I), in≺ (f ))) ≤ deg(S/I),
and deg(S/(I, f )) < deg(S/I) if I is an unmixed radical ideal and f ∈
/ I.
Remark 3.6. Let I ⊂ S be an unmixed graded ideal of dimension 1. If f ∈ Sd , then (I : f ) = I
if and only if dim(S/(I, f )) = 0. In this case deg(S/(I, f )) could be greater than deg(S/I).
Lemma 3.7. [24, Lemma 3.2] Let X be a finite subset of Ps−1 over a field K and let I(X) ⊂ S
be its graded vanishing ideal. If 0 6= f ∈ S is homogeneous, then the number of zeros of f in X
is given by
deg S/(I(X), f ) if (I(X) : f ) 6= I(X),
|VX (f )| =
0
if (I(X) : f ) = I(X).
Corollary 3.8. Let I = I(X) be the vanishing ideal of a finite set X of projective points, let
f ∈ F≺,d , and in≺ (f ) = ta11 · · · tas s . If in≺ (I) is generated by td22 , . . . , tds s , then there is r ≥ 2 such
that ar ≥ 1, ai ≤ di − 1 for i ≥ r, ai = 0 if 2 ≤ i < r, and
|VX (f )| ≤ d2 · · · ds − (d2 − a2 ) · · · (ds − as ).
Proof. As f is a zero-divisor of S/I, by Lemma 2.1, ta = in≺ (f ) is a zero-divisor of S/in≺ (I).
Hence, there is r ≥ 2 such that ar ≥ 1 and ai = 0 if 2 ≤ i < r. Using that ta is a standard
monomial of S/I, we get that ai ≤ di − 1 for i ≥ r. Therefore, using Lemma 3.7 together with
Lemmas 3.3 and 3.5, we get
|VX (f )| = deg(S/(I(X), f )) ≤ deg(S/(in≺ (I(X)), in≺ (f )))
= d2 · · · ds − (d2 − a2 ) · · · (ds − as ).
✷
Lemma 3.9. Let I be a graded ideal and let ≺ be a monomial order. Then the minimum
distance function δI is independent of ≺.
Proof. Fix a positive integer d. Let Fd be the set of all homogeneous zero-divisors of S/I not
in I of degree d and let f be an element of Fd . Pick a Gröbner basis
P g1 , . . . , gr of I. Then,
by the division algorithm [8, Theorem 3, p. 63], we can write f = ri=1 ai gi + h, where h is a
10
JOSÉ MARTÍNEZ-BERNAL, YURIKO PITONES, AND RAFAEL H. VILLARREAL
homogeneous standard polynomial of S/I of degree d. Since (I : f ) = (I : h), we get that h is in
F≺,d . Hence, as (I, f ) = (I, h), we get the equalities:
δI (d) = deg(S/I) − max{deg(S/(I, f ))| f ∈ F≺,d }
= deg(S/I) − max{deg(S/(I, f ))| f ∈ Fd },
that is, δI (d) does not depend on F≺,d .
Lemma 3.10. Let I be an unmixed graded ideal and ≺ a monomial order. The following hold.
(a) δI (d) ≥ fpI (d) and δI (d) ≥ 0 for d ≥ 1.
(b) fpI (d) ≥ 0 if in≺ (I) is unmixed.
(c) δI (d) ≥ 1 if I is radical.
Proof. If F≺,d = ∅, then clearly δI (d) = deg(S/I) ≥ 1, δI (d) ≥ fpI (d), and if in≺ (I) is unmixed,
then fpI (d) ≥ 0 (this follows from Lemma 3.5). Thus, (a), (b), and (c) hold. Now assume that
F≺,d 6= ∅. Pick a standard polynomial f ∈ Sd such that (I : f ) 6= I and
δI (d) = deg(S/I) − deg(S/(I, f )).
As I is unmixed, by Lemma 3.5, deg(S/(I, f )) ≤ deg(S/(in≺ (I), in≺ (f ))). On the other hand,
by Lemma 2.1, in≺ (f ) is a zero-divisor of S/in≺ (I). Hence δI (d) ≥ fpI (d). Using the second
inequality of Lemma 3.5 it follows that δI (d) ≥ 0, fpI (d) ≥ 0 if in≺ (I) is unmixed, and δI (d) ≥ 1
if I is radical.
Proposition 3.11. If I is an unmixed monomial ideal and ≺ is any monomial order, then
δI (d) = fpI (d) for d ≥ 1, that is, I is a Geil–Carvalho ideal.
Proof. The inequality δI (d) ≥ fpI (d) follows from Lemma 3.10. To show the reverse inequality
notice that M≺,d ⊂ F≺,d because one has I = in≺ (I). Also notice that M≺,d = ∅ if and only if
F≺,d = ∅, this follows from Lemma 2.1. Therefore one has fpI (d) ≥ δI (d).
Proposition 3.12. Let I ⊂ S be a graded ideal and let ≺ be a monomial order. Suppose that
in≺ (I) is a complete intersection of height s − 1 generated by tα2 , . . . , tαs , with di = deg(tαi ) and
di ≥ 1 for all i. The following hold.
(a) [25, Example 1.5.1] I is a complete intersection and dim(S/I) = 1.
P
(b) ([25, Example 1.5.1], [6, Lemma 3.5]) deg(S/I) = d2 · · · ds and reg(S/I) = si=2 (di − 1).
(c) 1 ≤ fpI (d) ≤ δI (d) for d ≥ 1.
Proof. (a): The rings S/I and S/init≺ (I) have the same dimension. Thus dim(S/I) = 1. As
≺ is a graded order, there are f2 , . . . , fs homogeneous polynomials in I with in≺ (fi ) = tαi for
i ≥ 2. Since in≺ (I) = (in≺ (f2 ), . . . , in≺ (fs )), the polynomials f2 , . . . , fs form a Gröbner basis of
I, and in particular they generate I. Hence I is a graded ideal of height s − 1 generated by s − 1
polynomials, that is, I is a complete intersection.
(b): Since I is a complete intersection
generated by the fi ’s, then the degree and regularity of
P
S/I are deg(f2 ) · · · deg(fs ) and si=2 (deg(fi ) − 1), respectively. This follows from the formula
for the Hilbert series of a complete intersection given in [32, Corollary 3.3].
(c) The ideal I is unmixed because, by part (a), I is a complete intersection; in particular
Cohen–Macaulay and unmixed. Hence the inequality δI (d) ≥ fpI (d) follows from Lemma 3.10.
Let ta be a standard monomial of S/I of degree d such that (in≺ (I) : ta ) 6= in≺ (I), that is, ta is
in M≺,d . Using Lemma 3.1, and the formulas for deg(S/(in≺ (I), ta )) given in Lemma 3.3 and
Lemma 3.4, we obtain that deg(S/(in≺ (I), ta )) < deg(S/I). Thus fpI (d) ≥ 1.
COMPLETE INTERSECTIONS
11
Proposition 3.13. [24, Proposition 5.7] Let 1 ≤ e1 ≤ · · · ≤ em and 0 ≤ bi ≤ ei − 1 for
i = 1, . . . , m be integers. If b0 ≥ 1, then
!
m
k+1
m
X
X
Y
(ei − bi ) − (k − 1) − b0 −
(ei − bi ) ≥
bi ek+2 · · · em
(3.5)
i=1
i=1
i=k+2
for k = 0, . . . , m − 1, where ek+2 · · · em = 1 and
We come to the main result of this paper.
Pm
i=k+2 bi
= 0 if k = m − 1.
Theorem 3.14. Let I ⊂ S be a graded ideal and let ≺ be a graded monomial order. If the initial
ideal in≺ (I) is a complete intersection of height s−1 generated by tα2 , . . . , tαs , with di = deg(tαi )
and 1 ≤ di ≤ di+1 for i ≥ 2, then δI (d) ≥ fpI (d) ≥ 1 for d ≥ 1 and
s
P
(di − 1) − 1,
(dk+2 − ℓ)dk+3 · · · ds if d ≤
i=2
fpI (d) =
s
P
if d ≥
(di − 1) ,
1
i=2
where 0 ≤ k ≤ s − 2 and ℓ are integers such that d =
Pk+1
i=2
(di − 1) + ℓ and 1 ≤ ℓ ≤ dk+2 − 1.
Proof. Let ta be any standard monomial
Ps of S/I of degree d which is a zero-divisorPofs S/in≺ (I),
a
that is, t is in M≺,d . Thus d = i=1 ai , where a = (a1 , . . . , as ). We set r = i=2 (di − 1).
Ps
P
If we substitute −ℓ = k+1
i=1 ai in the expression (dk+2 − ℓ)dk+3 · · · ds , it follows
i=2 (di − 1) −
that for d < r the inequality
fpI (d) ≥ (dk+2 − ℓ)dk+3 · · · ds
is equivalent to show that
!
s
k+2
X
X
a
ai dk+3 · · · ds
(di − ai ) − k − a1 −
(3.6)
deg(S/I) − deg(S/(in≺ (I), t )) ≥
i=2
ta
i=k+3
Ps
for any in M≺,d , where by convention i=k+3 ai = 0 and dk+3 · · · ds = 1 if k = s − 2. Recall
that, by Proposition 3.12, one has that fpI (d) ≥ 1 for d ≥ 1. By Lemma 3.1, and by permuting
variables and changing I, ≺, and ta accordingly, one has the following two cases to consider.
Case (i): Assume that in≺ (I) = (td22 , . . . , tds s ) with 1 ≤ di ≤ di+1 for i ≥ 2. Then, as ta is in
M≺,d , we can write ta = ta11 · · · tar r · · · tas s , r ≥ 2, ar ≥ 1, ai = 0 if 2 ≤ i < r, and ai ≤ di − 1 for
i ≥ r. By Lemma 3.3 we get
(3.7)
deg(S/(in≺ (I), ta )) = d2 · · · ds − (d2 − a2 ) · · · (ds − as )
d2 −1
for any ta in M≺,d . If d ≥ r, setting tc = td−r
· · · tds s −1 , one has tc ∈ M≺,d . Then, using
1 t2
c
Eq. (3.7), it follows that deg(S/(in≺ (I), t )) = d2 · · · ds − 1. Thus fpI (d) ≤ 1 and equality
dk+1 −1 ℓ
fpI (d) = 1 holds. We may now assume d ≤ r − 1. Setting tb = td22 −1 · · · tk+1
tk+2 , one has
b
t ∈ M≺,d . Then, using Eq. (3.7), we get
deg(S/(in≺ (I), tb )) = d2 · · · ds − (dk+2 − ℓ)dk+3 · · · ds .
Hence fpI (d) ≤ (dk+2 − ℓ)dk+3 · · · ds . Next we show the reverse inequality by showing that
the inequality of Eq. (3.6) holds for any ta ∈ M≺,d . By Eq. (3.7) it suffices to show that the
following equivalent inequality holds
!
s
k+2
X
X
ai dk+3 · · · ds
(di − ai ) − k − a1 −
(d2 − a2 ) · · · (ds − as ) ≥
i=2
i=k+3
12
JOSÉ MARTÍNEZ-BERNAL, YURIKO PITONES, AND RAFAEL H. VILLARREAL
for any a = (a1 , . . . , as ) such that ta ∈ M≺,d . This inequality follows from Proposition 3.13 by
making m = s − 1, ei = di+1 , bi = ai+1 for i = 1, . . . , s − 1 and b0 = 1 + a1 .
c
c
d
d
p
p+1
p+2
, tpp tp+1
Case (ii): Assume that in≺ (I) = (td12 , . . . , tp−1
, tp+2
, . . . , tds s ) for some p ≥ 1 such that
1 ≤ cp ≤ cp+1 and 1 ≤ di ≤ di+1 for all i, where dp+1 = cp + cp+1 .
d −1 d−r+c
−1 d
c
−1
p
p p+1
p+2
If d ≥ r, setting tc = td12 −1 · · · tp−1
tp
tp+1 tp+2
· · · tds s −1 , we get that tc ∈ M≺,d .
Then, using the first formula of Lemma 3.4, it follows that deg(S/(in≺ (I), tc )) = d2 · · · ds − 1.
Thus fpI (d) ≤ 1 and the equality fpI (d) = 1 holds.
We may now assume d ≤ r − 1. The inequality fpI (d) ≤ (dk+2 − ℓ)dk+3 · · · ds follows from
Lemma 3.4(B). To show that fpI (d) ≥ (dk+2 −ℓ)dk+3 · · · ds we need only show that the inequality
of Eq. (3.6) holds for any ta in M≺,d . Take ta in M≺,d . Then we can write ta = ta11 · · · tas s with
ai < di+1 for i < p and ai < di for i > p + 1. There are three subcases to consider.
Subcase (ii.1): Assume ap ≥ cp . Then cp+1 > ap+1 because ta is a standard monomial of S/I,
and by Lemma 3.4 we get
deg(S/(in≺ (I), ta )) = d2 · · · ds − (cp+1 − ap+1 )
p−1
Y
(di+1 − ai )
(cp+1 − ap+1 )
p−1
Y
(di+1 − ai )
(di − ai )
i=p+2
i=1
≥
s
Y
(di − ai ).
i=p+2
i=1
Therefore the inequality of Eq. (3.6) is equivalent to
s
Y
s
k+2
X
X
(di − ai ) − k − a1 −
ai
i=2
i=k+3
!
dk+3 · · · ds ,
and this inequality follows at once from Proposition 3.13 by making m = s − 1, ei = di+1
for i = 1, . . . , m, bi = ai for P
1 ≤ i ≤ p − 1,
Psbp = ap+1 + cp , bi = ai+1 for p < i ≤ m, and
m
b0 = ap − cp + 1. Notice that i=0 bi = 1 + i=1 ai .
Subcase (ii.2): Assume ap < cp , ap+1 ≥ cp+1 . By Lemma 3.4 we get
a
deg(S/(in≺ (I), t )) = d2 · · · ds − (cp − ap )
p−1
Y
(di+1 − ai )
(cp − ap )
p−1
Y
(di+1 − ai )
(di − ai )
i=p+2
i=1
≥
s
Y
(di − ai ).
i=p+2
i=1
Therefore the inequality of Eq. (3.6) is equivalent to
s
Y
s
k+2
X
X
(di − ai ) − k − a1 −
ai
i=2
i=k+3
!
dk+3 · · · ds ,
and this inequality follows from Proposition 3.13 by making m = s−1, ei = di+1 for i = 1, . . . , m,
bi = ai for 1P
≤ i ≤ p − 1, P
bp = cp+1 + ap , bi = ai+1 for p < i ≤ m, and b0 = ap+1 − cp+1 + 1.
m
Notice that i=0 bi = 1 + si=1 ai .
Subcase (ii.3): Assume ap < cp , ap+1 ≤ cp+1 − 1. By Lemma 3.4 we get
a
deg(S/(in≺ (I), t )) = d2 · · · ds − (dp+1 − ap − ap+1 )
p−1
Y
i=1
(di+1 − ai )
s
Y
(di − ai ).
i=p+2
COMPLETE INTERSECTIONS
13
Therefore the inequality of Eq. (3.6) is equivalent to
(dp+1 − ap − ap+1 )
p−1
Y
(di+1 − ai )
i=1
≥
s
Y
(di − ai )
i=p+2
s
k+2
X
X
(di − ai ) − k − a1 −
ai
i=2
i=k+3
!
dk+3 · · · ds ,
and this inequality follows from Proposition 3.13 by making m = s−1, ei = di+1 for i = 1, . . . , m,
bi = ai forP1 ≤ i ≤ p − P
1, bp = ap + ap+1 , bi = ai+1 for p < i ≤ m, and b0 = 1. Notice that in
s
this case m
b
=
1
+
i
i=0
i=1 ai .
4. Applications and examples
This section is devoted to give some applications and examples of our main result. As the
two most important applications to algebraic coding theory, we recover the formula for the
minimum distance of an affine cartesian code [14, 21], and the fact that the homogenization of
the corresponding vanishing ideal is a Geil–Carvalho ideal [5].
We begin with a basic application for complete intersections in P1 .
Corollary 4.1. If X is a finite subset of P1 and I(X) is a complete intersection, then
|X| − d if 1 ≤ d ≤ |X| − 2,
δI(X) (d) = fpI(X) (d) =
1
if d ≥ |X| − 1.
Proof. Let f be the generator of I(X). In this case d2 = deg(f ) = |X| and reg(S/I(X)) = |X|− 1.
By Proposition 2.5 and Theorem 3.14 one has
δX (d) = δI(X) (d) ≥ fpI(X) (d) = |X| − d for 1 ≤ d ≤ |X| − 2,
and δX (d) = 1 for d ≥ |X| − 1. Assume that 1 ≤ d ≤ |X| − 2. Pick [P1 ], . . . , [Pd ] points in P1 .
By Lemma 2.7, the vanishing ideal I[Pi ] of [Pi ] is a principal ideal generated by a linear form
hi . Notice that VX (hi ), the zero-set of hi in X, is equal to {[Pi ]}. Setting h = h1 · · · hd , we get a
homogeneous polynomial of degree d with exactly d zeros. Thus δX (d) ≤ |X| − d.
As another application we get the following uniform upper bound for the number of zeroes of
all polynomials f ∈ Sd that do not vanish at all points of X.
Corollary 4.2. Let X be a finite subset of Ps−1 , let I(X) be its vanishing ideal, and let ≺
be a monomial order. If the initial ideal in≺ (I(X)) is a complete intersection generated by
tα2 , . . . , tαs , with di = deg(tαi ) and 1 ≤ di ≤ di+1 for i ≥ 2, then
(4.1)
|VX (f )| ≤ deg(S/I(X)) − (dk+2 − ℓ) dk+3 · · · ds ,
for any f ∈ SdPthat does not vanish at all point of X, where 0 ≤ k ≤ s − 2 and ℓ are integers
such that d = k+1
i=2 (di − 1) + ℓ and 1 ≤ ℓ ≤ dk+2 − 1.
Proof. It follows from Corollary 2.6, Eq. (2.2), and Theorem 3.14.
We leave as an open question whether this uniform bound is optimal, that is, whether the
equality is attained for some polynomial f . Another open question is whether Corollary 4.2 is
true if we only assume that I(X) is a complete intersection. This is related to the following
conjecture of Tohǎneanu and Van Tuyl.
14
JOSÉ MARTÍNEZ-BERNAL, YURIKO PITONES, AND RAFAEL H. VILLARREAL
Conjecture 4.3. [33, Conjecture 4.9] Let X be a finite set of points in Ps−1 . If I(X) is a complete
intersection generated by f1 , . . . , fs−1 , with ei = deg(fi ) for i = 1, . . . , s − 1, and 2 ≤ ei ≤ ei+1
for all i, then δX (1) ≥ (e1 − 1)e2 · · · es−1 .
Notice that by Corollary 4.2 this conjecture is true if in≺ (I(X)) is a complete intersection,
and it is also true for s = 2 (see Corollary 4.1).
Affine cartesian codes and coverings by hyperplanes. Given a collection of finite subsets
A2 , . . . , As of a field K, we denote the image of
X ∗ = A2 × · · · × As
under the map As−1 7→ Ps−1 , x 7→ [(1, x)], by X = [1 × A2 × · · · × As ]. The affine Reed-Mullertype code CX∗ (d) of degree d is called an affine cartesian code [21]. The basic parameters of the
projective Reed-Muller-type code CX (d) are equal to those of CX∗ (d) [22].
A formula for the minimum distance of an affine cartesian code is given in [21, Theorem 3.8]
and in [14, Proposition 5]. A short and elegant proof of this formula was given by Carvalho in
[5, Proposition 2.3], where he shows that the best way to study the minimum distance of an
affine cartesian code is by using the footprint. As an application of Theorem 3.14 we also recover
the formula for the minimum distance of an affine cartesian code by examining the underlying
vanishing ideal and show that this ideal is Geil–Carvalho.
Corollary 4.4. [5, 14, 21] Let K be a field and let CX (d) be the projective Reed-Muller type code
of degree d on the finite set X = [1 × A2 × · · · × As ] ⊂ Ps−1 . If 1 ≤ di ≤ di+1 for i ≥ 2, with
di = |Ai |, and d ≥ 1, then the minimum distance of CX (d) is given by
s
P
(dk+2 − ℓ) dk+3 · · · ds if d ≤
(di − 1) − 1,
i=2
δX (d) =
s
P
1
if d ≥
(di − 1) ,
i=2
and I(X) is Geil–Carvalho, that is, δI(X) (d) = fpI(X) (d) for d ≥ 1, where k ≥ 0, ℓ are the unique
P
integers such that d = k+1
i=2 (di − 1) + ℓ and 1 ≤ ℓ ≤ dk+2 − 1.
Proof.QLet ≻ be the reverse lexicographical order on S with t2 ≻ · · · ≻ ts ≻ t1 . Setting
fi = γ∈Ai (ti − γt1 ) for i = 2, . . . , s, one has that f2 , . . . , fs is a Gröbner basis of I(X) whose
initial ideal is generated by td22 , . . . , tds s (see [21, Proposition 2.5]). By Theorem 2.6 one has the
equality δX (d) = δI(X) (d) for d ≥ 1. Thus the inequality “≥” follows at once from Theorem 3.14.
This is the difficult part of the proof. The rest of the argument reduces to finding an appropriate
polynomial f where equality occurs, and to using that the minimum distance δX (d) is 1 for d
greater than or equal to reg(S/I(X)).
P
We set r = si=2 (di − 1). By Propositions 2.5 and 3.12, the regularity and the degree of
S/I(X) are r and |X| = d2 · · · ds , respectively. Assume that d < r. To show the inequality “≤”
notice that there is a polynomial f ∈ Sd which is a product of linear forms such that |VX (f )|,
the number of zeros of f in X, is equal to
d2 · · · ds − (dk+2 − ℓ)dk+3 · · · ds ,
see [21, p. 15]. Hence δX (d) is less than or equal to (dk+2 − ℓ)dk+3 · · · ds . Thus the required
equality holds. If d ≥ r, by Proposition 2.5, δX (d) = 1 for d ≥ r. Therefore, by Theorem 3.14,
I(X) is Geil–Carvalho.
COMPLETE INTERSECTIONS
15
The next result is an extension of a result of Alon and Füredi [1, Theorem 1] that can be
applied to any finite subset of a projective space whose vanishing ideal has a complete intersection
initial ideal relative to a graded monomial order.
Corollary 4.5. Let X be a finite subset of a projective space Ps−1 and let ≺ be a monomial
order such that in≺ (I(X)) is a complete intersection generated by tα2 , . . . , tαs , with di = deg(tαi )
and 1 ≤ di ≤ di+1 for all i. If the hyperplanes H1 , . . . , Hd in Ps−1 avoid P
a point [P ] in X but
otherwise cover all the other |X| − 1 points of X, then d ≥ reg(S/I(X)) = si=2 (di − 1).
Proof. LetPh1 , . . . , hd be the linear forms in S1 that define H1 , . . . , Hd , respectively. Assume
that d < si=2 (di − 1). Consider the polynomial h = h1 · · · hd . Notice that h ∈
/ I(X) because
h(P ) 6= 0, and h(Q) = 0 for all [Q] ∈ X with [Q] 6= [P ]. By Theorem 3.14, δX (d) ≥ fpI(X) (d) ≥ 2.
Hence, h does not vanish in at least two points of X, a contradiction.
Example 4.6. Let S be the polynomial ring F3 [t1 , t4 , t3 , t2 ] with the lexicographical order
t1 ≺ t4 ≺ t3 ≺ t2 , and let I = I(X) be the vanishing ideal of
X = {[(1, 0, 0, 0)], [(1, 1, 1, 0)], [(1, −1, −1, 0)], [(1, 1, 0, 1)],
[(1, −1, 1, 1)], [(1, 0, −1, 1)], [(1, −1, 0, −1)], [(1, 0, 1, −1)], [(1, 1, −1, −1)]}.
Using the procedure below in Macaulay2 [17] and Theorem 3.14, we obtain the following
information. The ideal I(X) is generated by t2 − t3 − t4 , t33 − t3 t21 , and t34 − t4 t21 . The regularity
and the degree of S/I(X) are 4 and 9, respectively, and I(X) is a Geil–Carvalho ideal whose
initial ideal is a complete intersection generated by t2 , t33 , t34 . The basic parameters of the ReedMuller-type code CX (d) are shown in the following table.
d
|X|
HX (d)
δX (d)
fpI(X) (d)
1
9
3
6
6
2
9
6
3
3
3
9
8
2
2
4
9
9
1
1
By Corollary 4.5, if the hyperplanes H1 , . . . , Hd in P3 avoid a point [P ] in X but otherwise
cover all the other |X| − 1 points of X, then d ≥ reg(S/I(X)) = 4.
S=ZZ/3[t2,t3,t4,t1,MonomialOrder=>Lex];
I1=ideal(t2,t3,t4),I2=ideal(t4,t3-t1,t2-t1),I3=ideal(t4,t1+t3,t2+t1)
I4=ideal(t4-t1,t4-t2,t3),I5=ideal(t4-t1,t3-t1,t2+t1),I6=ideal(t2,t1-t4,t3+t1)
I7=ideal(t3,t1+t4,t1+t2),I8=ideal(t2,t4+t1,t3-t1),I9=ideal(t1+t4,t3+t1,t2-t1)
I=intersect(I1,I2,I3,I4,I5,I6,I7,I8,I9)
M=coker gens gb I, regularity M, degree M
h=(d)->degree M - max apply(apply(apply(apply (toList
(set(0..q-1))^**(hilbertFunction(d,M))-(set{0})^**(hilbertFunction(d,M)),
toList),x->basis(d,M)*vector x),z->ideal(flatten entries z)),
x-> if not quotient(I,x)==I then degree ideal(I,x) else 0)--this
--gives the minimum distance in degree d
apply(1..3,h)
Example 4.7. Let S be the polynomial ring S = F3 [t1 , t2 , t3 ] with the lexicographical order
t1 ≻ t2 ≻ t3 , and let I = I(X) be the vanishing ideal of
X = {[(1, 1, 0)], [(1, −1, 0)], [(1, 0, 1)], [(1, 0, −1)], [(1, −1, −1)], [(1, 1, 1)]}.
16
JOSÉ MARTÍNEZ-BERNAL, YURIKO PITONES, AND RAFAEL H. VILLARREAL
As in Example 4.6, using Macaulay2 [17], we get that I(X) is generated by
t22 t3 − t2 t23 , t21 − t22 + t2 t3 − t23 .
The regularity and the degree of S/I(X) are 3 and 6, respectively, I is a Geil–Carvalho ideal,
and in≺ (I) is a complete intersection generated by t22 t3 and t21 . The basic parameters of the
Reed-Muller-type code CX (d) are shown in the following table.
d
|X|
HX (d)
δX (d)
fpI(X) (d)
1
6
3
3
3
2
6
5
2
2
3
6
6
1
1
By Corollary 4.5, if the hyperplanes H1 , . . . , Hd in P2 avoid a point [P ] in X but otherwise
cover all the other |X| − 1 points of X, then d ≥ reg(S/I(X)) = 3.
Next we give an example of a graded vanishing ideal over a finite field, which is not Geil–
Carvalho, by computing all possible initial ideals.
Example 4.8. Let X = P2 be the projective space over the field F2 and let I = I(X) be the
vanishing ideal of X. Using the procedure below in Macaulay2 [17] we get that the binomials
t1 t22 −t21 t2 , t1 t23 −t21 t3 , t2 t23 −t22 t3 form a universal Gröbner basis of I, that is, they form a Gröbner
basis for any monomial order. The ideal I has exactly six different initial ideals and δX 6= fpI
for each of them, that is, I is not a Geil–Carvalho ideal. The basic parameters of the projective
Reed-Muller code CX (d) are shown in the following table.
d
|X|
HX (d)
δX (d)
fpI(X) (d)
1
7
3
4
4
2
7
6
2
1
3
7
7
1
1
load "gfaninterface.m2"
S=ZZ/2[symbol t1, symbol t2, symbol t3]
I=ideal(t1*t2^2-t1^2*t2,t1*t3^2-t1^2*t3,t2*t3^2-t2^2*t3)
universalGroebnerBasis(I)
(InL,L)= gfan I, #InL
init=ideal(InL_0)
M=coker gens gb init
f=(x)-> if not quotient(init,x)==init then degree ideal(init,x) else 0
fp=(d) ->degree M -max apply(flatten entries basis(d,M),f)
apply(1..regularity(M),fp)
Acknowledgments. We thank the referee for a careful reading of the paper and for the
improvements suggested.
References
[1] N. Alon and Z. Füredi, Covering the cube by affine hyperplanes, European J. Combin. 14 (1993), 79–83.
[2] M. F. Atiyah and I. G. Macdonald, Introduction to Commutative Algebra, Addison-Wesley, Reading, MA,
1969.
COMPLETE INTERSECTIONS
17
[3] A. Bishnoi, P. L. Clark, A. Potukuchi and J. R. Schmitt, On zeros of a polynomial in a finite grid, Combin.
Probab. Comput., to appear.
[4] A. Bishnoi, P. L. Clark, A. Potukuchi and J. R. Schmitt, On the Alon-Füredi bound, Electron. Notes Discrete
Math. 54 (2016), 57–62.
[5] C. Carvalho, On the second Hamming weight of some Reed-Muller type codes, Finite Fields Appl. 24 (2013),
88–94.
[6] M. Chardin and G. Moreno-Socı́as, Regularity of lex-segment ideals: some closed formulas and applications,
Proc. Amer. Math. Soc. 131 (2003), no. 4, 1093–1102 (electronic).
[7] CoCoATeam, CoCoA: a system for doing Computations in Commutative Algebra. Available at
http://cocoa.dima.unige.it.
[8] D. Cox, J. Little and D. O’Shea, Ideals, Varieties, and Algorithms, Springer-Verlag, 1992.
[9] P. Delsarte, J. M. Goethals and F. J. MacWilliams, On generalized Reed-Muller codes and their relatives,
Information and Control 16 (1970), 403–442.
[10] I. M. Duursma, C. Renterı́a and H. Tapia-Recillas, Reed-Muller codes on complete intersections, Appl.
Algebra Engrg. Comm. Comput. 11 (2001), no. 6, 455–462.
[11] D. Eisenbud, Commutative Algebra with a view toward Algebraic Geometry, Graduate Texts in Mathematics
150, Springer-Verlag, 1995.
[12] W. Fulton, Algebraic curves, Advanced Book Classics, Addison-Wesley Publishing Company, Advanced
Book Program, Redwood City, CA, 1989, An introduction to algebraic geometry, Notes written with the
collaboration of Richard Weiss, Reprint of 1969 original.
[13] O. Geil, On the second weight of generalized Reed-Muller codes, Des. Codes Cryptogr. 48 (2008), 323–330.
[14] O. Geil and C. Thomsen, Weighted Reed–Muller codes revisited, Des. Codes Cryptogr. 66 (2013), 195–220.
[15] L. Gold, J. Little and H. Schenck, Cayley-Bacharach and evaluation codes on complete intersections, J. Pure
Appl. Algebra 196 (2005), no. 1, 91–99.
[16] M. González-Sarabia, C. Renterı́a and H. Tapia-Recillas, Reed-Muller-type codes over the Segre variety,
Finite Fields Appl. 8 (2002), no. 4, 511–518.
[17] D. Grayson and M. Stillman, Macaulay2, 1996. Available via anonymous ftp from math.uiuc.edu.
[18] G. M. Greuel and G. Pfister, A Singular Introduction to Commutative Algebra, 2nd extended edition, Springer,
Berlin, 2008.
[19] J. Harris, Algebraic Geometry. A first course, Graduate Texts in Mathematics 133, Springer-Verlag, New
York, 1992.
[20] M. Kreuzer and L. Robbiano, Computational Commutative Algebra 2, Springer-Verlag, Berlin, 2005.
[21] H. H. López, C. Renterı́a and R. H. Villarreal, Affine cartesian codes, Des. Codes Cryptogr. 71 (2014), no.
1, 5–19.
[22] H. H. López, E. Sarmiento, M. Vaz Pinto and R. H. Villarreal, Parameterized affine codes, Studia Sci. Math.
Hungar. 49 (2012), no. 3, 406–418.
[23] F. J. MacWilliams and N. J. A. Sloane, The Theory of Error-correcting Codes, North-Holland, 1977.
[24] J. Martı́nez-Bernal, Y. Pitones and R. H. Villarreal, Minimum distance functions of graded ideals and ReedMuller-type codes, J. Pure Appl. Algebra 221 (2017), 251–275.
[25] J. C. Migliore, Introduction to liaison theory and Deficiency Modules, Progress in Mathematics 165,
Birkhäuser Boston, Inc., Boston, MA, 1998.
[26] C. Renterı́a, A. Simis and R. H. Villarreal, Algebraic methods for parameterized codes and invariants of
vanishing ideals over finite fields, Finite Fields Appl. 17 (2011), no. 1, 81–104.
[27] SAGE Mathematical Software, http://www.sagemath.org.
[28] M. Sala, T. Mora, L. Perret, S. Sakata and C. Traverso (eds.), Gröbner Bases, Coding, and Cryptography,
RISC Book Series, Springer, Heidelberg, 2009.
[29] E. Sarmiento, M. Vaz Pinto and R. H. Villarreal, The minimum distance of parameterized codes on projective
tori, Appl. Algebra Engrg. Comm. Comput. 22 (2011), no. 4, 249–264.
[30] W. M. Schmidt, Equations over finite fields, An elementary approach, Lecture Notes in Mathematics 536,
Springer-Verlag, Berlin-New York, 1976.
[31] A. Sørensen, Projective Reed-Muller codes, IEEE Trans. Inform. Theory 37 (1991), no. 6, 1567–1576.
[32] R. Stanley, Hilbert functions of graded algebras, Adv. Math. 28 (1978), 57–83.
[33] S. Tohǎneanu and A. Van Tuyl, Bounding invariants of fat points using a coding theory construction, J. Pure
Appl. Algebra 217 (2013), no. 2, 269–279.
[34] M. Tsfasman, S. Vladut and D. Nogin, Algebraic geometric codes: basic notions, Mathematical Surveys and
Monographs 139, American Mathematical Society, Providence, RI, 2007.
18
JOSÉ MARTÍNEZ-BERNAL, YURIKO PITONES, AND RAFAEL H. VILLARREAL
[35] R. H. Villarreal, Monomial Algebras, Second Edition, Monographs and Research Notes in Mathematics,
Chapman and Hall/CRC, 2015.
[36] W. Vogel, Lectures on results on Bezout’s theorem, Tata Institute of Fundamental Research Lectures on
Mathematics and Physics 74, Springer-Verlag, Berlin, 1984.
Departamento de Matemáticas, Centro de Investigación y de Estudios Avanzados del IPN,
Apartado Postal 14–740, 07000 Mexico City, D.F.
E-mail address: [email protected]
Departamento de Matemáticas, Centro de Investigación y de Estudios Avanzados del IPN,
Apartado Postal 14–740, 07000 Mexico City, D.F.
E-mail address: [email protected]
Departamento de Matemáticas, Centro de Investigación y de Estudios Avanzados del IPN,
Apartado Postal 14–740, 07000 Mexico City, D.F.
E-mail address: [email protected]
| 7cs.IT
|
arXiv:1604.00795v1 [math.AC] 4 Apr 2016
Méthodes Matricielles
Introduction à la
Complexité Algébrique
ABDELJAOUED Jounaı̈di
Maı̂tre Assistant
Université de Tunis
LOMBARDI Henri
Maı̂tre de Conférences
Université de Franche-Comté, Besançon
Version mise à jour en, avril 2012
Avant-Propos
L’algèbre linéaire nous a semblé être le terrain idéal pour une introduction simple et pédagogique aux outils modernes de la complexité
algébrique développés durant les trois dernières décennies.
Le tournant en matière de complexité algébrique en algèbre linéaire
fut la découverte par Strassen [86], en 1969, d’un fait d’une simplicité
déconcertante, mais d’une portée considérable, à savoir que la multiplication de deux matrices carrées d’ordre deux pouvait se faire avec
seulement sept multiplications (non commutatives) au lieu de huit dans
l’anneau de base. Ce qui ramenait la complexité asymptotique de la
multiplication de deux matrices carrées d’ordre n à O(nlog2 7 ) au lieu
de O(n3 ) et faisait descendre pour la première fois l’exposant de n audessous de 3, alors que les recherches antérieures n’avaient réussi qu’à
réduire le coefficient de n3 dans le nombre d’opérations arithmétiques
nécessaires pour calculer le produit de deux matrices carrées d’ordre n
(cf. [18]).
Depuis, de nombreux outils ont été développés. Des notions nouvelles
sont apparues comme celles de complexité bilinéaire et de rang tensoriel utilisées de manière intensive notamment par Bini, Pan, Schönhage,
Strassen, Winograd et d’autres (cf. [7, 8, 19, Pan, 82, 90]) pour réduire
l’exposant α : à l’heure actuelle, on sait que α < 2, 376. Il est cependant conjecturé que la borne inférieure des exposants α acceptables
serait 2, c’est-à-dire que pour tout ε > 0 le produit de deux matrices carrées d’ordre n pourrait être calculé par un circuit arithmétique
de taille O(n2+ε ) et de profondeur O(log n). Cependant ces méthodes,
d’un intérêt théorique certain, sont à l’heure actuelle inapplicables à
cause notamment de la constante démesurée que le (( grand O )) cache
(cf. [Knu] § 4.6.4). Par contre la méthode de Strassen a pu trouver une
implémentation concrète [14], et elle commence à battre la multiplication
usuelle (dite conventionnelle) à partir de n = 70.
ii
Avant-Propos
Le calcul parallèle est une technique, en plein développement, qui
distribue un calcul à faire sur un grand nombre de processeurs travaillant
au même moment, en parallèle. Pour la multiplication rapide de matrices
carrées, si le nombre de processeurs disponibles est suffisamment grand
(de l’ordre de O(nα )), le temps de calcul est alors extrêmement faible
(de l’ordre de O(log n) pour des matrices sur un corps fini).
La multiplication rapide des matrices carrées a de nombreuses applications en algèbre linéaire sur les corps, par exemple l’inversion d’une
matrice carrée peut se faire en O(nα ) avec le même exposant. Cependant, contrairement à la multiplication rapide des matrices, ces algorithmes ne sont pas bien adaptés au calcul parallèle. Ainsi l’agorithme
d’inversion d’une matrice carrée auquel on vient de faire allusion, et que
nous étudierons dans la section 8.2, ne voit jamais son temps de calcul
descendre en dessous d’un O(n log n).
C’est sur la base de résultats parfois anciens qu’on a pu exhiber,
en algèbre linéaire, des algorithmes bien adaptés au calcul parallèle,
s’appuyant sur la multiplication rapide des matrices. Ces algorithmes
sont en outre des algorithmes sans divisions (ou presque) et s’appliquent
donc à des anneaux commutatifs.
C’est le cas en particulier de la méthode développée en 1847 par
l’astronome français Le Verrier améliorée, un siècle plus tard, par Souriau, Frame et Faddeev qui l’utilisent pour le calcul des déterminants,
du polynôme caractéristique, pour l’inversion des matrices, et pour la
résolution des systèmes linéaires. Cette méthode s’est avérée porteuse
d’un algorithme très bien adapté au calcul parallèle, dû à Csanky, qui
en 1976 a construit, dans le cas d’un anneau commutatif contenant le
corps des rationnels, une famille de circuits arithmétiques, pour calculer
en O(log2 n) étapes parallèles les coefficients du polynôme caractéristique.
Une autre méthode, dite de partitionnement ([Gas] pp. 291–298) et
attribuée à Samuelson [79] (1942), a eu un regain d’intérêt avec l’algorithme de Berkowitz [6], qui fournit un calcul rapide, parallèle et
sans division, du polynôme caractéristique. Cet algorithme a permis de
généraliser aux anneaux commutatifs arbitraires le résultat de Csanky
concernant la complexité parallèle, par une voie tout à fait différente.
Nous en présenterons une version parallèle améliorée (section 10.2).
La version séquentielle la plus simple de l’algorithme de Berkowitz
n’utilise pas de produits de matrices mais seulement des produits d’une
Avant-Propos
iii
matrice par un vecteur.
Elle s’est avérée tout à fait efficace sur les ordinateurs usuels, et particulièrement bien adaptée au cas des matrices creuses.
Nous présentons dans cet ouvrage les principaux algorithmes en algèbre linéaire, et donnons plus particulièrement un aperçu détaillé des
différentes méthodes utilisées pour le calcul du polynôme caractéristique,
avec des résultats récents.
L’intérêt porté au polynôme caractéristique d’une matrice est justifié
par le fait que la détermination de ses coefficients suffit à connaı̂tre le
déterminant de cette matrice et à calculer son adjointe. Dans le cas des
corps cela permet de calculer son inverse et de résoudre les systèmes d’équations linéaires. Il réside également dans les renseignements que cela
donne sur une forme quadratique, comme par exemple sa signature dans
le cas du corps des réels.
Plan de l’ouvrage
Nous faisons quelques rappels d’algèbre linéaire dans le chapitre 1.
Le chapitre 2 contient quelques méthodes classiques couramment utilisées pour le calcul du polynôme caractéristique : l’algorithme de Jordan-Bareiss, la méthode de Hessenberg, la méthode d’interpolation de
Lagrange, l’algorithme de Le Verrier et son amélioration par SouriauFaddeev-Frame, la méthode de Samuelson modifiée à la Berkowitz, en
général la plus efficace, la méthode de Chistov qui a des performances
voisines, et enfin des méthodes reliées aux suites récurrentes linéaires,
les plus efficaces sur les corps finis.
Le chapitre 3 développe le formalisme des circuits arithmétiques (ou
programmes d’évaluation) pour une description formelle des calculs algébriques. Nous y expliquons la technique importante d’élimination des
divisions, elle aussi inventée par Strassen.
Dans le chapitre 4 nous donnons un aperçu des principales notions
de complexité les plus couramment utilisées. Ces notions constituent une
tentative de théoriser les calculs sur ordinateur, leur temps d’exécution
et l’espace mémoire qu’ils occupent.
Dans le chapitre 5 nous expliquons la stratégie générale (( diviser
pour gagner )), bien adaptée au calcul parallèle. Nous donnons quelques
exemples de base.
Le chapitre 6 est consacré à la multiplication rapide des polynômes,
avec la méthode de Karatsuba et la Transformée de Fourier Discrète.
iv
Avant-Propos
Le chapitre 7 est consacré à la multiplication rapide des matrices.
Nous y abordons notamment les notions fondamentales de complexité
bilinéaire, de rang tensoriel et de calculs bilinéaires approximatifs.
Le chapitre 8 est consacré à des algorithmes dans lesquels intervient
la multiplication rapide des matrices, mais sans que l’ensemble de l’algorithme soit bien adapté au calcul parallèle.
On obtient ainsi en général les procédures les plus rapides connues
en ce qui concerne le temps séquentiel asymptotique, pour la plupart des
problèmes classiques liés à l’algèbre linéaire. Ces performances sont en
général obtenues uniquement sur les corps. Seule la dernière section du
chapitre, consacrée à l’algorithme de Kaltofen-Wiedemann concerne le
calcul sur un anneau commutatif arbitraire.
Le chapitre 9 présente les parallélisations de la méthode de Le Verrier,
qui s’appliquent dans tout anneau commutatif où les entiers sont non
diviseurs de zéro et où la division par un entier, quand elle est possible,
est explicite.
Le chapitre 10 est consacré aux méthodes parallèles de Chistov et de
Berkowitz qui s’appliquent en toute généralité.
Le chapitre 11 présente tout d’abord quelques tableaux récapitulatifs
des complexités des différents algorithmes étudiés, séquentiels ou parallèles, pour le calcul du déterminant et celui du polynôme caractéristique.
Nous donnons ensuite les résultats des tests expérimentaux concernant
quelques méthodes séquentielles du calcul du polynôme caractéristique.
Ces résultats montrent des performances supérieures pour les algorithmes de Chistov et de Berkowitz avec un léger avantage pour ce dernier.
Les deux derniers chapitres sont consacrés aux travaux de Valiant
sur un analogue algébrique de la conjecture P =
6 N P, dans lesquels le
déterminant et le permanent occupent une place centrale. Bien qu’on
ait très peu d’idées sur la manière de résoudre la conjecture de Valiant
VP 6= VN P, celle-ci semble quand même moins hors de portée que la
conjecture algorithmique dont elle s’inspire.
L’annexe contient les codes Maple des algorithmes expérimentés.
Nous avons choisi le logiciel de Calcul Formel Maple essentiellement
pour des raisons de commodité. Le langage de programmation qui lui
est rattaché est proche de celui de nombreux autres langages classiques,
permettant de définir et de présenter de manière lisible et efficace les
algorithmes considérés. Les autres langages de calcul formel généralistes
auraient pu aussi bien faire l’affaire. Il n’y aura d’ailleurs aucun mal à
Avant-Propos
v
implémenter dans un de ces langages les algorithmes présentés dans ce
livre. Une liste récapitulative en est donnée dans la table page 355.
L’esprit dans lequel est écrit cet ouvrage
Nous avons en général donné des preuves complètes de nos résultats,
en accordant une grande place aux exemples. Mais il nous est aussi arrivé
de ne donner qu’une idée de la preuve, ou de ne la donner complètement
que sur un exemple, ou de renvoyer à une référence. Nous assumons très
consciemment ce que nous avons sacrifié de la rigueur formelle au profit
de la compréhension de (( ce qui se passe )). Nous avons essayé de donner
dessins et figures pour illustrer notre texte, tout en ayant conscience
d’en avoir fait bien trop peu.
Nous avons aussi essayé de rapprocher cet exposé de la pratique
concrète des algorithmes, en développant chaque fois que nous l’avons
pu des calculs de complexité dans lesquels nous explicitons les constantes
(( cachées dans le grand O )), sans la connaissance desquelles les résultats
théoriques n’ont pas de réelle portée pratique, et peuvent être trompeurs.
Le niveau requis pour lire ce livre est seulement une bonne familiarité avec l’algèbre linéaire. Le mieux serait évidemment d’avoir lu
auparavant cette perle rare qu’est le livre de Gantmacher [Gan]. On
peut recommander aussi le grand classique (toujours disponible) [LT]
de Lancaster & Tismenetsky. Il est naturellement préférable, mais pas
indispensable, d’avoir une idée des concepts de base de la complexité
binaire pour lesquels nous recommandons les ouvrages [BDG] et [Ste].
Enfin, sur les algorithmes en général, si vous n’avez pas lu le livre de
Knuth [Knu] parce que vous comprenez mal l’anglais ou que vous êtes
plutôt habitués à la langue de Voltaire, avant même de commencer la
lecture de notre ouvrage, écrivez une lettre à tous les éditeurs scientifiques en leur demandant par quelle aberration la traduction en français
n’a pas encore été faite.
Pour aller au delà en Calcul Formel nous recommandons les livres
de von zur Gathen & Gerhard [GG], Bini & Pan [BP], Bürgisser, Clausen & Shokrollahi [BCS], Bürgisser [Bur] et le Handbook of Computer
Algebra [GKW].
Nous espérons que notre livre contribuera à mieux faire saisir l’importance de la complexité algébrique à un moment où les mathématiques
constructives et les solutions algorithmiques se développent de manière
rapide et commencent à occuper de plus en plus une place essentielle
vi
Avant-Propos
dans l’enseignement des Mathématiques, de l’Informatique et des Sciences
de l’ingénieur.
Remerciements Nous remercions Marie-Françoise Roy et Gilles Villard
pour leur relecture attentive et leurs suggestions pertinentes, ainsi que
Peter Bürgisser pour son aide concernant les deux derniers chapitres. Et
enfin François Pétiard qui nous a fait bénéficier avec une patience infinie
de son expertise en LaTeX.
Table des matières
Avant-Propos
i
Table des matières : vous y êtes !
vii
1 Rappels d’algèbre linéaire
1.1 Quelques propriétés générales . . . . . . . . . . . .
1.1.1 Notations . . . . . . . . . . . . . . . . . . .
1.1.2 Formule de Binet-Cauchy . . . . . . . . . .
1.1.3 Rang, déterminant et identités de Cramer .
1.1.4 Identités de Sylvester . . . . . . . . . . . .
1.2 Polynôme caractéristique . . . . . . . . . . . . . .
1.2.1 Matrice caractéristique adjointe . . . . . . .
1.2.2 Formule de Samuelson . . . . . . . . . . . .
1.2.3 Valeurs propres de f (A) . . . . . . . . . . .
1.3 Polynôme minimal . . . . . . . . . . . . . . . . . .
1.3.1 Sous-espaces de Krylov . . . . . . . . . . .
1.3.2 Cas de matrices à coefficients dans Z. . . .
1.4 Suites récurrentes linéaires . . . . . . . . . . . . . .
1.4.1 Polynôme générateur, opérateur de décalage
1.4.2 Matrices de Hankel . . . . . . . . . . . . . .
1.5 Polynômes symétriques et relations de Newton . .
1.6 Inégalité de Hadamard et calcul modulaire . . . . .
1.6.1 Normes matricielles . . . . . . . . . . . . .
1.6.2 Théorème chinois et applications . . . . . .
1.7 Résolution uniforme des systèmes linéaires . . . .
1.7.1 L’inverse de Moore-Penrose . . . . . . . . .
1.7.2 Généralisation sur un corps arbitraire . . .
1
1
1
4
5
9
10
11
13
14
16
16
20
21
21
23
25
30
30
32
34
35
42
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
viii
Table des matières
2 Algorithmes de base en algèbre linéaire
2.1 Méthode du pivot de Gauss . . . . . . . . . . .
2.1.1 Transformations élémentaires . . . . . .
2.1.2 La LU - décomposition . . . . . . . . . .
2.1.3 Recherche de pivot non nul . . . . . . .
2.2 Méthode de Jordan-Bareiss . . . . . . . . . . .
2.2.1 Formule de Dodgson-Jordan-Bareiss . .
2.2.2 Méthode de Jordan-Bareiss modifiée . .
2.2.3 La méthode de Dodgson . . . . . . . .
2.3 Méthode de Hessenberg . . . . . . . . . . . . .
2.4 Méthode d’interpolation de Lagrange . . . . . .
2.5 Méthode de Le Verrier et variantes . . . . . . .
2.5.1 Le principe général . . . . . . . . . . . .
2.5.2 Méthode de Souriau-Faddeev-Frame . .
2.5.3 Méthode de Preparata & Sarwate . . . .
2.6 Méthode de Samuelson-Berkowitz . . . . . . . .
2.6.1 Principe général de l’algorithme . . . .
2.6.2 Version séquentielle . . . . . . . . . . .
2.7 Méthode de Chistov . . . . . . . . . . . . . . .
2.7.1 Le principe général . . . . . . . . . . . .
2.7.2 La version séquentielle . . . . . . . . .
2.8 Méthodes reliées aux suites récurrentes linéaires
2.8.1 L’algorithme de Frobenius . . . . . . . .
2.8.2 Algorithme de Berlekamp/Massey . . .
2.8.3 Méthode de Wiedemann . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. .
. .
. .
3 Circuits arithmétiques
3.1 Circuits arithmétiques et programmes d’évaluation
3.1.1 Quelques définitions . . . . . . . . . . . . .
3.1.2 Circuit arithmétique vu comme un graphe .
3.1.3 Circuits arithmétiques homogènes . . . . .
3.1.4 Le problème des divisions . . . . . . . . . .
3.2 Élimination des divisions à la Strassen . . . . . . .
3.2.1 Le principe général . . . . . . . . . . . . . .
3.2.2 Coût de l’élimination des divisions . . . . .
3.3 Calcul des dérivées partielles . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
51
53
53
56
62
65
66
70
71
74
81
82
82
84
88
90
90
91
93
93
95
97
98
108
109
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
111
. 112
. 112
. 116
. 118
. 119
. 120
. 121
. 125
. 126
Table des matières
ix
4 Notions de complexité
4.1 Machines de Turing et Machines à Accès Direct .
4.2 Complexité binaire, les classes P, N P et #P . .
4.2.1 Calculs faisables . . . . . . . . . . . . . .
4.2.2 Quand les solutions sont faciles à tester .
4.2.3 Problèmes de comptage . . . . . . . . . .
4.3 Complexités arithmétique et binaire . . . . . . .
4.3.1 Complexité arithmétique . . . . . . . . .
4.3.2 Complexité binaire . . . . . . . . . . . . .
4.4 Familles uniformes de circuits . . . . . . . . . . .
4.5 Machines parallèles à accès direct . . . . . . . . .
4.5.1 Une idéalisation des calculs parallèles . .
4.5.2 PRAM-complexité et Processeur-efficacité
4.5.3 Le principe de Brent . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
129
. 129
. 134
. 134
. 135
. 141
. 142
. 142
. 143
. 149
. 151
. 152
. 153
. 156
5 Diviser pour gagner
159
5.1 Le principe général . . . . . . . . . . . . . . . . . . . . . . 159
5.2 Circuit binaire équilibré . . . . . . . . . . . . . . . . . . . 162
5.3 Calcul parallèle des préfixes . . . . . . . . . . . . . . . . . 163
6 Multiplication rapide des polynômes
6.1 Méthode de Karatsuba . . . . . . . . . . . . . .
6.2 Transformation de Fourier discrète usuelle . . .
6.3 Transformation de Fourier discrète rapide . . .
6.3.1 Cas favorable . . . . . . . . . . . . . . .
6.3.2 Cas d’un anneau commutatif arbitraire
6.4 Produits de matrices de Toeplitz . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
171
. 172
. 175
. 177
. 177
. 180
. 183
7 Multiplication rapide des matrices
185
7.1 Analyse de la méthode de Strassen . . . . . . . . . . . . . 187
7.1.1 La méthode et sa complexité . . . . . . . . . . . . 187
7.1.2 Une famille uniforme de circuits arithmétiques . . 191
7.2 Inversion des matrices triangulaires . . . . . . . . . . . . . 195
7.3 Complexité bilinéaire . . . . . . . . . . . . . . . . . . . . . 197
7.3.1 Rang tensoriel d’une application bilinéaire . . . . . 198
7.3.2 Exposant de la multiplication des matrices carrées 204
7.3.3 Complexités bilinéaire et multiplicative . . . . . . 205
7.3.4 Extension du corps de base . . . . . . . . . . . . . 207
7.4 Calculs bilinéaires approximatifs . . . . . . . . . . . . . . 209
7.4.1 Méthode de Bini . . . . . . . . . . . . . . . . . . . 209
x
Table des matières
7.4.2
7.4.3
7.4.4
Une amélioration décisive de Schönhage . . . . . . 215
Sommes directes d’applications bilinéaires . . . . 221
L’inégalité asymptotique de Schönhage . . . . . . . 225
8 Algèbre linéaire séquentielle rapide
8.1 L’Algorithme de Bunch & Hopcroft .
8.2 Calcul du déterminant et de l’inverse
8.3 Forme réduite échelonnée en lignes .
8.4 Méthode de Keller-Gehrig . . . . . .
8.5 Méthode de Kaltofen-Wiedemann . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
229
231
235
236
243
245
9 Parallélisations de la méthode de Leverrier
255
9.1 Algorithme de Csanky . . . . . . . . . . . . . . . . . . . . 255
9.2 Amélioration de Preparata et Sarwate . . . . . . . . . . . 259
9.3 Amélioration de Galil et Pan . . . . . . . . . . . . . . . . 266
10 Polynôme caractéristique sur un anneau arbitraire
10.1 Méthode générale de parallélisation . . . . . . . . . .
10.2 Algorithme de Berkowitz amélioré . . . . . . . . . .
10.3 Méthode de Chistov . . . . . . . . . . . . . . . . . .
10.4 Applications des algorithmes . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
271
271
272
283
287
11 Résultats expérimentaux
11.1 Tableaux récapitulatifs des complexités . . . . . . . . . .
11.2 Présentation des tests . . . . . . . . . . . . . . . . . . .
11.3 Tableaux de Comparaison . . . . . . . . . . . . . . . . .
291
. 291
. 295
. 296
12 Le déterminant et les expressions arithmétiques
12.1 Expressions, circuits et descriptions . . . . . . . . .
12.2 Parallélisation des expressions et des circuits . . .
12.3 La plupart des polynômes sont difficiles à évaluer .
12.4 Le caractère universel du déterminant . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
13 Le permanent et la conjecture P =
6 NP
13.1 Familles d’expressions et de circuits booléens . . . . . .
13.2 Booléen versus algébrique . . . . . . . . . . . . . . . . .
13.2.1 Évaluation booléenne des circuits arithmétiques
13.2.2 Simulation algébrique des circuits et expressions
booléennes . . . . . . . . . . . . . . . . . . . . .
13.2.3 Formes algébriques déployées . . . . . . . . . . .
13.3 Complexité binaire versus complexité booléenne . . . .
303
303
309
313
315
321
. 321
. 328
. 328
. 330
. 333
. 335
Table des matières
xi
13.4 Le caractère universel du permanent . . . . . . . . . . . . 339
13.5 La conjecture de Valiant . . . . . . . . . . . . . . . . . . . 340
Annexe : codes Maple
Tables, bibliographie, index.
Liste des algorithmes, circuits
Liste des Figures . . . . . . .
Bibliographie . . . . . . . . .
Index des notations . . . . . .
Index des termes . . . . . . .
343
et programmes d’évaluation
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
355
355
357
359
371
371
1. Rappels d’algèbre
linéaire
Introduction
Ce chapitre est consacré à des rappels d’algèbre linéaire insistant
sur quelques identités algébriques liées aux déterminants et polynômes
caractéristiques. Notre but est double. D’une part fixer les notations
et donner les formules qui justifieront les algorithmes de calcul de ces
objets. D’autre part, donner une idée de l’étendue des applications qui
pourront être tirées de ces calculs.
La section 1.1 fixe les notations et rappelle la formule de BinetCauchy ainsi que les identités de Cramer et de Sylvester. La section
1.2 est consacrée au polynôme caractéristique et à la formule de Samuelson. Dans la section 1.3 nous étudions le polynôme minimal et les
sous-espaces de Krylov. La section 1.4 est consacrée aux suites récurrentes linéaires. Nous rappelons les identités liées aux sommes de Newton
dans la section 1.5. La section 1.6 aborde les méthodes du calcul modulaire. Enfin la section 1.7 est consacrée à l’inverse de Moore-Penrose et
à ses généralisations.
1.1
1.1.1
Quelques propriétés générales
Notations
Dans cet ouvrage A est un anneau commutatif et unitaire 1 et K
un corps commutatif. Pour deux entiers positifs quelconques m et n,
Am×n désigne l’ensemble des matrices de m lignes et n colonnes à
coefficients dans A.
1. Si A n’est pas unitaire, on peut toujours le plonger dans un anneau avec unité
(cf. [Jac]).
2
1. Rappels d’algèbre linéaire
Soit A = (aij ) ∈ An×n une matrice carrée d’ordre n ≥ 2 à coefficients dans A et soit un entier r (1 ≤ r ≤ n). On adopte les notations
suivantes :
– Ir est la matrice unité d’ordre r.
– det(A) désigne le déterminant de A : par définition, c’est le polynôme en les aij défini par la même formule que dans le cas d’une
matrice à coefficients dans un corps, autrement dit det(A) est
donné par :
n
X
Y
det(A) =
ε(σ)
ai,σ(i)
σ
i=1
où σ parcourt l’ensemble des permutations de l’ensemble d’indices
{1, . . . , n} et où ε(σ) est la signature de la permutation σ. Lorsque
cela ne prête pas à confusion, nous noterons parfois |A| au lieu de
det(A) le déterminant de A .
– Tr(A) est la trace de A, c’est-à-dire la somme de ses éléments
diagonaux.
– La comatrice de A est la matrice (dij )1≤i,j≤n où chaque dij est
le cofacteur de l’élément en position (i, j) dans A, c’est-à-dire
dij = (−1)i+j det(Bij )
où Bij est la matrice obtenue à partir de A en supprimant la
i - ème ligne et la j - ème colonne.
– On a alors les formules de développement de det(A) (suivant la
i - ème ligne ou suivant la j - ème colonne) valables sur un anneau
commutatif arbitraire :
det(A) =
n
X
k=1
aik dik =
n
X
akj dkj
(1 ≤ i, j ≤ n) .
k=1
– Adj(A) désigne la matrice adjointe de A. C’est la transposée de
la comatrice de A. Rappelons qu’elle vérifie la double égalité :
A Adj(A) = Adj(A) A = det(A) In .
(1.1)
Maintenant A = (aij ) ∈ Am×n désigne une matrice quelconque à coefficients dans A et r un entier ∈ {1, . . . , min(m, n)}.
– Un mineur d’ordre r de A est le déterminant d’une matrice carrée
extraite de A en supprimant m − r lignes et n − r colonnes.
1.1. Quelques propriétés générales
3
– Ai..j,h..k désigne la matrice extraite de A sur les lignes de i à j et
sur les colonnes de h à k (i ≤ j ≤ m, h ≤ k ≤ n). Si j − i = k − h
on note ah..k
i..j = det(Ai..j,h..k ) le mineur correspondant.
– Plus généralement, si α = {α1 , . . . , αr } avec 1 ≤ α1 < · · · < αr ≤
m et β = {β1 , . . . , βs } avec 1 ≤ β1 < · · · < βs ≤ n on notera
Aα,β la matrice extraite de A sur les lignes (resp. les colonnes)
dont les indices sont en ordre croissant dans α (resp. dans β).
– Les sous-matrices principales sont les sous-matrices dont la diagonale principale est extraite de celle de A. On appellera mineurs
principaux de A les déterminants des sous-matrices principales de
A. Comme cas particulier, Ar désigne la sous-matrice A1..r,1..r :
nous dirons que c’est une sous-matrice principale dominante de
A. Son déterminant est appelé un mineur principal dominant.
– Ai,h..k = (aih , . . . , aik ) = Ai..i,h..k est la matrice-ligne extraite de
la i - ème ligne de A sur les colonnes de h à k. On pose de même
Ai..j,h = t (aih , . . . , ajh ) = Ai..j,h..h .
Ar
A1..r,j
(r)
– On définit a ij =
pour tous entiers r, i, j tels
Ai,1..r
aij
que 1 ≤ r ≤ min(m, n) − 1 et r < i ≤ m, r < j ≤ n, le mineur
d’ordre r + 1 de A obtenu en bordant la sous-matrice principale dominante Ar par les coefficients correspondants de la i - ème
ligne et de la j - ème colonne de A, ce qui fait par exemple que
(r−1)
(0)
(−1)
|Ar | = a rr . Par convention on pose a ij = aij et a 00 = 1.
Parmi les propriétés du déterminant qui restent valables dans un
anneau commutatif arbitraire, on doit citer en premier la linéarité par
rapport à chaque ligne et par rapport à chaque colonne. La deuxième
propriété la plus importante est son caractère alterné, c’est-à-dire qu’il
s’annule si deux lignes ou si deux colonnes sont égales. On déduit de ces
propriétés que le déterminant d’une matrice carrée ne change pas si on
ajoute à une ligne (resp. à une colonne) une combinaison linéaire des
autres lignes (resp. des autres colonnes).
Plus généralement, on peut citer toutes les propriétés qui relèvent
d’identités algébriques. Par exemple l’égalité det(AB) = det(A)det(B),
ou encore le théorème de Cayley-Hamilton (qui peut être vu, pour une
matrice carrée d’ordre n, comme une famille de n2 identités algébriques). Ces identités sont vérifiées lorsque les coefficients sont réels, elles
sont donc vraies dans tous les anneaux commutatifs (un polynôme en k
variables à coefficients entiers est identiquement nul si et seulement si il
s’annule sur Zk ).
4
1. Rappels d’algèbre linéaire
1.1.2
Formule de Binet-Cauchy
C’est une formule qui généralise l’égalité det(AB) = det(A)det(B)
au cas d’un produit AB avec A ∈ Am×n et B ∈ An×m ( m < n ) : pour
chaque m-uple β = β1 , . . . , βm extrait en ordre croissant de {1, . . . , n} ,
on considère A1..m,β la matrice extraite de A sur les colonnes β1 , . . . , βm
et Bβ,1..m la matrice extraite 2 de B sur les lignes β1 , . . . , βm , alors on
a la formule de Binet-Cauchy (cf. [Gan] p. 9) :
det(AB) =
X
det(A1..m,β ) det(Bβ,1..m )
β
(1.2)
(la somme comporte Cm
n termes).
Pour le vérifier, on pose A = (aij ) ∈ Am×n , B = (bij ) ∈ An×m et
on utilise les propriétés élémentaires des déterminants pour obtenir la
suite d’égalités immédiates suivantes (avec des notations évidentes) :
Pn
a1, i1 bi1 ,1
..
.
...
..
.
Pn
a1, im bim ,m
..
.
Pn
am, i1 bi1 ,1 . . .
Pn
am, im bim ,m
i1 =1
det(AB) =
i1 =1
=
a1, i1 bi1 ,1
..
.
X
1≤i1 , i2 , ... ,im ≤n
=
X
1≤i1 , i2 , ... , im ≤n
im =1
im =1
...
..
.
a1, im bim ,m
..
.
am, i1 bi1 ,1 . . . am, im bim ,m
a1, i1
..
bi1 ,1 ×bi2 ,2 ×· · ·×bim ,m ×
.
am, i1
...
..
.
a1, im
..
.
.
. . . am, im
Parmi les nm termes de cette somme, il n’y a que m! Cm
n termes qui
risquent de ne pas être nuls 3 . Ce sont, pour chacun des Cm
n multi-indices
β = ( β1 , β2 , . . . , βm ) avec 1 ≤ β1 < β2 < . . . < βm ≤ n , les m! termes
correspondant aux multi-indices (i1 , i2 , . . . , im ) tels que :
{ i1 , i2 , . . . , im } = { β1 , β2 , . . . , βm } .
2. A1..m,β et Bβ,1..m sont des matrices carrées d’ordre m.
3. à cause du fait qu’un déterminant ayant deux colonnes identiques est nul ; c’est
d’ailleurs la raison simple pour laquelle det(AB) = 0 lorsque m > n.
1.1. Quelques propriétés générales
5
On regroupe ces termes en Cm
n sommes partielles. Chaque somme partielle correspond à une valeur β du multi-indice (β1 , β2 , . . . , βm ) et
comporte m! termes dans lesquels on peut mettre en facteur :
a1, β1
..
.
...
..
.
a1, βm
..
.
am, β1
. . . am, βm
= det(A1..m,β ) .
Il suffit alors d’utiliser la définition du déterminant de la matrice Bβ,1..m
pour voir que la somme partielle correspondant au multi-indice β n’est
bβ1 , 1 . . . bβ1 , m
..
..
..
autre que :
det(A1..m,β ). Ce qui donne :
.
.
.
bβm , 1 . . . bβm , m
X
det(AB) =
det(A1..m,β ) × det(Bβ,1..m ) .
2
1≤β1 <β2 <...<βm ≤n
1.1.3
Rang, déterminant et identités de Cramer
Une matrice carrée est dite régulière si elle est inversible, c’est-à-dire
si son déterminant est inversible dans A, et singulière si son déterminant est nul. Une matrice (non nécessairement carrée) est dite fortement
régulière si toutes ses sous-matrices principales dominantes sont régulières (i.e. tous les mineurs principaux dominants sont inversibles dans
l’anneau de base considéré).
Lorsque A est supposé intègre, on désignera par rg(M ) le rang
d’une matrice M quelconque à coefficients dans A, c’est-à-dire l’ordre
maximum des mineurs non nuls de M.
Utilisant les notations ci-dessus, nous rappelons maintenant quelques
résultats élémentaires d’algèbre linéaire dont certains seront accompagnés de brèves démonstrations.
Comme nous travaillerons souvent avec un anneau commutatif arbitraire A, nous aurons besoin de la notion de A – module, qui est la
généralisation aux anneaux de la notion d’espace vectoriel sur un corps.
Un A – module M est par définition un groupe abélien (la loi de groupe
est notée +) muni d’une loi externe A × M → M, (a, x) 7→ a.x qui
vérifie les axiomes usuels (pour tous a, b ∈ A et x, y ∈ M ) :
1.x
a.(b.x)
(a + b).x
a.(x + y)
=
=
=
=
x
(ab).x
a.x + b.x
a.x + a.y
6
1. Rappels d’algèbre linéaire
Nous ne considérerons dans cet ouvrage que des modules libres de dimension finie, c’est-à-dire isomorphes à An , ou parfois le module AN .
Dans le cas où il est intègre, l’anneau A peut être plongé dans
son corps des fractions qui est noté FA , et tout A – module libre de
dimension finie (isomorphe à un module An ) peut être considéré comme
inclus dans un FA – espace vectoriel (isomorphe à (FA )n ). Le rang d’une
matrice est alors égal à son rang usuel si on considère que ses coefficients
sont dans le corps FA .
Dans la suite, chaque fois que l’hypothèse d’intégrité sur A doit
intervenir, elle sera clairement soulignée.
Propriété 1.1.1 Soit A un anneau intègre. Pour toutes matrices carrées M et N d’ordre n ≥ 2, à coefficients dans A, on a :
(i) M N = 0 =⇒ rg(M ) + rg(N ) ≤ n;
(ii) Adj(M ) = 0 ⇐⇒ rg(M ) ≤ n − 2.
Preuve. (i) provient du fait que si u et v sont deux endomorphismes d’un FA – espace vectoriel de dimension finie n, alors : dim(Im u) +
dim(Ker u) = n et (u ◦ v = 0 ⇒ Im v ⊆ Ker u).
(ii) découle du fait que si Adj(M ) = 0, tous les mineurs d’ordre n − 1
de M sont nuls, et réciproquement.
t
u
Propriété 1.1.2 A étant un anneau intègre, le rang de la matrice
adjointe de toute matrice carrée singulière M ∈ An×n (n ≥ 2) est au
plus égal à 1, et on a les équivalences :
Adj(M ) 6= 0 ⇐⇒ rg(M ) = n − 1 ⇐⇒ rg(Adj(M )) = 1.
Preuve. C’est une conséquence de la propriété 1.1.1 sachant que, par
hypothèse, M Adj(M ) = det(M ) In = 0.
t
u
Propriété 1.1.3 Soit A un anneau commutatif arbitraire. Pour tous
entiers n, p ≥ 2 et toutes matrices P ∈ Ap×n , M ∈ An×n et Q ∈ An×p ,
on a l’implication :
det(M ) = 0 =⇒ det(P Adj(M ) Q) = 0
Preuve.
• Supposons tout d’abord l’anneau A intègre.
La propriété (1.1.2) nous permet alors d’écrire :
det(M ) = 0 ⇒ rg(Adj(M )) ≤ 1 ⇒ rg(P Adj(M ) Q) ≤ 1 < p
et de conclure.
1.1. Quelques propriétés générales
7
• Voyons maintenant le cas où A n’est pas intègre, et commençons par
le cas générique. Considérons pour cela l’anneau Z[a1 , a2 , . . . , at ] = Z[a]
où t = n2 + 2np et où les ai sont des variables représentant les entrées
des matrices M, P et Q. Il est bien connu que det(M ) ∈ Z[a] est un
polynôme irréductible dans cet anneau et que donc Z[a] /hdet(M )i est
un anneau intègre. Comme dans cet anneau det(M ) = 0 , on est ramené
au cas précédent et l’on a : det(P Adj(M ) Q) = 0 dans Z[a] /hdet(M )i .
Ceci prouve l’existence d’un polynôme f (a) ∈ Z[a] vérifiant l’identité :
det(P Adj(M ) Q) = f (a) det(M )
(le polynôme f (a) peut être calculé par un algorithme de division exacte
dans l’anneau Z[a]). Cette identité algébrique est vérifiée dans tout anneau commutatif, ce qui permet de conclure dans le cas général.
t
u
Proposition 1.1.4 (Identités de Cramer) Soit A un anneau commutatif arbitraire et A ∈ Am×n . Notons Cj la j- ème colonne de A (Cj =
A1..m,j ) et Bj la matrice extraite de A en supprimant la colonne Cj .
1. Si n = m + 1 et si µj = det(Bj ) on a :
Xn
j=1
(−1)j µj Cj = 0
(1.3)
2. Supposons n ≤ m et que tous les mineurs d’ordre n de A sont
nuls. Soit α = α1 , . . . , αn−1 extrait en ordre croissant de {1, . . . ,
m}. Soit Dj = (Bj )α,1..n−1 ∈ A(n−1)×(n−1) la matrice extraite de
Bj en gardant les lignes de α, et soit νj = det(Dj ). Alors on a :
Xn
j=1
(−1)j νj Cj = 0
(1.4)
Preuve.
Pn
j
Pour le premier point : la coordonnée no k de
j=1 (−1) µj Cj est
égale au déterminant de la matrice obtenue en collant au dessous de A
la ligne no k de A (ceci se voit en développant ce déterminant selon la
dernière ligne). Cette coordonnée est donc nulle.
Le deuxième point se prouve de manière analogue.
t
u
Ces deux égalités peuvent être relues sous la forme (( solution d’un
système linéaire )). Pour la première on considère A ∈ An×n et V ∈
An×1 on note Ej la matrice obtenue à partir de A en remplaçant la
8
1. Rappels d’algèbre linéaire
colonne Cj par V , et µ0j = det(Ej ). Alors (1.3) se relit sous la forme
plus classique :
0
µ1
Xn
.. = A · Adj(A) · V
0
µ Cj = A
(1.5)
det(A) V =
.
j=1 j
0
µn
On peut également relire (1.4) comme suit. On considère une matrice
A ∈ Am×n et V ∈ Am×1 avec m > n. On suppose que tous les mineurs
d’ordre n + 1 de (A|V ) sont nuls. On note toujours Cj = A1..m,j la j ème colonne de A. On choisit α = α1 , . . . , αn extrait en ordre croissant
de {1, . . . , m}. On considère la matrice Fj obtenue à partir de Aα,1..n
en remplaçant la j - ème colonne par V et on pose να,j = det(Fj ). Si
on applique (1.4) avec la matrice (A|V ) ∈ Am×(n+1) on obtient :
Xn
det(Aα,1..n ) V =
να,j Cj
(1.6)
j=1
Proposition 1.1.5 Soit A un anneau commutatif arbitraire non trivial, c’est-à-dire dans lequel 1A 6= 0A , et A ∈ Am×n . Les propriétés
suivantes sont équivalentes :
1. Pour tout V ∈ Am×1 il existe X ∈ An×1 tel que A X = V .
Autrement dit, l’application linéaire ϕ : An → Am définie par A
est surjective.
2. Il existe B ∈ An×m tel que A B = Im .
3. On a n ≥ m et il existe une combinaison linéaire des mineurs
d’ordre m de A qui est égale à 1.
Preuve.
(1) ⇒ (2) Soit ej le j - ème vecteur de la base canonique de Am et soit
Xj un vecteur de An tel que A Xj = ej . On prend pour matrice B la
matrice dont les colonnes sont les Xj .
(2) ⇒ (1) On prend X = B V .
(2) ⇒ (3) Montrons que n < m est impossible. Si tel est le cas on rajoute
m − n colonnes nulles à droite de A et m − n lignes nulles en dessous
de B, on obtient deux matrices carrées A0 et B 0 pour lesquelles on a
det(A0 ) = det(B 0 ) = 0 et A0 · B 0 = A · B = Im , ce qui donne 0A = 1A .
Pour la combinaison linéaire, on applique la formule de Binet-Cauchy
(1.2) avec A B = Im . P
(3) ⇒ (2) Supposons
β cβ det(A1..m,β ) = 1. La somme est étendue
à tous les β = {β1 , . . . , βm } où 1 ≤ β1 < · · · < βm ≤ n. On a
A1..m,β = A · (In )1..n,β . Posons Bβ = (In )1..n,β · Adj(A
P 1..m,β ). Alors
A · Bβ = det(A1..m,β ) Im . Il suffit donc de prendre B = β cβ Bβ .
t
u
1.1. Quelques propriétés générales
1.1.4
9
Identités de Sylvester
Propriété 1.1.6 (Une identité de Sylvester)
Soit A un anneau commutatif quelconque. Pour tout entier n ≥ 2 et
toute matrice A = (aij ) ∈ An×n , on a :
det(A) = ann det(An−1 ) − An,1..n−1 [Adj(An−1 )] A1..n−1,n .
Preuve. Pour obtenir cette formule, il suffit de développer
det(A) =
An−1
A1..n−1,n
An,1..n−1
ann
suivant la dernière ligne puis chacun des cofacteurs des éléments de
An,1..n−1 intervenant dans ce développement suivant la dernière colonne
qui n’est autre que A1..n−1,n .
t
u
Nous allons voir maintenant que la propriété précédente peut être généralisée à d’autres partitions de A. Étant donnés en effet deux entiers
r et n avec n ≥ 2 et 1 ≤ r < n, on associe à toute matrice A ∈ An×n
la partition suivante de la matrice A en blocs :
Ar A12
A=
A21 A22
où A12 = A1..r,r+1..n ∈ Ar×(n−r) , A21 = Ar+1..n,1..r ∈ A(n−r)×r et
A22 = Ar+1..n,r+1..n ∈ A(n−r)×(n−r) .
On a alors le résultat suivant, valable pour tout anneau commutatif
et unitaire A.
Proposition 1.1.7 (Identités de Sylvester) Avec les notations cidessus, et pour tous entiers n et r tels que 1 ≤ r ≤ n − 1, on a
les identités suivantes, dans lesquelles on a posé Br = AdjAr :
Ar
0
|Ar | Ir
Br A12
|Ar | A =
(1.7)
A21 In−r
0
|Ar | A22 − A21 Br A12
|Ar |n−r−1 |A| = det(|Ar | A22 − A21 Br A12 )
(r)
a(r−1)
rr
n−r−1
a(n−1)
=
nn
=
(r)
ar+1,r+1 . . . ar+1,n
..
..
.
.
(r)
(r)
an,r+1 . . . an,n
(n−2)
(n−3)
an−2,n−2 a(n−1)
nn
(n−2)
(1.9)
(n−2)
an−1,n−1 an−1,n
an,n−1
(1.8)
(n−2)
a n,n
(1.10)
10
1. Rappels d’algèbre linéaire
Preuve.
– L’égalité matricielle (1.7) résulte de l’identité Ar Br = |Ar | Ir .
– Pour démontrer (1.8) qui est une identité algébrique, on peut se restreindre au cas où les coefficients aij de A sont des indéterminées. Les
mineurs de A peuvent alors être vus comme des polynômes non nuls
dans l’anneau intègre Z[(aij )1≤i,j≤n ]. L’égalité (1.8) est alors obtenue
en prenant les déterminants des deux membres de l’égalité (1.7) et en
simplifiant par le polynôme |Ar |r+1 .
– Le premier membre de l’identité (1.9) est le même que celui de (1.8).
L’égalité des seconds membres provient du fait que l’élément de la (i−r) ème ligne et (j − r) - ème colonne (r + 1 ≤ i, j ≤ n) de la matrice :
det(Ar ) A22 − A21 [Adj(Ar )] A12 ∈ A(n−r)×(n−r)
est égal à :
aij det(Ar ) − Ai,1..r [Adj(Ar )] A1..r,j
(r)
qui n’est autre que a ij d’après la propriété (1.1.6) appliquée à la
Ar
A1..r,j
matrice
.
Ai,1..r
aij
– L’égalité (1.10) est un cas particulier de (1.9) pour r = n − 2.
t
u
Remarque. Si l’on fait r = n − 1 et par conséquent
An−1
A1..n−1,n
A=
,
An,1..n−1
ann
l’égalité (1.8) donne exactement la formule de la propriété (1.1.6), ce qui
permet d’affirmer que celle-ci est une identité de Sylvester particulière.
Les identités de Sylvester seront utilisées dans la section 2.2 pour le
calcul des déterminants par la méthode de Jordan-Bareiss.
1.2
Polynôme caractéristique
On appelle matrice caractéristique d’une matrice A ∈ An×n la matrice
A − XIn ∈ (A[X])n×n ( X désigne une indéterminée sur A ).
Le polynôme caractéristique de A est, par définition, le déterminant de
sa matrice caractéristique ( 4 ). On le notera PA :
PA (X) = det(A − XIn ) = p0 X n + p1 X n−1 + · · · + pn−1 X + pn .
4. Il serait en fait plus pratique de définir comme le fait Bourbaki le polynôme
caractéristique de A comme le déterminant de XIn − A, mais nous nous en tenons
à l’usage le plus répandu.
1.2. Polynôme caractéristique
11
Notons que p0 = (−1)n , pn = det(A) et que pour 1 ≤ k ≤ n − 1,
le coefficient pk est le produit par (−1)n−k de la somme de tous les
mineurs diagonaux d’ordre k de A ( 5 ). En particulier :
(−1)n−1 p1 =
n
X
aii = Tr(A).
i=1
1.2.1
Matrice caractéristique adjointe
On appelle matrice caractéristique adjointe de A ∈ An×n la matrice :
Q(X) = Adj(XIn − A) = (−1)(n−1) Adj(A − XIn ) .
C’est, à un signe près, l’adjointe de la matrice caractéristique de A . Elle
peut être vue comme un polynôme matriciel de degré n − 1 en X, à
coefficients dans An×n : en effet, la matrice des cofacteurs de XIn − A
est une matrice n × n dont les éléments diagonaux sont des polynômes
de degré n − 1 en X et les autres des polynômes de degré n − 2 en X .
Ce qui fait que la matrice caractéristique adjointe de A s’écrit :
Q(X) = B0 X n−1 + B1 X n−2 + · · · + Bn−2 X + Bn−1 ∈ An×n [X] .
(1.11)
D’après l’équation (1.1) page 2, on a l’égalité :
(XIn − A) Q(X) = P (X) In où P (X) = (−1)n PA (X).
On posera :
P (X) = X n − [ c1 X n−1 + · · · + cn−1 X + cn ] .
Ainsi le polynôme P (X) In est divisible par le polynôme (XIn − A) au
sens de la division euclidienne dans l’anneau de polynômes An×n [X].
Pour obtenir les coefficients (matriciels) du quotient Q(X) dans cette
division, on applique la procédure de Horner 6 au polynôme matriciel
5. Pour s’en convaincre, on peut examiner la formule qui donne par définition
PA (−X) = det(A + XIn ) et voir quels sont les produits qui contiennent X n−k . On
peut aussi faire une preuve par récurrence sur n en développant det(A + XIn ) suivant la première colonne.
6. La procédure (ou encore schéma) de Horner n’est rien d’autre qu’une
P misei
en forme algorithmique de la division euclidienne d’un polynôme P (X) = n
i ci X
par un polynôme X − a. Le reste, égal à P (a), est alors obtenu sous la forme c0 +
a (c1 + a (c2 + · · · + a cn ) · · ·). Ceci consitue une évaluation efficace de P (a), utilisant
un minimum de multiplications. Cette méthode est en fait identique à celle de Ch’in
Chiu-Shao employée en Chine médiévale. Elle a été redécouverte par Ruffini (1802)
et Horner (1819). Voir l’Encyclopedia of Mathematics, chez Kluwer (1996).
12
1. Rappels d’algèbre linéaire
P (X) In ∈ An×n [X] avec la constante A ∈ An×n . Le reste Bn = P (A)
est nul ; ce qui donne l’identité PA (A) = 0 et fournit en passant une
démonstration (élégante) du théorème de Cayley-Hamilton.
Le procédé de Horner peut être représenté par le schéma suivant :
P (X) In
A
In
0
−c1 In
B0
avec : B0 = In
et
−c2 In
B1
...
...
−cn−1 In
Bn−2
−cn In
Bn−1
Bk = ABk−1 − ck In (1 ≤ k ≤ n) .
Bn = 0
(1.12)
Cela donne une méthode rapide et efficace pour calculer, à partir
des coefficients du polynôme caractéristique, la matrice caractéristique
adjointe, et fournit les relations détaillées suivantes utilisées pour établir
la formule de Samuelson (§ 1.2.2 page ci-contre) :
B1 = A − c1 In
2
B2 = A − c1 A − c2 In
..
..
..
.
.
.
(1.13)
k − c Ak−1 − . . . − c
B
=
A
1
k
k−1 A − ck In
..
..
..
.
.
.
Bn = An − c1 An−1 − . . . − cn−1 A − cn In = 0
Notons qu’à la fin de la procédure de Horner, on obtient Bn = 0
c’est-à-dire A Bn−1 − cn In = 0 ou encore :
A Bn−1 = cn In = (−1)n−1 det(A) In .
Si det(A) est inversible dans A, alors A possède un inverse qui peut
être calculé par la formule A−1 = (cn )−1 Bn−1 .
Notons que Bn−1 = Q(0) = (−1)n−1 Adj(A). Donc
Adj(A) = (−1)(n−1) [ An−1 − c1 An−2 − · · · − cn−2 A − cn−1 In ] .
Ainsi le calcul de la matrice caractéristique adjointe nous donne l’adjointe de A. Il nous permet aussi d’obtenir l’inverse de A (s’il existe)
notamment dans les cas où la méthode du pivot de Gauss s’avérerait impraticable, ce qui se produit lorsque l’anneau contient des diviseurs de
zéro. La matrice caractéristique adjointe de A sert également, comme
nous le verrons plus loin (voir § 2.5.2 page 84 et suivantes), à calculer
le polynôme caractéristique de A par la méthode de Faddeev-SouriauFrame et, dans certains cas, des vecteurs propres non nuls.
Nous allons à présent l’utiliser pour établir un résultat important
pour la suite et faisant l’objet du paragraphe suivant.
1.2. Polynôme caractéristique
1.2.2
13
Formule de Samuelson
Utilisant l’égalité 1.11 et les relations 1.13 dans lesquelles on remplace
les coefficients ci par les coefficients pi du polynôme caractéristique de
A , sachant que pi = (−1)n+1 ci , on obtient l’identité algébrique :
Adj(A − X In ) = −
n−1
X
X
k
j=0
k=0
pj A
k−j
X n−1−k .
(1.14)
Cette égalité nous sert à démontrer la formule de Samuelson [79] (voir
[Gas], méthode de partitionnement pp. 291–298, [FF]) .
Proposition 1.2.1 (Formule de Samuelson)
Soit A un anneau commutatif arbitraire,
≥ 2 , A = (aij ) ∈
Prn un entier
i
n×n
A
et r = n − 1 . Notons Pr (X) = i=0 qr−i X = det(Ar − XIr ) le
polynôme caractéristique de la sous-matrice principale dominante Ar .
Posons Rr := An,1..r et Sr := A1..r,n de sorte que la matrice A est
partitionnée comme suit :
Ar Sr
A :=
.
Rr ann
Alors on a :
PA (X) = (an,n − X) Pr (X) +
r−1 X
X
k
qj (Rr Ak−j
r Sr )
j=0
k=0
X r−1−k .
C’est-à-dire encore :
(
(an,n − X) Pn−1 (X) +
n−2−k
PA (X) =
Pn−2
k
k=0 q0 (Rr Ar Sr ) + · · · + qk (Rr Sr ) X
(1.15)
Preuve. Tout d’abord on applique l’identité de Sylvester donnée en
1.1.6 à la matrice (A − XIn ). On obtient :
PA (X) = (an,n − X) det(Ar − XIr ) − Rr Adj(Ar − XIr ) Sr .
Ensuite on applique (1.14) en remplaçant A par Ar et n par r :
Adj(Ar − XIr ) = −
r−1 X
X
k
k=0
qj Ak−j
r
j=0
X r−1−k .
2
La formule de Samuelson sera utilisée dans l’algorithme de Berkowitz.
14
1.2.3
1. Rappels d’algèbre linéaire
Valeurs propres de f (A)
Supposons l’anneau A intègre et soit FA son corps des fractions
et L une extension de FA dans laquelle PA se décompose en produit
de facteurs du premier degré. Une telle extension est ce qu’on appelle
un corps de décomposition de PA . Tout zéro de PA dans L est appelé
valeur propre de A, et sa multiplicité est, par définition, la multiplicité algébrique de la valeur propre de A. Plus généralement, même si A
n’est pas intègre, il est parfois utile de l’envoyer dans un corps K par un
homomorphisme d’anneaux unitaires ϕ : A → K. Les valeurs propres de
la matrice A dans K seront alors par définition les zéros du polynôme
Xn
def
ϕ(PA ) =
ϕ(pi )X n−i ∈ K[X] .
i=1
m
= a0 X + a1 X m−1 + · · · + am−1 X + am ∈ L[X].
f (A) = a0 Am + a1 Am−1 + · · · + am−1 A + am In ∈
Soit un polynôme f
Considérons la matrice
Ln×n et les polynômes caractéristiques PA et Pf (A) de A et de f (A).
Le lemme suivant exprime alors en particulier le lien entre les valeurs
propres de la matrice A ∈ An×n et celles de la matrice f (A), dans le
cas où l’anneau de base A est intègre.
Lemme 1.2.2 Soit A un anneau intègre, L une extension du corps
des fractions de A, et f un polynôme de L[X]. Si le polynôme caractéristique de A s’écrit :
PA (X) = (−1)n (X − λ1 ) (X − λ2 ) · · · (X − λn )
avec les λi ∈ L, alors le polynôme caractéristique de f (A) s’écrit :
Pf (A) = (−1)n (X − f (λ1 )) (X − f (λ2 )) · · · (X − f (λn )).
Pn
k
En
particulier,
Tr(f
(A))
=
i=1 f (λi ) et pour tout k ∈ N, Tr(A ) =
Pn
k
i=1 λi .
Preuve. Il suffit de montrer le premier point. Dans le corps L, qui
contient toutes les valeurs propres de la matrice A, celle-ci peut être
ramenée à une forme triangulaire. C’est-à-dire qu’il existe une matrice triangulaire A0 ∈ Ln×n avec les λi sur la diagonale et une matrice
M ∈ Ln×n inversible telles que A0 = M −1 AM . Comme f est un polynôme, et que A0 est triangulaire de la forme
λ1 0 · · · 0
.
.. ..
.
. ..
×
0
A = . .
.
.
.
.
.
.
. 0
× · · · × λn
1.2. Polynôme caractéristique
15
la matrice f (A0 ) sera aussi triangulaire, de la forme
f (λ1 ) 0 · · ·
0
..
.. ..
.
.
×
.
f (A0 ) = .
.
.. ..
..
.
.
0
×
· · · × f (λn )
De plus, puisque ΘM : A 7−→ M −1 A M est un automorphisme 7 de
n×n , on a f (A0 ) = M −1 f (A)M , et par suite P
l’anneau
f (A) = Pf (A0 ) =
QnL
n
(−1)
(X
−
f
(λ
)).
t
u
i
i=1
En fait, il n’est pas nécessaire que les deux matrices A et A0 soient
semblables pour conclure que Pf (A) = Pf (A0 ) . Il suffit pour cela que
PA = PA0 , comme l’indique le résultat suivant, valable dans un anneau
commutatif arbitraire A.
Propriété 1.2.3 Soient A et A0 deux matrices carrées à coefficients
dans A ayant même polynôme caractéristique PA = PA0 . Alors, pour
tout f ∈ A[X] on a : Pf (A) = Pf (A0 ) .
Démonstration.
Soit B la matrice compagnon du polynôme unitaire (−1)n PA (X) =
X n − (c1 X n−1 + · · · + cn−1 X + cn ), c’est-à-dire la matrice
0 · · · · · · 0 cn
1 0 · · · 0 cn−1
.
..
.
.
n×n
B = 0 . . . . ..
.
. ∈A
. .
.
.
..
.. .. 0
..
0
···
0
1
c1
On vérifie sans difficulté que PA (X) = PB (X). Il suffit de montrer que
Pf (A) = Pf (B) pour conclure (puisqu’alors on a aussi Pf (A0 ) = Pf (B) ).
On peut écrire
Pf (B) = Qn (c1 , . . . , cn , X) = (−1)n X n +
Xn−1
i=0
qn,i (c1 , . . . , cn ) X i .
7. Précisément, B = Ln×n est une L - algèbre, c’est-à-dire un anneau muni d’une
loi externe (x, A) 7→ x.A (produit d’une matrice par un scalaire) qui vérifie les identités (x+y).A = x.A+y.A, x.(A+B) = x.A+x.B, x.(y.A) = xy.A, x.(AB) = (x.A)B
et 1.A = A. Et ΘM est un automorphisme de cette structure : un homomorphisme
bijectif d’anneau qui vérifie en plus ΘM (x.A) = x.ΘM (A). On en déduit que pour
tout polynôme f ∈ A[X] on a : ΘM (f (A)) = f (ΘM (A)).
16
1. Rappels d’algèbre linéaire
Il s’agit alors de montrer que
n
n
Pf (A) = (−1) X +
n−1
X
qn,i (c1 , c2 , . . . , cn )X i .
i=0
Si l’on considère les coefficients de f et les entrées de A comme indéterminées x1 , . . . , x` sur Z (on a ` = n2 + 1 + deg f ), établir l’égalité précédente revient à démontrer n identités algébriques dans Z[x1 , . . . , x` ].
Ces identités algébriques sont ensuite valables dans tout anneau commutatif A, en remplaçant les variables formelles xi par des éléments ξi
de A et elles donnent le résultat souhaité.
Or, pour démontrer ces identités algébriques, il suffit de les vérifier sur
un ouvert U de C` , c’est-à-dire lorsqu’on substitue à (x1 , . . . , x` ) un
élément (ξ1 , . . . , ξ` ) arbitraire de U .
Pour cela, on considère par exemple l’ouvert correspondant à des matrices (( suffisamment proches )) de la matrice diagonale suivante :
1
0
.
..
0
0
2
..
.
···
···
..
.
..
.
0
0
..
.
.
0
n
Ces matrices sont diagonalisables puisque leurs valeurs propres restent
distinctes dans l’ouvert considéré. Dans ce cas-là, A et B sont diagonalisables avec les mêmes valeurs propres, et le résultat est trivial. t
u
1.3
Polynôme minimal
Soit K un corps, E un K – espace vectoriel de dimension n > 0
et ϕ : E −→ E un opérateur linéaire, représenté en général par une
matrice A dans une base donnée de E.
1.3.1
Polynôme minimal et sous-espaces de Krylov
Le polynôme minimal de ϕ (ou de A) est par définition le polynôme
unitaire Pϕ ∈ K[X] engendrant l’idéal des polynômes f ∈ K[X] tels
que f (ϕ) = 0. A priori, ce polynôme peut être calculé par les méthodes
usuelles d’algèbre linéaire en cherchant la première relation de dépendance linéaire entre les (( vecteurs )) successifs : IdE , ϕ, ϕ2 , ϕ3 . . . , de
l’espace vectoriel End(E) ' Kn×n des endomorphismes de E.
1.3. Polynôme minimal
Pϕ
17
Puisque Pϕ (ϕ) = 0 (théorème de Cayley-Hamilton), on obtient que
divise Pϕ .
Par ailleurs, étant donné un vecteur v ∈ E, le sous-espace ϕ-engendré
par v (on dit aussi sous-espace de Krylov pour le couple (ϕ, v) ) est, par
définition, le sous-espace de E engendré par le système de vecteurs
(ϕi (v)) i∈N . On le note Kr(ϕ, v).
Sa dimension n’est autre que le degré du polynôme unitaire Pϕ,v ∈
K[X] qui engendre l’idéal {f ∈ K[X] | f (ϕ)(v) = 0} de K[X]. Ce polynôme s’appelle le polynôme ϕ-minimal de v. Les deux polynômes Pϕ
et Pϕ appartiennent évidemment à cet idéal et on a :
Kr(ϕ, v) = E ⇐⇒ do Pϕ,v = n ⇐⇒ Pϕ,v = Pϕ = (−1)n Pϕ .
Ainsi, l’existence d’un vecteur v qui ϕ-engendre E suffit pour que le
polynôme caractéristique de l’endomorphisme ϕ de E soit égal (à un
signe près) à son polynôme minimal :
∃v ∈ E, do Pϕ,v = n =⇒ Pϕ = (−1)n Pϕ .
Mais la réciproque est aussi vraie, comme nous allons le voir bientôt.
Remarquons que, comme le polynôme Pϕ , le polynôme Pϕ,v peut
être calculé par les méthodes usuelles d’algèbre linéaire.
Rappelons maintenant la propriété classique suivante relative à la
décomposition de E en sous espaces ϕ-stables :
Propriété 1.3.1 Soit ϕ un endomorphisme de E et f ∈ K[X] tel que
f = f1 f2 · · · fr avec pgcd(fi , fj ) = 1 pour i 6= j et f (ϕ) = 0. Posons
gi = f /fi , θi = fi (ϕ), ψi = gi (ϕ), Ei = Ker θi . Chaque Ei est un
sous-espace ϕ -stable, on note ϕi : Ei → Ei la restriction de ϕ. Alors :
a) E = E1 ⊕ · · · ⊕ Er .
b) Ei = Im ψi , et la restriction de ψi à Ei induit un automorphisme
de l’espace vectoriel Ei .Q
Q
c) On a les égalités Pϕ = ri=1 Pϕi et Pϕ = ri=1 Pϕi .
d) Si (v1 ,Q
. . . , vr ) ∈ E1 × · · · ×Er , et v = v1 + · · · + vr , on a l’égalité
Pϕ,v = ri=1 Pϕi ,vi .
Preuve. Tout d’abord on remarque que deux endomorphismes de la
forme a(ϕ) et b(ϕ) commutent toujours puisque a(ϕ) ◦ b(ϕ) = (ab)(ϕ).
Ensuite il est clair que tout sous-espace du type Ker a(ϕ) ou Im a(ϕ)
est ϕ-stable.
18
1. Rappels d’algèbre linéaire
La preuve des points a) et b) est alors basée sur les égalités ϕi ◦ ψi =
ψi ◦ ϕi = f (ϕ) = 0E et sur l’identité de Bézout u1 g1 + · · · + ur gr = 1,
qui implique u1 (ϕ) ◦ g1 (ϕ) + · · · + ur (ϕ) ◦ gr (ϕ) = IdE , ce qui se lit :
ψ1 ◦ α1 + · · · + ψr ◦ αr = IdE , où αi = ui (ϕ).
Si on choisit dans chaque Ei une base Bi leur réunion est une base
B de E et la matrice de ϕ sur B est diagonale parQblocs, chaque bloc
étant la matrice de ϕi sur Bi . Ceci implique Pϕ = ri=1 Pϕi .
Soit hi = P ϕi . Il est clair que (h1 · · · hr )(ϕ) = 0E (car nul sur chaque
Ei ), et que les hi sont premiers entre eux 2 à 2 (car hi divise fi ). Si
g(ϕ) = 0, a fortiori g(ϕi ) = 0, donc g est multiple des hi , et par suite
multiple de h = h1 · · · hr . Ceci termine la preuve du point c).
Et la preuve du point d) est analogue, en remplaçant hi = P ϕi par
ϕ
i
P ,vi .
t
u
Propriété 1.3.2 Soit Pϕ = P1m1 P2m1 . . . Prmr la décomposition du polynôme minimal de ϕ en facteurs premiers distincts P1 , P2 , . . ., Pr .
On reprend les notations de la propriété 1.3.1 avec fi = Pimi . On pose
en outre Qi = Pϕ /Pi , Gi = Ker(Qi (ϕ)), Fi = Ker(Pimi −1 (ϕ)), pour
1 ≤ i ≤ r. Alors E = E1 ⊕ · · · ⊕ Er , chaque Fi est strictement inclus
dans Ei et pour tout v = v1 + · · · + vr (vi ∈ Ei ) on a les équivalences
suivantes :
v∈
/
r
[
i=1
Gi ⇐⇒ P
ϕ,v
=P
ϕ
⇐⇒
r
^
i=1
P
ϕi ,vi
=P
ϕi
⇐⇒
r
^
vi ∈
/ Fi .
i=1
Preuve. Le fait que E = E1 ⊕· · ·⊕Er résulte de la propriété 1.3.1 a). La
deuxième équivalence résulte des propriétés 1.3.1 c) et d). La première
équivalence est claire : la première condition signifie exactement que le
polynôme ϕ-minimal de v ne divise pas strictement le polynôme minimal de ϕ. Même chose pour la dernière équivalence (tout diviseur strict
de Pimi divise Pimi −1 ). Cette remarque montre aussi que l’inclusion
Fi ⊂ Ei est stricte.
t
u
Corollaire 1.3.3 Il existe toujours un vecteur v tel que Pϕ,v = Pϕ . En
particulier, si le polynôme minimal de ϕ a pour degré la dimension de E
(autrement dit si, au signe près, il est égal au polynôme caractéristique)
il existe des vecteurs qui ϕ-engendrent E.
Preuve. Il suffit de prendre v = v1 + · · · + vr avec chaque vi dans
Ei \ Fi .
t
u
1.3. Polynôme minimal
19
Ainsi, sauf exception, c’est-à-dire si l’on choisit v en dehors de la
réunion d’un petit nombre de sous-espaces vectoriels stricts de E, les
Ker(Qi (ϕ)), (cette réunion ne remplit jamais l’espace, même si le corps
de base est fini), le vecteur v convient.
Notons que la preuve précédente est peu satisfaisante parce qu’en
général on ne sait pas calculer la décomposition en facteurs premiers
d’un polynôme. Il s’ensuit que notre preuve de l’existence d’un vecteur
qui ϕ-engendre E reste plus théorique que pratique.
Voici une manière de contourner cet obstacle.
Premièrement on établit le lemme suivant.
Lemme 1.3.4 Si on connaı̂t un polynôme Q qui est un facteur strict
d’un polynôme P de K[X] on peut, ou bien décomposer P en un produit P1 P2 de deux polynômes étrangers, ou bien écrire P et Q sous
la forme P1k et P1` (1 ≤ ` < k).
La procédure (que nous ne donnons pas en détail) consiste à partir
de la factorisation intitiale P = QQ1 et à la raffiner au maximum en
utilisant les pgcd’s.
Ensuite on rappelle que les polynômes Pϕ,v et Pϕ peuvent facilement
être calculés par les méthodes classiques d’algèbre linéaire.
On démarre alors avec un v non nul arbitraire. Si Pϕ,v est égal à Pϕ
on a terminé. Sinon, on applique le lemme précédent avec P = Pϕ et
Q = Pϕ,v . Dans le premier cas, on applique la propriété 1.3.1 c) et d)
avec les polynômes P1 et P2 . On est ramené à résoudre le même problème séparément dans deux espaces de dimensions plus petites. Dans
le deuxième cas, on choisit un nouveau v dans le complémentaire de
Ker(P1k−1 (ϕ)), ce qui fait que le degré de son polynôme minimal augmente strictement.
Lk
Remarque. Toute décomposition E =
i=1 Ei en sous-espaces Ei
ϕ -stables donne une forme réduite de ϕ, c’est-à-dire la représentation
de ϕ dans une base convenable par une matrice diagonale par blocs de
la forme :
A1 0 · · · 0
..
.
.
0 A2 . .
A= .
.
.
.
..
.. 0
..
0 · · · 0 Ak
où A1 , A2 , . . . , Ak sont les matrices des endomorphismes ϕi induits
par ϕ dans les sous-espaces Ei . Certaines de ces formes réduites sont
20
1. Rappels d’algèbre linéaire
dites canoniques, comme la réduction de Jordan dans le cas où le polynôme caractéristique se factorise en facteurs linéaires sur K. Il existe
d’autres formes réduites canoniques entièrement rationnelles, c’est-à-dire
qui n’utilisent pour le changement de base que des expressions rationnelles en les coefficients de la matrice de départ. Sur le sujet des formes
normales réduites (et sur bien d’autres) nous recommandons le livre
de Gantmacher ([Gan]) dont on attend toujours la réédition à un prix
abordable.
1.3.2
Cas de matrices à coefficients dans Z.
Dans certains algorithmes que nous aurons à développer par la suite,
nous partirons d’une matrice C à coefficients dans un anneau intègre A
et bien souvent il sera avantageux qu’aucun des calculs intermédiaires
ne produise des éléments qui seraient dans le corps des fractions de A
sans être dans A.
Se pose alors naturellement la question suivante : les polynômes PC
et PC,v que nous pouvons être amenés à envisager comme résultats de
calculs intermédiaires en vue de trouver le polynôme caractéristique de
C, sont-ils toujours des polynômes à coefficients dans A ?
Dans le cas de matrices carrées à coefficients dans Z, ou dans un
anneau de polynômes Z[x1 , . . . , xn ] ou Q[x1 , . . . , xn ], la réponse est positive.
Ce résultat n’est pas évident. Il est basé sur les définitions et les propriétés qui suivent, pour lesquelles on peut consulter les livres classiques
d’algèbre (par exemple [Gob] ou [MRR]).
Définition 1.3.5 Un anneau intègre A est dit intégralement clos si
tout diviseur unitaire dans FA [X] d’un polynôme unitaire de A[X] est
dans A[X].
Avec de tels anneaux les polynômes PC et PC,v sont donc automatiquement à coefficients dans A.
Définition 1.3.6 Un anneau intègre A est dit anneau à pgcds si tout
couple d’éléments (a, b) admet un pgcd, c’est-à-dire un élément g ∈ A
tel que :
∀x ∈ A ((x divise a et b) ⇐⇒ x divise g) .
Propriété 1.3.7 Pour qu’un anneau intègre soit intégralement clos, il
suffit que la propriété qui le définit soit vérifiée pour les diviseurs de
1.4. Suites récurrentes linéaires
21
degré 1. Autrement dit, tout zéro dans FA d’un polynôme unitaire de
A[X] est dans A.
Propriété 1.3.8 Tout anneau à pgcds est intégralement clos.
Propriété 1.3.9 Si A est un anneau à pgcds il en va de même pour
A[x1 , . . . , xn ].
1.4
1.4.1
Suites récurrentes linéaires
Polynôme générateur, opérateur de décalage
Soit E un K – espace vectoriel (resp. un A – module). On considère
une suite (an )n∈N d’éléments de E et un entier p ∈ N. Une relation de
récurrence linéaire d’ordre p pour cette suite est définie par la donnée
de p + 1 éléments c0 , c1 , . . . , cp de K (resp. de A) vérifiant :
∀n ∈ N
c0 an + c1 an+1 + · · · + cp an+p = 0
(1.16)
Pp
i
Le polynôme h(X) =
i=0 ci X dans K[X] (resp. dans A[X]) est
appelé un polynôme générateur de la suite (an ). Lorsque le coefficient
cp est inversible, la suite est alors déterminée par la donnée de a0 ,
a1 , . . . , ap−1 car elle peut ensuite être construite par récurrence : elle
est en quelque sorte (( engendrée )) par le polynôme h, ce qui justifie la
terminologie adoptée. Une suite récurrente linéaire dans E est une suite
(an )n∈N d’éléments de E qui possède un polynôme générateur dont le
coefficient dominant est inversible.
On interprète cette situation de la manière suivante en algèbre linéaire. On appelle S le K – espace vectoriel (resp. le A – module) E N
formé de toutes les suites (un )n∈N à valeurs dans E. On note Φ : S → S
l’opérateur de décalage qui donne pour image de (un )n∈N la suite décalée
d’un cran (un+1 )n∈N . Il est clair que l’opérateur de décalage est un
opérateur linéaire. Dire que la suite a = (an )n∈N vérifie la relation de
récurrence linéaire (1.16) se traduit exactement, en langage un peu plus
abstrait, par :
a ∈ Ker(h(Φ)) .
Cela montre que les polynômes générateurs d’une suite récurrente linéaire donnée forment un idéal de K[X] (resp. de A[X]). Dans le cas d’un
corps et d’une suite récurrente linéaire, comme K[X] est un anneau
principal, cet idéal (non nul) est engendré par un polynôme unitaire
22
1. Rappels d’algèbre linéaire
unique qu’on appelle le polynôme générateur minimal ou simplement le
polynôme minimal de la suite. Nous le noterons Pa . Nous allons voir
plus loin que ce polynôme peut effectivement être calculé, dès qu’on
connaı̂t un polynôme générateur de la suite récurrente linéaire.
Considérons maintenant un polynôme unitaire fixé
p
f (X) = X −
p−1
X
bi X i
i=0
dans K[X] (resp. dans A[X]). Le K – espace vectoriel (resp. le A – module) Ker(f (Φ)) formé des suites récurrentes linéaires dans E pour
lesquelles f est un polynôme générateur sera noté Sf . Il est isomorphe
à Kp (resp. Ap ) et une base canonique est fournie par les p suites 8
e(i) (i = 0, . . . , p − 1) telles que e(i) (j) = δij pour j = 0, . . . , p − 1, où
δij est le symbole de Kronecker (δij = 1 si i = j et δij = 0 si i 6= j).
Pour une suite récurrente linéaire arbitraire a = (an )n∈N dans Sf on a
P
(j)
alors : a = p−1
j=0 aj e .
Il est clair que Sf est stable par Φ. Notons Φf la restriction de
Φ à Sf . On constate immédiatement que la matrice de Φf sur la base
canonique (e(p−1) , . . . , e(1) , e(0) ) est la matrice compagnon du polynôme f :
0 · · · · · · 0 b0
0 ··· 0
b0
1 0 · · · 0 b1
b1
..
..
.. ..
=
.
.
Cf =
0
.
. .
.
.. . .
Ip−1
.
..
.. ... 0
..
bp−1
0 · · · 0 1 bp−1
(1)
En particulier PΦf ,e = PΦf = PΦf = f . En outre, par simple application des définitions on obtient PΦf ,a = Pa .
Comme exemples importants de suites récurrentes linéaires, on peut
citer :
– la suite récurrente linéaire formée des puissances (An )n∈N d’une
matrice A ∈ Km×m dont le polynôme générateur minimal n’est autre
que le polynôme minimal PA de la matrice ;
– pour des vecteurs donnés u, v ∈ Km×1 , les suites récurrentes linéaires (An v)n∈N et ( t u An v)n∈N dont les polynômes minimaux, notés
A,v
respectivement PA,v et PA,v
divise PA,v et
u , sont alors tels que : Pu
PA,v divise PA .
8. Chacune de ces suites est évidemment définie par ses p premiers termes.
1.4. Suites récurrentes linéaires
1.4.2
23
Suites récurrentes linéaires et matrices de Hankel
Comme le montre la discussion qui suit, l’étude des suites récurrentes
linéaires est étroitement liée à celle des matrices de Hankel.
Une matrice de Hankel est une matrice (pas nécessairement carrée)
H = (vij ) dont les coefficients sont constants sur les diagonales montantes : vij = vhk si i + j = h + k.
Les matrices de Hankel fournissent un exemple de matrices structurées. L’autre exemple le plus important est celui des matrices de
Toeplitz, celles dont les coefficients sont constants sur les diagonales descendantes : vij = vhk si i − j = h − k.
Remarquons qu’une matrice de Hankel carrée d’ordre n est une matrice symétrique et que les produits H Jn et Jn H d’une matrice de
Hankel H carrée d’ordre n par la matrice de Hankel particulière Jn :
Jn =
0 0 ···
0 0 ···
.. ..
. .
0 1 ···
1 0 ···
0 1
1 0
.. ..
. .
0 0
0 0
sont des matrices de Toeplitz. Cette matrice de permutation d’ordre n
permet de renverser l’ordre des n colonnes (resp. des n lignes) d’une
matrice lorsque celle-ci est multipliée à droite (resp. à gauche) par la
matrice Jn : c’est pourquoi on l’appelle matrice de renversement ou
encore matrice d’arabisation du fait qu’elle permet d’écrire de droite à
gauche les colonnes que l’on lit de gauche à droite et inversement.
Inversement, les produits Jn T et T Jn d’une matrice de Toeplitz T
carrée d’ordre n par la matrice Jn sont des matrices de Hankel.
Une matrice structurée est déterminée par la donnée de beaucoup
moins de coefficients qu’une matrice ordinaire de même taille. Par exemple une matrice de Hankel (resp. de Toeplitz) de type (n, p) est déterminée par la donnée de n + p − 1 coefficients : ceux des première
ligne et dernière (resp. première) colonne. Cela rend ces matrices particulièrement importantes pour les (( grands calculs )) d’algèbre linéaire.
Si a = (an )n∈N est une suite arbitraire et si i, r, p ∈ N nous noterons
24
1. Rappels d’algèbre linéaire
a
H i,r,p la matrice de Hankel suivante, qui possède r lignes et p colonnes :
ai
ai+1 ai+2 . . . ai+p−1
ai+1 ai+2
ai+p
a
H i,r,p = ai+2
.
..
..
.
.
ai+r−1 ai+r . . . . . . ai+r+p−2
Le fait suivant est une simple constatation.
Fait 1.4.1 Reprenant les notations ci-dessus, une suite a est une suite
récurrente linéaire avec f comme polynôme générateur si et seulement
si sont vérifiées pour tous i, r ∈ N les équations matricielles
a
a
H i,r,p Cf = H i+1,r,p
(1.17)
a
a
Ou ce qui revient au même, en transposant, t Cf H i,p,r = H i+1,p,r . Naturellement, il suffit que ces équations soient vérifées lorsque r = 1.
On en déduit
a
a
H i,r,p (Cf )k = H i+k,r,p
(1.18)
et donc on a aussi :
a
Fait 1.4.2 Sous les mêmes hypothèses, dans toute matrice H i,r,p+k les
k dernières colonnes sont combinaisons linéaires des p premières. Et,
a
par transposition, dans toute matrice H i,p+k,s les k dernières lignes
sont combinaisons linéaires des p premières.
On en déduit la proposition suivante.
Proposition 1.4.3 Avec les notations ci-dessus, et dans le cas d’un
corps K, si a est une suite récurrente linéaire qui admet f pour polynôme générateur, le degré d de son polynôme générateur minimal Pa est
a
égal au rang de la matrice de Hankel H 0,p,p . Les coefficients de Pa (X) =
P
i
X d − d−1
i=0 gi X ∈ K[X] sont l’unique solution de l’équation
a
a
H 0,d,d CPa = H d,d,1
c’est-à-dire encore l’unique solution du système linéaire
a0
a1 a2 · · · ad−1
ad
g0
.
.
a1
.
a
a
2
d
g1 ad+1
..
. ..
g2 ad+2
.
a2
.
.
.
=
.. ..
..
.
.
.
..
. .
.
.. ..
a2d−1
gd−1
ad−1 ad · · · · · · a2d−2
(1.19)
.
1.5. Polynômes symétriques et relations de Newton
25
Preuve. Considérons la première relation de dépendance linéaire entre
a
les colonnes de H 0,p,p , et appelons g le polynôme unitaire correspondant, qui vérifie
a
a
t
Cg H 0,d,p = H 1,d,p .
Cela donne l’équation
récurrence sur k :
tC
a
g
a
H 0,d,p = H 0,d,p Cf d’où immédiatement par
a
a
a
( t Cg )k H 0,d,p = H 0,d,p (Cf )k = H k,d,p ,
et donc
t
a
a
a
a
Cg H k,d,p = H k+1,d,p .
A fortiori, pour tout k on a :
t
Cg H k,d,1 = H k+1,d,1 ,
et cela signifie que g est également un polynôme générateur pour la
suite a.
On laisse le soin à la lectrice et au lecteur de finir la preuve.
t
u
1.5
Polynômes symétriques et relations de Newton
Soit A un anneau commutatif unitaire et A [x1 , . . . , xn ] l’algèbre
sur A des polynômes à n indéterminées x1 , . . . , xn .
Tout polynôme f ∈ A [x1 , . . . , xn ] s’écrit de manière unique comme
une somme finie de monômes distincts aJ xJ = aJ xj11 · · · xjnn où J =
(j1 , . . . , jn ) ∈ Nn et aJ ∈ A :
X
f=
aJ xj11 · · · xjnn
J
où la somme porte sur une partie finie de Nn . On a souvent intérêt
à donner un bon ordre sur les termes xJ = xj11 · · · xjnn pour faire des
preuves par induction. Il suffit par exemple d’ordonner les indéterminées
(x1 < x2 < · · · < xn ) pour définir un bon ordre sur les termes xJ , par
exemple l’ordre lexicographique, ou l’ordre lexicographique subordonné
au degré total.
On écrit souvent aussi le polynôme comme somme de ses composantes homogènes
Xn
f=
fh ,
h=0
26
1. Rappels d’algèbre linéaire
où n est le degré total de f et fh la composante homogène de degré
h de f . Une manière simple de voir les composantes homogènes d’un
polynôme f ∈ A [x1 , . . . , xn ] est de considérer une nouvelle indéterminée z et le polynôme en z : g(z) = f (x1 z, . . . , xn z) ∈ A [x1 , . . . , xn ] [z].
La composante homogène fh n’est autre que le coefficient de z h dans
g(z).
Désignant par Sn le groupe des permutations de {1, 2, . . . , n}, un
polynôme f ∈ A [x1 , . . . , xn ] est dit symétrique si son stabilisateur par
l’action
τ
f = f (x1 , . . . , xn ) 7−→ τ f = f (xτ (1) , . . . , xτ (n) )
de Sn sur A [x1 , . . . , xn ] est le groupe Sn tout entier. C’est-à-dire encore si les monômes d’une même orbite de Sn figurent dans l’expression
de f avec le même coefficient.
Des polynômes symétriques importants sont les sommes de Newton
à n indéterminées :
Xn
Sk (x1 , . . . , xn ) =
xki ∈ A [x1 , . . . , xn ] (k ∈ N)
(1.20)
i=1
On notera A [x1 , . . . , xn ]sym l’ensemble des polynômes symétriques
en x1 , . . . , xn sur A. C’est une sous-algèbre propre de A [x1 , . . . , xn ]. Il
est bien connu (la démonstration peut se faire par récurrence en utilisant l’ordre lexicographique) que tout polynôme symétrique s’exprime de
manière unique comme polynôme en σ1 , σ2 , . . . , σn où les σp (1 ≤ p ≤ n)
sont les polynômes symétriques élémentaires en x1 , x2 , . . . , xn :
X
σp =
x i1 x i2 · · · x ip ·
1≤i1 <i2 <···<ip ≤n
Cela signifie que l’homomorphisme de A –algèbres
ϕ : A [y1 , . . . , yn ] −→ A [x1 , . . . , xn ]sym
défini par
ϕ(f (y1 , . . . , yn )) := f (σ1 , σ2 , . . . , σn )
est un isomorphisme.
En outre, lorsque A est intègre, FA désignant le corps des fractions
de A, cet isomorphisme ϕ se prolonge de manière unique en un isomorphisme de FA – algèbres, de FA (y1 , . . . , yn ) vers FA (x1 , . . . , xn )sym , qui
est par définition la sous-algèbre de FA (x1 , . . . , xn ) formée des fractions
rationnelles invariantes par permutation des variables. Autrement dit,
1.5. Polynômes symétriques et relations de Newton
27
toute fraction rationnelle sur FA , symétrique en x1 , . . . , xn , s’écrit de
manière unique comme une fraction rationnelle en σ1 , . . . , σn sur FA . On
exprime ce fait en disant que (σ1 , . . . , σn ) est un système fondamental
de polynômes symétriques en x1 , . . . , xn sur le corps FA . Plus généralement :
Définition 1.5.1 Étant donné un anneau commutatif unitaire A (resp.
un corps K), et n indéterminées x1 , . . . , xn sur cet anneau (resp. ce
corps), on appelle système fondamental de polynômes symétriques en
x1 , . . . , xn sur l’anneau A (resp. fractions rationnelles symétriques en
x1 , . . . , xn sur le corps K) tout système (f1 , . . . , fn ) de n éléments de
A [x1 , . . . , xn ]sym (resp. de K(x1 , . . . , xn )sym ) vérifiant A [x1 , . . . , xn ]sym
= A [f1 , . . . , fn ] (resp. K(x1 , . . . , xn )sym = K(f1 , . . . , fn )).
Attention à l’ambiguité de langage : un système fondamental sur le
corps K n’est pas nécessairement un système fondamental sur l’anneau
K, même s’il est formé de polynômes. Par contre, un système fondamental sur l’anneau intègre A est toujours un système fondamental sur le
corps FA .
La définition d’un système fondamental sur un corps K implique
l’indépendance algébrique du système (f1 , f2 , . . . , fn ) et garantit l’unicité de l’expression rationnelle, dans ce système fondamental, de toute
fraction rationnelle symétrique sur K.
Les relations dites de Newton permettent d’exprimer les sommes de
Newton dans le système fondamental des polynômes symétriques élémentaires.
Proposition 1.5.2 (Relations de Newton) Les polynômes de Newton
à n indéterminées (Sk )k∈N sont reliés aux n polynômes symétriques
élémentaires (σk )1≤k≤n par les relations suivantes :
(i) S0 = n ;
P
i
k
(ii) pour 1 ≤ k ≤ n : SkP
+ k−1
i=1 (−1) σi Sk−i + (−1) k σk = 0 ;
n
i
(iii) pour k > n : Sk + i=1 (−1) σi Sk−i = 0.
Preuve. On pose αk = (−1)k+1 σk (k = 1 . . . , n). Les polynômes symétriques élémentaires apparaissent dans le développement du polynôme
Q(X) =
n
Y
i=1
(1 − xi X) = 1 −
n
X
k=1
αk X k .
28
1. Rappels d’algèbre linéaire
Par dérivation logarithmique formelle, on obtient les égalités suivantes
dans l’algèbre de séries formelles FA (x1 , . . . , xn ) [[X, X −1 ]] :
n
n
∞
Q0 (X) X −xi
1
1 X
1 X
=
1−
Sk X k
=−
=
Q(X)
1 − xi X
X
1 − xi X
X
i=1
i=1
k=1
ou encore dans FA (x1 , . . . , xn ) [[X]] :
0
− X Q (X) = Q(X)
∞
X
Sk X k
(1.21)
k=1
avec Q(X) = 1−[α1 X + . . . + αn X n ] et −XQ0 (X) = α1 X+2 α2 X 2 +
. . . + n αn X n . Identifiant dans l’équation (1.21) les termes en X k pour
1 ≤ k ≤ n, on obtient les formules (ii). Les formules (iii) sont obtenues
par identification des termes de degré supérieur à n.
Remarque. En notant αi = (−1)i+1 σi les relations de Newton s’écrivent
sous la forme matricielle suivante :
α1
S1
1
0
... ...
0
..
..
S1
S2
α
.
2
.
2
..
.. ..
..
..
..
..
.
.
.
.
.
.
.
.
.
..
..
Sn−2
.
.
n
−
1
0
.
(1.22)
= S
αn
S1
n
Sn−1 Sn−2
n
Sn Sn−1
S1
Sn+1
..
..
..
.
.
.
Sn+k
Sk+1
Sn+k+1
..
..
..
..
..
..
.
.
.
.
.
.
Les relations (ii) (qui correspondent aux n premières lignes dans la
matrice infinie ci-dessus) et (iii) (qui correspondent aux lignes no n et
suivantes) donnent la même formule si l’on fait k = n. D’autre part,
les relations (iii) peuvent être obtenues directement : en multipliant par
X k−n (k > n) le polynôme
P (X) =
n
Y
n
(X − xi ) = X −
i=1
on obtient
X k−n P (X) = X k −
n
X
αi X n−i ,
i=1
Xn
i=1
αi X k−i .
1.5. Polynômes symétriques et relations de Newton
29
Lorsque k > n les identités xk−n
P (xj ) = 0 donnent alors par sommaj
tion :
n
X
j=1
xkj −
n
X
i=1
αi
hXn
j=1
n
i
X
xk−i
=
S
−
αi Sk−i = 0 .
k
j
i=1
Corollaire 1.5.3 Si A est un anneau commutatif où les entiers 1, 2,
. . . , n sont inversibles, alors les sommes de Newton (Sk )1≤k≤n en les
n indéterminées x1 , . . . , xn forment un système fondamental de polynômes symétriques en x1 , . . . , xn sur l’anneau A.
Preuve. Le système triangulaire formé par les n premières équations
dans (1.22) admet clairement une solution unique en les αi .
t
u
Corollaire 1.5.4 Soit P = X n − a1 X n−1 + . . . + an−1 X + an un polynôme unitaire à une indéterminée sur un anneau intègre A, et soient
λ1 , λ2 , . . . , λn les n racines de ce polynôme (distinctes ou non) dans
un corps
P de décomposition de P (une extension de FA ). Si l’on pose
sk = ni=1 λki pour 1 ≤ k ≤ n, on a les relations :
s 1 = a1
s2 = s1 a1 + 2 a2
(1.23)
.. ..
..
.
.
.
sn = sn−1 a1 + . . . + s1 an−1 + n an
Ainsi les coefficients du polynôme P sont déterminés de manière
unique dans A par la donnée de ses sommes de Newton si n ! est non
diviseur de zéro dans A. Et si la division exacte par chacun des entiers
2, . . . , n est explicite lorsqu’elle est possible, le calcul des si à partir des
ai est lui aussi explicite.
Dans la définition suivante, on généralise les sommes de Newton pour
les zéros du polynôme P au cas d’un anneau commutatif arbitraire.
Définition 1.5.5 Soit P = X n − a1 X n−1 + . . . + an−1 X + an un polynôme unitaire à une indéterminée sur un anneau commutatif A. Alors
les éléments si donnés par les équations (1.23) s’appellent les sommes
de Newton de P (ou de tout polynôme u P avec u non diviseur de zéro
dans A).
Un fait important est le suivant.
30
1. Rappels d’algèbre linéaire
Lemme 1.5.6 Soit A un anneau commutatif arbitraire, A une matrice carrée d’ordre n sur A et PA son polynôme caractéristique. Les
sommes de Newton de PA données dans la définition précédente sont
les traces des puissances de la matrice A : sk = Tr(Ak ) pour 1 ≤ k ≤ n.
Preuve. On remarque que les égalités à démontrer sont des identités
algébriques en les entrées de la matrice A (considérées comme des indéterminées). Il suffit donc de traiter le cas d’un anneau intègre. Dans
ce cas, le résultat est donné par le corollaire 1.5.4 et le lemme 1.2.2. t
u
1.6
Inégalité de Hadamard et calcul modulaire
Nous décrivons d’abord quelques majorations utiles en algèbre linéaire.
1.6.1
Normes matricielles
Si A = (aij ) ∈ Am×n est une matrice à coefficients réels ou complexes, on définit classiquement les normes suivantes (cf. [Cia], [GL])
hXm
hXn
i
i
k A k1 = max
|aij | , k A k∞ = max
|aij |
1≤j≤n
i=1
k A kF =
1≤i≤m
rX
m
i=1
Xn
j=1
j=1
|aij |2
Chacune de ces normes vérifie les relations classiques
k c A k = |c| k A k ,
kA + B k ≤ kAk + kB k
(si A et B ont mêmes dimensions) et
kA B k ≤ kAk kB k
(si le produit A B est défini).
Considérons maintenant des matrices à coefficients entiers. La taille
d’un entier x est l’espace qu’il occupe lorsqu’on l’implante sur machine.
Si le codage des entiers est standard, cela veut dire que la taille de x
est correctement appréciée par 1 + dlog2 (1 + |x|)e.
Notation 1.6.1 Dans tout cet ouvrage log x dénote max(1, log2 |x|).
1.6. Inégalité de Hadamard et calcul modulaire
31
Lorsque x est entier, cela représente donc la taille de x à une constante près. Si λ(A) = log(k A k) avec l’une des normes précédentes, la
taille de chaque coefficient de A est clairement majorée par λ(A) (à une
constante additive près) et en outre les relations précédentes impliquent
immédiatement que
λ(A B) ≤ λ(A) + λ(B),
λ(A + B) ≤ max(λ(A), λ(B)) + 1.
Ces relations sont souvent utiles pour calculer des majorations de la
taille des entiers qui interviennent comme résultats de calculs matriciels.
L’inégalité de Hadamard
L’inégalité de Hadamard s’applique aux matrices à coefficients réels.
La valeur absolue du déterminant représente le volume (n - dimensionnel)
du parallélépipède construit sur les vecteurs colonnes de la matrice, (( et
donc )) elle est majorée par le produit des longueurs de ces vecteurs :
v
n uX
Y
u n
t
| det((aij )1≤i,j≤n )| ≤
a2ij
(1.24)
j=1
i=1
Il y a évidemment des preuves rigoureuses de ce fait intuitif. Par
exemple le processus d’orthogonalisation de Gram-Schmidt remplace la
matrice par une matrice de même déterminant dont les vecteurs-colonnes
sont deux à deux orthogonaux et ne sont pas plus longs que ceux de
la matrice initiale. La signification géométrique de cette preuve est la
suivante : le processus d’orthogonalisation de Gram-Schmidt remplace
le parallélépipède (construit sur les vecteurs colonnes) par un parallélépipède droit de même volume dont les cotés sont devenus plus courts. Ce
même raisonnement donne l’inégalité dans le cas d’une matrice à coefficients complexes en remplaçant a2ij par |aij |2 (mais l’interprétation
géométrique directe disparaı̂t).
Avec les normes k k1 et k k∞ on obtient pour une matrice carrée :
| det(A)| ≤ (k A k1 )n ,
| det(A)| ≤ (k A k∞ )n
(1.25)
Avec la norme de Frobenius k A kF on obtient la majoration suivante
(un produit de n réels positifs dont la somme est constante est maximum
lorsqu’ils sont tous égaux) :
k A kF n
√
| det(A)| ≤
(1.26)
n
32
1.6.2
1. Rappels d’algèbre linéaire
Le théorème chinois et son application aux calculs
modulaires
Soient p1 , p2 , . . . , pr des entiers positifs deux à deux premiers entre
eux. On pose M = p1 p2 · · · pr . Pour toute suite x1 , x2 , . . . , xr de r
entiers relatifs, il existe un entier x (unique modulo M ) vérifiant :
x ≡ xi [mod pi ] (i = 1, . . . , r). On peut calculer cet entier (modulo M )
en remarquant que, pour tout i compris entre 1 et r, les nombres pi
et qi = M/pi sont premiers entre eux et que par conséquent il existe
des entiers ui et vi (relation de Bézout) tels que P
pi ui + qi vi = 1 (i =
r
1, . . . , r). Le nombre x recherché n’est autre que
i=1 xi qi vi (modulo
M ). Il est facile de vérifier qu’il répond bien à la question.
L’une des conséquences importantes du Théorème chinois en calcul
formel est son utilisation pour le calcul de coefficients entiers x ∈ Z dont
on sait majorer la valeur absolue par un entier B strictement positif.
Il arrive souvent que les calculs intermédiaires, lorsqu’ils sont effectués
(avec ou sans division) dans Z, donnent des coefficients dont la taille
explose rapidement, ce qui risque de rendre ces calculs impraticables ou
trop coûteux, alors que la taille du résultat final est bien plus petite.
Supposons que l’on ait à calculer un x ∈ Z tel que −B ≤ x ≤ B par
un algorithme sans divisions.
On commence par choisir des entiers positifs p1 , p2 , . . . , pr deux à
deux premiers entre eux dont le produit dépasse strictement 2B. Au
lieu de calculer directement x, on effectue tous les calculs modulo pi
séparément pour chaque i (i = 1, . . . , r). Les résultats x1 , x2 , . . . , xr
ainsi obtenus sont tels que x est dans la classe de xi modulo pi (pour
i = 1 . . . r). Utilisant les mêmes notations que ci-dessus pour les coefficients de Bézout relatifs aux couples (pi , qi ), on récupère ensuite le
résultat principal x à partir des résultats partiels x1 , x2 , . . . , xr en remarquant
que x est l’entier relatif de plus petite valeur absolue congru
Pr
à
x
q
i=1 i i vi modulo p1 p2 · · · pr (puisque −B ≤ x ≤ B). Dans le cas
d’un algorithme avec divisions, les facteurs pi doivent être choisis de
manière à ce qu’ils soient premiers avec les diviseurs intervenant dans
les calculs.
Pour le calcul des déterminants de matrices à coefficients entiers,
par exemple, on peut utiliser l’inégalité de Hadamard (1.24) pour faire
fonctionner la méthode modulaire. On prendra pour borne B = M n nn/2
où M = max1≤i,j≤n |aij |. Si cela s’avère préférable, on peut choisir une
des bornes données dans les équations (1.25) et (1.26).
Il en est de même pour le calcul du polynôme caractéristique PA
1.6. Inégalité de Hadamard et calcul modulaire
33
d’une matrice A ∈ Zn×n car chacun des coefficients de PA est une
somme de mineurs diagonaux de la matrice A. On peut donc là aussi
utiliser l’inégalité de Hadamard pour majorer les valeurs absolues des
coefficients en vue du traitement modulaire.
Plus précisément, si l’on prend M = max1≤i,j≤n |aij | comme cidessus et si l’on désigne par mk,j (1 ≤ j ≤ Cnk ) les mineurs de A
diagonaux d’ordre k (pour k donné entre 1 et n), alors le coefficient
µk du terme de degré n − k de PA est majoré en valeur absolue comme
suit :
|µk | =
X
k
n
mk,j ≤ Cnk M k k 2 ≤ (2M )n n 2
(1.27)
k}
j∈{1,...,Cn
puisque Cnk ≤ 2n .
Quelques considérations pratiques
L’idée principale dans l’utilisation du calcul modulaire est de remplacer un algorithme dans Z permettant de résoudre un problème donné
par plusieurs algorithmes modulo des nombres premiers.
Pour être vraiment efficace, cette méthode doit être appliquée avec
des listes de nombres premiers p1 , p2 , . . . , pr déjà répertoriées et pour
lesquelles on a déjà calculé les coefficients qi vi correspondants qui permettent de récupérer x à partir des xi .
Ces nombres premiers peuvent être choisis par rapport à la taille
des mots traités par les processeurs. Par exemple, pour des processeurs
qui traitent des mots à 64 bits, on prend des nombres premiers compris entre 263 et 264 − 1 : il y en a suffisamment (bien plus que 1017
nombres !) pour résoudre dans la pratique tous les problèmes de taille
humainement raisonnable et réaliste [GG]. En outre on possède des tests
rapides pour savoir si un nombre est premier, et cela a permis d’établir
des listes p1 , p2 , . . . , pr avec la liste des coefficients qi vi correspondants,
qui répondent à tous les cas qui se posent en pratique.
Chaque opération arithmétique élémentaire modulo un tel nombre
premier se fait alors en temps constant, ce qui réduit considérablement
le temps de calcul. En outre la décomposition du problème en algorithmes modulaires offre la possibilité d’utiliser plusieurs processeurs en
parallèle.
34
1.7
1. Rappels d’algèbre linéaire
Résolution uniforme des systèmes linéaires
Nous expliquons ici comment le polynôme caractéristique permet de
déterminer le rang d’une matrice et de résoudre uniformément (avec une
seule formule, du type Cramer) les systèmes linéaires ayant un format
donné et un rang fixé. Et ceci sur un corps arbitraire.
Cette solution uniforme (cf. [24]) constitue une extension d’un résultat
de Mulmuley [73] qui ne traite que la question du rang.
Naturellement, le rang d’une matrice peut être calculé par la méthode du pivot de Gauss. Mais la méthode n’est pas uniforme et, a priori,
ne se laisse pas bien paralléliser.
Les applications des formules et algorithmes que nous allons décrire
ici seront de deux ordres : d’une part en calcul parallèle, d’autre part
lorsqu’on doit traiter des systèmes linéaires dépendant de paramètres.
Dans ce deuxième cas de figure, la méthode du pivot de Gauss produit un arbre de calcul qui risque de comporter un très grand nombre de
branches, correspondant à un grand nombre de formules distinctes, lorsque les paramètres prennent toutes les valeurs possibles. Le cas extrême
est celui où toutes les entrées d’une matrice sont des paramètres indépendants. Par exemple avec une matrice de rang maximum de format
n × 2n la solution du système linéaire correspondant par la méthode du
pivot de Gauss dépend du mineur maximal non nul qu’on extrait, et ce
n > 2n mineurs d’ordre n de
dernier peut être n’importe lequel des C2n
la matrice.
En analyse numérique matricielle, avec des matrices à coefficients
réels ou complexes, une formule uniforme compacte en rang fixé est obtenue par l’utilisation des coefficients de Gram de la matrice correspondant
au système linéaire homogène : dans le cas réel, le Gram d’ordre k d’une
matrice A est égal à la somme des carrés de tous les mineurs d’ordre
k de A, son annulation signifie que le rang de la matrice n’excède pas
k − 1.
Les identités que nous allons obtenir sont des généralisations directes
des formules usuelles qui expriment l’inverse de Moore-Penrose en fonction des coefficients de Gram de la matrice. L’étonnant est que, même
sur un corps fini, un petit nombre de sommes de carrés de mineurs suffit
à contrôler le rang d’une matrice, et que des formules semblables aux
formules usuelles fonctionnent encore.
Il y a cependant un prix à payer, non négligeable, qui est d’introduire
un paramètre supplémentaire dans les calculs.
1.7. Résolution uniforme des systèmes linéaires
1.7.1
35
Les coefficients de Gram et l’inverse de Moore-Penrose dans le cas réel ou complexe
Théorie générale
Dans toute la section 1.7.1 A est une matrice dans Km×n , avec
K = C ou R, représentant sur des bases orthonormées une application
linéaire ϕ : E → F entre espaces vectoriels hermitiens ou euclidiens de
dimension finie. Nous noterons hx, yi le produit scalaire des vecteurs
x et y. Nous notons A? la transposée de la conjuguée de A (dans le
cas réel on a A? = tA). La matrice A? représente sur les mêmes bases
l’application linéaire adjointe 9 ϕ? , caractérisée par :
∀x ∈ E
∀y ∈ F
hϕ(x), yiF = hx, ϕ? (y)iE
(1.28)
Les matrices AA? et A? A sont des matrices carrées hermitiennes positives (symétriques réelles positives dans le cas réel), en général non
régulières. Si H est un sous-espace vectoriel de E nous noterons πH
la projection orthogonale de E sur H, vue comme application linéaire
de E dans E.
D’un point de vue de pure algèbre linéaire tous les résultats de la
(( théorie générale )) qui suit sont basés sur la décomposition des espaces
E et F en sommes directes de noyaux et d’images de ϕ et ϕ? .
Lemme 1.7.1 Nous avons deux sommes directes :
Im ϕ ⊕ Ker ϕ? = F, Ker ϕ ⊕ Im ϕ? = E
(1.29)
Cela résulte du fait que Ker ϕ? (resp. Ker ϕ) est le sous-espace orthogonal de Im ϕ (resp. Im ϕ? ), ce qui est une conséquence directe de
l’égalité (1.28).
Nous en déduisons les faits suivants.
Fait 1.7.2
1. L’application linéaire ϕ se restreint en un isomorphisme ϕ0 de
Im ϕ? sur Im ϕ et ϕ? se restreint en un isomorphisme ϕ?0 de
Im ϕ sur Im ϕ? .
2. En outre :
Im ϕ = Im ϕϕ? ,
Ker ϕ? = Ker ϕϕ? ,
Ker ϕ = Ker ϕ? ϕ ,
Im ϕ? = Im ϕ? ϕ .
(1.30)
9. A ne pas confondre avec la matrice adjointe Adj(A). Cette ambigüité dans la
terminologie, en français, est ennuyeuse.
36
1. Rappels d’algèbre linéaire
3. Soit ϕ1 : Im ϕ → Im ϕ l’automorphisme linéaire défini par ϕ1 =
ϕ0 ϕ?0 . C’est la restriction de ϕ ϕ? à Im ϕ. Nous avons :
det(IdIm ϕ + Z ϕ1 ) = det(IdF + Z ϕϕ? )
= 1 + a1 Z + · · · + ar Z r .
où r = rg(ϕ) = rg(ϕ1 ) et ar =
6 0. De la même façon, nous avons
l’automorphisme ϕ?1 = ϕ?0 ϕ0 de Im ϕ? et
det(IdIm ϕ? + Z ϕ?1 ) = det(IdE + Z ϕ? ϕ)
= 1 + a1 Z + · · · + ar Z r .
Ce sont des conséquences directes du lemme 1.7.1.
Peut-être cela sera plus clair si nous représentons ϕ? et ϕ dans les
sommes orthogonales (1.29) :
ϕ=
ϕ0
0I ? ,K ?
0K,I
0K,K ?
,
?
ϕ =
ϕ?0
0I,K
0K ? ,I ?
0K ? ,K
(1.31)
où K = Ker ϕ, K ? = Ker ϕ? , I = Im ϕ, I ? = Im ϕ? .
Définition 1.7.3 Les coefficients de Gram de A (ou de ϕ) sont les
Gk (A) = Gk (ϕ) = ak donnés par la formule
det(Im + Z AA? ) = 1 + a1 Z + · · · + am Z m .
Nous définissons aussi G0 (A) = 1
et
G` (A) = 0
pour
(1.32)
` > m.
Notez que le polynôme caractéristique de B = AA? est égal à
(−1)m Z m Q(1/Z) où Q(Z) = det(Im + Z B). Les coefficients de Gram
de ϕ sont donc, au signe près, les coefficients du polynôme caractéristique de ϕϕ? .
Lemme 1.7.4 (Conditions de Gram pour le rang)
1. L’application linéaire ϕ est de rang ≤ r si et seulement si Gk (ϕ) =
0 pour r < k ≤ n. Elle est de rang r si en outre Gr (ϕ) 6= 0.
2. Le coefficient de Gram Gk (A) est un nombre réel positif ou nul,
égal à la somme des carrés des modules des mineurs d’ordre k de
la matrice A. En conséquence, Gr+1 (ϕ) = 0 suffit pour certifier
que le rang est ≤ r.
1.7. Résolution uniforme des systèmes linéaires
37
Preuve. Le premier point est une conséquence directe du fait 1.7.2-3.
Il pourrait aussi être vu comme une conséquence du second point, que
nous démontrons maintenant.
Le coefficient ak est la somme des mineurs principaux d’ordre k de
AA? . Chaque mineur principal d’ordre k est obtenu comme déterminant
de la matrice correspondante, qui est égale à Aα (Aα )? où α désigne un
k-uple α1 < · · · < αk extrait de {1, . . . , m} et Aα est la matrice
extraite de A en gardant seulement les k lignes correspondant à α. La
formule de Binet-Cauchy (1.2) nous indique alors que ce déterminant est
la somme des carrés des modules des mineurs d’ordre k extraits de Aα .
t
u
Nous supposons désormais r = rg(ϕ) (donc ar = Gr (ϕ) 6= 0).
Puisque
det(IdIm ϕ + Z ϕ1 ) = det(IdF + Z ϕϕ? ) = 1 + a1 Z + · · · + ar Z r
le théorème de Cayley-Hamilton nous donne
+ · · · + (−1)r ar IdIm ϕ = 0 .
ϕr1 − a1 ϕr−1
1
(1.33)
Par suite, on obtient en remplaçant ϕ1 par ϕϕ? dans la formule précédente
(ϕϕ? )r − a1 (ϕϕ? )r−1 + · · · + (−1)r ar πIm ϕ = 0 .
Ainsi :
Lemme 1.7.5 (projections orthogonales sur l’image et sur le noyau)
1. La projection orthogonale πI sur le sous-espace I = Im ϕ ⊆ F
est égale à :
a−1
ar−1 ϕ ϕ? − ar−2 (ϕ ϕ? )2 + · · · + (−1)r−1 (ϕ ϕ? )r . (1.34)
r
2. La projection orthogonale πI ? sur le sous-espace I ? = Im ϕ? ⊆ E
est égale à :
a−1
ar−1 ϕ? ϕ − ar−2 (ϕ? ϕ)2 + · · · + (−1)r−1 (ϕ? ϕ)r .
(1.35)
r
Et la projection orthogonale sur le noyau de ϕ est IdE − πI ? .
En outre l’équation (1.33) implique que l’inverse de ϕ1 est donné
par
ϕ−1
= a−1
ar−1 IdIm ϕ − ar−2 ϕ1 + · · · + (−1)r−1 ϕr−1
.
r
1
1
38
1. Rappels d’algèbre linéaire
De même on a
?
r−1 ? r−1
?
ϕ?1 −1 = a−1
a
Id
−
a
ϕ
+
·
·
·
+
(−1)
ϕ
.
r−1 Im ϕ
r−2 1
r
1
et puisque ϕ?1 = ϕ?0 ϕ0 , cela donne
(
∀y ∈ Im ϕ? ϕ?1 −1 (y) =
a−1
ar−1 IdE − ar−2 (ϕ? ϕ) + · · · + (−1)r−1 (ϕ? ϕ)r−1 (y)
r
(1.36)
Définition 1.7.6 Supposons que ϕ est de rang r. L’inverse de Moore-Penrose de ϕ (en rang r) est l’application linéaire ϕ[−1]r : F → E
définie par :
∀y ∈ F
ϕ[−1]r (y) = ϕ−1
0 (πIm ϕ (y)).
Remarque. Nous n’avons pas écrit ϕ[−1]r = ϕ−1
0 ◦ πIm ϕ parce que le
deuxième membre est a priori (( mal défini )) : πIm ϕ est une application
est une application de Im ϕ dans Im ϕ? et ϕ[−1]r
de F dans F , ϕ−1
0
est une application de F dans E.
D’après (1.29) et (1.31) on voit que
∀y ∈ F
∀y ∈ F
πIm ϕ (y) = ϕ?0 −1 (ϕ? (y))
? −1 ?
ϕ[−1]r (y) = ϕ−1
0 (ϕ0 (ϕ (y))).
? −1 = ϕ? −1 nous obtenons
et puisque ϕ−1
1
0 ◦ ϕ0
∀y ∈ F
ϕ[−1]r (y) = ϕ?1 −1 (ϕ? (y))
(1.37)
En appliquant (1.36) on obtient alors une formule uniforme en rang
r qui donne une solution des systèmes linéaires en analyse numérique
matricielle :
Proposition 1.7.7 (Inverse de Moore-Penrose) Soit v ∈ F . Soit ϕ ⊕ v
l’application linéaire E ⊕ K → F définie par (ϕ ⊕ v)(x, λ) = ϕ(x) + λv.
1) L’inverse de Moore-Penrose ϕ[−1]r ∈ L(F, E) est donné par :
−1
ar ar−1 IdE − ar−2 (ϕ? ϕ) + · · · + (−1)r−1 (ϕ? ϕ)r−1 ϕ?
=
(1.38)
−1 ?
ar ϕ ar−1 IdF − ar−2 (ϕϕ? ) + · · · + (−1)r−1 (ϕϕ? )r−1
où ak = Gk (ϕ).
2) Nous avons v ∈ Im(ϕ) si et seulement si Gr+1 (ϕ ⊕ v) = 0 si et
seulement si
v = ϕ ϕ[−1]r (v)
(1.39)
Dans ce cas x = ϕ[−1]r (v) est l’unique solution dans Im(ϕ? ).
1.7. Résolution uniforme des systèmes linéaires
39
Remarque 1.7.8 Voici la formulation matricielle du lemme 1.7.5 et de
la proposition 1.7.7. Soient m, n, r > 0 dans N avec r ≤ min(m, n) et
A ∈ Km×n une matrice de rang r. Posons ak = Gk (A). Soit V ∈ Km×1 .
1) La matrice de la projection orthogonale sur le sous-espace Im A ⊆ Km
est égale à
P = a−1
ar−1 AA? − ar−2 (AA? )2 + · · · + (−1)r−1 (AA? )r .
r
2) La matrice de la projection orthogonale sur le sous-espace Im A? ⊆
Kn est égale à :
P ? = a−1
ar−1 A? A − ar−2 (A? A)2 + · · · + (−1)r−1 (A? A)r .
r
Et celle de la projection orthogonale sur le noyau de A est In − P ? .
3) La matrice A[−1]r ∈ Kn×m (inverse de Moore-Penrose de A en rang
r) est égale à :
a−1
ar−1 In − ar−2 A? A + · · · + (−1)r−1 (A? A)r−1 A?
(1.40)
r
4) Le système linéaire AX = V admet une solution si et seulement si
Gr+1 (A|V ) = 0 ((A|V ) est la matrice obtenue en juxtaposant la colonne
V à droite de la matrice A) si et seulement si on a l’égalité :
V = A A[−1]r V
(1.41)
Dans ce cas X = A[−1]r V est l’unique solution dans l’espace Im A? .
Notez que la matrice A[−1]r ∈ Kn×m est bien définie par la formule
(1.40) dès que A est de rang ≥ r. Cela est utile en analyse numérique et
de manière plus générale chaque fois que les coefficients de A sont des
réels connus avec seulement une précision finie (ce qui peut introduire
une incertitude sur le rang de la matrice).
Cas des matrices hermitiennes
Lorsque E = F et ϕ = ϕ? , l’endomorphisme ϕ est dit hermitien.
Alors on a une décomposition orthogonale E = Ker ϕ ⊕ Im ϕ et la
restriction ϕ0 de ϕ à Im ϕ est un automorphisme linéaire de Im ϕ.
Nous posons
det(IdE + Z ϕ) = det(Im + Z A) = 1 + b1 Z + · · · + bn Z n . (1.42)
40
1. Rappels d’algèbre linéaire
Au signe près, les bi sont donc les coefficients du polynôme caractéristique de ϕ. Si le rang de ϕ est égal à r alors br 6= 0, br+1 = . . . = bn = 0
et
det(IdE + Z ϕ) = det(IdIm ϕ + Z ϕ0 ) = 1 + b1 Z + · · · + br Z r .
Ainsi par Cayley-Hamilton
ϕr0 − b1 ϕr−1
+ b2 ϕ0r−2 + · · · + (−1)r−1 br−2 ϕ0 + (−1)r−1 br IdIm ϕ = 0
0
Et en remplaçant ϕ0 par ϕ nous obtenons :
ϕr − b1 ϕr−1 + b2 ϕr−2 + · · · + (−1)r−1 br−2 ϕ + (−1)r−1 br πIm ϕ = 0
Ceci donne, pour le cas des matrices hermitiennes, une version simplifiée des résultats précédents plus généraux. Elle se trouve dans l’ouvrage [BP] de Bini et Pan.
Proposition 1.7.9 (inverse de Moore-Penrose, cas hermitien)
1) La projection orthogonale πIm ϕ sur le sous-espace Im ϕ est égale à :
b−1
br−1 ϕ − br−2 ϕ2 + · · · + (−1)r−1 b1 ϕr−1 + (−1)r ϕr .
(1.43)
r
2) L’inverse de Moore-Penrose ϕ[−1]r ∈ L(E, E) est égal à :
b−1
br−1 πIm ϕ − br−2 ϕ + · · · + (−1)r−1 b1 ϕr−2 + (−1)r ϕr−1
r
(1.44)
Remarquez que l’équation (1.34) peut être déduite de (1.30) et (1.43).
Interprétation géométrique
Si λ1 ≥ λ2 ≥ · · · ≥ λr sont les valeurs singulières non nulles de ϕ,
c’est-à-dire les racines carrées des valeurs propres > 0 de ϕϕ? il existe
des bases orthonormées de E et F par rapport auxquelles la matrice
de ϕ est égale à L :
λ1 0 · · · · · · 0 · · · 0
..
.
0 λ2 0
. .
..
.
.
.. .. ..
..
.
.
..
.
L = ..
.. λ
0
.
r
0 ··· ··· 0 0
0
.
..
..
.
0
···
···
···
0
1.7. Résolution uniforme des systèmes linéaires
41
(cf. [Cia, GL, LT]).
Matriciellement on obtient A = U LV où U et V sont des matrices
unitaires (orthogonales dans la cas réel) convenables. Ceci s’appelle la
décomposition de A en valeurs singulières (la SVD en anglais).
On voit que ϕ transforme la sphère unité de E en un ellipsoı̈de
dans Im ϕ avec pour longueurs des axes principaux 2λ1 , . . . , 2λr . Dans
ces conditions la matrice de ϕ? est égale à L? = tL et celle de ϕ[−1]r
est égale à
[−1]r
L
=
(λ1 )−1
0
···
0
..
.
..
.
(λ2 )−1
..
.
0
0
..
.
···
0
···
..
.
..
.
···
···
..
.
(λr )−1
0
···
0 ··· 0
..
.
..
.
..
0
.
0
0
..
.
··· 0
de même format que tL.
Bien que les matrices L et A[−1]r soient attachées de manière unique à A et dépendent continument de A (sous l’hypothèse que le
rang est fixé), il n’en va pas de même pour les matrices U et V de
la décomposition en valeurs singulières, qui sont fondamentalement instables.
Que le vecteur v appartienne ou non à Im ϕ on a toujours ϕ[−1]r (v) ∈
Im ϕ? , qui est le sous-espace orthogonal à Ker ϕ et ϕ(ϕ[−1]r (v)) est
la projection orthogonale de v sur Im ϕ. Ainsi lorsque v n’est pas
dans l’image, l’inverse de Moore-Penrose fournit une solution approchée
x = ϕ[−1]r (v) qui donne pour ϕ(x) la meilleure approximation (au sens
des moindres carrés) de v. En outre x est la plus petite en norme (parmi
les solutions qui réalisent cette meilleure approximation).
Ce qui est remarquable est qu’on arrive à calculer (essentiellement à
l’aide du polynôme caractéristique de ϕϕ? ) les projections orthogonales
et l’inverse de Moore-Penrose par une formule uniforme (plus exactement, par une formule qui ne dépend que du rang, lequel se lit sur le
polynôme caractéristique en question) sans qu’on ait besoin de calculer
les bases orthonormées dans lesquelles se révèle la géométrie de l’application linéaire ϕ.
42
1.7.2
1. Rappels d’algèbre linéaire
Généralisation sur un corps arbitraire
Théorie générale
Dans le cas réel ou complexe, on a vu qu’en termes d’algèbre linéaire tout le paragraphe (( théorie générale )) est gouverné par les sommes
directes (1.29) entre les noyaux et images de ϕ et ϕ? (lemme 1.7.1) :
Im ϕ ⊕ Ker ϕ? = F, Ker ϕ ⊕ Im ϕ? = E .
Il suffit en effet, lorsqu’on parle de projection orthogonale, de remplacer
par exemple l’expression projection orthogonale sur Im ϕ par projection
sur Im ϕ parallèlement à Ker(ϕ? ).
Nous allons voir maintenant que ces relations (1.29) peuvent être
réalisées de manière automatique sur un corps arbitraire K à condition
d’introduire, à la place de A? , une matrice A◦ à coefficients dans le
corps K(t) où t est une indéterminée.
Pour cela nous nous limitons au point de vue purement matriciel,
(c’est le point de vue où des bases ont été fixées dans E et F ). Nous
considérons une forme quadratique Φt,n sur E 0 = K(t)n et une forme
quadratique Φt,m sur F 0 = K(t)m :
Φt,n (ξ1 , . . . , ξn ) = ξ1 2 + t ξ2 2 + · · · + tn−1 ξn 2
Φt,m (ζ1 , . . . , ζm ) = ζ1 2 + t ζ2 2 + · · · + tm−1 ζm 2
Nous notons les (( produits scalaires )) correspondants par h·, ·itE 0 et
h·, ·itF 0 . Nous notons Qn et Qm les matrices (diagonales) de ces formes
sur les bases canoniques.
Toute application linéaire ϕ : E → F donne lieu à une application
linéaire E 0 → F 0 que nous notons encore ϕ et qui est définie par la
même matrice sur les bases canoniques. Il existe alors une unique application linéaire ϕ◦ : F 0 → E 0 qui réalise les égalités (1.28) dans ce
nouveau contexte :
∀x ∈ E 0
∀y ∈ F 0
hϕ(x), yitF 0 = hx, ϕ◦ (y)itE 0
La matrice A◦ de ϕ◦ sur les bases canoniques est alors
A◦ = Qn −1 tA Qm ,
puisqu’on doit avoir pour tous X ∈ K(t)n×1 , Y ∈ K(t)m×1
t
(A X) Qm Y = tX Qn (A◦ Y ) .
(1.45)
1.7. Résolution uniforme des systèmes linéaires
43
En pratique si A = (ai,j ) on obtient A◦ = (tj−i aj,i ), par exemple :
a11
t a21
t2 a31
t−1 a12
a22
t a32
a11 a12 a13 a14 a15
−2
◦
−1
a33
A = a21 a22 a23 a24 a25 , A =
t a13 t a23
t−3 a14 t−2 a24 t−1 a34
a31 a32 a33 a34 a35
t−4 a15 t−3 a25 t−2 a35
Comme nous l’avons déjà indiqué, pour pouvoir reproduire (avec les
légères variations nécessaires) le fait 1.7.2, les définitions 1.7.3 et 1.7.6,
les lemmes 1.7.4, 1.7.5 et la proposition 1.7.7 il nous suffit de démontrer
l’analogue du lemme 1.7.1.
Lemme 1.7.10 Avec les notations ci-dessus on a pour toute matrice
M ∈ Km×n , des sommes directes orthogonales dans les espaces F 0 =
K(t)m et E 0 = K(t)n
Im ϕ ⊕ Ker ϕ◦ = F 0 , Ker ϕ ⊕ Im ϕ◦ = E 0
(1.46)
Preuve. Les dimensions conviennent et il suffit de montrer que l’intersection est réduite à 0. Prenons par exemple la première. La relation
(1.45) implique que l’orthogonal de Im ϕ au sens de la forme bilinéaire
h·, ·iF 0 est égal à Ker ϕ◦ . Il nous suffit donc de montrer que si H est
un sous-espace vectoriel de F 0 = K(t)m défini sur K , son orthogonal
H ⊥ dans K(t)m au sens du produit scalaire h·, ·itF 0 ne le coupe qu’en
0. Soit donc (p1 (t), . . . , pm (t)) ∈ H ∩ H ⊥ . Il existe P
v1 , . . . , vr ∈ H et
a1 (t), . . . , ar (t) ∈ K(t) tels que (p1 (t), . . . , pm (t)) = i ai (t) vi . Quitte
à multiplier par le produit des dénominateurs on peut supposer que
les ai sont des polynômes et donc aussi les pi . On peut introduire
une nouvelle indéterminée u et travailler dans K[t, u]. Alors
P puisque
P
a
(t)
v
est
orthogonal
à
tous
les
v
il
est
orthogonal
à
i
i
i
i ai (u) vi =
i
(p1 (u), . . . , pm (u)) et cela donne
Xm
P (t, u) =
pi (t) pi (u) ti−1 = 0 .
i=1
Il nous reste à voir que cette relation implique que les pi sont tous nuls.
Supposons l’un des pi non nul. Soit d ≥ 0 le plus grand des degrés
des pi . Soit k le plus grand indice pour lequel deg pk = d et ak le
coefficient dominant de pk . Alors on vérifie facilement que le coefficient
de ud td+k−1 dans P (t, u) est égal à a2k , et donc P est non nul.
t
u
Nous nous contentons maintenant de reproduire les définitions et
résultats dans notre nouveau cadre.
44
1. Rappels d’algèbre linéaire
Fait 1.7.11
1. L’application linéaire ϕ : E 0 → F 0 se restreint en un isomorphisme ϕ0 de Im ϕ◦ sur Im ϕ et ϕ◦ se restreint en un isomorphisme
ϕ◦0 de Im ϕ sur Im ϕ◦ .
2. En outre :
Im ϕ = Im ϕϕ◦ ,
Ker ϕ◦ = Ker ϕϕ◦ ,
Ker ϕ = Ker ϕ◦ ϕ ,
Im ϕ◦ = Im ϕ◦ ϕ .
(1.47)
3. Soit ϕ1 : Im ϕ → Im ϕ l’automorphisme linéaire défini par ϕ1 =
ϕ0 ϕ◦0 . C’est la restriction de ϕ ϕ◦ à Im ϕ. Nous avons :
det(IdIm ϕ + Z ϕ1 ) = det(IdF + Z ϕϕ◦ )
= 1 + a1 Z + · · · + ar Z r .
où r = rg(ϕ) = rg(ϕ1 ) et ar 6= 0. De la même façon, nous avons
l’automorphisme ϕ◦1 = ϕ◦0 ϕ0 de Im ϕ◦ et
det(IdIm ϕ◦ + Z ϕ◦1 ) = det(IdE + Z ϕ◦ ϕ)
= 1 + a1 Z + · · · + ar Z r .
Les coefficients de la matrice A A◦ sont des polynômes de Laurent,
autrement dit des éléments de K[t, 1/t].
Définition 1.7.12 Les polynômes de Gram (généralisés) de A sont les
polynômes de Laurent Gk0 (A)(t) = ak (t) ∈ K[t, 1/t], et les coefficients de
0 (A) = a
Gram généralisés de A sont les coefficients Gk,`
k,` donnés par
la formule
det(Im + Z AA◦ ) = 1 + a1 (t) Z + · · · + am (t) Z m
P
(1.48)
k(m+n−2k)
`
ak (t) = t−k(n−k)
a
t
k,`
`=0
Nous définissons aussi G00 (A) = 1
et
G`0 (A) = 0
pour
` > m.
On vérifie aisément que Gk0 (A)(t) = Gk0 ( tA)(1/t).
Dans la suite de cette section, nous dirons, pour abréger, (( polynôme )) à la place de (( polynôme de Laurent )) en laissant au lecteur le soin
de déterminer selon le contexte si des puissances négatives de la variable
sont présentes ou non.
Notons que les coefficients de Gram usuels sont donnés par
X
Gk (A) = Gk0 (A)(1) =
ak,`
(1.49)
`
1.7. Résolution uniforme des systèmes linéaires
45
Les coefficients de Gram généralisés sont des sommes de carrés de
mineurs et ils permettent de contrôler le rang de la matrice en vertu du
lemme suivant, qui est l’analogue du lemme 1.7.4.
Lemme 1.7.13 (Conditions de Gram généralisées pour le rang) Soit
A ∈ Km×n ⊆ K(t)m×n .
(1) La matrice A est de rang ≤ r si et seulement si les polynômes
Gk0 (A)(t) pour k > r sont identiquement nuls. Elle est de rang r si en
outre Gr0 (A) 6= 0.
0 (A) est égal à la somme des
(2) Le coefficient de Gram ak,` = Gk,`
carrés des mineurs µα,β d’ordre k de la matrice A extraits sur les
lignes et les colonnes correspondant aux multi-indices α = (α1 , . . . , αk )
et β = (β1P
, . . . , βk ) pour toutes les paires (α, β) qui vérifient l’égalité
P
k
k
i=1 αi −
j=1 βj = ` − k(n − k).
En particulier Gk0 (A) = 0 si k > p = inf(m, n).
En posant p = inf(m, n) et p0 = sup(m, n) le nombre total des
coefficients de Gram généralisés est égal à :
p
X
k=1
(k (m + n − 2k) + 1) = p +
1
1
p (p + 1) (3p0 − p − 2) ≤ p(p + 1)p0 .
6
2
Nous avons les analogues du lemme 1.7.5 et de la proposition 1.7.7.
Lemme 1.7.14 (projections sur l’image et sur le noyau)
Soient m, n, r > 0 dans N avec r ≤ min(m, n), A ∈ Km×n une
matrice de rang r. Posons ak (t) = Gk0 (A).
(1) La matrice de la projection sur le sous-espace Im A ⊆ K(t)m parallèlement à Ker A◦ est égale à
P = a−1
ar−1 AA◦ − ar−2 (AA◦ )2 + · · · + (−1)r−1 (AA◦ )r
r
(1.50)
(2) La matrice de la projection sur le sous-espace Im A◦ ⊆ K(t)n parallèlement à Ker A est égale à
P • = a−1
ar−1 A◦ A − ar−2 (A◦ A)2 + · · · + (−1)r−1 (A◦ A)r
r
(1.51)
Et la matrice de projection sur le noyau de A parallèlement à Im A◦
est In − P • .
46
1. Rappels d’algèbre linéaire
Notez qu’il s’agit de projections orthogonales par rapport aux formes
bilinéaires h., .itE 0 et h., .itF 0 .
Remarque. En fait chaque formule peut être spécialisée en remplaçant
t par n’importe quelle valeur τ ∈ K \ {0} qui n’annule pas le dénominateur ar (t) (ce qui est toujours possible si le corps possède au moins
r(m + n − 2r) + 1 éléments).
Définition 1.7.15 Supposons que ϕ est de rang r. L’inverse de Moore-Penrose généralisé de ϕ (en rang r) est l’application linéaire ϕ[−1]r,t :
F 0 → E 0 définie par :
∀y ∈ F 0
ϕ[−1]r,t (y) = ϕ−1
0 (πIm ϕ (y)).
Proposition 1.7.16 (inverse de Moore-Penrose généralisé)
Soient m, n, r > 0 dans N avec r ≤ min(m, n), A ∈ Km×n de rang r,
V ∈ Km×1 .
(1) L’inverse de Moore-Penrose généralisé de A en rang r est la matrice A[−1]r,t ∈ K(t)n×m égale à
a−1
ar−1 Im − ar−2 A◦ A + · · · + (−1)r−1 (A◦ A)r−1 A◦
(1.52)
r
où ak = Gk0 (A).
(2) Le système linéaire AX = V admet une solution si et seulement si
0
le polynôme Gr+1
(A|V ) est identiquement nul si et seulement si
V = A A[−1]r,t V
(1.53)
Dans ce cas X = A[−1]r,t V est l’unique solution dans Im(A◦ ).
Remarque. Si ϕ est injective et v ∈ Im ϕ est représenté par un vecteur
colonne V ∈ Km×1 l’élément A[−1]r,t V est l’unique solution du système linéaire correspondant. En conséquence, il ne dépend pas de t et les
fractions rationnelles données par le calcul des coordonnées de A[−1]r,t V
se simplifient en des constantes.
Cas des matrices symétriques
Dans ce paragraphe E 0 = F 0 = K(t)n×n , ϕ est défini par une matrice
symétrique A = tA ∈ Kn×n et rg(ϕ) = r. Soit λ l’automorphisme linéaire de E 0 defini par Qn par rapport à la base canonique.
Définissons ϕ
b = λ−1 ◦ ϕ, ϕ
e = ϕ ◦ λ.
1.7. Résolution uniforme des systèmes linéaires
47
b = Q−1 A, celle de ϕ
e = A Qn .
La matrice de ϕ
b est A
e est A
n
Puisque A est symétrique, on a :
b = t (A)(1/t)
e
A(t)
et ϕ◦ = λ−1 ◦ ϕ ◦ λ = ϕ
b ◦ λ = λ−1 ◦ ϕ
e.
On en déduit :
Im ϕ
b = Im ϕ◦ ,
Ker ϕ
b = Ker ϕ,
Im ϕ
e = Im ϕ,
Ker ϕ
e = Ker ϕ◦ , Im ϕ = λ(Im ϕ◦ ), Ker ϕ = λ(Ker ϕ◦ )
Donc l’équation (1.46) peut être réécrite comme deux décompositions
orthogonales (par rapport à la forme bilinéaire h., .itE 0 ) :
Im ϕ
b ⊕ Ker ϕ
b = Im ϕ◦ ⊕ Ker ϕ = I ◦ ⊕ K = E 0 ,
Im ϕ
e ⊕ Ker ϕ
e = Im ϕ ⊕ Ker ϕ◦ = I ⊕ K ◦ = E 0 .
Nous notons ϕ
e0 l’automorphisme de I obtenu par restriction de ϕ.
e
Les applications linéaires ϕ, ϕ
e et ϕ
e0 ont même rang r, et la somme
directe Im ϕ
e ⊕ Ker ϕ
e = I ⊕ K ◦ = E 0 entraı̂ne que :
det(In + Z ϕ)
e = det(IdI + Z ϕ
e0 ) = 1 + b1 (t) Z + · · · + br (t)Z r
(1.54)
avec br 6= 0 et br+1 = . . . = bn = 0. Les bi (t) sont au signe près les
coefficients du polynôme caractéristique de ϕ.
e
On vient de démontrer la version simplifiée du lemme 1.7.13. Ceci
constitue le résultat clé de Mulmuley [73].
Lemme 1.7.17 (Conditions de Mulmuley pour le rang d’une matrice
e = A Qn avec A ∈ Kn×n symétrique. Soit ck = ck (t)
symétrique) Soit A
e
le coefficient de Z n−k dans le polynôme caractéristique PAe(Z) de A.
Alors, la matrice A est de rang ≤ r si et seulement si les polynômes
ck (t) pour k > r sont identiquement nuls.
Elle est de rang r si en outre cr (t) 6= 0.
Puisqu’on a la somme orthogonale Im ϕ
e ⊕ Ker ϕ
e = E 0 on peut reproduire les calculs donnés dans le cas des matrices symétriques réelles
et on obtient le résultat suivant, qui simplifie ceux obtenus dans le cas
d’une matrice arbitraire, de façon similaire à la proposition 1.7.9.
Proposition 1.7.18 (Un inverse généralisé d’une matrice symétrique)
Soit A ∈ Kn×n symétrique de rang r, E 0 = K(t)n et ϕ : E 0 → E 0
e = A Qn .
l’application linéaire définie par A. On considère la matrice A
48
1. Rappels d’algèbre linéaire
Les coefficients bi sont définis par l’égalité (1.54). Dans la suite l’orthogonalité s’entend par rapport à la forme bilinéaire h., .itE 0 .
(1) La projection orthogonale πIm ϕ sur le sous-espace Im ϕ de E 0 a
pour matrice :
e − br−2 A
e 2 + · · · + (−1)r−1 b1 A
e r−1 + (−1)r A
er .
P = b−1
b
A
r−1
r
(2) L’inverse de Moore-Penrose généralisé ϕ
e [−1]r,t ∈ L(E 0 , E 0 ) de ϕ
e a
pour matrice :
e [−1]r,t = br −1 br−1 P − br−2 A
e + · · · + (−1)r−1 b1 A
e r−2 + (−1)r A
e r−1
A
(3) L’endomorphisme ψ = λ ◦ ϕ
e [−1]r,t de E 0 , dont la matrice est B =
e [−1]r,t , est un inverse généralisé de ϕ au sens suivant :
Qn A
ϕ◦ψ◦ϕ=ϕ
et
ψ◦ϕ◦ψ = ψ.
L’application linéaire ϕ ◦ ψ est la projection sur Im ϕ parallèlement à
Ker ψ et ψ ◦ ϕ est la projection sur Im ψ parallèlement à Ker ϕ.
Pour tout vecteur colonne V le système linéaire AX = V admet une
solution si et seulement si ABV = V et en cas de réponse positive, BV
est l’unique solution dans l’espace Im ψ.
Preuve. Il reste à prouver le point 3. Posons θ = ϕ
e [−1]r,t . On sait que
ϕ
e◦θ ◦ϕ
e=ϕ
e et
θ◦ϕ
e◦θ = θ.
Puisque ϕ
e = ϕ◦λ et ψ = λ◦θ, cela donne tout de suite les deux égalités
demandées pour ψ et ϕ. Tout le reste suit sans difficulté.
t
u
Pour la théorie des inverses généralisés nous recommandons les livres
[Bha] et [LT].
Interprétation par les identités de Cramer
Supposons la matrice A de rang r et V dans l’espace engendré par
les colonnes de A. Appelons Cj la j - ème colonne de A. Soit µα,β =
det(Aα,β ) le mineur d’ordre r de la matrice A extrait sur les lignes
α = {α1 , . . . , αr } et les colonnes β = {β1 , . . . , βr }. Pour j = 1, . . . , r
soit να,β,j le déterminant de la même matrice extraite, à ceci près que la
colonne j a été remplacée par la colonne extraite de V sur les lignes α.
1.7. Résolution uniforme des systèmes linéaires
49
Alors on obtient pour chaque couple (α, β) de multi-indices une identité
de Cramer :
Xr
να,β,j Cβj
(1.55)
µα,β V =
j=1
due au fait que le rang de la matrice (A1..m,β |V ) est inférieur ou égal à
r (cf. l’égalité 1.6 page 8). Ceci peut se relire comme suit :
να,β,1
Cβ1 . . . Cβr · ... =
µα,β V =
να,β,r
v α1
Cβ1 . . . Cβr · Adj(Aα,β ) · ... =
=
v αr
= A · (In )1..n,β · Adj(Aα,β ) · (Im )α,1..m · V
P
P
Notons |α| = ri=1 αi , |β| = ri=1 βi . Rappelons que
X
Gr0 (A)(t) =
t|α|−|β| µ2α,β .
(1.56)
α,β
Si nous multiplions chaque égalité (1.56) par µα,β t|α|−|β| et si nous additionnons toutes ces égalités nous obtenons une expression de la forme :
X
0
|α|−|β|
Gr (A)·V = A·
µα,β t
· (In )1..n,β · Adj(Aα,β ) · (Im )α,1..m ·V
α,β
Cette formule ressemble beaucoup trop à (1.53) donnée dans la proposition 1.7.16 :
V = A A[−1]r,t V
pour ne pas être due à une égalité
X
Gr0 (A) A[−1]r,t =
µα,β t|α|−|β| (In )1..n,β ·Adj(Aα,β )·(Im )α,1..m . (1.57)
β,α
Ainsi l’inverse de Moore-Penrose généralisé peut être interprété comme
une somme pondérée d’identités de Cramer.
Nous ne prouverons cependant pas cette dernière égalité. On peut
la trouver, démontrée dans un cadre différent (plus général) et formulée
différemment, comme l’égalité 2.13 dans [76] ou, avec la même formulation qu’ici, dans [25].
2. Algorithmes de base en
algèbre linéaire
Introduction
Il s’agit dans ce chapitre de décrire et d’analyser certaines méthodes
séquentielles, plus ou moins classiques, pour le calcul du déterminant et
du polynôme caractéristique à coefficients dans un anneau commutatif.
L’objectif recherché est de comparer ces algorithmes séquentiels et de
dégager le meilleur possible, c’est-à-dire le plus rapide théoriquement et
pratiquement, occupant le moins d’espace mémoire possible (en évitant
notamment l’explosion de la taille des résultats intermédiaires), le plus
facilement implémentable sur machine séquentielle et le plus général,
c’est-à-dire applicable dans un anneau commutatif arbitraire.
Nous introduirons plus loin (chapitre 4) des notions précises de complexité. Dans ce chapitre nous nous contenterons de la notion informelle
de complexité arithmétique donnée par le compte du nombre d’opérations arithmétiques dans l’anneau de base lors de l’exécution de l’algorithme considéré. Nous ferons également quelques commentaires, souvent informels, sur le bon contrôle (ou non) de la taille des résultats
intermédiaires.
Nous commençons par l’algorithme du pivot de Gauss pour le calcul
du déterminant. C’est l’algorithme d’algèbre linéaire le plus classique. Il
fonctionne sur un corps et possède de nombreuses applications (solutions
de systèmes linéaires , calcul de l’inverse, LU-décomposition . . .). La méthode du pivot pour la résolution des systèmes linéaires est en fait due
aux savants chinois : on pourra consulter à ce sujet la notice historique
du chapitre 3 dans l’ouvrage de Schrijver [Sch] ainsi que l’étude plus
récente de Karine Chemla 1 [15, 16].
1. Dès le troisième siècle de notre ére, on trouve dans les commentaires de Liu
Hui sur le texte classique Les neuf Chapitres ce qu’il semble légitime d’appeler une
52
2. Algorithmes de base en algèbre linéaire
Nous continuons avec un algorithme qui a pour mérite sa très grande
simplicité : l’algorithme de Jordan-Bareiss qui peut être vu comme une
adaptation de la méthode du pivot de Gauss, avec un meilleur comportement des coefficients intermédiaires. Cet algorithme fonctionne sur un
anneau commutatif intègre, à condition que les divisions exactes ne soit
pas trop coûteuses. Dans le cas du calcul du polynôme caractéristique, il
devient un algorithme sans division et s’applique sur un anneau commutatif arbitraire. C’est ce que nous appelons la méthode de Jordan-Bareiss
modifiée. Une variante de l’algorithme de Jordan-Bareiss due à Dodgson
(alias Lewis Caroll) offre des perspectives intéressantes dans le cas des
matrices structurées.
Nous étudions ensuite l’algorithme de Hessenberg, couramment utilisé en analyse numérique. Il utilise des divisions par des éléments non
nuls arbitraires (on suppose donc qu’on travaille sur un corps). Mais en
calcul formel, où l’on veut des résultats exacts, se pose sérieusement le
problème de la croissance de la taille des résultats intermédiaires.
Nous signalons la méthode d’interpolation de Lagrange dans laquelle
le calcul du polynôme caractéristique dépend du calcul de plusieurs déterminants.
Nous examinons ensuite des méthodes qui utilisent des divisions uniquement par des nombres entiers (de petite taille). Il s’agit de la méthode
de Le Verrier et de son amélioration à la Souriau-Faddeev-Frame.
Nous continuons avec les méthodes sans division de Samuelson-Berkowitz et de Chistov. Plus sophistiquées et nettement plus efficaces que
la méthode de Jordan-Bareiss modifiée elles fonctionnent également sur
un anneau commutatif arbitraire. L’algorithme de Chistov présente les
mêmes caractéristiques que celui de Samuelson-Berkowitz mais révèle un
léger handicap par rapport à ce dernier dans les tests expérimentaux.
Nous terminons avec les méthodes qui utilisent les suites récurrentes
linéaires. Celle que nous appelons méthode de Frobenius a de bonnes
caractéristiques tant du point de vue du nombre d’opérations arithmétiques que de la taille des résultats intermédiaires. Elle ne s’applique
cependant pas en toute généralité, et, hormis le cas des corps finis, elle
est en pratique surpassée par l’algorithme de Berkowitz, sans doute parce
que ce dernier n’utilise pas de division, et a besoin de moins d’espace
mémoire (meilleur contrôle des résultats intermédiaires). Nous exposons
également une variante due à Wiedemann.
Dans ce chapitre, nous nous intéressons seulement à des versions
preuve de correction de l’algorithme du pivot de Gauss présenté dans ce texte ancien.
2.1. Méthode du pivot de Gauss
53
assez simples des algorithmes.
Nous dirons qu’une version d’un algorithme est élémentaire si les
multiplications de matrices, de polynômes ou de nombres entiers qui
interviennent en son sein sont exécutées selon la méthode classique usuelle (dite parfois (( naı̈ve ))). Pour les matrices et les polynômes la multiplication usuelle consiste à appliquer simplement la formule définissant
le produit. Pour la multiplication des entiers, il s’agit de l’algorithme
qu’on apprend à l’école primaire.
D’autre part, nous parlons de versions séquentielles dans la mesure
où les méthodes qui cherchent à accélérer l’exécution lorsque de nombreux processeurs sont utilisés en parallèle ne sont pas envisagées.
Dans ce chapitre, nous ne développons que des versions séquentielles
élémentaires.
Rappelons enfin la convention importante suivante : dans tout cet
ouvrage la notation log n signifie max(1, log2 n).
2.1
Méthode du pivot de Gauss
C’est la méthode la plus répandue et la plus courante aussi bien pour
le calcul exact que pour le calcul approché des déterminants lorsque les
coefficients appartiennent à un corps K dans lequel les opérations de
base ( + , − , × , / ) ainsi que le test d’égalité à 0 s’effectuent par des
algorithmes.
Son intérêt réside non seulement dans le fait qu’elle possède plusieurs
variantes (symboliques ou numériques) jouant un rôle important dans la
réduction et l’inversion des matrices et dans la résolution des systèmes
linéaires, mais aussi dans le fait que la technique du pivot est utilisée
dans d’autres méthodes de réduction comme celle de Jordan-Bareiss, ou
pour le calcul du polynôme caractéristique comme nous le verrons plus
loin avec, par exemple, les méthodes de (( Jordan-Bareiss modifiée )) ou
de Hessenberg.
2.1.1
Transformations élémentaires
Une matrice est dite triangulaire supérieure (resp. triangulaire inférieure) si les éléments situés au-dessous de (resp. au dessus de) la
diagonale principale sont nuls. On dit matrice triangulaire lorsque le
contexte rend clair de quelle variante il s’agit. Une matrice triangulaire
54
2. Algorithmes de base en algèbre linéaire
est dite unitriangulaire si les coefficients sur la diagonale principale sont
tous égaux à 1.
Basée sur l’idée des éliminations successives des inconnues dans la
résolution d’un système linéaire, la méthode du pivot de Gauss consiste
à réduire une matrice A ∈ Km×n à une matrice triangulaire supérieure par une succession de transformations élémentaires sur les lignes (et
éventuellement sur les colonnes) de A.
Les transformations élémentaires sur les lignes d’une matrice sont
de trois types :
(i) multiplier une ligne par un élément non nul de K ;
(ii) échanger deux lignes ;
(iii) ajouter à une ligne le produit d’une autre ligne par un élément
de K.
On définit de manière analogue les transformations élémentaires sur
les colonnes.
On associe à toute transformation élémentaire (sur les lignes ou sur
les colonnes) d’une matrice A ∈ Km×n la matrice (dite élémentaire) obtenue en effectuant cette même transformation élémentaire de la matrice
unité (matrice unité d’ordre m ou n selon le cas). Toute transformation
élémentaire sur les lignes (resp. colonnes) de A revient alors à multiplier à gauche (resp. à droite) la matrice A par la matrice élémentaire
correspondante. Ceci est dû simplement au fait que si L1 , L2 ∈ K1×n ,
on a pour tout λ ∈ K :
L1
λ 0
λL1
L1
0 1
L2
=
;
=
;
0 1
L2
L2
1 0
L2
L1
et
1 0
λ 1
L1
L2
=
L1
L2 + λL1
· ( 2)
Il est clair que l’inverse d’une transformation élémentaire sur les
lignes (resp. colonnes) est une transformation élémentaire de même type
sur les lignes (resp. colonnes). Précisément :
λ 0
0 1
−1
=
et
1 0
λ 1
λ−1 0
0 1
−1
1 0
−λ 1
=
,
2. Égalités analogues pour les colonnes.
0 1
1 0
−1
=
0 1
1 0
pour tout λ ∈ K.
2.1. Méthode du pivot de Gauss
55
Une matrice (et l’application linéaire correspondante) est dite unimodulaire si elle est de déterminant 1. Lorsqu’on veut se limiter aux
transformations élémentaires qui correspondent au produit par une matrice unimodulaire, on a droit seulement à celles du troisième type.
Néanmoins, il est facile de voir qu’une succession de trois telles transformations permet d’obtenir un échange signé de lignes (ou de colonnes) du
type (Li , Lj ) ← (Lj , −Li ), qui est considéré comme la variante unimodulaire des transformations élémentaires du deuxième type. Les échanges
signés et les transformations élémentaires du troisème type sont appelées
transformations élémentaires unimodulaires.
L’élimination de Gauss proprement dite que nous considérons ici
est essentiellement une succession de transformations élémentaires du
troisième type sur les lignes : des échanges de lignes ou de colonnes
n’interviennent que s’il y a lieu de chercher un pivot non nul pour le
ramener au bon endroit. Chaque étape de l’algorithme de Gauss consiste
donc à traiter le pivot (non nul) issu de l’étape précédente, en faisant
apparaı̂tre des zéros au-dessous de ce pivot, et à déterminer ensuite
le pivot de l’étape suivante pour le placer sur la diagonale consécutivement au pivot précédent. Si on remplaçait les échanges (de lignes ou
de colonnes) signés, on obtiendrait donc une réduction n’utilisant que
des transformations élémentaires unimodulaires.
En fait, il est bien connu, et c’est une conséquence de la méthode
du pivot de Gauss, que toute matrice carrée inversible est égale à un
produit de matrices élémentaires. Et qu’en conséquence toute matrice
(de n’importe quel format) peut être ramenée par manipulations élémentaires de lignes et de colonnes à une forme canonique du type suivant
Ir | 0
− − −
0 | 0
avec la possibilité de lignes ou de colonnes vides.
Cette réduction est d’une importance théorique capitale. Citons par
exemple Gabriel & Roiter [GR] page 5, qui donnent d’ailleurs dans leur
chapitre 1 des extensions très intéressantes de la méthode : [. . .] en
dépit de son évidence et de sa simplicité, ou peut-être grâce à elles,
cette réduction est très utile, et son usage répété conduit à des résultats
profonds.
Si on se limite aux transformations élémentaires unimodulaires, alors
la forme réduite est la même que ci-dessus dans le cas d’une matrice rectangulaire ou carrée non inversible, et pour une matrice carrée inversible
56
2. Algorithmes de base en algèbre linéaire
il faut modifier la forme réduite en prenant son dernier coefficient diagonal non nécessairement égal à 1.
2.1.2
La LU - décomposition
Lorsque le processus de triangulation d’une matrice A ∈ Km×n
aboutit sans qu’aucune permutation de lignes ou de colonnes n’intervienne — ce qui a lieu si les r premières sous-matrices principales dominantes de A, r étant le rang de A, sont régulières — et si l’on garde
en mémoire les matrices élémentaires associées aux transformations effectuées, la méthode du pivot de Gauss permet d’obtenir, en même temps
que la triangulation de A, ce qu’il est convenu d’appeler une LU -décomposition, c’est-à-dire une façon d’écrire A sous la forme : A = LU , où
U ∈ Km×n est une matrice triangulaire supérieure (c’est la forme triangulaire recherchée de A), et L ∈ Km×m une matrice unitriangulaire
inférieure : L n’est autre que l’inverse du produit des matrices élémentaires correspondant aux transformations successives effectuées sur les
lignes de A.
Pour une matrice carrée régulière, l’existence d’une telle décomposition équivaut au fait que le processus de triangulation arrive à son
terme sans aucun échange de lignes ni de colonnes. Elle équivaut aussi
à la complète régularité de la matrice puisque les mineurs principaux
de la matrice considérée ne sont autres que les produits successifs des
pivots rencontrés au cours du processus. Enfin, toujours dans le cas
d’une matrice carrée régulière, l’existence de la décomposition implique
son unicité. Cela ne serait plus le cas pour une matrice singulière comme
on peut le voir ici
1 0 0
1 3 1
1 3 1
1 0 0
1 3 1
2 1 0 0 0 1 = 2 6 3 = 2 1 0 0 0 1 .
3 4 1
0 0 1
3 9 8
3 0 1
0 0 5
Nous donnons maintenant à voir le résultat de la méthode du pivot
de Gauss avec des matrices à coefficients entiers.
Exemples 2.1.1 Nous montrons deux exemples caractérisitiques, où tous les pivots qui se présentent sur la diagonale sont non nuls. Nous donnons les matrices L et
U . Le premier est celui d’une matrice dont les coefficients entiers ne prennent pas plus
2.1. Méthode du pivot de Gauss
57
que 2 chiffres. Sur la première ligne les matrices M1 et L1 , ensuite la matrice U1 .
1
0
0
0 0
19
1
0
0 0
9
7
8 11 13
4
9
19
35
−313
4
56
84
73
10
1
0 0
35 62 −13 17 23 11 ,
,
9
97
20
3
6
7
5
9 20
113 −5562
1 0
49 23
50 42
2 17
9
97
7963
136 −4694 −21433
49
1
9
97
7963
244718
9
7
8
11
13
4
352
547
410
14
0 −97
9
9
9
9
9
7963
16523
11586
45
0
0
U1 =
.
97
97
97
97
51521
−10965
244718
0
0
0
7963
7963
7963
−15092695 −1665525
0
0
0
0
244718
244718
Le deuxième exemple est celui d’une matrice à coefficients dans Q. Le numérateur et
le dénominateur n’ont qu’un chiffre, mais la croissance de la taille des coefficients est
spectaculaire. Sur la première ligne M2 et U2 , sur la seconde L2 .
1
3
−9
7
−7
3
−9
7
−7
1
6
2
5
6
6
6
2
5
6
6
3
−9
5
2
1
86
−137
193
0 −117
2
8
7
9
8
5
14
18
−1
2
7
1
9
6763
−4175
35446
9
3
6
8
5
0
0
7
3510
19656
15795
3
1
−7 −4
,
8
−959257
−3391183
4
4
9
3
0
0
0
−1
1704276
486936
−7
−4
6
−1
25849022797
6
9
7 0
3
0
0
0
9
10254937392
−1
2
−5
9
0
0
0
0
8
2
5
9
8 0
0
0
0
0
0
2
−8
4
−3
−1
3
7
9
7
1
0
0
0
0
0 0
9
1
0
0
0
0 0
−2 −40
1
0
0
0 0
3
351
21
19
4635
1
0
0
0
4
39
6763
,
L2 =
−16
−8475
−969733
−2
1
0 0
117
6763
6782366
27
85
381
8966489
251177120859
1
0
117
13526
13564732
258490227970
4
400
−54066 −3752551
64239864618
4
0 1
819
47341
23738281
129245113985
58
2. Algorithmes de base en algèbre linéaire
Nous allons comprendre ces comportements typiques en exprimant
précisément les coefficients calculés dans la méthode du pivot de Gauss
en fonction de déterminants extraits de la matrice initiale. Nous avons
pour cela besoin de préciser les notations.
Notation 2.1.2 Soit A ∈ Km×n une matrice de rang r. On suppose
que la triangulation de Gauss aboutit à son terme sans échange de ligne
ni de colonne. Dans ces conditions on pose A[0] = A, on note A[p] la
matrice transformée de A à l’issue de l’étape p (p ≤ r) et on note L[p]
le produit des matrices élémentaires correspondant aux transformations
effectuées au cours de l’étape p, de sorte que A[p] = L[p] A[p−1] et L[p]
est une matrice qui ne diffère de la matrice unité Im que des éléments
de la p - ème colonne situés au-dessous de la diagonale principale. On
[p]
[p]
note a ij l’élément en position (i, j) de la matrice A[p] et l ij celui de
la matrice L[p] .
(p)
Le symbole de Kronecker est défini page 22 et la notation a ij page
3. On a alors :
[p]
[p]
Propriété 2.1.3 Avec les notations précédentes, les éléments l ij , a ij
(p)
et a ij sont liés par les relations suivantes (dans (2.1) on a 1 ≤ p ≤ r,
p < j ≤ n et p < i ≤ m) :
[p]
a ij
=
[p−1]
a ij
[p]
l ip
−
=
[p−1]
ip
[p−1]
a pp
a
[p−1]
a pj
[p−1]
ip
− [p−1]
a pp
a
=
[p−1] [p−1]
[p−1] [p−1]
ij a pp − a ip a pj
.
[p−1]
a pp
a
si i > p,
[p]
l ij = δij sinon.
[0] [1]
(p−1)
a 11 a 22 · · · a[p−1]
pp = a pp .
(2.1)
(2.2)
(2.3)
(p)
[p]
a ij
[p]
l ip
=
(p−1)
ip
− (p−1)
a pp
a
=
a ij
a
(p−1)
pp
si i > p,
.
(2.4)
[p]
(2.5)
l ij = δij sinon.
2.1. Méthode du pivot de Gauss
59
Preuve. Les deux premières équations correspondent exactement aux
affectations de l’algorithme de Gauss. Les deux suivantes correspondent
au fait que les déterminants des sous matrices correspondantes de A
sont inchangés par les transformations élémentaires de lignes utilisées
dans l’algorithme. La dernière résulte de la deuxième et la quatrième. t
u
[r−1]
Il est clair que la matrice U = A
(où r = rg(A)) obtenue à
l’issue de la dernière étape de l’algorithme de Gauss dans ce cas, est
bien la forme triangulaire supérieure recherchée de la matrice A, et que
A = LU où
h
i−1 h
i−1
h
i−1
L = [lij ] = L[1]
L[2]
· · · L[r−1]
(2.6)
est une matrice triangulaire inférieure avec en outre
(j−1)
l ij =
[j]
−l ij
=
a ij
(j−1)
a jj
si m ≥ i > j ≥ 1,
l ij = δij sinon.
(2.7)
−1
[p]
ne diffère de L[p] que des éléments l ip pour
En effet la matrice L[p]
1 ≤ p < i ≤ m qui doivent être remplacés par leurs opposés, et la mul
−1
−1 [p+1] −1
tiplication à gauche par L[p−1]
du produit L[p]
L
···
[r−1] −1
n’affecte que la (p−1) - ème colonne de ce dernier (identique à
L
la (p − 1) - ème colonne de Im ) et revient tout simplement à la remplacer
−1
par la (p − 1) - ème colonne de L[p−1] .
Remarquons aussi que la relation (2.3) montre comment l’algorithme du pivot de Gauss permet de calculer les mineurs principaux de la
matrice A (et donc son déterminant lorsqu’elle est carrée).
Nous comprenons maintenant dans le cas d’une matrice initiale à
coefficients entiers le comportement typique de la taille des coefficients
calculés dans la méthode du pivot de Gauss (cf. la matrice M1 de
l’exemple précédent). On voit sur les relations (2.4), (2.2) et (2.7) que
tous ces coefficients peuvent être écrits comme des fractions dont le numérateur et le dénominateur sont des mineurs de la matrice initiale. En
outre les mineurs sont majorés (en valeur absolue, donc aussi en taille si
ce sont des entiers) en utilisant l’inégalité de Hadamard. Grosso modo,
en partant d’une matrice à k lignes avec des coefficients de taille τ , on
obtient dans l’algorithme du pivot de Gauss des coefficients de taille kτ .
Pour ce qui concerne une matrice à coefficients dans Q (comme M2 ),
pour obtenir une majoration de la taille des coefficients calculés, nous
devons remplacer M2 par une matrice à coefficients entiers M20 = c M2
60
2. Algorithmes de base en algèbre linéaire
(où c est le ppcm des dénominateurs). Grosso modo, en partant d’une
matrice à k lignes avec des coefficients dont le dénominateur et le numérateur sont de taille τ , on obtient maintenant dans l’algorithme du
pivot de Gauss des coefficients de taille kτ 2 .
Algorithme du pivot de Gauss simplifié
L’algorithme simplifié pour la méthode du pivot de Gauss s’applique
pour les matrices fortement régulières. Dans ce cas, il n’y a pas de recherche de pivot et la matrice est de rang maximum inf(m, n). Cet algorithme remplace la matrice A par une matrice de mêmes dimensions
dont la partie supérieure (diagonale principale comprise) est celle de la
matrice U et la partie inférieure (sans la diagonale) 3 celle de la matrice
L de la LU -décomposition de A. On obtient l’algorithme 2.1.
Algorithme 2.1 Algorithme du pivot de Gauss simplifié (sans
recherche de pivot) et LU-décomposition.
Entrée : Une matrice A = (aij ) ∈ Km×n fortement régulière.
Sortie : La matrice A transformée ainsi que les matrices L et U comme
expliqué ci-dessus.
Début
Variables locales : i, j, p ∈ N ; piv ∈ A ;
pour p de 1 à inf(m, n) faire
piv := app ;
pour i de p + 1 à m faire
aip := aip /piv ;
pour j de p + 1 à n faire aij := aij − aip ∗ apj
fin pour
fin pour
fin pour
Fin.
En fait la dernière étape ( p = inf(m, n)) de la boucle principale ne
s’exécute que si m > n et elle ne modifie alors que les valeurs des ain
pour i > n. On aurait donc pu écrire pour p de 1 à inf(m, n) − 1
faire . . . mais il aurait fallu rajouter à la fin :
3. Et sans les éléments nuls en position (i, j) avec i > j > n lorsque m ≥ n + 2.
2.1. Méthode du pivot de Gauss
61
si m > n alors
piv := ann ;
pour i de n + 1 à m faire
ain := ain /piv ;
fin pour
fin si
Un calcul élémentaire donne le résultat suivant.
Proposition 2.1.4 Le nombre d’opérations arithmétiques dans K lorsqu’on exécute l’algorithme du pivot de Gauss simplifié, est majoré par :
n(m − 1)(2n − 2m + 1) + 61 m(m − 1)(4m − 5)
ce qui donne pour m = n la majoration
2 3
3n
− 12 n2 − 61 n.
Si la matrice A est de rang r et si les r premiers mineurs principaux
dominants sont non nuls, l’algorithme précédent, modifié pour s’arrêter
lorsque le pivot piv est nul, fournit encore la LU -décomposition de A.
Cela donne l’algorithme 2.2 page suivante.
Exemple 2.1.5 Voici une matrice M3 ∈ Z6×5 de rang 4, suivie des matrices L3
et U3 obtenues à partir de l’algorithme 2.2.
1
0
0
−21
1
0
73
−73 −53 −30
45 −58
−72
8123
21 −54 −11
1
0
−1
73
5055
72 −59
52 −23
77
, −33 −2266 220594
33
55
66 −15
62
73
5055
272743
−41 −95 −25
51 −54
41
4762
52277
14
55
35
−5
25
73
5055
272743
−14 −1091
83592
73
1685
272743
−73
−53
−30
45
−58
−5055 −1433
945
−1291
0
73
73
73
73
272743
196
243716
0
0
5055
337
5055
2911532 −3038698
0
0
0
272743
272743
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
1
1193
949
1052
949
0
1
0
62
2. Algorithmes de base en algèbre linéaire
Algorithme 2.2 Deuxième algorithme du pivot de Gauss simplifié (sans recherche de pivot) et LU-décomposition.
Entrée : Une matrice A = (aij ) ∈ Km×n .
Sortie : La matrice A transformée ainsi que le rang r de A lorsque celui-ci
est égal à l’ordre du dernier mineur principal dominant non nul. On obtient
également, dans ce cas, la LU-décomposition de la matrice A comme dans
l’algorithme 2.1.
Début
Variables locales : i, j, p, r ∈ N ; piv ∈ K ;
p := 1 ; r := inf (m, n) ;
tant que p ≤ inf (m, n) faire
piv := app ;
si piv = 0 alors r := p − 1 ; p := inf (m, n) sinon
pour i de p + 1 à m faire
aip := aip /piv ;
pour j de p + 1 à n faire
aij := aij − aip ∗ apj
fin pour
fin pour
fin si ;
p := p + 1
fin tant que
Fin.
2.1.3
Algorithmes avec recherche de pivot non nul
Si on rencontre un pivot nul sur la diagonale principale au cours
du processus de triangulation on doit procéder à des échanges de lignes
et/ou de colonnes pour ramener un pivot en position convenable (s’il
reste un élément non nul dans le coin sud-est). Alors ce n’est pas une
LU -décomposition de A que l’on obtient avec la méthode du pivot de
Gauss, mais une P̃LUP -décomposition (voir par exemple [AHU, BP]),
c’est-à-dire une LU -décomposition du produit à droite et à gauche de la
matrice A par des matrices de permutation.
De manière plus précise, si à l’issue de l’étape p − 1 du processus de
[p−1]
triangulation, on obtient un pivot nul (a pp = 0), alors de deux choses
[p−1]
l’une : ou bien a ij = 0 pour tous i, j ≥ p auquel cas le rang de A
est égal à p − 1, et le processus est terminé, ou bien on peut trouver des
2.1. Méthode du pivot de Gauss
63
[p−1]
entiers i et j ≥ p et tels que a ij 6= 0. Dans ce cas, une permutation
de lignes et/ou de colonnes doit intervenir pour remplacer le pivot nul
[p−1]
par l’élément a ij : ce qui revient à remplacer la matrice A[p−1] par
la matrice Eip (m) A[p−1] Ejp (n) où Ekl (h) désigne la matrice élémentaire obtenue à partir de Ih par échange des lignes k et l (ou, ce qui
revient au même, par échange des colonnes k et l). Cette opération,
qui prépare A[p−1] à subir avec succès l’étape p, n’altère pas les p − 1
premières lignes et les p − 1 premières colonnes de cette matrice. Plus
précisément, elle commute avec les opérations de type (( traitement d’un
pivot )) déjà effectuées (qui correspondent au produit à gauche par une
matrice triangulaire inférieure). Par exemple, si sur une matrice 6 × 6
on doit faire des échanges de lignes et de colonnes avant de traiter les
pivots no 3 et 5 on obtiendra la décomposition suivante
L[5] Q5 L[4] L[3] Q3 L[2] L[1] = L[5] L̃[4] L̃[3] L̃[2] L̃[1] Q5 Q3 = L5 Q5 Q3
où
L̃[4] = Q5 L[4] Q5 , L̃[3] = Q5 L[3] Q5 ,
L̃[2] = Q5 Q3 L[2] Q3 Q5 , L̃[1] = Q5 Q3 L[1] Q3 Q5
et
L5 = L[5] L̃[4] L̃[3] L̃[2] L̃[1] ,
et donc
A = Q3 Q5 (L5 )−1 U P5 P3 = P̃ L U P .
Ainsi le processus de triangulation de Gauss, lorsqu’une recherche
de pivots intervient, se ramène à un processus sans recherche de pivot
sur le produit à droite et à gauche de la matrice A par des matrices
de permutation. Cela montre aussi que l’algorithme du pivot de Gauss,
appliqué à la matrice A, donne, en même temps que sa P̃LUP -décomposition, le rang de la matrice A.
Notons aussi que la méthode avec recherche du pivot permet de calculer dans tous les cas le déterminant de la matrice A si elle est carrée.
Il suffit de garder en mémoire et de mettre à jour à chaque étape la
parité des permutations de lignes et de colonnes déjà effectuées.
LUP-décomposition d’une matrice surjective
Un cas particulier est donné par les matrices surjectives. Un pivot
non nul existe toujours sur la ligne voulue. Cela donne l’algorithme 2.3
page suivante.
64
2. Algorithmes de base en algèbre linéaire
Algorithme 2.3 LUP-décomposition d’une matrice surjective.
Entrée : Une matrice A = (aij ) ∈ Km×n surjective.
Sortie : La matrice A transformée (elle donne les matrices L et U comme
dans l’algorithme 2.1), la matrice de permutation P et sa signature e ∈
{−1, 1}.
Début
Variables locales : i, j, p ∈ N ; piv ∈ K ;
P := In ; e := 1 ;
pour p de 1 à m faire
piv := app ; j := p ;
si piv = 0 alors
tant que piv = 0 faire
j := j + 1 ; piv := apj ;
fin tant que ;
EchCol(A, p, j) ; EchCol(P, p, j) ; e := −e ;
# EchCol(A, p, j) est une procédure qui échange
# les colonnes p et j de la matrice A.
fin si ;
pour i de p + 1 à m faire
aip := aip /piv ;
pour j de p + 1 à n faire aij := aij − aip ∗ apj
fin pour
fin pour
fin pour
Fin.
Ainsi lorsqu’une matrice A ∈ Km×n est surjective (c’est-à-dire si
son rang est égal au nombre de ses lignes), on peut la décomposer en un
produit de trois matrices L, U, P où L ∈ Kn×n est une matrice triangulaire inférieure avec des 1 sur la diagonale, U ∈ Km×n une matrice
triangulaire supérieure et P ∈ Kn×n une matrice de permutation.
La LUP -décomposition permet de résoudre des problèmes comme le
calcul du déterminant ou la résolution d’un système d’équations linéaires. En effet, pour résoudre le système A x = b avec A = L U P , on
commence par résoudre le système L z = b puis le système U y = z et
enfin le système P x = y. Les deux premiers systèmes sont des systèmes
triangulaires que l’on peut résoudre par substitutions successives des
2.2. Méthode de Jordan-Bareiss
65
inconnues (en O(n2 ) opérations arithmétiques donc) et le dernier système est une simple permutation des inconnues. Enfin det A = ± det U
(selon la parité de la permutation représentée par la matrice P ).
Il faut remarquer qu’une matrice non surjective n’admet
pas toujours
0 0
de LUP -décomposition comme par exemple la matrice
.
1 1
Par ailleurs la LUP -décomposition d’une matrice surjective
n’est
pas
2 1 1
unique comme on peut le voir sur la matrice A =
qui
0 3 4
admet les deux LUP -décompositions A = L U P avec L = I2 , P = I3
et U = A, ou encore A = L U P avec
0 1 0
1
2 1
1 0
et P = 1 0 0 .
, U=
L=
0 −6 1
3 1
0 0 1
Notons enfin que la matrice U obtenue dans la décomposition A =
L U P est une matrice surjective et fortement régulière.
Résolution de systèmes linéaires et calcul de l’inverse
La méthode du pivot de Gauss permet de résoudre un ou plusieurs
systèmes linéaires associés à la même matrice, en triangulant la matrice
élargie aux seconds membres.
Dans sa variante (( Gauss-Jordan )), qui consiste à poursuivre le processus d’élimination de Gauss (( de bas en haut )) et (( de droite à gauche ))
sur les lignes de la matrice U de façon à annuler les éléments au dessus
de la diagonale principale, la méthode du pivot de Gauss sert également
à calculer l’inverse d’une matrice carrée inversible lorqu’on l’applique à
cette matrice élargie (à droite) avec la matrice unité de même ordre,
moyennant un coût légèrement supérieur qui fait passer la constante
dans O(n3 ) de 23 à 43 .
2.2
Méthode de Jordan-Bareiss
La méthode du pivot de Gauss est une méthode de traitement automatique des systèmes d’équations linéaires dont les coefficients et les
inconnues sont dans un corps donné K. Cette méthode fonctionne bien
dans le cas de matrices à coefficients dans un corps fini (et dans une
moindre mesure, dans le cas du corps Q). Mais hormis le cas des corps
finis, elle possède l’inconvénient majeur de nécessiter une simplification
66
2. Algorithmes de base en algèbre linéaire
systématique des fractions calculées si on ne veut pas voir la taille des
coefficients exploser, ce qui entraı̂ne souvent un temps de calcul prohibitif, par exemple lorsqu’on travaille avec un corps de fractions rationnelles
à plusieurs variables. En outre cette méthode utilise des divisions et ne
s’applique donc pas si la matrice a ses coefficients dans un anneau arbitraire.
Nous allons voir dans cette section que la méthode connue aujourd’hui sous le nom de (( méthode de Bareiss )), qui peut être considérée
comme une adaptation de la méthode du pivot de Gauss classique, permet dans une certaine mesure de pallier aux inconvénients présentés par
cette dernière.
La méthode de Bareiss (cf. [4], 1968) était connue de Jordan (cf.
[Dur]), et elle semble avoir été découverte par Dodgson (plus connu sous
le nom de Lewis Caroll) qui en a donné une variante dans [26]. Nous la
désignerons désormais sous le nom de méthode de Jordan-Bareiss.
Nous réservons le nom de méthode de Dodgson à la variante de Lewis
Caroll que nous exposons à la fin de la section.
La méthode de Jordan-Bareiss est valable dans le cas d’un anneau
intègre A où l’égalité peut être testée par un algorithme, et l’addition, la
multiplication et la division (( exacte )) (quand il y a un quotient exact)
peuvent être effectuées par des algorithmes. Cela signifie, pour la division
exacte, qu’il y a un algorithme prenant en entrée un couple (a, b) ∈
A2 , b 6= 0, et donnant en sortie l’unique élément x ∈ A vérifiant ax = b,
dans le cas où il existe.
2.2.1
Formule de Dodgson-Jordan-Bareiss et variantes
Soit A une matrice dans Am×n . Reprenant les relations données à la
[p]
(p) (p−1)
propriété 2.1.3, et puisque tous les coefficients a ij s’écrivent a ij /a pp
(relation (2.4)) avec le même dénominateur pour un p fixé, l’idée est de
calculer directement les numérateurs de manière récursive. L’équation
(2.1) se relit alors sous la forme
(p)
a ij
=
a
(p−1) (p−1)
(p−1) (p−1)
a pp − a ip a pj
ij
(p−2)
ap−1,p−1
(2.8)
C’est ce que nous appellerons la formule de Dodgson-Jordan-Bareiss.
On peut obtenir ce même résultat en appliquant l’identité de Sylvester
2.2. Méthode de Jordan-Bareiss
67
(1.10) de la proposition 1.1.7 page 9 à la matrice
Ap
A1..p,j
Ai,1..p
aij
avec p ∈ [1.. min(m, n) − 1] , i ∈ [p + 1..m] , j ∈ [p + 1..n] . Cela donne :
Proposition 2.2.1 (Formule de Dodgson-Jordan-Bareiss)
Soit A un anneau commutatif arbitraire. Pour toute matrice A = (aij ) ∈
Am×n , on a la relation :
(p)
avec les conventions usuelles a
(−1)
00
(p−1)
apj
(p−1)
aij
app
(p−2)
aij × ap−1,p−1 =
aip
(p−1)
(2.9)
(p−1)
(0)
= 1 et aij = aij .
On a également la variante suivante.
la formule (1.9)
Si l’on applique
Ap
A1..p,j
, on obtient :
de la proposition 1.1.7 à la matrice
Ai,1..p
aij
Proposition 2.2.2 (Formule de Bareiss à plusieurs étages)
Soit A un anneau commutatif arbitraire. Pour toute matrice A ∈ Am×n
et tout entier p ≥ 2, on a lorsque 1 ≤ r ≤ p − 1, p + 1 ≤ i ≤ m, et
p+1≤j ≤n :
(r)
a(r−1)
rr
p−r
(p)
aij =
ar+1,r+1
..
.
(r)
ap,r+1
(r)
ai,r+1
(r)
(r)
. . . ar+1,n ar+1,j
..
..
..
.
.
.
(r)
(r)
. . . ap,p
ap,j
...
(r)
ai,p
.
(2.10)
(r)
ai,j
Dans son article, Bareiss a remarqué qu’on pouvait utiliser cette
(p)
identité avec p − r = 2 pour calculer les aij de proche en proche,
lorsque l’anneau est intègre et possède un algorithme de division exacte.
En fait la (( méthode de Bareiss )) couramment utilisée aujourd’hui
est plutôt basée sur la première formule (celle de Dodgson-Jordan-Ba(p)
reiss). L’équation (2.8) permet en effet de calculer les aij de proche en
proche.
La méthode de Jordan-Bareiss est donc une adaptation de la méthode
du pivot de Gauss qui garantit, tout au long du processus de triangulation de la matrice traitée, l’appartenance des coefficients à l’anneau de
base. L’efficacité de cet algorithme tient à ce que les coefficients calculés
sont tous des déterminants extraits de la matrice initiale, et donc restent
de taille raisonnable pour la plupart des anneaux usuels.
68
2. Algorithmes de base en algèbre linéaire
Algorithme de Jordan-Bareiss
En utilisant la relation (2.8) on obtient l’algorithme de Jordan-Bareiss 2.4, dans sa version de l’élimination à un seul étage déchargée de
la recherche du pivot.
(−1)
(p−1)
Rappelons les conventions a 00 = 1 et a pp
= 0 pour p >
inf(m, n).
Algorithme 2.4 Algorithme de Jordan-Bareiss
Entrée : Une matrice A = (aij ) ∈ Am×n . L’anneau A est supposé intègre
avec un algorithme de division exacte.
Sortie : La matrice A transformée. Si les r premiers mineurs principaux
dominants sont non nuls, et si le (r+1)-ème est nul, elle contient en position
(p)
(i, j) le mineur ai,j avec p = inf(r, i−1, j −1). L’entier r est aussi calculé.
Si en outre r = rg(A) on retrouve facilement la LU-décomposition de A
à partir de la sortie, comme expliqué avant l’exemple 2.2.3.
Début
Variables locales : i, j, p ∈ N ; piv, den, coe ∈ A ;
p := 1 ; den := 1 ; r := inf(m, n) ;
tant que p < inf(m, n) faire
piv := app ;
si piv = 0 alors p := inf(m, n) ; r := p − 1 sinon
pour i de p + 1 à m faire
coe := aip ;
pour j de p + 1 à n faire
aij := (piv ∗ aij − coe ∗ apj ) / den
fin pour
fin pour
fin si ;
p := p + 1 ;
den := piv
fin tant que
Fin.
On retrouve facilement la LU -décomposition de A à partir de la
matrice retournée par l’algorithme précédent en utilisant les formules
(2.4) (propriété 2.1.3) et (2.7) page 59 : notons cij les coefficients de
cette matrice ; alors pour la matrice L on a lij = cij /cjj si 1 ≤ j <
2.2. Méthode de Jordan-Bareiss
69
i ≤ m (lij = δij sinon) et pour la matrice U on a uij = cij /ci−1,i−1 si
1 ≤ i ≤ j (uij = 0 sinon). On peut le voir sur l’exemple suivant.
Exemple 2.2.3 Dans cet exemple on reprend la matrice M3 de l’exemple 2.1.5
et on donne ses transformées par les algorithmes de Jordan-Bareiss et de Gauss.
−73 −53 −30
45 −58
21 −54 −11
0
−1
72 −59
52 −23
77
M3 =
33
55
66 −15
62
−41 −95 −25
51 −54
14
55
35
−5
25
−73
−53
−30
45
−58
21
5055
1433
−945
1291
72
8123
272743
2940
243716
33 −2266 220594 2911532 −3038698
−41
4762
52277 3660124
0
14 −3273
83592 3227536
0
−73
−53
−30
45
−58
−21 −5055 −1433
945
−1291
73
73
73
73
73
−72
8123
272743
196
243716
73
5055
5055
337
5055
−33 −2266 220594 2911532 −3038698
73
5055
272743
272743
272743
41
4762
52277
1193
0
73
5055
272743
949
−14 −1091
83592
1052
0
73
1685
272743
949
Comparons l’algorithme de Jordan-Bareiss à l’algorithme du pivot
de Gauss dans le cas de l’anneau A = Z[X, Y ].
Lorsqu’on utilise l’algorithme du pivot de Gauss dans le corps des
fractions FA = Q(X, Y ) sans réduire les fractions au fur et à mesure
qu’elles sont calculées (ce qui est très coûteux), il n’est pas difficile de voir
que les degrés des numérateurs et dénominateurs ont un comportement
exponentiel. Avec l’algorithme de Jordan-Bareiss, par contre, les degrés
ont seulement une croissance linéaire.
Remarque. Dans le cas non intègre, le fonctionnement de l’algorithme de Jordan-Bareiss sans recherche du pivot reste possible si tous les
(p−1)
mineurs principaux rencontrés app
au cours du processus de triangulation sont non diviseurs de 0, et si les divisions exactes :
(p−1) (p−1)
(p−1) (p−1)
aij
− aip apj
(p−2)
ap−1,p−1
app
peuvent se faire algorithmiquement .
70
2. Algorithmes de base en algèbre linéaire
Cette condition est satisfaite lorsqu’on remplace la matrice carrée A
par sa matrice caractéristique A − XIn ∈ A [X]n×n où tous les pivots
rencontrés sont des polynômes unitaires (au signe près). Donc l’algorithme de Jordan-Bareiss appliqué à A − XIn ne fait intervenir que la
structure d’anneau de A et ne nécessite aucune division dans A. C’est
l’objet du paragraphe suivant.
2.2.2
Cas d’un anneau commutatif arbitraire : méthode
de Jordan-Bareiss modifiée
C’est la méthode de Jordan-Bareiss appliquée à la matrice caractéristique A − XIn d’une matrice carrée A ∈ An×n . Les coefficients
de A − XIn sont dans l’anneau A [X]. Même si A n’est pas intègre,
les divisions exactes requises sont ici des divisions par des polynômes
unitaires qui ne nécessitent par conséquent aucune division dans A, mais
uniquement des additions, soustractions, et multiplications. En particulier, aucune permutation de lignes ou de colonnes n’intervient au cours
du processus de triangulation.
La méthode de Jordan-Bareiss modifiée permet donc de calculer le
polynôme caractéristique de la matrice A, et par conséquent son déterminant, son adjointe, et, au cas où elle est inversible, son inverse.
Cette méthode a été proposée en 1982 par Sasaki & Murao [80]. Les
auteurs remarquent également que dans un calcul de base de l’algorithme
(du type (( produit en croix divisé par le pivot précédent ))) :
f (X) :=
a(X) c(X) − b(X) d(X)
,
e(X)
les degrés en X sont égaux à k ou k + 1 pour f, à k ou k − 1 pour
a, b, c, d et à k − 1 pour e. On peut donc se passer de calculer les coefficients des monômes de degré < k−1 dans ac−bd et le calcul du quotient
ne doit pas non plus s’encombrer des termes de degrés < k − 1 dans
les restes successifs (pour l’algorithme usuel de division des polynômes).
Ceci conduit précisément aux résultats suivants.
– Les coefficients des monômes de degré k − 1 à 2k dans le produit
de deux polynômes de degré k se calculent (en utilisant la méthode usuelle) en k 2 + 2 k − 2 opérations arithmétiques.
– La division exacte d’un polynôme de degré 2k par un polynôme
unitaire de degré k − 1 se calcule (en utilisant la méthode usuelle)
en k 2 + 3 k − 1 opérations arithmétiques.
2.2. Méthode de Jordan-Bareiss
71
c
On en déduit qu’une affectation f := a d−b
dans l’algorithme de Jore
dan-Bareiss modifié, lorsque e = e(X) est le pivot unitaire de degré
k − 1, consomme 3 k 2 + O(k) opérations arithmétiques dans l’anneau
de base (et en tout cas au plus 3 k 2 + 8 k − 4).
Pour l’ensemble de l’algorithme on obtient un nombre total d’opérations arithmétiques inférieur à
n
X
(3 k 2 + 8 k − 4) (n − k + 1)2 ≤
k=2
1 5 7 4 7 3
n + n + n .
10
6
3
Proposition 2.2.4 Soit A ∈ An×n une matrice carrée sur un anneau
commutatif arbitraire. L’algorithme de Jordan-Bareiss appliqué à la matrice caractéristique A−X In s’exécute (en utilisant la méthode usuelle)
1 5
en 10
n + O(n4 ) opérations arithmétiques dans l’anneau A.
2.2.3
La méthode de Dodgson
La méthode de Dodgson est une variante élégante et symétrique de
la méthode de Jordan-Bareiss. Cependant son but n’est pas le calcul de
la LU -décomposition d’une matrice, mais seulement celui de ses mineurs
connexes, c’est-à-dire les mineurs ah..k
i..j (avec j−i = k−h). En particulier
elle peut être utilisée pour le calcul du déterminant d’une matrice carrée.
Une variante de la formule 2.9 (après un échange de lignes et un
échange de colonnes) est la formule suivante concernant les mineurs
connexes
ah..k−1
ah+1..k
i..j−1
i..j−1
h..k
ah+1..k−1
·
a
=
(2.11)
i..j
i+1..j−1
h+1..k
ah..k−1
a
i+1..j
i+1..j
Cela donne les affectations correspondantes dans l’algorithme de Dodgson. Mais ce dernier fonctionne uniquement si tous les mineurs connexes
appelés à servir de dénominateur sont non nuls : contrairement à la méthode du pivot de Gauss et à la méthode de Jordan-Bareiss, la méthode
de Dodgson ne possède pas de variante connue efficace dans le cas où
une affectation x := 0/0 est produite par l’algorithme( 4 ).
4. Lewis Caroll propose dans sa communication d’opérer des permutations circulaires sur les lignes et les colonnes de la matrice. Voici un contre-exemple
montrant que
1 0 1 1
1 0 1 0
la méthode de Dodgson ne s’applique pas toujours. La matrice A =
0 1 0 0
1 0 0 1
est une matrice inversible de déterminant 1, lequel ne peut pas se calculer par la
méthode de Lewis Carrol, même lorsqu’on effectue des permutations circulaires de
lignes et de colonnes.
72
2. Algorithmes de base en algèbre linéaire
Pour voir plus clairement ce que signifie l’équation de Lewis Caroll
(2.11) appelons B la matrice extraite Ai+1..j−1,h+1..k−1, et notons p,
q, u, v les indices i + 1, j − 1, h + 1, k − 1. L’équation se réécrit alors :
ai,h
| B | · Ap..q,h
aj,h
ai,h
Ap..q,h
Ai,u..v
B
Aj,u..v
B
Ai,u..v
·
Aj,u..v
B
ai,k
Ap..q,k =
aj,k
Ap..q,h
Ap..q,k
−
aj,h
aj,k
Ai,u..v
B
·
B
Aj,u..v
ai,k
Ap..q,k
Un exemple :
c2
d2
b1
c1
d1
b1
c3
c1
·
d3
d1
e1
b2
c2
d2
b2
c2
d2
e2
b3
c2
c 3 · d2
d3
e2
b3
c3
d3
e3
c3
d3
e3
b4
c4
=
d4
e4
c4
c1
d4 − d1
e4
e1
c2
d2
e2
c3
b2
d3 · c2
e3
d2
b3
c3
d3
b4
c4
d4
Dans la méthode de Jordan-Bareiss sans recherche de pivot on calcule
(p)
à l’étape no p tous les mineurs aij ((i, j > p) d’une matrice A. Dans la
[p]
(p)
(p−1)
méthode du pivot de Gauss on calcule les quotients aij = aij /ap,p .
Si la matrice a une (( structure interne )) comme dans le cas des matrices
de Hankel ou de Toeplitz la structure est perdue dès la première étape.
Dans la méthode de Dodgson, on calcule à l’étape no p tous les mineurs connexes d’ordre p + 1 de la matrice A. Il s’ensuit que dans le
cas d’une matrice structurée, les matrices intermédiaires calculées par
la méthode de Dodgson sont également structurées. Ceci diminue très
sérieusement le nombre d’opérations arithmétiques à effectuer et le fait
passer de O(n3 ) à O(n2 ).
Dans le cas d’un anneau intègre où les divisions exactes sont faisables
par un algorithme, on obtient les mêmes avantages que dans l’algorithme
de Jordan-Bareiss concernant la taille des coefficients intermédiaires.
Algorithme de Dodgson pour une matrice de Hankel
Nous donnons ici une version précise de l’algorithme de Dodgson
pour les matrices de Hankel dont tous les mineurs connexes sont non
nuls. C’est l’algorithme 2.5 page ci-contre.
L’entrée est une liste L = (ai ) contenant les m + n − 1 coefficients de
la matrice de Hankel initiale H ∈ Am×n (hi,j = ai+j−1 ). La sortie est
un tableau T = (tr,j ) (r = 0, . . . , inf(m, n), j = r, . . . , m + n − r) qui
2.2. Méthode de Jordan-Bareiss
73
contient tous les mineurs connexes de la matrice H, calculés en suivant
l’algorithme de Dodgson. Pour l’initialisation, sur la ligne 0 il y a des 1
(les (( mineurs connexes d’ordre 0 ))) et sur le ligne 1 les coefficients de
H (les (( mineurs connexes d’ordre 1 ))). Sur la ligne r ≥ 2 il y a les coefficients de la matrice de Hankel formée par les mineurs connexes d’ordre
r de H. Dans la colonne j il y a les déterminants des sous-matrices
carrées de H qui ont le coefficient aj sur leur diagonale ascendante.
Algorithme 2.5 Algorithme de Dodgson pour une matrice de
Hankel
Entrée : Deux entiers m, n ∈ N et une liste L = (ai ) ∈ Am+n−1 . Cette
liste contient les coefficients d’une matrice de Hankel H ∈ Am×n . L’anneau
A est supposé intègre avec un algorithme de division exacte.
Sortie : Un tableau T = (tr,j ) rempli d’éléments de A pour r ∈
{0, . . . , inf(m, n)}, j ∈ {r, . . . , m + n − r}. Il contient sur la ligne r les
mineurs connexes d’ordre r de la matrice H, supposés tous non nuls.
Début
Variables locales : r, j, q ∈ N ;
q := inf(m, n) ;
T :=TableauVide(0..q, 1..m + n − 1) ;
# on a créé T tableau vide de taille voulue
pour j de 1 à m + n − 1 faire
t0,j := 1 ; t1,j := aj ;
fin pour ;
# fin de l’initialisation
pour r de 1 à q − 1 faire
pour j de r + 1 à m + n − r − 1 faire
tr+1,j := (tr,j−1 tr,j+1 − t2r,j )/tr−1,j
fin pour
fin pour
Fin.
L’algorithme est pratiquement le même dans le cas d’une matrice de
Toeplitz Z (il suffit de changer le signe dans l’affectation de tr+1,j ) et
il peut s’appliquer pour le calcul du polynôme caractéristique.
Exemples 2.2.5
Dans le premier exemple, on considère la matrice de Hilbert d’ordre 5, qui est un
exemple classique de matrice de Hankel mal conditionnée (le déterminant de la ma-
74
2. Algorithmes de base en algèbre linéaire
trice, 1/266716800000, est l’inverse d’un entier très grand).
A=
1
1/2
1/3
1/4
1/5
1/2
1/3
1/4
1/5
1/6
1/3
1/4
1/5
1/6
1/7
1/4
1/5
1/6
1/7
1/8
1/5
1/6
1/7
1/8
1/9
Voici alors la sortie de l’algorithme de Dodgson 2.5 (on a supprimé la ligne des 1) :
1
2
1
12
1
1
3
1
72
1
2160
1
4
1
240
1
43200
1
6048000
1
5
1
600
1
378000
1
423360000
1
266716800000
1
6
1
1260
1
2116800
1
10668672000
1
7
1
2352
1
8890560
1
8
1
4032
1
9
Voici ensuite un exemple de la sortie de l’algorithme avec une matrice de Hankel
carrée d’ordre 7 à coefficients entiers (lisibles sur la première ligne) :
1
7
−42
2.3
7
−42
−330
1
13
−85
−1165
2
−2
24
373
−1671
2
4
2
−85
−442
−41
4
−10
−14
17
−119
259
870
3
11
13
−23
−42
889
5
−16
6
−1
157
3
26
4
−41
7
−43
−175
2
24
4
Méthode de Hessenberg
Toutes les matrices considérées ici sont à coefficients dans un corps
commutatif K.
Matrices quasi-triangulaires
Définition 2.3.1 Une matrice carrée H = (hij ) ∈ Kn×n (n ∈ N∗ ) est
dite quasi-triangulaire supérieure (resp. quasi-triangulaire inférieure) si
hij = 0 dès que i − j ≥ 2 (resp. dès que j − i ≥ 2). On dit encore que
H est une matrice de Hessenberg.
2.3. Méthode de Hessenberg
Une matrice quasi-triangulaire supérieure H
la forme :
h11 h12 . . . h1,n−1
h21 h22 . . . h2,n−1
..
.
H=
0 h32
..
.
.
..
..
..
.
.
0 . . . 0 hn,n−1
75
est donc une matrice de
h1n
h2n
h3n
(2.12)
..
.
hnn
On démontre par une récurrence immédiate sur k (1 ≤ k ≤ n) la
propriété suivante des matrices de Hessenberg :
Proposition 2.3.2 Soit H = (hij ) une matrice de Hessenberg (supérieure ou inférieure). On désigne par Hk (1 ≤ k ≤ n) la sous-matrice
principale dominante d’ordre k de H et par Dk le déterminant de
Hk . On pose D0 = 1. La suite (Dk )1≤k≤n (des mineurs principaux
dominants de H) vérifie alors la relation de récurrence :
Dk = hkk Dk−1 +
k−1
X
(−1)k−i hk,k−1 hk−1,k−2 . . . hi+1,i hik Di−1 ·
i=1
Pour le voir, il suffit de développer Dk suivant la dernière ligne (resp.
la dernière colonne) de Hk si celle-ci est une matrice quasi-triangulaire
supérieure (resp. inférieure).
Appliquant ce résultat à la matrice H − XIn , elle-même quasi-triangulaire, dont les mineurs principaux dominants sont les polynômes
caractéristiques Pk (X) des sous-matrices principales dominantes Hk
de H (1 ≤ k ≤ n), on obtient les relations de récurrence suivantes dites
relations de Hessenberg permettant de calculer de proche en proche les
polynômes caractéristiques Pk (X) de Hk pour 2 ≤ k ≤ n sachant que
P0 (X) = 1, P1 (X) = h11 − X ; et
(
Pk (X) =
(hkk − X) Pk−1 (X) +
i
Pk−1 hQk
(−h
)
h
P
(X)
j,j−1
ik i−1
i=1
j=i+1
(2.13)
La méthode de Hessenberg
Elle consiste à calculer le polynôme caractéristique d’une matrice
carrée A d’ordre n ≥ 2, dont les éléments aij appartiennent à un
corps K, en la réduisant à la forme (2.12) c’est-à-dire à une matrice
76
2. Algorithmes de base en algèbre linéaire
de Hessenberg H semblable à A dont les éléments hij appartiennent
également à K.
Algorithme 2.6 Algorithme de Hessenberg (K est un corps)
Entrée : Un entier n ≥ 2 et une matrice A = (aij ) ∈ Kn×n .
Sortie : Le polynôme caractéristique de A : PA (X).
Variables locales : jpiv, ipiv, iciv, i, m ∈ N ; piv, c, ∈ K ;
H := (hij ) ∈ Kn×n : les matrices transformées successives de A ;
P = (Pi ) : liste des polynômes caractéristiques successifs dans K[X] ;
Début
P0 := 1 ; H := A ;
# Initialisations
# Réduction de H à la forme de Hessenberg
pour jpiv de 1 à n − 2 faire
ipiv := jpiv + 1 ; iciv := ipiv ; piv := hiciv,jpiv ;
tant que piv = 0 et iciv < n faire
iciv := iciv + 1 ; piv := hiciv,jpiv
fin tant que ;
si piv 6= 0 alors
si iciv > ipiv alors
EchLin(H, ipiv, iciv) ; # Echange de lignes
EchCol(H, ipiv, iciv) # Echange de colonnes
fin si ;
pour i de iciv + 1 à n faire
c := hi,jpiv /piv ;
AjLin(H, ipiv, i, −c) ; # Manipulation de lignes
AjCol(H, i, ipiv, c) # Manipulation de colonnes
fin si
fin pour ;
# Calcul du polynôme caractéristique
pour m de 1 à n faire
Pm := (hmm − X) · Pm−1 ; c := 1 ;
pour i de 1 à m − 1 faire
c := −c · hm−i+1,m−i ; Pm := Pm + c · hm−i,m · Pm−i−1
fin pour
fin pour ;
PA (X) := Pn (X)
# le polynôme caractéristique de A.
Fin.
2.3. Méthode de Hessenberg
77
Pour ce faire, on applique la méthode du pivot de Gauss aux lignes
de la matrice donnée A en prenant comme pivots les éléments sousdiagonaux de la matrice traitée, et en prenant bien soin d’effectuer les
transformations (( inverses )) sur les colonnes de A pour que la matrice
et sa transformée soient semblables.
Plus précisément, l’étape p (1 ≤ p ≤ n − 2) consiste tout d’abord à
voir si l’élément en position (p + 1, p) est nul, auquel cas il faut chercher
un élément non nul au-dessous de lui (sur la colonne p) : si un tel élément
n’existe pas, on passe à l’étape suivante p+1. Sinon par une permutation
de lignes, on ramène le pivot non nul au bon endroit, c’est-à-dire à la
position (p + 1, p), ce qui revient à multiplier à gauche la matrice traitée
par la matrice de permutation Ei,p+1 (i > p + 1) obtenue en permutant
les lignes i et p + 1 (ou les colonnes i et p + 1, ce qui revient au même)
de la matrice unité In . On multiplie à droite par la même matrice de
permutation (qui est ici égale à son inverse) afin que la matrice obtenue
à l’issue de chaque étape reste semblable à la matrice de départ.
Le pivot non nul étant alors au bon endroit, on achève l’étape p en
utilisant ce pivot pour faire apparaı̂tre des zéros au-dessous de lui dans
sa colonne, ce qui revient à multiplier à gauche la matrice traitée par
une matrice du type :
1 ...
.. . .
.
.
0 ...
L=
0 ...
0 ...
.. . .
.
.
p
↓
p+1
↓
0
..
.
0
..
.
1
0
0
1
0 lp+2,p+1
..
..
.
.
0 ... 0
ln,p+1
0 ... 0
..
..
.
.
0 ... 0
← p
0 ... 0
← p+1
1 ... 0
← p+2
.. . . ..
. .
.
0 ... 1
et multiplier ensuite à droite la matrice traitée par la matrice L−1 (obtenue à partir de L en changeant le signe des éléments sous la diagonale).
Il est clair que ces opérations, qui définissent l’étape p et qui sont
effectuées sur la matrice provenant de l’étape précédente (appelonsla A(p−1) ), n’affectent pas les p − 1 premières colonnes de A(p−1) et
donnent une matrice A(p) semblable à A(p−1) .
78
2. Algorithmes de base en algèbre linéaire
Ceci donne l’algorithme de Hessenberg : Primo, calculer, à l’aide de
la procédure décrite ci-dessus, une matrice de Hessenberg H semblable
à la matrice donnée A ∈ Kn×n . Secundo, calculer le polynôme caractéristique de H (qui est aussi celui de A) en utilisant les relations de
Hessenberg (2.13).
On obtient ainsi l’algorithme 2.6 page 76. Dans cet algorithme la procédure AjLin(H, i, j, c) opère une manipulation de lignes sur la matrice
H : on ajoute à la ligne j la ligne i multipliée par c.
Exemple 2.3.3 Dans cet exemple on montre une ((
entiers et sa réduction à la forme de Hessenberg.
−3
3
0 −2 −3 1
1
2
1
1
2 3
2
2
3 −3
3 0
−2
0
1 −2
0 0
A :=
3 −3
3
3
2 3
3
2
3 −3
1 2
−3
3
3
2
3 1
0 −1 −3 −1 −1 1
petite )) matrice à coefficients
−2
2
−2
−1
0
−1
−2
−1
2
0
−3
2
3
−2
0
−3
Voici la liste des lignes de la forme réduite de Hessenberg. Nous n’avons pas indiqué
les 0 en position (i, j) lorsque i > j + 1.
−122 −26680 −4080544 4747626797 109259596132466
−3 , 7 ,
,
,
,
,
,2
7
6639
1522773 1757764263 234026268743849
87 24037 17521799 1473144559
1 , 11 ,
,
,
,
, 2, 0
7 13278 7613865 1757764263
−415 −54333 −1294739 689762552 −2270125812893340
7,
,
,
,
,
, −3
7
4426
2537955 585921421
234026268743849
13278 1670911 13199211 −11965124859 79329636778655517 107
,
,
,
,
,
49
30982
1973965
4101449947
1638183881206943
7
−17765685 −2532182353 2798215923779 6108776229950083011 25553
,
,
,
,
19589476
2246597766 2593288209346 1035800265460275674 4426
−2593288209346 954443884297868 17689012510838333947 600431
,
,
,
1288243116405 1487042200034055 28283244709037870895 2537955
13198847530884339751 −10729114442300396518997896 −64207585234
,
,
5149558673799888615
2056815059005858366341435
26366463945
306462654496531416683963262645 481086736521535
,
54768294462168235404375334801 234026268743849
On voit apparaı̂tre des fractions de grande taille : les coefficients de la matrice initiale
sont majorés par 3 en valeur absolue, et le numérateur le plus grand dans la matrice
transformée est environ égal à 361,8 . Une étude expérimentale dans les mêmes conditions avec des matrices carrées d’ordre n variant entre 8 et 32 donne une taille des
2.3. Méthode de Hessenberg
79
coefficients intermédiaires de type quadratique : le numérateur ou dénominateur de
2
taille maximum est de l’ordre de 32(n−2) . Il s’agit donc ici d’un cas typique d’une
méthode qui ne s’applique efficacement de manière directe, en calcul formel, que dans
le cas d’un corps fini.
Remarque 2.3.4 Une matrice triangulaire A ∈ Kn×n est une matrice
de Hessenberg particulière qui a ses valeurs propres dans K. Mais une
matrice de Hessenberg qui a ses valeurs propres dans K n’est pas nécessairement triangulaire ni semblable
à
une matrice triangulaire, comme
1 0 1
on le voit avec la matrice 1 1 1 .
0 0 1
Nombre d’opérations arithmétiques
• La phase 1 de réduction à la forme de Hessenberg est composée de
n − 2 étapes. Chacune des étapes p (1 ≤ p ≤ n − 2) comporte un
travail sur les lignes avec (n − p − 1) divisions, (n − p − 1) (n − p) multiplications et autant d’additions. L’opération inverse sur les colonnes
comporte (n − p − 1) n multiplications et autant d’additions.
Ce qui donne 61 (n − 1) (n − 2) (5n + 3) 56 n3 multiplications/divisions
et 56 n (n − 1) (n − 2) additions/soustractions, c’est-à-dire un nombre
total d’opérations arithmétiques dans K qui est asymptotiquement de
l’ordre de 53 n3 .
• La phase 2 qui consiste à calculer les polynômes caractéristiques
Pk (X) (2 ≤ k ≤ n) des sous-matrices principales dominantes de la réduite de Hessenberg s’effectue par récurrence sur k à partir de P0 (X) =
1 et P1 (X) = h11 − X. Si l’on désigne par S(k) le nombre de multiplications/divisions (resp. additions/soustractions) permettant de calculer
le polynôme caractéristique Pk (X) de Hk , l’utilisation des relations de
Hessenberg conduit aux relations de récurrence suivantes, vraies pour
2≤k≤n :
k−1
X
S(k) = S(k − 1) + (k − 1) +
i
i=1
pour les multiplications/divisions
k−2
X
S(k)
=
S(k
−
1)
+
2
(k
−
1)
+
i
i=1
pour les additions/soustractions
80
2. Algorithmes de base en algèbre linéaire
c’est-à-dire, dans les deux cas (que S(k) désigne le nombre de multiplications/divisions ou celui des additions/soustractions) :
S(k) = S(k − 1) +
1
(k − 1) (k + 2) .
2
Comme S(1) = 0, cela donne par sommation : 61 n (n − 1) (n + 4)
multiplications/divisions et autant d’additions/soustractions dans
le corps K (la phase de quasi-triangularisation est donc la plus coûteuse,
asymptotiquement cinq fois plus chère en nombre d’opérations arithmétiques que la phase de calcul du polynôme caractéristique de la matrice
quasi-triangulaire).
D’où le résultat :
1
6
n3
Proposition 2.3.5 L’algorithme de Hessenberg calcule les polynômes
caractéristiques de toutes les sous-matrices principales dominantes d’une
matrice n × n sur un corps K avec moins de (n + 1) (n − 1)2 multiplications/divisions et n (n − 1)2 additions/soustractions, soit en tout
2 n3 − 3 n2 + 1 opérations arithmétiques.
Remarque 2.3.6 Ce que la méthode de Hessenberg gagne en complexité arithmétique par rapport aux précédentes méthodes de calcul
du polynôme caractéristique, elle le perd sur un aspect essentiel au plan
pratique. Celui de l’absence de contrôle raisonnable de la taille des coefficients intermédiaires. La formule permettant d’exprimer dans la méthode du pivot de Gauss chaque coefficient intermédiaire comme quotient de
deux déterminants extraits de la matrice de départ (voir propriété 2.1.3
page 58), ne s’applique plus dans le processus de quasi-triangularisation
de Hessenberg. En effet les transformations subies par les lignes sont ici
suivies par des transformations inverses sur les colonnes. Et on ne dispose
pas actuellement pour la méthode de Hessenberg, pourtant la plus rapide
en temps séquentiel si on ne prend en compte que le nombre d’opérations
arithmétiques, d’une formule analogue qui permette de conclure sur la
question de la taille des coefficients intermédiaires. Cela est confirmé
par les résultats expérimentaux que nous avons pu avoir (voir l’exemple
2.3.3 et le chapitre 11). Dans le cas de matrices à coefficients entiers, il y
a la possibilité de remédier à ce problème en utilisant le calcul modulaire
(cf. section 1.6 page 32).
Remarque 2.3.7 Signalons l’existence d’une version modifiée récente
de l’algorithme de Hessenberg sur un anneau intègre, développée dans
2.4. Méthode d’interpolation de Lagrange
81
[3], qui permet de garder les coefficients intermédiaires, tout au long des
calculs, dans l’anneau de base, supposé intègre. Elle semble bien adaptée
au calcul modulaire sur les anneaux de polynômes à coefficients entiers.
2.4
Méthode d’interpolation de Lagrange
Elle ramène le calcul du polynôme caractéristique d’une matrice carrée A ∈ An×n au calcul de n + 1 déterminants.
On est donc supposé être dans une situation où le calcul des déterminants ne pose pas problème : cela peut être le cas par exemple lorsque
la méthode du pivot de Gauss ne se heurte pas à des problèmes graves
de simplification de fractions, ou lorsque l’on dispose d’un algorithme
efficace et sans division pour le calcul des déterminants (comme celui du
développement suivant une ligne ou une colonne si la matrice donnée est
creuse). La méthode consiste à appliquer la formule d’interpolation de
Lagrange au polynôme caractéristique PA (X) = det(A − XIn ), c’est-àdire la formule bien connue :
i6Y
=k
n
X
X − xi
P (xi )
PA (X) =
xk − xi
i=0
i∈{0,...,n}
où x0 , x1 , . . . , xn sont n + 1 éléments distincts de A, avec la restriction
suivante : les (xi − xj ) (pour i 6= j) doivent être non diviseurs de zéro
dans A et on doit disposer d’un algorithme de division exacte par les
(xi − xj ) dans A.
C’est par exemple le cas avec xi = i × 1A lorsque A est de caractéristique nulle, ou finie étrangère à n!, ou plus généralement lorsque la
division exacte par les les entiers de A inférieurs ou égaux à n (c’està-dire les éléments 1A , 1A + 1A , . . ., n 1A ), si elle est possible, est unique
et réalisable par un algorithme.
En effet, si l’on choisit xk = k pour 0 ≤ k ≤ n, la formule d’interpolation s’écrit :
i6Y
=k
n
X
det(A
−
k
I
)
n
(−1)k
det(A − XIn ) =
(X − i)
k ! (n − k) !
k=0
i∈{0,...,n}
ce qui exige la possibilité d’effectuer des divisions (exactes) par les entiers
de A inférieurs ou égaux à n.
82
2. Algorithmes de base en algèbre linéaire
En fait, PA (X) = (−1)n X n + Q(X) avec deg(Q) ≤ n − 1 et il
suffit d’appliquer la méthode d’interpolation de Lagrange à Q(X), ce
qui revient à calculer la valeur de PA en n points au lieu de n + 1.
Le nombre d’opérations arithmétiques lors de l’exécution de cet algorithme est à peu près n fois celle du calcul d’un déterminant d’ordre
n. Si, pour le calcul des déterminants P (xi ) = det(A − xi In ), on choisit
d’utiliser l’algorithme du pivot de Gauss (ou l’algorithme de JordanBareiss si on est dans une situation où il s’avère être préférable à l’algorithme de Gauss 5 ), on obtient donc pour la méthode d’interpolation
de Lagrange un O(n4 ). En fait les meilleurs algorithmes sans division
dont on dispose actuellement pour calculer les déterminants passent par
le calcul du polynôme caractéristique, ce qui rend caduque la méthode
d’interpolation de Lagrange. Celle-ci, avec le calcul du déterminant dans
l’anneau de base abandonné à la sagacité de Maple, sera comparée à
ces autres algorithmes sur quelques exemples testés sur machine (voir
chapitre 11).
2.5
Méthode de Le Verrier et variantes
Cette méthode, découverte en 1848 par l’astronome français Le Verrier [65], repose sur les relations de Newton entre les sommes de Newton
et les polynômes symétriques élémentaires dans l’algèbre des polynômes
à n indéterminées x1 , . . . , xn sur un anneau commutatif A.
De manière générale, la méthode de Le Verrier appliquée à une matrice n × n réclame qu’on soit dans un anneau où les entiers 1, 2, . . . , n
sont non diviseurs de zéro. Les seules divisions requises sont des divisions
exactes par l’un de ces entiers.
2.5.1
Le principe général
La méthode de Le Verrier consiste précisément à déduire le calcul
des coefficients du polynôme caractéristique du calcul de ses sommes de
Newton. Celles-ci sont en effet égales aux traces des puissances de A
comme le montre le lemme 1.5.6 page 30.
Rappelons que l’anneau de base n’a pas besoin d’être intègre, puisque
les sommes de Newton peuvent être définies sans recours aux valeurs
propres, en utilisant les équations (1.23) page 29 (cf. définition 1.5.5).
5. Dans les deux cas, nous avons vu que le nombre d’opérations arithmétiques est
O(n3 ).
2.5. Méthode de Le Verrier et variantes
83
Ceci donne l’algorithme 2.7.
Algorithme 2.7 Algorithme de Le Verrier
Début
Étape 1 :
Calculer les puissances A2 , . . . , An−1 de la matrice A ainsi que les
éléments diagonaux de la matrice An ;
Étape 2 :
Calculer les traces des matrices A1 , A2 , . . . , An ;
Étape 3 :
Calculer les coefficients pk (1 ≤ k ≤ n) en utilisant les équations (1.23).
Fin.
Nombre d’opérations arithmétiques
Pour un anneau A fixé par le contexte, nous noterons µM (n) le
nombre d’opérations arithmétiques nécessaires pour la multiplication de
deux matrices carrées d’ordre n (on trouvera une définition plus précise
dans la notation 7.2.1 page 195). Lorsqu’on utilise la méthode usuelle de
multiplication des matrices carrées on a µM (n) = n2 (2n − 1).
Pour l’algorithme de Le Verrier, le compte est le suivant :
– l’étape 1 utilise (n−2) µM (n)+n (2n−1) opérations arithmétiques
n
X
– les étapes 2 et 3 utilisent n2 + 2
(k − 1) = 2 n2 − n opérations.
k=1
Proposition 2.5.1 Le nombre total d’opérations arithmétiques lors de
l’exécution de l’algorithme de Le Verrier, si on utilise la multiplication
usuelle des matrices, est 2 n4 + O(n3 ) = O(n4 ) (précisément égal à
2 n (n − 1/2) (n2 − 2n + 2)).
Des algorithmes dérivés de l’algorithme de Le Verrier ont été proposés par de nombreux auteurs, avec des améliorations concernant la
complexité (cf. [22, 32, 77, 84] et [FF]). Nous les étudions dans la suite
de cette section et dans le chapitre 9.
84
2.5.2
2. Algorithmes de base en algèbre linéaire
Méthode de Souriau-Faddeev-Frame
Cette méthode, découverte séparément par Faddeev & Sominskii
[FS], Souriau [84] et Frame [31], est une amélioration astucieuse de l’algorithme de Le Verrier.
Comme dans le cas de l’algorithme de Le Verrier, l’anneau A est
supposé tel que la division par un entier, quand elle est possible, est
unique (autrement dit, les entiers ne sont pas des diviseurs de zéro) et
réalisable par un algorithme. Cette méthode permet de calculer :
– le polynôme caractéristique PA d’une matrice carrée A ∈ An×n ;
– l’adjointe de la matrice A et son inverse (s’il existe) ;
– un vecteur propre non nul relatif à une valeur propre donnée de
A si l’on suppose de plus que l’anneau A est intègre.
Posant P (X) = (−1)n PA (X) = X n − c1 X n−1 + · · · + cn−1 X + cn , la
méthode consiste à calculer les coefficients ck pour en déduire le polynôme caractéristique de A . On utilise pour cela le calcul de la matrice
caractéristique adjointe de A tel que développé dans 1.2.1. Rappelons la
définition de la matrice caractéristique
adjointe de A : c’est la matrice
P
n−1−k (formule 1.11 page 11) dans
B
X
Q(X) = Adj(XIn − A) = n−1
k=0 k
laquelle les matrices Bk sont données par les relations 1.12 (page 12) :
Bk = ABk−1 − ck In (1 ≤ k ≤ n)
avec
B0 = In .
On démontre, en utilisant les relations de Newton (1.23 page 29), que :
ck =
1
Tr(A Bk−1 )
k
pour 1 ≤ k ≤ n .
En effet, partant des équations suivantes (voir 1.13 page 12) qui découlent
des relations 1.12 rappelées ci-dessus :
Bk = Ak − c1 Ak−1 − . . . − ck−1 A − ck In pour tout entier k ∈ {1, . . . , n} ,
on considère les traces des deux membres dans chacune de ces n égalités
matricielles pour obtenir :
Tr(Bk ) = sk − c1 sk−1 − . . . − ck−1 s1 − nck
(1 ≤ k ≤ n) .
Mais sk = c1 sk−1 + . . . + ck−1 s1 + kck (ce sont les relations de Newton
pour le polynôme P (X)). Comme Tr(Bk ) = Tr(ABk−1 ) − nck (à cause
de l’égalité Bk = ABk−1 − ck In ), on obtient Tr(ABk−1 ) = kck .
t
u
Notons par ailleurs (comme nous l’avons fait au § 1.2.1 page 11) que
Bn = A Bn−1 −cn In = 0, c’est-à-dire A Bn−1 = cn In = (−1)n−1 det(A) In .
2.5. Méthode de Le Verrier et variantes
85
Ce qui montre que si det(A) est inversible dans A, alors A possède un
inverse donné par A−1 = (cn )−1 Bn−1 .
Rappelons également que Bn−1 = (−1)n−1 Adj(A) ce qui donne,
sans autre calcul, l’adjointe de la matrice A .
Nous traduisons la méthode de Souriau-Faddeev-Frame qui vient
d’être développée par l’algorithme 2.8, dans lequel B désigne successivement les matrices B0 = In , B1 ,. . ., Bn , et C les matrices A, A B1 ,. . .,
A Bn .
Algorithme 2.8 Algorithme de Souriau-Faddeev-Frame
Entrée : Un entier n et une matrice A ∈ An×n . L’anneau A est supposé
avoir un algorithme de division exacte par les entiers ≤ n.
Sortie : Le polynôme caractéristique PA de A.
Début
Variables locales : k ∈ N ; c ∈ A ; C, B ∈ An×n ; P ∈ A[X].
Id := In ; B := Id ; P := X n ;
pour k de 1 à n − 1 faire
C := B · A ; c := Tr(C)/k ;
P := P − c · X n−k ; B := C − c · Id
fin pour ;
c := Tr(B · A)/n ; P := P − c ;
PA := (−1)n P
Fin.
Calcul de vecteurs propres
Dans le cas où A est intègre, si λ est une valeur propre simple de
A (c’est-à-dire une racine simple de P (X)), le même calcul (donnant
entre autres la matrice caractéristique adjointe Q(X) = Adj(XIn − A))
nous permet d’obtenir un vecteur propre non nul associé à λ.
En effet Q(X) =
n−1
X
Bi X n−1−i donc Tr(Q(X)) =
i=0
n−1
X
Tr(Bi ) X n−1−i .
i=0
Mais Tr(Bi ) = (i − n) ci , donc
Tr(Q(X)) = −
n−1
X
i=0
(n − i) ci X n−i−1 = P 0 (X)
86
2. Algorithmes de base en algèbre linéaire
où P 0 (X) désigne le polynôme dérivé de P (X). Ainsi Tr(Q(λ)) =
P 0 (λ) 6= 0 puisque λ est une racine simple de P . Par conséquent la
matrice Q(λ) n’est pas nulle.
Mais l’égalité (XIn − A) Q(X) = P (X) In donne (A − λIn ) Q(λ) =
−P (λ) In = 0. Ce qui prouve que n’importe quelle colonne non nulle v
de Q(λ) vérifie Av = λv et c’est donc un vecteur propre non nul de A
relatif à la valeur propre λ.
Si l’on désigne par ` le numéro de la colonne présumée non nulle de
la matrice Q(λ) = B0 λn−1 + B1 λn−2 + · · · + Bn−2 λ + Bn−1 et par bk`
la colonne no ` de Bk , le calcul de ce vecteur propre peut se faire de la
manière suivante :
• Poser v0 = e` (colonne numéro ` de la matrice In ) ;
• Faire vk = λ vk−1 + bk` (pour k allant de 1 à n − 1).
Le vecteur propre recherché n’est autre que v = vn−1 .
Plus généralement, si la multiplicité géométrique 6 de λ est égale à 1,
la matrice Q(λ) n’est pas nulle (elle est de rang 1) et n’importe quelle
colonne non nulle de Q(λ) représente un vecteur propre non nul de A
pour la valeur propre λ.
Si par contre la multiplicité géométrique (et par conséquent la multiplicité algébrique 7 ) de la valeur propre λ est supérieure ou égale à 2,
non seulement la trace, mais la matrice Q(λ) elle-même est nulle d’après
la propriété 1.1.2 page 6, puisque le rang de la matrice singulière A−λIn
est, dans ce cas, au plus égal à n − 2. La matrice Q(λ), dans ce cas, ne
donne donc aucun vecteur propre non nul de A.
On montre alors que ce sont les matrices dérivées successives (par
rapport à X) de la matrice Q(X) qui permettent de calculer des vecteurs propres non nuls relatifs à λ.
Considérons en effet pour k ∈ N l’opérateur ∆[k] : A[X] −→ A[X],
P 7→ ∆[k] P , où ∆[k] P (X) = P [k] (X) est défini par l’identité
X
P (X + Y ) =
P [k] (X) Y k .
k≥0
Remarquons que P [0] (X) = P (X), P [1] (X) = P 0 (X) et qu’en ca1
ractéristique nulle ∆[k] = k!
Dk où Dk est l’opérateur de dérivation
1 (k)
d’ordre k dans A [X] (P [k] (X) = k!
P (X)). Il est facile de voir qu’en
[k]
caractéristique quelconque, ∆ est un A - endomorphisme de l’algèbre
6. La multiplicité géométrique d’une valeur propre est par définition la dimension
du sous-espace propre correspondant.
7. C’est la multiplicité de λ en tant que zéro du polynôme caractéristique.
2.5. Méthode de Le Verrier et variantes
87
A [X] qui vérifie une formule analogue à la formule de Leibnitz, mais
plus simple :
k
X
∆[k] (P1 P2 ) =
∆[i] (P1 ) ∆[k−i] (P2 )
i=0
En outre un polynôme P admet λ comme racine d’ordre k ≥ 1 si
et seulement si P [0] (λ) = P [1] (λ) = . . . = P [k−1] (λ) = 0 et P [k] (λ) 6= 0.
Appliquant successivement les opérateurs ∆[k] pour k allant de 1 à m
(où m est la multiplicité algébrique de la valeur propre λ) à l’égalité
matricielle :
(XIn − A) Q(X) = P (X) In ,
on obtient la suite d’égalités :
Q[k−1] (X) + (XIn − A) Q[k] (X) = P [k] (X) In
(1 ≤ k ≤ m).
Remplaçant dans ces égalités X par la valeur propre λ, et tenant
compte du fait que Q(λ) = P (λ) = P [1] (λ) = · · · = P [m−1] (λ) = 0, on
obtient le système :
Q(λ) = 0
(λIn − A) Q[1] (λ) = 0
Q[1] (λ) + (λIn − A) Q[2] (λ) = 0
.. .. .. .. ..
. . . . .
Q[m−2] (λ) + (λIn − A) Q[m−1] (λ) = 0
[m−1]
Q
(λ) + (λIn − A) Q[m] (λ) = P [m] (λ) In .
Soit r ∈ N le plus petit entier tel que Q[r] (λ) 6= 0.
Alors r < m car sinon, on aurait Q(λ) = Q[1] (λ) = · · · = Q[m−1] (λ) = 0
et (λIn − A) Q[m] (λ) = P [m] (λ) In avec P [m] (λ) 6= 0, ce qui contredit
(nous sommes dans un anneau intègre) le fait que la matrice λIn − A
est singulière.
Donc (λIn − A) Q[r] (λ) = 0 et toute colonne non nulle de Q[r] (λ)
est un vecteur propre non nul de A pour la valeur propre multiple λ.
Nombre d’opérations arithmétiques
L’algorithme de Souriau-Faddeev-Frame consiste à calculer, pour k
allant de 1 à n, le produit matriciel Ak = ABk−1 , le coefficient ck =
1
k Tr(Ak ) et enfin la matrice Bk = Ak − ck In .
Rappelons qu’on désigne par µM (n) le nombre d’opérations arithmétiques dans l’anneau de base pour la multiplication de deux matrices
88
2. Algorithmes de base en algèbre linéaire
carrées d’ordre n. Le coût de l’algorithme de Souriau-Faddeev-Frame
s’élève à
(n − 2) µM (n) + n (2n − 1) + 2 n (n − 1) .
C’est à très peu près le même coût que pour l’algo de Le Verrier (on
gagne n opérations arithmétiques).
Outre la plus grande simplicité, l’avantage est que l’on a aussi calculé
la matrice adjointe. En particulier le calcul de la matrice inverse, si elle
existe ne coûte que n2 divisions supplémentaires dans A. Enfin le calcul
d’un vecteur propre non nul relatif à une valeur propre donnée de multiplicité géométrique égale à 1 se fait moyennant 2 n (n − 1) opérations
arithmétiques supplémentaires.
Proposition 2.5.2 Avec la méthode de Faddeev-Souriau-Frame, le calcul du polynôme caractéristique de la matrice A, de son déterminant,
de sa matrice adjointe, de son inverse quand elle existe, ainsi que des
sous-espaces propres de dimension 1 (quand on connaı̂t la valeur propre
correspondante) se fait en 2 n4 + O(n3 ) = O(n4 ) opérations arithmétiques. Pour le calcul du seul polynôme caractéristique on en effectue
précisément 2 n (n − 1) (n2 − 3n/2 + 1/2) opérations.
2.5.3
Méthode de Preparata & Sarwate
La méthode de Preparata & Sarwate est une accélération astucieuse
de la méthode de Le Verrier, basée sur la remarque simple suivante.
Pour calculer la trace d’un produit AB de deux matrices carrées d’ordre
2
n,
P il suffit d’éxécuter 2n opérations arithmétiques puisque Tr AB =
k,` ak,` b`,k . Or le calcul le plus coûteux dans la méthode de Le Verrier
est celui des traces des puissances successives de la matrice A dont on
veut calculer le polynôme caractéristique.
√
Posons donc r = d n e, B0 = C0 = In , B1 = A, et calculons les
Bi = Ai pour i = 2, . . . , r, puis les Cj = Br j pour j = 1, . . . , r − 1. Ce
calcul consomme (2r − 3)(n3 − n2 ) ' 2 n3,5 opérations arithmétiques
dans A.
P
On a alors Tr Arj+i = Tr Bi Cj = k,` bi,k,` cj,`,k , et les valeurs rj + i
pour 0 ≤ i, j ≤ r −1 parcourent l’intervalle [0, r2 −1]. Si r2 = n on doit
calculer en outre Sn = Tr C1 Cr−1 . On obtient donc toutes les sommes
de Newton Sm = Tr Am , 1 ≤ m ≤ n pour un peu moins que 2n3
opérations arithmétiques supplémentaires dans A.
2.5. Méthode de Le Verrier et variantes
89
Ceci donne l’algorithme 2.9. Comme la récupération des coefficients
du polynôme caractéristique à partir des sommes de Newton réclame
O(n2 ) opérations arithmétiques on obtient la proposition suivante.
Algorithme 2.9 Algorithme de Preparata & Sarwate, version
séquentielle simple.
Entrée : Une matrice carrée A ∈ An×n où A est un anneau vérifiant les
hypothèses de l’algorithme de Le Verrier.
P
Sortie : Le polynôme caractéristique PA = (−1)n (X n + nk=1 pk X n−k ).
Début
Variables locales : i, j, r ∈ N ; Bi , Cj ∈ An×n (i, j = 1..r − 1), Si ∈ A
(i = 1..n) ;
√
Étape 1 : Calcul des puissances Ai pour i < r = d n e.
√
r := d n e ; B1 := A ; S0 := n ; S1 := Tr A ;
pour i de 1 à r − 2 faire
Bi+1 := A Bi ; Si+1 := Tr Bi+1
fin pour ;
Étape 2 : Calcul des puissances Arj pour j < r.
C1 := A Br−1 ; Sr := Tr C1 ;
pour j de 1 à r − 2 faire
Cj+1 := C1 Cj ; S(j+1)r := Tr Cj+1
fin pour ;
Étape 3 : Calcul des sommes de Newton.
pour i de 1 à r − 1 faire
pour j de 1 à r − 1 faire
Sjr+i := Tr Bi Cj
fin pour
fin pour
si n = r2 alors Sn := Tr C1 Cr−1 fin si ;
Étape 4 : Calcul des coefficients de PA .
Calculer les coefficients pk (1 ≤ k ≤ n) en utilisant les équations (1.23).
Fin.
Proposition 2.5.3 Supposons que l’anneau commutatif A satisfasse
les hypothèses de l’algorithme de Le Verrier : la division par un entier,
quand elle est possible, est unique, et réalisable par un algorithme. Le
nombre total d’opérations arithmétiques lors de l’exécution de l’algorith-
90
2. Algorithmes de base en algèbre linéaire
me de Preparata & Sarwate, si on utilise la multiplication usuelle des
matrices, est égal à 2 n3,5 + O(n3 ) = O(n3,5 ).
2.6
Méthode de Samuelson-Berkowitz
Elle est basée sur la méthode de partitionnement ([Gas] pp. 291–298,
[FF]), attribuée à Samuelson [79], et elle a l’avantage de s’appliquer à
un anneau commutatif arbitraire.
Berkowitz [6] en donne une version parallèle de laquelle nous extrayons une méthode séquentielle particulièrement simple et efficace. Elle
montre l’intérêt pratique de cet algorithme pour les machines séquentielles, en le plaçant parmi les algorithmes les plus performants actuellement pour le calcul sans division du polynôme caractéristique (cf. les
test expérimentaux présentés au chapitre 11).
2.6.1
Principe général de l’algorithme
Soit A = (aij ) ∈ An×n une matrice carrée d’ordre n ≥ 2 sur un anneau commutatif arbitraire A. Conformément aux notations introduites
dans la section 1.1, pour tout entier r (1 ≤ r ≤ n − 1), on désigne par
Ar la sous-matrice principale dominante d’ordre r de A. On partitionne
comme suit la matrice Ar+1 :
Ar
A1..r,r+1
Ar
Sr
Ar+1 =
=
.
Ar+1,1..r ar+1,r+1
Rr ar+1,r+1
Le polynôme caractéristiquePPr+1 (X) de Ar+1 est relié au polynôme caractéristique Pr (X) = ri=0 pr−i X i de Ar par la formule de
Samuelson (1.15) (proposition 1.2.1 page 13) que l’on peut réécrire sous
la forme suivante :
(
(ar+1,r+1 − X) Pr (X) +
r+1−k (2.14)
Pr+1 =
Pr+1
k−2 S )p + · · · + (R S )p
r 0
r r k−2 X
k=2 (Rr Ar
Notons Qr+1 le polynôme :
−X r+1 + ar+1,r+1 X r + Rr Sr X r−1 + Rr Ar Sr X r−2 + · · · + Rr Ar−1
r Sr .
Pour calculer Pr+1 (X) selon la formule de Samuelson on peut :
– effectuer le produit Pr Qr+1 ;
– supprimer les termes de degré < r ;
2.6. Méthode de Samuelson-Berkowitz
91
– et enfin diviser par X r .
On peut aussi décrire ce calcul sous la forme :
−−→
−
→
Pr+1 = Toep(Qr+1 ) × Pr
(2.15)
−
→
où Pr est le vecteur colonne t (p0 , p1 , . . . , pr ) des coefficients du polynôme Pr et Toep(Qr+1 ) ∈ A(r+2)×(r+1) est la matrice de Toeplitz
suivante définie à partir du polynôme Qr+1 :
−1
0
··· ···
0
..
..
a
.
−1
.
r+1,r+1
..
..
..
..
.
.
.
Rr Sr
.
Toep(Qr+1 ) =
..
..
..
..
.
.
.
0
.
..
..
R Ar−2 S
.
.
−1
r
r
r
r−2
Rr Ar−1
r Sr Rr Ar Sr . . . Rr Sr ar+1,r+1
Algorithme 2.10 Algorithme de Berkowitz, principe général.
Entrée : Une matrice A ∈ An×n .
Sortie : Le polynôme caractéristique PA (X) de A.
Début
Étape 1 :
Pour k < r dans {1, . . . , n} calculer les produits Rr (Ar )k Sr ,
ce qui donne les polynômes Qr+1 et les matrices Toep(Qr+1 ),
Étape 2 :
−
→
Calculer le produit Toep(Qn ) Toep(Qn−1 ) · · · Toep(Q2 ) P1 :
−
→
on obtient PA .
Fin.
−
→
Avec P1 =
2.6.2
−1
a1,1
, on obtient l’algorithme de Berkowitz informel 2.10.
Version séquentielle
Dans la version séquentielle la plus simple de l’algorithme de Berkowitz, le calcul des coefficients de la matrice Toep(Qr+1 ) se fait naturellement par l’utilisation exclusive de produits scalaires ou de produits de
92
2. Algorithmes de base en algèbre linéaire
matrices par des vecteurs. De même dans l’étape 2, le produit s’effectue
de droite à gauche, donc n’utilise que des produits matrice par vecteur.
Cela donne l’algorithme 2.11.
Algorithme 2.11 Algorithme de Berkowitz, version séquentielle simple.
Entrée : Un entier n ≥ 2 et une matrice A = (aij ) ∈ An×n .
Sortie : Le polynôme caractéristique de A : PA (X).
Début
Variables locales : i, j, k, r ∈ N ; v = (vi ), c = (ci ), s = (si ), P = (pi ) :
listes de longueur variable r (1 ≤ r ≤ n + 1) dans A.
# initialisation de c et de v
c1 := −1 ; v := (−1, a11 ) ;
# Calcul des polynômes caractéristiques des matrices principales
# dominantes d’ordre ≥ 2 (les listes successives dans P = (pi ))
pour r de 2 à n faire
(si )i=1..r−1 := (air )i=1..r−1 ;
c2 := arr ;
pour i de
1 à r − 2 faire
Pr−1
ci+2 := j=1 arj sj ;
P
pour j de 1 à r − 1 faire pj := r−1
k=1 ajk sk fin pour ;
(sj )j=1..r−1 := (pj )j=1..r−1
fin pourP
;
cr+1 := r−1
j=1 arj sj ;
pour i de 1 à r + 1 faire
Pmin (r,i)
pi := j=1
ci+1−j vj
fin pour ;
(vi )i=1..r+1 := (pi )i=1..r+1
fin pour ; P
PA (X) := ni=0 vi+1 X n−i
Fin.
Ainsi, sans calculer des puissances de matrices Ak−1
(3 ≤ k ≤ r) on
r
commence par calculer Rr Sr puis successivement, pour k allant de 2
à r, le produit (matrice par vecteur) Ak−1
Sr suivi du produit scalaire
r
3 + r 2 − 3r + 1 opérations arithRr Ak−1
S
,
ce
qui
se
traduit
par
2r
r
r
métiques pour chaque r (1 ≤ r ≤ n − 1). On en déduit que le nombre
d’opérations arithmétiques (dans l’anneau de base A) intervenant dans
2.7. Méthode de Chistov
93
ce calcul est égal à :
n−1
X
r=1
1
2
3
8
(2r3 + r2 − 3r + 1) = n4 − n3 − n2 + n − 1 .
2
3
2
3
Il en est de même pour la multiplication des matrices de Toeplitz
Toep(Qr ). On commence par multiplier la première matrice Toep(Q1 ) =
−
→
−
→
P1 à gauche par la matrice Toep(Q2 ) pour obtenir le vecteur P2 qui
−
→
−−−→
est un vecteur 3 × 1 et ainsi de suite jusqu’à Pn = Toep(Qn ) × Pn−1 .
−
→
−−→
Comme chaque multiplication Pr = Toep(Qr ) × Pr−1 (d’une matrice
sous-triangulaire (r + 1) × r avec des −1 sur la diagonale par un vecteur
−
→
r × 1) coûte r (r − 1) opérations arithmétiques dans A, le calcul de Pn
n
X
1
se fait en
(r2 − r) = (n3 − n) opérations arithmétiques de base 8 .
3
r=2
Proposition 2.6.1 Le coût total de l’algorithme séquentiel simple de
Berkowitz s’élève à
1 4 1 3 3 2 7
1 4 1 3
n − n − n + n−1 ≤
n − n
2
3
2
3
2
3
opérations arithmétiques dans l’anneau de base.
2.7
2.7.1
Méthode de Chistov
Le principe général
La méthode de Chistov [17] consiste à calculer le polynôme caractéristique PA (X) d’une matrice carrée A ∈ An×n (n ≥ 2) en le ramenant
à l’inversion du polynôme formel Q(X) = det(In − XA) dans l’anneau
des séries formelles A[[X]].
Ce polynôme est, à un signe près, le polynôme réciproque du polynôme caractéristique puisque
1
1
n n
n
(−1) X Q
= (−X) det In − A = det(A − XIn ) = PA (X).
X
X
Comme les polynômes PA et Q, résultat final du calcul, sont de degré
n, tous les calculs peuvent se faire modulo X n+1 dans l’anneau des séries
8. Ce calcul peut être accéléré en utilisant une multiplication rapide des polynômes (cf. chapitre 6), mais cela ne change pas substantiellement le résultat global
qui reste de O(n4 ) opérations arithmétiques avec la même constante asymptotique.
Nous n’avons pas implémenté cette amélioration lors de nos tests expérimentaux.
94
2. Algorithmes de base en algèbre linéaire
formelles A[[X]], c’est-à-dire encore peuvent se faire
l’anneau des
dans
n+1
développements limités à l’ordre n sur A : A[X] X
.
Dans la suite, nous noterons souvent cet anneau An .
L’algorithme de Chistov utilise le fait que, pour toute matrice carrée
B d’ordre n régulière, le n - ème élément de la diagonale de la matrice
inverse B −1 , noté (B −1 )n,n , est égal à :
(B −1 )n,n =
det Bn−1
det B
(Bn−1 )n,n =
ou encore
det Bn−1
,
det Bn
où Br désigne la sous-matrice principale dominante d’ordre r de B
(avec la convention det B0 = 1). Ceci permet d’écrire lorsque B est
fortement régulière :
Bn−1
−1
× Bn−1
n,n
n−1,n−1
× · · · × B1−1
1,1
=
1
.
detB
Appliquant ce fait à la matrice B = In − XAn ∈ A[X]n×n qui est
fortement régulière puisque tous ses mineurs principaux dominants sont
des polynômes de terme constant égal à 1 et sont donc inversibles dans
l’anneau A[[X]], on obtient :
−1
Q(X)
−1
= [ det(In − XAn ) ]
=
n
Y
B −1
r
r,r
.
(2.16)
r=1
Mais on a un isomorphisme canonique A[[X]]r×r ' Ar×r [[X]] et la
matrice Br = Ir − XAr est aussi inversible dans l’algèbre des séries
formelles sur l’anneau de matrices Ar×r , et son inverse est la matrice :
Br−1 = Ir +
∞
X
(Ar )k X k ∈ Ar×r [[X]] .
(2.17)
k=1
Donc en notant Er la r - ème colonne de Ir :
Br−1
r,r
mod X n+1 = 1 +
n
X
t
Er (Ar )k Er X k .
(2.18)
k=1
e
Par conséquent en notant Q(X)
= Q(X)−1 mod X n+1 on obtient :
"
#
n
n
Y
X
t
e
Q(X)
=
1+
Er (Ar )k Er X k mod X n+1
r=1
k=1
2.7. Méthode de Chistov
95
et donc,
(
Q(X) =
n
Y
r=1
"
1+
n
X
t
#)−1
Er (Ar )k Er X k
dans A[X]
X n+1
k=1
Ainsi Q(X) est l’inverse modulo X n+1 du produit modulo X n+1
de n polynômes de terme constant égal à 1 et de degré inférieur ou égal
à n.
Rappelons que le polynôme caractéristique à calculer PA (X) est le
produit par (−1)n du polynôme réciproque à l’ordre n de Q(X). On
obtient alors l’algorithme de Chistov 2.12.
Algorithme 2.12 Algorithme de Chistov, principe général.
Entrée : la matrice A ∈ An×n .
Sortie : le polynôme caractéristique PA (X) de A.
Début
Étape 1 :
Calculer pour r, k ∈ {1, . . . , n} les produits t Er (Ar )k Er ,
ce qui donne les polynômes B −1
r r,r (formule (2.18)).
Étape 2 :
Calculer le produit des n polynômes précédents modulo X n+1 ,
ce qui donne Q(X)−1 mod X n+1 (formule (2.16)).
Étape 3 :
Inverser modulo X n+1 le polynôme précédent : on obtient Q(X).
Étape 4 :
Prendre le polynôme réciproque à l’ordre n du polynôme Q(X).
On obtient PA (X) en multipliant par (−1)n .
Fin.
Nous détaillons maintenant la version séquentielle élémentaire de cet
algorithme.
2.7.2
La version séquentielle
Dans la version séquentielle la plus simple on obtient l’algorithme 2.13 page suivante.
On démontre maintenant que le coût en nombre d’opérations arithmétiques dans l’anneau de base pour cette version élémentaire est asymptotiquement de l’ordre de (2/3) n4 .
96
2. Algorithmes de base en algèbre linéaire
Algorithme 2.13 Algorithme de Chistov, version séquentielle
simple.
Entrée : Un entier n ≥ 2 et une matrice A = (ai,j ) ∈ An×n .
Sortie : Le polynôme caractéristique de A, PA ∈ A[X].
Début
Variables locales : i, j, k, r ∈ N ; q = (qi ), b = (bi ), c = (ci ) ∈ An+1 où
0 ≤ i ≤ n ; v = (vi ) , w = (wi ) ∈ Ar où 1 ≤ i ≤ r, Q ∈ A[X].
q := (1)i=0..n ; c := q ; (initialisation)
pour i de 1 à n faire qi := qi−1 a1,1 fin pour ;
pour r de 2 à n faire
v := (ai,r )i=1..r ; c1 := vr ;
pour i de 2 à n − 1 faire
P
pour j de 1 à r faire wj := rk=1 aj,k vk fin pour ;
v := w ; ci := vr
fin pour
P;
cn := rk=1 ar,k vk ;
P
pour j de 0 à n faire bj := jk=0 cj−k qk fin pour ;
q := b
fin pour P
n
k mod X n+1 ;
Q := 1/
k=0 qk X
PA := (−1)n X n Q(1/X)
Fin.
Reprenons en effet les quatre étapes dans l’algorithme de Chistov
général 2.12 page précédente.
• L’étape 1, la plus coûteuse, se ramène en fait à calculer successivement les produits (Ar )k Er pour 1 ≤ r ≤ n et pour 1 ≤ k ≤ n,
puisque t Er (Ar )k Er n’est autre que la r - ème composante du vecteur
(Ar )k Er .
Pour chaque valeur de r (1 ≤ r ≤ n), on commence par calculer
Ar Er puis, pour k allant de 2 à n, le produit de la matrice Ar par le
vecteur (Ar )k−1 Er , ce qui se traduit par n (2r2 −r) opérations arithmétiques (additions/soustractions et multiplications) pour chaque r compris
entre 1 et n. On en déduit que le nombre d’opérations dans ce calcul est
égal à :
n
n
X
r=1
(2r2 − r) =
1 2
n (n + 1)(4n − 1) = (2/3) n4 + O(n3 ) .
6
2.8. Méthodes reliées aux suites récurrentes linéaires
97
• L’étape 2 revient à calculer le produit tronqué de n polynômes
de degré au plus égal à n, ce qui se fait en O(n3 ) opérations de base.
• L’étape 3 consiste à calculer le produit tronqué à l’ordre n des
dlog ne
dlog ne polynômes 1 + R, 1 + R2 , . . . , 1 + R2
eux-mêmes obtenus
à l’issue de dlog ne élévations successives au carré, tronquées à l’ordre
n, de polynômes de degré n. Cela fait un nombre d’opérations arithmétiques de l’ordre de n2 dlog ne.
• L’étape 4 a un coût négligeable.
Proposition 2.7.1 Le coût total de l’algorithme séquentiel élémentaire de Chistov s’élève à (2/3) n4 + O(n3 ) opérations arithmétiques dans
l’anneau de base.
Nous présentons ci-dessous un résumé de la discussion sur le nombre
d’opérations arithmétiques :
Etape
Etape
Etape
Etape
Etape
1
2
3
4
Coût
(2/3) n4 + O(n3 )
O(n3 )
O(n2 log n)
négligeable
Tableau 2.7.2
Complexité de la version séquentielle de l’algorithme de Chistov
2.8
Méthodes reliées aux suites récurrentes linéaires
Dans la section 2.8.1 nous donnons un algorithme de calcul du polynôme caractéristique d’une matrice A basé sur la considération des
transformés successifs de vecteurs de la base canonique par A.
Dans la section 2.8.2 nous présentons un algorithme dû à Berlekamp
qui permet de calculer le polynôme générateur minimal d’une suite récurrente linéaire dans un corps lorsqu’on sait qu’elle vérifie une relation
de récurrence linéaire d’ordre n et qu’on connaı̂t les 2n premiers termes
de la suite.
Dans la section 2.8.3 on décrit l’algorithme de Wiedemann qui utilise
celui de Berlekamp pour trouver avec une bonne probabilité le polynôme
caractéristique d’une matrice sur un corps fini.
98
2. Algorithmes de base en algèbre linéaire
2.8.1
L’algorithme de Frobenius
Nous donnons ici un algorithme qui est basé sur une description de
nature géométrique pour un endomorphisme d’un K – espace vectoriel.
Comme conséquence, on calcule le polynôme caractéristique de l’endomorphisme avec essentiellement le même nombre d’opérations arithmétiques que dans la méthode du pivot de Gauss (qui ne calcule que le
déterminant), sans les inconvénients que présentait la méthode de Hessenberg (hormis le cas des corps finis) concernant la taille des coefficients
intermédiaires.
Le cas usuel
Nous aurons besoin de la procédure 2.14 page ci-contre (dérivée de
l’algorithme de Jordan-Bareiss) à laquelle nous donnons le nom de JorBarSol. Elle calcule, à la Jordan-Bareiss, la relation de dépendance linéaire exprimant la dernière colonne en fonction des premières dans une
matrice fortement régulière ayant une colonne de plus que de lignes. La
fin du calcul, après la triangulation, reste dans l’anneau A si la relation
de dépendance linéaire est à coefficients dans A.
Considérons une matrice carrée A ∈ Z5×5 d’ordre 5 prise au hasard,
donnée par exemple par Maple. Elle définit un endomorphisme hA de
Q5 . On note (fi )1≤i≤6 le premier vecteur de la base canonique de Q5 et
ses 5 transformés successifs par A. Ceci fournit une matrice B ∈ Z5×6 .
Voici un exemple typique
A =
57
−7
−35
−73
43
−82
58
−14
−91
−4
−48
−94
−9
1
−50
−11
−68
−51
5
50
38
14
−73
−86
67
,
1
57
7940
55624
−46831857
−22451480858
0
B =
0
0
−7
8051
1071926
199923276
−35
−998
−245490
54032957
−73
−7622
−1648929
−128141849
14745797441
9123769947
.
−10372211183
0
43
3460
209836
−58008810
−15808793525
En général les vecteurs (fi )1≤i≤5 sont indépendants, et même, la
matrice B est fortement régulière. C’est le cas ici.
Concernant la taille des coefficients, ceux de la matrice initiale sont
majorés par 100 en valeur absolue, et ceux de la matrice B dans la
2.8. Méthodes reliées aux suites récurrentes linéaires
99
k -ème colonne, sont majorés par M k−1 si M est une des normes de
A décrite en section 1.6, par exemple M < 270 pour la norme de
Frobenius.
Algorithme 2.14 Algorithme JorBarSol
Entrée : Une matrice A = (aij ) ∈ An×(n+1) fortement régulière. L’anneau
A est supposé intègre avec un algorithme de division exacte.
Sortie : P
L = (`j ) ∈ An : on a, en notant Cj la j-ème colonne de A,
Cn+1 = nj=1 `j Cj . La fin du calcul reste dans A si les `i sont dans A.
Début
Variables locales : i, j, p, q ∈ N ; piv, den, coe ∈ A ;
# LU-décomposition à la Jordan-Bareiss
den := 1 ; m := n + 1 ;
pour p de 1 à n − 1 faire
piv := app ; fortement régulière
pour i de p + 1 à n faire
coe := aip ;
pour j de p + 1 à m faire
aij := (piv ∗ aij − coe ∗ apj ) / den
fin pour
fin pour
den := piv
fin pour
# calcul des coefficients `i
pour q de 1 à n − 1 faire
p := n − q ; `p := apm /app ;
pour i de 1 à p − 1 faire
aim := aim − `p aip
fin pour
fin pour
Fin.
P
On peut calculer la relation
de dépendance linéaire f6 = 5i=1 αi fi ,
P
qui se relit hA 5 (f1 ) = 4i=0 αi hA i (f1 ). La matrice de hA sur la base
(f1 , . . . , f5 ) estPalors clairement la matrice compagnon du polynôme
P (X) = X 5 − 4i=0 αi X i (cf. page 15) et on obtient le polynôme caractéristique de A par la formule PA (X) = (−1)5 P (X).
Cet algorithme qui calcule le polynôme caractéristique de A fonc-
100
2. Algorithmes de base en algèbre linéaire
tionne lorsque le polynôme générateur minimal P de la suite récurrente
linéaire (An f1 )n∈N dans Q5 est de degré ≥ 5. Dans ce cas le polynôme
P est en effet égal au polynôme minimal et au polynôme caractéristique
de A (au signe près).
P
Pour calculer la relation de dépendance linéaire f6 = 5i=1 αi fi on
applique la procédure JorBarSol. Elle commence par la triangulation
à la Jordan-Bareiss de la matrice B. Cette triangulation (en fait, une
LU -décomposition) ne change pas les deux premières colonnes et donne
les 4 dernières suivantes.
7940
55624
−46831857
−22451480858
8051
1071926
199923276
14745797441
288771
39235840
6619083961
452236520806
641077
−110921281313
−32874613863452
−5984786805270056
−370413
−114147742050
−8244227015780803785
−1467472408808983073730
Si on traite une matrice carrée A d’ordre n dont une norme est majorée par M , le coefficient en position (i, j) dans la matrice ainsi obtenue
(k−1)
est égal au mineur bi,j
de la matrice B = [f1 |Af1 | · · · |An f1 ] avec k =
min(i, j). A priori (comme dans l’exemple ci-dessus, d’ailleurs) le plus
grand coefficient serait en position (n, n+1), majoré par M 1+···+(n−2)+n ,
2
c’est-à-dire M (n −n+2)/2 , ce qui reste raisonnable, en tout cas bien meilleur que dans l’algorithme de Hessenberg.
L’algorithme termine en donnant la combinaison linéaire recherchée
par le calcul successif des coefficients α5 , α4 , . . . , α1 .
Une matrice de Frobenius d’ordre n est un autre nom donné à
une matrice compagnon d’un polynôme P (X) de degré n. On peut
l’interprèter comme la matrice de l’application linéaire (( multiplication
par x )) (la classe de X) dans l’algèbre quotient K[X] /hP (X)i sur la
base canonique 1, x, . . . , xn−1 .
L’algorithme que nous venons de décrire n’ayant pas de nom officiel,
nous l’appellerons algorithme de Frobenius, c’est l’algorithme 2.15 page
suivante.
Nombre d’opérations arithmétiques
Avec une matrice carrée d’ordre n l’algorithme de Frobenius donne
un calcul en O(n3 ) opérations arithmétiques, ce qui est du même ordre
de grandeur que pour la méthode du pivot de Gauss.
Plus précisément l’algorithme 2.15 se décompose en deux grandes
étapes. La première étape crée la matrice B et exécute (n − 1) n2 mul-
2.8. Méthodes reliées aux suites récurrentes linéaires
101
Algorithme 2.15 Algorithme de Frobenius (le cas simple)
Entrée : Une matrice A = (aij ) ∈ An×n . L’anneau A est supposé intègre
avec un algorithme de division exacte.
Sortie : Le polynôme caractéristique PA (X). L’algorithme ne fonctionne
que si le premier vecteur de base et ses n − 1 transformés successifs par A
sont linéairement indépendants.
Début
Variables locales : i, k, m ∈ N ; B = (bi,j ) ∈ An×(n+1) ;
V = (vi ) ∈ An×1 , L = (`j ) ∈ An ;
m := n + 1 ; V := première colonne de A ;
B := 0 dans An×(n+1) ; b1,1 := 1 ;
2-ème colonne de B := V ;
pour k de 3 à m faire
V := A V ;
k-ème colonne de B := V
fin pour ;
L := JorBarSol(B) ; P
PA := (−1)n (X n − nk=1 `k X k−1 )
Fin.
tiplications et (n − 1)3 additions. La deuxième étape applique l’algorithme JorBarSol à la matrice B. Vue la première colonne de celle-ci,
cet algorithme utilise
n−1
X
(n − p) (n − p + 1) +
p=2
n−1
X
(p − 1) =
p=1
1 3 1 2 5
n − n − n+1
3
2
6
additions/soustractions et
3
n−1
X
n−1
X
p=2
p=1
(n − p) (n − p + 1) +
(p − 1) + n − 2 = n3 −
5 2 3
n + n−1
2
2
multiplications/divisions (les divisions sont toutes exactes). Ceci donne
le résultat suivant.
Proposition 2.8.1 L’algorithme de Frobenius (dans le cas usuel simple)
appliqué à une matrice carrée d’ordre n sur un anneau intègre dans lequel les divisions exactes sont explicites demande en tout
10 3
11
n − 7 n2 +
n−1
3
3
102
2. Algorithmes de base en algèbre linéaire
opérations arithmétiques dans l’anneau. Plus précisément cet algorithme
4
7
13
7
3
exécute n3 − n2 + n additions/soustractions et 2 n3 − n2 + n−1
3
2
6
2
2
multiplications/divisions.
En pratique, sur un corps fini, les algorithmes de Hessenberg et de
Frobenius s’avèrent meilleurs que tous les autres, ce qui correspond
au fait qu’ils fonctionnent en exécutant seulement O(n3 ) opérations
arithmétiques. Mais dès qu’on passe à des matrices à coefficients dans
Z, l’algorithme de Berkowitz devient plus performant, car ses O(n4 )
opérations arithmétiques sont exécutées sur des entiers de taille mieux
contrôlée. Si on passe à des anneaux tels que Z[t, u], l’algorithme de
Berkowitz bénéficie en plus du fait qu’il n’utilise pas de divisions. Enfin
sur des anneaux non intègres, les algorithmes de Hessenberg et de Frobenius, même dans leurs variantes avec recherche de pivot non nuls, ne
fonctionnent plus en toute généralité.
Le cas difficile : triangularisation par blocs
La méthode que nous décrivons maintenant est l’adaptation de la
précédente pour le cas le plus difficile, qui se présente cependant rarement. Cette méthode, comme celle décrite pour le cas usuel (celui où le
polynôme caractéristique de A est égal à son polynôme minimal et où le
premier vecteur de base A -engendre l’espace Kn ) fait partie de l’usage,
et nous ne savons pas à qui l’attribuer.
Soit A une matrice carrée dans Kn×n . Notons a = (e1 , . . . , en ) la
base canonique de Kn (on identifiera Kn×1 avec Kn ) et hA l’endomorphisme de Kn ayant pour matrice A dans cette base.
Nous allons construire une nouvelle base b = (f1 , . . . , fn ) dans laquelle l’endomorphisme hA aura une matrice suffisamment sympathique,
dont le polynôme caractéristique sera facile à calculer. Il s’agit précisément de réduire la matrice de hA à une forme triangulaire par blocs
avec des blocs diagonaux ayant la forme de Frobenius.
Voyons ce qui se passe sur un exemple.
Un exemple dans K7 = Q7 :
Soit (e1 , e2 , e3 , e4 , e5 , e6 , e7 ) la base canonique de K7 , v un vecteur
de K7 et A la matrice carrée d’ordre 7 ci-dessous. Rappelons que le
sous-espace de Krylov 9 associé au couple (A, v) (ou au couple (hA , v) ),
9. On dit parfois aussi sous espace cyclique.
2.8. Méthodes reliées aux suites récurrentes linéaires
103
noté KrA, v , est le sous-espace de K7 engendré par la suite récurrente
linéaire (An v)n∈N .
1
5
43
683
794
206
−1
0
A= 1
1
2
−2
−26
−458
−554
−148
−7
0
1
14
18
5
3
24
387
469
125
4
21
300
357
92
5
53
888
1082
292
−23
−163
−2547
−3074
−813
268
−186
6
157 ,
119
363
−1028
Pour obtenir une base du sous-espace KrA, e1 (de dimension k1 ) nous
calculons successivement les vecteurs e1 , Ae1 , A2 e1 , . . . en nous arrêtant
au dernier vecteur qui ne soit pas combinaison linéaire de ceux qui le
précèdent ou, ce qui revient au même, nous construisons successivement
les matrices [e1 ], [e1 | Ae1 ], [e1 | Ae1 | A2 e2 ] , . . . en nous arrêtant à la
dernière matrice dont le rang est égal au nombre de colonnes. Dans
notre cas, cela donne la suite de matrices :
1
0
0
0
0
0
0
,
1
0
0
0
0
0
0
1
−1
0
1
1
2
−7
,
1
0
0
0
0
0
0
1
−1
0
1
1
2
−7
9
−5
0
5
5
10
−35
, ...
Il faut s’arrêter à la deuxième matrice car la matrice [ e1 | Ae1 | A2 e1 ]
est de rang 2. On remarque en effet que A2 e2 = 4 e1 + 5 Ae1 . Une base
du sous-espace V1 = KrA, e1 est donc formée du couple (e1 , Ae1 ) et
dim V1 = k1 = 2. La matrice correspondante est notée U1 = [ e1 | Ae1 ].
Passons au second vecteur de la base canonique. On remarque que
e2 n’est pas dans le sous-espace V1 et l’on poursuit la construction
de la base recherchée avec les matrices [ U1 | e2 ] puis [ U1 | e2 | Ae2 ] puis
[ U1 | e2 | Ae2 | A2 e2 ] . . . jusquà obtenir une matrice dont la dernière colonne est combinaison linéaire des autres.
Ici c’est le vecteur A3 e2 qui est combinaison linéaire de ceux qui
le précèdent (c’est-à-dire qu’il appartient au sous-espace V2 = V1 +
104
2. Algorithmes de base en algèbre linéaire
e2 , Ae2 , A2 e2 ) et on obtient la suite de matrices :
1
0
0
0
0
0
0
1
−1
0
1
1
2
−7
0
1
0
0
0
0
0
,
1
0
0
0
0
0
0
1
−1
0
1
1
2
−7
0
1
0
0
0
0
0
5
−2
0
3
4
5
−23
, U2 =
1
0
0
0
0
0
0
1
−1
0
1
1
2
−7
0
1
0
0
0
0
0
5
−2
0
3
4
5
−23
86
−53
1
50
48
103
−347
.
Et la matrice [ e1 | Ae1 | e2 | Ae2 | A2 e2 | A3 e2 ] = [ U2 | A3 e2 ] :
1
1
0
5
86
348
0
−1
1
−2
−53
0
0
0
0
1
0
1
0
3
50
0
1
0
4
48
0
2
0
5
103
0
−7
0
−23
−347
−200
−2
209
214
411
−1471
est de rang 5 puisque sa dernière colonne A3 e2 est combinaison des
autres. On peut le voir par exemple par la méthode du pivot de Gauss,
qui fournit la relation de dépendance linéaire A3 e2 = 209e1 + 306Ae1 +
2e2 + Ae2 − 2A2 e2 . On passe ensuite au troisième vecteur de la base
canonique. On remarque que e3 n’est pas dans le sous-espace V2 . On
construit alors une base de V3 = V2 + KrA, e3 . On poursuit donc la
construction de la base recherchée avec les nouvelles matrices [ U2 | e3 ]
et [ U2 | e3 | Ae3 ] :
1
1
0
5
86
0
−1
1
−2
−53
0
0
0
0
1
0
1
0
3
50
0
1
0
4
48
0
2
0
5
103
0
−7
0
−23
−347
0
1
0
5
86
0
0
0
1
0
0 , 0
0
0
0 0
−1
1
−2
−53
0
0
0
0
1
1
1
0
3
50
0
1
0
4
48
0
2
0
5
103
0
−26
1
24
21
53
0
−7
0
−23
−347
0
−163
1
0
43
La dernière matrice, que nous notons U3 = U , est de rang 7. C’est la
matrice de passage de la base canonique à la base que nous venons de
construire b = (e1 , Ae1 , e2 , Ae2 , A2 e2 , e3 , Ae3 ).
Dans cette nouvelle base, il est clair que la matrice de l’endomorphisme hA est une matrice triangulaire supérieure par blocs, les blocs
diagonaux étant formés de matrices de Frobenius.
2.8. Méthodes reliées aux suites récurrentes linéaires
105
On peut d’ailleurs le vérifier, en calculant le produit matriciel U −1 A U
pour obtenir :
0
4
0
0
209
0
179
−1
U AU =
1
5
0
0
306
0
0
0
0
0
2
0
0
0
1
0
1
0
0
0
0
1
−2
0
0
0
0
0
0
0
291
6
−17
−4
1
0
0
0
0
0
1
5
dont le polynôme caractéristique (celui aussi de A ) est égal au produit
des polynômes caractéristiques
des blocs diagonaux
de Frobenius,
c’est
2
3
2
2
à-dire : X − 5 X − 4 X + 2 X − X − 2 X − 5 X − 1 .
En fait la matrice U −1 A U , et par suite les polynômes caractéristiques des blocs diagonaux de Frobenius, peuvent être retrouvés à partir
des relations de dépendance linéaires déjà calculées et de la relation qui
exprime le vecteur A2 e3 comme combinaison linéaire des vecteurs de
la base b, qui peut être obtenue en appliquant la méthode du pivot de
Gauss à la matrice [ U3 | A2 e3 ].
Description générale de l’algorithme
On prend f1 = e1 puis f2 = Ae1 , sauf si Ae1 est colinéaire avec e1 ,
auquel cas on prend f2 = e2 .
Précisément, on définit l’entier k1 ∈ {1, . . . , n} comme suit : les
vecteurs e1 , Ae1 , . . . , Ak1 −1 e1 sont indépendants, mais Ak1 e1 dépend
linéairement des précédents. Ceci définit le début
(f1 , . . . , fk1 ) = (e1 , Ae1 , . . . , Ak1 −1 e1 )
de notre base.
Les tests de dépendance linéaire dont nous avons eu besoin peuvent
être obtenus en appliquant le pivot de Gauss, avec éventuels échanges
de lignes mais sans échange de colonnes, sur les matrices successives
(e1 , Ae1 ), (e1 , Ae1 , A2 e1 ) etc. Cette méthode fournit aussi la relation de
dépendance linéaire lorsque l’entier k1 est atteint. Notez aussi que nous
n’avons pas besoin de calculer les puissances successives de la matrice
A mais seulement les transformés successifs du vecteur e1 par A.
Si k1 = n notre base b est trouvée, et en exprimant Afk1 sur la
base b nous obtenons en même temps la matrice de hA sur b sous forme
106
2. Algorithmes de base en algèbre linéaire
d’une matrice de Frobenius.
0
1
..
.
0
· · · 0 a0
. . ..
. .
a1
..
..
. 0
.
· · · 1 an−1
dont le polynôme caractéristique est, au signe près,
P (X) = X n − (an−1 X n−1 + · · · + a1 X + a0 )
Si k1 < n, nous cherchons le premier vecteur ei (i > 1) linéairement
indépendant de f1 , . . . , fk1 . Ceci nous fournit le vecteur fk1 +1 . Le calcul
de l’indice i et donc du vecteur fk1 +1 peut de nouveau être obtenu par
la méthode du pivot de Gauss (sans échange de colonnes) appliquée aux
matrices (f1 , . . . , fk1 , ei ). On définit ensuite l’entier k2 ∈ {1, . . . , n − k1 }
comme suit : les vecteurs
f1 , . . . , fk1 +1 , Afk1 +1 , . . . , Ak2 −1 fk1 +1
sont indépendants, mais Ak2 fk1 +1 dépend linéairement des précédents.
Ceci définit le nouveau début de notre base,
(f1 , . . . , fk1 +k2 ) = (f1 , . . . , fk1 +1 , Afk1 +1 , . . . , Ak2 −1 fk1 +1 ).
Si k1 + k2 = n notre base b est trouvée, et en exprimant Afk1 +k2
sur la base b nous obtenons en même temps la matrice de hA sur b sous
forme d’une matrice triangulaire par blocs, ayant pour blocs diagonaux
deux matrices de Frobenius.
0 ··· 0
a0 0 · · · · · · 0
c0
.
..
..
.
1 . . ..
a
.
.
c
1
1
..
..
..
..
.. . .
. 0
.
.
.
.
.
0 · · · 1 an−1 0 · · · · · · 0 cn−1
0 ··· ···
0
0 ··· ··· 0
b0
.
..
..
..
..
.
.
1
.
b1
.
..
..
. . . . ..
..
.
. .
.
0
.
.
..
.. . . . .
..
..
.
. 0
.
.
.
0 ··· ···
0
0 · · · 0 1 bm−1
2.8. Méthodes reliées aux suites récurrentes linéaires
107
dont le polynôme caractéristique est, au signe près,
Xm−1
Xn−1
bj X j .
ai X i · X m −
P (X) = X n −
i=0
j=0
Si k1 +k2 < n nous cherchons le vecteur fk1 +k2 +1 parmi les vecteurs
restants de la base canonique, et nous continuons le processus. En fin de
compte, en ayant calculé un nombre relativement restreint (certainement
< 2n) de produits du type matrice fois vecteur Ag, et en ayant appliqué
le pivot de Gauss un nombre relativement restreint de fois nous avons
obtenu une nouvelle base b ainsi que la matrice de hA sur cette base
sous la forme d’une matrice triangulaire par blocs, ayant sur la diagonale
des blocs formés de matrices de Frobenius. Le polynôme caractéristique
de la matrice est donc égal au produit des polynômes caractéristiques
Pi (X) des blocs diagonaux, qui sont donnés par simple lecture de la
dernière colonne de la matrice de Frobenius.
Notons pour terminer qu’il est facile de vérifier sur une telle forme
réduite
que chacun des vecteurs fj est annulé par l’endomorphisme
Q
i Pi (hA ), ce qui fournit une preuve géométrique élémentaire du théorème de Cayley-Hamilton. Pour la preuve de ce théorème il suffit d’ailleurs
de constater le fait pour le vecteur f1 , car celui-ci est simplement le
premier vecteur d’une base, et donc n’importe quel vecteur non nul a
priori.
Domaine de validité et nombre d’opérations arithmétiques
Dans cet algorithme le nombre d’opérations arithmétiques est encore
un O(n3 ).
Son domaine de validité est celui des corps, et plus généralement celui
des anneaux intègres et intégralement clos, que nous avons envisagés à
l’occasion de l’étude du polynôme minimal (voir section 1.3.2 page 20).
En effet avec un tel anneau, si C ∈ An×n les polynômes PC et
PC,v sont automatiquement à coefficients dans A. Il s’ensuit que les
procédures du type JorBarSol que nous utilisons au cours de l’algorithme ne calculent que des éléments dans A.
Dans le cas d’une matrice à coefficients dans Z on a les mêmes
majorations des coefficients intermédiaires que celles que nous avons
esquissées dans le cas facile le plus usuel.
108
2.8.2
2. Algorithmes de base en algèbre linéaire
Algorithme de Berlekamp/Massey
On donne dans un corps K les 2n premiers éléments d’une suite
récurrente linéaire (ak )k∈N pour laquelle on sait qu’il existe un polynôme générateur de degré n. Le problème est de calculer le polynôme
générateur minimal g de la suite.
Une telle solution est donnée par l’algorithme de Berlekamp/Massey
[27] qui donne en sortie le degré d ainsi que les coefficients d’un polynôme f = cd g associé au polynôme g. Ce polynôme g est alors obtenu
en divisant par cd .
L’algorithme de Berlekamp/Massey utilise les propriétés de la suite
des triplets (Ri , Ui , Vi ) formée des restes et des multiplicateurs de Bézout
successifs dans l’algorithme d’Euclide étendu
pour le couple de polynôP2n−1
2n
mes (R−1 , R0 ) où R−1 = X et R0 = i=0 ai X i .
Posant V−1 = U0 = 0 et U−1 = V0 = 1, ces triplets vérifient, pour
tout i ≥ 0, les relations :
Ri−1
Ui+1
Vi+1
Ri
de plus : Ui Vi−1 − Vi Ui−1
=
=
=
=
=
Ri Qi + Ri+1
où do Ri+1 < do Ri
Ui−1 − Qi Ui ,
Vi−1 − Qi Vi ,
d’où :
Ui R−1 + Vi R0 ,
(−1)i
et
do Ri < 2n − do Vi .
Les deux dernières relations se vérifient facilement par récurrence
sur i.
On arrête le processus au premier reste, disons Rm , de degré plus
bas que n, pour obtenir :
Um X 2n + Vm R0 = Rm
avec
do Rm < n.
Posons d = sup(do Vm , 1+do Rm ) et P = X d Vm (1/X). Alors on peut
montrer que P divisé par son coefficient dominant est le polynôme générateur minimal de la suite (ak ) (cf. [GG] et [27]). Par exemple dans
le cas où do Vm = n et Vm (0) 6= 0, en écrivant que les termes de degré
compris entre n et 2n − 1 du polynôme Vm (X) R0 (X) sont nuls, on
constate que P (X) est bien un polynôme générateur de la suite (ak ).
Ceci donne précisément l’algorithme 2.16 page suivante (dans lequel
cd(P ) désigne le coefficient dominant de P ).
Cet algorithme est dû à Berlekamp, mais sous une forme où la relation avec l’algorithme d’Euclide étendu était invisible. C’est Massey qui
a fait le rapprochement. Pour plus de détails sur la relation entre cet
algorithme et l’algorithme d’Euclide étendu, on pourra consulter [27].
2.8. Méthodes reliées aux suites récurrentes linéaires
109
Algorithme 2.16 Algorithme de Berlekamp-Massey
Entrée : Un entier n ≥ 1. Une liste non nulle d’éléments du corps K,
[a0 , a1 , . . . , a2n−1 ] : les 2n premiers termes d’une suite récurrente linéaire,
sous l’hypothèse qu’elle admet un polynôme générateur de degré ≤ n.
Sortie : Le polynôme générateur minimal P de la suite récurrente linéaire.
Début
Variables locales : R, R0 , R1 , V, V0 , V1 , Q : polynômes en X
# initialisationP
i
R0 := X 2n ; R1 := 2n−1
i=0 ai X ; V0 := 0 ; V1 := 1 ;
# boucle
tant que n ≤ deg(R1 ) faire
(Q, R) := quotient et reste de la division de R0 par R1 ;
V := V0 − Q ∗ V1 ;
V0 := V1 ; V1 := V ; R0 := R1 ; R1 := R ;
fin tant que
# sortie
d := sup(deg(V1 ), 1 + deg(R1 )) ; P := X d V1 (1/X) ;
Retourner P := P/cd(P ).
Fin.
2.8.3
Méthode de Wiedemann
L’algorithme de Wiedemann [96] pour la résolution des systèmes linéaires sur un corps K est un algorithme probabiliste, avec divisions,
qui est basé sur la théorie des suites récurrentes linéaires. Il est particulèrement efficace dans le cas des matrices creuses sur les corps finis.
Il utilise le fait que si le polynôme minimal PA d’une matrice A ∈
est de degré n (donc égal, à un signe près, au polynôme caractéristique PA de A ) , alors il existe toujours un vecteur v ∈ Kn×1 pour
lequel le polynôme générateur minimal de la suite récurrente linéaire
(Ak v) k∈N est égal à PA . Il suffit en effet, comme nous l’avons vu dans
la section 1.3 (corollaire 1.3.3), de prendre un vecteur de Kn en dehors
d’une réunion finie de sous-espaces de Kn .
L’algorithme de Wiedemann choisit au hasard une forme linéaire
π : Kn → K et un vecteur v ∈ Kn puis il calcule les 2n premiers
termes de la suite récurrente linéaire (π(Ak v)) k∈N dans K. Enfin le
polynôme générateur minimal de cette suite est obtenu par l’algorithme
Kn×n
110
2. Algorithmes de base en algèbre linéaire
de Berlekamp-Massey.
Dans le cas étudié par Wiedemann, le corps K est fini de cardinal q, et on a une mesure de probabilité naturelle, en postulant une
équiprobabilité des éléments du corps. Si le polynôme minimal de la
matrice A est égal à son polynôme caractéristique, la probabilité pour
trouver un vecteur v convenable après k essais successifs est supérieure
q k−1
à 1 − log qk−1
≥ 1 − qk−11 −1 (cf. [96]).
−1
Si on compare avec l’algorithme de Frobenius, on voit que l’on doit
calculer 2n vecteurs Ak v au lieu de n. Par contre le calcul du polynôme générateur minimal est ensuite beaucoup plus rapide. En outre,
dans le cas des matrices creuses, même le calcul des 2n vecteurs Ak v
est très rapide.
Notons enfin que les algorithmes de Frobenius et de Wiedemann
peuvent être accélérés très significativement au moyen de la multiplication rapide des polynômes et de la multiplication rapide des matrices
(cf. sections 8.4 et 8.5).
3. Circuits arithmétiques
Introduction
Dans ce chapitre nous introduisons la notion fondamentale de circuit
arithmétique qui est le cadre général dans lequel se situe l’analyse des
algorithmes développés dans cet ouvrage.
La complexité algébrique peut être vue comme une théorie qui cherche à analyser les algorithmes qui acceptent de se mettre sous forme de
familles de circuits arithmétiques.
Dans un circuit arithmétique les instructions de branchement ne sont
pas autorisées, ce qui semble une limitation assez sévère. Les algorithmes
usuels d’algèbre linéaire sont en effet ordinairement écrits en utilisant
des tests d’égalité à 0. Néanmoins, il s’avère que dans beaucoup de cas,
cette limitation apparente n’en est pas une, notamment en raison de
la procédure d’élimination des divisions de Strassen que nous exposons
dans la section 3.2. Par contre le cadre un peu strict fourni par les circuits arithmétiques s’avère très fécond. C’est grâce à lui que l’on peut
mettre en place la stratégie générale (( diviser pour gagner )).
Lorsqu’on envisage les algorithmes liés à la géométrie algébrique
réelle, la nécessité des tests de signe, et donc des instructions de branchement, devient souvent impérieuse, et une autre branche de la complexité
algébrique est nécessaire, avec la théorie des réseaux arithmétiques que
nous ne développerons pas ici.
Dans la section 3.1 nous donnons les définitions précises des circuits
arithmétiques et de leur variante (( programmée )), les programmes d’évaluation (straight-line programs en anglais). C’est l’occasion d’introduire
quelques mesures de complexité pour ces algorithmes.
Dans la section 3.2 nous introduisons l’élimination des divisions selon
la méthode de Strassen et nous établissons quelques uns des résultats les
plus importants qui la concernent.
Dans la section 3.3 nous donnons une méthode qui transforme un cir-
112
3. Circuits arithmétiques
cuit arithmétique Γ qui calcule une fraction rationnelle f en un circuit
arithmétique Γ0 de taille comparable (la taille est multipliée par au plus
5), qui calcule à la fois la fonction f et toutes ses dérivées partielles.
3.1
Circuits arithmétiques et programmes d’évaluation
Un circuit arithmétique constitue une façon naturelle et simple de
représenter les calculs algébriques dans un anneau arbitraire, dans le
cas où un algorithme n’utilise pas d’instructions de branchements, et
uniquement des boucles du type
pour i de m à n faire . . . fin pour.
Si la taille de l’entrée est fixée, ces boucles peuvent être (( mises à plat ))
et on obtient un programme dont les seules instructions sont des affectations.
La plupart des algorithmes présentés dans le chapitre 2 sont de ce
type et donnent donc lieu, lorsque les dimensions des matrices sont fixées,
à des programmes d’évaluation.
3.1.1
Quelques définitions
Par exemple l’idée d’un circuit arithmétique est donnée par le calcul
du déterminant d’une matrice carrée par l’algorithme du pivot de Gauss
simplifié, dans le cas des matrices fortement régulières, pour des matrices
de taille fixée. Le calcul est alors toujours exactement le même et peut
être représenté comme une suite d’affectations qu’on peut disposer séquentiellement ou dessiner au moyen d’un graphe plan.
Par exemple pour une matrice 4 × 4, en donnant un nom différent
à chaque résultat de calcul élémentaire (addition, soustraction, multi[k]
plication ou division), et en reprenant la notation aij introduite à la
section 2.1, on obtient la mise à plat sous la forme du programme d’évaluation 3.1 page ci-contre dans lequel toutes les affectations situées à une
même profondeur peuvent en principe être exécutées simultanément.
Pour une profondeur donnée, les calculs sont faits avec des variables définies aux étages précédents. Ce calcul comprend 37 opérations
arithmétiques, et sa profondeur est égale à 10, sa largeur est égale à 9.
3.1. Circuits arithmétiques et programmes d’évaluation
113
Programme d’évaluation 3.1 Calcul du déterminant et de la
LU-décomposition d’une matrice carrée d’ordre 4 par la méthode du pivot de Gauss (sans recherche de pivot).
Entrée : Une matrice A = (aij ) ∈ K4×4 à coefficients dans un corps K.
Sortie : Les coefficients lij en dessous de la diagonale de la matrice L, les
[i−1]
coefficients uij = aij (j ≥ i) de la matrice U , le déterminant d4 de A.
Début
profondeur 1 : Traitement du premier pivot
l21 := a21 /a11 ; l31 := a31 /a11 ; l41 := a41 /a11
profondeur 2 : largeur 9
[1]
[1]
[1]
b22 := l21 a12 ; b23 := l21 a13 ; b24 := l21 a14 ;
[1]
[1]
[1]
b32 := l31 a12 ; b33 := l31 a13 ; b34 := l31 a14 ;
[1]
[1]
[1]
b42 := l41 a12 ; b43 := l41 a13 ; b44 := l31 a44
profondeur 3 : largeur 9
[1]
[1]
[1]
[1]
[1]
[1]
a22 := a22 − b22 ; a23 := a23 − b23 ; a24 := a24 − b24 ;
[1]
[1]
[1]
[1]
[1]
[1]
a32 := a32 − b32 ; a33 := a33 − b33 ; a34 := a33 − b34 ;
[1]
[1]
[1]
[1]
[1]
[1]
a42 := a42 − b42 ; a43 := a43 − b43 ; a44 := a43 − b44
profondeur 4 : Traitement du deuxième pivot
[1] [1]
[1] [1]
[1]
l32 := a32 /a22 ; l42 := a42 /a22 ; d2 := a11 a22
profondeur 5 : largeur 4
[2]
[1]
[2]
[1]
b33 := l32 a23 ; b34 := l32 a24 ;
[2]
[1]
[2]
[1]
b43 := l42 a23 ; b44 := l42 a24
profondeur 6 : largeur 4
[2]
[1]
[2]
[2]
[1]
[2]
a33 := a33 − b33 ; a34 := a34 − b34 ;
[2]
[1]
[2]
[2]
[1]
[2]
a43 := a43 − b43 ; a44 := a44 − b44
profondeur 7 : Traitement du troisième pivot
[2] [2]
[2] [2]
[2]
l43 := a43 /a33 ; l44 := a44 /a33 ; d3 := d2 a33
profondeur 8 :
[3]
[2]
b44 := l43 a44
profondeur 9 :
[3]
[2]
[3]
a44 := a44 − b44
profondeur 10 :
[3]
d4 := d3 a44
Fin.
Plus généralement.
114
3. Circuits arithmétiques
Définition 3.1.1 Un programme d’évaluation arithmétique P sans division (resp. avec division) avec constantes dans C, où C est une partie
(codée) d’un anneau A ou d’un corps K, est la donnée :
— d’un ensemble de variables xp,u où p est un entier ≥ 0 donnant la
profondeur de la variable et (p, u) est l’identificateur de la variable,
— d’une suite d’instructions d’affectations de l’un des types suivants :
– xp,u := a ◦ b où a et b désignent ou bien une variable xq,u
avec q < p ou bien une constante c ∈ C, et ◦ ∈ {+, −, ×} (resp.
◦ ∈ {+, −, ×, /}).
– xp,u := a (avec les mêmes conventions pour a).
Mises à part les variables x0,u qui sont les entrées du programmes d’évaluation, toutes les variables xp,u sont affectées exactement une fois dans
le programme. Ce sont les variables d’affectation du programme.
Les constantes sont considérées comme de profondeur nulle. En conséquence on note prof(a) = 0 si a ∈ C et prof(xp,u ) = p.
Quelques commentaires sur cette définition.
Dans le cas d’un programme d’évaluation avec divisions, l’évaluation
peut échouer pour certaines valeurs des variables d’entrée dans le corps
K. Souvent l’ensemble C est vide ou réduit à {0, 1}. Le programme
peut alors être évalué sur un corps arbitraire (sur un anneau arbitraire
s’il est sans division).
Naturellement, tous les identificateurs doivent être distincts. Les affectations du type xp,u := a sont prévues uniquement pour le cas où on
désire respecter certaines contraintes dans une gestion précise des étapes
parallèles.
Ordinairement, on demande que dans une affectation xp,u := a ◦ b
on ait prof(xp,u ) = 1 + max(prof(a), prof(b)) et dans une affectation
xp,u := a on ait prof(xp,u ) = 1 + prof(a). On peut aussi demander que
dans une affectation xp,u := a ◦ b on ait
a ∈ C ou p = 1 + prof(a), et b ∈ C ou p = 1 + prof(b).
Le texte du programme d’évaluation doit normalement préciser quelles sont les variables représentant les sorties. Mais on peut demander
que les sorties soient exclusivement les variables de profondeur maximum. On peut demander aussi que toute variable de profondeur non
maximum soit utilisée.
Remarque 3.1.2 De manière plus générale, un programme d’évaluation peut être défini pour n’importe quel type de structure algébrique,
une fois qu’ont été précisés les opérateurs de base dans la structure,
3.1. Circuits arithmétiques et programmes d’évaluation
115
qui peuvent être de n’importe quelle arité. Par exemple un programme d’évaluation booléen correspond à la structure d’algèbre de Boole
avec les opérateurs booléens usuels. Autre exemple, dans les anneaux
commutatifs, on peut définir une notion de programme d’évaluation avec
déterminants si on introduit en tant qu’opérateurs de base les detn
comme opérations d’arité n2 qui donnent le déterminant d’une matrice
n × n en fonction de ses entrées.
Définition 3.1.3 Nous utiliserons la terminologie suivante concernant
les programmes d’évaluation :
– Le nombre des entrées dans un programme d’évaluation est en général contrôlé par un ou plusieurs paramètres qu’on appelle les paramètres d’entrée du programme. Par exemple, dans un programme d’évaluation qui calcule le produit de deux matrices n × n, on
prend l’entier n comme paramètre d’entrée et dans un programme
d’évaluation associé à la résolution d’un système de m équations
polynomiales à n indéterminées, de degré maximum d, écrites en
représentation dense 1 , les paramètres d’entrée sont m, n, d.
– Dans une affectation du type xp,u := a ◦ b, a et b sont les antécédents de xp,u .
– La profondeur du programme d’évaluation est la profondeur maximum de ses variables d’affectation ; notée prof(P ), elle correspond
au nombre d’étapes parallèles du programme d’évaluation P .
– La taille ou longueur du programme d’évaluation désignera le nombre
total de toutes les opérations arithmétiques, c’est-à-dire les affectations du type xp,u := a ◦ b.
– Pour chaque étape p (1 ≤ p ≤ prof(P )), on considère le nombre τp
d’opérations effectuées durant cette étape. On appelle largeur du
programme d’évaluation le plus grand de ces nombres , c’est-à-dire
max {τp | 1 ≤ p ≤ prof(P )}.
– Lors de l’évaluation d’un programme d’évaluation arithmétique sur
un anneau ou sur un corps dont les éléments sont codés, les entrées
xi ont a priori n’importe quelle taille tandis que les constantes du
programme ont une taille fixée une fois pour toutes. Du point de
vue du calcul concret sur des objets codés, on est donc souvent en
1. Un polynôme est codé en représentation dense lorsque le codage donne la liste
de tous les coefficients des monômes en dessous d’un degré donné, dans un ordre
convenu. Il est codé en représentation creuse lorsque le codage donne la liste des
paires (am , m) où m code un monôme (par exemple x2 y 3 z 5 peut être codé par
(2, 3, 5) et am son coefficient (non nul) dans le polynôme.
116
3. Circuits arithmétiques
droit d’estimer que seules importent vraiment les affectations sans
scalaires, c’est-à-dire celles du type xp,u := a±b et xp,u := a×b ou
aucun des deux antécédents n’est une constante, ainsi que xp,u :=
a/b où b n’est pas une constante. Ceci donne lieu aux notions de
longueur stricte et de profondeur stricte, dans lesquelles seules
sont prises en compte les affectations sans scalaires. Une multiplication ou division sans scalaire dans un programme d’évaluation
arithmétique est encore dite essentielle.
– Variation sur le thème précédent. Dans la mesure où on consière
que les additions ainsi que les multiplications ou divisions par
des constantes sont relativement peu coûteuses (ou éventuellement
pour des raisons plus profondes d’ordre théorique) on est intéressé
par la longueur multiplicative d’un programme d’évaluation et par
sa profondeur multiplicative qui sont définies comme la longueur
et la profondeur mais en ne tenant compte que des multiplications
et divisions essentielles.
Par exemple le programme d’évaluation 3.1 a une profondeur multiplicative égale à 6 et une largeur multiplicative égale à 9.
3.1.2
Circuit arithmétique vu comme un graphe
On peut également représenter un programme d’évaluation sous forme
d’un dessin plan. Par exemple pour le calcul du déterminant par l’algorithme du pivot de Gauss avec une matrice fortement régulière 3 × 3,
on peut le représenter par le dessin du circuit 3.2 page suivante.
Pour une matrice n × n, on obtiendra un circuit de profondeur 3n − 2
avec un nombre de portes, en tenant compte des n − 1 affectations
[p−1]
dp := dp−1 a pp
(2 ≤ p ≤ n) qui donnent les mineurs principaux
dominants de la matrice, égal à :
(n − 1) +
Xn−1
k=1
k (2k + 1) =
1
(n − 1) (4n2 + n + 6) .
6
Si on veut formaliser ce genre de dessin qui visualise bien la situation,
on peut adopter la définition suivante.
Définition 3.1.4 Un circuit arithmétique avec divisions (resp. sans
division) (( avec constantes dans C )), où C est une partie (codée) d’un
anneau A ou d’un corps K, est un graphe acyclique orienté et étiqueté,
chaque nœud qui n’est pas une porte d’entrée ayant exactement deux
3.1. Circuits arithmétiques et programmes d’évaluation
a [0]
3,3
a [0]
2,3
a [0]
1,3
a [0]
3,2
a [0]
3,1
a [0]
2,1
a [0]
1,1
[1]
3,3
b [1]
2,3
b [1]
2,2
a [1]
2,2
b [1]
3,2
a [1]
3,2
a [1]
2,3
a [1]
3,3
a [0]
1,2
l 2,1
l 3,1
b
a [0]
2,2
117
l 3,2
b [2]
3,3
d2
a [2]
3,3
d3
(
: division ; ⊗ : multiplication ;
: soustraction)
Pour la division et la soustraction
le brin entrant gauche représente le premier terme de l’opération
Circuit 3.2: Pivot de Gauss simplifié pour une matrice 3 × 3
antécédents. Le circuit est étiqueté de la manière suivante :
— chaque porte d’entrée est étiquetée par un triplet (0, n, c) où 0 est
la profondeur, (0, n) est le nom qui identifie le nœud et c ∈ {x} ∪ C
(avec x ∈
/ C). Ici un triplet (0, n, x) représente la variable xn et un
triplet (0, n, c) avec c ∈ C représente l’élément (codé par) c de A.
— chaque nœud interne et chaque porte de sortie est étiqueté par un
triplet (m, n, ◦) où m est sa profondeur, (m, n) est son identificateur,
et ◦ ∈ {+, −, ×, /} (resp. ◦ ∈ {+, −, ×}) désigne une opération arithmétique.
— enfin, dans le cas des opérateurs / et − (et dans le cas de l’opérateur
× si le circuit est destiné à être évalué dans un anneau non commutatif ) il faut étiqueter de manière à les distinguer (gauche, droite) les deux
arcs qui aboutissent à un nœud correspondant à cet opérateur.
— les portes de sortie correspondant aux résultats du calcul sont spécifiées
par une marque distinctive dans leur identificateur.
En fait, dans toute la suite, nous utiliserons indifféremment (( circuit arithmétique )) et (( programme d’évaluation arithmétique )), tout
en sous-entendant que, pour ce qui est du codage, nous choisissons toujours un codage correspondant à la définition d’un programme d’évaluation arithmétique. De la même manière, nous considérerons comme
118
3. Circuits arithmétiques
synonymes programme d’évaluation booléen et circuit booléen.
Dans un circuit arithmétique on peut interpréter chaque nœud comme représentant un polynôme de A [(xi )i∈I ] ou K [(xi )i∈I ] (dans le cas
sans division) ou une fraction rationnelle de K((xi )i∈I ) (dans le cas avec
division).
3.1.3
Circuits arithmétiques homogènes
Définition 3.1.5 On appelle circuit arithmétique homogène un circuit
arithmétique sans division dont tous les noeuds représentent des polynômes homogènes et qui a la structure suivante.
– Les polynômes de degré d sont calculés après ceux de degrés strictement inférieurs.
– Le calcul des polynômes de degré d se fait en deux phases. Dans
la première phase, en une seule étape parallèle, on effectue des
produits de polynômes précédemment calculés (de degrés d0 < d
et d − d0 ). Dans la deuxième phase on calcule des combinaisons
linéaires des précédents.
Proposition 3.1.6 Tout circuit arithmétique sans division qui calcule
une famille de polynômes de degré ≤ d peut être réorganisé en un circuit arithmétique homogène qui calcule toutes les composantes homogènes des polynômes en sortie. Le circuit arithmétique homogène obtenu
est de profondeur multiplicative d − 1. Par rapport au circuit initial, la
profondeur a été multipliée par O(log d), la longueur multiplicative a
été au plus multipliée par d(d − 1)/2, et la longueur totale a été au plus
multipliée par (d + 1)2 .
Preuve. Chaque noeud yj du circuit initial représente un polynôme en
les entrées xi qu’on décompose en somme de composantes homogènes.
[0]
[1]
[d]
yj := yj + yj + · · · + yj + des composantes sans importance
On analyse alors le calcul qui est fait sur les composantes homogènes de
degré ≤ d.
Lorsqu’on a dans le circuit arithmétique original une affectation correspondant à une addition y` := yh + yk on obtient sur les composantes homogènes au plus d + 1 additions qui peuvent être exécutées en
dlog(d + 1)e étapes parallèles.
3.1. Circuits arithmétiques et programmes d’évaluation
119
Lorsqu’on a dans le circuit arithmétique original une affectation correspondant à une multiplication essentielle y` := yh yk on obtient
[0]
[0]
[0]
y` = yh yk
[1]
[0] [1]
[1] [0]
y` = yh yk + yh yk
[2]
[0] [2]
[1] [1]
[2] [0]
y` = yh yk + yh yk + yh yk
.. ..
..
. .
.
[d]
[0] [d]
[d−1] [1]
[d] [0]
= yh yk + yh
yk + · · · + yh yk
y`
ce qui correspond à (au plus) d(d − 1)/2 multiplications essentielles
entre les composantes homogènes, 2d + 1 multiplications scalaires, et
d(d + 1)/2 additions, soit (d + 1)2 opérations arithmétiques en tout.
Par ailleurs on peut réorganiser l’ensemble du calcul de manière que tous
les polynômes homogènes de degré k soient calculés après ceux de degré
< k.
t
u
3.1.4
Le problème des divisions dans les circuits arithmétiques
Certains circuits arithmétiques avec division comportent une division
par une fraction rationnelle identiquement nulle, et ne représentent plus
aucun calcul raisonnable. Implicitement on suppose toujours qu’on n’est
pas dans ce cas.
Le cas de l’algorithme de Jordan-Bareiss sans recherche de pivot est
un peu plus subtil. Il correspond à un circuit arithmétique (( avec divisions exactes )), c’est-à-dire que, lorsqu’on le regarde comme produisant
à chaque porte un élément du corps des fractions rationnelles, on reste
en fait toujours dans l’anneau des polynômes : les divisions ont toujours
pour résultat un polynôme et non une fraction rationnelle.
Si les portes de sortie d’un circuit arithmétique sont des polynômes
(en les entrées) il est a priori préférable que le circuit soit sans division.
Il pourra en effet être évalué dans n’importe quel anneau.
Dans le cas d’un circuit avec divisions évalué dans un corps, il se
peut que certaines divisions soient impossibles, non parce qu’on doit
diviser par une fraction rationnelle identiquement nulle, mais parce que
les valeurs des xi annulent la fraction rationnelle du dénominateur. C’est
encore une raison qui fait qu’on préfère les circuits sans divisions.
Une autre raison est que si le corps K est de caractéristique nulle ou
s’il contient des éléments transcendants, l’addition de deux fractions est
120
3. Circuits arithmétiques
une affaire bien encombrante. L’addition dans Q par exemple, réclame,
dans Z tout d’abord 3 multiplications et une addition, suivies d’une
simplification de fraction, qui réclame un calcul de pgcd, donc les divisions successives de l’algorithme d’Euclide. Ainsi, lorsque les entrées
sont dans Z par exemple, on préfère que tout le calcul reste dans Z.
Si on essaie d’évaluer un circuit avec divisions dans un anneau arbitraire A, la situation est encore un peu compliquée. Toute division par
un diviseur de zéro est impossible. Et si on divise par un non diviseur
de zéro, on se retrouve naturellement dans l’anneau total des fractions
de A, défini de la même manière que le corps des fractions d’un anneau intègre, mais en autorisant comme dénominateurs uniquement des
non diviseurs de zéro dans A. Naturellement, les calculs dans ce nouvel
anneau A0 sont nettement plus compliqués que ceux dans A (cf. la
discussion à propos de la méthode du pivot dans Q.)
La profondeur d’un circuit est un paramètre pertinent à plus d’un
titre. Tout d’abord, la profondeur représente en quelque sorte le (( temps
de calcul parallèle )) si on donne une unité de temps pour chaque opération arithmétique et si on dispose de suffisamment de (( processeurs ))
entre lesquels on répartit les calculs à faire. Ensuite, la profondeur permet un contrôle de la taille des objets intermédiaires lorsque le calcul
est effectué comme une évaluation par exemple dans Z, Q, ou Q(x, y).
Grosso modo, la taille double au maximum lorsque la profondeur augmente de 1. Dans le cas des circuits sans divisions évalués par exemple
dans Z ou Z[x, y] la profondeur multiplicative est de loin la plus importante pour le contrôle de la taille des objets intermédiaires.
Tout ceci a conduit à attacher une importance toute particulière aux
circuits sans division et de faible profondeur.
3.2
Élimination des divisions à la Strassen
Lorqu’on dispose d’une procédure utilisant les divisions dans le corps
des fractions rationnelles pour calculer un polynôme de degré déterminé
à coefficients dans un anneau intègre, une technique de Strassen ([87])
basée sur une idée très simple permet d’éliminer toutes les divisions dans
cette procédure.
3.2. Élimination des divisions à la Strassen
3.2.1
121
Le principe général
L’idée de base est que la division par un polynôme de la forme (1−u)
où u = u((xi )) peut être remplacée par le produit par la série formelle
1 + u + · · · + um + · · ·
à condition d’être dans une situation où on sait qu’on peut ne considérer
qu’une partie finie bien contrôlée des séries formelles en jeu.
Nous allons expliquer cette idée fondamentale sur l’exemple du calcul
du déterminant d’une matrice carrée A par l’algorithme du pivot de
Gauss simplifié (c’est-à-dire sans recherche de pivot) mis sous forme
d’un circuit arithmétique pour les matrices n × n pour une valeur fixée
de n. On considère ce circuit comme un programme d’évaluation dans
un anneau arbitraire A (on peut se limiter au sous-anneau engendré par
les coefficients de la matrice, ou plutôt à l’anneau total des fractions de
ce sous-anneau). Naturellement un obstacle apparaı̂t éventuellement lors
d’une affectation vh := vk /v` si v` est diviseur de zéro. Il y a cependant
des cas où cet obstacle n’apparaı̂t pas du tout, le plus simple est celui
où la matrice de départ est égale à In : toutes les divisions se font par
1 ! Cette remarque d’apparence anodine est cependant la clé de l’élimination des divisions. En effet, il suffit de faire le changement de variable
F := A − In et de décider d’évaluer le circuit pour l’entrée In + zF ,
où z est une nouvelle variable, dans l’anneau A[z] z n+1 : l’anneau
des développements limités à l’ordre n à coefficients dans A, que nous
noterons souvent An .
Quelle que soit la matrice F à coefficients dans A prise en entrée,
chaque nœud v` intervenant dans une division est maintenant un développement limité du type
v` = 1 + c`,1 z + · · · + c`,n z n ,
(c`,1 , . . . , c`,n ∈ A)
c’est-à-dire un élément inversible de An = A[z] z n+1 . A la fin du
calcul on récupère donc det(In + zF ) dans An , c’est-à-dire en fait : on
récupère det(In + zF ) dans A[z]( 2 ). Et il suffit de faire z = 1 pour
2. Dans le cas présent, il serait donc plus astucieux, d’appliquer la procédure avec
la matrice A à la place de la matrice F , car on obtient ici à très peu près le polynôme caractéristique de F . Dans le cas présent la procédure d’élimination des divisions
est donc très proche de la méthode de Jordan-Bareiss modifiée. Cette dernière est
cependant un peu plus simple, car dans la méthode de Strassen on manipule très
rapidement des polynômes de degré n (l’ordre de la matrice). Pour terminer notons
que c’est un fait d’expérience assez curieux que les procédures (( rapides )) de calcul
sans division du déterminant passent toutes par le calcul du polynôme caractéristique.
122
3. Circuits arithmétiques
obtenir det(A). En fait la division dans An d’un élément a(z) par un
élément inversible b(z) = 1−zu(z) ne nécessite que des additions et multiplications dans An : on peut en effet faire une division en puissances
croissantes de a(z) par b(z) jusqu’à l’ordre n en z. On peut également
invoquer la formule (valable dans An )
k
(1 − zu(z))−1 = (1 − w)−1 = (1 + w)(1 + w2 )(1 + w4 ) · · · (1 + w2 ) (3.1)
si 2k+1 ≥ n + 1 (il suffit de prendre k = dlog (n + 1)e − 1).
Ainsi toutes les affectations (correspondant à l’évaluation du circuit)
dans An se ramènent à des additions et multiplications de polynômes
tronqués, c’est-à-dire encore à des additions et multiplications dans A.
Le théorème suivant est maintenant clair :
Théorème 3.1 La procédure d’élimination des divisions de Strassen
peut être appliquée à tout circuit arithmétique pourvu qu’on soit dans le
cas suivant : on connaı̂t un point (ξ1 , . . . , ξn ) de (( l’espace des entrées ))
tel que, lorsque le circuit est évalué en ce point, toutes les divisions qui
doivent être exécutées le sont par des éléments inversibles de l’anneau
de base (on rajoute alors ces éléments et leurs inverses à l’ensemble des
constantes C du circuit).
En particulier l’élimination des divisions est toujours possible si l’anneau
de base est un corps infini.
Définition 3.2.1 Éliminer les divisions (à la Strassen) dans un circuit
arithmétique à partir du point (ξ1 , . . . , ξn ), c’est lui appliquer la procédure d’élimination des divisions de Strassen en utilisant (ξ1 , . . . , ξn )
comme point en lequel le circuit est évalué sans divisions. Nous appellerons ce point le centre d’élimination des divisions.
Sur un corps infini, l’existence d’un centre d’élimination des divisions
pour un circuit arithmétique résulte du fait qu’on peut toujours éviter
l’ensemble des zéros d’une famille finie de polynômes non (formellement)
nuls : leur produit est un polynôme non formellement nul et un tel polynôme définit une fonction non identiquement nulle sur Kn (n est le
nombre de variables) si K est infini.
Un exemple d’élimination des divisions
Donnons à titre d’exemple le résultat de l’élimination des divisions
pour l’algorithme du pivot de Gauss simplifié, dans le cas n = 3, pour
3.2. Élimination des divisions à la Strassen
123
une matrice I3 +zF . Le circuit initial est donné par le programme d’évaluation 3.3.
Programme d’évaluation 3.3 Calcul du déterminant de la matrice carrée I3 + zF par la méthode du pivot de Gauss.
Entrée : Les coefficients fij de la matrice F dans un anneau commutatif
arbitraire A.
Sortie : Le déterminant d3 = det(I3 + zF ). Le calcul est correct si on se
situe dans un anneau B contenant z et A et dans lequel tous les éléments
de la forme 1 + zb sont inversibles. Les opérations arithmétiques de ce programme d’évaluation sont effectuées dans B. Notez que les coefficients de
I3 + zF sont les éléments zfij pour i 6= j et les éléments (1 + zfii ) pour
i = j.
Début
profondeur 1 : Traitement du premier pivot
l21 := z f 21 /(1 + zf 11 ) ; l31 := z f 31 /(1 + zf 11 )
profondeur 2 :
[1]
[1]
b 22 := l21 z f 12 ; b 23 := l21 z f 13 ;
[1]
[1]
b 32 := l31 z f 12 ; b 33 := l31 z f 13
profondeur 3 :
[1]
[1]
[1]
[1]
f 22 := z f 22 − b 22 ; f 23 := z f 23 − b 23 ;
[1]
[1]
[1]
[1]
f 32 := z f 32 − b 32 ; f 33 := z f 33 − b 33
profondeur 4 : Traitement du deuxième pivot
[1]
[1]
[1]
l32 := f 32 /(1 + f 22 ) ; d2 := (1 + z f 11 ) (1 + f 22 )
profondeur 5 :
[2]
[1]
b 33 := l32 f 23
profondeur 6 :
[2]
[1]
[2]
f 33 := f 33 − b 33
profondeur 7 :
[2]
d3 := d2 (1 + f 33 )
Fin.
Pour passer de l’ancien circuit (programme d’évaluation 3.3) au nouveau (programme d’évaluation 3.4 page suivante), chaque porte yij ou
di (sauf les portes d’entrée) a été remplacée par les portes yijk ou dik ,
avec k = 0, . . . , 3, qui donnent les quatre premiers coefficients de la
série formelle en z.
124
3. Circuits arithmétiques
Dans l’algorithme transformé 3.4, nous n’avons pas écrit les portes
nulles (pour les bas degrés) et nous n’avons pas mentionné les yijk qu’il
est inutile d’évaluer pour obtenir le résultat final.
Il faut remarquer que pour le déterminant et même le polynôme caractéristique des matrices 3 × 3, les formules directes sont bien entendu
préférables.
Programme d’évaluation 3.4 Calcul du déterminant d’une matrice carrée F par la méthode du pivot de Gauss après élimination des divisions à la Strassen.
Entrée : Les coefficients fij de la matrice F dans un anneau commutatif
arbitraire.
Sortie : Le déterminant det(F ). En fait, on calcule même d3 = det(I3 +
zF ) = 1 + d31 z + d32 z 2 + d33 z 3 . L’algorithme fonctionne (( en ligne droite ))
et sans aucune hypothèse restrictive. Les opérations arithmétiques de ce
programme d’évaluation sont effectuées dans A.
Début
[1]
[1]
Renommages : l211 = f21 , l311 = f31 , f221 = f22 , f231 = f23 ,
[1]
[1]
l321 = f321 = f32 , f331 = f33 .
profondeur 1 : Traitement du premier pivot
l212 := −f21 f11 ; l312 := −f31 f11 ; d21 := f11 + f22
profondeur 2 :
[1]
[1]
f222 := −l211 f12 ; f232 := −l211 f13 ;
[1]
[1]
[1]
f322 := −l311 f12 ; f332 := −l311 f13 ; f333 := −l312 f13
profondeur 3 :
[1]
[1]
d22 := f11 f221 + f222
profondeur 4 : Traitement du deuxième pivot
[1]
[1] [1]
l322 := f322 − f221 f321
profondeur 6 :
[2]
[1]
[2]
[1]
[1]
b332 := l321 f231 ; b333 := l321 f232 + l322 f231
profondeur 7 :
[2]
[1]
[2]
[1]
[2]
[2]
[1]
[2]
f331 := f331 ; f332 := f332 − b332 ; f333 := f333 − b333
profondeur 8 :
[2]
d31 := d21 + f331
profondeur 9 :
[2]
[2]
[2]
[2]
[2]
d32 := d22 + d21 f331 + f332 ; d33 := d22 f331 + d21 f332 + f333
Fin.
3.2. Élimination des divisions à la Strassen
3.2.2
125
Coût de l’élimination des divisions
Quel est le coût de la transformation d’un circuit avec division en
un circuit sans division, lorsque les sorties sont des polynômes de degré
≤ n en les entrées ?
Tout d’abord si on utilise les algorithmes usuels pour les opérations
arithmétiques dans An , la taille du circuit sera en gros multipliée par n2
(ce qui fait qu’on reste dans le cadre des circuits de taille polynomiale).
Par exemple le produit de deux éléments de An réclame (n+1)(n+2)/2
multiplications et n (n + 1)/2 additions, tandis que la division de a(z)
par 1−zu(z) nécessite n (n+1)/2 multiplications et autant d’additions
si on effectue la division en puissance croissante.
Si on applique ces constatations dans le cas du calcul du déterminant (et du polynôme caractéristique par la même occasion) d’une matrice carrée par élimination des divisions dans l’algorithme du pivot de
Gauss comme nous l’avons
3.2.1, on trouve une taille de
Pn−1 vu2 à la section
2 c’est-à-dire un 1 n5 + O(n4 ), à
n
(n
−
k)
circuit équivalente à
k=1
3
1 5
n + O(n4 ) que nous avons obtenu pour l’algorithme de
comparer au 10
Jordan-Bareiss modifié.
Notons aussi que la multiplication dans An par l’algorithme usuel
se fait naturellement en profondeur O(log n) tandis que la division par
puissance croissante est en profondeur O(n log n). On peut pallier ce
dernier inconvénient en utilisant la formule (3.1) qui donne un circuit
de taille O(n2 log n) et de profondeur O(log2 n).
Il existe par ailleurs des procédures de multiplication rapide pour
les polynômes : les opérations arithmétiques +, −, × et division par un
élément f vérifiant f (0) = 1 dans An peuvent être exécutées par des
circuits de taille O(n log n log log n) et de profondeur O(log n) (voir
[13] et, infra, le théorème 6.2 page 182).
Plus généralement nous utiliserons la notation suivante.
Notation 3.2.2 Pour un anneau A fixé par le contexte, nous noterons
µP (n) le nombre d’opérations arithmétiques nécessaires pour la multiplication de deux polynômes de degré n en profondeur O(log n).
Strassen obtient alors précisément le résultat suivant :
Théorème 3.2 Lorsqu’on élimine les divisions à la Strassen pour l’évaluation d’une famille de polynômes de degrés ≤ n la profondeur du circuit est multipliée par O(log n) et sa taille par O(µP (n)).
126
3. Circuits arithmétiques
Notons aussi le résultat suivant simple et intéressant concernant les
circuits arithmétiques qui évaluent des familles de polynômes du second
degré.
Proposition 3.2.3 Lorsqu’on élimine les divisions à la Strassen pour
l’évaluation d’une famille de polynômes de degré ≤ 2, la longueur multiplicative du circuit arithmétique est inchangée.
Preuve. Lorsqu’on applique la procédure d’élimination des divisions,
supposons qu’on ait f = f0 + zf1 + z 2 f2 , et g = g0 + zg1 + z 2 g2 dans
l’anneau des développements limités à l’ordre 2 en z sur A [(xi )] (ici
on suppose sans perte de généralité que (0, . . . , 0) est le centre d’élimination des divisions et donc que les fj et gj sont homogènes de
degré j en les entrées xi ). On obtient pour le produit h = f g modulo
z 3 , h = h0 + zh1 + z 2 h2 avec h0 = f0 g0 , h1 = f1 g0 + f0 g1 et h2 =
f2 g0 + f1 g1 + f0 g2 avec la seule multiplication essentielle f1 g1 puisque
f0 et g0 sont des constantes. Et on a un calcul analogue pour k = f /g.
k0 = f0 /g0 , k1 = f1 /g0 − g1 (f0 /g02 ) et k2 = f2 /g0 − k1 g1 /g0 − g2 k0 /g0
avec la seule multiplication essentielle k1 g1 .
t
u
On pourrait généraliser avec un circuit arithmétique calculant une
famille de polynômes de degrés ≤ d.
3.3
Calcul de toutes les dérivées partielles d’un
polynôme ou d’une fraction rationnelle
Nous donnons une méthode pour transformer un circuit arithmétique Γ qui calcule une fraction rationnelle f en un circuit arithmétique
Γ0 de taille comparable (la taille est multipliée par au plus 5), qui calcule
à la fois la fonction f et toutes ses dérivées partielles.
Si le circuit arithmétique Γ est sans division, il en est de même pour
0
Γ . La méthode est due à Baur & Strassen [5]. Nous suivons l’exposé
simple et constructif que Morgenstern en fait dans [72].
Une application importante de ce résultat concerne le calcul de l’adjointe d’une matrice avec un coût voisin de celui de son déterminant. En
effet les coefficients bij de la matrice adjointe de A sont donnés par :
bij = (−1)i+j det(A ↓ji ) =
∂ det(A)
∂ aji
3.3. Calcul des dérivées partielles
127
où det(A ↓ji ) est le mineur d’ordre n − 1 obtenu en supprimant la j ème
ligne et la ième colonne de la matrice A.
Nous montrons le résultat par récurrence sur la longueur du programme d’évaluation qui calcule la fonction.
Supposons donc par exemple qu’un polynôme f (x1 , . . . , xn ) soit calculé par un programme d’évaluation Γ sans division de longueur s. On
peut numéroter xn+1 , . . . , xn+s les variables du programme. La variable
xn+1 représente un polynôme g(x1 , . . . , xn ) de l’un des 4 types suivants :
(1) xi + xj , (2) xi × xj , (3) c + xi , (4) c × xi
avec 1 ≤ i, j ≤ n et c une constante. On a aussi
f (x1 , . . . , xn ) = f1 (x1 , . . . , xn , g(x1 , . . . , xn ))
où le polynôme f1 (x1 , . . . , xn , xn+1 ) est calculé par le programme d’évaluation évident Γ1 (( extrait )) de Γ et de longueur s − 1.
Par hypothèse de récurrence f1 et les n + 1 dérivées partielles de f1
peuvent être calculées par un programme d’évaluation Γ01 de longueur
`1 ≤ 5(s − 1).
On considère alors les formules qui permettent de calculer les dérivées
partielles de f à partir de celles de f1 dans les 4 cas envisagés précédemment :
(1) ∂f /∂xh = ∂f1 /∂xh si h 6= i, j,
∂f /∂xi = ∂f1 /∂xi + ∂f1 /∂xn+1 ,
∂f /∂xj = ∂f1 /∂xj + ∂f1 /∂xn+1 .
(2) ∂f /∂xh = ∂f1 /∂xh si h 6= i, j,
∂f /∂xi = ∂f1 /∂xi + xj × ∂f1 /∂xn+1 ,
∂f /∂xj = ∂f1 /∂xj + xi × ∂f1 /∂xn+1 .
(3) ∂f /∂xh = ∂f1 /∂xh si h 6= i,
∂f /∂xi = ∂f1 /∂xi + ∂f1 /∂xn+1 .
(4) ∂f /∂xh = ∂f1 /∂xh si h 6= i,
∂f /∂xi = ∂f1 /∂xi + c × ∂f1 /∂xn+1 .
C’est le deuxième cas qui consomme le plus d’instructions nouvelles :
4 en tout (2 instructions pour calculer ∂f /∂xi et 2 pour ∂f /∂xj ). Il
faut par ailleurs rajouter l’instruction qui permet de calculer xn+1 en
fonction des xi précédents.
Ceci nous permet donc de construire à partir de Γ01 un programme d’évaluation Γ0 pour calculer f et ses n dérivées partielles. Ce
programme d’évaluation Γ0 a une longueur majorée par
`1 + 1 + 4 ≤ 5(s − 1) + 5 = 5s.
Par ailleurs l’initialisation de la récurrence est immédiate.
Le cas d’un programme d’évaluation avec divisions se traite de la
même manière et aboutit à la même majoration.
4. Notions de complexité
Introduction
Ce chapitre est consacré aux notions de complexité binaire d’une
part, directement issue de la modélisation du travail des ordinateurs,
et de complexité arithmétique d’autre part, en relation avec le nombre
d’opérations arithmétiques exécutées par un algorithme.
Les deux premières sections sont consacrées à la complexité binaire
et constituent une présentation rapide en guise de (( rappels )).
Les trois dernières sections décrivent de manière précise la complexité arithmétique des familles de circuits arithmétiques, elles servent donc
de base de travail pour les calculs de complexité développés dans tout
le reste de l’ouvrage.
Dans la section 4.3 nous introduisons les classes importantes de complexité arithmétique SD(f (n), g(n)). Nous discutons le rapport entre
complexité arithmétique (le nombre d’opérations arithmétiques exécutées) et complexité binaire (le temps d’exécution effectivement utilisé
lorsqu’on travaille avec des entrées représentant les éléments de l’anneau
A convenablement codés).
Ceci nous conduit à la notion de famille uniforme de circuits arithmétiques et aux classes N C k qui sont discutées dans la section 4.4.
Enfin dans la section 4.5 nous discutons brièvement un modèle de
machine parallèle (les PRAMs) correspondant aux circuits arithmétiques
et assez proche de la pratique des architectures parallèles.
4.1
Machines de Turing et Machines à Accès
Direct
Nous donnons ici quelques indications succinctes sur les modèles de
calcul algorithmique dans lesquels est prise en compte la taille des objets
130
4. Notions de complexité
à manipuler. Par exemple le temps utilisé pour additionner deux entiers
écrits en base 10 est manifestement du même ordre de grandeur que la
place occupée par l’écriture de ces deux entiers, tandis que l’algorithme
usuel pour la multiplication de deux entiers de tailles k et ` utilise un
temps du même ordre de grandeur que k × `.
Lorsque dans les années 30 des mathématiciens et logiciens ont réfléchi à la manière de décrire en termes précis ce qu’est un calcul algorithmique, ils ont abouti à des résultats assez variés quant à la forme,
mais identiques quant au fond. Tous les modèles élaborés ont abouti à
la même notion de (( fonction calculable de N vers N )).
La Machine de Turing abstraite
Cependant, c’est Alan Turing qui a emporté la conviction par la
simplicité de son modèle et par son caractère vraiment mécanique. Il est
parti de l’idée qu’un calcul doit pouvoir être exécuté par une machine
idéale qui, à l’instar d’un calculateur humain, dispose d’une feuille de
papier et d’un crayon, et procède selon une suite d’opérations élémentaires bien répertoriées une fois pour toutes, exécutées conformément
à un plan de travail détaillé ne laissant place à aucune ambigüité. Ce
modèle est basé sur la notion d’opération élémentaire. Une telle opération doit être suffisamment simple pour ne consommer qu’une quantité
fixe de temps et d’énergie. On imagine donc que la machine dispose
d’un alphabet fini fixé une fois pour toutes, et qu’une opération élémentaire consiste à lire, écrire ou effacer une lettre à un endroit précis
(la feuille de papier doit être divisée en cases, par exemple on prend
du papier quadrillé), ou encore à se déplacer vers une case voisine sur
la feuille de papier. Naturellement on n’autorise qu’un nombre fini de
lettres distinctes. Dans le premier modèle, Turing utilise une feuille de
papier constituée d’une simple succession de cases sur une seule ligne
potentiellement infinie : la bande de la machine de Turing. Par la suite,
il a semblé plus naturel d’utiliser pour modèle une Machine de Turing qui
utilise plusieurs bandes pour son travail. Quant au crayon (muni d’une
gomme), il est représenté par ce qu’il est convenu d’appeler une tête de
lecture 1 qui se déplace le long de la bande. Il y a une tête de lecture
pour chacune des bandes. Au départ, certaines bandes doivent contenir
l’entrée de l’algorithme (convenablement codée), tandis que les autres
sont entièrement vides. Lorsque la machine s’arrête, on lit le résultat à
1. Il serait plus correct mais plus lourd de parler d’une tête de lecture/effaçage/écriture.
4.1. Machines de Turing et Machines à Accès Direct
131
un endroit convenu. Une tête de lecture est capable de reconnaı̂tre si
la case lue est vide, d’y écrire alors une lettre, si elle n’est pas vide de
lire la lettre qui s’y trouve et éventuellement de l’effacer. Pour plus de
détails nous renvoyons à l’ouvrage [Tur] où sont traduits et commentés
les articles originaux de Turing, ainsi qu’aux ouvrages [Ste] et [BDG].
Le caractère très élémentaire du fonctionnement abstrait de la Machine de Turing en a fait un candidat naturel, non seulement pour les
questions de calculabilité théorique, mais également pour les questions
de complexité, et en particulier pour la question de l’appréciation du
temps et de l’espace nécessaires à l’exécution d’un algorithme. Une fois
l’algorithme traduit dans le modèle de la Machine de Turing, le temps
d’exécution est simplement mesuré par le nombre d’opérations élémentaires qui sont effectuées avant d’aboutir à l’arrêt. L’espace nécessaire à
l’exécution est représenté par le nombre de cases réellement utilisées.
Programmes élémentaires
On peut donner un modèle équivalent à la machine de Turing en
termes de programmes exécutables, sans doute plus parlant pour quiconque a déjà écrit un programme informatique. On considère des programmes de nature très simple. Ils sont écrits en utilisant des variables
entières N1 , . . . , Nr (les entiers sont supposés écrits en binaire) ou booléennes B1 , . . . , Bs (∈ {0, 1}). Un (( programme élémentaire )) est une
suite finie d’instructions numérotées de l’un des types suivants :
(A) Affectations
(1) Bj ← Ni mod 2
(2) Ni ← Ni div 2
(3) Ni ← 2Ni + Bj
(4) Bj ← 0
(5) Bj ← 1
(B) Branchements
(1) Direct : aller à l’instruction no . . .
(2) Conditionnel booléen : si Bj = 0 aller à l’instruction no . . .
(3) Conditionnel entier : si Ni = 0 aller à l’instruction no . . .
(S) Arrêt.
Les variables sont toutes initialisées à 0 sauf celles qui représentent les
entrées du programme.
Puisque les entiers sont écrits en binaire, on voit que chaque affectation ou branchement peut correspondre à un travail réalisé en consommant un temps et une énergie indépendantes de l’état des variables.
132
4. Notions de complexité
Le temps d’exécution est donc raisonnablement estimé comme étant le
nombre d’instructions exécutées avant d’aboutir à l’arrêt.
Machines à Accès Direct
Comme tout modèle abstrait, la machine de Turing est une idéalisation. Le point le plus contestable est l’hypothèse implicite selon laquelle
une opération élémentaire est équivalente à une autre quel que soit l’état
de la bande (dans la version Machine) ou des variables (dans la version
programme informatique élémentaire). Une telle conception se heurte à
des limitations physiques. Elle n’est en tout cas pas conforme à ce qui
se passe concrètement dans les ordinateurs actuels.
Alan Turing participa à l’aventure des premiers ordinateurs. Les ordinateurs ont une conception globale qui diffère sensiblement de la Machine de Turing abstraite. Les données ne sont pas traitées (( là où elles
sont )), comme dans l’image du crayon qui se déplace sur la feuille de
papier, mais elles sont transférées depuis la périphérie (un disque dur
par exemple) vers le centre où elles sont traitées, c’est-à-dire vers un microprocesseur, avant d’être renvoyées vers la périphérie. Ces transferts
permanents prennent d’autant plus de temps que les données sont plus
éloignées et que l’espace nécessaire à leur stockage est plus grand.
Ceci a donné lieu à un autre modèle de calcul, le modèle MAD des
Machines à Accès Direct (RAM en version anglaise abrégée), avec de
nombreuses variantes. Dans un modèle MAD, on doit considérer une infinité potentielle de (( registres )) (correspondant au stockage des données
en mémoire, ou aux cases d’une bande de Machine de Turing). Il serait
logique (mais ce n’est pas en général l’option choisie), de considérer que
chaque registre ne contient qu’une information dont la taille est fixée
une fois pour toutes. Pour traiter le registre dont l’adresse est l’entier n,
on considère que l’opération de transfert vers l’unité centrale requiert un
temps égal à la taille en binaire de l’entier n. Dans le modèle de Turing,
le temps correspondant peut être nul mais aussi beaucoup plus grand
que log n, selon la position des têtes de lecture sur chaque bande.
En fin de compte, selon l’algorithme utilisé (et selon le modèle MAD
choisi), les temps d’exécution T et T 0 obtenus dans le modèle MT (Machine de Turing à plusieurs bandes) ou dans les modèles MAD pour une
entrée de taille t sont soumis à des majorations respectives du type
suivant (voir un exemple précis dans [Ste] chapitre 2, sections 5.5 et
5.6) :
a T 0 ≤ T ≤ b T 02 ), c T ≤ T 0 ≤ d T (T + t)2 .
4.1. Machines de Turing et Machines à Accès Direct
133
Signalons le terme d’accumulateur qui dans le modèle MAD désigne
le microprocesseur.
L’espace de travail proprement dit
Nous terminons cette section avec un commentaire et une définition
plus précise de (( l’espace de travail )) utilisé dans les modèles MT ou
MAD. Dans le modèle MT nous avons défini (( l’espace nécessaire ))
comme le nombre total de cases effectivement utilisées au cours de
l’exécution de l’algorithme. En fait, si on veut étudier l’espace de travail
proprement dit utilisé par un algorithme, il est judicieux d’opérer une
distinction entre l’espace nécessaire aux données d’entrée-sortie d’une
part, et l’espace nécessaire au travail proprement dit d’autre part. On
convient dans ce cas que les bandes contenant les entrées sont utilisées
en lecture uniquement et qu’elles sont lues en une seule passe. De même,
les bandes contenant les sorties sont utilisées en écriture uniquement, et
elles sont écrites en une seule passe.
Par exemple lorsqu’on veut faire la preuve par 9 pour un produit
a × b = c où a, b et c sont considérées comme des entrées écrites en
base 10, il suffit de lire en une seule passe les données et aucun stockage
des résultats intermédiaires n’est nécessaire. On donne à la fin le résultat
(oui, ou non) sans avoir utilisé aucun espace pour le travail proprement
dit 2 . Si on écrivait cela sous forme d’un programme informatique élémentaire du type que nous avons décrit ci-dessus, cela signifierait que les
variables de travail sont toutes booléennes, que les variables représentant
les entrées sont seulement utilisées en lecture (elles ne peuvent être utilisées que via les affectations A1 et A2) et les variables représentant la
sortie sont seulement utilisées en écriture (elles ne peuvent être utilisées
que via les affectations A3).
Ainsi certains algorithmes utilisent un espace de travail nul (dans le
cas optimal) ou nettement inférieur à la taille des entrées-sorties. Pour
les études de complexité d’algorithmes on est particulièrement intéressé
par ceux qui n’utilisent aucun espace de travail d’une part, par ceux qui
utilisent un espace de travail linéaire par rapport à la taille de l’entrée
2. De même si on veut additionner deux entiers il suffit de les lire en une seule
passe et d’écrire au fur et à mesure le résultat sur la bande de sortie. Cependant
les entrées et la sortie ne sont pas écrites dans le même sens. En effet, pour pouvoir
enchaı̂ner des algorithmes, la convention naturelle est que la tête de lecture sur chaque
entrée doit être au départ à l’extrémité droite de l’entrée, et la tête de lecture sur
chaque sortie doit être à la fin à l’extrémité droite de la sortie
134
4. Notions de complexité
d’autre part et enfin par ceux qui utilisent un espace de travail de l’ordre
de grandeur de C log(n) où C est une constante et n est la taille de
l’entrée. On appelle ces derniers des algorithmes LOGSPACE .
4.2
4.2.1
Complexité binaire, les classes P, N P et #P
Calculs faisables
Malgré la grande abondance des modèles de calcul proposés, un
consensus a fini par s’établir sur ce qu’est un calcul faisable. On dit qu’un
calcul est faisable, ou encore qu’il est dans la classe P si on connaı̂t un
algorithme qui dans les modèles MT ou MAD nécessite un temps polynomial par rapport à la taille de l’entrée. Plus précisément, on ne dit
rien concernant tel calcul isolé (celui des 100.000 premières décimales de
π par exemple), mais on dit quelque chose concernant un calcul général
correspondant à des entrées de tailles variables et en tout cas arbitrairement grandes (celui de la k - ème décimale de π par exemple). On
demande que, pour un certain polynôme à coefficients positifs ou nul P ,
pour toute entrée de taille inférieure ou égale à n, l’algorithme donne sa
réponse en un temps majoré par P (n). Les algorithmes LOGSPACE sont
dans la classe P, et ils sont considérés à juste titre comme bien meilleurs
que les algorithmes qui travailleraient en temps et espace polynomial.
On voit que la notion d’algorithme de classe P est une notion asymptotique, qui peut être assez éloignée de la réalité des calculs. Un algorithme ayant un temps de calcul (( linéaire )) égal à n + 10100 correspond
en pratique à quelque chose d’infaisable, tandis que si son temps de
100
calcul est (( exponentiel )) majoré par sup(n, 2n /210 ) il reste facile à
exécuter pour toutes les entrées concrètement envisageables, alors même
qu’il n’est pas dans la classe P.
De nombreux auteurs distinguent les problèmes faisables (les entrées
sont des entiers, ou codées par des entiers, mais la sortie est du type
oui/non, donc codée par un booléen dans {0, 1}) des fonctions faisables
(la ou les sorties sont des entiers) et ils réservent le symbole P pour
les problèmes faisables. La classe des fonctions faisables (calculables en
temps polynomial) est alors notée FP. En fait une fonction f : N → N
est faisable si et seulement si d’une part la taille de la sortie est polynomialement majorée en fonction de la taille de l’entrée, et d’autre part,
le problème f (n) ≤ p ? est dans la classe P. Nous n’introduirons donc
pas deux notations distinctes et nous ferons confiance au contexte pour
lever les ambigüités éventuelles.
4.2. Complexité binaire, les classes P, N P et #P
135
Citons des problèmes de base qui ont reçu dans le passé une solution algorithmique satisfaisante, ce qui les mettait dans la classe P bien
avant qu’elle ne fût inventée. La résolution des systèmes linéaires par la
méthode chinoise du pivot, appelée en Occident méthode du pivot de
Gauss, donne un algorithme de classe P lorsque les coefficients et les
inconnues sont des nombres rationnels. Le calcul du nombre de racines
réelles d’un polynôme par la méthode de Sturm, qui avait été saluée
lors de sa découverte pour sa clarté et son élégance, fournit un algorithme en temps polynomial lorsque les coefficients du polynôme sont des
nombres rationnels. Le calcul du polynôme caractéristique d’une matrice carrée par la méthode de Leverrier est un autre exemple célèbre. Le
calcul intégral lui-même a un aspect algorithmique (pour le calcul automatique de certaines aires par exemple) qui frappa les contemporains
de Leibniz et Newton et qui est devenu aujourd’hui une des branches du
calcul formel.
4.2.2
Problèmes dont les solutions sont faciles à tester
La conjecture P =
6 N P est apparue dans les années 70 (Cook, [21]).
Elle correspond à l’idée intuitive suivante : il y a des problèmes dont les
solutions sont faciles à tester mais qui sont difficiles à résoudre. On pourrait dire a priori que la plupart des systèmes d’équations qu’on cherche
à résoudre correspondent à ce paradigme. Il est remarquable que cette
idée intuitive n’ait pu recevoir une forme mathématique précise qu’avec
l’avènement de la théorie de la complexité des algorithmes. Tard venue
dans le monde des conjectures mathématiques, la conjecture P 6= N P
apparaı̂t aujourd’hui comme l’une des plus importantes, l’une dont la
signification est la plus profonde. Elle a résisté à toutes les tentatives
d’en venir à bout, et beaucoup d’experts pensent qu’on ne dispose pas
aujourd’hui des concepts nécessaires à sa solution, alors même qu’elle a
quasiment la force d’une évidence. Nous allons en donner quelques commentaires relativement informels. Ils sont nécessaires pour aborder dans
les chapitres 12 et 13 l’analogue en complexité algébrique de la conjecture P =
6 N P en complexité binaire. Nous recommandons là encore sur
ce sujet les ouvrages [BDG] et [Ste].
Comme exemple de problème dont les solutions sont faciles à tester
mais qui sont difficiles à résoudre, nous allons considérer les problèmes
de programmation linéaire. Un tel problème est donné par une matrice
A (de type n × m) et un vecteur colonne b (de type m × 1) à coefficients
réels, et une solution du problème est un vecteur colonne x (de type
136
4. Notions de complexité
n × 1) tel que le vecteur y = Ax − b ait toutes ses coordonnées ≥ 0 ( 3 ).
Pour en faire un problème dont la nature algorithmique est bien précise,
nous nous limitons aux matrices A0 = [A|b] à coefficients entiers codés
en binaire. Quant aux solutions, nous avons le choix. Si nous demandons
des solutions en nombres rationnels, nous parlons de programmation linéaire en rationnels, et si nous demandons des solutions en nombres
entiers, nous parlons de programmation linéaire en entiers. Pour chacun
de ces deux problèmes une solution x éventuelle est facile à tester.
Un algorithme qui donne en général une solution rapide (s’il en existe
une) pour la programmation linéaire en rationnels a été mis au point
dans les années 50 (cf. [23]). Il est en général très performant et il est
encore aujourd’hui fréquemment utilisé, c’est l’agorithme de Dantzig.
L’inconvénient est que pour certaines matrices A0 , l’algorithme a un
mauvais comportement et son temps de calcul peut devenir exponentiel
par rapport à la taille de A0 . Dans les années 70 on a trouvé d’autres
algorithmes, qui dans la plupart des cas sont nettement plus lents que
celui de Dantzig, mais qui tournent en temps polynomial pour n’importe
quelles matrices A0 (cf. [55, 60, 61] et l’ouvrage [Sch]). Depuis, on sait
donc que la programmation linéaire en rationnels est dans la classe P.
Par contre pour ce qui concerne la programmation linéaire en entiers, on
n’est toujours pas capable de résoudre ce problème par un algorithme de
la classe P, même si on ne s’intéresse qu’aux solutions de petite taille.
En fait, on pense qu’on en sera à tout jamais incapable, car une réponse
dans l’autre sens signifierait que la conjecture P =
6 N P est fausse.
Pour expliquer comment est définie la classe N P, nous essayons
d’examiner avec un peu de recul ce que signifierait en général (( savoir
résoudre un problème dont on sait tester facilement les solutions )). Nous
commençons par remarquer que pour bien poser la question, il faut savoir
donner le problème sous une forme codée, qui puisse être prise comme
entrée d’un programme informatique (ou d’une Machine de Turing). On
peut donc toujours considérer que l’on a une suite infinie de problèmes
Pn où n est justement la forme codée en binaire du problème (les entiers
n qui ne coderaient pas correctement une instance de notre problème
doivent pouvoir être faciles à repérer). Quant aux solutions, elles doivent
également pouvoir être codées, données comme telles en entrée ou à la
sortie d’un programme informatique. Nous supposons donc sans perte
3. Un problème de programmation linéaire est en général énoncé sous forme d’un
problème d’optimisation. Nous en présentons ici une version équivalente plus facile à
discuter pour notre propos actuel.
4.2. Complexité binaire, les classes P, N P et #P
137
de généralité que la solution éventuelle est elle aussi codée par un entier x. Maintenant considérons la fonction ϕ : N × N → {0, 1} qui est
définie comme suit : ϕ(n, x) = 1 si x est le code d’une solution du problème Pn et ϕ(n, x) = 0 sinon. Supposer qu’on sait tester facilement
les solutions de notre famille Pn peut être raisonnablement interprété
comme signifiant que la fonction ϕ est dans la classe P. Tandis que
supposer que le problème est intrinsèquement difficile à résoudre peut
être raisonnablement interprété comme signifiant que la question
∃x ∈ N
ϕ(n, x) = 1 ?
n’a pas de réponse dans la classe P. Maintenant, nous devons apporter
une restriction. Il se peut que le problème soit intrinsèquement difficile à résoudre pour une trop bonne raison, à savoir que les solutions
éventuelles sont de taille trop grande. Plus exactement que la taille de
toute solution x du problème no n croisse trop vite par rapport à celle
de n. Nous notons dans la suite de cette section |x| la taille de l’entier naturel x, c’est-à-dire la longueur de son écriture en binaire. Nous
pouvons maintenant énoncer ce qu’est un problème dans la classe N P.
C’est répondre à une question du type suivant :
Existe-t-il une solution x(n) de taille raisonnable pour telle famille Pn
de problèmes dont les solutions sont faciles à tester ?
Plus précisément une famille de problèmes codée dans N est dite dans
la classe N P si sa solution revient à résoudre une question du type
∃x ∈ N
(|x| ≤ a + |n|k et ϕ(n, x) = 1)
?
(4.1)
où a et k sont deux entiers positifs donnés et où ϕ : N × N → {0, 1}
est dans la classe P. Autrement dit si on pose
n
o
ψ(n) = sup ϕ(n, x) ; |x| ≤ a + |n|k
(4.2)
et si la fonction ϕ est dans la classe P, alors la fonction ψ est dans
la classe N P. On peut d’ailleurs supposer sans perte de généralité que
ϕ(n, x) = 0 si |x| > a + |n|k .
Le N de N P est mis pour non déterministe. La raison en est la
suivante. La fonction ψ ci-dessus pourrait être calculée en temps polynomial par une machine dont le fontionnement serait (( non déterministe )).
Plus précisément, en utilisant nos programmes informatiques élémentaires ci-dessus, on admettrait des intructions de branchement non déterministe : aller à l’instruction no . . . ou . . . (selon l’humeur du moment).
138
4. Notions de complexité
Le programme peut alors aboutir à plusieurs résultats différents selon
le chemin choisi lors de son exécution. Et ce programme serait réputé
calculer (pour une entrée x fixée) la plus grande des valeurs qu’il peut
délivrer en sortie. L’acronyme N P vaut alors pour : calculable en temps
polynomial par une machine à fonctionnement non déterministe.
Notez que si on avait P = N P (ce que personne ne croit), on pourrait
non seulement calculer, dans l’exemple ci-dessus, la fonction ψ en temps
polynomial, mais également, dans le cas d’une réponse positive ψ(n) =
1, trouver une solution x pour ϕ(n, x) = 1 en temps polynomial. En
effet, on pourrait calculer un tel x par dichotomie en temps polynomial
en posant un nombre polynomial de fois la question
∃x x ≤ p et ϕ(n, x) = 1 ?
qui serait résoluble en temps polynomial sur les entrées n, p (on démark
rerait avec p = 2a+|n| ).
Certains problèmes qui peuvent sembler a priori être dans la classe
N P sont ramenés dans la classe P lorsque quelqu’un découvre un algorithme rapide pour les résoudre. Des succès spectaculaires ont été à la
fin du 20ème siècle la solution en temps polynomial des systèmes d’équations linéaires à coefficients et inconnues entières, celle des problèmes
de programmation linéaire en rationnels et la détermination de (( petits
vecteurs )) dans un réseau (qui conduit notamment à la factorisation en
temps polynomial des polynômes sur Q[X]).
Cook a montré (cf. [21]) que certains problèmes de la classe N P sont
universels : si on démontre pour l’un d’entre eux qu’il est dans la classe
P, alors P = N P. Un tel problème est dit N P - complet. Par exemple la
programmation linéaire en entiers est un problème N P - complet, même
si on limite a priori la taille des solutions par un entier fixe.
Nous pouvons expliquer informellement pourquoi il existe des problèmes N P - complets.
Un ordinateur qui ne serait soumis à aucune limitation physique
de temps et d’espace serait une machine universelle en ce sens qu’il
est capable d’exécuter n’importe quel programme qu’on lui soumet (en
faisant abstraction des limitations physiques). Un des premiers théorèmes d’Alan Turing était l’existence d’une Machine de Turing universelle. Une conséquence importante de l’existence d’une Machine de Turing
universelle est, via le processus diagonal de Cantor, l’existence de problèmes bien posés (pour les Machines de Turing) mais qui ne pourront
être résolus par aucun procédé mécanique du type Machine de Turing :
l’ensemble des (codes Turing de) fonctions mécaniquement calculables
4.2. Complexité binaire, les classes P, N P et #P
139
de N vers N (au sens des Machines de Turing) n’est pas mécaniquement
calculable (au sens des Machines de Turing). L’existence de problèmes
N P - complets est un résultat de nature similaire.
Introduisons la notation hx1 , . . . , xk i pour un code dans N d’un
k-uple d’entiers 4 . En termes de programmes informatiques élémentaires, l’existence d’une Machine de Turing universelle signifie qu’on sait
écrire un programme élémentaire (( universel )) en ce sens qu’il remplit
le contrat suivant :
– Il prend en entrée 2 entiers binaires n, x et un entier bâton t
( 5 ), où n est un texte de programme élémentaire Qn codé en
binaire, x est un code pour la liste des entrées pour Qn et t est
le nombre d’étapes élémentaires pendant lequel on désire que soit
exécuté Qn .
– Il donne en sortie une description instantanée de (c’est-à-dire un
codage binaire U (n, x, t) qui décrit de manière exacte) l’état où se
trouve la machine qui exécute le programme Qn après l’exécution
de t étapes élémentaires de calcul sur l’entrée x : la valeur de
chacune des variables xi du programme d’une part, le numéro h
de l’instruction en cours d’autre part (codés par hx1 , . . . , x` , hi).
Si le temps d’exécution est t0 on demande que pour t > t0 on ait
U (n, x, t) = U (n, x, t0 ). Nous supposons aussi sans perte de généralité
que les variables de sortie sont en écriture seulement, c’est-à-dire ne sont
utilisées que via les affectations de type A3.
Il n’est pas très difficile de vérifier qu’un programme élémentaire universel écrit de manière naturelle calcule la fonction universelle U en
temps polynomial.
Comme conséquence on obtient quelque chose qui pouraı̂t être compris comme une énumération dans la classe P de tous les programmes
dans la classe P s’exécutant sur une entrée de taille polynomialement
majorée. Expliquons nous.
Tout d’abord notons (n, x, t) 7→ V (n, x, t) la fonction (dans la classe
P) qui donne l’état de la variable en sortie (ou, s’il y a plusieurs sorties
4. On considère un codage naturel, de sorte que les fonctions de codage (x1 , . . . , xk ) 7→ hx1 , . . . , xk i et celles de décodage (hx1 , . . . , xk i 7→ xi et
hx1 , . . . , xk i 7→ k) sont dans la classe P. On suppose aussi sans perte de généralité que hx1 , . . . , xk i ≥ xi .
5. Un entier bâton sert de compteur, il est codé en binaire par 2t − 1, c’est-à-dire
(si t ≥ 1) le mot formé de t fois la lettre 1. Ici il est nécessaire de prendre pour t
un entier bâton parce qu’on veut que la fonction universelle soit calculable en temps
polynomial par rapport à la taille de ses entrées.
140
4. Notions de complexité
prévues, l’état de la première d’entre elles).
Soit maintenant ϕ : N × N → {0, 1} une fonction dans la classe P
qui vérifie
∀x |x| > a + |n|k ⇒ ϕ(n, x) = 0
(4.3)
Alors la fonction ψ(n) = sup {ϕ(n, x) ; x ∈ N} résoud un poblème dans
la classe N P.
Vu que ϕ est dans la classe P et vu le caractère universel de V il
existe un entier m0 et deux entiers b, ` tels que
ϕ(n, x) = V (m0 , hn, xi , t)
avec
t ≤ b + |n|`
si
|x| ≤ a + |n|k
Définissons par ailleurs (avec z, t des entiers bâtons, et n et x des
entiers binaires,)
Φ(p, x) =
inf(1, V (m, hn, xi , t)) si |x| ≤ z
0
sinon
avec
p = hm, n, z, ti .
C’est naturellement une fonction N × N → {0, 1} dans la classe P pour
laquelle on a
∀x |x| > |p| ⇒ Φ(p, x) = 0
et à partir de laquelle on peut définir
Ψ(p) = sup {Φ(p, x) ; |x| ≤ |p|}
(4.4)
qui est dans la classe N P. Maintenant il est clair que si on pose
D
E
λ(n) = m0 , n, a + |n|k , b + |n|`
alors la fonction λ est dans la classe P et
ψ(n) = Ψ(λ(n)).
Ceci montre le caractère universel de la fonction Ψ au sens où nous
le souhaitions. En écrivant cette preuve en détail, on peut donner des
précisions supplémentaires sur la manière dont le problème N P associé
à ϕ à été réduit en temps polynomial à celui associé à Φ. En particulier on peut fabriquer une variante où la réduction est dans la classe
LOGSPACE .
4.2. Complexité binaire, les classes P, N P et #P
4.2.3
141
Problèmes de comptage
Si E est un ensemble fini, nous noterons #E le nombre d’éléments
de E.
Lorsqu’on a une famille de problèmes dont les solutions sont faciles
à tester et de taille polynomialement majorées, on peut se poser non
seulement la question de savoir si une solution existe, mais également
combien de solutions existent.
Précisément si ϕ : N × N → {0, 1} est une fonction dans la classe P
qui vérifie
∀x |x| > a + |n|k ⇒ ϕ(n, x) = 0
alors la fonction
θ(n) = # {x | x ∈ N, ϕ(n, x) = 1 } =
X
ϕ(n, x)
(4.5)
|x|≤a+|n|k
compte le nombre de solutions (pour la question codée par n). A priori
cette fonction est plus difficile à calculer que la fonction définie par
l’équation (4.2) ψ(n) = sup {ϕ(n, x) ; x ∈ N} (qui est dans la classe
N P). La taille de θ(n) est polynomialement majorée en fonction de
celle de n. Les fonctions θ obtenues de cette manière définissent une
nouvelle classe de complexité, les fonctions de comptage pour les problèmes dont les solutions sont faciles à tester, que l’on note #P (prononcer
dièse P). Cette classe a été introduite par Valiant dans [93].
Si on veut que la classe #P soit une classe de problèmes plutôt
qu’une classe de fonctions, on la définit comme la classe des problèmes
k
du type θ(n) ≤ p ?. En effet, puisque θ(n) ≤ 2a+|n| il est facile de
calculer par dichotomie, en temps polynomial, la fonction θ à partir des
tests θ(n) ≤ p ?.
On conjecture que les deux inclusions
P ⊂ N P ⊂ #P
sont strictes.
De même qu’il existe des problèmes N P - complets, il existe des fonctions #P-complètes. En fait la réduction que nous avons esquissée dans
le cas N P ci-dessus fonctionne aussi pour les fonctions de comptage.
Définissons en effet
X
Θ(p) = # {x | x ∈ N, Φ(p, x) = 1 } =
Φ(p, x)
(4.6)
|x|≤|p|
alors, avec la même fonction λ que ci-dessus, on obtient θ(n) = Θ(λ(n)).
142
4.3
4.3.1
4. Notions de complexité
Complexité arithmétique et complexité binaire des circuits
Complexité arithmétique
La taille (en fait le nombre d’opérations arithmétiques) et la profondeur d’un circuit arithmétique ou d’un programme d’évaluation sont les
deux paramètres qui mesurent ce qu’on appelle la complexité arithmétique de ce circuit arithmétique ou de ce programme d’évaluation.
Ce sont des fonctions de ce que nous avons appelé les paramètres
d’entrée du circuit arithmétique. Comme on s’intéresse souvent à la
complexité asymptotique des algorithmes (c’est-à-dire à leur comportement quand ces paramètres tendent vers l’infini), nous allons utiliser
les notations classiques O, o, Θ, Ω définies de la manière suivante :
Notation 4.3.1 Étant données deux fontions f et g de N∗ dans R∗+ ,
on dit que :
• g ∈ O(f ) et l’on écrira g(n) = O(f (n)) s’il existe une constante
réelle c > 0 telle que ∀n, n ∈ N∗ ⇒ g(n) ≤ c f (n).
• g ∈ o(f ) et l’on écrira g(n) = o(f (n)) si pour tout réel ε > 0, il
existe k ∈ N∗ tel que ∀n, (n ∈ N∗ et n > k) ⇒ g(n) ≤ ε f (n).
• g ∈ Ω(f ) et l’on écrira g(n) = Ω(f (n)) si f (n) = O(g(n)).
• g ∈ Θ(f ) et l’on écrira g(n) = Θ(f (n)) si g(n) = O(f (n)) et
f (n) = O(g(n)). On dit dans ce cas que f est du même ordre que g.
Remarquons que pour montrer que g ∈ O(f ), il suffit de trouver une
constante réelle K0 et un entier n0 ∈ N∗ tels que g(n) ≤ K0 f (n) pour
tout n ≥ n0 . Nous appellerons une telle constante K0 une constante
asymptotique (cachée dans le grand O). Dans la suite chaque fois que
ce sera possible nous nous appliquerons à faire apparaı̂tre la constante
asymptotique cachée dans le grand O dans l’étude de complexité des
algorithmes. Et l’entier n0 sera parfois précisé.
Notation 4.3.2 (complexité arithmétique d’une famille de circuits)
On écrira qu’un algorithme est (dans la classe) SD(f (n), g(n)) pour
dire qu’il correspond à une famille de circuits arithmétiques de taille
t(n) = O(f (n)) et de profondeur p(n) = O(g(n)).
Par exemple, l’algorithme simplifié du pivot de Gauss, tel qu’il a été
développé dans la section 2.1, est SD(n3 , n).
4.3. Complexités arithmétique et binaire
143
Un algorithme est dit optimal lorsqu’il n’y a pas d’algorithme asymptotiquement plus performant, du point de vue de la taille.
Il y a des problèmes dont on connaı̂t la complexité séquentielle, c’està-dire l’ordre asymptotique exact du nombre d’opérations arithmétiques
nécessaires pour le résoudre, comme par exemple le problème de l’évaluation d’un polynôme à une indéterminée sur un anneau commutatif
quelconque 6 . D’autres problèmes, par contre, comme celui de la multiplication des matrices, sont des problèmes dont on ignore la complexité exacte à cause de l’écart entre les bornes inférieure et supérieure
asymptotiques que l’on connaı̂t 7 .
Il faut remarquer que le grand O de la notation introduite ci-dessus
présente l’inconvénient majeur de (( cacher )) la constante asymptotique qui permet de le définir. Elle a pourtant une importance pratique
considérable puisque deux algorithmes permettant par exemple de résoudre
respectivement le même problème avec 100n3 et 109 n2 opérations arithmétiques sont tels que le second a une complexité asymptotique nettement meilleure que le premier (il peut arriver qu’il soit aussi optimal)
alors que le second, asymptotiquement moins performant, reste plus rapide tant que le nombre d’opérations à effectuer n’a pas atteint la borne
astronomique de 1023 .
4.3.2
Complexité binaire
On raconte que l’inventeur du jeu d’échec demanda comme récompense un grain de blé sur la première case, deux sur la deuxième, quatre
sur la troisième et ainsi de suite jusqu’à la soixante-quatrième. Cela fait
a priori un circuit arithmétique de profondeur 64. Mais pour calculer
6
264 − 1 = 22 − 1 = 18.446.744.073.709.551.615 un circuit arithmétique
de taille (et de profondeur) 6+1 suffit :
Début
v0 := x
(porte d’entrée, on évaluera avec x = 2)
v1 := v0 × v0
v2 := v1 × v1
v3 := v2 × v2
v4 := v3 × v3
v5 := v4 × v4
6
v6 := v5 × v5 (v6 = 22 )
6. L’algorithme de Horner est optimal pour ce problème, cf. page 11.
7. Le problème de la multiplication des matrices est Ω(n2 ) et O(n2,376 ).
144
4. Notions de complexité
v7 := v6 − 1
Fin.
De même, un circuit arithmétique de taille 20 évalué sur l’entrée 2
20
permet de calculer 22 = 21.048.576 = 6, 7411 . . . 10315.652 . Ceci montre
clairement qu’il y a une différence considérable entre la taille d’un circuit
et celle des objets qu’il peut produire lorsqu’on l’évalue sur des entiers
codés en binaire.
La complexité binaire d’un circuit (ou d’une famille de circuits) est
par définition la complexité du calcul d’évaluation qu’il produit lorsqu’on
prend ses entrées dans un anneau fixé avec un codage fixé. L’exemple le
plus simple et le plus important est l’anneau des entiers codés en binaire.
Naturellement, si on accepte de coder un entier par un circuit arithmétique sans division ayant pour seules entrées des constantes déterminées a priori (−1, 0, 1, 2 par exemple) et si on note Zpreval l’anneau des
entiers ainsi codé, on voit que l’évaluation d’un circuit arithmétique sans
division dans Zpreval est en temps linéaire (il suffit de mettre les circuits
bout à bout en changeant seulement certaines profondeurs et certains
identificateurs). Le problème avec Zpreval est alors reporté du côté du
test de signe, de la division euclidienne, ou de l’évaluation des circuits
avec divisions exactes.
Il est donc crucial de préciser à la fois l’anneau et le codage choisi
pour cet anneau lorsqu’on veut parler de la complexité binaire d’un circuit arithmétique.
Signalons à ce sujet qu’en géométrie algébrique, la notion usuelle de
degré d’un polynôme peut être souvent remplacée avantageusement par
la notion de profondeur d’un programme d’évaluation arithmétique qui
lui correspond. Il s’agit là d’un sujet de recherche actif et prometteur
(cf. [39, 40]).
Un exemple : complexité binaire de l’algorithme du pivot de
Gauss
Elle est mesurée par le nombre d’opérations booléennes nécessaires
pour exécuter l’algorithme avec des entrées codées sous forme de suites
de bits. Cette complexité dépend de manière importante du corps K et
du codage choisi pour les éléments de K.
Si le corps K est un corps fini, la complexité binaire est proportionnelle à la complexité arithmétique. C’est (( le bon cas )) pour l’algorithme.
Appliqué dans le cadre de calculs numériques (ce qui constitue aujourd’hui une partie importante du travail des ordinateurs), l’algorithme
4.3. Complexités arithmétique et binaire
145
est en général exécuté avec des nombres en virgule flottante, codés par
des suites de bits de longueur fixe, et la complexité binaire est de nouveau proportionnelle à la complexité arithmétique. Mais naturellement,
on ne travaille pas vraiment avec les éléments du corps des réels. D’où la
nécessité de garantir les résultats avec une précision demandée. L’analyse
numérique matricielle remplit des rayons entiers de bibliothèques.
Dans cet ouvrage, nous ne prenons en compte que les calculs exacts
(en précision infinie dit-on parfois), et nous ne ferons guère d’autre allusion aux aspects proprement numériques des algorithmes que nous
commenterons (voir cependant page 148).
La méthode du pivot de Gauss appliquée dans le corps des rationnels
réserve quelques désagréables surprises. Même si les entrées sont des
nombres entiers (supposés codés en binaire de la manière usuelle), on
doit immédiatement passer au corps des fractions. Un rationnel est alors
codé par un couple d’entiers, le numérateur avec un signe et le dénominateur strictement positif. Avec les rationnels ainsi codés (ce qui est
le codage binaire naturel), on est alors devant l’alternative suivante :
simplifier les nouvelles entrées de la matrice dès qu’elles sont calculées,
ou ne jamais simplifier. La deuxième solution est désastreuse, car les
fractions successives voient en général les tailles de leur numérateur et
dénominateur croı̂tre de manière exponentielle. La première solution,
quoique moins désastreuse, est néanmoins coûteuse, car elle implique
des calculs systématiques de pgcd. La dernière formule donnée dans la
[p]
propriété 2.1.3 permet d’exprimer a ij comme quotient de deux déterminants extraits de la matrice de départ (et elle se généralise au cas où
des permutations de lignes ou de colonnes sont effectuées). On a donc
la garantie que toutes les fractions qui sont calculées au cours de l’algorithme restent de taille raisonnable (O(n (t + log n)) si on part d’une
matrice n × n à coefficients entiers majorés par t en taille binaire, i.e.
majorés par 2t en valeur absolue). Le nombre d’opérations arithmétiques dans Z doit donc être multiplié par un facteur nt pour tenir
compte du calcul de simplification des fractions. La complexité binaire,
elle, a une majoration fort décevante en O(n5 t2 ) (à des facteurs logarithmiques près) si on utilise les algorithmes usuels pour la multiplication
ou la division de deux entiers.
Appliquée avec le corps des fractions de Z[X] ou Z[X, Y, Z] la méthode du pivot de Gauss se heurte au même type de difficultés, mais très
nettement aggravées, car les calculs de pgcd de polynômes, surtout en
plusieurs variables, sont très coûteux.
146
4. Notions de complexité
Situations dans lesquelles la complexité binaire d’un circuit est
en rapport étroit avec sa complexité arithmétique
Nous signalerons ici trois situations de ce type.
Le premier cas est celui d’une famille de circuits arithmétiques
évalués dans un anneau avec un codage pour lequel les opérations arithmétiques produisent des objets de taille bien contrôlée, du fait même de
la structure du circuit arithmétique.
Proposition 4.3.3 Considérons une famille de circuits arithmétiques
Γn de taille σn et de profondeur πn (n est un paramètre contrôlant le
nombre d’entrées du circuit Γn . Supposons en outre que la production
du circuit Γn réclame un temps τn . Soit enfin A un anneau donné
dans un codage pour lequel les opérations arithmétiques sont en temps
polynomial O(N k ) avec k > 1 et la taille t(x) des objets vérifie l’inégalité t(a ◦ b) ≤ t(a) + t(b).
Alors la production puis l’exécution de ce circuit réclame, dans le modèle
MAD, un temps majoré par τn +σn ·O((2πn N )k ) (N > n est la taille de
la liste des entrées). En particulier si σn = O(nh ), πn ≤ ` log n et τn =
O(nc ) (pour des constantes convenables h, ` et c) alors l’exécution de
l’algorithme correspondant à la famille Γn est (globalement) en temps
polynomial, précisément en O(nc + nh+`k N k ).
Preuve. Dans le modèle MAD, on peut utiliser un registre distinct pour
chacune des variables du programme d’évaluation. La taille de tous les
résultats intermédiaires est majorée par 2πn N puisqu’elle double au maximum quand la profondeur augmente d’une unité.
Les transferts entre les registres de travail et l’accumulateur représentent
un temps de l’ordre de σn · (2πn N + log(σn )) qui est négligeable devant
l’estimation du temps d’exécution des opérations arithmétiques proprement dites : σn · O((2πn N )k ).
t
u
Remarque 4.3.4 Dans le modèle des machines de Turing, on obtient
les mêmes majorations pour n fixé. Par contre, lorsque n varie, se pose
le problème de la gestion d’un nombre non fixé a priori de variables de
travail, alors qu’une telle machine n’a, quant à elle, qu’un nombre fixé
a priori de bandes de travail. Les transferts de données entre d’une part
la bande où est stockée la liste des (contenus des) variables de travail et
d’autre part les bandes où sont exécutées les opérations arithmétiques
prennent normalement un temps de l’ordre de (σn )2 (2πn N + log(σn ))
4.3. Complexités arithmétique et binaire
147
car la bande de stockage doit être relue pour chacune des σn opérations arithmétiques, et sa taille est seulement majorée par σn · (2πn N +
log(σn )). Il s’ensuit que la majoration en temps obtenue peut parfois
être un peu moins bonne que celle indiquée pour le modèle MAD.
De nombreuses variantes de la situation précédente peuvent être utilisées. Par exemple, pour l’évaluation dans Z, c’est seulement la profondeur multiplicative qui doit être en O(log n) pour qu’on ait un bon
contrôle de la taille des objets produits, et donc de l’ensemble du calcul
d’évaluation.
Un algorithme est dit bien parallélisé lorsqu’il correspond à une famille de circuits arithmétiques (Γn ) dont la taille σn est optimale et
dont la profondeur est en O(log` (σn )) (pour un certain exposant ` > 0).
Si la taille est polynomiale en n, la profondeur est alors polylogarithmique , c’est-à-dire en O(log` (n)). En fait, nous utilisons dans cet ouvrage
le terme bien parallélisé avec un sens un peu plus libéral pour le mot
optimal. Pour les algorithmes en temps polynomial nous demandons
seulement que, en ce qui concerne la taille, l’exposant du n ne soit pas
très loin de celui du meilleur algorithme séquentiel connu (la profondeur
étant, elle polylogarithmique). C’est en ce sens que nous considérons
que les algorithmes de Csanky, de Chistov ou de Berkowitz sont bien
parallélisés.
Le deuxième cas est celui d’une famille de circuits arithmétiques
dont la profondeur n’est pas nécessairement logarithmique et pour laquelle on a un argument de nature algébrique qui permet de mieux majorer la taille des objets intermédiaires que l’argument de profondeur.
C’est par exemple le cas de l’algorithme du pivot de Gauss simplifié
(éventuellement modifié par élimination des divisions à la Strassen) ou
de l’algorithme de Jordan-Bareiss. Même dans le cas d’un algorithme
bien parallélisé comme celui de Berkowitz, exécuté dans Z, les majorations de taille obtenues par un argument algébrique direct sont meilleures
que celles obtenues par l’argument de profondeur.
Signalons un calcul de majoration simple qui permet souvent un
contrôle satisfaisant de la taille des objets intermédiaires dans le cas de
l’évaluation dans un anneau du style Matn (Z[x, y]) ( 8 ) codé en représentation dense (voir la note 1 page 115), les entiers étant eux-mêmes
codés en binaire. Si A = (aij ) est une matrice dans cet anneau, on note
dA le degré maximum d’une entrée aij (x, y) et `A := log(Σijhk |aijhk |),
8. Ceci désigne l’anneau des matrices n × n à coefficients dans Z[x, y].
148
4. Notions de complexité
où aijhk est le coefficient de xh y k dans aij (x, y). On a alors la taille de
A qui est majorée par n2 d2A `A et les formules suivantes sont faciles à
vérifier :
`A±B ≤ 1 + max (`A , `B ) `AB ≤ `A + `B
dA±B ≤ 1 + max (dA , dB ) dAB ≤ dA + dB
Ceci signifie que ce type d’anneau se comporte comme Z pour tous les
calculs de majoration de taille des objets produits lors de l’évaluation
d’un circuit arithmétique. En particulier si la taille du circuit no n est polynomiale en n et si sa profondeur multiplicative est logarithmique, alors
la taille des objets est polynomialement majorée. La plupart des algorithmes que nous examinons dans cet ouvrage ont pour le type d’anneau
que nous venons de signaler, une majoration polynomiale de la taille des
objets intermédiaires. Signalons en revanche le mauvais comportement
de l’algorithme de Hessenberg pour la taille des objets intermédiaires.
Le troisième cas est celui d’une famille de circuits arithmétiques
(sans divisions) évalués dans un cadre de calcul numérique bien contrôlé.
Lors de l’évaluation du circuit, les entrées sont des nombres dyadiques
interprétés comme des nombres réels pris avec une précision fixée. Toutes
les portes du circuit sont elles-mêmes évaluées avec une précision fixée.
Un calcul de majoration d’erreur est nécessaire pour que le résultat du
calcul ait un sens mathématique précis. Ce calcul dit une chose du genre
suivant : sachant que vous désirez les sorties avec une précision absolue
p (c’est-à-dire de p digits après la virgule), et que les entrées sont prises
sur l’intervalle contrôlé par le paramètre n, alors vous devez évaluer le
circuit Γn,p en effectuant tous les calculs intermédiaires avec la précision
ε(n, p) (en particulier les entrées doivent être prises avec cette précision).
Par exemple, on pourra imaginer une famille de circuits arithmétiques
évaluant en ce sens la fonction (x2 +1)/ln(1+x) sur l’intervalle ] 0, ∞ [ :
le circuit arithmétique Γn,p doit permettre d’évaluer cette fonction sur
l’intervalle [2−n , 2n ] avec la précision p, en exécutant tous les calculs
avec une précision ε(n, p).
Si la famille peut être produite en temps polynomial, et si la précision
requise ε(n, p) peut être majorée par un polynôme en (n, p) alors la
fonction réelle ainsi calculée est dite calculable en temps polynomial (cf.
[KKo, 45, 59, 63]). Cela signifie que cette fonction peut être évaluée avec
la précision p sur n’importe quel réel dans l’intervalle contrôlé par n en
un temps qui dépend polynomialement de (n, p). Il s’agit donc d’analyse
numérique entièrement sûre et parfaitement contrôlée.
Ce type d’algorithmes est en phase d’être implémenté sur machine,
4.4. Familles uniformes de circuits
149
cela peut être considéré comme une des tâches importantes à réaliser
par le Calcul Formel.
4.4
Familles uniformes de circuits arithmétiques
et booléens
Les algorithmes de calcul algébrique usuels ont un nombre d’entrées
et de sorties qui dépend d’un ou plusieurs paramètres entiers, comme par
exemple (( la multiplication de deux matrices )) (3 paramètres pour fixer
les tailles des deux matrices) ou (( le produit d’une liste de matrices ))
(une liste d’entiers pour paramètres) ou (( le déterminant d’une matrice ))
(un paramètre). Nous avons appelé ces paramètres des paramètres d’entrée. Comme nous l’avons déjà dit, ce n’est pas seulement la taille et
la profondeur du circuit (en fonction des paramètres d’entrée) qui sont
importantes, mais aussi son coût de production. Pour calculer le déterminant d’une matrice à coefficients entiers dans la situation la plus
générale possible, par exemple, on doit d’abord produire le texte du programme d’évaluation correspondant au circuit qu’on envisage, et ensuite
exécuter ce programme d’évaluation sur la liste d’entrées voulue. Si le
circuit est de faible profondeur et de faible taille mais que le coût de
la production du programme d’évaluation correspondant croı̂t très vite
lorsque le paramètre d’entrée augmente, on ne peut guère être satisfait
du résultat.
C’est la raison pour laquelle on a introduit la notion de famille uniforme de circuits arithmétiques. On dit qu’une famille de circuits arithmétiques (indexée par les paramètres d’entrée) est uniforme lorsque le
coût de production du circuit (en tant que texte d’un programme d’évaluation) dépend (( de manière raisonnable )) des paramètres d’entrée. Une
première notion d’uniformité consiste à demander que le coût de production du circuit soit dans la classe P, c’est-à-dire en temps polynomial.
Une deuxième notion, plus forte, consiste à demander qu’on soit dans
la classe LOGSPACE c’est-à-dire que l’espace de travail nécessaire à la
production du circuit soit logarithmique.
Ces notions d’uniformité sont relativement satisfaisantes mais elles
nécessiteraient d’être mieux explicitées dans chaque cas concret. Il est
clair qu’une famille de circuits dépendant d’un paramètre d’entrée n qui
aurait une profondeur en log n, une taille en n2 et un coût de production en n2001 ne serait pas un très bon cru pour l’année 2001. Dans la
littérature sur le sujet règne un silence discret. En fait tout le monde
150
4. Notions de complexité
considère apparemment qu’il est bien clair que le coût de production du
circuit n’a en général pas un ordre de grandeur bien supérieur à sa taille.
Nous nous contenterons de confirmer cette impression par l’étude
d’un cas d’école, la multiplication rapide des matrices à la Strassen. Nous
renvoyons pour cette étude au chapitre 7 section 7.1.2 théorème 7.2.
Classes de complexité N C
Pour définir les notions de taille et profondeur en complexité arithmétique parallèle on a utilisé des familles de circuits arithmétiques sans
exiger que ces familles soient uniformes.
En complexité binaire, les entrées et les sorties d’un algorithme sont
des mots écrits sur un alphabet fixé, par exemple l’alphabet {0, 1} (ou si
on préfère des entiers écrits en binaire). Il est alors naturel d’utiliser les
familles de circuits booléens pour définir les notions de taille et profondeur d’un algorithme parallèle. Dans un circuit booléen, chaque entrée
est un élément de {0, 1}, et les portes sont de trois sortes : ∨, ∧ (avec
deux antécédents) ou ¬ (à un seul antécédent). Pour chaque longueur
de l’entrée d’un algorithme parallèle, codée comme une suite finie de
booléens, le circuit booléen correspondant doit calculer la sortie, codée
de la même manière. Mais sans uniformité de la famille, on aboutirait
à des contre-sens intuitifs évidents, puisque toute fonction f de N vers
{0, 1} telle que f (n) ne dépend que de la longueur de n est réalisable
par une famille non uniforme de circuits booléens de taille n + 1 et
de profondeur 0 (a vrai dire, l’entrée du circuit no n ne sert à rien, et
aucune opération booléenne n’est exécutée). Or une telle fonction peut
ne pas être calculable.
Pour un entier naturel k donné, on note N C k la classe de toutes les
fonctions qui peuvent être calculées par une famille uniforme de circuits
booléens dans SD(nh , logk n) où h est un entier positif (par hypothèse,
le circuit Cn a un nombre de portes d’entrée polynomialement relié à
n). L’uniformité est prise ici au sens le plus fort que nous avons considéré
au début de cette section. C’est la LOGSPACE uniformité, c’est-à-dire,
pour une famille de circuits (Cn )n∈N , l’existence d’une machine de Turing qui, pour l’entrée n, donne en sortie le codage du circuit Cn en
utilisant un espace mémoire en O(log n).
S
On pose N C = k∈N N C k . Il s’agit d’un acronyme pour Nick’s Class
du nom de Nicholas Peppinger qui a proposé cette classification des algorithmes parallèles.
4.5. Machines parallèles à accès direct
151
Alors N C ⊆ P mais l’inclusion dans l’autre sens (c’est-à-dire l’égalité des deux classes) est un problème ouvert, et il est conjecturé que
l’inclusion est stricte.
On peut définir des notions analogues en complexité arithmétique
([34, BCS]). Il serait alors théoriquement nécessaire de distinguer dans
les notations la classe N C au sens de la complexité arithmétique de
celle définie précédemment. En outre, en complexité arithmétique on
peut exiger ou ne pas exiger l’uniformité de la famille de circuits, et on
peut aussi vouloir indiquer sur quel anneau commutatif on travaille.
Dans le cadre de cet ouvrage, nous ne désirons pas multiplier les
notations et nous garderons la notation N C k pour parler des familles uniformes de circuits arithmétiques en SD(nh , logk n), (où n est la
somme des paramètres d’entrée du circuit et h est un entier positif).
Nous demandons en outre que le degré de tous les polynômes évalués aux
noeuds du circuit soit majoré par un polynôme en n. Enfin, nous prendrons l’uniformité en un sens plus modeste : la famille des circuits doit
seulement être construite en temps polynomial. La seule vraie preuve
d’uniformité que nous faisons est d’ailleurs celle du théorème 7.2, et la
construction que nous donnons n’est pas LOGSPACE (par contre, notre
résultat est plus précis en ce qui concerne le temps de construction du
circuit arithmétique).
La plupart des autres algorithmes développés dans cet ouvrage ont
une preuve d’uniformité plus simple, ou alors analogue à celle donnée
pour le théorème 7.2.
Dans le cas des familles non nécessairement uniformes, qui ont été
intensivemnt étudiées par Valiant, nous utiliserons les notations VN C k
et VN C en complexité arithmétique et BN C k et BN C en complexité
booléenne. (voir chapitres 12 et 13).
4.5
Machines parallèles à accès direct
Nous présentons brièvement dans cette section quelques modèles de
(( machines )) susceptibles d’exécuter des familles de circuits, arithmétiques ou booléens. Nous ne développerons pas cependant les questions de
la programmation pour les machines parallèles concrètes.
Le principal objet de la conception d’algorithmes parallèles est la
réduction du temps de calcul permettant de résoudre un problème donné
moyennant un nombre suffisant mais raisonnable de processeurs.
152
4.5.1
4. Notions de complexité
Une idéalisation des calculs parallèles sur ordinateur
A défaut de modèle unique nous devons faire un choix. En algorithmique séquentielle la Machine à Accès Direct ou (( Random Access
Machine )) (RAM) est une abstraction de l’ordinateur séquentiel de Von
Neumann. Nous considérons ici le modèle analogue en algorithmique
parallèle, celui des machines parallèles à accès direct (Parallel Random
Access Machines) ou PRAM, qui constitue le modèle (( standard )) (cf.
[CT, 36, 57]).
Une machine parallèle à accès direct ou PRAM est une machine
virtuelle (et un modèle idéal abstrait) composée d’un nombre illimité
de processeurs partageant une mémoire commune, la mémoire globale ,
elle-même constituée d’un nombre illimité de registres 9 , auxquels ils ont
accès pour y lire ou pour y écrire des données ou des résultats de calcul.
Chaque processeur a sa propre mémoire locale supposée également
de taille illimitée, et inaccessible aux autres processeurs. Elle lui permet d’exécuter en une seule unité de temps ou étape de calcul la tâche,
considérée comme élémentaire, composée de la suite d’instructions suivantes :
– chercher ses opérandes dans la mémoire globale ;
– effectuer l’une des opérations arithmétiques {+, −, ×} (et éventuellement la division quand elle est permise) sur ces opérandes ;
– écrire le résultat dans un registre de la mémoire commune (ou globale).
Faisant abstraction de tous les problèmes d’accès à la mémoire globale, de communication et d’interconnexion entre processeurs, une unité
de temps ou étape de calcul parallèle dans un tel modèle abstrait correspond à l’exécution simultanée de cette tâche par un certain nombre de
processeurs, les processeurs actifs , d’autres processeurs pouvant rester
inactifs.
L’exécution des tâches par l’ensemble des processeurs actifs est synchronisée : une étape démarre dès que les opérandes sont disponibles,
c’est-à-dire au démarrage du processus, quand chaque processeur sollicité puise ses données dans la mémoire globale, ou dès la fin d’une
étape quand chaque processeur actif a livré le résultat de son calcul, le
déroulement de ce calcul étant lié aux contraintes de dépendance entre
données dans l’algorithme considéré.
9. Le nombre de processeurs ainsi que le nombre de registres de mémoire partagée
sont habituellement fonctions de la taille du problème à traiter.
4.5. Machines parallèles à accès direct
153
Il existe plusieurs variantes du modèle PRAM selon le mode d’accès
à la mémoire globale, concurrent ou exclusif.
Ce sera une PRAM-EREW 10 si la lecture ou l’écriture dans un même
registre n’est permise qu’à un seul processeur à la fois, une PRAMCREW 11 si la lecture est concurrente et l’écriture exclusive, une PRAMERCW si la lecture est exclusive et l’écriture concurrente, et une PRAMCRCW si la lecture et l’écriture simultanées dans un même registre de
la mémoire globale sont permises pour plusieurs processeurs à la fois.
Dans les deux derniers cas, il faut éviter que deux processeurs mettent
simultanément dans un même registre des résultats différents, ce qui
donne d’autres variantes de machines PRAM selon le mode de gestion
de la concurrence d’écriture (mode prioritaire, arbitraire, etc.).
Même s’il existe une hiérarchie entre ces différentes variantes, de
la (( moins puissante )) (EREW) à la (( plus puissante )) (CRCW prioritaire), ces modèles PRAM sont en fait équivalents, pour la classe des
problèmes qui nous intéressent, dans le sens où ils se ramènent l’un à
l’autre par des techniques de simulation (cf. [CT, 36, 57]).
Nous utiliserons pour la description et l’analyse des algorithmes qui
nous concernent, la variante PRAM-CREW dont la conception est
très proche de la notion de circuit arithmétique ou de programme d’évaluation, puisqu’une PRAM-CREW peut être représentée par un circuit
arithmétique dans lequel les nœuds d’entrée représentent les données du
problème, et chacun des autres nœuds (internes) représente aussi bien un
processeur actif (et l’opération qu’il exécute) que le contenu d’un registre
de la mémoire globale correspondant au résultat de cette opération.
Enfin la profondeur du circuit arithmétique ou du programme d’évaluation telle que nous l’avons définie précédemment (section 3.1) correspond au nombre d’étapes du calcul parallèle.
4.5.2
PRAM-complexité et Processeur-efficacité
Plusieurs paramètres permettent de mesurer ce que nous appellerons
la PRAM-complexité d’un algorithme donné. Ces paramètres sont :
– le temps parallèle qui est égal au nombre d’étapes du calcul parallèle et qui correspond au temps d’exécution de l’algorithme parallèle ;
c’est aussi ce que l’on appelle la complexité parallèle ou la profondeur
de l’algorithme ;
10. EREW comme (( Exclusive Read, Exclusive Write )).
11. CREW comme (( Concurrent Read, Exclusive Write )) etc.
154
4. Notions de complexité
– le nombre de processeurs c’est-à-dire le nombre maximum de processeurs simultanément actifs durant une étape quelconque du calcul, sachant qu’un processeur peut être sollicité durant une ou plusieurs étapes
successives ;
– le temps séquentiel de l’algorithme c’est-à-dire le nombre d’opérations
arithmétiques qui interviennent dans le calcul ou, ce qui revient au
même, le temps parallèle si on ne disposait que d’un seul processeur,
ou encore la somme des nombres de processeurs actifs durant toutes les
étapes du calcul parallèle. C’est ce que l’on appelle aussi la taille et
parfois même la surface de calcul [CT] ou la complexité séquentielle de
l’algorithme ;
– le travail potentiel ou la surface totale de l’algorithme qui est le produit
du nombre de processeurs utilisés par le nombre d’étapes du calcul parallèle, c’est-à-dire le temps séquentiel si tous les processeurs étaient actifs
durant toutes les étapes du calcul.
On peut résumer la parfaite analogie des paramètres jusqu’ici définis
entre PRAM-CREW, circuit arithmétique et programme d’évaluation
par le tableau suivant :
PRAM-CREW
Temps parallèle
Temps séquentiel
Nombre de processeurs
Programme d’Evaluation
Profondeur
Longueur
Largeur
Circuit Arithmétique
Profondeur
Taille
Largeur
Tableau 4.5.2
Le nombre de processeurs dans une PRAM est l’équivalent de la largeur
dans un programme d’évaluation, le temps séquentiel dans une PRAM
est l’analogue de la longueur (ou la taille) d’un programme d’évaluation,
et le temps parallèle correspond à la profondeur.
L’efficacité d’un algorithme est alors définie comme le rapport entre
le temps séquentiel et le travail potentiel de cet algorithme, ou encore le
rapport entre surface de calcul et surface totale de l’algorithme considéré.
Pour revenir à l’exemple de l’algorithme du pivot de Gauss (voir page
117), la PRAM-CREW qui réalise cet algorithme peut être représentée
par le tableau suivant (rectangle de gauche) dont les lignes correspondent
aux 7 étapes successives du calcul et les colonnes aux processeurs (ceux
marqués d’une croix sont les processeurs actifs au cours d’une étape
donnée) :
4.5. Machines parallèles à accès direct
155
4 processeurs
z
Etape 1
Etape 2
Etape 3
Etape 4
Etape 5
Etape 6
Etape 7
−·−
−·−
−→
−→
−→
−→
−→
−→
−→
×
×
×
×
×
×
×
·
·
}|
{
×
× × ×
× × ×
×
·
·
·
·
2 processeurs
·
·
z }| {
Etape
Etape
Etape
Etape
Etape
Etape
Etape
Etape
Etape
1
2
3
4
5
6
7
8
9
−→
−→
−→
−→
−→
−→
−→
−→
−→
×
×
×
×
×
×
×
×
×
×
×
×
×
×
×
Le même algorithme peut être simulé par une PRAM à deux processeurs (rectangle de droite) au lieu de quatre, moyennant une augmentation du nombre d’étapes (c’est-à-dire un (( ralentissement )) des calculs)
avec 9 étapes au lieu de 7.
Pour chaque rectangle, la surface marquée représente la surface de
calcul ou le temps séquentiel, la surface totale représentant le travail
potentiel ; la longueur et la largeur du rectangle représentent respectivement le temps parallèle et le nombre de processeurs. L’efficacité de cet
algorithme passe de 15/28 quand il est réalisé par la PRAM initiale à
15/18 avec la PRAM modifiée c’est-à-dire de 54 % à 83 % environ.
Nous introduisons maintenant la notation classique suivante pour la
PRAM-complexité qui sera utilisée dans la suite.
Notation 4.5.1 On note PRAM (p(n), t(n)) la classe des problèmes de
taille n résolus par un algorithme PRAM-CREW en O(t(n)) étapes,
avec O(p(n)) processeurs. Tout algorithme P qui, exécuté sur une telle
machine, permet de résoudre un problème de cette classe, est lui-même
considéré, par abus de langage, comme appartenant à cette classe, et on
dira que P est un algorithme PRAM (p(n), t(n)).
La Processeur-efficacité d’un algorithme représenté par une PRAMCREW est une notion relative [50, 51, 62] estimée à partir du temps
séquentiel d’un algorithme choisi comme algorithme de référence : il
s’agit en ce qui nous concerne, pour l’algèbre linéaire, de l’algorithme de
la multiplication des matrices carrées d’ordre n supposé être réalisé par
156
4. Notions de complexité
une PRAM-CREW en log n étapes, avec M(n) processeurs. On peut
évidemment supposer M(n) = O(n3 ) et M(n) = Ω(n2 ).
Définition 4.5.2 Un algorithme P est dit processeur-efficace (par rapport à un algorithme de référence de temps séquentiel S(n)) s’il existe
k, m ∈ N∗ tels que P soit dans PRAM (S(n) logm (n), logk (n)).
Nous verrons plus loin des exemples d’algorithmes (( processeur-efficaces )) (comme celui de l’inversion des matrices fortement régulières,
page 195) pour lesquels on prend comme algorithme de référence celui de
la multiplication usuelle (resp. rapide) des matrices carrées n×n réalisée
par un circuit arithmétique en SD(n3 , log n) (resp. SD(nα , log n) pour
α < 3).
4.5.3
Le principe de Brent
Le principe de Brent affirme qu’on peut répartir intelligemment le
travail entre les différentes étapes d’un calcul parallèle, afin de diminuer
de manière significative la proportion des processeurs inactifs (cf. [10]
lemme 2.4).
Proposition 4.5.3 Un algorithme parallèle dont le temps séquentiel sur
une PRAM est égal à s(n) et dont le temps parallèle est égal à t(n) peut
être simulé sur une PRAM utilisant p processeurs et bs(n)/pc + t(n)
étapes de calcul sans changer le temps séquentiel.
Preuve. Supposons, en effet, qu’un calcul parallèle peut être effectué
en t(n) étapes parallèles à raison de mi opérations arithmétiques de
base par étape. Si l’on implémente directement ce calcul sur une PRAM
pour être exécuté en t(n) étapes, le nombre de processeurs utilisés sera
alors égal à m = max {mi | 1 ≤ i ≤ t(n)}. En prenant p processeurs au
lieu de m avec p < m (pour le cas p ≥ m, la proposition est triviale)
on peut exécuter le même calcul en faisant effectuer les mi opérations
de base de la i - ème étape par les p processeurs en dmi /pe étapes, et
comme dmi /pe ≤ bmi /pc + 1 le nombre total d’étapes avec une PRAM
à p processeurs n’excèdera pas
X
t(n)
X
t(n)
(bmi /pc + 1) ≤ t(n) +
mi /p ≤ t(n) + bs(n)/pc .
i=1
i=1
t
u
4.5. Machines parallèles à accès direct
157
Ce principe est très utile lorsque le temps parallèle t(n) est négligeable (quand n → ∞) devant le temps séquentiel s(n) de l’algorithme puisqu’on peut pratiquement diviser le nombre de processeurs
par t(n) en doublant simplement le temps d’exécution parallèle de
l’algorithme : on prend p = ds(n)/t(n)e. Par exemple, un algorithme
SD(nα , logk (n)) où α est un réel positif et k un entier naturel quelconque, donne par application de ce principe de Brent un algorithme
PRAM (nω / logk (n), logk (n)).
Cela permet dans la pratique, au prix d’un ralentissement relatif
(multiplication du temps de calcul par une petite constante), d’améliorer
l’efficacité d’un algorithme parallèle en diminuant le temps d’inactivité
des processeurs par une réduction du rapport entre le travail potentiel (i.e. la surface totale) et le travail réel (i.e. la surface de calcul),
et ceci par une réorganisation des calculs dans le sens d’une meilleure
répartition des processeurs entre les étapes parallèles.
Nous en déduisons la propriété suivante qui relie la complexité des
circuits arithmétiques à celle des PRAM.
Proposition 4.5.4 Un algorithme parallèle en SD(f (n), g(n)) est un
algorithme PRAM (f (n)/g(n), g(n)). Inversement, tout algorithme dans
PRAM (p(n), t(n)) est un algorithme en SD(p(n)t(n), t(n)).
Remarque. Dire qu’un algorithme est processeur-efficace par rapport
à un algorithme de référence de temps séquentiel S(n) revient à dire
qu’il est SD(S(n) logm (n), logk (n)) pour un couple (m, k) ∈ N∗ × N∗ .
5. Diviser pour gagner
Introduction
Dans ce chapitre, nous présentons une approche bien connue sous le
nom de (( divide and conquer )) que l’on peut traduire par (( diviser pour
régner )) auquel nous préférons le concept (( diviser pour gagner )) parce
que mieux adapté, nous semble-t-il, au calcul parallèle.
Après en avoir donné le principe général nous l’utilisons pour étudier
deux problèmes classiques de l’algorithmique parallèle que nous serons
amenés à utiliser dans la suite :
– le calcul du produit de n éléments d’un monoı̈de ;
– le problème du calcul parallèle des préfixes ((( Parallel Prefix Algorithm )))
Pour ce dernier problème, nous développerons, en plus de l’algorithme classique, une méthode récursive due à Ladner & Fischer [64] pour
obtenir une famille de circuits de taille linéaire et de profondeur logarithmique. C’est le meilleur résultat connu à l’heure actuelle.
Nous appliquerons la stratégie (( diviser pour gagner )) en plusieurs
autres occasions dans les chapitres suivants, notamment pour les multiplications rapides de matrices et de polynômes et pour l’algèbre linéaire
rapide sur les corps.
5.1
Le principe général
L’approche (( diviser pour gagner )) s’applique pour résoudre une
famille de problèmes (Pn )n∈N . Elle consiste à (( diviser )) le problème
numéro n en q (q ≥ 2) sous-problèmes du style Pm avec m < n, auxquels on peut appliquer, en parallèle et de manière récursive, le même
algorithme que celui qui permet de résoudre le problème initial, pour
récupérer ensuite le résultat final à partir des solutions des sous-problèmes.
160
5. Diviser pour gagner
Le paramètre entier q représente le nombre des sous-problèmes qui
seront traités en parallèle. Lorsqu’il ne dépend pas de n, il s’appelle le
degré de parallélisme de l’algorithme ainsi obtenu.
Une telle approche récursive de conception d’algorithmes permet
souvent d’apporter une solution efficace à un problème dans lequel les
q sous-problèmes Pm sont des copies réduites du problème initial, et
avec m sensiblement égal (à dn/pe par exemple, où p est un entier
donné ≥ 2).
Cette méthode nous permet également d’analyser la complexité de
l’algorithme qu’elle produit et de calculer des majorants asymptotiques
de la taille et de la profondeur du circuit arithmétique correspondant,
avec une estimation précise de la constante cachée du (( grand O )).
En effet, supposons que le problème à traiter est le problème no n =
m0 pν (m0 , p, ν ∈ N∗ ) et qu’il peut être scindé en q sous-problèmes Pm
avec m = m0 pν−1 , suceptibles d’être traités en parallèle. Remarquons
tout de suite que q est un entier ≥ 2 dépendant éventuellement de ν :
c’est pourquoi on écrira, dans le cas général, q = q(ν).
Le coût κ̂(ν) = (τ (ν), π(ν)) de cet algorithme où τ (ν) (resp. π(ν))
désigne la taille (resp. la profondeur) du circuit correspondant, se calcule
par récurrence sur ν à l’aide des formules suivantes :
(
τ (ν) = q(ν) τ (ν − 1) + τ 0 (ν)
(5.1)
π(ν) = π(ν − 1) + π 0 (ν)
où τ 0 (ν) (resp. π 0 (ν)) représente la taille (resp. la profondeur) des circuits correspondant à la double opération de partitionnement du problème et de récupération de sa solution à partir des solutions partielles.
L’absence du facteur q dans l’égalité exprimant la profondeur π est
due au fait que les q sous-problèmes, de même taille, sont traités en
parallèle avec des circuits de profondeur maximum π(ν − 1).
Si l’on se donne τ (0) et π(0) le système (5.1) ci-dessus admet pour
solution :
ν hY
i
X
ν
τ
(ν)
=
q(1)
q(2)
·
·
·
q(ν)
τ
(0)
+
q(j)
τ 0 (i)
j=i+1
i=1
(5.2)
ν
X
0
π (i)
π(ν) = π(0) +
i=1
Dans le cas où q = q(ν) est une constante, sachant que la profondeur,
ne dépendant pas de q, reste la même, le système (5.2) devient (5.3)
5.1. Le principe général
161
ci-dessous. Nous rappelons précisément les hypothèses dans l’énoncé qui
suit.
Proposition 5.1.1 Soient m0 , p, q ∈ N∗ fixés et ν ∈ N∗ variable.
Nous supposons que le problème à traiter est le problème Pn avec n =
m0 pν et qu’il peut être scindé en q sous-problèmes de type Pm avec
m = m0 pν−1 , suceptibles d’être traités en parallèle. Nous notons τ 0 (ν)
(resp. π 0 (ν)) la taille (resp. la profondeur) des circuits correspondant à
la double opération de partitionnement du problème et de récupération
de sa solution à partir des solutions partielles. Enfin τ0 et π0 sont la
taille et la profondeur d’un circuit qui traite le problème Pm0 . Alors la
taille et profondeur d’un circuit produit en utilisant la méthode (( diviser
pour gagner )) sont :
(
P
τ (ν) = q ν τ0 + νi=1 q ν−i τ 0 (i)
P
π(ν) = π0 + νi=1 π 0 (i)
(5.3)
En particulier si τ 0 (ν) = O(nr ) avec r 6= log q et π 0 (ν) = O(ν ` ) on
obtient :
(
τ (ν) = O(q ν ) = O(nsup(r,log q) )
(5.4)
π(ν) = O(ν `+1 ) = O(log`+1 n)
Donnons un aperçu rapide sur quelques cas particuliers significatifs
que nous allons traiter dans la suite.
Dans le calcul parallèle des préfixes section 5.3, nous avons de manière
naturelle p = q = 2, r = 1 et ` = 0 ce qui conduit à une famille de
circuits en SD(n log n, log n), et nous verrons qu’on peut encore très
légèrement améliorer la borne sur la taille.
Dans la multiplication des polynômes à la Karatsuba section 6.1,
nous avons p = 2, q = 3, r = 1 et ` = 0 ce qui conduit à une famille
de circuits en SD(nlog 3 , log n).
Dans la multiplication rapide des matrices à la Strassen section 7.1,
nous avons p = 2, q = 7, r = 2 et ` = 0 ce qui conduit à une famille
de circuits en SD(nlog 7 , log n).
Enfin pour l’inversion des matrices triangulaires section 7.2, nous
avons p = 2, q = 2, r = α et ` = 1 ce qui conduit à une famille de
circuits en SD(nα , log2 n).
162
5.2
5. Diviser pour gagner
Circuit binaire équilibré
L’approche (( diviser pour gagner )), appliquée à ce premier problème, nous donne la construction d’un type particulier de circuits arithmétiques de taille linéaire et de profondeur dlog ne que l’on appelle les
circuits binaires équilibrés ((( Balanced Binary Trees ))).
Un circuit binaire équilibré est un circuit arithmétique prenant en
entrée une liste (x1 , x2 , . . . , xn−1 , xn ) de n éléments d’un monoı̈de M
(loi associative notée ∗ avec élément neutre noté 1) et donnant en sortie
le produit Π = x1 ∗ x2 ∗ · · · ∗ xn−1 ∗ xn .
On peut supposer n = 2ν où ν ∈ N∗ quitte à compléter la liste
donnée par 2dlog ne − n éléments égaux à 1, ce qui ne change pas le
résultat.
Le circuit est défini de manière récursive en divisant le problème en
deux sous-problèmes de taille 2ν−1 , qui correspondent à deux (( sous-circuits )) acceptant chacun en entrée une liste de taille moitié.
Ces deux sous-circuits calculent respectivement et en parallèle les
deux produits partiels Π1 = x1 ∗· · ·∗x2ν−1 et Π2 = x2ν−1 +1 ∗· · ·∗x2ν . On
récupère ensuite le produit Π en multipliant ces deux produits partiels.
Ainsi un circuit binaire équilibré pour une entrée de taille 2ν est
défini par récurrence sur ν : pour ν = 0 c’est le circuit trivial C0 de
taille profondeur nulles. Pour ν ≥ 1, le circuit Cν prend en entrée une
liste de longueur 2ν , fait agir deux copies du circuit Cν−1 pour calculer
Π1 et Π2 qu’il utilise pour récupérer le résultat final Π = Π1 ∗ Π2
(comme l’indique la figure 5.1 page ci-contre). Si l’on note τ (ν) et π(ν)
la taille et la profondeur du circuit Cν , on obtient les relations :
(
τ (ν) = 2τ (ν − 1) + 1 avec τ (0) = 0
π(ν) = π(ν − 1) + 1
qui admet la solution exacte :
(
τ (ν)
avec π(0) = 0
= 2ν − 1
π (ν) = ν .
(5.5)
Proposition 5.2.1 Un circuit binaire équilibré qui prend en entrée une
liste une liste (x1 , x2 , . . . , xn−1 , xn ) dans un monoı̈de M et donne en
sortie le produit Π = x1 ∗ x2 ∗ · · · ∗ xn−1 ∗ xn est un circuit arithmétique
de profondeur dlog ne. Il est de taille n − 1 si n est une puissance de
5.3. Calcul parallèle des préfixes
163
2 ν e n tré e s
2 ν − 1 e n tré e s
2 ν − 1 e n tré e s
....
....
ν−1
ν−1
π1
π2
π
Figure 5.1 – Construction récursive du circuit binaire équilibré Cν
(à partir du circuit binaire équilibré Cν−1 )
2, et cette taille est en tous cas majorée par 2n − 3 lorsque n n’est pas
une puissance de 2.
Notons qu’on peut trouver une majoration légèrement meilleure de la
taille pour n > 3.
5.3
Calcul parallèle des préfixes
Étant donnée une liste de n éléments x1 , x2 , . . . , xn (ou n -uplet)
d’un monoı̈de (M, ∗, 1) dont la loi (en général non commutative) est
notée multiplicativement et dont l’élément neutre est noté 1, le problème du calcul des préfixes consiste à calculer les produits partiels
Yk
Πk =
xi pour (1 ≤ k ≤ n).
i=1
La solution naı̈ve de ce problème donne un circuit de taille n − 1
(c’est la taille minimum) et de profondeur n − 1.
Première méthode de parallélisation
Il est facile de voir que ce calcul peut être parallélisé pour obtenir
un circuit de profondeur dlog ne.
164
5. Diviser pour gagner
On peut toujours supposer n = 2ν où ν = dlog ne ∈ N∗ , quitte à
compléter la liste donnée par 2dlog ne − n copies de l’élément neutre 1.
Ce problème peut se décomposer en deux sous-problèmes de taille
n/2 = 2ν−1 qui seront traités en parallèle :
• le calcul des préfixes pour la liste x1 , x2 , . . . , x2ν−1 ;
• le calcul des préfixes pour la liste y1 , y2 , . . . , y2ν−1 où yi = x2ν−1 +i
pour 1 ≤ i ≤ 2ν−1 .
La solution du problème principal est ensuite obtenue par multiplication du produit Π2ν−1 , faisant partie de la solution du premier sous-problème, par les 2ν−1 produits partiels des yi qui constituent la solution
du second sous-problème. Cette dernière étape de récupération augmente
par conséquent de 2ν−1 multiplications la taille du circuit et de 1 sa profondeur.
Pour le cas n = 7 par exemple (on prend n = 8 pour avoir une
puissance de 2 et on fait x8 = 1), on obtient le circuit 5.1 qui montre le
déroulement de cette procédure pour le calcul des sept (ou huit) produits
Π1 = x1 , Π2 = x1 ∗x2 , Π3 = x1 ∗x2 ∗x3 , . . . , Π7 = x1 ∗x2 ∗· · ·∗x7 (Π8 =
Π7 puisque x8 = 1). Appliquées à notre problème, les relations (5.3)
x1
x2
x3
x4
x5
x6
(x 8 = 1 )
x7
E ta p e 1
E ta p e 2
E ta p e 3
1
2
3
Circuit 5.1:
4
5
6
7
(
8
=
7)
Calcul Parallèle des Préfixes pour n = 7
donnent la taille et la profondeur du circuit arithmétique correspondant
au calcul parallèle des préfixes pour une liste donnée de taille 2ν .
Il suffit en effet de faire p = q = 2, τ 0 (i) = 2i−1 , π 0 (i) = 1 (pour
5.3. Calcul parallèle des préfixes
165
i ≥ 1) et τ (0) = π(0) = 0 pour obtenir :
τ (ν) =
ν
X
i=1
ν
2
ν−i i−1
2
ν−1
=ν2
X
n
= log n, et π(ν) =
1 = ν = log n.
2
i=1
Ainsi le problème du calcul des préfixes pour une liste de n éléments
se parallélise bien, et il admet une solution en SD(n log n, log n) ou
encore, en utilisant le principe de Brent (proposition 4.5.3), une solution
qui est PRAM (n, log n).
Ladner & Fischer [64] obtiennent un meilleur résultat en donnant
une construction récursive d’un circuit en SD(n, log n). C’est ce que
nous allons développer au paragraphe suivant.
Amélioration du calcul des préfixes (Ladner & Fischer)
Étant donnés un monoı̈de (M, ∗, 1), un entier n ≥ 2, et x1 , . . . , xn
dans M, nous allons construire, à l’instar de Ladner & Fischer [64]
deux familles de circuits (Pk (n)) n∈N∗ de tailles Sk (n) (k ∈ {0, 1})
majorées respectivement par 4n et 3n et de profondeurs respectives
D0 (n) = dlog ne et D1 (n) = dlog ne + 1 qui calculent les préfixes
Π1 , Π2 , . . . , Πn du n-uplet (x1 , x2 , . . . , xn ).
Cette construction se fait conjointement et de manière récursive à
partir du circuit trivial P0 (1) = P1 (1) réduit à une seule porte (la
porte d’entrée). La figure 5.2 page suivante montre le déroulement de
cette construction récursive conjointe des deux familles (P0 (n)) n∈N∗ et
(P1 (n)) n∈N∗ .
Construction de la famille (P0 (n)) n∈N∗
On définit récursivement le circuit P0 (n) à partir des circuits P1 (b n2 c)
et P0 (d n2 e) appliqués respectivement aux entrées (x1 , . . . , xb n2 c )
et
n
(xb 2 c+1 , . . . , xn ) qui forment une partition de la liste donnée (x1 , . . . , xn ).
Comme P1 (b n2 c) calcule Π1 , Π2 , . . . , Πb n2 c , il suffit d’effectuer en parallèle et en une seule étape les d n2 e multiplications de Πb n2 c par les d n2 e
sorties de P0 (d n2 e) pour avoir les préfixes Πb n2 c+1 , Πb n2 c+2 , . . . , Πn et
par conséquent tous les préfixes Π1 , Π2 , . . . , Πn de la liste (x1 , . . . , xn ).
Partant du circuit trivial P0 (1) = P1 (1), la figure 5.3 page 167
illustre cette construction.
La construction du circuit (P1 (n)), quant à elle, se fait à partir du
circuit P0 (b n2 c), elle est illustrée par la figure 5.4 page 168.
166
5. Diviser pour gagner
P0 (1)
P1 (1)
2 ≤ m0 < 4
P0 (m0)
P1 (m0)
4 ≤ m1 < 8
P0 (m0)
P1 (m1)
...
...
P0 (mp-1)
P1 (mp-1)
P0 (mp)
P1 (mp)
...
...
2p-1 ≤ mp-1 < 2p
2p ≤ mp < 2p+1
Figure 5.2 – Schéma de la construction récursive des circuits
(Pk (n)) n∈N
(k ∈ {0, 1}, (mp = 2 mp−1 pour 1 ≤ p ≤ dlog ne)
Construction de la famille (P1 (n)) n∈N∗
• On commence par calculer en parallèle (c’est-à-dire en une seule étape)
les produits x1 ∗ x2 , x3 ∗ x4 , · · · , x2p−1 ∗ x2p (où p = b n2 c) d’un élément
de rang impair par l’élément suivant (de rang pair) dans la liste donnée
(x1 , . . . , , xn ) (n = 2p si n est pair et n = 2p + 1 si n est impair).
• À ce p-uplet on applique le circuit P0 (b n2 c) = P0 (p) pour obtenir en
sortie les p préfixes de longueur paire : Π2 , Π4 , . . . , Π2p .
• On multiplie enfin, et en parallèle, les préfixes Π2 , Π4 , . . . , Π2p respectivement par les entrées (x3 , x5 , . . . , x2p−1 (et éventuellement x2p+1
si n est impair) pour obtenir, en plus de Π1 (Π1 = x1 est déjà donné),
les autres préfixes de longueur impaire : Π3 , Π5 , . . . , Π2p−1 (et éventu-
5.3. Calcul parallèle des préfixes
167
n entrées
n/2 entrées
n/2 entrées
....
....
P1 ( n/2 )
π1
π2
...
P0 ( n/2 )
π
n/2
...
πn
Figure 5.3 –
Construction récursive des circuits P0 (n).
ellement Π2p+1 si n est impair).
On obtient ainsi le circuit arithmétique parallèle P1 (n) à partir du
circuit P0 (b n2 c) en ajoutant au maximum deux étapes (à l’entrée et à
la sortie) comportant au total n − 1 opérations arithmétiques (2p − 1
si n est pair et 2p si n est impair).
Les circuits 5.2 page 169 sont des exemples de circuits P0 (n) et
P1 (n) pour quelques valeurs de n.
Analyse de la complexité des circuits
Si l’on note Sk (n) (resp. Dk (n)) la taille (resp. la profondeur) du circuit
Pk (n) pour n ≥ 2 et k ∈ {0, 1}, cette construction récursive donne les
relations suivantes :
– Pour la taille :
(
S1 (n) = S0 (b n2 c) + n − 1
(5.6)
S0 (n) = S1 (b n2 c) + S0 (d n2 e) + d n2 e
– Pour la profondeur :
(
D1 (n) ≤ D0 (b n2 c) + 2
D0 (n) = max {D0 (b 12 b n2 cc) + 1 , D0 (d n2 e)} + 1
(5.7)
168
5. Diviser pour gagner
x1
x2 x3
∗
n/2 entrées
x4
x5
...
...
x2p-1 x2p x2p+1
Première
Etape
∗
∗
...
Pk-1 ( n/2 )
π2p
π2p-2
...
...
n/2 sorties
π4
π2
π1
∗
∗
∗
π3
π5
...
π2p-1
∗
π2p+1
Dernière
Etape
Figure 5.4 – Construction du circuit P1 (n) à partir du circuit P0 (b n2 c)
(p = b n2 c et les 2 lignes en pointillé sont absentes si n est pair)
avec Sk (1) = Dk (1) = 0 pour tout k ∈ {0, 1}.
Il faut remarquer que l’inégalité D1 (n) ≤ D0 (b n2 c) + 2 dans (5.7)
peut être stricte (voir par exemple le circuit P1 (6) dans les circuits 5.2
page ci-contre pour s’en convaincre).
La deuxième équation dans (5.7) est justifiée par le fait que, dans
le circuit P0 (n), le nœud correspondant au produit Πb n2 c – dont on a
besoin pour calculer en une étape supplémentaire les autres préfixes – se
trouve exactement à la profondeur D0 (b 12 b n2 cc) + 1 dans le sous-circuit
P1 (b n2 c) de P0 (n) qui calcule ce produit.
Il est facile de voir, à partir des équations (5.7), par une récurrence
immédiate sur n, que les profondeurs Dk (n) des circuits Pk (n) pour
k ∈ {0, 1} vérifient :
D0 (n) = dlog ne
et
D1 (n) ≤ dlog ne + 1 .
Pour calculer les tailles des circuits à partir des équations (5.6), nous
allons d’abord considérer le cas où n est une puissance de 2 en faisant
n = 2ν où ν = dlog ne.
Posant τk (ν) = Sk (2ν ) avec τk (0) = 0 pour k ∈ {0, 1} les équations
(5.6) deviennent :
5.3. Calcul parallèle des préfixes
x1
x2
x1
x2
x3
169
x1 x2 x3 x4 x5
x
x
x
x1 x2 x3 x4 x5 x6
x6
x
x
x
x
x
x
Pk (2)
1
(k = 0, 1)
2
x
x
3
x
x
x
Pk (3)
1
(k = 0, 1)
2
3
4
5
x
6
x
x
P1 (6)
1
2
3
4
5
6
P0 (6)
x1 x2 x3 x4 x5 x6 x7 x8
x1 x2 x3 x4 x5 x6 x7
x
x1 x2 x3 x4 x5 x6 x7
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
1
2
3
4
5
P0 (7)
Circuit 5.2:
(
6
7
1
2
3
4
P1 (7)
5
6
7
1
2
3
4
5
6
7
8
P1 (8)
Circuits P0 (n) , P1 (n) pour quelques valeurs de n.
τ0 (ν) = τ0 (ν − 1) + τ0 (ν − 2) + 2ν − 1
τ1 (ν) = τ1 (ν − 1) + τ1 (ν − 2) + 3.2ν−2 .
(5.8)
Posant u0 (ν) = 4.2ν + 1 − τ0 (ν) et u1 (ν) = 3.2ν − τ1 (ν), les relations
(5.8) permettent de vérifier que uk (ν + 2) = uk (ν + 1) + uk (ν) (ν ∈ N,
k ∈ {0, 1}). Comme u0 (0) = 5 , u0 (1) = 8 , u1 (0) = 3 et u1 (1) = 5 , on
en déduit que :
u0 (ν) = F (ν + 5) et u1 (ν) = F (ν + 4)
où (F (ν))ν∈N est la suite de Fibonacci 1 . Par conséquent :
1. La suite de Fibonacci est définie par F (0) = 0, F (1) = 1 et la relation F (ν +
2) = F (ν + 1) + F (ν) pour tout ν ∈ N.
170
5. Diviser pour gagner
(
τ0 (ν) = 4.2ν + 1 − F (ν + 5)
τ1 (ν) = 3.2ν − F (ν + 4) .
(5.9)
qui donne, lorsque n est une puissance de 2, les majorations souhaitées.
Dans le cas contraire, il est facile – en utilisant directement les relations
(5.7) – d’obtenir, par récurrence sur n, les majorations suivantes vraies
pour tout n ≥ 2 :
S0 (n) ≤ 4 n − 7
et S1 (n) ≤ 3 n − 3 .
Ce qui donne le résultat suivant de Ladner & Fischer ([64]) qui
montre que le calcul des préfixes est PRAM (n/ log n, log n) :
Théorème 5.1 (Ladner & Fischer) Le calcul des préfixes d’une liste
de n éléments dans un monoı̈de (non nécessairement commutatif ) se
fait par un circuit arithmétique parallèle de profondeur dlog ne et de
taille inférieure à 4n et aussi par un circuit arithmétique parallèle de
profondeur 1 + dlog ne et de taille inférieure à 3n.
6. Multiplication rapide
des polynômes
Introduction
Soit A un anneau commutatif unitaire et A[X] l’anneau des polynômes à une indéterminée sur A.
P
P
i
Le produit de deux polynômes A = ni=0 ai X i et B = m
i=0 bi X
est défini par
C = AB =
m+n
X
k=0
ck X k
avec
ck =
k
X
ai bk−i pour 0 ≤ k ≤ m + n .
i=0
L’algorithme usuel pour le calcul des coefficients du polynôme C correspond à un circuit arithmétique de profondeur O(log m) (si l’on suppose m ≤ n) et de taille O(mn) avec précisément (m + 1) (n + 1) multiplications et mn additions dans l’anneau de base A. Pour m = n,
cela donne un algorithme en SD(n2 , log n).
Dans les trois premières sections nous exposons deux façons d’améliorer la multiplication des polynômes.
Dans la section 6.1 nous expliquons la méthode de Karatsuba, facile
à implémenter pour n’importe quel anneau commutatif, avec un résultat
en SD(nlog 3 , log n).
Un bien meilleur résultat est obtenu en SD(n log n, log n) grâce à la
transformation de Fourier discrète ([AHU, Knu]) pour un anneau auquel
s’applique une telle transformation. Ceci fait l’objet des sections 6.2 et
6.3.1.
Dans la section 6.3.2 nous exposons une amélioration due à Cantor
et Kaltofen [13] qui ont étendu le résultat à tout anneau commutatif
unitaire en exhibant un algorithme en SD(n log n log log n, log n) (avec
172
6. Multiplication rapide des polynômes
le même nombre de multiplications dans l’anneau de base, le facteur
log log n étant dû à l’augmentation du nombre d’additions). Pour réaliser
ce travail il a fallu l’adjonction de racines principales de l’unité à l’anneau
considéré. On peut comparer la borne obtenue avec la meilleure borne
inférieure actuellement connue, qui est O(n).
Dans la section 6.4 nous donnons le lien entre la multiplication des
polynômes et celle des matrices de Toeplitz triangulaires inférieures.
Nous en déduisons un résultat de complexité intéressant concernant le
produit d’une matrice de Toeplitz arbitraire par une matrice arbitraire.
6.1
Méthode de Karatsuba
Considérons deux polynômes arbitraires A et B et leur produit C.
Si les polynômes A et B sont de degré < d (déterminés chacun par
d coefficients), leur produit C = AB peut être calculé en appliquant
directement la formule qui le définit. Il y a alors d2 multiplications et
(d − 1)2 additions. Les d2 multiplications peuvent être calculées en une
seule étape de calcul parallèle et les 2d−1 coefficients de C sont ensuite
calculés en dlog de étapes parallèles (le coefficient réclamant l’addition
la plus longue est celui de degré d − 1).
Une première façon d’améliorer cette multiplication est d’adopter
une démarche récursive basée sur le fait que le produit de deux polynômes de degré 1 peut s’effectuer avec seulement 3 multiplications au lieu
de 4 (le nombre d’additions/soustractions passant de 1 à 4). En effet, on
peut calculer a + bX + cX 2 = (a1 + a2 X) (b1 + b2 X) en posant :
a = a1 b1 ,
c = a2 b2 ,
b = (a1 + a2 ) (b1 + b2 ) − (a + c),
(6.1)
ce qui correspond à un circuit arithmétique de profondeur totale 3, de
largeur 4 et de profondeur multiplicative 1.
Considérons maintenant deux polynômes arbitraires A et B et leur
produit C . Ces polynômes s’écrivent de manière unique, sous la forme :
A = A1 (X 2 ) + X A2 (X 2 )
B = B1 (X 2 ) + X B2 (X 2 )
C = C1 (X 2 ) + X C2 (X 2 )
avec C1 = A1 B1 + X A2 B2 et C2 = A1 B2 + A2 B1 . Si A1 , B1 , A2 , B2
sont de degrés ≤ k − 1 (avec k coefficients) alors A et B sont de degré
≤ 2k − 1 (avec 2k coefficients).
6.1. Méthode de Karatsuba
173
Supposons qu’un programme d’évaluation Kara(k) calcule les coefficients du produit de deux polynômes arbitraires de degré ≤ k − 1,
avec une profondeur multiplicative égale à µ(k) , une profondeur totale
égale à π (k) , une largeur égale à λ(k) , un nombre de multiplications
égal à m(k) , un nombre d’additions/soustractions égal à a(k) , et donc
avec pour nombre total d’opérations arithmétiques s(k) = a(k) + m(k) .
L’utilisation des équations (6.1) donne un circuit arithmétique Kara(2k)
que nous avons décrit schématiquement dans le programme d’évaluation 6.1.
Programme d’évaluation 6.1 Kara(2k)
Entrée : Les 4k coefficients dans A (un anneau commutatif arbitraire)
de deux polynômes de degré < 2k : A(X) = A1 (X 2 ) + X A2 (X 2 ) et
B(X) = B1 (X 2 ) + X B2 (X 2 ).
Sortie : Les coefficients du produit des deux polynômes : C(X) = C1 (X 2 )+
X C2 (X 2 ).
Début
profondeur 1 :
D1 := A1 + A2 ; D2 := B1 + B2
profondeur π (k) :
D3 := Kara(k) (A1 , B1 ) ; D4 := Kara(k) (A2 , B2 )
profondeur π (k) + 1 :
D5 := Kara(k) (D1 , D2 ) ; D6 := D3 + D4 ; C1 := D3 + XD4
profondeur π (k) + 2 :
C2 := D5 − D6
Fin.
Notez que la ligne écrite avec la profondeur π (k) représente la dernière
ligne des deux programmes d’évaluation Kara(k) (A1 , B1 ) et Kara(k) (A2 ,
B2 ), qui ont démarré en parallèle avec les deux affectations indiquées sur
la ligne de profondeur 1. Sur la ligne écrite avec la profondeur π (k) +1, la
première affectation correspond à la dernière ligne du programme d’évaluation Kara(k) (D1 , D2 ) qui a commencé à la profondeur 2, tandis que
les deux autres affectations sont effectuées à la profondeur π (k) + 1.
On constate donc que lorsqu’on passe de Kara(k) à Kara(2k) selon
la méthode décrite dans le programme d’évaluation 6.1 :
– la profondeur passe de π (k) à π (2k) = π (k) + 2,
– la profondeur multiplicative n’a pas changé (µ(2k) = µ(k) ),
174
6. Multiplication rapide des polynômes
– la largeur passe de λ(k) à λ(2k) = sup(3 λ(k) , λ(k) + 4k − 2),
– le nombre de multiplications est maintenant m(2k) = 3m(k) ,
– le nombre d’additions/soustractions ( 1 ) est a(2k) = k + k + 3a(k) +
(2k − 1) + (2k − 2) + (2k − 1) = 3a(k) + 8k − 4,
– et le nombre total d’opérations arithmétiques passe de s(k) à
s(2k) = 3s(k) + 8k − 4.
En comparaison, pour la multiplication usuelle des polynômes, le
nombre de multiplications passe de m̃(k) = k 2 à m̃(2k) = 4k 2 = 4m̃(k) , le
nombre d’additions/soustractions de ã(k) = (k−1)2 à ã(2k) = (2k−1)2 =
4ã(k) + 4k − 3 et le nombre total d’opérations arithmétiques de s̃(k) à
s̃(2k) = 4s̃(k) + 4k − 3.
Si on veut minimiser le nombre de multiplications on initialisera la
processus récursif avec Kara(1) (le produit de deux constantes) et on
mettra en place les circuits arithmétiques successifs Kara(2) , Kara(4) ,
ν
Kara(8) , . . ., Kara(2 ) selon la procédure décrite ci-dessus. Le circuit
ν
Kara(2 ) = Karaν est ensuite utilisé pour le produit de deux polynômes
de degrés < n = 2ν et ≥ 2ν−1 . Pour deux polynômes de degré exactement n−1 on aura ainsi remplacé le circuit arithmétique usuel qui utilise
4ν = n2 multiplications par un circuit arithmétique Kara(n) = Karaν
qui utilise 3ν = nlog 3 ' n1.585 multiplications 2 . Le gain concernant le
nombre total d’opérations arithmétiques est du même style. En notant
sν pour s(n) , on passe en effet de sν à sν+1 = 3 sν + 8.2ν − 4. Les
premières valeurs de sν sont s0 = 1, s1 = 7, s2 = 33 et la relation de
récurrence se résoud avec l’aide de Maple en :
sν = 7 · 3ν − 8 · 2ν + 2.
En fait sν devient meilleur que 4n + (2n − 1)2 à partir de ν = 4 (pour
des polynômes de degré 15). Enfin, concernant la largeur λν du circuit
arithmétique Karaν , la résolution de la récurrence donne λν = 2 · 3ν
pour ν ≥ 2.
Nous pouvons conclure avec la proposition suivante.
Proposition 6.1.1 La multiplication de deux polynômes de degré ≤ n
par la méthode de Karatsuba se fait en SD(nlog 3 , log n). Plus précisément, le produit de deux polynômes de degrés < 2ν = n peut être réalisé
1. On ne compte pas les opérations de substitution de X 2 à X ou vice-versa,
ni les multiplications par X ou par X 2 , qui reviennent en fait à des décalages de
coefficients.
2. log 3 = 1.58496250072115618145373894394.
6.2. Transformation de Fourier discrète usuelle
175
par un circuit arithmétique de profondeur multiplicative 1, de profondeur
totale 1+2 ν, de largeur 2·3ν = 2 nlog 3 , avec 3ν = nlog 3 multiplications
et 6 · 3ν − 8 · 2ν + 2 = 6nlog 3 − 8n + 2 additions/soustractions.
Notons que pour deux polynômes dont les degrés sont compris entre
et 2ν , on obtient seulement les majorations suivantes en appelant
n le plus grand degré : 3 + 2 log n pour la profondeur, 6 nlog 3 pour la
largeur et 21 nlog 3 − 8 n + 2 pour la taille du circuit.
Remarquons qu’on aurait pu envisager une autre partition de coefficients des polynômes A et B pour une application récursive, à savoir
2ν−1
A = A1 + X k A2
et
B = B1 + X k B2 .
avec Ai et Bi de degrés ≤ k − 1, A et B de degrés ≤ 2k − 1. Alors
C = A B = A1 B1 + X k (A1 B2 + A2 B1 ) + X 2k A2 B2 . Une procédure
récursive basée sur cette partition produirait des circuits arithmétiques
avec une estimation analogue à la précédente pour ce qui concerne la
taille mais une profondeur de 1 + 3 log n au lieu de 1 + 2 log n (pour
le produit de deux polynômes de degré n − 1 lorsque n = 2ν ).
6.2
Transformation de Fourier discrète usuelle
Un bien meilleur résultat, que nous exposons dans cette section et
la suivante, est obtenu en SD(n log n, log n) grâce à la transformation
de Fourier discrète pour un anneau auquel s’applique une telle transformation. La transformation de Fourier discrète, que nous désignerons ici
par le sigle TFD, est définie sur un anneau commutatif unitaire A, pour
un entier donné n ≥ 2, à condition de disposer dans A d’une racine
n - ème principale de 1, c’est-à-dire d’un élément ξ ∈ A vérifiant :
ξ 6= 1 , ξ n = 1 , et
n−1
X
ξ ij = 0 pour i = 1, . . . , n − 1 .
j=0
Dans un anneau intègre, toute racine primitive 3 n - ème de 1 est
principale, mais ceci peut-être mis en défaut dans un anneau contenant
des diviseurs de zéro. Dans un anneau intègre, s’il y a une racine primitive n - ème de 1, il y en a ϕ(n) où ϕ désigne l’indicatrice d’Euler.
Dans C, les racines n - èmes principales de 1 sont les nombres complexes
e2ikπ/n tels que 1 ≤ k < n et k premier avec n.
3. C’est un ξ tel que ξ n = 1 mais ξ m 6= 1 si 1 ≤ m < n.
176
6. Multiplication rapide des polynômes
Il est clair que si ξ est une racine n - ème principale de 1, alors il en
est de même de ξ −1 .
Définition 6.2.1 La transformation de Fourier discrète d’ordre n sur
A, associée à la racine principale ξ, est l’application linéaire
TFDn,ξ : An −→ An
définie, pour tout (a0 , a1 , . . . , an−1 ) ∈ An par :
TFDn,ξ (a0 , a1 , . . . , an−1 ) = (A(1), A(ξ), . . . , A(ξ n−1 ))
où A est le polynôme A(X) = a0 + a1 X + . . . + an−1 X n−1 .
Cette application peut aussi être vue comme un homomorphisme de
A-algèbres
TFDn,ξ : A[X] /hX n − 1i −→ An
qui à tout polynôme A de degré ≤ n − 1 associe le vecteur formé des
valeurs de A aux points 1, ξ, . . . , ξ n−1 . En effet, en notant
la loi
produit (coordonnée par coordonnée) de l’algèbre An , il est immédiat
de vérifier que :
TFDn,ξ (AB) = TFDn,ξ (A)
TFDn,ξ (B) .
En tant qu’application linéaire, TFDn,ξ est représentée dans les
bases canoniques par la matrice de Vandermonde particulière :
1
1
1
...
1
ξ
ξ2
. . . ξ n−1
1
2
4
1
ξ
ξ
. . . ξ 2(n−1) .
Wn,ξ =
.
..
..
..
..
..
.
.
.
.
2
1 ξ n−1 ξ 2(n−1) . . . ξ (n−1)
Si, de plus, n 1A est inversible dans l’anneau A (on désignera par n−1
son inverse), alors la matrice Wn,ξ est inversible dans An×n et on vérifie
qu’elle admet pour inverse la matrice
1
1
1
...
1
−1
−2
1−n
ξ
...
ξ
1 ξ
−2
−4
2(1−n)
−1
−1 1
= n−1 Wn,ξ−1 .
ξ
ξ
... ξ
Wn,ξ = n
.
.
.
.
..
..
..
..
..
.
2
1 ξ 1−n ξ 2(1−n) . . . ξ −(n−1)
6.3. Transformation de Fourier discrète rapide
177
Dans ce cas, modulo l’identification précédente, l’application TFDn,ξ est
un isomorphisme d’algèbres TFDn,ξ : A[X] /hX n − 1i −→ An . Nous
énonçons ce résultat.
Proposition 6.2.2 Supposons que l’anneau commutatif A possède une
racine n - ème principale de 1, notée ξ, et que n 1A est inversible dans
A. Alors la transformation de Fourier discrète TFDn,ξ : An −→ An
−1
est un isomorphisme de A – modules, et TFD−1
n,ξ = (n1A ) TFDn,ξ −1 .
n
Par ailleurs, si on identifie le A – module A source de l’application
linéaire TFDn,ξ avec A[X] /hX n − 1i (en choisissant le représentant
de degré ≤ n et en l’exprimant sur la base des monômes) alors TFDn,ξ
définit un isomorphisme de l’algèbre A[X] /hX n − 1i (munie de la multiplication des polynômes) vers l’algèbre An (munie de la multiplication
coordonnée par coordonnée). En bref, pour deux polynômes de degré
< n, on a :
−1
AB ≡ T F Dn,ξ
(T F Dn,ξ (A)
T F Dn,ξ (B)) modulo (X n − 1).
C’est la clé de l’algorithme de multiplication rapide, que nous explicitons
dans la section suivante.
6.3
6.3.1
Transformation de Fourier discrète rapide
Cas favorable
Le résultat énoncé dans la proposition
P2n−16.2.2 kprécédente peut être
de deux polynômes
appliqué
au
calcul
du
produit
AB
=
k=0 ck X
Pn−1
Pn−1
i
i
A = i=0 ai X et B = i=0 bi X à une indéterminée sur A, à condition que l’anneau A s’y prête. Nous supposons qu’il possède une racine
2n - ème principale ω de 1, et que (2n) 1A est inversible dans A, alors
la proposition 6.2.2 pour la TFD d’ordre 2n sur l’anneau A se traduit
par :
−1
AB = T F D2n,ω
(T F D2n,ω (A) T F D2n,ω (B)) .
car le calcul de AB modulo X 2n − 1 donne exactement AB. Le calcul
du produit AB de deux polynômes de degrés inférieur ou égal à n par
la TFD est résumé dans l’algorithme 6.2 page suivante.
Le lemme suivant nous permet tout d’abord de montrer comment
une TFD d’ordre 2ν peut être effectuée rapidement au moyen d’une
stratégie (( diviser pour gagner )).
178
6. Multiplication rapide des polynômes
Algorithme 6.2 Multiplication des polynômes via la Transformation de Fourier Discrète.
Entrée : Deux polynômes A et B de degrés < n sur un anneau A convenable (voir proposition 6.2.2).
Sortie : Le produit A B.
Début
Étape 1 :
Deux TFD d’ordre 2n appliquées à A et B.
Étape 2 :
Évaluation de 2n multiplications dans A pour obtenir la transformée
de Fourier discrète de A B.
Étape 3 :
Calcul de l’inverse d’une TFD d’ordre 2n pour obtenir A B.
Fin.
Lemme 6.3.1 Soit n un entier ≥ 2 et ν = dlog ne. La transformation
de Fourier discrète d’ordre n et son inverse, dans un anneau possédant
une racine 2ν - ème principale de 1 et dans lequel 2A est inversible,
se font en SD(n log n, log n). Plus précisément, la taille S(n) et la profondeur D(n) du circuit arithmétique correspondant sont respectivement
majorées par n (3 log n+3) et 2 log n+2 pour la transformation directe
et par n (3 log n + 4) et 2 log n + 3 pour la transformation inverse.
Pn−1
ai X i un polynôme de degré ≤ n − 1 à coefPreuve. Soit A = i=0
ficients dans A, ν = dlog ne (de sorte que 2ν−1 < n ≤ 2ν ) et ω une
racine 2ν - ème principale de 1. Il s’agit de calculer les valeurs de A aux
ν
points 1, ω, ω 2 , . . . , ω 2 −1 . Le polynôme A peut être mis sous la forme
A = A1 (X 2 ) + XA2 (X 2 ) avec deg A1 , deg A2 ≤ 2ν−1 − 1.
Remarquons que ξ = ω 2 est une racine 2ν−1 - ème principale de 1,
ν−1
que ω 2
= −1 et que A (ω i ) = A1 (ξ i )+ω i A2 (ξ i ) pour 0 ≤ i ≤ 2ν−1 −
ν−1
ν−1
1. Comme ω 2 +i = −ω i , on a aussi A (ω 2 +i ) = A1 (ξ i ) − ω i A2 (ξ i )
pour 0 ≤ i ≤ 2ν−1 − 1.
Ce qui donne toutes les valeurs recherchées de A et ramène récursivement l’évaluation de A en les 2ν points ω i (0 ≤ i ≤ 2ν − 1), c’est-à-dire
la TFD d’ordre 2ν , au calcul suivant :
• deux TFD d’ordre 2ν−1 appliquées à A1 et A2 et effectuées en
parallèle ;
6.3. Transformation de Fourier discrète rapide
179
• 2ν−1 multiplications (par les ± ω i avec 0 ≤ i ≤ 2ν−1 − 1) effectuées en parallèle et en une seule étape de calcul, suivies de 2ν additions dans l’anneau de base A effectuées également en une seule étape
parallèle.
Si S et D désignent respectivement la taille et la profondeur de l’algorithme récursif ainsi défini, on obtient les relations suivantes valables
pour tout entier ν ≥ 1 :
S(2ν ) ≤ 2 S(2ν−1 ) + 3 2ν−1
D(2ν ) ≤ D(2ν−1 ) + 2 .
Ce qui donne, par sommation, sachant que S(1) = D(1) = 0 :
S(2ν ) ≤ 3 ν 2ν−1
D(2ν ) ≤ 2 ν .
Comme 2ν−1 < n ≤ 2ν et par conséquent ν − 1 < log n, on en déduit
que S(n) < 3 n (1 + log n) et que D(n) < 2 (1 + log n).
−1 =
Pour la TFD inverse d’ordre n, nous avons vu que T F Dn,ω
(n 1A )−1 T F Dn,ω−1 . Cela signifie que l’on peut récupérer les coefficients
~ = (A(1),
du polynôme A de degré ≤ n − 1, à partir du vecteur A
n−1
A(ω), . . . , A(ω
)) formé des valeurs de ce polynôme aux points ω i ,
~ la TFD d’ordre n associée à la racine prinen effectuant sur le vecteur A
−1
n−1
cipale ω = ω
et en multipliant ensuite ce vecteur par (n 1A )−1 .
Par conséquent, la TFD inverse d’ordre n peut se faire par un circuit
arithmétique de taille S(n) + n et de profondeur D(n) + 1.
t
u
Ce résultat et l’algorithme 6.2 qui a introduit le lemme 6.3.1 nous
permettent d’estimer avec précision la complexité de l’algorithme de la
multiplication rapide des polynômes et d’énoncer le théorème suivant.
Théorème 6.1 On considère un anneau A possédant une racine 2ν+1 ème principale de 1 et dans lequel 2A est inversible.
Alors, en utilisant l’algorithme 6.2 avec l’évaluation récursive décrite
dans la preuve du lemme 6.3.1, la multiplication de deux polynômes de
degrés < n ≤ 2ν à coefficients dans A se fait à l’aide d’un circuit arithmétique de taille ≤ n (18 log n + 44) et de profondeur ≤ 4 log n + 10.
Preuve. Supposons d’abord n = 2ν . On exécute en parallèle deux TFD
d’ordre 2n suivies d’une étape parallèle avec 2n multiplications dans
l’anneau de base, et on termine par une transformation inverse d’ordre
180
6. Multiplication rapide des polynômes
2n. La preuve du lemme 6.3.1 donne la majoration de la taille par 9(ν +
1) 2ν +4n = 9 n log n+13 n et de la profondeur par 4 ν +6 = 4 log n+6.
Dans le cas général, il faut remplacer n par 2n et log n par 1 + log n.
t
u
Rappelons que pour un anneau A fixé par le contexte, nous notons
µP (n) le nombre d’opérations arithmétiques nécessaires pour la multiplication de deux polynômes de degré n en profondeur O(log n). Le
théorème précédent nous dit donc qu’on a µP (n) = O(n log n) si 2A
est inversible et si l’anneau possède des racines 2ν - èmes principales de
l’unité pour tout ν.
6.3.2
Algorithme de la TFD rapide pour un anneau commutatif arbitraire
L’algorithme que nous venons de développer n’est pas valable lorsque
2A divise zéro dans l’anneau A (puisque, dans un tel anneau, la division
par 2 ne peut pas être définie de manière unique, même lorsqu’elle est
possible). On peut essayer de contourner cette difficulté en remplaçant 2
par un entier s ≥ 2 tel que s 1A ne divise pas zéro dans A. Lorsqu’un tel
entier s ≥ 2 existe, et à supposer qu’on dispose d’une racine principale
s - ème de 1 dans A, il faut encore disposer d’un algorithme performant
pour la division par s (quand elle est possible) pour pouvoir effectuer
la transformation de Fourier inverse. En outre, un tel entier s n’existe
pas nécessairement.
Pour se débarrasser radicalement de ce problème, l’idée de CantorKaltofen dans [13] est de calculer séparément uAB et vAB avec deux
entiers u et v premiers entre eux, puis de récupérer AB en utilisant
une relation de Bezout entre u et v. Par exemple, on prend u = 2ν ≥ 2n
et v = 3µ ≥ 2n. On calcule sans aucune division 2ν AB par la formule
2ν AB = T F D2ν ,ω−1 (T F D2ν ,ω2,ν (A)
2,ν
T F D2ν ,ω2,ν (B)) .
(où ω2,ν est une racine 2ν - ème principale de 1). De même, on calcule
3µ AB par la formule
3µ AB = T F D3µ ,ω−1 (T F D3µ ,ω3,µ (A)
3,µ
T F D3µ ,ω3,µ (B))
(où ω3,µ est une racine 3µ - ème principale de 1).
Il reste néanmoins un obstacle de taille, qui consiste en la nécessité
de rajouter un substitut formel à ω2,ν (et ω3,µ ) lorsqu’on ne les a pas
6.3. Transformation de Fourier discrète rapide
181
sous la main dans l’anneau A. Or l’idée toute simple de faire les calculs dans l’anneau A[λ2,ν ], où λ2,ν est un susbstitut formel de ω2,ν
ne donne pas le résultat souhaité. En effet, une opération arithmétique
dans l’anneau A [λ2,ν ] correspond a priori à grosso modo n opérations
arithmétiques dans A, ce qui annule le bénéfice de la transformation de
Fourier discrète.
L’idée de Cantor et Kaltofen pour résoudre ce deuxième problème
est d’appliquer une stratégie (( diviser pour gagner )), un peu semblable
à celle du lemme 6.3.1.
La définition précise de l’anneau A [λ2,ν ] et la description de l’algorithme font appel aux polynômes cyclotomiques, dont nous rappelons
maintenant quelques propriétés.
Le n - ème polynôme cyclotomique est défini à partir d’une racine n ème primitive de 1, c’est-à-dire un générateur ωn du groupe multiplicatif
(cyclique) des racines n - èmes de 1 dans une clôture algébrique de Q,
par exemple dans C avec ωn = ei 2π/n .
Le n - ème polynôme cyclotomique est, par définition, le polynôme
Φn (X) =
Y
X − ωnh .
1≤h<n
(h, n) = 1
C’est un polynôme unitaire à coefficients entiers dont les zéros sont les
racines n - èmes primitives de 1 et dont le degré est égal à ϕ(n). C’est
aussi un polynôme réciproque : X ϕ (n) Φn (1/X) = Φn (X). Les polynômes cyclotomiques possèdent en outre les propriétés suivantes :
Q
• Φn (X) = d|n Φd (X) ;
(d|n signifie que d est un diviseur positif de n)
• Φp (X) = X p−1 + · · · + X + 1 pour tout nombre premier p ;
• Φmsk (X) = Φms (X s
k−1
) si k ≥ 2 ;
• Φm (X) Φmp (X) = Φm (X p ) si p premier ne divise pas m ;
• Φ2n (X) = Φn (−X) si n est impair ≥ 3.
On en déduit, en particulier, que :
p si n est une puissance d’un nombre premier p
• Φn (1) =
1 sinon.
Rajouter formellement une racine primitive sq - ème de 1 dans A revient à considérer l’anneau A[Y ] /hΦsq (Y )i = A[λs,q ]. Dans cet anneau,
182
6. Multiplication rapide des polynômes
une addition équivaut à ϕ(sq ) additions dans A. Pour une multiplicaq
tion, on peut travailler dans A[Y ] modulo (Y s − 1) puis réduire le
résultat obtenu modulo Φsq (Y ). Cette dernière opération est relativeq−1
ment peu coûteuse car Φsq (Y ) = Φs (Y s ) est un polynôme unitaire
qui a très peu de coefficents non nuls. Cette remarque permet de voir
que les multiplications dans A[λs,q ] ne sont pas tellement plus coûteuses
que les additions. Elle donne une idée de comment pourra être appliquée
une stratégie diviser pour gagner, de manière à rendre peu coûteux les
calculs dans l’anneau A[λs,q ]. L’algorithme de Cantor-Kaltofen donne
alors le résultat suivant :
Théorème 6.2 Il existe une famille uniforme de circuits arithmétiques de profondeur O(log n) qui calculent le produit de deux polynômes
de degré < n à coefficients dans un anneau commutatif arbitraire A
avec O(n log n) multiplications et µP (n) = O(n log n log log n) additions/soustractions.
Remarque 6.3.2 L’algorithme de Cantor-Kaltofen prend en entrée
deux polynômes A et B de degré < n et donne en sortie C = AB.
Il calcule tout d’abord sq11 C et sq22 C, où s1 et s2 sont deux petits
entiers premiers entre eux, et sq11 et sq22 ne sont pas trop grands par
rapport à n. La constante cachée du (( grand O )) dans l’estimation
O(n log n log log n) de la taille du circuit calculant sq C est de l’ordre
de 4s2 (3s + 1) si s est premier. Il s’ensuit qu’en utilisant les deux
valeurs optimales s1 = 2 et s2 = 3, l’algorithme de Cantor-Kaltofen
ne devient plus performant que l’algorithme en O(nlog 3 ) que pour les
valeurs de n qui sont de l’ordre de 6 104 .
Remarque 6.3.3 La multiplication rapide des polynômes est en fait
couramment utilisée en analyse numérique, en prenant des approximations numériques des racines de l’unité dans C. Cela laisse supposer
qu’une implémentation efficace de cette multiplication rapide est également possible en calcul formel avec des anneaux tels que Z ou un anneau
de polynômes sur Z. Il suffit en effet de faire le calcul numérique approché avec une précision suffisante pour que le résultat du calcul soit
garanti avec une précision meilleure que 1/2. Une autre solution voisine, mais où la précision est plus facile à contrôler, serait de faire un
calcul numérique approché non dans C mais dans un anneau d’entiers
p - adiques (voir par exemple [Ser]) : un tel anneau contient une racine
primitive (p − 1) - ème de l’unité, et (p − 1) y est inversible.
6.4. Produits de matrices de Toeplitz
6.4
183
Produits de matrices de Toeplitz
Nous signalons ici une interprétation matricielle du produit de deux
polynômes A et B de degrés m et n. On considère le A – module
libre Pm+n+1 ' Am+n+1 des polynômes de degré ≤ m + n muni de la
base canonique des monômes X k . La multiplication par A (resp. B,
resp. AB) tronquée au degré m + n est représentée sur cette base par
une matrice de Toeplitz triangulaire TA (resp. TB , resp. TAB ) et on a
TA TB = TAB . Par exemple avec m = 3, n = 2 on obtient le produit
a0 0 0 0 0 0
b0 0 0 0 0 0
a1 a0 0 0 0 0
b1 b0 0 0 0 0
a2 a1 a0 0 0 0
× b2 b1 b0 0 0 0
a3 a2 a1 a0 0 0
0 b2 b1 b0 0 0
0 a3 a2 a1 a0 0
0 0 b2 b1 b0 0
0
0
a3 a2 a1 a0
0
0
0
b2 b1 b0
qui est égal à la matrice de Toeplitz triangulaire inférieure dont la
première colonne est donnée par les coefficients du produit AB :
a0 0 0 0 0 0
b0
a0 b0
a1 a0 0 0 0 0
b1
a1 b0 + a0 b1
a2 a1 a0 0 0 0
b2
a2 b0 + a1 b1 + a0 b2
×
=
a3 a2 a1 a0 0 0
0
a3 b0 + a2 b1 + a1 b2
0 a3 a2 a1 a0 0
0
a3 b1 + a2 b2
0 0 a3 a2 a1 a0
0
a3 b2
Inversement, le produit de deux matrices de Toeplitz triangulaires
inférieures dans An×n peut s’interpréter comme le produit de deux polynômes de degrés ≤ n − 1, tronqué au degré n − 1 (c’est-à-dire encore
comme le produit dans l’anneau des développements limités An−1 =
A[X] /hX n i ). Par exemple
a0 0 0 0
b0
a0 b0
a1 a0 0 0
b1
a1 b0 + a0 b1
.
a2 a1 a0 0 × b2 =
a2 b0 + a1 b1 + a0 b2
b3
a3 b0 + a2 b1 + a1 b2 + a0 b3
a3 a2 a1 a0
En bref il n’y a pas de différence significative entre le produit de 2
polynômes, le produit de 2 matrices de Toeplitz triangulaires inférieures
carrées et le produit d’une matrice de Toeplitz triangulaire inférieure
par un vecteur.
184
6. Multiplication rapide des polynômes
Voyons maintenant la question du produit d’une matrice de Toeplitz
arbitraire par un vecteur. Par exemple
c
a3 a2 a1 a0
3
a4 a3 a2 a1
c4
b0
a5 a4 a3 a2
b1
= c5 .
×
c6
a6 a5 a4 a3
b2
c7
a7 a6 a5 a4
b3
c8
a8 a7 a6 a5
Il suffit d’insérer la première
matrice dans la matrice de la multiplication
P
par le polynôme A = 8i=0 ai X i , tronquée au degré 11, dans le A – module libre des polynômes de degrés ≤ 11 :
a0 0 0 0
a0 b0
a a
a b +a b
0 1
1 0 0 0
1 0
a a a
0
c
2
2 1 0
a a a a
c
3
3 2 1 0
a a a a
b0
c4
4 3 2 1
b1
c5
a5 a4 a3 a2
=
×
.
b2
c6
a6 a5 a4 a3
a7 a6 a5 a4
b3
c7
a8 a7 a6 a5
c8
0 a8 a7 a6
c9
0 0 a8 a7
a7 b3 + a8 b2
0 0 0 a8
a8 b3
On voit alors quePle calcul se ramène au produit du polynôme A par le
polynôme B = 3i=0 bi X i . On en déduit le résultat important suivant
où l’on voit que le produit par une matrice de Toeplitz n’est guère plus
cher que le produit par une matrice creuse.
Proposition 6.4.1 Le produit d’une matrice de Toeplitz et d’une matrice arbitraire, toutes deux carrées d’ordre n peut se faire par une famille de circuits arithmétiques en SD(n µP (n), log n)).
Remarque. Plus précisément supposons que dans l’anneau commutatif
A la multiplication d’un polynôme de degré ≤ n par un polynôme de
degré ≤ m soit en SD(µ(n, m), λ(n, m)). Alors le produit T B d’une
matrice de Toeplitz T ∈ An×m par une matrice B ∈ Am×p est en
SD(p µ(n + m, m), λ(n + m, m)). Ceci n’est qu’un exemple des résultats
de complexité arithmétique concernant les matrices de Toeplitz. Nous
renvoyons le lecteur intéressé par le sujet à l’ouvrage [BP].
7. Multiplication rapide
des matrices
Introduction
La multiplication des matrices à coefficients dans un anneau commutatif unitaire A a fait l’objet de multiples investigations durant les trente
dernières années en vue de réduire le nombre d’opérations arithmétiques
(dans A) nécessaires au calcul du produit d’une matrice m × n par une
matrice n × p, et d’améliorer la borne supérieure asymptotique de ce
nombre. Il s’est avéré que c’est le nombre de multiplications essentielles
qui contrôle la complexité asymptotique de la multiplication des matrices carrées, comme nous allons le voir tout d’abord à travers l’algorithme
de la multiplication rapide de Strassen.
L’algorithme conventionnel (dit usuel) pour le calcul du produit C =
(cij ) ∈ Am×p d’une matrice A = (aij ) ∈ Am×n par une matrice B =
(bij ) ∈ An×p se fait par mnp multiplications et mp (n − 1) additions en
calculant en parallèle (en une seule étape) les mnp produits aik bkj et
en calculant ensuite, en parallèle et en dlog ne étapes, les mp sommes
cij intervenant dans les formules
cij =
Xn
k=1
aik bkj pour 1 ≤ i ≤ m et 1 ≤ j ≤ p .
En particulier pour la multiplication de deux matrices carrées d’ordre
n, cet algorithme correspond à un circuit arithmétique de taille n2 (2n −
1) et de profondeur dlog ne + 1 avec n3 multiplications et n2 (n − 1)
additions.
Dans un premier temps, les investigations portaient sur la diminution
du nombre de multiplications en essayant d’y réduire le coefficient de
n3 sans s’occuper de l’exposant de n, et c’est Winograd qui réussit le
premier à réduire ce coefficient de moitié, mais en doublant presque le
186
7. Multiplication rapide des matrices
nombre d’additions, ce qui constitue, malgré ce prix, un progrès dans la
complexité asymptotique si l’on sait que dans une large classe d’anneaux
la multiplication est beaucoup plus coûteuse que l’addition 1 . Beaucoup
pensaient que ce résultat de Winograd serait optimal au sens que 12 n3
multiplications seraient nécessaires pour le calcul du produit de deux
matrices n × n (voir [Knu], page 481).
Mais une année plus tard (1969), Strassen montra que l’on pouvait
multiplier deux matrices n × n en utilisant seulement O(n2,8 ) multiplications. Ce résultat était basé sur le fait très simple que le produit de
deux matrices 2 × 2 à coefficients dans un anneau non nécessairement
commutatif pouvait être calculé avec seulement 7 multiplications au lieu
de 8, le nombre d’additions passant de 4 à 18, et il donna les relations
prouvant ce fait dans son fameux article Gaussian elimination is not
optimal [86]. Winograd donna un peu plus tard [97] une variante de la
multiplication rapide de Strassen avec seulement 15 additions.
Comme ces relations n’utilisent pas la commutativité de la multiplication, elles s’appliquent récursivement au calcul du produit de deux
matrices quelconques à coefficients dans A selon la stratégie (( diviser
pour gagner )).
La section 7.1 est consacrée à une analyse détaillée de la multiplication rapide des matrices dans la version Strassen-Winograd. Nous
étudions également l’uniformité de la construction de la famille de circuits arithmétiques qui correspond à la version originale de Strassen,
comme annoncé dans la section 4.4.
Dans la section 7.2 nous montrons que l’inversion des matrices triangulaires fortement régulières peut être réalisée par des circuits arithmétiques avec une taille de même ordre que les circuits de la multiplication
des matrices carrées et une profondeur d’ordre O(log2 n) au lieu de
O(log n) .
Dans la section 7.3 nous introduisons les notions de complexité bilinéaire, de complexité multiplicative et de rang tensoriel. Nous montrons
le rôle central joué par la notion de rang tensoriel dans la complexité
asymptotique de la multiplication des matrices carrées (théorème 7.4
dû à Strassen). Nous montrons également le résultat de Schönhage, qui
dit que l’exposant de la multiplication des matrices carrées ne dépend
que de la caractéristique du corps de base (on conjecture en fait que cet
1. Ce qui n’est pas vrai par exemple dans le corps des fractions rationnelles Q (X)
où l’addition est plus coûteuse.
7.1. Analyse de la méthode de Strassen
187
exposant est le même pour tous les corps et pour l’anneau des entiers
relatifs.)
Dans la section 7.4 nous nous attaquons à des algorithmes nettement
plus sophistiqués qui s’appuient sur la notion de calcul bilinéaire approximatif, introduite par Bini. Malgré leurs performances asymptotiques,
aucun des algorithmes de cette section ne semble devoir être implémenté
sur machine dans un proche avenir. Il nous a pourtant semblé que ce serait un crime contre la beauté que de ne pas dévoiler au moins en partie
les idées fascinantes qui y sont à l’œuvre. Nous n’avons cependant pas
exposé la (( méthode du laser )) due à Strassen (cf. [BCS, 90]), car nous
n’avons pas vu comment en donner une idée assez exacte en termes suffisamment simples. Cette méthode a conduit à la meilleure borne connue
pour l’exposant de la multiplication des matrices carrées. L’estimation
actuelle de cet exposant ω est de 2, 376 : Winograd & Coppersmith,
1987 ([19, 20]).
7.1
7.1.1
Analyse de la méthode de Strassen
La méthode de Strassen (version Winograd) et sa
complexité
On considère dans un anneau B (non nécessairement commutatif)
deux matrices A et B :
a11 a12
b11 b12
c11 c12
A=
B=
avec C = AB =
.
a21 a22
b21 b22
c21 c22
Alors la matrice C peut être obtenue par le calcul suivant :
m1 := a11 b11
m3 := (a11 − a21 ) (b22 − b12 )
m5 := (a21 + a22 − a11 ) (b22 − b12 + b11 )
m6 := (a11 + a12 − a21 − a22 ) b22
m7 := a22 (b22 − b12 + b11 − b21 )
c11 := m1 + m2
c21 := m1 + m3 + m5 − m7
m2 := a12 b21
m4 := (a21 + a22 ) (b12 − b11 )
c12 := m1 + m5 + m4 + m6
c22 := m1 + m3 + m4 + m5
Ces relations de Strassen (version Winograd), appliquées à l’anneau
des matrices carrées d’ordre 2k, ramènent le calcul du produit de deux
188
7. Multiplication rapide des matrices
matrices 2k × 2k (k ∈ N∗ ) à celui de sept produits de matrices k × k
et de 15 sommes de matrices de même type.
L’analyse de complexité faite à la section 5.1 montre que ce passage
de 8 à 7 multiplications est un avantage décisif, indépendamment du
nombre des additions utilisées par ailleurs. Cela tient à ce que 7 est le
degré de parallélisme dans la procédure (( diviser pour gagner )) tandis
que le nombre d’additions n’intervient que dans la constante du O(n2 )
opérations arithmétiques nécessaires pour, partant du problème initial
Pn , d’une part créer les 7 sous-problèmes de type Pdn/2e , et d’autre part
récupérer la solution du problème initial à partir des solutions, calculées
en parallèle, des 7 sous-problèmes (cf. proposition 5.1.1 page 161).
Posant :
A11 A12
B11 B12
C11 C12
=
A21 A22
B21 B22
C21 C22
où les Aij , Bij , Cij (1 ≤ i, j ≤ 2) sont des matrices k × k, on a un
schéma de programme d’évaluation comportant les instructions suivantes dans lesquelles les affectations des variables Mi (1 ≤ i ≤ 7) correspondent aux 7 multiplications et celles des variables Ni (1 ≤ i ≤ 11)
et Cij (1 ≤ i, j ≤ 2) correspondent aux 15 additions/soustractions( 2 ),
avec indication des étapes du calcul parallèle :
Appliqué récursivement à une matrice m2ν × m2ν (m ∈ N∗ , ν ∈ N)
ce programme donne un circuit arithmétique parallèle de taille S(m2ν )
et de profondeur D(m2ν ) dans l’anneau A, vérifiant les relations de
récurrence 3 :
S(m2ν ) = 7 S(m2ν−1 ) + 15 · m2 4ν−1
(7.1)
D(m2ν ) = D(m2ν−1 ) + 6 ·
La dernière équation est justifiée par le fait que les étapes où il n’y
a que des additions de matrices m2ν−1 × m2ν−1 ont une profondeur
égale à 1 (les m2 4ν−1 additions correspondantes dans A se faisant en
parallèle) alors que l’étape comprenant les multiplications de matrices
(Etape 4) est de profondeur D(m2ν−1 ).
Utilisant l’algorithme usuel pour la multiplication de deux matrices
m × m, on peut écrire S(m) = m2 (2m − 1) et D(m) = dlog me + 1.
2. Dans la suite, nous dirons simplement additions, en sous-entendant additions/soustractions.
3. Signalons que pour la version originale de Strassen avec 18 additions (cf. page
191), la profondeur vérifie la relation D(m2ν ) = D(m2ν−1 ) + 3.
7.1. Analyse de la méthode de Strassen
189
Algorithme 7.1 Multiplication de matrices par blocs, à la
Strassen-Winograd
Début
Étape 1 :
N1 := A11 − A21 ; N2 := A21 + A22 ;
N3 := B12 − B11 ; N4 := B22 − B12
Étape 2 :
N5 := N2 − A11 ; N6 := B22 − N3
Étape 3 :
N7 := A12 − N5 ; N8 := N6 − B21
Étape 4 : Les 7 multiplications
M1 := A11 B11 ; M2 := A12 B21 ; M3 := N1 N4 ; M4 := N2 N3 ;
M5 := N5 N6 ; M6 := N7 B22 ; M7 := A22 N8
Étape 5 :
C11 := M1 + M2 ; N9 := M1 + M5 ; N10 := M4 + M6
Étape 6 :
N11 := M3 + N9 ; C12 := N9 + N10
Étape 7 :
C21 := N11 − M7 ; C22 := M4 + N11
Fin.
Ce qui donne D(n) = D(m2ν ) = D(m)+6ν = 6 dlog ne+dlog me+1
comme résultat pour la profondeur du circuit arithmétique correspondant au calcul du produit de deux matrices n × n si l’on prend n = m2ν
(la version originale de Strassen donne D(n) = 3 dlog ne + dlog me + 1).
Concernant la taille, la première équation dans (7.1) donne successivement :
1×
7×
..
.
S(2ν m) =
S(2ν−1 m) =
..
.
7 S(2ν−1 m)
7 S(2ν−2 m)
7ν−1 ×
7ν ×
S(2m) =
S(m) =
7 S(m)
2
m (2m − 1)
−→
S(2ν m) =
7ν m2 (2m − 1)
+ 15 · 4ν−1 m2
+ 15 · 4ν−2 m2
..
.
+ 15 · m2
+ 5 m2 (7ν − 4ν )
190
7. Multiplication rapide des matrices
Ce qui donne comme résultat S(m2ν ) = 7ν m2 (2m + 4) − 5m2 4ν
pour la taille du circuit arithmétique correspondant au calcul par la méthode de Strassen (variante Winograd) 4 du produit de deux matrices
m2ν × m2ν . Ainsi :
S(m2ν ) = 2m2 (m + 2) 7ν − 5m2 4ν
(7.2)
D(m2ν ) = 6 dlog ne + dlog me + 1
En particulier, si n est une puissance de 2 (c’est-à-dire m = 1), et
comme 7ν = 2ν log 7 :
S(n) = 6 nlog 7 − 5 n2 et D(n) = 6 dlog ne
(on obtient 3 dlog ne+1 seulement pour la version originale de Strassen).
Mais le coefficient de nlog 7 ' n2.807 ( 5 ) dans S(n) peut être ramené
à 4, 15 lorsque n est une puissance de 2. En effet, si n = 32 on peut
vérifier directement que le nombre d’opérations arithmétiques dans la
multiplication usuelle des matrices n2 (2n−1) ne dépasse guère 3,9 nlog 7
et pour n ≥ 32, on pose log n = ν +5 ≥ 5 , de sorte que n = 32. 2ν (m =
32) .
La première des équations (7.2) donne alors :
S(n) =
<
≤
≤
≤
S(m2ν )
2m2 (m + 2) 7ν
211 ∗ 34 ∗ 7ν
211 ∗ 34 ∗ (1/7)5 ∗ 7log n
4, 15 nlog 7 .
(puisque 7ν = 7log n−5 )
(on remplace 7log n par nlog 7 )
Ceci conduit donc au résultat suivant dû à Strassen, mais dans lequel
nous intégrons la version (avec 15 additions) de Winograd :
Théorème 7.1 La multiplication de deux matrices n × n à coefficients
dans un anneau arbitraire A est dans la classe SD (nlog 7 , log n). Plus
précisément, lorsque n est une puissance de 2, elle se fait soit avec un
circuit arithmétique dont la taille et la profondeur sont respectivement
majorées par 4, 15 nlog 7 et 6 dlog ne, soit par un circuit dont la taille et
la profondeur sont respectivement majorées par 4, 61 nlog 7 et 3 dlog ne.
4. La version originale de Strassen donne S(m2ν ) = 7ν m2 (2m + 5) − 6m2 4ν .
5. log 7 ' 2.8073549220576041074.
7.1. Analyse de la méthode de Strassen
191
Notez aussi que la profondeur multiplicative de ces circuits est égale
à 1. En fait la conclusion dans le théorème précédent est non seulement qu’il existe une famille de circuits arithmétiques dans la classe
SD(nlog 7 , log n) qui réalise la multiplication des matrices carrées, mais
qu’on sait construire explicitement une famille uniforme de tels circuits
arithmétiques. Ceci est l’objet du paragaphe qui suit avec le théorème 7.2.
7.1.2
Un exemple de construction uniforme d’une famille
de circuits arithmétiques
Nous allons maintenant tenir une promesse que nous avions faite
dans la section 4.4. Celle d’analyser un exemple de construction récursive
uniforme typique d’une famille de circuits arithmétiques pour laquelle le
coût de production d’un circuit de la famille n’a pas un ordre de grandeur
bien supérieur à sa taille. Nous utiliserons pour cet exemple la multiplication rapide des matrices originale de Strassen [86] qui repose sur le
calcul suivant. On considère dans un anneau B (non nécessairement
commutatif) deux matrices A et B :
a11 a12
b11 b12
c11 c12
A=
B=
avec C = AB =
.
a21 a22
b21 b22
c21 c22
Alors la matrice C peut être obtenue par le calcul suivant, qui nécessite
18 additions/soustractions et 7 multiplications :
m1 := (a12 − a22 ) (b21 + b22 )
m3 := (a11 − a21 ) (b11 + b12 )
m5 := a11 (b12 − b22 )
m7 := (a21 + a22 ) b11
c11 := m1 + m2 − m4 + m6
c21 := m6 + m7
m2 := (a11 + a22 ) (b11 + b22 )
m4 := (a11 + a12 ) b22
m6 := a22 (b21 − b11 )
c12 := m4 + m5
c22 := m2 − m3 + m5 − m7
Ceci peut être réécrit sous forme d’un circuit arithmétique de profondeur 4. Concernant les variables en entrée, on note x0,i,j pour aij
et x0,2+i,2+j pour bij . On obtient le programme d’évaluation 7.2 page
suivante, que nous appelons P1 .
La méthode de Strassen consiste à utiliser ces formules de manière
récursive. Si on doit multiplier des matrices carrées à m = 2n lignes
et colonnes, on les partitionne chacune en 4 matrices carrées à 2n−1
192
7. Multiplication rapide des matrices
Programme d’évaluation 7.2 P1 : produit de deux matrices carrées d’ordre 2 sur un anneau non nécessairement commutatif,
à la Strassen.
Entrée : Les 8 coefficients x0,ij dans A (un anneau arbitraire) de deux
matrices carrées A et B d’ordre 2.
Sortie : Les coefficients x4,ij du produit : C = A B.
Début
profondeur 1 :
x1,1 := x0,12 − x0,22 ; x1,2 := x0,11 + x0,22 ;
x1,3 := x0,11 − x0,21 ; x1,4 := x0,11 + x0,12 ;
x1,5 := x0,11 ; x1,6 := x0,22 ; x1,7 := x0,21 + x0,22 ;
x1,8 := x0,43 + x0,44 ; x1,9 := x0,33 + x0,44 ;
x1,10 := x0,33 + x0,34 ; x1,11 := x0,44 ;
x1,12 := x0,34 − x0,44 ; x1,13 := x0,43 − x0,33 ; x1,14 := x0,33
profondeur 2 : Les 7 multiplications
x2,1 := x1,1 x1,8 ; x2,2 := x1,2 x1,9 ; x2,3 := x1,3 x1,10 ;
x2,4 := x1,4 x1,11 ; x2,5 := x1,5 x1,12 ; x2,6 := x1,6 x1,13 ;
x2,7 := x1,7 x1,14
profondeur 3 :
x3,1 := x2,1 + x2,2 ; x3,2 := x2,4 − x2,6 ;
x3,3 := x2,2 − x2,3 ; x3,4 := x2,5 − x2,7
profondeur 4 :
x4,11 := x3,1 − x3,2 ; x4,12 := x2,4 + x2,5 ;
x4,21 := x2,6 + x2,7 ; x4,22 := x3,3 + x3,4
Fin.
lignes et colonnes, qui jouent le rôle des aij et bij dans les formules
précédentes. On obtient en définitive un circuit arithmétique de profondeur 3n + 1 = 3 log(m) + 1 comportant 6m2 additions/soustractions et
7n = mlog(7) multiplications (la méthode usuelle donne un circuit de profondeur 1 + n comportant m3 additions/soustractions et 8n = m3 multiplications). Notre problème est de déterminer la complexité en temps
pour l’écriture du programme d’évaluation correspondant.
Supposons qu’on ait écrit le programme d’évaluation Pn pour la
multiplication de deux matrices carrées à m = 2n lignes et colonnes,
avec les entrées x0,ij et x0,2n +i,2n +j avec 1 ≤ i, j ≤ 2n , et les sorties
x3n+1,ij (1 ≤ i, j ≤ 2n ).
7.1. Analyse de la méthode de Strassen
193
Comment écrit-on le programme d’évaluation Pn+1 ?
Les entrées sont maintenant x0,i,j et x0,2n+1 +i,2n+1 +j avec 1 ≤ i, j ≤
2n+1 . Notons X0,uv (1 ≤ u, v ≤ 2 ou 3 ≤ u, v ≤ 4) les matrices extraites
à 2n lignes et colonnes, avec m = 2n et m2 = 2n+1 :
X0,11 [i, j] := x0,i,j
X0,21 [i, j] := x0,m+i,j
X0,33 [i, j] := x0,m2 +i,m2 +j
X0,43 [i, j] := x0,m2 +m+i,m2 +j
X0,12 [i, j] := x0,i,m+j
X0,22 [i, j] := x0,m+i,m+j
X0,34 [i, j] := x0,m2 +i,m2 +m+j
X0,44 [i, j] := x0,m2 +m+i,m2 +m+j
On commence par créer (conformément au programme P1 appliqué aux
matrices X0,uv ) les (( matrices X1,k )) pour 1 ≤ k ≤ 14, au moyen des
affectations matricielles :
X1,1 := X0,12 − X0,22
X1,3 := X0,11 − X0,21
X1,5 := X0,11
X1,7 := X0,21 + X0,12
X1,9 := X0,33 + X0,44
X1,11 := X0,44
X1,13 := X0,43 − X0,33
X1,2 := X0,11 + X0,22
X1,4 := X0,11 + X0,12
X1,6 := X0,22
X1,8 := X0,34 + X0,44
X1,10 := X0,33 + X0,34
X1,12 := X0,34 − X0,44
X1,14 := X0,33
Cela signifie précisément dans l’anneau de base B, avec X1,k [i, j] =
x1,k,i,j pour 1 ≤ i, j ≤ m = 2n :
profondeur 1 :
x1,1,i,j := x0,i,m+j − x0,m+i,m+j
x1,2,i,j := x0,i,j + x0,m+i,m+j
x1,3,i,j := x0,i,j − x0,m+i,j
x1,4,i,j := x0,i,j + x0,i,m+j
x1,5,i,j := x0,i,j
x1,6,i,j := x0,m+i,m+j
x1,7,i,j := x0,m+i,j + x0,i,m+j
x1,8,i,j := x0,m2 +i,m2 +m+j + x0,m2 +m+i,m2 +m+j
x1,9,i,j := x0,m2 +i,m2 +j + x0,m2 +m+i,m2 +m+j
x1,10,i,j := x0,m2 +i,m2 +j + x0,m2 +i,m2 +m+j
x1,11,i,j := x0,m2 +m+i,m2 +m+j
x1,12,i,j := x0,m2 +i,m2 +m+j − x0,m2 +m+i,m2 +m+j
x1,13,i,j := x0,m2 +m+i,m2 +j − x0,m2 +i,m2 +j
x1,14,i,j := x0,m2 +i,m2 +j
194
7. Multiplication rapide des matrices
Ensuite on crée les (( matrices X2,k )) pour 1 ≤ k ≤ 7. Pour cela il
s’agit d’écrire 7 fois, avec à chaque fois une renumérotation convenable, le
programme Pn . Pour k de 1 à 7, on réécrit Pn avec les transformations
suivantes :
– les variables d’entrées x0,i,j (1 ≤ i, j ≤ m) sont remplacées par
les variables x1,k,i,j ,
– les variables d’entrées x0,m+i,m+j (1 ≤ i, j ≤ m) sont remplacées
par les variables x1,7+k,i,j ,
– toute variable xp,u dans Pn avec une profondeur p ≥ 1 est remplacée par la variable xp+1,k,u .
En particulier, on obtient en sortie les variables, de profondeur 3n+2,
x3n+2,k,i,j (1 ≤ i, j ≤ m) qui sont les coefficients des matrices X2,k
(1 ≤ k ≤ 7).
Il reste enfin à réaliser les affectations matricielles :
profondeur 3 :
X3,1 := X2,1 + X2,2 ; X3,2 := X2,4 − X2,6
X3,3 := X2,2 − X2,3 ; X3,4 := X2,5 − X2,7
profondeur 4 :
X4,11 := X3,1 − X3,2 ; X4,12 := X2,4 + X2,5
X4,21 := X2,1 + X2,7 ; X4,22 := X3,3 + X3,4
Cela signifie précisément, avec 1 ≤ i, j ≤ m = 2n :
profondeur 3n + 3 :
x3n+3,1,i,j := x3n+2,1,i,j + x3n+2,2,i,j ;
x3n+3,2,i,j := x3n+2,4,i,j − x3n+2,6,i,j ;
x3n+3,3,i,j := x3n+2,2,i,j − x3n+2,3,i,j ;
x3n+3,4,i,j := x3n+2,5,i,j − x3n+2,7,i,j
profondeur 3n + 4 = 3(n + 1) + 1 :
x3n+4,11,i,j := x3n+3,1,i,j − x3n+3,2,i,j ;
x3n+4,12,i,j := x3n+2,4,i,j + x3n+2,5,i,j ;
x3n+4,21,i,j := x3n+2,1,i,j + x3n+2,7,i,j ;
x3n+4,22,i,j := x3n+3,3,i,j + x3n+3,4,i,j
Le programme qui, pour l’entrée n donne en sortie le texte du programme d’évaluation Pn est un programme du type (( loop program ))
(ou programme à boucles pour : (( pour u de 1 à r faire . . . )))
de structure simple. Lorsqu’on le réalise sous forme d’une machine de
Turing écrivant le texte Pn , la gestion des boucles occupe un temps
négligeable par rapport aux instructions qui permettent d’écrire successivement P1 , P2 , . . . , Pn . Il faut prévoir que, à la fin de l’étape no i, le
texte Pi doit être recopié sur une bande où il sera lu pendant l’étape i+1,
7.2. Inversion des matrices triangulaires
195
car durant cette étape, la première bande où a été écrite Pi sera effacée
par l’écriture de Pi+1 . Si t(n) est le temps d’exécution pour l’écriture de
Pn et s(n) la taille de Pn , on obtient les formules récurrentes suivantes,
où les ci sont des constantes :
s(n + 1) ≤ c0 n m2 + 7 s(n)
t(n + 1) ≤ c1 n
m2
et
+ t(n) + c2 s(n) + c3 s(n + 1) ,
d’où, puisque nm2 = n22n = n4n est négligeable devant 7s(n) ≥ 7n ,
s(n) = O(7n )
et
t(n) = O(7n ) .
Nous pouvons résumer comme suit (rappelons que nous notons log k
pour max(log2 k, 1)).
Théorème 7.2 Lorsqu’on utilise la méthode récursive de Strassen pour
construire une famille de circuits arithmétiques pour la multiplication
des matrices carrées d’ordre m = 2n , on peut construire une machine
de Turing qui écrit le code du programme d’évaluation Qm = Pn en un
temps du même ordre de grandeur que la taille de sa sortie : O(mlog 7 ).
Naturellement, comme d’habitude le résultat sur le temps de calcul
O(mlog 7 ) est encore valable lorsque m n’est pas une puissance de 2, en
complétant les matrices dans Am×m par des lignes et colonnes de 0.
7.2
Inversion des matrices triangulaires
Les notations que nous précisons maintenant concernant la multiplication des matrices carrées seront utilisées dans toute la suite de l’ouvrage quand nous aurons à faire des calculs de complexité.
Notation 7.2.1 Nous supposerons que le calcul du produit de deux matrices n×n se fait par un circuit arithmétique de taille µM (n) = Cα nα
de profondeur γM (n) = Kα log n et de largeur λM (n) = Lα nα / log n
où 2 < α ≤ 3, Kα et Lα sont des constantes réelles positives ≥ 1 et
Cα ≥ 3 ( 6 ).
6. Certains calculs de complexité dans la suite de l’ouvrage conduiraient à des
formules légèrement différentes pour les cas α > 2 et α = 2. C’est la raison pour
laquelle nous avons préféré exclure cette dernière valeur, qui n’est de toute manière
pas d’actualité. L’hypothèse Cα ≥ 3 qui est vérifiée pour la multiplication rapide de
Strassen et pour toutes les autres multiplications rapides connues, n’est pas non plus
restrictive et simplifie quelques calculs.
196
7. Multiplication rapide des matrices
L’approche (( diviser pour gagner )) donne un algorithme qui montre
que le problème de l’inversion d’une matrice triangulaire inversible (autrement dit, fortement régulière) admet une solution en SD(nα , log2 n)
avec une constante asymptotique de l’ordre de 4 Cα pour la taille et de
l’ordre de Kα pour la profondeur du circuit.
Proposition 7.2.2 Soit A un anneau arbitraire, n un entier ≥ 2 et
A ∈ An×n une matrice triangulaire inversible.
Alors l’inverse de A peut être calculée par une famille uniforme de
circuits arithmétiques de taille τ (n) et de profondeur π(n) vérifiant :
τ (n) ≤ 4 Cα nα et π(n) ≤ Kα log2 n + O(log n) .
ν
ν
Preuve. On peut toujours supposer A ∈ A2 ×2 où ν = dlog ne (i.e.
2ν−1 < n ≤ 2ν ) quitte à rajouter 2ν − n lignes et 2ν − n colonnes
de zéros à la matrice A, et remplir la partie (( sud-est )) restante par
la matrice unité
à remplacer la matrice A par la
I2ν−n , ce qui revient
A
02ν−n,n
ν
ν
matrice A0 =
∈ A2 ×2 où 0p,q désigne, pour tous
0n,2ν−n I2ν−n
entiers naturels p et q, la matrice nulle à p lignes et q colonnes.
Le calcul de A−1 , si A est inversible, se ramène évidemment à celui
de A0−1 puisque dans ce cas A0 est inversible et
−1 −1
A
02ν−n,n
A
02ν−n,n
0−1
A
=
=
.
0n,2ν−n I2ν−n
0n,2ν−n I2ν−n
Ainsi, remplacée par A0 , la matrice A peut être considérée comme une
ν
ν
matrice
2 × 2 et s’écrire (si elle est triangulaire inférieure) :
A1 0
ν−1
ν−1
A=
où A1 , A2 , A3 ∈ A2 ×2
avec A1 , A2 trianguA3 A2
laires inférieures. Donc :
A est fortement régulière ⇐⇒ A1 et A2 sont fortement régulières.
A−1
02ν−1 ,2ν−1
1
−1
.
De plus : A =
−1
−A−1
A−1
2 A3 A1
2
−1
Le calcul de A−1
1 et A2 se fait en parallèle avec un circuit arithméν−1
tique de taille τ (2 ) et de profondeur π(2ν−1 ). On récupère ensuite
le résultat, c’est-à-dire la matrice A−1 , à partir de A−1
et A−1
1
2 , en
−1
−1
ν−1
ν−1
calculant le produit A2 A3 A1 de trois matrices 2
×2 .
Ce qui donne les relations de récurrence vraies pour tout ν ≥ 1 avec
τ (1) = π(1) = 1 :
(
τ (2ν ) = 2 τ (2ν−1 ) + 2 Cα 2(ν−1)α
π(2ν ) = π(2ν−1 ) + 2 Kα ν .
7.3. Complexité bilinéaire
197
On obtient par sommation lorsque n = 2ν , avec a = 2α−1 :
(
1
τ (n) = (Cα nα − (Cα + 1 − a) n) / (a − 1) ≤ a−1
Cα nα
π(n) = Kα (log n + 1) log n + 1 .
(ici on a utilisé sur la première ligne l’hypothèse Cα ≥ 3).
Pour le cas général, on remplace n par 2dlog ne < 2 n, log n par
1 + log n et on obtient les majorations (2 < a ≤ 4) :
(
2a
τ (n) ≤ a−1
Cα nα ≤ 4 Cα nα
π(n) ≤ Kα (log2 n + 3 log n + 2) + 1 .
7.3
2
Complexité bilinéaire
Soit un corps K et trois K – espaces vectoriels E, F , G de dimensions finies. Rappelons qu’une application bilinéaire ψ : (x, y) 7→ ψ(x, y)
de E × F vers G est une application qui est séparément linéaire en x
et en y.
Retour sur les égalités de Strassen-Winograd
Réécrivons les égalités de Strassen-Winograd données dans la section
7.1, sous une forme où nous isolons les multiplications :
α1 := α11
β1 := β11
α2 := α12
β2 := β21
α3 := α11 − α21
β3 := β22 − β12
α4 := α21 + α22
β4 := β12 − β11
α5 := α21 + α22 − α11
β5 := β22 − β12 + β11
α6 := α12 − α21 − α22 + α11 β6 := β22
α7 := α22
β7 := β22 − β12 + β11 − β21
µi := αi βi
(i = 1, . . . , 7)
γ11
γ12
γ21
γ22
:= µ1 + µ2
:= µ1 + µ4 + µ5 + µ6
:= µ1 + µ3 + µ5 − µ7
:= µ1 + µ3 + µ5 + µ4
Ici nous avons considéré avec trois matrices A, B, C les entrées
de A comme des formes linéaires αij , celles de B comme des formes
linéaires βjk , celles de C comme des formes linéaires γik . Ces formes
linéaires sont définies sur l’espace des matrices carrées d’ordre 2 sur un
anneau A. Les 8 affectations qui définissent le produit C := AB,
γik := αi1 β1k + αi2 β2k
198
7. Multiplication rapide des matrices
ont été remplacées par d’autres affectations, avec l’avantage de n’avoir
que 7 multiplications.
L’analyse de complexité nous a montré que ce passage de 8 à 7 était
un avantage décisif, indépendamment du nombre des additions utilisées
par ailleurs.
Nous avons utilisé 7 formes linéaires α` sur l’espace où vit la matrice
A, 7 formes linéaires β` sur l’espace où vit la matrice B, effectué les 7
produits µ` = α` β` et récupéré les γik comme combinaisons linéaires
des µ` .
Si nous appelons (c11 , c12 , c21 , c22 ) la base canonique de l’espace où
vit la matrice C, nous pouvons écrire
C := µ1 c1 + µ2 c2 + · · · + µ7 c7
i.e.,
C :=
X7
`=1
α` · β` · c`
où les c` sont des combinaisons linéaires suivantes des cij :
c1 := c11 + c12 + c21 + c22 c2 := c11 c3 := c21 + c22 c4 := c12 + c22
c5 := c12 + c21 + c22
c6 := c12 c7 := −c21
Bilan des courses : 7 formes linéaires (( en A )), 7 formes linéaires (( en
B )) et 7 vecteurs (( en C )). En mathématiques un peu plus savantes on
réécrit ceci en utilisant la notation tensorielle. L’application bilinéaire
(A, B) 7→ C = AB correspond au tenseur suivant, (le premier membre
de l’égalité provient directement de la définition)
X
i,j,k∈{1,2}
αij ⊗ βjk ⊗ cik
=
X7
`=1
α` ⊗ β` ⊗ c`
On peut considérer, au choix, que ces tenseurs appartiennent à un espace
tensoriel abstrait construit à partir des trois espaces E, F , G où vivent
les matrices A, B, C, ou bien qu’ils sont dans l’espace des applications
bilinéaires de E × F vers G. Dans ce dernier cas un tenseur élémentaire
α ⊗ β ⊗ c est égal par définition à l’application bilinéaire
(A, B) 7−→ α(A) · β(B) · c
7.3.1
Rang tensoriel d’une application bilinéaire
Considérons plus généralement un corps K, trois K – espaces vectoriels E, F , G de dimensions finies. Soient (ei )i∈I , (fj )j∈J , (g` )`∈L des
bases de E, F , G et notons (e?i )i∈I , (fj? )j∈J , (g`? )`∈L les bases duales.
Toute application bilinéaire ψ de E ×F vers G est alors une somme de
7.3. Complexité bilinéaire
199
tenseurs élémentaires : ψ est complètement déterminée par les images
qu’elle donne
P pour les vecteurs des bases canoniques de E et F , et si
ψ(ei , fj ) = ` γij` g` on obtient ipso facto
X
γij` e?i ⊗ fj? ⊗ g`
ψ=
i,j,`
Les γij` peuvent être appelées les coordonnées de ψ sur les trois bases
(ei )i∈I , (fj )j∈J , (g` )`∈L .
L’important du point de vue du calcul sont les règles de manipulation
des tenseurs, qui disent qu’on a le droit d’utiliser ⊗ comme (( n’importe
quel )) produit (en utilisant la linéarité par rapport à chacune des entrées,
l’associativité, mais pas la commutativité).
On peut par exemple supprimer les symboles ⊗ et calculer avec des
variables formelles xi , yj , z` à la place des e?i , fj? , g` à condition de ne
pas autoriser la commutation de deux variables entre elles (par contre,
elles commutent avec les éléments de K). L’objet abstrait correspondant
à ce calcul s’appelle l’anneau des polynômes non commutatifs à coefficients dans K.
Définition 7.3.1 (Rang tensoriel d’une application bilinéaire) Soient
K un corps, E, F , G trois K – espaces vectoriels de dimension finie.
On note Bil(E, F ; G) l’espace des applications bilinéaires de E ×F vers
G. Soit ψ ∈ Bil(E, F ; G). On appelle rang tensoriel de ψ le plus petit
entier r tel que ψ puisse s’écrire sous forme
Xr
ε` ⊗ ϕ` ⊗ g`
`=1
où les ε` sont dans E ? , les ϕ` sont dans F ? et les g` sont dans G.
Autrement dit encore c’est le plus petit entier r tel que ψ puisse s’écrire
comme composée de trois applications selon le format suivant
ε×ϕ
µr
g
E × F −→ Kr × Kr −→ Kr −→ G
où ε : E → Kr , ϕ : F → Kr et g : Kr → G sont des applications
linéaires et µr : Kr ×Kr → Kr est le produit coordonnée par coordonnée.
Le programme d’évaluation arithmétique correspondant s’appelle un calcul bilinéaire de ψ. Le rang tensoriel de ψ est encore appelé la complexité bilinéaire de ψ. Nous le noterons R(ψ), ou s’il y a ambiguité
RK (ψ).
L’importance du rang tensoriel dans les questions de complexité algébrique a été soulignée par Gastinel ([33]) et Strassen ([87, 90, 91]).
200
7. Multiplication rapide des matrices
Remarque 7.3.2 Nous laissons libre choix pour l’interprétation du tenseur ε` ⊗ ϕ` ⊗ g` . Pour les gens savants cet objet vit dans un espace
tensoriel abstrait E ? ⊗ F ? ⊗ G, canoniquement isomorphe à l’espace
des applications bilinéaires E × F → G. Mais on peut considérer aussi
que cet objet est égal par définition à l’application bilinéaire (x, y) 7→
ε` (x) · ϕ` (y) · g` .
Remarque 7.3.3 Le lecteur ou la lectrice peut donner la définition
analogue pour le rang tensoriel d’une application linéaire, ou celui d’une
forme bilinéaire, et vérifier qu’on retrouve la notion usuelle de rang pour
ces objets.
Remarque 7.3.4 Nous pourrions remplacer dans la définition 7.3.1
le corps K par un anneau commutatif arbitraire A, à condition de
considérer des espaces convenables analogues aux espaces vectoriels. Une
possibilité est de considérer que E, F et G doivent être des A – modules libres, c’est-à-dire des modules (isomorphes à) Ae , Af et Ag . Par
exemple pour le produit matriciel, le cadre le plus naturel serait de choisir de travailler sans aucune hypothèse précise, c’est-à-dire sur l’anneau
Z.
Remarque 7.3.5 Contrairement au rang d’une application linéaire, le
rang tensoriel d’une application bilinéaire est en général difficile à déterminer. Il ne semble pas qu’on connaisse d’algorithme qui réalise ce travail, sauf pour quelques classes de corps particuliers (les corps finis ou
les corps algébriquement clos par exemple). Mais même dans ces cas,
les algorithmes sont impraticables. La détermination du rang tensoriel
des applications bilinéaires sur un corps fini fixé est un problème N Pcomplet (cf. [43]) : le problème, en prenant pour entrées un entier r et
une application bilinéaire ψ donnée par ses coordonnées sur trois bases,
est de déterminer si le rang tensoriel de ψ est ≤ r ou non.
Rang tensoriel de la multiplication des matrices
La notation tensorielle n’a pas seulement l’avantage de l’élégance.
Elle a aussi le mérite de nous aider à réfléchir sur les calculs mis en
œuvre. La meilleure synthèse de l’idée de Strassen est peut-être de dire
que le miracle s’est produit quand il a pu écrire le tenseur de la multiplication des matrices carrées d’ordre 2 comme somme de 7 tenseurs
élémentaires.
7.3. Complexité bilinéaire
201
Chaque fois qu’on arrive à écrire la multiplication des matrices carrées d’ordre k comme une somme de h tenseurs élémentaires, avec une
bonne valeur de log h/ log k on obtient immédiatement que la multiplication des matrices tombe dans la classe SD(nlog h/ log k , log n) car le
calcul de complexité de la section 7.1 pourra fonctionner à l’identique.
Plus précisément, la possibilité d’effectuer des produits de matrices carrées d’ordre k ` par blocs de taille k `−1 implique récursivement
que ce produit matriciel est représenté par une somme de h` tenseurs
élémentaires. En outre si la profondeur du programme d’évaluation correspondant au produit des matrices d’ordre k est un entier K alors
celle du programme d’évaluation correspondant au produit des matrices
d’ordre k ` est égale à ` K, et sa profondeur multiplicative (c’est-à-dire
la profondeur mesurée en ne tenant compte que des multiplications essentielles, cf. définition 3.1.3 page 116) est égale à `.
Depuis la découverte de Strassen, un nouveau sport a été créé, auquel
ont participé quelques grands noms de la complexité algébrique : faire diminuer log h/ log k en élaborant des identités algébriques inédites, pour
des valeurs de k de plus en plus grandes.
Un aspect fascinant de la notation tensorielle pour les applications
bilinéaires est qu’elle établit une symétrie entre les trois espaces E, F ,
G en jeu (rappelons qu’il s’agit ici d’espaces de matrices). Symétrie
qui n’est pas directement visible sur la définition. En fait, il n’y aurait vraiment symétrie que si nous considérions notre tenseur comme
représentant l’application trilinéaire
(A, B, C) 7−→
X
i,j,k
αij βjk γki
L’écriture tensorielle permet de traiter des arguments de dualité sous
forme scripturale. Pour montrer que ce jeu d’écriture est bien plus qu’un
jeu, prenons de nouveau les égalités de Strassen-Winograd que nous
réécrivons avec des tenseurs où nous ne marquons pas la différence entre
formes linéaires et vecteurs. Cela donne alors pour le produit matriciel
l’égalité
X
i,j,k∈{1,2}
aij ⊗ bjk ⊗ cik
=
X7
`=1
a` ⊗ b` ⊗ c`
202
7. Multiplication rapide des matrices
avec
a1
a2
a3
a4
a5
a6
a7
:= a11
:= a12
:= a11 − a21
:= a21 + a22
:= a21 + a22 − a11
:= a12 − a21 − a22 + a11
:= a22
b1
b2
b3
b4
b5
b6
b7
:= b11
:= b21
:= b22 − b12
:= b12 − b11
:= b22 − b12 + b11
:= b22
:= b22 − b12 + b11 − b21
c1
c2
c3
c4
c5
c6
c7
:= c11 + c12 + c21 + c22
:= c11
:= c21 + c22
:= c12 + c22
:= c12 + c21 + c22
:= c12
:= −c21
Pour réaliser une symétrie par permutation circulaire dans la définition,
échangeons les indices i et k dans les cik . Alors, vu l’invariance par
permutation circulaire nous pouvons remplacer partout a, b et c par
b, c et a. Et finalement nous permutons à nouveau les indices i et k
dans les (nouveaux) cik pour revenir à la définition. Ceci nous donne
d’autres égalités, qui peuvent tout aussi bien servir que les premières :
a1
a2
a3
a4
a5
a6
a7
:= a11 + a21 + a12 + a22
:= a11
:= a12 + a22
:= a21 + a22
:= a21 + a12 + a22
:= a21
:= −a12
b1
b2
b3
b4
b5
b6
b7
:= b11
:= b12
:= b11 − b21
:= b21 + b22
:= b21 + b22 − b11
:= b12 − b21 − b22 + b11
:= b22
c1
c2
c3
c4
c5
c6
c7
:= c11
:= c12
:= c22 − c21
:= c21 − c11
:= c22 − c21 + c11
:= c22
:= c22 − c21 + c11 − c12
Naturellement, dans le cas présent, on obtient seulement sans fatigue
un nouveau système d’identités algébriques pour traiter le même produit matriciel. Mais si nous étions parti d’un produit de matrices rectangulaires non carrées, la permutation circulaire deviendrait un outil
vraiment efficace, produisant des identités correspondant à un cas de figure vraiment nouveau. Cette remarque importante remonte à 1972 (cf.
[46, 75]).
Notez que nous avons une situation familière analogue si nous considérons
le cas des applications linéaires de E vers F . La dualité nous dit que le
passage de ϕ à t ϕ est un isomorphisme. En termes de matrices c’est
une banale transposition. En termes d’écriture tensorielle, c’est un jeu
d’écriture. Les tenseurs remplacent les matrices lorsqu’il y a plus que
deux espaces en cause.
Notation 7.3.6 (Rang tensoriel de la multiplication des matrices)
Soient m, n, p trois entiers > 0 et K un corps. On note hm, n, piK (ou
hm, n, pi) l’application bilinéaire
(A, B) 7−→ AB
où
A ∈ Km×n , B ∈ Kn×p et AB ∈ Km×p
On note donc le rang tensoriel par R hm, n, pi, ou si on doit préciser
RK hm, n, pi.
7.3. Complexité bilinéaire
203
Proposition 7.3.7 (Rang tensoriel de la multiplication des matrices)
(1) Si m ≤ m0 , n ≤ n0 et p ≤ p0 alors R hm, n, pi ≤ R hm0 , n0 , p0 i
0 0 0
(2) R hmm0 , nn0 , pp0 i ≤ R hm, n, pi · R
Phm , n , p i
(3) R hm1 + m2 , n1 + n2 , p1 + p2 i ≤ i,j,k∈{1,2} R hmi , nj , pk i
(4) R n` , n` , n` ≤ (R hn, n, ni)`
(5) R hm, n, pi est invariant par permutation des entiers m, n, p.
(6) R h1, n, 1i = n et R hm, n, 1i = mn
Preuve. Le point (1) est facile : on peut compléter des matrices correspondant au format (m, n, p) par des 0 pour en faire des matrices au
format (m0 , n0 , p0 ).
Les points (2), (3) résultent de la possibilité de faire des produits de
matrices par blocs. Et (4) résulte de (2).
Le point (5) a été expliqué avant la proposition (voir page 201). On peut
redire à peu près la même chose sous la forme suivante un peu plus abstraite, qui décrit peut-être mieux l’essence du résultat. Si E1 , E2 , E3
sont trois K – espaces vectoriels de dimensions finies p, n, m, alors la
multiplication des matrices correspondante hm, n, pi : (A, B) 7→ AB est
un élément canonique θ de l’espace
Bil(Hom(E2 , E3 ), Hom(E1 , E2 ); Hom(E1 , E3 )) = Bil(E, F ; G)
Si nous notons L3 (E, F, G) l’espace des formes trilinéaires sur E×F ×G,
nous avons un isomorphisme canonique
Bil(E, F ; G) ' L3 (E, F, G? )
ψ 7−→ ϕ = ((x, y, γ) 7→ γ(ψ(x, y))
Dans la situation présente, il y a aussi une dualité canonique entre
Hom(E1 , E3 ) = G et Hom(E3 , E1 ) donnée sous forme matricielle par
(C, D) 7→ Tr(CD), ce qui fournit un isomorphisme canonique entre
Hom(E3 , E1 ) et G? . Une fois mis bout à bout tous ces isomorphismes
canoniques, on voit que l’élément canonique
θ ∈ Bil(Hom(E2 , E3 ), Hom(E1 , E2 ); Hom(E1 , E3 ))
correspond à l’élément canonique
θ0 ∈ L3 (Hom(E2 , E3 ), Hom(E1 , E2 ), Hom(E3 , E1 ))
donné sous forme matricielle par (A, B, D) 7→ Tr(ABD). Maintenant il
est bien connu que Tr(ABD) = Tr(BDA) = Tr(DAB) et Tr(ABD) =
Tr( t (ABD)) = Tr( t D t B t A). Ceci établit les symétries demandées.
204
7. Multiplication rapide des matrices
Voyons maintenant le point (6). Nous reprenons les notations précédentes
avec E1 = K, donc on identifie Hom(E1 , Ei ) à Ei (i = 2, 3). Regardons l’espace Bil(E, F ; G) sous la forme (canoniquement équivalente)
Hom(E, Hom(F, G)). Si on écrit le produit d’une matrice par un vecteur
colonne (A, X) 7→ AX sous forme
Xr
Xi=m,j=n
α` ⊗ β` ⊗ c`
αij ⊗ βj1 ⊗ ci1 =
i,j=1
`=1
on voit que l’application
linéaire correspondante de E vers Hom(F, G)
T
est nulle sur rl=1 Ker(α` ). Mais dans le cas présent, modulo les identifications précédentes, cette application linéaire n’est autre que l’application identique de E. Son noyau est donc réduit à {0} et r est au moins
égal à la dimension de E c’est-à-dire à mn.
t
u
7.3.2
Exposant de la multiplication des matrices carrées
Définition 7.3.8 On dit que α est un exposant acceptable pour la multiplication des matrices carrées si celle-ci peut être réalisée en SD(nα ,
log n). La borne inférieure des exposants acceptables est appelée l’exposant de la multiplication des matrices carrées et elle est notée ω.
A priori on devrait mettre en indice le corps K pour les exposants α
et ω. Les résultats concernant ces exposants dont nous rendons compte
sont cependant indépendants du corps K considéré.
Théorème 7.3 (Rang tensoriel et exposant de la multiplication des
matrices carrées)
(1) S’il existe n et r tels que R hn, n, ni = r alors l’exposant α =
log r
log n est acceptable pour la multiplication des matrices carrées.
(2) S’il existe m, n, p et r tels que R hm, n, pi = r alors l’exposant
r
α = 3 loglogmnp
est acceptable pour la multiplication des matrices
carrées.
Preuve. Comme nous l’avons déjà remarqué le point (1) résulte du
même calcul de complexité que celui fait dans la section 7.1.
Le point (2) résulte du point (1) puisque d’après les points (2) et (5) de
la proposition 7.3.7 on a (avec N = mnp)
R hN, N, N i ≤ R hm, n, pi R hn, p, mi R hp, m, ni = (R hm, n, pi)3
t
u
En fait la conclusion dans le théorème précédent est non seulement qu’il existe une famille de circuits arithmétiques dans la classe
7.3. Complexité bilinéaire
205
SD(nα , log n) qui réalise la multiplication des matrices carrées, mais
qu’on sait construire explicitement une telle famille uniforme de circuits
arithmétiques. En outre le temps de construction du circuit arithmétique numéro n est proportionnel à sa taille, selon les lignes de la preuve
du théorème 7.2.
Le point (1) du théorème ci-dessus peut être précisé comme suit (par
le même calcul qu’à la section 7.1).
Proposition 7.3.9 (Précision pour le théorème 7.3 (1)) Supposons que
l’application bilinéaire hn, n, ni puisse être calculée par un circuit arithmétique de profondeur ` contenant r multiplications essentielles et
s autres opérations arithmétiques (addition, soustraction, multiplication par une constante), avec r > n2 . Alors l’application bilinéaire
hnν , nν , nν i peut être calculée par un circuit arithmétique de profonν −n2ν
deur ν ` contenant rν multiplications essentielles et s r r−n
autres
2
opérations arithmétiques.
7.3.3
Complexité bilinéaire versus complexité multiplicative
Soient K un corps, H, G deux K – espaces vectoriels de dimension
finie. Une application quadratique de H vers G est par définition une
application de la forme Ψ : x 7→ ψ(x, x) où ψ ∈ Bil(H, H; G). Si (hi )i∈I
et (g` )`∈L sont des bases de H et G il revient au même de dire que
chaque coordonnée de Ψ(x) est une forme quadratique en x, c’est-àdire un polynôme homogène du second degré en les coordonnées de x.
Si les coordonnées de x sont prises comme variables, on peut alors
considérer les programmes d’évaluation arithmétiques sans division qui
permettent de calculer les coordonnées de Ψ(x). La complexité multiplicative de Ψ est alors définie comme la plus petite longueur multiplicative d’un tel programme d’évaluation. Nous la noterons M (Ψ).
Comme les changements de base ne coûtent rien en longueur multiplicative cette définition ne dépend pas du choix des bases. Le lemme suivant
est une paraphrase de la proposition 3.1.6 dans le cas d’une application
quadratique.
Lemme 7.3.10 Avec les notations précédentes la complexité multiplicative d’une application quadratique Ψ est aussi égale au plus petit entier
r tel que Ψ puisse s’écrire comme composée de trois applications selon
206
7. Multiplication rapide des matrices
le format suivant
(η,ζ)
µr
g
H −→ Kr × Kr −→ Kr −→ G
où η : H → Kr , ζ : H → Kr et g : Kr → G sont des applications linéaires et µr : Kr × Kr → Kr est le produit coordonnée par
coordonnée. Un programme d’évaluation arithmétique correspondant à
cette décomposition s’appelle un calcul quadratique de Ψ.
Remarque 7.3.11 Si on considérait des circuits arithmétiques avec
division on ne pourrait pas diminuer pour autant la longueur multiplicative pour évaluer une application quadratique, au moins dans le cas
d’un corps infini, d’après le théorème 3.1 et la proposition 3.2.3.
Proposition 7.3.12 Soient K un corps, E, F , G trois K – espaces
vectoriels de dimension finie. Soit ψ ∈ Bil(E, F ; G) et H = E × F .
Alors ψ est une application quadratique de H vers G. Sa complexité
bilinéaire R(ψ) et sa complexité multiplicative M (ψ) sont reliées par
M (ψ) ≤ R(ψ) ≤ 2M (ψ)
Preuve. La première inégalité est évidente. Pour la seconde considérons
un programme quadratique comme dans le lemme 7.3.10 qui calcule
ψ(u, v) avec m = M (ψ) multiplications essenteielles. On a donc
Xm
ψ(u, v) =
α` (u, v) · β` (u, v) · g`
`=1
où les α` et β` sont dans H ? . Remarquons qu’on a
α` (u, v) · β` (u, v) = α` (u, 0) · β` (u, 0) + α` (0, v) · β` (0, v)+
α` (u, 0) · β` (0, v) + α` (0, v) · β` (u, 0)
Puisque ψ(u, v) est bilinéaire on a ψ(u, 0) = 0 et on peut supprimer
les termes α` (u, 0) · β` (u, 0) · g` dont la somme est nulle. Même chose
avec ψ(0, v) = 0 et finalement on obtient
Xm
ψ(u, v) =
(α` (u, 0) · β` (0, v) + β` (u, 0) · α` (0, v)) · g`
`=1
et ceci montre que R(ψ) ≤ 2r.
t
u
On en déduit le résultat suivant qui relie le rang tensoriel et l’exposant de la multiplication des matrices carrées.
7.3. Complexité bilinéaire
207
Théorème 7.4 L’exposant ω de la multiplication des matrices carrées
est égal à la borne inférieure des exposants α qui vérifient, pour au
moins un entier n, l’inégalité R hn, n, ni ≤ nα . On a aussi
ω = lim
n→∞
log R hn, n, ni
log M (hn, n, ni)
= lim
n→∞
log n
log n
où chaque suite converge vers sa borne inférieure.
Preuve. Il est clair d’après la proposition 7.3.12 que les deux suites
considérées ont la même borne inférieure β.
On a le résultat direct plus précis dans le théorème 7.3 : tout exposant α
est acceptable
strictement supérieur à la borne inférieure des log Rloghn,n,ni
n
pour la multiplication des matrices carrées.
Pour la réciproque on considère un α > β. Pour un n0 assez grand
on a un programme d’évaluation sans division de longueur < nα0 /2 qui
calcule l’application quadratique hn0 , n0 , n0 i. A fortiori sa longueur multiplicative est < nα0 /2 et on a R hn0 , n0 , n0 i ≤ 2M (hn0 , n0 , n0 i) < nα0 .
t
u
Malgré la relation très étroite entre M (ψ) et R(ψ), c’est seulement
la considération du rang tensoriel qui permet de démontrer les résultats
de base concernant l’exposant de la multiplication des matrices. Cela
tient à ce que la proposition 7.3.7 ne serait pas vraie en remplaçant
le rang tensoriel par la longueur multiplicative. Le fait d’interdire la
commutation dans les tenseurs est ce qui permet de traiter correctement
le produit des matrices par blocs.
7.3.4
Extension du corps de base
Soient K un corps, E, F , G trois K – espaces vectoriels de dimension
finie. Soit ψ ∈ Bil(E, F ; G). Si L est une extension de K on peut étendre
ψ à L de manière naturelle. Nous nous en tiendrons ici à un point de
vue pragmatique et purement calculatoire. Si (ei )i∈I , (fj )j∈J , (g` )`∈L
sont des bases de E, F , G et si
X
ψ =
γij` e?i ⊗ fj? ⊗ g`
i,j,`
nous considérons trois L – espaces vectoriels EL , FL , GL ayant les
mêmes bases et l’extension ψL de ψ est définie par la même égalité.
Comme tout calcul bilinéaire dans K est aussi un calcul bilinéaire dans
L on a nécessairement l’inégalité RL (ψL ) ≤ RK (ψ), mais il se peut
208
7. Multiplication rapide des matrices
que l’utilisation de constantes dans L puisse faciliter le calcul de ψ et
l’inégalité peut être stricte.
Nous allons cependant voir dans ce paragraphe que l’exposant de la
multiplication des matrices ne peut pas changer lorsqu’on passe d’un
corps K à une extension L.
Lemme 7.3.13 Avec les notations précédentes
1) Si L est une extension finie de K de degré n il existe un entier m ≤
n3 tel que pour toute application bilinéaire ψ on a RK (ψ) ≤ m RL (ψL ).
En particulier l’exposant de la multiplication des matrices ne change pas
lorsqu’on passe de K à L.
2) Si L = K(t) (corps des fractions rationnelles en t) et si K est infini,
on a l’égalité RL (ψL ) = RK (ψ).
Preuve. Dans le cas 2) la famille finie des constantes cs (t) dans L
utilisées par le circuit arithmétique peut être remplacée par des constantes cs (a) où a ∈ K est choisi de manière à n’annuler aucun des
dénominateurs.
Dans le cas 1) considérons une base b = (1, b2 , . . . , bn ) de L lorsqu’on
le voit comme K – espace vectoriel. La multiplication dans L représente
une application bilinéaire sur K lorsqu’elle est traduite dans les coordonées sur la base b. Cette application bilinéaire L × L → L peut
être réalisée par m ≤ n3 multiplications essentielles dans K. En fait la
constante m peut être prise égale au rang tensoriel de cette application
bilinéaire, qui est en général noté RK (L).
Tout calcul bilinéaire dans L peut alors être mimé par un calcul bilinéaire dans K de la manière suivante. Chaque variable xi sur L est
remplacée par n variables xi,n sur K qui représentent les coordonnées
de xi sur la base b. Seules les multiplications essentielles du calcul
dans L produisent des multiplications essentielles dans K. Dans cette
simulation, le nombre de multiplications essentielles est multiplié par la
constante m.
Si maintenant α > ωL il existe un entier n tel que RL hn, n, ni < nα ,
donc pour une puissance convenable N = n` on a RK hN, N, N i ≤
RK (L) · RL hN, N, N i < N α , donc α ≥ ωK .
t
u
On en déduit le résultat suivant dû à Schönhage ([82]).
Proposition 7.3.14 L’exposant de la multiplication des matrices carrées sur un corps K ne dépend que de la caractéristique de K.
Preuve. Il suffit de prouver que l’exposant ne change pas lorsqu’on
passe d’un corps premier K (Q ou l’un des Fp ) à l’une de ses extensions
7.4. Calculs bilinéaires approximatifs
209
L. Supposons qu’on ait RL (hn, n, ni) ≤ r ≤ nα , c’est-à-dire qu’on ait
sur le corps L une égalité
Xr
X
α` ⊗ β` ⊗ c`
αij ⊗ βjk ⊗ cik =
i,j,k∈{1,...,n}
`=1
On peut considérer les coordonnées des α` , β` , c` sur les bases αij , βij ,
cij comme 3rn2 indéterminées zs . L’égalité des deux tenseurs ci-dessus
signifie que ces indéterminées vérifient un système de n6 équations polynomiales de degré 3 dont tous les coefficients sont égaux à 1 ou 0.
Maintenant, le lecteur ou la lectrice connaı̂t peut-être le beau résultat
suivant 7 : lorsqu’un système d’équations polynomiales sur un corps K
admet une solution dans une extension L, alors il admet une solution
dans une extension finie de K. On est donc ramené au premier cas du
lemme précédent.
t
u
Remarque 7.3.15 L’exposant ωK peut donc être étudié en prenant
pour K la clôture algébrique de Q ou de Fp . Lorsqu’on a affaire à un
corps algébriquement clos K le rang tensoriel RK (hn, n, ni) est calculable en principe (sinon en pratique) car savoir si RK (hn, n, ni) ≤ m
revient à déterminer si un système d’équations algébriques admet ou
non une solution (comme dans la preuve de la proposition 7.3.14). Et
on sait, en principe, répondre à ce genre de questions par un algorithme
d’élimination. On ne sait cependant pas grand chose concernant ωK .
Cet exposant mythique est un nombre compris entre 2 et 2, 38. Mais on
ne sait apparemment toujours rien sur la vitesse avec laquelle la suite
log RK (hn, n, ni)/ log n (cf. théorème 7.4) converge vers ωK . Il se pourrait que la vitesse de convergence soit si lente que le nombre ωK serait
définitivement impossible à calculer.
7.4
7.4.1
Accélération par la méthode des calculs bilinéaires approximatifs
Méthode de Bini
La méthode des calculs bilinéaires approximatifs est inspirée des
méthodes numériques approchées, elle a été inventée par Bini et elle
7. Ce résultat admet de nombreuses preuves, dont certaines tout à fait explicites. Essentiellement c’est un résultat de la théorie de l’élimination. On peut le faire
découler du Nullstellensatz de Hilbert, du lemme de normalisation de Noether ou
encore de la théorie des bases de Gröbner. Il se trouve dans les bons livres d’algèbre.
210
7. Multiplication rapide des matrices
a quelque parenté avec l’élimination des divisions de Strassen. Elle a
débloqué la situation pour l’exposant ω de la multiplication des matrices.
Un exemple est le produit de deux matrices A, B ∈ A2×2 avec la
première qui a son coefficient a2,2 nul. On schématise ce produit sous
la forme suivante (figure 7.1). Ce produit matriciel à trous correspond
Figure 7.1 – Produit matriciel à trou de Bini
à un tenseur de rang 6 qui s’écrit avec la notation des polynômes non
commutatifs
ψ = a11 b11 c11 + a12 b21 c11 + a21 b11 c21 + a11 b12 c12 + a12 b22 c12 + a21 b12 c22
On introduit (pour les mêmes variables) un tenseur de rang 5 perturbé
par des ε xij (avec x = a, b ou c).
ϕ(ε) = (a12 + εa11 ) (b12 + εb22 ) c21 + (a21 + εa11 ) b11 (c11 + εc12 )
−a12 b12 (c11 + c21 + εc22 ) − a12 (b11 + b12 + εb21 ) c11
+(a12 + a21 ) (b12 + εb21 ) (c11 + εc22 )
Lorsqu’on développe on obtient
ϕ(ε) = ε ψ + ε2 θ(ε)
Numériquement on a donc lorsque ε est suffisamment petit ψ ' ϕ(ε)/ε.
On dit que ϕ constitue une approximation d’ordre 1 de ψ. On peut
transformer ceci en un calcul purement formel dans l’anneau des développements limités à l’ordre 1 en ε comme lorsqu’on élimine les divisions
à la Strassen. Naturellement, il n’y a pas de miracle, cela ne donne pas
une écriture de ψ comme somme de 5 tenseurs élémentaires. Mais il
y a néanmoins quelque chose à gagner en prenant un peu de recul et
en analysant en détail ce qui se passe. Tout d’abord en appliquant le
schéma suivant (où on fait un produit par blocs non rectangulaires !) on
constate que le produit matriciel h3, 2, 2i peut être réalisé de manière
approximative (à l’ordre 1) par une somme de 10 tenseurs élémentaires
au lieu de 12. Grâce au produit de matrices par blocs rectangulaires on
pourra alors réaliser h12, 12, 12i de manière approximative (à un ordre
7.4. Calculs bilinéaires approximatifs
211
Figure 7.2 – Produit matriciel plein
convenable, nous allons définir cela un peu plus loin) comme somme de
103 tenseurs élémentaires au lieu des 123 nécessaires dans la méthode
usuelle. Enfin il reste à réaliser que lorsqu’on passe (grâce au produit par
blocs) à h12n , 12n , 12n i l’ordre d’approximation ne croı̂t pas trop vite
et que le coût du décryptage d’un calcul bilinéaire approximatif en un
calcul bilinéaire exact devient négligeable devant 10εn pour n’importe
quel ε > 0. Tout ceci ramène l’exposant ω à
3 log(10)/ log(12) < 2, 78 < 3 log(7)/ log(8) = 2, 807
Nous devons maintenant donner des définitions et énoncés plus précis
pour vérifier que ce plan de travail fonctionne bien.
Définition 7.4.1 Soient K un corps, E, F , G trois K – espaces vectoriels de dimensions finies. Soit ψ ∈ Bil(E, F ; G). Soit L = K[ε] l’anneau
des polynômes en la variable ε sur K. Un élément ϕ(ε) de Bil(EL , FL ;
GL ) est appelé une approximation d’ordre q de ψ si on a
ϕ(ε) ≡ εq ψ
modulo
εq+1
Un calcul bilinéaire de ϕ(ε) est appelé un calcul bilinéaire approximatif
de ψ à l’ordre q. On appelle rang tensoriel marginal de ψ à l’ordre q
le plus petit rang possible pour un calcul bilinéaire approximatif de ψ à
l’ordre q. On le note R(ψ, q). Enfin, le rang tensoriel marginal de ψ
est le plus petit des R(ψ, q) et il est noté R(ψ). Nous dirons aussi plus
simplement le rang marginal de ψ.
Remarque 7.4.2 Nous utilisons ici des calculs bilinéaires sur un anneau, comme il était indiqué dans la remarque 7.3.4. De même l’extension de ψ à l’anneau K[ε] se fait comme dans le cas d’une extension du
corps de base (cf. page 207).
Remarque 7.4.3 Il est clair que lorsque q augmente, le rang tensoriel
marginal à l’ordre q d’une application bilinéaire ne peut que diminuer,
212
7. Multiplication rapide des matrices
autrement dit
R(ψ) = R(ψ, 0) ≥ R(ψ, 1) ≥ · · · ≥ R(ψ, q) ≥ R(ψ, q + 1) · · ·
Le rang tensoriel marginal à l’ordre q d’une application bilinéaire est a
priori nettement plus difficile à calculer que son rang tensoriel. Le rang
marginal est encore plus difficile à établir. En fait on est satisfait quand
on a établi une bonne majoration du rang marginal en explicitant un
calcul bilinéaire approximatif.
Quel est le coût d’un décryptage d’un calcul bilinéaire approximatif à l’ordre q ? Nous avons déjà fait un calcul analogue dans la preuve
de la proposition 3.1.6 qui concernait la possibilité d’une mise en forme
simplifiée des circuits arithmétiques sans division.
On commence par considérer que le calcul bilinéaire de ϕ(ε) se passe
non pas sur l’anneau L = K[ε] mais sur l’anneau des développements
limités à l’ordre q. Ensuite on simule toute variable Z, qui représente un
élément de K[ε] modulo εq , par q +1 variables Z [k] dans K (0 ≤ k ≤ q)
qui représentent les coefficients de Z en dessous du degré q. Quand on
doit calculer le coefficient de εq dans un tenseur X(ε) Y (ε) Z(ε) on voit
qu’on doit faire la somme des X [i] Y [j] Z [k] pour tous les triplets (i, j, k)
dont la somme vaut q. Il y a au plus (q + 1)(q + 2)/2 triplets de ce type.
En termes de rang tensoriel, cela signifie donc que le rang tensoriel de
ψ est majoré par (q + 1)(q + 2)/2 fois son rang marginal à l’ordre q.
Nous avons donc établi le lemme suivant.
Lemme 7.4.4 Soit ψ une application bilinéaire définie sur un corps K.
Si un calcul bilinéaire approximatif à l’ordre q de ψ a une complexité
bilinéaire `, on en déduit par décryptage un calcul bilinéaire de ψ de
complexité bilinéaire ≤ ` · (q + 1)(q + 2)/2. En bref
R(ψ) ≤
(q + 1)(q + 2)
R(ψ, q)
2
≤ q 2 R(ψ, q)
si
q≥4
Maintenant nous devons examiner comment se comporte le rang marginal du produit matriciel lorsqu’on utilise des produits par blocs.
Proposition 7.4.5 (Rang tensoriel marginal de la multiplication des
matrices)
(1) Le rang marginal R(hm, n, pi , q) est une fonction croissante de
chacun des entiers m, n, p et décroissante de l’entier q.
7.4. Calculs bilinéaires approximatifs
213
(2) R(hmm0 , nn0 , pp0 i , q+q 0 ) ≤ R(hm, n, pi , q)·R(hm0 , n0 , p0 i , q 0 ). En
particulier avec N = mnp on a
n, pi , q))3
R(hN, N, N i , 3q) ≤ (R(hm,
P
(3) R(hm1 + m2 , n1 + n2 , p1 + p2 i , q) ≤ i,j,k∈{1,2} R(hmi , nj , pk i , q)
(4) R n` , n` , n` , `q ≤ R(hn, n, ni , q)`
(5) R(hm, n, pi , q) est invariant par permutation des entiers m, n, p.
Preuve. Tout se passe comme avec le rang tensoriel usuel dans la preuve
de la proposition 7.3.7. Le seul point qui demande un peu d’attention
est le point (2). La meilleure manière de le comprendre est (encore une
fois) de prendre du recul. Il faut prendre du recul sur ce que représente
le tenseur hm1 m2 , n1 n2 , p1 p2 i par rapport aux tenseurs hm1 , n1 , p1 i et
hm2 , n2 , p2 i. Lorsque nous voyons une matrice A de type m1 m2 × n1 n2
comme une matrice de type m1 × n1 ayant pour entrées des matrices
Aij de type m2 × n2 nous repérons une entrée de la grosse matrice par
deux paires d’indices ((i1 , j1 ), (i2 , j2 )) correspondant au couple d’indices
(i1 (m2 − 1) + i2 , j1 (n2 − 1) + j2 ) comme dans l’exemple décrit par la
figure 7.4.1 avec (m1 , n1 ) = (5, 6), (m2 , n2 ) = (3, 4), (i1 , i2 ) = (3, 1) et
(j1 , j2 ) = (4, 2).
(4, 2)
(3, 1)
s
Figure 7.3 – Numérotation par blocs
Cependant la mise en ligne de la paire (i1 , i2 ) sous forme i1 (m2 −1)+
i2 , si elle est indispensable au dessin et à une première compréhension
des choses, est plutôt un obstacle pour ce qui concerne la compréhension
du calcul (( emboité )) que représente un produit par blocs. Prenons en
214
7. Multiplication rapide des matrices
effet les indices dans la grande matrice sous forme des couples (i, j) =
((i1 , i2 ), (j1 , j2 )) comme dans la figure 7.4.1 (et non pas (i1 (m2 − 1) +
i2 , j1 (n2 − 1) + j2 ) ni non plus ((i1 , j1 ), (i2 , j2 ))). Nous obtenons, en
notation de polynômes non commutatifs :
X
hm1 m2 , n1 n2 , p1 p2 i =
ai,j bj,k ci,k
i,j,k
(où la somme est prise sur i ∈ I1 × I2 , j ∈ J1 × J2 , k ∈ K1 × K2 pour
des ensembles d’indices de cardinalités convenables). Alors nous avons
l’égalité :
hm1 m2 , n1 n2 , p1 p2 i = hm1 , n1 , p1 i hm2 , n2 , p2 i
où
P
hm1 , n1 , p1 i =
a
b
c
et
Pi1 ,j1 ,k1 0i1 ,j1 0j1 ,k1 0i1 ,k1
hm2 , n2 , p2 i =
i2 ,j2 ,k2 ai2 ,j2 bj2 ,k2 ci2 ,k2
(nous avons mis des 0 pour le cas où les ensembles d’indices dans le
premier tenseur ne seraient pas disjoints de ceux du second) à condition
de respecter les règles de calcul suivantes
x(i1 ,i2 ),(j1 ,j2 ) = xi1 ,j1 x0i2 ,j2
xi1 ,j1 yi02 ,j2 = yi02 ,j2 xi1 ,j1
(x vaut pour a, b ou c)
(idem avec x et y 6= x)
Une fois ceci constaté, nous n’avons même plus besoin de penser au
calcul emboité que représente le produit par blocs. Nos nouvelles règles
de calcul fonctionnent toutes seules et produisent automatiquement le
résultat (2) aussi bien dans la proposition 7.3.7 que dans la proposition
7.4.5. Nous sommes en effet ramenés maintenant à la constatation banale suivante concernant les développements limités : le premier terme
non nul du produit de deux développements limités est le produit des
premiers termes non nuls de chacun des deux développements limités.
Et les ordres des deux développements limités s’ajoutent.
t
u
Le raisonnement fait au début de ce paragraphe, en tenant compte
de la proposition 7.4.5 et du lemme 7.4.4 donne alors le résultat de Bini.
Théorème 7.5
(1) S’il existe n et r tels que R hn, n, ni ≤ r alors
log r
ω
n ≤ r, c’est-à-dire ω ≤ log
n.
(2) S’il existe m, n, p tels que R hm, n, pi ≤ r alors (mnp)ω/3 ≤ r,
r
c’est-à-dire ω ≤ 3 loglogmnp
.
7.4. Calculs bilinéaires approximatifs
215
En bref, pour ce qui concerne l’exposant ω, une inégalité R hm, n, pi
≤ r donne le même résultat qu’une inégalité R hm, n, pi ≤ r.
On pourra remarquer que la preuve du théorème 7.5 est tout à fait
explicite. Si on connait un calcul bilinéaire approximatif à l’ordre q qui
utilise r multiplications essentielles pour le produit matriciel hm, n, pi
r
et si α > 3 loglogmnp
alors on sait construire un entier N et un calcul
bilinéaire pour le produit matriciel hN, N, N i qui utilise moins de N α
multiplications essentielles.
Corollaire 7.4.6 On a pour l’exposant de la multiplication des matrices
carrées ω ≤ 3 log 10/ log 12 < 2.7799.
7.4.2
Une première amélioration décisive de Schönhage
La méthode de Bini n’a pas donné dans un premier temps une amélioration très importante de l’exposant ω mais elle a ouvert la voie aux
améliorations suivantes, beaucoup plus substantielles.
Dans la méthode de Strassen on remplace pour calculer le produit
matriciel h2, 2, 2i le calcul bilinéaire avec 8 multiplications (correspondant à la définition du produit) par un calcul bilinéaire avec seulement
7 multiplications essentielles, et on obtient ω ≤ 3 log 7/ log 8. Dans la
méthode de Bini, on utilise un produit de matrices à trous dans lesquel
les 6 multiplications qui interviennent dans la définition du produit matriciel peuvent être remplacées (dans un calcul bilinéaire approximatif)
par seulement 5 multiplications essentielles. Cependant au lieu d’aboutir
à ω ≤ 3 log 5/ log 6 comme dans la méthode de Strassen, on a abouti
à ω ≤ 3 log 10/ log 12. Schönhage a pensé qu’il y avait là quelque chose
d’immoral et il a obtenu dans un travail mémorable (voir [82]) l’amélioration décisive suivante.
Théorème 7.6 Si dans un produit de matrices à trous, on est capable
de remplacer, dans un calcul bilinéaire approximatif, les ρ multiplications qui interviennent dans la définition du produit matriciel par seulog θ
lement θ multiplications essentielles, alors ω ≤ 3 log
ρ . En particulier
5
ω ≤ 3 log
log 6 ≤ 2, 695.
Le reste de ce paragraphe est consacré à la preuve de ce théorème,
selon les lignes de [82]. La preuve est faite sur un corps K infini, ce qui est
légitime d’après la proposition 7.3.14. Le plus simple est de commencer
sur un exemple. Nous allons voir directement sur l’exemple de Bini quelle
est la machinerie mise en œuvre par Schönhage. La méthode itérative de
216
7. Multiplication rapide des matrices
Strassen donne des produits matriciels à trous successifs du type suivant
(figures 7.4 et 7.5). Le produit matriciel à trous itéré une fois de la figure
Figure 7.4 – Bini itéré une fois
7.4 peut être obtenu par un calcul bilinéaire approximatif d’ordre 2 et
de rang 52 (au lieu de 62 ). Ceci se démontre comme le point (2) dans la
proposition 7.4.5. Le produit matriciel à trous itéré deux fois (figure 7.5)
Figure 7.5 – Bini itéré deux fois
peut être obtenu par un calcul bilinéaire approximatif d’ordre 3 et de
rang 53 . Ceci se démontre aussi comme le point (2) dans la proposition
7.4.5.
Plus généralement, la même preuve donne.
Proposition 7.4.7 Notons ϕ l’application bilinéaire qui correspond à
un produit matriciel à trou (A, B) 7→ A B où certaines entrées fixées
de A et B sont nulles et les autres sont des variables inépendantes.
Notons ϕ⊗k le produit matriciel à trou obtenu en itérant k − 1 fois le
produit ϕ. Alors on a
R ϕ⊗k ≤ (R(ϕ))k
7.4. Calculs bilinéaires approximatifs
217
et
R ϕ⊗k , kq ≤ (R(ϕ, q))k
De même si on emboı̂te dans un produit par blocs deux produits matriciels
à trous ϕ et ψ et qu’on note ϕ ⊗ ψ le produit matriciel à trous que
l’on obtient, on a les inégalités
R (ϕ ⊗ ψ) ≤ R(ϕ) R(ψ)
et
R ϕ ⊗ ψ, q + q 0 ≤ R(ϕ, q) R(ψ, q 0 )
Revenons à notre exemple. Dans le produit de la figure 7.5 nous pouvons sélectionner les 3 colonnes 2, 3, 5 de la première matrice, qui contiennent chacune 4 véritables entrées et les lignes 2, 3, 5 de la deuxième,
qui contiennent 8 entrées. On obtient le produit à trous U × V = W
suivant (figure 7.6). Du point de vue du calcul bilinéaire approximatif,
cette extraction de lignes et de colonnes revient simplement à remplacer des variables par des 0 et donc ne peut que le simplifier. Si nous
Figure 7.6 – Produit à trous extrait de (( Bini itéré deux fois ))
considérons maintenant une matrice fixée G ∈ K4×8 l’application linéaire µG : U 7→ GU est en fait une application linéaire entre deux espaces
vectoriels de dimension 12. Admettons un moment que les coefficients
de G peuvent être choisis de manière que µG soit un isomorphisme
(lemme de compression). En posant GU = U 0 on voit que le produit
matriciel sans trou h4, 3, 8i est réalisé sous forme (U 0 , V ) 7→ U 0 × V
par un calcul bilinéaire approximatif d’ordre 3 et de rang 53 : décrypter
218
7. Multiplication rapide des matrices
U 0 pour obtenir U (sans aucune multiplication essentielle) puis calculer
U ×V.
De manière plus générale, nous pouvons considérer le produit à trous
Ak × Bk obtenu en itérant k − 1 fois le processus de Bini. La matrice
Ak est de plus en plus creuse. Dans chaque colonne, le nombre d’entrées
véritables est égal à une puissance de 2. En sélectionnant les colonnes
ayant un même nombre d’entrées (disons mk colonnes avec 2hk entrées
non nulles), on obtient un produit à trous Uk × Vk = Wk à l’intérieur du
format 2k , mk , 2k . Chaque colonne de Uk a exactement 2hk entrées
véritables. En appliquant le lemme de compression, nous choisissons une
h
k
matrice convenable Gk ∈ K2 k ×2 nous remplaçons Uk par Uk0 = Gk Uk
et nous obtenons un produit matriciel sans trou 2hk , mk , 2k sous la
forme (Uk0 , Vk ) 7→ Uk0 × Vk par un calcul bilinéaire approximatif d’ordre
k et de rang 5k .
Quel est le comportement asymptotique de ce calcul ? On peut facilement se convaincre que le produit mk 2hk est obtenu comme l’un des
termes du développement de (1 + 2)k selon la formule du binôme. Cela
tient à ce que la matrice à trous initiale possède une colonne à deux
entrées et une autre à une entrée. Comme la formule du binôme est une
somme de (k + 1) termes, le plus grand de ces termes est certainement
supérieur à 3k /(k + 1). Donc par un choix optimal de hk nous obtenons
Nk = 2hk · mk · 2k ≥
6k
k+1
Donc en appliquant la proposition 7.4.7
R (hNk , Nk , Nk i , 3k) ≤ 53k
d’où en appliquant le lemme 7.4.4
R hNk , Nk , Nk i ≤ 9k 2 53k
5
ce qui donne bien par passage à la limite ω ≤ 3 log
log 6 .
Avant de passer à la preuve dans le cas général, nous montrons le
lemme de compression.
Lemme 7.4.8 (lemme de compression) Soit A = (aij ) une (( matrice
à trous )) de format m × n dont les entrées sont ou bien nulles ou bien
des variables indépendantes. Nous supposons que la matrice possède p
variables et m − p entrées nulles dans chaque colonne. Si on spécialise
les variables dans le corps K on obtient un espace vectoriel EA de
7.4. Calculs bilinéaires approximatifs
219
dimension n p. On suppose le corps K infini. Alors il existe une matrice
G ∈ Kp×m telle que l’application linéaire
µG,A : EA −→ Kp×n ,
U 7−→ G U
soit bijective.
Preuve. Les colonnes de U sont transformées indépendamment les unes
des autres. Chaque colonne Uj de la matrice U , en ne gardant que les
entrées non nulles, est transformée selon le schéma
Uj 7−→ Gj Uj
où Gj est une matrice carrée extraite de G en ne gardant que p colonnes de G. Il s’ensuit que l’application linéaire µG,A est bijective si et
seulement si les matrices Gj sont inversibles. Pour cela, il suffit que p
colonnes distinctes de G soient toujours indépendantes. Ce problème de
géométrie combinatoire admet toujours une solution sur un corps ayant
suffisamment d’éléments. Si on a déjà construit une matrice convenable
G avec ` ≥ p colonnes, pour rajouter une colonne, il faut choisir un
vecteur en dehors des hyperplans définis par n’importe quel système de
p − 1 colonnes extraites de G.
t
u
Passons maintenant à la preuve du cas général. Nous supposons que
nous avons un produit de matrice à trous A × B par exemple du style
suivant (figure 7.7) qui peut être réalisé par un calcul bilinéaire approxi-
Figure 7.7 – Un exemple arbitraire de produit matriciel à trous
matif de manière économique. Supposons que les colonnes successives de
A, au nombre de t contiennent respectivement m1 , m2 , . . ., mt entrées
véritables. Supposons que les t lignes successives de B contiennent
respectivement n1 , n2 , . . ., nt entrées véritables. A priori ce produit à
trous réclame
ρ = m1 n1 + · · · + mt nt
multiplications : le tenseur qui correspond à sa définition est une somme
de ρ tenseurs élémentaires (dans l’exemple ci-dessus, t = 3, (m1 , m2 ,
220
7. Multiplication rapide des matrices
m3 ) = (2, 2, 1), (n1 , n2 , n3 ) = (1, 2, 4) et ρ = 10). Supposons qu’un calcul bilinéaire approximatif à l’ordre ` et de rang θ permette de réaliser
ce produit à trous. Si on itère k − 1 fois ce calcul bilinéaire approximatif, on obtient un nouveau produit de matrices à trous Ak × Bk . Par
exemple pour A2 × B2 , on obtient le produit à trous suivant (figure 7.8).
Une colonne de Ak doit être indexée par un k-uple j = (j1 , . . . , jk )
Figure 7.8 – L’exemple précédent itéré une fois
d’éléments de {1, . . . , t}. Une telle colonne contient alors
mj1 · · · mjk = mu1 1 · · · mut t
entrées non nulles, où chaque ui est égal au nombre des js égaux à i. De
même une ligne de Bk doit être indexée par un k-uple j = (j1 , . . . , jk )
d’éléments de {1, . . . , t} et elle contient nj1 · · · njk entrées non nulles.
Parmi toutes les colonnes de Ak on décide de sélectionner toutes celles
qui fournissent une certaine liste d’exposants (u1 , . . . , ut ). En particulier
elles ont toutes le même nombre µk = mu1 1 · · · mur t d’entrées non nulles
(avec u1 + · · · + ut = k). Le nombre des colonnes en question est égal
au coefficient multinomial
k
k!
λk =
=
u1 , . . . , ut
u1 ! · · · ut !
De même, nous sélectionnons parmi les lignes de Bk toutes celles correspondant aux mêmes indices (qui sont des k-uples j = (j1 , . . . , jk )).
Elles ont toutes le même nombre d’entrées non nulles νk = nu1 1 · · · nut t .
Nous obtenons de cette manière un produit de matrice à trous Uk × Vk .
Comme les colonnes de Uk ont toutes le même nombre µk d’entrées
non nulles, on peut utiliser le lemme de compression. Même chose pour
Vk en tenant compte du fait que toutes les lignes ont le même nombre
7.4. Calculs bilinéaires approximatifs
221
νk d’entrées non nulles. En définitive nous obtenons un produit matriciel sans trou de type hµk , λk , νk i qui est réalisé par un calcul bilinéaire
approximatif d’ordre k` et de rang ≤ θk .
A quoi est égal µk · λk · νk ? C’est l’un des termes du développement
multinomial de (m1 n1 + · · · + mt nt )k = ρk . Si on choisit le terme le plus
grand dans cette somme on obtient donc
Mk = µk · λk · νk ≥
ρk
≥
k+t
t−1
ρk
(k + 1)t−1
car il y a k+t
t−1 termes dans cette somme. On termine comme dans le
cas particulier examiné au début :
R (hMk , Mk , Mk i , 3k`) ≤ θ3k ,
R hMk , Mk , Mk i ≤ 9 k 2 `2 θ3k
log θ
et par passage à la limite en appliquant le théorème 7.4, ω ≤ 3 log
ρ.
Remarque 7.4.9 Dans [82] Schönhage indique des produits matriciels
à trous avec un rang marginal plus avantageux que celui de Bini, ce
qui donne ω ≤ 2, 6087. Mais ce dernier résultat est surpassé par la
formule asymptotique qu’il obtient ensuite et que nous exposons dans le
paragraphe suivant.
Remarque 7.4.10 Dans le lemme 7.4.4 il est possible de remplacer (q+
1)(q + 2)/2 par 1 + 6q. Même avec cette amélioration, c’est uniquement
pour des entiers N très grands que la procédure de Bini aussi bien
que celle de Schönhage fournissent un meilleur calcul bilinéaire pour
hN, N, N i que celui qui découle de la procédure originale de Strassen.
Ces méthodes ne sont donc pas implémentées sur machine.
7.4.3
Sommes directes d’applications bilinéaires
Approfondissant son analyse des produits de matrices à trous, Schönhage a remarqué que certains produits du type ci-dessous (figure 7.9)
permettent de construire à partir d’un calcul bilinéaire approximatif des
calculs bilinéaires exacts donnant un meilleur exposant pour la multiplication des matrices carrées que celui établi dans le théorème 7.6.
L’exemple de la figure 7.9 correspond à la somme disjointe (on peut
dire aussi somme directe ou encore juxtaposition) des deux applications
bilinéaires h1, 2, 1i et h3, 1, 3i. De manière générale, la somme directe de
222
7. Multiplication rapide des matrices
Figure 7.9 – Somme directe de deux produits matriciels
deux applications bilinéaires ϕ1 : E1 × F1 → G1 et ϕ2 : E2 × F2 → G2
est l’application bilinéaire
ϕ : (E1 ⊕ E2 ) × (F1 ⊕ F2 ) −→ (G1 ⊕ G2 )
définie par ϕ((x1 , y1 ), (x2 , y2 )) = (ϕ1 (x1 , y1 ), ϕ2 (x2 , y2 )). Du point de
vue des calculs bilinéaires, un calcul bilinéaire possible pour la somme
disjointe consiste à faire seulement les deux calculs en parallèle avec
toutes les variables distinctes.
Notation 7.4.11 On note ϕ1 ⊕ ϕ2 la somme directe des applications
bilinéaires ϕ1 et ϕ2 . On note ` ϕ pour la somme directe de ` exemplaires de ϕ.
On fait alors les remarques suivantes. Le premier lemme est à la fois
simple et crucial.
Lemme 7.4.12 Supposons R hf, f, f i ≤ s et R(s
Alors R hf m, f n, f pi ≤ r.
hm, n, pi) ≤ r.
Preuve. L’application bilinéaire hf m, f n, f pi peut être réalisée comme
un produit par blocs, chacune des deux matrices A et B qu’on multiplie
étant découpée en f 2 blocs de même format. Les f 3 multiplications
correspondantes de type hm, n, pi qui sont a priori nécessaires pour ce
produit par blocs peuvent être remplacées par seulement s produits
(entre combinaisons linéaires convenables des blocs), selon le schéma
fourni par le calcul bilinéaire qui montre R hf, f, f i ≤ s.
t
u
Lemme 7.4.13
(1) R(ss0 hmm0 , nn0 , pp0 i) ≤ R(s hm, n, pi) · R(s0 hm0 , n0 , p0 i). En
particulier avec N = mnp on a
R(s3 hN, N, N i) ≤ (R(s hm, n, pi))3
(2)
R(ss0 hmm0 , nn0 , pp0 i , q+q 0 ) ≤ R(s hm, n, pi , q)·R(s0 hm0 , n0 , p0 i , q 0 ).
En particulier avec N = mnp on a
R(s3 hN, N, N i , 3q) ≤ (R(s hm, n, pi , q))3
7.4. Calculs bilinéaires approximatifs
223
Preuve. C’est toujours la méthode du produit par blocs, appliquée avec
les produits matriciels à trous correspondants. On peut considérer qu’il
s’agit d’un cas particulier de la proposition 7.4.7.
t
u
On en déduit la proposition suivante qui généralise le théorème 7.5.
Proposition 7.4.14 S’il existe s, m, n, p, r tels que R(s hm, n, pi) ≤
r alors s (mnp)ω/3 ≤ r, c’est-à-dire ω ≤ 3 log(r/s)
log mnp .
Autrement dit, pour ce qui concerne l’exposant ω, l’inégalité R(s
hm, n, pi) ≤ r donne le même résultat qu’une inégalité R hm, n, pi ≤
r/s.
Preuve. Si R(s hm, n, pi , q) ≤ r, en appliquant le lemme 7.4.13 on
obtient avec N = mnp
R s3 hN, N, N i , 3q ≤ r3
puis aussi
R s3`
et donc
R s3`
D
E
N ` , N ` , N ` , 3`q ≤ r3`
D
N `, N `, N `
E
≤ 9 `2 q 2 r3` .
(7.3)
Par passage à la limite, cela nous ramène au cas où on connaı̂t des
entiers s, m, r tels que R(s hm, m, mi) ≤ r. On veut alors montrer
log(r/s)
ω ≤ log(r/s)
log m . Posons λ = log m .
Supposons tout d’abord qu’on connaisse un calcul bilinéaire qui montre
que R hf, f, f i ≤ s et posons α0 = log s/ log f (α0 est un exposant
acceptable). Si α0 ≤ λ on n’a rien à faire. Si α0 > λ le lemme 7.4.13
nous dit que R hf m, f m, f mi ≤ r. Donc l’exposant
α1 = log r/ log f m = α0 log r/(log s + α0 log m)
est acceptable pour la multiplication des matrices carrées. Un calcul
simple montre alors que λ < α1 < α0 . Donc nous avons amélioré la
situation en passant de α0 à α1 .
Nous voyons maintenant le travail qui nous reste à faire.
Primo, montrer que si on a R hf, f, f i ≤ s0 = f α0 avec un entier s0 6= s,
cela n’est pas trop grave, car on peut utiliser R s`
m` , m` , m` ≤ r`
α
0
avec s0k ≤ s` et le rapport de leurs
et R f k , f k , f k ≤ s0k = f k
logarithmes aussi proche qu’on veut de 1. Donc par le lemme 7.4.13
R f k m` , f k m` , f k m` ≤ r` , ce qui conduit à un exposant acceptable
α10 = log r`/ log f k m` = α00 log r/(log s + α00 log m)
224
7. Multiplication rapide des matrices
avec α00 aussi proche qu’on veut de α0 .
Secundo, montrer que si on recommence, les exposants successifs αn
qu’on obtient convergent bien vers λ.
Nous ne ferons pas ce travail, car les détails techniques deviennent vraiment trop lourds.
t
u
La conjecture additive de Strassen
On a évidemment
R(ϕ1 ⊕ ϕ2 ) ≤ R(ϕ1 ) + R(ϕ2 )
et
R(ϕ1 ⊕ ϕ2 , q) ≤ R(ϕ1 , q) + R(ϕ2 , q) .
Une conjecture de Strassen est que la première inégalité est en fait toujours une égalité. On appelle cette conjecture la conjecture additive (pour
le rang tensoriel des applications bilinéaires). Bien que plausible, cette
conjecture a été ébranlée par Schönhage qui a montré que la variante avec
(( rang tensoriel marginal )) à la palce de (( rang tensoriel )) est fausse,
d’après le résultat du lemme 7.4.15. Si la conjecture additive est vraie, ou
même si seulement R(s hm, m, mi) = s R hm, m, mi pour s et m arbitraires, la preuve de la proposition 7.4.14 est beaucoupsimplifiée, car on
déduit de l’équation 7.3 directement R N ` , N ` , N ` ≤ 9 `2 q 2 (r/s)3` .
Mais cela ne fournirait les calculs bilinéaires demandés que si on était capable de trouver un calcul bilinéaire de rang convenable pour hm, m, mi
à partir d’un calcul bilinéaire pour s hm, m, mi .
Lemme 7.4.15 Pour k > 1 et m = (k − 1)2 :
R(hk, 1, ki ⊕ h1, m, 1i) = k 2 + 1 < k 2 + m = R hk, 1, ki + R h1, m, 1i
Preuve. Nous montrons seulement R(hk, 1, ki ⊕ h1, m, 1i) ≤ k 2 + 1.
Nous représentons le produit hk, 1, ki par le polynôme non commutatif
P
k
i=1 ai bj cj,i et le produit h1, m, 1i par le polynôme non commutatif
Pm
`=1 u` v` w. Pour simplifier les écritures qui suivent, nous prenons ` =
(i, j) avec 1 ≤ i, j ≤ k − 1. Nous introduisons en outre les notations
ui,k = vk,j = 0,
uk,j = −
k−1
X
ui,j ,
vi,k = −
i=1
k−1
X
j=1
de sorte que
m
X
`=1
u` v` =
k−1
X
i=1,j=1
ui,j vi,j =
k
X
i=1,j=1
ui,j vi,j
vi,j
7.4. Calculs bilinéaires approximatifs
225
On considère alors le polynôme non commutatif suivant (qui correspond
à un calcul bilinéaire approximatif avec k 2 + 1 multiplications essentielles)
! k
k
k
X
X
X
(ai + εui,j ) (bj + εvi,j ) (ε2 cj,i + w) −
bj w
ai
i=1,j=1
i=1
j=1
qui, une fois développé donne
k
X
ε2
(ai bj cj,i + ui,j vi,j w) + ε3 Q .
i=1,j=1
t
u
7.4.4
L’inégalité asymptotique de Schönhage
Revenons au produit à trou de la figure 7.9 qui représente la juxtaposition h2, 1, 2i ⊕ h1, 3, 1i. Si nous itérons une fois (à la Strassen)
ce produit à trous, nous obtenons un nouveau produit à trou correspondant à la figure 7.10, qui peut être réorganisé, par changement de
Figure 7.10 – Somme directe, itérée une fois, de deux produits matriciels
numérotation des lignes et colonnes, en le produit à trou qui correspond
à la figure 7.11, et nous voyons clairement que cela signifie
(h2, 1, 2i ⊕ h1, 3, 1i)⊗2 ' h4, 1, 4i ⊕ 2
h2, 3, 2i ⊕ h1, 9, 1i
226
7. Multiplication rapide des matrices
Figure 7.11 – Somme directe, itérée une fois et réorganisée
Le lecteur ou la lectrice est invitée à réaliser par elle-même l’itération
une deuxième fois, et à vérifier que
(h2, 1, 2i ⊕ h1, 3, 1i)⊗3 ' h8, 1, 8i ⊕ 3
h4, 3, 4i ⊕ 3
h2, 9, 2i ⊕ h1, 27, 1i
avec la parenté évidente avec la formule du binôme. Cette parenté n’est
pas un hasard. C’est bien la même machinerie combinatoire qui est à
l’œuvre dans les deux cas. En itérant k − 1 fois on obtiendra
⊗k
(h2, 1, 2i ⊕ h1, 3, 1i)
'
k
X
k
i=1
i
D
2i , 3k−i , 2i
E
P
où le
indique une somme disjointe d’applications bilinéaires. En fait
nous avons une formule du multinôme générale, où les sommes indiquent
des sommes disjointes d’applications bilinéaires, et où l’isomorphisme
correspond à une organisation convenable des lignes et colonnes du produit matriciel à trous correspondant au premier membre
t
X
i=1
!⊗k
hmi , ni , pi i
'
X
(u1 ,...,ut )
k
u1 , . . . , ut
*Y
t
i=1
mui i ,
t
Y
i=1
nui i ,
t
Y
+
pui i
i=1
(la deuxième somme est prise sur tous les t-uples (u1 , . . . , ut ) tels que
P
ui = k). On en déduit la formule asymtotique suivante.
7.4. Calculs bilinéaires approximatifs
227
Théorème 7.7 (formule asymptotique de Schönhage) Supposons qu’on
ait
!
t
t
M
X
R
hmi , ni , pi i ≤ r et
(mi ni pi )β = r
i=1
i=1
Alors on obtient pour l’exposant de la multiplication des matrices carrées
ω ≤ 3 β.
Preuve. Notons d’abord que le théorème 7.6 donne
2≤ω≤
log r
Pt
log
i=1 mi ni pi )
donc log r ≥ 2 log t. En appliquant la formule du multinôme et l’inégalité
de la proposition 7.4.7 on obtient
+!
*Y
t
t
t
Y
Y
k
R
mui i ,
nui i ,
pui i
≤ rk
u1 , . . . , ut
i=1
i=1
i=1
Pour un choix particulier de u1 , . . . , ut , nous notons ceci sous la forme
R(Sk
hMk , Nk , Pk i) ≤ rk
ce qui nous donne, d’après la proposition 7.4.14
ω≤3
log(rk /Sk )
.
log Mk Nk Pk
Quel est le choix optimal de u1 , . . . , ut ? Nous considérons l’égalité
!k
!β
Y
t
t
X
X
k
(mi ni pi )β
= rk =
mui i nui i pui i
u1 , . . . , ut
i=1
(u1 ,...,ut )
i=1
rk
≤ (k+1)
La somme de droite a k+t−1
t−1 termes et donc pour le plus
t−1
grand d’entre eux on obtient
!β
Y
t
k
rk
ui ui ui
mi ni pi
= Sk (Mk Nk Pk )β ≥
(k + 1)t−1
u1 , . . . , ut
i=1
ce qui donne
log rk /(Sk (k + 1)t−1 )
(t − 1) log(k + 1)
ω≤3
≤ 3β + 3
log(Mk Nk Pk )
log(Mk Nk Pk )
D’où le résultat par passage à la limite : β log(Mk Nk Pk ) est équivalent
à log rk /Sk , et on a Sk < tk et r ≥ t2 .
t
u
228
7. Multiplication rapide des matrices
Corollaire 7.4.16 L’exposant de la multiplication des matrices carrées
vérifie ω ≤ 2, 5479.
Preuve. On applique la formule asymptotique avec la somme disjointe
h4, 1, 4i ⊕ h1, 9, 1i du lemme 7.4.15.
t
u
8. Algèbre linéaire
séquentielle rapide
Introduction
Une conséquence importante de la multiplication rapide des matrices
est la recherche de méthodes de calcul permettant de ramener les problèmes classiques d’algèbre linéaire à une complexité algébrique du même
ordre que celle de la multiplication des matrices.
Bien que nous utilisions systématiquement la multiplication rapide
des matrices, qui est obtenue par un algorithme très bien parallélisé, les
algorithmes obtenus dans ce chapitre ne sont pas eux-mêmes bien parallélisés. Leur profondeur est en général en O(n), ce qui explique le titre
du chapitre (algèbre linéaire séquentielle rapide)
Nous avons déjà vu à la section 7.2 que l’inverse d’une matrice triangulaire d’ordre n peut se calculer par une famille uniforme de circuits
arithmétiques de taille O(nα ) et de profondeur O(log2 n).
Nous allons dans ce chapitre montrer que, pour autant qu’on travaille
sur un corps et qu’on ait droit à la division 1 , des familles de circuits
arithmétiques ayant des tailles voisines peuvent être construites pour
résoudre les principaux problèmes de l’algèbre linéaire sur un corps.
Mais dans tous les algorithmes que nous exhiberons, le temps parallèle
(la profondeur du circuit) n’est plus polylogarithmique.
En outre, comme ce sont des circuits avec divisions, ils ne peuvent
pas être exécutés sur toutes les entrées, et nous donnerons en général
une version sous la forme d’un algorithme (( avec branchements )) (les
branchements sont gouvernés par des tests d’égalité à 0 dans le corps).
Dans ces algorithmes (qui ne correspondent plus à des circuits arith1. Ceci est légitime si la division n’est pas trop coûteuse en termes de complexité
binaire.
230
8. Algèbre linéaire séquentielle rapide
métiques proprement dits), nous aurons pour le temps séquentiel et le
temps parallèle des estimations très voisines de celles obtenues pour les
circuits arithmétiques avec divisions.
Par exemple le calcul du déterminant et de l’inverse d’une matrice
carrée (si elle est inversible) peuvent être réalisés par une famille uniforme de circuits arithmétiques avec divisions de taille O(nα ) (voir section
8.2). Ceci est une conséquence de l’algorithme séquentiel de Bunch &
Hopcroft pour la LUP -décomposition que nous développons dans la section 8.1. Cet algorithme se présente naturellement sous la forme d’un
algorithme avec branchements.
En ce qui concerne le calcul du polynôme caractéristique plusieurs
méthodes d’accélération de l’algorithme de Frobenius (section 2.8.1) assez sophistiquées ont été mises au point par Keller-Gehrig. L’algorithme avec branchements, qui utilise un temps séquentiel en O(nα log n)
nécessite au préalable une méthode rapide pour la mise en forme (( échelonnée en lignes )) d’une matrice arbitraire. Ceci est expliqué dans les
sections 8.3 et 8.4.
Dans la dernière section 8.5, nous quittons le cadre de l’algèbre linéaire sur les corps, mais nous restons dans celui de l’algèbre linéaire
séquentielle accélérée grâce à la multiplication rapide des matrices. Nous
décrivons la méthode de Kaltofen, inspirée de l’algorithme probabiliste
de Wiedemann, très efficace pour les matrices creuses sur des corps finis.
Elle donne le meilleur temps séquentiel actuellement connu pour le calcul
du déterminant, du polynôme caractéristique et de l’adjointe d’une matrice carrée sur un anneau commutatif arbitraire. L’algorithme utilise la
multiplication rapide des polynômes et celle des matrices. Contrairement
à l’algorithme de Wiedemann, celui de Kaltofen n’a cependant pas encore
fait l’objet d’une implémentation satisfaisante.
Si les algorithmes développés dans ce chapitre sont théoriquement
plus rapides que les algorithmes (( usuels )) donnés au chapitre 2, il y a
encore malheureusement loin de la théorie à la pratique. En fait seule la
première forme de la multiplication rapide des matrices (celle de Strassen, correspondant à α = log 7 ' 2, 807) commence à être implémentée.
Outre la difficulté pratique d’implémenter d’autres algorithmes de multiplication rapide des matrices, les coefficients Cα pour de meilleures
valeurs de α sont trop grands. Leur implémentation ne se révèlerait
efficace que pour des matrices de tailles astronomiques.
8.1. L’Algorithme de Bunch & Hopcroft
8.1
231
L’Algorithme de Bunch & Hopcroft
pour la LUP-décomposition
des matrices surjectives
Dans la section 2.1 nous avons présenté l’algorithme 8 page 63 qui
est l’algorithme séquentiel usuel (par la méthode du pivot de Gauss)
pour la LUP -décomposition des matrices surjectives. La procédure de la
LUP -décomposition que nous allons développer ici fait appel à la multiplication rapide des matrices. Cette procédure , que nous noterons lup,
est due à Bunch et Hopcroft [11].
L’algorithme de Bunch & Hopcroft prend en entrée une matrice de
rang n, A ∈ Kn×p (1 ≤ n ≤ p) et donne en sortie un triplet (L, U, P ) tel
que L est une matrice unitriangulaire inférieure, U une matrice triangulaire supérieure fortement régulière et P une matrice de permutation.
On écrira : lup(A, n, p) = lupn,p (A) = (L, U, P ).
Pour n = 1, A est une matrice ligne de rang 1 : il existe donc un
élément non nul de A occupant la i - ème place de cette ligne (1 ≤ i ≤
p). Il suffit de prendre L = [1] et U = AP où P est la matrice de
permutation d’ordre p correspondant à l’échange des colonnes 1 et i.
On a donc lup(A, 1, p) = ([1], A P, P ) pour la matrice P ainsi définie.
Supposant la propriété vraie pour tout entier n compris entre 1 et
2ν−1 , on la démontre pour 2ν−1 < n ≤ 2ν (ν = dlog ne). On pose
n0 = 2ν−1 , n1 = n − n0 et p1 = p − n0 . Pour obtenir lup(A, n, p) avec
A ∈ Kn×p , on considère la partition suivante de la matrice A :
A1
A=
, A1 ∈ Kn0 ×p , A2 ∈ Kn1 ×p
(8.1)
A2
Si A est une matrice surjective, A1 et A2 le sont également. On commence par appeler lup(A1 , n0 , p) qui donne une LUP -décomposition
(L1 , U1 , P1 ) de A1 . On considère alors les partitions suivantes des matrices U1 et A2 P1−1 :
(
U1 = [ V1 | B ] ∈ Kn0 ×p et A2 P1−1 = [ C | D ] ∈ Kn1 ×p
(8.2)
V1 ∈ Kn0 ×n0 , B ∈ Kn0 ×p1 , C ∈ Kn1 ×n0 , D ∈ Kn1 ×p1
V1 étant triangulaire supérieure et inversible (puisque U1 est fortement
régulière). Posant C1 = C V1−1 et E = D − C1 B, on vérifie que :
L1 0
V1 B
A=
P1 .
(8.3)
C1 In1
0 E
232
8. Algèbre linéaire séquentielle rapide
Comme la matrice E satisfait à l’hypothèse de récurrence (elle est surjective puisque A l’est), on peut appliquer la procédure lup(E, n1 , p1 )
qui donne la LUP -décomposition E = L2 U2 P2 dans laquelle U2 est
une matrice n1 × p1 triangulaire
supérieure
fortement régulière.
In0 0
Il suffit de poser Q =
et B2 := B P2−1 pour obtenir la
0 P2
décomposition :
L1 0
V 1 B2
A=
Q P1 .
(8.4)
C1 L2
0 U2
Ce qui donne lup(A, n, p) = (L, U, P ) avec :
V1 B2
L1 0
, et P = Q P1 .
, U=
L=
0 U2
C1 L2
En résumé on obtient le schéma récursif de l’algorithme 8.1.
Algorithme 8.1 : lup(n,p) , LUP-décomposition à la Bunch &
Hopcroft pour une matrice surjective.
Entrée : Une matrice surjective A ∈ Kn×p (K est un corps).
Sortie : Les matrices L, U, P de la LUP-décomposition de A .
Début On utilise la partition donnée en (8.1)
Étape 1 : récurrence avec A1 ∈ Kn0 ×p , ν = dlog ne, n0 = 2ν−1 .
(L1 , U1 , P1 ) := lup(n0 ,p) (A1 )
Étape 2 : pas d’opération arithmétique ici
B2 := A2 P1−1 avec A2 ∈ Kn1 ×p , n1 = n − n0 .
Étape 3 : inversion d’une matrice triangulaire supérieure régulière
V2 := V1−1 avec V1 ∈ Kn0 ×n0 (cf. la partition (8.2))
Étapes 4, 5, 6 :
C1 := C V2 ; F := C1 B ; E := D − F .
Étape 7 : récurrence avec E ∈ Kn1 ×p1
(L2 , U2 , P2 ) := lup(n1 ,p1 ) (E)
Étape 8 : pas d’opération arithmétique ici
B2 := B P2−1 , P := Q P1 .
Fin.
L’algorithme obtenu est un algorithme avec branchements. Ceci est
inévitable puisque la sortie P dépend de manière discontinue de l’entrée
8.1. L’Algorithme de Bunch & Hopcroft
233
A. Les branchements sont tous commandés par le test d’égalité à 0
dans le corps K. Notons τ (n, p) le nombre d’opérations arithmétiques
exécutées par cet algorithme pour les matrices A ∈ Kn×p et π(n, p) son
temps parallèle arithmétique, c’est-à-dire sa profondeur si on ne prend
pas en compte les étapes de recherche d’éléments non nuls ni les produits
d’une matrice par une matrice de permutation.
On a alors en suivant le schéma récursif 8.1 page ci-contre les inégalités suivantes.
Tout d’abord concernant le nombre d’opérations arithmétiques :
τ (n0 , p)
+ 4 Cα n0 α + Cα n0 α +
(8.5)
τ (n, p) ≤
dp1 /n0 e Cα n0 α +
p1 n1
+ τ (n1 , p1 )
Le terme p1 n1 correspond à la soustraction E := D − F et le
terme Cα n0 α dp1 /n0 e correspond au calcul du produit C1 B dans lequel
B ∈ Kn0 ×p1 et C1 ∈ Kn1 ×n0 : on peut toujours compléter C1 par des
lignes de 0 pour en faire une matrice carrée et on découpe B en dp1 /n0 e
blocs carrés (après lui avoir éventuellement rajouté des colonnes de 0) ;
on effectue alors en parallèle dp1 /n0 e multiplications dans Kn0 ×n0 .
Ensuite, concernant le temps parallèle arithmétique, on obtient de
la même manière en utilisant le résultat de l’inversion des matrices triangulaires (section 7.2) :
π(n, p) ≤ π(n0 , p) + Kα [(ν − 1)2 + 5 (ν − 1) + 2] + 2 + π(n1 , p1 ) (8.6)
On en déduit précisément :
Théorème 8.1 La LUP-décomposition d’une matrice surjective de type
(n, p) sur K peut être effectuée par un algorithme (avec branchements)
qui exécute un nombre d’opérations arithmétiques égal à τ (n, p) en
temps parallèle (arithmétique) π(n, p) majorés par
τ (n, p) ≤
lpm
1
1 lpm 2
γα (
+1) nα +
n log n et π(n, p) ≤ 4 (5 Kα +1) n
2
n
2 n
1
où γα = Cα max 4 , α−2
·
2
−1
Notons que pour p = n , la taille du circuit correspondant à l’algorithme
de Bunch & Hopcroft est exactement majorée par γα nα + 12 n2 log n .
234
8. Algèbre linéaire séquentielle rapide
Preuve.
Le calcul de lup(A, n, p) se fait de manière récursive. Nous donnons les
majorations pour le cas où n = 2ν , et il est clair que si n < 2ν , le calcul
ne peut être que plus rapide.
Pour le temps parallèle arithmétique on a π(1, p) = 0 donc, vue la
récurrence (8.6), le résultat ne dépend pas de p et
π(n, p) = π(n) = π(2ν ) ≤ 2 π(2ν−1 ) + Kα [(ν − 1)2 + 5 (ν − 1) + 2] + 2 .
La relation de récurrence f (ν) = 2f (ν−1)+c [(ν−1)2 +5 (ν−1)]+2 (c+1)
avec f (0) = 0 est résolue par Maple en
f (ν) = (10 c + 2) (2ν − 1) − cν 2 − 7 c ν
majoré par (10 c + 2) 2ν = 2 (5 c + 1) n , ce qui donne le résultat.
Pour calculer le nombre d’opérations arithmétiques on pose rν = r =
dp/ne et on suppose sans perte de généralité que p = rn. L’inégalité
(8.5) se réécrit, puisque p1 = (2r − 1)n0 :
(
τ (2ν , p) ≤
τ (2ν−1 , p) + (4Cα + Cα + (2rν − 1) Cα ) 2(ν−1)α
+(2 r − 1) 22(ν−1) + τ (2ν−1 , p1 )
ce qui donne :
τ (2ν , p) ≤ 2 τ (2ν−1 , p) + (4Cα + 2 r Cα ) 2(ν−1)α + 22ν−1 r .
Dans le déroulement récursif de l’algorithme, lorsqu’on traite les matrices
de type 2 2κ × p , on a rκ = 2ν−κ rν . Et donc en ramenant à r = rν on
obtient les inégalités :
τ (2κ , p) ≤ 2 τ (2κ−1 , p) + 2(κ−1)α (4Cα + 2ν−κ+1 r Cα ) + 22κ−1 2ν−κ r .
Sachant que τ (1, p) = 0, on obtient par sommation (et simplification
de la solution d’une relation de récurrence) la majoration suivante :
τ (n, p) ≤
1 2
Cα na r
Cα na
n r log n + 2 a
+4 a
·
2
2 −4
2 −2
Ce qui donne le résultat annoncé.
2. La majoration vaut aussi pour les matrices de type 2κ × p0 avec p0 ≤ p.
t
u
8.2. Calcul du déterminant et de l’inverse
8.2
235
Calcul du déterminant et de l’inverse
d’une matrice carrée
La LUP -décomposition précédente permet un calcul séquentiel rapide du déterminant et de l’inverse d’une matrice carrée (inversible) en
ramenant ces problèmes à la multiplication rapide des matrices carrées
d’ordre n.
En effet, si l’on passe par la LUP -décomposition, le calcul du déterminant d’une matrice A ∈ Kn×n s’effectue avec le même ordre de complexité séquentielle que la multiplication des matrices n × n puisque si
A = L U P , alors det A = detU ce qui revient à calculer le produit
des n éléments diagonaux de la matrice triangulaire U ( = ±1 est la
signature de la permutation représentée par la matrice P ). Il y a donc,
après la LUP -décomposition, un calcul supplémentaire en SD(n, log n)
(par un circuit binaire équilibré).
Il en est de même pour le calcul de l’inverse de A, quand elle est
inversible, puisque A−1 = P −1 U −1 L−1 ce qui revient, en plus de la
LUP -décomposition, à inverser deux matrices triangulaires ( U et L )
et à effectuer un produit de matrices n × n.
A priori les algorithmes de calcul du déterminant et de l’inverse tels
que nous venons de les décrire sont des algorithmes avec branchements.
Dans cette perspective, le coût de la recherche des éléments non nuls
comme celui des permutations de lignes ou de colonnes, c’est-à-dire des
multiplications à gauche ou à droite par une matrice de permutation,
n’est pas pris en considération dans les comptes d’opérations arithmétiques aussi bien du point de vue de leur nombre total que de celui de
leur profondeur.
Néanmoins, on peut aussi prendre le point de vue selon lequel nous
avons construit des familles uniformes de circuits arithmétiques avec
divisions, qui calculent des fractions rationnelles formelles en les coefficients de la matrice donnée au départ. Il n’y a alors pas de LUP -décomposition mais seulement une LU -décomposition, sans aucun branchement. Naturellement la contrepartie est que l’algorithme ne peut pas
être exécuté concrètement sur un corps avec une matrice arbitraire. C’est
seulement pour une (( matrice générique )) que le circuit arithmétique
fonctionne : une telle matrice est une matrice qui, lorsqu’on lui applique
l’algorithme avec branchements, subit tous les tests x = 0 ? en donnant
une réponse négative.
Dans nos énoncés nous adoptons de préférence ce second point de vue.
236
8. Algèbre linéaire séquentielle rapide
Proposition 8.2.1 Le calcul du déterminant d’une matrice carrée d’ordre n sur un corps K est réalisé par une famille uniforme de circuits
arithmétiques avec divisions en SD(nα , n) .
Les constantes asymptotiques sont respectivement majorées par γα pour
la taille O(nα ) et par 4 (5 Kα +1) pour la profondeur O(n) (les mêmes
majorations des constantes que celles données au théorème 8.1).
Proposition 8.2.2 L’inversion d’une matrice carrée d’ordre n sur un
corps K est un problème résolu par une famille uniforme de circuits
arithmétiques avec divisions en SD(nα , n) avec la même estimation que
celle de la proposition 8.2.1 pour la constante asymptotique de la profondeur O(n) , et une constante asymptotique majorée par ζα = γα + 9 Cα
pour la taille O(nα ) .
Dans la constante ζα de la proposition ci-dessus, le terme γα correspond
à la LUP -décomposition et le terme 9 Cα à l’inversion de deux matrices
triangulaires suivie de la multiplication de deux matrices carrées.
8.3
Forme réduite échelonnée en lignes
Dans cette section nous donnons un aperçu sur une méthode récursive
permettant de réduire les matrices à coefficients dans un corps commutatif K, à la forme échelonnée en lignes avec une complexité séquentielle
du même ordre que celle de la multiplication des matrices.
Étant donnée une matrice A de type (n, p) sur K, la réduction de
A à la forme échelonnée en lignes consiste à transformer A, en ayant
exclusivement recours à des transformations élémentaires unimodulaires sur les lignes 3 , en une matrice de de type (n, p) sur K avec un
nombre de zéros strictement croissant apparaissant à gauche des lignes
successives de la matrice réduite. Si l’on note E la matrice unimodulaire correspondant à ces transformations 4 , cela revient à multiplier la
matrice A à gauche par la matrice E.
3. Rappelons (cf. page 55) qu’il s’agit d’une part de la transformation qui consiste
à ajouter à une ligne une combinaison linéaire des autres et d’autre part des échanges
signés de lignes du type (Li , Lj ) ← (Lj , −Li ).
4. C’est-à-dire la matrice obtenue en faisant subir à la matrice unité d’ordre n
les mêmes transformations.
8.3. Forme réduite échelonnée en lignes
237
Prenons par exemple la matrice carrée d’ordre 6
A=
1
1
1
1
1
3
2
2
2
1
3
6
3
3
3
2
4
9
4
6
2
1
7
10
5
7
3
1
9
11
6
8
1
1
6
20
.
On peut la réduire à la forme échelonnée en lignes en effectuant
des transformations du style pivot de Gauss sur les lignes. Ces mêmes
transformations, effectuées sur les lignes de la matrice unité d’ordre 6,
donnent la matrice unimodulaire E qui résume ces transformations :
E=
1
−1
−1
−2
−2
−4/5
0
0
0
0
0
1
0
0
1
−1
0
1
0
1
0
0
1
−3/5
0
0
0
0
1
−3/5
0
0
0
1
0
0
.
La matrice réduite échelonnée en lignes est alors donnée par le produit :
EA =
1
0
0
0
0
0
2
−1
0
0
0
0
3
−1
0
0
0
0
4
−3
−2
0
0
0
5
−4
−2
−2
0
0
6
−5
−5
7
−5
0
.
Comme nous l’avons fait pour la LUP -décomposition, il s’agit ici de
décrire une version rapide de la méthode du pivot de Gauss sur les
lignes. Mais contrairement à la LUP -décomposition, aucune hypothèse
supplémentaire n’est faite sur la matrice A et aucune permutation de
colonnes n’est permise. En contrepartie, dans la décomposition A = F U
qui résulte de cette méthode de réduction ( F = E −1 ), la matrice F
possède seulement la propriété d’être unimodulaire.
La forme échelonnée en lignes trouve sa justification et son application dans des problèmes comme la résolution des systèmes d’équations
linéaires ou la détermination d’une base pour un sous-espace de Kn
défini par un système générateur. Elle sera aussi utilisée dans la section
8.4 pour le calcul rapide du polynôme caractéristique sur un corps. La
méthode que nous allons exposer ci-dessous est due à Keller-Gehrig [58]
et elle est reprise dans [BCS].
238
8. Algèbre linéaire séquentielle rapide
Description de la procédure rapide
On considère une matrice A ∈ Kn×p . Pour la réduire à la forme
échelonnée en lignes, on peut supposer sans perte de généralité que n =
p = 2ν quitte à compléter la matrice A avec suffisamment de lignes
et/ou colonnes de zéros 5 .
La procédure principale que nous noterons Fel utilise les procédures
auxiliaires Fel1 , Fel2 et Fel3 suivantes.
Procédure Fel1 :
C’est une procédure récursive qui transforme une matrice carrée A ∈
K2n×n ( n = 2ν ) dont la moitié inférieure est triangulaire supérieure en
une matrice triangulaire supérieure.
A1
Plus précisément, si A =
avec A1 , A2 ∈ Kn×n et A2 trianA2
gulaire supérieure, la procédure Fel1 calcule une matrice unimodulaire
n×n triangulaire supérieure telles
E ∈ SL2n (K)
et une matrice T ∈ K
T
. Utilisant l’approche (( diviser pour gagner )) on divise
que E A =
0
la matrice A donnée en huit blocs 2ν−1 × 2ν−1 (si ν = 0, le traitement
de la matrice A est immédiat) et on applique de manière récursive la
procédure Fel1 aux blocs 2ν−1 × 2ν qui possèdent la même propriété
que A. On obtient, avec des notations évidentes, le déroulement suivant
de la procédure :
A12
A14
A22
A24
0
A11
0
0
0
A13
E1
A21
A11
E2
A013
A11
A13
A21
0
A11
0
E1
A13
−→
0
0
A012
A0014
E4
−→
A0022
0
0
A13
=
0
0
A11
=
0
0
A12
A11 A012
00
A014
E2
−→
0 A14
0
0 A022
A22
A24
0 A24
0
A11 A012
0 A000
14
avec
0
0
0
0
0
A14
A14
et
= E1
A022
A22
0
A12
A12
et
= E2
00
A14
A014
E3
−→
:
;
;
5. Les lignes et les colonnes ajoutées ne feront l’objet d’aucune manipulation.
8.3. Forme réduite échelonnée en lignes
A022
A24
E3
=
Posant
on a bien EA =
T
0
laire supérieure A011
et
E4
A0014
A0022
=
A000
14
0
I2ν−1
0
0
E2 0
E4
0
0
0 E3
0 I2ν−1
0
I2ν−1
0
E=
0
A0022
0
239
A011 A012
où T =
0 A000
14
000
et A14 le sont.
.
0
0
E1
0 ,
0 I2ν−1
est une matrice triangu-
Procédure Fel2 :
Elle prend en entrée une matrice carrée A ∈ Kn×n ( n = 2ν ) et retourne
une matrice unimodulaire E ∈ SLn (K) et une matrice triangulaire supérieure T vérifiant EA = T .
Là encore, on obtient avec l’approche (( diviser pour gagner )) et des notations analogues à celles utilisées précédemment, le déroulement suivant
de la procédure :
A11 A12
A11 A12
E1
−→
A21 A22
A021 A022
E2
−→
A011 A012
0 A0022
E3
−→
A011 A012
0 A000
22
où E1 est la matrice unimodulaire correspondant à l’algorithme Fel2
appliqué de manière récursive à la matrice A21 ( E1 A21 = A021 est donc
une matrice triangulaire supérieure, et l’on pose A022 = E1 A22 ) alors
que les matrices E2 et E3 correspondent
à l’application respective de
A11
l’algorithme Fel1 à la matrice
qui est de type (2ν , 2ν−1 ) et de
A021
l’algorithme Fel2 à la matrice A0022 qui est carrée d’ordre 2ν−1 . Cela se
traduit par le fait que E3 A0022 = A000
22 est triangulaire supérieure et que :
E2
A11
A021
Posant E =
0
A11
où T =
0
=
A011
0
avec
A012
A0022
= E2
A12
A022
.
I2ν−1 0
I2ν−1 0
T
E2
, on a bien EA =
0
E3
0
E1
0
0
A12
est une matrice triangulaire supérieure.
A000
22
240
8. Algèbre linéaire séquentielle rapide
Procédure Fel3 :
Elle prend en entrée une matrice triangulaire supérieure A ∈ Kn×n (avec
n = 2ν ) et donne en sortie une matrice unimodulaire E ∈ SLn (K) et
une matrice S sous forme échelonnée
en lignes
vérifiant EA = S.
A11 A12
On considère la partition A =
en blocs 2ν × 2ν de la
0 A22
matrice A donnée (A11 et A22 sont des matrices triangulaires supérieures).
Le déroulement de la procédure est alors illustré par le schéma suivant
dans lequel c’est d’abord l’algorithme
Fel3 qui est appliqué à la matrice
S11
A11 pour donner la matrice
où S11 est une matrice surjective 6
0
échelonnée
en lignes ; c’est ensuite Fel1 qui est appliqué à la matrice
A23
A023
pour donner la matrice
où A023 est triangulaire suA24
0
périeure ; et c’est enfin Fel3 qui, appliqué à la matrice A023 , donne la
matrice échelonnée en lignes S23 :
A11 A12
0 A22
S11 A012
S11 A012
S11 A012
E
E2
3
0 S23
−→ 0 A23 −→ 0 A023 −→
0
0
0
0
0 A24
E1
avec E1 A12 =
A012
A23
. Si maintenant on pose :
Ir 0 0
Ir 0
E1
0
E = 0 E3 0
0 E2
0 I2ν−1
0 0 Ir
S11 A012
où r est le rang de A11 , alors EA = S avec S = 0 S23 qui est
0
0
bien une matrice échelonnée en lignes puisque S11 et S23 le sont.
Procédure principale Fel :
Elle prend en entrée une matrice carrée A ∈ Kn×n ( n = 2ν ) et retourne
une matrice unimodulaire E ∈ SLn (K) et une matrice S sous forme
échelonnée en lignes vérifiant EA = S.
Le cas ν = 0 est trivial. Pour ν ≥ 1, on applique la procédure auxiliaire
6. Le nombre r de ses lignes est égal à son rang qui est aussi celui de A11 .
8.3. Forme réduite échelonnée en lignes
241
Fel2 pour transformer la matrice A en une matrice triangulaire supérieure T puis la procédure Fel3 pour transformer T en une matrice
échelonnée en lignes.
Analyse de complexité
L’étude de complexié de la procédure principale Fel passe par celle
des trois algorithmes auxiliaires Fel1 , Fel2 et Fel3 . Si l’on désigne par
τ1 , τ2 et τ3 les tailles et par π1 , π2 et π3 les profondeurs respectives
de ces trois algorithmes, on a les majorations suivantes dans lesquelles
les coefficients Cα et Kα sont les constantes intervenant dans la taille
et la profondeur des complexités arithmétiques de la multiplication des
matrices.
pour les tailles :
τ1 (2ν ) ≤ 4 τ1 (2ν−1 ) + 2 Cα 2να
τ2 (2ν ) ≤ τ1 (2ν−1 ) + 2 τ2 (2ν−1 ) + (2α + 1) Cα 2(ν−1)α
τ3 (2ν ) ≤ τ1 (2ν−1 ) + 2 τ3 (2ν−1 ) + Cα 2(ν−1)α
pour les profondeurs :
π1 (2ν ) ≤ 3 π1 (2ν−1 ) + 2 Kα ν
π2 (2ν ) ≤ π1 (2ν−1 ) + 2 π2 (2ν−1 ) + Kα (2ν − 1)
π3 (2ν ) ≤ π1 (2ν−1 ) + 2 π3 (2ν−1 ) + Kα (ν − 1)
E
2
Il faut remarquer que dans la procédure Fel1 les étapes −→
et
E3
−→ peuvent être exécutées en parallèle, ce qui explique la diminution
du coefficient (de 4 à 3) entre τ1 et π1 .
Utilisant les inégalités ci-dessus et le fait que :
τ (2ν ) = τ2 (2ν ) + τ3 (2ν )
et
π(2ν ) = π2 (2ν ) + π3 (2ν ) ,
nous allons montrer le résultat suivant concernant la complexité du problème de la réduction à la forme échelonnée en lignes.
Proposition 8.3.1 La réduction à la forme échelonnée en lignes d’une
matrice carrée d’ordre n sur un corps commutatif K est réalisée par
une famille uniforme de circuits arithmétiques de taille τ (n) et de profondeur π(n) avec les majorations suivantes :
τ (n) ≤
21 Cα
nα
−1
2α−2
et
π(n) ≤ (3 Kα + 2) nlog 3 .
242
8. Algèbre linéaire séquentielle rapide
Preuve. Les sommations des relations τ1 (2k ) ≤ 4 τ1 (2k−1 ) + 2 Cα 2kα
d’une part et des relations π1 (2k ) ≤ 3 π1 (2k−1 ) + 2 Kα k d’autre part
pour k allant de 0 à ν (avec τ1 (1) = π1 (1) = 1 ) donnent les majorations suivantes pour la taille et la profondeur du circuit arithmétique
correspondant à la procédure Fel1 :
3
3
2α+1 Cα να
ν
ν
2
et π1 (2 ) <
Kα + 1 3ν − Kα ν − Kα .
τ1 (2 ) < α−2
2
−1
2
2
Tenant compte de ces relations et du fait que τ2 (1) = π2 (1) = 1, les
sommations pour k allant de 0 à ν des inégalités relatives à la taille
et la profondeur du circuit arithmétique correspondant à la procédure
Fel2 nous donnent la majoration :
τ2 (2ν ) <
1
2να
1
Eα α−1
< Eα 2να
2
2
−1
2
dans laquelle
2α+1
25 Cα
Cα + (2α + 1) Cα < α−2
2α−2 − 1
2
−1
(avec 2 < α ≤ 3) et la majoration :
3
1
ν
ν
ν
π2 (2 ) <
Kα + 1 3 − Kα 2 + ν +
.
2
2
Eα =
On obtient, par des calculs analogues, les majorations suivantes pour la
taille et la profondeur du circuit arithmétique correspondant à la procédure Fel3 :
2να
1
1
Fα α−1
< Fα 2να et
2
2
−1
2
3
3
π3 (2ν ) <
Kα + 1 3ν − Kα (2ν+1 − 1)
2
2
α+1
2
17 Cα
où Fα =
+ 1 Cα < α−2
α−2
2
−1
2
−1
τ3 (2ν ) <
Le résultat annoncé découle des majorations ci-dessus et du fait que l’on
a : τ (2ν ) = τ2 (2ν ) + τ3 (2ν ), π(2ν ) = π2 (2ν ) + π3 (2ν ) et n = 2ν .
t
u
Remarque. Le fait de considérer des matrices carrées dont le nombre
de lignes (et de colonnes) est une puissance de 2 n’est pas une hypothèse
restrictive. On peut en effet plonger toute matrice A ∈ Kn×p dans une
matrice carrée d’ordre 2ν en prenant ν = max (dlog ne , dlog pe) et en
8.4. Méthode de Keller-Gehrig
243
complétant la matrice donnée par 2ν −n lignes et 2ν −p colonnes nulles.
Les rangées ajoutées, formées de zéros, ne subissent aucune transformation au cours du déroulement de la procédure décrite et le résultat énoncé
dans la proposition 8.3.1 reste valable à condition de remplacer n par
max (n, p).
8.4
Méthode de Keller-Gehrig
Les algorithmes de Keller-Gehrig [58] sont des versions accélérées de
l’algorithme de Frobenius que nous avons décrit à la section 2.8.1.
Dans la section présente nous ne décrirons en détail que le plus simple
de ces algorithmes. Nous reprenons les notations de la section 2.8.1.
La matrice A ∈ Kn×n définit l’endomorphisme hA de Kn . Nous
appelons a = (e1 , . . . , en ) la base canonique de Kn .
Accélération dans le cas simple
Nous examinons ici le cas le plus simple (et le plus fréquent) où
k1 = n c’est-à-dire le cas où b = (e1 , Ae1 , . . . , An−1 e1 ) est une base
de Kn .
Nous désignons par [S 0 ]S la matrice d’un système de vecteurs (ou
d’un vecteur ou d’un endomorphisme) S 0 dans une base S. Alors U =
[b]a est la matrice de passage de a à b, et on a :
0 ... 0
a0
.
.
a1
1 . . ..
−1
n−1
[hA ]b = U AU = [(e1 , Ae1 , . . . , A
e1 )]b = . .
..
.. 0
..
.
0 . . . 1 an−1
où a0 , a1 , . . . , an−1 sont les coefficients (dans K ) de la relation de dépendance An e1 = an−1 An−1 e1 + . . . + a1 Ae1 + a0 e1 . Ceci prouve que A est
semblable à une matrice de Frobenius et que son polynôme caractéristique est PA (X) = (−1)n (X n − (an−1 X n−1 + · · · + a1 X + a0 )).
L’algorithme de Keller-Gehrig, dans ce cas le plus simple, consiste à
calculer la matrice U puis le produit U −1 A U pour obtenir par simple
lecture de la dernière colonne les coefficients du polynôme caractéristique de A. Prenant ν = dlog ne, le calcul de U se fait en ν étapes.
L’étape no k ( 1 ≤ k ≤ ν ) consiste à :
k
k−1
— calculer la matrice A2 (élévation au carré de la matrice A2
calculée à l’étape précédente) ;
déjà
244
8. Algèbre linéaire séquentielle rapide
ν−1
k−1
— calculer la matrice A2 [e1 | Ae1 | . . . | A2 −1 e1 ] à partir de la mak−1
trice [e1 | Ae1 | . . . | A2 −1 e1 ] calculée à l’étape k − 1 pour obtenir la
k
matrice [e1 | Ae1 | . . . | A2 −1 e1 ] de l’étape k.
À la fin de ces ν étapes, on obtient la matrice
[e1 | Ae1 | . . . | A2
ν −1
ν
e1 ] ∈ Kn×2
qui admet comme sous-matrice la matrice recherchée
U = [e1 | Ae1 | . . . | An−1 e1 ] ∈ Kn×n
puisque 2ν − 1 ≥ n − 1.
On calcule ensuite la dernière colonne de U −1 A U en commençant
par inverser la matrice U (en passant par sa LUP -décomposition). Enfin on calcule la dernière colonne V de A U en multipliant A par la
dernière colonne de U , puis on calcule U −1 V .
L’analyse de complexité dans ce cas simple nous donne donc :
Proposition 8.4.1 On peut calculer le polynôme caractéristique d’une
matrice carrée d’ordre n à coefficients dans un corps K au moyen d’un
circuit arithmétique avec divisions en SD(nα log n, n log n), de taille
majorée plus précisément par
2 Cα nα dlog ne + ζα nα + O(n2 )
où Cα et ζα sont les constantes intervenant dans les complexités séquentielles de la multiplication des matrices et de l’inversion des matrices
carrées (voir proposition 8.2.2 page 236).
Le cas général
L’algorithme précédent fournit déjà une famille uniforme de circuits arithmétiques avec divisions qui calcule le polynôme caractéristique d’une matrice sur un corps, au sens des circuits avec divisions.
Autrement dit, le circuit arithmétique évalue correctement le polynôme caractéristique en tant que fraction rationnelle : en tant qu’élément
du corps K((aij )) où les coefficients aij de la matrice carrée sont pris
comme des indéterminées.
Mais il échoue à calculer le polynôme caractéristique de toute matrice
qui n’a pas un polynôme minimal de même degré que le polynôme caractéristique.
On est donc dans une situation pire que pour le calcul du déterminant à la Bunch & Hopcroft, car dans ce dernier cas, il suffit de
8.5. Méthode de Kaltofen-Wiedemann
245
multiplier à droite et à gauche la matrice par des matrices unimodulaires
(à petits coefficients entiers) prises au hasard pour obtenir une matrice
qui possède une LU -décomposition avec une très grande probabilité 7 . Et
ceci même si son déterminant est nul (cf. l’algorithme 2.2 page 62). L’algorithme de Bunch & Hopcroft sans branchement, avec le preprocessing
que nous venons d’indiquer n’échouera que dans le cas d’une matrice
n × n dont le rang est strictement inférieur à n − 1.
C’est donc en produisant un algorithme avec branchements qui fonctionne dans tous les cas que Keller-Gehrig réalise son véritable tour
de force. Et pour cela il lui fallait d’abord développer sa méthode de
réduction rapide d’une matrice à la forme échelonnée en lignes (sur un
corps). Dans cette réduction nous avons vu que la profondeur de l’algorithme (avec branchements) est un O(nlog 3 ). Keller-Gehrig obtient
précisément le résultat suivant :
Théorème 8.2 Le polynôme caractéristique d’une matrice carrée d’ordre
n sur un corps K peut être calculé par un algorithme avec branchements
qui a pour taille un O(nα log n).
Une version plus rapide pour les cas favorables
Notons que Keller-Gehrig propose une version plus rapide pour un
algorithme avec divisions mais sans branchements, qui calcule le polynôme caractéristique dans les mêmes conditions qu’à la proposition 8.4.1 :
Proposition 8.4.2 On peut calculer le polynôme caractéristique d’une
matrice carrée d’ordre n à coefficients dans un corps K au moyen d’un
circuit arithmétique avec divisions qui a pour taille un O(nα ).
Une version parallèle
Signalons enfin qu’une parallélisation de l’algorithme de Keller-Gehrig
a été obtenue par Giesbrecht [37, 38].
8.5
Méthode de Kaltofen-Wiedemann
Pour généraliser l’algorithme de Wiedemann (section 2.8.3) à un anneau commutatif arbitraire A en évitant les divisions qu’il contient et
7. Si cette méthode est seulement probabiliste en théorie, elle fonctionne toujours
en pratique.
246
8. Algèbre linéaire séquentielle rapide
le débarrasser en même temps de son aspect aléatoire, l’idée de Kaltofen
[49] est de lui appliquer la méthode de l’élimination des divisions de
Strassen (cf. le théorème 3.1 page 122). Il doit pour cela exhiber une
matrice particulière C ∈ Kn×n et un couple de vecteurs u, v ∈ An×1
pour lesquels l’algorithme de Wiedemann s’effectue sans divisions et
tels que le polynôme générateur minimal de la suite récurrente linéaire
( t u C i v) i∈N , qui est donné par l’algorithme de Berlekamp/Massey [27],
est de degré n (et n’est autre, par conséquent, que le polynôme minimal
PC et, à un signe près, le polynôme caractéristique PC de C ).
Kaltofen considère la suite de nombres entiers (ai ) ∈ NN définie par
(
an+1 = 2 an
si n est impair et
i
ai =
, avec
n+1
bi/2c
an+1 = 2 n+2 an si n estpair.
Les premiers termes sont 1, 1, 2, 3, 6, 10, 20, 35, 70, 126, 252, 462,
924, 1716, 3432, 6435, 12870, 24310, 48620, 92378, . . .
Il applique l’algorithme de Berlekamp/Massey aux 2n premiers termes :
2n − 1
a0 = 1, a1 = 1, a2 = 2, . . . , a2n−1 =
n−1
Il constate que les restes successifs dans l’algorithme d’Euclide étendu,
jusqu’au (n − 1) - ème, ont un coefficient dominant égal à ±1, avec un
degré ne diminuant que d’une seule unité à chaque pas (c’est-à-dire que
do Ri = 2n − 1 − i pour 1 ≤ i ≤ n − 1 ). Ce qui garantit le fait que
les polynômes Ri , Qi , Ui , Vi (1 ≤ i ≤ n) appartiennent à Z[X] et que
do Rn = n − 1. Il constate également que les multiplicateurs Vi (1 ≤ i ≤
n) ont un coefficient dominant et un terme constant égaux à ±1 et que,
par conséquent, dans la dernière égalité obtenue :
Un X 2n + Vn
2n−1
X
ai X i = Rn
( avec do Rn = n − 1) ,
i=0
Vn est un polynôme de degré n qui, à un signe près, s’écrit :
±Vn = X n − (cn−1 X n−1 + · · · + c1 X + c0 ) .
(avec c0 = ±1, cn−1 = 1) Kaltofen montre même, à partir de l’algorithme qui calcule les coefficients de Vn , que ces derniers sont en fait donnés
par la formule :
n+i
n−i−1
b
c
2
2
ci = (−1)
pour 0 ≤ i ≤ n − 1 .
i
8.5. Méthode de Kaltofen-Wiedemann
247
C’est donc le polynôme f (X) = X n −cn−1 X n−1 −. . .−c1 X −c0 ainsi obtenu qui est le polynôme minimal de la suite récurrente linéaire (a0i ) i∈N
dont les 2n premiers termes coı̈ncident avec les 2n premiers termes
a0 , a1 , . . . , a2n−1 de la suite (ai ).
Il considére alors la matrice C
du polynôme f (X) :
0 1
0 0
..
C = ...
.
0 0
c0 c1
Par exemple, pour n = 7 on
0
0
0
C :=
0
0
0
−1
transposée de la matrice compagnon
0
1
..
.
...
...
..
.
0
0
..
.
0 ...
1
c2 . . . cn−1
,
obtient
1
0
0
0
0
0
4
0
0
0
1
0
0
0
1
0
0
0
1
0
0
0
0
0
0
6 −10 −5
0
0
0
0
1
0
6
0
0
0
0
0
1
1
Le polynôme caractéristique de C n’est autre que PC = (−1)n f (X). Il
considère enfin les deux vecteurs :
1
a0
0
a1
n×1
V =
.
. de A
. . . et E1 =
..
an−1
0
On vérifie immédiatement que les suites récurrentes linéaires (a0i ) i∈N et
( t E1 C i V ) i∈N , qui admettent un polynôme générateur unitaire commun
de degré n, sont telles que a0i = t E1 C i V (= ai ) pour tout i compris
entre 0 et 2n − 1. On en déduit que a0i = t E1 C i V pour tout i ∈ N.
Ainsi, par construction même de C, l’algorithme de Wiedemann,
prenant C en entrée avec les deux vecteurs E1 et V , s’effectue avec les
seules opérations d’addition et de multiplication dans Z pour donner en
sortie le polynôme minimal de la suite récurrente linéaire ( t E1 C i V ) i∈N ,
et par conséquent le polynôme caractéristique de C.
248
8. Algèbre linéaire séquentielle rapide
Soit maintenant A = (aij ) une matrice carrée d’ordre n à coefficients dans A. Il s’agit de calculer le polynôme caractéristique de A en
n’utilisant que les opérations arithmétiques de A. Cela se fait par élimination des divisions dans l’algorithme de Wiedemann pour la matrice
A en prenant comme centre d’élimination des divisions le point formé
par la matrice C et les deux vecteurs auxilaires E1 et V . Comme les
coefficients du polynôme caractéristique de A (les sorties de l’algorithme de Wiedemann) sont des polynômes de degré ≤ n en les coefficients
(aij ), on utilise l’élimination des divisions de Strassen en degré n.
On considère donc une indéterminée z sur A .
On pose F = A − C,et on applique l’algorithme de Wiedemann
dans l’anneau An = A[z] z n+1 à la matrice B = C + zF avec les
vecteurs auxiliaires E1 et V . On récupère le polynôme caractéristique
de A en remplaçant z par 1 dans les sorties.
Cet algorithme calcule le polynôme générateur minimal gz (X) ∈
An [X] de la suite récurrente linéaire ( t E1 B i V ) i∈N . Comme les seules
divisions se font par des polynômes en z de terme constant égal à ±1,
l’ensemble du calcul se fait uniquement avec des additions et multiplications dans A.
D’où l’algorithme 8.2 page ci-contre de Kaltofen pour le calcul du
polynôme caractéristique d’une matrice carrée A ∈ An×n .
Complexité de l’algorithme
On utilise comme d’habitude la notation 3.2.2 page 125 ainsi que la
notation 7.2.1 page 195. L’étude de complexité donne le résultat suivant
dû à Kaltofen [49] :
Théorème 8.3 Le calcul du déterminant, du polynôme caractéristique
et de l’adjointe d’une matrice carrée d’ordre n sur un anneau commutatif arbitraire A se fait à l’aide d’une famille uniforme de circuits
√
α+3
arithmétiques en SD(n log n, n 2 µP (d n e)).
Si on utilise une multiplication rapide des polynômes en O(n log n) ou
en O(n log n log log n) opérations arithmétiques (selon l’anneau considéré),
α
α
cela fait donc, O(n 2 +2 log n) ou O(n 2 +2 log n log log n) opérations
arithmétiques pour l’algorithme de Kaltofen. Nous verrons au chapitre
10 que les algorithmes parallèles en profondeur log2 n font moins bien
dans le cas d’un anneau vraiment arbitraire (ils utilisent O(nα+1 log n)
8.5. Méthode de Kaltofen-Wiedemann
249
Algorithme 8.2 Algorithme de Kaltofen-Wiedemann
Entrée : Un entier n ≥ 2 et une matrice A = (aij ) ∈ An×n .
Sortie : Le polynôme caractéristique PA (X) de A .
Début
(on pose An = A[z] z n+1 )
Variables locales : i, k ∈ N ; V = (vi ) ∈ Zn×1 (vecteur du centre d’élimination des divisions) ; C = (cij ) ∈ Zn×n (matrice du centre d’élimination
des divisions) ; B ∈ (An )n×n ; (rk )k=0..2n−1 ∈ (An )2n .
Étape 1 : Calcul du centre d’élimination des divisions, et initialisation.
C := 0 ∈ An×n ;
pour i de 1
à n faire
n+i−1
n−i
i−1
b
c
2
2
; cn,i := (−1)
vi :=
i−1
i
−
1
2
fin pour ;
pour i de 1 à n − 1 faire ci,i+1 := 1 fin pour ;
B := C + z × (A − C) ;
Étape 2 : Calcul de la suite récurrente linéaire
pour k de 0 à 2n − 1 faire
rk := première coordonnée de B k × V dans An
fin pour ;
Étape 3 : Berlekamp-Massey
Appliquer la procédure de Berlekamp-Massey à la suite (rk )k=0..2n−1
puis remplacer z par 1 dans le polynôme générateur minimal trouvé.
Fin.
1
opérations arithmétiques) mais un peu mieux (O(nα+ 2 ) opérations arithmétiques) dans le cas d’un anneau où les entiers ≤ n sont non diviseurs
de zéro.
Dans le cours de la preuve qui suit nous ferons également l’analyse de
complexité de la version élémentaire de l’algorithme de Kaltofen. Nous
obtenons le résultat suivant.
Proposition 8.5.1 Dans la version séquentielle simple de l’algorithme
de Kaltofen, le calcul du déterminant, du polynôme caractéristique et
de l’adjointe d’une matrice carrée d’ordre n sur un anneau commutatif
arbitraire A se fait à l’aide d’une famille uniforme de circuits arithmétiques de taille O(n4 ) et plus précisément avec un nombre de multiplications égal à 4n4 + O(n3 ) et un nombre d’additions du même ordre de
grandeur. Le nombre de multiplications essentielles est de 2n4 + O(n3 ).
250
8. Algèbre linéaire séquentielle rapide
Preuve. On remarque tout d’abord que le coût de l’étape 1 est négligeable. Les entiers qu’elle calcule sont des constantes du circuit disponibles une fois pour toutes et leur calcul ne doit pas être pris en
compte (ils sont de toute façon calculables en O(n2 ) opérations arithmétiques). Quant à l’affectation B := C + z × (A − C) dans (An )n×n
elle signifie du point de vue des opérations arithmétiques dans A qu’on
effectue 2n − 1 soustractions qui peuvent être effectuées en une seule
étape parallèle.
L’étape 3 est pour l’essentiel un algorithme d’Euclide étendu. Elle se
fait avec un circuit arithmétique de profondeur O(n log n) et de taille
O(n2 µP (n)) où µP (n) est le nombre d’opérations arithmétiques nécessaires pour la multiplication de deux polynômes de degré n dans A[z]
en profondeur O(log n). Cela est dû au fait que l’algorithme d’Euclide
étendu utilisé comporte O(n) étapes avec chacune O(n) opérations
arithmétiques dans l’anneau des développements limités An (certaines
de ces opérations sont des divisions par des éléments inversibles).
Pour obtenir le résultat énoncé, reste l’étape 2, la plus coûteuse en
nombre d’opérations arithmétiques.
Voyons tout d’abord la version élémentaire. On calcule successivement les Vk = B k V pour k = 1, . . . , 2n − 1 par Vk+1 = BVk . Cela
fait en tout 2n3 − n2 multiplications et n(n − 1)(2n − 1) additions
dans An . Chacune des 2n3 − n2 multiplications est le produit d’une
entrée de B par une coordonnée de l’un des Vk . Or les entrées de B
sont des éléments de la forme c + bz où c est une constante (une des
entrées non nulles de C) et b est une entrée de A − C. Un tel produit
consomme donc n multiplications essentielles, n + 1 multiplications du
type (( produit d’un élément de A par une constante )) et n additions.
En résumé, l’étape 2 dans la version séquentielle élémentaire consomme
2n4 − n3 multiplications essentielles, 2n4 + O(n3 ) multiplications non
essentielles et 4n4 + O(n3 ) additions.
Voyons maintenant la version accélérée. On subdivise l’étape 2 en
quatre sous-étapes qui sont les suivantes, numérotées de 2.1 à 2.4, dans
√
lesquelles on pose r = d n e , s = d2n/re − 1, U0 = E1 et V0 = V :
Étape
Étape
Étape
Étape
2.1
2.2
2.3
2.4
:
:
:
:
pour j de 1 à r − 1 calculer Vj := B j V0
Calculer la matrice B r
pour k de 1 à s calculer Uk := ( t B r )k E1
pour j de 0 à r − 1 et pour k de 0 à s calculer
bkr+j (z) := t Uk (z) Vj (z) .
8.5. Méthode de Kaltofen-Wiedemann
251
Notez que r(s+1) ≥ 2n si bien que que les entiers kr +j parcourent
tout l’intervalle [0, 2n − 1].
Au cours des sous-étapes 2.1 et 2.2, les coefficients calculés sont
des polynômes en z de degré ≤ r (dans B j V , ils sont de degré ≤ j ),
c’est-à-dire que chaque multiplication de deux coefficients correspond à
un circuit arithmétique de profondeur O(log r) avec µP (r) opérations
de base dans A. Cela donne l’analyse suivante pour les différentes sousétapes.
• Sous-étape 2.1 : Pour obtenir tous les vecteurs B j V pour 1 ≤
j ≤ r − 1 on peut procéder en blog rc étapes parallèles où chaque étape
i−1
i ( i = 1, . . . , blog rc ) consiste à élever au carré la matrice B 2
puis à
i−1
la multiplier à droite par la matrice [ V | B V | . . . | B 2 −1 V ] qui est
i
une matrice n × 2i−1 pour obtenir la matrice [ V | B V | . . . | B 2 −1 V ]
qui est une matrice n × 2i dont les coefficients sont des polynômes de
A[z] de degré < 2i ≤ r.
Chacune de ces blog rc étapes correspond donc à un circuit arithmétique de profondeur O(log n log r) et de taille O(nα µP (r)), ce qui donne
au total, pour la sous-étape 2.1, un circuit arithmétique de profondeur
O(log3 n) et de taille O(nα µP (r) log n) .
• Sous-étape 2.2 : Si r est une puissance de 2, le calcul de B r se
fait en élevant au carré la matrice B r/2 déjà calculée. Sinon il faut faire
i−1
le produit de certaines des matrices B 2
: par exemple si r = 39 =
5
2
32 + 4 + 2 + 1, on a B 39 = B 2 B 2 B 2 B. Pour chaque produit les coefficients des matrices sont de degré ≤ r/2 dans A[z]. Ceci correspond de
nouveau à un circuit arithmétique de profondeur O(log3 n) et de taille
O(nα µP (r) log n). Pour la suite nous posons B1 = tB r
• Sous-étape 2.3 : Nous ne pouvons plus utiliser la technique de
l’étape 2.1 qui ici donnerait a priori une famille uniforme de circuits
arithmétiques dans SD(log3 n, nα µP (n) log n).
Partant du vecteur U0 = E1 , la sous-étape 2.3 de notre algorithme consiste à calculer, pour k allant de 1 à s, le vecteur Uk (z) =
B1 Uk−1 (z). Posons s1 = d(n + 1)/re. Notons
Uk−1 = B1k−1 E1 se
Ps1que
−1 r`
réécrit dans An sous la forme Uk−1 (z) = `=0 z Uk−1,` où chacun
des Uk−1,` est un vecteur dont les composantes sont des polynômes en z
de degré < r. On peut donc identifier Uk−1 (z) avec la matrice n × s1 :
Wk (z) = [ Uk−1,0 | Uk−1,1 | . . . | Uk−1,s1 −1 ] .
Le calcul du vecteur Uk (z) à n lignes et s1 colonnes se fait comme
suit. On calcule la matrice B1 Wk−1 (z) dont les entrées sont des poly-
252
8. Algèbre linéaire séquentielle rapide
nômes de degré ≤ 2r, puis on réorganise les sommes correspondantes
pour obtenir Uk (z) (ce qui nécessite au plus n2 additions dans A).
Le produit B1 Wk−1 (z) est celui d’une matrice n × n par une matrice
n × s1 , toutes les entrées étant de degré ≤ r. Ceci peut se faire avec
r2 multiplications parallèles de blocs s1 × s1 . Chaque multiplication de
blocs se fait en O(sα1 ) opérations arithmétiques sur des polynômes de
degré ≤ r. On obtient donc chaque Uk (z) en SD(log2 r, r2+α µP (r)).
Cela donne au total, pour la sous-étape 2.3, une famille uniforme de
circuits arithmétiques dans SD(s log2 n, s r2+α µP (r)) c’est-à-dire en√
3+α
1
core dans SD(n 2 log2 n, n 2 µP (d n e).
• Sous-étape 2.4 : Cette étape peut être également ramenée à la
multiplication d’une matrice (s + 1) × n par une matrice n × r :
t U (z)
0
t U (z)
1
..
.
t U (z)
s
× V0 (z) V1 (z) · · · Vr−1 (z) =
t U (z) V (z)
0
0
..
.
t U (z) V (z)
s
0
···
..
.
···
t U (z) V
0
r−1 (z)
..
.
t U (z) V
s
r−1 (z)
dont l’élément en position (k +1, j +1) pour 0 ≤ k ≤ s et 0 ≤ j ≤ r −1
n’est autre que le coefficient recherché : t Uk (z) Vj (z) = bkr+j (z).
Utilisant à nouveau la multiplication par blocs (s + 1) × (s + 1), nous
concluons que la sous-étape 2.4 correspond à un circuit arithmétique
√
α+2
de profondeur O(log2 n) et de taille O(n 2 µP (d n e)).
On peut résumer le calcul de complexité dans le tableau (8.5) suivant qui donne, pour chaque étape, la complexité arithmétique du circuit correspondant, en même temps que le résultat général. Nous avons
également indiqué la taille lorsqu’on exécute l’algorithme avec une multiplication accélérée des polynômes mais sans multiplication rapide des
matrices, sur les lignes (( avec α = 3 )).
8.5. Méthode de Kaltofen-Wiedemann
Etape
253
Profondeur
Taille
Etape 1
O(1)
Etape 2
O(n 2 log2 n)
négligeable
√
α+3
O(n 2 µP (d n e))
√
O(n3 µP (d n e) log n)
avec α = 3
1
···
Etape 3
O(n log n)
Total
O(n log n)
avec α = 3
···
O(n2 µP (n))
√
α+3
O(n 2 µP (d n e))
√
O(n3 µP (d n e) log n)
Tableau 8.5
Complexité de l’algorithme de Kaltofen-Wiedemann
t
u
Dans notre preuve c’est l’étape 3 qui détermine la profondeur du circuit arithmétique correspondant à l’algorithme de Kaltofen-Wiedemann.
Mais on peut réduire la profondeur de l’étape 3 par diverses méthodes.
Une première est de ne pas utiliser l’algorithme de Berlekamp/Massey
pour le calcul du polynôme minimal d’une suite récurrente linéaire. Une
telle méthode, développée dans [50] (voir aussi [BP]) ramène ce calcul à
la résolution d’un système linéaire qui a la forme de Toeplitz, en utilisant
le calcul du polynôme caractéristique de sa matrice par la méthode de
Le Verrier améliorée par Csanky (cf. section 9.1). On obtient un circuit
arithmétique de profondeur de O(log3 n) et de même taille, c’est-à-dire
O(n2 µP (n)) . L’inconvénient de cette amélioration est qu’elle s’applique
uniquement lorsque n! ne divise pas zéro dans l’anneau A.
Une deuxième méthode, qui ne se heurte pas à l’obstacle précédent,
consiste à utiliser une version parallélisée de l’algorithme d’Euclide étendu. Voir [71, 66] et [GG] corollaire 11.6 page 304.
Cependant, il ne suffit pas de réduire la profondeur de l’étape 3 pour
obtenir une profondeur polylogarithmique. Il faudrait le faire également
pour l’étape 2 et plus précisément la sous-étape 2.3.
On a donc à l’heure actuelle un problème ouvert : peut-on obtenir un
circuit, de la taille de cet algorithme et de profondeur polylogarithmique, permettant de calculer le polynôme caractéristique sur un anneau
commutatif arbitraire ?
L’algorithme de Kaltofen-Wiedemann obtient le résultat asymptoti√
α+3
que ci-dessus, à savoir O(n 2 µP (d n e)), le meilleur temps séquentiel
254
8. Algèbre linéaire séquentielle rapide
de tous les algorithmes connus pour le calcul du polynôme caractéristique sur un anneau commutatif arbitraire, grâce à la multiplication rapide
des matrices, bien sûr, mais aussi grâce à la multiplication rapide des
polynômes. Et pour les polynômes la multiplication rapide est désormais
couramment implémentée sur machine.
Ainsi lorsqu’on ne dispose pas d’une multiplication rapide des matrices, on obtient un temps séquentiel asymptotiquement meilleur que
tous les autres algorithmes fonctionnant sur un anneau commutatif arbitraire, dès qu’on accélère la multiplication des polynômes, ne serait-ce
que par la méthode de Karatsuba.
Notons que sur un anneau commutatif qui ne possède pas de racines principales de l’unité, la méthode qui utilise la transformation de
Fourier rapide est en O(n log n log log n) et elle ne devient plus performante que la méthode de Karatsuba en O(nlog 3 ) que pour n très
grand, de l’ordre de plusieurs milliers (cf. section 6.3.2 et notamment la
remarque 6.3.2 page 182).
Un vaste champ d’expérimentation s’ouvre donc, maintenant que
différentes multiplications rapides commencent à avoir une réelle portée
pratique en calcul formel.
Conclusion
Nous terminons ce chapitre en renvoyant le lecteur à deux surveys
récents d’Erich Katofen et Gilles Villard [53, 54] concernant la complexité aussi bien algébrique que binaire du calcul des déterminants
(nous nous intéressons plutôt au calcul du polynôme caractéristique dans
cet ouvrage).
Ils montrent à quel point l’algèbre linéaire est un sujet de recherche
actif en calcul formel et l’importance des méthodes modulaires et seminumériques pour le traitement des problèmes concrets.
9. Parallélisations de la
méthode de Leverrier
Introduction
Csanky [22] fut le premier à prouver que les problèmes du calcul des
déterminants, de l’inversion des matrices, de la résolution des systèmes
d’équations linéaires et du calcul du polynôme caractéristique, dans le
cas d’un anneau contenant le corps des rationnels, sont dans la classe
N C, c’est-à-dire dans la classe des problèmes qui peuvent être résolus
en temps parallèle polylogarithmique avec un nombre polynomial de
processeurs par une famille uniforme de circuits arithmétiques.
Il montre, en effet, que tous ces problèmes se ramènent au calcul du
polynôme caractéristique et que ce dernier se calcule en SD(nα+1 , log2 n).
En particulier ils sont dans la classe N C 2 .
Nous présentons le travail de Csanky dans la section 9.1. Dans la
section suivante nous donnons l’amélioration due à Preparata & Sarwate [77] qui montre que le calcul du polynôme caractéristique peut
être réalisé dans SD(nα+1/2 , log2 n). Dans la section 9.3 nous donnons
une meilleure estimation de la complexité théorique de l’algorithme
précédent, légèrement amélioré, due à Galil & Pan [32].
Dans le chapitre 10, nous examinerons des algorithmes qui résolvent
les mêmes problèmes sur un anneau commutatif arbitraire.
9.1
Algorithme de Csanky
Pour calculer le polynôme caractéristique , Csanky utilise la méthode
de Le Verrier en la parallélisant de la manière suivante.
On se donne un entier n, un corps K (ou plus généralement un
anneau dans lequel n! est inversible) et une matrice A ∈ Kn×n de po-
256
9. Parallélisations de la méthode de Leverrier
lynôme caractéristique :
P (X) = det(A − XIn ) = (−1)n [X n − c1 X n−1 − . . . − cn−1 X − cn ] .
On pose sk = Tr(Ak ) pour k = 1, 2, . . . , n.
La méthode de Le Verrier consiste à résoudre l’équation
S ~c = ~s
(9.1)
où
~c =
c1
..
.
..
.
cn
, ~s =
s1
..
.
..
.
sn
1
0
s1
2
.
.
et S =
..
..
sn−2
sn−1 sn−2
··· ··· 0
..
..
.
.
..
.. ..
.
. .
.
.. ..
. 0
.
· · · s1 n
Cette équation admet la solution unique ~c = S −1~s qui donne les
coefficients du polynôme caractéristique.
Ceci donne l’algorithme de Csanky 9.1 en quatre grandes étapes.
Algorithme 9.1 Algorithme de Csanky, principe général.
Entrée : Un entier n ∈ N et une matrice A ∈ An×n . L’anneau A contient
le corps Q.
Sortie : Les coefficients du polynôme caractéristique PA (X) de A.
Début
Étape 1 :
Calculer en parallèle les puissances A2 , A3 , . . . , An ;
Étape 2 :
Calculer en parallèle les traces s1 , s2 , . . . , sn des matrices
A, A2 , . . . , An .
Étape 3 :
Créer et inverser la matrice triangulaire S (équation 9.1).
Étape 4 :
Calculer le produit S −1~s = ~c.
Fin.
L’analyse de complexité pour cet algorithme utilise les résultats de
complexité de la technique (( diviser pour gagner )) et notamment son
9.1. Algorithme de Csanky
257
application au calcul parallèle de l’inverse d’une matrice triangulaire
que nous avons décrite au § 7.2.
La complexité de l’algorithme
• Le calcul en parallèle des puissances A2 , . . . , An de la matrice A se
ramène à un algorithme de calcul parallèle des préfixes représenté par
un circuit arithmétique parallèle de profondeur O(log n) et de taille
majorée par 4n (théorème 5.1 page 170), mais dont les nœuds internes
représentent eux-mêmes des circuits de multiplication de matrices n×n,
c’est-à-dire des circuits de taille O(nα ) et de profondeur O(log n). Ce
qui donne au total, pour réaliser l’étape 1 un circuit arithmétique en
SD(nα+1 , log2 n).
• On calcule ensuite les traces des matrices A, A2 , . . . , An , c’est-à-dire
les coefficients sk = Tr(Ak ) qui forment la matrice triangulaire S.
Ce sont des sommes de n éléments de A que l’on calcule en parallèle
pour 1 ≤ k ≤ n en SD(n2 , log n).
• Le calcul de S −1 se fait comme indiqué au § 7.2. La matrice S est
en effet triangulaire et fortement régulière. D’après la proposition 7.2.2,
le calcul de la matrice S −1 se fait par un circuit arithmétique parallèle
en SD(nα , log2 n).
• Enfin, le calcul de ~c = S −1~s, qui est le produit d’une matrice triangulaire par un vecteur, se fait en parallèle par un circuit arithmétique de
taille n2 et de profondeur dlog ne, la profondeur étant essentiellement
due aux additions.
En fait, on a un tout petit peu mieux.
Théorème 9.1 (Csanky)
Soit A un anneau vérifiant les hypothèses pour l’algorithme de Le Verrier : la division par n!, quand elle est possible, est unique et explicite.
Le calcul du polynôme caractéristique, de l’adjointe et l’inverse d’une
matrice carrée d’ordre n est en SD(nα+1 , log 2 n).
Preuve. Une légère modification de l’algorithme de Csanky pour le polynôme caractéristique d’une matrice carrée d’ordre n montre que l’hypothèse d’un anneau dans lequel n! est inversible, peut être remplacée
par l’hypothèse pour l’algorithme de Le Verrier. En effet soit A ∈ An×n
et S la matrice utilisée dans l’algorithme de Csanky pour le calcul du
polynôme caractéristique .
258
9. Parallélisations de la méthode de Leverrier
Au lieu de calculer S −1 (ce qui n’est possible que si n! est inversible
dans A ), on calcule n! S −1 . Il suffit pour cela de développer le polynôme
caractéristique de S en calculant le produit (X − 1)(X − 2) · · · (X − n),
ce qui revient à calculer les valeurs des polynômes symétriques élémentaires σ1 , σ2 , . . . , σn de n variables au point (1, 2, . . . , n ).
Le théorème de Cayley-Hamilton permet alors d’écrire :
(−1)n+1 n! S −1 = S n−1 +
n−1
X
(−1)k σk S n−k−1
k=1
ce qui ramène le calcul de n! S −1 à celui des puissances S 2 , S 3 , . . . , S n−1 .
Or ce calcul se fait en parallèle, d’après le calcul des préfixes par
exemple (proposition 5.1) en SD(nα+1 , log2 n).
Nous laissons le lecteur ou la lectrice terminer pour ce qui concerne
les calculs de l’adjointe et de l’inverse.
t
u
Variante de Schönhage
Signalons qu’il existe une variante de la méthode de Csanky/Le Verrier due à Schönhage [81] qui donne une famille uniforme de circuits
arithmétiques avec divisions calculant le polynôme caractéristique avec
une faible profondeur sur un corps de caractéristique finie.
Schönhage utilise le résultat suivant concernant les sommes de Newton (§ 1.5) connu sous le nom de critère de Kakeya [48] :
Proposition 9.1.1 Soit J une partie finie à n éléments de N et
(sj )j∈J le système correspondant de n sommes de Newton à n indéterminées sur un corps K de caractéristique nulle. Alors (sj )j∈J est
un système fondamental de polynômes symétriques sur K (cf. définition
1.5.1) si et seulement si N \J est stable pour l’addition dans N.
Par exemple, pour tout entier p positif, la partie J(p, n) ⊂ N \ p N
constituée des n premiers entiers naturels qui ne sont pas des multiples
de p, satisfait ce critère, et Schönage [81] l’utilise pour adapter la méthode de Le Verrier au calcul du polynôme caractéristique sur un corps
de caractéristique p > 0.
Notez qu’en caractéristique p l’égalité (x + y)p = xp + y p implique
que les sommes de Newton vérifient les égalités skp = sk p .
Prenons maintenant
exemple. Le polynôme général de degré 8
P8 un 8−i
8
est P (X) = X − i=1 ai X . Si nous sommes sur un corps de caractéristique 3, nous considérons les 8 premières relations de Newton qui
9.2. Amélioration de Preparata et Sarwate
donnent les
1
s1
s3
s4
s6
s7
s9
s10
259
sommes sj pour j ∈ N \ 3 N (cf. l’égalité (1.22) page 28) :
0 0 0 0 0 0 0
s1
a1
s2
2 0 0 0 0 0 0
a2
s4
s 2 s 1 1 0 0 0 0 a3
s3 s2 s1 2 0 0 0
a4 = s 5
(9.2)
s5 s4 s3 s2 s1 1 0 a5
s7
s8
s6 s5 s4 s3 s2 s1 2
a6
s10
s8 s7 s6 s5 s4 s3 s2
a7
s9 s8 s7 s6 s5 s4 s3
s11
a8
Compte tenu des relations
s3 = s1 3 , s6 = s2 3 , s9 = s1 9 ,
(9.3)
le déterminant de la matrice carrée est égal à
d = −s1 s32 s5 + s22 s24 + s31 s4 s5 + s51 s2 s5 + s31 s22 s5 − s21 s32 s4 +
s21 s25 + s41 s42 + s61 s32 − s51 s7 − s41 s8 − s34 − s1 s7 s22 + s12
1
6
2
8 2
8
2
−s41 s24 + s21 s52 + s10
1 s2 + s2 + s2 s8 − s1 s2 + s1 s4 − s1 s2 s8
+s4 s8 − s5 s7 + s41 s22 s4 + s2 s25 − s61 s2 s4 + s1 s2 s4 s5
Un point non trivial est que d n’est pas une fonction identiquement
nulle (si le corps de base est infini). En fait dans le cas générique, c’est-àdire si on considère les ai comme des indéterminées et les si (i = 1, 2, 4,
5, 7, 8, 10, 11) comme donnés par les relations (9.2) et (9.3), les éléments
si sont algébriquement indépendants. Cela implique alors que les ai
(i = 1, . . . , 8) peuvent s’exprimer comme fractions rationnelles en les si
(i = 1, 2, 4, 5, 7, 8, 10, 11) avec d pour dénominateur.
Un autre point non trivial consiste à résoudre les systèmes linéaires
du type (9.2) (lorsque le déterminant correspondant est non nul) par un
algorithme (avec divisions) bien parallélisé.
L’algorithme de Schönage [81] correspond à une famille de circuits
arithmétiques (avec divisions) dans SD(nα+1 , log2 n) (voir aussi le livre
[BP] Annexe C pages 372–377).
9.2
Amélioration de Preparata et Sarwate
Principe général
Considérons un anneau A vérifiant les hypothèses pour l’algorithme
de Le Verrier, et une matrice carrée A ∈ An×n . L’amélioration apportée
260
9. Parallélisations de la méthode de Leverrier
par Preparata & Sarwate [77] à l’algorithme de Csanky provient du fait
que pour calculer les traces sk = Tr(Ak ) (1 ≤ k ≤ n), on n’a pas besoin
de calculer toutes les puissances de A.
√
Il suffit en effet, si l’on pose p = d n e, de disposer des 2p matrices
2
In , A, . . . , Ap−1 et Ap = B, B 2 , . . . , B p = Ap , ce qui revient à calculer
√
2 d n e − 2 puissances de matrices n × n au lieu des n − 1 puissances
de A. Il est fait appel pour cela à deux procédures récursives notées
Powers(A, r) et Superpowers(A, r) permettant de calculer les puissances
successives d’une matrice carrée A jusqu’à l’ordre r.
Les traces des puissances de A seront alors obtenues en considérant
les matrices Uj (1 ≤ j ≤ n) définies de la manière suivante :
Uj = Lj Cj où Lj ∈ Ap×n est la matrice formée uniquement des j - èmes
lignes ( 1 ≤ j ≤ n ) des p matrices In , A, . . . , Ap−1 et où Cj ∈ An×p est
la matrice formée des j - èmes colonnes ( 1 ≤ j ≤ n ) des autres matrices
2
Ap , A2p , . . . , Ap .
Les matrices Uj (1 ≤ j ≤ n) sont des matrices carrées d’ordre p dont
les p2 coefficients ne sont autres que les j - èmes éléments diagonaux des
2
matrices Ap , Ap+1 , . . . , Ap +p−1 .
[j]
Plus précisément, l’élément ukl qui est position (k, l) dans la matrice Uj et qui est obtenu par multiplication de la j - ème ligne de la
matrice Ak−1 par la j - ème colonne de la matrice Apl est donc le j ème élément de la diagonale du produit Ak−1 Apl = Apl+k−1 , c’est-à-dire
[j]
[pl+k−1]
[m]
que ukl = a jj
pour 1 ≤ k, l ≤ p, si l’on désigne par a rs l’élément
en position (r, s) de la matrice Am .
Posant m = pl + k − 1 ( m prend toutes les valeurs comprises entre
p et p2 + p − 1 quand k et l varient de 1 à p ) on obtient, avec les
notations ci-dessus, et pour p ≤ m ≤ p2 + p − 1 :
Tr(Am ) =
n
X
j=1
[m]
a jj =
n
X
[j]
ukl
j=1
(où l et k − 1 sont respectivement le quotient et le reste euclidiens de
m par p ).
Comme les matrices A, . . . , Ap−1 sont déjà disponibles, cela nous
2
donne donc les traces de toutes les puissances A, . . . , Ap , . . . , Ap donc
√
celles de toutes les matrices A, . . . , An puisque p2 +p−1 ≥ n+ n−1 ≥
n.
D’où l’algorithme de Preparata & Sarwate qui comprend deux parties, la première pour le calcul du polynôme caractéristique de la matrice
9.2. Amélioration de Preparata et Sarwate
261
donnée A ∈ An×n , et la deuxième pour le calcul de l’adjointe et de l’inverse de cette matrice.
Calcul du polynôme caractéristique
Avant de donner l’algorithme 9.2 page suivante, voyons tout d’abord
les sous-procédures utilisées dans cet algorithme. Il s’agit essentiellement
de la procédure Superpowers qui est définie de manière récursive à partir
de la procédure Powers (elle-même définie de manière récursive) en vue
d’accélérer le calcul des puissances d’une matrice carrée donnée (dans
notre cas, c’est la matrice A ∈ An×n ).
Chacune de ces deux sous-procédures prend donc en entrée A et un
entier p > 1 et donne en sortie la matrice rectangulaire n × np formée
des p puissances de A :
Powers(A, p) = Superpowers(A, p) = [ A | A2 | . . . | Ap ].
Powers(A, s)
• m := ds/2e ;
• [ A | . . . | Am ] := Powers(A, m) ;
• pour i de m + 1 à s faire Ai := Abi/2c Adi/2e .
Superpowers(A, p)
• r := dlog pe ;
• s := bp/rc ; q := p − rs ;
• [ A | . . . | As ] := Powers(A, s) ;
• pour k de 1 à r − 1 faire Ask × Powers (A, s) ;
(cela donne toutes les puissances de A jusqu’à l’ordre rs)
• pour i de 1 à q faire Asr × [ A | . . . | Aq ] ;
(pour avoir les q = p − rs puissances restantes de A).
La complexité de l’algorithme
Nous utilisons comme d’habitude les notations 7.2.1 page 195. Nous
allons déterminer les paramètres de complexité de la famille de circuits
arithmétiques parallèles représentant l’algorithme de Preparata & Sarwate en commençant par la complexité des sous-procédures qu’il utilise.
Les paramètres de complexité pour l’algorithme principal 9.2 (représenté par la colonne PS(A, n) ) et les procédures auxiliaires Powers
(colonne PW(A, p) ) et Superpowers (colonne SPW(A, p) ) seront désignés, conformément au tableau suivant, respectivement par :
262
9. Parallélisations de la méthode de Leverrier
Algorithme 9.2 Algorithme de Preparata & Sarwate
Entrée : Un entier n et une matrice A ∈ An×n .
Sortie : Le vecteur ~c des coefficients du polynôme caractéristique de A.
√
Les étapes du calcul (avec p = d n e),
1. Calculer les puissances A, . . . , Ap en appelant Superpowers(A, p) ;
2
2. Calculer les puissances Ap , . . . , Ap en faisant Superpowers(Ap , p) ;
3. Calculer en parallèle les n produits Uj = Lj Cj ( 1 ≤ j ≤ n ) ;
4. Former le vecteur ~s et la matrice triangulaire S en calculant en pa[j]
rallèle, à partir des matrices Uj = (ukl ) obtenues à l’étape précédente,
P
[j]
les n traces sm = nj=1 ukl (1 ≤ m ≤ n). On prendra, pour chaque
valeur de m, l = bm/pc et k = m + 1 − lp ;
5. Calculer S −1 (en utilisant l’approche (( diviser pour gagner ))) ;
6. Calculer le produit S −1~s = ~c.
↓ Paramètre / Procédure →
Taille
Profondeur
Largeur
PS(A, n)
τ (n)
π(n)
λ(n)
PW(A, p)
τ1 (p)
π1 (p)
λ1 (p)
SPW(A, p)
τ2 (p)
π2 (p)
λ2 (p)
La définition de la procédure Powers nous donne les relations de récurrence :
τ1 (p) = τ1 (dp/2e) + bp/2c µM (n)
π1 (p) = π1 (dp/2e) + γM (n)
λ1 (p) = max { λ1 (dp/2e), bp/2c λM (n) }
on en déduit
τ1 (p)
π1 (p)
λ1 (p)
pour p ≥ 2 par sommation de 1 à r = dlog pe :
≤ (2p − 3) Cα nα
≤ Kα dlog pe log n
= bp/2c λM (n) ≤ 12 Lα p log n .
La définition de la procédure Superpowers dans laquelle r = dlog pe,
s = bp/rc et q := p − rs, permet d’écrire 1 :
τ2 (p) = τ1 (s) + (p − s) µM (n)
π2 (p) = π1 (s) + r γM (n)
λ2 (p) = max {λ1 (s), s λM (n)}
1. Le calcul préliminaire des entiers r, s, q n’intervient pas : il fait partie de la
construction du circuit arithmétique correspondant.
9.2. Amélioration de Preparata et Sarwate
qui donnent,
τ2 (p)
π2 (p)
λ2 (p)
263
avec les majorations précédentes :
≤
≤
=
(p + s − 3) Cα nα ≤ [ p + (p/ log p) − 3 ] Cα nα
2Kα (dlog pe log n)
bp/rc λM (n)
≤ (Lα p nα ) / (log p log n) .
L’algorithme utilise en plus des procédures ci-dessus une procédure
d’inversion de matrice triangulaire. Nous avons vu (proposition 7.2.2)
que l’inversion d’une matrice triangulaire fortement régulière se fait par
un circuit arithmétique parallèle de taille majorée par Cα (2n−1)α donc
par 8 Cα nα , de profondeur au plus égale à
Kα (log2 (n) + 3 log (n) + 2) + 1 ≤ 2 Kα (log (n) + 1)2 .
Et sa largeur est O(nα / log2 n) si on applique le principe de Brent.
Ceci permet d’établir la complexité de la première partie de l’algo√
rithme principal. Compte tenu du fait que p = d n e et que 2 ≤ α ≤ 3,
le tableau 9.2 indique le résultat des majorations pour la taille et la
profondeur et pour chaque étape.
Complexité de l’Algorithme de Preparata & Sarwate
Etapes
Etape 1
Etape 2
Etape 3
Etape 4
Etape 5
Etape 6
Total
Taille
[ p + (p/ log p) − 3 ] nα
[ p + (p/ log p) − 3 ] nα
n [pµM (p) + (p − 1)p2 ]
n (n − 1)
8 Cα nα
n2
1
τ (n) = O(nα+ 2 )
Profondeur
2 Kα dlog pe log n
2 Kα dlog pe log n
Kα log n
dlog ne
2 Kα (log (n) + 1)2
dlog ne + 1
π(n) = O(log2 n)
Tableau 9.2
On en déduit le résultat suivant de Preparata & Sarwate, dans lequel
nous avons également intégré, le calcul de l’adjointe et de l’inverse qui
constitue la deuxième partie de cet algorithme :
Théorème 9.2 Soit A un anneau vérifiant les hypothèses pour l’algorithme de Le Verrier. Le polynôme caractéristique, le déterminant,
l’adjointe et l’inverse (s’il existe) d’une matrice carrée A ∈ An×n se
fait par un circuit arithmétique de taille τ (n), de profondeur π(n) et
264
9. Parallélisations de la méthode de Leverrier
de largeur λ(n) majorées
τ (n)
π(n)
λ(n)
respectivement par :
1
1
≤ 4 Cα nα+ 2 + o(nα+ 2 )
≤ 5 Kα log2 n + O(log n)
1
≤ (2 Lα nα+ 2 ) / (log2 n)
où Cα , Kα , Lα désignent les constantes asymptotiques de la multiplication parallèle des matrices en SD(nα , log n).
Calcul de l’adjointe et de l’inverse
L’algorithme de Preparata & Sarwate ne calcule pas toutes les puissances de la matrice A. Par conséquent le calcul de l’adjointe de A à
partir de la formule de Cayley-Hamilton doit se faire en n’utilisant que
√
les 2 d n e puissances de A déjà calculées, avec en plus les coefficients
c1 , c2 , . . . , cn du polynôme caractéristique et les matrices Lj formées
des lignes des premières puissances de A également disponibles.
L’astuce est de considérer les p matrices
P
j
Bi−1 = p−1
j=0 cn−p (i−1)−j−1 A
(1 ≤ i ≤ p) formées avec les coefficients du polynôme caractéristique,
avec la convention c0 = −1 et ck = 0 si k < 0 (rappelons que n ≤ p2 ).
On calcule ensuite la somme
Pp−1
Pp−1 Pp−1
pk =
pk+j
j=0 cn−pk−j−1 A
k=0 Bk A
k=0
en répartissant les calculs sur dlog pe étapes parallèles avec au maximum
p / log p multiplications de matrices n × n (i.e. des produits du type
Bk × Apk ) par étape.
Pn
`−1 = AdjA , puisque d’une
Or cette somme est égale à
`=1 cn−` A
part cn−` = 0 si ` > n et que d’autre part, si ` est compris entre
1 et n, ` correspond de manière unique à un couple (k, j) tel que
1 ≤ j, k ≤ p − 1 et ` − 1 = p k + j (division euclidienne de ` − 1 par
p). Ce qui donne l’adjointe puis l’inverse.
Ainsi la deuxième partie de l’algorithme de Preparata & Sarwate
pour le calcul de l’adjointe et de l’inverse de A peut être détaillée comme
suit.
Entrées :
— Les puissances A, . . . , Ap de la matrice A, ainsi que les puissances
2
A2p , . . . , Ap de la matrice Ap , toutes disponibles à l’issue des deux premières étapes de l’algorithme principal 9.2 ;
2
— La matrice L = [ L1 | L2 | · · · | Ln ] ∈ Ap×n formée des n matrices
9.2. Amélioration de Preparata et Sarwate
265
Lk ( 1 ≤ k ≤ n ) déjà calculées ;
— Enfin la matrice C formée à partir des coefficients
c1 , c2 ,. . . , cn du
P
polynôme caractéristique PA (X) = (−1)n X n − ni=1 ci X n−i :
C=
cn−1
cn−2
cn−p−1
..
.
cn−p−2
..
.
···
···
..
.
cn−p
cn−p−p
..
.
∈ Ap×p .
cn−p (p−1)−1 cn−p (p−1)−2 · · · cn−p (p−1)−p
P
j
On a alors Bi−1 = p−1
j=0 cij A et il est facile de voir que la k - ème ligne
de cette matrice n’est autre que la i - ème ligne de la matrice Tk := C Lk
où Lk , rappelons-le, est la matrice formée des k - èmes lignes des matrices
In , A, . . . , Ap−1 .
Sortie :
L’adjointe et l’inverse de A, c’est-à-dire les matrices
AdjA = An−1 − c1 An−2 − . . . − cn−1 A − cn In et A−1 =
1
cn
AdjA.
Les étapes du calcul :
Faisant suite aux étapes (1 à 6) qui calculent le polynôme caractéristique,
elles seront numérotées de 7 à 10. On pose r = dlog pe et D0 = 0nn (la
matrice carrée d’ordre n nulle) et s = bp/rc :
7. Calculer le produit T = C L = [ CL1 | CL2 | · · · | CLn ] (ce qui
revient à calculer en parallèle les produits de la matrice C qui est
une matrice p × p par les n matrices CLk qui sont des matrices
p × n ).
Cette étape permet d’écrire les matrices Bi−1P( 1 ≤ i ≤ p ).
is−1
Bk Apk ;
8. pour k de 1 à r faire Dk := Dk−1 + k=(i−1)s
Pp−1
9. Calculer Adj A := −(Dr + k=rs Bk Apk ) ;
10. Calculer A−1 = c1n Adj A.
La complexité de cette deuxième partie de l’algorithme de Preparata & Sarwate possède les mêmes bornes que l’algorithme principal du
polynôme caractéristique.
L’étape 7 se fait en Kα log n étapes comportant au total n p Cα pα
opérations arithmétiques dans l’anneau de base, utilisant au maximum
n p Lα (pα / log p) processeurs.
Les étapes 8 et 9 sont les plus coûteuses. Elles correspondent à un
total de r + 1 étapes parallèles comportant p multiplications de matrices carrées d’ordre n c’est-à-dire p Cα nα opérations arithmétiques de
266
9. Parallélisations de la méthode de Leverrier
base, ce à quoi il faut rajouter des additions de matrices n × n. Cela fait
un circuit de profondeur
(r + 1) Lα log n + O(log n) ≤
1
(log n + 1) Lα log n + O(log n) .
2
Le nombre de processeurs utilisés au cours de ces r + 1 étapes parallèles
est égal à s Lα (nα / log n) ≤ (Lα p nα) / log n log p puisque s = bp/rc ≤
p / log p.
9.3
Amélioration de Galil et Pan
Galil & Pan [32] réduisent les étapes les plus coûteuses de l’algorithme précédent à quatre multiplications de matrices rectangulaires.
Il s’agit plus précisément des l’étapes 1, 2 et 3 de l’algorithme principal (calcul du polynôme caractéristique) d’une part et des étapes 8 et
9 du calcul de l’adjointe d’autre part.
Par une réorganisation des étapes 1 et 2 de l’algorithme principal
qui font intervenir les procédures récursives Powers et Superpowers, on
remplace l’appel à ces procédures par l’appel récursif à une procédure
unique permettant de calculer les matrices
[ A | A2 | . . . | Ap−1 ] et [ Ap | A2p | . . . | Ap(p−1) ]
à partir des matrices
[ A | A2 | . . . | As−1 ] et [ As | A2s | . . . | A(s−1) s ]
√
où s =
p . Cela se fait en effectuant le produit d’une matrice rectangulaire n (s − 1) × n par une matrice rectangulaire n × ns qui donne
les puissances restantes :
A
A2
h
i
2
.
.. × As A2s · · · As
.
As−1
L’étape 3 de l’algorithme principal calcule les n produits Uj =
Lj Cj ∈ Ap×p pour en déduire les traces des puissances de A. Il est possible de réduire cette étape au calcul d’un seul produit de deux matrices
√
rectangulaires de types respectifs p × n2 et n2 × p où p = d n e. En
effet, si on écrit les éléments de chaque matrice Ak (pour 0 ≤ k ≤ p−1 )
sur une seule ligne, de manière à la représenter par la suite ordonnée de
2
ses n lignes, c’est-à-dire par vk ∈ A1×n , et si l’on fait de même avec
9.3. Amélioration de Galil et Pan
267
les matrices Apk ( 1 ≤ k ≤ n ), mais en déroulant cette fois chacune
d’elles sur une seule colonne ( Apk sera donc représentée, dans l’ordre
2
de ses colonnes, par wk ∈ An ×1 ), le calcul des traces sk revient alors
à calculer le produit des deux matrices rectangulaires :
v0
v1
· · · × w1
vp−1
v 0 w1
v 1 w1
..
.
v 0 w2
v 1 w2
..
.
vp−1 w1 vp−1 w2
w2
···
···
..
.
···
wp
=
v 0 wp
v 1 wp
..
.
.
· · · vp−1 wp
Il est clair que l’élément vi−1 wj de la i - ème ligne et j - ème colonne
de cette matrice est égal à spj+i−1 = TrApj+i−1 (1 ≤ i, j ≤ p).
On modifie enfin les étapes 8 et 9 du calcul de l’adjointe de A en
√
prenant q = d 3 n e, t = b(n + 1) / qc de manière à avoir qt ≤ n + 1 <
q(t+1), et on change les dimensions de la matrice C en la remplaçant par
C ∗ = (cij ) ∈ A(t+1)×q (avec les mêmes notations et la même convention
pour les cij ) ainsi que les dimensions des matrices L1 , L2 , . . . , Ln en
les remplaçant par des matrices L∗1 , L∗2 , · · · , L∗n définies exactement de
la même façon mais à partir des lignes des matrices In , A, · · · , Aq−1 , ce
qui fait qu’elles sont de type q × n au lieu d’être de type p × n.
2
On calcule alors la matrice T ∗ ∈ A(t+1)×n en effectuant le produit
d’une matrice (t + 1) × q par une matrice q × n2 :
T ∗ = C ∗ [ L∗1 | L∗2 | · · · | L∗n ] = [ C ∗ L∗1 | C ∗ L∗2 | · · · | C ∗ L∗n ]
en tenant compte du fait que la (i + 1) - ème ligne du bloc C ∗ L∗j n’est
autre que la j - ème ligne de la matrice
Bi =
q−1
X
j=0
j
cn−1−qi−j A =
q−1
X
ci+1,j+1 Aqi+j
(ici 0 ≤ i ≤ t) .
j=0
Avec ces modifications, les étapes 8 et 9 se ramènent donc, comme
on peut le constater, au calcul du produit de deux matrices rectangulaires (avec les mêmes notations que ci-dessus) qui est un produit d’une
268
9. Parallélisations de la méthode de Leverrier
matrice n × n (t + 1) par une matrice n (t + 1) × n
In
Aq
2q
B0 B1 B2 · · · Bt × A
..
.
Atq
:
.
Posant ` = qi + j, ce dernier produit est en effet égal à
t
X
i=0
qi
Bi A =
q−1
t X
X
(t+1) q−1
cn−1−qi−j A
qi+j
=
i=0 j=0
X
cn−1−` A` .
`=0
Comme (t + 1) q − 1 > n (d’après la définition même de t et de q ) et
que cn−1−` = 0 pour ` ≥ n, la matriceP
ainsi obtenue est exactement
l’opposée de l’adjointe de A : AdjA = − nl=1 cn−` A`−1 .
Les calculs de ces quatre produits de matrices rectangulaires auxquels
Galil & Pan réduisent l’algorithme de Preparata & Sarwate, et qui sont
des multiplications
respectifs donnés par le tableau suivant où
√ d’ordres
√
√
p , q = d 3 n e , t = b(n + 1) / qc,
p = d ne, s =
Multiplication
1ère multiplication
2ème multiplication
3ème multiplication
4ème multiplication
1er facteur
n (s − 1) × n
p × n2
(t + 1) × q
n × n (t + 1)
2ème facteur
n × ns
n2 × p
q × n2
n (t + 1) × n
s’effectuent en O(log2 n) étapes parallèles.
On fait d’autre part appel aux résultats concernant les notions d’algorithme bilinéaire et de rang tensoriel (voir la section 7.3), pour améliorer la complexité théorique de l’algorithme de Preparata & Sarwate
ainsi remanié, en faisant passer l’exposant de n dans cette complexité
(en taille et en nombre de processeurs) de 2, 876 à 2, 851 (si on prend le
α ≈ 2, 376 de Winograd & Coppersmith [19]).
Rappelons (voir la section 7.3.1) que le rang tensoriel de l’application
bilinéaire
f : Am×n × An×p −→ Am×p
associée à la multiplication des matrices m × n par des matrices n × p
à coefficients dans A (on note hm, n, piA cette application bilinéaire) est défini comme le rang de l’algorithme bilinéaire ou du tenseur
9.3. Amélioration de Galil et Pan
269
définissant hm, n, piA , c’est-à-dire le nombre minimum de multiplications essentielles nécessaires au calcul bilinéaire correspondant. Ce rang
est noté R hm, n, pi (nous omettons A en indice dans la mesure où tous
les résultats cités s’appliquent à n’importe quel anneau).
Outre les propriétés établies dans la section 7.3, il y a un résultat dû à
Coppersmith [18] pour le cas des matrices rectangulaires qui nous occupe
ici. Il est utilisé par Galil & Pan pour établir qu’il existe une constante
positive β estimée dans un premier temps à β = (2 log 2) / (5 log 5) ≈
0, 172 puis à β ≈ 0, 197 qui vérifie la propriété
R hm, mβ , mi = O(m2+ )
pour tout
> 0.
Les modifications des étapes les plus coûteuses aboutissent à des
multiplications de matrices rectangulaires de rangs respectifs :
Multiplication
1ère multiplication
2ème multiplication
3ème multiplication
4ème multiplication
Rang tensoriel
R hn5/4 , n, n5/4 i
R hn1/2 , n2 , n1/2 i
R ht + 1, q, n2 i
R hn, n (t + 1), ni
où q n1/3 et t n2/3 vérifient aussi qt ≤ n + 1 < q (t + 1).
On a alors :
Théorème 9.3 (Galil & Pan)
Le calcul du polynôme caractéristique, de l’adjointe et l’inverse d’une
1
matrice carrée d’ordre n est en SD(nα+ 2 −δ , log 2 n) où δ est un réel
strictement positif dépendant de α.
En particulier, pour α ≈ 2, 376 la taille du circuit arithmétique est un
O(n2,851 ).
Il suffit en effet, pour établir ce résultat, d’évaluer les quatre rangs
tensoriels indiqués dans le tableau ci-dessus en utilisant la constante β
de la multiplication des matrices rectangulaires ( β < 1 ). Pour cela, on
pose m = n1 / (4−4β) et r = n1 / (4−4β) , ce qui donne les estimations
R hm, mβ , mi = O(n(2+) / (4−4β) )
et
R hr, r, ri = O(nα(4−5β) / (4−4β) )
qui, multipliées entre elles, donnent
R hn5/4 , n, n5/4 i = O(nρ )
où
1
ρ = α+ +δ1
2
et
δ1 =
− β(α − 2)
·
4 − 4β
270
9. Parallélisations de la méthode de Leverrier
Comme α > 2 et β < 1, on peut prendre 0 < < β(α − 2) et δ1 > 0,
ce qui établit le résultat pour la première multiplication.
Pour les trois autres multiplications, on remarque que :
• d’une manière générale R hm, m4 , mi = O(mα+3 ) (multiplication par
α+3
blocs m × m ) et que, par conséquent, R hn1/2 , n2 , n1/2 i = O(n 2 ) =
1
O(nα+ 2 +δ2 ) avec δ2 = α−2
2 > 0.
2
(α−3)
η+3 ) = O(nα+ 21 +δ3 ) si l’on prend t = nη ,
• R ht + 1, q, n i = O(n
q = n1−η , et δ3 = (1 − η)(3 − α) + 12 avec 0 < η < 1.
1
• R hn, n (t + 1), ni = O(nα+η ) = O(nα+ 2 +δ4 ) avec δ4 = 12 − η pour le
même η.
Prenant 0 < η < 12 et (1 − η)(α − 3) < 12 , ce qui correspond au
cas concret η = 13 , cela donne bien inf (δ1 , δ2 , δ3 , δ4 ) > 0 et établit
le résultat o(nα+1/2−δ ) pour n’importe δ > 0 strictement inférieur à
inf (δ1 , δ2 , δ3 , δ4 ). Le résultat numérique en découle pour α < 2, 376.
En fin de compte l’exposant de n dans la complexité asymptotique
pour le calcul du polynôme caractéristique et de l’adjointe par la méthode de Preparata & Sarwate est de 2, 876 au lieu de 2, 851 de Galil &
Pan pour α = 2, 376.
Conclusion
Les algorithmes de Csanky, de Preparata & Sarwate, de Galil & Pan
ne sont en fait que des variantes parallélisées de la méthode de Le Verrier (1840) mais elles ont le mérite d’avoir ingénieusement réduit, et
de manière spectaculaire, la complexité des circuits arithmétiques permettant de résoudre ces problèmes dans le cas d’un anneau commutatif
autorisant les divisions exactes par les entiers. Les estimations de ces
algorithmes parallèles dans le cas de tels anneaux restent les meilleures
connues à l’heure actuelle.
10. Calcul du polynôme
caractéristique sur un
anneau commutatif
arbitraire
Introduction
Dans ce chapitre, nous présentons des algorithmes bien parallélisés
de calcul du polynôme caractéristique sur un anneau commutatif arbitraire.
Le premier résultat de cette sorte, exposé dans la section 10.1, a été
obtenu en 1982. L’estimation de son temps séquentiel est pessimiste,
mais il reste d’un grand intérêt théorique
Dans les sections suivantes nous expliquons les algorithmes de Chistov et de Berkowitz (amélioré) qui sont dans SD(nα+1 log n, log2 n).
On notera que le résultat est cependant moins bon en temps séquentiel que pour l’algorithme de Preparata & Sarwate (qui réclame la
division par un entier arbitraire) ou celui de Kaltofen (qui n’est pas bien
parallélisé).
10.1
Méthode générale de parallélisation
Tout programme d’évaluation (donc tout circuit arithmétique) sans
division à n indéterminées (xi )i=1..n sur un anneau A calcule un polynôme de A [x1 , . . . , xn ]. Valiant, Skyum, Berkowitz et Rackoff [95]
démontrent le résultat important suivant. La preuve, délicate, est bien
expliquée dans [Bur].
272
10. Polynôme caractéristique sur un anneau arbitraire
Théorème 10.1 Soit Γ un circuit arithmétique sans division, de taille
`, qui calcule un polynôme f de degré d en n variables sur un anneau A. Alors il existe un circuit arithmétique homogène Γ0 de taille
O(`3 d6 ) et de profondeur O(log(` d) log d) qui calcule (les composantes
homogènes de) f ( 1 ). En outre la construction de Γ0 à partir de Γ est
LOGSPACE .
En particulier :
Corollaire 10.1.1 Toute famille (Q` ) de polynômes de degrés d =
O(`k ) qui peut être calculée au moyen d’une famille uniforme de circuits arithmétiques peut aussi être calculée dans la classe N C 2 .
En appliquant le théorème 10.1 à l’algorithme du pivot de Gauss
auquel on fait subir la procédure d’élimination des divisions à la Strassen,
et vu que le déterminant qu’il calcule est un polynôme de degré n, on
obtient le résultat suivant dû à Borodin, Hopcroft et Von zur Gathen [9] :
Proposition 10.1.2 Le déterminant d’une matrice n × n est calculé
par un programme d’évaluation de taille O n18 log3 n log3 log n et de
profondeur O(log2 n).
Dans la construction correspondant au théorème 10.1 est utilisée la
multiplication rapide des polynômes. Avec la multiplication
usuelle des
polynômes, la proposition 10.1.2 donne O n21 opérations arithmétiques dans l’anneau de base.
10.2
Algorithme de Berkowitz amélioré
Introduction
Utilisant la méthode de partitionnement [Gas, FF], attribuée à Samuelson ([79]), Berkowitz [6] a pu exhiber un circuit arithmétique parallèle de taille O(nα+1+ ) et de profondeur O(log2 n), où est un réel
positif quelconque.
1. On trouve dans [Bur] la majoration O(log(` d) log d + log n) pour la profondeur. Le terme log(n) supplémentaire est nécessaire lorsque d = 1 si on a log 1 = 0.
Mais la convention de notation 1.6.1 que nous avons choisie pour log d, conforme à
la longueur du code binaire de d, nous donne log 1 = 1.
10.2. Algorithme de Berkowitz amélioré
273
Il a ainsi amélioré de manière décisive la complexité asymptotique
du calcul des déterminants, polynômes caractéristiques, et adjointes de
matrices à coefficients dans un anneau commutatif quelconque A.
Nous allons donner une version légèrement améliorée de l’algorithme
de Berkowitz, due à Eberly [29], qui ramène sa taille à O(nα+1 log n)
sans en changer la profondeur. Pour cela nous donnons une version plus
simple de la récurrence utilisée pour le calcul des coefficients du polynôme caractéristique. Nous donnons également une estimation précise de
la constante qui intervient dans le (( grand O )) de la complexité séquentielle (cf. [1]).
Soit A = (aij ) ∈ An×n une matrice carrée d’ordre n ≥ 2 sur un anneau commutatif arbitraire A. Conformément aux notations introduites
dans la section 1.1, pour tout entier r (1 ≤ r ≤ n), on désigne par Ar la
sous-matrice principale dominante d’ordre r de A. On notera ici Rr la
matrice Ar+1,1..r ∈ A1×r et Sr la matrice A1..r,r+1 ∈ Ar×1 . Rappelons
la formule de Samuelson (2.14) vue à la section 2.6.
(
(ar+1,r+1 − X) Pr (X) +
r+1−k
Pr+1 =
Pr+1
k−2 S ) p + · · · + (R S ) p
r
0
r r
k−2 X
k=2 (Rr Ar
P
où Pr (X) = ri=0 pr−i X i . Notons Qr+1 le polynôme
−X r+1 + ar+1,r+1 X r + Rr Sr X r−1 + Rr Ar Sr X r−2 + · · · + Rr Ar−1
r Sr .
On peut aussi écrire la formule de Samuelson sous la forme (2.15) :
−−→
−
→
Pr+1 = Toep(Qr+1 ) × Pr
−
→
où Pr est le vecteur colonne t (p0 , p1 , . . . , pr ) des coefficients du polynôme P et Toep(Qr+1 ) ∈ A(r+2)×(r+1) est la matrice de Toeplitz suivante
définie à partir du polynôme Qr+1 :
−1
0
··· ···
0
..
..
a
.
−1
.
r+1,r+1
..
..
..
..
.
.
.
R r Sr
.
Toep(Qr+1 ) =
..
..
..
..
.
.
.
.
0
..
..
R Ar−2 S
.
.
−1
r
r r
Rr Arr−1 Sr Rr Ar−2
r Sr . . . Rr Sr ar+1,r+1
Le calcul du polynôme caractéristique consiste donc :
274
10. Polynôme caractéristique sur un anneau arbitraire
– à calculer d’abord les coefficients de la matrice Toep(Qr+1 ) – qui
interviennent dans l’égalité (2.15) – ou, ce qui revient au même,
la famille T = {RM i S}r−1
i=0 lorsque R, M, S sont respectivement
des matrices 1 × r, r × r, et r × 1, et lorsque r est un entier tel
que 2 ≤ r < n (M = Ar , R = Rr , S = Sr ) ;
– à calculer ensuite le polynôme Pn dont le vecteur des coefficients,
compte tenu de (2.15), est donné par :
−
→
Pn = Toep(Qn ) × Toep(Qn−1 ) × · · · × Toep(Q1 )
(10.1)
Dans son papier original [6], Berkowitz démontre que les familles
1/2
1/2
U = {RM i }ni=0 et V = {M j n
1/2
S}nj=0
peuvent être calculées par un circuit arithmétique parallèle en SD(nα+ ,
log2 n) pour en déduire que le calcul du polynôme caractéristique se fait
en SD(nα+1+ , log2 n) .
La version parallèle améliorée et sa complexité
Nous utilisons comme d’habitude la notation 7.2.1 page 195.
Proposition 10.2.1 On considère un entier r ≥ 2 et des matrices
R ∈ A1×r , M ∈ Ar×r , S ∈ Ar×1 .
La famille
T = {R M i S}r−1
i=0
peut être calculée par un circuit arithmétique dont la taille et la profondeur sont majorées respectivement par
Cα rα log r + O(rα )
et
Kα log2 r + O(log r) .
Preuve.
Soit r ≥ 2. On utilisera, pour
1 l’analyse
de complexité des algorithmes,
les entiers ν = dlog4 re = 2 log r et η = dlog re qui vérifient les
inégalités : 22ν−2 < r ≤ 22ν et 2η−1 < r ≤ 2η (on a aussi 1 ≤ ν ≤
2ν − 1 ≤ η ≤ 2ν).
Toute matrice
carrée A d’ordre r sera plongée, selon le cas, soit
dans
A 0
A 0
η
une matrice
carrée d’ordre 2 soit dans une matrice
0 0
0 0
2ν
carrée d’ordre 2 (chacun des 0 désignant ici une matrice nulle de
dimensions convenables).
10.2. Algorithme de Berkowitz amélioré
275
Il faut cependant
que, dans les deux cas, l’élévation au carré
remarquer
A 0
de la matrice
se fait à l’aide d’un circuit arithmétique de taille
0 0
2 2
A 0
A 0
α
Cα r et de profondeur Kα log r puisque
=
.
0 0
0 0
De même, le produit
d’une
matrice 2k × 22ν (k = 1, . . . , ν) par une
A 0
2ν
2ν
matrice du type
∈ A2 ×2 avec A ∈ Ar×r peut être obtenu
0 0
par le calculdu produit d’une matrice 2k × 2η par une matrice du
A 0
η
η
∈ A2 ×2 à cause du fait que, dans ces deux produits,
type
0 0
les r premières colonnes sont les mêmes alors que les colonnes restantes
sont nulles. Ce qui fait que le produit en question peut être obtenu par
22(η−k) multiplications en parallèle de blocs 2k ×2k et de 2η−k (2η−k −1)
additions en parallèle des blocs produits obtenus, c’est-à-dire par un circuit arithmétique de taille 2η−k [2η−k Cα 2kα + (2η−k − 1)22k ] ( 2 ) et de
profondeur (Kα + 1) k .
Considérons à présent, pour k = 1, . . . , ν , la matrice Uk dont les
2k −1
lignes sont les éléments de la famille {RM i }i=0
considérée comme une
matrice 2k × 2η et la matrice Vk dont les colonnes sont les éléments
k −1
ν
de la famille {M j2 S}2j=0
considérée comme une matrice 2η × 2k . La
ν
η
famille T s’obtient alors en calculant la matrice Uν ∈ A2 ×2 puis
η
ν
la matrice Vν ∈ A2 ×2 et enfin le produit matriciel Wν = Uν Vν .
La famille T = {R M i S}r−1
déterminée par la donnée
i=0 est entièrement
ν ×2ν
2
de la matrice Wν = (wij ) = Uν Vν ∈ A
puisque : R M k S = wij
2ν−1
ν
(0 ≤ k ≤ 2
) si et seulement si k = (i −
1)ν + (j − 1)2 i.e. si et
k
k
seulement si j = 2ν + 1 et i = k + 1 − 2ν 2 .
Le calcul de T se fait donc en deux phases : une première phase
de calcul des matrices Uν et Vν et une deuxième phase de calcul du
produit Wν = Uν Vν .
• Coût de la phase 1 :
Le calcul de Uν et de Vν sefait de proche en proche à partir de
S
η
η
U0 = R 0 ∈ A1×2 , V0 =
∈ A2 ×1 et des puissances de
0
M obtenues par élévations successives au carré, c’est-à-dire les matrices
s
M 2 ( 1 ≤ s ≤ 2ν − 1 ).
2. Le premier terme du crochet provient des multiplications de blocs 2k × 2k , et
le second terme indique le nombre d’additions dues aux additions des blocs.
276
10. Polynôme caractéristique sur un anneau arbitraire
On a en effet, pour k = 1, . . . , ν :
k
k−1 −1
{RM i }2i=0−1
{M
j2ν
= {RM i }2i=0
2k −1
= {M
S}j=0
j2ν
2k−1 −1
S}j=0
k−1
∪ {RM i+2
∪ {M
k−1 −1
}2i=0
2ν (j+2k−1 )
et
2k−1 −1
S}j=0
Ce qui donne, de manière plus précise, les relations matricielles suivantes
ν
(si on pose N = M 2 ) :
h
i
Uk−1
k
η
η
k
]
Uk = ] ∈ A2 ×2 et Vk = Vk−1 V
∈ A2 ×2
k−1
Uk−1
avec
2k−1
]
et
U
k−1 = Uk−1 × M
2k−1
]
V
× Vk−1 .
k−1 = N
D’où l’algorithme suivant pour le calcul de Uν et Vν (comportant
2ν étapes successives) à partir des données initiales U0 , V0 (c’est-à-dire
R , S) :
k
1. L’étape k (1 ≤ k ≤ ν) consiste à calculer Uk et M 2 ; pour cela
k−1
deux opérations seront exécutées en parallèle sur M 2
qui est
une matrice 2η ×2η (déjà calculée à l’étape k−1 ) : l’élever au carré
et la multiplier à gauche par Uk−1 qui est une matrice 2k−1 × 2η .
A la fin de ces ν étapes, on obtient la matrice Uν et la matrice
ν
N = M 2 (figure 10.1).
R
Etape 0
M
2
Etape 1
M
RM
2
RM
3
M
RM
RM
22
Etape 2
4
.
.
.
...
.
2ν −1− 1
.
.
.
RM
....
2ν−1
RM
2ν −1
Etape (ν − 1)
M
.
.
.
2ν
M =N
Etape ν
Figure 10.1 – Calcul de Uν
les liens en trait pointillé indiquent les multiplications
à effectuer au cours d’une étape pour passer à l’étape suivante
10.2. Algorithme de Berkowitz amélioré
277
k
2. L’étape ν + k (1 ≤ k ≤ ν) consiste à calculer Vk et N 2 : là
encore, il s’agit d’élever au carré une matrice 2η × 2η et de la
multiplier à droite par Vk−1 qui est une matrice 2η × 2k−1 .
A l’issue de ces ν nouvelles étapes, on obtient Vν et la matrice
ν
N 2 (figure 10.2).
S
données
N
21
Etape ν+1
N
NS
2
NS
22
3
Etape ν+2
N
NS
22
N S
.
.
.
...
.
.
.
.
2ν −1− 1
N
S
....
2ν−1
N
2ν −1
Etape 2ν −1
N
.
.
.
2ν
N
S
Etape 2ν
Figure 10.2 – Calcul de Vν
les liens en trait pointillé indiquent les multiplications
à effectuer au cours d’une étape pour passer à l’étape suivante
Si l’on utilise les multiplications par blocs 2k−1 ×2k−1 (ils sont ici au
nombre de 2η−k+1 blocs ), l’étape k (resp. ν + k) ci-dessus est réalisée
par un circuit de taille :
Cα rα + 2η−k+1 × [ 2η−k+1 Cα 2(k−1)α + (2η−k+1 − 1)22(k−1) ].
et de profondeur égale à max {Kα η, (k − 1) (Kα − 1) + η} = Kα η
(puisque Kα est supposé ≥ 1).
Tenant compte du fait qu’il y a 2ν étapes et que η ≤ 2ν < log r + 2 ,
l’algorithme calculant Uν et Vν est donc réalisé par un circuit de profondeur 2νKα η ≤ Kα (log r + 1)(log r + 2) et de taille majorée par :
P
Cα rα (log r + 2) + 2η+1 νk=1 (Cα 2(α−2)(k−1) + 1) et donc par :
P
(Cα rα + 2 r)(log r + 2) + 4 Cα r νk=1 2(α−2)(k−1) qui est égal à :
(Cα rα + 2 r)(log r + 2) + 4 Cα
2(α−2)ν − 1
r.
2(α−2) − 1
Cette taille est donc majorée par :
α
α
8 Cα
(Cα rα + 2 r)(log r + 2) + (α−2)
r 2 = Cα rα log r + 2 Cα rα + O (r 2 ) .
2
−1
278
10. Polynôme caractéristique sur un anneau arbitraire
qui est clairement O(rα log r) .
• Coût de la phase 2 :
Cette phase consiste à calculer le produit Uν ×Vν qui peut s’effectuer
par des multiplications de blocs 2ν × 2ν , en parallèle et en 2 grandes
étapes.
– La première étape consiste à calculer en parallèle 2η−ν produits de
blocs 2ν × 2ν , avec 2η−ν Cα 2να opérations arithmétiques dans l’anneau
de base, ce qui donne une profondeur totale de Kα ν ;
– Il s’agit dans la deuxième étape de calculer en parallèle la somme
des 2η−ν produits obtenus précédemment, faisant intervenir (2η−ν − 1) 22ν
additions dans l’anneau de base, à l’aide d’une famille de circuits binaires
équilibrés de profondeur η − ν .
Le nombre total d’opérations arithmétiques, dans l’anneau de base,
qui interviennent dans ces deux grandes étapes du calcul de T, correspondant à une profondeur totale de (Kα − 1) ν + η, est donc majoré
par :
2η−ν Cα 2αν + 2η+ν ≤ Cα 2(α+1)ν + 23ν ≤ Cα 2α+1 r
α+1
2
3
+ 8r2 .
α+1
puisque η ≤ 2ν , α ≤ 3 et Cα ≥ 1 . Ce qui fait aussi O(r 2 ) avec
une constante asymptotique égale à 2α+1 Cα . Ainsi, le calcul de T à
partir de Uν et Vν se fait par un circuit parallèle de taille O(r
de profondeur
(Kα − 1) ν + η ≤
puisque Kα ≥ 1, ν ≤
1
2
α+1
2
) et
1
(Kα + 1) log r + Kα ≤ Kα (1 + log r)
2
(2 + log r) et r ≥ 2.
Nous résumons dans le tableau ci-dessous l’analyse de complexité qui
vient d’être faite et qui établit le résultat annoncé.
Etapes
1ère phase
2ème phase
Total
Profondeur
Kα (log r + 1)(log r + 2)
Kα (log r + 1)
Kα (log r + 1)(log r + 3)
Taille
Cα
+ O (rα )
α+1
O(r 2 )
Cα rα log r + O (rα )
rα log r
La différence essentielle avec l’algorithme de Berkowitz [6] réside dans
la simplification de la récurrence permettant de calculer de proche en
proche les matrices Uν et Vν : à chaque pas, la multiplication par une
10.2. Algorithme de Berkowitz amélioré
279
seule matrice (au lieu de dn e matrices), avec recours à la multiplication par blocs, a permis de réduire le nombre d’opérations arithmétiques
dans l’anneau de base, en éliminant le facteur n .
La démonstration de la proposition 10.2.1 a permis de donner une
estimation précise de la constante asymptotique : cette constante est en
effet égale à Cα ; elle est la même que la constante asymptotique de la
multiplication des matrices.
Théorème 10.2 Les coefficients du polynôme caractéristique d’une matrice carrée d’ordre n peuvent être calculés par un circuit arithmétique dont la taille et la profondeur sont respectivement majorées par
1
α+1 log n + O(nα+1 ) et par 2 K log2 n + O(log n) .
α
α+1 Cα n
Preuve. Le polynôme caractéristique de la matrice A = (aij ) n’est
autre que le polynôme Pn donné par la formule (10.1) :
−
→
Pn = Toep(Qn ) × Toep(Qn−1 ) × · · · × Toep(Q1 ).
Le calcul des coefficients (de la forme R M i S ) du polynôme Qk+1 (pour
1 ≤ k ≤ n − 1 ) se fait, d’après la proposition 10.2.1, en O(k α log k) .
De manière plus précise, le calcul de la totalité des matrices Toep(Qk+1 )
se fait donc avec une profondeur majorée par Kα (log n + 1)(log n + 3)
et une taille majorée par :
Cα
n−1
X
α
[k α log k + 2k α + O(k 2 )] ;
k=1
c’est-à-dire par :
α
Cα α+1
2Cα α+1
log n +
+ O(n 2 +1 )
n
n
α+1
α+1
à cause du fait :
n−1
X
k=1
α
k log k < (
n−1
X
k=1
α
k ) log n
et
n−1
X
k=1
α
α+1
Z
k <n
0
1
xα dx =
nα+1
·
α+1
D’autre part, le produit (10.1) peut être calculé à l’aide d’un circuit
binaire équilibré avec O(nα+1 ) opérations arithmétiques de base et une
profondeur majorée par Kα log2 n.
280
10. Polynôme caractéristique sur un anneau arbitraire
Cela donne en fin de compte, dans l’anneau de base, un nombre total
d’opérations arithmétiques majoré par
α
Cα α+1
2 Cα α+1
Cα α+1
n
log n+
n
+O(n 2 +1 ) =
n
log n+O(nα+1 ),
α+1
α+1
α+1
avec un circuit arithmétique de profondeur majorée par
2 Kα (log n + 1)2 + Kα = 2 Kα log2 n + O(log n).
Proposition 10.2.2 Les coefficients des polynômes caractéristiques de
toutes les sous-matrices principales dominantes d’une matrice carrée
d’ordre n peuvent être calculés en SD(nα+1 log n, log2 n) (avec les mêmes estimations que celles du théorème 10.2 pour les constantes asymptotiques).
Preuve. En effet, on a A = An , et les coefficients du polynôme caractéristique Pr de la sous-matrice principale dominante Ar de A ( 1 ≤
r ≤ n ) sont donnés par les vecteurs :
−
→
Pr = Toep(Qr ) × Toep(Qr−1 ) × · · · × Toep(Q1 )
Ces vecteurs ne sont autres que les troncatures successives (pour r allant
de 2 à n ) du second membre de (10.1) : ils peuvent donc être calculés par
un algorithme parallèle des préfixes. Le circuit que nous avons représenté
(figure 10.3) correspond à l’une des solutions du (( Calcul parallèle des
préfixes )) [64], que nous avons présentées dans la section 5.3. Il s’agit
d’un circuit parallèle de profondeur dlog ne + 1 et de taille majorée par
3n. Comme il s’agit de multiplications matricielles, chaque nœud interne
du circuit (représenté par une croix dans la figure) correspond à un circuit de multiplication de matrices de profondeur Kα log n avec O(nα )
opérations arithmétiques dans l’anneau de base.
Le calcul des Pr ( 2 ≤ r ≤ n ) à partir des matrices Toep(Qk ) se
fait donc par un circuit de taille O(nα+1 ) et de profondeur majorée par
Kα (log n + 1) (log n + 2). On conclut de la même façon que le théorème
10.2 pour le produit des matrices de Toeplitz.
t
u
Corollaire 10.2.3 Le déterminant et l’adjointe d’une matrice carrée
d’ordre n se calculent en SD(nα+1 log n, log2 n) (avec les mêmes bornes
que celles du théorème 10.2 pour les constantes asymptotiques).
Preuve. Le déterminant de A n’est autre que PA (0). D’autre part, la
matrice adjointe de A est donnée par la formule :
Adj(A) = −(p0 An−1 + p1 An−2 + · · · + pn−2 A + pn−1 In )
10.2. Algorithme de Berkowitz amélioré
Toep(Q1) Toep(Q4)
Toep(Q8)
...
...
x
x
x
x
Toep(Q9) Toep(Q12)
Toep(Q16)
...
...
x
x
x
281
x
x
x
x
x
x
x
x
x
x
P1 P2
...
P5 P6 P7 P8
x
x
x
P9 P10
x
x
...
x
P13
x
x
...
x
P16
Figure 10.3 – Calcul des Pr pour 1 ≤ r ≤ n (ici n = 24 )
où PA (X) =
Pn
i=1 pn−i X
i.
Preparata & Sarwate (voir section 9.2) donnent un algorithme récursif
(Powers (A, n)) pour calculer les n premières puissances de A avec un
circuit arithmétique parallèle de profondeur Kα log n (log n + 1) et de
taille majorée par (2n − 3) Cα nα .( 3 )
Le résultat est alors obtenu en remarquant que Adj (A) se calcule à
partir des puissances de A en 1 + dlog ne étapes avec O(n3 ) opérations
arithmétiques de base.
t
u
Remarque 10.2.4 La méthode de Baur & Strassen [5] pour le calcul
des dérivées partielles (cf. section 3.3) montre que le calcul de l’adjointe
d’une matrice a toujours un coût voisin de celui de son déterminant.
La construction originale ne se préoccupe pas de la profondeur, mais le
résultat a été amélioré par Kaltofen et Singer [52] : tout circuit arithmétique de taille τ et de profondeur π calculant une fonction polynomiale
(sur un anneau) ou une fonction rationnelle (sur un corps) donne un circuit de taille 4τ et de profondeur O(π) qui calcule la fonction et toutes
3. Le (( parallel prefix algorithm )) (section 5.3) donne le même résultat pour la
profondeur mais une taille majorée par 3n Cα nα .
282
10. Polynôme caractéristique sur un anneau arbitraire
ses dérivées partielles, et ceci indépendamment du nombre de variables
d’entrées du circuit.
Remarque 10.2.5 Les théorèmes de complexité que nous venons d’établir ne citent que deux paramètres de complexité : la taille et la profondeur des circuits. Mais une analyse minutieuse des algorithmes étudiés
nous permet également d’avoir le nombre de processeurs utilisés par
ces algorithmes dans le modèle PRAM, c’est-à-dire la largeur du circuit arithmétique correspondant. Ce troisième paramètre peut être exprimé en fonction de la largeur d’un circuit arithmétique (de profondeur
Kα log n et de taille Cα nα ) qui calcule le produit de deux matrices carrées d’ordre n. Il est facile de vérifier que le résultat trouvé est le même
que celui obtenu par application directe du principe de Brent à cet algorithme parallèle, c’est-à-dire un nombre de processeurs de l’ordre de
O(nα+1 / log n).
Remarque 10.2.6 Concernant les questions d’uniformité et de coût
de construction des circuits, ainsi que la taille des coefficients intermédiaires, le travail de Matera & Turull Torres [69] donne, dans le cas de
l’anneau Z des entiers relatifs, une construction effective, avec une taille
bien contrôlée des coefficients, des circuits de base qui interviennent dans
l’algorithme de Berkowitz.
Traduisant les opérations arithmétiques de Z (addition et multiplication) par des circuits booléens de profondeur O(log b) où b est la taille
maximum de la représentation binaire des coefficients de la matrice
donnée A ∈ Zn×n , ils obtiennent :
– pour la multiplication de deux matrices n×n sur Z un circuit booléen
de taille O(n3 b2 ) et de profondeur O(log (bn)) ;
– pour la taille des coefficients intermédiaires calculés, une majoration
de l’ordre de O(n(b + log n)) ;
– pour l’algorithme de Berkowitz, une famille uniforme de circuits booléens de profondeur O(log (n) log (bn)) et de taille O(n6 b2 log2 (n)) .
Cette construction, appliquée à l’algorithme amélioré que nous avons
présenté, donne une famille uniforme de circuits booléens de même profondeur, avec la même majoration pour la taille des coefficients intermédiaires, mais de taille réduite à O(n5 b2 log2 n).
Le facteur n ainsi économisé provient essentiellement des étapes correspondant aux figures 10.1 page 276 et 10.2 page 277 de notre algorithme.
10.3. Méthode de Chistov
10.3
283
Méthode de Chistov
Introduction
On considère une matrice carrée A ∈ An×n on pose B = In − XA ∈
A[X]n×n , Br est la sous-matrice principale dominante d’ordre r de B
et Q(X) = det B.
L’algorithme est basé sur les formules suivantes (ce sont les équations
(2.16) et (2.18) établies à la section 2.7.1) valables
dans l’anneau des développements limités à l’ordre n, An = A[X] X n+1 :
Q(X)−1 = [ det(In − XA) ]−1 =
n
Y
B −1
r
.
(10.2)
Er (Ar )k Er X k .
(10.3)
r,r
r=1
et, en notant Er la r -ème colonne de Ir :
Br−1
r,r
mod X n+1 = 1 +
n
X
t
k=1
Rappelons alors le principe général de l’algorithme 2.12 donné en
section 2.7.1.
Algorithme de Chistov, principe général
Entrée : la matrice A ∈ An×n .
Sortie : le polynôme caractéristique P (X) de A.
Début
Étape 1 :
Calculer pour r, k ∈ {1, . . . , n} les produits t Er (Ar )k Er ,
ce qui donne les polynômes B −1
r r,r (formule (10.3)).
Étape 2 :
Calculer le produit des n polynômes précédents modulo X n+1 ,
ce qui donne Q(X)−1 mod X n+1 (formule (10.2)).
Étape 3 :
Inverser modulo X n+1 le polynôme précédent : on obtient Q(X).
Étape 4 :
Prendre le polynôme réciproque à l’ordre n du polynôme Q(X).
On obtient P (X) en multipliant par (−1)n .
Fin.
284
10. Polynôme caractéristique sur un anneau arbitraire
La version parallèle et sa complexité
Étudions pour chacune des étapes de cet algorithme, la taille et la
profondeur d’un circuit arithmétique correspondant qui tire le meilleur
parti de la multiplication rapide des matrices et permet d’obtenir un
temps parallèle en O(log2 n).
• Coût de l’étape 1 :
Chacun des éléments t Er (Ar )k Er qu’il s’agit de calculer est obtenu
en prenant la r - ème composante du vecteur-colonne (Ar )k Er . On doit
donc calculer simultanément, pour tous r compris entre 1 et n et pour
chaque r , les n produits (matrice × vecteur) (Ar )k Er (1 ≤ k ≤ n) .
Pour évaluer la complexité de ce calcul, considérons l’entier ν ∈ N
tel que 2ν−1 < r ≤ 2ν , c’est-à-dire ν = dlog re , et ramenons la matrice
Ar à une matrice 2ν ×2ν en remplissant de zéros les rangées supplémentaires. Ainsi, toutes nos matrices (Ar )k seront considérées comme des
matrices 2ν × 2ν , et Er comme une matrice 2ν × 1 : cela ne change pas
les produits t Er (Ar )k Er recherchés. Considérons d’autre part l’entier
η ∈ N vérifiant 2η−1 ≤ n < 2η c’est-à-dire η = blog nc + 1 .
On procède alors en η sous-étapes successives (numérotées de 0 à η −1),
chacune utilisant le résultat de la précédente.
j
À l’étape j (0 ≤ j ≤ η − 1), on élève au carré la matrice (Ar )2 puis
on la multiplie à droite par la matrice
j −1
(Er |Ar Er | . . . |(Ar )2
pour obtenir la matrice (Ar )2
j −1
(Er |Ar Er | . . . |(Ar )2
j+1
Er ) ∈ A2
ν ×2j
et la matrice
j
j+1 −1
Er |(Ar )2 Er | . . . |(Ar )2
ν ×2j+1
Er ) ∈ A2
.
À la fin de ces η étapes (faisant j = η − 1 ), on obtient la matrice
η −1
(Er |Ar Er | . . . |(Ar )2
Er ) ∈ A2
ν ×2η
dont les éléments de la r - ème ligne, plus précisément les n premiers
(on a n < 2η ), ne sont autres que les éléments t Er (Ar )k Er recherchés.
Pour chaque r (1 ≤ r ≤ n ≤ 2η − 1) , on a ainsi η sous-étapes,
chacune d’elles comportant l’élévation au carré d’une matrice 2ν × 2ν
(en fait d’une matrice r × r ), et la multiplication d’une matrice 2ν × 2ν
par une matrice 2ν × 2j . Utilisant pour cette dernière opération les multiplications par blocs 2j × 2j (quitte à plonger la matrice 2ν × 2ν dans
10.3. Méthode de Chistov
285
une matrice 2η × 2η et la matrice 2ν × 2j dans une matrice 2η × 2j , on
obtient pour chacune des η sous-étapes considérées un nombre d’opérations arithmétiques majoré par :
Cα rα + 22η [Cα 2(α−2)j + 1] ≤ Cα rα + 4n2 [Cα 2(α−2)j + 1].
Cela est dû au fait que η − 1 ≤ log n (et 2η ≤ 2n ) qui permet d’obtenir
les majorations suivantes :
η−1
X
α
α
Cα r ≤ Cα r log n
et
j=0
η−1
X
2(α−2)j ≤
j=0
2 nα−2
·
2α−2 − 1
D’où la majoration du nombre d’opérations arithmétiques intervenant
(pour chaque valeur de r) dans le calcul des n produits t Er (Ar )k Er
( 1 ≤ k ≤ n) :
Cα rα log n + 4n2 [ log n +
2 Cα nα−2
].
2α−2 − 1
Comme r varie de 1 à n , le calcul de l’étape 1 s’effectue à l’aide d’un
circuit arithmétique de taille O(nα+1 log n) et de profondeur O(log2 n).
Plus précisément, la taille est majorée par :
3
[ 4n +Cα
n
X
rα ] log n+
r=1
8 Cα
Cα α+1
nα+1 ≤
n
log n+O(nα+1 ) .
−1
α+1
2α−2
et la profondeur par :
Kα η log n +
η−1
X
j=0
(Kα j + η − j) ≤
3 Kα + 1
log2 n + O(log n).
2
On peut remarquer qu’avec la multiplication usuelle des matrices
(α = 3 ), l’étape 1 correspond à un circuit arithmétique parallèle de
profondeur O(log2 n) et de taille O(n4 log n) + O(n4 ) = O(n4 log n)
η−1
X
puisque
n2 2j = n2 (2η − 1) < 2n4 .
j=0
• Coût de l’étape 2 :
On doit calculer le produit tronqué à l’ordre n des n polynômes de
degré ≤ n calculés à l’étape précédente. Ce calcul se fait à l’aide d’un
circuit binaire équilibré en SD(n3 , log2 n).
286
10. Polynôme caractéristique sur un anneau arbitraire
• Coût de l’étape 3 :
Il s’agit d’inverser modulo X n+1 le polynôme Q̃ de degré ≤ n obtenu
à l’étape précédente. Ce polynôme est de la forme Q̃ = 1 − XR où R
est un polynôme de degré n en X. Inverser Q̃ modulo X n+1 revient à
calculer le produit
ν
ν
Q(X) = (1 + XR)(1 + X 2 R2 ) . . . (1 + X 2 R2 ) mod X n+1 .
Cela s’effectue en SD(n2 log n, log n log log n) à l’aide d’un circuit binaire équilibré.
On peut accélérer le calcul des deux étapes précédentes en utilisant
une multiplication rapide des polynômes (mais cela n’améliore pas sensiblement le résultat final).
• Coût de l’étape 4 :
Cette étape, de profondeur 1, n’intervient pas dans la complexité de
l’algorithme.
Nous donnons ci-dessous un tableau résumant l’analyse qui vient
d’être faite pour la complexité de l’algorithme de Chistov, montrant
que ce dernier est SD(nα+1 log n, log2 n) si l’on utilise la multiplication
rapide des matrices (α < 3) avec une estimation précise des constantes
asymptotiques pour la taille et pour la profondeur.
Etape
Etape 1
Profondeur
3Kα +1
2
log2 n + O(log n)
Taille
Cα
α+1
nα+1 log n + O(nα+1 )
Etape 2
log2 n + O(log n)
O(n3 )
Etape 3
O(log n log log n)
O(n2 log n)
Etape 4
1
négligeable
Tableau 10.3
Complexité de la version parallèle de l’algorithme de Chistov
Si l’on utilise la multiplication usuelle (α = 3), cela donne un algorithme en SD(n4 log n, log2 n) . Dans ce dernier cas, l’algorithme séquentiel élémentaire donné à la section 2.7.2 est donc préférable (sur une
machine séquentielle).
10.4. Applications des algorithmes
287
Théorème 10.3 L’algorithme de Chistov calcule les coefficients du
polynôme caractéristique d’une matrice carrée d’ordre n par un circuit
arithmétique de profondeur O(log2 n) et de taille O(nα+1 log n) avec des
constantes asymptotiques estimées respectivement à 32 (Kα + 1) pour la
1
profondeur et α+1
Cα pour la taille.
Enfin, comme dans le cas de l’algorithme de Berkowitz, on obtient
facilement le résultat complémentaire suivant.
Proposition 10.3.1 Les coefficients des polynômes caractéristiques de
toutes les sous-matrices principales dominantes d’une matrice carrée
d’ordre n peuvent être calculés en SD(nα+1 log n, log2 n) par un algorithme directement dérivé de celui correspondant au théorème 10.3.
Remarque 10.3.2 Remarquons que dans l’estimation de la taille des
circuits arithmétiques construits à partir des algorithmes de Chistov et
de Berkowitz amélioré, les termes en nα+1 log n sont les mêmes pour
les deux algorithmes alors que les termes en nα+1 sont respectivement
Cα
8 Cα
nα+1 pour Chistov et à seulement 2 α+1
estimés à 2α−2
nα+1 pour
−1
Berkowitz amélioré (le rapport du premier coefficient au second étant
strictement supérieur à 16).
10.4
Applications des algorithmes à des anneaux
commutatifs
Application en évaluation dynamique
Le calcul des déterminants et des polynômes caractéristiques de toutes les sous–matrices principales d’une matrice donnée trouve une application intéressante en évaluation dynamique.
Lorsqu’on travaille dans la clôture algébrique dynamique [28] d’un
corps K, on se trouve dans la situation standard suivante : on a des
variables x1 , . . . , xn qui représentent des éléments ξ1 , . . . , ξn algébriques sur K. On sait que ces éléments vérifient un système triangulaire
d’équations algébriques.
De sorte que le corps K[ξ1 , . . . , ξn ] est un quotient d’une K – algèbre
de dimension finie
AP1 ,...,Pn = K[x1 , . . . , xn ] /hP1 (x1 ), P2 (x1 , x2 ), Pn (x1 , . . . , xn )i .
288
10. Polynôme caractéristique sur un anneau arbitraire
Chaque Pi est unitaire en xi et cela donne la structure de l’algèbre de
manière explicite.
Néanmoins, cette algèbre peut contenir des diviseurs de zéro, ce qui
signifie que plusieurs situations différentes sont représentées par un seul
calcul dans AP1 ,...,Pn .
Lorsqu’on pose la question (( Q(x1 , . . . , xn ) = 0 ? )), le programme
doit calculer les coefficients sous-résultants de Pn et Q par rapport à
la variable xn (une discussion (( cas par cas )) s’ensuit).
Une solution est de calculer ces coefficients dans l’algèbre K[x1 , . . . , xn−1 ]
en utilisant l’algorithme des sous-résultants [41, 67, 68] qui nécessite des
divisions exactes et se situe naturellement dans le cadre d’un anneau
intègre, puis de les réduire modulo l’idéal hP1 , . . . , Pn−1 i.
Dès qu’on a trois xi le calcul s’avère très lourd. Une étude de complexité montre qu’on a un bien meilleur contrôle de la taille des objets
manipulés si on fait tous les calculs dans l’algèbre AP1 ,...,Pn−1 .
Malheureusement l’algorithme des sous-résultants ne peut plus s’appliquer. En effet, des divisions requises par l’algorithme peuvent s’avérer
impossibles, et même si l’algèbre est un corps, la division peut demander
un effort disproportionné par rapport aux multiplications.
Aussi semble-t-il que l’algorithme de Berkowitz (ou celui de Chistov),
appliqué à la matrice de Sylvester des polynômes Pn et Q offre la
meilleure solution (en l’état de l’art actuel) pour calculer ces coefficients
sous-résultants.
Il faut noter à cet égard que l’algorithme de Le Verrier-FadeevCsanky etc. (en caractéristique nulle) ou celui proposé par Kaltofen (cf.
section 8.5) en caractéristique arbitraire n’ont des performances supérieures à l’algorithme de Berkowitz que pour le calcul d’un déterminant
isolé, mais non pour le calcul de tous les mineurs principaux dominants
d’une matrice donnée.
Signalons aussi que dans le cas où on utilise l’évaluation dynamique
pour la clôture réelle d’un corps ordonné, certaines discussions (( cas par
cas )) font appel aux signes de tous les coefficients sous-résultants (cf.
[42]).
Une autre application de l’algorithme du calcul du polynôme caractéristique en évaluation dynamique est la détermination de la signature
d’une forme quadratique donnée par une matrice symétrique arbitraire
S. Dans ce cas, la seule connaissance des signes des mineurs principaux
dominants de la matrice S ne suffit pas toujours pour certifier le rang
10.4. Applications des algorithmes
289
et la signature 4 . On pourra consulter à ce sujet le livre [Gan]. Mais il
n’est pas difficile de voir que la connaissance des signes des coefficients
du polynôme caractéristique de la matrice S permet de calculer et de
certifier le rang de S et la signature de la forme quadratique qui lui est
associée.
Cas des matrices creuses
Signalons pour terminer que l’algorithme de Berkowitz et celui de
Chistov sont particulièrement bien adaptés au cas des matrices creuses,
notamment en version séquentielle élémentaire où le nombre d’opérations passe de O(n4 ) à O(n3 ) lorsque seulement O(n) coefficients de
la matrice sont non nuls.
Parmi les autres algorithmes étudiés, celui de Kaltofen-Wiedemann
peut également être adapté au cas des matrices creuses, avec une diminution similaire du nombre d’opérations arithmétiques.
4. Cela suffit dans le cas d’une matrice fortement régulière.
11. Résultats
expérimentaux
11.1
Tableaux récapitulatifs des complexités
Dans cette section, nous donnons les tableaux récapitulatifs des complexités arithmétiques théoriques pour les différents algorithmes étudiés.
Y figure notamment le tableau des complexités algébriques des algorithmes en version séquentielle élémentaire (c’est-à-dire n’utilisant que
la multiplication usuelle des matrices, des polynômes et des entiers) que
nous avons expérimentés.
Abbréviations utilisées
Le mot Cte signifie (( constante asymptotique )) (pour les estimations
de taille des circuits), et Val. signifie (( Domaine de validité )) :
– A.C.A. signifie (( anneau commutatif arbitraire )),
– A.I.A.D. signifie (( anneau intègre possédant un algorithme pour
les divisions exactes )),
– A.I.C. signifie (( anneau intègre et intégralement clos possédant un
algorithme pour les divisions exactes )),
– D. n! signifie (( la division par n! quand elle est possible, est unique
et explicite )).
– Prob. signifie (( algorithme de nature probabiliste )), il s’agit de
l’algorithme de Wiedemann, qui fonctionne sur les corps, avec des
variantes possibles dans le cas A.I.A.D.
Les sigles M.R.P. et M.U.P. désignent respectivement la multiplication rapide et la multiplication usuelle des polynômes.
Rappelons que nous notons µP (n) le nombre d’opérations arithmétiques dans la multiplication de deux polynômes de degré n en profondeur O(log n). En M.U.P. µP (n) = O(n2 ), avec la méthode de Karat-
292
11. Résultats expérimentaux
suba µP (n) = O(nlog 3 ), et en M.R.P. µP (n) = O(n log n log log n) ou
O(n log n) selon les anneaux.
Les initiales G et JB désignent les algorithmes de Gauss (sur un
corps) et de Jordan-Bareiss (sur un anneau intègre possédant un algorithme pour les divisions exactes) pour le calcul des déterminants.
Rappelons que l’algorithme de Jordan-Bareiss, qui consomme un peu
plus d’opérations arithmétiques, présente des avantages significatifs par
rapport à l’algorithme du pivot de Gauss, dans le nombreux anneaux
commutatifs (comme par exemple les anneaux de polynômes à coefficients entiers).
A : Calcul des Déterminants
Méthodes séquentielles simples
Algorithme
Taille
Cte
Val.
Date
Gauss
O(n3 )
2/3
Corps
< 1900
Jordan-Bareiss
O(n3 )
4/3
A.I.A.D.
< 1900
Gauss avec élimination
des divisions
O(n5 )
1/3
A.C.A.
1973
Jordan-Bareiss modifié
O(n5 )
1/10
A.C.A.
1982
Méthodes rapides en profondeur O(n)
Bunch&Hopcroft
O(nα )
γα (*)
Corps
1974
(*) Voir théorème 8.1 et proposition 8.2.1.
Dans le premier tableau ci-après nous avons rajouté la colonne Cr.
pour le traitement des matrices creuses : si une matrice C ∈ An×n a
environ k · n coefficients non nuls, certains algorithmes sont accélérés
et leur temps d’exécution séquentiel divisé par n/k. Nous avons indiqué
cette possibilité d’accélération par un (( oui )) dans la colonne Cr.
11.1. Tableaux récapitulatifs des complexités
293
B : Calcul du Polynôme Caractéristique
Versions séquentielles simples
Algorithme
Taille
Cte
Val.
Cr.
Wiedemann
O(n3 )
2
Prob.
oui
Hessenberg
O(n3 )
2
Corps
Frobenius
O(n3 )
10/3
A.I.C.
O(n3,5 )
2
D. n!
Berkowitz
O(n4 )
1/2
A.C.A.
oui
Chistov
O(n4 )
2/3
A.C.A.
oui
Faddeev-SouriauFrame (Le Verrier)
O(n4 )
2
D. n!
oui
Interpolation
(Lagrange)
O(n4 )
2/3 (G)
Corps
4/3 (JB)
A.I.A.D.
KaltofenWiedemann
O(n4 )
8
A.C.A.
Jordan-Bareiss modifié
O(n5 )
1/10
A.C.A.
Gauss avec élimination des divisions
O(n5 )
1/3
A.C.A.
Preparata & Sarwate
oui
Taille avec multiplication rapide des polynômes
Kaltofen-Wiedemann
√
O(n3 µP (d n e) log n)
Jordan-Bareiss modifié
O(n3 µP (n))
Gauss avec élimination des divisions
O(n3 µP (n))
294
11. Résultats expérimentaux
C : Calcul du Polynôme Caractéristique
Méthodes séquentielles rapides
Taille
Cte
Val.
Keller-Gehrig
O(nα log n)
···
Corps
Interpolation
(de Lagrange)
O(nα+1 )
γα
Corps
Faddeev-SouriauFrame
O(nα+1 )
Cα
D. n!
···
A.C.A.
Algorithme
KaltofenWiedemann
O(n
α+3
2
√
µP (d n e))
D : Calcul du Polynôme Caractéristique
Méthodes parallèles en profondeur O(log2 n)
Algorithme
Taille
Cte
K
Val.
Csanky 1976
O(nα+1 )
4 Cα
Kα
D. n!
Preparata &
Sarwate 1978
O(nα+ 2 )
1
4 Cα
5 Kα
D. n!
Galil & Pan
1989
O(nα+ 2 −δ(α) )
1
···
···
D. n!
B.H.G. (†)
1982
O(n18+ ) (*)
O(n21 )
(**)
···
···
A.C.A.
Berkowitz
1984
O(nα+1+ )
···
···
A.C.A.
Chistov 1985
O(nα+1 log n)
1
α+1
Cα
Berkowitz
amélioré 1985
O(nα+1 log n)
1
α+1
Cα
3
2
(Kα + 1)
3 Kα
A.C.A.
A.C.A.
(†) Borodin, Hopcroft & v. z. Gathen. (*) M.R.P. (**) M.U.P.
La colonne K donne la constante asymptotique du temps parallèle en
O(log2 n). Le nombre δ(α) > 0 dépend de α. Enfin est positif arbitrairement petit.
11.2. Présentation des tests
11.2
295
Présentation des tests
Les algorithmes considérés dans les tableaux de comparaison que
nous présentons ci-dessous ont été expérimentés à l’aide du logiciel de
Calcul Formel Maple et écrits dans le langage de programmation qui
lui est rattaché 1 .
Les algorithmes sont ceux du tableau B, c’est-à-dire les versions séquentielles simples pour le calcul du polynôme caractéristique. Nous
avons indiqué également dans la colonne (( linalpoly )) les performances
de l’algorithme donné par Maple dans la version V. Les versions plus
récentes du logiciel utilisent désormais l’algorithme de Berkowitz.
Chacun des tests de comparaison entre les différents algorithmes a été
effectué sur une même machine, avec le même échantillon de matrices.
Les matrices utilisées font partie de l’un des groupes suivants, selon
le type de l’anneau de base choisi :
• Groupe 1 : les matrices randmatrix(n, n) qui sont des matrices
carrées d’ordre n à coefficients pris au hasard (entre -99 et +99) dans
l’anneau Z des entiers relatifs ;
• Groupe 2 : les matrices M athard(n, x, y) dont les éléments sont des
polynômes en [x, y] de degré total ≤ 5. Les coefficients de ces polynômes de Z[x, y] sont aussi des entiers compris entre -99 et +99 ;
• Groupe 3 : les matrices M atmod(n, lisvar, Ideal, p) qui sont des
matrices carrées d’ordre n dont les coefficients sont des éléments choisis
au hasard dans l’anneau-quotient
Zp [lisvar] /hIdeali
où p est un entier positif (on le prendra premier), lisvar une liste
donnée de variables et Ideal une liste donnée de polynômes en lisvar
à coefficients dans Z. L’anneau de base est donc ici, sauf exception, un
anneau dans lequel la division n’est pas permise.
• Groupe 4 : les matrices Jou(n, x), carrées d’ordre n, à coefficients
dans Z[x], dont les coefficients sont donnés par :
[Jou] ij = x + x2 (x − ij)2 + (x2 + j)(x + i)2
pour
1 ≤ i, j ≤ n .
Quelle que soit la valeur de n, le rang de la matrice Jou(n, x) ne dépasse
pas 3 : c’est ce qui explique la supériorité, dans ce cas, des algorithmes
1. Les programmes ont tourné avec la version Maple V Release 3.
296
11. Résultats expérimentaux
de Souriau-Faddeev et de Jordan-Bareiss (nettement plus performants
pour les matrices de rang petit).
• Groupe 5 : ce sont des matrices creuses à coefficients entiers choisis
au hasard entre -99 et +99. Elles sont données par la procédure Maple
randmatrix(n,n,sparse) .
Quant aux machines utilisées, il s’agit essentiellement d’un DEC Alpha600 à 175 Mhz et 320 Mo de mémoire centrale. 2
Les matrices intervenant dans les comparaisons sont générées par des
codes maple : une procédure Matmod par exemple crée une matrice du
Groupe 3 à partir de la donnée de deux entiers positifs n (la taille de
la matrice) et p (on calcule modulo p), d’une liste de variables lisvar,
et d’une liste Ideal de polynômes en lisvar comprenant autant de
polynômes Pi que de variables xi , chacun des Pi étant un polynôme en [x1 , . . . , xi ], unitaire en xi . Ceci afin d’illustrer le genre d’application de l’algorithme de Berkowitz lorsqu’on se place dans l’algèbre
Zp [lisvar] /hIdeali , et la situation indiquée dans la section 10.4.
La procédure Matmod utilise comme sous-procédure la procédure
polmod (donnée dans l’annexe) qui prend en entrée un nombre entier p,
un polynôme P de Z[lisvar], et donne en sortie un représentant simple
de l’image canonique de P dans l’anneau-quotient Zp [lisvar] /hIdeali .
11.3
Tableaux de Comparaison
Nous donnons dans les trois pages qui suivent les tableaux correspondant aux cinq groupes de matrices que nous avons précédemment
indiqués.
Il ne s’agit que de quelques exemples, mais ils sont significatifs.
La comparaison entre le comportement pratique des algorithmes
montre un bon accord avec les calculs théoriques de complexité, surtout si on prend en compte la taille des objets intermédiaires créés par
les différents algorithmes. Sauf exception l’algorithme de Berkowitz est
le plus performant, suivi de près par celui de Chistov.
Les performances a priori meilleures pour les algorithmes de Hessenberg, Frobenius et Wiedemann ne se révèlent qu’avec des tests portant
sur des matrices à coefficients dans des corps finis. En effet l’avantage en
nombre d’opérations arithmétiques est contrebalancé par la plus mau2. Grâce notamment à l’hospitalité du Laboratoire GAGE (Ecole Polytechnique).
11.3. Tableaux de Comparaison
297
vaise taille des objets intermédiaires manipulés, par exemple dès que
l’anneau des coefficients contient Z. Il aurait fallu créer un autre groupe
de matrices pour mettre en évidence cet avantage.
Il serait également intéressant d’élargir l’expérimentation en implémentant la version séquentielle simple de l’algorithme de Preparata &
Sarwate.
Dans le groupe 1 nous avons pris des matrices carrées d’ordre n à
coefficients dans Z pour 10 valeurs de n comprises entre 16 et 128.
Dans le groupe 2, ce sont des matrices carrées d’ordre n à coefficients
dans Z[x, y] pour n ∈ {10, 12, 20} et des matrices à coefficients dans
Z[x] pour n ∈ {10, 15, 20, 25}.
Parmi les matrices du groupe 3, nous avons prisdes matrices carrées
d’ordre n ∈ {8, 10, 12, 16} à coefficients dans Z7 [x] x3 − 1 (pour lesquelles Faddeev ne s’applique pas) et des matrices à coefficients dans
Z17 [x, y] /hH, Li (pour lesquelles Faddeev s’applique).
Ici hH, Li est l’idéal engendré par les deux polynômes H = x5 − 5xy + 1
et L = y 3 − 2y + 1.
Dans le groupe 4, ce sont des matrices carrées d’ordre n ∈ {10, 15,
20, 25} à coefficients dans Z[x], mais de rang petit ≤ 3.
Enfin les matrices du groupe 5 (des matrices creuses à coefficients entiers choisis au hasard entre -99 et +99) ont été prises parmi les matrices
randmatrix(n,n,sparse) telles que n ∈ {32, 50, 64, 128, 200}.
linalpoly †
0’ 05”
1, 244 Mb
0’ 16”
1, 834 Mb
0’ 40”
1, 900 Mb
2’ 09”
1, 507 Mb
6’ 08”
2, 096 Mb
19’ 42”
2, 424 Mb
1H 17’ 17”
1, 900 Mb
71H 33’ 31”
–
berkosam
0’ 02 ”
3, 014 Mb
0’ 04”
3, 538 Mb
0’ 17”
1, 900 Mb
0’ 30”
1, 965 Mb
1’ 13”
4, 062 Mb
3’ 04”
4, 193 Mb
8’ 47”
4, 324 Mb
8H 17’ 38”
6, 552 Mb
Chistov
0’ 05”
2, 003 Mb
0’ 07”
1, 834 Mb
0’ 17”
1, 900 Mb
0’ 44 ”
1, 965 Mb
1’ 48”
2, 031 Mb
4’ 37”
2, 096 Mb
13’ 23”
2, 162 Mb
11H 22’ 53”
6, 814 Mb
Faddeev
0’ 12”
1, 376 Mb
0’ 28”
2, 096 Mb
1’ 07”
2, 489 Mb
3’ 16”
3, 014 Mb
17’ 58”
5, 241 Mb
2H 38’ 50”
9, 500 Mb
13H 15’ 45”
18, 870 Mb
≈ 170H
OUT∗
barmodif ‡
0’ 09”
1, 244 Mb
0’ 24”
2, 031 Mb
1’ 04”
1, 834 Mb
3’ 25”
2, 817 Mb
11’ 28”
5, 307 Mb
42’ 38”
8, 124 Mb
3h 54’ 46”
14, 284 Mb
≈ 7H
OUT∗
† La procédure Maple linalg[charpoly].
‡ La procédure correspondant à la Méthode de Jordan-Bareiss modifiée.
∗
OUT signifie qu’après un temps ((( CPU Time ))) plus ou moins long de calcul, un message d’erreur
(( Out of memory )) apparaı̂t lorsque la mémoire allouée ((( Mem. All. ))) a dépassé le seuil de 350 Mbytes.
matrice
CPU Time
n = 16
Mem. All.
CPU Time
n = 20
Mem. All.
CPU Time
n = 25
Mem. All.
CPU Time
n = 32
Mem. All.
CPU Time
n = 40
Mem. All.
CPU Time
n = 50
Mem. All.
CPU Time
n = 64
Mem. All.
CPU Time
n = 128
Mem. All.
(Matrices à coefficients entiers)
Premier Groupe : Matrices denses n × n
Hessenberg
0’ 03”
1, 969 Mb
0’ 20”
5, 254 Mb
1’ 28”
14, 468 Mb
≈ 2H
OUT∗
–
–
–
–
–
–
–
–
298
11. Résultats expérimentaux
0’ 42”
3, 341 Mb
≈ 2H 10’
OUT
–
–
–
–
CPU Time
Mem. All.
CPU Time
Mem. All.
CPU Time
Mem. All.
n = 15
&y = 1
n = 20
&y = 1
n = 25
&y = 1
11’ 56”
8, 583 Mb
3’ 26”
5, 110 Mb
berkosam
4’ 04”
8, 059 Mb
15’ 30”
13, 891 Mb
1H 23’ 24”
30, 403 Mb
0’ 05”
2, 031 Mb
linalpoly
48’ 35”
44, 359 Mb
≈ 1H 25’
OUT
–
–
0’ 47”
4, 062 Mb
matrice
CPU Time
n = 10
Mem. All.
CPU Time
n = 12
Mem. All.
CPU Time
n = 15
Mem. All.
CPU Time
n = 10
& y = 1 Mem. All.
Deuxième Groupe : Mathard(n, x, y)
24’ 52”
10, 025 Mb
7’ 14”
4, 979 Mb
1’ 33”
3, 603 Mb
Chistov
13’ 17”
10, 352 Mb
54’ 18”
13, 563 Mb
5H 28’ 28”
26, 733 Mb
0’ 10”
2, 227 Mb
3H 24’ 47”
5, 831 Mb
23H 54’ 55”
9, 107 Mb
≈ 1H 10’
OUT
22’ 06”
3, 800 Mb
barmodif
≈ 17’
OUT
–
–
–
–
1’ 13”
2, 162 Mb
44’ 30”
47, 111 Mb
5’ 33”
14, 480 Mb
Faddeev
53’ 31”
47, 898 Mb
≈ 1H 10’
OUT
–
–
0’ 27”
3, 276 Mb
11.3. Tableaux de Comparaison
299
linalpoly
0’ 14”
2, 162 Mb
2’ 03”
8, 059 Mb
≈ 1H 30’
OUT
berkomod
0’ 02”
1, 703 Mb
0’ 04”
1, 769 Mb
0’ 12”
1, 834 Mb
Chistov
0’ 04”
1, 769 Mb
0’ 07”
1, 769 Mb
0’ 19”
1, 900 Mb
linalpoly
38’ 13”
37, 872 Mb
≈ 2H
OUT
–
–
berkomod
0’ 28”
2, 162 Mb
0’ 50”
2, 293 Mb
2’ 11”
2, 424 Mb
Chistov
0’ 50”
3, 014 Mb
1’ 33”
3, 996 Mb
4’ 25”
7, 142 Mb
† on a : H = x5 − 5xy + 1 et L = y 3 − 3y + 1.
∗ signifie que Faddeev n’est pas applicable dans ce cas (puisque p < n ).
matrice
CPU Time
n = 10
Mem. All.
CPU Time
n = 12
Mem. All.
CPU Time
n = 16
Mem. All.
Pour lisvar = [x, y] ; Ideal = [H, L] † ; et p = 17 :
matrice
CPU Time
n = 10
Mem. All.
CPU Time
n = 12
Mem. All.
CPU Time
n = 16
Mem. All.
Pour lisvar = [x], Ideal = [x3 − 1] et p = 7 :
Troisième Groupe : Matmod(n, lisvar, Ideal, p)
barmodif
1H 06’ 01”
9, 369 Mb
5H 59’ 23”
14, 087 Mb
≈ 84H 40’
OUT
barmodif
0’ 43”
1, 900 Mb
1’ 52”
2, 031 Mb
8’ 31”
2, 424 Mb
Faddeev
1’ 25”
1, 638 Mb
2’ 43”
2, 096 Mb
8’ 03”
3, 276 Mb
*
*
*
Faddeev
300
11. Résultats expérimentaux
linalpoly
0’ 09”
1, 310 Mb
5’ 10”
5, 677 Mb
≈ 3H 30’
OUT
≈ 3H 30’
OUT
berkosam
0’ 07”
2, 489 Mb
1’ 00”
3, 603 Mb
5’ 09”
5, 700 Mb
19’ 01”
8, 714 Mb
Chistov
0’ 14”
2, 555 Mb
2’ 08”
3, 669 Mb
10’ 55”
5, 176 Mb
44’ 21”
9, 238 Mb
Faddeev
0’ 05”
1, 244 Mb
0’ 17”
1, 376 Mb
0’ 43”
1, 834 Mb
1’ 39”
2, 620 Mb
barmodif
0’ 06”
1, 310 Mb
0’ 24”
1, 703 Mb
1’ 07”
2, 293 Mb
2’ 32”
2, 948 Mb
Jou(n, x) est une matrice n × n dont les coefficients Jij (x) ∈ Z[x] sont donnés par la formule :
Jij (x) = x2 (x − ij)2 + (x2 + j)(x + i)2 + x pour 1 ≤ i, j ≤ n .
Son rang est ≤ 3 pour tout x et pour tout entier positif n . On remarque la supériorité des algorithmes
de Faddeev et de Jordan-Bareiss modifié dans ce cas exceptionnel.
†
n = 25
n = 20
n = 15
n = 10
matrice
CPU Time
Mem. All.
CPU Time
Mem. All.
CPU Time
Mem. All.
CPU Time
Mem. All.
(Matrices de rang petit)
Quatrième Groupe : Jou(n, x) †
11.3. Tableaux de Comparaison
301
matrice
CPU Time
n = 32
Mem. All.
CPU Time
n = 50
Mem. All.
CPU Time
n = 64
Mem. All.
CPU Time
n = 128
Mem. All.
CPU Time
n = 200
Mem. All.
linalpoly
0’ 59”
2, 424 Mb
5’ 32”
2, 555 Mb
14’ 46”
2, 620 Mb
4H 29’ 43”
9, 173 Mb
> 16 heures :
Calcul stoppé
(Matrices à coefficients entiers)
berkosam (creux)
0’ 05”
1, 834 Mb
0’ 21”
1, 965 Mb
0’ 42”
2, 031 Mb
6’ 27”
2, 489 Mb
29’ 20”
2, 555 Mb
Cinquième Groupe : Matrices creuses n × n
Chistov (creux)
0’ 35”
1, 769 Mb
3’ 55”
1, 834 Mb
12’ 00”
2, 031 Mb
6H 27’ 19”
3, 603 Mb
Calcul stoppé
Calcul stoppé
302
11. Résultats expérimentaux
12. Le déterminant et les
expressions
arithmétiques
Introduction
Ce chapitre et le suivant donnent quelques aperçus sur le travail de
Valiant (notamment [92, 93, 94]) dans lequel il décrit un analogue algébrique de la conjecture P =
6 N P. Notre exposé doit beaucoup au survey
de von zur Gathen [35] et au livre de Bürgisser [Bur]. Une autre référence
classique est le livre de Bürgisser, Clausen et Shokrollahi [BCS].
Dans la section 12.1 nous discutons différents codages possibles pour
un polynôme sur un anneau. La section 12.2 est consacrée pour l’essentiel
à la méthode de Brent pour la parallélisation des expressions arithmétiques. Dans la section 12.3 nous montrons pourquoi la plupart des polynômes sont difficiles à calculer. Enfin la section 12.4 expose le résultat
de Valiant sur le caractère universel du déterminant.
12.1
Expressions, circuits et descriptions
Nous nous intéressons dans cette section à différentes approches
concernant le codage d’un polynôme arbitraire sur un anneau commutatif A (le codage des éléments de A est supposé fixé).
Une première manière de coder un polynôme est de donner son degré
total, les noms de ses variables et la liste de ses coefficients, dans un ordre
convenu. C’est ce que nous avons appelé la représentation dense des polynômes. Il est raisonnable de penser que pour l’immense majorité des
polynômes il n’y a rien de mieux à faire, et nous donnerons un résultat
dans cette direction (voir le théorème 12.2).
304
12. Le déterminant et les expressions arithmétiques
Certains polynômes très utilisés ont relativement peu de coefficients
non nuls. On peut choisir pour leur codage une représentation creuse,
dans laquelle on donne la liste des couples (coefficient non nul, monôme)
effectivement présents dans le polynôme, chaque monôme étant codé
lui-même par la liste des exposants de chaque variable, écrits en binaire. Par exemple le polynôme αX 64 Y + βXY 33 Z 4 sera codé par
[[a, [1000000, 1, 0]], [b, [1, 100001, 100]]], où a et b désignent des codes
pour α et β.
La taille booléenne d’une représentation creuse ou dense est la longueur du mot qui code le polynôme. La taille peut également être appréciée d’un point de vue purement algébrique, auquel cas chaque constante
et chaque variable a conventionnellement la longueur 1. Le point faible
de la représentation creuse est que le produit d’un petit nombre de polynômes creux est un polynôme dense comme le montre l’exemple classique
suivant :
2
4
2n
(1 + X) × (1 + X ) × (1 + X ) × · · · × (1 + X ) =
2n+1
X−1
X k (12.1)
k=0
Un autre codage naturel est l’utilisation des expressions arithmétiques. Une expression arithmétique est un mot bien formé qui utilise
comme ingrédients de base les éléments de A et les symboles de variables
d’une part, les symboles +, × d’autre part, et enfin les parenthèses
ouvrante et fermante. D’un point de vue un peu plus abstrait, une expression est vue comme un arbre étiqueté. Aux feuilles de l’arbre, il
y a des éléments de A (les constantes) et des symboles de variables,
chaque nœud est étiqueté par + ou ×. En outre deux branches partent
exactement de chaque nœud. La racine de l’arbre représente l’expression
arithmétique.
La taille d’une expression peut être appréciée d’un point de vue
purement algébrique, on prend alors le nombre de nœuds dans l’arbre,
sans compter les feuilles (la taille est alors égale au nombre de feuilles
moins 1). Si on adopte un point de vue proprement informatique, il
faut prendre en compte pour la taille booléenne la longueur de l’écriture
explicite de l’expression dans un langage précis, où les constantes et les
variables ont des codes. Même si on ne travaille qu’avec un nombre fini
de constantes, la taille booléenne de l’expression ne peut être considérée
comme simplement proportionnelle à sa taille algébrique, ceci parce que
l’ensemble des variables n’est pas borné a priori.
12.1. Expressions, circuits et descriptions
a5
X
⊗
a4
⊕
X
⊗
a3
⊕
X
⊗
305
a2
⊕
X
⊗
a1
⊕
X
⊗
a0
⊕
Figure 12.1 – L’arbre de l’expression de Horner
La représentation dense peut naturellement être vue comme une représentation par expressions dans laquelle seules sont autorisées des écritures canoniques. Le nombre
de coefficients d’un polynôme de degré d en
d+n
n variables est égal à n . La représentation par expression arithmétique permet d’exprimer certains polynômes (une petite minorité, mais
ce sont les polynômes les plus utilisés) sous forme plus compacte, et plus
efficace en ce qui concerne leur évaluation. Donnons en trois exemples.
Le premier est celui de la représentation à la Horner d’un polynôme
en une variable. Dans les deux écritures ci-dessous
a5 X 5 + · · · + a1 X + a0 =
(12.2)
a0 + X (a1 + X (a2 + X (a3 + X (a4 + X a5 ))))
l’expression dense réclame pour son évaluation 15 multiplications et l’expression de Horner (dans le second membre)
en réclame seulement 5. En
degré d on obtient respectivement d+1
+ d et 2d opérations arith2
métiques respectivement pour l’expression développée et l’expression de
Horner.
Le deuxième exemple est celui d’un produit itéré. L’expression cidessous, qui est de taille 2n − 1
(X1 + Y1 ) × (X2 + Y2 ) × · · · × (Xn + Yn )
(12.3)
s’écrit comme une somme de 2n monômes, et a une taille de l’ordre de
n 2n en représentation creuse (et plus grande encore en représentation
dense).
306
12. Le déterminant et les expressions arithmétiques
Le troisième exemple, sur lequel nous reviendrons plus en détail est
celui du déterminant d’une matrice carrée dont les entrées sont n2 variables indépendantes. On ne sait pas si cette famille de polynômes peut
être ou non représentée par une famille d’expressions de taille polynomiale, c’est-à-dire dont la taille serait majorée par un C nk ≤ C 2k log n
(avec C et k fixés). On conjecture que c’est faux. Par contre nous verrons que le déterminant peut être représenté par une expression de taille
k
quasi-polynomiale, c’est-à-dire majorée par un C 2(log n) (avec C et k
fixés). Il est clair qu’en représentation dense comme en représentation
creuse, le déterminant a une taille ≥ n! ≥ 2n (pour n ≥ 5) donc asympk
totiquement beaucoup plus grande que C 2(log n) .
Notez par contre que la famille de polynômes de l’exemple (12.1)
occupe une taille exponentielle en représentation par expressions arithn
métiques, à cause du X 2 : le degré d’un polynôme ne peut pas être plus
grand que la taille d’une expression arithmétique qui l’exprime.
Un troisième codage naturel est celui que nous avons retenu pour
l’ensemble de cet ouvrage, le codage par les programmes d’évaluation
arithmétiques ou, ce qui revient au même, par les circuits arithmétiques.
Une expression arithmétique peut être vue comme un cas particulier
de circuit arithmétique. Sa taille en tant qu’expression arithmétique est
la même que celle du circuit arithmétique qui lui correspond, c’est-à-dire
est égale au nombre d’opérations arithmétiques lors de l’exécution du circuit. La représentation creuse peut également être simulée efficacement
par un circuit.
Pour un circuit, les paramètres pertinents sont à la fois la taille et la
profondeur. Un polynôme calculé par un circuit arithmétique de profondeur p a un degré majoré par 2p et on est particulèrement intéressé par
les familles de polynômes (Pn ) qui peuvent être évalués par des familles de circuits dont la profondeur est un O(log(deg(Pn ))). Il semble
cependant très improbable que le déterminant (comme polynôme de
degré n à n2 variables), qui est dans la classe SD(n4 , log2 n) puisse
être réalisé dans une classe SD(nk , log n) (pour un entier k).
Convention 12.1.1 Dans les chapitres 12 et 13 les circuits et les expressions arithmétiques que nous considérerons seront toujours sans division et sans soustraction. Rappelons que l’élimination des divisions à
la Strassen montre qu’il ne s’agit pas d’une restriction importante (surtout dans le cas des corps, voir théorèmes 3.1 et 3.2). La soustraction,
quant à elle, est simulée en deux opérations par x − y = x + (−1) × y.
12.1. Expressions, circuits et descriptions
307
Un dernier codage naturel que nous envisagerons est celui dans lequel
un polynôme P (x1 , . . . , xk ) est obtenu sous la forme
P (x1 , . . . , xk ) =
X
R(x1 , . . . , xk , e1 , . . . , e` )
(12.4)
e1 ,...,e` ∈{0,1}
le polynôme R étant lui-même donné par un circuit ou une expression
arithmétique. Ceci peut sembler a priori artificiel, mais nous verrons
dans le chapitre 13 que cette écriture condensée des polynômes est en
rapport assez étroit avec la conjecture P =
6 N P.
Nous donnons maintenant quelques définitions qui résultent de la
discussion précédente.
Définition 12.1.2 Soit (Pn ) une famille de polynômes (indexée par
n ∈ N ou N` ) à coefficients dans un anneau commutatif A. Notons vn
et dn le nombre de variables et le degré de Pn .
– Nous disons que la famille (Pn ) est p-bornée si vn et dn sont
majorés par un polynôme en n. On dit encore qu’il s’agit d’une
p-famille de polynômes.
– Nous disons qu’une famille (ϕn ) d’expressions arithmétiques est
p-bornée si la taille de ϕn est majorée par un polynôme en n.
– Nous disons que la famille (Pn ) est p-exprimable si elle est réalisable par une famille p-bornée d’expressions arithmétiques ϕn (en
particulier, (Pn ) est p-bornée).
– Nous disons qu’une famille (γn ) de circuits arithmétiques est pbornée en taille si la taille de γn est majorée par un polynôme en
n, p-bornée en degrés si les polynômes évalués à tous les noeuds
de γn sont majorés par un polynôme en n, p-bornée si la famille
est p-bornée en taille et en degrés.
– Nous disons que la famille (Pn ) est p-évaluable (ou encore p-calculable) si elle est réalisable par une famille p-bornée de circuits
arithmétiques (en particulier, (Pn ) est p-bornée).
– Nous disons que la famille (Pn ) est qp-exprimable si c’est une pfamille réalisable par une famille d’expressions arithmétiques dont
la taille est quasi-polynomiale en n (c’est-à-dire majorée par un
k
C 2(log n) avec C et k fixés).
– Nous disons que la famille (Pn ) est qp-évaluable (ou encore qpcalculable) si c’est une p-famille réalisable par une famille de circuits arithmétiques dont la taille est quasi-polynomiale en n.
308
12. Le déterminant et les expressions arithmétiques
– Nous disons qu’un polynôme R en les variables x1 , . . . , xk , y1 , . . . ,
y` est une description du polynôme P en les variables x1 , . . . , xk
si
X
(12.5)
P (x) =
` R(x, y)
y∈{0,1}
– Nous disons que la famille (Pn ) est p-descriptible s’il existe une
famille p-calculable de polynômes (Rn ), telle que chaque Rn est
une description de Pn .
– Nous disons que la famille (Pn ) est p-descriptible en expressions
s’il existe une famille p-exprimable de polynômes (Rn ), telle que
chaque Rn est une description de Pn .
Il faut souligner que toutes les notions introduites ici sont non uniformes, c’est-à-dire qu’on ne demande pas que les familles d’expressions
ou de circuits soient des familles uniformes (cf. section 4.4).
Nous utiliserons les notations suivantes pour décrire les classes (de
familles de polynômes) correspondant aux définitions précédentes. Le V
est mis pour Valiant, qui a établi la plupart des concepts et des résultats
des chapitres 12 et 13.
Notation 12.1.3
– La classe des familles de polynômes p-exprimables est notée VPe ,
celle des familles qp-exprimables VQPe .
– La classe des familles de polynômes p-calculables est notée VP,
celle des familles qp-calculables VQP.
– La classe des familles de polynômes p-descriptibles est notée VN P,
celle des familles p-descriptibles en expressions VN Pe .
– La classe des familles de polynômes évaluables par des familles pbornées de circuits arithmétiques de profondeur O(logk (n)) est
notée VN C k . La réunion des VN C k est notée VN C.
Ces classes sont définies relativement à un anneau commutatif fixé A.
Si on a besoin de préciser l’anneau on notera VPe (A), VP(A), etc. . .La
plupart des résultats sont cependant indépendants de l’anneau. Les conjectures sont énoncées en général pour des corps.
Remarque 12.1.4 Vue la proposition 3.1.6, s’il existe une famille pbornée en taille de circuits arithmétiques qui calcule une p-famille de
polynômes, alors il existe aussi une famille p-bornée de circuits arithmétiques qui calcule la même famille de polynômes. Pour la même raison
nous aurions pu demander, pour définir la classe VQP, que la famille de
12.2. Parallélisation des expressions et des circuits
309
circuits arithmétiques soit non seulement qp-bornée en taille mais aussi
p-bornée en degrés.
12.2
Parallélisation des expressions et des circuits
Parallélisation des expressions
Des expressions comme celles de Horner, qui sont optimales quant
à leur taille (i.e. pour le temps séquentiel d’évaluation), présentent un
défaut de parallélisme criant. Brent a découvert que n’importe quelle
expression arithmétique peut être remplacée par un circuit ou par une
expression dont la profondeur est logarithmique en la taille de l’expression initiale.
Théorème 12.1 (Brent [10]) Pour tout polynôme P la profondeur π
du meilleur circuit et la taille τ de la meilleure expression sont reliés
par
2
log(τ + 1) ≤ π ≤
log(τ + 1)
(12.6)
log 3/2
NB : On a
2
log 3/2
= 3, 4190 . . .. Un calcul plus précis (théorème 21.35
√
dans [BCS]) donne π ≤ log2 φ log(τ ) + 1 où φ est le nombre d’or 1+2 5
et log2 φ = 2, 8808 . . .
Preuve. Dans cette preuve nous notons t(ϕ) le nombre de feuilles de
l’arbre correspondant à l’expression ϕ (c’est la taille de l’expression +1)
et π(γ) la profondeur d’un circuit ou d’une expression γ.
La première inégalité est facile. Si P est une variable ou une constante la profondeur et la taille sont nulles. Sinon lorsque P est évalué
par un circuit γ on a γ = γ1 ◦ γ2 (où ◦ représente + ou ×) et si on
suppose avoir déjà réécrit γ1 et γ2 avec des expressions ϕ1 et ϕ2 on
obtient t(ϕ) = t(ϕ1 ) + t(ϕ2 ) et π(γ) = 1 + max(π(γ1 ), π(γ2 )). En fait ce
calcul correspond à une procédure qui déploie le circuit en une expression de même profondeur. Or l’arbre d’une expression de profondeur p
a au plus 2p feuilles.
La deuxième inégalité est nettement plus subtile. L’idée est la suivante. Appelons x1 , . . . , xm les variables de l’expression ϕ qui représente
P . Nous voyons cette expression comme un arbre. Si on considère un
nœud N de l’arbre, il représente une sous-expression α (voir figure 12.2
page suivante).
310
12. Le déterminant et les expressions arithmétiques
En remplaçant cette sous-expression (ce sous-arbre) par une nouvelle
variable y (c’est-à-dire une feuille), on obtient une expression (un arbre)
β qui représente un polynôme b0 + b1 y avec b0 , b1 ∈ A [x1 , . . . , xm ].
Les polynômes b0 et b1 correspondent à des arbres β0 et β1 qu’il est
facile de construire à partir de l’arbre β.
6
α
5
4
3
2
1
N
⊗
ϕ
⊕
⊗
⊗
⊗
⊕
Figure 12.2 – Parallélisation d’une expression, à la Brent.
En effet (voir figure 12.3 page suivante) pour β0 on substitue 0 à y
dans β, et on simplifie. Pour β1 on part de la racine de β on suit le
chemin jusqu’à y et on supprime les nœuds étiquetés +, (et avec eux,
la branche qui ne va pas à y).
On peut alors construire une expression dans laquelle on met d’abord
en parallèle les expressions α, β0 et β1 et où on termine en calculant
β0 + (β1 × α).
La profondeur π(γ) de cette expression γ est majorée par 2 +
max(π(β0 ), π(β1 ), π(α)).
Pour que cela soit efficace, il faut bien choisir le nœud N (de manière
que les tailles des trois expressions aient baissé dans une proportion
suffisante) et procéder de manière récursive, c’est-à-dire que chacune
des 3 expressions est ensuite soumise de nouveau au même traitement
(et ainsi de suite, cela va sans dire). Le choix du nœud N se fait comme
suit. Soit t0 = t(ϕ). Si t0 ≤ 4 on ne fait rien. Sinon on part de la racine
de l’arbre et on choisit à chaque nœud la branche la plus lourde ϕk .
Si tk = t(ϕk ) on a donc tk+1 ≥ tk /2. On s’arrête la dernière fois que
tk > (1/3)t0 (on aura fait un pas de trop lorsqu’on s’apercevra que le
12.2. Parallélisation des expressions et des circuits
5
4
⊗
β0
⊗
3
2
1
6
4
⊗
⊗
β1
⊕
311
⊗
3
2
⊗
Figure 12.3 – Parallélisation d’une expression, à la Brent ( β0 et β1 ).
seuil a éte franchi, et il faudra retourner un cran en arrière).
β1
α
β0
⊗
⊕
Figure 12.4 – Parallélisation d’une expression, à la Brent, β0 + β1 × α.
On a donc tk > (1/3)t0 ≥ tk+1 ≥ tk /2, on en déduit que tk et t0 − tk
sont tous deux ≤ (2/3)t0 . Et on a t(α) = tk et t(β0 ), t(β1 ) ≤ t0 − tk .
L’inégalité voulue est donc établie par récurrence en vérifiant qu’elle
fonctionne pour les expressions de taille ≤ 3.
t
u
Notez que les procédures décrites sont uniformes.
Remarque 12.2.1 Dans la première procédure, on transforme un circuit en une expression de même profondeur mais de taille peut-être beaucoup plus grande. La seconde procédure transforme toute expression
312
12. Le déterminant et les expressions arithmétiques
(( mal équilibrée )) de taille τ en une expression (( bien équilibrée )) dont
la taille τ 0 n’a pas trop augmenté 1 et dont la profondeur est devenue
logarithmique. Autrement dit la partie difficile du théorème de Brent
fonctionne entièrement au niveau des expressions.
a5
X
⊗
a4
⊕
X
X
X
⊗
⊗
a3
X
⊗
a2
X
⊗
X
a1
X
⊗
⊕
⊗
a0
X
⊕
⊗
⊗
⊕
⊕
Figure 12.5 – L’arbre de Horner, parallélisé à la Brent
Du théorème de Brent, on déduit le corollaire suivant (la version
uniforme serait également valable) :
Corollaire 12.2.2 On a VPe = VN C 1 .
On conjecture que par contre le déterminant n’est pas réalisable par
une expression de taille polynomiale, et donc que VP 6= VPe .
La partie facile du théorème de Brent et l’algorithme parallèle de
Berkovitz montrent également :
Fait 12.2.3 Le déterminant n × n est réalisable par une expression de
2
taille quasi-polynomiale, en 2O(log n) .
Parallélisation des circuits arithmétiques
Rappelons maintenant le théorème 10.1 (section 10.1) de Valiant et
al. [95] (voir aussi [47]). C’est en quelque sorte l’analogue pour les circuits
arithmétiques du théorème de Brent pour les expressions arithmétiques.
Il donne une procédure pour paralléliser n’importe quel circuit arithmétique à condition qu’il calcule un polynôme de degré raisonnable (le
2
1. Le théorème donne (1 + τ 0 ) ≤ (1 + τ ) log(3/2) . En fait, lors d’une étape de
parallélisation on a t(β0 + β1 × α) ≤ (5/3) t(ϕ) et cela conduit plus précisément à
log(5/2)
(1 + τ 0 ) ≤ (1 + τ ) log(3/2) ≤ (1 + τ )2.26 .
12.3. La plupart des polynômes sont difficiles à évaluer
313
n
circuit arithmétique purement séquentiel de taille n qui calcule x2 ne
peut pas être parallélisé, mais c’est à cause de son degré trop élevé).
Théorème 10.1 Soit Γ un circuit arithmétique sans division, de taille
`, qui calcule un polynôme f de degré d en n variables sur un anneau
A. Alors il existe un circuit arithmétique homogène Γ0 de taille O(`3 d6 )
et de profondeur O(log(` d) log d) qui calcule f .
Ce théorème implique immédiatement que VP = VN C 2 . Avec la
partie facile du théorème de Brent il implique aussi qu’une famille qpbornée de circuits arithmétiques peut être parallélisée en une famille de
circuits de profondeur polylogarithmique et donc en une famille d’expressions arithmétiques de taille quasi-polynomiale. En bref :
Corollaire 12.2.4 On a VP = VN C 2 = VN C et VQP = VQPe .
Remarque 12.2.5 1) Ainsi VQP est la classe des familles de polynômes réalisables par des circuits arithmétiques dont le nombre de variables
et les degrés sont p-bornés et la profondeur est polylogarithmique (ce
qui ne signifie pas pour autant qu’ils soient dans VN C). Pour VQPe cela
résultait déjà du théorème de Brent.
2) On conjecture a contrario que les inclusions N C 2 ⊂ N C ⊂ P sont
strictes.
12.3
La plupart des polynômes sont difficiles à
évaluer
Pour établir sous forme précise ce qui est annoncé dans le titre de
cette section nous avons besoin d’un résultat de théorie de l’élimination,
dont la signification est intuitivement évidente. Si vous paramétrez (( un
objet géométrique S )) dans l’espace de dimension 3 en donnant les 3
cooordonnées x, y, z comme fonctions polynomiales de deux paramètres u et v, l’objet que vous obtenez est en général une surface, exceptionnellement une courbe ou plus exceptionnellement encore un point,
mais jamais l’objet géométrique ainsi créé ne remplira l’espace. Plus
précisément, à partir des trois polynômes X(u, v), Y (u, v), Z(u, v) qui
paramètrent l’objet S il est possible de calculer un polynôme Q à trois
variables non identiquement nul tel que Q(X(u, v), Y (u, v), Z(u, v)) soit
identiquement nul. Autrement dit tous les points de S sont sur la surface
algébrique S1 d’équation Q(x, y, z) = 0. Donc, si le corps de base K est
infini (( la plupart )) des points de K3 sont en dehors de S. Précisément,
314
12. Le déterminant et les expressions arithmétiques
considérons un point M en dehors de S1 (puisque le corps est infini,
il en existe sûrement), alors toute droite passant par M ne coupe S1
qu’en un nombre fini de points, majoré par le degré de Q. Si le corps
de base est celui des réels ou celui des complexes, on en déduit que le
complémentaire de S1 est un ouvert dense, ce qui donne encore une
signification intuitive plus claire au terme (( la plupart )) utilisé dans la
phrase ci-dessus.
On peut montrer l’existence du polynôme Q comme suit. Supposons
les degrés de X, Y , Z majorés par d. Pour m fixé, les polynômes
, leur
X m1 Y m2 Z m3 avec m1 + m2 + m3 ≤ m sont au nombre de m+3
3
degré est majoré par dm, donc ils sont dans l’espace des
polynômes
de
degré ≤ dm en 2 variables, qui est de dimension dm+2
.
Pour
m
assez
2
dm+2
grand, m+3
>
,
d’où
une
relation
de
dépendance
linéaire non
3
2
m
m
m
1
2
3
triviale entre les X Y Z , ce qui donne le polynôme Q.
Nous énonçons maintenant le résultat général, qui peut se démontrer de
la même manière.
Proposition 12.3.1 Soit K un corps et (Pi )1≤i≤n une famille de polynômes en m variables y1 , . . . , ym avec m < n. Alors il existe un polynôme non identiquement nul Q(x1 , . . . , xn ) tel que Q(P1 , . . . , Pn ) est
identiquement nul. En termes plus géométriques, l’image d’un espace
Km dans un espace Kn (avec m < n) par une application polynomiale
est toujours contenue dans une hypersurface algébrique.
La proposition précédente a la signification intuitive que, au moins
en géométrie algébrique, ∞n > ∞m lorsque n > m.
Nous en déduisons notre théorème, dans lequel l’expresssion (( la
plupart )) doit être comprise au sens de la discussion qui précédait la
proposition 12.3.1.
Théorème 12.2 Soit K un corps infini, n et d des entiers fixés. L’ensemble des polynômes de degré ≤ d en n variables
x1 , . . . , xn est un
espace vectoriel E(n, d) sur K de dimension n+d
.
Soit
t une constante
d
arbitraire fixée. Notons A(n, d, t) la famille de tous les circuits arithmétiques qui représentent
des polynômes en x1 , . . . , xn de degré ≤ d, avec
n+d
au plus
− 1 constantes aux portes d’entrées, et dont la taille est
d
majorée par t. Alors (( la plupart )) des éléments de E(n, d) ne sont pas
représentés par un circuit arithmétique dans A(n, d, t).
En particulier, pour la plupart des P ∈ E(n, d) la taille τ du meilleur
circuit admet la minoration τ ≥ n+d
d .
12.4. Le caractère universel du déterminant
315
Preuve. Chaque circuit dans A(n, d, t) peut être interprété comme calculant un polynôme P (c1 , . . . , cs ; x1 , . . . , xn ) dans lequel les ci sont les
constantes du circuit. Le polynôme P en s + n variables correspondant fournit (lorsqu’on fait varier les constantes) une application de Ks
vers E(n, d) et chaque coordonnée de cette application est une fonction
polynomiale. Le fait de majorer la taille du circuit par t implique que
les polynômes correspondants (en s + n variables) sont en nombre fini.
Finalement les éléments de E(n, d) représentés par un circuit arithmétique dans A(n, d, t) sont contenus dans une réunion finie d’hypersurfaces
algébriques (d’après la proposition 12.3.1), qui est encore une hypersurface algébrique.
t
u
On trouvera dans [44, 88] des résultats plus précis sur ce sujet.
Remarque 12.3.2 Notez a contrario, que le circuit arithmétique qui
exprime un polynôme de E(n, d) directement comme somme de ses
constantes, et qu’il peut être écrit avec une taille
monômes utilise n+d
d
µ1
n+d
≤ 3 d . Tous les x1 · · · xµnn de degré ≤ d, qui sont au nombre de
n+d
n+d
− n, étapes (un produit de
d
d , peuvent en effet être calculés en
degré k > 1 est calculé en multipliant un produit de degré k − 1, déjà
calculé, par une variable). Il reste ensuite à multiplier chaque produit
par une constante convenable, puis à faire l’addition.
12.4
Le caractère universel du déterminant
Le but de cette section est de montrer que toute expression arithmétique peut être vue comme un cas particulier de l’expression (( déterminant )) dans laquelle les entrées de la matrice ont simplement été
remplacées par une des constantes ou une des variables de l’expression,
avec en outre le fait que le nombre de lignes de la matrice carrée est du
même ordre de grandeur que la taille de l’expression.
Ceci n’est pas très surprenant, au vu de l’exemple classique ci-dessous,
(inspiré de la matrice compagnon d’un polynôme) dans lequel nous
n’avons pas marqué les entrées nulles :
x
a4
−1 x
a3
2
3
4
−1 x
a2
det
= a4 + a3 x + a2 x + a1 x + a0 x .
−1 x a1
−1 a0
316
12. Le déterminant et les expressions arithmétiques
Projections
Nous introduisons maintenant formellement une notion précise appelée projection pour décrire le processus de substitution extrêmement
limité auquel nous allons avoir recours dans la suite.
Définition 12.4.1 Soit A un anneau commutatif fixé. Soient P ∈ A [x1 ,
. . . , xk ] et Q ∈ A [y1 , . . . , y` ]. Soient aussi (Pn ) et (Qm ) des p-familles
de polynômes à coefficients dans A.
(1) On dit que Q est une projection de P si Q est obtenu à partir
de P en substituant à chaque xi un yj ou un élément de A.
(2) On dit que la famille (Qm ) est une p-projection de la famille
(Pn ) s’il existe une fonction polynomialement majorée m 7→ ϕ(m)
telle que, pour chaque m, Qm est une projection de Pϕ(m)
(3) On dit que la famille (Qm ) est une qp-projection de la famille (Pn ) s’il existe une fonction quasi-polynomialement majorée
m 7→ ϕ(m) telle que, pour chaque m, Qm est une projection de
Pϕ(m) .
La proposition suivante est facile.
Proposition 12.4.2
(1) La composée de deux projections est une projection. Même chose
pour les p-projections, ou pour les qp-projections.
(2) Les classes VP et VPe sont stables par p-projection.
(3) La classe VQP = VQPe est stable par qp-projection.
Réécriture d’une expression comme déterminant
Dans le théorème de Valiant qui suit, la difficulté est de produire
une matrice ayant pour déterminant la somme des déterminants de deux
autres matrices. L’idée est de faire cette construction non pas pour n’importe quelles matrices, mais en respectant un certain format. C’est l’objet du lemme crucial qui suit. Le format des matrices qui interviennent
dans ce lemme est visualisé ci-dessous sur un exemple avec d = 4.
a1
1
0
0
0
a2
x1,2
1
0
0
a3
x1,3
x2,3
1
0
a4
x1,4
x2,4
x3,4
1
0
b1
α
b2 =
T
b3
b4
0
β
12.4. Le caractère universel du déterminant
317
(α est un vecteur ligne, β un vecteur colonne et T est carrée unitriangulaire supérieure). Notez que
P lorsqu’on développe le déterminant d’une
telle matrice sous la forme
1≤i,j≤4 ai bj cij , le polynôme cij en facteur
de ai bj est nul si la colonne de ai et la ligne de bj se coupent dans
la partie strictement supérieure de la matrice T , puisque le cofacteur
correspondant de T est nul.
Lemme 12.4.3 Soient (pour i = 1, 2) deux entiers di , deux matrices
carrées Ti ∈ Adi ×di unitriangulaires supérieures, soient deux vecteurs
lignes αi ∈ A1×di et deux vecteurs colonnes βi ∈ Adi ×1 . Considérons
les trois matrices suivantes
α1 α2 0
α1 0
α2 0
M1 =
, M2 =
, M = T1 0 β1 .
T1 β1
T2 β2
0 T2 β2
Alors on a
det M = (−1)d2 det M1 + (−1)d1 det M2
Preuve. Nous donnons seulement l’idée directrice de cette preuve un
peu technique. Lorsqu’on développe complètement le déterminant comme
indiqué avant le lemme, le polynôme en facteur d’un produit α2,i β1,j
est nul car la ligne et la colonne correspondante
secoupent dans la par
T1 0
(cf. le commentaire
tie strictement supérieure de la matrice
0 T2
juste avant le lemme). Pour voir que le polynôme en facteur d’un produit
α1,i β2,j est nul également, il suffit de considérer la matrice
α2
M 0 = T2
0
α1
0
T1
0
β2 .
β1
Son déterminant est identique (en tant qu’expression développée) à celui
de M et l’argument précédent s’applique. Il reste à considérer, dans
le développement complet du déterminant en somme de produits, les
produits contenant un facteur α1,i β1,j (et ceux contenant un facteur
α2,i β2,j ). Un examen attentif montre que les seuls produits non nuls de
ce type sont ceux qui empruntent la diagonale de T2 , donc on retrouve
exactement les facteurs présents dans det M1 au signe près. Ce signe
correspond à une permutation circulaire des d2 + 1 dernières colonnes
de M .
t
u
318
12. Le déterminant et les expressions arithmétiques
Théorème 12.3 Toute expression de taille n est la projection du déterminant d’une matrice d’ordre inférieur ou égal à 2n + 2.
Précisions : la matrice est dans le format décrit au lemme précédent,
ses entrées sont soit une constante de l’expression, soit une variable de
l’expression, soit 0, 1 ou −1, la dernière colonne (resp. la dernière ligne)
ne contient que 0, 1 ou −1, une colonne quelconque contient au plus une
variable ou une constante de l’expression.
Corollaire 12.4.4
(1) Toute famille p-exprimable est une p-projection de la famille
(( déterminant )) (detn est le déterminant d’une matrice carrée d’ordre
n donc un polynôme de degré n en n2 variables).
(2) VQP = VQPe coı̈ncide avec la classe des familles qui sont des
qp-projections de la famille déterminant.
Preuve. On construit la matrice en suivant l’arbre de l’expression. Pour
une feuille f de l’arbre (constante ou variable) on prend la matrice 2 × 2
f
1
0
1
qui répond bien aux spécifications souhaitées. Supposons qu’on a construit les matrices Ai (i = 1, 2) qui ont pour déterminants les polynômes
Pi . Voyons d’abord la matrice pour P1 +P2 . Quitte à changer la dernière
colonne βi en −βi on peut aussi avoir les déterminants opposés, et on
peut donc dans tous les cas appliquer le lemme 12.4.3. Donnons enfin la
matrice N pour P1 × P2
0
0
J =
...
0
N=
A1
J
0
A2
avec
···
···
···
0
0
..
.
0
1
0
.
..
.
0
Cette matrice N répond aux spécifications voulues, et comme elle est
triangulaire par blocs, son déterminant est égal à detA1 · detA2 .
t
u
Donnons par exemple la matrice construite comme indiqué dans la
preuve ci-dessus pour obtenir le déterminant x + (2 + y)z (nous n’avons
12.4. Le caractère universel du déterminant
319
pas mis les 0) :
x
1
.
.
.
.
2 y
. .
1 .
. 1
. .
. .
.
.
−1
−1
1
.
.
.
.
.
z
1
.
1
.
.
.
.
−1
Conclusion
Dans le corollaire 12.4.4 on a vu que toute famille p-exprimable
est une p-projection du déterminant et que toute famille qp-calculable
est une qp-projection du déterminant. Cette dernière propriété s’énonce
sous la forme suivante, qui ressemble à la N P - complétude.
La famille (detn ) est universelle pour VQP et les qp-projections.
Cependant le déterminant lui-même n’est probablement pas p-exprimable et il est p-calculable, donc mieux que qp-calculable.
Il se pose donc la question légitime de trouver une famille p-exprimable qui soit universelle dans VPe par rapport aux p-projections et celle de
trouver une famille p-calculable qui soit universelle dans VP par rapport
aux p-projections (le déterminant serait un candidat naturel, mais pour
le moment on ne connaı̂t pas la réponse à son sujet). Le premier de ces
deux problèmes a été résolu positivement par Fich, von zur Gathen et
Rackoff dans [30]. Le deuxième par Bürgisser dans [12].
La première question admet une réponse assez facile une fois connu
le théorème de Brent. En effet toute expression peut être obtenue comme
projection d’une expression de profondeur comparable extrêmement parallélisée qui combine systématiquement additions et multiplications.
Par exemple l’expression (de profondeur 2) ϕ3 = x1 + (x2 × x3 ) donne
par projection, au choix, l’une des deux expressions (de profondeur 1)
x1 + x2 ou x2 × x3 . Si maintenant on remplace chacun des xi par l’expression xi,1 + (xi,2 × xi,3 ) on obtient une expression ϕ9 à 9 variables
de profondeur 4 et on voit que toute expression de profondeur 2 est une
projection de ϕ9 .
En itérant le processus, toute expression de profondeur n est une projection de l’expression ϕ3n , qui est elle-même de profondeur 2n.
Donc, après parallélisation à la Brent d’une famille dans VPe , la famille parallélisée est clairement une p-projection de la famille (ϕ3n ).
Enfin la famille (ϕ3n ) est elle-même dans VPe car ϕ3n est de taille
320
⊗
12. Le déterminant et les expressions arithmétiques
⊕
⊗
⊕
⊗
⊕
⊗
⊗
⊕
⊗
⊕
⊗
⊕
⊗
⊗
⊕
⊕
⊗
⊕
⊗
⊕
⊕
ϕ27
⊗
⊕
⊗
⊕
Figure 12.6 – Une famille d’expressions p-universelle dans VPe
3n − 1 (on peut, si on a des scrupules, définir ϕk pour tout entier k en
posant ϕk = ϕ3κ où 3κ−1 < k ≤ 3κ ).
La deuxième question (trouver une famille p-calculable qui soit universelle dans VP par rapport aux p-projections) admet une réponse
positive du même style (cf. [12]), mais nettement plus embrouillée.
13. Le permanent et la
conjecture P =
6 NP
Introduction
Ce chapitre est dédié à la conjecture de Valiant. Nous ne démontrerons que les résultats les plus simples et nous souhaitons faire sentir
l’importance des enjeux.
Dans la section 13.1 nous faisons une étude rapide des classes de
complexité booléenne, qui constituent une variante non uniforme de la
complexité binaire.
Dans la section 13.2 nous mettons en évidence quelques liens étroits
et simples entre fonctions booléennes et polynômes, et entre complexité
booléenne et complexité algébrique.
Dans la section 13.3 nous faisons le lien entre complexité binaire
et complexité booléenne. Dans la section 13.4 nous donnons quelques
résultats sur le permanent. Dans la section finale, nous rappelons la
conjecture de Valiant et discutons brièvement sa portée.
Parmi les références utiles pour ce chapitre, il faut citer le livre [Weg]
et l’article [83], non encore signalés.
13.1
Familles d’expressions et de circuits booléens
Expressions, circuits et descriptions
L’analogue booléen de l’anneau de polynômes A [x1 , . . . , xn ] est l’algèbre de Boole
B [x1 , . . . , xn ] ' F2 [x1 , . . . , xn ]
x21 − x1 , . . . , x2n − xn
322
13. Le permanent et la conjecture P =
6 NP
avec dans B [x1 , . . . , xn ] les égalités a ∧ b = ab, a ∨ b = a + b + ab,
¬a = 1 + a et a + b = (¬a ∧ b) ∨ (a ∧ ¬b). Cette interprétation de
l’algèbre de Boole librement engendrée par n éléments comme quotient
d’un anneau de polynômes à n variables sur le corps F2 laisse penser
que les méthodes algébriques sont a priori pertinentes pour résoudre les
problèmes booléens.
L’analogue booléen d’une fonction polynôme à n variables est une
fonction booléenne f : {0, 1}n → {0, 1}. Nous aurons aussi à considérer
des applications booléennes g : {0, 1}n → {0, 1}m .
L’algèbre de Boole B [x1 , . . . , xn ] est isomorphe à l’algèbre des fonctions booléennes f : {0, 1}n → {0, 1}. L’isomorphisme fait correspondre
à l’élément xi de B [x] la i -ème fonction coordonnée : (a1 , . . . , an ) 7→
ai .
Rappelons que si p1 , . . . , pn sont les variables booléennes présentes
dans une expression booléenne, on appelle littéral l’une des expressions
p1 , ¬p1 , . . . , pn , ¬pn . Et qu’une expression est dite en forme normale
conjonctive (resp. en forme normale disjonctive) si elle est une conjonction de disjonctions de littéraux (resp. une disjonction de conjonctions
de littéraux).
Il y a plusieurs types d’écritures canoniques pour une fonction booléenne, en forme normale conjonctive, en forme normale disjonctive ou
sous forme d’un polynôme en représentation creuse (chaque variable intervenant avec un degré ≤ 1 dans chaque monôme). On peut aussi exprimer une fonction booléenne au moyen d’une expression booléenne ou
d’un circuit booléen.
Convention 13.1.1 Nous adopterons la convention qu’une expression
booléenne ou un circuit booléen n’utilisent que les connecteurs ∧, ∨
et ¬. En outre dans le cas d’une expression l’usage du connecteur ¬
sera seulement implicite : on utilisera les littéraux comme variables (aux
feuilles de l’arbre), et nulle part ailleurs n’apparaı̂tra le connecteur ¬.
La taille et la profondeur d’une expression booléenne ne prendront en
compte que les connecteurs ∧ et ∨ (les littéraux sont tous considérés
comme de profondeur nulle).
Cette convention n’a pas de conséquence importante en ce qui concerne
les circuits car autoriser d’autres connecteurs ne ferait diminuer la taille
et/ou la profondeur que d’un facteur constant. Par contre, en ce qui
concerne les expressions booléennes, il s’agit d’une restriction significative de leur pouvoir d’expression : par exemple si on admet en plus le
13.1. Familles d’expressions et de circuits booléens
323
def
connecteur a ⊕ b = (a ∧ ¬b) ∨ (¬a ∧ b) l’expression p1 ⊕ p2 ⊕ · · · ⊕ pn
réclamera très probablement une écriture nettement plus longue sans
l’utilisation de ⊕.
Classes de complexité booléenne
Nous sommes particulièrement interessés ici par les analogues booléens des classes VN C, VP, VPe , VN P et VN Pe (cf. section 12.1).
Définition 13.1.2 Soit fn : {0, 1}vn → {0, 1} une famille de fonctions
booléennes (indexée par n ∈ N ou N` ).
– Nous disons que la famille (fn ) est p-bornée si vn est majoré
par un polynôme en n. On dit encore qu’il s’agit d’une p-famille
de fonctions booléennes.
– Nous disons qu’une famille d’expressions booléennes (ϕn ) est pbornée si la taille de ϕn est majorée par un polynôme en n.
– Nous disons que la famille (fn ) est p-exprimable si elle est réalisable par une famille p-bornée d’expressions booléennes. La classe
des familles de fonctions booléennes p-exprimables est notée BPe .
– Nous disons qu’une famille de circuits booléens (γn ) est p-bornée
si la taille de γn est majorée par un polynôme en n.
– Nous disons que la famille (fn ) est p-évaluable (ou encore p-calculable) si elle est réalisable par une famille p-bornée de circuits
booléens. La classe des familles de fonctions booléennes p-calculables est notée BP = P/poly.
– Nous notons BN C k = N C k /poly la classe des familles de fonctions
booléennes réalisables par une famille de circuits booléens de taille
polynomiale et de profondeur en O(logk n), et BN C = N C/poly
dénote la réunion des BN C k .
– Nous disons qu’une fonction booléenne g en les variables p1 , . . . ,
pk , r1 , . . . , r` est une description de la fonction booléenne f en
les variables p1 , . . . , pk si
f (p) =
_
r∈{0,1}`
g(p, r)
(13.1)
– Nous disons que la famille (fn ) est p-descriptible s’il existe une
famille p-calculable de fonctions booléennes (gn ), telle que chaque
gn est une description de fn . La classe des familles de fonctions
booléennes p-descriptibles est notée BN P = N P/poly.
324
13. Le permanent et la conjecture P =
6 NP
– Nous disons que la famille (fn ) est p-descriptible en expressions
s’il existe une famille p-exprimable de fonctions booléennes (gn ),
telle que chaque gn est une description de fn . La classe des familles de fonctions booléennes p-descriptibles en expressions est
notée BN Pe .
Il faut souligner que toutes les notions introduites ici sont non uniformes, comme dans le cas algébrique.
La classe P/poly est clairement l’analogue booléen de la classe VP en
complexité algébrique. C’est aussi un analogue non uniforme de la classe
P. Ce dernier point sera plus clair après le théorème 13.7 page 337. De
même nous verrons que la classe N P/poly est un analogue non uniforme
de la classe N P.
Si on compare les définitions des descriptions dans le cas algébrique
et dans le cas booléen, on voit qu’on utilise maintenant une disjonction
à la place d’une somme (formules 12.5 et 13.1).
La notation P/poly (voir par exemple [BDG] ou [Weg]) s’explique
comme suit : une famille (fn ) dans P/poly peut être calculée en temps
polynomial si on a droit à (( une aide )) (sous forme d’une famille de circuits booléens γn qui calculent les fonctions fn ) qui n’est peut-être pas
uniforme mais qui est de taille polynomiale en n.
Signalons que Karp et Lipton, qui introduisent la classe P/poly dans
[56] donnent une définition générale pour une variante non uniforme
C/poly en complexité booléenne d’une classe de complexité binaire arbitraire C. Leur définition justifie aussi les égalités BN P = N P/poly et
BN C k = N C k /poly. Enfin la définition de Karp et Lipton ne semble rien
donner pour Pe /poly par absence de la classe Pe en complexité binaire.
La complexité booléenne des opérations arithmétiques dans Z
Le livre [Weg] de Wegener contient une étude précise et très complète
de la complexité des familles de fonctions booléennes. On y trouve notamment les résultats donnés dans le théorème qui suit concernant la
complexité booléenne des opérations arithmétiques dans N. En fait les
résultats sont uniformes et ils s’étendent immédiatement à Z.
Théorème 13.1 (théorèmes 1.3, 2.4 et 2.8 du chapitre 3 de [Weg])
L’addition et la multiplication dans N sont réalisables par des familles
de circuits booléens dans BN C 1 . Plus précisément :
1. L’addition de deux entiers de taille n ≥ 3 est réalisable par un
circuit booléen de taille ≤ 9 n et de profondeur 2 dlog ne + 8.
13.1. Familles d’expressions et de circuits booléens
325
2. Le produit de deux entiers de taille n est réalisable à la Karatsuba par un circuit booléen de taille O(nlog 3 ) et de profondeur
O(log n), ou même, en suivant Schönage et Strassen qui adaptent
la transformation de Fourier discrète rapide des polynômes au cas
des entiers, par un circuit booléen de taille O(n log n log log n) et
de profondeur O(log n).
Concernant la multiplication des entiers on lira aussi avec intérêt
l’exposé de Knuth dans [Knu].
Parallélisation des expressions booléennes
Nous avons pour les expressions booléennes un résultat analogue à
la parallélisation à la Brent des expressions arithmétiques (voir [85] ou
[Sav] théorème 2.3.3).
Théorème 13.2 Pour toute fonction booléenne f : {0, 1}n → {0, 1} la
profondeur π du meilleur circuit booléen et la taille τ de la meilleure
expression booléenne sont reliés par
log(τ + 1) ≤ π ≤
2
log(τ + 1)
log 3/2
Preuve. Cela marche de la même manière que la parallélisation à la
Brent des expressions arithmétiques. Le polynôme b0 y + b1 dans le cas
algébrique (cf. la preuve du théorème 12.1, page 310) doit être remplacé
par une expression (y ∧ b0 ) ∨ (¬y ∧ b1 ) dans le cas booléen.
t
u
L’analogue booléen du corollaire 12.2.2 (VPe = VN C 1 ) est :
Corollaire 13.1.3 On a BPe = BN C 1 .
Description des circuits booléens par des expressions booléennes
Le lemme suivant est facile et utile.
Lemme 13.1.4 Étant donné un circuit booléen γ de taille τ avec les
portes d’entrée p1 , . . . , pn , les portes internes r1 , . . . , r` et une seule
porte de sortie (donc τ = ` + 1), on peut construire une expression
booléenne en forme normale conjonctive ϕ(p, r) de taille ≤ 7 τ et de
profondeur ≤ 2 + log(3 τ ) telle que, pour tous p ∈ {0, 1}n on ait
l’équivalence :
_
γ(p) = 1
⇐⇒
(13.2)
` ϕ(p, r) = 1 .
r∈{0,1}
326
13. Le permanent et la conjecture P =
6 NP
En outre dans le second membre il y a une seule affectation des booléens
r1 , . . . , r` qui rend l’expression vraie (lorsque γ(p) = 1).
Preuve. On remplace chaque affectation du programme d’évaluation
défini par le circuit booléen par une expression booléenne qui est vraie
si et seulement si la valeur du booléen affecté est correcte. La conjonction
de toutes ces expressions booléennes donne l’expression ϕ. Une affectation c := ¬a est traduite par (c∨a)∧(¬c∨¬a). Une affectation c := a∨b
est traduite par (¬c∨a∨b)∧(c∨¬a)∧(c∨¬b). Une affectation c := a∧b
est traduite par (c ∨ ¬a ∨ ¬b) ∧ (¬c ∨ a) ∧ (¬c ∨ b).
t
u
On en déduit immédiatement.
Proposition 13.1.5 On a l’inclusion BP ⊂ BN Pe et l’égalité BN P =
BN Pe .
Signalons aussi le résultat important de Valiant (pour une preuve
voir [Bur]).
Théorème 13.3 On a pour tout corps VN P = VN Pe .
Expressions, circuits et descriptions : le cas des applications
booléennes
Nous pouvons reprendre avec les familles d’applications booléennes
les définitions données au début de cette section pour les familles de
fonctions booléennes. Notre objectif est surtout ici de définir l’analogue
non uniforme de la classe #P.
Définition 13.1.6 Soit fn : {0, 1}vn −→ {0, 1}wn une famille d’applications booléennes (indexée par n ∈ N ou N` ). Soit (fn,k ) la famille
double de fonctions booléennes fn,k : {0, 1}vn → {0, 1} qui donne la k ème coordonnée de fn si k ≤ wn .
– Nous disons que la famille (fn ) est p-bornée si vn et wn sont
majorés par un polynôme en n. On dit encore qu’il s’agit d’une
p-famille d’applications booléennes.
– Nous disons que la famille (fn ) est p-exprimable si elle est p-bornée et si la famille double (fn,k ) correspondante est p-exprimable.
Définition analogue pour une famille p-calculable, p-descriptible,
ou p-descriptible en expressions.
13.1. Familles d’expressions et de circuits booléens
327
– Nous disons qu’une famille gn : {0, 1}vn −→ N est dans la classe
#BP = #P/poly, ou encore qu’ elle compte les solutions d’une
famille p-calculable de fonctions booléennes si elle vérifie :
n
o
∀p gn (p) = # q ∈ {0, 1}`n | hn (p, q) = 1
(13.3)
où hn : {0, 1}vn +`n → {0, 1} est une famille p-calculable de fonctions booléennes. Si la famille (hn ) est p-exprimable on dira que
la famille (gP
n ) est dans la classe #BPe .
n
k−1 on dira que la famille (f ) est dans
Si gn (p) = w
n
k=1 fn,k (p) 2
#BP (resp. dans #BPe ) lorsque la famille (gn ) est dans #BP
(resp. dans #BPe ).
Une conséquence immédiate de la description des circuits booléens
par les expressions booléennes (lemme 13.1.4) est la proposition suivante,
analogue à la proposition 13.1.5.
Proposition 13.1.7
On a l’égalité #BP = #BPe .
Remarque 13.1.8 Il n’y a pas de différence de principe entre une famille d’applications booléennes et une famille de fonctions booléennes,
puisque donner une famille d’applications booléennes revient à donner
une famille double de fonctions booléennes. Si on veut définir directement la classe #BP = #P/poly comme une classe de fonctions booléennes, on pourra dire que le problème dans #P/poly associé à la famille
(hn ) ∈ P/poly est le problème suivant portant sur le couple (p, m) (où
m est codé en binaire) :
n
o
# q ∈ {0, 1}`n | hn (p, q) = 1 ≤ m ?
La plupart des fonctions booléennes sont difficiles à évaluer
On a aussi l’analogue suivant du théorème 12.2 page 314 : ici on
trouve qu’une famille de circuits booléens de taille quasi-polynomiale ne
peut calculer qune infime partie de toutes les fonctions booléennes.
Proposition 13.1.9 Soit VQPB (k) l’ensemble des familles (fn ) de
fonctions booléennes à n variables réalisables par une famille de cirk
cuits booléens de taille 2 log n . Soit ε > 0. Pour n assez grand seulement une proportion < ε de fonctions booléennes à n variables est dans
VQPB (k).
328
13. Le permanent et la conjecture P =
6 NP
Preuve. Faisons les comptes. Le nombre total de fonctions booléennes
n
à n variables est égal à 22 . Le nombre total de circuits booléens à n
variables et de taille t+1 est majoré par N (t+1) = 2 N (t) (t+2n)2 : en
effet un programme d’évaluation de taille t + 1 est obtenu en rajoutant
une instruction à un programme d’évaluation de taille t, instruction de
la forme xt+1 ← y ◦z avec ∧ ou ∨ pour ◦ , et y, z sont à choisir parmi
les littéraux ou parmi les xi (1 ≤ i ≤ t). Cette majoration conduit à
N (1) = 2 (2n)2 , N (2) = 22 (2n)2 (2n + 1)2 , . . ., N (t) < 2t ((2n + t)!)2 =
k
k+1
2O((n+t) log(n+t)) . Donc si t = 2 log n , log N (t) = O(2log n ) qui devient
négligeable devant 2n pour n grand.
t
u
On trouvera des résultats du même style mais nettement plus précis
dans le chapitre 4 du livre de Wegener [Weg].
13.2
Booléen versus algébrique (non uniforme)
13.2.1
Évaluation booléenne des circuits arithmétiques
Rappelons ici le problème, déjà évoqué à la section 4.3.2, de l’évaluation d’un circuit arithmétique sur un anneau A dont les éléments
sont codés en binaire. Si l’anneau A est fini, le temps parallèle ou séquentiel du calcul booléen correspondant à l’éxécution d’un circuit arithmétique est simplement proportionnel à la profondeur ou à la taille du
circuit arithmétique. Par ailleurs rappelons que VP = VN C = VN C 2 ,
VN C 1 = VPe et BN C 1 = BPe . On obtient donc :
Lemme 13.2.1 Si une p-famille de polynômes sur un anneau fini A
est dans la classe VP (resp. VPe , VN P) son évaluation booléenne est
donnée par une famille dans la classe booléenne BN C 2 ⊂ BP (resp. BPe ,
#BP).
Dans le cas d’un anneau infini, l’évaluation booléenne d’un circuit
arithmétique peut réserver quelques mauvaises surprises (voir l’exemple
de l’inventeur du jeu d’échec page 144). Il faudrait bannir toute constante
(même 1 !) d’un circuit arithmétique sur Z si on veut que l’évaluation booléenne (avec le codage naturel binaire de Z) ne produise pas
d’explosion (et Z est l’anneau infini le plus simple).
Une solution serait de coder les éléments de l’anneau par des circuits
arithmétiques n’ayant que des constantes en entrées 1 . Mais le test d’éga1. Du point de vue des calculs en temps polynomial on peut remarquer que le
codage binaire usuel de Z est équivalent à un codage par des expressions arithméti-
13.2. Booléen versus algébrique
329
lité, le test de signe et bien d’autres opérations simples sur Z semblent
alors sortir de la classe P.
Une autre solution serait d’apporter une restriction plus sévère aux
familles p-bornées de circuits arithmétiques. Avant d’y introduire la
moindre constante, même 1, la famille devrait être p-bornée (en taille
et en degrés). Ensuite seulement on remplacerait certaines variables par
des constantes.
De manière générale il faut avoir une majoration convenable de la
taille des objets à calculer.
Définition 13.2.2 Une famille de fonctions fn : Zvn → Z est dite pbornée en taille si vn est majoré par un polynôme en n et la taille
de fn (x1 , . . . , xvn ) est majorée par un polynôme en la taille de l’entrée
x1 , . . . , xvn (en utilisant les codages binaires usuels).
On a alors l’extension importante suivante du lemme 13.2.1 à l’anneau Z, sous une condition restrictive supplémentaire, qui est d’ailleurs
inévitable.
Lemme 13.2.3 On considère une p-famille (Pn ) de polynômes sur Z.
On suppose que la famille de fonctions fn : Zvn → Z définie par (Pn )
est p-bornée en taille. Alors si (Pn ) est dans VP(Z) = VN C 2 (Z) (resp.
VPe (Z) = VN C 1 (Z), VN P(Z)) son évaluation booléenne est donnée par
une famille de circuits booléens dans BN C 3 ⊂ BP (resp. BN C 2 , #BP).
Preuve. Supposons que (Pn ) est dans la classe VP(Z) et soit (Γn ) une
famille p-bornée de circuits arithmétiques correspondant à (Pn ). Pour
tous m, n entiers positifs on veut construire un circuit booléen γn,m qui
calcule (le code de) fn (x1 , . . . , xvn ) à partir des (codes des) xi lorsqu’ils
sont de taille ≤ m. On sait que la taille de la sortie y est majorée par
un entier p ≤ C (n + m)k . Il suffit alors de prendre les constantes de Γn
modulo 22p et d’exécuter les calculs indiqués par le circuit Γn modulo
22p pour récupérer y comme élément de Z à la fin du calcul. La taille du
circuit booléen γn,m correspondant est bien polynomialement majorée.
Quant à sa profondeur, comparée à celle de Γn , elle a été multipliée par
un O(log p) = O(log(m + n)) (cf. le théorème 13.1)
Le résultat pour (Pn ) dans la classe VN P(Z) se déduit immédiatement
du résultat pour (Pn ) dans la classe VP(Z).
t
u
ques n’ayant que les constantes 0, 1, ou −1 aux feuilles de l’arbre. Il n’est donc pas
artificiel de proposer un codage de Z par des circuits arithmétiques n’ayant que les
constantes 0, 1, ou −1 aux portes d’entrée, ce que nous avions noté Zpreval .
13. Le permanent et la conjecture P =
6 NP
330
Notez que si (Pn ) est dans VN C 1 , l’hypothèse que (fn ) est p-bornée en taille est automatiquement vérifiée si les constantes du circuit
Γn ont une taille majorée par un C n` . Dans la section suivante, tous
les circuits arithmétiques qui simulent des circuits booléens utilisent les
seules constantes 0, 1 et −1.
13.2.2
Simulation algébrique des circuits et expressions
booléennes
Nous nous intéressons dans cette section à la possibilité de simuler
algébriquement une fonction booléenne f : {0, 1}n → {0, 1}, ou une
application g : {0, 1}n → Z (par exemple codée par un circuit booléen
{0, 1}n → {0, 1}m ).
Nous disons que le polynôme P simule la fonction booléenne f en
évaluation s’il a le même nombre de variables, et s’il s’évalue de la même
manière que la fonction booléenne sur des entrées dans {0, 1}.
Définition analogue pour la simulation algébrique de l’application g
par un polynôme (l’anneau doit contenir Z).
Un résultat élémentaire
Le lemme suivant nous dit ce que donne la simulation naturelle d’un
circuit booléen par un circuit arithmétique : la profondeur et la taille
sont convenables mais les degrés peuvent réserver de mauvaises surprises.
Lemme 13.2.4 Un circuit booléen γ de taille τ et de profondeur π
peut être simulé en évaluation par un circuit arithmétique ψ de taille
≤ 4τ et de profondeur ≤ 3π (sa profondeur multiplicative reste égale à
π donc le degré des polynômes est ≤ 2π ). Cette simulation fonctionne
sur tout anneau commutatif (non trivial).
Preuve. Les seules valeurs des booléens sont 0 et 1, on a donc
x ∧ y = xy,
¬x = 1 − x,
x ∨ y = x + y − xy
sur n’importe quel anneau commutatif (non trivial) 2 .
t
u
2. Nous rappelons que dans les chapitres 12 et 13 les seules opérations arithmétiques autorisées sont + et × ce qui nous contraint à introduire des multiplications par
la constante −1 pour faire des soustractions. Ceci implique que le polynôme x+y −xy
est évalué par un circuit de profondeur égale à 3.
13.2. Booléen versus algébrique
331
Simulation d’une expression booléenne par une expression arithmétique
Le lemme suivant est une conséquence directe du théorème de parallélisation 13.2 et du lemme 13.2.4.
Lemme 13.2.5 Une expression booléenne ϕ de taille τ peut être simulée en évaluation par une expression arithmétique de profondeur majorée
par log63/2 log(τ + 1) ≤ 10, 26 log(τ + 1). Cette simulation fonctionne
sur tout anneau commutatif (non trivial).
En particulier la taille de l’expression arithmétique est ≤ (τ + 1)10,26
( 3 ). On en déduit :
Proposition 13.2.6 Toute famille dans BPe est simulée algébriquement par une famille dans VPe . Cette simulation fonctionne sur tout
anneau commutatif (non trivial).
Dans [Bur] la proposition précédente est énoncée avec une terminologie différente : (( BPe est contenu dans la partie booléenne de VPe )).
Une proposition analogue à la précédente et qui voudrait relier de
manière aussi simple les classes BP et VP échouerait parce que la traduction naturelle d’un circuit booléen en un circuit arithmétique donnée au
lemme 13.2.4 fournit en général un polynôme de degré trop grand. Autrement dit, on ne connaı̂t pas d’analogue satisfaisant du lemme 13.2.5
pour les circuits booléens.
Supposons maintenant que nous ayons démarré avec une p-famille double d’expressions booléennes (ϕn,k ) associée à une famille de
fonctions fn : {0, 1}vn → Z. La sortie est codée par exemple comme
suit dans {0, 1}m , le premier bit code le signe, et les bits suivants
codent l’entier sans signe en binaire (supposé < 2m−1 ). Par exemple avec
m = 8 les entiers 5, −11 et 69 sont respectivement codés par 00000101,
10001011 et 01000101. Il n’y a alors aucune difficulté à calculer par un
circuit arithmétique ou par une expression arithmétique de profondeur
O(log m) la sortie dans Z à partir de son code.
Nous pouvons alors énoncer la proposition suivante, qui généralise la
proposition 13.2.6, et qui résulte également du lemme 13.2.5.
Proposition 13.2.7 Soit (ϕn,k )1≤k≤a+nh une p-famille double dans
BPe qui code une famille de fonctions fn : {0, 1}vn → Z. Alors il existe
3. Le degré du polynôme est majoré par (τ + 1)3,419 .
13. Le permanent et la conjecture P =
6 NP
332
une p-famille (γn ) d’expressions arithmétiques dans VPe qui simule en
évaluation la famille (fn ) sur n’importe quel anneau contenant Z.
Description d’un circuit booléen par une expression arithmétique
Nous pouvons faire une synthèse des lemmes 13.1.4 et 13.2.4 pour
obtenir une description algébrique (au sens de la définition 12.1.2) d’un
circuit booléen.
Lemme 13.2.8 Soit γ un circuit booléen de taille τ = `+1 qui calcule
une fonction booléenne f : {0, 1}n → {0, 1}. Il existe une expression
arithmétique ψ(x1 , . . . , xn , y1 , . . . , y` ) de taille ≤ 14 τ et de profondeur
≤ 4 + dlog(3 τ )e vérifiant :
∀p ∈ {0, 1}n
f (p) =
X
r∈{0,1}`
ψ(p, r)
(13.4)
Cette expression arithmétique utilise les seules constantes 0, 1 et −1 et
l’égalité est valable sur tout anneau commutatif (non trivial).
Preuve. On applique la simulation donnée dans le lemme 13.2.4 à l’expression booléenne ϕ en forme normale conjonctive construite au lemme
13.1.4. On doit simuler algébriquement chacune des expressions booléennes de base qui sont du type (¬c ∨ a ∨ b) ∧ (c ∨ ¬a) ∧ (c ∨ ¬b) ou du
type (c ∨ ¬a ∨ ¬b) ∧ (¬c ∨ a) ∧ (¬c ∨ b). Dans ces expressions booléennes c est un littéral positif et a, b des littéraux positifs ou négatifs.
L’examen précis montre que la taille maximum pour une telle simulation
est 11. Il reste ensuite à faire le produit de 3τ expressions (chacune
correspond à l’un des composants dans les deux types ci-dessus). On
obtient alors une expression arithmétique ψ de taille ≤ 14 τ et de profondeur ≤ 4 + dlog(3 τ )e vérifiant :
_
n
∀p ∈ {0, 1}
f (p) = 1 ⇐⇒
ψ(p, r) = 1
(13.5)
`
r∈{0,1}
En outre dans le second membre il y a une seule affectation des variables
r1 , . . . , r` dans {0, 1}` qui rend l’expression ϕ vraie (lorsque f (p) = 1),
`
c’est-à-dire que ψ(p, r) est nulle pour tout
P r ∈ {0, 1} à l’exception de
cette valeur. D’où l’égalité 13.4 : f (p) = r∈{0,1}` ψ(p, r).
t
u
Nous en déduisons les corollaires suivants.
13.2. Booléen versus algébrique
333
Proposition 13.2.9 Toute famille fn : {0, 1}vn → {0, 1} dans BP est
simulée en évaluation par une famille dans VN Pe , ceci sur tout anneau
commutatif (non trivial). Toute famille gn : {0, 1}vn → N dans #BP,
est simulée en évaluation par une famille dans VN Pe sur tout anneau
commutatif contenant Z.
Théorème 13.4 Si VP(Z) = VN Pe (Z) alors BN C 3 = BP = BN P =
#BP.
Preuve. Supposons VP(Z) = VN Pe (Z) et soit (fn ) une famille dans
#BP. Remarquons que (fn ) est p-bornée en taille. Par la proposition
précédente, cette famille est simulée en évaluation par une famille dans
VN Pe (Z) donc par une famille dans VP(Z) = VN C 2 (Z). Or une telle
famille s’évalue par une famille dans BN C 3 d’après le lemme 13.2.3. t
u
En fait, en utilisant des techniques nettement plus subtiles, Bürgisser
a montré les résultats suivants (cf. [Bur]).
Théorème 13.5
1. Soit Fq un corps fini, si VP(Fq ) = VN Pe (Fq ) alors BN C 2 = BP =
BN P.
2. Soit K un corps de caractéristique nulle. Supposons que l’hypothèse de Riemann généralisée est vraie. Si VP(K) = VN Pe (K)
alors BN C 3 = BP = BN P.
13.2.3
Formes algébriques déployées
Forme algébrique déployée d’une fonction booléenne
Pour traiter les questions de taille d’expressions ou de circuits booléens il est a priori prometteur d’interpréter une fonction booléenne par
un polynôme algébrique usuel. Une traduction particulièrement simple
consiste à étaler certaines valeurs de la fonction booléenne : on remplace
la fonction booléenne à m + n variables f (p1 , . . . , pm , q1 , . . . , qn ) par le
polynôme suivant, en m variables, avec pour seuls exposants 0 ou 1
dans les monômes
X
µ
F (p1 , . . . , pm ) =
(13.6)
n f (p1 , . . . , pm , µ) x
µ∈{0,1}
où µ = µ1 , . . . , µn et xµ = xµ1 1 · · · xµnn .
Nous dirons que le polynôme F est la forme algébrique déployée
(sur les variables q1 , . . . , qn ) de la fonction booléenne f . Lorsque m 6= 0
13. Le permanent et la conjecture P =
6 NP
334
chaque coefficient de F est une fonction booléenne de p1 , . . . , pm qui
doit être simulée algébriquement. Lorsque m = 0 on a une forme algébrique déployée pure et les coefficients de F sont tous égaux à 0 ou 1.
Une définition analogue est également valable si on remplace f :
{0, 1}m+n → {0, 1} par une application g : {0, 1}m+n → Z.
Si la fonction booléenne f est facile à calculer, le polynôme correspondant F aura ses coefficients faciles à évaluer, mais il risque d’être
difficile à évaluer, puisqu’il y aura en général un nombre trop grand
(exponentiel en n) de coefficients non nuls.
On a alors comme conséquence des résultats précédents.
Lemme 13.2.10 Soit une fonction booléenne f (p1 , . . . , pm , q1 , . . . , qn )
évaluée par un circuit booléen γ de taille τ . Sa forme algébrique déployée F sur les variables q1 , . . . , qn admet une description (au sens
de la définition 12.1.2) par une expression arithmétique de profondeur
≤ 5+dlog(3τ )e et de taille ≤ 14τ +4n ≤ 18τ. Cette expression arithmétique utilise les seuls constantes 0, 1 et −1 et est valable sur tout anneau
commutatif (non trivial).
Preuve. Cela résulte du lemme 13.2.8 et de la constatation suivante.
On a pour µ1 , . . . , µn ∈ {0, 1}
Yn
(µi (xi − 1) + 1)
xµ1 1 · · · xµnn =
i=1
qui s’écrit comme une expression de profondeur ≤ 3 + dlog ne < 4 +
dlog(3τ )e et de taille 4n − 1. Donc si la fonction booléenne f est décrite
par l’expression arithmétique ψ(p, q, r) (lemme 13.2.8), le polynôme F
est égal à
X
Yn
ψ(p,
µ,
r)
·
(µi xi + 1 − µi )
n+`
(µ,r)∈{0,1}
i=1
et il admet pour description l’expression arithmétique à m + 2n + `
variables
Yn
(µi xi + 1 − µi )
θ(p, x, µ, r) = ψ(p, µ, r) ·
i=1
de profondeur ≤ 5 + dlog(3τ )e et de taille ≤ 14τ + 4n.
t
u
Forme algébrique déployée d’une famille de fonctions booléennes
Une famille de fonctions booléennes fn : {0, 1}vn +wn → {0, 1} admet
pour forme algébrique déployée (sur les wn dernières variables) la famille des polynômes Fn qui sont les formes algébriques déployées des
13.3. Complexité binaire versus complexité booléenne
335
fonctions fn . Même chose pour la forme algébrique déployée d’une famille gn : {0, 1}vn +wn → N.
On a comme corollaire du lemme 13.2.10.
Théorème 13.6 (Critère de Valiant) Toute famille de fonctions booléennes dans BP admet pour forme algébrique déployée une famille de
polynômes dans VN Pe , qui convient pour tout anneau commutatif (non
trivial). En conséquence une famille dans #BP admet pour forme algébrique déployée une famille dans VN Pe (Z), et cette famille convient
pour tout anneau contenant Z.
Dans le cas d’une fonction booléenne cela peut sembler un peu décevant, puisqu’a priori VN Pe est une classe réputée difficile à calculer
(elle simule #BPe = #BP), mais il y a une très bonne raison à cela.
En effet, supposons qu’on déploie toutes les variables, alors si on calcule Fn (1, . . . , 1) on trouve le nombre
total des solutions de l’équation
P
fn (p) = 1, c’est-à-dire la somme
p fn (p). Et ce n’est donc pas surprenant que Fn soit a priori plus difficile à calculer que ses coefficients. De
manière générale, on ne peut guère espérer que l’intégrale définie d’une
fonction soit en général aussi simple à calculer que la fonction elle-même.
Le critère de Valiant, malgré la simplicité de sa preuve, est un moyen
puissant pour fabriquer des familles dans VN Pe .
Comme toutes les preuves que nous avons données dans les chapitres
12 et 13, la preuve du critère de Valiant est clairement uniforme. Donc
si (gn ) est une famille dans #P (on prend pour entrée le mot formé par
l’entier bâton n suivi d’un 0 puis du mot p), alors la forme algébrique
déployée de (gn ) admet pour description une famille uniforme de circuits arithmétiques dans N C 1 qui utilise les seules constantes 0, 1 et
−1 et qui donne le résultat correct sur tout anneau contenant Z.
13.3
Complexité binaire versus complexité booléenne
Famille de fonctions booléennes associée à un problème algorithmique
Notons {0, 1}? l’ensemble des mots écrits sur l’alphabet {0, 1}. Nous
pouvons voir cet ensemble comme la réunion disjointe des {0, 1}n .
Considérons un problème algorithmique P qui est codé sous forme
binaire : autrement dit, toute instance de ce problème correspond à une
336
13. Le permanent et la conjecture P =
6 NP
question codée comme un élément de {0, 1}n pour un certain entier n
et la réponse à la question, du type oui ou non, est elle-même codée
comme un élément de {0, 1}.
On peut interpréter ce problème P comme fournissant, pour chaque
n, une fonction booléenne fn : {0, 1}n → {0, 1}. Nous dirons que la
famille (fn ) est la famille de fonctions booléennes associée au problème
algorithmique P.
Supposons maintenant que le problème P porte sur les graphes
orientés. Un code naturel pour un graphe orienté à n sommets est sa
matrice d’ajacence qui est une matrice dans {0, 1}n×n . Cette matrice
contient 1 en position (i, j) si et seulement si il y a une arète orientée
qui va de i à j dans le graphe considéré. Dans ce cas, on voit que la
famille de fonctions booléennes associée au problème P est plus natu2
rellement définie comme une famille fn : {0, 1}n → {0, 1}.
On dira que le problème algorithmique P est dans une classe de
complexité booléenne C si la famille de fonctions booléennes qui lui est
naturellement attachée est dans C.
Famille d’applications booléennes associée à une fonction algorithmique
Considérons maintenant une fonction algorithmique F, une fonction
qu’on aurait envie de faire calculer par un ordinateur : l’entrée et la sortie
sont codées en binaire, c’est-à-dire considérées comme des éléments de
{0, 1}? .
Supposons que t(n) est une majoration de la taille de la sortie en
fonction de la taille n de l’entrée et que la fonction t (( n’est pas plus
difficile à calculer )) que F.
Nous pouvons alors recalibrer la fonction F de manière que la taille
de sa sortie ne dépende que de la taille de son entrée. Par exemple nous
prenons la fonction G qui, pour un mot µ en entrée de taille n, calcule
le mot F (µ) précédé d’un 1, lui-même précédé du nombre de 0 nécessaire
pour atteindre la longueur 1 + t(n). Il est clair qu’on récupère facilement
F à partir de G.
Cette convention nous permet d’associer à toute fonction algorithmique F une famille d’applications booléennes
fn : {0, 1}n → {0, 1}1+m
où m = t(n) est une majoration de la taille de la sortie en fonction de
13.3. Complexité binaire versus complexité booléenne
337
la taille de l’entrée. La famille d’applications booléennes associée à F
dépend donc de la fonction de majoration t que l’on considère.
Définition 13.3.1 Dans les conditions ci-dessus, nous dirons que la famille (fn ) est la famille d’applications booléennes associée à la fonction
algorithmique F avec la fonction de majoration t. Si nous ne précisons
pas cette fonction de majoration, nous disons simplement que la famille
(fn ) est une famille d’applications booléennes associée à la fonction algorithmique F. On dira que la fonction algorithmique F est dans une
classe de complexité booléenne C si la famille (fn ) est dans C.
Lorsque la fonction F est calculable, dans une classe de complexité
binaire connue, on choisira toujours la fonction de majoration suffisamment simple, de façon que la fonction G reste dans la même classe de
complexité.
Tout ce que nous venons de dire s’applique par exemple à une fonction F de {0, 1}? ou N vers N ou Z, modulo des codages binaires naturels
convenables.
Familles uniformes de circuits booléens
Considérons un problème algorithmique P qui est codé sous forme
binaire : pour chaque n, une fonction booléenne fn : {0, 1}n → {0, 1}
donne la réponse pour les mots de longueur n. Cette famille (fn ) peut
être réalisée sous forme d’une famille d’expressions booléennes, ou sous
forme d’une famille de circuits booléens.
En complexité binaire on s’intéresse à la fois à la taille (de ces expressions ou de ces circuits) et à la difficulté proprement algorithmique
qu’il peut y avoir à produire l’expression (ou le circuit) no n en fonction
de l’entrée n (codée en unaire). Ce deuxième aspect correspond à la
question : la famille est-elle uniforme ?
Un théorème précis donne l’interprétation de la calculabilité en temps
polynomial en termes de familles de circuits booléens (cf. [BDG] théorème 5.19).
Théorème 13.7 Soit P un problème algorithmique codé sous forme
binaire.
(1) Si le problème P est résoluble en temps T par une Machine de
Turing à une seule bande, on peut construire en temps O(T 2 ) une
famille de circuits booléens qui représente la famille de fonctions
booléennes (fn ) associée à P .
338
13. Le permanent et la conjecture P =
6 NP
(2) Le problème P est résoluble en temps polynomial par une Machine de Turing si et seulement si il existe une famille uniforme
de circuits booléens (γn ) qui représente la famille de fonctions
booléennes (fn ) associée à P .
Ce théorème est important, même s’il a l’air de se mordre un peu la
queue, puisque la famille (γn ) doit être uniforme, c’est-à-dire calculable
en temps polynomial par une Machine de Turing.
Il n’est pas trop compliqué à démontrer. Il est relié à l’existence
d’une Machine de Turing universelle qui travaille en temps polynomial.
Le fait que le résultat du calcul (sur une entrée µ de taille n) au bout
de t étapes est bien celui affiché peut être vérifié en exécutant soi-même
le programme à la main, et on peut certifier la totalité du calcul en
certifiant le résultat de chaque étape intermédiaire. Quand au bout du
compte, on dit (( la sortie a été correctement calculée )), on peut aussi
l’écrire en détail sous forme d’un circuit booléen γn qui fonctionne pour
toute entrée de taille n. Il faut un peu d’attention pour vérifier que
tout ceci reste dans le cadre de la taille polynomiale. C’est le même
genre d’argument qui a permis à Cook de fournir le premier et le plus
populaire des problèmes N P - complets, celui de la satisfiabilité des expressions booléennes (étant donné une expression booléenne, existe-t-il
une façon d’affecter les variables booléennes en entrée qui donne à l’expression la valeur Vrai ?), problème plus parlant que le problème N P complet universel que nous avons exposé page 140 dans la section 4.2.
Le théorème 13.7 nous donne immédiatement.
Proposition 13.3.2 Soit P un problème algorithmique, codé sous forme binaire. Soit F : {0, 1}? → N une fonction algorithmique.
– Si le problème P est dans la classe P, alors il est dans P/poly.
– Si le problème P est dans la classe N P, alors il est dans N P/poly.
– Si la fonction F est dans la classe #P, alors elle est dans #P/poly.
La signification intuitive importante du théorème 13.7 est que la
classe P/poly est l’exact analogue non uniforme de la classe P : soit en
effet P un problème algorithmique qui correspond à une famille (fn )
de fonctions booléennes,
– le problème P est dans la classe P signifie que (fn ) est calculable
par une famille uniforme γn de circuits booléens (à fortiori la
taille de (γn ) est polynomiale en n),
13.4. Le caractère universel du permanent
339
– le problème P est dans la classe P/poly signifie que (fn ) est
calculable par une famille (γn ) de circuits booléens dont la taille
est polynomiale en n.
Étant donné que N P/poly et #P/poly sont définis à partir de P/poly
de manière similaire à la définition de N P et #P à partir de P, une
autre signification intuitive importante du théorème 13.7 est que les
classes N P/poly et #P/poly sont les exacts analogues non uniformes
des classes N P et #P.
La preuve qu’un problème algébrique donné P est N P-complet donne
en général un procédé uniforme de réduction d’une famille arbitraire
dans BN P (uniforme ou non) à la famille de fonctions booléennes attachée à P . En conséquence on obtient l’implication P = N P ⇒ BP =
BN P. Autrement dit la conjecture non uniforme BP 6= BN P est plus
forte que la conjecture classique P =
6 N P.
La même remarque vaut en remplaçant N P par #P.
13.4
Le caractère universel du permanent
Le permanent
Par définition le permanent d’une matrice carrée A = (aij )1≤i,j≤n
sur un anneau commutatif A est le polynôme en les aij , noté per(A),
défini par l’expression analogue à celle du déterminant obtenue en remplaçant les signes − par les + :
per(A) = pern ((aij )1≤i,j≤n ) =
X
σ∈Sn
Yn
i=1
ai,σ(i)
(13.7)
où σ parcourt toutes les permutations de {1, . . . , n}. Nous considérons
(pern ) comme une famille de polynômes à n2 variables sur l’anneau A.
On ne connaı̂t pas de manière rapide d’évaluer le permanent d’une
matrice à coefficients entiers, ni sur aucun corps de caractéristique distincte de 2 (en caractéristique 2 le permanent est égal au déterminant
et se laisse donc évaluer facilement).
Lorsque les coefficients sont tous égaux à 0 ou 1 on peut interpréter
la matrice A comme donnant le graphe d’une relation entre deux ensembles à n éléments F et G. Par exemple les éléments de F sont
des filles et ceux de G sont des garçons, et la relation est la relation
d’affinité (ils veulent bien danser ensemble). Alors le permanent de la
matrice correspondante compte le nombre de manières distinctes de rem-
340
13. Le permanent et la conjecture P =
6 NP
plir la piste de danse sans laisser personne sur le bord. Ainsi la famille
2
fn : {0, 1}n → N définie par fn (A) = pern (A) est une famille dans #P.
Le critère de Valiant (théorème 13.6 page 335) montre par ailleurs
que la famille de polynômes (pern ) est dans VN Pe sur n’importe quel
anneau commutatif : en effet la famille (pern ) n’est autre que la forme
algébrique déployée de la famille des fonctions booléennes qui testent si
une matrice dans {0, 1}n×n est une matrice de permutation.
Deux théorèmes de Valiant sur le permanent
Valiant a établi l’égalité VN P = VN Pe et il a montré le caractère
universel du permanent, à la fois en complexité binaire et en complexité
algébrique.
Théorème 13.8 Le calcul du permanent pour les matrices carrées à
coefficients dans {0, 1} est #P-complet.
Théorème 13.9 Sur un corps de caractéristique 6= 2, et plus généralement sur un anneau dans lequel 2 est inversible, la famille (pern )
est universelle pour la classe VN P : toute famille dans VN P est une
p-projection de la famille (pern ).
Les preuves de ces deux théorèmes sont délicates. Pour le deuxième
nous recommandons [Bur].
13.5
La conjecture de Valiant
Le petit tableau ci-après récapitule les analogies entre différentes
classes de complexité. Dans les colonnes Booléen et Algébrique interviennent des familles non uniformes d’expressions ou circuits. Dans
la colonne Sim nous indiquons si la simulation algébrique du cas booléen
est connue comme étant sur la même ligne : deux points d’interrogation
signifient qu’on ne le croit guère possible.
Rappelons que dans la première colonne (complexité binaire) toutes
les inclusions en descendant sont conjecturées être strictes, et que les
inclusions correspondantes dans le cas booléen (2ème colonne) sont aussi
conjecturées strictes.
13.5. La conjecture de Valiant
341
Petit récapitulatif
(Analogies entre complexité binaire, booléenne et algébrique)
Binaire
Booléen
Algébrique
Sim
N C1
BPe = BN C 1
VPe = VN C 1
oui
N C2
BN C 2
VN C 2
??
NC
BN C
VN C = VN C 2
??
P
BP
VP = VN C 2
??
VQP = VQPe
NP
BN P = BN Pe
#P
#BP = #BPe
VN P = VN Pe
oui
Valiant a proposé la conjecture :
Pour tout corps K,
VP 6= VN P.
Cette conjecture est un analogue algébrique non uniforme de la conjecture algorithmique P =
6 N P ou plus précisément de P =
6 #P.
Sur un corps de caractéristique 6= 2, vu le théorème 13.9, cette conjecture s’écrit purement en termes d’expressions arithmétiques :
Le permanent n’est pas une p-projection du déterminant.
C’est sur les corps finis que la conjecture semble le plus significative,
parce que la situation algébrique y est le plus proche du cas booléen : elle
n’est pas perturbée par la présence d’éléments de taille arbitrairement
grande dans le corps.
Si on disposait d’une procédure uniforme qui réduise la famille (pern )
à une famille dans VP, alors le calcul du permanent d’une matrice dans
2
{0, 1}n serait dans la classe P et donc on aurait P = #P par le théorème 13.8.
Plus généralement, le théorème 13.5 page 333 montre que P/poly 6=
#P/poly implique VP(Fq ) 6= VN P(Fq ) pour tout corps fini, et sous
l’hypothèse de Riemann généralisée, VP(K) 6= VN P(K) pour tout corps
de caractéristique nulle.
Par ailleurs si on avait P = #P, le calcul du permanent d’une matrice
2
dans {0, 1}n serait dans la classe P, donc a foriori dans P/poly et on
aurait #P/poly = P/poly, mais peut-être pas pour autant VP = VN P.
342
13. Le permanent et la conjecture P =
6 NP
La conjecture de Valiant est qu’il n’existe aucune procédure, même
sans l’hypothèse restrictive d’uniformité, qui réduise la famille (pern ) à
une famille dans VP.
L’avantage de la conjecture de Valiant est qu’elle est un problème
purement algébrique, qui parle uniquement de la taille de la représentation d’une certaine famille de polynômes par des familles de circuits
arithmétiques.
Comme un des aspects les plus mystérieux de la conjecture P 6=
N P (cela n’a pas toujours représenté un million de dollars 4 mais cela
a toujours semblé très excitant) tient à la question de l’uniformité des
familles de circuits booléens en jeu, on contournerait cet obstacle si on
démontrait la conjecture analogue non uniforme et plus forte P/poly 6=
N P/poly.
Et la forme purement algébrique VP 6= VN P serait plus à notre
portée. Une preuve de VP 6= VN P serait un pas important qui éclairerait
le chemin pour une preuve de P/poly 6= #P/poly, qui implique P =
6 #P.
Cela pourrait suggérer enfin une preuve de P =
6 N P.
Un petit ennui dans cette suite de considérations informelles : les
deux points d’interrogation sur la ligne P du petit tableau récapitualtif.
Comme VQP = VQPe et VN P = VN Pe , la conjecture de Valiant
étendue, à savoir :
Pour tout corps K,
VN P 6⊂ VQP
est regardée par certains auteurs comme encore plus instructive pour la
compréhension du problème algorithmique analogue P =
6 #P.
Sur un corps de caractéristique 6= 2, cela équivaut à :
Le permanent n’est pas une qp-projection du déterminant.
Notons que Bürgisser a démontré que VQP 6⊂ VN P sur les corps
de caractéristique nulle (voir [Bur]).
4. Un milliardaire américain qui aimerait devenir célèbre a proposé en l’an 2000 un
prix d’un million de dollars pour celui ou celle qui résoudrait le problème P = N P ?.
Six autres conjectures mathématiques importantes sont dotées d’un prix analogue.
Un million de dollars n’est d’ailleurs pas grand chose comparé à ce que gagne un bon
joueur de football, et rien du tout par rapport à un avion furtif. Ceci tendrait à dire
qu’un milliardaire peut espérer devenir célèbre avec un investissement très modeste.
Notez que si vous démontrez que #P =
6 N P, vous aurez droit à l’admiration de
tou(te)s les mathé/infor-maticien(ne)s, mais vous n’aurez pas le million de dollars
correspondant à P 6= N P. C’est certainement injuste, mais c’est ainsi.
Annexe : codes Maple
Nous donnons, dans les pages qui suivent, les codes Maple des algorithmes qui calculent le polynôme caractéristique et dont nous avons
testé les performances.
Les codes sont écrits ici dans la version Maple 6, mais les tests ont
été faits avec la version Maple 5. Les différences sont les suivantes.
Premièrement la version Maple 6 a grandement amélioré son calcul
standard de polynôme caractéristique (en le basant sur l’algorithme de
Berkowitz ?). Deuxièmement, dans Maple 6, le dernier objet calculé
est désigné par % alors que dans Maple 5 il était désigné par ". Enfin,
dans Maple 6 une procédure se termine par end proc: tandis que
dans Maple 5 elle se termine par end:
Les algorithmes que nous avons comparés sont ceux de Berkowitz
amélioré (noté berkodense), de Jordan-Bareiss modifié (barmodif), de
Faddeev-Souriau-Frame (faddeev), de Chistov (chistodense) et leurs
versions modulaires respectives (nous donnons ici berkomod), ainsi que
les algorithmes correspondant à la méthode d’interpolation de Lagrange,
celle de Hessenberg et celle de Kaltofen-Wiedemann (notés respectivement interpoly, hessenberg et kalto), en plus de la fonction charpoly
faisant partie du package linalg de Maple que nous avons notée
linalpoly dans nos tableaux de comparaison. Nous avons également
adapté berkodense et chistodense au cas des matrices creuses (voir
les codes berksparse et chisparse dérivés)
Les mesures du temps CPU et de l’espace-mémoire pour chaque algorithme testé sont prises à l’aide des fonctions time() et bytesalloc
du noyau de Maple.
344
Annexe
### Somme des éléments d’une liste ###
###
de fractions rationnelles
###
somme : = proc(suite : : list(ratpoly))
normal(convert(suite,‘+‘))
end proc:
###
##### Berkowitz dans le cas d’une matrice dense #####
berkodense : = proc(A : : matrix,X : : name)
local n,r,i,j,k,V,C,S,Q;
n : = coldim(A);
V : = table([1 = -1,2 = A[1,1]]); C[1] : = -1;
for r from 2 to n do
for i to r-1 do S[i] : = A[i,r] od; C[2] : = A[r,r];
for i from 1 to r-2 do
C[i+2] : = somme([seq(A[r,k]*S[k],k=1..r-1)]);
for j to r-1 do
Q[j] : = somme([seq(A[j,k]*S[k],k=1..r-1)])
od;
for j to r-1 do S[j] : = Q[j] od;
od;
C[r+1] : = somme([seq(A[r,k]*S[k],k=1..r-1)]);
for i to r+1 do
Q[i] : = somme([seq(C[i+1-k]*V[k],k=1..min(r,i))]);
od;
for i to r+1 do V[i] : = Q[i] od;
od;
somme([seq(V[k+1]*X^ (n-k),k=0..n)]);
collect(%,X)
end proc:
#####
Codes Maple
345
##### Berkowitz dans le cas d’une matrice creuse #####
berksparse : = proc(A : : matrix,X : : name)
local n,r,i,j,k,V,C,S,Q,N;
n : = coldim(A);
V : = table([1 = -1,2 = A[1,1]]);
N : = vector(n);
for i to n do N[i] : = {} od;
C[1] : = -1;
for r from 2 to n do
for i to n do
if A[i,r-1]<>0 then N[i] : = N[i] union {r-1} fi
od;
for i to r-1 do S[i] : = A[i,r] od; C[2] : = A[r,r];
for i from 1 to r-2 do
C[i+2] : = somme([seq(A[r,j]*S[j],j = N[r])]);
for j to r-1 do
Q[j] : = somme([seq(A[j,k]*S[k],k = N[j])]);
od;
for j to r-1 do S[j] : = Q[j] od;
od;
C[r+1] : = somme([seq(A[r,j]*S[j],j = N[r])]);
for i to r+1 do
Q[i] : = somme([seq(C[i+1-k]*V[k],k=1..min(r,i)]);
od;
for i to r+1 do V[i] : = Q[i] od;
od;
somme([seq(V[k+1]*X^ (n-k),k=0..n)]);
collect(%,X)
end proc;
#####
Nous avons également adapté les codes Maple ci-dessus, correspondant à l’algorithme amélioré de Berkowitz, au cas où les coefficients
appartiennent à un anneau-quotient du type Zp [lisvar] /hIdeali . On
obtient une procédure, notée berkomod dans nos tableaux de comparaison, qui prend en entrée un entier positif p , une liste d’indéterminées lisvar , une liste Ideal de polynômes en lisvar et la matrice
A ∈ (Zp [lisvar] /hIdeali )n×n pour donner en sortie le polynôme caractéristique de A .
La procédure berkomod ainsi que les versions modulaires des autres
346
Annexe
algorithmes utilisent comme sous-procédure la procédure polmod qui
prend en entrée un nombre entier p, un polynôme P de Z[lisvar] , et
donne en sortie un représentant simple de l’image canonique de P dans
l’anneau-quotient Zp [lisvar] /hIdeali .
##### Réduction d’un polyn^
ome modulo un idéal #####
polmod : =
proc(P : : polynom,lisvar : : list,Ideal : : list,p : : posint)
local i, Q;
if nops(lisvar)<>nops(Ideal) then
ERROR(‘The number of polynomials must
be equal to the number of variables‘)
fi;
Q : = P;
for i to nops(lisvar) do
Q : = rem(Q,Ideal[i],lisvar[i]);
Q : = Q mod p
od;
sort(Q);
end proc:
#####
On en déduit les deux calculs de base modulo l’idéal considéré, la
somme d’une liste et le produit de deux éléments.
##### Somme d’une liste modulo un idéal #####
sommod : = proc(s : : list(polynom),
lsv : : list(name),lsp : : list(polynom),p : : posint)
polmod(somme(s),lsv,lsp,p)
end proc:
#####
##### Evaluation d’un produit modulo un un ideal #####
promod : = proc(P,Q : : polynom,
lsv : : list(name),lsp : : list(polynom),p : : posint)
polmod(P*Q,lsv,lsp,p)
end proc:
#####
Il ne reste plus qu’à réécrire berkodense en y remplaçant les opérations
somme d’une liste de polynômes et produit de deux polynômes par les
calculs modulaires donnés par sommod et promod.
Codes Maple
347
##### Berkowitz modulaire #####
berkomod : = proc(A : : matrix,X : : name,lsv : : list(name),
lsp : : list(polynom),p : : posint)
local n,r,i,j,V,C,S,Q;
n : = coldim(A);
V : = table([1 = -1,2 = A[1,1]]); C[1] : = -1;
for r from 2 to n do
for i to r-1 do S[i] : = A[i,r] od; C[2] : = A[r,r];
for i from 1 to r-2 do
[seq(promod(A[r,k],S[k],lsv,lsp,p),k=1..r-1)];
C[i+2] : = sommod(%,lsv,lsp,p);
for j to r-1 do
[seq(promod(A[j,k],S[k],lsv,lsp,p),k=1..r-1)]
Q[j] : = sommod(%,lsv,lsp,p)
od;
for j to r-1 do S[j] : = Q[j] od;
od;
[seq(promod((A[r,k],S[k],lsv,lsp,p),k=1..r-1)];
C[r+1] : = sommod(%,lsv,lsp,p);
for i to r+1 do
[seq(promod((C[i+1-k],V[k],lsv,lsp,p),
k=1..min(r,i))];
Q[i] : = sommod(%,lsv,lsp,p);
od;
for i to r+1 do V[i] : = Q[i] od;
od;
somme([seq(V[k+1]*X^ (n-k),k=0..n)]);
collect(%,X)
end proc:
#####
Voici maintenant sans plus de commentaire les codes Maple des
algorithmes chistodense, chisparse, barmodif, faddeev, interpoly, hessenberg, kalto.
348
Annexe
##### (Chistov. Cas des matrices denses) #####
chistodense : = proc(A : : matrix,X : : name)
local n,r,i,j,k,a,b,C,V,W,Q;
n : = coldim(A);
a : = array(0..n,[1]); C : = array(0..n,[1]);
for i to n do a[i] : = normal(a[i-1]*A[1,1]) od;
for r from 2 to n do
for i to r do V[i] : = A[i,r] od; C[1] : = V[r];
for i from 2 to n-1 do
for j to r do
W[j] : = somme([seq(A[j,k]*V[k],k = 1..r)]);
od;
for j to r do V[j] : = W[j] od; C[i] : = V[r];
od;
[C[n] : = somme(seq(A[r,k]*V[k],k = 1..r)]);
for j from 0 to n do
b[j] : = somme([seq(C[j-k]*a[k],k = 0..j)]);
od;
for j from 0 to n do a[j] : = b[j] od;
od;
Q : = somme([seq(X^ k*a[k],k = 0..n)]);
Q : = X^ n*subs(X = 1/X,inversf(Q,X,n));
Q : = collect((-1)^ n*Q,X)
end proc:
#####
### Calcul de l’inverse modulo z (n+1) d’un polyn^
ome en z ###
inversf : = proc(P,z,n)
collect(convert(series(1/P,z,n+1),polynom),z,normal)
end proc:
### cette procédure utilisée dans les algorithmes
### de Chistov sera aussi utile dans l’algorithme kalto
Codes Maple
##### Chistov. Cas des matrices creuses #####
chisparse : = proc(A : : matrix,X : : name)
local n,r,i,j,k,a,b,C,N,V,W,Q;
n : = coldim(A);
a : = array(0..n,[1]); C : = array(0..n,[1]);
N : = array(1..n); for i to n do N[i] : = {} od;
## N[i]; éléments non nuls de la i-ème ligne
for i to n do
for j to n do
if A[i,j] <> 0 then N[i] : = N[i] union {j} fi
od
od; #### Fin de la construction de N
for i to n do a[i] : = normal(a[i-1]*A[1,1]) od;
for r from 2 to n do
for i to r do V[i] : = A[i,r] od;
C[1] : = V[r];
for i from 2 to n-1 do
for j to r do
[seq(A[j,k]*V[k],k = {$ 1..r} intersect N[j])];
W[j] : = somme(%)
od;
for j to r do V[j] : = W[j] od;
C[i] : = V[r];
od;
[seq(A[r,k]*V[k],k = {$ 1..r} intersect N[r])];
C[n] : = somme(%);
for j from 0 to n do
b[j] : = somme([seq(C[j-k]*a[k],k = 0..j)]);
od;
for j from 0 to n do a[j] : = b[j] od;
od;
Q : = somme([seq(X^ k*a[k],k = 0..n)]);
Q : = X^ n*subs(X = 1/X,inversf(Q,X,n));
Q : = collect((-1)^ n*Q,X)
end proc:
#####
349
350
Annexe
##### Jordan-Bareiss Modifié #####
barmodif : = proc(A : : matrix,X : : name)
local B,n,p,i,j,piv,dencoe;
den : = 1; n : = coldim(A); B : = copy(A);
B : = evalm(B-X*array(identity, 1..n,1..n)); piv : = B[1,1];
for p from 1 to n-1 do
for i from p+1 to n do
coe : = B[i,p];
for j from p+1 to n do
B[i,j] : = normal((piv*B[i,j]-coe*B[p,j])/den)
od
od;
den : = piv; piv : = B[p+1,p+1]
od;
sort(collect(piv,X),X)
end proc:
#####
##### Faddeev-Souriau-Frame #####
faddeev : = proc(A : : matrix,X : : name)
local n, k, a, C, B, Id, P;
n : = coldim(A); a : = array(1..n);
Id : = array(1..n,1..n,identity); B : = copy(Id);
for k from 1 to n do
C : = map(normal,multiply(A,B));
a[k] : = trace(C)/k;
B : = map(normal,evalm(C-a[k]*Id))
od;
P : = somme([seq(a[k]*X^ (n-k),k=1..n)]);
sort((-1)^ n*(X^ n-P,X);
end proc:
#####
##### Interpolation de Lagrange #####
interpoly : = proc(M : : matrix,X : : name)
local n,Id,i,j,N,d,L;
n : = coldim(M); Id : = array(identity, 1..n, 1..n);
for i to n+1 do d[i] : = det(evalm(M-(i-1)*Id)) od;
L : = [seq(d[j], j = 1..n+1)];
interp([‘$‘(0 .. n)], L, X);
end proc:
#####
Codes Maple
351
##### Méthode de Hessenberg #####
hessenberg : = proc(A : : matrix,X : : name)
local jpiv, ipiv, iciv, i, m, n, piv, c, H, P;
# Initialisations
n : = coldim(A); P[0] : = 1; H : = copy(A);
# Réduction de H à la forme de Hessenberg
for jpiv from 1 to n-2 do
ipiv : = jpiv+1; iciv : = ipiv; piv : = normal(H[iciv,jpiv]) ;
while piv = 0 and iciv < n do
iciv : = iciv+1 ; piv : = normal(H[iciv,jpiv])
od ;
if piv <> 0 then
if iciv > ipiv then
H : = swaprow(H,ipiv,iciv); # Echange de lignes
H : = swapcol(H,ipiv,iciv) # Echange de colonnes
fi;
for i from iciv+1 to n do
c : = normal(H[i,jpiv]/piv) ;
H : = addrow(H,ipiv,i,-c);# Manipulation de lignes
H : = addcol(H,i,ipiv,c) # Manipulation de colonnes
od;
H : = map(normal,H)
fi
od ;
# Calcul du polyn^
ome caractéristique
for m from 1 to n do
P[m] : = normal((H[m,m]-X) * P[m-1]) ; c : = 1 ;
for i from 1 to m-1 do
c : = normal(-c * H[m-i+1,m-i]) ;
P[m] : = normal(P[m]+c * H[m-i,m]* P[m-i-1])
od
od ;
collect(P[n],X) # le polyn^
ome caractéristique de A.
end proc;
#####
352
Annexe
#### developpement limité à l’ordre n ####
devlim : = proc(s : : ratpoly,u : : name,n : : integer)
convert(series(s,u,n+1),polynom); collect(",u,normal)
end proc;
#####
##### Kaltofen-Wiedemann #####
kalto : = proc(A : : matrix,X : : name)
local n,i,j,k,a,b,bv,bw,c,B,C,P,u;
n : = coldim(A);
### Initialisation
a : = stre(n); C : = stra(n);
b : = vector(2*n); B : = evalm(C+u*(A-C));
### Calcul des b i
b[1] : = a[1];
bv : = copy(a); bw : = vector(n);
for i from 2 to n+1 do
## multiplication de B par bv
for j to n do
bw[j] : = somme([seq(B[j,k]*bv[k],k = 1..n)]);
od;
for j to n do bv[j] : = bw[j] od;
b[i] : = bv[1]
od;
for i from n+2 to 2*n do
## multiplication de B par bv
for j to n do
bw[j] : = somme([seq(B[j,k]*bv[k],k = 1..n)]);
od;
for j to n do bv[j] : = bw[j] od;
b[i] : = devlim(bv[1],u,n);
od;
P : = polgenmin(b,X,u,n);
P : = sort(subs(u = 1,(-1)^ n*res),X);
P : = collect(P,X,normal)
end proc:
#####
Codes Maple
##### Sous-procédures utilisées dans kalto
353
#####
### polgenmin: Procédure de Berlekamp-Massey
###
### pour le calcul du polyn^
ome générateur minimal ###
###
d’une suite récurrente linéaire
###
###
Ici, l’anneau de base est l’anneau des
###
###
développements limités A[z]/<z^(n+1)>
###
polgenmin : = proc(b : : vector,X : : name,z : : name,n : : integer)
local i,lc,ilc,ill,R1,R2,R3,V1,V2,V3,Q;
R1 : = somme([seq(b[2*n-k]*X^ k,k = 0..2*n-1)]);
Q : = quo(X^(2*n),R,X,’R2’);
V1 : = 1; V2 : = -Q; ill : = 1;
for i from 2 to n do
### traiter R2
R2 : = collect(R2,X,normal);
lc : = lcoeff(R2,X);
ilc : = inversf(lc,z,n);
R2 : = devlim(ilc*R2,z,n);
Q : = quo(R1,R2,X,’R3’);
Q : = devlim(Q,z,n);
R3 : = devlim(R3,z,n);
V3 : = devlim(ill*V1-ilc*V2*Q,z,n);
ill : = ilc;
V1 : = V2; V2 : = sort(V3,X);
R1 : = R2; R2 : = sort(R3,X);
od;
V2 : = collect(V2,X);
lc : = lcoeff(V2,X); ilc : = inversf(lc,z,n);
V2 : = devlim(ilc*V2,z,n);
V2 : = collect(V2,z,normal)
end proc:
#####
354
Annexe
#### Vecteur du centre d’élimination des divisions ####
stre : = proc(n)
local i,a;
a : = vector(n);
for i to n do
a[i] : = binomial(i-1,floor((i-1)/2))
od;
eval(a)
end proc:
#####
#### Matrice du centre d’élimination des divisions ####
stra : = proc(n)
local i,C;
C : = array(1..n,1..n,sparse);
for i to n-1 do
C[n,i] : = (-1)^ floor((n-i)/2) *
binomial(floor((n+i-1)/2),i-1);
C[i,i+1] : = 1
od;
C[n,n] : = 1; evalm(C)
end proc:
#####
Liste des algorithmes,
circuits et programmes
d’évaluation
2.1 Algorithme du pivot de Gauss simplifié et LU -décomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Deuxième algorithme du pivot de Gauss simplifié . . . .
2.3 LUP -décomposition d’une matrice surjective . . . . . . .
2.4 Algorithme de Jordan-Bareiss . . . . . . . . . . . . . . .
2.5 Algorithme de Dodgson pour une matrice de Hankel . .
2.6 Algorithme de Hessenberg . . . . . . . . . . . . . . . . .
2.7 Algorithme de Le Verrier . . . . . . . . . . . . . . . . . .
2.8 Algorithme de Souriau-Faddeev-Frame . . . . . . . . . .
2.9 Algorithme de Preparata & Sarwate, version séquentielle
2.10 Algorithme de Berkowitz, principe général . . . . . . .
2.11 Algorithme de Berkowitz, version séquentielle . . . . .
2.12 Algorithme de Chistov, principe général . . . . . . . . .
2.13 Algorithme de Chistov, version séquentielle simple . . .
2.14 Algorithme JorBarSol . . . . . . . . . . . . . . . . . . .
2.15 Algorithme de Frobenius. . . . . . . . . . . . . . . . . .
2.16 Algorithme de Berlekamp-Massey . . . . . . . . . . . .
3.1 Programme d’évaluation du déterminant et de la LU -décomposition d’une matrice carrée d’ordre 4 . . . . . . .
3.2 Circuit de l’algorithme du pivot de Gauss simplifié . . .
3.3 Programme d’évaluation du déterminant et de la LU -décomposition de la matrice carrée I3 + zF . . . . . . . .
3.4 Programme d’évaluation du déterminant d’une matrice
carrée après élimination des divisions à la Strassen . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
60
62
64
68
73
76
83
85
89
91
92
95
96
99
101
109
. 113
. 117
. 123
. 124
356
Liste des algorithmes
5.1
5.2
Calcul Parallèle des Préfixes . . . . . . . . . . . . . . . . 164
Exemples de circuits Pk (n) (k ∈ {0, 1}) . . . . . . . . . . 169
6.1 Produit de deux polynômes à la Karatsuba . . . . . . . . 173
6.2 Transformation de Fourier Discrète . . . . . . . . . . . . . 178
7.1 Multiplication par blocs, à la Strassen-Winograd . . . . . 189
7.2 Produit à la Strassen de deux matrices carrées d’ordre 2 . 192
8.1 LUP -décomposition à la Bunch & Hopcroft . . . . . . . . 232
8.2 Algorithme de Kaltofen-Wiedemann . . . . . . . . . . . . 249
9.1 Algorithme de Csanky . . . . . . . . . . . . . . . . . . . . 256
9.2 Algorithme de Preparata & Sarwate . . . . . . . . . . . . 262
Liste des figures
5.1
5.2
5.3
5.4
Construction récursive des circuits binaires équilibrés
Schéma de la construction récursive des circuits Pk (n)
Construction récursive des circuits P0 (n) . . . . . . .
Construction récursive des circuits P1 (n) . . . . . . .
.
.
.
.
.
.
.
.
163
166
167
168
7.1
7.2
7.3
7.4
7.5
7.6
7.7
7.8
7.9
7.10
7.11
Produit matriciel à trou de Bini . . . . . . . . . . .
Produit matriciel plein . . . . . . . . . . . . . . . .
Numérotation par blocs . . . . . . . . . . . . . . .
Bini itéré une fois . . . . . . . . . . . . . . . . . . .
Bini itéré deux fois . . . . . . . . . . . . . . . . . .
Produit à trous extrait de (( Bini itéré deux fois )) .
Un exemple arbitraire de produit matriciel à trous
L’exemple précédent itéré une fois . . . . . . . . .
Somme directe de deux produits matriciels . . . . .
Somme directe itérée . . . . . . . . . . . . . . . . .
Somme directe itérée, variante . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
210
211
213
216
216
217
219
220
222
225
226
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
10.1 Calcul de Uν . . . . . . . . . . . . . . . . . . . . . . . . . 276
10.2 Calcul de Vν . . . . . . . . . . . . . . . . . . . . . . . . . 277
10.3 Calcul des produits partiels . . . . . . . . . . . . . . . . . 281
12.1
12.2
12.3
12.4
12.5
12.6
L’arbre de Horner . . . . . . .
Parallélisation d’une expression,
Parallélisation d’une expression,
Parallélisation d’une expression,
L’arbre de Horner parallélisé .
Une expression p-universelle . .
. .
1
2
3
. .
. .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
305
310
311
311
312
320
Bibliographie
[AHU] Aho A.V. & Hopcroft J. & Ulmann J. The Design and
Analysis of Computer Algorithms, Addison Wesley, Reading,
MA, 1974. 62, 171
[BDG] Balcázar J. L., Dı́az J. & Gabarró J. Structural complexity. I. Second edition. Texts in Theoretical Computer
Science. Springer, (1995). v, 131, 135, 324, 337
[Bha]
Bhaskara Rao K. The Theory of Generalized Inverses over a
Commutative Ring. Taylor & Francis Londres, 2002. 48
[BP]
Bini D. & Pan V. Polynomial and Matrix Computations, Vol.
1, Fundamental Algorithms, Birkhäuser, 1994. v, 40, 62, 184,
253, 259
[Bur]
Bürgisser P. Completeness and Reduction in Algebraic Complexity Theory, Springer, 2000. v, 271, 272, 303, 326, 331, 333,
340, 342
[BCS]
Bürgisser P., Clausen M. & Shokrollahi M. Algebraic
Complexity Theory, Springer, 1997. v, 151, 187, 237, 303, 309
[Cia]
Ciarlet P. Introduction à l’analyse numérique matricielle et
à l’optimisation, Dunod, 1998. 30, 41
[Coh]
Cohen H. A course in computational algebraic number theory,
Graduate Texts in Maths, vol. 138, Springer, 1993.
[CCS]
Cohen A., Cuypers H., Sterk H. (eds) Some tapas of computer algebra, Algorithms and Computation in Mathematics,
vol. 4, Springer, 1999. 365
[CT]
Cosnard M. & Trystram D. Algorithmes et architectures
parallèles, InterEditions, Paris, 1993. 152, 153, 154
[Dur]
Durand E. Solutions numériques des équations algébriques.
Tome II : Systèmes de plusieurs équations, Valeurs propres des
matrices. Masson, Paris, 1961. 66
360
Bibliographie
[FF]
Faddeev D.K. & Faddeeva V.N. Computational methods of
linear algebra , W.H. Freeman & Co., San Francisco, 1963. 13,
83, 90, 272
[FS]
Faddeev D. & Sominskii I. Collected Problems in Higher
Algebra, 1949, Problem no 979. 84
[GR]
Gabriel P. & Roiter A. Representations of finite-dimensional algebras, Springer, 1997. 55
[GG]
von zur Gathen J. & Gerhard J. Modern Computer Algebra, Cambridge University Press, 1999. v, 33, 108, 253
[Gan]
Gantmacher F. Théorie des Matrices, Tome 1, (Théorie
Générale), Dunod, Paris, 1966. v, 4, 20, 289
[Gas]
Gastinel N. Analyse numérique réelle, Hermann, Paris, 1966.
ii, 13, 90, 272
[Gob]
Goblot R. Algèbre commutative, Masson, Paris. 2ème édition,
2002. 20
[GL]
Golub G. & van Loan Ch. Matrix Computations, J. Hopkins
Univ. Press, Baltimore and London. 3ème édition, 1996. 30, 41
[GKW] Grabmeier J., Kaltofen E. & Weispfenning V. (Eds)
Computer Algebra Handbook : Foundations, Applications, Systems. Springer, 2003. v
[Jac]
Jacobson N. Lectures in abstract algebra, I. Basic concepts,
D. Van Nostrand & Co., Inc., Toronto, 1951. Réédition Springer
G.T.M. no 30, 1975. 1
[KKo]
Ker I. Ko Complexity theory of real functions, Birhaüser
(1991). 148
[Knu]
Knuth D. The art of computer programming, vol. 2 : Seminumerical algorithms, Second Edition, Addison Wesley Publishing
Co, 1981. i, v, 171, 186, 325
[Kou]
Koulikov L. Algèbre et Théorie des Nombres, Mir, Moscou,
1982.
[LT]
Lancaster P. & Tismenetsky M. The theory of matrices,
Second Edition, Academic Press, 1984. v, 41, 48
[Min]
Minc H. Permanents, Encyclopedia of Mathematics and its
Applications, Vol 6, Addison-Wesley, Reading MA, 1978.
[MRR] Mines R., Richman F. & Ruitenburg W. A Course in
Constructive Algebra, Universitext, Springer, 1988. 20
Bibliographie
361
[Pan]
Pan V. How to multiply matrices faster, Lecture Notes in Computer Science no 179, Springer Verlag, 1984. i
[Rob]
Robert Y. The impact of vector and parallel architectures on
the Gaussian elimination algorithm, Manchester Univ. Press,
Halsted Press, John Wiley & Sons, New-York Brisbane Toronto,
1990.
[Sav]
Savage J. The Complexity of Computing, Robert E. Krieger
Pub. Comp., Malabar, Florida, 1987. 325
[Sch]
Schrijver A. Theory of integer and linear programming, John
Wiley, New-York, 1985. 51, 136
[Ser]
Serre J.-P. Cours d’arithmétique, PUF, Paris, 1970. 182
[Ste]
Stern J. Fondements mathématiques de l’informatique. Mc
Graw-Hill, Paris, 1990. v, 131, 132, 135
[Tur]
Turing A., Girard J.-Y. La machine de Turing, Le Seuil.
Points Sciences, Paris, 1995. 131
[Weg]
Wegener I. The Complexity of Boolean Functions, WileyTeubner Series in Computer Science, Stuttgart, 1987. 321, 324,
328
Articles
[1] Abdeljaoued J. Algorithmes Rapides pour le Calcul du Polynôme Caractéristique, Thèse de l’Université de FrancheComté Besançon, 1997. 273
[2] Abdeljaoued J. Berkowitz Algorithm, Maple and computing
the characteristic polynomial in an arbitrary commutative ring,
Computer Algebra MapleTech, 4/3, Birkhauser Boston 1997.
[3] Abdeljaoued J. & Malaschonok G. Efficient algorithms
for Computing the characteristic Polynomial in a Domain,
Journal of Pure and Applied Algebra, 156, 2001, pp. 127–145.
81
[4] Bareiss E. Sylvester’s Identity and Multistep Integer-Preserving Gaussian Elimination, Math. Comput., 22, 1968, pp. 565–
578. 66
362
Bibliographie
[5] Baur W. & Strassen V. The complexity of partial derivatives, Theoretical Computer Science, 22, 1983, pp. 317–330.
126, 281
[6] Berkowitz S. On computing the determinant in small parallel
time using a small number of processors, Information Processing Letters, 18, 1984, pp. 147–150. ii, 90, 272, 274, 278
[7] Bini D. Relation between exact and approximate bilinear algorithms. Applications, Calcolo, 17, 1980, pp. 87–97. i
[8] Bini D., Capovani M., Lotti G. & Romani F. O(n2.7799 )
complexity of matrix multiplication, Inf. Proc. Letters, 8, 1979,
pp. 234–235. i
[9] Borodin A., von zur Gathen J. & Hopcroft J. Parallel matrix and GCD computations, Information & Control, 52,
1982, pp. 241–256. 272
[10] Brent R. The parallel evaluation of general arithmetic expressions, J. Assoc. Comp. Mach., 21, 1974, pp. 201–206. 156, 309
[11] Bunch J. & Hopcroft J. Triangular factorization and inversion by fast matrix multiplication, Math. Comput., 28/125,
1974, pp. 231–236. 231
[12] Bürgisser P. On the structure of Valiant’s complexity classes,
Discr. Math. Theoret. Comp. Sci., 3, 1999, pp. 73–94. 319, 320
[13] Cantor D. & Kaltofen E. On fast multiplication of polynomials over arbitrary rings, Acta Informatica, 28/7, 1991,
pp. 693–701. 125, 171, 180
[14] Chandra A. Maximal parallelism in matrix multiplication, Report RC – 6193, I.B.M. Watson Research Center, Yorktown
Heights, N.Y., 1976. i
[15] Chemla K. Résonances entre démonstration et procédure. Remarques sur le commentaire de Liu Hui (3ème siècle) aux
Neuf Chapitres sur les Procédures Mathématiques (1er siècle),
Extrême-Orient, Extrême-Occident, 14, 1992, pp. 91–129. 51
[16] Équipe de recherche (( Histoire des sciences et de la
philosophie arabes et médiévales )) Les problèmes comme
champ d’interprétation des algorithmes dans Les neuf chapitres
sur les procédures mathématiques et leurs commentaires. De la
résolution des systèmes d’équations linéaires, Oriens-Occidens,
3, 2000, pp. 189–234. 51
Bibliographie
363
[17] Chistov A. Fast parallel calculation of the rank of matrices
over a field of arbitrary characteristic, Proc. FCT ’85, Springer
Lecture Notes in Computer Science no 199, 1985, pp. 147–150.
93
[18] Coppersmith D. & Winograd S. On the asymptotic complexity of matrix multiplication, SIAM J. Comput., 11/3, 1982,
pp. 472–492. i, 269
[19] Coppersmith D. & Winograd S. Matrix multiplication via
arithmetic progressions, Proc. 19th Ann. ACM Symp. on the
Theory of Computing, 1987, pp. 1–6. i, 187, 268
[20] Coppersmith D. & Winograd S. Matrix multiplication via
arithmetic progressions, Proc. J. of Symbolic Computation, 9,
1990, pp. 251–280. 187
[21] Cook S. The complexity of theorem proving procedures, Proc.
3rd Ann. ACM Symp. on the Theory of Computing, 1971,
pp. 151–158. 135, 138
[22] Csanky L. Fast parallel inversion algorithms, SIAM J. Comput., 5/4, 1976, pp. 618–623. 83, 255
[23] Dantzig G. Maximization of a linear function of variables
subject to linear inequalities, in : Activity Analysis of Production and Allocation (Koopmans T. ed.) Wiley New-York, 1951,
pp. 359–373. 136
[24] Dı́az-Toca G., Gonzalez Vega L. & Lombardi H. Generalizing Cramer’s Rule : Solving uniformly linear systems of
equations, SIAM Journal on Matrix Analysis and Applications.
27 no 3 (2005), 621–637. 34
[25] Dı́az-Toca G., Gonzalez Vega L., Lombardi H. &
Quitté C. Modules projectifs de type fini, applications linéaires croisées et inverses généralisés, Journal of Algebra. 303
2 (2006), 450–475. 49
[26] Dodgson C. Condensation of determinants, being a new and
brief method for computing their arithmetic values, Proc. Royal
Soc. Lond., 15, 1866, pp. 150-155. 66
[27] Dornstetter J. On the equivalence between Berlekamp’s and
Euclid’s algorihms, IEEE Trans. Inform. Theory, 33/3, 1987,
pp. 428–431. 108, 246
364
Bibliographie
[28] Della Dora J., Dicrescenzo C. & Duval D. About a new
method for computing in algebraic number fields, Eurocal ’85,
vol 2, Springer Lecture Notes in Computer Science 204, 1985,
pp. 289–290. 287
[29] Eberly W. Very fast parallel matrix and polynomial arithmetic, Technical Report # 178/85, PhD Thesis, University of Toronto, Canada, 1985. 273
[30] Fich F., von zur Gathen J. & Rackoff C. Complete families of polynomials, Manuscript, 1986. 319
[31] Frame J.S. A simple recurrent formula for inverting a matrix,
(abstract) Bull. Amer. Math. Soc., 55, 1949, p. 1045. 84
[32] Galil N. & Pan V. Parallel evaluation of the determinant and
of the inverse of a matrix, Information Processing Letters, 30,
1989, pp. 41–45. 83, 255, 266
[33] Gastinel N. Sur le calcul de produits de matrices, Num.
Math., 17, 1971, pp. 222–229. 199
[34] von zur Gathen J. Parallel arithmetic computations : a survey, in Proc. 12th. Internat. Symp. on Mathematical Foundations of Computer Science, Lecture Notes in Computer Science,
no 233, Springer, Berlin, 1986, pp. 93–113. 151
[35] von zur Gathen J. Feasible Arithmetic Computations : Valiant’s Hypothesis, Journal of Symbolic Computation, 4, 1987,
pp. 137–172. 303
[36] von zur Gathen J. Parallel Linear Algebra, in Synthesis of
parallel algorithms (J. H. Reif Editor), Morgan Kaufmann publishers, San Mateo, Californie, 1993, pp. 573–617. 152, 153
[37] Giesbrecht, M. Fast algorithms for rational forms of integer
matrices, in International Symposium on Symbolic and Algebraic Computation, Oxford, UK, ACM Press, 1994, pp. 305–
317. 245
[38] Giesbrecht, M. Nearly optimal algorithms for canonical matrix forms, SIAM Journal on Computing, 24/5, 1995, pp. 948–
969. 245
[39] Giusti M. & Heintz J. La détermination des points isolés et
de la dimension d’une variété algébrique peut se faire en temps
polynomial, in Computational Algebraic Geometry & Commutative Algebra, eds : Eisenbud, Robbiano. Cambridge University Press, 1993, pp. 216–256. 144
Bibliographie
365
[40] Giusti M., Heintz J, Morais J.-E., Morgenstern, J. &
Pardo, L.-M. Straight-line programs in geometric elimination
theory, J. Pure Appl. Algebra, 124/1-3, 1998, 1, pp. 101–146.
144
[41] Gonzalez Vega L., Lombardi H., Recio T. & Roy M.F. Spécialisation de la suite de Sturm et sous-résultants, Informatique Théorique et Applications (R.A.I.R.O.), 24, 1990,
pp. 561–588. 288
[42] González-Vega L., Rouillier F., Roy M.-F. & Trujillo
G. Symbolic Recipes for Real Solutions, pp. 121-167 dans [CCS].
288
[43] Håstad J. Tensor rank is N P-complete, in Automata, Languages and Programming Proc. of the 16th International Colloquium, Springer Lecture Notes in Computer Science, 372,
1989, pp. 451–460. 200
[44] Heintz J. & Sieveking M. Lower bounds for polynomials with
algebraic coefficients, Theoretical Computer Science 11, 1980,
pp. 321–330. 315
[45] Hoover J. Feasible real functions and arithmetic circuits,
SIAM J. Comput., 19/1, 1990, pp. 182–204. 148
[46] Hopcroft J. & Musinski Duality applied to the complexity of
matrix multiplication and other bilinear forms, SIAM J. Comp.,
2, 1973, pp. 159–173. 202
[47] Hyafil L. On the parallel evaluation of multivariate polynomials, SIAM J. Comput., 8, 1979, pp. 120–123. 312
[48] Kakeya R. On fundamental systems of symmetric functions,
Jap. J. of Math., 2, 1927, pp. 69–80. 258
[49] Kaltofen E. On computing determinants of matrices without
divisions, I.S.S.A.C.’92, ACM, 1992, pp. 342–349. 246, 248
[50] Kaltofen E. & Pan V. Processor efficient parallel solution of
linear systems over an abstract field, in Proc. 3rd Ann. Parallel
Algo. Architectures, ACM Press (July 1991), pp. 180–191. 155,
253
[51] Kaltofen E. & Pan V. Processor efficient parallel solution
of linear systems II : the general case, in Proc. 33rd IEEE
Symp. Foundations Of Computer Science, Pittsburg, USA,
1992, pp. 714–723. 155
366
Bibliographie
[52] Kaltofen E. & Singer M. Size efficient parallel algebraic
circuits for partial derivatives, in Proc. IV Intern. Conf. on
Computer Algebra in Physical Research. World Scientific, Singapour, 1991, pp. 113–145. 281
[53] Kaltofen E. & Villard G. On the complexity of computing
determinants, In Proc. Fifth Asian Symposium on Computer
Mathematics (ASCM 2001) pp. 13–27. 254
[54] Kaltofen E. & Villard G. Computing the sign or the value
of the determinant of an integer matrix, a complexity survey,
à paraı̂tre dans Computational Applied Math. Special issue on
Congrès International Algèbre Linéaire et Arithmétique : Calcul
Numérique, Symbolique et Parallèle, Rabat, Maroc, May 2001.
17 pages. 254
[55] Karmakar N. A new polynomial time algorithm for linear
programming, Combinatorica, 4/4, 1984, pp. 373-395. 136
[56] Karp R., Lipton R. Turing Machines that take advice, Logic
and arithmetic, int. Symp., Zurich 1980, Monogr. L’Enseign.
Math. 30, 1982, pp. 255–273,. 324
[57] Karp R. & Ramachandran V. Parallel algorithms for
shared-memory machines, Handbook of Theoretical Computer
Science, Edited by J. van Leeuwen, Elsevier Science Publishers
B.V., 1990, pp. 871–941. 152, 153
[58] Keller-Gehrig W. Fast algorithms for the characteristic polynomial, Theoretical Computer Science, 36, 1985, pp. 309–317.
237, 243
[59] Ker-I. KO & Friedman H. Computational complexity of real
functions, Theoretical Computer Science, 20, 1982, pp. 323–
352. 148
[60] Khachiyan L. Polynomial algorithms in linear programming,
Soviet Math. Doklady, 20, 1979, pp. 191–194. 136
[61] Khachiyan L. A polynomial algorithm in linear programming,
Computational Mathematics and Mathematical Physics, 20,
1980, pp. 53–72. 136
[62] Kruskal C., Rudolph L. & Snir M. A complexity theory of
efficient parallel algorithms, Theoretical Computer Science, 71,
1990, pp. 95–132. 155
Bibliographie
367
[63] Labhalla S., Lombardi H. & Moutai E. Espaces métriques
rationnellement présentés et complexité, le cas de l’espace des
fonctions réelles uniformément continues sur un intervalle compact, Theoretical Computer Science, 250/1-2, 2001, pp. 265–
332. 148
[64] Ladner R. & Fischer M. Parallel Prefix Computation, Journal of the ACM, 27/4, 1980, pp. 831–838. 159, 165, 170, 280
[65] Le Verrier U. Sur les variations séculaires des éléments elliptiques des sept planètes principales : Mercure, Vénus, La Terre,
Mars, Jupiter, Saturne et Uranus, J. Math. Pures Appli., 4,
1840, pp. 220–254. 82
[66] Lickteig T., Roy M.-F. Sylvester-Habicht sequences and fast
Cauchy index computation, Journal of Symbolic Computation,
31, 2001, pp. 315–341. 253
[67] Lombardi H., Roy M.-F., Safey El Din M. New structure
theorems for subresultants, Journal of Symbolic Computation
29/3-4, 2000, pp. 663–690. 288
[68] R. Loos, Generalized polynomial remainder sequences, in :
Computer algebra, symbolic and algebraic computation,
Springer-Verlag, Berlin, 1982. 288
[69] Matera G. & Turull Torres J.-M. The space complexity of
elimination : upper bounds, Foundations of computational mathematics (Rio de Janeiro, 1997), pp. 267–276, Springer, Berlin.
282
[70] Miller G., Ramachandran V., Kaltofen E. Efficient Parallel Evaluation of Straight-Line Code and Arithmetic Circuits,
SIAM J. Comput. 17/4, 1988, pp. 687–695.
[71] Moenck Fast computation of GCDs, Proc. STOC’73, pp. 142–
151. 253
[72] Morgenstern J. How to compute fast a function and all its
derivatives (a variation on the theorem of Baur-Strassen), SIGACT News, 16/4, 1985, pp. 60–62. 126
[73] Mulmuley K. A fast parallel algorithm to compute the rank
of a matrix over an arbitrary field, Combinatorica, 7/1, 1987,
pp. 101–104. 34, 47
[74] Ostrowski A. On two problems in abstract algebra connected
with Horner’s rule, Studies in Math. and Mec. presented to
Richard von Mises, Academic Press N.Y, 1954, pp. 40–48.
368
Bibliographie
[75] Pan V. Computation schemes for a product of matrices and for
the inverse matrix, Uspehi Mat. Nauk, 5/167, 1972, pp. 249–
250. 202
[76] Prasad K., Bapat R. The generalized Moore-Penrose inverse.
Linear Algebra Appl. 165, 1992, 59–69. 49
[77] Preparata F. & Sarwate D. An improved parallel processor
bound in fast matrix inversion, Inf. Proc. Letters, 7/3, 1978,
pp. 148–150. 83, 255, 260
[78] Revol N. Complexité de l’évaluation parallèle de circuits arithmétiques, Thèse de l’Institut National Polytechnique de Grenoble, Août 1994.
[79] Samuelson P. A method for determining explicitely the characteristic equation, Ann. Math. Statist., 13, 1942, pp. 424–429.
ii, 13, 90, 272
[80] Sasaki T. & Murao H. Efficient Gaussian elimination method for symbolic determinants and linear systems, ACM Trans.
Math. Software, 8/4, 1982, pp. 277–289. 70
[81] Schönhage R. Fast parallel computation of characteristic polynomials by Le Verrier’s power sum method adapted to fields of
finite characteristic, Proc. 20th ICALP, Lecture Notes in Computer Science no 700, Springer, 1993, pp. 410-417. 258, 259
[82] Schönhage R. Partial and total matrix multiplication, Siam
J. Comp., 10, 1981, pp. 434–445. i, 208, 215, 221
[83] Skyum S. & Valiant L. A complexity theory based on Boolean
algebra, J. Assoc. Comp. Mach., 32/2, 1985, pp. 484–502. 321
[84] Souriau J.-M. Une méthode pour la décomposition spectrale
et l’inversion des matrices, C.R. Acad. Sciences, 227, 1948,
pp. 1010–1011. 83, 84
[85] Spira P. On Time hardware Complexity Tradeoffs for Boolean
Functions, Proceedings of Fouth Hawaii International Symposium on System Sciences, 1971, pp. 525–527. 325
[86] Strassen V. Gaussian elimination is not optimal, Numerische
Mathematik 13, 1969, pp. 354–356. i, 186, 191
[87] Strassen V. Vermeidung von divisionen, Crelle J. Reine Angew. Math., 264, 1973, pp. 184–202. 120, 199
[88] Strassen V. Polynomials with rational coefficients which are
hard to compute, Siam J. Comp., 3, 1974, pp. 128–149. 315
Bibliographie
369
[89] Strassen V. The work of Valiant, Proc. International
Congress of Mathematicians, Berkeley, USA, 1986.
[90] Strassen V. Relative bilinear complexity and matrix multiplication, Crelle J. Reine Angew. Math., 375/376, 1987, pp. 406–
443. i, 187, 199
[91] Strassen V. Algebraic complexity theory, in Handbook of
Theoretical Computer Science (ed. J. van Leeuwen) Vol A.
Chap. 11, Elsevier Science Publishers B.V., 1990, pp. 634–672.
199
[92] Valiant L. Completeness classes in algebra, in : Proc. 11th
ACM STOC, 1979, pp. 249–261. 303
[93] Valiant L. The complexity of computing the permanent, Theoretical Computer Science 8, 1979, pp. 189–201. 141, 303
[94] Valiant L. Reducibility by algebraic projections, in : Logic and
Algorithmic, Symposium in honour of Ernst Specker, Enseign.
Math 30, 1982, pp. 365–380. 303
[95] Valiant L., Skyum S., Berkowitz S. & Rackoff C. Fast
parallel computation of polynomials using few processors, SIAM
J. Comput., 12/4, 1983, pp. 641–644. 271, 312
[96] Wiedemann D. Solving sparse linear equations over finite
fields, I.E.E.E. Trans. Inf. Theory 32, 1986, pp. 54–62. 109,
110
[97] Winograd S. On the multiplication of 2 by 2 matrices, Linear
Algebra Appl., 4, 1971, pp. 381–388. 186
Index des termes
affectation sans scalaires, 116
A – module, 5
anneau des développements limités à l’ordre n, 94, 121
anneau des polynômes non commutatifs, 199
application
bilinéaire, 197
booléenne, 322
quadratique, 205
approximation d’ordre q d’une
application bilinéaire, 211
centre d’élimination des divisions,
122
Chistov
algorithme de, 94
circuit
arithmétique, 116
avec divisions, 116
homogène, 118
sans division, 116
booléen, 118, 322
code d’un k-uple d’entiers, 139
coefficient de Gram
d’une matrice, 37
généralisé, 45
comatrice, 2
compagnon
matrice, 16
complexité
arithmétique, 142
arithmétique d’une famille de
circuits , 142
bilinéaire, 199
binaire, 143
binaire d’une famille de circuits , 144
multiplicative, 205
conjecture addititve, 224
Cramer, 7
identifés de, 7
Berkowitz
algorithme de, 91
Berlekamp/Massey
algorithme de, 108
bien parallélisé
algorithme, 147
bilinéaire
application, 197
calcul, 199
complexité, 199
binaire
complexité, 144
calcul
bilinéaire, 199
bilinéaire approximatif, 211
quadratique, 205
Cayley-Hamilton
théorème de, 12
degré de parallélisme, 160
description
371
372
d’un polynôme, 308
d’une fonction booléenne, 323
instantanée, 139
Dodgson-Jordan-Bareiss
formule de, 67
élémentaire
matrice, 54
transformation, 54
élimination des divisions (Strassen), 122
éliminer les divisions
à la Strassen dans un circuit
arithmétique, 122
entier bâton, 139
essentielle
multiplication, 116
exposant
acceptable pour la multiplication des matrices carrées , 204
de la multiplication des matrices carrées , 204
expression
arithmétique, 304
booléenne, 322
famille uniforme
de circuits arithmétiques, 149
de circuits booléens, 150
fonction booléenne, 322
forme algébrique déployée
d’une famille de fonctions booléennes, 334
d’une fonction booléenne, 333
forme normale
conjonctive, 322
disjonctive, 322
formule
de Binet-Cauchy, 4
Index des termes
de Samuelson, 13
Frobenius
algorithme de, 100
matrice de, 100
Gauss
méthode du pivot, 53
Gram
coefficient de, 37
coefficient généralisé de, 45
polynôme de, 45
Hadamard
inégalité de, 31
Hankel
matrice de, 23
Hessenberg
algorithme de, 76
Horner
expression de, 305
parallélisation de l’expression,
312
schéma de, 12
inverse de Moore-Penrose
en rang r, 38
inverse Moore-Penrose
généralisé en rang r, 47
Jordan-Bareiss
méthode de, 65
méthode modifiée , 70
Kakeya
critère de, 258
Krylov
sous-espace de, 17
Le Verrier
algorithme de, 83
hypothèses de l’algorithme de,
89
Index des termes
littéral, 322
longueur
d’un programme d’évaluation,
115
multiplicative, 116
stricte, 116
LU -décomposition, 56
LUP -décomposition , 63
matrice
adjointe, 2
caractéristique, 11
caractéristique adjointe, 11
compagnon d’un polynôme
unitaire, 16
de Frobenius, 100
de Hankel, 23
de Toeplitz, 23
élémentaire , 54
fortement régulière, 5
régulière, 5
rang d’une, 5
singulière, 5
structurée, 23
trace d’une, 2
triangulaire, 53
unimodulaire, 55
unitriangulaire, 54
mineur
connexe, 71
d’une matrice, 2
principal, 3
module
libre, 6
sur un anneau, 5
Moore-Penrose
inverse de, 38
multiplicité
algébrique, 14
géométrique, 86
373
Newton
relation de, 28
somme de, 26
non déterministe, 137
N P - complet , 138
p-bornée, 307, 323, 326
famille d’applications booléennes, 326
famille d’expressions arithmétiques,
307
famille d’expressions booléennes,
323
famille de circuits arithmétiques,
307
famille de circuits booléens,
323
famille de fonctions booléennes, 323
famille de polynômes, 307
p-calculable, 307, 323
famille de fonctions booléennes, 323
famille de polynômes, 307
p-descriptible, 308, 323
en expressions, 308, 324
p-évaluable, 307
p-exprimable, 307, 323
famille de fonctions booléennes, 323
famille de polynômes, 307
p-famille
d’applications booléennes, 326
de fonctions booléennes, 323
de polynômes, 307
p-projection, 316
paramètre d’entrée
d’un programme d’évaluation,
115
P̃LUP-décomposition , 63
374
polylogarithmique, 147
polynôme
caractéristique, 11
cyclotomique, 181
générateur d’une suite, 21
générateur minimal d’une suite
récurrente linéaire, 22
minimal, 17
minimal d’une suite récurrente
linéaire, 22
réciproque, 95
polynôme de Gram
(généralisé), 45
Preparata&Sarwate
algorithme de, 89
produit matriciel à trous, 209
profondeur
d’un programme d’évaluation,
115
d’une expression booléenne,
322
multiplicative, 116
stricte, 116
programme d’évaluation
arithmétique, 114
booléen, 115
largeur d’un, 115
longueur d’un, 115
paramètre d’entrée d’un, 115
profondeur d’un, 115
taille d’un, 115
variable d’affectation d’un, 114
projection, 316
d’une expression, 316
qp-calculable, 307
qp-évaluable, 307
qp-exprimable, 307
qp-projection, 316
quadratique
Index des termes
application, 205
calcul, 205
quasi-polynomial, 306
racine de l’unité
primitive n - ème, 175
principale n - ème, 175
rang marginal, 211
rang tensoriel
d’une application bilinéaire,
199
marginal, 211
relation de Newton, 28
représentation
creuse, 115
dense, 115
par expressions, 305
Samuelson, 13
Schönhage
inégalité asymptotique, 225
variante de la méthode de
Csanky, 258
simulation
en évaluation, 330
somme de Newton, 26
somme directe (de deux applications bilinéaires), 221
somme disjointe (de deux applications bilinéaires), 221
Souriau-Faddeev-Frame
algorithme de, 85
sous-matrice
principale, 3
principale dominante, 3
suite récurrente linéaire, 21
Sylvester
identités de, 9
système fondamental
de polynômes symétriques, 27
Index des termes
de fractions rationnelles symétriques, 27
taille
d’un programme d’évaluation,
115
d’une expression artithmétique,
304
d’une expression booléenne,
322
tenseur
coordonnées d’un, 199
élémentaire, 198
Théorème chinois, 32
Toeplitz
matrice de, 23
transformation élémentaire, 54
unimodulaire, 55
triangulaire
matrice, 53
unimodulaire
transformation élémentaire,
55
unitriangulaire
matrice, 54
valeur propre, 14
multiplicité algébrique d’une,
14
multiplicité géométrique d’une,
86
Wiedemann
algorithme de, 109
375
| 0math.AC
|
1
Coded Caching Schemes with Reduced
Subpacketization from Linear Block Codes
arXiv:1706.00101v3 [cs.IT] 19 Feb 2018
Li Tang and Aditya Ramamoorthy
Department of Electrical and Computer Engineering
Iowa State University
Ames, IA 50010
Emails:{litang, adityar}@iastate.edu
Abstract—Coded caching is a technique that generalizes conventional caching and promises significant reductions in traffic
over caching networks. However, the basic coded caching scheme
requires that each file hosted in the server be partitioned into a
large number (i.e., the subpacketization level) of non-overlapping
subfiles. From a practical perspective, this is problematic as it
means that prior schemes are only applicable when the size of the
files is extremely large. In this work, we propose coded caching
schemes based on combinatorial structures called resolvable
designs. These structures can be obtained in a natural manner
from linear block codes whose generator matrices possess certain
rank properties. We obtain several schemes with subpacketization
levels substantially lower than the basic scheme at the cost
of an increased rate. Depending on the system parameters,
our approach allows us to operate at various points on the
subpacketization level vs. rate tradeoff.
Index Terms—coded caching, resolvable designs, cyclic codes,
subpacketization level
I. I NTRODUCTION
Caching is a popular technique for facilitating large scale
content delivery over the Internet. Traditionally, caching operates by storing popular content closer to the end users.
Typically, the cache serves an end user’s file request partially
(or sometimes entirely) with the remainder of the content
coming from the main server. Prior work in this area [1]
demonstrates that allowing coding in the cache and coded
transmission from the server (referred to as coded caching)
to the end users can allow for significant reductions in the
number of bits transmitted from the server to the end users.
This is an exciting development given the central role of
caching in supporting a significant fraction of Internet traffic.
In particular, reference [1] considers a scenario where a single
server contains N files. The server connects to K users over
a shared link and each user has a cache that allows it to
store M/N fraction of all the files in the server. Coded
caching consists of two distinct phases: a placement phase
and a delivery phase. In the placement phase, the caches of the
users are populated. This phase does not depend on the user
demands which are assumed to be arbitrary. In the delivery
This work was supported in part by the National Science Foundation
by grants CCF-1718470, CCF-1320416 and CCF-1149860. This paper was
presented in part at the 2016 IEEE Workshop on Network Coding and
Applications (NetCod) and at the 2017 IEEE International Symposium on
Information Theory (ISIT).
phase, the server sends a set of coded signals that are broadcast
to each user such that each user’s demand is satisfied.
The original work of [1] considered the case of centralized
coded caching, where the server decides the content that
needs to be placed in the caches of the different users.
Subsequent work considered the decentralized case where the
users populate their caches by randomly choosing parts of each
file while respecting the cache size constraint. Recently, there
have been several papers that have examined various facets
of coded caching. These include tightening known bounds
on the coded caching rate [2], [3], considering issues with
respect to decentralized caching [4], explicitly considering
popularities of files [5], [6], network topology issues [7], [8]
and synchronization issues [9], [10].
In this work, we examine another important aspect of the
coded caching problem that is closely tied to its adoption in
practice. It is important to note that the huge gains of coded
K
caching require each file to be partitioned into Fs ≈ KM
N
non-overlapping subfiles of equal size; Fs is referred to as the
subpacketization level. It can be observed that for a fixed cache
size M
N , Fs grows exponentially with K. This can be problematic in practical implementations. For instance,
suppose
64
that K = 64, with M
=
0.25
so
that
F
=
≈
4.8
× 1014
s
N
16
with a rate R ≈ 2.82. In this case, it is evident that at the bare
minimum, the size of each file has to be at least 480 terabits
for leveraging the gains in [1]. It is even worse in practice. The
atomic unit of storage on present day hard drives is a sector
of size 512 bytes and the trend in the disk drive industry is
to move this to 4096 bytes [11]. As a result, the minimum
size of each file needs to be much higher than 480 terabits.
Therefore, the scheme in [1] is not practical even for moderate
values of K. Furthermore, even for smaller values of K,
schemes with low subpacketization levels are desirable. This is
because any practical scheme will require each of the subfiles
to have some header information that allows for decoding at
the end users. When there are a large number of subfiles,
the header overhead may be non-negligible. For these same
parameters (K = 64, M/N = 0.25) our proposed approach in
this work allows us obtain, e.g., the following operating points:
(i) Fs ≈ 1.07×109 and R = 3, (ii) Fs ≈ 1.6×104 and R = 6,
(iii) Fs = 64 and R = 12. For the first point, it is evident
that the subpacketization level drops by over five orders of
magnitude with only a very small increase in the rate. Point
(ii) and (iii) show proposed scheme allows us to operate at
various points on the tradeoff between subpacketization level
and rate.
The issue of subpacketization was first considered in the
work of [12], [13] in the decentralized coded caching setting.
In the centralized case it was considered in the work of
[14]. They proposed a low subpacketization scheme based on
placement delivery arrays. Reference [15] viewed the problem
from a hypergraph perspective and presented several classes
of coded caching schemes. The work of [16] has recently
shown that there exist coded caching schemes where the
subpacketization level grows linearly with the number of users
K; however, this result only applies when the number of users
is very large. We elaborate on related work in Section II-A.
In this work, we propose low subpacketization level
schemes for coded caching. Our proposed schemes leverage
the properties of combinatorial structures known as resolvable
designs and their natural relationship with linear block codes.
Our schemes are applicable for a wide variety of parameter
ranges and allow the system designer to tune the subpacketization level and the gain of the system with respect to an
uncoded system. We note here that designs have also been used
to obtain results in distributed data storage [17] and network
coding based function computation in recent work [18], [19].
This paper is organized as follows. Section II discusses
the background and related work and summarizes the main
contributions of our work. Section III outlines our proposed
scheme. It includes all the constructions and the essential
proofs. A central object of study in our work are matrices
that satisfy a property that we call the consecutive column
property (CCP). Section IV overviews several constructions
of matrices that satisfy this property. Several of the longer
and more involved proofs of statements in Sections III and
IV appear in the Appendix. In Section V we perform an
in-depth comparison our work with existing constructions in
the literature. We conclude the paper with a discussion of
opportunities for future work in Section VI.
columns correspond to the points and blocks respectively. Let
i ∈ X and j ∈ A. Then,
(
1 if i ∈ j,
N (i, j) =
0 otherwise.
It can be observed that the transpose of an incidence matrix
also specifies a design. We will refer to this as the transposed
design. In this work, we will utilize resolvable designs which
are a special class of designs.
Definition 3. A parallel class P in a design (X, A) is a subset
of disjoint blocks from A whose union is X. A partition of A
into several parallel classes is called a resolution, and (X, A)
is said to be a resolvable design if A has at least one resolution.
For resolvable designs, it follows that each point also
appears in the same number of blocks.
Example 1. Consider a block design specified as follows.
X = {1, 2, 3, 4}, and
A = {{1, 2}, {1, 3}, {1, 4}, {2, 3}, {2, 4}, {3, 4}}.
Its incidence matrix is given
1 1
1 0
N =
0 1
0 0
below.
1
0
0
1
0
1
1
0
0
1
0
1
0
0
.
1
1
It can be observed that this design is resolvable with the
following parallel classes.
P1 = {{1, 2}, {3, 4}},
P2 = {{1, 3}, {2, 4}}, and
P3 = {{1, 4}, {2, 3}}.
In the sequel we let [n] denote the set {1, . . . , n}. We
emphasize here that the original scheme of [1] can be viewed
as an instance of the trivial design. For example, consider the
setting when t = KM/N is an integer. Let X = [K] and
A = {B : B ⊂ [K], |B| = t}. In the scheme of [1], the users
are associated with X and the subfiles with A. User i ∈ [K]
caches subfile Wn,B , n ∈ [N ] for B ∈ A if i ∈ B. The main
message of our work is that carefully constructed resolvable
designs can be used to obtain coded caching schemes with
low subpacketization levels, while retaining much of the rate
gains of coded caching. The basic idea is to associate the users
with the blocks and the subfiles with the points of the design.
The roles of the users and subfiles can also be interchanged
by simply working with the transposed design.
II. BACKGROUND , R ELATED W ORK AND S UMMARY OF
C ONTRIBUTIONS
We consider a scenario where the server has N files each
of which consist of Fs subfiles. There are K users each
equipped with a cache of size M Fs subfiles. The coded
caching scheme is specified by means of the placement scheme
and an appropriate delivery scheme for each possible demand
pattern. In this work, we use combinatorial designs [20] to
specify the placement scheme in the coded caching system.
Definition 1. A design is a pair (X, A) such that
Example 2. Consider the resolvable design from Example
1. The blocks in A correspond to six users U12 , U34 , U13 ,
U24 , U14 , U23 . Each file is partitioned into Fs = 4 subfiles
Wn,1 , Wn,2 , Wn,3 , Wn,4 which correspond to the four points
in X. The cache in user UB , denoted ZB is specified as Zij =
N
(Wn,i , Wn,j )N
n=1 . For example, Z12 = (Wn,1 , Wn,2 )n=1 .
We note here that the caching scheme is symmetric with
respect to the files in the server. Furthermore, each user caches
half of each file so that M/N = 1/2. Suppose that in the
delivery phase user UB requests file WdB where dB ∈ [N ].
1) X is a set of elements called points, and
2) A is a collection of nonempty subsets of X called
blocks, where each block contains the same number of
points.
A design is in one-to-one correspondence with an incidence
matrix N which is defined as follows.
Definition 2. The incidence matrix N of a design (X, A) is
a binary matrix of dimension |X| × |A|, where the rows and
2
the rate benefits of coded caching would be lost if Fs did not
scale exponentially in K. Following this work, the authors in
[14] introduced a technique for designing low subpacketization
schemes in the centralized setting which they called placement
delivery arrays. In [14], they considered the setting when
M/N = 1/q or M/N = 1 − 1/q and demonstrated a scheme
where the subpacketization level was exponentially smaller
than the original scheme, while the rate was marginally higher.
This scheme can be viewed as a special case of our work. We
discuss these aspects in more detail in Section V. In [15],
the design of coded caching schemes was achieved through
the design of hypergraphs with appropriate properties. In
particular, for specific problem parameters, they were able to
establish the existence
of schemes where the subpacketization
√
scaled as exp(c K). Reference [21] presented results in this
setting by considering strong edge coloring of bipartite graphs.
Very recently, [16] showed the existence of coded caching
schemes where the subpacketization grows linearly with the
number of users, but the coded caching rate grows as O(K δ )
where 0 < δ < 1. Thus, while the rate is not a constant, it
does not grow linearly with K either. Both [15] and [16] are
interesting results that demonstrate the existence of regimes
where the subpacketization scales in a manageable manner.
Nevertheless, it is to be noted that these results come with
several caveats. For example, the result of [16] is only valid in
the regime when K is very large and is unlikely to be of use
for practical values of K. The result of [15] has significant
restrictions on the number of users, e.g., in their paper, K
needs to be of the form na and q t na .
These demands can be satisfied as follows. We pick three
blocks, one each from parallel classes P1 , P2 , P3 and generate
the signals transmitted in the delivery phase as follows.
Wd12 ,3 ⊕ Wd13 ,2 ⊕ Wd23 ,1 ,
(1)
Wd12 ,4 ⊕ Wd24 ,1 ⊕ Wd14 ,2 ,
Wd34 ,1 ⊕ Wd13 ,4 ⊕ Wd14 ,3 , and
Wd34 ,2 ⊕ Wd24 ,3 ⊕ Wd23 ,4 .
The three terms in the in eq. (1) above correspond to
blocks from different parallel classes {1, 2} ∈ P1 , {1, 3} ∈
P2 , {2, 3} ∈ P3 . This equation has the all-but-one structure
that was also exploited in [1], i.e., eq. (1) is such that each
user caches all but one of the subfiles participating in the
equation. Specifically, user U12 contains Wn,1 and Wn,2 for
all n ∈ [N ]. Thus, it can decode subfile Wd12 ,3 that it needs.
A similar argument applies to users U13 and U23 . It can be
verified that the other three equations also have this property.
Thus, at the end of the delivery phase, each user obtains its
missing subfiles.
This scheme corresponds to a subpacketization level of 4
and a rate of 1. In contrast, the scheme of [1] would require
a subpacketization level of 63 = 20 with a rate of 0.75.
Thus, it is evident that we gain significantly in terms of the
subpacketization while sacrificing some rate gains.
As shown in Example 2, we can obtain a scheme by
associating the users with the blocks and the subfiles with
the points. In this work, we demonstrate that this basic idea
can be significantly generalized and several schemes with low
subpacketization levels that continue to leverage much of the
rate benefits of coded caching can be obtained.
B. Summary of Contributions
In this work, the subpacketization levels we obtain are typically exponentially smaller than the original scheme. However,
they still continue to scale exponentially in K, albeit with
much smaller exponents. However, our construction has the
advantage of being applicable for a large range of problem
parameters. Our specific contributions include the following.
• We uncover a simple and natural relationship between a
(n, k) linear block code and a coded caching scheme. We
first show that any linear block code over GF (q) and in
some cases Z mod q (where q is not a prime or a prime
power) generates a resolvable design. This design in turn
specifies a coded caching scheme with K = nq users
where the cache fraction M/N = 1/q. A complementary
cache fraction point where M/N = 1 − α/nq where α is
some integer between 1 and k + 1 can also be obtained.
Intermediate points can be obtained by memory sharing
between these points.
• We consider a class of (n, k) linear block codes whose
generator matrices satisfy a specific rank property. In
particular, we require collections of consecutive columns
to have certain rank properties. For such codes, we are
able to identify an efficient delivery phase and determine
the precise coded caching rate. We demonstrate that the
subpacketization level is at most q k (k + 1) whereas the
coded caching gain scales as k + 1 with respect to an
uncoded caching scheme. Thus, different choices of k
A. Discussion of Related Work
Coded caching has been the subject of much investigation in recent work as discussed briefly earlier on. We now
overview existing literature on the topic of low subpacketization schemes for coded caching. In the original paper [1],
for given problem parameters K (number of users) and M/N
(cache fraction), the authors showed that when N ≥ K, the
rate equals
R=
K(1 − M/N )
1 + KM/N
when M is an integer multiple of N/K. Other points are
obtained via memory sharing. Thus, in the regime when
KM/N is large, the coded caching rate is approximately
N/M − 1, which is independent of K. Crucially, though this
K
requires the subpacketization level Fs ≈ KM/N
. It can be
observed that for a fixed M/N , Fs grows exponentially with
K. This is one of main drawbacks of the original scheme and
for reasons outlined in Section I, deploying this solution in
practice may be difficult.
The subpacketization issue was first discussed in the work
of [12], [13] in the context of decentralized caching. Specifically, [13] showed that in the decentralized setting for any
subpacketization level Fs such that Fs ≤ exp(KM/N ) the
rate would scale linearly in K, i.e., R ≥ cK. Thus, much of
3
•
where u = [u0 , · · · , uk−1 ]. Let a∗ be such that ga∗ b 6= 0.
Consider the equation
X
ua gab = ∆b − ua∗ ga∗ b ,
allow the system designer significant flexibility to choose
the appropriate operating point.
We discuss several constructions of generator matrices
that satisfy the required rank property. We characterize
the ranges of alphabet sizes (q) over which these matrices
can be constructed. If one has a given subpacketization
budget in a specific setting, we are able to find a set of
schemes that fit the budget while leveraging the rate gains
of coded caching.
a6=a∗
where ∆b is fixed. For arbitrary values of ua , a 6= a∗ , this
equation has a unique solution for ua∗ , which implies that for
any ∆b , |Bb,∆b | = q k−1 and that Pb forms a parallel class.
Remark 1. A k × n generator matrix over GF (q) where q
is a prime power can also be considered as a matrix over
an extension field GF (q m ) where m is an integer. Thus,
one can obtain a resolvable design in this case as well; the
corresponding parameters can be calculated in an easy manner.
III. P ROPOSED LOW SUBPACKETIZATION LEVEL SCHEME
All our constructions of low subpacketization schemes will
stem from resolvable designs (cf. Definition 3). Our overall
approach is to first show that any (n, k) linear block code
over GF (q) can be used to obtain a resolvable block design.
The placement scheme obtained from this resolvable design
is such that M/N = 1/q. Under certain (mild) conditions on
the generator matrix we show that a delivery phase scheme
can be designed that allows for a significant rate gain over the
uncoded scheme while having a subpacketization level that
is significantly lower than [1]. Furthermore, our scheme can
be transformed into another scheme that operates at the point
M/N = 1 − k+1
nq . Thus, intermediate values of M/N can
be obtained via memory sharing. We also discuss situations
under which we can operate over modular arithmetic Zq = Z
mod q where q is not necessarily a prime or a prime power;
this allows us to obtain a larger range of parameters.
Remark 2. We can also consider linear block codes over Z
mod q where q is not necessarily a prime or a prime power.
In this case the conditions under which a resolvable design
can be obtained by forming the matrix T are a little more
involved. We discuss this in Lemma 4 in the Appendix.
Example 3. Consider a (4, 2) linear
with generator matrix
1 0 1
G=
0 1 1
A ={{0, 1, 2}, {3, 4, 5}, {6, 7, 8}, {0, 3, 6}, {1, 4, 7}, {2, 5, 8},
(2)
{0, 5, 7}, {1, 3, 8}, {2, 4, 6}, {0, 4, 8}, {2, 3, 7}, {1, 5, 6}}.
where the 1 × n vector c` represents the `-th codeword of
the code. Let X = {0, 1, · · · , q k − 1} be the point set and A
be the collection of all subsets Bi,l for 0 ≤ i ≤ n − 1 and
0 ≤ l ≤ q − 1, where
It can be observed that A has a resolution (cf. Definition 3)
with the following parallel classes.
P0 = {{0, 1, 2}, {3, 4, 5}, {6, 7, 8}},
P1 = {{0, 3, 6}, {1, 4, 7}, {2, 5, 8}},
Bi,l = {j : Ti,j = l}.
P2 = {{0, 5, 7}, {1, 3, 8}, {2, 4, 6}}, and
P3 = {{0, 4, 8}, {2, 3, 7}, {1, 5, 6}}.
Using this construction, we can obtain the following result.
Lemma 1. The construction procedure above results in a
design (X, A) where X = {0, 1, · · · , q k − 1} and |Bi,l | =
q k−1 for all 0 ≤ i ≤ n − 1 and 0 ≤ l ≤ q − 1.
Furthermore, the design is resolvable with parallel classes
given by Pi = {Bi,l : 0 ≤ l ≤ q − 1}, for 0 ≤ i ≤ n − 1.
B. A special class of linear block codes
We now introduce a special class of linear block codes
whose generator matrices satisfy specific rank properties. It
turns out that resolvable designs obtained from these codes
are especially suited for usage in coded caching.
Consider the generator matrix G of a (n, k) linear block
code over GF (q). The i-th column of G is denoted by gi .
Let z be the least positive integer such that k + 1 divides nz
(denoted by k + 1 | nz). We let (t)n denote t mod n.
In our construction we will need to consider various collections of k + 1 consecutive columns of G (wraparounds
Proof. Let G = [gab ], for 0 ≤ a ≤ k − 1, 0 ≤ b ≤ n − 1,
gab ∈ GF (q). Note that for ∆ = [∆0 ∆1 . . . ∆n−1 ] = uG,
we have
k−1
X
constructed as follows.
2 2 2
0 1 2
.
2 0 1
2 1 0
Using T, we generate the resolvable block design (X, A)
where the point set is X = {0, 1, 2, 3, 4, 5, 6, 7, 8}. For
instance, block B0,0 is obtained by identifying the column
indexes of zeros in the first row of T, i.e., B0,0 = {0, 1, 2}.
Following this, we obtain
Consider a (n, k) linear block code over GF (q). To avoid
trivialities we assume that its generator matrix does not have
an all-zeros column. We collect its q k codewords and construct
a matrix T of size n × q k as follows.
∆b =
1
.
2
Collecting the nine codewords, T is
0 0 0 1 1 1
0 1 2 0 1 2
T=
0 1 2 1 2 0
0 2 1 1 0 2
A. Resolvable Design Construction
T = [cT0 , cT1 , · · · , cTqk −1 ],
block code over GF (3)
ua gab ,
a=0
4
over the boundaries are allowed). For this purpose, let Ta =
{a(k + 1), · · · , a(k + 1) + k} (a is a non-negative integer) and
Sa = {(t)n | t ∈ Ta }. Let GSa be the k × (k + 1) submatrix
of G specified by the columns in Sa , i.e., g` is a column in
GSa if ` ∈ Sa . Next, we define the (k, k + 1)-consecutive
column property that is central to the rest of the discussion.
Algorithm 1: Placement Scheme
Input : Resolvable design (X, A) constructed from a
(n, k) linear block code. Let z be the least
positive integer such that k + 1 | nz.
k
1 Divide each file Wn , for n ∈ [N ] into q z subfiles. Thus,
s
k
Wn = {Wn,t : t ∈ {0, . . . , q − 1} and s ∈
{0, . . . , z − 1}} ;
2 User UB for B ∈ A caches
s
ZB = {Wn,t
: n ∈ [N ], t ∈ B and s ∈ {0, . . . , z − 1}} ;
Output: Cache content of user UB denoted ZB for
B ∈ A.
Definition 4. (k, k+1)-consecutive column property. Consider
zn
− 1.
the submatrices of G specified by GSa for 0 ≤ a ≤ k+1
We say that G satisfies the (k, k + 1)-consecutive column
property if all k × k submatrices of each GSa are full rank.
Henceforth, we abbreviate the (k, k+1)-consecutive column
property as (k, k + 1)-CCP.
Parallel Classes
Example 4. In Example 3 we have k = 2, n = 4 and hence
z = 3. Thus, S0 = {0, 1, 2}, S1 = {3, 0, 1}, S2 = {2, 3, 0}
and S3 = {1, 2, 3}. The corresponding generator matrix G
satisfies the (k, k + 1) CCP as any two columns of the each of
submatrices GSi , i = 0, . . . , 3 are linearly independent over
GF (3).
P0
0
We note here that one can also define different levels of the
consecutive column property. Let Taα = {aα, · · · , aα+α−1},
Saα = {(t)n | t ∈ Taα } and z is the least positive integer such
that α|nz.
PS0
Definition 5. (k, α)-consecutive column property Consider the
submatrices of G specified by GSaα for 0 ≤ a ≤ zn
α − 1. We
say that G satisfies the (k, α)-consecutive column property,
where α ≤ k if each GSaα has full rank. In other words, the
α columns in each GSaα are linearly independent.
P1
1
2
0
P2
1
2
PS1
0
1
PS2
P3
2
0
1
2
PS3
Recovery Sets
Fig. 1. Recovery set bipartite graph
In general, (see Algorithm 1) we have K = |A| = nq
users. Each file Wn , n ∈ [N ] is divided into q k z subfiles
s
s
| 0 ≤ t ≤ q k −1, 0 ≤ s ≤ z −1}. A subfile Wn,t
Wn = {Wn,t
is cached in user UB where B ∈ A if t ∈ B. Therefore, each
user caches a total of N q k−1 z subfiles. As each file consists
of q k z subfiles, we have that M/N = 1/q.
It remains to show that we can design a delivery phase
scheme that satisfies any possible demand pattern. Suppose
that in the delivery phase user UB requests file WdB where
dB ∈ [N ]. The server responds by transmitting several
equations that satisfy each user. Each equation allows k + 1
1
users from different parallel classes
to simultaneously obtain
a missing subfile. Our delivery scheme is such that the set of
transmitted equations can be classified into various recovery
sets that correspond to appropriate collections of parallel
classes. For example, in Fig. 1, PS0 = {P0 , P1 , P2 }, PS1 =
{P0 , P1 , P3 } and so on. It turns out that these recovery sets
zn
− 1 defined
correspond precisely to the sets Sa , 0 ≤ a ≤ k+1
earlier. We illustrate this by means of the example below.
As pointed out in the sequel, codes that satisfy the (k, α)CCP, where α ≤ k will result in caching systems that have
a multiplicative rate gain of α over an uncoded system.
Likewise, codes that satisfy the (k, k + 1)-CCP will have a
gain of k + 1 over an uncoded system. In the remainder of the
paper, we will use the term CCP to refer to the (k, k +1)-CCP
if the value of k is clear from the context.
C. Usage in a coded caching scenario
A resolvable design generated from a linear block code that
satisfies the CCP can be used in a coded caching scheme as
follows. We associate the users with the blocks. Each subfile is
associated with a point and an additional index. The placement
scheme follows the natural incidence between the blocks and
the points; a formal description is given in Algorithm 1 and
illustrated further in Example 5.
Example 5. Consider the resolvable design from Example
3, where we recall that z = 3. The blocks in A correspond
to twelve users U012 , U345 , U678 , U036 , U147 , U258 , U057 ,
U138 , U246 , U048 , U237 , U156 . Each file is partitioned into
Fs = 9 × z = 27 subfiles, each of which is denoted by
s
Wn,t
, t = 0, · · · , 8, s = 0, 1, 2. The cache in user Uabc ,
s
denoted Zabc is specified as Zabc = {Wn,t
| t ∈ {a, b, c}, s ∈
{0, 1, 2} and n ∈ [N ]}. This corresponds to a coded caching
system where each user caches 1/3-rd of each file so that
M/N = 1/3.
Example 6. Consider the placement scheme specified in
Example 5. Let each user UB request file WdB . The recovery
sets are specified by means of the recovery set bipartite graph
shown in Fig. 1, e.g., PS1 corresponds to S1 = {0, 1, 3}. The
outgoing edges from each parallel class are labeled arbitrarily
with numbers 0, 1 and 2. Our delivery scheme is such that
each user recovers missing subfiles with a specific superscript
from each recovery set that its corresponding parallel class
participates in. For instance, a user in parallel class P1 recovers
5
missing subfiles with superscript 0 from PS0 , superscript 1
from PS1 and superscript 2 from PS3 ; these superscripts are
the labels of outgoing edges from P1 in the bipartite graph.
It can be verified, e.g., that user U012 which lies in P0
recovers all missing subfiles with superscript 1 from the
equations below.
Wd1012 ,3 ⊕ Wd1036 ,2 ⊕ Wd0237 ,0 ,
Wd1012 ,6 ⊕ Wd1036 ,1 ⊕ Wd0156 ,0 ,
Wd1012 ,8 ⊕ Wd1258 ,0 ⊕ Wd0048 ,2 ,
Wd1012 ,5 ⊕ Wd1258 ,1 ⊕ Wd0156 ,2 .
Wd1012 ,4 ⊕ Wd1147 ,0 ⊕ Wd0048 ,1 ,
Algorithm 2: Signal Generation Algorithm for PSa
Input : For P ∈ PSa , E(P) = label(P − PSa ). Signal
set Sig = ∅.
1 while any user UB ∈ Pj , j ∈ Sa does not recover all its
missing subfiles with superscript E(Pj ) do
2
Pick blocks Bj,lj ∈ Pj for all j ∈ Sa and
lj ∈ {0, . . . , q − 1} such that ∩j∈Sa Bj,lj = ∅;
/* Pick blocks from distinct parallel classes
in PSa such that their intersection is
empty */
Wd1012 ,7 ⊕ Wd1147 ,2 ⊕ Wd0237 ,1 ,
3
Each of the equations above benefits three users. They are
generated simply by choosing U012 from P0 , any block from
P1 and the last block from P3 so that the intersection of all
these blocks is empty. The fact that these equations are useful
for the problem at hand is a consequence of the CCP. The
process of generating these equations can be applied to all
possible recovery sets. It can be shown that this allows all
users to be satisfied at the end of the procedure.
4
Let ˆls = ∩j∈Sa \{s} Bj,lj for s ∈ Sa ;
/* Determine the missing subfile index that
the user from Ps will recover */
E(Ps )
Add signal ⊕s∈Sa W
to Sig /* User UBs,ls
κs,ls ,l̂s
demands file Wκs,ls . This equation allows
it to recover the corresponding missing
subfile index l̂s . The superscript is
determined by the recovery set bipartite
graph */
In what follows, we first show that for the recovery set PSa
it is possible to generate equations that benefit k + 1 users
simultaneously.
5
Claim 1. Consider the resolvable design (X, A) constructed
as described in Section III.A by a (n, k) linear block code
that satisfies the CCP. Let PSa = {Pi | i ∈ Sa } for 0 ≤ a ≤
zn
k+1 − 1, i.e., it is the subset of parallel classes corresponding
to Sa . We emphasize that |PSa | = k + 1. Consider blocks
Bi1 ,li1 , . . . , Bik ,lik (where lij ∈ {0, . . . , q − 1}) that are
picked from any k distinct parallel classes of PSa . Then,
| ∩kj=1 Bij ,lij | = 1.
end
Output: Signal set Sig.
To see this consider the system of equations in variables
u0 , . . . , uk−1 .
k−1
X
ub gbi1 = li1 ,
b=0
..
.
k−1
X
Before proving Claim 1, we discuss its application in the
delivery phase. Note that the claim asserts that k blocks
chosen from k distinct parallel classes intersect in precisely
one point. Now, suppose that one picks k + 1 users from
k + 1 distinct parallel classes, such that their intersection
is empty. These blocks (equivalently, users) can participate
in an equation that benefits k + 1 users. In particular, each
user will recover a missing subfile indexed by the intersection
of the other k blocks. We emphasize here that Claim 1 is
at the core of our delivery phase. Of course, we need to
justify that enough equations can be found that allow all
users to recover all their missing subfiles. This follows from a
natural counting argument that is made more formally in the
subsequent discussion. The superscripts s ∈ {0, . . . , z −1} are
needed for the counting argument to go through.
ub gbik = lik .
b=0
By the CCP, the vectors gi1 , gi2 , . . . , gik are linearly independent. Therefore this system of k equations in k variables has
a unique solution over GF (q). The result follows.
We now provide an intuitive argument for the delivery
phase. Recall that we form a recovery set bipartite graph (see
Fig. 1 for an example) with parallel classes and recovery sets
as the disjoint vertex subsets. The edges incident on each
parallel class are labeled arbitrarily from 0, . . . , z − 1. For a
parallel class P ∈ PSa we denote this label by label(P −PSa ).
For a given recovery set PSa , the delivery phase proceeds
by choosing blocks from distinct parallel classes in PSa such
that their intersection is empty; this provides an equation that
benefits k + 1 users. It turns out that the equation allows a
user in parallel class P ∈ PSa to recover a missing subfile
with the superscript label(P − PSa ).
The formal argument is made in Algorithm 2. For ease of
notation in Algorithm 2, we denote the demand of user UBi,j
for 0 ≤ i ≤ n − 1, 0 ≤ j ≤ q − 1 by Wκi,j .
Proof. Following the construction in Section III.A, we note
that a block Bi,l ∈ Pi is specified by
Bi,l = {j : Ti,j = l}.
Now consider Bi1 ,li1 , . . . , Bik ,lik (where ij ∈ Sa , lij ∈
{0, . . . , q − 1}) that are picked from k distinct parallel classes
of PSa . W.l.o.g. we assume that i1 < i2 < · · · < ik . Let
I = {i1 , . . . , ik } and TI denote the submatrix of T obtained
by retaining the rows in I. We will show that the vector
[li1 li2 . . . lik ]T is a column in TI and only appears once.
Claim 2. Consider a user UB belonging to parallel class P ∈
PSa . The signals generated in Algorithm 2 can recover all the
missing subfiles needed by UB with superscript E(P).
Proof. Let Pα ∈ PSa . In the arguments below, we argue
that user UBα,lα that demands file Wκα,lα can recover all its
6
missing subfiles with superscript E(Pα ). Note that |Bα,lα | =
q k−1 . Thus, user UBα,lα needs to obtain q k − q k−1 missing
subfiles with superscript E(Pα ). Consider an iteration of
the while loop where block Bα,lα is picked in step 2. The
E(P )
equation in Algorithm 2 allows it to recover Wκ α,l̂ where
α,lα α
ˆlα = ∩j∈S \{α} Bj,l . This is because ∩j∈S Bj,l = ∅ and
j
a
j
a
because of Claim 1.
Next we count the number of equations that UBα,lα participates in. We can pick k − 1 users from some k − 1 distinct
parallel classes in PSa . This can be done in q k−1 ways.
Claim 1 ensures that the blocks so chosen intersect in a single
point. Next we pick a block from the only remaining parallel
class in PSa such that the intersection of all blocks is empty.
This can be done in q − 1 ways. Thus, there are a total of
q k−1 (q − 1) = q k − q k−1 equations in which user UBα,lα
participates in.
It remains to argue that each equation provides a distinct
subfile. Towards this end, let {i1 , . . . , ik } ⊂ Sa be an index
set such that α ∈
/ {i1 , . . . , ik }. Suppose that there exist
sets of blocks {Bi1 ,li1 , . . . , Bik ,lik } and {Bi1 ,li0 , . . . , Bik ,li0 }
1
k
such that {Bi1 ,li1 , . . . , Bik ,lik } 6= {Bi1 ,li0 , . . . , Bik ,li0 }, but
1
k
∩kj=1 Bij ,lij = ∩kj=1 Bij ,li0 = β. This is a contradiction since
j
T k+1
∩j=2 Bij ,li0 , which
this in turn implies that β ∈ ∩k+1
j=2 Bij ,lij
j
is impossible since two blocks from the same parallel class
have an empty intersection.
As the algorithm is symmetric with respect to all blocks
in parallel classes belonging to PSa , we have the required
result.
D. Obtaining a scheme for M/N = 1 −
The construction above works for a system where M/N =
1/q. It turns out that this can be converted into a scheme for
M
k+1
N = 1 − nq . Thus, any convex combination of these two
points can be obtained by memory-sharing.
Towards this end, we note that the class of coded caching
schemes considered here can be specified by an equationsubfile matrix. This is inspired by the hypergraph formulation
and the placement delivery array (PDA) based schemes for
coded caching in [15] and [14]. Each equation is assumed to
be of the all-but-one type, i.e., it is of the form Wdt1 ,Aj1 ⊕
Wdt2 ,Aj2 ⊕ · · · ⊕ Wdtm ,Ajm where for each ` ∈ [m], we have
the property that user Ut` does not cache subfile Wn,Aj` but
caches all subfiles Wn,Ajs where {js : s ∈ [m], s 6= `}.
The coded caching system corresponds to a ∆×Fs equationsubfile matrix S as follows. We associate each row of S with
an equation and each column with a subfile. We denote the
i-th row of S by Eqi and j-th column of S by Aj . The
value S(i, j) = t if in the i-th equation, user Ut recovers
subfile Wdt ,Aj , otherwise, S(i, j) = 0. Suppose that these
∆ equations allow each user to satisfy their demands, i.e., S
corresponds to a valid coded caching scheme. It is not too hard
to see that the placement scheme can be obtained by examining
S. Namely, user Ut caches the subfile corresponding to the j-th
column if integer t does not appear in the j-th column.
Example 7. Consider a coded caching system in [1] with
K = 4, ∆ = 4 and Fs = 6. We denote the four users as
U1 , U2 , U3 , U4 . Suppose that the equation-subfile matrix S for
this scheme is as specified below.
The overall delivery scheme repeatedly applies Algorithm
2 to each of the recovery sets.
A1
A2
A3
A4
A5
A6
3
2
0
1
0
0
Eq2
4
Eq3 0
Eq4
0
0
2
0
1
0
.
1
2
Eq1
Lemma 2. The proposed delivery scheme terminates and
allows each user’s demand to be satisfied. Furthermore the
transmission rate of the server is (q−1)n
k+1 and the subpacketization level is q k z.
Proof. See Appendix.
k+1
nq .
4
3
0
0
0
0
4
3
Upon examining S it is evident for instance that user U1
caches subfiles A1 , . . . , A3 as the number 1 does not appear
in the corresponding columns. Similarly, the cache placement
of the other users can be obtained. Interpreting this placement
scheme in terms of the user-subfile assignment, it can be verified that the design so obtained corresponds to the transpose of
the scheme considered in Example 1 (and also to the scheme
of [1] for K = 4, M/N = 1/2).
The main requirement for Lemma 2 to hold is that the
recovery set bipartite graph be biregular, where multiple edges
between the same pair of nodes is disallowed and the degree
of each parallel class is z. It is not too hard to see that this
follows from the definition of the recovery sets (see the proof
in the Appendix for details).
In an analogous manner, if one starts with the generator
matrix of a code that satisfies the (k, α)-CCP for α ≤ k, then
we can obtain the following result which is stated below. The
details are quite similar to the discussion for the (k, k+1)-CCP
and can be found in the Appendix (Section B).
Lemma 3. Consider a ∆×Fs equation-subfile matrix S whose
entries belong to the set {0, 1, . . . , K}. It corresponds to a
valid coded caching system if the following three conditions
are satisfied.
• There is no non-zero integer appearing more than once
in each column.
• There is no non-zero integer appearing more than once
in each row.
• If S(i1 , j1 ) = S(i2 , j2 ) 6= 0, then S(i1 , j2 ) = S(i2 , j1 ) =
0.
Corollary 1. Consider a coded caching scheme obtained by
forming the resolvable design obtained from a (n, k) code
that satisfies the (k, α)-CCP where α ≤ k. Let z be the least
positive integer such that α | nz. Then, a delivery scheme can
and
be constructed such that the transmission rate is (q−1)n
α
the subpacketization level is q k z.
Proof. The placement scheme is obtained as discussed earlier,
i.e., user Ut caches subfiles Wn,Aj if integer t does not appear
7
in column Aj . Therefore, matrix S corresponds to a placement
scheme.
Next we discuss the delivery scheme. Note that Eqi corresponds to an equation as follows.
By memory sharing any convex combination of these points
is achievable.
In a similar manner for the (n, k) linear block code that
satisfies the (k, α)-CCP over GF (q), the caching system
where M/N = 1/q can be converted into a system where
0
α
α
0
k zn
0
K 0 = nq, M
N 0 = 1 − nq , Fs = (q − 1)q α and R = (q−1)n
using the equation-subfile technique. The arguments presented
above apply with essentially no change.
Wdt1 ,Aj1 ⊕ Wdt2 ,Aj2 ⊕ · · · ⊕ Wdtm ,Ajm ,
where S(i, j1 ) = t1 , · · · , S(i, jm ) = tm . The above equation
can allow m users to recover subfiles simultaneously if (a) Ut`
does not cache Wn,Aj` and (b) Ut` caches all Wn,Ajs where
{js : s ∈ [m], s 6= `}. It is evident that Ut` does not cache
Wn,Aj` owing to the placement scheme. Next, to guarantee the
condition (b), we need to show that integer t` = S(i, j` ) will
not appear in column Ajs in S where {js : s ∈ [m], s 6= `}.
Towards this end, t` 6= S(i, js ) because of Condition 2. Next,
consider the non-zero entries that lie in the column Ajs but not
in the row Eqi . Assume there exists an entry S(i0 , js ) such that
S(i0 , js ) = S(i, j` ) = t` and i0 6= i, then S(i, js ) = ts 6= 0,
which is a contradiction to Condition 3. Finally, Condition 1
guarantees that each missing subfile is recovered only once.
IV. S OME CLASSES OF LINEAR CODES THAT SATISFY THE
CCP
At this point we have established that linear block codes that
satisfy the CCP are attractive candidates for usage in coded
caching. In this section, we demonstrate that there are a large
class of generator matrices that satisfy the CCP. For most of
the section we work with matrices over a finite field of order
q. In the last subsection, we discuss some constructions for
matrices over Z mod q when q is not a prime or prime power.
We summarize the constructions presented in this section in
Table I.
Lt
t
User Ut caches a fraction M
N = Fs where Lt is the number
of columns of S that do not have the entry t. Similarly, the
transmission rate is given by R = F∆s .
The crucial point is that the transpose of S, i.e., ST also
corresponds to a coded caching scheme. This follows directly
from the fact that ST also satisfies the conditions in Lemma
3. In particular, ST corresponds to a coded caching system
with K users and ∆0 subfiles. In the placement phase, the
M
cache size of Ut is Nt = ∆−F∆s +Lt . In the delivery phase, by
transmitting Fs equations corresponding to the rows of ST ,
all missing subfiles can be recovered. Then, the transmission
rate is R0 = F∆s .
Applying the above discussion in our context, consider the
equation-subfile matrix S corresponding to the coded caching
1
k
t
system with K = nq, M
N = q for 1 ≤ t ≤ nq, Fs = q z
nz
k
T
and ∆ = q (q − 1) k+1 . Then S corresponds to a system
0
k+1
k zn
0
with K 0 = nq, M
N = 1 − nq , Fs = (q − 1)q k+1 , and
k+1
transmission rate R0 = F∆s = (q−1)n
. The following theorem
is the main result of this paper.
A. Maximum-distance-separable (MDS) codes
(n, k)-MDS codes with minimum distance n − k + 1 are
clearly a class of codes that satisfy the CCP. In fact, for these
codes any k columns of the generator matrix can be shown to
be full rank. Note however, that MDS codes typically need
large field size, e.g., q + 1 ≥ n (assuming that the MDS
conjecture is true) [22]. In our construction, the value of
M/N = 1/q and the number of users is K = nq. Thus,
for large n, we will only obtain systems with small values of
M/N , or equivalently large values of M/N (by Theorem 1
above). This may be restrictive in practice.
B. Cyclic Codes
A cyclic code is a linear block code, where the circular shift
of each codeword is also a codeword [23]. A (n, k) cyclic
code
(q) is specified by a monic polynomial g(X) =
Pn−kover GF
i
i=0 gi X with coefficients from GF (q) where gn−k = 1
and g0 6= 0; g(X) needs to divide the polynomial X n − 1.
The generator matrix of the cyclic code is obtained as below.
g0 g1 · · · gn−k
0
· ·
0
0 g0 g1 · ·
·
gn−k 0 ·
0
G=.
.. .
..
.
Theorem 1. Consider a (n, k) linear block code over GF (q)
that satisfies the (k, k + 1) CCP. This corresponds to a coded
caching scheme with K = nq users, N files in
the server
1
where each user has a cache of size M ∈
q N, 1 −
k+1
N . Let z be the least positive integer such that
nq
k + 1 | nz. When
M
N
0
= 1q , we have
M
N
= (1 −
k+1
nq ),
·
0
g0
g1
·
·
·
gn−k
The following claim shows that for verifying the CCP for
a cyclic code it suffices to pick any set of k + 1 consecutive
columns.
(q − 1)n
R=
, and
k+1
Fs = q k z.
When
0
Claim 3. Consider a (n, k) cyclic code with generator matrix
G. Let GS denote a set of k + 1 consecutive columns of G.
If each k × k submatrix of GS is full rank, then G satisfies
the (k, k + 1)-CCP.
we have
k+1
, and
(q − 1)n
zn
Fs = (q − 1)q k
.
k+1
R=
Proof. Let the generator polynomial of the cyclic code be
g(X), where we note that g(X) has degree n − k. Let
8
Code type
Code construction
Notes
(n, k) MDS codes
Satisfy (k, k + 1)-CCP. Need q + 1 ≥ n.
(n, k) Cyclic codes
Existence depends on certain properties of the generator
polynomials. All cyclic codes satisfy the (k, k)-CCP. Need
additional conditions for the (k, k + 1)-CCP.
Kronecker product of z × α matrix satisfying the (z, z)-CCP with the identity matrix
It×t
Satisfy the (k, k)-CCP where k = tz.
Kronecker product of Vandermonde and
Vandermonde-like matrices with structured
base matrices
Satisfy the (k, k + 1)-CCP for certain parameters.
CCP matrix extension
Extends a k × n CCP matrix to a k × (n + s(k + 1)) CCP
matrix for integer s.
Codes over field GF (q)
Codes over ring Z mod q
Single parity-check (SPC) code
Satisfy the (k, k + 1)-CCP with n = k + 1.
Cyclic codes over the ring
Require that q = q1 × q2 × · · · × qd where qi ’s are prime.
Satisfy the (k, k)-CCP.
Kronecker product of z × α matrix satisfying the (z, z)-CCP with the identity matrix
It×t
Satisfy the (k, k)-CCP property where k = tz.
CCP matrix extension
Extends a k × n CCP matrix to a k × (n + s(k + 1)) CCP
matrix for integer s.
TABLE I
A SUMMARY OF THE DIFFERENT CONSTRUCTIONS OF CCP MATRICES IN S ECTION IV
GS = [g(a)n , g(a+1)n , · · · , g(a+k)n ] where we assume that
GS satisfies the (k, k + 1)-CCP. Let
•
GS\j =[g(a)n , . . . , g(a+j−1)n , g(a+j+1)n , . . . , g(a+k)n ], and
GS 0 \j =[g(a+i)n , . . . , g(a+j−1+i)n ,
g(a+j+1+i)n , . . . , g(a+k+i)n ].
We need to show that if GS\j has full rank, then GS 0 \j has
full rank, for any 0 ≤ j ≤ k.
As GS\j has full rank, there is no codeword c 6= 0 such
that c((a)n ) = · · · = c((a + j − 1)n ) = c((a + j + 1)n ) =
· · · = c((a + k)n ) = 0. By the definition of a cyclic code, any
circular shift of a codeword results in another codeword that
belongs to the code. Therefore, there is no codeword c0 such
that c0 ((a + i)n ) = · · · = c0 (a + j − 1 + i)n ) = c0 ((a + j +
1 + i)n ) = · · · = c0 ((a + k + i)n ) = 0. Thus, GS 0 \j has full
rank.
•
For 0 < j ≤ b k2 c, the submatrices
gn−k−1
gn−k
0
gn−k−2
g
g
n−k−1
n−k
..
Cj =
.
gn−k−j+1
·
·
gn−k−j
·
·
· ·
0 ·
·
·
0
0
..
.
·
gn−k
· gn−k−1
have full rank. In the above expression, gi = 0 if i < 0.
For b k2 c < j < k, the submatrices
g1 g2 · · ·
·
gk−j
g0 g1 · · ·
· gk−j−1
..
..
Cj = .
.
0
· · 0 g0 g1
g2
0
·
·· 0 g0
g1
have full rank.
Example 8. Consider the polynomial g(X) = X 4 + X 3 +
X + 2 over GF (3). Since it divides X 8 − 1, it is the generator
polynomial of a (8, 4) cyclic code over GF (3). The generator
matrix of this code is given below.
2 1 0 1 1 0 0 0
0 2 1 0 1 1 0 0
G=
0 0 2 1 0 1 1 0 .
0 0 0 2 1 0 1 1
Claim 3 implies a low complexity search algorithm to determine if a cyclic code satisfies the CCP. Instead of checking
zn
all GSa , 0 ≤ a ≤ k+1
− 1, in Definition 4, we only need
to check an arbitrary GS = [g(i)n , g(i+1)n , · · · , g(i+k)n ],
for 0 ≤ i < n. To further simplify the search, we choose
i = n − b k2 c − 1.
For this choice of i, Claim 4 shows that GS is such that
we only need to check the rank of a list of small-dimension
matrices to determine if each k × k submatrix of GS is full
rank (the proof appears in the Appendix).
It can be verified that the 4 × 5 submatrix which consists of
the two leftmost columns and three rightmost columns of G
is such that all 4 × 4 submatrices of it are full rank. Thus, by
Claim 3 the (4,5)-CCP is satisfied for G.
Claim 4. A cyclic code with generator matrix G satisfies the
CCP if the following conditions hold.
Remark 3. Cyclic codes form an important class of codes
that satisfy the (k, k)-CCP (cf. Definition 5). This is because,
9
it is well-known [23] that any k consecutive columns of the
generator matrix of a cyclic code are linearly independent.
Proof. The proof again leverages the idea that G can be
expressed succinctly by using Kronecker products. The arguments can be found in the Appendix.
C. Constructions leveraging properties of smaller base matrices
Consider the case when α = z + 1. We construct a (n, k)
linear code satisfy the CCP over GF (q) where q ≥ z. It can
be noted that the constraint of field size is looser than the
corresponding constraint in Claim 6.
It is well recognized that cyclic codes do not necessarily
exist for any choice of parameters. This is because of the
divisibility requirement on the generator polynomial. We now
discuss a more general construction of generator matrices that
satisfy the CCP. As we shall see, this construction provides a
more or less satisfactory solution for a large range of system
parameters.
Our first simple observation is that the Kronecker product
(denoted by ⊗ below) of a z ×α generator matrix that satisfies
the (z, z)-CCP with a t × t identity matrix, It×t immediately
yields a generator matrix that satisfies the (tz, tz)-CCP.
Claim 7. Consider nz = (z + 1) · (k + 1) . Consider a (n, k)
linear block code whose generator matrix (over GF (q)) is
specified as follows.
G=
It×t
0t×t
0t×t
It×t
···
···
0t×t
0t×t
0t×(t−1)
0t×(t−1)
1t
1t
..
.
b1 It×t
b2 It×t
..
.
0t×t
0t×t ···
It×t
0t×(t−1)
1t
bz−1 It×t
0(t−1)×t 0(t−1)×t ··· 0(t−1)×t I(t−1)×(t−1) 1t−1 C(c1 ,c2 )(t−1)×t ,
(4)
where t = n − k − 1. If q ≥ z, b1 , b2 , · · · , bz−1 are non-zero
and distinct, and c1 + c2 = 0, then G satisfies the CCP.
Claim 5. Consider a (n, k) linear block code over GF (q)
whose generator matrix is specified as G = A ⊗ It×t where
A is a z × α matrix that satisfies the (z, z)-CCP. Then, G
satisfies the (k, k)-CCP where k = tz and n = tα.
Proof. See Appendix.
Given a (n, k) code that satisfies the CCP, we can use it
obtain higher values of n in a simple manner as discussed in
the claim below.
Proof. The recovery set for A is specified as Saz =
{(az)α , · · · , (az+z−1)α } and the recovery set for G is specified as Sak = {(ak)n , · · · , (ak+k−1)n }. Since A satisfies the
(z, z)-CCP, ASaz has full rank. Note that GSak = ASaz ⊗ It×t .
Then det(GSak ) = det(ASaz ⊗ It×t ) = det(ASaz )t 6= 0.
Therefore, G satisfies the (k, k)-CCP.
Claim 8. Consider a (n, k) linear block code over GF (q)
with generator matrix G that satisfies the CCP. Let the first
k + 1 columns of G be denoted by the submatrix D. Then the
matrix G0 of dimension k × (n + s(k + 1)) where s ≥ 0
Remark 4. Let A be the generator matrix of a cyclic code
over GF (q), then G = A ⊗ It×t satisfies the (k, k)-CCP by
Claim 5.
G0 = [D| · · · |D |G]
| {z }
s
also satisfies the CCP.
Our next construction addresses the (k, k + 1)-CCP. In what
follows, we use the following notation.
T
• 1a : [1, · · · , 1] ;
| {z }
Proof. See Appendix.
Claim 8 can provide more parameter choices and more
possible code constructions. For example, given n, k, q, where
k +1+(n)k+1 ≤ q +1 < n, there may not exist a (n, k)-MDS
code over GF (q). However, there exists a (k +1+(n)k+1 , k)MDS code over GF (q). By Claim 8, we can obtain a (n, k)
linear block code over GF (q) that satisfies the CCP. Similarly,
combining Claim 4, Claim 6, Claim 7 with Claim 8, we can
obtain more linear block codes that satisfy the CCP.
A result very similar to Claim 8 can be obtained for the
(k, α)-CCP. Specifically, consider a (n, k) linear block code
with generator matrix G that satisfies the (k, α)-CCP and let
D be the first α columns of G. Then, G0 = [D| · · · |D |G] of
| {z }
a
• C(c1 , c2 )a×b :
a × b matrix where each row is the cyclic
shift (one place to the right) of the row above it and the
first row is [c1 c2 0 · · · 0]; and
• 0a×b : a × b matrix with zero entries.
Consider parameters n, k. Let the greatest common divisor
of n and k + 1, gcd(n, k + 1) = t. It is easy to verify that
is the smallest integer such that k + 1 | nz. Let
z = k+1
t
n = tα and k + 1 = tz. Claim 6 below constructs a (n, k)
linear code that satisfies the CCP over GF (q) where q > α.
Since α = nt , the required field size in Claim 6 is lower than
the MDS code considered in Section IV-A.
s
dimension k × (n + sα) also satisfies the (k, α)-CCP.
Claim 6. Consider a (n, k) linear block code over GF (q)
whose generator matrix is specified as eq. (3), where
b00
b01
···
b0(α−1)
b10
b11
···
b1(α−1)
..
..
.
.
b(z−1)0 b(z−1)1 · · · b(z−1)(α−1)
D. Constructions where q is not a prime or a prime power
We now discuss constructions where q is not a prime or a
prime power. We attempt to construct matrices over the ring Z
mod q in this case. The issue is somewhat complicated by the
fact that a square matrix over Z mod q has linear independent
rows if and only if its determinant is a unit in the ring [24]. In
general, this fact makes it harder to obtain constructions such
is a Vandermonde matrix and q > α. Then, G satisfies the
(k, k + 1)-CCP.
10
b00 It×t
b10 It×t
..
G=
.
b(z−2)0 It×t
C(b(z−1)0 , b(z−1)0 )(t−1)×t
b01 It×t
b11 It×t
···
···
b(z−2)1 It×t
C(b(z−1)1 , b(z−1)1 )(t−1)×t
···
···
as those in Claim 6 that exploit the Vandermonde structure of
the matrices. Specifically, the difference of units in a ring is not
guaranteed to be a unit. However, we can still provide some
constructions. It can be observed that Claim 5 and Claim 8
hold for linear block codes over Z mod q. We will use them
without proof in this subsection.
•
(3)
Let C denote the code over Z mod q. Let c(i) ∈ C i for
i = 1, . . . , d. The codeword c ∈ C is obtained as follows.
(1)
(d)
The j-th component of c, cj = ψ −1 (cj , . . . , cj )
Therefore, there are q1k1 q2k2 · · · qdkd codewords in C. It is
also evident that C is cyclic. As discussed in Section III-A,
we form the matrix T for the codewords in C. It turns out
that using T and the technique discussed in Section III-A,
we can obtain a resolvable design. Furthermore, the gain
of the system in the delivery phase can be shown to be
kmin = min{k1 , k2 , · · · , kd }. We discuss these points in detail
in the Appendix (Section C).
Claim 9. Let G = [Ik×k |1k ], i.e., it is the generator matrix of
a (k + 1, k) single parity check (SPC) code, where the entries
are from Z mod q. The G satisfies the (k, k + 1)-CCP and
the (k, k)-CCP. It can be used as base matrix for Claim 5.
Proof. It is not too hard to see that when G = [Ik×k |1k ], any
k × k submatrix of G has a determinant which is ±1, i.e., it is
a unit over Z mod q. Thus, the result holds in this case.
V. D ISCUSSION AND C OMPARISON WITH E XISTING
S CHEMES
A. Discussion
Claim 10. The following matrix with entries from Z mod q
satisfies the (k, k + 1)-CCP. Here k = 2t − 1 and n = 3t.
It×t
0
1t
It×t
G=
.
0
I(t−1)×(t−1) 1t C(1, −1)(t−1)×t
When the number of users is K = nq and the cache
1
fraction is M
N = q , we have shown in Theorem 1 that the
gain g = k + 1 and Fs = q k z. Therefore, both the gain and
the subpacketization level increase with larger k. Thus, for
our approach given a subpacketization budget Fs0 , the highest
coded gain that can be obtained is denoted by gmax = kmax +1
where kmax is the largest integer such that q kmax z ≤ Fs0 and
there exists a (n, kmax ) linear block code that satisfies the
CCP.
For determining kmax , we have to characterize the collection of values of k such that there exists a (n, k) linear
code satisfies the CCP over GF (q) or Z mod q. We use our
proposed constructions (MDS code, Claim 4, Claim 6, Claim
7, Claim 8, Claim 9, Claim 10) for this purpose. We call this
collection C(n, q) and generate it in Algorithm 3. We note here
that it is entirely possible that there are other linear block codes
that fit the appropriate parameters and are outside the scope
of our constructions. Thus, the list may not be exhaustive. In
addition, we note that we only check for the (k, k + 1)-CCP.
Working with the (k, α)-CCP where α ≤ k can provide more
operating points.
Proof. This can be proved by following the arguments in the
proof of Claim 7 while treating elements to be from Z mod q
and setting z = 2. We need to consider three different k ×(k +
1) submatrices for which we need to check the property. These
correspond to simpler instances of the submatrices considered
in Types I - III in the proof of Claim 7. In particular, the
corresponding determinants will always be ±1 which are units
over Z mod q.
Remark 5. We note that the general construction in Claim 7
can potentially fail in the case when the matrices are over Z
mod q. This is because in one of the cases under consideration
(specifically, Type III, Case 1), the determinant depends on the
difference of the bi values. The difference of units in Z mod q
is not guaranteed to be a unit, thus there is no guarantee that
the determinant is a unit.
Remark 6. We can use Claim 8 to obtain higher values of n
based on the above two classes of linear block codes over Z
mod q.
Example 9. Consider a caching system with K = nq = 12 ×
1
5 = 60 users and cache fraction M
N = 5 . Suppose that the
6
subpacketization budget is 1.5 × 10 . By checking all k < n
we can construct C(n, q) (see Table II). As a result, C(n, q) =
{1, 2, 3, 4, 5, 6, 7, 8, 9, 11}. Then kmax = 8, Fs ≈ 1.17 × 106
and the maximal coded gain we can achieve is gmax = 9. By
contrast, the scheme in [1] can achieve coded gain g = KM
N +
K
1 = 13 but requires subpacketization level Fs = KM
≈
N
12
1.4 × 10 .
We can achieve almost the same rate by performing
memory-sharing by using the scheme of [1] in this example. In
particular, we divide each file of size Ω into two smaller sub9
files Wn1 and Wn2 , where the size of Wn1 , |Wn1 | = 10
Ω and the
While most constructions of cyclic codes are over GF (q),
there has been some work on constructing cyclic codes over Z
mod q. Specifically, [25] provides a construction where q =
q1 × q2 · · · × qd and qi , i = 1, . . . , d are prime. We begin by
outlining this construction. By the Chinese remainder theorem
any element γ ∈ Z mod q has a unique representation in
terms of its residues modulo qi , for i = 1, . . . , d. Let ψ : Z
mod q → GF (q1 ) × · · · × GF (qd ) denote this map.
•
b0(α−1) It×t
b1(α−1) It×t
..
.
b(z−2)(α−1) It×t
C(b(z−1)(α−1) , b(z−1)(α−1) )(t−1)×t
Suppose that (n, ki ) cyclic codes over GF (qi ) exist for
all i = 1, . . . , d. Each individual code is denoted C i .
11
k
n0
z
α
Construction
Notes
11
12
1
1
(12, 11) SPC code
k + 1 = n0
10
12
11
12
-
-
9
12
5
6
Claim 7
α = z + 1 and q ≥ z
8
12
3
4
Claim 7
α = z + 1 and q ≥ z
7
12
2
3
Claim 7
6
12
7
12
Claim 4
5
6
1
1
(6,5) SPC code and Claim 8
Extend (6,5) SPC code to (12,5) code
4
7
5
7
Claim 4
Generator polynomial is X 8 + X 7 + 4X 6 + 3X 5 + 2X 3 + X 2 + 4X + 4
3
4
1
1
(4,3) SPC code and Claim 8
Extend (4,3) SPC code to (12,3) code
2
3
1
1
(3,2) SPC code and Claim 8
Extend (3,2) SPC code to (12,2) code
1
2
1
1
(2,1) SPC code and Claim 8
Extend (2,1) SPC code to (12,1) code
α = z + 1 and q ≥ z
Generator polynomial is X 6 + X 5 + 3X 4 + 3X 3 + X 2 + 4X + 3
TABLE II
L IST OF k VALUES FOR E XAMPLE 9. T HE VALUES OF n0 , α AND z ARE OBTAINED BY FOLLOWING A LGORITHM 3.
1
Ω. The scheme of [1] is then applied
size of Wn2 , |Wn2 | = 10
2
1
1
separately on Wn and Wn2 with M
N1 = 15 (corresponding to
M2
13
1
2
Wn ) and N2 = 15 (corresponding to Wn ). Thus, the overall
2
1
cache fraction is 0.9 × 15
+ 0.1 × 13
15 ≈ 5 . The overall coded
gain of this scheme is g ≈ 9. However, the subpacketization
level is FsM N = KMK1 /N1 + KMK2 /N2 ≈ 5 × 109 , which is
much greater than the subpacketization budget.
In Fig. 2, we present another comparison for system parameters K = 64 and different values of M/N . The scheme of
[1] works for all M/N such that KM/N is an integer. In Fig.
2, our plots have markers corresponding to M/N values that
our scheme achieves. For ease of presentation, both the rate
(left y-axis) and the logarithm of the subpacketization level
(right y-axis) are shown on the same plot. We present results
corresponding to two of our construction techniques: (i) the
SPC code and (ii) a smaller SPC code coupled with Claim
8. It can be seen that our subpacketization levels are several
orders of magnitude smaller with only a small increase in the
rate.
An in-depth comparison for general parameters is discussed
next. In the discussion below, we shall use the superscript ∗ to
refer to the rates and subpacketization levels of our proposed
scheme.
Fig. 2. A comparison of rate and subpacketization level vs. M/N for a
system with K = 64 users. The left y-axis shows the rate and the right
y-axis shows the logarithm of the subpacketization level. The green and the
blue curves correspond to two of our proposed constructions. Note that our
schemes allow for multiple orders of magnitude reduction in subpacketization
level and the expense of a small increase in coded caching rate.
determine λ and
∗
K(1 − MN )
K MN
R=λ
+
(1
−
λ)
∗
1 + KM
1 + K(1 −
N
M
M∗
M∗
=λ
+ (1 − λ) 1 −
.
N
N
N
Suppose that for given K, M and N , a given rate R can
be achieved by the memory sharing of the scheme in [1]
between the corner points (M1 , R1 ), (M2 , R2 ), · · · , (Md , Rd )
Mi
that satisfy the following equations.
∗
B. Comparison with memory-sharing within the scheme of [1]
K(1−
M∗
N
∗
0
M∗
N )
, and
(5)
(6)
Here, MN ≤ 12 , and M 0 = tKN , where t0 is the least integer
such that M 0 ≥ M ∗ .
To see this, consider the following argument. Suppose
that the above statement is not true. Then, there exists a
scheme that operates via memory sharing between
points
(M1 , R1 ), · · · , (Md , Rd ) such that Fs < KKM 0 . Note that
N
K
K
1
1
2
1
2
KM1
< KM
if M
< M
≤ 12 or M
> M
2
N
N
N
N ≥ 2 . By
N
N
the convexity of h(·), we can conclude that (M, R) is not in
the convex hull of the corner points (M1 , R1 ), · · · , (Md , Rd ).
This is a contradiction.
)
N
where Mi = tiKN for some integer ti . Then Ri =
,
KM
1+ N i
Pd
Pd
P
d
i
R = i=1 λi Ri , M/N = i=1 λi M
i=1 λi = 1. The
N and
P
d
K
MS
subpacketization level is Fs
= i=1 KMi . In addition,
N
we note that the function h(x) = K(1 − x)/(1 + Kx) is
convex in the parameter 0 ≤ x ≤ 1. This can be verified by a
simple second derivative calculation.
K
0 ,
We first argue that FsM S is lower bounded by KM
N
where M 0 is obtained as follows. For a given M/N , we first
12
Unfortunately, doing this analytically becomes quite messy
and does not yield much intuition. Instead, we illustrate the
reduction in subpacketization level by numerical comparisons.
Algorithm 3: C(n, q) Construction Algorithm
Input : n, q, C(n, q) = ∅
1 if q is a prime power then
2
for k = 1 : (n − 1) do
3
n0 ← (n)k+1 + k + 1;
k+1
4
z ← gcd(n
0 ,k+1) ;
0
5
α ← gcd(nn0 ,k+1) ;
6
if there exists a (n0 + i(k + 1), k) cyclic code
which satisfies the condition in Claim 4 for
some i such that n0 + i(k + 1) ≤ n then
7
C(n, q) ← k. Corresponding codes are
constructed by using Claim 8.
8
else
9
if z ≤ 2 then
10
C(n, q) ← k. Corresponding codes are
constructed by SPC code and Claim 8
when z = 1 or Claim 7 and Claim 8
when z = 2.
11
else
12
if q + 1 ≥ n0 then
13
C(n, q) ← k. Corresponding codes are
constructed by MDS code and Claim
8.
14
else
15
if α = z + 1 and q ≥ z then
16
C(n, q) ← k. Corresponding codes
are constructed by Claim 7 and
Claim 8.
17
end
18
if α > z + 1 and q > α then
19
C(n, q) ← k. Corresponding codes
are constructed by Claim 6 and
Claim 8.
20
end
21
end
22
end
23
end
24
end
25 end
26 if q is not a prime power then
27
for k = 1 : (n − 1) do
28
if z ≤ 2 then
29
C(n, q) ← k. Corresponding codes are
constructed by Claim 9 when z = 1 and
Claim 8 or Claim 10 and Claim 8 when
z = 2.
30
end
31
end
32 end
Output: C(n, q)
Example 10. Consider a (9, 5) linear block code over GF (2)
with generator matrix specified below.
1 0 0 0 0 1 1 0 0
0 1 0 0 0 1 0 1 0
G=
0 0 1 0 0 1 0 0 1 .
0 0 0 1 0 1 1 1 0
0 0 0 0 1 1 0 1 1
It can be checked that G satisfies the (5, 6)-CCP. Thus, it
corresponds to a coded caching system with K = 9 × 2 = 18
1
3
1
users. Our scheme achieves the point M
N = 2 , R1 = 2 ,
M2
2
2
∗
∗
Fs,1 = 64 and N = 3 , R2 = 3 , Fs,2 = 96.
1
3
1
On the other hand for M
N =
2 , R1 = 2 , by numerically
M1∗
M10
solving (5) and (6) we obtain
N ≈ 0.227 and therefore N =
18
5
MS
= 8568, which is much higher than
18 . Then Fs,1 ≥
5
M∗
∗
Fs,1 = 64. A similar calculation shows that N2 ≈ 41 and
M0
5
MS
. Thus Fs,2
is also at least as large as
therefore N2 = 18
∗
8568, which is still much higher than Fs,2
= 96.
The next set of comparisons are with other proposed
schemes in the literature. We note here that several of these
are restrictive in the parameters that they allow.
C. Comparison with [1], [14], [21], [15] and [16]
For comparison with [1], denote RM N and FsM N be the
rate and the subpacketization level of the scheme of [1],
respectively. For the rate comparison, we note that
R∗
1+n
M
1
=
, for
=
M
N
R
1+k
N
q
R∗
nq − k
M
1+k
=
, for
=1−
,
RM N
nq − n
N
nq
For the comparison of subpacketization level we have the
following results.
Claim 11. When K = nq, the following results hold.
1
M
• If N = q , we have
1
F MN
1
η
lim
log2 s ∗ = H2
− log2 q.
n→∞ K
Fs
q
q
•
results can be obtained for
M
N
=1−
M
N
=1−
k+1
nq ,
we have
1
F MN
η
η
log2 s ∗ = H2
− log2 q.
n→∞ K
Fs
q
q
lim
(8)
In the above expressions, 0 < η = k/n ≤ 1 and H2 (·)
represents the binary entropy function.
Proof.
results are simple consequences of approximating
Both
K
KH2 (p)
[26]. The derivations can be found in the
Kp ≈ 2
Appendix.
Next, we compare this lower bound on FsM S to the subpacketization level of our proposed scheme. In principle, we
can solve the system of equations (5) and (6) for R = n(q−1)
k+1
1
∗
1
and M
N = q and obtain the appropriate λ and M values .
1 Similar
If
(7)
It is not too hard to see that Fs∗ is exponentially lower than
Thus, our rate is higher, but the subpacketization level
is exponentially lower. Thus, the gain in the scaling exponent
of with respect to the scheme of [1] depends on the choice of R
FsM N .
k+1
nq
13
where a,b, m are positive integers and a + b ≤ m and (ii)
1
1
m
t
t M
K= m
t q , N = 1 − q t , Fs = q (q − 1) and R = (q−1)t ,
where q, t, m are positive integers.
Their scheme (i), is the special case of scheme in [21] when
β = 0. For the second
scheme, if we let t = 2, [15] shows that
q
K √
K
∗
2q
R ≈ R , Fs ≈ q
( q −1)2 and Fs∗ ≈ (q −1)q q −1 , which
means Fs is again better than Fs∗ . We emphasize here that
these results require somewhat restrictive parameter settings.
Finally, we consider the work of [16]. In their work, they
leveraged the results of [27] to arrive at coded caching schemes
where the subpacketization is linear in K. Specifically, they
show that for any constant M/N , there exists a scheme with
rate K δ , where δ > 0 can be chosen arbitrarily small by
choosing K large enough. From a theoretical perspective, this
is a positive result that indicates that regimes where linear
subpacketization scaling is possible. However, these results
are only valid when the value of K is very large. Specifically,
K = C n and the result is asymptotic in the parameter n. For
these parameter ranges, the result of [16] will clearly be better
as compared to our work.
1
Scaling exponent gain
= 0.25
= 0.5
0.8
=0.75
=0.99
0.6
0.4
0.2
0
0
5
10
15
20
q
Fig. 3. The plot shows the gain in the scaling exponent obtained using our
techniques for different value of M/N = 1/q. Each curve corresponds to a
choice of η = k/n.
and the value of M/N . In Fig. 3 we plot this value of different
values of R and q. The plot assumes that codes satisfying the
CCP can be found for these rates and corresponds to the gain
in eq. (7).
In [14] a scheme for the case when M/N = 1/q or M/N =
1−1/q with subpacketization level exponentially smaller with
respect to [1] was presented. This result can be recovered a
special case of our work (Theorem 1) when the linear block
code is chosen as a single parity check code over Z mod q.
In this specific case, q does not need to be a prime power.
Thus, our results subsume the results of [14].
In a more recent preprint,
referencem[21],
proposed a caching
m
system
with
K
=
,
F
=
and M/N = 1 −
s
a
b
m
a m−a
/
.
The
corresponding
rate
is
λ
b−β
b
m
m − (a + b − 2β)
a + b − 2β
a+b−2β
min
,
,
R=
m
λ
a−β
b
VI. C ONCLUSIONS AND F UTURE W ORK
In this work we have demonstrated a link between specific
classes of linear block codes and the subpacketization problem
in coded caching. Crucial to our approach is the consecutive column property which enforces that certain consecutive
column sets of the corresponding generator matrices are fullrank. We present several constructions of such matrices that
cover a large range of problem parameters. Leveraging this
approach allows us to construct families of coded caching
schemes where the subpacketization level is exponentially
smaller compared to the approach of [1].
There are several opportunities for future work. Even though
our subpacketization level is significantly lower than [1],
it still scales exponentially with the number of users. Of
course, the rate of growth with the number of users is much
smaller. There have been some recent results on coded caching
schemes that demonstrate the existence of schemes where the
subpacketization scales sub-exponentially in the number of
users. It would be interesting to investigate whether some of
these ideas can be leveraged to obtain schemes that work for
practical systems with tens or hundreds of users.
where m, a, b, β are positive integers and 0 < a < m, 0 <
b < m, 0 ≤ β ≤ min{a, b}. While a precise comparison
is somewhat hard, we can compare the schemes for certain
parameter choices, that were also considered in [21].
Let a = 2, β = 1, m = 2b. This corresponds to a coded
b−1
1
caching system with K = b(2b − 1) ≈ 2b2 , M
N = 2b−1 ≈ 2 ,
2b
2b
Fs = b ≈ 2 , R = b. For comparison with our scheme
we keep the transmission rates of both schemes roughly the
same and let n = b2 , q = 2, k = b − 1. We assume that the
corresponding linear block code exists. Then Fs∗ ≈ 2b , which
is better than Fs .
On the other hand if we let β = 0, a = 2, m = 2qb, we
1
, M
obtain a coded caching system with K = m(m−1)
2
N ≈ q,
m
Fs = m
≈ (2q) 2q , RY AN = (2q − 1)2 . For keeping the
m
R EFERENCES
[1] M. Maddah-Ali and U. Niesen, “Fundamental limits of caching,” IEEE
Trans. on Info. Th., vol. 60, no. 5, pp. 2856–2867, May 2014.
[2] H. Ghasemi and A. Ramamoorthy, “Improved lower bounds for coded
caching,” IEEE Trans. on Info. Th., vol. 63, no. 7, pp. 4388–4413, 2017.
[3] A. Sengupta, R. Tandon, and T. C. Clancy, “Improved approximation of
storage-rate tradeoff for caching via new outer bounds,” in IEEE Intl.
Symposium on Info. Th., 2015, pp. 1691–1695.
[4] M. Maddah-Ali and U. Niesen, “Decentralized coded caching attains
order-optimal memory-rate tradeoff,” IEEE/ACM Trans. Netw., vol. 23,
no. 4, pp. 1029–1040, 2015.
[5] U. Niesen and M. A. Maddah-Ali, “Coded caching with nonuniform
demands,” IEEE Trans. on Info. Th., vol. 63, no. 2, pp. 1146–1158, Feb
2017.
[6] J. Hachem, N. Karamchandani, and S. Diggavi, “Multi-level coded
caching,” in IEEE Intl. Symposium on Info. Th., 2014, pp. 56–60.
[7] L. Tang and A. Ramamoorthy, “Coded caching for networks with the
resolvability property,” in IEEE Intl. Symposium on Info. Th., 2016.
2q
rates the same, we let n =
m(m−1)
4q(2q−1)
m2
8q 2
m(m−1)
,
2q
k=
m(m−1)
4q(2q−1)
− 1 so that
Fs∗ ≈ q
≈ q . In this regime, the subpacketization
level of [21] will typically be lower.
The work of [15] proposed caching schemes with paramem
M
(m−a
(a+b
)
m
b )
ters (i) K = m
a , N = 1 − (m) , Fs = b and R = (m) ,
b
b
14
that there exists a ga∗ b which is relatively prime to qi , i.e.,
ga∗ b is a unit in the ring Z mod qi .
Note that for ∆ = [∆0 ∆1 . . . ∆n−1 ] = uG, we have
[8] M. Ji, M. F. Wong, A. M. Tulino, J. Llorca, G. Caire, M. Effros, and
M. Langberg, “On the fundamental limits of caching in combination
networks,” in IEEE 16th International Workshop on Signal Processing
Advances in Wireless Communications (SPAWC), June 2015, pp. 695–
699.
[9] H. Ghasemi and A. Ramamoorthy, “Asynchronous coded caching,” in
IEEE Intl. Symposium on Info. Th., 2017, pp. 2438–2442.
[10] U. Niesen and M. A. Maddah-Ali, “Coded caching for delay-sensitive
content,” in IEEE Intl. Conf. Comm., 2015, pp. 5559–5564.
[11] M. E. Fitzpatrick, “4K Sector Disk Drives: Transitioning to the Future with Advanced Format Technologies,” Toshiba 4K White Paper, 2011, [Online] Available:http://cdaweb01.storage.toshiba.com/docs/
services-support-documents/toshiba 4kwhitepaper.pdf.
[12] K. Shanmugam, M. Ji, A. M. Tulino, J. Llorca, and A. G. Dimakis,
“Finite length analysis of caching-aided coded multicasting,” in 52nd
Annual Allerton Conference on Communication, Control, and Computing, Sept 2014, pp. 914–920.
[13] K. Shanmugam, M. Ji, A. M. Tulino, J. Llorca, and A. G. Dimakis.,
“Finite-length analysis of caching-aided coded multicasting,” IEEE
Trans. on Info. Th., vol. 62, no. 10, pp. 5524–5537, 2016.
[14] Q. Yan, M. Cheng, X. Tang, and Q. Chen, “On the placement delivery
array design for centralized coded caching scheme,” IEEE Trans. on
Info. Th., vol. 63, no. 9, pp. 5821–5833, 2017.
[15] C. Shangguan, Y. Zhang, and G. Ge, “Centralized coded caching
schemes: A hypergraph theoretical approach,” preprint, 2016, [Online]
Available: https://arxiv.org/abs/1608.03989.
[16] K. Shanmugam, A. M. Tulino, and A. G. Dimakis, “Coded caching
with linear subpacketization is possible using Ruzsa-Szeméredi graphs,”
in IEEE Intl. Symposium on Info. Th., June 2017, pp. 1237–1241.
[17] O. Olmez and A. Ramamoorthy, “Fractional repetition codes with
flexible repair from combinatorial designs,” IEEE Trans. on Info. Th.,
vol. 62, no. 4, pp. 1565 –1591, 2016.
[18] A. S. Tripathy and A. Ramamoorthy, “Capacity of sum-networks for
different message alphabets,” in IEEE Intl. Symposium on Info. Th.,
2015, pp. 606–610.
[19] ——, “”Sum-networks from incidence structures: construction and capacity analysis,” IEEE Trans. on Info. Th., 2017 (to appear).
[20] D. R. Stinson, Combinatorial Designs: Construction and Analysis.
Springer, 2003.
[21] Q. Yan, X. Tang, Q. Chen, and M. Cheng, “Placement delivery array
design through strong edge coloring of bipartite graphs,” IEEE Communications Letters, 2017 (to appear).
[22] R. M. Roth, Introduction to Coding Theory. Cambridge University
Press, 2006.
[23] S. Lin and D. J. Costello, Error Control Coding, 2nd Ed. Prentice Hall,
2004.
[24] D. S. Dummit and R. M. Foote, Abstract algebra. Wiley, 3rd Ed.,
2003.
[25] I. F. Blake, “Codes over certain rings,” Information and Control, vol. 20,
no. 4, pp. 396–404, 1972.
[26] R. L. Graham, D. E. Knuth, and O. Patashnik, Concrete mathematics:
a foundation for computer science (2nd ed.). Addison-Wesley Professional, 1994.
[27] N. Alon, A. Moitra, and B. Sudakov, “Nearly complete graphs decomposable into large induced matchings and their applications,” in Proc.
of the 44-th Annual ACM symposium on Theory of computing (STOC),
2012, pp. 1079–1090.
[28] R. A. Horn and C. R. Johnson, Topics in matrix analysis. Cambridge
University Press, 1991.
∆b =
k−1
X
ua gab ,
(9)
a=0
where u = [u0 , · · · , uk−1 ]. We consider eq. (9) over the ring
Z mod qi and rewrite eq. (9) as
X
∆b − ua∗ ga∗ b =
ua gab ,
a6=a∗
For arbitrary ua , a 6= a∗ , this equation has a unique
solution for ua∗ since ga∗ b is a unit in Z mod qi . This
implies that there are qik−1 distinct solutions for (9) over Z
mod qi . Using the Chinese remainder theorem, eq. (9) has
q1k−1 × q2k−1 × · · · qdk−1 = q k−1 solutions over Z mod q and
the result follows.
Remark 7. From Lemma 4, it can be easily verified that a
linear block code over Z mod q can construct a resolvable
block design if one of the following conditions for each
column gi of the generator matrix is satisfied.
• At least one non-zero entry of gi is a unit in Z mod q,
or
• all non-zero entries in gi are zero divisors but their
greatest common divisor is 1.
For the SPC code over Z mod q, all the non-zero entries in
the generator matrix are 1, which is an unit. Therefore, the
construction always results in a resolvable design.
Proof of Lemma 2
First, we show that the proposed delivery scheme allows
each user’s demand to be satisfied. Note that Claim 2 shows
that each user in a parallel class that belongs to the recovery
set Sa recovers all missing subfiles with a specified superscript
from it. Thus, we only need to show that if signals are
generated (according to Claim 2) for each recovery set, we are
done. This is equivalent to showing that the bipartite recovery
set graph is such that each parallel class has degree z and
multiple edges between nodes are disallowed.
Towards this end, consider the parallel class and we claim
that there exist exactly z solutions (aα , bα ) for integer values
of α = 1, . . . , z to the equation
aα (k + 1) + bα = j + n(α − 1)
(10)
such that aα1 6= aα2 for α1 6= α2 and j < n. The existence of
the solution for each equation above follows from the division
algorithm. Note that aα < nz/(k + 1) as the RHS < nz.
Furthermore, note that for 1 ≤ α1 ≤ z and 1 ≤ α2 ≤ z, we
cannot have solutions to eq. (10) such that aα1 = aα2 as this
would imply that |bα1 − bα2 | ≥ n which is a contradiction.
This shows that each parallel class Pj participates in at least
z different recovery sets.
The following facts follow easily from the construction
of the recovery sets. The degree of each recovery set in
nz
of them;
the bipartite graph is k + 1 and there are k+1
multiple edges between a recovery set and a parallel class
A PPENDIX
A. Resolvable design over Z mod q
Lemma 4. A (n, k) linear block code over Z mod q
with generator matrix G = [gab ] can construct a resolvable block design by the procedure in Section III-A if
gcd(q, g0b , g1b , · · · , g(k−1)b ) = 1 for 0 ≤ b < n.
Proof. Assume q = q1 × q2 × · · · × qd where qi , 1 ≤ i ≤ d is a
prime or a prime power. If the gcd(q, g0b , g1b , · · · , g(k−1)b ) =
1, then it is evident that gcd(qi , g0b , g1b , · · · , g(k−1)b ) = 1 for
1 ≤ i ≤ d. As qi is either a prime or a prime power, it follows
15
are disallowed. Therefore, the total number of edges in the
bipartite graph is nz. As each parallel class participates in
at least z recovery sets, by this argument, it participates in
exactly z recovery sets. .
Finally, we calculate the rate of the delivery phase. In total,
zn
equations, where the symbol
the server transmits q k (q −1) k+1
transmitted has the size of a subfile. Thus, the rate is
and
gn−k
gn−k−1
Ej =
..
.
gn−k−b k c+j+1
0
gn−k
·
2
· ·
0 ·
·
·
0
0
..
.
.
gn−k
Matrices Aj and Ej have full rank as they are respectively
upper triangular and lower triangular, with non-zero entries on
the diagonal (as g0 and gn−k are non-zero in a cyclic code).
Therefore, GS\(a+j)n has full rank if Cj has full rank. For
b k2 c < j < k, GS\(a+j)n can be partitioned into a similar
form and the result in Claim 4 follows.
zn 1
R = q (q − 1)
k + 1 qk z
(q − 1)n
=
.
k+1
k
Proof of Claim 4
Proof of Claim 6
The matrix GS is shown below.
g0
0
.
.
.
0
0
0
.
.
.
0
g1
g0
·
g1
·
·
gd k e−1
2
gd k e−2
0
0
·
·
·
·
0
0
·
·
·
0
·
·
g0
0
0
g1
g0
0
0
gn−k
·
0
gn−k−1
gn−k
·
·
0
0
·
·
·
·
0
0
gn−k−b k c
·
·
·
2
2
We need to argue that all k × k submatrices of GSa where
0
≤
a < α are full rank. In what follows we argue that all
k × k submatrices of GS0 are full rank. The proof for any
GSa is similar. Note that GS0 can be written compactly as
follows by using Kronecker products.
A ⊗ It
GS0 =
,
B ⊗ C(t−1)×t (1, 1)
where
gn−k
b00
···
b0(z−1)
..
if i is
A = ...
.
0
0
..
.
0
0
0
..
.
In the above expression and the subsequent discussion
such that i < 0, we set gi = 0.
By Claim 3, a cyclic code with generator matrix G satisfies
the CCP if all submatrices
b(z−2)0
g(a+j+1)n , · · · , g(a+k)n ],
where a = n − b k2 c − 1, 0 ≤ j ≤ k, have full rank. In what
follows, we argue that this is true. Note that in the generator
matrix of cyclic code, any k consecutive columns are linearly
independent [23]. Therefore for j = 0 and k, GS\(a+j)n has
full rank, without needing the conditions of Claim 4. For 0 <
j ≤ b k2 c, GS\(a+j)n is as eq. (11).
Rewriting GS\(a+j)n in block form, we get
Aj
Bj
GS\(a+j)n =
Cj 0 ,
Dj Ej
where A0 and A00 are the first z − 1 columns and last
column of A respectively. Likewise, B0 and B00 are the first
z − 1 components and last component of B. The matrices ∆1
and ∆2 are obtained by deleting the j-th column of It and
C(t−1)×t (1, 1) respectively. Then, using the Schur determinant
identity [28], we have
0
det(GS0 \i ) = det(A0 ⊗ It ) det(B00 ⊗ ∆2 −
where
gn−k−1
gn−k−2
..
.
Cj =
gn−k−j+1
gn−k−j
g1
g0
·
g1
·
·
·
0
gn−k
0
gn−k−1
gn−k
·
·
·
·
gd k e−1
2
gd k e−2
2
,
..
.
g0
· ·
0 ·
·
·
b(z−2)(z−1)
and B = [b(z−1)0 , · · · , b(z−1)(z−1) ].
Next, we check the determinant of submatrices GS0 \i
obtained by deleting i-th column of GS0 . W.l.o.g, we let
i = (z − 1)t + j where 0 ≤ j < t. The block form of the
resultant matrix GS0 \i can be expressed as
A0 ⊗ It
A00 ⊗ ∆1
GS0 \i =
,
B0 ⊗ C(t−1)×t (1, 1) B00 ⊗ ∆2
GS\(a+j)n =[g(a)n , g(a+1)n , · · · , g(a+j−1)n ,
g0
0
Aj = .
..
0
···
(1)
B0 ⊗ C(t−1)×t (1, 1) · (A0 ⊗ It )−1 · A00 ⊗ ∆1 )
= det(A0 ⊗ It ) det(B00 ⊗ ∆2 −
(2)
B0 A0−1 A00 ⊗ C(t−1)×t (1, 1)∆1 )
= det(A0 ⊗ It ) det((B00 − B0 A0−1 A00 ) ⊗ ∆2 ),
0
0
..
.
·
gn−k
· gn−k−1
where (1) holds by the properties of the Kronecker product
[28] and (2) holds since C(t−1)×t (1, 1)∆1 = ∆2 . Next note
that det(∆2 ) 6= 0. This is because ∆2 can be denoted as
A 0
,
∆2 =
0 B
,
16
g0
0
.
..
0
0
0
.
.
.
0
0
0
.
.
.
0
g1
g0
·
·
gd k e−1
2
gd k e−2
0
0
·
·
·
·
·
·
·
·
·
·
·
g0
0
·
g1
g0
0
0
gn−k
·
0
gn−k−1
gn−k
·
·
0
·
·
·
·
·
·
·
·
·
·
·
·
0
0
0
gn−k−j+1
gn−k−j
gn−k−j−1
·
·
·
gn−k
gn−k−1
gn−k−2
0
0
gn−k
·
·
·
·
·
0
gn−k−b k c
·
gn−k−b k c+j−1
gn−k−b k c+j+1
2
2
2
where A is a j × j upper-triangular matrix;
1 1 0 ··· 0 0
0 1 1 · · · 0 0
..
A = ...
.
0 0 0 · · · 1 1
0 0 0 ··· 0 1
and B is a (t − 1 − j) × (t − 1 − j)
1 0 ··· 0
1 1 · · · 0
B = ...
0 0 · · · 1
0 0 ··· 0
Next, we define the matrix
b00
b01
b10
b11
F= .
..
b(z−1)0
b(z−1)1
•
•
lower-triangular matrix;
0 0
0 0
.. .
.
1 0
1 1
2
·
0
0
..
.
0
0
0
.
.
..
0
0
0
..
.
(11)
gn−k
Type I
When a = 0, it is easy to verify that any k × k submatrix
of GS0 has full rank since GS0 has the form [Ik×k |1k ],
which is the generator matrix of the SPC code.
Type II
When a = 1, GS1 has the form in eq. (12),
Case 1: Suppose that we delete any of first (z − 1)t
columns in GS1 (this set of columns is depicted by the
underbrace in eq. (12)), say i-th column of GS1 , where
z1 t ≤ i < (z1 + 1)t and 0 ≤ z1 ≤ z − 2. Let i1 = i − z1 t,
i2 = (z1 + 1)t − i − 1. The resultant matrix GS1 \i can
be expressed as follows.
A C
,
GS1 \i =
B D
where
···
···
b0(z−1)
b1(z−1)
..
.
···
A = Ii×i ,
.
B = 0(k−i)×i ,
0t×(k−t−i)
0t×(k−t−i)
..
C=
.
0t×(k−t−i)
0i1 ×(k−t−i)
b(z−1)(z−1)
Another application of the Schur determinant identity yields
det(F)
det(A0 )
6 0,
=
det(B00 − B0 A0−1 A00 ) =
b1 It×t
b2 It×t
..
.
bz1 It×t
bz1 +1 Ii1 ×i1 0i1 ×(i2 +1)
,
and D has the form in eq. (13).
Note
that
if
z1
=
0,
C
=
[0i1 ×(k−t−i) b1 Ii1 ×i1 0i1 ×(i2 +1) ] and if z1 = z − 2,
01×i2
01×i1 bz−1
01×i2
0i2 ×(i1 +1)
bz−1 Ii2 ×i2 .
D = Ii2 ×i2
0(t−1)×i2
C(c1 , c2 )(t−1)×t
0
since det(F) and det(A ) are both non-zero as their columns
have the Vandermonde form. In F, the columns correspond to
distinct and non-zero elements from GF (q); therefore, q > z.
Note however, that the above discussion focused only GS0 . As
the argument needs to apply for all GSa where 0 ≤ a < α,
we need q > α.
To verify GS1 \i has full rank, we just need to check D
has full rank (as A is full rank). Checking that D has full
rank can be further simplified as follows. As bz1 +1 6= 0,
we can move the corresponding column that has bz1 +1
as its first entry so that it is the first column of D.
Following this, consider C(c1 , c2 )(t−1)×t \ ci1
which is obtained by deleting the i1 -th column of
C(c1 , c2 )(t−1)×t .
D1 D2
C(c1 , c2 )(t−1)×t \ ci1 =
,
D3 D4
Proof of Claim 7
Note that the matrix in eq. (4) is the generator matrix of
(n, k) linear block code over GF (q) where nz = (z + 1)(k +
1). Since z and z+1 are coprime, z is the least positive integer
such that k + 1 | nz. To show G satisfies the CCP, we need
to argue that all k × k submatrices of GSa where 0 ≤ a ≤ z
are full rank. It is easy to check that Sa = {0, · · · , n − 1} \
{t(z − a), t(z − a) + 1, · · · , t(z − a) + t − 1}. We verify three
types of matrix GSa as follows: I. a = 0 II. a = 1 III. a > 1.
17
It×t
0t×t
..
.
0t×t
0t×t
It×t
···
···
0t×t
0t×t
0t×t
It×t
0(t−1)×t
0(t−1)×t
{z
···
···
G S1
D=
=
|
···
···
01×t
0i2 ×t
0t×t
0(t−1)×t
0(t−1)×i2
where D1 is a i1 × i1 matrix as follows
c1 c2 0
0 ··· 0
0 c1 c2 0 · · · 0
D1 = ...
0 0 0 ···
0 c1
0 0 0 ···
0
0
D4 is a i2 × i2 matrix as follows
c2 0
0 ··· 0
c1 c2 0 · · · 0
D4 = ...
0 0 ···
0 c1
0 0 ···
0
0
0
0
c2
c1
···
···
01×t
0i2 ×t
It×t
0(t−1)×t
0
0
.. ,
.
c2
c1
0
0
.. ,
.
0
c2
A
B
C
D
(12)
Case 2
01×i1 bz1 +1
01×i2
0i2 ×(i1 +1)
bz1 +1 Ii2 ×i2
..
.
bz−1 It×t
C(c1 , c2 )(t−1)×t
1
(13)
0
···
0
···
0
bz−a+1
det(F) = (bz−a+1 −bz1 +1 )bt−1
z−a+1 . Since z1 6= z −a and
then bz1 +1 6= bz−a+1 , the above matrix has full rank and
det(GSa \i ) = ±(bz−a+1 − bz1 +1 )bt−1
z−a+1 6= 0, so that
GSa \i has full rank.
Case 2: By deleting i-th column of GSa , where z1 t ≤
i < (z1 + 1)t, z − a ≤ z1 ≤ z − 3, the proof that the
resultant matrix has full rank is similar to the case that
z1 ≤ z − a − 1 and we omit it here.
Case 3: By deleting i-th column of GSa , where (z−2)t ≤
i ≤ (z −1)t−2, i1 = i−(z −2)t and i2 = (z −1)t−2−i,
the resultant matrix is as follows,
A C
GSa \i =
,
B D
where A = I(z−1)t×(z−1)t , B = 0(t−1)×(z−1)t ,
C is obtained by deleting the (i − (z − 1)t)-th
column of matrix [b1 It×t b2 It×t · · · bz−1 It×t ]T
and D is C(c1 , c2 )(t−1)×t \ ci−(z−1)t . Since
i−(z−1)t zt−i−1
det(D) = c1
c2
6= 0, det(GS1 \i ) =
i−(z−1)t zt−i−1
c2
6= 0 and therefore GS1 \i has full
±c1
rank.
Type III when a > 1, GSa has the form in eq. (14). As
before we perform a case analysis. Each of the cases is
specified by the corresponding underbrace in eq. (14).
Case 1: By deleting the i-th column of GSa , where z1 t ≤
i < (z1 + 1)t, z1 ≤ z − a − 1, i1 = i − z1 t, and i2 =
(z1 + 1)t − i − 1, the block form of the resultant matrix
GSa \i can be expressed as follows,
GSa \i =
}
C = [0i1 ×i2 0i1 ×t · · · 0i1 ×(t−1) 1i1 b1 Ii1 ×i1 0i1 ×(i2 +1) ]
and if z1 = z − a − 1, D has the form in (17).
To verify that GSa \i has full rank, we just need to check
D has full rank. Owing to the construction of D, we have
to check the determinant of the following (t+1)×(t+1)
matrix.
1
0
0
· · · bz1 +1 · · ·
0
1 bz−a+1
0
···
0
···
0
1
0
bz−a+1 · · ·
0
···
0
F=
;
..
..
.
.
and D2 and D3 are i1 × i2 and i2 × i1 all zero matrices
respectively. Then det(C(c1 , c2 )(t−1)×t \ ci1 ) = ci11 ci22
and det(GS1 \i ) = ±bz1 +1 ci11 ci22 6= 0.
Case 2: By deleting any of last t columns in GS1 , say i-th
column of GS1 , where (z − 1)t ≤ i < zt, the block form
of resultant matrix GS1 \i can be expressed as follows.
A C
,
GS1 \i =
B D
•
0(t−1)×t
Case 1
01×i2
Ii2 ×i2
..
.
0t×i2
b1 It×t
b2 It×t
..
.
bz−1 It×t
C(c1 , c2 )(t−1)×t
|
{z
}
where
A = I(z−a)t×(z−a)t
B = 0(k−(z−a)t)×(z−a)t
0t×t · · · 0t×(t−2)
..
C= .
0t×t · · ·
0t×t
D=
,
where A = Ii×i , B = 0(k−i)×i , C and D has the form
in eq. (15) and eq. (16), respectively. Note that if z1 = 0,
It×t
.
.
.
0(t−1)×t
0t×(t−2)
···
···
···
1t
1t
b1 It×t
..
.
bz−a It×t
0t×(t−2)
0t×(t−2)
Ii1 ×i1
01×i1
0i2 ×i1
0i1 ×i2
01×i2
Ii2 ×i2
1t
1t
.
.
.
bz−a+1 It×t
bz−a+2 It×t
1t−1
C(c1 , c2 )(t−1)×t
To verify GSa \i has full rank, we need to check the
determinant of D. Owing to the construction of D, the
18
GSa
It×t
..
.
0t×t
0t×t
0t×t
.
..
0t×t
=
0(t−1)×t
|
···
···
···
···
···
···
{z
0t×t
..
.
0t×t
0t×t
It×t
.
..
0t×t
0(t−1)×t
0(t−1)×t
}
Case 1
C=
D=
D=
···
0t×t
..
.
It×t
0t×t
0t×t
.
..
0t×t
0t×i2
..
.
0t×i2
0i1 ×i2
···
···
···
Case 2
···
0t×t
···
···
0t×t
0i1 ×t
1t
0t×(t−1)
0i1 ×(t−1)
1t
1i1
I(t−1)×(t−1)
{z
}
Case 3
Case 4
b1 It×t
..
.
bz1 It×t
bz1 +1 Ii1 ×i1 0i1 ×(i2 +1)
b1 It×t
..
.
bz−a It×t
bz−a+1 It×t
bz−a+2 It×t
.
..
bz−1 It×t
C(c1 , c2 )(t−1)×t
{z
}
|
(15)
···
···
···
···
···
···
···
···
···
0t×t
0t×t
···
···
0t×t
It×t
···
···
0(t−1)×i2
0(t−1)×t
···
0(t−1)×t
···
I(t−1)×(t−1)
1t−1
C(c1 , c2 )(t−1)×t
01×i2
Ii2 ×i2
0t×i2
0t×i2
.
..
01×t
0i2 ×t
0t×t
0t×t
···
···
···
···
···
···
0t×t
It×t
···
···
···
···
01×(t−1)
0i2 ×(t−1)
0t×(t−1)
0t×(t−1)
1
1i2
1t
1t
0(t−1)×i2
0(t−1)×t
···
0(t−1)×t
···
I(t−1)×(t−1)
1t−1
01×i2
01×i1 bz−a
0i2 ×(i1 +1)
bz−a Ii2 ×i2
bz−a+1 It×t
bz−a+2 It×t
.
..
C(c1 , c2 )(t−1)×t
0
0
0
···
0
0
..
.
,
··· 0 0 0 ··· bz−a+1
0 c1 c2 0 ···
0
01×(t−1)
0i2 ×(t−1)
0t×(t−1)
..
.
0t×(t−1)
0t×(t−1)
(14)
Case 5
01×t
0i2 ×t
It×t
1 bz−a+1
0
··· 0 0 0 ···
1
0
bz−a+1 ··· 0 0 0 ···
1
1
0t×(t−1)
|
1t
..
.
1t
1t
1t
..
.
1t
1t−1
| {z }
01×i2
Ii2 ×i2
0t×i2
..
.
0t×i2
0t×i2
..
.
following matrix is required to be full rank,
1t
bz−a+1 It×t
0
D =
1 C(c1 , c2 )(t−1)×t (i1 )
.
=
..
···
· · · 0(t−1)×(t−1)
{z
}
|
0t×(t−1)
..
.
0t×(t−1)
0t×(t−1)
0t×(t−1)
..
.
0t×(t−1)
0t×t
..
.
0t×t
0t×t
0t×t
.
..
It×t
1
1i2
1t
01×i1 bz1 +1
01×i2
0i2 ×(i1 +1)
bz1 +1 Ii2 ×i2
bz1 +2 It×t
1t
1t
bz−a+1 It×t
bz−a+2 It×t
..
.
(16)
.
(17)
Case 5: By deleting i-th column of GSa , where (z−1)t ≤
i < zt and i1 = i−(z−1)t, the block form of the resultant
matrix GSa \i can be expressed as eq. (19).
where bs It×t \ ci1 denotes the submatrix obtained by
deleting i1 -th column of bs It×t and C(c1 , c2 )(t−1)×t \ci1
denotes the submatrix obtained by deleting i1 -th column
of C(c1 , c2 )(t−1)×t \ ci1 . To verify GSa \i has full rank,
we just need to check [1t |bz−a+1 It×t \ ci1 ] has full rank.
Since [1t |bz−a+1 It×t ] has the following form,
1 bz−a+1
0
···
0
1
0
bz−a+1 · · ·
0
..
.. ,
.
.
1
0
0
· · · bz−a+1
where C(c1 , c2 )(t−1)×t (i1 ) denotes the i1 -th row of
C(c1 , c2 )(t−1)×t , 0 ≤ i ≤ t − 2.
det D0 = det(bz−a+1 It×t ) · det(1−
C(c1 , c2 )(t−1)×t (i1 ) · (b−1
z−a+1 It×t ) · 1t )
=btz−a+1 (1 − b−1
z−a+1 (c1 + c2 ))
by deleting any column of above matrix, it is obvious that
det([1t |bz−a+1 It×t \ci1 ]) = ±bt−1
z−a+1 and det(GSa \i ) 6=
0.
Since bz−a+1 6= 0 and c1 +c2 = 0, det D0 6= 0 and D0 has
full rank. Then det(D) = btz−a+1 (1−b−1
z−a+1 (c1 +c2 )) 6=
0 and thus GSa \i is full rank.
Case 4: By deleting i-th column of GSa , where i = (z −
1)t − 1, the block form of the resultant matrix GSa \i
can be expressed as eq. (18). Evidently, det(GSa \i ) =
±btz−a+1 , so that GSa \i has full rank.
Proof of Claim 8
Let z be the least integer such that k + 1 | nz. First, we
argue that z is the least integer such that k+1 | (n+s(k+1))z.
Assume that this is not true, then there exists z 0 < z such that
19
GSa \i
GSa \i
It×t
..
.
0
t×t
= 0t×t
0
t×t
..
.
0(t−1)×t
It×t
..
.
0t×t
0t×t
0t×t
..
.
=
0(t−1)×t
···
0t×t
0t×t
···
0t×(t−1)
···
···
···
It×t
0t×t
0t×t
0t×t
0t×t
It×t
···
···
···
0t×(t−1)
0t×(t−1)
0t×(t−1)
···
0(t−1)×t
0(t−1)×t
···
I(t−1)×(t−1)
b1 It×t
..
.
bz−a It×t
bz−a+1 It×t
bz−a+2 It×t
..
.
C(c1 , c2 )(t−1)×t
···
0t×t
0t×t
···
0t×(t−1)
1t
···
···
···
It×t
0t×t
0t×t
0t×t
0t×t
It×t
···
···
···
0t×(t−1)
0t×(t−1)
0t×(t−1)
1t
1t
1t
···
0(t−1)×t
0(t−1)×t
···
I(t−1)×(t−1)
1t−1
k + 1 | (n + s(k + 1))z 0 . As n ≥ k + 1 and k + 1 | s(k + 1)z 0
this implies that k + 1 | nz 0 which is a contradiction.
Next we argue that G0 satisfies the CCP, i.e., all k × k
submatrices of each G0Sa0 , where Ta0 = {a(k + 1), · · · , a(k +
nz
1)+k} and Sa0 = {(t)n+s(k+1) |t ∈ Ta0 } and 0 ≤ a ≤ k+1
+sz,
0
are full rank. Let n = n+s(k +1). We argue it in three cases.
0
• Case 1. The first column of GS 0 lies in the first s(k + 1)
a
0
columns of G .
Suppose ln0 ≤ a(k + 1) < ln0 + s(k + 1) where 0 ≤
l < z − 1. By the construction of G0 , G0Sa = D. Since
D = GS0 , all k × k submatrices of D have full rank and
so does G0Sa .
0
• Case 2. The first and last column of GS 0 lie in the last
a
0
n columns of G .
Suppose ln0 + s(k + 1) ≤ a(k + 1) and a(k + 1) + k <
(l + 1)n0 where 0 ≤ l < z − 1. As n0 > s(k + 1), a(k +
1)−(l+1)s(k+1) > 0 and k+1|a(k+1)−(l+1)s(k+1).
Let a0 = a − (l + 1)s, then G0Sa = GSa0 and hence all
k × k submatrices of G0Sa have full rank.
0
• Case 3. The first column of GS 0 lies in the last n columns
a
0
of G but the last column lies in the first (k + 1) columns
of G0 .
Suppose ln0 + s(k + 1) ≤ a(k + 1) and a(k + 1) +
k > (l + 1)n0 where 0 ≤ l < z − 2. Again, we can get
k + 1|a(k + 1) − (l + 1)s(k + 1) and let a0 = a − (l + 1)s.
Let S 01 = {(a(k + 1))n0 , · · · , (ln0 + n0 − 1)n0 } and S 1 =
{(a0 (k+1))n , · · · , (ln+n−1)n }. As (ln0 +n0 −1)−a(k+
1) = (ln + n − 1) − a0 (k + 1), G0S 01 = GS 1 . Let S 02 =
{(ln0 + n0 )n0 , · · · , (a(k + 1) + k)n0 } and S 2 = {(ln +
n)n , · · · , (a0 (k + 1) + k)n }. By the construction of G0 ,
GS 2 = G0S 02 . Then G0Sa = [G0S 01 G0S 02 ] = [GS 1 GS 2 ] =
GSa0 and hence all k × k submatrices of G0Sa have full
rank.
b1 It×t \ ci1
..
.
bz−a It×t \ ci1
bz−a+1 It×t \ ci1
bz−a+2 It×t \ ci1
..
.
C(c1 , c2 )(t−1)×t \ ci1
(18)
(19)
Using the fact that z ≤ k+1 and taking limits as n → ∞,
we get that
1
FsM N
1
η
lim
log2
= H2
− log2 q.
∗
n→∞ K
Fs
q
q
•
M
N
=1−
k+1
nq .
We have
1
F MN
1
K
k+1
log2 s ∗ = log2
−
log2 q
K
Fs
K
k+1
K
1
zn
−
log2
.
K
k+1
Using the fact that z ≤ k+1 and taking limits as n → ∞,
we get that
1
F MN
η
η
lim
log2 s ∗ = H2
− log2 q.
n→∞ K
Fs
q
q
B. Discussion on coded caching systems constructed by generator matrices satisfying the (k, α)-CCP where α ≤ k
Consider the (k, α)-CCP (cf. Definition 5) where α ≤ k.
Let z be the least integer such that α | nz, and let Taα =
{aα, · · · , aα + α − 1)} and Saα = {(t)n | t ∈ Taα }. Let
GSaα = [gi0 , · · · , giα−1 ] be the submatrix of G specified by
the columns in Saα , i.e, gij ∈ GSaα if ij ∈ Saα . We demonstrate
that the resolvable design generated from a linear block code
that satisfies the (k, α)-CCP can also be used in a coded
caching scheme. First, we construct a (X, A) resolvable design
as described in Section III.A., which can be partitioned into n
parallel classes Pi = {Bi,j : 0 ≤ j < q}, 0 ≤ i < n. By the
constructed resolvable design, we partition each subfile Wn
s
into q k z subfiles Wn = {Wn,t
| 0 ≤ t < q k , 0 ≤ s < z} and
operate the placement scheme in Algorithm 1. In the delivery
phase, for each recovery set, several equations are generated,
each of which benefit α users simultaneously. Furthermore,
the equations generated by all the recovery sets can recover
Proof of Claim 11
M
1
all the missing subfiles. In this section, we only show that
• N = q . We have
for the recovery set PSaα , it is possible to generate equations
1
FsM N
1
K
1
k
which benefit α users and allow the recovery of all of missing
log2
=
log2
−
log2 z −
log2 q.
subfiles with given superscript. The subsequent discussion
K
Fs∗
K
K/q
K
K
20
By the assumed condition, it is evident that this system of α0
equations in α0 variables has a unique solution for a given
vector v = [uα0 , · · · , uk−1 ] over GF (q). Since there are
0
q k−α possible v vectors, the result follows.
exactly mirrors the discussion in the (k, k + 1)-CCP case and
is skipped.
Towards this end, we first show that picking α users from
α distinct parallel classes can always form q k−α+1 − q k−α
signals. More specifically, consider blocks Bi1 ,li1 , . . . , Biα ,liα
(where lij ∈ {0, . . . , q − 1}) that are picked from α distinct
k−α+1
parallel classes of PSaα . Then, | ∩α−1
and
j=1 Bij ,lij | = q
α
k−α
| ∩j=1 Bij ,lij | = q
.
As in the case of the (k, k +1)-CCP, we form a recovery set
bipartite graph with parallel classes and recovery sets as the
disjoint vertex subsets, and the edges incident on each parallel
class are labeled arbitrarily from 0 to z −1. For a parallel class
P ∈ PSaα we denote this label by label(P − PSaα ). For a given
recovery set PSaα , the delivery phase proceeds by choosing
blocks from α distinct parallel classes in PSaα and it provides
q k−α+1 −q k−α equations that benefit α users. Note that in the
(k, α)-CCP case, randomly picking α blocks from α parallel
classes in PSaα will always result in q k−α intersections, which
is different from (k, k +1)-CCP. It turns out that each equation
allows a user in P ∈ PSaα to recover a missing subfile with
superscript label(P − PSaα ).
Let the demand of user UBi,j for i ∈ n−1, 0 ≤ j ≤ q−1 by
Wκi,j . We formalize the argument in Algorithm 4 and prove
that equations generated in each recovery set PSaα can recover
all missing subfile with superscript label(P − PSaα ).
Claim 12. Consider the resolvable design (X, A) constructed
by a (n, k) linear block code that satisfies the (k, α) CCP. Let
PSaα = {Pi | i ∈ Saα } for 0 ≤ a < zn
α , i.e., it is the set
of parallel classes corresponding to Saα . We emphasize that
|PSaα | = α ≤ k. Consider blocks Bi1 ,li1 , . . . , Biα0 ,li 0 (where
α
lij ∈ {0, . . . , q − 1}) that are picked from any α0 distinct
0
parallel classes of PSaα where α0 ≤ α. Then, | ∩α
j=1 Bij ,lij | =
0
q k−α .
The above argument implies that any α −1 blocks from any
α − 1 distinct parallel classes of PSaα have q k−α+1 points in
common and any α blocks Bi1 ,li1 , Biα ,liα from any α distinct
parallel classes of PSaα have q k−α points in common. These
blocks (or users) can participate in q k−α+1 − q k−α equations,
each of which benefits α users. In particular, each user will
recover a missing subfile indexed by an element belonging to
the intersection of the other α − 1 blocks in each equation.
A very similar argument to Lemma 2 can be made to justify
enough equations can be found that allow all users to recover
all missing subfiles.
Algorithm 4: Signal Generation Algorithm for PSaα
1
2
Proof. Recall that by the construction in Section III.A, block
Bi,l ∈ Pi is specified as follows,
/* Pick blocks from distinct parallel classes
in PSaα . The cardinality of their
intersection is always q k−α */
Bi,l = {j : Ti,j = l}.
Let G = [gab ], for 0 ≤ a < k, 0 ≤ b < n.
Now consider Bi1 ,li1 , . . . , Biα0 ,li 0 (where ij ∈ Saα , lij ∈
α
{0, . . . , q −1}) that are picked from α0 distinct parallel classes
of PSaα . W.l.o.g. we assume that i1 < i2 < · · · < iα0 . Let
I = {i1 , . . . , iα0 } and TI denote the submatrix of T obtained
by retaining the rows in I. We will show that the vector
0
[li1 li2 . . . liα0 ]T is a column in TI and appears q k−α times
in it.
We note here that by the (k, α)-CCP, the vectors
gi1 , gi2 , . . . , giα are linearly independent and thus the subset
of these vectors, gi1 , · · · , giα0 are linearly independent. W. l.
o. g., we assume that the top α0 × α0 submatrix of the matrix
[gi1 gi2 . . . giα0 ] is full-rank. Next, consider the system of
equations in variables u0 , . . . , uα0 −1 .
0
αX
−1
b=0
0
αX
−1
b=0
ub gbi1 = li1 −
ub gbi2 = li2 −
k−1
X
3
b=0
Find set L̂s = ∩j∈Saα \{s} Bj,lj \ ∩j∈Saα Bj,lj for
s ∈ Saα ;
/* Determine the missing subfile indices that
the user from Psα will recover. Note that
|L̂s | = q k−α+1 − q k−α */
4
E(Ps )
,
s,ls ,L̂s [t]
k−α
Add signals ⊕s∈Saα Wκ
0 ≤ t < q k−α+1 − q
, to Sig;
/* User UBs,ls demands file Wκs,ls . This
equation allows it to recover the
corresponding missing subfile index L̂s [t],
which is the t-th element of L̂s [t]. The
superscript is determined by the recovery
set bipartite graph */
5
ub gbi1 ,
end
Output: Signal set Sig.
b=α0
k−1
X
For the sake of convenience we argue that user UBβ,lβ
that demands Wκβ,lβ can recover all its missing subfiles with
superscript E(Pβ ). Note that Bβ,lβ = q k−1 . Thus user UBβ,lβ
needs to obtain q k − q k−1 missing subfiles with superscript
E(Pβ ). The delivery phase scheme repeatedly picks α users
from different parallel classes of PSaα . The equations in AlgoE(P )
rithm 4 allow UBβ,lβ to recover all Wκ β,L̂ [t] where L̂β =
ub gbi2 ,
b=α0
..
.
0
αX
−1
Input : For P ∈ PSaα , E(P) = label(P − PSaα ). Signal
set Sig = ∅.
while any user UB ∈ Pj , j ∈ Saα does not recover all its
missing subfiles with superscript E(P) do
Pick blocks Bj,lj ∈ Pj for all j ∈ Saα and
lj ∈ {0, . . . , q − 1};
ub gbiα0 = liα0 −
k−1
X
ub gbiα0 .
b=α0
β,lβ
21
β
∩j∈Saα \{β} Bj,lj \ ∩j∈Saα Bj,lj and t = 1, · · · , q k−α+1 − q k−α .
This is because of Claim 12.
Next, we count the number of equations that UBβ,lβ participates in. We can pick α − 1 users from α − 1 parallel classes
in PSaα . There are totally q α−1 ways to pick them, each of
which generate q k−α+1 − q k−α equations. Thus there are a
total of q k − q k−1 equations in which user UBβ,lβ participates
in.
It remains to argue that each equation provides a
distinct file part of user UBβ,lβ . Towards this end,
let {i1 , · · · , iα−1 } ⊂ Saα be an index set such that
β∈
/ {i1 , · · · , iα−1 } but β ∈ Saα . Note that when we pick the
same set of blocks {Bi1 ,li1 , · · · , Biα−1 ,liα−1 }, it is impossible
E(Pβ )
β,lβ ,L̂β [t1 ]
that the recovered subfiles Wκ
E(Pβ )
β,lβ ,L̂β [t2 ]
and Wκ
where each c(i) is a codeword of C i (the cyclic code over
GF (qi )). Thus, the b-th component ∆b can be mapped to
(1) (2)
(d)
(cb , cb , . . . , cb )
(i)
i
Let G = [gab ] represent the generator matrix of the code
C i . Based on prior arguments, it is evident that there are qiki −1
Pki −1
(i)
distinct solutions over GF (qi ) to the equation a=0
ua gab =
(i)
cb . In turn, this implies that ∆b appears q1k1 −1 q2k2 −1 · · · qdkd −1
times in the b-th row of T and the result follows.
Next we show any α blocks from distinct parallel classes of
PS kmin have q1k1 −α q2k2 −α · · · qdkd −α intersections, where α ≤
a
kmin and Sakmin = {(akmin )n , (akmin + 1)n , · · · , (akmin +
kmin − 1)n }
Towards this end consider Bi1 ,li1 , . . . , Biα ,liα (where ij ∈
Sakmin , lij ∈ {0, . . . , q − 1}) that are picked from α distinct
parallel classes of PS kmin . W.l.o.g. we assume that i1 < i2 <
a
· · · < iα . Let I = {i1 , . . . , iα } and TI denote the submatrix
of T obtained by retaining the rows in I. We will show that
the vector [li1 li2 . . . liα ]T is a column in TI and appears
q1k1 −α q2k2 −α · · · qdkd −α times.
Let ψm (lij ) for m = 1, . . . , d represent the m-th component
of the map ψ. Consider the (n, k1 ) cyclic code over GF (q1 )
and the system of equations in variables u0 , . . . , uα−1 that lie
in GF (q1 ).
are
the same since the points in L̂β are distinct. Next, suppose
that there exist sets of blocks {Bi1 ,li1 , · · · , Biα−1 ,liα−1 }
and
{Bi1 ,li0 , · · · , Biα−1 ,li0
}
such
that
1
α−1
{Bi1 ,li1 , · · · , Biα−1 ,liα−1 } 6= {Bi1 ,li0 , · · · , Biα−1 ,li0
},
1
α−1
α−1
0 .
but γ ∈ ∩α−1
j=1 Bij ,lij \ Bβ,lβ and γ ∈ ∩j=1 Bij ,li0 j \ Bβ,lβ
This is a contradiction
T α since this in turn implies that
γ ∈ ∩α
∩j=2 Bij ,li0 , which is impossible since
j=2 Bij ,lij
j
two blocks from the same parallel class have an empty
intersection.
Finally we calculate the transmission rate. In Algorithm 4,
for each recovery set, we transmit q k+1 − q k equations and
there are totally zn
α recovery sets. Since each equation has size
equal to a subfile, the rate is given by
R = (q k+1 − q k ) ×
α−1
X
b=0
α−1
X
zn
1
× k
α
zq
b=0
(1)
kX
1 −1
(1)
b=α
kX
1 −1
ub gbi1 = ψ1 (li1 ) −
ub gbi2 = ψ1 (li2 ) −
(1)
ub gbi1 ,
(1)
ub gbi2 ,
b=α
..
.
n(q − 1)
.
α
The (n, k) linear block codes that satisfy the (k, α)-CCP
over GF (q) correspond to a coded caching system with K =
n(q−1)
1
k
nq, M
. Thus, the
N = q , Fs = zq and have a rate R =
α
rate of this system is a little higher compared to the (k, k +1)CCP system with almost the same subpacketization level.
However, by comparing Definitions 4 and 5 it is evident that
the rank constraints of the (k, α)-CCP are weaker as compared
to the (k, k + 1)-CCP. Therefore, in general we can find more
instances of generator matrices that satisfy the (k, α)-CCP. For
example, a large class of codes that satisfy the (k, k)-CCP are
(n, k) cyclic codes since any k consecutive columns in their
generator matrices are linearly independent [23]. Thus, (n, k)
cyclic codes always satisfy the (k, k)-CCP but satisfy (k, k +
1)-CCP if they satisfy the additional constraints discussed in
Claim 3.
=
α−1
X
b=0
(1)
ub gbiα = ψ1 (liα ) −
kX
1 −1
(1)
ub gbiα .
b=α
By arguments identical to those made in Claim 12 it can be
seen that this system of equations has q1k1 −α solutions. Applying the same argument to the other cyclic codes we conclude
that the vector [li1 , li2 , · · · , liα ] appears q1k1 −α q2k2 −α · · · qdkd −α
times in TI and the result follows.
C. Cyclic codes over Z mod q [25]
First, we show that matrix T constructed by constructed
by the approach outlined in Section IV-D still results in a
resolvable design. Let ∆ = [∆0 ∆1 · · · ∆n−1 ] be a codeword of the cyclic code over Z mod q, denoted C where
q = q1 q2 · · · qd , and qi , i = 1, . . . , d are prime. By using the
Chinese remaindering map ψ (discussed in Section IV-D), ∆
can be uniquely mapped into d codewords c(i) , i = 1, . . . , d
Li Tang received his B.E. degree in mechanical engineering from Beihang
University, Beijing, China in 2011, and M.S. degree in electrical and information engineering from Beihang University, Beijing, China in 2014. He is
currently working towards the Ph.D degree in the Department of Electrical
and Computer Engineering at Iowa State University, Ames, IA, USA. His
research interests include network coding and channel coding.
22
Aditya Ramamoorthy (M’05) received the B.Tech. degree in electrical
engineering from the Indian Institute of Technology, Delhi, in 1999, and
the M.S. and Ph.D. degrees from the University of California, Los Angeles
(UCLA), in 2002 and 2005, respectively. He was a systems engineer with
Biomorphic VLSI Inc. until 2001. From 2005 to 2006, he was with the Data
Storage Signal Processing Group of Marvell Semiconductor Inc. Since fall
2006, he has been with the Electrical and Computer Engineering Department
at Iowa State University, Ames, IA 50011, USA. His research interests
are in the areas of network information theory, channel coding and signal
processing for bioinformatics and nanotechnology. Dr. Ramamoorthy served
as an editor for the IEEE Transactions on Communications from 2011 – 2015.
He is currently serving as an associate editor for the IEEE Transactions on
Information Theory. He is the recipient of the 2012 Early Career Engineering
Faculty Research Award from Iowa State University, the 2012 NSF CAREER
award, and the Harpole-Pentair professorship in 2009 and 2010.
23
| 7cs.IT
|
Journal of Machine Learning Research X (2017) XX-XX
Submitted 12/17; Published XX/XX
f -Divergence constrained policy improvement
arXiv:1801.00056v1 [cs.LG] 29 Dec 2017
Boris Belousov
Jan Peters
[email protected]
[email protected]
Department of Computer Science
Technische Universität Darmstadt
FG IAS, Hochschulstr. 10, 64289 Darmstadt, Germany
Editor: XX
Abstract
To ensure stability of learning, state-of-the-art generalized policy iteration algorithms augment the policy improvement step with a trust region constraint bounding the information
loss. The size of the trust region is commonly determined by the Kullback-Leibler (KL)
divergence, which not only captures the notion of distance well but also yields closed-form
solutions. In this paper, we consider a more general class of f -divergences and derive the
corresponding policy update rules. The generic solution is expressed through the derivative
of the convex conjugate function to f and includes the KL solution as a special case. Within
the class of f -divergences, we further focus on a one-parameter family of α-divergences to
study effects of the choice of divergence on policy improvement. Previously known as well
as new policy updates emerge for different values of α. We show that every type of policy
update comes with a compatible policy evaluation resulting from the chosen f -divergence.
Interestingly, the mean-squared Bellman error minimization is closely related to policy
evaluation with the Pearson χ2 -divergence penalty, while the KL divergence results in the
soft-max policy update and a log-sum-exp critic. We carry out asymptotic analysis of
the solutions for different values of α and demonstrate the effects of using different divergence functions on a multi-armed bandit problem and on common standard reinforcement
learning problems.
Keywords: Reinforcement Learning, Policy Search, Bandit Problems
1. Introduction
Many state-of-the art reinforcement learning algorithms, including natural policy gradients (Kakade, 2001; Bagnell and Schneider, 2003; Peters et al., 2003), trust region policy optimization (TRPO) (Schulman et al., 2015) and relative entropy policy search (REPS) (Peters et al., 2010), impose a Kullback-Leibler (KL) divergence constraint between successive
policies during parametric policy iteration to avoid large steps towards unknown regions of
the state space. Similar objective functions with an entropy-like term have been proposed in
the context of linearly-solvable optimal control (Kappen, 2005; Todorov, 2006) and inverse
reinforcement learning (Ziebart et al., 2008). In all these approaches, the objective function
has the form of the free energy, therefore they all can be viewed as performing free energy
minimization (Still and Precup, 2012).
In this paper, we explore the implications of using a generic f -divergence to constrain the
policy improvement. The objective function in this case resembles the free energy objective,
c 2017 Boris Belousov and Jan Peters.
License: CC-BY 4.0, see https://creativecommons.org/licenses/by/4.0/. Attribution requirements are provided
at http://jmlr.org/papers/vX/XXX.html.
Belousov and Peters
commonly encountered in variational methods (Wainwright and Jordan, 2007), but with the
KL divergence replaced by an f -divergence. The idea of using f -divergence penalties for
general non-linear problems goes back to the early work by Teboulle (1992). Generalizations
of the KL divergence—apart from being very useful at providing a unified view of existing algorithms, e.g., statistical inference can be seen as f -divergence minimization (Altun
and Smola, 2006) and various message passing algorithms can be understood as minimizing
different α-divergences (Minka, 2005),—pave the way to new algorithms that are better
suited for particular problems—e.g., the generative adversarial networks (GANs) (Goodfellow et al., 2014) minimize a particular f -divergence, the Jensen-Shannon divergence,
which is well-suited for measuring similarity between real-world images (Nowozin et al.,
2016). This paper is, to our knowledge, the first to shed light on policy improvement in
reinforcement learning from the f -divergence point of view.
2. Background
This section provides the background on the f -divergence, the α-divergence, and the convex
conjugate function, highlighting the key properties required for our derivations.
The f -divergence (Csiszár, 1963; Morimoto, 1963; Ali and Silvey, 1966) generalizes
many similarity measures between probability distributions (Sason and Verdu, 2016). For
two distributions π and q on a finite set A, the f -divergence is defined as
X
π(a)
Df (πkq) ,
q(a)f
,
q(a)
a∈A
y-axis
where f is a convex function on (0, ∞) such that f (1) = 0. For example, the KL divergence
corresponds to fKL (x) = x log x. Note that π must be absolutely continuous with respect to q to avoid division by zero, 4 y = α(α − 1) 2
Neyman χ
Pearson χ2
i.e., q(a) = 0 implies π(a) = 0 for all a ∈ A. We addition- 3
ally assume f to be continuously differentiable, which includes
2
all cases of interest for us. The f -divergence can be general1
ized to unnormalized distributions. For example, the generKL
alized KL divergence (Zhu and Rohwer, 1995) corresponds to 0 Reverse KL
f1 (x) = x log x − (x − 1). The derivations in this paper bene-−1
Hellinger distance
0
1
2
3
fit from employing unnormalized distributions and subsequently −2 −1 Parameter
α
imposing the normalization condition as a constraint.
Figure 1: The α-divergence
The α-divergence (Chernoff, 1952; Amari, 1985) is a one- smoothly connects several
parameter family of f -divergences generated by the α-function prominent divergences.
fα (x) with α ∈ R. The particular choice of the family of functions fα is motivated by generalization of the natural logarithm (Cichocki and Amari, 2010).
The α-logarithm logα (x) , (xα−1 − 1)/(α − 1) is a power function for α 6= 1 that turns
into the natural logarithm for α → 1. Replacing the natural logarithm in the derivative of
the KL divergence f10 = log x by the α-logarithm and integrating fα0 under the condition
that fα (1) = 0 yields the α-function
fα (x) ,
(xα − 1) − α(x − 1)
.
α(α − 1)
2
(1)
f -Divergence constrained policy improvement
Divergence
KL
Reverse KL
2
α
1
0
Pearson χ
2
Neyman χ2
−1
Hellinger
1
2
f 0 (x)
f (x)
x log x − (x − 1)
− log x + (x − 1)
log x
1 +1
−x
1 (x − 1)2
2
(x−1)2
2x
2
√
(f ∗ )0 (y)
x−1
− 12 + 1
2
2x
2 − √2
x−1 2
x
y
e
1
1−y
y+1
√ 1
1−2y
4
(2−y)2
f ∗ (y)
dom f ∗
y
e −1
− log(1 − y)
1 (y
2
2
R
y<1
1
2
+ 1) −
√
− 1 − 2y + 1
2y
2−y
y > −1
y< 1
2
y<2
Table 1: Function fα , its convex conjugate fα∗ , and their derivatives for some values of α
The α-divergence generalizes the KL divergence, reverse KL divergence, Hellinger distance,
Pearson χ2 -divergence, and Neyman (reverse Pearson) χ2 -divergence. Figure 1 displays
well-known α-divergences as points on the parabola y = α(α − 1). For every divergence,
there is a reverse divergence symmetric with respect to the point α = 0.5, corresponding to
the Hellinger distance.
The convex conjugate of f (x) is defined as f ∗ (y) = supx∈dom f {hy, xi − f (x)}, where
the angle brackets hy, xi denote the dot product (Boyd and Vandenberghe, 2004). The
key property (f ∗ )0 = (f 0 )−1 relating the derivatives of f ∗ and f yields Table 1, which lists
common functions fα together with their convex conjugates and derivatives. In the general
case (1), the convex conjugate and its derivative are given by
α
1
1
(1 + (α − 1)y) α−1 − ,
αp
α
(fα∗ )0 (y) = α−1 1 + (α − 1) y,
fα∗ (y) =
for y(1 − α) < 1.
(2)
Function fα is convex, non-negative, and attains minimum at x = 1 with fα (1) = 0.
Function (fα∗ )0 is positive on its domain with (fα∗ )0 (0) = 1. Function fα∗ has the property
fα∗ (0) = 0. The linear inequality constraint (2) on the dom fα∗ follows from the requirement dom fα = (0, ∞). Another result from convex analysis crucial to our derivations is
Fenchel’s equality
f ∗ (y) + f (x? (y)) = hy, x? (y)i,
(3)
where x? (y) = arg supx∈dom f {hy, xi − f (x)}. We will occasionally put the conjugation
symbol at the bottom, especially for the derivative of the conjugate function f∗0 , (f ∗ )0 .
3. Policy improvement with f -divergence for multi-armed bandits
In order to develop intuition regarding the influence of entropic penalties on policy improvement, we first consider a simplified version of the reinforcement learning problem—namely,
the stochastic multi-armed bandit problem (Bubeck and Cesa-Bianchi, 2012). The resulting
algorithm is closely related to the family of Exp3 algorithms (Auer et al., 2003), which are
extremely good on adversarial bandit problems but have fallen behind the state-of-the-art
for stochastic bandit problems. We nevertheless focus on the stochastic bandit problems
to illustrate crucial effects of entropic penalties on policy improvement in a clean intuitive
way. This section is meant to serve as a foundation for the reinforcement learning section.
At every time step t ∈ {1, . . . , T }, an agent chooses among K actions a ∈ A and after
every choice at = a receives a noisy reward rt = r(at , t ) drawn from a distribution with
3
Belousov and Peters
mean Q(a)
P = E [r(a, )]. The goal of the agent is to maximize the expected total reward
J , E[ Tt=1 rt ]. If the agent knew the true value of each action Q(a), the optimal strategy
would be to always choose the action with the highest value, a∗t = arg maxa Q(a). However,
since the agent does not know Q in advance, it faces the exploration-exploitation dilemma
at every time step. It can either pick the best action according to the current estimate of
action values Q̃t or try an exploratory action that may turn out to be better than arg max Q̃t .
A generic way to encode the exploration-exploitation trade-off is by introducing a policy
πt , that is, a distribution over a ∈ A from which the agent draws an action at ∼ πt .
Thus, the question becomes: given the current policy πt and the current estimate of action
values Q̃t , what should the policy πt+1 at the next time step be? Unlike the choice of the
best action under perfect information, such sampling policies are hard to derive from first
principles (Ghavamzadeh et al., 2015).
In this section, we formalize the exploration-exploitation trade-off by choosing the next
stochastic policy to maximize the expectation of Q̃t while regularizing the change in the policy through the f -divergence. This formalization always yields both a convex optimization
problem and, as its solution, the next stochastic policy. For the α-divergence, we obtain
both classical policy updates (e.g., -greedy, softmax, etc.) as well as novel ones resulting
from different α’s. Empirical evaluations on a simulated multi-armed bandit environment
illustrate the differences in policy updates corresponding to various penalty functions.
3.1 f -Divergence constrained policy optimization problem for bandits
A policy improvement at time step t is derived by maximizing the expected return J(π) ,
Ea∼π [Q(a)] while punishing the policy change between the old policy q(a) , πt (a) and the
new policy π(a) , πt+1 (a) by an f -divergence as a kind of regularization. This regularization
is added to the objective function weighted by a ‘temperature’ η to trade-off exploration
vs exploitation. For simplicity, the current estimate of action values is denoted by Q(a) ,
Q̃t (a). A policy π that achieves a higher expected return than q is a solution of the following
optimization problem
X
X
π(a)
max Jη (π) =
Q(a)π(a) − η
q(a)f
,
π
q(a)
a
a
X
(4)
s.t.
π(a) = 1,
a
π(a) ≥ 0, ∀a ∈ A.
Different formulations of Problem (4) with the KL divergence in place of the f -divergence
have been proposed. An inequality constraint (Peters et al., 2010) bounding the information loss by certain fixed amount can automatically yield an η as a Lagrange multiplier
found during optimization. Alternating between maximization of the expected return at
fixed information loss and subsequent information loss minimization at constant expected
return (Still and Precup, 2012) yields algorithms that allow for automatic adaptation of .
In (4), the divergence is added as a penalty, as suggested by Azar and Kappen (2012). Since
the treatment of these three mathematical problems differs only in the sets of hyperparameters, our results straightforwardly generalize to the other two formulations.
4
f -Divergence constrained policy improvement
Differentiating the Lagrangian of problem (4)
Lη (π, λ, κ) =
X
a
Q(a)π(a) − η
X
a
q(a)f (π(a)/q(a)) − λ
X
a
π(a) − 1
!
+
X
κ(a)π(a)
a
with respect to π(a) and equating the derivative to zero, we find an expression for the
optimal new policy π ? through the derivative of the convex conjugate function
Q(a) − λ + κ(a)
?
0
.
(5)
π (a) = q(a)f∗
η
Substituting π ? back into the Lagrangian and using Fenchel’s equality (3), we arrive at the
dual problem
X
∗ Q(a) − λ + κ(a)
min g(λ, κ) = η
q(a)f
+ λ,
λ, κ
η
a
(6)
s.t. κ(a) ≥ 0, ∀a ∈ A,
arg f ∗ ∈ rangex ≥ 0 f 0 (x),
∀a ∈ A.
The constraint on the argument of the convex conjugate has a particularly simple linear
form (2) for the α-function fα (y); see Section 2 for the necessary background. In the very
special KL divergence case it completely disappears because range f10 = R, see Table 1 for
well-known divergences. In the multi-armed bandit scenario, one is free to choose whether
to solve the optimization problem in the primal form (4) or in the dual form (6). However, as
we will see later, the dual formulation (6) is much better suited for reinforcement learning.
When the Lagrange multipliers κ disappear. When choosing function f , it is important to consider whether 0 ∈ dom f 0 (x), as it implies that the improved policy can put zero
mass on some actions. Although f is not required to have a finite value at x = 0 by definition, for many divergences it does (e.g., for any α-divergence for α > 1) or it has a finite
limit (e.g., for the KL divergence). However, even those functions f that are defined at zero
do not necessarily have a finite derivative at zero (e.g., for the KL-divergence f10 (x) = log x).
If 0 ∈
/ dom f 0 (x), the new policy π cannot put zero probability mass on actions that were
possible under the old policy q. In other words, if q(a) > 0, then π(a) > 0, and the complementary slackness condition κ(a)π(a) = 0 implies κ(a) = 0. Therefore, we can ignore the
non-negativity constraint if f 0 has a singularity at x = 0. Thereby, one commonly omits
the non-negativity constraint when using the KL divergence or the reverse KL. In general,
κ(a) = 0 for any fα with α ≤ 1. On the other hand, if 0 ∈ dom f 0 (x), some actions may get
zero probability under the new policy, and κ(a) 6= 0. This case emerges for fα with α > 1.
3.2 Effects of the exploration-exploitation-tradeoff and the divergence type
The exploration-exploitation parameter η controls the step size, while the divergence type α
affects the search direction.
Greedy policy vs no change: effects of the temperature parameter η. Independent of the f -divergence chosen, the temperature parameter η controls the magnitude of
deviation of the new policy from the old one. The optimal Lagrange multiplier λ? plays
5
Belousov and Peters
the role of the baseline similar as in policy gradient methods (Sutton et al., 2000; Peters
and Schaal, 2006). For low temperatures η → 0, the policy π ? (a) → δ(a, arg maxa Q(a))
?
tends towards greedy action selection (Sutton and Barto, 1998)
Pwith λ → maxa Q(a).
?
?
For η → ∞, both the policy π (a) → q(a) and the baseline λ → a q(a)Q(a) do no longer
change. For other values of η, the agent chooses an intermediate solution between focusing
on a single best action and remaining close to the old policy.
From -greedy to -elimination: effects of α on policy improvement. Given the
functional form of the policy update (5), we can find the Lagrange multipliers λ and κ
and thus obtain a closed-form solution of Problem (4) in two cases: when f∗0 is exponential or linear, which corresponds to α = 1 and α = 2. For other α’s, we have to
solve the optimization problem numerically. Table 2 summarizes the Table 2: Effects of α on
effects of α on policy improvement. For α 0, we recover the - the policy ∀η > 0
greedy action selection method (Sutton and Barto, 1998), which
Policy π
α
assigns probability 1− to the best action and spreads the remaining
greedy
−∞
probability mass uniformly among the rest; parameter η is propor-greedy
α 0
tional to , see Section 3.2. Values of α ≤ 1 correspond to -soft poli-soft
α≤1
soft-max
α=1
cies (Sutton and Barto, 1998); parameter α controls the distribution
linear
α=2
of probability mass among suboptimal actions: large negative values
-elimin.
α 1
of α correspond to more uniform distributions, whereas smaller α’s
elimination +∞
distribute the probability mass according to values. When α = 1,
we arrive at a form of the soft-max policy (Bridle, 1990) with the log-sum-exp baseline
1
?
π (a) = q(a)e
1
(Q(a)−λ? )
η
Q(a)
q(a)e η
=P
,
1
Q(b)
η
b q(b)e
P
with λ? = η log a q(a) exp(Q(a)/η), where the regular soft-max arises for a uniform distribution q(a) = U(a) over all arms. That the soft-max policy results from using the KL
divergence as a measure of distance has been suggested several times in the literature (Azar
and Kappen, 2012; Peters et al., 2010; Still and Precup, 2012). Further increasing α to
α = 2, we obtain a linear policy
Q(a) − λ? + κ(a)
π ? (a) = q(a)
+1 ,
η
X
?
λ =
q(a)(Q(a) + κ(a)).
a
As pointed out in Section 3.1, some actions may get zero probability under π ? when α > 1.
Therefore, in general one needs to keep track of the Lagrange multipliers κ. However, we
can simplify analysis by dropping κ for sufficiently large η. Indeed, if q(a) > 0 for all a ∈ A,
there exists a temperature η > ηmin for which π ? (a) > 0, and thus the inequality constraints
are inactive. This minimum temperature
ηmin = | mina A(a)|
is defined in terms of the advantage function
A(a) , Q(a) − J(q) = Q(a) −
6
P
b q(b)Q(b)
f -Divergence constrained policy improvement
which measures agent’s gain from taking action a instead of following policy q. The baseline
λ? equals the expected return J(q) under policy q when κ = 0. Thus, for any η > ηmin , we
obtain a linearly advantage-reweighted policy
1 + A(a)
?
π (a) = q(a)
.
η
For α 1, we obtain a new policy, that we call -elimination. Under this policy, all
actions get equal probability except for the set of the worst actions that gets probability
zero; the size of the set depends on α and η.
From Qmax to Qmin : effects of α on the baseline. To complete the picture, we
describe the asymptotic behavior of λ? when the divergence type α is varied. Consider the
limit α → −∞ of large negative α. As established in Section 3.1, Lagrange multipliers κ
vanish for α < 1. For sufficiently large α, the solution hits the inequality constraint in (6),
from which the lower bound on λ?
Table 3: Effects of α on
λ? ≥ Q(a) − η,
the baseline ∀η > 0
∀a ∈ A
with = 1/(1 − α) can be obtained. Given that it holds for all a,
it also holds for the maximum over a. Since the solution lies at the
constraint, in the limit one obtains
n
o
lim λ? = lim max Q(a) − η = max Q(a).
α→−∞
→0
a
a
Baseline λ
α
Qmax
(Q̄, Qmax )
log-sum-exp
Q̄
(Qmin , Q̄]
Qmin
−∞
α≤1
α=1
α=2
α>1
+∞
Recall from Section 3.2 that for a fixed α the baseline λ? shifts
towards the interior of the feasible set when the temperature η is increased, i.e., limη→∞ λ? =
Ea∼q [Q(a)] , Q̄. Therefore, the optimal baseline λ? for any divergence α < 1 lies in the
range λ? ∈ (Q̄, Qmax ). The same analysis for α > 1 yields λ? ∈ (Qmin , Q̄]. Table 3
summarizes these asymptotic results.
3.3 Empirical evaluation of policy improvement with f -divergence in bandits
The effects of the divergence type are demonstrated on a simulated stochastic multi-armed
bandit problem. Policy improvement in a single iteration as well as regret over a period of
time are compared for different values of α.
Regret
Effects of α on policy improvement. Figure 3 shows the effects of α-divergence on
policy update. We consider a 10-armed bandit problem with arm values Q(a) ∼ N (0, 1).
The temperature is kept fixed at η = 2 for all values of α. Initial
150
policy is uniform and the arm values Q(a) are fixed. Several iterations are shown in the figure for comparison. Extremely large
100
α=0
positive and negative values of α correspond to -elimination and
α = 20
50
α = −10
-greedy policies respectively. Small values of α, in contrast,
α = 10
α = −20
weigh actions according to their values. Policies for α < 1 are
UCB
0
0
200
400
600
800
peaked and heavy-tailed, eventually turning into -greedy poliTime step
cies when α → −∞. Policies for α ≥ 1 are more uniform, but Figure 2: Average regret
they put zero mass on bad actions, eventually turning into - for various values of α.
elimination policies when α → ∞. For α ≥ 1, policy iteration may spend a lot of time in
the end deciding between two best actions, whereas for α < 1 the final convergence is faster.
7
Belousov and Peters
α = 20
1
Arm values Q(a)
1.0
0
−1
0.8
0.0
Action probabilities π(a)
α=2
1
0
−1
1
α=1
Iteration 1
0.1
0
−1
1
0.1
0.0
0.2
Iteration 3
α = −10
Iteration 25
0.5
0.1
0.0
0.0
0.0
0.2
0.5
0.0
0.0
0.2
0.5
0.6
0.0
0.0
0.0
0.2
0.2
0.4
0.1
0.0
0.0
0.0
0.0
0.50
0.0
0.2
0.25
0.25
0.5
0.0
0.00
0.00
0.0
1.0
1 2 3 4 5 6 7 8 9100.8
1 2 3 4 5 6 7 8 910
0.2
0.2
1 2 3 4 5 6 7 8 910
Iteration 4
0.1
0.2
0.1
0
−1
Iteration 2
0.0
0.0
0.01 2 3 4 5 6 7 8 910
0.2 1 2 3 4 5 6 7 8 910 0.4 1 2 3 4 5 6 7 8 910 0.6
Arm number a
Arm number a
Figure 3: Effects of α on policy improvement. Each row corresponds to a fixed α. First four
iterations of policy improvement together with iteration 25 are shown in each row. Large
positive α’s eliminate bad actions one by one, keeping the exploration level equal among
the rest. Small α’s weigh actions according to their values; actions with low value get zero
probability for α > 1, but remain possible with small probability for α ≤ 1. Large negative
α’s focus on the best action, exploring the remaining ones with equal probability.
Regret
Pn−1
rt ] is shown in Figure 2 for
Effects of α on regret. Average regret Cn = nQmax −E[ t=0
different values of α as a function of the time step n with 95% confidence error bars. We also
plot performance of the UCB algorithm (Bubeck and Cesa-Bianchi, 2012) for comparison.
The presented results are obtained in a 20-armed bandit environment where rewards have
Gaussian distribution r(a) ∼ N (Q(a), σ 2 ) with variance σ 2 = 0.5. Arm values are estimated
from observed rewards and the policy is updated every 20 time steps. The temperature
parameter η is decreased starting from η = 1 after every policy update according to the
schedule η + = βη with β = 0.8. Results are averaged over 400 runs. In general, extreme α’s
accumulate more regret. However, they eventually focus on a single action and flatten out.
Small α’s accumulate less regret, but they may keep exploring sub-optimal actions longer.
Values of α ∈ [0, 2] perform comparably with UCB after around 400 steps, once reliable
estimates of values have been obtained.
120
after 50 steps
after 100 steps
Figure 4 shows the average regret after a given number of
100
after 150 steps
after 200 steps
time steps as a function of the divergence type α. Best results
80
are achieved for small values of α. Large negative α’s correspond
to -greedy policies, which oftentimes prematurely converge to a
60
sub-optimal action, failing to discover the optimal action for a
40
long time if the exploration probability is small. Large positive
−20
−10
0
10
20
α’s correspond to -elimination policies, which may by mistake
Parameter α
completely eliminate the best action or spend a lot of time de- Figure 4: Regret after a
ciding between two options in the end of learning, accumulating fixed time as a function of α.
8
f -Divergence constrained policy improvement
more regret. The optimal value of α depends on the time horizon for which one wishes to
optimize the policy. The minimum of the curves shifts from slightly negative α’s towards
the range α ∈ [0, 2] with increasing time horizon.
4. Policy iteration with state-action f -divergence for ergodic MDPs
The methodology described in the previous section can be generalized from the bandit
scenario to reinforcement learning. We focus on the average-reward setting (Sutton and
Barto, 1998). At every time step, an agent finds itself in a state s ∈ S and takes an
action a ∈ A drawn from a stochastic policy π(a|s). Subsequently, the agent transitions
from state s to state s0 with probability p(s0 |s, a) and receives a reward r(s, a) ∈ R. The
fraction of time the agent spends on average in each state s as a result of following policy π
is given by the steady state distribution µ(s), which satisfies the stationarity condition
µ(s0 ) =
X
s,a
µ(s)π(a|s)p(s0 |s, a).
(7)
The steady state distribution is sometimes denoted by µπ (s) to emphasize its dependence
on the policy. The agent seeks a policy π that maximizes the expected return
X
J(π) ,
µ(s)π(a|s)r(s, a)
(8)
s,a
subject to the constraint (7), provided that both π and µ are probability distributions.
4.1 f -Divergence constrained policy optimization for ergodic MDPs
We can generalize the bandit problem (4) by adding a single additional constraint here—
namely, the stationarity condition (7). This constraint directly yields the value function (Peters et al., 2010), i.e., the policy evaluation step, as the corresponding dual variable. However, for a more straightforward Lagrangian, we rewrite (7) in terms of both policy and
state distribution. Assuming we have the state-action distribution q(s, a) = µπ0 (s)π0 (a|s)
under the old policy π0 (a|s) to which we aim to stay similar to with the new distribution
µ(s)π(a|s) while achieving a higher expected return, we obtain the following optimization
problem
X
X
µ(s)π(a|s)
max Jη (µ, π) =
µ(s)π(a|s)r(s, a) − η
q(s, a)f
µ,π
q(s, a)
s,a
s,a
X
X
s.t.
µ(s0 )π(a0 |s0 ) =
µ(s)π(a|s)p(s0 |s, a), ∀s0 ∈ S,
0
(9)
s,a
a
X
µ(s)π(a|s) = 1,
s,a
µ(s)π(a|s) ≥ 0,
∀s ∈ S, ∀a ∈ A.
A similar Lagrangian as for problem (4) can be obtained where the value function v arises
as an additional Lagrangian multiplier.
9
Belousov and Peters
The Lagrangian of problem (9) can be written as
L(µ, π, v, λ, κ) =
X
s,a
−λ
µ(s)π(a|s)Av (s, a) − η
X
s,a
µ(s)π(a|s) − 1
!
+
X
q(s, a)f
s,a
X
µ(s)π(a|s)
q(s, a)
κ(s, a)µ(s)π(a|s),
s,a
where A is a function of the dual variables v corresponding to the stationarity constraint (7)
Av (s, a) , r(s, a) +
X
s0
p(s0 |s, a)v(s0 ) − v(s).
(10)
We note that v can be identified with the value function and A with the advantage function.
The solution of Problem (9) can be expressed as a function of the dual variables
Av (s, a) − λ + κ(s, a)
?
?
0
µ (s)π (a|s) = q(s, a)f∗
,
(11)
η
which allows extracting the optimal next policy by Bayes’ rule
q(s, a)f∗0 ((Av (s, a) − λ + κ(s, a)) /η)
π ? (a|s) = P
.
0
b q(s, b)f∗ ((Av (s, b) − λ + κ(s, b)) /η)
(12)
The optimal values of the dual variables {v, λ, κ} are found by solving the dual problem
X
Av (s, a) − λ + κ(s, a)
+λ
min g(v, λ, κ) = η
q(s, a)f ∗
v,λ, κ
η
s,a
(13)
s.t. κ(s, a) ≥ 0, ∀s ∈ S, ∀a ∈ A,
arg f ∗ ∈ rangex ≥ 0 f 0 (x),
∀s ∈ S, ∀a ∈ A.
The constraint on arg f ∗ is linear because derivative of a convex function is monotone (Boyd
and Vandenberghe, 2004). Problem (13) altogether is a convex optimization problem, therefore can be solved efficiently.
A generic policy iteration algorithm employing an f -divergence penalty proceeds in two
steps. Policy evaluation: the dual problem (13) is solved for a given q. Policy improvement:
the analytic solution (12) of the primal problem (9) is used to update π. In the following
subsection we consider several particular choices of the divergence function in more detail.
4.2 Two important special cases: α = 1 and α = 2
Similar to the bandit scenario (see Section 3.2), there are two special choices of the f divergence for which we can analytically find the Lagrange multipliers λ and κ and thus
obtain the policy evaluation and policy improvement steps in the simplest form.
KL divergence (α = 1). With f = f1 , policy evaluation (13) turns into
X
Av (s, a)
?
min g(v) = λ (v) = η log
q(s, a) exp
v
η
s,a
10
(14)
f -Divergence constrained policy improvement
where g(v) , L(µ? , π ? , v, λ? , κ? ) is the Lagrangian upon substitution of all variables except
for v. Remarkably, the baseline λ? equals the Lagrangian dual function g, which is only
true in the KL divergence case. The policy improvement step (12) takes the form
q(s, a) exp (Av? (s, a)/η)
π ? (a|s) = P
.
b q(s, b) exp (Av ? (s, b)/η)
(15)
Note that the baseline λ? cancels out in the policy but would be included in the state-action
distribution.
Pearson χ2 divergence (α = 2). When f = f2 , one needs to be careful with the dual
variables κ ensuring the non-negativity of probabilities because they are not necessarily all
zero (see Section 3.1). For notational simplicity, we employ differential advantage Ãv (s, a)
defined as a difference between the advantage (10) and the expected return (8)
Ãv (s, a) = Av (s, a) − J(π0 ),
following Sutton and Barto (1998). It is straightforward to check that if the temperature
η is bigger than the absolute value of the minimum differential advantage, η > ηmin =
| mins,a Ãv (s, a)|, then the non-negativity constraint is satisfied and κ ≡ 0. The baseline λ?
turns into the average return
X
λ? =
q(s, a)Av (s, a) = J(π0 ),
s,a
becoming independent of v, in contrast to the KL divergence case. The policy evaluation
step (13) corresponds to minimization of the squared differential advantage (scaled by 2η)
min g(v) =
v
1 X
q(s, a)Ã2v (s, a) + J(π0 ),
2η s,a
(16)
and the policy improvement step (12) corresponds to the linear in the differential advantage
reweighting of the old policy
q(s, a) 1 + Ãv? (s, a)/η
.
π ? (s, a) = P
(17)
? (s, b)/η
q(s,
b)
1
+
Ã
v
b
A careful reader shall notice that the solution for the Pearson χ2 -divergence (16, 17) can
be obtained as the first-order approximation of the KL divergence solution (14, 15) in
the high-temperature limit η → ∞. It turns out, an even more general statement holds.
Solution (12, 13) with any f -divergence turns into the linear-quadratic Pearson χ2 solution
in the limit of high temperatures (small policy update steps). The underlying reason for this
is that the Pearson χ2 -divergence is a quadratic approximation of any f -divergence around
unity. Since pushing η towards infinity puts more weight on the divergence penalty in the
optimization objective (9), the distance between the old state-action distribution and the
new one becomes smaller, justifying the Taylor expansion. Therefore, one should expect
to see pronounced differences between various f -divergence penalties either for big update
steps or when approximations of the value function and the policy are used.
11
Belousov and Peters
4.3 Practical algorithm: sample-based policy iteration with f -divergence
Solving problem (13) and computing the optimal next policy (12) would require full knowledge of the environment {p(s0 |s, a), r(s, a)}. In practice, a reinforcement learning agent
would need to estimate p and r from experience, which is prohibitively expensive in large
state-action spaces. Fortunately, we can rephrase the algorithm entirely in terms of sample
averages, thus side-stepping model estimation. We replace expectations by sample averages
using the transitions (si , ai , s0i , ri ) gathered under the current policy. The agent does not
need to visit all states—local information under the current policy suffices to compute the
averages. Algorithm 1 summarizes the main steps of the f -divergence constrained policy
iteration.
Function approximation. If the dimensionality of the state space is large, one can resort
to function approximation methods. We note that relaxing the steady state constraint (7)
to a “steady features constraint”
X
X
0
0
µ(s
)φ(s
)
=
µ(s)π(a|s)p(s0 |s, a)φ(s0 ).
(18)
0
0
s
s,a,s
naturally introduces a linear function approximation vθ (s) = θT φ(s) with fixed features
φ : S → Rm while the parameters θ are the new Lagrangian multipliers (Peters et al., 2010).
To our knowledge, such a constraint is the only way to introduce function approximation in
the same way into the primal optimization problem (9), the next policy (12), and the dual
optimization problem (13).
Implementation details. Several practical improvements make Algorithm 1 faster and
robuster. First, the temperature parameter η should be decayed with iterations in order to
guarantee convergence to a solution of the original non-penalized problem. In our experiments, simple exponential decay was sufficient; however, more complex adaptive schemes
can be devised. Second, for α-divergences with α ≤ 1, one can omit the dual variables
κ(s, a) because they are all equal to zero (see Section 3.1). Moreover, one can omit κ(s, a)
even for α > 1 if η is sufficiently big. The exact value of ηmin above which κ can be ignored
depends on the scale of the rewards; in the special case α = 2, there is an explicit condition
on ηmin (see Section 4.2). Third, the estimate of the advantage Av? (s, a) for actions that
were not sampled under the current policy can be set to the baseline λ? , which serves as
a proxy for the expected reward; in our experiments this strategy performed best. Fourth,
introducing a little slack in the linear constraint on the domain of the convex conjugate
function improves robustness and helps the optimizer.
4.4 Empirical results on ergodic MDPs
We evaluate our policy iteration algorithm with f -divergence on standard grid-world reinforcement learning problems from OpenAI Gym (Brockman et al., 2016). The environments
that terminate or have absorbing states are restarted during data collection in order to ensure ergodicity. Figure 5 demonstrates learning dynamics on different environments for
various choices of the divergence function. Parameter settings and other implementation
details can be found in Appendix A. In summary, one can either promote risk averse behavior by choosing α < 0, which may, however, result in suboptimal exploration, or one can
12
f -Divergence constrained policy improvement
Algorithm 1: Policy iteration with any type of f -divergence
Input: Initial policy π0 (a|s), divergence function f , temperature η
define Dual Function ĝ(v, λ, κ):
for every sample i ← 1 to N do
Â(si , ai ) ← Â(si , ai ) + (ri + v(s0i ) − v(si ))
n(si , ai ) ← n(si , ai ) + 1
for every state-action pair (s, a) ∈ D do
Â(s, a) ← Â(s, a)/n(s, a)
ŷ(s, a) ← (Â(s, a) − λ + κ(s, a))/η
P
∗
return Nη N
i=1 f (ŷ(si , ai )) + λ
foreach policy update k do
Sampling: Gather D = {(si , ai , s0i , ri ), i = 1, N } under current policy πk (a|s)
Policy Evaluation: Minimize dual function ĝ defined above
{v ? , λ? , κ? } ← arg minv,λ,κ ĝ(v, λ, κ) s.t. κ(s, a) ≥ 0, ŷ(s, a) ∈ rangex ≥ 0 f 0 (x)
Policy Improvement: for every state-action pair (s, a) ∈ D do
0
? ? ?
∗ (ŷ(s,a; v ,λ ,κ ))
πk+1 (a|s) ← Pπkπ(a|s)f
0 (ŷ(s,b; v ? ,λ? ,κ? ))
(b|s)f
∗
b k
Output: Optimal policy π ? (a|s)
promote risk seeking behavior with α > 1, which may lead to overly aggressive elimination
of options. Our experiments suggest that the optimal balance should be found in the range
α ∈ [0, 1]. It should be pointed out that the effect of the α-divergence on policy iteration
is not linear and not symmetric with respect to α = 0.5, contrary to what one could have
expected given the symmetry of the α-divergence as a function of α pointed out in Section 2. That is, switching from α = −3 to α = −2 may have little effect on policy iteration,
whereas switching from α = 3 to α = 4 may have a much more pronounced influence on
the learning dynamics.
4.5 Relation to the mean squared Bellman error minimization
We have seen in Section 4.2 that policy evaluation with the Pearson χ2 -divergence penalty,
corresponding to the dual optimization problem (16), is equivalent to minimization of the
mean squared differential advantage (MSDA). Here, we establish the relation between the
MSDA objective and the classical mean squared Bellman error (MSBE) minimization (Sutton and Barto, 1998).
In order to state precisely how the MSDA and MSBE objectives are related, we need
to recall several definitions. Assume that the value function v : S → R is parameterized by
a vector θ as discussed in Section 4.3, and let an agent follow a policy π0 which yields a
state-action distribution q(s, a) = µπ0 (s)π0 (a|s). In discounted infinite horizon setting, one
defines the temporal difference (TD) error as δθ (s, a, s0 ) = r(s, a) + γvθ (s0 ) − vθ (s) where
γ ∈ (0, 1) is a discount factor. In the average-reward setting, this TD error is typically
replaced by the differential TD error
δ̃θ (s, a, s0 ) = r(s, a) − r̄ + vθ (s0 ) − vθ (s)
13
Belousov and Peters
Chain
4
2
0
5
-1.0
0.0
0.5
1.0
2.0
-4.0
-2.0
0.0
1.0
3.0
5.0
-10.0
0.0
1.0
10.0
10
15
20
25
30 0
5
10
15
20
25
30 0
5
10
15
20
25
30
Expected reward
CliffWalking
5
-1.0
0.0
0.5
1.0
2.0
-4.0
-2.0
0.0
1.0
3.0
5.0
-10.0
0.0
1.0
10.0
0
0
10
20
30
40 0
10
20
30
40 0
10
20
30
40
FrozenLake 8x8
0.06
-10.0
0.0
1.0
10.0
0.04
-4.0
-2.0
0.0
1.0
3.0
5.0
0.02
-1.0
0.0
0.5
1.0
2.0
0.00
0
10
20
30
40
50 0
10
20
30
40
50 0
10
20
30
40
50
Iteration
Figure 5: Effects of α-divergence on policy iteration. Each row corresponds to a given
environment. Results for different values of α are split into three subplots within each row,
from the more extreme α’s on the left to the more refined values on the right. In all cases,
more negative values α < 0 initially show faster improvement because they immediately
jump to the mode and keep the exploration level low; however, after a certain umber of
iterations they get overtaken by moderate values α ∈ [0, 1] that weigh advantage estimates
more evenly. Positive α > 1 demonstrate high variance in the learning dynamics because
they clamp the probability of good actions to zero if the advantage estimates are overly
pessimistic, never being able to recover from such a mistake. Large positive α’s may even
fail to reach the optimum altogether, as exemplified by α = 10 in the plots. The most stable
and reliable α-divergences lie between the reverse KL (α = 0) and the KL (α = 1), with the
Hellinger distance (α = 0.5) outperforming both of them on the FrozenLake environment.
that differs from the “usual” TD error in two ways: first, there is no discount factor, and
second, the expected reward under the current policy
r̄ =
X
q(s, a)r(s, a) =
s,a
X
µπ0 (s)π0 (a|s)r(s, a) = J(π0 )
s,a
is subtracted in order to ensure that the mean of the differential TD error is zero
X
s,a
µπ0 (s)π0 (a|s)p(s0 |s, a)δ̃θ (s, a, s0 ) = 0.
14
f -Divergence constrained policy improvement
In the following, we only consider differential versions of the errors. By averaging over the
next states s0 in the differential TD error, one obtains the differential advantage
X
Ãθ (s, a) =
p(s0 |s, a)δ̃θ (s, a, s0 ).
s0
By integrating out the action a in the differential advantage, one obtains the differential
Bellman error
X
ε̃θ (s) =
π0 (a|s)Ãθ (s, a).
a
In total, there are three different error signals: differential TD error (DTDE), differential
advantage (DA), and the differential Bellman error (DBE). Each of them can be used for
learning. Creating a mean squared (MS) objective out of each of them, we obtain three
objective functions:
h
i2
X
MSDTDE(θ) =
µπ0 (s)π0 (a|s)p(s0 |s, a) δ̃θ (s, a, s0 ) ,
s,a,s0
MSDA(θ) =
X
s,a
MSDBE(θ) =
X
s
π0
µ (s)π0 (a|s)
"
X
s0
0
0
p(s |s, a)δ̃θ (s, a, s )
#2
,
2
X
µπ0 (s)
π0 (a|s)p(s0 |s, a)δ̃θ (s, a, s0 ) .
a,s0
Thus, the novel MSDA objective fits in-between the TD error and the Bellman error minimization. Performing stochastic gradient descent (SGD) on these objectives leads to various
flavors of the residual gradient (RG) algorithm (Baird, 1995), recently surveyed by Dann
et al. (2014). Curiously, all three objectives result in the naive residual gradient algorithm (Sutton and Barto, 1998) when expectations are “naively” replaced by one-sample
estimates.
We realize that the dual for α = 2, i.e., (16) is exactly the MSDA objective. Interestingly, since the dual objective (13) tends towards the MSDA objective in the limit of high
temperatures η → ∞ for any choice of the divergence function f , dual minimization can be
viewed as a continuous generalization of the MSDA minimization to finite temperatures;
therefore, SGD on the dual objective (13) is a continuous generalization of the RG algorithm
to finite temperatures.
5. Conclusion
In this paper, we have developed a framework for deriving new and established policy
improvement and policy iteration algorithms by constraining policy change using the f divergence. First, we described the key idea in the bandit scenario. Looking at the smooth
curve of α-divergences in the space of f -divergences, we were able to see the connection
between many classical action selection strategies as well as discover several new ones.
The optimal choice of α depends on the time horizon, since different α’s corresponds to
different exploration strategies. The soft-max policy update rule, employed by the Exp3
15
Belousov and Peters
algorithm (Bubeck and Cesa-Bianchi, 2012) and the gradient bandit algorithm (Sutton
and Barto, 1998), is a special case of the f -divergence constrained policy improvement
corresponding to the KL divergence.
After the bandit case, we considered the average-reward reinforcement learning problem,
for which we found both (i) an efficient sample-based solution and (ii) the important new
insight that the choice of the f -divergence does not only imply the type of policy improvement step but also the compatible policy evaluation step. In particular, we have shown that
employing the soft-max policy for the actor implies that the log-sum-exp function is used for
policy evaluation by the critic (somewhat similar to REPS (Peters et al., 2010)). Similarly,
in our framework, policy evaluation by minimizing the mean squared differential advantage,
which is closely related to the mean squared Bellman error, results from the choice of the
α = 2-divergence and comes along with a new, compatible policy update rule. Thus, it is
likely that this approach will still yield a larger number of new policy improvements along
with their compatible critics derived by just choosing another f -divergence.
Acknowledgments
This project has received funding from the European Union’s Horizon 2020 research and
innovation programme under grant agreement No 640554.
16
f -Divergence constrained policy improvement
Appendix A.
In all experiments, the temperature parameter η is exponentially decayed ηi+1 = η0 ai in
each iteration i = 0, 1, . . . . The choice of η0 and a depends on the scale of the rewards
and the number of samples collected per policy update. Tables for each environment list
these parameters along with the number of samples per policy update, the number of
policy iteration steps, and the number of runs for averaging the results. Where applicable,
environment-specific settings are also listed.
Parameter
Value
Number of states
Action success probability
Small and large rewards
Number of runs
Number of iterations
Number of samples
Temperature parameters (η0 , a)
8
0.9
(2.0, 10.0)
10
30
800
(15.0, 0.9)
Table 4: Chain environment.
Parameter
Value
Punishment for falling from the cliff
Reward for reaching the goal
Number of runs
Number of iterations
Number of samples
Temperature parameters (η0 , a)
-10.0
100
10
40
1500
(50.0, 0.9)
Table 5: CliffWalking environment.
Parameter
Value
Action success probability
Number of runs
Number of iterations
Number of samples
Temperature parameters (η0 , a)
0.8
10
50
2000
(1.0, 0.8)
Table 6: FrozenLake environment.
17
Belousov and Peters
References
S. M. Ali and Samuel Daniel Silvey. A General Class of Coefficients of Divergence of One
Distribution from Another. Journal of the Royal Statistical Society. Series B (Methodological), 28:131–142, 1966. ISSN 0035-9246.
Yasemin Altun and Alex Smola. Unifying Divergence Minimization and Statistical Inference
via Convex Duality. Learning Theory, 4005:1–15, 2006. ISSN 03029743. doi: 10.1007/
11776420 13.
Shun-Ichi Amari. Differential-Geometrical Methods in Statistics. Springer New York, 1985.
doi: 10.1007/978-1-4612-5056-2.
Peter Auer, Nicolo Cesa-Bianchi, Yoav Freund, and Robert Schapire. The Non-Stochastic
Multi-Armed Bandit Problem. SIAM Journal on Computing, 32(1):48–77, 2003.
Mohammad Gheshlaghi Azar and Bert Kappen. Dynamic Policy Programming. Journal of
Machine Learning Research, 13:3207–3245, 2012. ISSN 15324435.
J. Andrew Bagnell and Jeff Schneider. Covariant policy search. In IJCAI, pages 1019–1024,
2003.
Leemon Baird. Residual Algorithms: Reinforcement Learning with Function Approximation. Proceedings of the 12th International Conference on Machine Learning, (July):
30–37, 1995. ISSN 00043702. doi: 10.1.1.48.3256.
Stephen Boyd and Lieven Vandenberghe. Convex Optimization, volume 25. 2004. ISBN
9780511804441. doi: 10.1017/CBO9780511804441.
John S. Bridle. Training Stochastic Model Recognition Algorithms as Networks can Lead
to Maximum Mutual Information Estimation of Parameters. In NIPS, 1990. ISBN 155860-100-7.
Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie
Tang, and Wojciech Zaremba. OpenAI Gym. arXiv preprint arXiv:1606.01540, 2016.
Sébastien Bubeck and Nicolò Cesa-Bianchi. Regret Analysis of Stochastic and Nonstochastic
Multi-armed Bandit Problems. Foundations and Trends in Machine Learning, 5(1):1–122,
2012. ISSN 9781601986269. doi: 10.1561/2200000024.
Herman Chernoff. A measure of asymptotic efficiency for tests of a hypothesis based on the
sum of observations. The Annals of Mathematical Statistics, pages 493–507, 1952.
Andrzej Cichocki and S. Amari. Families of alpha- beta- and gamma- divergences: Flexible
and robust measures of Similarities. Entropy, 12(6):1532–1568, 2010. ISSN 10994300.
doi: 10.3390/e12061532.
Imre Csiszár. Eine informationstheoretische Ungleichung und ihre Anwendung auf den
Beweis der Ergodizität von Markoffschen Ketten. Publ. Math. Inst. Hungar. Acad. Sci.,
8:85–108, 1963.
18
f -Divergence constrained policy improvement
Christoph Dann, Gerhard Neumann, and Jan Peters. Policy Evaluation with Temporal
Differences: A Survey and Comparison. In Journal of Machine Learning Research, volume 15, pages 809–883, 2014. ISBN 1532-4435.
Mohammad Ghavamzadeh, Shie Mannor, Joelle Pineau, and Aviv Tamar. Bayesian Reinforcement Learning: A Survey. Foundations and Trends in Machine Learning, 8(5-6):
359–483, 2015. ISSN 1935-8237. doi: 10.1561/2200000049.
Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil
Ozair, Aaron Courville, and Yoshua Bengio. Generative Adversarial Nets. In NIPS, 2014.
ISBN 1406.2661.
Sham Machandranath Kakade. A Natural Policy Gradient. In NIPS, pages 1531–1538,
2001. ISBN 9780874216561. doi: 10.1.1.19.8165.
Hilbert J Kappen. Path integrals and symmetry breaking for optimal control theory. Journal
of Statistical Mechanics: Theory and Experiment, 2005. ISSN 1742-5468. doi: 10.1088/
1742-5468/2005/11/P11011.
Tom Minka. Divergence measures and message passing. Technical report, Microsoft Research, 2005.
Tetsuzo Morimoto. Markov processes and the H-theorem. Journal of the Physical Society
of Japan, 18(3):328–331, 1963.
Sebastian Nowozin, Botond Cseke, and Ryota Tomioka. f-GAN: Training Generative Neural
Samplers using Variational Divergence Minimization. In NIPS, pages 271–279, 2016.
Jan Peters and Stefan Schaal. Policy gradient methods for robotics. In Intelligent Robots and
Systems, 2006 IEEE/RSJ International Conference on, pages 2219–2225. IEEE, 2006.
Jan Peters, Sethu Vijayakumar, and Stefan Schaal. Reinforcement Learning for Humanoid
Robotics. In Humanoids, 2003. doi: 10.1163/1568553041719519.
Jan Peters, Katharina Mülling, and Yasemin Altun. Relative Entropy Policy Search. In
AAAI, pages 1607–1612, 2010.
Igal Sason and Sergio Verdu. F-divergence inequalities. IEEE Transactions on Information
Theory, 62(11):5973–6006, 2016. ISSN 00189448. doi: 10.1109/TIT.2016.2603151.
John Schulman, Sergey Levine, Michael Jordan, and Pieter Abbeel. Trust Region Policy
Optimization. In ICML, 2015. ISBN 0375-9687. doi: 10.1063/1.4927398.
Susanne Still and Doina Precup. An information-theoretic approach to curiosity-driven
reinforcement learning. Theory in Biosciences, 131(3):139–148, 2012. ISSN 14317613.
doi: 10.1007/s12064-011-0142-z.
Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction. MIT
press Cambridge, 1998.
19
Belousov and Peters
Richard S Sutton, David A McAllester, Satinder P Singh, and Yishay Mansour. Policy
gradient methods for reinforcement learning with function approximation. In Advances
in neural information processing systems, pages 1057–1063, 2000.
Marc Teboulle. Entropic Proximal Mappings with Applications to Nonlinear Programming.
Mathematics of Operations Research, 17(3):670–690, 1992. ISSN 0364-765X. doi: 10.
1287/moor.17.3.670.
Emanuel Todorov. Linearly-solvable Markov decision problems. In NIPS, 2006. ISBN
9780262195683.
Martin J Wainwright and Michael I Jordan. Graphical Models, Exponential Families, and
Variational Inference. Foundations and Trends in Machine Learning, 1(1–2):1–305, 2007.
ISSN 19358237. doi: 10.1561/2200000001.
Huaiyu Zhu and Richard Rohwer. Information geometric measurements of generalisation.
Technical report, Aston University, 1995.
Brian D. Ziebart, Andrew Maas, J. Andrew Bagnell, and Anind K. Dey. Maximum entropy
inverse reinforcement learning. In AAAI, pages 1433–1438, 2008.
20
| 2cs.AI
|
EULERIAN GRADED D-MODULES
arXiv:1210.8402v2 [math.AC] 15 Jul 2016
LINQUAN MA AND WENLIANG ZHANG
Abstract. Let R = K[x1 , . . . , xn ] with K a field of arbitrary characteristic and D be the ring of differential
operators over R. Inspired by Euler formula for homogeneous polynomials, we introduce a class of graded
D-modules, called Eulerian graded D-modules. It is proved that a vast class of D-modules, including all
local cohomology modules HJi11 · · · HJiss (R) where J1 , . . . , Js are homogeneous ideals of R, are Eulerian. As
an application of our theory of Eulerian graded D-modules, we prove that all socle elements of each local
i0 i1
cohomology module Hm
HJ1 · · · HJiss (R) must be in degree −n in all characteristic. This answers a question
raised in [Zha12]. It is also proved that graded F -modules are Eulerian and hence the main result in [Zha12]
is recovered. An application of our theory of Eulerian graded D-modules to the graded injective hull of
R/P , where P is a homogeneous prime ideal of R, is discussed as well.
1. Introduction
Let R = K[x1 , . . . , xn ] be a polynomial ring in n indeterminates over a field K with the standard grading,
[j]
i.e. deg(xi ) = 1 for each xi and deg(c) = 0 for each nonzero c ∈ K. Let ∂i denote the j-th order
j
1 ∂
differential operator · j with respect to xi for each 0 ≤ i ≤ n and j ≥ 1. By [Gro67, Théorème 16.11.2],
j! ∂xi
[j]
D = Rh∂i |1 ≤ i ≤ n, 1 ≤ ji is the ring of K-linear differential operators of R (note if K has characteristic
0, D is the same as the Weyl algebra Rh∂1 , . . . , ∂n i). The ring of K-linear differential operators D has a
[j]
[j]
natural Z-grading given by deg(xi ) = 1, deg(∂i ) = −j, and deg(c) = 0 for each xi , ∂i and each nonzero
c ∈ K.
The classical Euler formula for homogeneous polynomials says that
n
X
xi ∂i f = deg(f )f
i=1
for each homogeneous polynomial f ∈ R. Inspired by Euler formula, we introduce a class of D-modules
called Eulerian graded D-modules: the graded D-modules whose homogeneous elements satisfy a series of
“higher order Euler formulas” (cf. Definition 2.1). One of our main results concerning Eulerian graded
D-modules is the following (proved in Section 2 and Section 5):
Theorem 1.1. Let R = K[x1 , . . . , xn ], m = (x1 , . . . , xn ), and J1 , . . . , Js be homogeneous ideals of R. Then
(1) R(ℓ) is Eulerian if and only if ℓ = 0.
(2) Let ∗ E be the graded injective hull of R/m. Then ∗ E(ℓ) is Eulerian if and only if ℓ = n.
(3) Each local cohomology module HJi11 (· · · (HJiss (R))) is Eulerian for all i1 , . . . , is .
As an application of our theory of Eulerian graded D-modules, we have the following result on local
cohomology (proved in Section 5):
i0
Theorem 1.2. Let notations be as in the previous theorem. Then all socle elements of each Hm
(HJi11 (· · · (HJiss (R)))
i1
is
i0
must have degree −n, and consequently each Hm (HJ1 (· · · (HJs (R))) is isomorphic (as a graded D-module)
to a direct sum of copies of ∗ E(n).
This result is characteristic-free, in particular it gives a positive answer to a question stated in [Zha12]
and recovers the main theorem in [Zha12].
The paper is organized as follows. In Section 2, Eulerian graded D-modules are defined over an arbitrary
field and some basic properties of these modules are discussed. In Section 3 and Section 4, we consider
The second author was partially supported by the NSF grant DMS #1068946.
1
Eulerian graded D-modules in characteristic 0 and characteristic p, respectively; in particular, we show in
Section 4 that each graded F -module (introduced in [Zha12]) is Eulerian. In Section 5, we apply our theory
of Eulerian graded D-modules to local cohomology modules; Theorem 1.2 is proved in this section. Finally,
in Section 6, an application of our theory to graded injecitve hull of R/P , where P is a homogeneous prime
ideal, is considered.
We finish our introduction by fixing our notation throughout the paper as follows. R = K[x1 , . . . , xn ]
1 ∂j
denotes the polynomial ring in n indeterminates over a field K. The j-th order differential operator · j
j! ∂xi
[j]
[j]
with respect to xi is denoted by ∂i and D = Rh∂i |1 ≤ i ≤ n, 1 ≤ ji denotes the ring of differential
operators over R. It follows from [Lyu11, Corollary 2.2] that every element of D may be uniquely written
as a linear combinations of monomials in x’s and ∂’s. The natural Z-grading on R and D is given by
[j]
deg(xi ) = 1, deg(∂i ) = −j, deg(c) = 0
[j]
for each xi , ∂i and nonzero c ∈ K (it is evident that R is a graded D-module).
A graded D-module is a left D-module with a Z-grading that is compatible with the natural Z-grading on
D. Given any graded D-module M , the module M (ℓ) denotes M with degree shifted by ℓ, i.e. M (ℓ)i = Mℓ+i
for each i.
The irrelevant maximal ideal (x1 , . . . , xn ) of R is denoted by m. The graded injective hull of R/m is
1
denoted by ∗ E. It may be identified with the K-vector space with a basis { xe1 ···x
en |e1 , . . . , en ≥ 1} and it
n
1
has a natural D-module structure given by
1
−ej
∂j · e1
= e1
.
x1 · · · xenn
x1 · · · xej +1 · · · xenn
1
has degree 0 (cf. [BS98, Example 13.3.9]).
x1 · · · xn
a · (a − 1) · · · (a − b + 1)
a
(note
For each integer a and a nonnegative integer b, we will use
to denote
b!
b
that this number is still well-defined when char(K) = p > 0).
∗
E is graded to the effect that the element
Acknowledgement
The authors would like to thank Professor Gennady Lyubeznik for carefully reading a preliminary version
of the paper and for his valuable suggestions which improved the paper considerably. The authors are also
grateful to the anonymous referee for her/his comments on the paper.
2. Eulerian graded D-modules
In this section, we introduce Eulerian graded D-modules and discuss some of their basic properties. We
begin with the following definition.
Definition 2.1. The r-th Euler operator, denoted by Er , is defined as
X
[i ]
Er :=
xi11 · · · xinn ∂1 1 · · · ∂n[in ] .
i1 +i2 +···+in =r,i1 ≥0,...,in ≥0
Pn
In particular E1 is the usual Euler operator i=1 xi ∂i .
A graded D-module M is called Eulerian, if each homogeneous element z ∈ M satisfies
deg(z)
(2.1.1)
Er · z =
·z
r
for every r ≥ 1.
We start with an easy lemma.
[s] [t]
Lemma 2.2. For all positive integers s and t, we have ∂i ∂i =
2
s + t [s+t]
.
∂i
s
Proof. It is easy to check that (in all characteristic)
∂s ∂t
s + t ∂is+t
s + t [s+t]
[s] [t]
∂i ∂i = i i =
=
.
∂i
s! t!
(s + t)!
s
s
The following lemma is a special case of [Lyu11, Proposition 2.1], which will be needed in the sequel.
Lemma 2.3. For all positive integers s and t, we have
min{s,t}
X
t t−j [s−j]
[s]
xi ∂i
∂i xti =
j
j=0
for each i.
The following proposition indicates a connection among Euler operators.
Proposition 2.4. For every r ≥ 1, we have E1 · Er = (r + 1)Er+1 + rEr .
[j+1]
[j]
Proof. By Lemma 2.2 we know ∂i ∂i = (j + 1)∂i
X
. Now we have
[i ] [i ]
E1 · Er =
X
=
X
xi11 · · · (xj ∂j xjj ) · · · xinn ∂1 1 ∂2 2 · · · ∂n[in ]
j
X
i1 +i2 +···+in =r
=
X
X
xi1i · · · (xjj
xj ∂j ·
i1 +i2 +···+in =r
j
j
i1 +i2 +···+in =r
+
i +1
X
[i ] [i ]
i
∂j + ij xjj ) · · · xinn ∂1 1 ∂2 2 · · · ∂n[in ] by lemma 2.3
X
i1 +i2 +···+in =r
i +1
xi11 · · · xjj
[i ]
[i ]
· · · xinn ∂1 1 · · · (∂j ∂j j ) · · · ∂n[in ]
j
X
i1 +i2 +···+in =r
=
[i ] [i ]
i
i1 +i2 +···+in =r
X
=
xi11 xi22 · · · xinn ∂1 1 ∂2 2 · · · ∂n[in ]
X
[i ]
ij xi11 · · · xinn ∂1 1 · · · ∂n[in ]
j
X
i +1
(ij + 1)xi11 · · · xjj
[i ]
[i +1]
· · · xinn ∂1 1 · · · ∂j j
· · · ∂n[in ] + rEr
j
= (r + 1)Er+1 + rEr
Some remarks are in order.
Remark 2.5.
(1) If M is an Eulerian graded D-module, then M (ℓ) is Eulerian if and only if ℓ = 0
(equivalently, for any graded D-module
M, M (ℓ) is Eulerian graded for at most one ℓ). Because
a
b
in any characteristic, we have
=
for all r ∈ N if and only if a = b. This is trivial in
r
r
characteristic 0 (take r = 1). We give a short argument in characteristic p > 0. If both a and b are
positive, then by a theorem of Lucas [Luc78], we have
Y
ai
a
=
ri
r
P
P i
i
where
a= ai p and r = ri p are the p-adic decompositions of a and r. In particular, if we apply
a
b
=
to r = pi , we get ai = bi for every i, hence a = b. When a and b are both negative, we
r
r
P
P
look at the p-adic decomposition of −a − 1 = ai pi and −b − 1 = bi pi . Taking r = pi , we get
i
i
i
−a + pi − 1
a
b
−b + pi − 1
=
(−1)p · (ai + 1) = (−1)p
=
=
= (−1)p · (bi + 1).
i
i
i
i
p
p
p
p
3
Hence again we have ai = bi for every i, −a − 1 = −b − 1 so a = b. Therefore it suffices to
j
show a and b must have the same sign. But if a > 0 and b < 0, we can
− b,
pick
r = p ≫
a
a
b
then direct computation (or using the theorem of Lucas [Luc78]) gives
= 0 while
=
pj
pj
j
j
−b + pj − 1
= (−1)p , which is a contradiction.
(−1)p
pj
(2) Definition 2.1 does not depend on the characteristic of K. However we will see in section 3 that, in
characteristic 0, we only need to consider E1 , the usual Euler operator.
(3) R is Eulerian and a proof goes as follows. For each monomial xj11 xj22 · · · xjnn , where jn ’s are arbitrary
integers (we allow negative integers), we have (for each r ≥ 1)
X
[i ] [i ]
Er · xj11 xj22 · · · xjnn = (
xi11 xi22 · · · xinn ∂1 1 ∂2 2 · · · ∂n[in ] ) · (xj11 xj22 · · · xjnn )
i1 +i2 +···+in =r
jn
j1
· xj11 xj22 · · · xjnn
···
i
i
n
1
i1 +i2 +···+in =r
j1 + · · · + jn
· xj11 xj22 · · · xjnn
=
r
=
X
We explain the last equality: we use an induction argument. It is clear when r = 1, now suppose we
have the equality for r − 1. Then
X
jn
j1
···
in
i
1
i1 +i2 +···+in =r
r
X
X
jn−1
j1
jn
···
=
in−1
in i +i +···+i =r−i i1
in =0
1
2
n−1
n
r
X
j1 + · · · + jn−1
jn
=
r − in
i
n
in =0
j1 + · · · + jn
=
r
where the second equality is by induction hypothesis for r − 1, and the last equality is by the
Chu-Vandermonde identity (see [Ask75])
X
r
a
j
a−j
=
.
r
k
r−k
k=0
jn in R
and the Er ’s clearly preserve addition, we can
Since xj11 xj22 · · · xjnn clearly has degree j1 + · · ·+
deg(z)
see from the computation above that Er · z =
· z for every homogenous z ∈ R. Therefore,
r
R is Eulerian.
(4) DPis not Eulerian.PIt is clear that the identity 1 ∈ D is homogeneous with degree 0, but E1 · 1 =
( ni=1 xi ∂i ) · 1 = ni=1 xi ∂i 6= 0.
One of the main results in this section is that, to check whether a graded D-module is Eulerian, it suffices
to check whether each element of any set of homogeneous generators satisfies (2.1.1).
Theorem 2.6. Let M be a graded D-module. Assume that {g1 , g2 , . . . } is a set of homogeneous D-generators
of M . Then, M is Eulerian if and only if each gj satisfies Euler formula (2.1.1) for every r.
Proof. If M is Eulerian, then it is clear that each gj satisfies Euler formula (2.1.1) for every r. Assume that
each gj satisfies Euler formula (2.1.1) for every r and we wish to prove that M is Eulerian. To this end, it
suffices to show that, if a homogeneous element z ∈ M satisfies Euler formula (2.1.1) for every r, then so
[j]
[j ]
[j ]
does xs11 · · · xsnn ∂1 1 · · · ∂n n · z. And it is clear that it suffices to consider xsi ∂i · z. Without loss of generality,
4
[j]
we may assume i = 1. We will prove this in two steps; first we consider ∂1 · z and then xs1 · z (once we finish
our first step, we may replace ∂ [j] · z by z and then our second step will finish the proof).
[j]
First we will use induction on r to show that ∂1 · z satisfies Euler formula (2.1.1) for each r. When r = 1,
we compute
[j]
E1 · (∂1 z) =
n
X
[j]
xi ∂i · (∂1 z)
i=1
[j]
[j]
= x1 ∂1 ∂1 z + ∂1
X
xi ∂i · z
i≥2
[j]
[j−1]
= ∂1 x1 ∂1 z − ∂1
[j]
∂1 z + ∂1
X
xi ∂i · z
i≥2
[j]
= ∂1
n
X
[j]
xi ∂i · z − j∂1 z
i=1
[j]
= (deg(z) − j) · ∂1 z
[j]
Now for general r, suppose we know that Er−k · (∂1 z) =
deg(z) − j
[j]
· ∂1 z for every 1 ≤ k ≤ r − 1. Then
r−k
we have
(
X
[j]
[i ] [i ]
xi11 xi22 · · · xinn ∂1 1 ∂2 2 · · · ∂n[in ] ) · (∂1 z)
i1 +i2 +···+in =r
=
[j]
[i ] [i ]
X
(xi11 ∂1 )xi22 · · · xinn ∂1 1 ∂2 2 · · · ∂n[in ] · z
X
(∂1 xi11 )xi22 · · · xinn ∂1 1 ∂2 2 · · · ∂n[in ] · z
i1 +i2 +···+in =r
(1)
=
[j]
[i ] [i ]
i1 +i2 +···+in =r
−
X
i1 +i2 +···+in =r
min{i1 ,j}
X
k=1
i1 i1 −k [j−k] i2
[i ] [i ]
x
∂1
x2 · · · xinn ∂1 1 ∂2 2 · · · ∂n[in ] · z
k 1
min{i1 ,j}
deg(z)
[j]
=
· (∂1 z) −
r
i
X
X
1 +i2 +···+in =r
(2)
k=1
X
deg(z)
[j]
· (∂1 z) −
r
i
=
(4)
j
X
deg(z)
j
[j]
=
· (∂1 z) −
r
k
1 +i2 +···+in =r
k=1
X
[i −k] [i2 ]
∂2
x1i1 −k xi22 · · · xinn ∂1 1
[k] [j−k]
· · · ∂n[in ] · (∂1 ∂1
z)
k=1
min{i1 ,j}
X
(3)
(6)
i1 [i1 ] [i2 ]
[j−k]
z)
∂ )∂2 · · · ∂n[in ] · (∂1
k 1
· · · xinn (
min{i1 ,j}
deg(z)
[j]
=
· (∂1 z) −
r
i
1 +i2 +···+in =r
(5)
x1i1 −k xi22
X
[i −k] [i2 ]
∂2
x1i1 −k xi22 · · · xinn ∂1 1
k=1
i′
X
[i′ ] [i ]
· · · ∂n[in ] ·
j
[j]
(∂1 z)
k
[j]
x11 xi22 · · · xinn ∂1 1 ∂2 2 · · · ∂n[in ] · (∂1 z)
i′1 +i2 +···+in =r−k
j
X
j
deg(z) − j
deg(z)
[j]
[j]
=
· (∂1 z)
· (∂1 z) −
k
r−k
r
k=1
deg(z) − j
[j]
=
· (∂1 z)
r
[j]
deg(∂1 z)
[j]
=
· (∂1 z)
r
5
where (1) follows from Lemma 2.3:
[j]
[j]
xi11 ∂1 = ∂1 xi11 −
min{i1 ,j}
X
k=1
(2)
(4)
(5)
(6)
i1 i1 −k [j−k]
x
∂1
;
k 1
and (3) follow from Lemma 2.2;
is obtained by setting i′1 = i1 − k;
is true by induction on r;
follows from the Chu-Vandermonde identity ([Ask75, page 59])
X
j
a
j
a−j
=
.
r
k
r−k
k=0
This completes our first step.
Next we consider x1 · z and we have
X
[i ] [i ]
xi11 xi22 · · · xinn ∂1 1 ∂2 2 · · · ∂n[in ] ) · (x1 · z)
(
i1 +i2 +···+in =r
X
=(
[i ]
[i ]
xi11 xi22 · · · xinn (∂1 1 x1 )∂2 2 · · · ∂n[in ] ) · z
i1 +i2 +···+in =r
X
=(
[i ]
[i −1]
xi11 xi22 · · · xinn (x1 ∂1 1 + ∂1 1
[i ]
)∂2 2 · · · ∂n[in ] + x1
= x1 (
xi11 xi22
[i ]
xi22 · · · xinn ∂2 2 · · · ∂n[in ] ) · z
i2 +···+in =r
i1 +i2 +···+in =r,i1 ≥1
X
X
[i ] [i ]
· · · xinn ∂1 1 ∂2 2
i1 +i2 +···+in =r
· · · ∂n[in ]
X
+
[i −1] [i2 ]
∂2
x1i1 −1 xi22 · · · xinn ∂1 1
· · · ∂n[in ] ) · z
i1 +i2 +···+in =r,i1 ≥1
deg(z)
deg(z)
= x1 (
+
)z
r
r−1
deg(z) + 1
=
(x1 · z)
r
deg(x1 · z)
=
(x1 · z)
r
This finishes our second step in the case when s = 1.
Now we consider xs1 · z when s ≥ 2. By an easy induction we may assume x1s−1 z satisfies (2.1.1), now we
have
deg(x1 (xs−1 z))
deg(xs1 z)
s−1
s
s−1
Er (x1 z) = Er (x1 x1 z) =
(x1 (x z)) =
(xs1 z).
r
r
where the second equality is the case when s = 1 (because we assume x1s−1 z satisfies (2.1.1)). This completes
the proof of our theorem.
An immediate consequence of Theorem 2.6 on cyclic D-modules is the following.
D
Proposition 2.7. Let J be a homogeneous left ideal in D. Then
is Eulerian if and only if Er =
J
X
[i
]
[i
]
xi11 xi22 · · · xinn ∂1 1 ∂2 2 · · · ∂n[in ] ∈ J for every r.
i1 +i2 +···+in =r
D
D
Proof. According to Theorem 2.6,
is Eulerian if and only if 1 ∈
satisfies Euler’s formula (2.1.1). Since
J
J
D
D
1 has degree 0, 1 ∈
satisfies Euler’s formula (2.1.1) if and only if Er · 1 = Er = 0 ∈ , which holds if and
J
J
only if Er ∈ J for every r.
Proposition 2.8. If a graded D-module M is Eulerian, so are each graded submodule of M and each graded
quotient of M .
6
Proof. Let N be a graded submodule of M . Since each homogeneous element is also a homogeneous element
in M , it is clear that N is also Eulerian. Given a D-linear degree-preserving surjection ψ : M → M ′ and
a homogeneous element z ′ ∈ M ′ , there is a homogeneous z ∈ M with the same degree such that ψ(z) = z ′
and hence we have (for every r)
deg(z)
deg(z ′ )
′
Er · z = Er · ψ(z) = ψ(Er · z) = ψ(
· z) =
· z′.
r
r
This proves that M ′ is also Eulerian.
We end this section with the following result which is one of the key ingredients for our application to
local cohomology.
Theorem 2.9.
(1) The graded D-module R(ℓ) is Eulerian if and only if ℓ = 0.
(2) The graded D-module ∗ E(ℓ) is Eulerian if and only ℓ = n.
Proof. By Remark 2.5 (1), it suffices to show that R(0) and ∗ E(n) are Eulerian graded. It is clear that
R = R(0) is Eulerian by Remark 2.5 (3). Since ∗ E(ℓ) is spanned over K by xj11 xj22 · · · xjnn with each jt ≤ −1.
By the computation in Remark 2.5 (3), it is clear that ∗ E(n) is Eulerian graded (because in ∗ E(n), the
element xj11 xj22 · · · xjnn has degree j1 + · · · + jn ).
3. Eulerian graded D-module in characteristic 0
Throughout this section K will be a field of characteristic 0. In this section we collect some properties of
Eulerian graded D-modules when char(K) = 0. The main result is that, if a graded D-module M is Eulerian,
then so is Mf for each f ∈ R. This is one of the ingredients for our application to local cohomology in Section
5. First we observe that, in characteristic 0, if each homogeneous element z in a graded D-module M satisfies
(2.1.1) for r = 1 (instead of for all r ≥ 1), then M is Eulerian.
Proposition 3.1. Let M be a graded D-module. If E1 · z = deg(z) · z for every homogeneous element z ∈ M .
Then M is Eulerian.
deg(z)
Proof. We prove by induction that Er · z =
· z for every r ≥ 1. When r = 1 this is exactly
r
deg(z)
E1 · z = deg(z) · z which is given. Now suppose we know Er · z =
· z. By Proposition 2.4, we know
r
1
that Er+1 =
(E1 · Er − rEr ). So we have
r+1
1
(E1 · Er − rEr ) · z
Er+1 · z =
r+1
1
deg(z)
deg(z)
=
(E1 ·
·z−r·
· z)
r+1
r
r
deg(z)
1
·
· (deg(z) · z − rz)
=
r+1
r
1
deg(z)
=
·
(deg(z) − r) · z
r+1
r
deg(z)
=
·z
r+1
where the second equality uses the induction hypothesis. This finishes the proof.
Remark 3.2. As we have seen, Lemma 2.4 is quite useful when char(K) = 0. Unfortunately, this is no longer
the case once we are in characteristic p. For instance, when r ≡ −1(mod p), we can’t link Er+1 and Er via
Lemma 2.4. This is one of the reasons that we treat characteristic 0 and characteristic p separately in two
different sections.
n
X
D
xi ∂i ∈ J.
is Eulerian if and only if
Corollary 3.3. Let J be a homogeneous left ideal in D. Then
J
i=1
7
Proof. This is clear from Proposition 3.1 and (the proof of) Proposition 2.7.
Proposition 3.4. If M is an Eulerian graded D-module, so is S −1 M for each homogeneous multiplicative
system S ⊆ R. In particular, Mf is Eulerian for each homogeneous polynomial f ∈ R.
z
Proof. By Proposition 3.1, it suffices to show for each homogeneous f ∈ S and z ∈ M , we have E1 · t =
f
z
z
deg( t ) · t . Now we compute
f
f
E1 ·
z
ft
=
n
X
xi ∂i ·
=
n
X
xi ·
i=1
i=1
=
=
=
=
z
ft
f t · ∂i (z) − ∂i (f t ) · z
f 2t
n
n
X
X
1
t
xi ∂i (f t ) · z)
x
∂
(z)
−
(f
i
i
f 2t
i=1
i=1
1
1
· deg(z) · z − 2t · deg(f t ) · f t · z
ft
f
z
(deg(z) − deg(f t )) · t
f
z
z
deg( t ) · t
f
f
This finishes the proof.
Remark 3.5.
(1) It turns out that Eulerian graded D-modules are not stable under extension because
of the following short exact sequence of graded D-modules:
P
n
D
D
D
i=1 xi ∂i
0 → Pn
−−−
−−−→ Pn
→ Pn
→ 0,
2
h i=1 xi ∂i i
h( i=1 xi ∂i ) i
h i=1 xi ∂i i
Pn
P
D
D
i=1 xi ∂i
is the multiplication by ni=1 xi ∂i , i.e. a 7→
−−−
−−−→ Pn
where the map Pn
h i=1 xi ∂i i
h( i=1 xi ∂i )2 i
P
a · ( ni=1 xi ∂i ).
D
D
(2) Since dim( Pn
) = 2n − 1 and Pn
is Eulerian, finitely generated (even cyclic)
h i=1 xi ∂i i
h i=1 xi ∂i i
Eulerian graded D-modules may not be holonomic when n ≥ 2.
(3) When n = 1, it is rather straightforward to check that each finitely generated Eulerian graded
D-module is holonomic.
(4) As we will see in section 5, in characteristic 0, a vast class of graded D-modules (namely local
cohomology modules of R) are both Eulerian and holonomic.
4. Eulerian graded D-module in characteristic p > 0
Throughout this section K will be a field of characteristic p > 0. In this section we prove that being
Eulerian is preserved under localization. The proof is quite different from that in characteristic 0. We also
show that each graded F -module is always an Eulerian graded D-module, which will enable us to recover
the main result in [Zha12] in Section 5.
Proposition 4.1. If M is an Eulerian graded D-module, so is S −1 M for each homogeneous multiplicative
system S ⊆ R. In particular, Mf is Eulerian for each homogeneous polynomial f ∈ R.
[j]
e
Proof. First notice that, ∂i is Rp -linear if pe ≥ j + 1. So we have
e
e
z
[j] z
[j]
[j]
∂i (z) = ∂i (f p · pe ) = f p · ∂i ( pe )
f
f
8
[j]
This tells us that, if pe ≥ r +1 and f ∈ S, then ∂i (
we have
Er ·
z
1 [j]
∂ (z) for every j ≤ r in S −1 M . In particular
e ) =
p
f
f pe i
1
z
Er · z
e =
p
f
f pe
z
∈ S −1 M , we can multiply both the numerator and denominator by a large power
ft
e
z
f p −t z
of f and write t =
for some pe ≥ max{r + 1, t}. So we have
f
f pe
For any homogeneous
Er ·
z
ft
e
=
=
=
=
e
f p −t z
1
= pe Er · f p −t z
e
p
f
f
e
e
1 deg(f p −t ) + deg(z)
· f p −t z
e
p
f
r
e
pe −t
f
z
p · deg(f ) − deg(f t ) + deg(z)
·
f pe
r
z
deg( t )
z
f
· t
f
r
Er ·
where the last equality is because pe ≥ r + 1 and we are in characteristic p > 0. This finishes the proof.
Recall the definition of a graded F -module as follows.
Definition 4.2 (cf. Definitions 2.1 and 2.2 in [Zha12]). For each integer e ≥ 1, let e R denote the R-module
that is the same as R as a left R-module and whose right R-module structure is given by r′ · r = rp r′
for all r′ ∈ e R and r ∈ R. An F -module is an R-module M equipped with an R-module isomorphism
θ : M → F (M ) = 1 R ⊗R M . An F -module (M, θ) is called a graded F -module if M is graded and θ is
degree-preserving.
Remark 4.3. It is clear from the definition that, if (M, θ) is an F -module, the map
θ
F 2 (θ)
F (θ)
αe : M −
→ F (M ) −−−→ F 2 (M ) −−−−→ · · · → F e (M )
induced by θ is also an isomorphism.
This induces a D-module structure on M . To specify the induced D-module structure, it suffices to
[i ]
[i ]
specify how ∂1 1 · · · ∂n n acts on M . Choose
e such that pe ≥ (i1 + · · · + in ) + 1. Given each element z, we
P
consider αe (z) and we will write it as
yj ⊗ zj with yj ∈ e R and zj ∈ M . And we define
X [i ]
[i ]
∂1 1 · · · ∂n[in ] yj ⊗ zj ).
∂1 1 · · · ∂n[in ] z := α−1
e (
See [KLZ12, §1] for more details.
When an F -module (M, θ) is graded, the induced map αe is also degree-preserving. And hence M is
naturally a graded D-module. It turns out that each graded F -module is Eulerian as a graded D-module.
Theorem 4.4. If M is a graded F -module, then M is Eulerian graded as a D-module.
deg(z)
Proof. Pick any homogeneous element z ∈ M , we want to show Er · z =
· z for each r ≥ 1.
r
e
Pick e such that p ≥ r + 1. Since M is a graded F -module, we have a degree-preserving isomorphism
P
αe
M are homogeneous
and
FRe (M ). Assume αe (z) = i yi ⊗ zi where yi ∈ R and zi ∈
M −→
deg(z) =
deg(yi )
deg(z)
e
deg(yi ⊗ zi ) = p deg(zi ) + deg(yi ) for each i. In particular we have
=
for every i
r
r
9
(because we are in characteristic p > 0). So we know
X
(Er · yi ) ⊗ zi )
Er · z = α−1
e (
i
=
=
=
X deg(yi )
yi ⊗ zi )
r
i
deg(z) X
yi ⊗ zi )
αe−1 (
r
i
deg(z)
·z
r
α−1
e (
This finishes the proof.
5. An application to local cohomology
Let R be an arbitrary commutative Noetherian ring and I be an ideal of R. We recall that if I is generated
by f1 , . . . , fl ∈ R and M is any R-module, we have the Čech complex:
M
M
Mfj →
Mfj fk → · · · → Mf1 ···fl → 0
0→M →
j
j,k
whose i-th cohomology module is HIi (M ). Here the map Mfj1 ···fji → Mfk1 ···fki+1 induced by the corresponding differential is the natural localization (up to sign) if {j1 , . . . , ji } is a subset of {k1 , . . . , ki+1 } and is 0
otherwise.
When R is graded and I is a homogeneous ideal (i.e.f1 , . . . , fl are homogeneous elements in R) and M is
a graded R-module, each differential in the Čech complex is degree-preserving because natural localization
is so. It follows that each cohomology module HJi11 (· · · (HJiss (R))) is a graded R-module.
When R = K[x1 , . . . , xn ] with K a field of characteristic p > 0 and m = (x1 , . . . , xn ), it is proven in
[Zha12] (using the theory of graded F -modules) that
Theorem 5.1 (Theorem 3.4 in [Zha12]). Let R = K[x1 , . . . , xn ] be a polynomial ring over a field K of characi0
teristic p > 0 and J1 , . . . , Js be homogeneous ideals of R. Each local cohomology module Hm
(HJi11 · · · (HJiss (R)))
∗
i1
1
i0
is isomorphic to a direct sum of copies of E(n) (i.e. all socle elements of Hm (HJ1 · · · (HJiss (R))) must have
degree −n).
It is a natural question (and is asked in [Zha12]) whether the same result holds in characteristic 0. Using
our theory of Eulerian graded D-module, we can give a characteristic-free proof of the same result. In
particular we answer the question in characteristic 0 in the affirmative.
We begin with the following easy observation.
Proposition 5.2. Let J1 , . . . , Js be homogeneous ideals of R, then each local cohomology module HJi11 (· · · (HJiss (R)))
is a graded D-module.
Proof. Since natural localization map is D-linear (and so is each differential in the Čech complex), our
proposition follows immediately from the Čech complex characterization of local cohomology.
Theorem 5.3. Let J1 , . . . , Js be homogeneous ideals of R, then each local cohomology module HJi11 (· · · (HJiss (R)))
(considered as a graded D-module) is Eulerian.
Proof. This follows immediately from Propositions 3.4, 4.1 and 2.8, and the Čech complex characterization
of local cohomology.
Proposition 5.4 (cf. Proposition 2.3 in [Lyu93] in characteristic 0 and Lemma (b) on page 208 in [Lyu00] in
characteristic p > 0). Let R = K[x1 , . . . , xn ] and m = (x1 , . . . , xn ). There is a degree-preserving isomorphism
D/Dm → ∗ E .
1when s = 1, this is also proved in [Zha11, page 615]
10
Proof. It is proven in Proposition 2.3 in [Lyu93] in characteristic 0 and Lemma (b) on page 208 in [Lyu00]
in characteristic p > 0 that the map D/Dm → ∗ E given by
(5.4.1)
[i ]
1 −1
∂1 1 · · · ∂n[in ] 7→ (−1)i1 +···+in x−i
· · · xn−in −1
1
[i ]
[i ]
is an isomorphism. The grading on D/Dm is induced by the one on D and hence deg(∂1 1 · · · ∂n n ) = −(i1 +
−i1 −1
−1
· · · + in ). Since in ∗ E, the socle element x−1
· · · xn−in −1 ) =
1 · · · xn has degree 0, it follows that deg(x1
−(i1 + · · · + in ). Therefore it follows that (5.4.1) defines a degree-preserving isomorphism D/Dm → ∗ E.
Proposition 5.5 (cf. Theorem 2.4(a) in [Lyu93] in characteristic 0 and Lemma (c) on page 208 in [Lyu00]
in characteristic
p >L0). Let M be a graded D-module. If SuppR (M ) = {m}, then as a graded D-module
L D
M∼
(n
)
= j Dm j ∼
= j ∗ E(nj ).
Proof. M is a graded D-module hence also graded as an R-module. We first claim that the socle of M can
be generated by homogeneous elements and we reason
Pt as follows. Pick a generator g of the socle, we can
write it as a sum of homogeneous elements g = i=1 gi where each gi has a different degree. For every
Pt
xj ∈ m, we have i=1 xj · gi = xj · g = 0 (since g is killed by m), hence xj · gi = 0 for every i (because each
xj · gi has a different degree). Therefore gi is killed by every xj , hence is killed by m, so gi is in the socle for
each i. This proves our claim. We also note that since the socle is killed by m, a minimal homogeneous set
of generators is actually a homogeneous K-basis.
Let {ej } be a homogeneous K-basis of the socle of M with deg(ej ) = −nj . There is a degree-preserving
L D
homomorphism of D-modules j Dm
(nj ) → M which sends 1 of the j-th copy to ej . This map is injective
L D
because it induces an isomorphism on socles and
j Dm (nj ) is supported only at m (as an R-module).
L D
L
L
L D
∗
∼
N where N is some
By 5.4,
j Dm (nj )
j E(nj ) is an injective R-module. So M =
j Dm (nj ) =
graded
R-module
supported
only
at
m.
Since
the
map
on
the
socles
is
an
isomorphism,
N = 0, so M =
L D
L ∗
∼
E(n
).
(n
)
=
j
j
j Dm
j
Theorem 5.6. Let M be an Eulerian graded D-module. If SuppR (M ) = {m}, then M is isomorphic (as a
graded D-module) to a direct sum of copies of ∗ E(n).
L
Proof. Since M is supported only at m, we know it is isomorphic to j ∗ E(nj ) as a graded D-module by
L ∗
Proposition 5.5. By our assumption, M is Eulerian, so is
j E(nj ). It follows from Theorem 2.9 that
nj = n for each j, i.e. M is isomorphic (as a graded D-module) to a direct sum of copies of ∗ E(n). This
finishes the proof.
i0 i1
Corollary 5.7. Let J1 , . . . , Js be homogeneous ideals of R, then Hm
HJ1 · · · HJiss (R) is isomorphic (as a
∗
i0 i1
graded D-module) to a direct sum of copies of E(n) (or equivalently, all socle elements of each Hm
HJ1 · · · HJiss (R)
must have degree −n ).
Proof. This follows immediately from Theorems 5.3 and 5.6.
Remark 5.8. It is proven in [Lyu93] (resp, [Lyu97]) that every HJi11 (· · · (HJiss (R))) is holonomic (resp, F finite) as a D-module (resp, F -module) in characteristic 0 (resp, characteristic p > 0). Therefore in any case
we know HJi11 (· · · (HJiss (R))) has finite Bass numbers (cf. Theorem 3.4(d) in [Lyu93] and Theorem 2.11 in
i0 i1
HJ1 · · · HJiss (R) ∼
[Lyu97]). It follows from this and Corollary 5.7 that Hm
= ∗ E(n)c for some integer c < ∞.
6. Remarks on the graded injective hull of R/P when P is a homogeneous prime ideal
We have seen in Theorem 2.9 that ∗ E(ℓ) = ∗ E(R/m)(ℓ) is Eulerian graded if and only if ℓ = n. In this
section we wish to extend this result to ∗ E(R/P ) where P is a non-maximal homogeneous prime ideal (here
∗
E(R/P ) denotes the graded injective hull of R/P , see cf. [BS98, Chapter 13.2]]). To this end, we will
discuss in detail the graded structures of ∗ E(R/P ) as an R-module and as a D-module. The underlying idea
is that, there does not exist a canonical choice of grading on ∗ E(R/P ) when it is considered as a graded
R-module; however, there is a canonical grading when it is considered as a graded D-module.
11
Remark 6.1. [ ∗ E(R/P ) as a graded R-module] Since P 6= m, there is at least one xi that is not contained
in P . Hence the multiplication by xi induces an automorphism on ∗ E(R/P ), and consequently we have a
degree-preserving isomorphism
·xi ∗
∗
E(R/P ).
E(R/P )(−1) −−→
It follows immediately that
E(R/P ) ∼
= ∗ E(R/P )(m)
for each integer m in the category of graded R-modules. In other words, we have
∗
∼ ∗ E(R/P )(j)
E(R/P )(i) =
∗
for all integers i and j.
In some sense, this tells us that ∗ E(R/P ) does not have a canonical grading when considered merely as
a graded R-module.
However, as we will see, ∗ E(R/P ) is equipped with a natural Eulerian graded D-module structure, and
from this point of view there is indeed a unique natural grading on ∗ E(R/P ). The D-module structure
on ∗ E(R/P ) is obtained via considering HPht P (R)(P ) where (·)(P ) denotes homogeneous localization with
respect to P (i.e. inverting all homogeneous elements not in P ), which has a natural grading as follows.
Remark 6.2 (Grading on HPht P (R)(P ) ). Choose a set of homogeneous generators f1 , . . . , ft of P and consider
the Čech complex
M
M
Rfi1 ···fij → · · · → Rf1 ···ft → 0.
Rfi → · · · →
C • (P ) : 0 → R →
i1 <i2 <···<ij
i
Then, since each module has a natural grading and each differential is degree-preserving, H h (C • (P )) also has
a natural grading (h = ht P ), hence so is H h (C • (P ))(P ) . We will identify HPht P (R)(P ) with H h (C • (P ))(P )
with its natural grading.
Proposition 6.3.
∗
E(R/P ) ∼
= HPht P (R)(P ) in the category of graded R-modules.
Proof. We have a graded injective resolution of R (or a *-injective resolution of R, cf. [BS98, Chapter 13])
M
∗
(6.3.1)
0 → R → ∗ E(R)(d0 ) → · · · →
E(R/Q)(dQ ) → · · · → ∗ E(R/m)(dm ) → 0
ht Q=s
where each dQ is an integer depending on Q. Notice that when Q 6= m, ∗ E(R/Q)(i) ∼
= ∗ E(R/Q)(j) for all
integers i and j by Remark 6.1. So the above resolution can be written as
M
∗
(6.3.2)
0 → R → ∗ E(R) → · · · →
E(R/Q) → · · · → ∗ E(R/m)(dm ) → 0
ht Q=s
HPh (R)(P )
Let h = ht P . Then
is the homogeneous localization of the h-th homology of (6.3.2) when we
apply ΓP (·). But when we apply ΓP (·), (6.3.2) becomes
M
∗
(6.3.3)
0 → 0 → · · · → ∗ E(R/P ) → ΓP (
E(R/Q)) → · · · → ΓP ( ∗ E(R/m)(dm )) → 0
ht Q=h+1
and when we do homogeneous localization at P to (6.3.3), we get
0 → · · · → 0 → ∗ E(R/P ) → 0 → · · · → 0
n
So the h-th homology is exactly ∗ E(R/P ) (and when P = m, we get Hm
(R) ∼
= ∗ E(R/m)(dm ), so actually
dm = n). This finishes the proof.
Remark 6.4 (D-module structure on ∗ E(R/P )). Since HPh (R)(P ) has a natural graded D-module structure,
it follows from Proposition 6.3 that ∗ E(R/P ) also has a natural graded D-module structure.
Since ∗ E(R/P ) is a graded D-module, it is natural to ask the following question.
Question 6.5. Let P be a homogeneous prime ideal in R. Is there a natural grading on ∗ E(R/P ) making it
Eulerian graded?
12
Remark 6.6. HPht P (R)(P ) is always Eulerian graded by Theorem 5.3, Propositions 3.4, and 4.1. From Remark
2.5(1), we know that, in the category of graded D-modules, ∗ E(R/P )(ℓ) is Eulerian graded for exactly one
ℓ, we will identify this “canonical” ℓ.
Contrary to the case when we consider ∗ E(R/P ) as a graded R-module, we can see that in the category
of graded D-modules we have
∗
∼ ∗ E(R/P )(j) if and only if i = j.
E(R/P )(i) =
(Otherwise we would have ∗ E(R/P )(ℓ) ∼
= ∗ E(R/P )(ℓ + j − i) for every ℓ, and hence there would be more
∗
than one choice of ℓ such that E(R/P )(ℓ) is Eulerian.)
We wish to find the natural grading on ∗ E(R/P ) that makes it Eulerian, and we need the following lemma
(which may be well-known to experts).
Lemma 6.7. We have a canonical degree-preserving isomorphism (h = ht P )
Exth (R/P, R) ∼
= HomR (R/P, H h (R))
P
R
Proof. HPh (R) is the h-th homology of (6.3.2) when we apply ΓP (·), which is the h-th homology of (6.3.3),
which is the kernel of ∗ E(R/P ) → ΓP (⊕ht Q=h+1 ∗ E(R/Q)). Since HomR (R/P, ·) is left exact, we know that
HomR (R/P, HPh (R)) is isomorphic to the kernel of
∗
E(R/P ) → HomR (R/P, ΓP (⊕ht Q=h+1 ∗ E(R/Q))) ∼
= HomR (R/P, ⊕ht Q=h+1 ∗ E(R/Q)))
But this is exactly the h-th homology of (6.3.2) when we apply HomR (R/P, ·), which by definition is
ExthR (R/P, R). And we want to emphasize here that the isomorphism obtained does not depend on the grading on ∗ E(R/P ) as long as ∗ E(R/P ) is equipped with the same grading when we calculate ExthR (R/P, R)
and HomR (R/P, HPh (R)) as above.
Definition 6.8. For a d-dimensional graded K-algebra S with irrelevant maximal ideal m, the a-invariant
of S is defined to be
d
a(S) = max{t ∈ Z|Hm
(S)t 6= 0}.
Proposition 6.9. We have min{t|(AnnHPht P (R) P )t 6= 0} = −a(R/P )−n. Hence we have a degree-preserving
inclusion
R/P ֒→ HPht P (R)(−a(R/P ) − n) ֒→ HPht P (R)(P ) (−a(R/P ) − n).
Proof. Let h = ht P and let s = min{t|(AnnHPht P (R) P )t 6= 0}. By lemma 6.7, we know
ExthR (R/P, R) ∼
= HomR (R/P, HPh (R)) ∼
= AnnHPh (R) P
so we know
s = min{t|(ExthR (R/P, R)t 6= 0} = min{t|(ExthR (R/P, R(−n))t 6= 0} − n
by graded local duality
n−h
min{t|(ExthR (R/P, R(−n))t 6= 0} = − max{t ∈ Z|Hm
(R/P )t 6= 0} = −a(R/P )
Hence we get s = −a(R/P ) − n. The second statement follows from the first one by sending 1 in R/P to
any element in AnnHPht P (R) P of degree −a(R/P ) − n.
Remark 6.10. From what we have discussed so far, we can see that HPht P (R)(P ) is a graded injective module
(or *-injective module) and there is a degree-preserving inclusion R/P ֒→ HPht P (R)(P ) (−a(R/P )−n), always
sending 1 in R/P to the lowest degree element in AnnHPht P (R) P . Therefore we propose a “canonical” grading
on ∗ E(R/P ) to the effect that ∗ E(R/P ) can be identified with HPht P (R)(P ) (−a(R/P )−n) (where the grading
on HPht P (R)(P ) is obtained via Čech complex).
We end with the following proposition.
Proposition 6.11 (Compare with Theorem 13.2.10 and Lemma 13.3.3 in [BS98]). Given the grading on
∗
E(R/P ) as proposed in Remark 6.10, we have that
(1) ∗ E(R/P )(ℓ) is Eulerian if and only if ℓ = a(R/P ) + n;
13
(2) the minimal graded injective resolution (or *-injective resolution) of R can be written as
M
∗
0 → R → ∗ E(R) → · · · →
E(R/P )(a(R/P ) + n) → · · · → ∗ E(R/m)(n) → 0.
ht P =j
Proof. (1). This is clear by our grading on ∗ E(R/P ) and the fact that there is a unique grading on
HPht P (R)(P ) that makes it Eulerian.
(2). This follows immediately from the calculation of HPht P (R)(P ) using the minimal *-injective resolution
of R (note that a(R) = −n and a(R/m) = 0).
References
[Ask75] R. Askey: Orthogonal polynomials and special functions, Society for Industrial and Applied Mathematics, Philadelphia, Pa., 1975. 0481145 (58 #1288)
[BS98] M. P. Brodmann and R. Y. Sharp: Local cohomology: an algebraic introduction with geometric applications, Cambridge Studies in Advanced Mathematics, vol. 60, Cambridge University Press, Cambridge, 1998. MR1613627 (99h:13020)
[Gro67] A. Grothendieck: Éléments de géométrie algébrique. IV. Étude locale des schémas et des morphismes de schémas
IV, Inst. Hautes Études Sci. Publ. Math. (1967), no. 32, 361. MR0238860 (39 #220)
[KLZ12] M. Katzman, G. Lyubeznik, and W. Zhang: Two interesting examples of D-modules in characteristic p > 0, Bull.
Lond. Math. Soc. 44 (2012), no. 6, 1116–1122. 3007644
[Luc78] E. Lucas: Sur les congruences des nombres eulériens et les coefficients différentiels des functions trigonométriques
suivant un module premier, Bull. Soc. Math. France 6 (1878), 49–54.
[Lyu93] G. Lyubeznik: Finiteness properties of local cohomology modules (an application of D-modules to commutative
algebra), Invent. Math. 113 (1993), no. 1, 41–55. 1223223 (94e:13032)
[Lyu97] G. Lyubeznik: F -modules: applications to local cohomology and D-modules in characteristic p > 0, J. Reine Angew.
Math. 491 (1997), 65–130. MR1476089 (99c:13005)
[Lyu00] G. Lyubeznik: Injective dimension of D-modules: a characteristic-free approach, J.Pure Appl.Algebra 149 (2000),
205–212.
[Lyu11] G. Lyubeznik: A characteristic-free proof of a basic result on D-modules, J. Pure Appl. Algebra 215 (2011), no. 8,
2019–2023. 2776441 (2012b:13068)
[Zha11] W. Zhang: Lyubeznik numbers of projective schemes, Adv. Math. 228 (2011), no. 1, 575–616. 2822240
[Zha12] Y. Zhang: Graded F -modules and local cohomology, Bull. Lond. Math. Soc. 44 (2012), no. 4, 758–762. 2967243
Department of Mathematics, University of Michigan, Ann Arbor, Michigan 48109
E-mail address: [email protected]
Department of Mathematics, University of Nebraska, Lincoln, Nebraska 68588
E-mail address: [email protected]
14
| 0math.AC
|
Free Applicative Functors
Paolo Capriotti
Ambrus Kaposi
University of Nottingham, United Kingdom
University of Nottingham, United Kingdom
[email protected]
[email protected]
Applicative functors [6] are a generalisation of monads. Both allow the expression of effectful computations into an otherwise pure language, like Haskell [5]. Applicative functors are to be preferred
to monads when the structure of a computation is fixed a priori. That makes it possible to perform
certain kinds of static analysis on applicative values. We define a notion of free applicative functor,
prove that it satisfies the appropriate laws, and that the construction is left adjoint to a suitable forgetful functor. We show how free applicative functors can be used to implement embedded DSLs which
can be statically analysed.
1
Introduction
Free monads in Haskell are a very well-known and practically used construction. Given any endofunctor
f, the free monad on f is given by a simple inductive definition:
data Free f a
= Return a
| Free (f (Free f a))
The typical use case for this construction is creating embedded DSLs (see for example [10], where
Free is called Term). In this context, the functor f is usually obtained as the coproduct of a number
of functors representing “basic operations”, and the resulting DSL is the minimal embedded language
including those operations.
One problem of the free monad approach is that programs written in a monadic DSL are not amenable to
static analysis. It is impossible to examine the structure of a monadic computation without executing it.
In this paper, we show how a similar “free construction” can be realised in the context of applicative
functors. In particular, we make the following contributions:
• We give two definitions of free applicative functor in Haskell (section 2), and show that they are
equivalent (section 5).
• We prove that our definition is correct, in the sense that it really is an applicative functor (section
6), and that it is “free” in a precise sense (section 7).
• We present a number of examples where the use of free applicative functors helps make the code
more elegant, removes duplication or enables certain kinds of optimizations which are not possible
when using free monads. We describe the differences between expressivity of DSLs using free
applicatives and free monads (section 3).
• We compare our definition to other existing implementations of the same idea (section 10).
N.R. Krishnaswami and P.B. Levy (Eds.):
Mathematically Structured Functional Programming 2014 (MSFP 2014).
EPTCS 153, 2014, pp. 2–30, doi:10.4204/EPTCS.153.2
c P. Capriotti & A. Kaposi
This work is licensed under the
Creative Commons Attribution License.
P. Capriotti & A. Kaposi
3
Applicative functors can be regarded as monoids in the category of endofunctors with Day convolution
(see for instance [3], example 3.2.2). There exists a general theory for constructing free monoids in
monoidal categories [4], but in this paper we aim to describe the special case of applicative functors
using a formalism that is accessible to an audience of Haskell programmers.
Familiarity with applicative functors is not required, although it is helpful to understand the motivation
behind this work. We make use of category theoretical concepts to justify our definition, but the Haskell
code we present can also stand on its own.
The proofs in this paper are carried out using equational reasoning in an informally defined total subset
of Haskell. In sections 8 and 9 we will show how to interpret all our definitions and proofs in a general
(locally presentable) cartesian closed category, such as the category of sets.
1.1 Applicative functors
Applicative functors (also called idioms) were first introduced in [6] as a generalisation of monads that
provides a lighter notation for expressing monadic computations in an applicative style.
They have since been used in a variety of different applications, including efficient parsing (see section
1.4), regular expressions and bidirectional routing.
Applicative functors are defined by the following type class:
class Functor f ⇒ Applicative f where
pure :: a → f a
( <*> ) :: f (a → b) → f a → f b
The idea is that a value of type f a represents an “effectful” computation returning a result of type a. The
pure method creates a trivial computation without any effect, and ( <*> ) allows two computations to be
sequenced, by applying a function returned by the first, to the value returned by the second.
Since every monad can be made into an applicative functor in a canonical way,1 the abundance of monads
in the practice of Haskell programming naturally results in a significant number of practically useful
applicative functors.
Applicatives not arising from monads, however, are not as widespread, probably because, although it is
relatively easy to combine existing applicatives (see for example [7]), techniques to construct new ones
have not been thoroughly explored so far.
In this paper we are going to define an applicative functor FreeA f for any Haskell functor f, thus
providing a systematic way to create new applicatives, which can be used for a variety of applications.
The meaning of FreeA f will be clarified in section 7, but for the sake of the following examples, FreeA f
can be thought of as the “simplest” applicative functor which can be built using f.
1.2 Example: option parsers
To illustrate how the free applicative construction can be used in practice, we take as a running example
a parser for options of a command-line tool.
For simplicity, we will limit ourselves to an interface which can only accept options that take a single
argument. We will use a double dash as a prefix for the option name.
For example, a tool to create a new user in a Unix system could be used as follows:
1 To
be precise, there are two canonical ways to turn a monad into an applicative functor, with opposite orderings of effects.
Free Applicative Functors
4
create_user --username john \
--fullname "John Doe" \
--id 1002
Our parser could be run over the argument list and it would return a record of the following type:
data User = User
{ username :: String
, fullname :: String
, id :: Int}
deriving Show
Furthermore, given a parser, it should be possible to automatically produce a summary of all the options
that it supports, to be presented to the user of the tool as documentation.
We can define a data structure representing a parser for an individual option, with a specified type, as a
functor:
data Option a = Option
{ optName :: String
, optDefault :: Maybe a
, optReader :: String → Maybe a}
deriving Functor
We now want to create a DSL based on the Option functor, which would allow us to combine options for
different types into a single value representing the full parser. As stated in the introduction, a common
way to create a DSL from a functor is to use free monads.
However, taking the free monad over the Option functor would not be very useful here. First of all,
sequencing of options should be independent: later options should not depend on the value parsed by
previous ones. Secondly, monads cannot be inspected without running them, so there is no way to obtain
a summary of all options of a parser automatically.
What we really need is a way to construct a parser DSL in such a way that the values returned by the
individual options can be combined using an Applicative interface. And that is exactly what FreeA
will provide.
Thus, if we use FreeA Option a as our embedded DSL, we can interpret it as the type of a parser
with an unspecified number of options, of possibly different types. When run, those options would be
matched against the input command line, in an arbitrary order, and the resulting values will be eventually
combined to obtain a final result of type a.
In our specific example, an expression to specify the command line option parser for create_user
would look like this:
userP :: FreeA Option User
userP = User
<$> one (Option "username" Nothing Just)
<*> one (Option "fullname" (Just "") Just)
<*> one (Option "id" Nothing readInt)
readInt :: String → Maybe Int
where we need a “generic smart constructor”:
P. Capriotti & A. Kaposi
5
one :: Option a → FreeA Option a
which lifts an option to a parser.
1.3 Example: web service client
One of the applications of free monads, exemplified in [10], is the definition of special-purpose monads,
allowing to express computations which make use of a limited and well-defined subset of IO operations.
Given the following functor:
data WebService a =
GET {url :: URL , params :: [String] , result :: (String → a)}
| POST {url :: URL , params :: [String] , body :: String , cont :: a}
deriving Functor
the free monad on WebService allows the definition of an application interacting with a web service
with the same convenience as the IO monad, once “smart constructors” are defined for the two basic
operations of getting and posting:
get :: URL → [String] → Free WebService String
get url params = Free (GET url params Return)
post :: URL → [String] → String → Free WebService ()
post url params body = Free (POST url params body (Return ()))
For example, one can implement an operation which copies data from one server to another as follows:
copy :: URL → [String] → URL → [String] → Free WebService ()
copy srcURL srcPars dstURL dstPars = get srcURL srcPars >>= post dstURL dstPars
For some applications, we might need to have more control over the operations that are going to be executed when we eventually run the embedded program contained in a value of type Free WebService a.
For example, a web service client application executing a large number of GET and POST operations might
want to rate limit the number of requests to a particular server by putting delays between them, and, on
the other hand, parallelise requests to different servers. Another useful feature would be to estimate the
time it would take to execute an embedded Web Service application.
However, there is no way to achieve that using the free monad approach. In fact, it is not even possible
to define a function like:
count :: Free WebService a → Int
which returns the total number of GET/POST operations performed by a value of type Free WebService a.
To see why, consider the following example, which updates the email field in all the blog posts on a
particular website:
updateEmails :: String → Free WebService ()
updateEmails newEmail = do
entryURLs ← get "myblog.com" ["list_entries"]
forM (words entryURLs) $ λ entryURL →
post entryURL ["updateEmail"] newEmail
Free Applicative Functors
6
Now, the number of POST operations performed by updateEmails is the same as the number of blog
posts on myblog.com which cannot be determined by a pure function like count.
The FreeA construction, presented in this paper, represents a general solution for the problem of constructing embedded languages that allow the definition of functions performing static analysis on embedded programs, of which count :: FreeA WebService a → Int is a very simple example.
1.4 Example: applicative parsers
The idea that monads are “too flexible” has also been explored, again in the context of parsing, by Swierstra and Duponcheel [9], who showed how to improve both performance and error-reporting capabilities
of an embedded language for grammars by giving up some of the expressivity of monads.
The basic principle is that, by weakening the monadic interface to that of an applicative functor (or, more
precisely, an alternative functor), it becomes possible to perform enough static analysis to compute first
sets for productions.
The approach followed in [9] is ad-hoc: an applicative functor is defined, which keeps track of first
sets, and whether a parser accepts the empty string. This is combined with a traditional monadic parser,
regarded as an applicative functor, using a generalised semi-direct product, as described in [7].
The question, then, is whether it is possible to express this construction in a general form, in such a
way that, given a functor representing a notion of “parser” for an individual symbol in the input stream,
applying the construction one would automatically get an Applicative functor, allowing such elementary
parsers to be sequenced.
Free applicative functors can be used to that end. We start with a functor f, such that f a describes an
elementary parser for individual elements of the input, returning values of type a. FreeA f a is then a
parser which can be used on the full input, and combines all the outputs of the individual parsers out of
which it is built, yielding a result of type a.
Unfortunately, applying this technique directly results in a strictly less expressive solution. In fact, since
FreeA f is the simplest applicative over f, it is necessarily just an applicative, i.e. it cannot also have an
Alternative instance, which in this case is essential.
The Alternative type class is defined as follows:
class Applicative f ⇒ Alternative f where
empty :: f a
( <|> ) :: f a → f a → f a
An Alternative instance gives an applicative functor the structure of a monoid, with empty as the unit
element, and <|> as the binary operation. In the case of parsers, empty matches no input string, while
<|> is a choice operator between two parsers.
We discuss the issue of Alternative in more detail in section 11.
2
Definition of free applicative functors
To obtain a suitable definition for the free applicative functor generated by a functor f, we first pause
to reflect on how one could naturally arrive at the definition of the Applicative class via an obvious
generalisation of the notion of functor.
Given a functor f, the fmap method gives us a way to lift unary pure functions a → b to effectful
functions f a → f b, but what about functions of arbitrary arity?
P. Capriotti & A. Kaposi
7
For example, given a value of type a, we can regard it as a nullary pure function, which we might want
to lift to a value of type f a.
Similarly, given a binary function h :: a → b → c, it is quite reasonable to ask for a lifting of h to
something of type f a → f b → f c.
The Functor instance alone cannot provide either of such liftings, nor any of the higher-arity liftings
which we could define.
It is therefore natural to define a type class for generalised functors, able to lift functions of arbitrary
arity:
class Functor f ⇒ MultiFunctor f where
fmap0 :: a → f a
fmap1 :: (a → b) → f a → f b
fmap1 = fmap
fmap2 :: (a → b → c) → f a → f b → f c
It is easy to see that a higher-arity fmapn can now be defined in terms of fmap2 . For example, for n = 3:
fmap3 :: MultiFunctor f
⇒ (a → b → c → d)
→fa→fb→fc→fd
fmap3 h x y z = fmap2 ( $ ) (fmap2 h x y) z
However, before trying to think of what the laws for such a type class ought to be, we can observe that
MultiFunctor is actually none other than Applicative in disguise.
In fact, fmap0 has exactly the same type as pure, and we can easily convert fmap2 to ( <*> ) and vice
versa:
g <*> x = fmap2 ( $ ) g x
fmap2 h x y = fmap h x <*> y
The difference between ( <*> ) and fmap2 is that ( <*> ) expects the first two arguments of fmap2 , of
types a → b → c and f a respectively, to be combined in a single argument of type f (b → c).
This can always be done with a single use of fmap, so, if we assume that f is a functor, ( <*> ) and
fmap2 are effectively equivalent.
Nevertheless, this roundabout way of arriving to the definition of Applicative shows that an applicative
functor is just a functor that knows how to lift functions of arbitrary arities. An overloaded notation to
express the application of fmapi for all i is defined in [6], where it is referred to as idiom brackets.
Given a pure function of arbitrary arity and effectful arguments:
h : b1 → b2 → · · · → bn → a
x1 : f b1
x2 : f b2
···
xn : f bn
the idiom bracket notation is defined as:
J h x1 x2 · · · xn K = pure h <*> x1 <*> x2 <*> · · · <*> xn
Free Applicative Functors
8
We can build such an expression formally by using a PureL constructor corresponding to pure and a
left-associative infix ( :*: ) constructor corresponding to ( <*> ):
PureL h :*: x1 :*: x2 :*: · · · :*: xn
The corresponding inductive definition is:
data FreeAL f a
= PureL a
| ∀b.FreeAL f (b → a) :*: f b
infixl 4 :*:
The MultiFunctor typeclass, the idiom brackets and the FreeAL definition correspond to the left parenthesised canonical form2 of expressions built with pure and ( <*> ). Just as lists built with concatenation
have two canonical forms (cons-list and snoc-list) we can also define a right-parenthesised canonical form
for applicative functors — a pure value over which a sequence of effectful functions are applied:
x : b1
h1 : f (b1 → b2 )
h2 : f (b2 → b3 )
···
hn : f (bn → a)
hn <*> (· · · <*> (h2 <*> (h1 <*> pure x)) · · · )
Replacing pure with a constructor Pure and ( <*> ) by a right-associative infix ( :$: ) constructor gives
the following expression:
hn :$: · · · :$: h2 :$: h1 :$: Pure x
The corresponding inductive type:
data FreeA f a
= Pure a
| ∀b.f (b → a) :$: FreeA f b
infixr 4 :$:
FreeAL and FreeA are isomorphic (see section 5); we pick the right-parenthesised version as our official
definition since it is simpler to define the Functor and Applicative instances:
instan e Functor f ⇒ Functor (FreeA f) where
fmap g (Pure x) = Pure (g x)
fmap g (h :$: x) = fmap (g ◦ ) h :$: x
The functor laws can be verified by structural induction, simply applying the definitions and using the
functor laws for f.
instan e Functor f ⇒ Applicative (FreeA f) where
pure = Pure
2 Sometimes
called simplified form because it is not necessarily unique.
P. Capriotti & A. Kaposi
9
Pure g <*> y = fmap g y
(h :$: x) <*> y = fmap uncurry h :$:
(( , ) <$> x <*> y)
In the last clause of the Applicative instance, h has type f (x → y → z), and we need to return a value
of type FreeA f z. Since ( :$: ) only allows us to express applications of 1-argument “functions”, we
uncurry h to get a value of type f ((x , y) → z), then we use ( <*> ) recursively (see section 8 for a
justification of this recursive call) to pair x and y into a value of type FreeA f (x , y), and finally use the
( :$: ) constructor to build the result. Note the analogy between the definition of ( <*> ) and ( ++ ) for
lists.
3
Applications
3.1 Example: option parsers (continued)
By using our definition of free applicative, we can compose the command line option parser exactly
as shown in section 1.2 in the definition of userP. The smart constructor one which lifts an option (a
functor representing a basic operation of our embedded language) to a term in our language can now be
implemented as follows:
one :: Option a → FreeA Option a
one opt = fmap const opt :$: Pure ()
A function which computes the global default value of a parser can also be defined:
parserDefault :: FreeA Option a → Maybe a
parserDefault (Pure x) = Just x
parserDefault (g :$: x) =
optDefault g <*> parserDefault x
In section 7 we show that our definition is a free construction which gives us general ways to structure
programs. Specifically, we are able to define a generic version of one which works for any functor.
By exploiting the adjunction describing the free construction we are able to shorten the definition of
parserDefault, define a function listing all possible options and a function parsing a list of command
line arguments given in arbitrary order (section 7.1).
3.2 Example: web service client (continued)
In section 1.3 we showed an embedded DSL for web service clients based on free monads does not
support certain kinds of static analysis.
However, we can now remedy this by using a free applicative, over the same functor WebService. In
fact, the count function is now definable for FreeA WebService a. Moreover, this is not limited to this
particular example: it is possible to define count for the free applicative over any functor.
count :: FreeA f a → Int
count (Pure ) = 0
count ( :$: u) = 1 + count u
Free Applicative Functors
10
Static analysis of the embedded code now also allows decorating requests with parallelization instructions statically as well as rearranging requests to the same server.
Of course, the extra power comes at a cost. Namely, the expressivity of the corresponding embedded
language is severely reduced.
Using FreeA WebService, all the URLs of the servers to which requests are sent must be known in
advance, as well as the parameters and content of every request.
In particular, what one posts to a server cannot depend on what has been previously read from another
server, so operations like copy cannot be implemented.
3.3 Summary of examples
Applicative functors are useful for describing certain kinds of effectful computations. The free applicative construct over a given functor specifying the “basic operations” of an embedded language gives
rise to terms of the embedded DSL built by applicative operators. These terms are only capable of
representing a certain kind of effectful computation which can be described best with the help of the
left-parenthesised canonical form: a pure function applied to effectful arguments. The calculation of the
arguments may involve effects but in the end the arguments are composed by a pure function, which
means that the effects performed are fixed when specifying the applicative expression.
In the case of the option parser example userP, the pure function is given by the User constructor and
the “basic operation” Option is defining an option. The effects performed depend on how an evaluator is defined over an expression of type FreeA Option a and the order of effects can depend on the
implementation of the evaluator.
For example, if one defines an embedded language for querying a database, and constructs applicative
expressions using FreeA, one might analyze the applicative expression and collect information on the
individual database queries by defining functions similar to the count function in the web service example. Then, different, possibly expensive duplicate queries can be merged and performed at once instead
of executing the effectful computations one by one. By restricting the expressivity of our language we
gain freedom in defining how the evaluator works.
One might define parts of an expression in an embedded DSL using the usual free monad construction,
other parts using FreeA and compose them by lifting the free applicative expression to the free monad
using the following function:
liftA2M :: Functor f ⇒ FreeA f a → Free f a
liftA2M (Pure x) = Return x
liftA2M (h :$: x) = Free
(fmap (λ f → fmap f (liftA2M x)) h)
In the parts of the expression defined using the free monad construction, the order of effects is fixed
and the effects performed can depend on the result of previous effectful computations, while the free
applicative parts have a fixed structure with effects not depending on each other. The monadic parts of
the computation can depend on the result of static analysis carried out over the applicative part:
test :: FreeA FileSystem Int → Free FileSystem ()
test op = do
...
let n = count op -- result of static analysis
n′ ← liftA2M op -- result of applicative computation
P. Capriotti & A. Kaposi
11
max ← read "max"
when (max > n + n′ ) $ write "/tmp/test" "blah"
...
The possibility of using the results of static analysis instead of the need of specifying them by hand (in
our example, this would account to counting certain function calls in an expression by looking at the
code) can make the program less redundant.
4
Parametricity
In order to prove anything about our free applicative construction, we need to make an important observation about its definition.
The ( :$: ) constructor is defined using an existential type b, and it is clear intuitively that there is no
way, given a value of the form g :$: x, to make use of the type b hidden in it.
More specifically, any function on FreeA f a must be defined polymorphically over all possible types b
which could be used for the existentially quantified variable in the definition of ( :$: ).
To make this intuition precise, we assume that some form of relational parametricity [8] [11] holds in
our total subset of Haskell. In particular, in the case of the ( :$: ) constructor, we require that:
( :$: ) :: ∀b.f (b → a) → (FreeA f b → FreeA f a)
is a natural transformation of contravariant functors. The two contravariant functors here could be defined, in Haskell, using a newtype:
newtype F1 f a x = F1 (f (x → a))
newtype F2 f a x = F2 (FreeA f x → FreeA f a)
instan e Functor f ⇒ Contravariant (F1 f a) where
contramap h (F1 g) = F1 $ fmap ( ◦ h) g
instan e Functor f ⇒ Contravariant (F2 f a) where
contramap h (F2 g) = F2 $ g ◦ fmap h
The action of F1 and F2 on morphisms is defined in the obvious way. Note that here we make use of the
fact that FreeA f is a functor.
Naturality of ( :$: ) means that, given types x and y, and a function h : x → y, the following holds:
∀g :: f (y → a), u :: FreeA f x.
fmap ( ◦ h) g :$: u ≡ g :$: fmap h u
(1)
where we have unfolded the definitions of contramap for F1 and F2, and removed the newtypes.
Note that the results in [8] do not actually imply naturality of ( :$: ) at this generality, since f is a type
variable with an arbitrary Functor instance, not a concrete positive type expression together with its
canonical instance. However, in the interpretation given in sections 8 and 9, FreeA will be defined in
such a way that equation 1 holds automatically.
Free Applicative Functors
12
5
Isomorphism of the two definitions
In this section we show that the two definitions of free applicatives given in section 2 are isomorphic.
First of all, if f is a functor, FreeAL f is also a functor:
instan e Functor f ⇒ Functor (FreeAL f) where
fmap g (PureL x) = PureL (g x)
fmap g (h :*: x) = (fmap (g ◦ ) h) :*: x
Again, the functor laws can be verified by a simple structural induction.
For the ( :*: ) constructor, a free theorem can be derived in a completely analogous way to deriving
equation 1. This equation states that ( :*: ) is a natural transformation:
∀h :: x → y, g :: FreeAL f (y → a), u :: f x.
fmap ( ◦ h) g :*: u ≡ g :*: fmap h u
(2)
We define functions to convert between the two definitions:
r2l :: Functor f ⇒ FreeA f a → FreeAL f a
r2l (Pure x) = PureL x
r2l (h :$: x) = fmap (flip ( $ )) (r2l x) :*: h
l2r :: Functor f ⇒ FreeAL f a → FreeA f a
l2r (PureL x) = Pure x
l2r (h :*: x) = fmap (flip ( $ )) x :$: l2r h
We will also need the fact that l2r is a natural transformation:
∀h :: x → y, u :: FreeAL f x.
l2r (fmap h u) ≡ fmap h (l2r u)
(3)
Proposition 1. r2l is an isomorphism, the inverse of which is l2r.
Proof. First we prove that ∀u :: FreeA f a.l2r (r2l u) ≡ u. We compute using equational reasoning
with induction on u:
l2r (r2l (Pure x))
≡ h definition of r2l i
l2r (PureL x)
≡ h definition of l2r i
Pure x
l2r (r2l (h :$: x))
≡ h definition of r2l i
l2r (fmap (flip ( $ )) (r2l x) :*: h)
≡ h definition of l2r i
fmap (flip ( $ )) h :$:
l2r (fmap (flip ( $ )) (r2l x))
P. Capriotti & A. Kaposi
13
≡ h equation 3 i
fmap (flip ( $ )) h :$:
fmap (flip ( $ )) (l2r (r2l x))
≡ h inductive hypothesis i
fmap (flip ( $ )) h :$: fmap (flip ( $ )) x
≡ h equation 1 i
fmap ( ◦ (flip ( $ ))) (fmap (flip ( $ )) h) :$: x
≡ h f is a functor i
fmap (( ◦ (flip ( $ ))) ◦ flip ( $ )) h :$: x
≡ h definition of flip and ( $ ) i
fmap id h :$: x
≡ h f is a functor i
h :$: x
Next, we prove that ∀u :: FreeAL f a.r2l (l2r u) ≡ u. Again, we compute using equational reasoning
with induction on u:
r2l (l2r (PureL x))
≡ h definition of l2r i
r2l (Pure x)
≡ h definition of r2l i
PureL x
r2l (l2r (h :*: x))
≡ h definition of l2r i
r2l (fmap (flip ( $ )) x :$: l2r h)
≡ h definition of r2l i
fmap (flip ( $ )) (r2l (l2r h)) :*: fmap (flip ( $ )) x
≡ h inductive hypothesis i
fmap (flip ( $ )) h :*: fmap (flip ( $ )) x
≡ h equation 2 i
fmap ( ◦ (flip ( $ ))) (fmap (flip ( $ )) h) :*: x
≡ h FreeAL f is a functor i
fmap (( ◦ (flip ( $ ))) ◦ flip ( $ )) h :*: x
≡ h definition of flip and ( $ ) i
fmap id h :*: x
≡ h FreeAL f is a functor i
h :*: x
In the next sections, we will prove that FreeA is a free applicative functor. Because of the isomorphism
of the two definitions, these results will carry over to FreeAL.
Free Applicative Functors
14
6
Applicative laws
Following [6], the laws for an Applicative instance are:
pure id <*> u ≡ u
(4)
pure ( ◦ ) <*> u <*> v <*> x ≡ u <*> (v <*> x)
(5)
pure f <*> pure x ≡ pure (f x)
u <*> pure x ≡ pure ( $ x) <*> u
We introduce a few abbreviations to help make the notation lighter:
uc = uncurry
pair x y = ( , ) <$> x <*> y
Lemma 1. For all
u :: y → z
v :: FreeA f (x → y)
x :: FreeA f x
the following equation holds:
fmap u (v <*> x) ≡ fmap (u ◦ ) v <*> x
Proof. We compute:
fmap u (Pure v <*> x)
≡ h definition of ( <*> ) i
fmap u (fmap v x)
≡ h FreeA f is a functor i
fmap (u ◦ v) x
≡ h definition of ( <*> ) i
Pure (u ◦ v) <*> x
≡ h definition of fmap i
fmap (u ◦ ) (Pure v) <*> x
fmap u ((g :$: y) <*> x)
≡ h definition of ( <*> ) i
fmap u (fmap uc g :$: pair y x)
≡ h definition of fmap i
fmap (u ◦ ) (fmap uc g) :$: pair y x
≡ h f is a functor i
fmap (λ g → u ◦ uc g) g :$: pair y x
≡ h f is a functor i
fmap uc (fmap ((u ◦ ) ◦ ) g) :$: pair y x
≡ h definition of ( <*> ) i
(6)
(7)
P. Capriotti & A. Kaposi
15
(fmap ((u ◦ ) ◦ ) g :$: y) <*> x
≡ h definition of fmap i
fmap (u ◦ ) (g :$: y) <*> x
Lemma 2. Property 5 holds for FreeA f, i.e. for all
u :: FreeA f (y → z)
v :: FreeA f (x → y)
x :: FreeA f x,
pure ( ◦ ) <*> u <*> v <*> x ≡ u <*> (v <*> x)
Proof. Suppose first that u = Pure u0 for some u0 :: y → z:
Pure ( ◦ ) <*> Pure u0 <*> v <*> x
≡ h definition of ( <*> ) i
Pure (u0 ◦ ) <*> v <*> x
≡ h definition of ( <*> ) i
fmap (u0 ◦ ) v <*> x
≡ h lemma 1 i
fmap u0 (v <*> x)
≡ h definition of ( <*> ) i
Pure u0 <*> (v <*> x)
To tackle the case where u = g :$: w, for
g :: f (w → y → z)
w :: FreeA f w,
we need to define a helper function
t :: ((w , x → y) , x) → (w , y)
t ((w , v) , x) = (w , v x)
and compute:
pure ( ◦ ) <*> (g :$: w) <*> v <*> x
≡ h definition of pure and ( <*> ) i
(fmap (( ◦ ) ◦ ) g :$: w) <*> v <*> x
≡ h definition of composition i
(fmap (λ g w v → g w ◦ v) g :$: w) <*> v <*> x
≡ h definition of ( <*> ) i
(fmap uc (fmap (λ g w v → g w ◦ v) g) :$: pair w v)
<*> x
≡ h f is a functor and definition of uc i
(fmap (λ g (w , v) → g w ◦ v) g :$: pair w v) <*> x
Free Applicative Functors
16
≡ h definition of ( <*> ) i
fmap uc (fmap (λ g (w , v) → g w ◦ v) g) :$:
pair (pair w v) x
≡ h f is a functor and definition of uc i
fmap (λ g ((w , v) , x) → g w (v x)) g :$:
pair (pair w v) x
≡ h definition of uc and t i
fmap (λ g → uc g ◦ t) g :$: pair (pair w v) x
≡ h f is a functor i
fmap ( ◦ t) (fmap uc g) :$: pair (pair w v) x
≡ h equation 1 i
fmap uc g :$: fmap t (pair (pair w v) x)
≡ h lemma 1 (3 times) and FreeA f is a functor (3 times) i
fmap uc g :$: (pure ( ◦ ) <*> fmap ( , ) w <*> v <*> x)
≡ h induction hypothesis for fmap ( , ) w i
fmap uc g :$: (fmap ( , ) w <*> (v <*> x))
≡ h definition of ( <*> ) i
(g :$: w) <*> (v <*> x)
Lemma 3. Property 7 holds for FreeA f, i.e. for all
u :: FreeA f (x → y)
x :: x,
u <*> pure x ≡ pure ( $ x) <*> u
Proof. If u is of the form Pure u0 , then the conclusion follows immediately.
Let’s assume, therefore, that u = g :$: w, for some w :: w, g :: f (w → x → y), and that the lemma is true
for structurally smaller values of u:
(g :$: w) <*> pure x
≡ h definition of ( <*> ) i
fmap uc g :$: pair w (pure x)
≡ h definition of pair i
fmap uc g :$: (fmap ( , ) w <*> pure x)
≡ h induction hypothesis for fmap ( , ) w i
fmap uc g :$: (pure ( $ x) <*> fmap ( , ) w)
≡ h FreeA f is a functor i
fmap uc g :$: fmap (λ w → (w , x)) w)
≡ h equation 1 i
fmap (λ g w → g (w , x)) (fmap uc g) :$: w
≡ h f is a functor i
fmap (λ g w → g w x) g :$: w
≡ h definition of fmap for FreeA f i
fmap ( $ x) (g :$: w)
P. Capriotti & A. Kaposi
17
≡ h definition of ( <*> ) i
pure ( $ x) <*> (g :$: w)
Proposition 2. FreeA f is an applicative functor.
Proof. Properties 4 and 6 are straightforward to verify using the fact that FreeA f is a functor, while
properties 5 and 7 follow from lemmas 2 and 3 respectively.
7
FreeA as a Left adjoint
We are now going to make the statement that FreeA f is the free applicative functor on f precise.
First of all, we will define a category A of applicative functors, and show that FreeA is a functor
FreeA : F → A ,
where F is the category of endofunctors of Hask.
Saying that FreeA f is the free applicative on f, then, amounts to saying that FreeA is left adjoint to the
forgetful functor A → F .
Definition 1. Let f and g be two applicative functors. An applicative natural transformation between f
and g is a polymorphic function
t :: ∀a.f a → g a
satisfying the following laws:
t (pure x) ≡ pure x
(8)
t (h <*> x) ≡ t h <*> t x.
(9)
We define the type of all applicative natural transformations between f and g, we write, in Haskell,
type AppNat f g = ∀a.f a → g a
where the laws are implied.
Similarly, for any pair of functors f and g, we define
type Nat f g = ∀a.f a → g a
for the type of natural transformations between f and g.
Note that, by parametricity, polymorphic functions are automatically natural transformations in the categorical sense, i.e, for all
t :: Nat f g
h :: a → b
x :: f a,
t (fmap h x) ≡ fmap h (t x).
It is clear that applicative functors, together with applicative natural transformations, form a category,
which we denote by A , and similarly, functors and natural transformations form a category F .
Free Applicative Functors
18
Proposition 3. FreeA defines a functor F → A .
Proof. We already showed that FreeA sends objects (functors in our case) to applicative functors.
We need to define the action of FreeA on morphisms (which are natural transformations in our case):
liftT :: (Functor f , Functor g)
⇒ Nat f g
→ AppNat (FreeA f) (FreeA g)
liftT (Pure x) = Pure x
liftT k (h :$: x) = k h :$: liftT k x
First we verify that liftT k is an applicative natural transformation i.e. it satisfies laws 8 and 9. We use
equational reasoning for proving law 8:
liftT k (pure x)
≡ h definition of pure i
liftT k (Pure x)
≡ h definition of liftT i
Pure x
≡ h definition of pure i
pure x
For law 9 we use induction on the size of the first argument of ( <*> ) as explained in section 8. The
base cases:
liftT k (Pure h <*> Pure x)
≡ h definition of ( <*> ) i
liftT k (fmap h (Pure x))
≡ h definition of fmap i
liftT k (Pure (h x))
≡ h definition of liftT i
Pure (h x)
≡ h definition of fmap i
fmap h (Pure x)
≡ h definition of ( <*> ) i
Pure h <*> Pure x
≡ h definition of liftT i
liftT k (Pure h) <*> liftT k (Pure x)
liftT k (Pure h <*> (i :$: x))
≡ h definition of ( <*> ) i
liftT k (fmap h (i :$: x))
≡ h definition of fmap i
liftT k (fmap (h ◦ ) i :$: x)
≡ h definition of liftT i
k (fmap (h ◦ ) i) :$: liftT k x
≡ h k is natural i
P. Capriotti & A. Kaposi
19
fmap (h ◦ ) (k i) :$: liftT k x
≡ h definition of fmap i
fmap h (k i :$: liftT k x)
≡ h definition of ( <*> ) i
Pure h <*> (k i :$: liftT k x)
≡ h definition of liftT i
liftT k (Pure h) <*> liftT k (i :$: x)
The inductive case:
liftT k ((h :$: x) <*> y)
≡ h definition of ( <*> ) i
liftT k (fmap uncurry h :$: (fmap ( , ) x <*> y)
≡ h definition of liftT i
k (fmap uncurry h) :$: liftT k (fmap ( , ) x <*> y)
≡ h inductive hypothesis i
k (fmap uncurry h) :$:
(liftT k (fmap ( , ) x) <*> liftT k y)
≡ h liftT k is natural i
k (fmap uncurry h) :$:
(fmap ( , ) (liftT k x) <*> liftT k y)
≡ h k is natural i
fmap uncurry (k h) :$:
(fmap ( , ) (liftT k x) <*> liftT k y)
≡ h definition of ( <*> ) i
(k h :$: liftT k x) <*> liftT k y
≡ h definition of liftT i
liftT k (h :$: x) <*> liftT k y
Now we need to verify that liftT satisfies the functor laws
liftT id ≡ id
liftT (t ◦ u) ≡ liftT t ◦ liftT u.
The proof is a straightforward structural induction.
We are going to need the following natural transformation (which will be the unit of the adjunction 11):
one :: Functor f ⇒ Nat f (FreeA f)
one x = fmap const x :$: Pure ()
which embeds any functor f into FreeA f (we used a specialization of this function for Option in section
1.2).
Lemma 4.
g :$: x ≡ one g <*> x
Free Applicative Functors
20
Proof. Given
h :: a → (() , a)
h x = (() , x)
it is easy to verify that:
( ◦ h) ◦ uncurry ◦ const ≡ id,
(10)
so
one g <*> x
≡ h definition of one i
(fmap const g :$: Pure ()) <*> x
≡ h definition of ( <*> ) and functor law for f i
fmap (uncurry ◦ const) g :$: fmap h x
≡ h equation 1 and functor law for f i
fmap (( ◦ h) ◦ uncurry ◦ const) g :$: x
≡ h equation 10 i
g :$: x
Proposition 4. The FreeA functor is left adjoint to the forgetful functor A → F . Graphically:
lower
−→
HomF (FreeA f, g) ∼
= HomA (f, g)
←−
(11)
raise
Proof. Given a functor f and an applicative functor g, we define a natural bijection between Nat f g and
AppNat (FreeA f) g as such:
raise :: (Functor f , Applicative g)
⇒ Nat f g
→ AppNat (FreeA f) g
raise (Pure x) = pure x
raise k (g :$: x) = k g <*> raise k x
lower :: (Functor f , Applicative g)
⇒ AppNat (FreeA f) g
→ Nat f g
lower k = k ◦ one
A routine verification shows that raise and lower are natural in f and g. The proof that raise k
satisfies the applicative natural transformation laws 8 and 9 is a straightforward induction having the
same structure as the proof that liftT k satisfies these laws (proposition 3). To show that f and g are
inverses of each other, we reason by induction and calculate in one direction:
raise (lower t) (Pure x)
≡ h definition of raise i
pure x
P. Capriotti & A. Kaposi
21
≡ h t is an applicative natural transformation i
t (pure x)
≡ h definition of pure i
t (Pure x)
raise (lower t) (g :$: x)
≡ h definition of raise i
lower t g <*> raise (lower t) x
≡ h induction hypothesis i
lower t g <*> t x
≡ h definition of lower i
t (one g) <*> t x
≡ h t is an applicative natural transformation i
t (one g <*> x)
≡ h lemma 4 i
t (g :$: x)
The other direction:
lower (raise t) x
≡ h definition of lower i
raise t (one x)
≡ h definition of one i
raise t (fmap const x :$: Pure ())
≡ h definition of raise i
t (fmap const x) <*> pure ()
≡ h t is natural i
fmap const (t x) <*> pure ()
≡ h fmap h ≡ ((pure h) <*> ) in an applicative functor i
pure const <*> t x <*> pure ()
≡ h t is natural i
pure ( $ ()) <*> (pure const <*> t x)
≡ h applicative law 5 i
pure ( ◦ ) <*> pure ( $ ()) <*> pure const <*> t x
≡ h applicative law 6 applied twice i
pure id <*> t x
≡ h applicative law 4 i
tx
7.1 Example: option parsers (continued)
With the help of the adjunction defined above by raise and lower we are able to define some useful
functions. In the case of command-line option parsers, for example, it can be used for computing the
global default value of a parser:
Free Applicative Functors
22
parserDefault :: FreeA Option a → Maybe a
parserDefault = raise optDefault
or for extracting the list of all the options in a parser:
allOptions :: FreeA Option a → [String]
allOptions = getConst ◦ raise f
where
f opt = Const [optName opt]
allOptions works by first defining a function that takes an option and returns a one-element list with
the name of the option, and then lifting it to the Const applicative functor.
The raise function can be thought of as a way to define a “semantics” for the whole syntax of the
DSL corresponding to FreeA f, given one for just the individual atomic actions, expressed as a natural
transformation from the functor f to any applicative functor g.
When defining such a semantics using raise, the resulting function is automatically an applicative
natural transformation. In some circumstances, however, it is more convenient to define a function by
pattern matching directly on the constructors of FreeA f, like when the target does not have an obvious
applicative functor structure that makes the desired function an applicative natural transformation.
For example, we can write a function that runs an applicative option parser over a list of command-line
arguments, accepting them in any order:
matchOpt :: String → String
→ FreeA Option a
→ Maybe (FreeA Option a)
matchOpt
(Pure ) = Nothing
matchOpt opt value (g :$: x)
| opt ≡ ’-’ : ’-’ : optName g
= fmap ( <$> x) (optReader g value)
| otherwise
= fmap (g :$: ) (matchOpt opt value x)
The matchOpt function looks for options in the parser which match the given command-line argument,
and, if successful, returns a modified parser where the option has been replaced by a pure value.
Clearly, matchOpt opt value is not applicative, since, for instance, equation 8 is not satisfied.
runParser :: FreeA Option a
→ [String]
→ Maybe a
runParser p (opt : value : args) =
ase matchOpt opt value p of
Nothing → Nothing
Just p′ → runParser p′ args
runParser p [ ] = parserDefault p
runParser
= Nothing
Finally, runParser calls matchOpt with successive pairs of arguments, until no arguments remain, at
which point it uses the default values of the remaining options to construct a result.
P. Capriotti & A. Kaposi
8
23
Totality
All the proofs in this paper apply to a total fragment of Haskell, and completely ignore the presence of
bottom. The Haskell subset we use can be given a semantics in any locally presentable cartesian closed
category.
In fact, if we assume that all the functors used throughout the paper are accessible, all our inductive
definitions can be regarded as initial algebras of accessible functors.
For example, to realise FreeA f, assume f is κ -accessible for some regular cardinal κ . Then define a
functor:
A : Funcκ (C , C ) → Funcκ (C , C ),
where Funcκ is the category of κ -accessible endofunctors of C , which is itself locally presentable by
proposition 5.
The inductive definition of FreeA f above can then be regarded as the initial algebra of A, given by:
(AG)a = a +
Z b:C
F[b, a] × Gb,
(12)
where [−, −] denotes the internal hom (exponential) in C . Since F and G are locally presentable, and
C is cocomplete, the coend exists by lemma 8, and AG is κ -accessible by lemma 7, provided κ is large
enough.
Furthermore, the functor A itself is accessible by proposition 6, hence it has an initial algebra. Equation
1 is then a trivial consequence of this definition.
As for function definitions, most use primitive recursion, so they can be realised by using the universal
property of the initial algebra directly.
One exception is the definition of ( <*> ):
(h :$: x) <*> y = fmap uncurry h :$: (( , ) <$> x <*> y)
which contains a recursive call where the first argument, namely ( , ) <$> x, is not structurally smaller
than the original one (h :$: x).
To prove that this function is nevertheless well defined, we introduce a notion of size for values of type
FreeA f a:
size :: FreeA f a → N
size (Pure ) = 0
size ( :$: x) = 1 + size x
To conclude that the definition of ( <*> ) can be made sense of in our target category, we just need to
show that the size of the argument in the recursive call is smaller than the size of the original argument,
which is an immediate consequence of the following lemma.
Lemma 5. For any function f :: a → b and u :: FreeA f a,
size (fmap f u) ≡ size u
Proof. By induction:
size (fmap f (Pure x))
≡ h definition of fmap i
Free Applicative Functors
24
size (Pure (f x))
≡ h definition of size i
0
≡ h definition of size i
size (Pure x)
size (fmap f (g :$: x))
≡ h definition of fmap i
size (fmap (f ◦ ) g :$: x)
≡ h definition of size i
1 + size x
≡ h definition of size i
size (g :$: x)
In most of our proofs using induction we carry out induction on the size of the first argument of ( <*> )
where size is defined by the above size function.
9
Semantics
In this section, we establish the results about accessible functors of locally presentable categories that we
used in section 8 to justify the inductive definition of FreeA f.
We begin with a technical lemma:
Lemma 6. Suppose we have the following diagram of categories and functors:
A
i
F
B
K
⑥
⑥
⑥L
// C
⑥>>
D
where B, and C are locally presentable, F is accessible, and i is the inclusion of a dense small full
subcategory of compact objects of B. Then the pointwise left Kan extension L of F along K exists and is
equal to the left Kan extension of Fi along Ki.
If, furthermore, D is locally presentable and K is accessible, then L is accessible.
Proof. Let κ be a regular cardinal such that B and C are κ -locally presentable.
The pointwise left Kan extension L can be obtained as a colimit:
Ld = colim Fb,
b:B
g:Kb→d
(13)
Where the indices range over the comma category (K ↓ d). To show that L exists, it is therefore enough
to prove that the colimit 13 can be realised as the small colimit:
colim Fa.
a:A
f :Ka→d
P. Capriotti & A. Kaposi
25
For any b : B, and g : Kb → d, we can express b as a canonical κ -filtered colimit of compact objects:
b∼
= colim a.
a:A
h:a→b
Since F preserves κ -filtered colimits, we then get a morphism:
Fb → colim Fa → colim Fa.
a:A
h:a→b
a:A
f :Ka→d
This gives a cocone for the colimit 13, and a straightforward verification shows that it is universal.
As for the second statement, suppose D is also κ -locally presentable. By possibly increasing κ , we can
assume that Ka is κ -compact for all a : A (such a κ exists because A is small, and every object of D is
λ -compact for some λ ).
Then, by the first part:
Ld =
Z a:A
Fa · D(Ka, d).
Now, a filtered colimit in d commutes with D(Ka, −) because Ka is compact, it commutes with Fa · −
because copowers are left adjoints, and it commutes with coends because they are both colimits.
Therefore, L is accessible.
From now on, let B and C be categories with finite products, and F, G : B → C be functors.
Definition 2. The Day convolution of F and G, denoted F ∗ G, is the pointwise left Kan extension of the
diagonal functor in the following diagram:
F×G //
C ×C
❑❑❑
❑❑❑
×
❑❑❑ ×
❑%%
B ❴ ❴ ❴ ❴ ❴// C
B × B❑
F∗G
Note that the Day convolution of two functors might not exist, but it certainly does if B is small and C
is cocomplete.
Lemma 7. Suppose that B and C are locally presentable and F and G are accessible. Then the Day
convolution of F and G exists and is accessible.
Proof. Immediate consequence of lemma 6.
Lemma 8. Suppose that B is cartesian closed. Then the Day convolution of F and G can be obtained
as the coend:
(F ∗ G)b =
Z y:B
F[y, b] × Gy
Free Applicative Functors
26
Proof. By coend calculus:
(F ∗ G)b
=
Z xy:B
=
Z y:B
Fx × Gy · C (x × y, b)
Z y:B Z x:B
=
Fx · C (x × y, b) × Gy
Z y:B Z x:B
=
Fx · C (x, [y, b]) × Gy
F[y, b] × Gy
Proposition 5. Let κ be a regular cardinal, and B and C be locally κ -presentable. Then the category
Funcκ (B, C ) of κ -accessible functors is locally κ -presentable.
Proof. Let A be a dense small full subcategory of B. The obvious functor Funcκ (B, C ) → Func(A , C )
is an equivalence of categories (its inverse is given by left Kan extensions along the inclusion A → B),
and Func(A , C ) is locally κ -presentable (see for example [2], corollary 1.54).
Proposition 6. Let κ be a regular cardinal such that B and C are locally κ -presentable, and the Day
convolution of any two κ -accessible functors is κ -accessible (which exists by lemma 7).
Then the Day convolution operator
Funcκ (B, C ) × Funcκ (B, C ) → Funcκ (B, C )
is itself a κ -accessible functor.
Proof. It is enough to show that ∗ preserves filtered colimits pointwise in its two variables separately.
But this is clear, since filtered colimits commute with finite products, copowers and coends.
We can recast equation 12 in terms of Day convolution as follows:
AG = Id + F ∗ G.
(14)
Equation 14 makes precise the intuition that free applicative functors are in some sense lists (i.e. free
monoids). In fact, the functor A is exactly the one appearing in the usual recursive definition of lists, only
in this case the construction is happening in the monoidal category of accessible endofunctors equipped
with Day convolution.
We also sketch the following purely categorical construction of free applicative (i.e. lax monoidal)
functors, which is not essential for the rest of the paper, but is quite an easy consequence of the machinery
developed in this section.
The idea is to perform the “list” construction in one step, instead of iterating individual Day convolutions
using recursion. Namely, for any category C , let C ∗ be the free monoidal category generated by C . The
objects (resp. morphisms) of C ∗ are lists of objects (resp. morphisms) of C . Clearly, C ∗ is accessible if
C is.
If C has finite products, there is a functor
ε :C∗ → C
P. Capriotti & A. Kaposi
27
which maps a list to its corresponding product. Note that ε is accessible. Furthermore, the assigment
C 7→ C ∗ extends to a 2-functor on Cat which preserves accessibility of functors.
Now, the free applicative G on a functor F : C → C is simply defined to be the Kan extension of ε ◦ F ∗
along ε :
F∗
// C ∗
❈❈
❈❈
ε
❈❈ ε
❈!!
C ❴ ❴ ❴// C
C ∗❈
G
The functor G is accessible by an appropriate generalisation of lemma 6, and it is not hard to see that it
is lax monoidal (see for example [7], proposition 4). We omit the proof that G is a free object, which can
be obtained by diagram chasing using the universal property of Kan extensions.
10 Related work
The idea of free applicative functors is not entirely new. There have been a number of different definitions of free applicative functor over a given Haskell functor, but none of them includes a proof of the
applicative laws.
The first author of this paper published a specific instance of applicative functors3 similar to our example
shown in section 1.2. The example has later been expanded into a fully-featured Haskell library for
command line option parsing.4
Tom Ellis proposes a definition very similar to ours,5 but uses a separate inductive type for the case
corresponding to our ( :$: ) constructor. He then observes that law 6 probably holds because of the
existential quantification, but does not provide a proof. We solve this problem by deriving the necessary
equation 1 as a “free theorem”.
Gergő Érdi gives another similar definition6 , but his version presents some redundancies, and thus fails
to obey the applicative laws. For example, Pure id <*> x can easily be distinguished from x using a
function like our count above, defined by pattern matching on the constructors.
However, this is remedied by only exposing a limited interface which includes the equivalent of our
raise function, but not the Pure and Free constructors. It is probably impossible to observe a violation
of the laws using the reduced interface, but that also means that definitions by pattern matching, like the
one for our matchOpt in section 7.1, are prohibited.
The free package on hackage7 contains a definition essentially identical to our FreeAL, differing only
in the order of arguments.
Another approach, which differs significantly from the one presented in the paper, underlies the definition contained in the free-functors package on hackage,8 and uses a Church-like encoding (and
the ConstraintKinds GHC extension) to generalise the construction of a free Applicative to any
superclass of Functor.
3 http://paolocapriotti.com/blog/2012/04/27/applicative-option-parser
4 http://hackage.haskell.org/package/optparse-applicative
5 http://web.jaguarpaw.co.uk/
~tom/blog/posts/2012-09-09-towards-free-applicatives.html
6 http://gergo.erdi.hu/blog/2012-12-01-static_analysis_with_applicatives/
7 http://hackage.haskell.org/package/free
8 http://hackage.haskell.org/package/free-functors
Free Applicative Functors
28
The idea is to use the fact that, if a functor T has a left adjoint F, then the monad T ◦ F is the codensity
monad of T (i.e. the right Kan extension of T along itself). By taking T to be the forgetful functor
A → F , one can obtain a formula for F using the expression of a right Kan extension as an end.
One problem with this approach is that the applicative laws, which make up the definition of the category
A , are left implicit in the universal quantification used to represent the end.
In fact, specializing the code in Data.Functor.HFree to the Applicative constraint, we get:
data FreeA′ f a = FreeA′ {
runFreeA :: ∀g.Applicative g
⇒ (∀x.f x → g x) → g a}
instan e Functor f ⇒ Functor (FreeA′ f) where
fmap h (FreeA′ t) = FreeA′ (fmap h ◦ t)
instan e Functor f ⇒ Applicative (FreeA′ f) where
pure x = FreeA′ (λ → pure x)
FreeA′ t1 <*> FreeA′ t2 =
FreeA′ (λ u → t1 u <*> t2 u)
Now, for law 4 to hold, for example, we need to prove that the term λ u → pure id <*> t u is equal
to t. This is strictly speaking false, as those terms can be distinguished by taking any functor with an
Applicative instance that does not satisfy law 4, and as t a constant function returning a counterexample for it.
Intuitively, however, the laws should hold provided we never make use of invalid Applicative instances. To make this intuition precise, one would probably need to extend the language with quantification over equations, and prove a parametricity result for this extension.
Another problem of the Church encoding is that, like Érdi’s solution above, it presents a more limited
interface, and thus it is harder to use. In fact, the destructor runFreeA is essentially equivalent to our
raise function, which can only be used to define applicative natural transformation. Again, a function
like matchOpt, which is not applicative, could not be defined over FreeA′ in a direct way.
11 Discussion and further work
We have presented a practical definition of free applicative functor over any Haskell functor, proved its
properties, and showed some of its applications. As the examples in this paper show, free applicative
functors solve certain problems very effectively, but their applicability is somewhat limited.
For example, applicative parsers usually need an Alternative instance as well, and the free applicative
construction does not provide that. One possible direction for future work is trying to address this issue
by modifying the construction to yield a free Alternative functor, instead.
Unfortunately, there is no satisfactory set of laws for alternative functors: if we simply define an alternative functor as a monoid object in A , then many commonly used instances become invalid, like the
one for Maybe. Using rig categories and their lax functors to formalise alternative functors seems to be a
workable strategy, and we are currently exploring it.
Another direction is formalizing the proofs in this paper in a proof assistant, by embedding the total
subset of Haskell under consideration into a type theory with dependent types.
Our attempts to replicate the proofs in Agda have failed, so far, because of subtle issues in the interplay
between parametricity and the encoding of existentials with dependent sums.
P. Capriotti & A. Kaposi
29
In particular, equation 1 is inconsistent with a representation of the existential as a Σ type in the definition
of FreeA. For example, terms like const () :$: Pure 3 and id :$: Pure () are equal by equation 1,
but can obviously be distinguished using large elimination.
This is not too surprising, as we repeatedly made use of size restrictions in sections 8 and 9, and those
will definitely need to be somehow replicated in a predicative type theory like the one implemented by
Agda.
A reasonable compromise is to develop the construction only for containers [1], for which one can prove
that the free applicative on the functor S ⊲ P is given, using the notation at the end of section 9, by
S∗ ⊲ (ε ◦ P∗ ), where S is regarded as a discrete category.
Another possible further development of the results in this paper is trying to generalise the construction
of a free applicative functor to functors of any monoidal category. In section 9 we focused on categories
with finite products, but it is clear that monoidal categories are the most natural setting, as evidenced by
the appearance of the corresponding 2-comonad on Cat.
Furthermore, an applicative functor is defined in [6] as a lax monoidal functor with a strength, but we
completely ignore strengths in this paper. This could be remedied by working in the more general setting
of V -categories and V -functors, for some monoidal category V .
12 Acknowledgements
We would like to thank Jennifer Hackett, Thorsten Altenkirch, Venanzio Capretta, Graham Hutton, Edsko
de Vries and Christian Sattler, for helpful suggestions and insightful discussions on the topics presented
in this paper.
References
[1] Michael Abott, Thorsten Altenkirch & Neil Ghani (2005): Containers - Constructing Strictly Positive Types.
Theoretical Computer Science 342, pp. 3–27, doi:10.1016/j.tcs.2005.06.002. Applied Semantics: Selected Topics.
[2] J. Adámek & J. Rosický (1994): Locally Presentable and Accessible Categories. Cambridge University
Press, doi:10.1017/CBO9780511600579.
[3] Brian J. Day (1970): Construction of Biclosed Categories. Ph.D. thesis, University of New South Wales.
[4] G. M. Kelly (1980): A unified treatment of transfinite constructions for free algebras, free monoids, colimits,
associated sheaves, and so on. Bulletin of the Australian Mathematical Society 22, pp. 1–83, doi:10.1017/
S0004972700006353.
[5] Simon Marlow (2010): Haskell 2010 Language Report.
[6] Conor McBride & Ross Paterson (2008): Applicative programming with effects. Journal of Functional Programming 18(1), pp. 1–13, doi:10.1017/S0956796807006326.
[7] Ross Paterson (2012): Constructing Applicative Functors. In: Mathematics of Program Construction, Lecture
Notes in Computer Science 7342, Springer-Verlag, pp. 300–323, doi:10.1007/978-3-642-31113-0_15.
[8] John C. Reynolds (1983): Types, Abstraction and Parametric Polymorphism. In: IFIP Congress, pp. 513–523.
[9] S. Doaitse Swierstra & Luc Duponcheel (1996): Deterministic, Error-Correcting Combinator Parsers. In:
Advanced Functional Programming, Lecture Notes in Computer Science 1129, Springer-Verlag, pp. 184–
207, doi:10.1007/3-540-61628-4_7.
[10] Wouter Swierstra (2008): Data types à la carte. Journal of Functional Programming 18(4), pp. 423–436,
doi:10.1017/S0956796808006758.
30
Free Applicative Functors
[11] Philip Wadler (1989): Theorems for free! In: Functional Programming Languages and Computer Architecture, ACM Press, pp. 347–359, doi:10.1145/99370.99404.
| 6cs.PL
|
arXiv:1709.04673v2 [cs.SY] 10 Dec 2017
Conditions for Stability and Convergence of
Set-Valued Stochastic Approximations:
Applications to Approximate Value and Fixed
point Iterations
Arunselvan Ramaswamy
∗
Dept. of Electrical Engineering and Information Technology
Paderborn University
Paderborn - 33908, Germany.
Shalabh Bhatnagar
†
Dept. of Computer Science and Automation
Indian Institute of Science
Bengaluru - 560012, India.
December 12, 2017
Abstract
The main aim of this paper is the development of easily verifiable sufficient conditions for stability (almost sure boundedness) and convergence
of stochastic approximation algorithms (SAAs) with set-valued meanfields, a class of model-free algorithms that have become important in
recent times. In this paper we provide a complete analysis of such algorithms under three different, yet related sets of sufficient conditions, based
on the existence of an associated global/local Lyapunov function. Unlike
previous Lyapunov function based approaches, we provide a simple recipe
for explicitly constructing the Lyapunov function, needed for analysis.
Our work builds on the works of Abounadi, Bertsekas and Borkar (2002),
Munos (2005), and Ramaswamy and Bhatnagar (2016). An important
motivation for the flavor of our assumptions comes from the need to understand dynamic programming and reinforcement learning algorithms,
that use deep neural networks (DNNs) for function approximations and
parameterizations. These algorithms are popularly known as deep learning algorithms. As an important application of our theory, we provide a
complete analysis of the stochastic approximation counterpart of approximate value iteration (AVI), an important dynamic programming method
∗ email:[email protected]
† email:[email protected]
1
designed to tackle Bellman’s curse of dimensionality. Further, the assumptions involved are significantly weaker, easily verifiable and truly
model-free. The theory presented in this paper is also used to develop
and analyze the first SAA for finding fixed points of contractive set-valued
maps.
1
Introduction
Stochastic approximation algorithms (SAAs) are an important class of iterative schemes that are used to solve problems arising in stochastic optimization,
stochastic control, machine learning and financial mathematics, among others.
SAAs constitute a powerful tool, due to their model free approach, to solving problems. The first stochastic approximation algorithm was developed by
Robbins and Monro [16] in 1951 to solve the root finding problem. Important contributions to modern stochastic approximations theory were made by
Benaı̈m (1996) [3], Benaı̈m and Hirsch (1996) [4], Borkar (1997) [8], Borkar and
Meyn (1999) [10] and Benaı̈m, Hofbauer and Sorin (2005) [5], to name a few.
An important aspect of the analysis of SAAs lies in verifying the almost sure
boundedness (stability) of the iterates. This can be hard in many applications.
In this paper we present easily verifiable sufficient conditions for both stability
and convergence of SAAs with set-valued mean-fields. Specifically, we consider
the following iterative scheme:
xn+1 = xn + a(n) (yn + Mn+1 ) ,
(1)
where xn ∈ Rd for all n; {a(n)}n≥0 is a step-size sequence; yn ∈ H(xn ) for all
n such that H : Rd → {subsets of Rd } is a Marchaud map; and {Mn+1 }n≥0
is the noise sequence. We present three different yet overlapping sets of (easily
verifiable) conditions for stability (almost sure boundedness) and convergence (to
a closed connected internally chain transitive invariant set of ẋ(t) ∈ H(x(t)))
of (1). The reader is referred to Section 4 for the analysis.
The problem of stability for SAAs with set-valued mean-fields has been previously studied by Ramaswamy and Bhatnagar [15]. They developed the first
set of sufficient conditions for the stability and convergence of (1) by extending
the ideas of Borkar and Meyn [10]. Their sufficient conditions are based on
the “limiting properties” of an associated scaled differential inclusion. On the
contrary, the conditions presented in this paper are based on “local properties”
of the associated differential inclusion. We believe that the stability criterion
presented here is applicable to scenarios that are in some sense “orthogonal”
to those that readily use the stability criterion of [15]. Our work contributes
to the literature of SAAs with set-valued mean-fields by presenting the first set
of Lyapunov function based sufficient conditions for stability and convergence
of (1). These are also the only sets of sufficient conditions for stability and
convergence of (1), after the ones presented by us in [15].
In this paper, we present Lyapunov function based stability conditions for
the analyses of SAAs with set-valued mean-fields. An important motivation for
2
doing this lies in the proliferation of dynamic programming and reinforcement
learning methods based on value iteration and policy gradients that use deep
neural networks (DNNs) for function approximations and parameterizations,
respectively. The use of DNNs often causes the algorithms to “blow-up”, in
finite time, and find sub-optimal solutions. In this paper we present sufficient
conditions which guarantee that no finite-time blow-up occurs. Further, these
conditions guarantee that the suboptimal solutions found are very close to the
optimal solution. The reader is referred to Sections 7 and 8 for details. Our work
builds on the work of Abounadi, Bertsekas and Borkar [1] and Ramaswamy and
Bhatnagar [15]. Our stability criterion is dependent in part on the possibility
of comparing various instances of an algorithm being analyzed. For the exact
nature of this comparison, the reader is referred to assumption (A5) of Section 3.
As stated earlier, we present three sets of assumptions that are qualitatively
different yet overlapping, for stability and convergence. As a consequence, the
framework developed herein is semantically rich enough to cover a multitude
of scenarios encountered in reinforcement learning, stochastic optimization and
other applications.
We answer the following important question: Does the mere existence of
a Lyapunov function for ẋ(t) ∈ H(x(t)) imply the almost sure boundedness
(stability) of (1)? We will show that the existence of a global/local Lyapunov
function allows us to construct an “inward directing set”, see Proposition 2 for
details. We then use this inward directing set to develop a partner projective
scheme (to (1)). This scheme is shown to converge to a point inside the previously constructed inward directing set. In order to show the stability of (1), we
compare the same to the aforementioned partner projective scheme. The exact
nature of this comparison is outlined in (A5). It is imperative that (1) and it’s
partner projective scheme are comparable. In other words, it seems that the
mere existence of a Lyapunov function is insufficient to ensure stability. Additional assumptions such as (A5) of Section 3 are needed for this purpose. We
demonstrate the verifiability of our assumptions by using our framework to comprehensively analyze two important problems: (i) approximate value iteration
methods with possibly biased approximation errors and (ii) SAAs for finding
fixed points of contractive set-valued maps. It is worth noting that our analysis
of approximate value iteration methods does not distinguish between biased and
unbiased approximation errors.
In Section 7, as an application of our main results, we present a complete
analysis of approximate value iteration methods, an important class of dynamic
programming algorithms, under significantly weaker set of assumptions. Value
iteration is an important dynamic programming method used to numerically
compute the optimal value function of a Markov decision process (MDP). However, it is well known that for many important applications, it suffers from
Bellman’s curse of dimensionality. Approximate Value Iteration (AVI) methods endeavor to address Bellman’s curse of dimensionality by introducing an
approximation operator that allows for approximations at every step of the
classical value iteration method. If the approximation errors are allowed to be
unbounded then the algorithm may not converge, see [6] for details. AVIs with
bounded approximation errors have been previously studied in [6, 11, 13]. Bert-
3
sekas and Tsitsiklis [6] studied scenarios wherein the approximation errors are
uniformly bounded over all states. Munos [13] extended the analysis of [6], allowing for approximation errors that are bounded in the weighted p-norm sense,
for the infinite horizon discounted cost problem. However the convergence analysis of [13] requires that the transition probabilities or future state distributions
be “smooth”. For a detailed comparison of our results concerning AVI methods
to those already present in the literature, see Section 7.2.
An important contribution of this paper is in providing a convergence analysis
of AVIs without the aforementioned restriction on transition probabilities or future distributions (cf. Section 7). Our analysis encompasses both the stochastic
shortest path and the discounted cost infinite horizon problems. When analyzing
stochastic iterative AVIs (see (11) in Section 7 for details on stocastic iterative
AVIs), stability (almost sure boundedness) of the iterates is normally assumed
to hold. As stated before, stability is a hard assumption to verify. Further, it
is unclear if the introduction of an approximation operator leads to unstable
iterates. Thus, an important contribution of this paper is in showing stability
of stochastic iterative AVIs, under weak, verifiable conditions. In Section 7,
it is shown that a stochastic iterative AVI converges to a possibly suboptimal
cost-to-go vector J∞ which belongs to a “small neighborhood” of the optimal
vector, J ∗ . Further it is shown that the size of this neighborhood is directly
proportional to the magnitude of the approximation errors, see Theorems 3 and
4 for details.
Thus, in Section 7 we provide a complete analysis (stability and convergence)
of general AVI methods under weak, easily verifiable, set of sufficient conditions.
We eliminate all previous restrictions on the “smoothness” of transition probabilities and future distributions. We also allow for more general “operational
noise” as compared to previous literature. An important aspect of our analysis is that it encompasses both stochastic shortest path and infinite horizon
discounted cost problems. We provide a unified analysis for stability and convergence of AVI methods wherein the approximation errors are bounded with
respect to multiple norms. Finally, we believe that the theory developed herein
is useful in providing the theoretical foundation for understanding reinforcement
learning and dynamic programming algorithms that use DNNs, an area that has
garnered significant interest recently.
In Section 8, as another important application of our framework, we develop
and analyze, for the first time, general SAAs for finding fixed points of set-valued
maps. Fixed point theory is an active area of research due to its applications
in a multitude of disciplines. Our contribution on this front is in analyzing
stochastic approximation algorithms for finding fixed points of contractive setvalued maps , see Section 8 for details. As mentioned before, we show that such
algorithms are bounded almost surely and that they converge to a sample path
dependent fixed point of the set-valued map under consideration. To the best
of our knowledge ours is the first SAA, complete with analysis, for finding fixed
points of set-valued maps.
4
1.1
Organization of this paper
In Section 2 we list the definitions and notations used in this paper. In Section 3
we present three sets of sufficient conditions for stability and convergence (to a
closed connected internally chain transitive invariant set of the associated DI) of
SAAs with set-valued mean-fields. Through Sections 4, 5 and 6 we present our
main results, Theorems 1 and 2. In Section 7 we analyze stochastic iterative AVI
methods, see Theorems 3 and 4. In Section 8 we develop and analyze a SAA for
finding fixed points of contractive set-valued maps, see Theorem 5 for the main
result here. A detailed discussion on assumption (A5), which is crucial to our
analysis, is provided in Section 9. Finally Section 10 provides the concluding
remarks.
2
Definitions and Notations
The definitions and notations encountered in this paper are listed in this section.
[Upper-semicontinuous map] We say that H is upper-semicontinuous,
if given sequences {xn }n≥1 (in Rn ) and {yn }n≥1 (in Rm ) with xn → x,
yn → y and yn ∈ H(xn ), n ≥ 1, then y ∈ H(x).
[Marchaud Map] A set-valued map H : Rn → {subsets of Rm } is
called Marchaud if it satisfies the following properties: (i) for each x
∈ Rn , H(x) is convex and compact; (ii) (point-wise boundedness) for each
x ∈ Rn , sup kwk < K (1 + kxk) for some K > 0; (iii) H is upperw∈H(x)
semicontinuous.
Let H be a Marchaud map on Rd . The differential inclusion (DI) given
by
ẋ ∈ H(x)
(2)
is guaranteed to have at least one solution that is absolutely continuous.
P
The reader is referred to [2] for more details. We say that x ∈
if x is
an absolutely continuous map that satisfies (2). The set-valued semiflow
d
Φ associated with (2)
Pis defined on [0, +∞) × R as:
Φt (x) = {x(t) | x ∈ , x(0) = x}. Let B × M ⊂ [0, +∞) × Rd and define
[
ΦB (M ) =
Φt (x).
t∈B, x∈M
[Limit set of a solution & ω-limit-set]
The limit set of a solution x
T
with x(0) = x is given by L(x) = t≥0 x([t, +∞)). Let M ⊆ Rd , the
T
ω-limit-set be defined by ωΦ (M ) = t≥0 Φ[t,+∞) (M ). Similarly the limit
T
set of a solution x is given by L(x) = t≥0 x([t, +∞)).
[Invariant set]PM ⊆ Rd is invariant if for every x ∈ M there exists a
trajectory, x ∈ , entirely in M with x(0) = x, x(t) ∈ M , for all t ≥ 0.
Note that the definition of invariant set used in this paper, is the same as
that of positive invariant set in [5] and [9].
5
[Open and closed neighborhoods of a set] Let x ∈ Rd and A ⊆ Rd ,
then d(x, A) := inf{ka − yk | y ∈ A}. We define the δ-open neighborhood
of A by N δ (A) := {x | d(x, A) < δ}. The δ-closed neighborhood of A is
defined by N δ (A) := {x | d(x, A) ≤ δ}. The open ball of radius r around
the origin is represented by Br (0), while the closed ball is represented by
B r (0).
[Internally chain transitive set] M ⊂ Rd is said to be internally chain
transitive if M is compact and for every x, y ∈ M , ǫ > 0 and T > 0 we have
the following: There exists n and Φ1 , . . . , Φn that are n solutions to the differential inclusion ẋ(t) ∈ H(x(t)), points x1 (= x), . . . , xn+1 (= y) ∈ M and
n real numbers t1 , t2 , . . . , tn greater than T such that: Φiti (xi ) ∈ N ǫ (xi+1 )
and Φi[0,ti ] (xi ) ⊂ M for 1 ≤ i ≤ n. The sequence (x1 (= x), . . . , xn+1 (= y))
is called an (ǫ, T ) chain in M from x to y.
[Attracting set & fundamental neighborhood] A ⊆ Rd is attracting
if it is compact and there exists a neighborhood U such that for any
ǫ > 0, ∃ T (ǫ) ≥ 0 with Φ[T (ǫ),+∞) (U ) ⊂ N ǫ (A). Such a U is called
the fundamental neighborhood of A. In addition to being compact if the
attracting set is also invariant then it is called an attractor. The basin of
attraction of A is given by B(A) = {x | ωΦ (x) ⊂ A}.
[Lyapunov stable] The above set A is Lyapunov stable if for all δ > 0,
∃ ǫ > 0 such that Φ[0,+∞) (N ǫ (A)) ⊆ N δ (A).
3
Assumptions
Consider the following iteration in Rd :
xn+1 = xn + a(n) [yn + Mn+1 ] ,
(3)
where yn ∈ H(xn ) for all n ≥ 0 with H : Rd → {subsets of Rd }, {a(n)}n≥0 is
the given step-size sequence and {Mn+1 }n≥0 is the given noise sequence.
We make the following assumptions 1 :
(A1) H : Rd → {subsets of Rd } is a Marchaud map. For all x ∈ Rd sup kzk ≤
z∈H(x)
K(1 + kxk), where K > 0 is the given Marchaud constant.
(A2) The step-size sequence {a(n)}n≥0 is such that ∀n a(n) ≥ 0,
P
∞ and n≥0 a(n)2 < ∞.
(A3)
P
n≥0
a(n) =
(i) For all n ≥ 0 kMn+1 k ≤ D, where D > 0 is some constant.
(ii) lim
mP
T (k)
k→∞ n=k
a(n)Mn+1 = 0 for all T , where mT (k) := min {m ≥ k |
Pm
n=k
1 Note that assumption (A4a) presents a set of Lyapunov conditions on the associated
DI. Two more sets of alternative conditions, viz., (A4b) and (A4c) will also be presented
subsequently in this section.
6
a(n) ≥ T }.
(A4a) Associated with the differential inclusion (DI) ẋ(t) ∈ H(x(t))
is a compact
set Λ, a bounded open neighborhood U Λ ⊆ U ⊆ Rd and a function
V : U → R+ such that
(i) ∀t ≥ 0 Φt (U) ⊆ U i.e., U is strongly positively invariant.
(ii) V −1 (0) = Λ.
(iii) V is a continuous function such that for all x ∈ U \ Λ and y ∈ Φt (x)
we have V (x) > V (y), for any t > 0.
(A5) Let {xn }n≥0 and {x̂n }n≥0 be two sequences generated by (3) on a common probability space with the same noise sequence {Mn+1 }n≥0 . Then
sup kxn − x̂n k < ∞ a.s.
n
Note that it follows from Proposition 3.25 of Benaı̈m, Hofbauer and Sorin [5]
that Λ contains a Lyapunov stable attracting set. Further there exists an attractor contained in Λ whose basin of attraction contains U, with V −1 ([0, r)) as
fundamental neighborhoods for small values of r.
Note that assumptions on the noise, (A3), will be weakened to include more
general noise sequences later, see Section 6. Also, (A5) is similar to an assumption used by Abounadi et. al. [1] in the case of regular SAAs (with point-to-point
maps). The reader is referred to Section 9 for more detailed discussions on this
assumption.
We define open sets Ba and Ca such that the following conditions are satisfied:
(a)Λ ⊆ Ba ⊆ Ca (b) Ca ⊆ U and (c) Ca is a fundamental neighborhood of the
attractor contained in Λ. Since V is continuous, it follows that Vr := {x |
V (x) < r} is T
open relative to U , for all r > 0. Further, assumption (A4a)(ii)
implies that
Vr = Λ and that Vr ⊂ U for small values of r. It may be noted
r>0
that the closure of Vr is Vr := {x | V (x) ≤ r}. Let Ba := VRb and Ca := VRc ,
where 0 < Rb < Rc . Here Rc is chosen small enough to ensure that VRc ⊆ U and
the above conditions (b) and (c) hold. Note that condition (a) is automatically
satisfied since Rb < Rc .
We now present an alternative to assumption (A4a).
(A4b) Associated with ẋ(t) ∈ H(x(t)) is a compact set Λ, a bounded open neighborhood U and a function V : U → R+ such that
(i) ∀t ≥ 0 Φt (U) ⊆ U i.e., U is strongly positively invariant.
(ii) V −1 (0) = Λ.
(iii) V is an upper semicontinuous function such that for all x ∈ U \ Λ
and y ∈ Φt (x) we have V (x) > V (y), where t > 0.
(iv) V is bounded on U. In other words, sup kV (x)k < ∞.
x∈U
The difference between statements (A4a)(iii) and (A4b)(iii) contributes to the
qualitative difference between assumptions (A4a) and (A4b). It follows from
7
Proposition 3.25 of Benaı̈m, Hofbauer and Sorin [5] that Λ contains an attractor
set whose basin of attraction contains U. As in the case of (A4a), we define open
sets Bb and Cb satisfying the above stated (a) and (b). But first, we prove that
sets of the form {x | V (x) < r} are open relative to U, as expected.
Proposition 1. For any r < sup V (u), the set Vr = {x | V (x) < r} is open
u∈U
relative to U . Further, Vr = {x | V (x) ≤ r}.
Proof. PROOF: Suppose Vr is not open relative to U, then there exists {xn }n≥0
such that xn → x, as n → ∞, with xn ∈ U \Vr for every n and x ∈ Vr . It follows
from the boundedness of V , i.e., (A4b)(iv) that there exists {m(n)}n≥0 ⊆ {n ≥
0} such that V (xm(n) ) → v as m(n) → ∞ for some v ∈ R+ . Since xm(n) → x,
it follows from the upper semicontinuity of V that v = V (x). Since v < r and
inf V (xm(n) ) ≥ r, we get a contradiction.
m(n)
To prove the second part of the proposition, it is enough to show that V (x) =
r for every x ∈ ∂Vr . Since Vr is open relative to U and x ∈ ∂Vr , we have
V (x) ≥ r. It is left to show that V (x) ≤ r. Since x ∈ ∂Vr , there exists
xn ∈ Vr such that xn → x. It follows from the upper semicontinuity and the
boundedness of V that V (x) ≤ lim sup V (xn ) ≤ r.
n→∞
As in the case of (A4a), we have Λ =
T
Vr and Vr ⊂ U for small values of
r>0
r. We are now ready to define Bb and Cb . Define Cb := Vr such that Vr ⊂ U,
possible for small values of r. Further, choose Bb such that Bb is open and
Λ ⊂ Bb ⊂ Bb ⊂ Cb . This is possible since Λ is compact and Cb is open.
Remark 1. Let us suppose we are given a differential inclusion ẋ(t) ∈ H(x(t)),
an associated attractor set A and a strongly positive invariant neighborhood
U of A. We define an upper-semicontinuous Lyapunov function V , as found
in Remark 3.26, Section 3.8 of [5]. In other words, V : U → R+ given by
V (x) := max {d(y, A)g(t) | y ∈ Φt (x), t ≥ 0}, where g is an increasing function
such that 0 < c < g(t) < d for all t ≥ 0. We claim that V satisfies (A4b). To
verify this claim we consider the following. U is strongly positive invariant and
V (x) ≤ sup d(u, A) × d for x ∈ U, hence sup V (u) < ∞. It follows from the
u∈U
u∈U
upper semicontinuity of V that sup V (u) < ∞ i.e., (A4b)(iv) is satisfied. It is
u∈U
left to show that (A4b)(iii) is also satisfied. Fix x ∈ U and t > 0. It follows
from the definition of a semi-flow that Φs (y) ⊆ Φt+s (x) for any y ∈ Φt (x),
where s > 0. Further,
V (x) ≥ max{d(z, A)g(t + s) | z ∈ Φs (y), s ≥ 0} and
max{d(z, A)g(t + s) | z ∈ Φs (y), s ≥ 0} > max{d(z, A)g(s) | z ∈ Φs (y), s ≥ 0}.
The RHS of the above equation is V (y) i.e., V (x) > V (y).
We consider one final alternative to (A4a) and (A4b) below.
(A4c)
(i) A is the global attractor of ẋ(t) ∈ H(x(t)).
8
(ii) V : Rd → R+ is an upper semicontinuous function such that V (x) >
V (y) for all x ∈ Rd \ A, y ∈ Φt (x) and t > 0.
(iii) V (x) ≥ V (y) for all x ∈ A, y ∈ Φt (x) and t > 0.
As with (A4a) and (A4b), we define open sets Bc and Cc satisfying conditions
(a) and (b), see below the statement of (A4a). Recall that Vr = {x | V (x) < r}.
Define Cc := Vr and Bc := Vs for appropriate r and s satisfying sup V (x) > r >
x∈Rd
s > sup V (x). Suppose we are unable to find such an s then we may choose Bc to
x∈A
be any open set satisfying the required properties once Cc is fixed as mentioned
before.
Remark 2. In Remark 1, we explicitly constructed a local Lyapunov function
satisfying (A4b). Similarly, here, we construct a global Lyapunov function satisfying (A4c). Define the function V : Rd → R+ as V (x) := max {d(y, A)g(t) | y ∈ Φt (x), t ≥ 0},
where g(· ) is defined in Remark 1. This Lyapunov function, V , satisfies (A4c).
The proof is similar to the one found in Remark 1.
Inward directing sets: Given a differential inclusion ẋ(t) ∈ H(x(t)), an open
set O is said to be an inward directing set with respect to the aforementioned
differential inclusion, if Φt (x) ⊆ O, t > 0, whenever x ∈ O. Clearly inward
directing sets are invariant. It also follows that any solution starting at the
boundary of O is “directed inwards”, into O.
Proposition 2. The open sets Ca , Cb and Cc , constructed in accordance to
assumptions (A4a), (A4b) and (A4c) respectively, are inward directing sets with
respect to ẋ(t) ∈ H(x(t)).
Proof. PROOF: Recall that the set Ca is constructed such that V (x) = Rc for
every x ∈ ∂Ca and V (y) < Rc for every y ∈ Ca . Since Ca ⊆ U, it follows from
(A4a)(iii) that V (x) > V (y), where x ∈ ∂Ca , y ∈ Φt (x) and t > 0. In other
words, V (y) < Rc and Φt (x) ⊂ Ca for t > 0 and x ∈ ∂Ca . It is left to show that
Φt (x) ⊂ Ca for x ∈ Ca and t > 0. This follows directly from the observation that
Rc > V (x) > V (y) for every y ∈ Φt (x) and t > 0. Cb and Cc can be similarly
shown to be inward directing.
In what follows, we use assumptions (A1)-(A3), (A5) and the existence of
an inward directing set with respect to the associated DI, to prove the stability
of (3). As a consequence of Proposition 2, we may verify one among (A4a),
(A4b) and (A4c) to ensure the existence of such an inward directing set. It
may be noted that these assumptions are qualitatively different. However, their
primary role is to help us find one of the aforementioned inward directing sets.
Depending on the nature of the iteration being analyzed, it may be easier to
verify one or the others.
9
4
Analysis of the projective scheme
We begin this section with a minor note on notations. Since the roles of (A4a),
(A4b) and (A4c) are indistinguishable, we shall refer to them generically as
(A4). In a similar manner, Ba/b/c and Ca/b/c are generically referred to as B
and C respectively. Note that B ⊆ C. We also define the projection map,
d
d
B,C : R → {subsets of R }, as follows:
(
{x}, if x ∈ C
.
B,C (x) :=
{y | d(y, x) = d(x, B), y ∈ B}, otherwise
⊓
⊓
As in [1], in order to prove stability, (3) is compared to the following projective
scheme.
x̃(n + 1) = xn + a(n) [yn + Mn+1 ] ,
xn+1 = zn , where zn ∈
(4)
⊓B,C (x̃n+1 ),
⊓
where yn ∈ H(xn ) and x0 ∈ B,C (x̃0 ), with x̃0 ∈ Rd . Note that the initial point
x̃0 is first projected before starting the projective scheme. The above equation
can be rewritten as
xn+1 = xn + a(n) [yn + Mn+1 ] + gn ,
(5)
where gn = zn − (xn + a(n) [yn + Mn+1 ]). Let us construct a linearly interpolated trajectory using (5). We begin by dividing [0, ∞) into diminishing intervals
n−1
P
a(m) for n ≥ 1. The
using the step-size sequence. Let t0 := 0 and tn :=
m=0
linearly interpolated trajectory Xl (t) is defined as follows:
(
xn , for t = tn
Xl (t) :=
t−tn
n
1 − t−t
a(n) xn + a(n) x̃n+1 , for t ∈ [tn , tn+1 ).
.
(6)
The above constructed trajectory is right continuous with left-hand limits, i.e.,
Xl (t) = lim Xl (s) and lim Xl (s) exist. Further the jumps occur exactly at those
s↓t
s↑t
tn ’s for which the corresponding gn−1 ’s are non-zero. We also define three
piece-wise constant trajectories Xc (· ), Yc (· ) and Gc (· ) as follows: Xc (t) := xn ,
n−1
P
Yc (t) := yn and Gc (t) :=
gm for t ∈ [tn , tn+1 ). The trajectories Xc (· ), Yc (· )
m=0
and Gc (· ) are also right continuous with left-hand limits. We define a linearly
interpolated trajectory associated with {Mn+1 }n≥0 as follows:
n−1
P a(m)Mm+1 for t = tn
m=0
.
Wl (t) :=
1 − t−tn Wl (tn ) + t−tn Wl (tn+1 ), for t ∈ [tn , tn+1 )
a(n)
a(n)
We define a few “left-shifted trajectories” using the above constructed trajectories. For t ≥ 0,
Xln (t) := Xl (tn + t),
10
Xcn (t) := Xc (tn + t),
Ycn (t) := Yc (tn + t),
Gnc (t) := Gc (tn + t) − Gc (tn ),
Wln (t) := Wl (tn + t) − Wl (tn ).
Lemma 1. Xln (t) = Xln (0) +
Proof.
Rt
0
Ycn (τ ) dτ + Wln (t) + Gnc (t) for t ≥ 0.
Fix s ∈ [tm , tm+1 ) for some m ≥ 0. We have the following
s − tm
s − tm
xm +
x̃m+1 ,
Xl (s) = 1 −
a(m)
a(m)
s − tn
s − tm
xm +
(xm + a(m) [ym + Mm+1 ]) ,
Xl (s) = 1 −
a(m)
a(m)
PROOF:
Xl (s) = xm + (s − tm ) [ym + Mm+1 ] .
Xln (t)
Let us express
in the form of the above equation. Note that tn + t ∈
[tn+k , tn+k+1 ) for some k ≥ 0. Then we have the following:
Xln (t) = xn+k + (tn + t − tn+k ) [yn+k + Mn+k+1 ] .
Unfolding xn+k , in the above equation till xn (Xln (0)), yields:
Xln (t) = Xln (0)+
n+k−1
X
(a(l) [yl + Ml+1 ] + gl )+(tn +t−tn+k ) [yn+k + Mn+k+1 ] .
l=n
(7)
We make the following observations:
n+k−1
P
Gnc (t) =
gl ,
l=n
Wln (tn+k
− tn ) =
n+k−1
P
a(l)Ml+1 ,
l=n
Wln (t) = Wln (tn+k − tn ) + (tn + t − tn+k )Mn+k+1 and
n+k−1
Rt n
P
a(l)yl + (tn + t − tn+k )yn+k .
Yc (τ ) dτ =
0
l=n
As a consequence of the above observations, (7) becomes:
Xln (t)
=
Xln (0)
+
Zt
Ycn (τ ) dτ + Wln (t) + Gnc (t).
0
Fix T > 0. If {Xln ([0, T ]) | n ≥ 0} and {Gnc ([0, T ]) | n ≥ 0} are viewed as subsets
of D([0, T ], Rd) equipped with the Skorohod topology, then we may use the
Arzela-Ascoli theorem for D([0, T ], Rd ) to show that they are relatively compact,
see Billingsley [7] for details. The Arzela-Ascoli theorem for D([0, T ], Rd) states
the following: A set S ⊆ D([0, T ], Rd ), is relatively compact if and only if the
following conditions are satisfied:
11
sup
sup kx(t)k < ∞,
x(·)∈S t∈[0,T ]
lim sup
sup
δ→0 x(·)∈S t1 ≤t≤t2 , t2 −t1 ≤δ
lim sup
sup
δ→0 x(·)∈S t1 ,t2 ∈[0,δ)
lim sup
sup
min {kx(t) − x(t1 )k, kx(t2 ) − x(t)k} = 0,
kx(t2 ) − x(t1 )k = 0 and
kx(t2 ) − x(t1 )k = 0.
δ→0 x(·)∈S t1 ,t2 ∈[T −δ,T )
If {Xln ([0, T ]) | n ≥ 0} and {Gnc ([0, T ]) | n ≥ 0} are point-wise bounded and any
two of their discontinuities are separated by at least ∆, for some fixed ∆ > 0,
then the above four conditions will be satisfied, see [7] for details.
Lemma 2. {Xln ([0, T ]) | n ≥ 0} and {Gnc ([0, T ]) | n ≥ 0} are relatively compact
in D([0, T ], Rd) equipped with the Skorohod topology.
Proof. PROOF: Recall from (A3)(i) that kMn+1 k ≤ D, n ≥ 0. Since H is Markyk ≤ C1 for some C1 > 0 and that sup kx̃n+1 −
chaud, it follows that
sup
n≥0
x∈C, y∈H(x)
xn k ≤ sup a(m) (C1 + D). Further, gn ≤ kx̃n+1 − xn k + d(xn , ∂B) ≤ C2
m≥0
for some constant C2 that is independent of n. In other words, we have that
the sequences {Xln ([0, T ]) | n ≥ 0} and {Gnc ([0, T ]) | n ≥ 0} are point-wise
bounded. It remains to show that any two discontinuities are separated. Let
d := min d(x, B) and D1 := D + sup sup kyk. Clearly d > 0. Define
x∈∂C
x∈C y∈H(x)
(
m(n) = max j > 0 |
n+j
X
a(k) < d/D1
k=n
)
.
If there is a jump at tn , then xn ∈ ∂B. It follows from the definition of m(n)
that xk ∈ C for n ≤ k ≤ m(n). In other words, there are no discontinuities in
d
d
the interval [tn , tn+m(n) ) and tn+m(n) − tn ≥ 2D
. If we fix ∆ := 2D
> 0, then
1
1
any two discontinuities are separated by at least ∆.
Since T is arbitrary, it follows that {Xln ([0, ∞)) | n ≥ 0} and {Gnc ([0, ∞)) | n ≥
0} are relatively compact in D([0, ∞), Rd ). Since {Wln ([0, T ]) | n ≥ 0} is pointwise bounded (assumption (A3)(i)) and continuous, it is relatively compact in
D([0, T ], Rd). It follows from (A3)(ii) that any limit of {Wln ([0, T ]) | n ≥ 0}, in
D([0, T ], Rd), is the constant
of
( 0 function. Suppose we consider sub-sequences
)
T
R
{Xln ([0, T ]) | n ≥ 0} and Xln (0) + Ycn (τ ) dτ + Gnc (T ) | n ≥ 0 along which
0
the aforementioned noise trajectories converge, then their limits are identical.
m(n)
m(n)
([0, T ])}n≥0 and {Xl
([0, T ])}n≥0
Consider {m(n)}n≥0 ⊆ N along which {Gc
converge in D([0, T ], Rd ). Further, let gm(n)−1 = 0 for all n ≥ 0. Suppose the
12
m(n)
limit of {Gc
([0, T ])}n≥0 is the constant 0 function, then it can be shown that
m(n)
the limit of {Xl
([0, T ])}n≥0 is
X(t) = X(0) +
Zt
Y (τ )dτ,
0
where X(0) ∈ C and Y (t) ∈ H(X(t)) for t ∈ [0, T ]. The proof of this is along
the lines of the proof of Theorem 2 in Chapter 5.2 of Borkar [9]. Suppose every
m(n)
limit of {Gc
([0, T ])}n≥0 is the constant 0 function whenever gm(n)−1 = 0
m(n)
for all n ≥ 0, then every limit of {Xl
([0, T ])}n≥0 is a solution to Ẋ(t) ∈
H(X(t)). Suppose we show that the aforementioned statement is true for every
T > 0. Then, along with (A5) the stability of (3) is implied. Note that the set
K := {n | gn = 0} has infinite cardinality since any two discontinuities are at
least ∆ > 0 apart.
Lemma 3. Let K = {n | gn = 0}. Without loss of generality let {Xln ([0, T ])}n∈K
and {Gnc ([0, T ])}n∈K be convergent, as n → ∞, in D([0, T ], Rd). Then Xln (t) →
Rt
Y (s) ds + G(t) for t ∈ [0, T ] and G(· ) ≡ 0.
X(0) +
τ =0
Proof.
PROOF:
We begin by making the following observations:
(a) X(0) ∈ C.
(b) Any two discontinuities of X(· ) are at least ∆ apart.
(c) G(0) = 0.
(d) Solutions to ẋ(t) ∈ H(x(t)) with starting points in C will not hit the
boundary, ∂C, later, i.e., they remain in the interior of C. This observation
is a consequence of Proposition 2.
It follows from the above observations that (i) X(t) ∈ C for small values of t,
(ii) τ := inf{t | t > 0, X(t+ ) 6= X(t− )} and τ > 0. It follows from the nature of
convergence that ∃ τn′ > τ > τn , n ≥ 0 such that
τn′ − τn → 0,
kXln (τn′ ) − X(τ + )k → 0 and
kXln (τn ) − X(τ − )k → 0.
For large values of n, Xln (· ) has exactly one jump (point of discontinuity) at
τ̂n ∈ [τn , τn′ ], let us call this point of discontinuity as τ̂n . Let δ := kX(τ + ) −
X(τ − )k > 0, then for large values of n we have
kXln (τ̂n+ ) − Xln (τ̂n− )k ≥ δ/2.
Also, Xln (τ̂n− ) is not in C and Xln (τ̂n+ ) is in ∂B. Further, since τ̂n− − τn → 0, as
n → ∞, it follows that
Xln (τ̂n− ) − X(τ − ) → 0.
Hence, X(τ − ) ∈
/ C. Similarly, we have that X(τ + ) ∈ ∂B. Observe that X([0, τ ))
is a solution to ẋ(· ) ∈ H(x(· )) such that X(0) ∈ C, since G(t) = 0 for t ∈ [0, τ ).
Further, since C is inward directing, we have that X(t) ∈ C for t ∈ [0, τ ). Since
X(t) ∈ C for t < τ and X(τ − ) ∈
/ C we have X(τ − ) ∈ ∂C.
13
Since X(0) ∈ C, we have that V (X(0)) ≤ R, for some 0 < R < ∞. As a consequence of our choice of C (C is Ca /Cb /Cc within the context of (A4a)/(A4b)/(A4c))
we have V (x) = V (y) for any x, y ∈ ∂C, hence we may fix R := V (x) for any
x ∈ ∂C. Fix τ0 ∈ (0, τ ), it follows from Proposition 2 that V (X(τ0 )) < R.
Let tn ↑ τ such that tn ∈ (τ0 , τ ) for n ≥ 1. Without loss of generality,
X(tn ) → X(τ − ) and V (X(tn )) → V (X(τ − )), as tn → τ (else we may choose
a sub-sequence of {tn }n≥0 along which V (X(tn )) is convergent). Thus, ∃ N
such that V (X(tn )) > V (X(τ0 )) for n ≥ N . Since X([τ0 , tn ]) is a solution to
ẋ(t) ∈ H(x(t)) with starting point X(τ0 ), the aforementioned conclusion contradicts (A4a)(iii)/(A4b)(iii)/(A4c). In other words, X(τ − ) ∈ C and ∈
/ ∂C.
Thus we have shown that there is no jump at τ , i.e., X(τ + ) = X(τ − ).
Suppose (A4a) / (A4b) holds, then it follows from Proposition 3.25 of Benaı̈m,
Hofbauer and Sorin [5] that there is an attracting set A ⊆ Λ such that C a / C b
is within the basin of attraction. Suppose (A4c) holds, then A is the globally
attracting set of ẋ(t) ∈ H(x(t)).
Lemma 4. The projective stochastic approximation scheme given by (4) converges to the attractor A.
Proof. PROOF: We begin by noting that T of Lemma 3 is arbitrary. Since
G ≡ 0, after a certain number of iterations of (4), there are no projections, i.e.,
x̃n = xn for n ≥ N . Here N could be sample path dependent. Further, it
follows from Lemma 3 that the projective scheme given by (4) tracks a solution
to ẋ(t) ∈ H(x(t)). In other words, the projective scheme given by (4) converges
to a limit point of the DI, ẋ(t) ∈ H(x(t)).
The iterates given by (4) are within C after sometime and they track a
solution to ẋ(t) ∈ H(x(t)). Since C is within the basin of attraction of A, the
iterates converge to A.
5
Main results: stability and convergence
In this section we show that the iterates given by (3) are stable (bounded almost
surely) and converge to a closed connected internally chain transitive invariant
set associated with ẋ(t) ∈ H(x(t)). We show the stability of (3), by comparing
the iterates generated by (3) to those generated by (4). As stated in (A5), we
assume that the noise sequence is exactly the same for both (3) and (4).
Theorem 1. Under (A1)-(A5), the iterates given by (3) are stable (bounded
almost surely) and converge to a closed, connected, internally chain transitive,
invariant set associated with ẋ(t) ∈ H(x(t)).
Proof. PROOF: For notational convenience we use {x̂n }n≥0 for the iterates generated by the projective scheme and {xn }n≥0 for the iterates generated by (3).
Recall that x̂n → A, where A ⊂ B, see Lemma 4. In other words there exists
N , possibly sample path dependent, such that x̂n ∈ B for all n ≥ N . In other
14
words, sup kx̂n k ≤ supkxk < ∞. It follows from (A5) that sup kxn k < ∞. Aln≥N
x∈B
n≥N
most surely, there exists N , possibly sample path dependent, such that x̂n ∈ B
for all n ≥ N . This directly leads to the stability of (3).
The iterates given by (3) satisfy assumptions (A1)-(A3). Further, we have
shown above that they are also stable. It follows from Theorem 3.6 and Lemma
3.8 of Benaı̈m [5] that the iterates converge to a closed connected internally
chain transitive invariant set associated with ẋ(t) ∈ H(x(t)).
6
General noise sequences
The restriction (A3) on the noise sequence, is rather strict since it only allows
for “bounded noise”. In this section, we show that Theorem 1 will continue to
hold even when the noise sequence is more generally a square integrable Martingale difference sequence. Analysis of regular SAAs has been shown in [1] under
this assumption on the noise sequence.
(A3a) (Mn , Fn )n≥1 is a square integrable Martingale difference sequence E[Mn+1 |
2
Fn ] = 0 and EMn+1 < ∞, n ≥ 0 such that
E kMn+1 k2 | Fn ≤ K(1 + kxn k2 ), where n ≥ 0 and K > 0.
In the above, F0 := σhx0 i and Fn := σ hx0 , x1 , . . . , xn , M1 , . . . , Mn i for n ≥ 1.
Further without loss of generality, we may assume that K in (A3a) and (A1)
are equal (otherwise we can use maximum of the two constants).
In the analysis of the projective scheme given by (4), assumption (A3) is used
in Lemma 2. Specifically, (A3)(i) is used to show that any two discontinuities
of {Xnl ([0, T ])}n≥0 and {Gnc ([0, T ])}n≥0 are separated by at least ∆ > 0. We
show that the aforementioned property holds when (A3) is replaced by (A3a).
First, we prove an auxiliary result.
Lemma 5. Let {tm(n) , tl(n) }n≥0 be such that tl(n) > tm(n) , tm(n+1) > tl(n) and
lim tl(n) − tm(n) = 0. Fix an arbitrary c > 0 and consider the following:
n→∞
l(n)−1
ψn :=
X
a(i)Mi+1 .
i=m(n)
Then P ({ψn > c} i.o.) = 0 within the context of the projective scheme given by
(4).
P
P (ψn > c) < ∞. It follows from
Proof. PROOF: We need to show that
n≥0
Chebyshev’s inequality that
P (ψn > c) ≤
Eψn2
c2
=
E
15
l(n)−1
P
2
a(i)Mi+1
i=m(n)
c2
.
Since {Mn+1 }n≥0 is a Martingale difference sequence, we get:
P (ψn > c) ≤
l(n)−1
P
i=m(n)
a(i)2 E kMi+1 k2
.
c2
(8)
Within the context of the projective scheme given by (4), almost surely
∀ n xin ∈
h
2
C, i.e., sup kxn k ≤ C1 < ∞ a.s. It follows from (A3a) that E kMn+1 k ≤
n≥0
h
i
2
K 1 + Ekxn k2 . Hence, E kMn+1 k ≤ K 1 + C12 . Equation (8) becomes
P (ψn > c) ≤
l(n)−1
P
i=m(n)
a(i)2 K 1 + C12
c2
Since tl(n) > tm(n) and tm(n+1) > tl(n) , we have
P
.
l(n)−1
P
a(i)2 ≤
X
P (ψn > c) ≤
P
2
a(n)
n≥0
!
K 1 + C12
c2
n≥0
a(n)2 .
n≥0
n≥0 i=m(n)
Finally we get,
P
< ∞.
Let us consider the scenario in which we cannot find ∆, the separation between
any two points of discontinuity. In other words, there exists {t(m(n)), t(l(n))}n≥0
such that tl(n) > tm(n) , tm(n+1) > tl(n) and lim tl(n) − tm(n) = 0. Without loss
n→∞
of generality we assume that there are no jumps between t(m(n)) and t(l(n)).
−
Note that Xl (t+
/ C for all n ≥ 0. We have
m(n) ) ∈ ∂B and Xl (tl(n) ) ∈
l(n)−1
−
Xl (tl (n) ) = Xl (tm (n)) +
X
a(i) (yi + Mi+1 ) .
i=m(n)
We have that sup kyn k ≤ D′ for some 0 < D′ < ∞. The above equation
n≥0
becomes
l(n)−1
l(n)−1
kXl (tl (n)− ) − Xl (tm (n))k ≤
X
a(i)D′ +
i=m(n)
X
a(i)Mi+1 ,
i=m(n)
l(n)−1
d ≤ kXl (tl (n)− ) − Xl (tm (n))k ≤ tl(n) − tm(n) D′ +
Since tl(n) − tm(n) → 0, for large n
l(n)−1
P
a(i)Mi+1
X
a(i)Mi+1 .
i=m(n)
> d/2. This directly
i=m(n)
contradicts Lemma 5. Hence we can always find ∆ > 0 separating any two
points of discontinuity.
16
In Lemma 4, (A3) is used to ensure the convergence of (4) to the attractor A.
In Theorem 1, (A3) is used to ensure the convergence of (3) to a closed connected
internally chain transitive invariant set of the associated DI. Specifically, it is
n−1
P
(A3)(ii) that ensures these convergences. Let us define ζn :=
a(k)Mk+1 ,
k=0
n ≥ 1. If {ζn }n≥1 converges, then it trivially follows that the Martingale
noise sequence satisfies (A3)(ii). To show convergence, it is enough to show
that the corresponding quadratic variation
converges almost
surely.
P process
a(n)2 E kMn+1 k2 |Fn < ∞ a.s or
In other words, we need to show that
n≥0
P
E a(n)2 kMn+1 k2 < ∞. Consider the following:
n≥0
X
a(n)2 EkMn+1 k2 =
n≥0
X
n≥0
X
a(n)2 E E kMn+1 k2 | Fn ≤
a(n)2 K(1+Ekxn k2 ).
n≥0
(9)
Convergence of the quadratic variation process in the context of Lemma 4 follows
from (9) and the fact that Ekxn k2 ≤ sup kxk2 . In other words,
x∈C
X
n≥0
X
a(n)2 E kMn+1 k2 ≤
a(n)2 K(1 + sup kxk2 ) < ∞.
n≥0
x∈C
Similarly, for convergence in Theorem 4, it follows from (9) and stability of the
iterates (supkxn k < ∞ a.s.) that
n≥0
X
n≥0
X
a(n)2 E kMn+1 k2 |Fn ≤
a(n)2 K(1 + sup kxn k2 ) < ∞ a.s.
n≥0
n≥0
In other words, both in Lemma 4 and Theorem 1, assumption (A3)(ii) is satisfied. The following generalized version of Theorem 1 is a direct consequence of
the observations made above.
Theorem 2. Under (A1), (A2), (A3a), (A4) and (A5), the iterates given by (3)
are stable (bounded almost surely) and converge to a closed connected internally
chain transitive invariant set associated with ẋ(t) ∈ H(x(t)).
7
Application: Approximate value iteration methods
In this section we present an analysis of the recursion
Jn+1 = AT Jn ,
(10)
where Jn ∈ Rd for all n, T is the Bellman operator, and A is the approximation operator. Specifically, we consider the following stochastic approximation
counterpart of the above scheme:
Jn+1 = Jn + a(n) [T Jn − Jn + ǫn + Mn+1 ] , where
(i) T is the Bellman operator.
(ii) {a(n)}n≥0 is the step-size sequence, satisfying (A2).
17
(11)
(iii) ǫn = AT Jn − T Jn is the approximation error at stage n.
(iv) {Mn+1 }n≥0 is the Martingale difference noise sequence.
Let us call (11) as a stochastic iterative AVI. It is worth noting that we do
not distinguish between stochastic shortest path and infinite horizon discounted
cost problems. The definition of the Bellman operator T changes appropriately.
We make the following assumptions:
(AV1) The Bellman operator T is contractive with respect to some weighted
max-norm, k· kν , i.e., kT x − T ykν ≤ αkx − ykν for some 0 < α < 1.
(AV2) T has a unique fixed point J ∗ and J ∗ is the unique globally asymptotically
˙ = T J(t) − J(t).
stable equilibrium point of J(t)
(AV3) kǫn kν ≤ ǫ for some fixed ǫ > 0.
Recall the definition of the weighted
ν = (ν1 , . . . , νd ) such that
o
n max-norm: Given
|xi |
νi > 0 for 1 ≤ i ≤ d, kxkν = max νi | 1 ≤ i ≤ d , where x = (x1 , x2 , . . . , xd ) ∈
Rd . Later in this section, we analyze (11) where the approximation errors are
bounded in the general weighted p-norm sense (weighted Euclidean norms).
(AV 1) is readily satisfied in many applications, see Section 2.2 of Bertsekas
and Tsitsiklis [6] for details. First, let us consider a couple of technical lemmas.
Lemma 6. B ǫ := {y | kykν ≤ ǫ} is a convex compact subset of Rd , where ǫ > 0.
Proof. PROOF: First we show that B ǫ is convex. Given y1 , y2 ∈ B ǫ and y =
λy1 + (1 − λ)y2 , where λ ∈ (0, 1), we need show that y ∈ B ǫ . kykν ≤ λky1 kν +
(1 − λ)ky2 kν = λǫ + (1 − λ)ǫ.
Now we show that B ǫ is compact. Define νmax := max {νi | 1 ≤ i ≤ d}.
Since kyk∞ ≤ νmax kykν , it follows that B ǫ is a bounded set. It is left to show
that B ǫ is closed. Let yn → y and yn ∈ B ǫ for every n. Since lim inf kyn kν ≥
n→∞
kykν , it follows that y ∈ B ǫ .
Lemma 7. The set-valued map T̃ given by T̃ x 7→ T x + B ǫ is a Marchaud map.
Proof. PROOF: Since B ǫ is a compact convex set, it follows that T̃ x is compact
and convex. Since T is a contraction map, we have:
kT xkν ≤ kT 0kν + kx − 0kν .
Let νmin := min {ν1 , ν2 , . . . , νd } and νmax := max {ν1 , ν2 , . . . , νd }, then observe
that kzkν ≤ kzk/νmin , where z ∈ Rd and k· k is the standard Euclidean norm.
Also, observe that
q
kzk = z12 + . . . + zd2 ≤ (d νmax ) kzkν .
Hence we have
kT xk ≤ (d νmax )kT xkν ≤ (d νmax ) (kT 0kν + kxkν ) and that
kT xk ≤ K ′ (1 + kxk),
18
(12)
where K ′ :=
d νmax
νmin
∨ (d νmax kT 0kν ). We have that sup kzk ≤ kT xk +
z∈T̃ x
sup kzk. It follows from Lemma 6 that K := K ′ ∨ sup kzk is finite, hence
z∈B ǫ
z∈B ǫ
sup kzk ≤ K(1 + kxk).
z∈T̃ x
We now show that T̃ is upper semicontinuous. Let xn → x, yn → y and
yn ∈ T̃ xn for n ≥ 0. Since T is continuous, we have that T xn → T x. Hence,
(yn − T xn ) → (y − T x) and ǫ ≥ lim inf kyn − T xn kν ≥ ky − T xkν . In other
n→∞
words, y ∈ T̃ x.
Let us define Hν , the Hausdorff metric with respect to the weighted max-norm
as follows: Given A, B ⊆ Rd , Hν (A, B) := max dν (x, B) ∨ max dν (y, A), where
x∈A
y∈B
dν (x, B) := min{kx − ykν | y ∈ B} and dν (y, A) := min{kx − ykν | x ∈ A}.
Given x, y ∈ Rd , there exist x∗ ∈ T̃ x and y ∗ ∈ T̃ y such that kx∗ − y ∗ kν =
Hν (T̃ x, T̃ y). For any x0 ∈ T̃ x and y0 ∈ T̃ y we have
kx0 − y0 kν ≤ kx0 − x∗ kν + kx∗ − y ∗ kν + ky ∗ − y0 kν = 2ǫ + H(T̃ x, T̃ y). (13)
For any z ∈ Rd we have
νmin kzkν ≤ kzk ≤ (d νmax )kzkν .
(14)
Consider the set of all equilibrium points of T̃ : A := {J | kT J − Jkν ≤ ǫ}. For
small values of ǫ, it follows from the above set of inequalities that A belongs
to a small neighborhood of J ∗ . Similarly for small values of ǫ, the mean-field
T̃ is a minor perturbation of T . Recall that J ∗ is the globally asymptotic
˙
stable equilibrium point of J(t)
= T J(t) − J(t). It follows from the upper
semicontinuity of attractors that for small values of ǫ there exists A′ within a
small neighborhood of J ∗ such that A′ is the global attractor of the perturbed
˙
system J(t)
∈ T̃ J(t) − J(t). We will show that (11) converges to A ∩ A′ and
′
that A ∩ A 6= φ.
˙ ∈
We may construct a global Lyapunov function for the attractor A′ of J(t)
T̃ J(t) − J(t) as illustrated in Remark 2. In other words, (11) satisfies (A4c).
Hence we can find sets B and C such that A′ ⊆ B and B ⊆ C. Let us consider
the following projective approximate value iteration:
Jn+1 = Jˆn + a(n) T Jˆn − Jˆn + ǫ̂n + Mn+1 ,
Jˆn+1 =
⊓B,C (Jn+1).
(15)
It is worth noting that the noise sequences in (11) and (15) are identical and
that ǫ̂n ≤ ǫ for all n. Following the analysis in Section 4, we conclude that
Jˆn → A′ . We are now ready to analyze (11).
Theorem 3. Under (AV 1)-(AV 3), (A2), and (A3a), (11) is stable and converges to some point in {J | kT J − Jkν ≤ ǫ}, where ǫ is the norm-bound on the
approximation errors.
19
Proof. PROOF: We start by showing that (11) satisfies assumption (A5). Earlier
we showed that Jˆn → A′ . This implies that there exists N , possibly sample path
dependent, such that Jˆn ∈ B for all n ≥ N . For k ≥ 0 and n ≥ N ,
kJn+k+1 −Jˆn+k+1 kν ≤ Jn+k − Jˆn+k + a(n + k) (T Jn+k + ǫn+k ) − (T Jˆn+k + ǫ̂n+k ) − (Jn+k − Jˆn+k )
ν
.
Grouping terms of interest in the above inequality we get:
kJn+k+1 −Jˆn+k+1 kν ≤ (1−a(n+k))kJn+k −Jˆn+k kν +a(n+k)k(T Jn+k +ǫn+k )−(T Jˆn+k +ǫ̂n+k )kν .
As a consequence of (13) the above equation becomes
kJn+k+1 −Jˆn+k+1 kν ≤ (1−a(n+k))kJn+k −Jˆn+k kν +a(n+k) 2ǫ + Hν (T̃ Jn+k , T̃ Jˆn+k ) .
(16)
We now consider the following two cases:
Case 1. 2ǫ ≤ (1 − α)kJn+k − Jˆn+k kν :
In this case (16) becomes
kJn+k+1 −Jˆn+k+1 kν ≤ (1−a(n+k))kJn+k −Jˆn+k kν +a(n+k) (1 − α)kJn+k − Jˆn+k kν + αkJn+k − Jˆn+k kν .
Simplifying the above equation, we get
kJn+k+1 − Jˆn+k+1 kν ≤ kJn+k − Jˆn+k kν .
Case 2. 2ǫ > (1 − α)kJn+k − Jˆn+k kν :
In this case (16) becomes
2ǫ
2ǫ
ˆ
.
kJn+k+1 − Jn+k+1 kν ≤ (1 − a(n + k))
+ a(n + k) 2ǫ + α
1−α
1−α
Simplifying the above equation, we get
kJn+k+1 − Jˆn+k+1 kν ≤
2ǫ
.
1−α
We may thus conclude the following:
kJn+k+1 − Jˆn+k+1 kν ≤ kJn+k − Jˆn+k kν ∨
2ǫ
1−α
.
Applying the above set of arguments to kJn+k − Jˆn+k kν and proceeding recursively to kJN − JˆN kν we may conclude that for any n ≥ N
2ǫ
ˆ
ˆ
kJn − Jn kν ≤ kJN − JN kν ∨
.
(17)
1−α
In other words, we have sup kJn − Jˆn kν < ∞ a.s. It follows from (14) that
n≥0
sup kJn − Jˆn k < ∞ a.s. (in the Euclidean norm). The above arguments
n≥0
are inspired by Abounadi, Bertsekas and Borkar [1] and Jaakkola, Jordan and
Singh [12].
20
We have shown that (11) satisfies (A1), (A2), (A3), (A4c) and (A5). It
follows from Theorem 2 that the iterates given by (11) track a solution to
˙ ∈ T̃ J(t) − J(t) and that Jn → S, where S is a closed connected internally
J(t)
˙ ∈ T̃ J(t)−J(t). Since A′ is a global attractor
chain transitive invariant set of J(t)
˙ ∈ T̃ J(t) − J(t), it follows that S ⊆ A′ .
of J(t)
Now we show that Jn → A, the set of equilibrium points of T̃ J − J, implying
that Jn → A ∩ A′ and A ∩ A′ 6= φ. Before proceeding further, we consider the
following theorem from Aubin and Cellina [2]:
[Theorem 2, Chapter 6 [2]] Let F be an upper semicontinuous map from a
closed subset K ⊂ X to X with compact convex values and x(· ) be a solution
trajectory of ẋ(t) ∈ F (x(t)) that converges to some x∗ in K. Then x∗ is an
equilibrium of F .
We have already shown that supkJn k < ∞ a.s. In other words, there exists
n≥0
a large compact convex set K ⊆ Rd , possibly sample path dependent, such that
Jn ∈ K for all n ≥ 0. Further, K can be chosen such that the “tracking solution”
of ẋ(t) ∈ T̃ J(t) − J(t) is also inside K, asymptotically. It follows that the conditions of the above stated theorem are satisfied, hence every limit point of (11)
is an equilibrium point of T̃ J − J. In other words, Jn → {J | kT J − Jkν ≤ ǫ}.
Remark 3. As the bound on the approximation errors decreases, the size of the
limiting set of the corresponding approximate value iteration given by (11) also
decreases. Specifically, as ǫ ↓ 0, {J | kT J − Jkν ≤ ǫ} ↓ J ∗ .
7.1
Approximation errors bounded in the weighted p-norm
sense
The approximation errors encountered hitherto in this section are bounded in
the weighted max-norm sense. As stated earlier, these errors are often a consequence of approximation operators that are used to counter Bellman’s curse of
dimensionality. In a model-free setting, typically one is given data of the form
(xn , vn ), where vn s are unbiased estimates of the objective function at xn ’s.
Note that the role of an approximation operator may be played by a supervised
learning algorithm. This algorithm would return a good fit g from within a
class of functions T . The objective for these algorithms would be to minimize
the approximation errors. Previously, we considered approximation operators
that minimize errors in the weighted max-norm sense. This is relevant to largescale applications where it may not be possible to approximate over all the states
uniformly.
In many applications the approximation operators work by minimizing the
errors in the ℓ1 and ℓ2 norms, see Munos [13] for details. In this section, we
consider the general case of approximation errors being bounded in the weighted
21
p-norm sense. Specifically, we analyze (11) with kǫn kω,p ≤ ǫ for some fixed ǫ > 0.
Recall the definition of the weighted p-norm of a given z ∈ Rd :
kzkω,p =
d
X
i=1
|ωi zi |
p
!1/p
,
where ω = (ω1 , . . . , ωd ) is such that ωi > 0, 1 ≤ i ≤ d and p ≥ 1.
Recall that the Bellman operator T is contractive with respect to k· kν . First,
we establish the relationship between the weighted p-norm and the weighted
max-norm. Again, fix z ∈ Rd . We have that |zi | ≤ νmax kzkν , where νmax
is defined in Lemma 6. Hence kzkω,p ≤ (d ωmax νmax )kzkν , where ωmax :=
kzkω,p
max {ωi | 1 ≤ i ≤ d}. Similarly, kzkν ≤ νmin
. We have the following
inequality:
νmin kzkν ≤ kzkω,p ≤ (d ωmax νmax )kzkν .
(18)
In this section we consider the following stochastic iterative AVI scheme:
Jn+1 = Jn + a(n) (T Jn − Jn + ǫ̃n + Mn+1 )
(19)
where kǫ̃n kω,p ≤ ǫ for all n ≥ 0. If we define T J := T J + {x | kxkω,p ≤ ǫ}, as in
the previous subsection, we can show that T is a Marchaud map. Further, we
may state an identical theorem for (19).
Theorem 4. Under (AV 1)-(AV 3), (A2), and (A3a), (19) is stable and converges to some point in {J | kT J − Jkω,p ≤ ǫ}, where ǫ is the norm-bound on
the approximation errors.
Proof. PROOF: As in the proof of Theorem 3, we need to show that (19) satisfies
(A5). It follows from (18) that
kǫn kω,p ≤ ǫ =⇒ kǫn kν ≤ ǫ/νmin .
We are now in the setting of Theorem 3, and it follows that (19) satisfies (A5).
With regards to convergence, using similar arguments, we can show that (19)
converges to the set of equilibrium points of T J − J given by {J | kT J − Jkω,p ≤
ǫ}. As in the previous subsection,
{J | kT J − Jkω,p ≤ ǫ} ↓ J ∗ as ǫ ↓ 0.
7.2
Comparison with previous literature
An important contribution in understanding the convergence of approximate
value iteration (AVI) methods has been due to Munos in 2005 [13]. This paper analyzed AVI methods (for the infinite horizon discounted case problem)
wherein the approximation errors are bounded in the weighted p-norm sense,
a significant improvement over [6] that only considered max norms. However,
22
in [13], the basic procedure considered is a numerical AVI scheme where complete knowledge of the ‘system model’, i.e., the transition probabilities is assumed. In addition, in [13], convergence is shown under one of the following
two assumptions on transition probabilities. Let µ be a distribution over the
state space.
(MA1) ∃C > 0 such that for all states x, y and policy π, we have P π (x, y) ≤
Cµ(y).
(MA2) ∃ a distribution ρ and co-efficients c(m) ∀m ≥ 1 and policies π1 , . . . , πm
the dissuch that ρP π1 . . . P πm ≤ c(m)µ. The smoothness constand
P C of
mγ m−1 c(m),
counted future state distribution is given by C := (1−γ)2
m≥1
where γ is the discount factor.
These smoothness requirements on the transition probabilities are strong and
will not hold for a large class of systems. It is worth noting that such smoothness
requirements are eliminated in our analysis.
We consider a stochastic approximation counterpart of AVI that involves an
‘operational noise’ component in addition, that we model as a general martingale
difference noise sequence. An algorithm such as (11) arises for instance in the
Q-learning procedure (where the Bellman operator T is in fact the Q-Bellman
operator) or when the same is a Bellman operator corresponding to a given
stationary policy, see Chapter 5 of [6]. In such a model, no information is
assumed known about the system transition probabilities, i.e., the setting in
such a case is ‘model-free’. Thus our convergence analysis works in the case
of AVI schemes for which (a) information on the transition probabilities is not
known, (b) no such restrictions as (MA1) or (MA2) are imposed on the transition
probabilities, and (c) there is a measurement error (albeit bounded) that may
arise for instance from the use of function approximation. We require that J ∗
˙
be the unique globally asymptotically stable equilibrium of J(t)
= T J(t) −
J(t), where T is the Bellman operator that is a contraction map. Since J ∗ is
the unique solution to the equation T J = J, it is natural to expect that the
aforementioned requirement holds true.
It is also important to note that our analysis works for both stochastic shortest path and infinite horizon discounted cost problems.
Value iteration is an important reinforcement learning algorithm. As stated
in Section 1, in the case of large-scale problems, AVI methods are used to
obtain suboptimal solutions (arising from say the use of function approximation
techniques). Showing that the stochastic approximation counterpart is bounded
almost surely can be hard in many reinforcement learning applications. Thus,
one of the most significant contributions of this paper, not addressed in previous
literature, lies in the development of easily verifiable sufficient conditions for the
almost sure boundedness of AVI methods involving set-valued dynamics.
23
8
Application: Finding fixed points of set-valued
maps
In Section 7 we showed that the stochastic iterative AVI given by (11) converges
to a vector that belongs to a small neighborhood of the optimal cost-to-go vector
J ∗ . To do this we started by observing that the fixed points of the perturbed
Bellman operator belong to a small neighborhood of J ∗ as a consequence of the
upper semicontinuity of attractor sets. Then we showed that (11) converges
to a fixed point of the perturbed Bellman operator, thereby showing that (11)
converges to a small neighborhood of J ∗ . In this section, we generalize the ideas
of Section 7 to develop and analyze a SAA for finding fixed points of contractive
set-valued maps.
Suppose that we are given a set-valued map T : Rd → {subsets of Rd }. We
present sufficient conditions under which the following stochastic approximation
algorithm is bounded a.s. and converges to a fixed point of T :
xn+1 = xn + a(n) [yn + Mn+1 ] ,
(20)
where
(ii) yn ∈ T xn − xn for all n ≥ 0.
(ii) {a(n)}n≥0 is the given step-size sequence satisfying (A2).
(iii) {Mn+1 }n≥0 is the Martingale difference noise sequence satisfying (A3a).
Definitions: Given a metric space (Rd , ρ), we define the Hausdorff metric with
respect to ρ as follows:
Hρ (A, B) := min ρ(x, B) ∨ min ρ(y, A) ,
x∈A
y∈B
where A, B ⊂ Rd and ρ(u, C) := min{ρ(u, v) | v ∈ C} for any u ∈ Rd and
C ⊆ Rd . We call a set-valued map T as contractive if and only if Hρ (T x, T y) ≤
αρ(x, y), where x, y ∈ Rd and 0 < α < 1. We say that T is of bounded diameter
if and only if diam(T x) ≤ D, where x ∈ Rd and 0 < D < ∞. We define
diam(A) := sup{ρ(z1 , z2 ) | z1 , z2 ∈ A} for any A ⊂ Rd .
We impose the following restrictions on (20):
(AF1) T is a Marchaud map that is of bounded diameter and contractive with
respect to some metric ρ.
(AF2) The metric ρ is such that kx − yk ≤ C ρ(x, y) for x, y ∈ Rd , C > 0.
(AF3) Let F := {x | x ∈ T x} denote the set of fixed points of T . There exists a
compact subset F ′ ⊆ F along with a strongly positive invariant bounded
open neighborhood.
OR
F is the unique global attractor of ẋ(t) ∈ T x(t) − x(t).
Since T is assumed to be contractive with respect to ρ, it follows from Theorem 5
24
of Nadler [14] that T has at least one fixed point. Assumption (AF 2) is readily
satisfied by the popular metric norms such as the weighted p-norms and the
weighted max-norms among others. Assumption (AF 3) is imposed to ensure
that (20) satisfies (A4b) or (A4c). Specifically, (AF 3) is imposed to ensure the
existence of an inward directing set associated with ẋ(t) ∈ T x(t) − x(t), see
Proposition 2 for details. In other words, we can find bounded open sets CF
and BF such that CF is inward directing and BF ⊂ CF .
As in Section 7, we compare (20) with it’s projective counterpart given by:
x̃n+1 = x̂n + a(n) (yn + Mn+1 ) ,
x̂n+1 ∈
⊓B
(21)
F ,CF
(x̃n+1 ),
where yn ∈ T x̂n − x̂n , {Mn+1 }n≥0 is identical for both (20) and (21) and
BF ,CF (· ) is the projection operator defined at the beginning of Section 4.
The analysis of the above projective scheme proceeds in an identical manner as
in Section 4. Specifically, we may show that every limit point of the projective
scheme (21) belongs to BF . The following theorem is immediate.
⊓
Theorem 5. Under (AF 1)-(AF 3) and (A3a), the iterates given by (20) are
bounded almost surely. Further, any limit point of (20) (as n → ∞) is a fixed
point of the set-valued map T .
Proof. PROOF: The proof of this theorem proceeds in a similar manner to that
of Theorem 3. We only provide an outline here to avoid repetition. We begin
by showing that (20) is bounded almost surely (stable) by comparing it to (21).
Since the limit points of (21) belong to BF , there exists N , possibly sample path
dependent, such that x̂n ∈ CF for all n ≥ N . For k ≥ 0 we have the following
inequality:
ρ(xn+k+1 , x̂n+k+1 ) ≤ (1−a(n+k))ρ(xn+k , x̂n+k )+a(n+k) (2D + Hρ (T xn+k , T x̂n+k )) ,
where diam(T x) ≤ D for every x ∈ Rd . Recall that 0 < α < 1 is the contraction
parameter of the set-valued map T . We consider two possible cases.
Case 1. 2D ≤ (1 − α)ρ(xn+k , x̂n+k ) : In this case, it can be shown that
ρ(xn+k+1 , x̂n+k+1 ) ≤ ρ(xn+k , x̂n+k ).
Case 2. 2D > (1 − α)ρ(xn+k , x̂n+k ) : In this case, it can be shown that
ρ(xn+k+1 , x̂n+k+1 ) ≤
2D
.
1−α
We conclude the following:
ρ(xn , x̂n ) ≤
2D
1−α
∨ ρ(xN , x̂N ), n ≥ N.
It follows from the above inequality and (AF 2) that (20) satisfies assumption
(A5). Hence, we get that {xn }n≥0 is bounded almost surely (stable).
25
Since the iterates are stable, it follows from [Theorem 2, Chapter 6, [2]]
that every limit point of (20) is an equilibrium point of the set-valued map
x 7→ T x − x. In other words, if x∗ is a limit point of (20), then 0 ∈ T x∗ − x∗ ,
i.e., x∗ ∈ T x∗ . Hence we have shown that every limit point of (20) is a fixed
point of the set-valued map T .
Remark 4. It is assumed that T is of bounded diameter, see (AF 1). The primary task of this assumption is in showing the almost sure boundedness of (20).
Specifically, it is used to show that (A5) is satisfied. Depending on the problem
at hand, one may wish to do away with this “bounded diameter” assumption.
For example, we may have sup diam(T xn ) < ∞ a.s., then the bounded diameter
n≥0
assumption can be dispensed with.
Since T is Marchaud, it is point-wise bounded, i.e., sup kzk ≤ K(1 + kxk),
z∈T x
where K > 0. In other words, diam(T x) ≤ 2K(1 + kxk). In theory, the pointwise boundedness of T allows for unbounded diameters, i.e., diam(T x) ↑ ∞ as
kxk ↑ ∞. Our bounded diameter assumption prevents this scenario from happening. In applications that use “approximate operators”, it is often reasonable
to assume that the errors (due to approximations) are bounded. Then the “associated set-valued map” is naturally of bounded diameter. The reader is referred
to Section 7 for an example of this setting.
9
A note on assumption (A5)
In this section, we investigate the following question: Under what conditions
is the stability of a SAA guaranteed, provided it is known a priori that the
corresponding projective scheme is convergent. In Lemma 2.1 of Abounadi et.
al. [1] this question is answered for simple iterative schemes. Here, we make an
investigation for the setting of SAAs with set-valued mean-fields. We prove the
following lemma.
Lemma 8. Let B and C be open bounded subsets of Rd such that B ⊂ C.
Consider the algorithm
xn+1 ∈ Gn (xn , ξn ), n ≥ 0.
We make the following assumptions:
1. {ξn }n≥0 is a random sequence that constitutes noise.
2. Gn is of bounded diameter and contractive (in the first co-ordinate with
second co-ordinate fixed) with respect to some metric ρ, for n ≥ 0. In
other words, Hρ (Gn (x, ξ), Gn (y, ξ)) ≤ αρ(x, y) for some 0 < α < 1 and
sup
ρ(u, v) ≤ D, where 0 < D < ∞ and x ∈ Rd .
u,v∈Gn (x,ξ)
26
3. There exists C > 0 such that kx − yk ≤ C ρ(x, y).
4. The sequence {x̃n }n≥0 generated by
x̃n+1 ∈ Gn
converges to some vector x∗ ∈ B.
⊓B,C (x̃n ), ξn
Then {xn }n≥0 is bounded almost surely.
Proof. PROOF: Since x̃n → x∗ as n → ∞, there exists N such that x̃n ∈ B for
all n ≥ N . For k ≥ 0 we have the following:
ρ(xn+k+1 , x̃n+k+1 ) ≤ 2D + ρ(xn+k , x̃n+k ).
The details of the above inequality are identical to the inequality given by (13).
Unfolding the right hand side down to stage n we get the following:
ρ(xn+k+1 , x̃n+k+1 ) ≤ 1 + α + . . . + αk 2D + ρ(xn , x̃n ),
ρ(xn+k+1 , x̃n+k+1 ) ≤
2D
+ ρ(xn , x̃n ).
1−α
In other words, we have
sup ρ(xn , x̃n ) ≤
n≥N +1
2D
+ ρ(xN , x̃N ).
1−α
Hence we have that {xn }n≥0 is bounded almost surely.
10
Conclusions
We now briefly provide a summary of our results, hitherto presented.
1. In this paper we developed three sets of sufficient conditions for the stability (almost sure boundedness) and convergence of stochastic approximation algorithms with set-valued mean-fields, under general conditions
on noise.
2. The Lyapunov function based assumptions presented here are general, easily verifiable and truly model-free. In previous Lyapunov function based
approaches, the verifiability of the assumptions suffers from the need to
explicitly construct the said Lyapunov function. However, here, we have
provided a recipe to explicitly construct the required Lyapunov function,
see Remarks 1 and 2 for details. Moreover our Lyapunov function based
stability conditions are the first in literature for stochastic approximations
with set-valued maps.
27
3. Our framework lends itself naturally to the analysis of the stochastic iterative counterpart of AVI methods. Such algorithms have become important
in recent years due to the proliferation of DNNs for function approximations and parameterizations. As a consequence of our framework, AVI
methods can be analyzed under significantly relaxed set of assumptions
than previous literature. We showed, in particular, that the stochastic iterative AVI is bounded almost surely and converges to a vector belonging
to a small neighborhood of the optimal J ∗ . An interesting consequence
of our analysis is the fact that the limiting set of these AVI methods are
the fixed points of the perturbed Bellman operator. It is worth noting
that our framework can be used to analyze approximate policy
gradient methods as well. These will pertain to the case where unlike [17] (that assumes complete knowledge of policy gradients), we only
have knowledge of approximate policy gradients, i.e., gradients with errors.
4. Finally we demonstrated the generality of our theory, by developing and
analyzing the first SAA for finding fixed points of contractive set-valued
maps. To the best of our knowledge there is no prior work in this direction,
and ours is the first such algorithm for finding fixed points of set-valued
maps.
References
[1] J. Abounadi, D.P. Bertsekas, and V. Borkar. Stochastic approximation for
nonexpansive maps: Application to q-learning algorithms. SIAM Journal
on Control and Optimization, 41(1):1–22, 2002.
[2] J. Aubin and A. Cellina. Differential Inclusions: Set-Valued Maps and
Viability Theory. Springer, 1984.
[3] M. Benaı̈m. A dynamical system approach to stochastic approximations.
SIAM J. Control Optim., 34(2):437–472, 1996.
[4] M. Benaı̈m and M. W. Hirsch. Asymptotic pseudotrajectories and chain
recurrent flows, with applications. J. Dynam. Differential Equations, 8:141–
176, 1996.
[5] M. Benaı̈m, J. Hofbauer, and S. Sorin. Stochastic approximations and
differential inclusions. SIAM Journal on Control and Optimization, pages
328–348, 2005.
[6] D.P. Bertsekas and J.N. Tsitsiklis. Neuro-Dynamic Programming. Athena
Scientific, 1st edition, 1996.
[7] P. Billingsley. Convergence of probability measures. John Wiley & Sons,
2013.
[8] V. S. Borkar. Stochastic approximation with two time scales. Syst. Control
Lett., 29(5):291–294, 1997.
28
[9] V. S. Borkar. Stochastic Approximation: A Dynamical Systems Viewpoint.
Cambridge University Press, 2008.
[10] V. S. Borkar and S.P. Meyn. The O.D.E. method for convergence of stochastic approximation and reinforcement learning. SIAM J. Control Optim,
38:447–469, 1999.
[11] D.P. De Farias and B. Van Roy. On the existence of fixed points for approximate value iteration and temporal-difference learning. Journal of Optimization theory and Applications, 105(3):589–608, 2000.
[12] T. Jaakkola, M. I. Jordan, and S.P. Singh. Convergence of stochastic iterative dynamic programming algorithms. In Advances in neural information
processing systems, pages 703–710, 1994.
[13] R. Munos. Error bounds for approximate value iteration. In Proceedings of
the National Conference on Artificial Intelligence, volume 20, page 1006,
2005.
[14] S. Nadler. Multi-valued contraction mappings. Pacific Journal of Mathematics, 30(2):475–488, 1969.
[15] A. Ramaswamy and S. Bhatnagar. A generalization of the Borkar-Meyn
theorem for stochastic recursive inclusions. Mathematics of Operations Research, 2016.
[16] Herbert Robbins and Sutton Monro. A stochastic approximation method.
The annals of mathematical statistics, pages 400–407, 1951.
[17] R.S. Sutton, D. A. McAllester, S.P. Singh, and Y. Mansour. Policy gradient methods for reinforcement learning with function approximation. In
Advances in neural information processing systems, pages 1057–1063, 2000.
29
| 3cs.SY
|
1
Sensor Selection for Target Tracking in
Wireless Sensor Networks with Uncertainty
Nianxia Cao, Student Member, IEEE, Sora Choi,
arXiv:1510.01993v1 [cs.SY] 7 Oct 2015
Engin Masazade, Member, IEEE, Pramod K. Varshney, Fellow, IEEE
Abstract
In this paper, we propose a multiobjective optimization framework for the sensor selection problem
in uncertain Wireless Sensor Networks (WSNs). The uncertainties of the WSNs result in a set of sensor
observations with insufficient information about the target. We propose a novel mutual information
upper bound (MIUB) based sensor selection scheme, which has low computational complexity, same
as the Fisher information (FI) based sensor selection scheme, and gives estimation performance similar
to the mutual information (MI) based sensor selection scheme. Without knowing the number of sensors
to be selected a priori, the multiobjective optimization problem (MOP) gives a set of sensor selection
strategies that reveal different trade-offs between two conflicting objectives: minimization of the number
of selected sensors and minimization of the gap between the performance metric (MIUB and FI) when
all the sensors transmit measurements and when only the selected sensors transmit their measurements
based on the sensor selection strategy. Illustrative numerical results that provide valuable insights are
presented.
Index Terms
Target tracking, sensor selection, Fisher information, mutual information, information fusion,
multiobjective optimization, wireless sensor networks.
N. Cao, S. Choi, and P. K. Varshney are with the Department of Electrical Engineering and Computer Science, Syracuse
University, Syracuse, NY, 13244, USA, email: {ncao, schoi101, varshney}@syr.edu
E. Masazade is with Department of Electrical and Electronics Engineering, Yeditepe University, Istanbul, 34755, Turkey,
email: [email protected]
The work of N. Cao, S. Choi, and P. K. Varshney was supported by U.S. Air Force Office of Scientific Research (AFOSR)
under Grant No. FA9550-10-1-0458. The work of E. Masazade was supported by the Scientific and Technological Research
Council of Turkey (TUBITAK) under Grant 113E220. A preliminary version of this paper appears in the IEEE International
Conference on Information Fusion 2013.
October 8, 2015
DRAFT
2
I. I NTRODUCTION
A wireless sensor network (WSN) is composed of a large number of densely deployed sensors,
where sensors are battery-powered devices with limited signal processing capabilities. When
programmed and networked properly, WSNs are very useful in many application areas including
battlefield surveillance [1], environment monitoring and target tracking [2], industrial processes
[3] and health monitoring and control [4]. In our work presented in this paper, the task of the
WSN is to track a target emitting or reflecting energy in a given region of interest (ROI), and
the sensors send their observations regarding the target to a central node called the fusion center,
which is responsible for the final inference.
Target tracking problems often require coverage of broad areas and a large number of sensors
that can be densely deployed over the ROI. This results in new challenges when the resources
(bandwidth and energy) are limited. In such situations, it is inefficient to utilize all the sensors in
the ROI including the uninformative ones, which hardly contribute to the tracking task at hand
but still consume resources. This issue has been investigated and addressed via the development
of sensor selection schemes, whose goal is to select the best non-redundant set of sensors for
the tracking task while satisfying some performance and/or resource constraints [5]. The sensor
selection problem for target localization and target tracking has been considered in [6]–[16]
among others, where the sensor sets are selected to get the desired information gain or reduction
in estimation error about the target state. In [6]–[9], the mutual information (MI) or entropy
is considered as the performance metric, and in [10], [11], the sensors that have the lowest
posterior Cramer-Rao lower bound (PCRLB), which is the inverse of the Fisher information
(FI), are selected. In [12], the authors compared the two sensor selection criteria namely MI and
PCRLB for the sensor selection problem based on quantized data, and showed that the PCRLB
based sensor selection scheme achieves similar mean square error (MSE) with significantly
less computational effort. In [13], the sensor selection problem was formulated as an integer
programming problem, which was relaxed and solved through convex optimization. In [14], a
multi-step sensor selection strategy by reformulating the Kalman filter was proposed, which
was able to address different performance metrics and constraints on available resources. In
[15], the authors aimed to find the optimal sparse collaboration topologies subject to a certain
information or energy constraint in the context of distributed estimation. For a more complete
October 8, 2015
DRAFT
3
literature review on sensor management for target tracking, see [16] and references therein.
The previous research on sensor selection assumes that the WSNs operate reliably during the
target tracking process without any interruptions. The fact is that, in some situations, the sensor
observations are quite uncertain [17]–[21]. For example, sensors may have temporary failure,
there may be abrupt changes in the operating environment [19], [20], or other interference
such as traffic or birds/animals that may change the power received by the sensors. Moreover,
some random interruptions may appear over the communication channels in the system, and
adversaries may jam wireless communications using different attack strategies [21]. These types
of uncertainties would result in the set of sensor observations with insufficient information about
the target at the fusion center. In other words, in such an uncertain WSN, sensor observations may
contain useful information regarding the target only with a certain probability. It is important to
investigate the sensor selection problem in such an uncertain environment. In our work here, we
study the uncertainty caused by occlusions, i.e., the sensors may not be able to observe the target
when blocked by some obstacles. Regarding the representation of this type of uncertainty, the
authors in [17] and [18] introduced a stochastic model for sensor measurements. Furthermore,
the work in [19] and [20] generalized the model in [17], [18] to multiple sensors by considering
a more realistic viewpoint in that the sensors have different uncertainty at different time instants.
For the problems involving uncertain WSNs, even though there are studies about the Kalman
filter for target tracking [18], [22]–[24], and about the target localization problem with non-ideal
channels [25], [26], the sensor selection problem in WSNs with uncertain sensor observations
has not been considered in the literature and is the subject of this paper.
In the aforementioned literature, sensor selection schemes require a priori information about
the number of sensors to be selected at each time, denoted as A, and computationally efficient
algorithms are developed in order to find the optimal A sensors that achieve the maximum
performance gain. Realistically, in many applications like target tracking, it is unlikely that the
number of sensors that need to be selected at each time step of tracking is known to the system
designer before operation begins. Therefore, it is quite necessary and important to investigate
sensor selection strategies that determine the optimal number of sensors to be selected as well
as which sensors to select based on the WSN conditions.
Sensor network design usually involves consideration of multiple conflicting objectives, such
as maximization of the lifetime of the network or the inference performance, while minimizing
October 8, 2015
DRAFT
4
the cost of resources such as energy, communication or deployment costs [27]–[30]. The
problems that investigate the trade-offs among such conflicting objective functions are called
Multiobjective Optimization Problems (MOPs). In our preliminary work [31], a sensor selection
method utilizing FI as the performance metric in an MOP framework has been presented under
the assumption that the sensors in the WSN are all reliable. Our work in [31] optimized two
objectives simultaneously: minimization of the total number of sensors selected at each time,
and minimization of the information gap between the FI when all the sensors transmit their
measurements and the FI when only the selected sensors transmit their measurements. In this
work, we investigate the sensor selection problem in an uncertain WSN, and generalize the
approach presented in [31] by addressing the issues that arise due to uncertainty. As we will see
in the paper, the FI based selection scheme (FISS) tends to select sensors which are relatively
close to the target, while the MI based selection scheme (MISS) selects sensors that have high
sensing probabilities, and achieves better performance. The better performance of MISS comes
along with high computational complexity. Thus, we propose to use a mutual information upper
bound (MIUB) as the performance metric for the sensor selection problem. The complexity of
computing MIUB is similar to that of evaluating FI, and is much lower than that of computing
MI. We also show through simulation experiments that the MIUB based selection scheme
(MIUBSS) hardly degrades the tracking performance. Furthermore, we consider our sensor
selection problem with uncertainty under the MOP framework, where the Nondominating Sorting
Genetic Algorithm-II (NSGA-II) is applied to dynamically select an optimal set of sensors at
each time step. Numerical results show that MIUBSS selects more sensors than FISS under the
MOP framework. We also compare our framework with some other sensor selection methods,
e.g., weighted sum method and convex optimization method, and show that NSGA-II with the
compromise solution (to be discussed later in the paper) adaptively decides the optimal number
of sensors at each time step of tracking and achieves satisfactory estimation performance while
obtaining savings in terms of number of sensors.
The rest of this paper is organized as follows: In Section II, we introduce the uncertain
WSN system model. The target tracking framework using a particle filter is given in Section
III. In Section IV, the performance metric, FI and MI, for sensor selection are introduced, and
comparisons are performed through numerical experiments. Then in Section V, we review the
fundamentals of MOP and apply NSGA-II to solve the proposed MOP; we also investigate the
October 8, 2015
DRAFT
5
performance of the MOP framework through simulations in this section. Section VI is devoted
to our conclusions and future research directions.
II. S YSTEM
MODEL
We consider a target tracking problem, where a moving target emitting (or reflecting) a signal
over an area of interest is tracked by a WSN consisting of N sensors. The target state is assumed
to be a 4-dimensional vector xt = [xt , yt , ẋt , ẏt ] where xt and yt are the target positions, and
ẋt and ẏt are the target velocities in the horizontal and vertical directions. Even though the
approaches developed in this paper are applicable to more complex dynamic models, here we
assume a linear dynamic model
xt+1 = Fxt + wt .
(1)
where F is the state transition matrix and wt is the Gaussian process noise with zero mean and
covariance matrix Q:
D3
3
0
D2
2
0
0 1 0 D
0
F=
,Q = q 2
0 0 1 0
D
2
0
0 0 0 1
D3
3
0
D2
2
0
D
1 0 D
0
D2
2
0
,
0
D
(2)
where D is the sampling interval and q is the process noise parameter. It is assumed that the
signal emitted by the target follows a power attenuation model [25]. Thus, the signal power
received by sensor i which is located at (xi , yi ) is
Pi,t (xt ) =
P0
1 + αdni,t
(3)
where P0 is the emitted signal power from the target at distance zero, n is the signal decay
exponent and α is a scaling parameter. In (3), di,t is the distance between the target and the ith
p
sensor at time step t, i.e., di,t = (xt − xi )2 + (yt − yi )2 .
A. Uncertainty Model of Sensor Observations
As discussed earlier, sensor observations may be uncertain due to sensor failures, natural
interference or some random interruptions. Regarding different uncertainties, there are different
probabilistic models [32]. In this paper, we consider the scenario that the sensor observation
October 8, 2015
DRAFT
6
uncertainty is caused by some obstacles, and assume the following probabilistic measurement
model, which has been proposed in [17] and generalized in [19] and [20]: the sensor observation
is assumed to contain only noise if the sensor cannot sense the target due to obstacles, and since
such uncertainty may happen at any time for any sensor, the sensing probability may not be
identical across the sensors in the WSN, i.e.,
h (x ) + v , with probability p(i)
s
i,t
t
i,t
zi,t =
(i)
vi,t ,
with probability 1 − ps
(i)
where ps is the sensing probability of sensor i, hi,t (xt ) =
p
(4)
Pi,t (xt ) represents the signal
amplitude received by sensor i at time step t, and vi,t is the measurement noise, which is
assumed to be independent across time steps and across sensors, follows a Gaussian distribution
with parameters N (0, σ 2). The likelihood function for sensor measurements zt = [zi,t , . . . , zN,t ]T
given the target state xt is simply the product of each sensor i’s likelihood function. Given xt , zi,t
(i)
follows the Gaussian distribution N (hi,t (xt ), σ 2 ) with probability ps , and follows the Gaussian
(i)
distribution N (0, σ 2 ) with probability 1 − ps , i.e.,
2
p(zi,t |xt ) = ps(i) N (hi,t (xt ), σ 2 ) + (1 − p(i)
s ) N (0, σ )
(5)
For communication between the fusion center and the sensors, we consider the following
two practical scenarios: 1) the sensors directly send their analog measurements zt to the fusion
center; and, 2) the sensors quantize their analog measurements to M bits, and then transmit the
quantized data to the fusion center for tracking. Analog sensor measurements contain complete
information about the observation, at the expense of high communication cost; on the other
hand, quantized measurements save communication burden, but lose some information about the
target.
The quantized measurement of sensor i at time step t, Di,t , is defined as:
Di,t =
0
η0 ≤ zi,t ≤ η1
1
..
.
η1 ≤ zi,t ≤ η2
L−1
(6)
η(L−1) ≤ zi,t ≤ ηL
where η = [η0 , η1 , . . . , ηL ]T is the set of quantization thresholds with η0 = −∞ and ηL = ∞
October 8, 2015
DRAFT
7
Algorithm 1 SIR Particle Filter for target tracking
1: Set t = 1. Generate initial particles xs0 ∼ p(x0 ) with ∀s , w0s = Ns−1 .
2:
3:
4:
while t ≤ Ts do
xst = Fxst−1 + υ t (Propagating particles)
P s
s
p(xt |z1:t ) = N1s N
s=1 δ(xt − xt )
5:
Obtain sensor data zt
6:
wts ∝ p(zt |xst ) (Updating weights through obtained data)
7:
wts =
ws
PNst
Ps=1
Ns
wts
(Normalizing weights)
wts xst
8:
x̂t =
9:
{xst , Ns−1 } = Resampling(xst , wts )
10:
11:
s=1
t=t+1
end while
and L = 2M is the number of quantization levels. The probability that Di,t takes the value l is
p(Di,t = l|xt ) = Pr(ηl ≤ zi,t ≤ ηl+1 |xt )
= ps(i) Pr(ηl ≤ zi,t ≤ ηl+1 |zi,t ∼ N (hi,t(xt ), σ 2 ))
+ (1 − ps(i) ) Pr(ηl ≤ zi,t ≤ ηl+1 |zi,t ∼ N (0, σ 2))
ηl+1 − hi,t (xt )
ηl − hi,t (xt )
(i)
−Q
= ps Q
σ
σ
η
ηl
l+1
+ (1 − ps(i) ) Q
−Q
σ
σ
(7)
where Q(·) denotes the complementary distribution of the standard Gaussian distribution with
zero mean and unit variance
Q(x) =
Z
x
∞
1
y2
√ exp{− }dy
2
2π
(8)
Since the sensor measurements are conditionally independent, the likelihood function of Dt =
[D1,t , D2,t , ..., DN,t ]T can be written as the product of each sensor i’s likelihood function.
III. PARTICLE F ILTERING
FOR
TARGET T RACKING
The target tracking problem requires the estimation of the target state using a sequence of
sensor measurements. For nonlinear systems, extended Kalman filter (EKF) provides suboptimal
October 8, 2015
DRAFT
8
solutions. However, when the sensor measurements are quantized, even for linear and Gaussian
systems, the EKF fails to provide an acceptable performance especially when the number of
quantization levels is small [33]. Thus, we employ a sequential importance resampling (SIR)
particle filter to solve our nonlinear target tracking problem with analog and quantized sensor
measurements [34], [35]. The SIR algorithm is based on the Monte Carlo method, and can
be used for recursive Bayesian filtering problems under very weak assumptions [35]. The main
idea of the particle filter is to find a discrete representation of the posterior distribution p(xt |z1:t )
(p(xt |D1:t )) by using a set of particles xst with associated weights wts ,
p(xt |z1:t ) ≈
Ns
X
s=1
wts δ(xt − xst ),
(9)
where, δ(·) is the Dirac delta measure, and Ns denotes the total number of particles. When
the number of particles is large enough, the weighted sum of the particles based on the Monte
Carlo characterization will be an equivalent representation of the posterior distribution. The
resampling step in the SIR particle filter avoids the situation that all but one of the importance
weights are close to zero after a few iterations, which is known as the degeneracy phenomenon
in the particle filter. Algorithm 1 provides a summary of the SIR particle filtering algorithm for
the target tracking problem with analog data zt , where Ts denotes the number of time steps over
which the target is tracked, and zt is replaced by Dt if quantized data is utilized for transmission.
IV. S ENSOR S ELECTION C RITERIA
FOR
U NCERTAIN WSN S
In this section, we present and investigate three performance metrics, FI, MI, and MIUB,
for the sensor selection problem in an uncertain WSN. After formulating the three performance
metrics mathematically for the analog data and quantized data respectively, we compare them
with respect to the resulting tracking performance.
A. Fisher Information
Posterior Cramer-Rao Lower Bound (PCRLB) provides the theoretical performance limit for a
Bayesian estimator [36]. Let p(zt , xt ) denote the joint probability density function of the sensor
measurements and the target state, and let x̂t denote the estimate of xt . The PCRLB on the
estimation error is represented as [36],
E [x̂t − xt ][x̂t − xt ]T
October 8, 2015
≥ Jt−1
(10)
DRAFT
9
where Jt is the Fisher information (FI) matrix. It has been shown in [37] that, the FI matrix for
Bayesian estimation is composed of two parts: the FI obtained from the sensor measurements
and the FI corresponding to a priori information. Furthermore, under the assumption that the
sensor measurements are conditionally independent given the target state xt , the FI obtained
from the measurements of multiple sensors can be written as the summation of each sensor’s FI
plus the FI from the prior information,
Jt ,
N Z
X
i=1
S
Ji,t
(xt )p(xt )dxt + JtP
(11)
xt
S
where JtP is the FI matrix of the a priori information, and Ji,t
(xt ) represents the standard FI of
each sensor as a function of the target state xt ,
S
Ji,t
(xt )
=
Z
zi,t
∂p(zi,t |xt )
p(zi,t |xt )
∂xt
1
2
dzi,t
(12)
1) Fisher information for the analog sensor measurement model: The FI for analog data is
obtained by substituting the likelihood function p(zi,t |xt ) given in (5) into (12). The derivative
of p(zi,t |xt ) is
∂p(zi,t |xt )
∂xt
(zi,t − h(xt ))2 ∂hi,t (xt )
(i) zi,t − hi,t (xt )
√
= ps
exp −
2σ 2
∂xt
σ 2 2πσ 2
(13)
where
n−2
∂hi,t (xt )
αn hi,t (xt )di,t
=
∂xt
2 1 + αdni,t
October 8, 2015
xt − xi
y −y
i
t
0
0
(14)
DRAFT
10
SA
Substituting (5), (13), and (14) into (12) and letting Ji,t
(xt ) denote the standard FI matrix for
SA
analog data, Ji,t
(xt ) is obtained as follows:
SA
Ji,t
(xt )
=
2
(p(i)
s )
κA
i,t (xt )
∂hi,t (xt )
∂xt
∂hi,t (xt )
∂xt
T
(i)
=
κA
i,t (xt )
2n−4
(ps )2 α2 n2 h2i,t (xt )di,t
4(1 + αdni,t)2
(xi − xt )2
(x − x )(y − y )
t
i
t
i
×
0
0
(xi − xt )(yi − yt ) 0 0
(yi − yt )2
0
0
(15)
0 0
0 0
0 0
where
κA
i,t (xt ) =
Z
n (z − h(x ))2 o2
1
zi,t − hi,t (xt )
i,t
t
√
exp −
dzi,t
p(zi,t |xt ) σ 2 2πσ 2
2σ 2
(16)
zi,t
2) Fisher information for the quantized sensor measurement model: The FI of quantized data
is calculated by replacing the likelihood function p(zi,t |xt ) given in (12) with p(Di,t |xt ) in (7).
Since the derivative of the likelihood function of the quantized observations is
(i)
ps
n (η − h (x ))2 o
∂p(Di,t |xt )
l
i,t
t
= √ exp −
2
∂xt
2σ
σ 2π
n (η − h (x ))2 o ∂h (x )
i,t t
l+1
i,t
t
− exp −
2σ 2
∂xt
October 8, 2015
(17)
DRAFT
11
we derive the FI for quantized data by substituting (14) into (17) as follows:
SQ
Ji,t
(xt )
=
X
Di,t
∂p(Di,t |xt )
p(Di,t |xt )
∂xt
1
κQ
i,t (xt )
∂hi,t (xt )
∂xt
∂hi,t (xt )
∂xt
=
2
(p(i)
s )
=
(i)
2n−4
(ps )2 α2 n2 h2i,t (xt )di,t
Q
κi,t (xt )
4(1 + αdni,t)2
(xi − xt )2
(x − x )(y − y )
t
i
t
i
×
0
0
2
T
(18)
(xi − xt )(yi − yt ) 0 0
(yi − yt )2
0
0
0 0
0 0
0 0
where
κQ
i,t (xt )
=
X
Di,t
n (η − h (x ))2 o
1
l
i,t
t
exp
−
2πσ 2 p(Di,t |xt )
2σ 2
− exp
n
2
(ηl+1 − hi,t (xt ))2 o
−
2σ 2
(19)
Thus, we get the FI for the analog observation model in (15), and for the quantized observation
model in (18).
B. Mutual Information
Information-theoretic sensor management for target tracking seeks to minimize the uncertainty
in the estimate of the target state conditioned on the sensor measurements [38]. Entropy, which
is defined by Shannon [39], represents the uncertainty or randomness in the estimate of the target
state xt . Moreover, because of the relationship between the entropy and the MI [40], the sensor
selection problem for target tracking can be solved by maximizing the MI between the target
state and the sensor measurements.
Given the distribution of the target state and the likelihood function of the sensor measureOctober 8, 2015
DRAFT
12
ments, the MI for the analog data can be written as [8], [12]
I(xt , zt ) = H(zt ) − H(zt |xt )
Z
Z
=−
p(zt |xt )p(xt )dxt
zt
xt
Z
p(zt |xt )p(xt )dxt dzt
log2
xt
+
N Z
X
i=1
xt
Z
zi,t
(20)
p(zi,t |xt ) log2 p(zi,t |xt )dzi,t p(xt )dxt
where H(zt ) is the entropy of the sensor measurements zt , and H(zt|xt ) is the conditional
entropy of the sensor measurements zt given the target state xt . Similarly, the MI for the quantized
sensor measurements can be written as
I(xt , Dt ) = H(Dt ) − H(Dt |xt )
Z
X
p(Dt |xt )p(xt )dxt
=−
Dt xt
Z
log2
p(Dt |xt )p(xt )dxt
xt
+
N Z
X
i=1
xt
(21)
X
p(Di,t |xt ) log2 p(Di,t |xt )p(xt )dxt
Di,t
where the summation over Dt is taken over all possible combinations of the quantized
measurements of the set of sensors.
C. Mutual Information Upper Bound (MIUB)
The computational complexity of evaluating the MI for a set of A sensors increases
exponentially with the number of sensors A, so that it becomes impractical to compute the
MI in (20) and (21) when the number of sensors to be selected is large [8] [41]. The chain rule
October 8, 2015
DRAFT
13
25
20
15
31 0.93
32 0.92
33 0.97
34 0.57
35 0.99
36 0.19
Sensor
Target Track
25 0.94
26 0.92
27 0.95
28 0.98
29 0.15
30 0.66
19 0.90
20 0.93
21 0.96
22 0.16
23 0.68
24 0.52
13 0.50
14 0.98
15 0.13
16 0.64
17 0.91
18 0.93
9 0.65
10 0.96
11 0.97
12 0.99
3 0.52
4 0.97
5 0.94
6 0.96
10
5
0
−5
t=8
−10
−15
7 0.96
t=3
−20
−25
8 0.21
1 0.14 t=1 2 0.66
−20
−10
0
10
20
Fig. 1. WSN with 36 unreliable sensors. Numbers above the stars indicate sensor index (left) and its sensing probability (right).
for the MI is described as follows (we only show the MI for analog data, results for quantized
data are similar):
I(zt ; xt ) =
N
X
i=1
I(zi,t ; xt |zi−1,t , · · · , z1,t )
(22)
Since z1,t , · · · , zN,t are conditionally independent given the target state xt , zi,t → xt → zj,t
(i, j ∈ {1, · · · , N}) form a Markov chain, and we have the following data processing inequality
[40]:
I(zi,t ; xt |zi−1,t , · · · , z1,t ) ≤ I(zi,t ; xt |zi−1,t , · · · , z2,t )
(23)
· · · ≤ I(zi,t ; xt )
Thus,
PN
i=1
I(zi,t ; xt ) is an upper bound on I(zt ; xt ). We use this mutual information upper
bound (MIUB) as the performance metric for our sensor selection problem. It can be easily
shown that the computational complexity of evaluating MIUB for selecting A out N sensors
increases linearly with A, which is the same with that of computing FI.
D. Comparison of Performance Metrics for Sensor Selection by Numerical Experiments
In this subsection, we compare the performance of the above three performance metrics, FI,
MI, and MIUB, for the sensor selection problem through some numerical experiments.
October 8, 2015
DRAFT
14
a) Simulation setting: In our simulations, we consider the WSN shown in Fig. 1, which
has N = 6 × 6 = 36 sensors deployed in the ROI of area b2 = 50 × 50 m2 . In the current
work, we assume that the sensing probabilities of the sensors are already known to the fusion
center, research on how to learn the probabilities iteratively is an interesting problem and can be
considered in the future. Generally, if the sensors around the target tracks have higher sensing
probabilities compared to other sensors in the WSN, it is highly likely that the algorithm will
select those sensors owing to both higher signal power and sensing probability. Our interest is in
considering more challenging cases to test the performance of our algorithm. Thus, we assume
that the sensors around the target track have relatively low sensing probabilities as shown in
the figure. Moreover, the sensing probabilities may be identical for some sensors if they are in
the same environment, however, if the sensors have the same sensing probability, the selection
results would be similar to our preliminary work [31]. Thus, we consider the scenario in which
the sensors in the WSN all have different sensing probabilities.
For the linear dynamical model of the target given in (1), the time interval is D = 1.25 seconds
and the process noise parameter q = 2.5×10−3 . The source power is P0 = 1000 and the variance
of the measurement noise is selected as σ = 0.2. The sensors quantize their observations to M
bits for quantized data, and the quantization thresholds [η1 , · · · , ηL−1 ] are selected to be the
√
values which evenly partition the interval [−σ, σ + P0 ]. The prior distribution about the state
of the target, p(x0 ), is assumed to be Gaussian with mean µ0 = [−23 −24 2 2]T and covariance
Σ0 = diag[σx2 σx2 0.01 0.01] where we select σx = 6. The initial Ns = 5000 particles are drawn
from p(x0 ). The mean square error (MSE) is used to measure errors between the ground truth
and the estimates, and the MSE of the estimation at each time step of tracking is averaged over
Ttotal trials as,
MSEt =
1
Ttotal
TX
total
tr=1
tr
2
tr
tr
2
(x̂tr
t (1) − xt (1)) + (x̂t (2) − xt (2))
(24)
th
where x̂tr
trial.
t and xt are the estimated and the actual target states at time t of the tr
b) Sensors with highest MI or FI at different time steps: We first consider analog and two
quantization communication schemes (M = 5 and 2) for one Monte Carlo run. The sensors
with highest MI or FI are listed in Table I. Note that, 1) since the FI in our paper is a matrix,
we consider the determinant of the FI matrix, which corresponds to the area of the uncertainty
ellipsoid [42]; 2) we are interested in the effect of the sensors’ distances from the target and the
October 8, 2015
DRAFT
15
TABLE I
S ENSORS WITH THE MOST SIGNIFICANT MI OR FI AT
DIFFERENT TIME STEPS
Quantized data
Quantized data
M =5
M=2
Sensor 2,7
Sensor 2,7
Sensor 2,7
t = 1 FI
Sensor 2,7
Sensor 2,7
Sensor 2,7
t = 3 MI
Sensor 8
Sensor 8
Sensor 2,7,14
t = 3 FI
Sensor 8
Sensor 8
Sensor 2,7,9
t = 8 MI
Sensor 16
Sensor 16
Sensor 10,16,21
t = 8 FI
Sensor 15
Sensor 15
Sensor 10,16
Time step
Analog data
t = 1 MI
sensing probabilities on the performance metrics, thus we compute the performance metric for
each sensor instead of focusing on different sets of multiple sensors; 3) for individual sensors,
the MI and MIUB are identical.
Generally, quantized data contains less information compared with the analog data. We first
discuss the results for analog data and 5-bit quantized data. We observe from Table I that the
sensors with highest MI or FI are identical for Analog data and 5-bit quantized data, which means
that 5-bit quantization preserves most information of the analog data as far as sensor selection
is concerned. Additionally, we investigate three distinct time steps to compare the results:
•
At time step 1, the target is relatively close to sensors 2 and 7 with a similar distance from
the target, so that sensors 2 and 7 have the most significant MI and FI. Sensors 1 and 8 have
very low sensing probabilities though they have similar distance to the target as sensors 2
and 7, and therefore have low MI and FI.
•
At time step 3, the target is much closer to sensor 8 than the other sensors, so that sensor
8 has the highest MI and FI even though it has a low sensing probability.
•
At time step 8, sensor 15 is the closest one to the target with very low sensing probability,
and sensor 16 is the second closest with higher sensing probability. In this case, sensor 15
has the highest FI while sensor 16 has the highest MI.
The 2-bit quantized data contains much less information about the target compared to the analog
data and the 5-bit quantized data, so that the sensing probability of sensors affects the FI and MI
more with the 2-bit quantized data. Thus, the sensors with relatively higher sensing probabilities
October 8, 2015
DRAFT
16
20
16
14
90
Percentage of reliable sensors (%)
18
MSE
12
10
8
6
4
80
70
60
50
40
30
20
2
0
2−bit FI
2−bit MI
5−bit MI
Analog−MI
5−bit FI
Analog−FI
100
2−bit FI
2−bit MI
5−bit FI
5−bit MI
Analog−FI
Analog−MI
0
5
10
Time step t
15
20
10
0
(a)
Fig. 2.
5
10
Time Step t
15
20
(b)
Target tracking performance with analog data, 5-bit quantized data, and 2-bit quantized data, (a) MSE performance;
(b) average number of reliable sensors selected.
have higher FI and MI than the other sensors for the 2-bit quantized data case as shown in Table
I.
Therefore, we conclude that for analog data or quantized data with a large number of
quantization levels, MI is more affected by the sensing probabilities of the sensors than FI;
for quantized data with small number of quantization levels, both MI and FI are considerably
affected by the sensing probabilities. Moreover, FISS tends to select sensors which are closer
to the target compared to MISS, which can be explained from Equation (15) and (18) with
the corresponding parameters, i.e., the distance between the target and the sensors dominates FI.
However, such an explanation cannot be found for MI. In other words, the sensor’s distance from
the target, sensing probability, and the number of quantization levels are all important factors
that determine the tracking performance of the WSNs.
c) Tracking performance: In Fig. 2, we show the performance of the WSN given in Fig.
1 when only one sensor is selected at each time step over Ttotal = 500 Monte Carlo runs.
Fig. 2(a) shows that MISS has better MSE performance than FISS with both analog data and
5-bit quantized data. We explain the result by investigating the percentage of reliable sensors
(the fusion center treats a sensor as unreliable if it’s amplitude is quite close to noise1 ) among
1
In the experiments, we check if it is within the region [−3σ, 3σ].
October 8, 2015
DRAFT
17
1
10
MSE
MI
MIUB
0
10
−1
Percentage of reliable sensors (%)
10
Fig. 3.
0
5
10
15
20
10
Time Step t
15
20
90
MI
MIUB
80
70
60
50
40
0
5
Target tracking performance for MI and MIUB, A = 2.
the selected ones over 500 Monte Carlo trials in Fig. 2(b). We observe that, in 500 Monte
Carlo trials, around 60% of the sensors selected by MISS are reliable, and only around 40%
of the sensors selected by FISS are reliable, which explains the better estimation performance
of MISS. Although the sensor selection scheme with 2-bit quantized data selects even more
reliable sensors, there is no improvement with respect to the MSE performance because of the
significant information loss in the quantization process. As is shown in Fig. 2(a), the sensor
selection scheme based on analog data has the best tracking performance; 5-bit quantized data
based sensor selection scheme achieves performance that is close to that with the analog data;
and 2-bit quantized data based sensor selection scheme performs much worse. We only show
simulation results for the 5-bit quantized data in the following simulation experiments.
d) Performance of MIUBSS: The complexity of computing MIUB for selecting A out of
N sensors is the same as that of computing FI (both increase linearly with A), and is much less
than that of evaluating the MI (increases exponentially with A). Fig. 3 shows the results of MISS
and MIUBSS when A = 2 sensors are selected, and we observe similar performance for MISS
and MIUBSS in terms of both the percentage of reliable sensors selected by the schemes and
the MSE performance. In other words, MIUBSS obtains performance similar to MISS but with
October 8, 2015
DRAFT
18
much lower computational complexity. Thus, in the next section, we utilize MIUBSS, instead
of MISS, in the multiobjective optimization framework, and compare it with FISS.
V. M ULTIOBJECTIVE O PTIMIZATION
BASED
S ENSOR S ELECTION
In this section, we utilize the MOP framework to find the sensor selection strategy that can
determine the optimal sensor set.
The mathematical description of an n-objective optimization problem is given as
{f1 (α), f2 (α), . . . , fn (α)}
min
α
(25)
subject to a ≤ αi ≤ b, h(α) = 0, g(α) ≤ 0
where α is the vector of decision variables with elements αi , a and b define the bounds on
decision variables, functions h(.) and g(.) represent the equality and inequality constraints of
the problem respectively. For the MOP, the solutions satisfying the constraints of (25) form the
feasible set C. In an optimization problem involving the minimization of all the objectives, the
solution α1 dominates the solution α2 (α1 ≻ α2 ) if and only if
fu (α1 ) ≤ fu (α2 ) ∀u ∈ {1, 2, . . . n}
1
∗
(26)
2
fv (α ) < fv (α ) ∃v ∈ {1, 2, . . . n}
α is called a Pareto optimal solution if and only if there is no α in C that dominates α∗ ,
and the set of Pareto optimal outcomes is called the Pareto front. A well-known technique for
solving MOPs is to minimize a weighted sum of the objectives, which yields a single solution
corresponding to the weights used. With this approach, if a uniform spread of weights is employed
to obtain different solutions, it rarely produces a uniform spread of points on the Pareto front.
Some of the optimal solutions may become closely spaced and hence reducing the number of
design alternatives [43].
In our work, the sensor selection strategies reflect different trade-offs between two objective
functions: the estimation performance and the number of selected sensors, which are dependent
on the binary decision variables.
A. Objective Functions based on Fisher Information (FI) and Mutual Information Upper Bound
(MIUB)
1) FI based objective functions: Let αt = [α1,t , . . . , αN,t ] be the sensor selection strategy at
time step t. The elements of αt are binary variables, i.e, αi,t = 1, if sensor i is selected and
October 8, 2015
DRAFT
19
αi,t = 0 otherwise. Then, A =
PN
i=1
αi,t is the number of sensors selected at time step t. Based
on the sensor selection strategy αt , the FI matrix at time step t can be written as,
Jt (αt ) =
N
X
D
αi,t Ji,t
+ JtP
(27)
i=1
We determine the sensor selection strategy from the solution of the MOP where the objective
functions are: minimization of the information gap between the FI based on the measurements
of all the sensors and the FI based on the sensor set selected by strategy αt ,
f1 (αt ) =
P
P
N
N
D
P
D
P
α
J
+
J
J
+
J
−
log
det
log det
t
t
i=1 i,t i,t
i=1 i,t
P
N
D
P
log det
i=1 Ji,t + Jt
(28)
and minimization of the normalized number of selected sensors,
f2 (αt ) =
N
1 X
αi,t
N i=1
(29)
2) MIUB based objective functions: The objective functions based on MIUB are very similar
to that with the FI: minimization of the normalized information gap between the total MIUB
based on all the sensors and the MIUB based on the sensor selection strategy αt ,
PN (i) PN
I −
αi I (i)
f1 (αt ) = i=1 PN i=1
(i)
i=1 I
(30)
where I (i) denotes I(zi,t ; xt ) (or I(Di,t ; xt )), and minimization of the normalized number of
selected sensors (the same as (29)).
B. NSGA-II
In this paper, we solve the above MOP which has binary decision variables using a state-of-theart multiobjective evolutionary algorithm, Nondominating sorting genetic algorithm (NSGA)-II
[44]. This algorithm yields all the solutions on the Pareto front that explore all the possible
tradeoffs between conflicting objectives.
NSGA-II [44] first generates an initial population of size P where each solution in the
population is a feasible solution of the MOP. In our problem, a solution in the population
is represented as a vector of N elements where each element is a binary variable. NSGA-II is an
elitist algorithm where good solutions are always preserved in the population. The values of the
October 8, 2015
DRAFT
20
objective functions for each solution in the population form the fitness values of the solution.
Then all the solutions in the population are sorted based on their non-domination. As an example,
solutions with Rank 1 consist of all non-dominated solutions, then solutions with Rank 2 consist
of all the solutions which are dominated by only one of the solutions in the population and so
on. If two solutions in the population have the same fitness value, then they are sorted based on
their crowding distance, which is a closure measure of each solution to its neighbors.
NSGA-II uses the rank of a solution to create the mating population. The offspring solutions
are generated by using binary tournament selection [44]. If both of the selected solutions have
the same fitness value, then the solution with larger crowding distance is selected. In our problem
where we have binary decision variables, we use a real-parameter recombination operator called
uniform crossover (UX), where offspring solutions c1 and c2 are obtained from parent solutions
p1 and p2 according to,
c1 = ξp1 + (1 − ξ)p2
(31)
c2 = (1 − ξ)p1 + ξp2
where ξ is defined by a random number q between [0, 1] [45]
ξ=1
q ≤ 0.5
ξ=0
q > 0.5
(32)
Along with UX, the uniform mutation procedure is employed. In uniform mutation, an offspring
solution cl is obtained from the parent solution pl according to
cl = δ(1 − cl ) + (1 − δ)cl
(33)
where δ is also determined according to (32). Then the new population with all the parents and
offsprings are sorted again based on their non-dominance and the population size is decreased to
the original population size P by eliminating all the lower rank solutions. Remaining solutions
are then fed to a binary tournament selection operator and so on. After several generations G,
the population will preserve solutions near or on the Pareto optimal front.
C. Solution Selection from the Pareto-optimal Front
Since NSGA-II provides P non-dominated solutions, it is necessary to select one particular
solution from the Pareto-front which can yield the desired trade-off between the conflicting
October 8, 2015
DRAFT
21
t=3
t=6
0.4
t=3
0.25
0.35
0.2
t=6
1
1
0.9
0.9
0.8
0.8
0.7
0.7
0.3
0.2
0.1
0.15
← Knee point Sol.
0.6
SMI Gap
0.6
SMI Gap
0.15
FI Gap
FI Gap
0.25
← Knee point Sol.
0.5
0.5
0.4
0.4
0.3
0.3
← Knee point Sol.
0.1
0.05
0.2
0.05
0
Compromise Sol. and
← Knee point Sol.
0
0.5
Sensor Selection
← Compromise Sol.
1
0
0
0.2
← Compromise Sol.
0.1
0.5
Sensor Selection
0
1
0
0.5
Sensor Selection
(a)
Fig. 4.
← Compromise Sol.
0.1
0
1
0
0.5
Sensor Selection
1
(b)
Pareto optimal front obtained by using NSGA-II at time step t = 3 and t = 6, (a) FI; (b) MIUB.
objectives. In [46], the knee of the trade-off curve is introduced as the solution where a small
decrease in one objective is associated with a large increase in the other. Let αa and αb be
two adjacent (neighboring) solutions on the Pareto-optimal front where f1 (αa ) > f1 (αb ) and
f2 (αa ) < f2 (αb ). Then we can compute the slope of the curve between solutions αa and αb
from,
f1 (αa ) − f1 (αb )
slope{α } = 180 − arctan
f2 (αa ) − f2 (αb )
b
180
π
(34)
For our problem, we define α1 as the all zero solution where none of the sensors are selected,
so that f1 (α1 ) = 1 and f2 (α1 ) = 0. Similarly, we define αP is the all one solution which yields
f1 (αP ) = 0 and f2 (αP ) = 1. We call the Pareto-optimal solution which maximizes (34) as the
knee point solution given by,
αt = arg max slope{αρ }
α2 ,...,αP
(35)
where αρ (ρ ∈ {2, 3, . . . P }) represents the solutions on or near the Pareto-optimal front.
Alternatively, the utopia point F ∗ of a MOP is defined as [43],
F ∗ = [f1∗ , . . . , fn∗ ]T
(36)
where fj∗ is the individual minima of objective fj (j ∈ {1, . . . , n}) defined as,
fj∗ = min{fj (α) | α ∈ C}
(37)
α
October 8, 2015
DRAFT
22
MOP−MIUB, Compromise Sol.
MOP−FI, Compromise Sol.
MOP−MIUB, Knee point Sol.
MOP−FI, Knee ponit Sol.
1
MSE
10
0
10
−1
Number of active sensors
10
Fig. 5.
0
5
10
15
20
MOP−MIUB, Compromise Sol.
MOP−FI, Compromise Sol.
MOP−MIUB, Knee point Sol.
MOP−FI, Knee ponit Sol.
10
8
6
4
2
0
0
5
10
Time step t
15
20
Tracking performance at each time step with different solution selection methods.
and let F (αρ ) = [f1 (αρ ), . . . , fn (αρ )]T where ρ ∈ {1, 2, . . . , P }. In [43], the point which is
closest to the utopia point has been defined as the compromise solution. In this paper, we use
the Euclidean distance to find the compromise solution as,
v
uX
u n
2
αt = arg min t
fj∗ − fj (αρ )
α1 ,...,αP
(38)
j=1
In the next section, we present some numerical results.
D. Numerical Experiments for the MOP Framework
In this section, we conduct some simulation experiments to investigate the performance of
the multiobjective optimization method. The WSN considered in this subsection is the same as
shown in Fig. 1 in Section IV-D, and the system parameters are also the same as Section IV-D.
Note that, for NSGA-II, the population size is chosen as P = 100. We choose the number of
generations according to the diversity metric introduced in [44]. The diversity metric measures
October 8, 2015
DRAFT
23
MIUB−NSGA−II
MIUB−Weighted Sum
MIUB−A=2
MIUB−A=9
All
1
1
MSE
10
MSE
10
FI−NSGA−II
FI−Weighted Sum
FI−A=1
FI−A=4
All
0
0
10
10
−1
10
−1
0
5
10
Time step t
15
10
20
(a)
0
5
10
Time step t
15
20
(b)
Fig. 6. Tracking performance for MOP with NSGA-II, convex relaxation, and weighted sum methods (a) MSE for MIUB; (b)
MSE for FI.
the extent of spread achieved among the obtained solutions, which is defined as
PN −1 (E) ¯(E)
(E)
(E)
|di − d |
df + dl + i=1
∆=
(E)
(E)
d + d + (N − 1)d¯(E)
f
(E)
where df
(E)
and dl
(39)
l
are the Euclidean distances between the extreme solutions and the boundary
solutions of the obtained nondominated set. We observe that for both FISS and MIUBSS, the
diversity metric converges after 100 generations for all the 20 time steps. Thus, in our simulation
experiments, we set the number of generations as G = 100. Also, before running NSGA-II, we
include the two extreme solutions, i.e, all zero and all one solutions to the initial population.
a) Pareto optimal front: In Fig. 4, we present the Pareto optimal front for our MOP obtained
using NSGA-II, where Fig. 4(a) is for FISS and Fig. 4(b) shows the result for MIUBSS. It is
interesting to note that at the end of G generations, NSGA-II yields N + 1 different solutions on
the Pareto-optimal front where each solution corresponds to the optimal selection of A sensors
out of N sensors where A ∈ {0, 1, . . . , N}. We know from (15), (18) and Table I that the distance
between the target and the sensor plays a more important role than the sensing probability for
FISS. At time step t = 3, the target is relatively close to sensor 8, and sensor 8 itself is able
to achieve significant FI gain. At time step t = 6 the target is not relatively close to any of
the sensors in the network and the fusion center has relatively large uncertainty about the target
location. Thus, the Pareto front for FISS at t = 3 is steeper than that at t = 6. However, compared
October 8, 2015
DRAFT
24
with FISS, MIUBSS prefers the sensors with high sensing probability and selects more sensors,
so that the Pareto front of MIUBSS at t = 3 or t = 6 is not as steep as that for FISS. Moreover,
we observe that the compromise solution and the knee point solution are identical when the
Pareto front is relatively steep.
b) Solution selection method: The solution, i.e., the sensor selection strategy, that we choose
from the Pareto optimal front determines the overall tracking performance. In Fig. 5, we compare
the average number of active sensors2 at each time step of tracking and the MSE performance
using the knee point solution (35) and the compromise solution (38) with MIUBSS and FISS
under the MOP framework. We observe similar results for MIUB and FISS that the knee point
solution always selects one sensor for target tracking, and thus gives poorer MSE performance.
However, the sensor selection strategy using the compromise solution in (38) selects the sensors
which balance the tradeoff between the performance gain (MIUB and FI) and the total number
of selected sensors. Thus, in the rest of our simulations, we use the compromise solution to
choose the sensor selection strategy from the Pareto optimal front.
Recall the results shown in Fig. 2 and Fig. 3 that MIUBSS selects more reliable sensors when
the number of sensors to be selected is given. Furthermore, Fig. 5 shows that when the number
of sensors to be selected is not known, MIUBSS tends to select more sensors than FISS under
the MOP framework, such that the MSE performance of MIUBSS is better than FISS.
c) NSGA-II, convex optimization, and weighted sum methods: In Fig. 6, we compare the
tracking performance based on NSGA-II and (38) with the convex relaxation based sensor
selection method similar to [13], [37] which always chooses A sensors out of N sensors at
each time step of tracking. We apply the convex relaxation method to select the minimum and
maximum number of sensors selected by NSGA-II with compromise solution in Fig. 5. With the
minimum number of sensors, the convex relaxation based sensor selection method gives poor
tracking performance. On the other hand, selecting the maximum number of sensors or all the
sensors through convex relaxation method negligibly improves the MSE performance compared
to the MOP approach. Thus, compared to the convex relaxation method, the multiobjective
optimization method gives satisfactory tracking performance while saving in terms of the number
2
We show the number of active sensors (the selected sensors) to investigate the energy cost of each solution selection method,
because selecting more sensors for data transmission incurs more energy cost.
October 8, 2015
DRAFT
25
1
10
MIUB
MSE
MIUB, pth=0.5
s
th
s
MIUB, p =0.15
0
10
−1
10
0
5
10
15
20
1
10
FI
th
MSE
FI, ps =0.5
th
FI, ps =0.15
0
10
−1
10
Fig. 7.
0
5
10
Time step t
15
20
Turn off sensors with relatively low sensing probabilities.
of selected sensors with both MIUBSS and FISS. We also compare the MSE performance of the
MOP framework with the weighted sum approach where the sensor selection scheme chooses
those sensors which minimize the summation of both objectives, i.e. w1 f1 (αt ) + (1 − w1 )f2 (αt )
with w1 = 0.5. Simulation results show that for MIUBSS (Fig. 6(a)), the NSGA-II method
obtains similar MSE performance with weighted sum method, while for FISS (Fig. 6(b)), the
weighted sum method achieves much worse MSE performance.
d) A Naive strategy: We consider a naive sensor selection method in which the fusion
center turns off the sensors with relatively low sensing probabilities before sensor selection.
In Fig. 7, we present the results when the fusion center turns off the sensors whose sensing
th
th
probabilities are lower than some threshold pth
s , where ps = 0.5 and ps = 0.15 are considered.
Note that, for the WSN in Fig. 1, the sensors that are relatively close to the target will be turned
off because they have low sensing probabilities. As shown in the previous results, MIUBSS
prefers to select more reliable sensors, and FISS selects the sensors that are close to the target.
Turning off sensors before selection performs worse for MIUBSS because it reduces the selection
alternatives, and it performs better for FISS because more reliable sensors are selected when the
closest sensors with low sensing probabilities are no longer available.
October 8, 2015
DRAFT
26
FI−With uncertainty
MIUB−With uncertainty
FI−Without uncertainty
MIUB−Without uncertainty
MSE
0
10
−2
Number of active sensors
10
Fig. 8.
0
5
10
15
20
FI−With uncertainty
MIUB−With uncertainty
FI−Without uncertainty
MIUB−Without uncertainty
10
8
6
4
2
0
0
5
10
Time step t
15
20
Comparison the performance for WSNs with and without uncertainty.
e) Comparison with the performance when there is no uncertainty: In Fig. 8, we present
the target tracking performance when the sensors are all reliable, e.g., ps = 1 for all the sensors,
and compare with the results with uncertain observations. We observe that, with uncertain
observations, both FISS and MIUBSS achieve worse MSE performance though they both tend
to select more sensors. Moreover, compared with FISS, MIUBSS selects many more sensors
with uncertain observations, and therefore achieves better MSE performance.
f) WSN with another instance of sensing probabilities: In Fig. 9, the sensors’ sensing
probabilities are distributed in a reverse manner as compared with Fig. 1, i.e., the sensors that
are around the target track have relatively high sensing probabilities. In this condition, MIUBSS
and FISS select similar number of sensors with similar MSE performance. The reason is that,
under this scenario, MIUBSS and FISS both select the sensors around the target track with
high sensing probabilities. We also have conducted experiments for the following scenarios: 1)
the sensors’ sensing probabilities are all uniformly distributed between 0 and 1; 2) the sensor
measurements have higher noise; and 3) the sensor measurements are quantized to 3 bits, since
the results do not provide any new insights, we do not show the results in the paper.
October 8, 2015
DRAFT
27
MOP−MIUB
MOP−FI
0
MSE
10
−1
10
Number of active sensors
0
Fig. 9.
5
10
15
20
8
MOP−MIUB
MOP−FI
6
4
2
0
0
5
10
Time step t
15
20
Tracking performance of MIUB and FI, sensors’ sensing probabilities are reversely deployed.
VI. C ONCLUSION
In this paper, we have proposed a multiobjective optimization method for the sensor selection
problem in an uncertain wireless sensor network (WSN) for target tracking. We have considered
the three performance metrics, Fisher information (FI), mutual information (MI), and mutual
information upper bound (MIUB), as objective functions in characterizing the estimation
performance for the multiobjective optimization problem (MOP). Numerical results show that
the MIUB based selection scheme (MIUBSS) selects more reliable sensors compared with
the FI based selection scheme (FISS) while saving computational cost compared with the MI
based selection scheme (MISS). Furthermore, for the MOP framework, we have shown that the
compromise solution on the Pareto front of the MOP achieves good estimation performance
while obtaining savings in terms of the number of selected sensors.
In this work, we were interested in finding the sensor selection strategy with a multiobjective
optimization method in uncertain WSNs. Future work will consider the application of the
multiobjective optimization method in the multitarget tracking problem in uncertain WSNs.
October 8, 2015
DRAFT
28
R EFERENCES
[1] T. Bokareva, W. Hu, S. Kanhere, B. Ristic, N. Gordon, T. Bessell, M. Rutten, and S. Jha, “Wireless sensor networks for
battlefield surveillance,” in Proc. Conf. Land Warfare, 2006, pp. 1–8.
[2] J. Yick, B. Mukherjee, and D. Ghosal, “Wireless sensor network survey,” Computer networks, vol. 52, no. 12, pp. 2292–
2330, 2008.
[3] V. C. Gungor and G. P. Hancke, “Industrial wireless sensor networks: Challenges, design principles, and technical
approaches,” IEEE Trans. Ind. Electron., vol. 56, no. 10, pp. 4258–4265, 2009.
[4] A. Milenković, C. Otto, and E. Jovanov, “Wireless sensor networks for personal health monitoring: Issues and an
implementation,” Computer communications, vol. 29, no. 13, pp. 2521–2533, 2006.
[5] H. Rowaihy, S. Eswaran, M. Johnson, D. Verma, A. Bar-Noy, T. Brown, and T. La Porta, “A survey of sensor selection
schemes in wireless sensor networks,” in Defense and Security Symposium. International Society for Optics and Photonics,
2007, pp. 65 621A–65 621A.
[6] H. Wang, K. Yao, G. Pottie, and D. Estrin, “Entropy-based sensor selection heuristic for target localization,” in Proc. 3rd
int. symposium on Information processing in sensor networks.
ACM, 2004, pp. 36–45.
[7] J. L. Williams, J. W. Fisher, and A. S. Willsky, “Approximate dynamic programming for communication-constrained sensor
network management,” IEEE Trans. Signal Process., vol. 55, no. 8, pp. 4300–4311, 2007.
[8] G. M. Hoffmann and C. J. Tomlin, “Mobile sensor network control using mutual information methods and particle filters,”
IEEE Trans. Autom. Control, vol. 55, no. 1, pp. 32–47, Jan. 2010.
[9] F. Zhao, J. Shin, and J. Reich, “Information-driven dynamic sensor collaboration,” IEEE Trans. Signal Process., vol. 19,
no. 1, pp. 61–72, May 2002.
[10] L. Zuo, R. Niu, and P. K. Varshney, “Posterior CRLB based sensor selection for target tracking in sensor networks,” in
IEEE Int. Conf. Acoustics, Speech and Signal Process. (ICASSP), vol. 2, Apr. 2007, pp. 1041–1044.
[11] ——, “A sensor selection approach for target tracking in sensor networks with quantized measurements,” in IEEE Int.
Conf. Acoustics, Speech and Signal Process. (ICASSP), Mar 31-Apr 4 2008, pp. 2521–2524.
[12] E. Masazade, R. Niu, P. K. Varshney, and M. Keskinoz, “Energy aware iterative source localization for wireless sensor
networks,” IEEE Trans. Signal Process., vol. 58, no. 9, pp. 4824–4835, 2010.
[13] S. Joshi and S. Boyd, “Sensor selection via convex optimization,” IEEE Trans. Signal Process., vol. 57, no. 2, pp. 451
–462, Feb. 2009.
[14] Y. Mo, R. Ambrosino, and B. Sinopoli, “Sensor selection strategies for state estimation in energy constrained wireless
sensor networks,” Automatica, vol. 47, pp. 1330–1338, Jul. 2011.
[15] S. Liu, S. Kar, M. Fardad, and P. K. Varshney, “Sparsity-aware sensor collaboration for linear coherent estimation,” IEEE
Trans. Signal Process., vol. 63, no. 10, pp. 2582–2596, 2015.
[16] S. Liu, M. Fardad, E. Masazade, and P. Varshney, “Optimal periodic sensor scheduling in networks of dynamical systems,”
IEEE Trans. Signal Process., vol. 62, no. 12, pp. 3055–3068, June 2014.
[17] N. E. Nahi, “Optimal recursive estimation with uncertain observation,” IEEE Trans. Inf. Theory,, vol. 15, no. 4, pp. 457–462,
1969.
[18] M. Hadidi and S. Schwartz, “Linear recursive state estimators under uncertain observations,” IEEE Trans. Autom. Control,
vol. 24, no. 6, pp. 944–948, 1979.
[19] F. O. Hounkpevi and E. E. Yaz, “Robust minimum variance linear state estimators for multiple sensors with different
failure rates,” Automatica, vol. 43, no. 7, pp. 1274–1280, 2007.
October 8, 2015
DRAFT
29
[20] H. Zhang, Y. Shi, and A. S. Mehr, “Robust weighted h filtering for networked systems with intermittent measurements of
multiple sensors,” Int. J. Adapt. Control and Signal Process., vol. 25, no. 4, pp. 313–330, 2011.
[21] W. Xu, K. Ma, W. Trappe, and Y. Zhang, “Jamming sensor networks: attack and defense strategies,” IEEE Network, vol. 20,
no. 3, pp. 41–47, 2006.
[22] M. Mariton, Jump linear systems in automatic control.
CRC Press, 1990.
[23] O. L. V. Costa and S. Guerra, “Stationary filter for linear minimum mean square error estimator of discrete-time markovian
jump systems,” IEEE Trans. Autom. Control, vol. 47, no. 8, pp. 1351–1356, 2002.
[24] B. Sinopoli, L. Schenato, M. Franceschetti, K. Poolla, M. I. Jordan, and S. S. Sastry, “Kalman filtering with intermittent
observations,” IEEE Trans. Autom. Control, vol. 49, no. 9, pp. 1453–1464, 2004.
[25] O. Ozdemir, R. Niu, and P. K. Varshney, “Channel aware target localization with quantized data in wireless sensor networks,”
IEEE Trans. Signal Process., vol. 57, no. 3, pp. 1190–1202, 2009.
[26] E. Masazade, R. Niu, P. K. Varshney, and M. Keskinoz, “Channel aware iterative source localization for wireless sensor
networks,” in Proc. IEEE 13th Int. Conf. Information Fusion (FUSION). IEEE, 2010, pp. 1–7.
[27] E. Masazade, R. Rajagopalan, P. K. Varshney, C. Mohan, G. Kiziltas Sendur, and M. Keskinoz, “A multi-objective
optimization approach to obtain decision thresholds for distributed detection in wireless sensor networks,” IEEE Trans.
Syst., Man, Cybern. - Part B, vol. 40, no. 2, pp. 444–457, Apr. 2010.
[28] R. Rajagopalan, C. Mohan, P. K. Varshney, and K. Mehrotra, “Multi-objective mobile agent routing in wireless sensor
networks,” in The IEEE Congress on Evolutionary Computation, vol. 2, Sept. 2005, pp. 1730 – 1737 Vol. 2.
[29] R. Rajagopalan, “Multi-objective optimization algorithms for sensor network design,” in IEEE Annual Wireless and
Microwave Technology Conference (WAMICON), Apr. 2010, pp. 1 –4.
[30] M. Nasir, S. Sengupta, S. Das, and P. Suganthan, “An improved multi-objective optimization algorithm based on fuzzy
dominance for risk minimization in biometric sensor network,” in IEEE Congress on Evolutionary Computation (CEC),
June 2012, pp. 1 –8.
[31] N. Cao, E. Masazade, and P. K. Varshney, “A multiobjective optimization based sensor selection method for target tracking
in wireless sensor networks,” in Proc. IEEE 16th Int. Conf. Information Fusion (FUSION), 2013, pp. 974–980.
[32] W.-C. Lee, “Uncertainty in wireless sensor networks,” in Workshop on AFRL, 2010.
[33] Y. Ruan, P. Willett, A. Marrs, F. Palmieri, and S. Marano, “Practical fusion of quantized measurements via particle filtering,”
IEEE Trans. Aerosp. Electron. Syst., vol. 44, no. 1, pp. 15 –29, January 2008.
[34] N. J. Gordon, D. J. Salmond, and A. F. Smith, “Novel approach to nonlinear/non-gaussian bayesian state estimation,” in
IEEE Proceedings F (Radar and Signal Processing), vol. 140, no. 2.
IET, 1993, pp. 107–113.
[35] M. S. Arulampalam, S. Maskell, N. Gordon, and C. Tim, “A tutorial on particle filters for online nonlinear/ non-gaussian
bayesian tracking,” IEEE Trans. Signal Process., vol. 50, no. 2, pp. 174–188, Feb. 2002.
[36] H. V. Trees, Detection, Estimation, and Linear Modulation Theory, Part I.
Wiley Interscience, 2001.
[37] E. Masazade, R. Niu, and P. K. Varshney, “Dynamic bit allocation for object tracking in wireless sensor networks,” IEEE
Trans. Signal Process., vol. 60, no. 10, pp. 5048–5063, Sep. 2012.
[38] A. Ryan, “Information-theoretic tracking control based on particle filter estimate,” in AIAA Guidance Navigation and
Control Conference, 2008, pp. 1–15.
[39] C. E. Shannon, “A mathematical theory of communication,” ACM SIGMOBILE Mobile Computing and Communications
Review, vol. 5, no. 1, pp. 3–55, 2001.
[40] T. M. Cover and J. A. Thomas, Elements of information theory.
October 8, 2015
John Wiley & Sons, 2012.
DRAFT
30
[41] Y. Zhang and Q. Ji, “Efficient sensor selection for active information fusion,” IEEE Trans. Syst., Man, and Cybern., Part
B: Cybern., vol. 40, no. 3, pp. 719–728, 2010.
[42] Y. Bar-Shalom, P. K. Willett, and X. Tian, Tracking and Data Fusion: A Handbook of Algorithms.
YBS Publishing,
Storrs, CT, 2011.
[43] R. T. Marler and J. S. Arora, “Survey of multi-objective optimization methods for engineering,” Struct Multidisc Optim,
vol. 26, pp. 369–395, 2004.
[44] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist multiobjective genetic algorithm: Nsga-ii,” IEEE
Trans. Evol. Comput., vol. 6, no. 2, pp. 182–197, Apr. 2002.
[45] N. Issimakis and M. Adam, “A genetic algorithm for the multidimensional knapsack problem,” Journal of Heuristics, no. 4,
pp. 63–86, 1998.
[46] S. Boyd and L. Vandenberghe, Convex Optimization.
October 8, 2015
Cambridge: Cambridge University Press, 2004.
DRAFT
| 3cs.SY
|
A norm knockout method on indirect reciprocity to
reveal indispensable norms∗
Hitoshi Yamamoto1,* , Isamu Okada2,+ , Satoshi Uchida3,+ , and Tatsuya Sasaki4,+
arXiv:1703.03943v1 [physics.soc-ph] 11 Mar 2017
1 Rissho
University, Department of Business Administration, Tokyo, 141-8602, Japan
University, Department of Business Administration, Tokyo, 192-8577, Japan
3 RINRI Institute, Research Center for Ethiculture Studies, Tokyo, 102-8561, Japan
4
University of Vienna, Faculty of Mathematics, Vienna, 1090, Austria
* [email protected]
+ these authors contributed equally to this work
2 Soka
ABSTRACT
Although various norms for reciprocity-based cooperation have been suggested that are evolutionarily stable against invasion
from free riders, the process of alternation of norms and the role of diversified norms remain unclear in the evolution of
cooperation. We clarify the co-evolutionary dynamics of norms and cooperation in indirect reciprocity and also identify the
indispensable norms for the evolution of cooperation. Inspired by the gene knockout method, a genetic engineering technique,
we developed the norm knockout method and clarified the norms necessary for the establishment of cooperation. The results
of numerical investigations revealed that the majority of norms gradually transitioned to tolerant norms after defectors are
eliminated by strict norms. Furthermore, no cooperation emerges when specific norms that are intolerant to defectors are
knocked out.
Introduction
Reciprocity is a fundamental mechanism that underlies all cooperative societies. Theoretically it is well known that direct
reciprocity, typified by the “I’ll help you if you help me” attitude, promotes cooperative regimes1, 2 . However, in recent
societies that have high relational mobility, indirect reciprocity such as “I’ll help you and somebody else will help me” plays
a more important role in promoting cooperation. Indirect reciprocity has therefore been the focus of much research in the
interdisciplinary fields in recent decades3–7 .
Many theoretical studies on indirect reciprocity have explored norms that become evolutionarily stable against defection
and the invasion of free riders, and several typical norms have been proposed8–11 . These approaches have clarified the robust
norms that can maintain the cooperative regime. The norms in the studies on the indirect reciprocity are regarded as assessment
rules that label the other’s action as either Good or Bad. They include tolerant norms that assess cooperative behaviors toward
defectors as good4 and strict norms that assess such behaviors as bad12. Other theoretical studies analysing the global dynamics
of norms assume that at most a few robust norms are shared in the population13–15 .
Their approaches have clarified the robustness of the norms against invasion of other norms including free riders when the
norms are acceptable in the population. However, little is known about a process by which gradual changes toward cooperation
occur as new norms emerge and compete, which is to say, the co-evolutionary process of norm-diversity and cooperation. A
study on the indirect reciprocity has dealt with co-existing different norms and has analysed their frequencies in the population
as a consequence of a dynamical process16 . In the study, each individual keeps a private image of everyone else and errors
of perception and implementation are included in a limited strategy space. Although they have considered some action rules
and assessment rules, all possible norms in indirect reciprocity have not been studied all-together. How cooperation evolves
cannot be fully understood unless the evolution of norms is also considered.
It is thus a challenging task to theoretically understand how cooperation can be formed even under a collection of norms
in a social system. How is the co-existence of cooperation and diversity possible at all? Are there any indispensable norms
needed to facilitate the evolution of cooperation in the melting pot of norms, even though some norms never become dominant?
Do norms that could be accepted as a result of the co-evolutionary process have common aspects? These questions can be
addressed only if all possible norms are considered, and the combination of norms governing a group can evolve.
Here we explore the dynamics of co-evolution of cooperation by using different social norms. The process of the evolution
∗ The
final version was published in Scientific Rreports. How to cite this article: Yamamoto, H., Okada, I., Uchida, S., & Sasaki, T. A norm knockout
method on indirect reciprocity to reveal indispensable norms. Sci. Rep. 7, 44146; doi: 10.1038/srep44146 (2017).
of norms has a transition from stricter to tolerant norms. Additionally, we find a set of norms that seem not to have an impact
on promoting cooperation, but are fundamental to allow a transition to a cooperative regime from a defective regime.
Results
Agent-based simulations17, 18 are an optimal tool to tackle the challenge outlined above. See Methods for the details of our
agent-based model described by the ODD protocol19. Using an evolutionary game theoretical framework and constructing an
interaction model based on players’ private rules and local information, we model a giving game to elucidate the dynamics
of the evolution of cooperation amid the coexistence of diverse norms (Fig. 1). We conducted numerical simulations of
all 16 possible norm combinations that could react to the four combinations of assessment criteria to clarify the dynamics
of the evolution of cooperation from the melting pot of diverse norms. Figure 2a shows time-series graphs of each norm’s
population and cooperation ratio. As shown, the majority undergo an alternation from strict to tolerant norms, mostly in the
order of SH → SJ → ST. Figure 3a shows the transition in the norm with the greatest population ratio. In many cases, the
majority transitioned from the state where strict SH11 was the majority to SJ5, 12 . Afterwards, the majority norm changed to
tolerant ST4, 20, 21 and ALLG. In contrast, as shown in Figs. 2b and 3b, in an environment with errors, alternation from strict
norms to tolerant norms was observed. However, the likelihood of going through SJ decreased. Alternation paths through
IS22, 23 , which could not be seen in an environment without errors24, increased. It is important to note here that similar
paths toward cooperation are observed when only ALLB-individuals are initially assumed. New norms are created during
the evolutionary process at the same time cooperation evolves. This indicates that cooperation and diversity of norms jointly
evolve in the model.
Why does the alternation of norms emerge? For one thing, in states in which defection is dominant, ALLB (BBBB) and
SH (GBBB) coexist and jointly form the majority. However, BGBB and IS (GGBB) continue to exist as the minority. The
characteristic of these groups is having the evaluation rule of **BB. Evaluation rule **BB assesses donors that took D as
B, regardless of the evaluation of the recipient. In states in which defection is dominant, those who adopt **BB strategies
consider many partners as B. As a result, cooperation does not occur for the most part. The ALLB and SH norms thus survive
because they do not lower their own cost. On the other hand, after cooperation is achieved, ALLG (GGGG), ST (GGBG),
IS (GGBB), and GGGB coexist. The common characteristic of these norms is having the evaluation rule of GG**. Thus,
reciprocally cooperating norms survive. Because SJ (GBBG), which becomes the majority temporarily when the cooperation
ratio rises in an environment without errors, does not belong to either group, it cannot stably exist. Also, it is rare that SJ
makes up the majority temporarily in an environment with errors. Meanwhile, because IS belongs to both norm groups with
**BB and GG**, IS can constantly exist.
We discover several norms that are indispensable to the evolution of cooperation. Reputation-based cooperation cannot
emerge without indispensable norms. To elucidate indispensable norms for the evolution of cooperation, we propose a novel
analysis using the norm knockout method. This method enables us to determine which norms are indispensable for the
evolution of cooperation. The norm knockout method is inspired by the targeted gene knockout technique used in genetic
engineering25. Gene knockout, a genetic technique in which one of an organism’s genes is made inoperative, is used to research
genes whose sequences are known but whose functions are not well-understood. Researchers infer the gene’s function from
differences between the knockout animal and a normal animal. For simulating evolution, we utilized a method that removed
only one particular norm from the population to understand whether that norm is an indispensable one that plays a critical role
in the evolution of cooperation.
Figure 4 shows the cooperation ratio when a particular norm is knocked out. Regardless of whether there is an error, if SH
or IS is knocked out, cooperat
ion does not evolve at all. We define indispensable norms in the evolution of cooperation as the norms that, when knocked
out, have an average cooperation ratio of less than 0.1 after 1,000 generations. In an environment with no errors, SH and IS
are indispensable norms. In an environment with errors, SH, IS, and ST are indispensable norms.
When an indispensable norm is knocked out, cooperation does not evolve. When cooperation evolves, alternation from
strict norms to tolerant norms was observed, as shown in Figs. 2 and 3. To analyse whether alternation also occurs when a
norm is knocked out, the population ratio of norms when typical norms are knocked out is displayed as time-series graphs
(see Fig. 5). Figure 5 shows the results in the cases where SH or IS were knocked out. We discovered that the first condition
for the necessary process when cooperation evolves is whether SH can antagonize ALLB. No norm that resists the invasion
of ALLB appears in a society in which SH does not exist. Also, in a society in which IS does not exist, SH cannot antagonize
ALLB. We found that IS is a norm indispensable for SH to resist ALLB.
2/15
Discussion
Our model offered two major findings on the evolution of cooperation on indirect reciprocity. On the one hand, the most
essential contribution is the discovery of indispensable norms by the norm knockout method. By using the norm knockout
method, we were able to elucidate the existence of norms indispensable for the evolution of cooperation from a melting pot
of norms. Regardless of the existence of errors, SH and IS were indispensable norms. In addition, in an environment with
errors, ST is an indispensable norm. Interestingly, SH and IS are reconciled to the minorities after the cooperative regime
emerges while they temporarily become major norms in the process of dynamics. We call such minority norms required for
the evolution of cooperation “unsung hero norms”. The results clearly illustrate the two roles of norms: one to catalyse a
cooperative regime and the other to maintain the regime. Norms having the GG** for the evaluation rule play the latter role.
On the other hand, we discovered alternation of norms. Recent analysis of evolutionary stability against the invasion of free
riders could identify neither superiority among norms nor the process on the path to cooperation. Among studies on indirect
reciprocity, ours is the first exhaustive theoretical analysis on all possible norms, although several studies have addressed the
comparison of two types of reciprocal norms14, 15, 26 . Others analyse the alternation of norms in direct reciprocity27–29 . We
find the alternation of the norms and also discover the indispensable norms that are required to foster indirect reciprocity.
An empirical study30 supports the co-existence of various norms in the cooperative regime and indicates that the ST norm
plays a more important role in human cooperation than SJ, which is consistent with our simulation. This is because we show
that the SJ norm cannot survive in the cooperative regime, while the ST one can. Our approach may provide deep insight
on the evolution of cooperation because several norms absolutely play an essential role in order to evolve cooperation even
though, on the surface, it seems as though they are not directly leading to the evolution of the cooperation.
The present work considers a single action rule (cooperate with Good, defect with Bad) to stress the role of multiple
assessment rules. However, the other papers stress the role of multiple co-existing action rules13, 31–33 . Integrating the multiple
assessment and action rules may be a useful extension of this paper. We analyse what happens when one norm is absent from
the population; however, we have not analysed all the indispensable combinations of norms yet. Extending the norm knockout
method to combinations of norms may also be a useful extension of this paper.
Methods
In this section, we describe the details of our agent-based model that uses the norm knockout method. The following model
description follows the ODD protocol19.
Purpose
The aim of the model is to understand the dynamics of norms during the evolution of cooperation, and to find indispensable
norms without which cooperative societies could never emerge. In particular, we reveal the effect of these indispensable
norms on indirect reciprocity using a new methodology we call the ”norm knockout method”. We utilize the giving game
framework34 for simulation.
Entities, state variables, and scales
The entities in the model are agents who play as donor and recipient in the giving game with no spatial structures. The donor
chooses cooperation or defection with a recipient using an image that the donor has to the recipient. An image is either Good
or Bad. If a donor’s image to a recipient is Good, the donor cooperates with the recipient. If the image is Bad, the donor
defects. The group size of the model is N. Each agent has it’s own norm and a list of images to other agents. The agent also
has a probability of errors and a payoff of the game.
The norm of an agent is denoted as one of four possible ”assessment combinations”, and there are two possible ”alleles
(G/B)” at the ”locus” for each of the four assessment combinations. The first locus of the gene represents an assessment rule
to an agent who cooperates with a Good recipient. The second locus represents an assessment rule to an agent who cooperates
with a Bad recipient. The third locus represents an assessment rule to an agent who defects with a Good recipient. The fourth
locus represents an assessment rule to an agent who defects with a Bad recipient. Incidentally, all agents evaluate themselves
as Good. For instance, ALLG always assesses others as Good, and thus, its ”genotype” is GGGG using the above mentioned
definition of the four loci. Similarly, ALLB is described as BBBB, IS as GGBB, ST as GGBG, and SJ as GBBG.
Each agent has two different types of errors: one, the probability that the agent’s updating of its evaluation of others,
Good/Bad, is inverted (errors in perception), described as p, and two, the probability to perform an action differently from the
one prescribed by its action rule (errors in implementation), described as q. The evolution process of norms involves adopting
a genetic algorithm35. State variables and initialization in the simulation are shown in Table 1.
3/15
Process overview and scheduling
Our simulation runs throughout G generations. A generation consists of R rounds. The agents play the giving game R times as
donor in each generation. At the end of a generation, they evolve their own norms using accumulated payoffs that are obtained
in the generation. One round has two phases: (A) a phase to play giving games and (B) a phase to update images. After all
agents play giving games in phase (A), all agents update their images in phase (B). In phase (A), each agent becomes a donor
and each donor randomly chooses a recipient from N − 1 players excluding itself. The donor chooses whether to give benefits
to the recipient or not. At that time, the action of the donor is inverted with the probability q. The donor who cooperates pays
cost c and the recipient receives benefit b (b > c > 0). In phase (B), each agent (set to i) evaluates and updates an image to
the other agent (set to j). The new image to j from i’s viewpoint depends on j’s action (C/D) as a donor in the last round and
depends on the image to k from the viewpoint of i (G/B), where k was a recipient of j in the last round. At that time, the image
to j is inverted with the probability p. In the first round in a generation, j’s action (C/D) is regarded as random.
After R rounds of the giving game are played in every generation, agents evolve their norm. The evolution process of
norms involves adopting the genetic algorithm35. Because each locus of norm has an independent meaning for assessment
to others, the adaptive process should contain a combination of these elements rather than a string of norms. We have
modeled a process of updating norms not as a string of norms but rather as four different assessment rules, which enables
the norms to be interpreted as different situations depending on the norm genotype. The first, second, third, and fourth loci
represent the assessment for pro-social behavior, tolerant behavior, anti-social behavior, and justified defection (punishment),
respectively. Each agent randomly selects two agents from N agents (including itself) to become its parents. For choosing
parents, we adopt a roulette selection method. This roulette selection sets a probability distribution of all agents as Πi =
(Ui −Umin )2 / ∑ j (U j −Umin )2 , where Ui denotes the agent i’s accumulated payoff in a generation given by Ui = bW − cV , with
W being the number of donations i received in the generation and V the number of donations i gave. Umin means a minimum
value of the accumulated payoffs among all. Finally, each agent updates its norm using a uniform crossover technique. With
mutation rate m, each locus is inverted for maintaining the diversity of the norm space.
Design concepts
Basic principles An agent-based simulation is utilized to study indirect reciprocity. We explore how different combinations
of norms interact to produce an evolutionary progression towards cooperation.
Emergence A cooperative regime in the situation of social dilemma emerges from interactions among agents who have
various social norms.
Adaptation The agents of the model play the giving game using their images to others. The agents update their images to
others using their norms every round. They evolve their own norms using accumulated payoffs that are obtained in the
generation. A norm that can obtain a higher payoff can increase the population through the generation.
Objectives The objective of all agents is to maximize their own payoff. To maximize payoff, they change their own norm at
the end of each generation.
Learning The agents change their norms in each generation using a genetic algorithm. The fitness of each agent is calculated
from the accumulated payoff in the generation. To select the parents of the agent, the model utilizes a roulette selection
method.
Interaction The interaction between the agents is one to one interaction. The giving game consists of the donor and the
recipient. There are no spatial structures in the society.
Stochasticity The interaction between agents is a stochastic process because interaction partners are chosen randomly from
the society. At the start of the simulation, each agent is randomly assigned a norm of all 16 norms.
Observation Three indexes are used for observation: average cooperation ratio in the society, the transition of norms with
the greatest populations, and population ratio of each norm.
Initialization
At the start of simulation, the norm of each agent is chosen randomly from all 16 possible norm combinations. In the first
round of each generation, the evaluation of all agents is initialized as Good11, 23, 36 and payoff of the agents is initialized as 0.
Input data
After initialization, the model does not include any external inputs, i.e., the number of agents (N), error ratio (p, q), benefit (b),
and cost (c) are constant.
4/15
Submodels
The norm knockout method
The norm knockout method is implemented as follows. When we knock out a particular norm, that norm is removed in the
first round of each generation. Concretely, if the norm of an agent evolves into a norm that is knocked out as a result of the
adopting process, the norm of the agent is changed to one of the other 15 norms randomly. In other words, the norm that is
knocked out will never exist at all in the society.
References
1. Trivers, R. L. The evolution of reciprocal altruism. Q. Rev. Biol. 46, 35–57 (1971).
2. Axelrod, R. & Hamilton, W. D. The evolution of cooperation. Science 211, 1390–1396 (1981).
3. Alexander, R. The Biology of Moral Systems (Aldine de Gruyter, New York, 1987).
4. Sugden, R. The Economics of Rights, Cooperation and Welfare (Basil Blackwell, Oxford, 1986).
5. Kandori, M. Social norms and community enforcement. Rev. Econ. Stud. 59, 63–80 (1992).
6. Wedekind, C. & Milinski, M. Cooperation through image scoring in humans. Science 288, 850–852 (2000).
7. Panchanathan, K. & Boyd, R. Indirect reciprocity can stabilize cooperation without the secondorder free rider problem.
Nature 432, 499–502 (2004).
8. Ohtsuki, H. & Iwasa, Y. How should we define goodness? -reputation dynamics in indirect reciprocity. J. Theor. Biol.
231, 107–120 (2004).
9. Nowak, M. A. & Sigmund, K. Evolution of indirect reciprocity. Nature 437, 1291–1298 (2005).
10. Ohtsuki, H. & Iwasa, Y. The leading eight: social norms that can maintain cooperation by indirect reciprocity. J. Theor.
Biol. 239, 435–44 (2006).
11. Takahashi, N. & Mashima, R. The importance of subjectivity in perceptual errors on the emergence of indirect reciprocity.
J. Theor. Biol. 243, 418–36 (2006).
12. Pacheco, J. M., Santos, F. C. & Chalub, F. A. C. Stern-judging: A simple, successful norm which promotes cooperation
under indirect reciprocity. PLoS Comput. Biol. 2, e178 (2006).
13. Ohtsuki, H. & Iwasa, Y. Global analyses of evolutionary dynamics and exhaustive search for social norms that maintain
cooperation by reputation. J. Theor. Biol. 244, 518 – 531 (2007).
14. Uchida, S. & Sigmund, K. The competition of assessment rules for indirect reciprocity. J. Theor. Biol. 263, 13–19 (2010).
15. Uchida, S. Effect of private information on indirect reciprocity. Phys. Rev. E 82, 036111 (2010).
16. Brandt, H. & Sigmund, K. The logic of reprobation: Assessment and action rules for indirect reciprocation. J. Theor. Biol.
231, 475–486 (2004).
17. Gilvert, N. & Troitzsch, K. Simulation for the Social Scientist (Open University Press, 1999).
18. Roberts, G. Evolution of direct and indirect reciprocity. Proc. R. Soc. Lond. B 275, 173–179 (2008).
19. Grimm, V. et al. The odd protocol: A review and first update. Ecol. Modell. 221, 2760–2768 (2010).
20. Leimar, O. & Hammerstein, P. Evolution of cooperation through indirect reciprocity. Proc. R. Soc. Lond. B 268, 745–753
(2001).
21. Panchanathan, K. A tale of two defectors: the importance of standing for evolution of indirect reciprocity. J. Theor. Biol.
224, 115–126 (2003).
22. Nowak, M. A. & Sigmund, K. Evolution of indirect reciprocity by image scoring. Nature 393, 573–577 (1998).
23. Nowak, M. A. & Sigmund, K. The dynamics of indirect reciprocity. J. Theor. Biol. 194, 561–574 (1998).
24. Lotem, A., Fishman, M. & Stone, L. Evolution of cooperation between individuals. Nature 400, 226–227 (1999).
25. Strepp, R., Scholz, S., Kruse, S., Speth, V. & Reski, R. Plant nuclear gene knockout reveals a role in plastid division for
the homolog of the bacterial cell division protein ftsz, an ancestral tubulin. Proc. Natl. Acad. Sci. USA 95, 4368–4373
(1998).
26. Matsuo, T., Jusup, M. & Iwasa, Y. The conflict of social norms may cause the collapse of cooperation: indirect reciprocity
with opposing attitudes towards in-group favoritism. J. Theor. Biol. 346, 34–46 (2014).
5/15
27. Lindgren, K. Evolutionary phenomena in simple dynamics. In Langton, C. G., Taylor, C., Farmer, J. D. & Rasmussen, S.
(eds.) Artificial Life II, 295–312 (Addison-Wesley, Redwood City, CA, 1992).
28. Zagorsky, B. M., Reiter, J. G., Chatterjee, K. & Nowak, M. A. Forgiver triumphs in alternating Prisoner’s Dilemma. PLoS
ONE 8, e80814 (2013).
29. Berg, P. V. D. & Weissing, F. J. The importance of mechanisms for the evolution of cooperation. Proc. R. Soc. Lond. B
282, 20151382 (2015).
30. Swakman, V., Molleman, L., Ule, A. & Egas, M. Reputation-based cooperation: Empirical evidence for behavioral
strategies. Evol. Hum. Behav. 37, 230–235 (2016).
31. Brandt, H. & Sigmund, K. Indirect reciprocity, image scoring, and moral hazard. Proc. Natl. Acad. Sci. USA 102,
2666–2670 (2005).
32. Santos, F. P., Santos, F. C. & Pacheco, J. M. Social norms of cooperation in small-scale societies. PLoS Comput. Biol. 12,
e1004709 (2016).
33. Sasaki, T., Okada, I. & Nakai, Y. The evolution of conditional moral assessment in indirect reciprocity. Sci. Rep. 7, 41870
(2017).
34. Sigmund, K. The calculus of selfishness (Princeton University Press, 2010).
35. Holland, J. Adaptation in Natural and Artificial Systems (University of Michigan Press, 1975).
36. Ohtsuki, H., Iwasa, Y. & Nowak, M. A. Reputation effects in public and private interactions. PLoS Comput. Biol. 11,
e1004527 (2015).
Acknowledgements
HY acknowledges Grant-in-Aid for Scientific Research (C) 15KT0133 and 26330387. HY also acknowledges Prof. Kurihara (UEC, Japan) for providing the computational resources. IO acknowledges Grant-in-Aid for Scientific Research (B)
16H03120. TS acknowledges the Austrian Science Fund (FWF): P27018-G11.
Author contributions statement
HY initiated and performed the project. HY, IO, SU and TS designed the project, wrote the paper, and approved the submission.
All authors reviewed the manuscript.
Additional information
The authors declare no competing financial interests.
6/15
Tables
Table 1. State variables and initialization in the simulation
Variable
Agent
Norm
Image
Payoff
p
q
Environment
N
G
R
b
c
m
Description
Type of variable
Norm of agent
Images to other agents
Accumulated payoff of the giving game
Errors in perception
Errors in implementation
16 types
Binary (G/B)
Real Number
Constant
Constant
Number of agents
Generations of simulation
Times of playing giving game per generation
The benefit of giving game
The cost of giving game
The mutation ratio
Constant
Constant
Constant
Constant
Constant
Constant
Initial value
chosen randomly
G
0
{0, 0.001}
{0, 0.001}
500
1000
500
[3.0, 6.0]
1
0.01
7/15
Figures
A Giving game
B The norms of the giving game
(1) Playing giving game phase
Donor
Recipient
C for good recipient
or D for bad recipient
(2) Updating image phase
Recipient
Donor
Observer
Observer updates donor’s image
Donor
G/B
C
D
D
Bad
Good
Bad
Updated donor’s score
G/B
G/B
G/B
G/B
C
C
D
D
Good
Bad
Good
Bad
B
G
G
G
G
G
B
B
B
G
G
G
B
B
B
B
B
G
B
B
G
B
G
G
C Typical norms
C/D
G/B
Donor’s action
C
Recipient’s image from observers in
Good
last interaction
Donor’s action
Recipient’s image from observers in
last interaction
All Bad [ALLB]
Shunning [SH]
Stern Judging [SJ]
Image Scoring [IS]
Simple Standing [ST]
All Good [ALLG]
C/D
Recipient
Figure 1. The norms of cooperation and simulation framework. a, (1) If the donor’s image of a recipient is Good, the
donor gives the recipient something with personal cost c and the recipient receives benefit b. Nothing happens otherwise. (2)
In the Updating image phase, the observer updates the evaluation to the donor on the basis of the donor’s action (Cooperation
[C] / Defection [D]) and the observer’s evaluation (Good [G] / Bad [B]) of the recipient. b, Each agent adopts an evaluation
rule of the donor that depends on the donor’s action and the recipient’s image. This combination of Good/Bad is the norm
held by the agent. There are a total of 24 = 16 possible norms. In this phase, each agent evaluates and updates its image to all
donors. c, Typical norms can be expressed in the manner shown in this table. Typical norms include Shunning [SH] = GBBB,
Stern Judging [SJ] = GBBG, Image Scoring [IS] = GGBB, and Simple Standing [ST] = GGBG. SH is a strict norm where
any action for a Bad recipient is assessed as Bad. ST is a tolerant norm where any action for a Bad recipient is assessed as
Good. SJ is an intermediately strict norm where cooperation for a Bad recipient is assessed as Bad while defection is Good.
In contrast, IS does not use an image to recipient but uses only donor’s action. If the donor’s previous action is C, then IS
evaluates the donor as Good, otherwise IS evaluates the donor as Bad.
8/15
A With no error
Defective regime
Cooperative regime
B With errors
Defective regime
Cooperative regime
Figure 2. Time series of typical simulation runs with all norms. With no error (left panel) and with errors (right
panel). a, The average frequencies of 16 norms and the cooperation of the overall society. The black dotted line is the
cooperation ratio. Parameters: b = 5, c = 1, N = 500, R = 500, G = 1000, p = 0, q = 0. When SH and ALLB coexist,
cooperation does not emerge. When ALLB is completely driven out by SH, SJ invades and the cooperation ratio abruptly
rises. At the same time, SH is driven out by SJ. After cooperation is completely achieved, SJ permits the invasion of ST, and
also coexists with other tolerant norms (IS, ST, GGGB, and ALLG). Finally, strategies whose norm is expressed as GG** (in
other words, norms that constantly cooperate if cooperation has been selected in the past by the recipient) coexist. In b, both
errors in perception and implementation were introduced, and simulation similar to a was run
(b = 5, c = 1, N = 500, R = 500, G = 1000, p = 0.001, q = 0.001). As in a, when SH and ALLB coexist, cooperation does
not emerge. However, cooperation is achieved without going through SJ.
A With no error
B With errors
Figure 3. The alternation patterns of the majority of norms with 50 replications. With no error (left panel) and with
errors (right panel). a, The panel shows the transition of norms with the greatest populations in a round of 20 generations,
before the cooperation ratio exceeds 0.8, and 100 generations, after the cooperation ratio exceeds 0.8 (for a total of 120
generations). For the sake of visibility, in a replication, we stop calculation when ALLG becomes the majority norm. This is
because in a state in which tolerant norms coexist, the norms with the greatest population frequently change place. The
thickness of the arrows corresponds to the number of times alternation of norms occurred. (See the Supplementary
Information for details.) The alternation of norms SH → SJ → ST → ALLG was observed to be stable. In b , both errors in
perception and implementation were introduced, and simulations similar to a were run
(b = 5, c = 1, N = 500, R = 500, G = 1000, p = 0.001, q = 0.001). As shown in b, the transition of majority norms is not
distinct compared to the times when there were no errors.
9/15
A With no error
B With errors
Figure 4. The cooperation ratio in the norm knockout method. Each graph shows the average cooperation ratio of 50
replications when a typical norm is knocked out. The basic parameter set is c = 1, N = 500, R = 500, G = 1000. To confirm
the effects of errors in perception and errors in implementation, two simulations with and without error were executed. See
the Supplementary Information for knockout analysis of all norms. a, The case when errors in perception (p) and errors in
implementation (q) are 0. When SH or IS is knocked out, cooperation does not evolve at all. Also, when SJ, which becomes
the majority for only a brief round during the process of alternation, is knocked out, cooperation evolves to the extent of only
30 percent, even when b is large. Furthermore, when ST is knocked out, the range in which cooperation is achieved becomes
narrow. Only when b is sufficiently large can cooperation evolve. b, The case where p = q = 0.001. The indispensable norm
is ST in addition to SH and IS. Conversely, when SJ is knocked out, cooperation evolves when b is sufficiently large in the
same manner as ST in a.
A GBBB [SH] is knocked out
B GGBB [IS] is knocked out
Figure 5. Time series of typical simulation runs in norm knockout method. The parameters are
b = 5, c = 1, N = 500, R = 500, p = q = 0. a, When SH is knocked out, the strategy to eliminate ALLB does not exist. b,
When IS is knocked out, SH exists only in a small population and cannot gain superiority over ALLB.
10/15
Supplementary Information
This section includes:
• Supplementary Text S1
• Supplementary Tables S1 – S6
• Supplementary Figure S1
Text S1
The details of alternation of norms
We present the details of alternating the majority of norms, as shown in Fig. 3 (in the Main Text). Table S1 shows the
alternation with no error and Table S2 shows with errors.
The population of each norm
We show the average population ratio at the 1,000th generation (50 replications). As shown in Table S3, persistent norms are
able to be described as GG**.
The details of results of norm knockout method
We comprehensively explore the indispensable norms by knocking out each of the 16 norms. Table S4 shows the cooperation
ratio in which each norm is knocked out. Without errors, SH and IS are the indispensable norms. If SJ or ST is knocked out,
their standard deviations (S.D.) are larger than other cases. The reason for this is that to knock out SJ or ST produces two
contrasting results: cooperation dominant or defection dominant. When indispensable norms are knocked out, no cooperative
regime appears at all, and the standard deviation has a very small value.
The analysis of the alternation of norms after cooperative regime achieved
In this section, we have analysed the transition of norms with the greatest populations after a cooperation ratio exceeds 0.9
until the end of generations. For the sake of understanding the mechanism of co-evolution of norms and cooperation, we
have focused on the duration of regime changes from defection to cooperation in the main text. Therefore, we stopped the
calculation of the transition of the majority norm when ALLG becomes the majority norm in Fig. 3. Intuitively, it seems
impossible to transit from a majority of ALLG to a majority of any other norm; however, it is well-known that a state where
everyone cooperates indiscriminately is easily replaced by a state where everyone refuses to cooperate.
To clarify whether ALLG is a stable state and what would happen if more generations were considered, we show the
transition of norms with the greatest population after a cooperative regime is achieved. The results (Tables S5 and S6, Fig. S1)
show that the cooperation regime is maintained robustly and tolerant norms such as ALLG, GGGB, and ST coexist. Although
ALLG forms the majority of the population, the other norms including GGGB, ST (GGBG), and IS (GGBB) protect against
invasion from defective norms.
11/15
Tables
Table S1. The alternation patterns of dominant norms with no error correspond to the left panel of Fig. 3 (in the
Main Text). Each row shows the transition of norms with the greatest populations in a period of 20 generations, before the
cooperation ratio exceed s 0.8, and 100 generations, after the cooperation ratio exceeds 0.8 (for a total of 120 generations). For
the sake of visibility, we stop calculation when ALLG becomes the majority norm. Fifty replications are conducted. During
this time, alternation in majority norms for a total of 156 times could be observed. A cooperative regime with a cooperation
ratio exceeding 0.8 was achieved in 46 replications. For example, over 50 replications, the number of times the transition
of greatest population followed (SH → SJ → ST → ALLG) was 31. Moreover, four (indicated by the dash) never had a
cooperation ratio exceeding 0.8.
Transition pattern of dominant strategies
SH → SJ → ST → ALLC
SH → SJ → ST → SJ → ST → ALLC
−
SH → GBGB → GGGB → ALLC
SH → GBGB → SH → SJ → ST → ALLC
SH → IS → GGGB → ALLC
SH → IS → GGGB → SH → ALLD → SH → ALLD → SH → ALLD
SH → GGGB → ALLC
SH → GBGB → ST → ALLC
No.
31
5
4
3
2
2
1
1
1
Table S2. The alternation patterns of dominant norms with errors correspond to the right panel of Fig. 3 (in the Main
Text). The setting of the table is the same as Table S1.
Transition pattern of dominant strategies
SH → IS → GGGB → ALLC
SH → GBGB → GGGB → ALLC
SH → SJ → ST → ALLC
SH → GGGB → ALLC
SH → IS → ST → ALLC
SH → ST → ALLC
SH → GBGB → GGGB → SH → GBGB → GGGB → SH → IS → GGGB → ALLC
SH → IS → ST → IS → GGGB → ALLC
SH → GBGB → SH → IS → GGGB → ALLC
SH → IS → GGGB
SH → GBGB → IS → GGGB → GBGB → GGGB → ALLC
SH → SJ → IS → ST → ALLC
SH → GBGB → IS → GGGB → ALLC
SH → GBGB → SH → IS → ST → ALLC
SH → GBGB → GGGB → IS → GGGB → ALLC
No.
13
8
7
6
5
2
1
1
1
1
1
1
1
1
1
12/15
Table S3. The average population ratio at the 1,000th generation (b = 5). Each column shows without/with errors. All
cells are obtained by averaging the results of 50 replications. These results show the population of each norm in which all
norms exist (i.e., the norm knockout method is not used). The second row shows the average cooperation ratio (Cratio ) and
standard deviation at the 1,000th generation. Below the fourth row is shown the population of each norm and its standard
deviation. The norms described as GG** can coexist stably while any norm that is not GG** can barely exist. SH, which is
an indispensable norm, also cannot survive. IS, which is included in four persistent norms GG**, is the most in minority of
the four.
cratio (S.D.)
Norms
BBBB [ALLB]
BBBG
BBGB
BBGG
BGBB
BGBG
BGGB
BGGG
GBBB [SH]
GBBG [SJ]
GBGB
GBGG
GGBB [IS]
GGBG [ST]
GGGB
GGGG [ALLG]
p=q=0
0.939 (0.187)
population (S.D)
0.015 (0.085)
0.001 (0.002)
0.001 (0.003)
0.000 (0.001)
0.003 (0.007)
0.003 (0.003)
0.004 (0.004)
0.005 (0.004)
0.026 (0.109)
0.009 (0.006)
0.020 (0.013)
0.024 (0.012)
0.132 (0.043)
0.165 (0.073)
0.271 (0.093)
0.322 (0.090)
p = q = 0.001
0.980 (0.006)
population (S.D)
0.000 (0.001)
0.000 (0.000)
0.000 (0.001)
0.000 (0.001)
0.002 (0.002)
0.002 (0.002)
0.004 (0.003)
0.005 (0.003)
0.002 (0.003)
0.005 (0.004)
0.007 (0.005)
0.012 (0.006)
0.148 (0.040)
0.201 (0.071)
0.271 (0.079)
0.341 (0.064)
Table S4. Analysis of norm knockout method for each of the 16 norms. The table shows the cooperation ratio at the
1,000th generation in which each of the 16 norms is knocked out (b = 5). Each value shows the average cooperation ratio from
50 replications and standard deviation. The cells in which the average cooperation ratio is less than 0.1 are shown in red. In
this paper, we call these norms “indispensable norms”. With no error, SH and IS are indispensable norms. With errors, these
two plus ST are indispensable norms.
Knockouted norm
BBBB [ALLB]
BBBG
BBGB
BBGG
BGBB
BGBG
BGGB
BGGG
GBBB [SH]
GBBG [SJ]
GBGB
GBGG
GGBB [IS]
GGBG [ST]
GGGB
GGGG [ALLG]
Without Knockout
p=q=0
Mean (S.D.)
0.816 (0.354)
0.980 (0.008)
0.978 (0.012)
0.923 (0.226)
0.920 (0.225)
0.982 (0.008)
0.959 (0.134)
0.979 (0.011)
0.025 (0.004)
0.120 (0.287)
0.982 (0.007)
0.941 (0.188)
0.023 (0.006)
0.412 (0.432)
0.915 (0.225)
0.897 (0.179)
0.939 (0.187)
p = q = 0.001
Mean (S.D.)
0.745 (0.399)
0.979 (0.007)
0.979 (0.007)
0.961 (0.134)
0.922 (0.225)
0.977 (0.006)
0.978 (0.007)
0.959 (0.134)
0.026 (0.005)
0.616 (0.457)
0.977 (0.006)
0.978 (0.006)
0.022 (0.004)
0.055 (0.060)
0.961 (0.010)
0.371 (0.431)
0.980 (0.006)
13/15
Table S5. The number of transitions of norms with the greatest populations with no error after a cooperation ratio
exceeds 0.9 until the end of generations. The simulation runs 50 replications. The transition is counted when the most
majority norm is superseded by other norms. Parameters: b = 5, c = 1, N = 500, R = 500, G = 1000, p = 0, q = 0. For
example, the transition from ALLG to GGGB occurs 929 times. During this time, alternation of norms with the greatest
populations for a total of 2,324 times could be observed.
From
ALLG
GGGB
ST
ALLG
SJ
ST
IS
SH
ALLB
ALLG
IS
ST
GGGB
GGGB
SH
GGGB
ST
IS
To
GGGB
ALLG
ALLG
ST
ST
SJ
GGGB
ALLB
SH
IS
ST
IS
ST
SH
SJ
IS
GGGB
ALLG
No.
929
920
225
185
35
6
3
3
3
3
3
2
2
1
1
1
1
1
Table S6. The number of transitions of norms with the greatest populations with errors after cooperation ratio exceeds
0.9 until the end of generations. The setting of the table is the same as Table S5. Parameters: b = 5, c = 1, N = 500, R =
500, G = 1000, p = 0.001, q = 0.001.
From
GGGB
ALLG
ST
ALLG
GGGB
IS
ST
GGGB
IS
ALLG
ST
IS
To
ALLG
GGGB
ALLG
ST
IS
GGGB
GGGB
ST
ST
IS
IS
ALLG
No.
968
950
434
426
41
40
15
14
12
10
9
8
14/15
Figure
Figure S1. The transition diagram of norms with the greatest populations with no error (A) and with errors (B) after
cooperation ratio exceeds 0.9 until the end of generations. Panel A is drawn using the data of Table S5 and panel B is
drawn using the data of Table S6. Both panels show that the tolerant norms (such as ALLG, GGGB, and ST) coexist as the
majority.
15/15
| 9cs.NE
|
ON FEEBLY COMPACT SHIFT-CONTINUOUS TOPOLOGIES ON THE
SEMILATTICE expn λ
arXiv:1702.06105v2 [math.GN] 5 May 2017
OLEG GUTIK AND OLEKSANDRA SOBOL
Abstract. We study feebly compact topologies τ on the semilattice (expn λ, ∩) such that (expn λ, τ )
is a semitopological semilattice and prove that for any shift-continuous T1 -topology τ on expn λ the
following conditions are equivalent: (i) τ is countably pracompact; (ii) τ is feebly compact; (iii) τ is
d-feebly compact; (iv) (expn λ, τ ) is an H-closed space.
Dedicated to the memory of Professor Vitaly Sushchanskyy
We shall follow the terminology of [6, 8, 9, 13]. If X is a topological space and A ⊆ X, then by
clX (A) and intX (A) we denote the closure and the interior of A in X, respectively. By ω we denote
the first infinite cardinal and by N the set of positive integers.
A subset A of a topological space X is called regular open if intX (clX (A)) = A.
We recall that a topological space X is said to be
• quasiregular if for any non-empty open set U ⊂ X there exists a non-empty open set V ⊂ U
such that clX (V ) ⊆ U;
• semiregular if X has a base consisting of regular open subsets;
• compact if each open cover of X has a finite subcover;
• countably compact if each open countable cover of X has a finite subcover;
• countably compact at a subset A ⊆ X if every infinite subset B ⊆ A has an accumulation point
x in X;
• countably pracompact if there exists a dense subset A in X such that X is countably compact
at A;
• feebly compact (or lightly compact) if each locally finite open cover of X is finite [3];
• d-feebly compact (or DFCC ) if every discrete family of open subsets in X is finite (see [12]);
• pseudocompact if X is Tychonoff and each continuous real-valued function on X is bounded.
According to Theorem 3.10.22 of [8], a Tychonoff topological space X is feebly compact if and only if X
is pseudocompact. Also, a Hausdorff topological space X is feebly compact if and only if every locally
finite family of non-empty open subsets of X is finite [3]. Every compact space and every sequentially
compact space are countably compact, every countably compact space is countably pracompact, and
every countably pracompact space is feebly compact (see [2]), and every H-closed space is feebly
compact too (see [10]). Also, it is obvious that every feebly compact space is d-feebly compact.
A semilattice is a commutative semigroup of idempotents. On a semilattice S there exists a natural
partial order: e 6 f if and only if ef = f e = e. For any element e of a semilattice S we put
↑e = {f ∈ S : e 6 f } .
A topological (semitopological) semilattice is a topological space together with a continuous (separately continuous) semilattice operation. If S is a semilattice and τ is a topology on S such that (S, τ )
is a topological semilattice, then we shall call τ a semilattice topology on S, and if τ is a topology on S
such that (S, τ ) is a semitopological semilattice, then we shall call τ a shift-continuous topology on S.
Date: February 26, 2018.
2010 Mathematics Subject Classification. Primary 22A26, 22A15, Secondary 54D10, 54D30, 54H12.
Key words and phrases. Topological semilattice, semitopological semilattice, compact, countably compact, feebly
compact, H-closed, semiregular space, regular space.
1
2
OLEG GUTIK AND OLEKSANDRA SOBOL
For an arbitrary positive integer n and an arbitrary non-zero cardinal λ we put
expn λ = {A ⊆ λ : |A| 6 n} .
It is obvious that for any positive integer n and any non-zero cardinal λ the set expn λ with the
binary operation ∩ is a semilattice. Later in this paper by expn λ we shall denote the semilattice
(expn λ, ∩).
This paper is a continuation of [11] where we study feebly compact topologies τ on the semilattice expn λ such that (expn λ, τ ) is a semitopological semilattice. Therein, all compact semilattice
T1 -topologies on expn λ were described. In [11] it was proved that for an arbitrary positive integer n
and an arbitrary infinite cardinal λ every T1 -semitopological countably compact semilattice (expn λ, τ )
is a compact topological semilattice. Also, there we construct a countably pracompact H-closed
quasiregular non-semiregular topology τfc2 such that (exp2 λ, τfc2 ) is a semitopological semilattice with
the discontinuous semilattice operation and show that for an arbitrary positive integer n and an arbitrary infinite cardinal λ a semiregular feebly compact semitopological semilattice expn λ is a compact
topological semilattice.
In this paper we show that for any shift-continuous T1 -topology τ on expn λ the following conditions
are equivalent: (i) τ is countably pracompact; (ii) τ is feebly compact; (iii) τ is d-feebly compact;
(iv) (expn λ, τ ) is an H-closed space.
The proof of the following lemma is similar to Lemma 4.5 of [5] or Proposition 1 from [1].
Lemma 1. Every Hausdorff d-feebly compact topological space with a dense discrete subspace is countably pracompact.
We observe that by Proposition 1 from [11] for an arbitrary positive integer n and an arbitrary infinite
cardinal λ every shift-continuous T1 -topology τ on expn λ is functionally Hausdorff and quasiregular,
and hence it is Hausdorff.
Proposition 1. Let n be an arbitrary positive integer and λ be an arbitrary infinite cardinal. Then for
every d-feebly compact shift-continuous T1 -topology τ on expn λ the subset expn λ \ expn−1 λ is dense
in (expn λ, τ ).
Proof. Suppose to the contrary that there exists a d-feebly compact shift-continuous T1 -topology τ on
expn λ such that expn λ \ expn−1 λ is not dense in (expn λ, τ ). Then there exists a point x ∈ expn−1 λ
of the space (expn λ, τ ) such that x ∈
/ clexpn λ (expn λ \ expn−1 λ). This implies that there exists an open
neighbourhood U(x) of x in (expn λ, τ ) such that U(x) ∩ expn λ \ expn−1 λ = ∅. The definition of
the semilattice expn λ implies that every maximal chain in expn λ is finite and hence there exists a
point y ∈ U(x) such that ↑y ∩ U(x) = {y}. By Proposition 1(iii) from [11], ↑y is an open-and-closed
subset of (expn λ, τ ) and hence ↑y is a d-feebly compact subspace of (expn λ, τ ).
It is obvious that the subsemilattice ↑y of expn λ is algebraically isomorphic to the semilattice expk λ
for some positive integer k 6 n. This and above arguments imply that without loss of generality we
may assume that y is the isolated zero of the d-feebly compact semitopological semilattice (expn λ, τ ).
Hence we assume that τ is a d-feebly compact shift-continuous topology on expn λ such that the
zero 0 of expn λ is an isolated point of (expn λ, τ ). Next we fix an arbitrary infinite sequence {xi }i∈N
of distinct elements of cardinal λ. For every positive integer j we put
aj = xn(j−1)+1 , xn(j−1)+2 , . . . , xnj .
Then aj ∈ expn λ and moreover aj is a greatest element of the semilattice expn λ for each positive
integer j. Also, the definition of the semilattice expn λ implies that for every non-zero element a
of expn λ there exists at most one element aj such that aj ∈ ↑a. Then for every positive integer j
by Proposition 1(iii) of [11], aj is an isolated point of (expn λ, τ ), and hence the above arguments
imply that {a1 , a2 , . . . , aj , . . .} is an infinite discrete family of open subset in the space (expn λ, τ ).
This contradicts the d-feeble compactness of the semitopological semilattice (expn λ, τ ). The obtained
contradiction implies the statement of our proposition.
ON FEEBLY COMPACT SHIFT-CONTINUOUS TOPOLOGIES ON THE SEMILATTICE expn λ
3
The following example show that the converse statement to Proposition 1 is not true in the case of
topological semilattices.
Example 1. Fix an arbitrary cardinal λ and an infinite subset A in λ such that |λ \ A| > ω. By
π : λ → exp1 λ : a 7→ {a} we denote the natural embedding of λ into exp1 λ. On exp1 λ we define a
topology τdm in the following way:
(i) all non-zero elements of the semilattice exp1 λ are isolated points in (exp1 λ, τdm ); and
(ii) the family Bdm = {UB = {0} ∪ π(B) : B ⊆ A and A \ B is finite} is the base of the topology
τdm at zero 0 of exp1 λ.
Simple verifications show that τdm is a Hausdorff locally compact semilattice topology on exp1 λ which
is not compact and hence by Corollary 8 of [11] it is not feebly compact.
Remark 1. We observe that in the case when λ = ω by Proposition 13 of [11] the topological space
(exp1 λ, τdm ) is collectionwise normal and it has a countable base, and hence (exp1 λ, τdm ) is metrizable
by the Urysohn Metrization Theorem [14]. Moreover, if |B| = ω then the space (exp1 λ, τdm ) is
metrizable for any infinite cardinal λ, as a topological sum of the metrizable space (exp1 ω, τdm ) and
the discrete space of cardinality λ.
Remark 2. If n is an arbitrary positive integer > 3, λ is any infinite cardinal and τcn is the unique
compact semilattice topology on the semilattice expn λ defined in Example 4 of [11], then we construct
n
on expn λ them τcn in the following way. Fix an arbitrary element x ∈ expn λ
more stronger topology τdm
such that |x| = n − 1. It is easy to see that the subsemilattice ↑x of expn λ is isomorphic to exp1 λ,
and by h : exp1 λ → ↑x we denote this isomorphism.
Fix an arbitrary subset A in λ such that |λ \ A| > ω. For every zero element y ∈ expn λ \ ↑x we
n
n
assume that the base Bdm
(y) of the topology τdm
at the point y coincides with the base of the topology
n
τc at y, and assume that ↑x is an open-and-closed subset and the topology on ↑x is generated by the
n
map h : (exp2 λ, τfc2 ) → ↑x. We observe that (expn λ, τdm
) is a Hausdorff locally compact topological
space, because it is the topological sum of a Hausdorff locally compact space ↑x (which is homeomorphic
to the Hausdorff locally compact space (exp1 λ, τdm ) from Example 1) and an open-and-closed subspace
n
expn λ\↑x of (expn λ, τcn ). It is obvious that the set expn λ\expn−1 λ is dense in (expn λ, τdm
). Also, since
n
↑x is an open-and-closed subsemilattice with zero x of (expn λ, τdm ), the continuity of the semilattice
n
n
operations in (expn λ, τdm
) and (expn λ, τcn ) and the property that the topology τdm
is more stronger
n
n
n
) is not
them τc , imply that (expn λ, τdm ) is a topological semilattice. Moreover, the space (expn λ, τdm
d-feebly compact, because it contains an open-and-closed non-d-feebly compact subspace ↑x.
Arguments presented in the proof of Proposition 1 and Proposition 1(iii) of [11] imply the following
corollary.
Corollary 1. Let n be an arbitrary positive integer and λ be an arbitrary infinite cardinal. Then for
every d-feebly compact shift-continuous T1 -topology τ on expn λ a point x is isolated in (expn λ, τ ) if
and only if x ∈ expn λ \ expn−1 λ.
Remark 3. We observe that the example presented in Remark 2 implies there exists a locally compact
n
non-d-feebly compact semitopological semilattice (expn λ, τdm
) with the following property: a point x
n
is isolated in (expn λ, τdm ) if and only if x ∈ expn λ \ expn−1 λ.
The following proposition gives an amazing property of the system of neighbourhoodd of zero in a
T1 -feebly compact semitopological semilattice expn λ.
Proposition 2. Let n be an arbitrary positive integer, λ be an arbitrary infinite cardinal and τ be a
shift-continuous feebly compact T1 -topology on the semilattice expn λ. Then for every open neighbourhood U(0) of zero 0 in (expn λ, τ ) there exist finitely many x1 , . . . , xm ∈ λ such that
expn λ \ clexpn λ (U(0)) ⊆ ↑x1 ∪ · · · ∪ ↑xm .
4
OLEG GUTIK AND OLEKSANDRA SOBOL
Proof. Suppose to the contrary that there exists an open neighbourhood U(0) of zero in a Hausdorff
feebly compact semitopological semilattice (expn λ, τ ) such that
expn λ \ clexpn λ (U(0)) 6⊆ ↑x1 ∪ · · · ∪ ↑xm
for any finitely many x1 , . . . , xm ∈ λ.
We fix an arbitrary y1 ∈ λ such that expn λ \ clexpn λ (U(0)) ∩↑y1 6= ∅. By Proposition 1(iii) of [11]
the set ↑y1 is open in (expn λ, τ ) and hence the set expn λ \ clexpn λ (U(0)) ∩ ↑y1 is open in (expn λ, τ )
too. Then by Proposition 1 there exists an isolated point m1 ∈ expn λ \ expn−1 λ in (expn λ, τ ) such
that m1 ∈ expn λ \ clexpn λ (U(0)) ∩ ↑y1 . Now, by the assumption there exists y2 ∈ λ such that
expn λ \ clexpn λ (U(0)) ∩ (↑y2 \ ↑y1 ) 6= ∅.
Again, since by Proposition 1(iii) of [11] both sets ↑y1 and ↑y2 are open-and-closed in (expn λ, τ ),
Proposition 1 implies that there exists an isolated point m2 ∈ expn λ \ expn−1 λ in (expn λ, τ ) such that
m2 ∈ expn λ \ clexpn λ (U(0)) ∩ (↑y2 \ ↑y1 ) .
Hence by induction we can construct a sequence {yi : i = 1, 2, 3, . . .} of distinct points of λ and a
sequence of isolated points {mi : i = 1, 2, 3, . . .} ⊂ expn λ \ expn−1 λ in (expn λ, τ ) such that for any
positive integer k the following conditions hold:
(i) expn λ \ clexpn λ (U(0)) ∩ (↑yk \ (↑y1 ∪ · · · ∪ ↑yk−1 )) 6= ∅; and
(ii) mk ∈ expn λ \ clexpn λ (U(0)) ∩ (↑yk \ (↑y1 ∪ · · · ∪ ↑yk−1 )).
Then similar arguments as in the proof of Proposition 1 imply that the following family
{{mi } : i = 1, 2, 3, . . .}
is infinite and locally finite, which contradicts the feeble compactness of (expn λ, τ ). The obtained
contradiction implies the statement of the proposition.
Proposition 1(iii) of [11] implies that for any element x ∈ expn λ the set ↑x is open-and-closed
in a T1 -semitopological semilattice (expn λ, τ ) and hence by Theorem 14 from [3] we have that for
any x ∈ expn λ the space ↑x is feebly compact in a feebly compact T1 -semitopological semilattice
(expn λ, τ ). Hence Proposition 2 implies the following proposition.
Proposition 3. Let n be an arbitrary positive integer, λ be an arbitrary infinite cardinal and τ be a
shift-continuous feebly compact T1 -topology on the semilattice expn λ. Then for any point x ∈ expn λ
and any open neighbourhood U(x) of x in (expn λ, τ ) there exist finitely many x1 , . . . , xm ∈ ↑x \ {x}
such that
↑x \ clexpn λ (U(x)) ⊆ ↑x1 ∪ · · · ∪ ↑xm .
The main results of this paper is the following theorem.
Theorem 1. Let n be an arbitrary positive integer and λ be an arbitrary infinite cardinal. Then for
any shift-continuous T1 -topology τ on expn λ the following conditions are equivalent:
(i) τ is countably pracompact;
(ii) τ is feebly compact;
(iii) τ is d-feebly compact;
(iv) the space (expn λ, τ ) is H-closed.
Proof. Implications (i) ⇒ (ii) and (ii) ⇒ (iii) are trivial and implication (iii) ⇒ (i) follows from
Proposition 1 of [11], Lemma 1 and Proposition 1.
Implication (iv) ⇒ (ii) follows from Proposition 4 of [10].
(ii) ⇒ (iv) We shall prove this implication by induction.
By Corollary 2 from [11] every feebly compact T1 -topology τ on the semilattice exp1 λ such that
(exp1 λ, τ ) is a semitopological semilattice, is compact, and hence (exp1 λ, τ ) is an H-closed topological
space.
ON FEEBLY COMPACT SHIFT-CONTINUOUS TOPOLOGIES ON THE SEMILATTICE expn λ
5
Next we shall show that if our statements holds for all positive integers j < k 6 n then it holds
for j = k. Suppose that a feebly compact T1 -semitopological semilattice (expk λ, τ ) is a subspace of
Hausdorff topological space X. Fix an arbitrary point x ∈ X and an arbitrary open neighbourhood
V (x) of x in X. Since X is Hausdorff, there exist disjoint open neighbourhoods U(x) ⊆ V (x) and U(0)
of x and zero 0 of the semilattice expk λ in X, respectively. Then clX (U(0)) ∩ U(x) = ∅ and hence by
Proposition 2 there exists finitely many x1 , . . . , xm ∈ λ such that
expk λ ∩ U(x) ⊆ ↑x1 ∪ · · · ∪ ↑xm .
But for any x ∈ λ the subsemilattice ↑x of expk λ is algebraically isomorphic to the semilattice expk−1 λ.
Then by Proposition 1(iii) of [11] and Theorem 14 from [3], ↑x is a feebly compact T1 -semitopological
semilattice, and the assumption of our induction implies that ↑x1 , · · · , ↑xm are closed subsets of X.
This implies that
W (x) = U(x) \ (↑x1 ∪ · · · ∪ ↑xm )
is an open neighbourhood of x in X such that W (x) ∩ expk λ = ∅. Thus, (expk λ, τ ) is an H-closed
space. This completes the proof of the requested implication.
The following theorem gives a sufficient condition when a d-feebly compact space is feebly compact.
Theorem 2. Every quasiregular d-feebly compact space is feebly compact.
Proof. Suppose to the contrary that there exists a quasiregular d-feebly compact space X which is not
feebly compact. Then there exists an infinite locally finite family U0 of non-empty open subsets of X.
By induction we shall construct an infinite discrete family of non-empty open subsets of X.
Fix an arbitrary U1 ∈ U0 and an arbitrary point x1 ∈ U1 . Since the family U0 is locally finite
there exists an open neighbourhood U(x1 ) ⊆ U1 of the point x1 in X such that U(x1 ) intersects finitely
many elements of U0 . Also, the quasiregularity of X implies that there exists a non-empty open subset
V1 ⊆ U(x1 ) such that clX (V1 ) ⊆ U(x1 ). Put
U1 = {U ∈ U0 : U(x1 ) ∩ U = ∅} .
Since the family U0 is locally finite and infinite, so is U1 . Fix an arbitrary U2 ∈ U1 and an arbitrary
point x2 ∈ U2 . Since the family U1 is locally finite, there exists an open neighbourhood U(x2 ) ⊆ U2
of the point x2 in X such that U(x2 ) intersects finitely many elements of U1 . Since X is quasiregular,
there exists a non-empty open subset V2 ⊆ U(x2 ) such that clX (V2 ) ⊆ U(x2 ). Our construction implies
that the closed sets clX (V1 ) and clX (V2 ) are disjoint and hence so are V1 and V2 . Next we put
U2 = {U ∈ U1 : U(x2 ) ∩ U = ∅} .
Also, we observe that it is obvious that U(x1 ) ∩ U = ∅ for each U ∈ U1 .
Suppose for some positive integer k > 1 we construct:
(a) a sequence of infinite locally finite subfamilies U1 , . . . , Uk−1 in U0 of non-empty open subsets
in the space X;
(b) a sequence of open subsets U1 , . . . , Uk in X;
(c) a sequence of points x1 , . . . , xk in X and a sequence of their corresponding open neighbourhoods
U(x1 ), . . . , U(xk ) in X;
(d) a sequence of disjoint non-empty subsets V1 , . . . , Vk in X
such that the following conditions hold:
(i) Ui is a proper subfamily of Ui−1 ;
(ii) Ui ∈ Ui−1 and Ui ∩ U = ∅ for each U ∈ Uj with i 6 j 6 k;
(iii) xi ∈ Ui and U(xi ) ⊆ Ui ;
(iv) Vi is an open subset of Ui with clX (Vi ) ⊆ U(xi ),
for all i = 1, . . . , k, and
(v) clX (V1 ), . . . , clX (Vk ) are disjoint.
6
OLEG GUTIK AND OLEKSANDRA SOBOL
Next we put
Uk = {U ∈ Uk−1 : U(x1 ) ∩ U = . . . = U(xk ) ∩ U = ∅} .
Since the family Uk−1 is infinite and locally finite, there exists a subfamily Uk in Uk−1 which is infinite
and locally finite. Fix an arbitrary Uk+1 ∈ Uk and an arbitrary point xk+1 ∈ Uk+1 . Since the family
Uk is locally finite, there exists an open neighbourhood U(xk+1 ) ⊆ Uk+1 of the point xk+1 in X such
that U(xk+1 ) intersects finitely many elements of Uk . Since the space X is quasiregular, there exists
a non-empty open subset Vk+1 ⊆ U(xk+1 ) such that clX (Vk+1 ) ⊆ U(xk+1 ). Simple verifications show
that the conditions (i) − (iv) hold in the case of the positive integer k + 1.
Hence by induction we construct the following two infinite countable families of open non-empty
subsets of X:
U = {Ui : i = 1, 2, 3, . . .}
and
V = {Vi : i = 1, 2, 3, . . .}
such that clX (Vi ) ⊆ Ui for each positive integer i. Since U is a subfamily of U0 and U0 is locally finite
in X, U is locally finite in X as well. Also, above arguments imply that V and
V = {clX (Vi ) : i = 1, 2, 3, . . .}
are locally finite families in X too.
Next we shall show that the family VSis discrete in X. Indeed, since the family V is locally finite
S in
X, by Theorem 1.1.11 of [8] the union V Sis a closed subset of X, and hence any point x ∈ X \ V
has an open neighbourhood O(x) = X \ V which does not intersect the elements of the family
V . If x ∈ clX (Vi ) for some positive integer i, then our construction implies that U(xi ) is an open
neighbourhood of x which intersects only the set Vi ∈ V . Hence X has an infinite discrete family V of
non-empty open subsets in X, which contradicts the assumption that the space X is d-feebly compact.
The obtained contradiction implies the statement of the theorem.
We finish this note by some simple remarks about dense embedding of an infinite semigroup of
matrix units and a polycyclic monoid into d-feebly compact topological semigroups which follow from
the results of the paper [5].
Let λ be a non-zero cardinal. On the set Bλ = (λ × λ) ∪ {0}, where 0 ∈
/ λ × λ, we define the
semigroup operation “ · ” as follows
(a, d), if b = c;
(a, b) · (c, d) =
0,
if b 6= c,
and (a, b) · 0 = 0 · (a, b) = 0 · 0 = 0 for a, b, c, d ∈ λ. The semigroup Bλ is called the semigroup of
λ×λ-matrix units (see [7]).
The bicyclic monoid C (p, q) is the semigroup with the identity 1 generated by two elements p and
q subjected only to the condition pq = 1 [7]. For a non-zero cardinal λ, the polycyclic monoid Pλ on
λ generators is the semigroup with zero given by the presentation:
D
E
−1
−1
Pλ = {pi }i∈λ , p−1
|
p
p
=
1,
p
p
=
0
for
i
=
6
j
i
i
i
i
j
i∈λ
(see [5]). It is obvious that in the case when λ = 1 the semigroup P1 is isomorphic to the bicyclic
semigroup with adjoined zero.
By Theorem 4.4 from [5] for every infinite cardinal λ the semigroup of λ×λ-matrix units Bλ does
not densely embed into a Hausdorff feebly compact topological semigroup, and by Theorem 4.5 from
[5] for arbitrary cardinal λ > 2 there exists no Hausdorff feebly compact topological semigroup which
contains the λ-polycyclic monoid Pλ as a dense subsemigroup. These theorems and Lemma 1 imply
the following two corollaries.
Corollary 2. For every infinite cardinal λ the semigroup of λ×λ-matrix units Bλ does not densely
embed into a Hausdorff d-feebly compact topological semigroup.
Corollary 3. For arbitrary cardinal λ > 2 there exists no Hausdorff d-feebly compact topological
semigroup which contains the λ-polycyclic monoid Pλ as a dense subsemigroup.
ON FEEBLY COMPACT SHIFT-CONTINUOUS TOPOLOGIES ON THE SEMILATTICE expn λ
7
The proof of the following corollary is similar to Theorem 5.1(5) from [4].
Corollary 4. There exists no Hausdorff topological semigroup with the d-feebly compact square which
contains the bicyclic monoid C (p, q) as a dense subsemigroup.
Acknowledgements
We acknowledge Alex Ravsky and the referee for their comments and suggestions.
References
[1] A. V. Arhangel’skii, On spaces with point-countable base, Topological Spaces and Their Mappings, Riga, 1985,
P. 3–7 (in Russian).
[2] A. V. Arkhangel’skii, Topological Function Spaces, Kluwer Publ., Dordrecht, 1992.
[3] R. W. Bagley, E. H. Connell, and J. D. McKnight, Jr., On properties characterizing pseudo-compact spaces, Proc.
Amer. Math. Soc. 9:3 (1958), 500–506.
[4] T. Banakh, S. Dimitrova, and O. Gutik, Embedding the bicyclic semigroup into countably compact topological
semigroups, Topology Appl. 157:18 (2010), 2803–2814.
[5] S. Bardyla and O. Gutik, On a semitopological polycyclic monoid, Algebra Discr. Math. 21:2 (2016), 163–183.
[6] J. H. Carruth, J. A. Hildebrant and R. J. Koch, The Theory of Topological Semigroups, Vol. I, Marcel Dekker, Inc.,
New York and Basel, 1983; Vol. II, Marcel Dekker, Inc., New York and Basel, 1986.
[7] A. H. Clifford and G. B. Preston, The Algebraic Theory of Semigroups, Vols. I and II, Amer. Math. Soc. Surveys
7, Providence, R.I., 1961 and 1967.
[8] R. Engelking, General Topology, 2nd ed., Heldermann, Berlin, 1989.
[9] G. Gierz, K. H. Hofmann, K. Keimel, J. D. Lawson, M. W. Mislove, and D. S. Scott, Continuous Lattices and
Domains, Cambridge Univ. Press, Cambridge, 2003.
[10] O. V. Gutik and O. V. Ravsky, Pseudocompactness, products and topological Brandt λ0 –extensions of semitopological
monoids, Math. Methods and Phys.-Mech. Fields 58:2 (2015), 20–37; reprinted version: J. Math. Sci. 223:1 (2017),
18-38.
[11] O. Gutik and O. Sobol, On feebly compact topologies on the semilattice expn λ, Mat. Stud. 46:1 (2016), 29–43.
[12] M. Matveev, A survey of star covering properties, Topology Atlas preprint, April 15, 1998.
[13] W. Ruppert, Compact Semitopological Semigroups: An Intrinsic Theory, Lect. Notes Math., 1079, Springer, Berlin,
1984.
[14] P. Urysohn, Zum Metrisationsproblem, Math. Ann. 94 (1925), 309–315.
Faculty of Mechanics and Mathematics, National University of Lviv, Universytetska 1, Lviv, 79000,
Ukraine
E-mail address: o [email protected], [email protected], [email protected]
| 4math.GR
|
Unit Interval Editing is Fixed-Parameter Tractable
Yixin Cao∗
arXiv:1504.04470v2 [cs.DS] 31 Dec 2016
Abstract
Given a graph G and integers k1 , k2 , and k3 , the unit interval editing problem asks whether G can be
transformed into a unit interval graph by at most k1 vertex deletions, k2 edge deletions, and k3 edge
additions. We give an algorithm solving this problem in time 2O(k log k) · (n + m), where k := k1 + k2 + k3 ,
and n, m denote respectively the numbers of vertices and edges of G. Therefore, it is fixed-parameter
tractable parameterized by the total number of allowed operations.
Our algorithm implies the fixed-parameter tractability of the unit interval edge deletion problem, for
which we also present a more efficient algorithm running in time O(4k · (n + m)). Another result is an
O(6k · (n + m))-time algorithm for the unit interval vertex deletion problem, significantly improving the
algorithm of van ’t Hof and Villanger, which runs in time O(6k · n6 ).
1
Introduction
A graph is a unit interval graph if its vertices can be assigned to unit-length intervals on the real line such
that there is an edge between two vertices if and only if their corresponding intervals intersect. Most
important applications of unit interval graphs were found in computational biology [5, 16, 18], where
data are mainly obtained by unreliable experimental methods. Therefore, the graph representing the
raw data is very unlikely to be a unit interval graph, and an important step before understanding the
data is to find and fix the hidden errors. For this purpose various graph modification problems have been
formulated: Given a graph G on n vertices and m edges, is there a set of at most k modifications that
make G a unit interval graph. In particular, edge additions, also called completion, and edge deletions are
used to fix false negatives and false positives respectively, while vertex deletions can be viewed as the
elimination of outliers. We have thus three variants, which are all known to be NP-complete [21, 30, 16].
These modification problems to unit interval graphs have been well studied in the framework of
parameterized computation, where the parameter is usually the number of modifications. Recall that a
graph problem, with a nonnegative parameter k, is fixed-parameter tractable (FPT) if there is an algorithm
solving it in time f(k) · (n + m)O(1) , where f is a computable function depending only on k [12]. The
problems unit interval completion and unit interval vertex deletion have been shown to be FPT by Kaplan
et al. [18] and van Bevern et al. [3] respectively. In contrast, however, the parameterized complexity of the
edge deletion version remained open to date, which we settle here. Indeed, we devise single-exponential
linear-time parameterized algorithms for both deletion versions.
Theorem 1.1. The problems unit interval vertex deletion and unit interval edge deletion can be solved in
time O(6k · (n + m)) and O(4k · (n + m)) respectively.
Our algorithm for unit interval vertex deletion significantly improves the currently best parameterized
algorithm for it, which takes O(6k · n6 ) time [17]. Another algorithmic result of van ’t Hof and Villanger
[17] is an O(n7 )-time 6-approximation algorithm for the problem, which we improve to the following.
Theorem 1.2. There is an O(nm + n2 )-time approximation algorithm of approximation ratio 6 for the
minimization version of the unit interval vertex deletion problem.
The structures and recognition of unit interval graphs have been well studied and well understood
[11]. It is known that a graph is a unit interval graph if and only if it contains no claw, S3 , S3 , (as depicted
in Fig. 1,) or any hole (i.e., an induced cycle on at least four vertices) [26, 29]. Unit interval graphs are
thus a subclass of chordal graphs, which are those graphs containing no holes. Modification problems
∗ Department of Computing, Hong Kong Polytechnic University, Hong Kong, China. [email protected]. Supported in part
by the Hong Kong Research Grants Council (RGC) under grant 252026/15E, the National Natural Science Foundation of China
(NSFC) under grants 61572414 and 61420106009, the Hong Kong Polytechnic University (PolyU) under grant 4-ZZEZ, and the
European Research Council (ERC) under grant 280152.
1
u3
v2
u1
(a) claw
v3
u3
v1
v3
(b) S3
u2
v1 u1
u2 v2
(d) W5
(c) S3
(e) C∗5 = W5
(f) C6
Figure 1: Small forbidden induced graphs.
to chordal graphs and unit interval graphs are among the earliest studied problems in parameterized
computation, and their study had been closely related. For example, the algorithm of Kaplan et al. [18]
for unit interval completion is a natural spin-off of their algorithm for chordal completion, or more
specifically, the combinatorial result of all minimal ways to fill holes in. A better analysis was shortly
done by Cai [7], who also made explicit the use of bounded-depth search in disposing of finite forbidden
induced subgraphs. This observation and the parameterized algorithm of Marx [24] for the chordal vertex
deletion problem immediately imply the fixed-parameter tractability of the unit interval vertex deletion
problem: One may break first all induced claws, S3 ’s, and S3 ’s, and then call Marx’s algorithm. Here we
are using the hereditary property of unit interval graphs,—recall that a graph class is hereditary if it is
closed under taking induced subgraphs. However, neither approach can be adapted to the edge deletion
version in a simple way. Compared to completion that needs to add Ω(`) edges to fill a C` (i.e., a hole
of length `) in, an arbitrarily large hole can be fixed by a single edge deletion. On the other hand, the
deletion of vertices leaves an induced subgraph, which allows us to focus on holes once all claws, S3 ’s, and
S3 ’s have been eliminated; however, the deletion of edges to fix holes of a {claw, S3 , S3 }-free graph may
introduce new claws, S3 ’s, and/or S3 ’s. Therefore, although a parameterized algorithm for the chordal
edge deletion problem has also been presented by Marx [24], there is no obvious way to use it to solve
the unit interval edge deletion problem.
Direct algorithms for unit interval vertex deletion were later discovered by van Bevern et al. [3] and
van ’t Hof and Villanger [17], both using a two-phase approach. The first phase of their algorithms
breaks all forbidden induced subgraphs on at most six vertices. Note that this differentiates from the
aforementioned simple approach in that it breaks not only claws, S3 ’s, and S3 ’s, but all C` ’s with ` 6 6.
Although this phase is conceptually intuitive, it is rather nontrivial to efficiently carry it out, and the simple
brute-force way introduces an n6 factor to the running time. Their approaches diverse completely in the
second phase. Van Bevern et al. [3] used a complicated iterative compression procedure that has a very
high time complexity, while van ’t Hof and Villanger [17] showed that after the first phase, the problem
is linear-time solvable. The main observation of van ’t Hof and Villanger [17] is that a connected {claw,
S3 , S3 , C4 , C5 , C6 }-free graph is a proper circular-arc graph, whose definition is postponed to Section 2.
In the conference presentation where Villanger first announced the result, it was claimed that this settles
the edge deletion version as well. However, the claimed result has not been materialized: It appears
neither in the conference version [28] (which has a single author) nor in the significantly revised and
extended journal version [17]. Unfortunately, this unsubstantiated claim did get circulated.
Although the algorithm of van ’t Hof and Villanger [17] is nice and simple, its self-contained proof is
excruciatingly complex. We revisit the relation between unit interval graphs and some subclasses of proper
circular-arc graphs, and study it in a structured way. In particular, we observe that unit interval graphs
are precisely those graphs that are both chordal graphs and proper Helly circular-arc graphs. As a matter
of fact, unit interval graphs can also be viewed as “unit Helly interval graphs” or “proper Helly interval
graphs,” thereby making a natural subclass of proper Helly circular-arc graphs. The full containment
relations are summarized in Fig. 2; the reader unfamiliar with some graph classes in this figure may
turn to the appendix for a brief overview. These observations inspire us to show that a connected {claw,
S3 , S3 , C4 , C5 }-free graph is a proper Helly circular-arc graph. It is easy to adapt the linear-time certifying
recognition algorithms for proper Helly circular-arc graphs [22, 9] to detect an induced claw, S3 , S3 , C4 ,
or C5 if one exists. Once all of them have been completely eliminated and the graph is a proper Helly
circular-arc graph, it is easy to solve the unit interval vertex deletion problem in linear time. Likewise,
using the structural properties of proper Helly circular-arc graphs, we can derive a linear-time algorithm
for unit interval edge deletion on them. It is then straightforward to use simple branching to develop the
parameterized algorithms stated in Theorem 1.1, though some nontrivial analysis is required to obtain the
time bound for unit interval edge deletion.
Van Bevern et al. [3] showed that the unit interval vertex deletion problem remains NP-hard on
2
Normal Helly Circular-arc
Proper Helly Circular-arc
Chordal
Interval
Unit Helly Circular-arc
Unit interval = Proper interval
Figure 2: Containment relation of related graph classes.
Normal Helly circular-arc ∩ Chordal = Interval.
Proper Helly circular-arc ∩ Chordal = Unit Helly circular-arc ∩ Chordal = Unit interval.
{claw, S3 , S3 }-free graphs. After deriving a polynomial-time algorithm for the problem on {claw,
S3 , S3 , C4 , C5 , C6 }-free graphs, van ’t Hof and Villanger [17] asked for its complexity on {claw, S3 , S3 , C4 }free graphs. (It is somewhat intriguing that they did not mention the {claw, S3 , S3 , C4 , C5 }-free graphs.)
Note that a {claw, S3 , S3 , C4 }-free graph is not necessarily a proper (Helly) circular-arc graph, evidenced
by the W5 (Fig. 1d). We answer this question by characterizing connected {claw, S3 , S3 , C4 }-free graphs
that are not proper Helly circular-arc graphs. We show that such a graph must be like a W5 : If we keep
only one vertex from each twin class (all vertices in a twin class have the same closed neighborhood) of
the original graph, then we obtain a W5 . It is then routine to solve the problem in linear time.
Theorem 1.3. The problems unit interval vertex deletion and unit interval edge deletion can be solved in
O(n + m) time on {claw, S3 , S3 , C4 }-free graphs.
We remark that the techniques we developed in previous work [8] can also be used to derive
Theorems 1.1 and 1.2. Those techniques, designed for interval graphs, are nevertheless far more
complicated than necessary when applied to unit interval graphs. The approach we used in the current
work, i.e., based on structural properties of proper Helly circular-arc graphs, is tailored for unit interval
graphs, hence simpler and more natural. Another benefit of this approach is that it enables us to devise a
parameterized algorithm for the general modification problem to unit interval graphs, which allows all
three types of operations. This formulation generalizes all the three single-type modifications, and is also
natural from the viewpoint of the aforementioned applications for de-noising data, where different types
of errors are commonly found coexisting. Indeed, the assumption that the input data contain only a single
type of errors is somewhat counterintuitive. Formally, given a graph G, the unit interval editing problem
asks whether there are a set V− of at most k1 vertices, a set E− of at most k2 edges, and a set E+ of at
most k3 non-edges, such that the deletion of V− and E− and the addition of E+ make G a unit interval
graph. We show that it is FPT, parameterized by the total number of allowed operations, k := k1 + k2 + k3 .
Theorem 1.4. The unit interval editing problem can be solved in time 2O(k log k) · (n + m).
By and large, our algorithm for unit interval editing again uses the two-phase approach. However,
we are not able to show that it can be solved in polynomial time on proper Helly circular-arc graphs.
Therefore, in the first phase, we use brute force to remove not only claws, S3 ’s, S3 ’s, and C4 ’s, also all
holes of length at most k3 + 3. The high exponential factor in the running time is due to purely this phase.
After that, every hole has length at least k3 + 4, and has to be fixed by deleting a vertex or edge. We
manage to show that an inclusion-wise minimal solution of this reduced graph does not add edges, and
the problem can then be solved in linear time.
The study of general modification problems was initiated by Cai [7], who observed that the problem
is FPT if the objective graph class has a finite number of minimal forbidden induced subgraphs. More
challenging is thus to devise parameterized algorithms for those graph classes whose minimal forbidden
induced subgraphs are infinite. Prior to this paper, the only known nontrivial graph class on which the
general modification problem is FPT is the chordal graphs [10]. Theorem 1.4 extends this territory by
including another well-studied graph class. As a corollary, Theorem 1.4 implies the fixed-parameter
tractability of the unit interval edge editing problem, which allows both edge operations but not vertex
deletions [6]. To see this we can simply try every combination of k2 and k3 as long as k2 + k3 does not
exceed the given bound.
3
Organization. The rest of the paper is organized as follows. Section 2 presents combinatorial and
algorithmic results on {claw, S3 , S3 , C4 }-free graphs. Sections 3 and 4 present the algorithms for unit
interval vertex deletion and unit interval edge deletion respectively (Theorems 1.1–1.3). Section 5 extends
them to solve the general editing problem (Theorem 1.4). Section 6 closes this paper by discussing some
possible improvement and new directions. The appendix provides a brief overview of related graph classes
as well as their characterizations by forbidden induced subgraphs.
2 {Claw, S3 , S3 , C4 }-free graphs
All graphs discussed in this paper are undirected and simple. A graph G is given by its vertex set V(G)
and edge set E(G), whose cardinalities will be denoted by n and m respectively. All input graphs in this
paper are assumed to be nontrivial (n > 1) and connected, hence n = O(m). For ` > 4, we use C` to
denote a hole on ` vertices; if we add a new vertex to a C` and make it adjacent to no or all vertices in the
hole, then we end with a C∗` or W` , respectively. For a hole H, we have |V(H)| = |E(H)|, denoted by |H|.
The complement graph G of a graph G is defined on the same vertex set V(G), where a pair of vertices u
and v is adjacent if and only if uv 6∈ E(G); e.g., W5 = C∗5 , and depicted in Fig. 1f is the complement of C6 .
An interval graph is the intersection graph of a set of intervals on the real line. A natural way to extend
interval graphs is to use arcs and a circle in the place of intervals and the real line, and the intersection
graph of arcs on a circle is a circular-arc graph. The set of intervals or arcs is called an interval model or
arc model respectively, and it can be specified by their 2n endpoints. In this paper, all intervals and arcs
are closed, and no distinct intervals or arcs are allowed to share an endpoint in the same model; these
restrictions do not sacrifice any generality. In a unit interval model or a unit arc model, every interval or
arc has length one. An interval or arc model is proper if no interval or arc in it properly contains another
interval or arc. A graph is a unit interval graph, proper interval graph, unit circular-arc graph, or proper
circular-arc graph if it has a unit interval model, proper interval model, unit arc model, or proper arc
model, respectively. The forbidden induced subgraphs of unit interval graphs have been long known.
Theorem 2.1 ([29]). A graph is a unit interval graph if and only if it contains no claw, S3 , S3 , or any hole.
Clearly, any (unit/proper) interval model can be viewed as a (unit/proper) arc model leaving some
point uncovered, and hence all (unit/proper) interval graphs are always (unit/proper) circular-arc graphs.
A unit interval/arc model is necessarily proper, but the other way does not hold true in general. A
well-known result states that a proper interval model can always be made unit, and thus these two graph
classes coincide [26, 29].1 This fact will be heavily used in the present paper; e.g., most of our proofs
consist in modifying a proper arc model into a proper interval model, which represents the desired unit
interval graph. On the other hand, it is easy to check that the S3 is a proper circular-arc graph but not
a unit circular-arc graph. Therefore, the class of unit circular-arc graphs is a proper subclass of proper
circular-arc graphs.
An arc model is Helly if every set of pairwise intersecting arcs has a common intersection. A circular-arc
graph is proper Helly if it has an arc model that is both proper and Helly.
Theorem 2.2 ([27, 22, 9]). A graph is a proper Helly circular-arc graph if and only if it contains no claw,
S3 , S3 , W4 , W5 , C6 , or C∗` for ` > 4.
The following is immediate from Theorems 2.1 and 2.2.
Corollary 2.3. If a proper Helly circular-arc graph is chordal, then it is a unit interval graph.
From Theorems 2.1 and 2.2 one can also derive the following combinatorial result. But since we will
prove a stronger result in Theorem 2.5 that implies it, we omit its proof here.
Proposition 2.4. Every connected {claw, S3 , S3 , C4 , C5 }-free graph is a proper Helly circular-arc graph.
Note that in Proposition 2.4, as well as most combinatorial statements to follow, we need the graph
to be connected. Circular-arc graphs are not closed under taking disjoint unions. If a (proper Helly)
circular-arc graph is not chordal, then it is necessarily connected. In other words, a disconnected (proper)
circular-arc graph must be a (unit) interval graph.
1 The reason we choose “unit” over “proper” in the title of this paper is twofold. On the one hand, the applications we are
interested in are more naturally represented by unit intervals. On the other hand, we want to avoid the use of “proper interval
subgraphs,” which is ambiguous.
4
Proposition 2.4 can be turned into an algorithmic statement. We say that a recognition algorithm (for
a graph class) is certifying if it provides a minimal forbidden induced subgraph of the input graph G when
G is determined to be not in this class. Linear-time certifying algorithms for recognizing proper Helly
circular-arc graphs have been reported by Lin et al. [22] and Cao et al. [9], from which one can derive a
linear-time algorithm for detecting an induced claw, S3 , S3 , C4 , or C5 from a graph that is not a proper
Helly circular-arc graph.
This would suffice for us to develop most of our main results. Even so, we would take pain to
prove slightly stronger results (than Proposition 2.4) on F-free graphs, where F denotes the set {claw,
S3 , S3 , C4 }. The purpose is threefold. First, they enable us to answer the question asked by van ’t Hof
and Villanger [17], i.e., the complexity of unit interval vertex deletion on F-free graphs, thereby more
accurately delimiting the complexity border of the problem. Second, as we will see, the disposal of
C5 ’s would otherwise dominate the second phase of our algorithm for unit interval edge deletion, so
excluding them enables us to obtain better exponential dependency on k in the running time. Third, the
combinatorial characterization might be of its own interest.
A (true) twin class of a graph G is an inclusion-wise maximal set of vertices in which all have the same
closed neighborhood. A graph is called a fat W5 if it has precisely six twin classes and it becomes a W5
after we remove all but one vertices from each twin class.2 By definition, vertices in each twin class induce
a clique. The five cliques corresponding to hole in the W5 is the fat hole, and the other clique is the hub,
of this fat W5 .
Theorem 2.5. Let G be a connected graph.
(1) If G is F-free, then it is either a fat W5 or a proper Helly circular-arc graph.
(2) In O(m) time we can either detect an induced subgraph of G in F, partition V(G) into six cliques
constituting a fat W5 , or build a proper and Helly arc model for G.
Proof. We prove only assertion (2) using the algorithm described in Fig. 3, and its correctness implies
assertion (1). The algorithm starts by calling the certifying algorithm of Cao et al. [9] for recognizing
proper Helly circular-arc graphs (step 0). It enters one of steps 1–4, or 6 based on the outcome of step 0.
Here the subscripts of vertices in a hole C` should be understood to be modulo `.
If the condition of any of steps 1–4 is satisfied, then either a proper and Helly arc model or a subgraph
in F is returned. The correctness of steps 1–3 is straightforward. Step 4.1 can find the path because G is
connected; possibly v = x, which is irrelevant in steps 4.2–4.4. Note also that ` > 4 in step 4.
By Theorem 2.2, the algorithm passes steps 1–4 only when the outcome of step 0 is a W5 ; let H be its
hole and let v be the other vertex. Steps 5–7 either detect an induced subgraph of G in F or partition
V(G) into six cliques constituting a fat W5 . Step 6 scans vertices not in the W5 one by one, and proceeds
based on the adjacency between x and H. In step 6.1, H and x make a C∗5 , which means that we can
proceed exactly the same as step 4. Note that the situation of step 6.4 is satisfied if x is adjacent to four
vertices of H. If none of the steps 6.1 to 6.5 applies, then x has precisely three neighbors in H and they
have to be consecutive. This is handled by step 6.6.
Steps 0 and 4 take O(m) time. Steps 1, 2, 3, 5, and 7 need only O(1) time. If the condition in step 6.1
is true, then it takes O(m) time but it always terminates the algorithm after applying it. Otherwise,
step 6.1 is never called, and the rest of step 6 scans the adjacency list of each vertex once, and hence
takes O(m) time in total. Therefore, the total running time of the algorithm is O(m). This concludes the
proof.
Implied by Theorem 2.5, a connected {claw, S3 , S3 , C4 , W5 }-free graph is a proper Helly circular-arc
graph, which in turns implies Proposition 2.4. At this point a natural question appealing to us is the
relation between connected {claw, S3 , S3 , C4 , C5 }-free graphs and unit Helly circular-arc graphs. Recall
that the class of unit interval graphs is a subclass of unit Helly circular-arc graphs, on which we have a
similar statement as Corollary 2.3, i.e., a unit Helly circular-arc graph that is chordal is a unit interval
graph. However, a connected {claw, S3 , S3 , C4 , C5 }-free graph that is not a unit Helly circular-arc graph
can be constructed as follows: Starting from a C` with ` > 6, for each edge hi hi+1 in the hole add a new
vertex vi and two new edges vi hi , vi hi+1 . (This is actually the CI(`, 1) graph defined by Tucker [27]; see
also [22].) Therefore, Proposition 2.4 and Theorem 2.5 are the best we can expect in this sense.
2 Interestingly, in the study of proper circular-arc graphs that are chordal, Bang-Jensen and Hell [1] showed that if a connected
{claw, S3 , C4 , C5 }-free graph contains an S3 , then it must be a fat S3 , which is defined analogously as a fat W5 .
5
Algorithm recognize-F-free(G)
Input: a connected graph G.
Output: a proper and Helly arc model, a subgraph in F, or six cliques making a fat W5 .
0. call the recognition algorithm for proper Helly circular-arc graphs [9];
1. if a proper and Helly arc model A is found then return it;
2. if a claw, S3 , or S3 is found then return it;
3. if a W4 , C∗4 , or C6 is found then return a C4 contained in it;
4. if a C∗` with hole H and isolated vertex v is found then
4.1.
use breadth-first search to find a shortest path v · · · xyhi from v to H;
4.2.
if y has a single neighbor hi in H then return claw {hi , y, hi−1 , hi+1 };
4.3.
if y has only two neighbors on H that are consecutive, say, {hi , hi+1 } then
return S3 {x, y, hi−1 , hi , hi+1 , hi+2 };
4.4.
return claw {y, x, hj , hj 0 }, where hj , hj 0 are two nonadjacent vertices in N[y] ∩ H;
\\ The outcome of step 0 must be a W5 ; let it be H and v. All subscripts of hi and Ki are modulo 5.
5. K0 ← {h0 };K1 ← {h1 };K2 ← {h2 }; K3 ← {h3 };K4 ← {h4 };Kv ← {v};
6. for each vertex x not in the W5 do
6.1.
if x is not adjacent to H then similar as step 4 (H and x make a C∗5 );
6.2.
if x has a single neighbor hi in H then return claw {hi , x, hi−1 , hi+1 };
6.3.
if x is only adjacent to hi , hi+1 in H then
if xv ∈ E(G) then return claw {v, hi−1 , hi+2 , x};
else return S3 {x, hi , hi+1 , hi−1 , v, hi+2 };
6.4.
if x is adjacent to hi−1 , hi+1 but not hi then return xhi−1 hi hi+1 as a C4 ;
6.5.
if x is adjacent to all vertices in H then
if xy 6∈ E(G) for some y ∈ Kv or Ki then return xhi−1 yhi+1 as a C4 ;
else add x to Kv ;
\\x is adjacent to all vertices in the six cliques.
6.6.
else
\\ Hereafter |N(x) ∩ H| = 3; let them be hi−1 , hi , hi+1 .
if xy 6∈ E(G) for some y ∈ Ki or Kv then return xhi−1 yhi+1 as a C4 ;
if xy 6∈ E(G) for some y ∈ Ki−1 then return claw {v, y, hi+2 , x};
if xy 6∈ E(G) for some y ∈ Ki+1 then return claw {v, hi−2 , y, x};
if xy ∈ E(G) for some y ∈ Ki−2 then return xhi+1 hi+2 y as a C4 ;
if xy ∈ E(G) for some y ∈ Ki+2 then return xhi−1 hi−2 y as a C4 ;
else add x to Ki ;
\\Kv , Ki−1 , Ki , Ki+1 ⊆ N(x) and Ki−2 , Ki+2 ∩ N(x) = ∅.
7. return the six cliques.
Figure 3: Recognizing F-free graphs.
Note that a C4 is a proper Helly circular-arc graph. Thus, the algorithm of Theorem 2.5 is not yet a
certifying algorithm for recognizing F-free graphs. To detect an induced C4 from a proper Helly circulararc graph, we need to exploit its arc model. If a proper Helly circular-arc graph G is not chordal, then the
set of arcs for vertices in a hole necessarily covers the circle, and it is minimal. Interestingly, the converse
holds true as well—note that this is not true for chordal graphs.
Proposition 2.6. [22, 9] Let G be a proper Helly circular-arc graph. If G is not chordal, then at least four
arcs are needed to cover the whole circle in any arc model for G.
Proposition 2.6 forbids among others two arcs from having two-part intersection.3
Corollary 2.7. Let G be a proper Helly circular-arc graph that is not chordal and let A be an arc model for G.
A set of arcs inclusion-wise minimally covers the circle in A if and only if the vertices represented by them
induce a hole of G.
Therefore, to find a shortest hole from a proper Helly circular-arc graph, we may work on an arc model
of it, and find a minimum set of arcs covering the circle in the model. This is another important step of
our algorithm for the unit interval editing problem. It has the detection of C4 ’s as a special case, because
a C4 , if existent, must be the shortest hole of the graph.
3A
model having no such intersection is called normal; see the appendix for more discussion.
6
Lemma 2.8. There is an O(m)-time algorithm for finding a shortest hole of a proper Helly circular-arc graph.
Before proving Lemma 2.8, we need to introduce some notation. In an interval model, the interval I(v)
for vertex v is given by [lp(v), rp(v)], where lp(v) and rp(v) are its left and right endpoints respectively. It
always holds lp(v) < rp(v). In an arc model, the arc A(v) for vertex v is given by [ccp(v), cp(v)], where
ccp(v) and cp(v) are its counterclockwise and clockwise endpoints respectively. All points in an arc model
are assumed to be nonnegative; in particular, they are between 0 (inclusive) and ` (exclusive), where ` is
the perimeter of the circle. We point out that possibly ccp(v) > cp(v); such an arc A(v) necessarily passes
through the point 0. Note that rotating all arcs in the model does not change the intersections among
them. Thus we can always assume that a particular arc contains or avoids the point 0. We say that an
arc model (for an n-vertex circular-arc graph) is canonical if the perimeter of the circle is 2n, and every
endpoint is a different integer in {0, 1, . . . , 2n − 1}. Given an arc model, we can make it canonical in linear
time: We sort all 2n endpoints by radix sort, and replace them by their indices in the order.
Each point α in an interval model I or arc model A defines a clique, denoted by KI (α) or KA (α)
respectively, which is the set of vertices whose intervals or arcs contain α. There are at most 2n distinct
cliques defined as such. If the model is Helly, then they include all maximal cliques of this graph [15].
Since the set of endpoints is finite, for any point ρ in an interval or arc model, we can find a small positive
value such that there is no endpoint in [ρ − , ρ) ∪ (ρ, ρ + ],—in other words, there is an endpoint in
[ρ − , ρ + ] if and only if ρ itself is an endpoint. Note that the value of should be understood as a
function, depending on the interval/arc model as well as the point ρ, instead of a constant.
Let G be a non-chordal graph and let A be a proper and Helly arc model for G. If uv ∈ E(G),
then exactly one of ccp(v) and cp(v) is contained in A(u) (Proposition 2.6). Thus, we can define a
“left-right relation” for each pair of intersecting arcs, which can be understood from the viewpoint of
an observer placed at the center of the model. We say that arc A(v) intersects arc A(u) from the left
when cp(v) ∈ A(u), denoted by v → u. Any set of arcs whose union is an arc not covering the circle
(the corresponding vertices induce a connected unit interval graph) can be ordered in a unique way such
that vi → vi+1 for all i. From it we can find the leftmost (most counterclockwise) and rightmost (most
clockwise) arcs.
For any vertex v, let h(v) denote the length of the shortest holes through v, which is defined to be +∞
if no hole of G contains v. The following is important for the proof of Lemma 2.8.
Lemma 2.9. Let A be a proper and Helly arc model for a non-chordal graph G. Let v1 , v2 , . . . , vp be a
sequence of vertices such that for each i = 2, . . . , p, the arc A(vi ) is the rightmost of all arcs containing
cp(vi−1 ). If v1 is contained in some hole and vi v1 6∈ E(G) for all 2 < i 6 p, then there is a hole of length
h(v1 ) containing v1 , v2 , . . . , vp as consecutive vertices on it.
Proof. Suppose that there is no such a hole, then there exists a smallest number i with 2 6 i 6 p such that
no hole of length h(v1 ) contains v1 , v2 , . . . , vi . By assumption, there is a hole v1 · · · vi−1 ui ui+1 · · · uh(v1 )
of length h(v1 ) with ui 6= vi ; let it be H. In case that i = 2, we assume that H is given in the way
that v1 → u2 . By Corollary 2.7, the set of arcs for H cover the circle in A. Since vi−1 → ui and by the
assumption on vi , the arc A(vi ) covers [cp(vi−1 ), ccp(ui+1 )] (note that i < h(v1 ) as otherwise vi → v1 ).
Therefore, the arcs for V(H) \ {ui } ∪ {vi } cover the circle as well. By Corollary 2.7, a subset of these
vertices induces a hole; since vi is not adjacent to v1 from the left, this subset has to contain v1 . But a
hole containing v1 cannot be shorter than H, and hence it contains v1 , . . . , vi , a contradiction. Therefore,
such an i does not exist, and there is a hole of length h(v1 ) containing v1 , v2 , . . . , vp .
Since vj and vj+1 are adjacent for all 1 6 j 6 p − 1, vertices v1 , v2 , . . . , vp have to be consecutive on
this hole.
Let α be any fixed point in a proper and Helly arc model. According to Corollary 2.7, every hole needs
to visit some vertex in KA (α). Therefore, to find a shortest hole in G, it suffices to find a hole of length
min{h(x) : x ∈ KA (α)}.
Proof of Lemma 2.8. The algorithm described in Fig. 4 finds a hole of length min{h(x) : x ∈ KA (cp(v) +
0.5)}. Step 1 creates |KA (cp(v) + 0.5)| arrays, each starting with a distinct vertex in KA (cp(v) + 0.5),
and they are ordered such that their (counter)clockwise endpoints are increasing. The main job of this
algorithm is done in step 3. During this step, w is the new vertex to be processed, and U is the current
array. Each new vertex is added to at most one array, while each array is either dropped or extended.
We use ⊥ as a dummy vertex, which means that no new vertex has been met after the last one has been
put into the previous array. Step 3.1 records the last scanned arc. Once the clockwise endpoint of the
last vertex z of the current array U is met, w is appended to U (step 3.4); note that A(w) is the most
7
Algorithm shortest-hole(G, A)
Input: a proper and Helly arc model A for a non-chordal graph G.
Output: a shortest hole of G.
0. make A canonical where 0 is ccp(v) for some v;
1. for i = 1, . . . , cp(v) − 1 do
1.1.
if i is ccp(x) then create a new array {x};
\\ these arrays are circularly linked so that the next of the last array is the first one.
2. w ← ⊥; U ← the first array;
3. for i = cp(v) + 1, . . . , 2n − 1 do
3.0.
z ← the last vertex of U;
3.1.
if i is ccp(x) then w ← x; continue† ;
3.2.
if i 6= cp(z) then continue;
3.3.
if w = ⊥ then delete U; U ← the next array of U;
3.4.
if w 6= ⊥ then append w to U; w ← ⊥; U ← the next array of U;
4. for each U till the last array do
4.1.
if the first and last vertices of U are adjacent then return U;
5. return U ∪ {v}.
\\U is the last array.
†: i.e., ignore the rest of this iteration of the for-loop and proceed to the next iteration.
Figure 4: Finding a shortest hole in a proper Helly circular-arc graph.
clockwise arc that contains cp(z). On the other hand, if w = ⊥, then we drop this array from further
consideration (step 3.3). If after step 3, one of the arrays already induces a hole (i.e., the first and last
vertices are adjacent), then it is returned in step 4.1. Otherwise, U does not induces a hole, and step 5
returns the hole induced by U and v.
We now verify the correctness of the algorithm. It suffices to show that the length of the found hole is
min{h(x) : x ∈ KA (cp(v) + 0.5)}. The following hold for each array U:
(1) For any pair of consecutive vertices u, w of U, the arc A(w) is the rightmost of all arcs containing
cp(u).
(2) At the end of the algorithm, if U is not dropped, then 0 < cp(z) < cp(v), where z is the last vertex of
U.
Let v1 , v2 , . . . , vp be the vertices in U. From (1) and (2) it can be inferred that for 1 < i < p, the vertex
vi is adjacent to only vi−1 , vi+1 in U, while v1 and vp may or may not be adjacent. If v1 vp ∈ E(G), then
vertices in U induce a hole of G; otherwise U ∪ {v} induces a hole of G. By Lemma 2.9, this hole has
length h(v1 ).
Some of the |KA (cp(v) + 0.5)| arrays created in step 1 may be dropped in step 3. Note that step 3
processes arrays in a circular order starting from the first one, and each array is either deleted (step 3.3)
or extended by adding one vertex (step 3.4).
(3) At any moment of the algorithm, the sizes of any two arrays differ by at most one. In particular, at
the end of step 3, if the current array U is not the first, then U, as well as all the succeeding arrays,
has one less element than its predecessor(s).
This ensures that the hole returned in step 4.1 or 5 is the shortest among all the holes decided by the
remaining arrays after step 3.
It remains to argue that for any array U deleted in step 3.3, the found hole is not longer than h(x),
where x is the first vertex of U. All status in the following is referred at the moment before U is deleted
(i.e., before step 3.3 that deletes U). Let z be the last vertex of U. Let U 0 be the array that is immediately
preceding U, and let z 0 be its last vertex. Note that there is no arc with a counterclockwise endpoint
between cp(z 0 ) and cp(z), as otherwise w 6= ⊥ and U would not be deleted. Therefore, any arc intersecting
A(z) from the right also intersects A(z 0 ) from the right. By Lemma 2.9, there is a hole H that has length
h(x) and contains U ∪ {z 0 }. We find a hole through U 0 of the same length as follows. If U is not the first
array, then |U 0 | = |U| + 1, and we replace U ∪ {z 0 } by U 0 . Otherwise, |U 0 | = |U|, and we replace U ∪ {z 0 } by
{v} ∪ U 0 . It is easy to verify that after this replacement, H remains a hole of the same length.
8
We now analyze the running time of the algorithm. Each of the 2n endpoints is scanned once, and
each vertex belongs to at most one array. Using a linked list to store an array, the addition of a new vertex
can be implemented in constant time. Using a circularly linked list to organize the arrays, we can find the
next array or delete the current one in constant time. With the (endpoints of) all arcs given, the adjacency
between any pair of vertices can be checked in constant time, and thus step 4 takes O(n) time. It follows
that the algorithm can be implemented in O(m) time.
3
Vertex deletion
We say that a set V− of vertices is a hole cover of G if G − V− is chordal. The hole covers of proper Helly
circular-arc graphs are characterized by the following lemma.
Lemma 3.1. Let A be a proper and Helly arc model for a non-chordal graph G. A set V− ⊆ V(G) is a hole
cover of G if and only if it contains KA (α) for some point α in A.
Proof. For any vertex set V− , the subgraph G − V− is also a proper Helly circular-arc graph, and the set of
arcs {A(v) : v ∈ V(G) \ V− } is a proper and Helly arc model for G − V− . For the “if” direction, we may
rotate A to make α = 0, and then setting I(v) = A(v) for each v ∈ V(G) \ V− gives a proper interval
model for G − V− . For the “only if” direction, note thatS
if there is no α with KA (α) ⊆ V− , then we can find
a minimal set X of vertices from V(G) \ V− such that v∈X A(v) covers the whole circle in A. According
to Corollary 2.7, X induces a hole of G, which remains in G − V− .
Noting that any “local” part of a proper and Helly arc model “behaves similarly” as an interval model,
Lemma 3.1 is an easy extension of the clique separator property of interval graphs [14]. On the other
hand, to get a unit interval graph out of a fat W5 , it suffices to delete a smallest clique from the fat hole.
Therefore, Theorem 2.5 and Lemma 3.1 imply the following linear-time algorithm.
Corollary 3.2. The unit interval vertex deletion problem can be solved in O(m) time (1) on proper Helly
circular-arc graphs and (2) on F-free graphs.
We are now ready to prove the main results of this section.
Theorem 3.3. There are an O(6k · m)-time parameterized algorithm for the unit interval vertex deletion
problem and an O(nm)-time approximation algorithm of approximation ratio 6 for its minimization version.
Proof. Let (G, k) be an instance of unit interval vertex deletion; we may assume that G is not a unit
interval graph and k > 0. The parameterized algorithm calls first Theorem 2.5(2) to decide whether it
has an induced subgraph in F, and then based on the outcome, it solves the problem by making recursive
calls to itself, or calling the algorithm of Corollary 3.2. If an induced subgraph F in F is found, it calls
itself |V(F)| times, each with a new instance (G − v, k − 1) for some v ∈ V(F); since we need to delete at
least one vertex from V(F), the original instance (G, k) is a yes-instance if and only if at least one of the
instances (G − v, k − 1) is a yes-instance. Otherwise, G is F-free and the algorithm calls Corollary 3.2
to solve it. The correctness of the algorithm follows from discussion above and Corollary 3.2. On each
subgraph in F, which has at most 6 vertices, at most 6 recursive calls are made, all with parameter value
k − 1. By Theorem 2.5, each recursive call is made in O(m) time; each call of Corollary 3.2 takes O(m)
time. Therefore, the total running time is O(6k · m).
The approximation algorithm is adapted from the parameterized algorithm as follows. For the subgraph
F found by Theorem 2.5, we delete all its vertices. We continue the process until the remaining graph is
F-free, and then we call Corollary 3.2 to solve it optimally. Each subgraph in F has 4 or 6 vertices, and
thus at most n/4 such subgraphs can be detected and deleted, each taking O(m) time, hence O(nm) in
total. Corollary 3.2 takes another O(m) time. The total running time is thus O(nm) + O(m) = O(nm),
and the approximation ratio is clearly 6.
4
Edge deletion
Inspired by Lemma 3.1, one may expect a similarly nice characterization—being “local” to some point in
an arc model for G—for a minimal set of edges whose deletion from a proper Helly circular-arc graph G
makes it chordal. This is nevertheless not the case; as shown in Fig. 5, they may behave in a very strange
or pathological way.
9
v3
v2
u3
v4
Figure 5: The set of all 30 edges (both
u2
u4
u1
u5
v1
u6
v5
v6
solid and dashed) spans a proper Helly
circular-arc graph. After the set of 19
dashed edges deleted (we rely on the
reader to verify its minimality), the remaining graph on the 11 solid edges is
a unit interval graph.
Note that four edges would suffice, e.g.,
{u2 u3 , u2 v3 , v2 u3 , v2 v3 }.
Recall that v → u means arc A(v) intersecting arc A(u) from the left, or cp(v) ∈ A(u). For each
point α in a proper and Helly arc model A, we can define the following set of edges:
→
−
E A (α) = {vu : v ∈ KA (α), u 6∈ KA (α), v → u}.
(1)
→
−
One may symmetrically view E A (α) as {vu : v 6∈ KA (β), u ∈ KA (β), v → u}, where β := max{cp(x) : x ∈
→
−
KA (α)} + . It is easy to verify that the following gives a proper interval model for G − E A (0):
[ccp(v), cp(v) + `] if v ∈ KA (0),
I(v) :=
(2)
[ccp(v), cp(v)]
otherwise,
→
−
where ` is the perimeter of the circle in A; see Fig. 6. For an arbitrary point α, the model G − E A (α) can
be given analogously, e.g., we may rotate the model first to make α = 0.
α(0)
β
0
(a) An arc model for G.
`
−
→
(b) The interval model for G − E A (α) given by (2).
Figure 6: Illustration for Proposition 4.1.
Proposition 4.1. Let A be a proper and Helly arc model for a non-chordal graph G. For any point α in A,
→
−
the subgraph G − E A (α) is a unit interval graph.
The other direction is more involved and very challenging. A unit interval graph G is called a
spanning unit interval subgraph of G if V(G) = V(G) and E(G) ⊆ E(G); it is called maximum if it has the
largest number of edges among all spanning unit interval subgraphs of G. To prove that all maximum
spanning unit interval subgraphs have a certain property, we use the following argument by contradiction.
Given a spanning unit interval subgraph G not having the property, we locally modify a unit interval
model I for G to a proper interval model I 0 such that the represented graph G 0 satisfies E(G 0 ) ⊆ E(G)
and |E(G 0 )| > |E(G)|. Recall that we always select in a way that there cannot be any endpoint in
[ρ − , ρ) ∪ (ρ, ρ + ], and thus an arc covering ρ + or ρ − must contain ρ.
Lemma 4.2. Let A be a proper and Helly arc model for a non-chordal graph G. For any maximum spanning
→
−
unit interval subgraph G of G, the deleted edges, E(G) \ E(G), are E A (ρ) for some point ρ in A.
Proof. Let I be a unit interval model for G, and let E− = E(G) \ E(G), i.e., the set of deleted edges from
G. We find first a vertex v satisfying at least one of the following conditions.
(C1) The sets NG [v] and KA (ccp(v) − ) are disjoint and all edges between them are in E− .
(C2) The sets NG [v] and KA (cp(v) + ) are disjoint and all edges between them are in E− .
Recall that a vertex u ∈ KA (ccp(v) − ) if and only if u → v, and a vertex w ∈ KA (cp(v) + ) if and
only if v → w; see Fig. 7a. These two conditions imply NG [v] ⊆ KA (cp(v)) and NG [v] ⊆ KA (ccp(v))
10
w
u
v
cp(v) +
ccp(v)
(a)
v0
v0
v
w
u
u0
u0
(b)
v
w
(c)
w0
w
u
lp(u)
y0
u
y
v
v
u
u0
w
w0
lp(w0 ) rp(u0 )
rp(u)
(d)
(e)
Figure 7: Illustration for the proof of Lemma 4.2.
respectively: All edges between v itself, which belongs to NG [v], and KA (ccp(v) − ) or KA (cp(v) + )
are in E− .
Let I(v0 ) be the leftmost interval in I. Note that by Proposition 2.6, arcs for NG [v0 ] cannot cover
the whole circle. If NG [v0 ] is a separate component in G, then we can take v to be the vertex from
NG [v0 ] with the leftmost arc; it satisfies condition (C1). Otherwise, let I(u) be the last interval containing
rp(v0 ) and let I(u 0 ) be the next interval (i.e., I(u 0 ) is the leftmost interval that does not intersect I(v0 )).
See Fig. 7b. Intervals I(u) and I(u 0 ) intersect (NG [v0 ] is not isolated), but v0 u 0 6∈ E(G) because G is
maximum: Moving I(v0 ) to the right to intersect I(u 0 ) would otherwise make a unit interval model that
represents a subgraph of G having one more edge than G.
• If u 0 → u, then u 0 → u → v0 . We argue by contradiction that there cannot be vertices v 0 ∈
KA (cp(v0 ) + ) and w 6∈ NG (v0 ) with uv 0 , v 0 w ∈ E(G). If uv 0 ∈ E(G) ⊆ E(G), then by the position
of A(v0 ), we must have u → v 0 and v0 → v 0 . Then v0 u 0 6∈ E(G) excludes the possibility u 0 → v 0 ; on
the other hand, v 0 → u 0 is excluded by Proposition 2.6: The arcs for u 0 , u, v 0 cannot cover the whole
circle. Therefore, v 0 u 0 6∈ E(G), and likewise uw 6∈ E(G). They cannot be in E(G) either, but this is
impossible; See Fig. 7b. Let v be the vertex in NG [v0 ] such that A(v) is the rightmost; it satisfies
condition (C2).
• If u → u 0 , then v0 → u → u 0 . Similarly as above, the vertex v in NG [v0 ] such that A(v) is the
leftmost satisfies condition (C1).
Noting that conditions (C1) and (C2) are symmetric, we assume that the vertex v found above satisfies
condition (C2), and a symmetric argument would apply to condition (C1). Note that by the selection of v,
which has the leftmost arc among NG [v0 ], we have NG [v] ⊆ NG [v 0 ] for every v 0 satisfying lp(v 0 ) < lp(v);
thus setting their intervals to be [lp(v) + , rp(v) + ] would produce another unit interval model for G.
In the rest of the proof we may assume without loss of generality that I(v) is the first interval in I.
Since the model A is proper and Helly, no arc in A can contain both ccp(v) and cp(v) + . In other
→
−
words, KA (ccp(v)) and KA (cp(v) + ) are disjoint, and E A (ccp(v)) comprises precisely edges between
→
−
them; see Fig. 7a. By Proposition 4.1, |E− | 6 | E A (ccp(v))|. If KA (ccp(v)) is not adjacent to KA (cp(v) + )
→
−
in G, then E A (ccp(v)) ⊆ E− , and they have to be equal. In this case the proof is complete: ρ = ccp(v).
We are hence focused on edges between KA (ccp(v)) and KA (cp(v) + ).
Claim 1. Let u ∈ KA (ccp(v)). If u is adjacent to KA (cp(v) + ) in G, then I(u) does not intersect the
intervals for NG [v].
11
Proof. Recall that uv ∈
/ E(G) by condition (C2) and the fact that u is adjacent to KA (cp(v) + ) in G. Let
u 0 be the vertex in NG [v] with the rightmost interval, and let I(y) be the leftmost interval not intersecting
I(v). Note that vy 6∈ E(G): Otherwise moving I(v) to the right to intersect I(y) would make a unit interval
model that represents a subgraph of G with one more edge than G. Suppose to the contrary of this claim
that I(u) intersects some interval for NG [v], then it intersects I(u 0 ). See Fig. 7c. Since v satisfies condition
(C2) and by Proposition 2.6, u 0 → v and y → u, u 0 .
Let w be the vertex in KA (cp(v) + ) that has the leftmost interval. Since v satisfies condition (C2),
lp(w) > rp(u 0 ); by the assumption that u is adjacent to KA (cp(v) + ) in G, we have lp(w) < rp(u).
By Proposition 2.6, yw 6∈ E(G). Let A(y 0 ) be the leftmost arc such that y 0 ∈ KA (ccp(u)) and I(y 0 )
lies in (rp(v), lp(w)); this vertex exists because y itself is a candidate for it. Again by Proposition 2.6,
y 0 w 6∈ E(G).
Let X denote the set of vertices x with lp(x) < lp(w). We make a new interval model by resetting
intervals for these vertices. Since every vertex in X is adjacent to at least one of u and u 0 , by Proposition 2.6,
the union of arcs for X do not cover the circle in A. These arcs can thus be viewed as a proper interval
model for G[X]. The new intervals for X, adapted from these arcs, are formally specified as follows. The
left endpoint of each x ∈ X is set as lp 0 (x) = lp(w) − (ccp(w) − ccp(x)). For each vertex x ∈ X \ NG (w),
we set
I 0 (x) = [lp(w) − (ccp(w) − ccp(x)), lp(w) − (ccp(w) − cp(x))].
By the selection of w, we have x → w for all x ∈ X ∩ NG (w). Arcs for X ∩ NG (w) are thus pairwise
intersecting; by Proposition 2.6, they cannot cover the whole circle. We can thus number vertices in
X ∩ NG (w) as u1 , . . . , up such that ui → ui+1 for each i = 1, . . . , p − 1. Their right endpoints are set as
rp 0 (u1 ) = max lp(w) + , rp(u1 ) ,
and
0
0
rp (ui ) = max rp (ui−1 ) + , rp(ui )
for i = 2, . . . , p.
Let I 0 denote the resulting new interval model. To see that I 0 is proper, note that (a) no new interval
can contain or be contained by an interval I(z) for z ∈ V(G) \ X; and (b) the left and right endpoints
of the intervals for X have the same ordering as the counterclockwise and clockwise endpoints of the
arcs {A(x) : x ∈ X}, hence necessarily proper. Let G 0 denote the proper interval graph represented by I 0 .
We want to argue that E(G) ⊂ E(G 0 ) ⊆ E(G), which would contradict that G is a maximum unit interval
subgraph of G, and conclude the proof of this claim.
By construction, G 0 − X is the same as G − X, while G 0 [X] is the same as G[X]. Thus, we focus on
edges between X and V(G) \ X, which are all incident to X ∩ NG (w). For each i = 1, . . . , p, we have
rp 0 (ui ) > rp(ui ), and thus E(G) ⊆ E(G 0 ); on the other hand, they are not equal because uv ∈ E(G 0 )\E(G).
We show by induction that for every i = 1, . . . , p, the edges incident to ui in G 0 is a subset of G. The
base case is clear: NG 0 (u1 ) \ X is either {w} or NG (u1 ) \ X. For the inductive step, if rp 0 (ui ) = rp(ui ),
then NG 0 (ui ) \ X = NG (ui ) \ X ⊆ NG (ui ); otherwise, NG 0 (ui ) \ X ⊆ NG (ui−1 ) \ X, which is a subset of
NG (ui ) because ui−1 → ui . This verifies E(G 0 ) ⊆ E(G).
y
We consider then edges deleted from each vertex in KA (ccp(v)), i.e., edges in E− that is incident to
KA (ccp(v)).
Claim 2. Let u ∈ KA (ccp(v)). If u is adjacent to KA (cp(v) + ) in G, then there are strictly more edges
→
−
incident to u in E− than in E A (ccp(v)) (i.e., between u and KA (cp(v) + ).).
Proof. The vertices in NG (u) consists of three parts, NG [v], those in KA (cp(v) + ), and others. By Claim 1,
→
−
edges between u and all vertices in NG [v] are in E− ; they are not in E A (ccp(v)). All edges between u
→
−
→
−
and KA (cp(v) + ) are in E A (ccp(v)). In E A (ccp(v)) there is no edge between u and the other vertices.
Therefore, to show the claim, it suffices to show |NG [u] ∩ KA (cp(v) + )| 6 |NG (v)| < |NG [v]|.
Let uw be an edge of G with w ∈ KA (cp(v) + ). Since v satisfies condition (C2), both uv and vw are
in E− . We show only the case that I(w) intersects I(u) from the right; a similar argument works for the
other case: Note that all the intervals used below are disjoint from I(v).
Consider first that there exists another vertex w 0 ∈ KA (cp(v) + ) with interval I(w 0 ) intersecting
I(u) from the left. See Fig. 7d. Any interval that intersects I(u) necessarily intersects at least one of
I(w) and I(w 0 ), and thus by Proposition 2.6, its vertex must be in NG (v). As a result, setting I(v) to
[lp(u) − , rp(u) − ] gives another proper interval model and it represents a subgraph of G. Since G is
maximum, we can conclude |NG [u]| 6 |NG (v)| < |NG [v]|. In this case the proof of this claim is concluded.
In the second case there is no vertex in KA (cp(v) + ) whose interval intersects I(u) from the left.
Note that every interval intersecting [lp(w), rp(u)] represents a vertex in NG [v]. Let w 0 be vertex in
12
KA (cp(v)) ∩ KI (rp(u)) that has the leftmost interval, and let u 0 be vertex in KA (ccp(v)) ∩ KI (lp(w)) that
has the rightmost interval; these two vertices exist because u and w are candidates for them, respectively.
See Fig. 7e. There cannot be any vertex whose interval contains [lp(w 0 ), rp(u 0 )]: Such a vertex, if it
exists, is in NG (v), but then it contradicts the selection of u 0 and w 0 . Also by the selection of u 0 and
w 0 , no interval contains [lp(w 0 ) − , rp(u 0 ) + ]. Thus, setting I(v) to [lp(w 0 ) − , rp(u 0 ) + ] gives
another proper interval model and it represents a subgraph of G. Since G is maximum, we can conclude
|NG [u] ∩ KA (cp(v) + )| 6 |NG (v)| < |NG [v]|.
y
Therefore, for every vertex u ∈ KA (ccp(v)), there are no less edges incident to u in E− than in
→
−
E A (ccp(v)). Moreover, as there is at least one vertex in KA (ccp(v)) adjacent to KA (cp(v)+) in G, (noting
→
−
→
−
that no edge in E A (ccp(v)) is incident to two vertices in KA (ccp(v)),) it follows that |E− | > | E A (ccp(v))|,
contradicting that G is a maximum unit interval subgraph of G.
It is worth stressing that a thinnest place in an arc model with respect to edges is not necessarily a
thinnest place with respect to vertices; see Fig. 8 for an example. There is a linear number of different
places to check, and thus the edge deletion problem can also be solved in linear time on proper Helly
circular-arc graphs. The problem is also simple on fat W5 ’s.
α
β
Figure 8: The thinnest points for vertices
and edges are α and β respectively:
KA (α) = 2 < KA (β) = 3; while
→
−
→
−
E A (α) = 8 > E A (β) = 6.
Theorem 4.3. The unit interval edge deletion problem can be solved in O(m) time (1) on proper Helly
circular-arc graphs and (2) on F-free graphs.
Proof. For (1), we may assume that the input graph G is not an unit interval graph; according to
Corollary 2.3, it is not chordal. We build a proper and Helly arc model A for G; without loss of generality,
assume that it is canonical. According to Lemma 4.2, the problem reduces to finding a point α in A such
→
−
that E A (α) is minimized. It suffices to consider the 2n points i + 0.5 for i ∈ {0, . . . , 2n − 1}. We calculate
→
−
→
−
→
−
first E A (0.5), and then for i = 1, . . . , 2n − 1, we deduce E A (i + 0.5) from E A (i − 0.5) as follows. If i is a
→
−
→
−
clockwise endpoint of some arc, then E A (i + 0.5) = E A (i − 0.5). Otherwise, i = ccp(v) for some vertex
→
−
→
−
v, then the difference between E A (i + 0.5) and E A (i − 0.5) is the set of edges incident to v. In particular,
→
−
→
−
→
−
→
−
{uv : u → v} = E A (i − 0.5) \ E A (i + 0.5), while {uv : v → u} = E A (i + 0.5) \ E A (i − 0.5). Note that
→
−
the initial value E A (0.5) can be calculated in O(m) time, and then each vertex and its adjacency list is
scanned exactly once. It follows that the total running time is O(m).
For (2), we may assume that the input graph G is connected, as otherwise we work on its components
one by one. According to Theorem 2.5(1), G is either a proper Helly circular-arc graph or a fat W5 . The
former case has been considered above, and now assume G is a fat W5 . Let K0 , . . . , K4 be the five cliques
in the fat hole, and let K5 be the hub. We may look for a maximum spanning unit interval subgraph G
of G such that NG [u] = NG [v] for any pair of vertices u, v in Ki , where i ∈ {0, . . . , 5}. We now argue the
existence of such a subgraph. By definition, NG [u] = NG [v]. Let G 0 be a maximum spanning unit interval
subgraph of G where NG 0 [u] 6= NG 0 [v], and assume without loss of generality, |NG 0 [u]| > |NG 0 [v]|. We
may change the deleted edges that are incident to v to make another subgraph where v has the same
neighbors as NG 0 [u]; this graph is clearly a unit interval graph and has no less edges than G 0 . This
operation can be applied to any pair of u, v in the same twin class, and it will not violate an earlier pair.
Repeating it we will finally end with a desired maximum spanning unit interval subgraph. Therefore,
there is always some i ∈ {0, . . . , 4} (all subscripts are modulo 5) such that deleting all edges between
cliques Ki and Ki+1 together with edges between one of them and K5 leaves a maximum spanning unit
interval subgraph. Once the sizes of all six cliques have been calculated, which can be done in O(m) time,
the minimum set of edges can be decided in constant time. Therefore, the total running time is O(m).
The proof is now complete.
13
Indeed, it is not hard to see that in the proof of Theorems 4.3(2), every maximum spanning unit
interval subgraph of a fat W5 keeps its six twin classes, but we are satisfied with the weaker statement
that is sufficient for our algorithm.
Theorems 4.3 and 2.5 already imply a branching algorithm for the unit interval edge deletion problem
running in time O(9k · m). Here the constant 9 is decided by the S3 , which has 9 edges. However, a closer
look at it tells us that deleting any single edge from an S3 introduces either a claw or a C4 , which forces
us to delete some other edge(s). The disposal of an S3 is similar. The labels for an S3 and a S3 used in the
following proof are as given in Fig. 1.
Proposition 4.4. Let G be a spanning unit interval subgraph of a graph G, and let F = E(G) \ E(G).
(1) For an S3 of G, there must be some i = 1, 2, 3 such that F contains at least two edges from the triangle
involving ui .
(2) For an S3 of G, the set F contains either an edge ui vi for some i = 1, 2, 3, or at least two edges from the
triangle u1 u2 u3 .
Proof. (1) We consider the intervals for v1 , v2 , and v3 in a unit interval model I for G. If v1 v2 v3 remains a
triangle of G, then the interval I(v1 ) ∪ I(v2 ) ∪ I(v3 ) has length less than 2, and it has to be disjoint from
I(ui ) for at least one i ∈ {1, 2, 3}. In other words, both edges incident to this ui are in F. Otherwise, v1 v2 v3
is not a triangle of G. Assume without loss of generality lp(v1 ) < lp(v2 ) < lp(v3 ). Then v1 v3 6∈ E(G) and
u2 cannot be adjacent to both v1 and v3 . Therefore, at least two edges from the triangle u2 v1 v3 are in F;
other cases are symmetric.
(2) If F contains none of the three edges v1 u1 , v2 u2 , and v3 u3 , then it contains at least two edges from
the triangle u1 u2 u3 : Otherwise there is a claw.
This observation and a refined analysis will yield the running time claimed in Theorem 1.1. The
algorithm goes similarly as the parameterized algorithm for unit interval vertex deletion used in the proof
of Theorem 3.3.
Theorem 4.5. The unit interval edge deletion problem can be solved in time O(4k · m).
Proof. The algorithm calls first Theorem 2.5(2) to decide whether there exists an induced subgraph in F,
and then based on the outcome, it solves the problem by making recursive calls to itself, or calling the
algorithm of Theorem 4.3. When a claw or C4 is found, the algorithm makes respectively 3 or 4 calls to
itself, each with a new instance with parameter value k − 1 (deleting one edge from the claw or C4 ). For
an S3 , the algorithm branches on deleting two edges from a triangle involving a vertex ui with i = 1, 2, 3.
Since there are three such triangles, and each has three options, the algorithm makes 9 calls to itself, all
with parameter value k − 2. For an S3 , the algorithm makes 6 calls to itself, of which 3 with parameter
value k − 1 (deleting edge vi ui for i = 1, 2, 3), and another 3 with parameter value k − 2 (deleting two
edges from the triangle v1 v2 v3 ).
To verify the correctness of the algorithm, it suffices to show that for any spanning unit interval
subgraph G of G, there is at least one recursive call that generates a graph G 0 satisfying E(G) ⊆ E(G 0 ) ⊆
E(G). This is obvious when the recursive calls are made on a claws or C4 . It follows from Proposition 4.4
when the recursive calls are made on an S3 or S3 . With standard technique, it is easy to verify that O(4k )
recursive calls are made, each in O(m) time. Moreover, the algorithm for Theorem 4.3 is called O(4k )
times. It follows that the total running time of the algorithm is O(4k · m).
What dominates the branching step is the disposal of C4 ’s. With the technique the author developed
in [23], one may (slightly) improve the running time to O(ck · m) for some constant c < 4. To avoid
blurring the focus of the present paper, we omit the details.
5
General editing
Let V− ⊆ V(G), and let E− and E+ be a set of edges and a set of non-edges of G − V− respectively. We say
that (V− , E− , E+ ) is an editing set of G if the deletion of E− from and the addition of E+ to G − V− create
a unit interval graph. Its size is defined to be the 3-tuple (|V− |, |E− |, |E+ |), and we say that it is smaller
than (k1 , k2 , k3 ) if all of |V− | 6 k1 and |E− | 6 k2 and |E+ | 6 k3 hold true and at least one inequality is
strict. The unit interval editing problem is formally defined as follows.
14
Input:
Task:
A graph G and three nonnegative integers k1 , k2 , and k3 .
Either construct an editing set (V− , E− , E+ ) of G that has size at most (k1 , k2 , k3 ),
or report that no such set exists.
We remark that it is necessary to impose the quotas for different modifications in the stated, though
cumbersome, way. Since vertex deletions are clearly preferable to both edge operations, the problem
would be computationally equivalent to unit interval vertex deletion if we have a single budget on the
total number of operations.
By and large, our algorithm for the unit interval editing problem also uses the same two-phase
approach as the previous algorithms. The main discrepancy lies in the first phase, when we are not
satisfied with a proper Helly circular-arc graph or an F-free graph. In particular, we also want to dispose
of all holes C` with ` 6 k3 + 3, which are precisely those holes fixable by merely adding edges (recall
that at least ` − 3 edges are needed to fill a C` in). In the very special cases where k3 = 0 or 1, a fat W5
is F ∪ {C` : ` 6 k3 + 3}-free. It is not hard to solve fat W5 ’s, but to make the rest more focused and also
simplify the presentation, we also exclude these cases by disposing of all C5 ’s in the first phase.
A graph is called reduced if it contains no claw, S3 , S3 , C4 , C5 , or C` with ` 6 k3 + 3. By Proposition 2.4,
a reduced graph G is a proper Helly circular-arc graph. Hence, if G happens to be chordal, then it must be
a unit interval graph (Corollary 2.3), and we terminate the algorithm. Otherwise, our algorithm enters
the second phase. Now that G is reduced, every minimal forbidden induced subgraph is a hole C` with
` > k3 + 3, which can only be fixed by deleting vertices and/or edges. Here we again exploit a proper
and Helly arc model A for G. According to Lemma 3.1, if there exists some point ρ in the model such
that |KA (ρ)| 6 k1 , then it suffices to delete all vertices in KA (ρ), which results in a subgraph that is a unit
interval graph. Therefore, we may assume hereafter that no such point exists, then G remains reduced
and non-chordal after at most k1 vertex deletions. As a result, we have to delete edges as well.
Consider an (inclusion-wise minimal) editing set (V− , E− , E+ ) to a reduced graph G. It is easy to verify
that (∅, E− , E+ ) is an (inclusion-wise minimal) editing set of the reduced graph G − V− . In particular, E−
needs to intersect all holes of G − V− . We use A − V− as a shorthand for {A(v) ∈ A : v 6∈ V− }, an arc
model for G − V− that is proper and Helly. One may want to use Lemma 4.2 to find a minimum set E−
→
−
of edges (i.e., E A−V− (α) for some point α) to finish the task. However, Lemma 4.2 has not ruled out
the possibility that we delete less edges to break all long holes, and subsequently add edges to fix the
incurred subgraphs in {claw, S3 , S3 , C4 , C5 , C` } with ` 6 k3 + 3. So we need the following lemma.
Lemma 5.1. Let (V− , E− , E+ ) be an inclusion-wise minimal editing set of a reduced graph G. If |E+ | 6 k3 ,
then E+ = ∅.
Proof. We may assume without loss of generality V− = ∅, as otherwise it suffices to consider the inclusionwise minimal editing set (∅, E− , E+ ) to the still reduced graph G − V− . Let A be a proper and Helly arc
0
0
model for G. Let E−
be an inclusion-wise minimal subset of E− such that for every hole in G − E−
, the
0
union of arcs for its vertices does not cover the circle of A. We argue the existence of E− by showing
that E− itself satisfies this condition. Suppose for contradiction that there exists in G − E− a hole whose
arcs cover the circle of A. Then we can find a minimal subset of them that covers the circle of A. By
Corollary 2.7, this subset has at least k3 + 4 vertices, and thus the length of the hole in G − E− is at least
k3 + 4. But then it cannot be fixed by the addition of the at most k3 edges from E+ .
0
Now for the harder part, we argue that G := G − E−
is already a unit interval graph. Together with
0
the inclusion-wise minimality, it would imply E− = E− and E+ = ∅.
Suppose for contradiction that G[X] is a claw, S3 , S3 , or a hole for some X ⊆ V(G). We find three
0
vertices u, v, w ∈ X such that uw ∈ E−
and uv, vw ∈ E(G) as follows. By Corollary 2.7 and the fact that
G is {C4 , C5 }-free, at least six arcs are required to cover the circle. As a result, if the arcs for a set Y of at
most six vertices
S covers the circle, then G[Y] must be a C6 , and its subgraph G[Y] cannot be a claw, S3 , or
S3 . Therefore, v∈X A(v) cannot cover the whole circle when G[X] is a claw, S3 , or S3 . On the other hand,
0
from the selection of E−
, this is also true when G[X] is a hole. Thus, G[X] is a unit interval graph, and we
0
can find two vertices x, z from X having xz ∈ E−
. We find a shortest x-z path in G[X]. If the path has more
than one inner vertex, then it makes a hole together with xz; as G[X] is a unit interval graph, this would
0
0
imply that there exists an inner vertex y of this path such that xy ∈ E−
or yz ∈ E−
. We consider then the
new pair x, y or y, z accordingly. Note that their distance in G[X] is smaller than xz, and hence repeating
this argument (at most |X| − 3 times) will end with two vertices with distance precisely 2 in G[X]. They
are the desired u and w, while any common neighbor of them in G[X] can be v.
0
By the minimality of E−
, in G + uw there exists a hole H such that arcs for its vertices cover the circle
in A. This hole H necessarily passes uw, and we denote it by x1 x2 · · · x`−1 x` , where x1 = u and x` = w.
15
Note that A(u) intersects A(w), and since A is proper and Helly, A(u), A(v), A(w) cannot cover the circle;
moreover, it cannot happen that A(v) intersects all the arcs A(xi ) for 1 < i < ` simultaneously. From
x1 x2 · · · x`−1 x` we can find p and q such that 1 6 p < p + 1 < q 6 ` and vxp , vxq ∈ E(G) but vxi 6∈ E(G)
for every p < i < q. Here possibly p = 1 and/or q = `. Then vxp · · · xq makes a hole of G, and the union
0
of its arcs covers the circle, contradicting the definition of E−
. This concludes the proof.
Therefore, a yes-instance on a reduced graph always has a solution that does not add any edge.
By Lemma 4.2, for any editing set (V− , E− , ∅), we can always find some point α in the model and use
→
−
E A−V− (α) to replace E− . After that, we can use the vertices “close” to this point to replace V− . Therefore,
the problem again boils down to find some “weak point” in the arc model. This observation is formalized
in the following lemma. We point out that this result is stronger than required by the linear-time algorithm,
and we present in the current form for its own interest (see Section 6 for more discussions).
Lemma 5.2. Given a proper Helly circular-arc graph G and a nonnegative integer p, we can calculate in
O(m) time the minimum number q such that G has an editing set of size (p, q, 0). In the same time we can
find such an editing set.
Proof. We may assume that G is not chordal; otherwise, by Corollary 2.3, G is a unit interval graph and
the problem becomes trivial because an empty set will suffice. Let us fix a proper and Helly arc model
A for G. The lemma follows from Lemma 3.1 when there is some point ρ satisfying KA (ρ) 6 p. Hence,
we may assume that no such point exists, and for any subset V− of at most p vertices, G − V− remains
a proper Helly circular-arc graph and is non-chordal. Hence, q > 0. For each point ρ in A, we can
ρ
ρ
define an editing set (V−
, Eρ− , ∅) by taking the p vertices in KA (ρ) with the most clockwise arcs as V−
and
→
−
ρ
ρ
E A−V− (ρ) as E− . We argue first that the minimum cardinality of this edge set, taken among all points in
A is the desired number q. See Fig. 9.
∗
Let (V−
, E∗− , ∅) be an editing set of G with size (p, q, 0). According to Lemma 4.2, there is a point α
→
−
0
∗
0
such that the deletion of E−
:= E A−V−∗ (α) from G − V−
makes it a unit interval graph and |E−
| 6 |E∗− |.
∗
We now consider the original model A. Note that a vertex in V−
is in either KA (α) or {v 6∈ KA (α) : u →
v, u ∈ KA (α)}; otherwise replacing this vertex by any end of an edge in E∗− , and removing this edge from
∗
E∗− gives an editing set of size (p, q − 1, 0). Let V− comprise the |V−
∩ KA (α)| vertices of KA (α) whose
∗
arcs are the most clockwise in them, as well as the first |V− \ KA (α)| vertices whose arcs are immediately
→
−
to the right of α. And let E− := E A−V− (α). It is easy to verify that |E− | 6 |E∗− | = q and (V− , E− , ∅) is also
an editing set of G (Lemma 4.1). Note that arcs for V− are consecutive in A. Let v be the vertex in V−
with the most clockwise arc, and then ccp(v) + is the desired point ρ.
We give now the O(m)-time algorithm for finding the desired point, for which we assume that A is
0.5
canonical. It suffices to consider the 2n points i + 0.5 for i ∈ {0, . . . , 2n − 1}. We calculate first the V−
0.5
0.5
and E− , and maintain a queue that is initially set to be V− . For i = 1, . . . , 2n − 1, we deduce the new
i+0.5
i−0.5
sets V−
and Ei+0.5
from V−
and Ei−0.5
as follows. If i is a clockwise endpoint of some arc, then
−
−
i+0.5
i−0.5
i+0.5
i−0.5
V−
= V−
and E−
= E− . Otherwise, i = ccp(v) for some vertex v, then we enqueue v, and
i+0.5
dequeue u. We set V−
to be the vertices in the queue, whose size remains p. The different edges
i+0.5
i−0.5
between E−
and E−
are those incident to u and v. In particular, Ei+0.5
\ Ei−0.5
= {ux : u → x, x 6∈
−
−
i+0.5
i−0.5
i+0.5
i−0.5
0.5
V− }, while E−
\ E−
= {xv : x → v, x 6∈ V− }. Note that the initial sets V−
and E0.5
− can be
found in O(m) time, and then each vertex and its adjacency is scanned exactly once. The total running
time is O(m). This concludes the proof.
u
v
β
α
Figure 9: Illustration for the proof of
α
Lemma 5.2. Here p = 2 and q = 3. Then V−
α
consists of the two thick arcs, and |E− | = 3.
Moving from point α to β (= α + 1) gives
α
|Eβ
− | = |E− | + 4 − 2 = 5.
Again, one should note that in the general case (when both p, q > 0), the point identified by Lemma 5.2
may not be the thinnest point for vertices or the thinnest point for edges, as specified by respectively
16
Lemmas 3.1 and 4.2. Indeed, for different values of p, the thinnest points found by Lemma 5.2 may be
different.
The mixed hole covers consists of both vertices and edges, and thus the combinatorial characterization
given in Lemma 5.2 extends Lemmas 3.1 and 4.2. The algorithm used in the proof is similar as that of
Theorem 4.3. Recall that a reduced graph is a proper Helly circular-arc graph. Thus, Lemmas 5.1 and
5.2 have the following consequence: It suffices to call the algorithm with p = k1 , and returns the found
editing set if q 6 k2 , or “NO” otherwise.
Corollary 5.3. The unit interval editing problem can be solved in O(m) time on reduced graphs.
Putting together these steps, the fixed-parameter tractability of unit interval editing follows. Note that
to fill a hole, we need to add an edge whose ends have distance 2.
Proof of Theorem 1.4. We start by calling Theorem 2.5. If a subgraph in F or W5 is detected, then we
branch on all possible ways of destroying it or the contained C5 . Otherwise, we have in our disposal a
proper and Helly arc model for G, and we call Lemma 2.8 to find a shortest hole C` . If ` 6 k3 + 3, then
we either delete one of its ` vertices and ` edges, or add one of ` edges hi hi+2 (the subscripts are modulo
`). One of the three parameters decreases by 1. We repeat these two steps until some parameter becomes
negative, then we terminate the algorithm by returning “NO”; or the graph is reduced, and then call the
algorithm of Corollary 5.3 to solve it. The correctness of this algorithm follows from Lemma 2.8 and
Corollary 5.3. In the disposal of a subgraph of F, at most 21 recursive calls are made, while 3` for a C` ,
each having a parameter k − 1. Therefore, the total number of instances (with reduced graphs) made in
the algorithm is O((3k3 + 21)k ). It follows that the total running time of the algorithm is 2O(k log k) · m.
It is worth mentioning that Lemma 5.2 actually implies a linear-time algorithm for the unit interval
deletion problem (which allows k1 vertex deletions and k2 edge deletions) on the proper Helly circular-arc
graphs and an O(10k1 +k2 · m)-time algorithm for it on general graphs. The constant 10 can be even
smaller if we notice that (1) the problem is also easy on fat W5 ’s, and (2) the worst cases for vertex
deletions (S3 ’s and S3 ’s) and edge deletions (C4 ’s) are different.
6
Concluding remarks
All aforementioned algorithms exploit the characterization of unit interval graphs by forbidden induced
subgraphs [29]. Very recently, Bliznets et al. [4] used a different approach to produce a subexponentialtime parameterized algorithm for unit interval completion (whose polynomial factor is however not
linear). Using a parameter-preserving reduction from vertex cover [21], one can show that the vertex
deletion version cannot be solved in 2o(k) · nO(1) time, unless the Exponent Time Hypothesis fails [7].
Now that the edge deletion version is FPT as well, one may want to ask to which side it belongs. The
evidence we now have is in favor of the hard side: In all related graph classes, the edge deletion versions
seem to be harder than their vertex deletion counterparts.
As said, it is not hard to slightly improve the constant c in the running time O(ck · m), but a significant
improvement would need some new observation(s). More interesting would be to fathom their limits. In
particular, can the deletion problems be solved in time O(2k · m)?
Polynomial kernels for unit interval completion [2] and unit interval vertex deletion [13] have been
known for a while. Using the approximation algorithm of Theorem 1.2, we [19] recently developed an
O(k4 )-vertex kernel for unit interval vertex deletion, improving from the O(k53 ) one of Fomin et al. [13].
We conjecture that the unit interval edge deletion problem also has a small polynomial kernel.
The algorithm for unit interval editing is the second nontrivial FPT algorithm for the general editing
problem. The main ingredient of our algorithm is the characterization of the mixed deletion of vertices
and edges to break holes. A similar study has been conducted in the algorithm for the chordal editing
problem [10]. In contrast to that, Lemmas 5.1 and 5.2 are somewhat stronger. For example, we have
shown that once small forbidden subgraphs have been all fixed, no edge additions are further needed.
Together with Marx, we had conjectured that this is also true for the chordal editing problem, but we failed
to find a proof. Very little study had been done on the mixed deletion of vertices and edges [25]. We hope
that our work will trigger more studies on this direction, which will further deepen our understanding of
various graph classes.
We point out that although we start by breaking small forbidden induced subgraphs, our major
proof technique is instead manipulating (proper/unit) interval models. The technique of combining
17
(a) F1
(f) K2,3
(b) F2
(d) †
(c) F3
(g) F4
(h) F5
(i) F6
(j) F7
(e) ‡
(k) F8
(l) F9
Figure 10: Forbidden induced graphs.
Circular-arc
Normal circular-arc
Helly circular-arc
∗
claw S3 W5 F1 F2 F6 F7 S3∗ C`∗ C2` C2`−1
F2 F3 F8 F9 K2,3 C6 F4 F5 † ‡ C`∗ W`
C`
Proper circular-arc
Normal Helly circular-arc
Chordal
W4 S3
{CI(n, k) : n > 2k}
claw
C`
F2 F3 † ‡
Proper Helly circular-arc
Unit circular-arc
W4
Interval
{CI(n, k) : n > 3k}
Unit Helly circular-arc
claw
C`
Unit interval = Proper interval
Figure 11: Forbidden induced subgraphs and containment relations of related graph classes (` > 4)
(constructive) interval models and (destructive) forbidden induced subgraphs is worth further study on
related problems.
Appendix
For the convenience of the reader, we collect related graph classes and their containment relations in
Fig. 11, which is adapted from Lin et al. [22] (note that some of these graph classes are not used in
the present paper). The CI(n, k) graphs are defined by Tucker [27]; see also [22]. Other subgraphs
that have not been introduced in the main text are depicted in Fig. 10. The relations in Fig. 11 can be
viewed both from the intersection models, arcs or intervals, and forbidden induced subgraphs, every
minimal forbidden induced subgraph of a super-class being a (not necessarily minimal) forbidden induced
subgraph of its subclass. For example, Proposition 2.6 and Corollary 2.7 are actually properties of normal
Helly circular-arc graphs. For a normal Helly circular-arc graph that is not chordal, every arc model has to
be normal and Helly [22, 9]. This is also true for the subclass of proper Helly circular-arc graphs, but an
arc model for a proper Helly circular-arc graph may not be proper.
A word of caution is worth on the definition of proper Helly circular-arc graphs. One graph might
admit two arc models, one being proper and the other Helly, but no arc model that is both proper and
18
Helly, e.g., the S3 and the W4 . Therefore, the class of proper Helly circular-arc graphs does not contain all
those graphs being both proper circular-arc graphs and Helly circular-arc graphs, but a proper subclass of
it. A similar remark applies to normal Helly circular-arc graphs.
For the three classes at the top of Fig. 11, their characterizations by minimal forbidden induced
subgraphs are still open. At the third level, the minimal forbidden induced subgraphs for proper circulararc graphs and normal Helly circular-arc graphs are completely determined by Tucker [27] and Cao et
al. [9]. For all the classes at lower levels, their forbidden induced subgraphs with respect to its immediate
super-classes are given. From them we are able to derive all the minimal forbidden induced subgraphs for
each of these classes.
For example, the characterization of unit interval graphs (Theorem 2.1) follows from the characterization of interval graphs and that we can find a claw in an F2 , an F3 , a † that is not an S3 , or a ‡ that is
not an S3 . Likewise, the minimal forbidden induced subgraphs of proper Helly circular-arc graphs stated
in Theorem 2.2 can be derived from those of proper circular-arc graphs and by Corollary 5 of [22]: A
proper circular-arc graph that is not a proper Helly circular-arc graph must contain a W4 or S3 . Clearly, S∗3
contains an S3 . To see that each of F1 , F2 , F6 , F7 , and C2` , C∗2`−1 for ` > 4 contain a W4 , it is equivalent
to check that each of F1 , F2 , F6 , F7 , and C2` , C∗2`−1 for ` > 4 contains a W4 , i.e., two non-incident edges
and another independent vertex v. This can be directly read from Fig. 10 for F1 , F2 , F6 , F7 . Let h1 , h2 , . . .
denote the vertices in the hole of C2` and C∗2`−1 . Then edges h1 h2 and h4 h5 are non-incident. In a C∗7 ,
the vertex not in the hole can be the v, while in all other holes longer than 7, the vertex h7 can be the v.
References
[1] Jørgen Bang-Jensen and Pavol Hell. On chordal proper circular arc graphs. Discrete Mathematics, 128(1):395–
398, 1994. doi:10.1016/0012-365X(94)90130-9.
[2] Stéphane Bessy and Anthony Perez. Polynomial kernels for proper interval completion and related problems.
Information and Computation, 231:89–108, 2013. doi:10.1016/j.ic.2013.08.006.
[3] René van Bevern, Christian Komusiewicz, Hannes Moser, and Rolf Niedermeier. Measuring indifference: Unit
interval vertex deletion. In Dimitrios M. Thilikos, editor, Graph-Theoretic Concepts in Computer Science (WG),
volume 6410 of LNCS, pages 232–243. Springer, 2010. doi:10.1007/978-3-642-16926-7_22.
[4] Ivan Bliznets, Fedor V. Fomin, Marcin Pilipczuk, and Michał Pilipczuk. A subexponential parameterized algorithm
for proper interval completion. SIAM Journal on Discrete Mathematics, 29(4):1961–1987, 2015. A preliminary
version appeared in ESA 2014. doi:10.1137/140988565.
[5] Hans L. Bodlaender and Babette van Antwerpen-de Fluiter. On intervalizing k-colored graphs for DNA physical
mapping. Discrete Applied Mathematics, 71(1-3):55–77, 1996. A preliminary version appeared in ICALP 1995.
doi:10.1016/S0166-218X(96)00057-1.
[6] Pablo Burzyn, Flavia Bonomo, and Guillermo Durán. NP-completeness results for edge modification problems.
Discrete Applied Mathematics, 154(13):1824–1844, 2006. doi:10.1016/j.dam.2006.03.031.
[7] Leizhen Cai. Fixed-parameter tractability of graph modification problems for hereditary properties. Information
Processing Letters, 58(4):171–176, 1996. doi:10.1016/0020-0190(96)00050-6.
[8] Yixin Cao. Linear recognition of almost interval graphs. In Krauthgamer [20], pages 1096–1115. Full version
available at arXiv:1403.1515. doi:10.1137/1.9781611974331.ch77.
[9] Yixin Cao, Luciano N. Grippo, and Martı́n D. Safe. Forbidden induced subgraphs of normal Helly circular-arc
graphs: Characterization and detection. Discrete Applied Mathematics, 216:67–83, 2017. doi:10.1016/j.dam.
2015.08.023.
[10] Yixin Cao and Dániel Marx. Chordal editing is fixed-parameter tractable. Algorithmica, 75(1):118–137, 2016. A
preliminary version appeared in STACS 2014. doi:10.1007/s00453-015-0014-x.
[11] Xiaotie Deng, Pavol Hell, and Jing Huang. Linear-time representation algorithms for proper circular-arc
graphs and proper interval graphs. SIAM Journal on Computing, 25(2):390–403, 1996. doi:10.1137/
S0097539792269095.
[12] Rodney G. Downey and Michael R. Fellows. Fundamentals of Parameterized Complexity. Undergraduate texts in
computer science. Springer, 2013. doi:10.1007/978-1-4471-5559-1.
[13] Fedor V. Fomin, Saket Saurabh, and Yngve Villanger. A polynomial kernel for proper interval vertex deletion.
SIAM Journal on Discrete Mathematics, 27(4):1964–1976, 2013. A preliminary version appeared in ESA 2012.
doi:10.1137/12089051X.
[14] Delbert R. Fulkerson and Oliver A. Gross. Incidence matrices and interval graphs. Pacific Journal of Mathematics,
15(3):835–855, 1965. doi:10.2140/pjm.1965.15.835.
19
[15] Fănică Gavril. Algorithms on circular-arc graphs. Networks, 4:357–369, 1974. doi:10.1002/net.3230040407.
[16] Paul W. Goldberg, Martin C. Golumbic, Haim Kaplan, and Ron Shamir. Four strikes against physical mapping of
DNA. Journal of Computational Biology, 2(1):139–152, 1995. doi:10.1089/cmb.1995.2.139.
[17] Pim van ’t Hof and Yngve Villanger. Proper interval vertex deletion. Algorithmica, 65(4):845–867, 2013.
doi:10.1007/s00453-012-9661-3.
[18] Haim Kaplan, Ron Shamir, and Robert Endre Tarjan. Tractability of parameterized completion problems on
chordal, strongly chordal, and proper interval graphs. SIAM Journal on Computing, 28(5):1906–1922, 1999. A
preliminary version appeared in FOCS 1994. doi:10.1137/S0097539796303044.
[19] Yuping Ke, Yixin Cao, Xiating Ouyang, and Jianxin Wang. Unit interval vertex deletion: Fewer vertices are
relevant. arXiv:1607.01162, 2016.
[20] Robert Krauthgamer, editor. Proceedings of the 27th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA).
SIAM, 2016. doi:10.1137/1.9781611974331.
[21] John M. Lewis and Mihalis Yannakakis. The node-deletion problem for hereditary properties is NP-complete.
Journal of Computer and System Sciences, 20(2):219–230, 1980. Preliminary versions independently presented
in STOC 1978. doi:10.1016/0022-0000(80)90060-4.
[22] Min Chih Lin, Francisco J. Soulignac, and Jayme L. Szwarcfiter. Normal Helly circular-arc graphs and its
subclasses. Discrete Applied Mathematics, 161(7-8):1037–1059, 2013. doi:10.1016/j.dam.2012.11.005.
[23] Yunlong Liu, Jianxin Wang, Jie You, Jianer Chen, and Yixin Cao. Edge deletion problems: Branching facilitated
by modular decomposition. Theoretical Computer Science, 573:63–70, 2015. doi:10.1016/j.tcs.2015.01.049.
[24] Dániel Marx. Chordal deletion is fixed-parameter tractable. Algorithmica, 57(4):747–768, 2010. A preliminary
version appeared in WG 2006. doi:10.1007/s00453-008-9233-8.
[25] Dániel Marx, Barry O’Sullivan, and Igor Razgon. Finding small separators in linear time via treewidth reduction.
ACM Transactions on Algorithms, 9(4):30.1–30.35, 2013. A preliminary version appeared in STACS 2010.
doi:10.1145/2500119.
[26] Fred S. Roberts. Indifference graphs. In Frank Harary, editor, Proof Techniques in Graph Theory (Proc. Second
Ann Arbor Graph Theory Conf., 1968), pages 139–146. Academic Press, New York, 1969.
[27] Alan C. Tucker. Structure theorems for some circular-arc graphs. Discrete Mathematics, 7(1-2):167–195, 1974.
doi:10.1016/S0012-365X(74)80027-0.
[28] Yngve Villanger. Proper interval vertex deletion. In Venkatesh Raman and Saket Saurabh, editors, Parameterized
and Exact Computation - IPEC 2010, volume 6478 of LNCS, pages 228–238. Springer, 2010. doi:10.1007/
978-3-642-17493-3_22.
[29] Gerd Wegner. Eigenschaften der Nerven homologisch-einfacher Familien im Rn . PhD thesis, Universität Göttingen,
1967.
[30] Mihalis Yannakakis. Computing the minimum fill-in is NP-complete. SIAM Journal on Algebraic and Discrete
Methods, 2(1):77–79, 1981. doi:10.1137/0602010.
20
| 8cs.DS
|
arXiv:1709.07565v1 [cs.CV] 22 Sep 2017
NOVEL EVALUATION METRICS FOR SEAM CARVING BASED IMAGE RETARGETING
Tam V. Nguyen
Guangyu Gao
Department of Computer Science
University of Dayton
Email: [email protected]
School of Software
Beijing Institute of Technology
Email: [email protected]
ABSTRACT
Image retargeting effectively resizes images by preserving the
recognizability of important image regions. Most of retargeting methods rely on good importance maps as a cue to retain
or remove certain regions in the input image. In addition, the
traditional evaluation exhaustively depends on user ratings.
There is a legitimate need for a methodological approach for
evaluating retargeted results. Therefore, in this paper, we conduct a study and analysis on the prominent method in image retargeting, Seam Carving. First, we introduce two novel
evaluation metrics which can be considered as the proxy of
user ratings. Second, we exploit salient object dataset as a
benchmark for this task. We then investigate different types
of importance maps for this particular problem. The experiments show that humans in general agree with the evaluation
metrics on the retargeted results and some importance map
methods are consistently more favorable than others.
Fig. 1. The flowchart of Seam Carving on a given image with
the importance map from different methods, namely, edge detector, human fixation predictor, and salient object detector.
The removal map is later generated by highlighting the least
important seams. The red lines are represented the removal
seams. The accordingly retargeted images are finally constructed by removing the red lines to reach the desired size.
Index Terms— Seam Carving, Image Retargeting, Visual
Saliency
1. INTRODUCTION
Image retargeting, sometimes referred as image cropping,
thumbnailing, or resizing, is beneficial for some practical
scenarios, i.e., facilitating large image viewing in small size
displays, particularly on mobile devices. This is a very
challenging task since it requires preserving the relevant information while maintaining an aesthetically pleasing image
for viewers. The premise of this task is to remove indistinct
regions and retain the context with the most salient regions.
In the pioneering work, Setlur et al. [1] propose using an
importance map of the source image obtained from saliency
and face detection. In the importance map, the pixels with
higher values are most likely preserved and vice versa. If the
specified size contains all the important regions, the source
image is simply cropped. Otherwise, the important regions
are removed from the image, and fill the resulting “holes”
using the background creation technique. Later, Avidan et
al. [2] propose the Seam Carving method based on the importance map computed from gradient magnitude. Seam
Carving functions by constructing a number of seams (paths
of least importance) in an image and automatically removes
seams to reduce image size. Zhang et al. [3] present an image
resizing method that attempts to ensure that important local
regions undergo a geometric similarity transformation, and
at the same time, image edge structure is preserved. Suh et
al. [4] propose a general thumbnail cropping method based
on a saliency model that finds the informative portion of images and cuts out the non-core part of images. Marchesotti et
al. [5] propose a framework for image thumbnailing based on
visual similarity. Their underlying assumption is that images
sharing their global visual appearance are likely to share similar saliency values. While other works are dedicated to still
images, Chamaret and Le Meur [6] propose a video retargeting algorithm. Meanwhile, Rubinstein et al. [7] extend Seam
Carving [2] into video retargeting.
To date, the existing evaluation scheme mostly depends
on user ratings. However, it is not always feasible to recruit
a large pool of participants for the evaluation. Also, there
is mostly impossible to get the same participant pool of a
previous work to make a fair comparison. Thus there is a
legitimate need of an automatic way to evaluate these retargeting methods. In this paper, we revisit and further analyze
Fig. 2. The two novel metrics, namely, Mean Area Ratio, and Mean Sum of Squared Distances. From left to right: Original
image, (a) the ground truth saliency map, (b) the shape points of the ground truth map, (c) the retargeted ground truth map from
COV [8], (d) the shape points of the retargeted ground truth map, (e) the mean area ratio map, (f) the mapping between two
correspondence sets.
the most popular method, Seam Carving, for image retargeting. Our contribution is two-fold. First, we propose two novel
metrics to systematically evaluate the retargeting algorithms,
namely, Mean Area Ratio (MAR) and Mean Sum of Squared
Distances (MSSD). Our novel metrics focus on how much
shape of the salient object(s) is distorted after the retargeting process. Second, we evaluate various types of importance
map, namely, fixation prediction map, salient object map, and
edge map, with the newly proposed metrics.
2. SEAM CARVING REVISIT AND PROPOSED
EVALUATION METRICS
2.1. Seam Carving Revisit
Seam Carving, the most popular method in image retargeting,
aims to automatically retarget the images into a certain size to
facilitate the viewing purpose as aforementioned. Let I be an
m × n image. As illustrated in Figure 1, the first step is the
computation of an importance map S, which quantifies the
importance of every pixel in the image. Every pixel in the importance map is assigned a value within [0, 1], where higher
values mean higher importance. Assume I is a landscape image where n > m, we aim to reduce its width. The vertical
seam s, an 8-connected path in the image from the top to the
bottom containing one pixel per row, is defined as below:
m
s = {si }m
i=1 = {(i, y(i))}i=1 , s.t.∀i, |y(i) − y(i − 1)| ≤ 1,
(1)
where y(i) is the corresponding column of row i within
the seam. Our goal is to find the optimal seam that minimizes:
s∗ = min
s
m
X
S(si ),
(2)
i=1
where S(si ) is the importance value of one seam pixel.
Eqn. (2) can be solved by dynamic programming. This optimal seam is later removed out of the input image. This
process repeats until the image reaches its desired dimension.
It is worth noting that the recent years witness the rapid
popularity of smartphones and tablets that equips people with
imaging capabilities. In fact, people are taking photos in different ways. Traditional filmmakers take more photos about
the landscape than human figures. However, on a mobile
phone, people prefer to take pictures in the portrait mode.
Due to this difference in people’s preferences, applications
like Instagram have been developed which meets the demands
of both groups of people by asking them to crop the image to
the square size. In the social media, most of profile images are
in the square form, i.e., Facebook and Twitter. One reasonable
explanation is that squared photos display well in a feed format. In this work, we utilize the Seam Carving method into
an application, so called Make-It-Square, which automatically retargets images into the square size. In particular, the
Seam Carving process loops for n − m times until the landscape image reaches its expected square size. For the portrait
image, we transpose the image and use the same function to
find the optimal vertical seam.
2.2. Proposed Evaluation Metrics
In order to mitigate the dependency of user ratings, we propose two additional metrics to systematically evaluate the retargeting algorithms, namely, Mean Area Ratio and Mean
Sum of Squared Distances. Our motivation is that the users
prefer the shape of the salient object(s) is preserved after the
image retargeting process as discussed in [2]. As shown in
Fig. 1, the distorted boxes in the first two rows (retargeted
images) are not entertained by the viewers.
Our first metric, the Mean Area Ratio, measures how
much the salient object(s) can be preserved after the image
retargeting. We simultaneously remove seams on both the
original image and its ground truth saliency map. Obviously,
the retargeted groundtruth map has the exactly same size with
the retargeted image. For each input image, the area ratio is
computed as the ratio between the salient regions in the retargeted ground truth map and the ground truth salient areas, as
Fig. 3. From left to right: Original image, and importance maps from 6 different methods: (a) Sobel edge map [9], (b)
Structured edge map [10], (c) boolean map based saliency (BMS [11]), (d) saliency based on region covariance (COV [8]), (e)
high-dimensional color transform (HDCT [12]), (f) discriminative regional feature integration (DRFI [13]).
shown in Fig. 2e. The area ratio is 1 when the whole salient
regions are retained. The Mean Area Ratio, MAR, for a set of
input images is computed over the area ratios of all images.
Our second metric, the Mean Sum of Squared Distances,
evaluates the shape similarity of the salient regions before and
after the image retargeting. We adopt Shape Contexts [14] to
measure the shape similarity. For each image, Shape Contexts
compute the shape correspondences of two given silhouettes
(the ground truth map and the retargeted ground truth maps
as shown in Fig. 2b, d). Next, the distances between two correspondence sets are summed as illustrated in Fig. 2f. The
sum of squared distances is 0 when two shapes are identical.
Eventually the Mean Sum of Squared Distances, MSSD, is
computed across over all images.
Actually, the two proposed evaluation metrics are complementary to each other. MAR measures how much salient object(s) are maintained, whereas MSSD measures the amount
of distortion after the image retargeting process.
2.3. Selection of Importance Map
In literature, the edge map is first introduced as the importance map for image retargeting problem [2]. Additionally,
the importance level can be measured by visual saliency values. There exist two popular outputs of visual saliency prediction, namely, the predicted human fixation map for fixation prediction, and the salient object map for salient object/region detection. In literature, there also exist many efforts to predict visual saliency with different cues, i.e., depth
matters [15], audio source [16], touch behavior [17], object
proposals [18, 19], and semantic priors [20]. In this paper, we
consider three types of importance maps as follows.
Edge map is retrieved from the edge detection process, a fundamental task in computer vision since the early
1970’s [21, 22]. Early works [9, 23] focused on the detection
of intensity or color gradients. For example, the popular Sobel detector [9] computes an approximation of the gradient
of the image intensity function. Recently, Dollar et al. [10]
proposed structured edge detection (SE) by formulating the
problem of edge detection as predicting local segmentation
masks given input image patches. In this work, we consider
different edge detectors [9, 10].
Fixation prediction map is obtained from trained models
which are constructed originally to understand human viewing patterns. Actually, these models aim to predict points that
people look at (freeviewing of natural scenes usually for 35 seconds). The typical ground-truth fixation map includes
several fixation points smoothened by a Gaussian kernel. We
consider using two state-of-the-art models, namely, Boolean
Map based Saliency (BMS [11]) and saliency based on region
covariance (COV [8]) for the later evaluation.
Salient object map is computed from models which
aim to detect and segment the most salient object(s) as a
whole. Note that a typical pixel-accurate ground-truth map
usually contains several regions marked by humans. As recommended in the extensive survey [24], we consider two
state-of-the-art models, namely, saliency based on Discriminative Regional Feature Integration (DRFI [13]) and HighDimensional Color Transform (HDCT [12]).
Fig. 3 shows the importance maps generated from different computational methods. Note that edge maps and fixation prediction maps are of low resolution and highlight edges
whereas the salient object maps focus on the entire objects.
3. EVALUATION
It is obvious that the benchmark of image retargeting task requires a set of input images with their corresponding saliency
map. This requirement elegantly fits the settings of salient object datasets. Therefore, we exploit the popular MSRA-1000
dataset [25], which contains 1, 000 images with the annotated
pixel-wise ground truth of salient regions, for the evaluation.
We first show the visual comparison of retargeted images
from different importance maps. As observed from Fig. 4,
the retargeted results from salient object detection methods
well preserve the main salient objects without distortion.
Though fixation prediction is in general biologically plausible and suggests important regions as the way as humans
look at, their retargeted images lose details. Meanwhile, the
retargeted images from edge-based importance map lose both
details and layout structure.
Next, we conduct a user study to evaluate the performance of retargeted images from different input saliency
maps on previously mentioned MSRA-1000 dataset [25]. We
run Make-It-Square on the dataset to obtain 1, 000 retargeted
squared images. 40 participants (14 are female) who are
Fig. 4. Visual comparison of retargeted images from different importance maps on MSRA-1000 dataset [25]. From left to right:
Original image, the ground truth saliency map, the pairs of retargeted image and the retargeted groundtruth saliency map with
the importance maps from Sobel, Structured Edge (SE), BMS, VOC, HDCT, DRFI, respectively. (Please view in high 400%
resolution for best visual effect).
Table 1. The performance of different importance maps on
image retargeting.
Importance Map
Sobel [9]
Structured Edge [10]
COV [8]
BMS [11]
HDCT [12]
DRFI [13]
User Ratings
1.3
1.9
3.5
3.2
5.4
5.7
MAR
0.8976
0.9132
0.9581
0.9638
0.9840
0.9877
MSSD
0.0406
0.0402
0.0395
0.0395
0.0387
0.0389
university staff/students are involved in this experiment, and
a set of images is provided to each participant. Note that
every image set contains 50 random images and six other
retargeted results where each method is randomly labeled
from 1 to 6 to hide identities. The participant is requested
to rate all methods with the scores (1-6), where 1 means bad
viewing experience and 6 means excellent viewing experience. As shown in Table 1, users prefer the salient object
map methods, HDCT [12] and DRFI [13], whereas the retargeted results from edge map, Sobel [9], Structured Edge [10],
receive the least rating.
We then compute two evaluation metrics, MAR and
MSSD, and the results are generally similar with user ratings. Also shown in Table 1, the retargeted images obtained
from the salient object map source are consistently more favorable than others, namely, achieving the highest MAR and
the lowest MSSD. On the contrary, the retargeted results of
edge maps receive the lowest MAR and the highest MSSD.
In addition, we further compute the Pearson coefficient
correlations (CC) (defined in [24]) between user ratings and
the two novel metrics. Note that the correlation of one met-
Table 2. The Pearson coefficient correlation [24] among
three metrics, user ratings, MAR and MSSD.
User Ratings
MAR
- MSSD
User Ratings
1
0.955
0.977
MAR
0.955
1
0.981
- MSSD
0.977
0.981
1
ric score and itself is 1. As shown in Table 2, the CCs between user ratings and MAR and negative MSSD are 0.955
and 0.977, respectively. This demonstrates those two metrics
are highly correlated with users’s responses. Hence, the proposed metrics can be used as the proxy of user ratings.
4. CONCLUSION AND FUTURE WORK
In this paper, we introduce two novel metrics to automatically evaluate Seam Carving for the image retargeting task.
We utilized salient object dataset as a benchmark and showed
that the newly proposed metrics are highly correlated with the
user ratings across six different importance maps. We also
found that the retargeted results, with the salient object map
used as the importance map, are consistently more favorable
than others. We believe that the new benchmark type and our
evaluation measures will lead to improved retargeting algorithms, as well as better understanding of image retargeting
problem.
For future work, we aim to investigate other image retargeting operators apart from Seam Carving. We also would
like to extend our work by considering additional cues, e.g.,
the depth in RGBD images or motion information in videos.
5. REFERENCES
[1] Vidya Setlur, Saeko Takagi, Ramesh Raskar, Michael
Gleicher, and Bruce Gooch, “Automatic image retargeting,” in International Conference on Mobile and Ubiquitous Multimedia, 2005, pp. 59–68.
[2] Shai Avidan and Ariel Shamir, “Seam carving for
content-aware image resizing,” ACM Trans. Graph., vol.
26, no. 3, pp. 10, 2007.
[3] Guo-Xin Zhang, Ming-Ming Cheng, Shi-Min Hu, and
Ralph R Martin, “A shape-preserving approach to image
resizing,” in Computer Graphics Forum, 2009, vol. 28,
pp. 1897–1906.
[4] Bongwon Suh, Haibin Ling, Benjamin B Bederson, and
David W Jacobs, “Automatic thumbnail cropping and
its effectiveness,” in ACM UIST, 2003, pp. 95–104.
[5] Luca Marchesotti, Claudio Cifarelli, and Gabriela
Csurka, “A framework for visual saliency detection with
applications to image thumbnailing,” in International
Conference on Computer Vision, 2009, pp. 2232–2239.
[6] Christel Chamaret and Olivier Le Meur, “Attentionbased video reframing: validation using eye-tracking,”
in International Conference on Pattern Recognition,
2008, pp. 1–4.
[7] Michael Rubinstein, Ariel Shamir, and Shai Avidan,
“Improved seam carving for video retargeting,” ACM
TOG., vol. 27, no. 3, 2008.
[8] Erkut Erdem and Aykut Erdem, “Visual saliency estimation by nonlinearly integrating features using region
covariances,” Journal of Vision, vol. 13, no. 4, pp. 1–20,
2013.
[9] I. Sobel and G. Feldman, “A 3x3 Isotropic Gradient
Operator for Image Processing,” 1968.
[10] Piotr Dollár and C. Lawrence Zitnick, “Fast edge detection using structured forests,” Transactions on Pattern
Analysis and Machine Intelligence, vol. 37, no. 8, pp.
1558–1570, 2015.
[11] Jianming Zhang and Stan Sclaroff, “Saliency detection:
A boolean map approach,” in International Conference
on Computer Vision, 2013, pp. 153–160.
[12] Jiwhan Kim, Dongyoon Han, Yu-Wing Tai, and Junmo
Kim, “Salient region detection via high-dimensional
color transform,” in Conference on Computer Vision and
Pattern Recognition, 2014, pp. 883–890.
[13] Huaizu Jiang, Zejian Yuan, Ming-Ming Cheng, Yihong
Gong, Nanning Zheng, and Jingdong Wang, “Salient
object detection: A discriminative regional feature integration approach,” in Conference on Computer Vision
and Pattern Recognition, 2013, pp. 2083–2090.
[14] Serge J. Belongie, Jitendra Malik, and Jan Puzicha,
“Shape matching and object recognition using shape
contexts,” IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 24, no. 4, pp. 509–522, 2002.
[15] Congyan Lang, Tam V. Nguyen, Harish Katti, Karthik
Yadati, Mohan S. Kankanhalli, and Shuicheng Yan,
“Depth matters: Influence of depth cues on visual
saliency,” in European Conference on Computer Vision,
2012, pp. 101–115.
[16] Yanxiang Chen, Tam V. Nguyen, Mohan S. Kankanhalli,
Jun Yuan, Shuicheng Yan, and Meng Wang, “Audio
matters in visual attention,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 24, no. 11,
pp. 1992–2003, 2014.
[17] Bingbing Ni, Mengdi Xu, Tam V. Nguyen, Meng Wang,
Congyan Lang, ZhongYang Huang, and Shuicheng Yan,
“Touch saliency: Characteristics and prediction,” IEEE
Transactions on Multimedia, vol. 16, no. 6, pp. 1779–
1791, 2014.
[18] Tam V. Nguyen, “Salient object detection via objectness proposals,” in AAAI Conference on Artificial Intelligence, 2015, pp. 4286–4287.
[19] Tam V. Nguyen and Jose Sepulveda, “Salient object
detection via augmented hypotheses,” in International
Joint Conference on Artificial Intelligence, 2015, pp.
2176–2182.
[20] Tam V. Nguyen and Luoqi Liu, “Salient object detection
with semantic priors,” in International Joint Conference
on Artificial Intelligence, 2017.
[21] Richard O Duda, Peter E Hart, et al., Pattern classification and scene analysis, vol. 3, Wiley New York, 1973.
[22] Guner Robinson, “Color edge detection,” in 20th
Annual Technical Symposium. International Society for
Optics and Photonics, 1976, pp. 126–133.
[23] John Canny, “A computational approach to edge detection,” Transactions on Pattern Analysis and Machine
Intelligence, vol. 8, no. 6, pp. 679–698, 1986.
[24] Ali Borji, Ming-Ming Cheng, Huaizu Jiang, and Jia Li,
“Salient object detection: A benchmark,” Transactions
on Image Processing, vol. 24, no. 12, pp. 5706–5722,
2015.
[25] Radhakrishna Achanta, Sheila S. Hemami, Francisco J.
Estrada, and Sabine Süsstrunk,
“Frequency-tuned
salient region detection,” in Conference on Computer
Vision and Pattern Recognition, 2009, pp. 1597–1604.
| 1cs.CV
|
The inflation technique solves completely the classical inference problem
Miguel Navascués1 and Elie Wolfe2
1
Institute for Quantum Optics and Quantum Information (IQOQI) Vienna, Austrian Academy of Sciences, Boltzmanngasse 3,
1090 Vienna, Austria
2
Perimeter Institute for Theoretical Physics, 31 Caroline St. N, Waterloo, Ontario, Canada, N2L 2Y5
arXiv:1707.06476v1 [quant-ph] 20 Jul 2017
The causal inference problem consists in determining whether a probability distribution over a
set of observed variables is compatible with a given causal structure. In (Wolfe et al., 2016), one
of us introduced a hierarchy of necessary linear programming constraints which all the observed
distributions compatible with the considered causal structure must satisfy. In this work, we prove
that the inflation hierarchy is complete, i.e., any distribution of the observed variables which
does not admit a realization within the considered causal structure will fail one of the inflation
tests. More quantitatively,
we show that any distribution of measurable events satisfying the nth
1
√
-close in Euclidean norm to a distribution realizable within the given causal
inflation test is O
n
structure. In addition, we show that the corresponding nth -order relaxation of thedual problem
2
consisting in maximizing a kth degree polynomial on the observed variables is O kn -close to
the optimal solution.
I. INTRODUCTION
A Bayesian network or causal structure is a directed acyclic graph where vertices represent random variables, each of
which is generated by a non-deterministic function depending on the value of its parents. Nowadays, causal structures
are commonly used in bioinformatics, medicine, image processing, sports betting, risk analysis and experiments of
quantum nonlocality. It is important to remark that, while some of these variables may be directly observable, some
others are not. Those are called hidden or latent variables.
Due to the presence of latent variables, determining whether a causal structure may be behind the statistics of the
set of observable events, the so-called inference problem, is a very difficult mathematical question. Given a function
on the probabilities of the measurable events, the dual of the causal inference problem is the task of computing its
maximum value when evaluated over probability distributions compatible with the considered causal structure. While
there are a number of more or less effective heuristics to search for probabilistic models compatible with our causality
assumptions or optimizing functionals thereof (Koller and Friedman, 2009), the problem of proving the impossibility
to accommodate experimental data within a given network or bounding the values of a function of such data is still
open. In recent years, though, there have been many advances, see (Fritz, 2012; Fritz and Chaves, 2013; Chaves et al.,
2014; Chaves et al., 2014; Chaves, 2016; Bohr Brask and Chaves, 2017).
In (Wolfe et al., 2016), the authors presented the inflation technique, a hierarchy of necessary constraints, verifiable
via linear programming, which any distribution realizable within the considered causal structure must satisfy. Notably,
the inflation technique allowed the authors to derive polynomial inequalities for the triangle scenario (Fritz and Chaves,
2013), one of the simplest causal structures for which the inference problem is not solved. More generically, the inflation
technique’s versatiliy and practical performance make it a prominent tool to attack the causal inference problem. The
inflation technique also leads naturally to a simple sequence of linear programming relaxations of the dual problem
when the function to evaluate is a polynomial on the probabilities of the observed variables.
In this paper, we will first study the performance of the inflation method on a particular type of causal structures
called causal networks. For these structures, we will show that the hierarchies of relaxations on the inference and
the dual problems provided by the inflation technique are complete. In other words: any distribution of observed
variables that passes all inflation tests must be realizable within the considered causal network, and the sequence of
upper bounds obtained via inflation on the solution f ? of the dual problem converges to f ? asymptotically. Next, we
will show that any inference (dual) problem in an arbitrary causal structure can be mapped to an inference (dual)
problem in a causal network. Put together, these two results imply that the inflation technique, far from being a
relaxation, is an alternative way of understanding general causal structures.
This paper is organized as follows: in Section II, we will formulate the concepts of causal networks and causal
structures, and introduce the inference and dual problems. In Section III, we will review the inflation technique to
solve the causal inference and dual problems. In Section IV we will prove an extension of the finite de Finetti theorem
(Diaconis and Freedman, 1980) for distributions admitting an “inflated” extension.
This theorem will allow us, in
Section V, to prove that any distribution P passing the nth inflation test is O √1n -close in Euclidean norm to a
feasible distribution within the considered causal network. From this theorem it will also follow straightforwardly
2
that the nth inflation relaxation of the dual problem differs from the optimal value on at most O kn , where k is
2
FIG. 1 A generic causal network. The independent latent variables λ1 , λ2 , ... influence the observed variables A1 , A2 , ....
FIG. 2 The triangle scenario.
the degree of the considered polynomial. In Section VI we will describe the mapping from causal structures to causal
networks, hence extending our results to general causal structures. Finally, we will present our conclusions.
II. CAUSAL NETWORKS, CAUSAL STRUCTURES AND THE CAUSAL INFERENCE PROBLEM
A causal network is a type of causal structure with just two layers: a bottom layer of independently distributed
latent random variables {λ1 , λ2 , ..., λL } and a top layer of observable random variables {A1 , A2 , ...}, see Figure 1.
The observable distribution P (A1 , A2 , ..., Am ) is generated via non-deterministic functions Ax = Ax (λ̄Lx ), with λ̄ =
(λ1 , ..., λL ) and Lx ⊂ {1, ..., L}. Here (and in the following) the notation v̄ S , where v̄ is a vector with N entries and
S ⊂ {1, ..., N }, will represent the vector with entries v s , s ∈ S.
Consider, for example, the causal network dubbed “the triangle scenario”, with m = L = 3, see Figure 2. Denoting
A1 , A2 , A3 respectively by A, B, C, we have that a probability distribution P (A, B, C) is realizable in the triangle
scenario if A, B, C are generated via the non-deterministic functions A(λ1 , λ2 ), B(λ2 , λ3 ), C(λ3 , λ1 ). Alternatively,
P (A, B, C) is realizable in the triangle scenario iff it admits a decomposition of the form:
P (A, B, C) =
X
P (λ1 )P (λ2 )P (λ3 )P (A|λ1 , λ2 )P (B|λ2 , λ3 )P (C|λ3 , λ1 ).
(1)
λ1 ,λ2 ,λ3
In a general causal structure, the non-deterministic functions giving rise to the observed variables can also depend
on other observed variables. An example is given by the instrumental scenario (Figure 3, left), where A and λ
are, respectively, a free observable and a latent variable, and the observed variables B and C are generated via the
non-deterministic functions B = B(A, λ), C = C(B, λ).
3
FIG. 3 Left: the instrumental scenario.
re-parenting.
Right: the instrumental scenario after re-wiring, unpacking and
The causal inference problem consists in, given the observable distribution P (A1 , ..., Am ), determine whether it
admits a realization within the considered causal structure. More formally:
Definition 1. The causal inference problem
Given a causal structure C and a probability distribution over the observed variables P (A1 , ..., Am ), decide if there
exists a probability distribution Q for all the random variables in C (observed or latent), compatible with this causal
structure and such that the marginal distribution Q(A1 , ..., Am ) coincides with P (A1 , ..., Am ).
The above definition of the causal inference problem is different from the conventional one, where, given a probability
distribution over observed variables, one asks which causal models can/cannot accommodate it. However, both
problems are oracle-wise equivalent. In the following, we will hence stick to Definition 1.
For illustration, consider this example: if P (A, B, C) is the input of the problem and the causal structure to
test is the triangle scenario, then the causal inference problem is solved either by providing probability distributions P (λ1 ), P (λ2 ), P (λ3 ), P (A|λ1 , λ2 )P (B|λ2 , λ3 )P (C|λ3 , λ1 ) such that eq. (1) holds, or by proving that no such
distributions exist.
In contrast, what we will call the dual inference problem consists in, given a function F on the probabilities of the
observed events, to maximize its value among all distributions compatible with a given causal structure C. That is:
Definition 2. The dual inference problem
Given a causal structure C and a real function F of the distribution P over observed variables A1 , ..., Am , solve the
optimization problem
f ? = max F (P ),
s.t. P admits a realization in C,
(2)
Note that this definition does not coincide with the standard one in the literature of probabilistic graphical models.
There the dual problem consists in, given a causal structure C, identifying the set of all restrictions which affect any
distribution over observed variables compatible with C. Even for the simplest causal structures, the output of such
a problem can be too large to store in a normal computer. Thus in this paper we will focus on the above restricted
notion of dual.
Coming back to the triangle scenario, an instance of the dual problem (as defined in this paper) would be maximizing
−(P (000)−1/2)2 −(P (111)−1/2)2 over all distributions P (A, B, C) with A, B, C ∈ {0, 1}, realizable within the triangle
scenario.
There exist a number of variational algorithms to solve problem (2) (Koller and Friedman, 2009). Similarly, there
exist many heuristics which scan possible causal realizations of a given distribution of observed variables. However,
general practical tools to demonstrate the irrealizability of a probability distribution or to derive upper bounds on
the solution of the dual problem are scarce. One of them is the inflation technique, which we describe next.
III. A QUICK OVERVIEW OF THE INFLATION TECHNIQUE
Let P (A, B, C) be a distribution realizable in the triangle scenario, and suppose that we generate n independently
distributed copies of λ1 , λ2 , λ3 , that is, the variables {λ1i , λ2i , λ3i : i = 1, ..., n}. Then we could define the random
variables
Ai,j ≡ A(λ1i , λ2j ), Bi,j ≡ B(λ2i , λ3j ), Ci,j ≡ C(λ3i , λ1j ).
These variables follow a probability distribution Qn ({Ai,j }, {Bk,l }, {Cp,q }) with the property
(3)
4
FIG. 4 A second-order inflation of the triangle scenario.
Qn ({Aij = aij , Bkl = bkl , Cpq = cpq }) = Qn ({Aij = aπ(i)π0 (j) , Bkl = bπ0 (k)π00 (l) , Cpq = cπ00 (p)π(q) }),
(4)
for all permutations of n elements π, π 0 , π 00 . Moreover, the marginal distribution over the “diagonal variables”
{Aii , Bii , Cii }i also satisfies the identities
Qn ({Aii = ai , Bii = bi , Cii = Ci }i ) =
n
Y
P (ai , bi , ci ).
(5)
i=1
See Figure 4 for the resulting causal network when n = 2.
Given an arbitrary distribution P (A, B, C), the inflation technique consists in demanding the existence of a distribution Qn satisfying (4) and (5). Any such Qn will be called an nth order inflation of P . Clearly, if P (A, B, C)
does not admit a nth order inflation for some n, then it cannot be realizable in the triangle scenario. Deciding the
existence of an nth order inflation can be cast as a linear program (Alevras, 2001).
Notice that, for any distribution Qn satisfying (4), Qn ({Aπ(i)π0 (i) = ai , Bπ0 (i)π00 (i) = bi , Cπ00 (i)π(i) = ci }i ) =
Qn ({Aii = ai , Bii = bi , Cii = Ci }i ) holds for all permutations π, π 0 , π 00 . Therefore, any distribution Qn subject
to the constraints (4), (5) must be such that
Qn ({Aπ(i)π0 (i) = ai , Bπ0 (i)π00 (i) = bi , Cπ00 (i)π(i) = ci }i ) =
n
Y
i=1
P (ai , bi , ci ),
(6)
5
FIG. 5 The bilocality scenario.
for all permutations π, π 0 , π 00 of n elements. Actually, the original description of the inflation technique in (Wolfe
et al., 2016) imposes the constraints (6) rather than (4), (5) over the distribution Qn .
Demanding the existence of a distribution Qn satisfying condition (6) can be shown to enforce over P (A, B, C)
exactly the same constraints as demanding the existence of a distribution satisfying (4) and (5). Indeed, as noted in
(Wolfe et al., 2016), any distribution Qn satisfying (6) can be twirled or symmetrized (see below) to a distribution
Q̃n satisfying eqs. (4), (5). For convenience, from now on we will just refer to the formulation of the inflation
technique involving the symmetries (4). This formulation has the added advantage that the symmetry constraints
can be exploited to reduce the time and memory complexity of the corresponding linear program, see (Gent et al.,
2006).
The inflation technique is easy to generalize to relax the property of admitting an explanation in terms of arbitrary
causal networks (remember, though, that causal networks are just a particular type of causal structures). Namely,
one just adds to each observable variable as many subindices as latent variables it depends on, and makes the
total probability distribution invariant under independent permutations of the same type of indices. The symmetry
condition to impose is
m
1
1 L1
m
m Lm
Qn ({A1ī1 = a1 (ī1 ), ...., Am
(īm )) : ī1 , ..., īm }),
īm = a (īm ) : ī1 , ..., īm }) = Qn ({Aī1 = a (π̄ (ī1 )), ...., Aīm = a (π̄
(7)
for all vectors π̄ = (π 1 , ..., π L ) of L independent permutations (one for each latent variable or index type). Here īx
denotes the tuple of subindices on which variable Ax depends. In addition, one must enforce that Qn satisfies the
compatibility conditions
m
Qn ({A1i...i = a1i , ..., Am
i...i = ai }i ) =
n
Y
P (a1i , ..., am
i ).
(8)
i=1
For further elucidation, consider another causal network. In the bilocality scenario, Fig. 5, we again have three
random variables A, B, C which are defined, respectively, via the non-deterministic functions A(λ1 ), B(λ1 , λ2 ), C(λ2 ).
As before, we assume that the latent variables λ1 , λ2 are independently distributed.
In this scenario, an nth -order inflation corresponds to a distribution Qn over the variables Ai , Bjk , Cl , where i, j, k, l
range from 1 to n. Qn must satisfy the linear constraints:
Qn ({Ai = ai , Bjk = bjk , Cl = cl }) = Qn ({Ai = aπ(i) , Bjk = bπ(j)π0 (k) Cl = cπ0 (l) }),
(9)
for all permutations of n elements π, π 0 . It is also subject to the identities
Qn ({Ai = ai , Bii = bi , Ci = ci }i ) =
n
Y
P (ai , bi , ci ).
(10)
i=1
Note that the inflation technique also suggests a simple method to solve the dual problem, see Definition
2. Indeed, for any probability distribution q(X), let q ⊗k (X1 , ..., Xk ) represent the k-variate distribution gen1
m
erated by k independent samples of q(X), and let Qkn (a11 , ..., am
1 , ..., ak , ..., ak ) denote the marginal probability
6
Qn ({Axi,...,i = axi : x = 1, ..., m, i = 1, ..., k}). Now, let F be a polynomial of degree k on the probabilities of
measurable events, and consider the problem f ? ≡ maxP F (P ), where the maximization is understood over the set of
all distributions P realizable in the considered causal network. From (8), we can relax this problem to
f n = max F̃ (Qkn ),
s.t. Qn , satisfies (7),
(11)
where F̃ is any linear functional such that F̃ (q(X)⊗k ) = F (q). It is immediate that f n ≥ f ? for all n and that the
above can be cast as a linear program.
In fact, as shown in (Wolfe et al., 2016), the inflation technique partially solves as well the conventional dual
problem, where one asks which constraints any distribution compatible with the considered causal network is subject
to. This is achieved by deriving, via combinatorial tools, the facets or linear inequalities which define the set of
diagonal marginals of all distributions Qn satisfying eq. (9). When applied to a distribution of the form P ⊗n , each
of such linear inequalities translates to a polynomial inequality to be satisfied by any distribution P (A1 , ..., Am ) over
the observed variables admitting an nth -order inflation.
In the next two sections, we will
prove
that, conversely, any distribution P for the set of measurable events that
1
th
admits an n -order inflation is O √n -close in Euclidean norm to a distribution realizable in the considered causal
2
network. Similarly, we will show that f n = f ? + O kn .
IV. A GENERALIZATION OF THE DE FINETTI THEOREM
The purpose of this section is to prove the following result:
1
m
Theorem 3. Let Qn be any distribution satisfying the symmetry constraints (7), and call Qkn (a11 , ..., am
1 , ..., ak , ..., ak )
x
x
the marginal probability Qn ({Ai,...,i = ai : x = 1, ..., m, i = 1, ..., k}). Then, there exist normalized probability
P
distributions Qµ (a1 , ..., am ), achievable in the considered causal network, and probabilities pµ ≥ 0, µ pµ = 1 such
that
!
D
Qkn ,
X
pµ Q⊗k
µ
µ
≤O
Lk 2
n
,
(12)
where
D(q, r) denotes the statistical distance between the probability distributions q(X), r(X), i.e., D(q, r) =
P
|q(x)
− r(x)|.
x
Proof. We will prove the result for the triangle scenario; the generalization will be obvious. Given an arbitrary
distribution of 3n2 variables Qn ({Ai,j , Bk,l , Cp,q }), consider its symmetrization Q̃n , defined by
Q̃n ({Aij = aij , Bkl = bkl , Cpq = cpq }) =
1
n!3
X
Qn ({Aij = aπ(i)π0 (j) , Bkl = bπ0 (k)π00 (l) , Cpq = cπ00 (p)π(q) }).
π,π 0 ,π 00 ∈Sn
(13)
˜
Note that Q̃n = Q̃n . In addition, any distribution Qn satisfying the symmetry condition (4) fulfills Q̃n = Qn and any
symmetrized distribution satisfies (4).
Let δ({â(i, j), b̂(k, l), ĉ(p, q)}) be the deterministic distribution assigning the values â(i, j), b̂(k, l), ĉ(p, q) to the random variables Ai,j , Bk,l , Cp,q , for i, j, k, l, p, q ∈ {1, ..., n}. Since any distribution is a convex combination of deterministic points, it follows that any distribution satisfying eq. (4) can be expressed as a convex combination of symmetric
distributions of the form δ({â(i, j), ^
b̂(k, l), ĉ(p, q)}). For ease of notation, we will assume from now on that the values
{â(i, j), b̂(k, l), ĉ(p, q)} are fixed and denote the latter distribution simply by Q̃.
Call Q̃1 the marginal Q̃(A1,1 , B1,1 , C1,1 ). It can be verified, by symmetry, that it is given by the formula:
Q̃1 (a, b, c) =
n
1 X
δ(â(i, j), a)δ(b̂(j, k), b)δ(ĉ(k, i), c).
n3
i,j,k=1
(14)
7
Notice that Q̃1 (a, b, c) can be reproduced in the triangle scenario. Indeed, the latent variables are i, j, k, they can take
values in {1, ..., n} and are uniformly distributed.
Consider now the marginal distribution Q̃k ≡ Q̃(A1,1 , B1,1 , C1,1 , ..., Ak,k , Bk,k , Ck,k ). By symmetry, it is expressed
as:
Q̃k (a1 , b1 , c1 , ..., ak , bk , ck ) =
k
X Y
1
δ(â(ix , jx ), ax )δ(b̂(jx , kx ), bx )δ(ĉ(kx , ix ), cx ),
n3 (n − 1)3 ...(n − k + 1)3
x=1
(15)
hī,j̄,k̄i
where the sum is taken over all tuples ī, j̄, k̄ ∈ {1, ..., n}k such that ix = iy (jx = jy ) [kx = ky ] implies x = y. That
is, there are no repeated indices when we move from one block of variables (ax , bx , cx ) to another.
Now, compare Q̃k with (Q̃1 )⊗k . It is straightforward that
(Q̃1 )⊗k (a1 , b1 , c1 , ..., ak , bk , ck ) =
k
Y
Q̃1 (ax , bx , cx ) =
x=1
k
1 XY
δ(â(ix , jx ), ax )δ(b̂(jx , kx ), bx )δ(ĉ(kx , ix ), cx ), (16)
n3k
x=1
ī,j̄,k̄
where, this time, the sum contains all possible tuples ī, j̄, k̄ ∈ {1, ..., n}k . The statistical distance between the two
1
1
distributions is bounded by n3 (n−1)3 ...(n−k+1)
3 − n3k times the number of tuples with non-repeated indices (namely,
3
3
3
3k
n (n − 1) ...(n − k + 1) ), plus 1/n times the number of tuples with repeated indices (namely, n3k − n3 (n − 1)3 ...(n −
k + 1)3 ). The result is
n3 (n − 1)3 ...(n − k + 1)3
D Q̃k , (Q̃1 )⊗k ≤ 2 1 −
.
n3k
(17)
P
P
Finally, let Qn be any distribution satisfying eq. (4). Then, Qn = µ pµ Q̃µ , where pµ ≥ 0, µ pµ = 1 and Q̃µ is a
distribution of the form δ({â(i, j),^
b̂(i, j), ĉ(p, q)}) for every µ. By convexity of the statistical distance, we have that
!
D
Qkn ,
X
µ
pµ (Q̃1µ )⊗k
≤
X
µ
pµ D
Q̃kµ , (Q̃1µ )⊗k
n3 (n − 1)3 ...(n − k + 1)3
≤2 1−
n3k
=O
3k 2
n
.
(18)
Extending this result to general causal networks is straightforward, so we will just sketch the proof. First, the
action of the corresponding symmetrization over a deterministic distribution equals a distribution Q̃ whose 1marginal Q̃1 (a1 , ..., am ) is a uniform mixture over the tuple of indices ī of deterministic distributions of the form
Qm
x x Lx
1
k
x=1 δ(a , a (ī )). It thus follows that Q̃ is realizable within the causal network. The k-site marginal Q̃ is also a
uniform mixture of deterministic distributions of a similar type, but where no repeated indices are allowed between
the different blocks of variables. The statistical difference between Q̃k and (Q̃1 )⊗k is thus bounded by
nL (n − 1)L ...(n − k + 1)L
2 1−
nkL
=O
Lk 2
n
.
(19)
V. PROOF OF CONVERGENCE OF THE INFLATION TECHNIQUE
Now we are ready to prove our first main result. Let P be a probability distribution over the observed
P variables, and
suppose that P admits an nth -order inflation Qn . Define the second-degree polynomial N (R) = ā (R(ā) − P (ā))2 ,
and let Ñ be a linear functional such that Ñ (q ⊗2 ) = N (q) for all distributions q. Note that, due to conditions (8),
Ñ (Q2n ) = N (P ) = 0. Invoking the theorem in the previous section, we have that
8
!
0 = N (P )
=Ñ (Q2n )
= Ñ
X
pµ Q⊗2
µ
+O
µ
L
=
n
L
=
n
µ
X
L
pµ N (Qµ ) + O
,
n
µ
X
pµ Ñ (Q⊗2
µ )+O
(20)
where {Qµ }µ are
The above implies that, for some µ,
distributions realizable within the considered causal structure.
th
N (Qµ ) ≤ O L
.
In
other
words:
for
any
distribution
P
admitting
an
n
-order
inflation, there exists a realizable
n
q
L
distribution Q that is O
n -close to P in the Euclidean norm. Since the set of compatible distributions is closed
(Rosset et al., 2016), taking the limit n → ∞ it follows that any distribution P passing all inflation tests must be
realizable.
Finally, notice that the previous argument can also be used to prove the convergence of the sequence of linear
programs (11). In effect, let F be a polynomial of degree k, with f ? = maxP F (P ), and let f n be defined as in (11).
Then, by the extended de Finetti theorem, we have that
2
2
X
Lk
Lk
?
n
?
f ≤f =
pµ F (Qµ ) + O
≤f +O
.
(21)
n
n
µ
In certain practical cases, we may not know the full probability distribution of the observable variables, but only
the probabilities of a restricted set E of measurable events. To apply the inflation technique to those cases, rather
than fixing the value of all probability products, like in eq. (8), we would impose the constraint
n X
X
m
Qn ({A1i...i = a1i , ..., Am
i...i = ai }i ) =
i=1 ~
ai ∈ei
n
Y
P (ei ),
(22)
i=1
where e1 , ..., en ∈ E. Any distribution Qn satisfying both (7) and (22) will also be dubbed an nth order inflation of
the distribution of measurable events.
For example, consider again the triangle scenario (Fig. 2), and assume that our experimental setup just allows us
to detect events of the form e(a) ≡ {(A, B, C) : A = B = C = a}. Then our set of measurable events is E = ∪a {e(a)}
and the input of the causal inference problem is the distribution {P (e), e ∈ E}. An nth order inflation Qn of P (e)
would satisfy eq. (4) and the linear conditions
Qn ({Aii = ai , Bii = ai , Cii = ai }i ) =
n
Y
P (e(ai )).
(23)
i=1
The proofs P
of convergence presented above easily extend to this scenario. Indeed, choosing the polynomial N such
that N (R) = e∈E (P (e) − R(e))2 , and following the same derivation as in eq. (20), we conclude that a distribution
q
L
of measurable events admitting an nth order inflation is O
n -close in Euclidean norm to a realizable distribution.
Similarly, one can bound the speed convergence of the inflation technique when applied to maximize polynomials of
a probability distribution of measurable events.
VI. EXTENSION OF OUR RESULTS TO GENERAL CAUSAL STRUCTURES
So far we have been just referring to causal networks, i.e., those causal structures where all observed variables
just depend on a number of independent latent variables. However, in a general causal structure, the value of a
given variable can depend, not only on latent variables, but also on the values of other observed variables. Can the
inflation technique solve the inference and dual problems for these general structures as well? In the following, we
will show that any causal (dual) inference problem in an arbitrary causal structure can be mapped to an inference
9
(dual) problem with a non-trivial set of measurable events in an extended causal network. Hence we can solve it via
the inflation technique.
To go from general causal structures to causal networks, we will follow the following three steps.
FIG. 6 First step: re-wiring.
The first step is called re-wiring, see Figure 6. Note that, if in the original structure there is an edge from variable
X to the latent variable λ, if we delete the edge and replace it by edges from variable X to the direct successors of
λ we obtain a causal structure with the same predictive power. Indeed, in the first structure λ could carry a copy
of X to each of its direct successors: this implies that any probability distribution of observed variables realizable in
the re-wired causal structure can also be realized in the original causal structure. Conversely, suppose that in the
original causal structure λ depends on the variables X̄ (we group them all in a vector) as λ = λ(X̄, µ), where λ is a
deterministic function and µ is an internal random variable. Then one can simulate the same probability distribution
of observed variables in the re-wired causal structure by distributing µ to the direct successors of λ and making them
compute locally the value of λ, given X̄. Sequentially re-wiring the edges pointing to a latent variable, we end up
with an equivalent causal structure where the latent variables just have edges pointing from them and not to them.
FIG. 7 Second step: unpacking.
For the second step, unpacking (Figure 7), we will draw inspiration from the notion of interruption, described in
(Wolfe and Sainz, 2017). Let X be an observed variable depending on the latent variables λ̄ and the observed variables
Ȳ , and suppose that Ȳ can take d different values, e.g.: Ȳ ∈ {1, ..., d}. In this step, we will break the edges between Ȳ
and X and “unpack” variable X into the observed variables X(1), ..., X(d), defined via the expression X(s) ≡ X(s, λ̄).
Unpacking all the variables, we arrive at a causal structure with a few disconnected random variables Ā, the observable
parents of the causal graph, and a number of observed variables B̄ depending on just the latent variables, which by
the previous step are independent from each other. The probabilities of the observed variables in the original causal
structure can be obtained from the probabilities of a set of measurable events in this new causal structure via the
relation
A
B
P (Ā = ā, B̄ = b̄) = P ({Ā = ā} ∩ {B i (āIi , b̄Ii ) = bi }i ),
(24)
where IiA , IiB are, respectively, the superindices of the variables Ā, B̄ who are direct predecessors of B i in the original
causal structure.
10
FIG. 8 Third step: re-parenting.
In the last step, re-parenting (Figure 8), we will introduce a new latent variable λA for each observable parent
A of the original graph, and an edge from λA to A. At the end of this step, we have a layer of observed variables
depending only on the another layer of independent latent variables. This is none other but a causal network! The
original causal inference (dual) problem is hence mapped to an inference (dual) problem in a causal network with a
set E of measurable events given by
n
o
A
B
E = {Ā = ā} ∩ {B i (āIi , b̄Ii ) = bi }i
.
(25)
ā,b̄
When dealing with an inference problem with observable parents Ā, an alternative to the re-parenting step is simply
to erase the observed parents from the structure and consider the resulting causal network with constraints
A
B
P ({B i (āIi , b̄Ii ) = bi }i ) = P (B = b|Ā = ā).
(26)
Applying the inflation technique to this second network is computationally cheaper, since it has less observed variables.
Note though that this mapping does not allow us to optimize polynomials of P (Ā, B̄), but of P (B̄|Ā), so one cannot
use it to solve the causal dual problem.
For illustration, consider again the instrumental scenario, see Fig. 3. Suppose that the random variables A, B can
just take two possible values, 0 or 1. Then we can define the vector variable (B(0), B(1)) via the non-deterministic
function (B(0, λ), B(1, λ)). Similarly, we can unpack C into the vector (C(0), C(1)) ≡ (C(0, λ), C(1, λ)). The resulting causal network after re-wiring, unpacking and re-parenting is depicted on the right-hand side of Fig. 7.
Let P̂ (A, B(0), B(1), C(0), C(1)) be the distribution of the observed variables of the new network. Then, our set
of measurable events is {A = a, B(a) = b, C(b) = c}a,b,c , and the correspondence with the original distribution is
P (A = a, B = b, C = c) = P̂ (A = a, B(a) = b, C(b) = c).
Alternatively, we can erase variable A from the causal network and take the distribution P̃ (B(0), B(1), C(0), C(1))
of observed variables to be our fundamental object. Then, the set of measurable events is {B(a) = b, C(b) = c}a,b,c ,
and the corresponding probabilities are P (B = b, C = c|A = a) = P̃ (B(a) = b, C(b) = c).
VII. CONCLUSION
In this paper we have proven that the hierarchy of tests proposed in (Wolfe et al., 2016) to bound the set of
distributions admitting a representation in a given causal network is complete, in the sense that any distribution that
does not admit such a representation
will fail one of the tests. More quantitatively, we showed that any distribution
1
th
P passing the n test is O √n -close in Euclidean norm to a distribution Q realizable within the considered causal
network. We also proved that the nth linear programming relaxation provided by the inflation technique to the
solution of the dual problem is O n1 away from optimality.
To top it up, we showed that any causal inference (dual) problem in a general causal structure can be mapped to a
causal inference (dual) problem in an extended causal network with a non-trivial set of measurable events. This is a
regime for which we also proved the convergence of the inflation technique. Put together, these two results thus show
that the inflation technique is much more than a useful machinery to derive statistical limits. It is an alternative way
to define causal structures!
For future work, it would be interesting to adapt our results to the quantum case, for which the inflation technique can also be applied (Wolfe et al., 2016). That would require an extension of the finite quantum de Finetti
11
theorem (Koenig and Renner, 2005) that held for certain positive-semidefinite structures subject to certain symmetry
constraints. An arduous task!
Acknowledgements
This work was not supported by the European Research Council. This research was supported in part by Perimeter
Institute for Theoretical Physics. Research at Perimeter Institute is supported by the Government of Canada through
the Department of Innovation, Science and Economic Development Canada and by the Province of Ontario through
the Ministry of Research, Innovation and Science.
References
E. Wolfe, R. W. Spekkens, and T. Fritz, “The Inflation Technique for Causal Inference with Latent Variables,” arXiv:1609.00672
(2016).
D. Koller and N. Friedman, Probabilistic Graphical Models: Principles and Techniques - Adaptive Computation and Machine
Learning (The MIT Press, 2009).
T. Fritz, “Beyond Bell’s theorem: correlation scenarios,” New J. Phys. 14, 103001 (2012).
T. Fritz and R. Chaves, “Entropic Inequalities and Marginal Problems,” IEEE Trans. Info. Theo. 59, 803 (2013).
R. Chaves, L. Luft, and D. Gross, “Causal structures from entropic information: geometry and novel scenarios,” New J. Phys.
16, 043001 (2014).
R. Chaves, L. Luft, T. O. Maciel, D. Gross, D. Janzing, and B. Schölkopf, “Inferring latent structures via information
inequalities,” in Proc. of the 30th Conference on Uncertainty in Artificial Intelligence (AUAI, 2014) pp. 112–121.
R. Chaves, “Polynomial Bell inequalities,” Phys. Rev. Lett. 116, 010402 (2016).
J. Bohr Brask and R. Chaves, “Bell scenarios with communication,” J. Phys. A 50, 094001 (2017).
P. Diaconis and D. Freedman, “Finite exchangeable sequences,” Annals of Probability 8, 745764 (1980).
P. M. W. Alevras, Dimitris, Linear Optimization and Extensions (Springer Berlin Heidelberg, 2001).
I. P. Gent, K. E. Petrie, and J.-F. Puget, “Chapter 10 - Symmetry in Constraint Programming,” in Handbook of Constraint
Programming, Foundations of Artificial Intelligence, Vol. 2, edited by F. Rossi, P. van Beek, and T. Walsh (Elsevier, 2006)
pp. 329 – 376.
D. Rosset, N. Gisin, and E. Wolfe, “All n-local set are semialgebraic,” unpublished (2016).
E. Wolfe and A. B. Sainz, “Interruption Technique for Causal Inference and the Quantum Instrumental Scenario,” unpublished
(2017).
R. Koenig and R. Renner, “A de finetti representation for finite symmetric quantum states,” J. Math. Phys. 46, 122108 (2005).
| 10math.ST
|
June 2015 (revised September 2015)
Report LIDS - 3176
Incremental Aggregated Proximal and Augmented
Lagrangian Algorithms
arXiv:1509.09257v2 [cs.SY] 4 Nov 2015
Dimitri P. Bertsekas†
Abstract
We consider minimization of the sum of a large number of convex functions, and we propose an incremental aggregated version of the proximal algorithm, which bears similarity to the incremental aggregated
gradient and subgradient methods that have received a lot of recent attention. Under cost function differentiability and strong convexity assumptions, we show linear convergence for a sufficiently small constant
stepsize. This result also applies to distributed asynchronous variants of the method, involving bounded
interprocessor communication delays.
We then consider dual versions of incremental proximal algorithms, which are incremental augmented
Lagrangian methods for separable equality-constrained optimization problems. Contrary to the standard
augmented Lagrangian method, these methods admit decomposition in the minimization of the augmented
Lagrangian, and update the multipliers far more frequently. Our incremental aggregated augmented Lagrangian methods bear similarity to several known decomposition algorithms, most of which, however, are
not incremental in nature: the augmented Lagrangian decomposition algorithm of Stephanopoulos and Westerberg [StW75], and the related methods of Tadjewski [Tad89] and Ruszczynski [Rus95], and the alternating
direction method of multipliers (ADMM) and more recent variations. We compare these methods in terms
of their properties, and highlight their potential advantages and limitations.
We also address the solution of separable inequality-constrained optimization problems through the
use of nonquadratic augmented Lagrangiias such as the exponential, and we dually consider a corresponding
incremental aggregated version of the proximal algorithm that uses nonquadratic regularization, such as an
entropy function. We finally propose a closely related linearly convergent method for minimization of large
differentiable sums subject to an orthant constraint, which may be viewed as an incremental aggregated
version of the mirror descent method.
1.
INCREMENTAL GRADIENT, SUBGRADIENT, AND PROXIMAL METHODS
We consider optimization problems with a cost function that consists of additive components:
minimize
def
F (x) =
m
X
i=1
subject to x ∈ X,
fi (x)
(1.1)
† Dimitri Bertsekas is with the Dept. of Electr. Engineering and Comp. Science, and the Laboratory for Infor-
mation and Decision Systems, M.I.T., Cambridge, Mass., 02139.
1
where fi : ℜn 7→ ℜ, i = 1, . . . , m, are convex real-valued functions, and X is a closed convex set. We focus
on the case where the number of components m is very large, and there is an incentive to use incremental
methods that operate on a single component fi at each iteration, rather than on the entire cost function F .
Problems of this type arise often in various practical contexts and have received a lot of attention recently.
Suitable algorithms include the incremental subgradient method (abbreviated IS), where a cost com˜ i (xk ) of fi is used in place of a full
ponent fik is selected at iteration k, and an arbitrary subgradient ∇f
k
k
subgradient of F at xk :†
˜ i (xk ) ,
xk+1 = PX xk − αk ∇f
k
(1.2)
where αk is a positive stepsize, and PX (·) denotes projection on X. It is important here that all components
are taken up for iteration with equal long-term frequency, using either a cyclic or a random selection scheme.
Methods of this type and their properties have been studied for a long time, and the relevant literature,
beginning in the 60’s, is too voluminous to list here. The author’s survey [Ber10] discusses the history of this
algorithm, its convergence properties, and its connections with stochastic approximation methods. Generally,
a diminishing stepsize αk is needed for convergence, even when the components fi are differentiable. Moreover
the convergence rate properties are generally better when the index ik is selected by randomization over the
set {1, . . . , m} than by a deterministic cyclic rule, as first shown by Nedić and Bertsekas [NeB01]; see also
[BNO03].
Another method, introduced by the author in [Ber10] and further studied in [Ber11], [Ber12], is the
incremental proximal method (abbreviated IP),
1
kx − xk k2 .
xk+1 ∈ arg min fik (x) +
x∈X
2αk
(1.3)
This method relates to the proximal algorithm (Martinet [Mar70], Rockafellar [Roc76a]) in the same way
that the IS method (1.2) relates to the classical nonincremental subgradient method. Similar to the IS
method, it is important that all components are taken up for iteration with equal long-term frequency. The
theoretical convergence properties of the IS and IP algorithms are similar, but it is generally believed that
IP is more robust, a property inherited from its nonincremental counterpart.
It turns out that the structures of the IS and IP methods (1.2) and (1.3) are quite similar. An important
fact in this regard is that the IP method (1.3) can be equivalently written as
˜ i (xk+1 ) ,
xk+1 = PX xk − αk ∇f
k
(1.4)
† Throughout the paper, we will operate within the n-dimensional space ℜn with the standard Euclidean norm,
denoted k · k. All vectors are considered column vectors and a prime denotes transposition, so x′ x = kxk2 . The scalar
coordinates of an optimization vector such as x are denoted by superscripts, x = (x1 , . . . , xn ), while sequences of
˜ (x) to denote a subgradient of a convex function f at a vector x ∈ ℜn ,
iterates are indexed by subscripts. We use ∇f
˜ (x)′ (z − x) for all z ∈ ℜn . The choice of ∇f
˜ (x) from within the set of all
i.e, a vector such that f (z) ≥ f (x) + ∇f
˜ (x) is the gradient ∇f (x).
subgradients at x will be clear from the context. If f is differentiable at x, ∇f
2
˜ i (xk+1 ) is a special subgradient of fi at the new point xk+1 (see Bertsekas [Ber10], Prop. 2.1,
where ∇f
k
k
[Ber11], Prop. 1, or [Ber15], Prop. 6.4.1). This special subgradient is determined from the optimality
conditions for the proximal maximization (1.3). For example if X = ℜn , we have
˜ i (xk+1 ) = xk − xk+1 ,
∇f
k
αk
˜ i (xk+1 ) may be a difficult
which is consistent with Eq. (1.4). Thus determining the special subgradient ∇f
k
problem, and in most cases it is preferable to implement the iteration in the proximal form (1.3) rather than
the projected form (1.4). However, the equivalent form of the IP iteration (1.4), when compared with the
IS iteration (1.2), suggests the close connection between the IS and IP iterations. In fact this connection
is the basis for a combination of the two methods to provide flexibility for the case where some of the cost
components fi are well suited for the proximal minimization of Eq. (1.3), while others are not; see [Ber10],
[Ber11], [Ber12].
Incremental Aggregated Gradient and Subgradient Methods
Incremental aggregated methods aim to provide a better approximation of a subgradient of the entire cost
function F , while preserving the economies accrued from computing a single component subgradient at each
iteration. In particular, the aggregated subgradient method (abbreviated IAS), has the form
!
m
X
˜ i (xℓ ) ,
∇f
xk+1 = PX xk − αk
i
(1.5)
i=1
˜ i (xℓ ) is a “delayed” subgradient of fi at some earlier iterate xℓ . We assume that the indexes ℓi
where ∇f
i
i
satisfy
k − b ≤ ℓi ≤ k,
∀ i, k,
(1.6)
where b is a fixed nonnegative integer. Thus the algorithm uses outdated subgradients from previous iterations for the components fi , i 6= ik , and need not compute a subgradient of these components at iteration
k.
The IAS method was first proposed, to our knowledge, by Nedić, Bertsekas, and Borkar [NBB01]. It
was motivated primarily by distributed asynchronous solution of dual separable problems, similar to the ones
to be discussed in Section 2 (in a distributed asynchronous context, it is natural to assume that subgradients
are used with some delays). A convergence result was shown in [NBB01] assuming that the stepsize sequence
{ak } is diminishing, and satisfies the standard conditions
∞
X
k=0
∞
X
αk = ∞,
k=0
α2k < ∞.
(1.7)
This result covers the case of iteration (1.5) for the case X = ℜn ; the more general case where X 6= ℜn
admits a similar analysis. We note that distributed algorithms that involve bounded delays in the iterates
3
have a long history, and are common in various distributed asynchronous computation contexts, including
gradient-like and coordinate descent methods; see [Ber89], Sections 7.5-7.8.
˜ i (xℓ ),
Note a limitation of this iteration over the IS iteration: one has to store the past subgradients ∇f
i
i 6= ik . Moreover, whatever effect the use of previously computed subgradients has, it will not be fully
manifested until a subgradient of each component has been computed; this is significant when the number of
components m is large. We note also that there are other approaches for approximating a full subgradient
of the cost function, which aim at computational economies, such as ǫ-subgradient methods (see Nedić and
Bertsekas [NeB10] and the references quoted there), and surrogate subgradient methods (see Bragin et. al
[BLY15] and the references quoted there).
The IAS method (1.5) contains as a special case the incremental aggregated gradient method (abbreviated IAG) for the case where the components fi are differentiable:
xk+1 = xk − αk
m
X
i=1
∇fi (xℓi ),
(1.8)
where ℓi ∈ [k − b, k] for all i and k. This method has attracted considerable attention thanks to a particularly
interesting convergence result. For the favorable case where the component gradients ∇fi are Lipschitz
continuous and F is strongly convex, it has been shown that the IAG method is linearly convergent to the
solution with a sufficiently small but constant stepsize αk ≡ α. This result was first given by Blatt, Hero,
and Gauchman [BHG08], for the case where the cost components fi are quadratic and the delayed indexes
ℓi satisfy certain restrictions that are consistent with a cyclic selection of components for iteration (see also
[AFB06]). The linear convergence result has been subsequently extended for nonquadratic problems and
for various forms of the method by several other authors, including Schmidt, Le Roux, and Bach [SLB13],
Mairal [Mai13], [Mai14], and Defazio, Caetano, and Domke [DCD14]. Several schemes have been proposed
˜ i (xℓ ), i 6= ik . Moreover, several
to address the limitation of having to store the past subgradients ∇f
i
experimental studies have confirmed the theoretical convergence rate advantage of the IAG method over the
corresponding incremental gradient method under the preceding favorable conditions. The use of arbitrary
indexes ℓi ∈ [k − b, k] in the IAG method was introduced in the paper by Gurbuzbalaban, Ozdaglar, and
Parillo [GOP15], who gave an elegant and particularly simple linear convergence analysis.
Incremental Aggregated Proximal Algorithm
In this paper, we consider an incremental aggregated proximal algorithm (abbreviated IAP), which has the
form
xk+1 ∈ arg min
x∈X
fik (x) +
X
i6=ik
1
2
˜ i (xℓ )′ (x − xk ) +
,
kx
−
x
k
∇f
k
i
2αk
(1.9)
˜ i (xℓ ) is a “delayed” subgradient of fi at some earlier iterate xℓ . We assume that the indexes ℓi
where ∇f
i
i
4
satisfy the boundedness condition ℓi ∈ [k − b, k], cf. Eq. (1.6). Intuitively, the idea is that the term
X
i6=ik
˜ i (xℓ )′ (x − xk )
∇f
i
in the proximal minimization (1.9) is a linear approximation to the term
X
fi (x)
i6=ik
[minus the constant
P
i6=ik
fi (xk )], which would be used in the standard proximal algorithm
xk+1
1
2
kx − xk k .
∈ arg min F (x) +
x∈X
2αk
It is straightforward to verify the following equivalent form of the IAP iteration (1.9):
1
xk+1 ∈ arg min fik (x) +
kx − zk k2 ,
x∈X
2αk
(1.10)
(1.11)
where
zk = xk − αk
X
i6=ik
˜ i (xℓ ).
∇f
i
(1.12)
In this form the algorithm is executed as a two-step process: first use xk and preceding subgradients to
compute zk via Eq. (1.12), and then execute an IP iteration starting from zk . Note a limitation of this
iteration over the IP iteration, which is shared with other incremental aggregated methods: to keep updating
˜ i (xℓ ), i 6= ik .
the vector zk , one has to store the past subgradients ∇f
i
Similar to the IP iteration (1.4), the IAP iteration (1.9) and its equivalent form (1.11)-(1.12) can be
written as
˜ i (xk+1 ) ,
xk+1 = PX zk − αk ∇f
k
(1.13)
˜ i (xk+1 ), which can be used in
so when executing the iteration, we typically can obtain the subgradient ∇f
k
subsequent IAP iterations. For example, in the unconstrained case where X = ℜn , from Eq. (1.13), we see
that
˜ i (xk+1 ) = zk − xk+1 .
∇f
k
αk
It is possible to prove various convergence results for the IAP iteration (1.9), or its equivalent forms
(1.11)-(1.12) and (1.12)-(1.13), for the case where the stepsize αk is diminishing and satisfies the standard
conditions (1.7). These results are in line with similar results for the IP method, given in [Ber10], [Ber11],
and for the IAS method (1.5), given in [NBB01]. Since the difference between the IAP and IAS methods
˜ i (xk+1 ) in IAP in place of ∇f
˜ i (xℓ ) in IAS, intuitively, for a diminishing stepsize, the
is the use of ∇f
k
k
i
k
asymptotic performance of the two methods should be similar, and indeed the convergence proofs for the
two methods are fairly similar, under comparable assumptions. We will thus not go into this convergence
analysis.
5
Incremental Aggregated Proximal Algorithm for Unconstrained Problems
In the unconstrained case where X = ℜn and the component functions fi are differentiable, the IAP iteration
(1.13) can be written as
xk+1 = xk − αk ∇fik (xk+1 ) +
X
i6=ik
∇fi (xℓi ) .
(1.14)
In this case, one may expect similar convergence behavior for the IAP and IAG methods, under favorable
conditions which allow the use of a constant stepsize αk ≡ α. In particular, we prove the following for the
IAP method.
Proposition 1.1:
Assume that X = ℜn and that the functions fi are convex and differentiable,
and satisfy
∇fi (x) − ∇fi (z) ≤ Li kx − zk,
for some constants Li . Assume further that the function F =
∀ x, z ∈ ℜn ,
Pm
i=1
fi is strongly convex with unique
minimum denoted x∗ . Then there exists α > 0 such that for all α ∈ (0, α], the sequence {xk } generated
by the IAP iteration (1.14) with constant stepsize αk ≡ α converges to x∗ linearly, in the sense that
kxk − x∗ k ≤ γρk for some scalars γ > 0 and ρ ∈ (0, 1), and all k.
The proof, given in Section 3, follows closely the one of [GOP15] for the IAG iteration, and relies on the
similarity of the iterations (1.14) and (1.8) [the use of the term ∇fik (xk+1 ) in place of the term ∇fik (xℓi )].
k
A key idea is to view the IAP iteration (1.14) as a gradient method with errors in the calculation of the
gradient, i.e.,
where ∇F (xk ) =
Pm
i=1
∇fi (xk ), and
xk+1 = xk − αk ∇F (xk ) + ek ,
ek = ∇fik (xk+1 ) − ∇fik (xk ) +
X
i6=ik
∇fi (xℓi ) − ∇fi (xk ) ,
(1.15)
(1.16)
and then to appropriately bound the size of the errors ek . This is similar to known lines of convergence
proofs for gradient and subgradient methods with errors. The proof of Section 3 applies also to a diagonally
scaled version of IAP, where a separate but constant stepsize is used for each coordinate.
We note that the line of proof of Prop. 1.1 does not readily extend to the constrained case when X 6= ℜn ,
nor is it clear whether and under what conditions linear convergence can be proved. In Section 4, however,
we will consider an incremental aggregated proximal algorithm that uses a nonquadratic regularization term
and seems to cope better with the case of nonnegativity constraints, i.e., X = {x | x ≥ 0}.
6
We finally return to the similarity of the IAP method (1.9) with the IAS method (1.5), and note that
the two methods admit similar distributed asynchronous implementations, which was described in the paper
[NBB01]. In this context, we have a central processor executes the proximal iteration (1.9) for some selected
component fik , while other processors compute subgradients for other components fi at points xℓi , which are
supplied by the central processor. These subgradients involve a “delay” that may be unpredictable, hence
the asynchronous character of the computation.
Local Versions of Proximal Algorithms
While the analysis of this paper requires that fi and X are convex, there is a straightforward way to extend
our incremental proximal methods to nonconvex problems involving twice differentiable functions, which we
will describe briefly. The idea is to use a local version of the proximal algorithm, proposed in the author’s
paper [Ber79] and based on a local version of the Fenchel duality framework given in [Ber78]. The algorithm
applies to the problem
minimize
f (x)
(1.17)
subject to g(x) = 0,
where f : ℜn 7→ ℜ and g : ℜn 7→ ℜr are twice continuously differentiable functions, such that f is “locally
convex” over the set x | g(x) = 0 (this is defined in terms of assumptions that relate to second order
sufficiency conditions of nonlinear programming; see [Ber78], [Ber79]). The local proximal algorithm has the
form
xk+1 ∈ arg min
g(x)=0
1
2
kx − xk k ,
f (x) +
2αk
(1.18)
where αk is sufficiently small to ensure that the function minimized in Eq. (1.18) is convex over ℜn [not just
locally over the set x | g(x) = 0 ]. A Newton-like version of this algorithm was also given in [Ber79].
There is an incremental version of the local proximal iteration (1.18) for problems involving sums of
functions. In particular, consider the problem
m
X
minimize
fi (x)
i=1
(1.19)
subject to g(x) = 0,
where fi : ℜn 7→ ℜ and g : ℜn 7→ ℜr are twice continuously differentiable functions, such that each fi is
“locally convex” over the set x | g(x) = 0 , for all i. This incremental local proximal iteration is
1
fik (x) +
kx − xk k2 ,
2αk
g(x)=0
xk+1 ∈ arg min
(1.20)
where ik is the index of the cost component that is iterated on. One may also consider an aggregated form
of this incremental iteration. The convergence properties of these algorithms are an interesting subject for
investigation, which lies, however, outside the scope of the present paper.
7
There is also another way to combine local proximal and incremental ideas for the case of the (nonconvex) separable problem in the vector x = (x1 , . . . , xm ),
minimize
m
X
def
f (x) =
fi (xi )
i=1
m
X
def
subject to g(x) =
(1.21)
gi
(xi )
= 0,
i=1
where fi : ℜni 7→ ℜ and gi : ℜni 7→ ℜr are twice continuously differentiable functions, and are such that
the problem admits a solution-Lagrange multiplier pair (x∗ , λ∗ ) satisfying standard second order sufficiency
conditions. In this approach, also developed in [Ber78], [Ber79], the problem (1.21) is converted to the
equivalent problem
minimize
def
φγ (z) =
min
g(x)=0
1
f (x) +
kx − zk2
2γ
subject to z ∈ ℜn1 +···+nm ,
where γ is sufficiently small so that for fixed z, f (x) +
1
2γ kx
suitably small neighborhood of x∗ , i.e., γ should be such that
(1.22)
− zk2 is convex in x locally, for all x in a
1
γI
+ ∇2 f (x∗ ) is positive definite. Since the
minimization problem (1.22), which defines φγ (z), is separable of the form
m
X
1 i
i
2
i
kx − z k
fi (x ) +
2γ
i=1
minimize
subject to x ∈ ℜn1 +···+nm ,
m
X
(1.23)
gi (xi ) = 0,
i=1
and locally convex in x, for fixed z and suitably small values of γ, it can be solved using the augmented
Lagrangian-based methods of the next section. Denoting x(z, γ) the optimal solution of this problem for
given z and γ, it is shown in [Ber79] (Prop. 2.1) (see also [Ber78], Prop. 2) that φγ is differentiable and
∇φγ (z) =
1
z − x(z, γ) .
γ
Thus the gradient algorithm
zk+1 = zk − γ∇φγ (zk ),
(1.24)
can be written as zk+1 = x(zk , γ) or equivalently, using Eqs. (1.22) and (1.23), in the (local) proximal form
xk+1
(m
)
X
1
fi (xi ) +
∈ arg Pm min
.
kxi − xik k2
2γ
gi (xi )=0 i=1
(1.25)
i=1
Note that the above minimization is amenable to decomposition, including solution using the incremental
aggregated augmented Lagrangian and ADMM methods of the next section, assuming γ is sufficiently small to
induce the required amount of convexification to make problem (1.25) convex (locally within a neighborhood
of x∗ ).
8
The convergence properties of this algorithm are developed in [Ber79], based on a local theory of
conjugate functions and Fenchel duality developed in [Ber78]. We refer to these papers for a discussion of
the local aspects of the minimization (1.25), as well as for the implementation of the Newton iteration
−1
zk+1 = zk − ∇2 φγ (zk )
∇φγ (zk ),
(1.26)
in analogy with the gradient method (1.24). A further analysis is again outside the scope of the present
paper, and is an interesting subject for investigation.
2.
INCREMENTAL AUGMENTED LAGRANGIAN METHODS
A second objective of this paper is to consider the application of the IP and IAP methods in a dual setting,
where they take the form of incremental augmented Lagrangian algorithms for the separable constrained
optimization problem
minimize
m
X
hi (y i )
i=1
subject to y i ∈ Yi , i = 1, . . . , m,
m
X
i=1
(2.1)
(Ai y i − bi ) = 0,
as shown in [Ber15], Section 6.4.3. Here hi : ℜni 7→ ℜ are convex functions (ni is a positive integer, which
may depend on i), Yi are nonempty closed convex subsets of ℜni , Ai are given r × ni matrices, and bi ∈ ℜr
are given vectors. The optimization vector is y = (y 1 , . . . , y m ), and our objective is to consider algorithms
that allow decomposition in the minimization of the augmented Lagrangian, so that m separate augmented
Lagrangian minimizations are performed, each with respect to a single component y i . Note that the problem
Pm
(2.1) is unaffected by redefinition of the scalars bi , as long as i=1 bi is not changed. It may be beneficial
to adjust the scalars bi so that the residuals Ai y i − bi are small near the optimal, and this may in fact be
attempted in the course of some algorithms as a form of heuristic.
Following a standard analysis, the dual function for problem (2.1) is given by
)
(m
X
hi (y i ) + λ′ (Ai y i − bi ) ,
Q(λ) =
inf
y i ∈Yi , i=1,...,m
(2.2)
i=1
where λ ∈ ℜr is the dual vector. By decomposing the minimization over the components y i , Q can be
expressed in the additive form
Q(λ) =
m
X
qi (λ),
i=1
where qi is the concave function
qi (λ) = inf
y i ∈Y
i
hi (y i ) + λ′ (Ai y i − bi ) ,
9
i = 1, . . . , m.
(2.3)
Dual Gradient-Like Methods for Separable Problems
Assuming that the dual function components qi are real-valued (which is true for example if Yi is compact),
the dual function Q(λ) can be minimized with the classical subgradient method.† This method takes the
form
λk+1 = λk + αk
m
X
i=1
˜ i (λk ),
∇q
(2.4)
˜ i (λk ) are obtained as
where αk > 0 is the stepsize and the subgradients ∇q
˜ i (λk ) = Ai y i − bi ,
∇q
k+1
i = 1, . . . , m,
with all components y i updated according to
i
yk+1
∈ arg min hi (y i ) + λ′k (Ai y i − bi ) ,
y i ∈Yi
i = 1, . . . , m.
The additive form of the dual function Q makes it suitable for application of incremental methods,
including the IAS method described in Section 1, which in fact was proposed in [NBB01] with the separable
problem (2.1) in mind. In the case where the components qi are differentiable [which is true if the infimum
in the definition (2.3) is attained uniquely for all λ], one may also use the IAG method with a constant
but sufficiently small stepsize. This is an incremental aggregated version of a classical dual gradient method
proposed in the 60s and often attributed to Everett [Eve63]. It takes the form
λk+1 = λk + α ∇qik (λk ) +
X
i6=ik
∇qi (λℓi ) ;
(2.5)
cf. Eq. (1.8). The gradient of the dual function component qi is given by
∇qi (λ) = Ai y i (λ) − bi ,
where y i (λ) is the minimizer over Yi of
fi (y i ) + λ′ Ai y i ,
which is assumed to be unique for differentiability of qi . By streamlining the computations using the preceding
relations, we see that the iteration has the following form.
† In the case where qi is not real-valued, the dual function can be maximized over the set ∩m
i=1 Λi , where Λi =
λ | qi (λ > −∞ . This can be done by using incremental constraint projection methods involving projection or
proximal maximization over a single set Λi at a time. Methods of this type have been proposed in [Ber11], [Ned11],
[WaB13], [WaB15], but their discussion is beyond the scope of the present paper.
10
Incremental Aggregated Dual Gradient Iteration (IADG)
Select a component index ik , and update the single component y ik according to
i
k
yk+1
∈ arg min
y ik ∈Yi
k
hik (y ik ) + λ′k Aik y ik ,
(2.6)
i
while keeping the others unchanged, yk+1
= yki for all i 6= ik . Then update λ according to
λk+1 = λk +
ik
α Aik yk+1
+
X
i6=ik
Ai yℓii − b .
(2.7)
The convergence properties of the method are governed by the known results for the IAG method,
which were noted in Section 1. In particular, we obtain linear convergence with a constant sufficiently small
stepsize α, assuming Lipschitz continuity of ∇qi and strong convexity of Q, and that the long-term frequency
of updating y i is the same for all i. Note, however, that this linear convergence result cannot be used when
the primal problem (2.1) has additional convex inequality constraints, because then the corresponding dual
problem involves nonnegativity constraints.
Augmented Lagrangian-Based Algorithms for Separable Problems
The nonincremental and incremental subgradient and gradient methods just described are convenient for the
purposes of decomposition, but their convergence properties tend to be fragile. On the other hand, the more
stable augmented Lagrangian methods have a major drawback: when a quadratic penalty term is added to
the Lagrangian function, the resulting augmented Lagrangian
m
X
i=1
m
αk X
hi (y i ) + λ′ (Ai y i − bi ) +
(Ai y i − bi )
2 i=1
2
is not separable any more, and is not amenable to minimization by decomposition. This is a well-known
limitation of the augmented Lagrangian approach that has been addressed by a number of authors with
various algorithmic proposals, which we will now survey.
The first proposal of this type was the paper by Stephanopoulos and Westerberg [StW75], which was
based on enforced decomposition: minimizing the augmented Lagrangian separately with respect to each
component vector y i , while holding the other components fixed at some estimated values. Minimization over
the components y i is followed by a multiplier update (using the standard augmented Lagrangian formula).
11
The decomposition method of [StW75] attracted considerable attention and motivated further research,
including the similarly structured methods by Tadjewski [Tad89] and by Ruszczynski [Rus95], which include
convergence analyses and give references to earlier works. Our incremental aggregated proximal algorithm
bears similarity with the methods of [StW75], [Tad89], and [Rus95]. We note, however, that the methods
of [StW75] and [Tad89] were motivated by nonconvex separable problems for which there is a duality gap,
while our analysis requires a convex programming structure, where there is no duality gap. The method of
[Rus95] is applied to convex separable problems, including linear programming.
Another method for convex separable problems that uses augmented Lagrangian minimizations is
given by Deng, Lai, Peng, and Yin [DLP14], who give several related references, including the paper by
Chen and Teboulle [ChT94]. The method is based on the use of primal proximal terms in the augmented
Lagrangian (in addition to the quadratic penalty term). This is in the spirit of Rockafellar’s proximal
method of multipliers [Roc76b], and involves two separate penalty parameters, which for convergence should
satisfy certain restrictions. The papers by Hong and Luo [HoL13], and Robinson and Tappenden [RoT15]
also propose algorithms that use primal proximal terms and two penalty parameters, but differ from the
algorithm of [DLP14] in that they update the primal variables in Gauss-Seidel rather than Jacobi fashion,
while requiring additional assumptions (see also Dang and Lan [DaL15] for a related algorithm). GaussSeidel updating is somewhat similar to the incremental mode of iteration of this paper, and based on the
results of experiments in [WHM13] and [RoT15], it appears to be beneficial.
A different possibility to deal with nonconvex separable problems is based on the convexification provided by the local proximal algorithm that was discussed at the end of the preceding section. Its application
to nonconvex separable problems is described in [Ber79]; see also Tanikawa and Mukai [TaM85], who proposed
a method that aims at improved efficiency relative to the approach of [Ber79]. A discussion of additional
proposals of decomposition methods that use augmented Lagrangians is given in the recent paper by Hamdi
and Mishra [HaM11].
Still another approach that has been used to exploit the structure of the separable problem (2.1) is
the alternating direction method of multipliers (ADMM), a popular method for convex programming, first
proposed by Glowinskii and Morocco [GIM75], and Gabay and Mercier [GaM76], and further developed by
Gabay [Gab79], [Gab83]. This method applies to the problem
minimize
f1 (x) + f2 (z)
(2.8)
subject to x ∈ ℜn , z ∈ ℜm , Ax = z,
where f1 : ℜn 7→ (−∞, ∞] and f2 : ℜm 7→ (−∞, ∞] are closed convex functions, and A is a given m × n
matrix. The method is better suited than the augmented Lagrangian method for exploiting special structures,
including separability, and is capable of decoupling the vectors x and z in the augmented Lagrangian
f1 (x) + f2 (z) + λ′ (Ax − z) +
12
α
kAx − zk2 .
2
For a discussion of the properties and the many applications of the method, we refer to its extensive literature,
including the books [BeT89], Section 3.4.4, [Ber15], Section 5.4, and [BPC11], which give many references.
The form of the ADMM for separable problems to overcome the coupling of variables in the augmented
Lagrangian minimization was first derived in Bertsekas and Tsitsiklis [BeT89], Section 3.4, pp. 249-254 (see
also [Ber15], Section 5.4.2). We will describe the form of this specialized ADMM later in this section.
We will now consider the incremental proximal methods IP [cf. Eq. (1.3)] and IAP [cf. Eq. (1.9)] for
Pm
maximizing the dual function i=1 qi (λ). Taking into account the concavity of the components qi , the IP
method takes the form
1
λk+1 ∈ arg maxr qik (λ) −
kλ − λk k2 ,
λ∈ℜ
2αk
(2.9)
where ik is the index of the component chosen for iteration and αk is a positive parameter. This method was
given in [Ber15], Section 6.4.3, where it was shown that it can be implemented through the use of decoupled
augmented Lagrangian minimizations, each involving a single component vector y i . The IAP method takes
the form
λk+1 ∈ arg maxr
λ∈ℜ
qik (λ) +
X
i6=ik
˜ i (λℓ )′ (λ − λk ) − 1 kλ − λk k2 ,
∇q
i
2αk
(2.10)
and has not been considered earlier within the dual separable constrained optimization context of this section.
The convergence results noted in Section 1 apply to this method. In particular, by Prop. 1.1, the IAP method
(2.10) is convergent with a sufficiently small constant stepsize, assuming that each qi is differentiable with
Lipschitz continuous gradient and Q is strongly concave. Of course, the differentiability of qi is a restrictive
assumption, and it amounts to attainment of the minimum at a unique point y i ∈ Yi in the definition (2.3)
of qi (λ) for all λ ∈ ℜr .
We will now describe how the incremental proximal methods IP and IAP can be implemented in terms
of augmented Lagrangian minimizations, which decompose with respect to components y i and have an
incremental character. To this end, we will review the well-known Fenchel duality relation between proximal
and augmented Lagrangian iterations, given first by Rockafellar [Roc73], [Roc76b], and subsequently in many
sources, including the author’s monograph and textbook accounts [Ber82], Chapter 5, and [Ber15], Section
5.2.1.
Duality Between Proximal and Augmented Lagrangian Iterations
Given a proper convex function P : ℜr 7→ (−∞, ∞], let Q : ℜr 7→ [−∞, ∞) be the closed proper concave
function defined by†
Q(λ) = inf r P (u) + λ′ u .
u∈ℜ
(2.11)
† Here and later, for concave functions Q, we use terminology used for convex functions as applied to −Q.
13
This is a conjugacy relation, since Q(λ) = −P ⋆ (−λ), where P ⋆ is the conjugate convex function of P .
Moreover, if P is closed, it can be recovered from Q using the conjugacy theorem,
P (u) = P ⋆⋆ (u) = sup
λ∈ℜr
λ′ u + Q(−λ) ,
(2.12)
where P ⋆⋆ is the conjugate convex function of P ⋆ (see, e.g., [Ber09], Prop. 1.6.1).
A key fact, assuming that P is closed, is that the proximal iteration
λk+1
1
2
kλ − λk k ,
∈ arg maxr Q(λ) −
λ∈ℜ
2αk
(2.13)
can be equivalently implemented in two steps as
n
o
αk
uk+1 ∈ arg minr P (u) + λ′k u +
kuk2 ,
u∈ℜ
2
(2.14)
λk+1 = λk + αk uk+1 ;
(2.15)
followed by
see, e.g., [Ber15], Section 5.2.1. Moreover, uk+1 is a subgradient of Q at λk+1 :
˜
uk+1 = ∇Q(λ
k+1 ).
(2.16)
These relations are shown by straightforward application of the Fenchel duality theorem to the maximization
of Eq. (2.13), which involves the sum of the concave functions Q and −(1/2αk )kλ − λk k2 . The closedness of
P is used both to ensure that the duality relation (2.12) holds, and to guarantee that the minimum in Eq.
(2.14) is attained. Note that Eq. (2.14) has the form of an augmented Lagrangian minimization relating to
the (somewhat contrived) problem of minimizing P subject to the equality constraint u = 0.
Augmented Lagrangian Method
We will now translate the duality between the proximal and augmented Lagrangian iterations just described
to the constrained optimization context, setting the stage for using this duality in an incremental context.
Consider a generic convex programming problem of the form
minimize
H(y)
(2.17)
subject to y ∈ Y,
Ay − b = 0,
where H : ℜn 7→ ℜ is a convex function, Y is a convex set, A is an r × n matrix, and b ∈ ℜr . Consider also
the corresponding primal and dual functions
P (u) =
inf
y∈Y, Ay−b=u
H(y),
Q(λ) = inf
y∈Y
14
H(y) + λ′ (Ay − b) ,
which are convex and concave, respectively. We assume that P is closed and proper, and that the optimal
value of the problem is finite, so that Q is also closed proper and concave, and there is no duality gap (see
[Ber09], Section 4.2).
There is a well-known relation between the primal and dual functions. In particular, Q has the equivalent form
Q(λ) = inf r
inf
u∈ℜ y∈Y, Ay−b=u
H(y) + λ′ (Ay − b) = inf r P (u) + λ′ u ,
u∈ℜ
so P and Q satisfy the conjugacy relation (2.11). Based on the preceding discussion [cf. (2.11)-(2.16)], it
follows that the proximal iteration (2.13) can be equivalently written as the two-step process (2.14)-(2.15)
n
o
αk
uk+1 ∈ arg minr P (u) + λ′k u +
kuk2 ,
u∈ℜ
2
(2.18)
λk+1 = λk + αk uk+1 .
(2.19)
followed by
Moreover, from Eqs. (2.15) and (2.16), we have
uk+1 =
λk+1 − λk
˜
= ∇Q(λ
k+1 ).
αk
(2.20)
We will now write the iteration (2.18)-(2.19) in terms of the augmented Lagrangian, and obtain the
classical (first order) augmented Lagrangian method. Using the definition of the primal function P , we see
that the minimization in Eq. (2.18) can be written as
n
o
αk
inf r P (u) + λk ′ u +
kuk2
u∈ℜ
2
αk
′
2
inf
H(y) + λk u +
= inf r
kuk
y∈Y, Ay−b=u
u∈ℜ
2
o
n
αk
kAy − bk2
H(y) + λ′k (Ay − b) +
= inf r
inf
u∈ℜ y∈Y, Ay−b=u
2
n
o
α
k
= inf H(y) + λ′k (Ay − b) +
kAy − bk2
y∈Y
2
= inf Lαk (y, λk ),
y∈Y
where for any α > 0, Lα is the augmented Lagrangian function
Lα (y, λ) = H(y) + λ′ (Ay − b) +
α
kAy − bk2 ,
2
y ∈ ℜn , λ ∈ ℜr .
(2.21)
From the preceding calculation it also follows that for any yk+1 ∈ Y that minimizes the augmented Lagrangian over Y :
yk+1 ∈ arg min Lαk (y, λk ),
y∈Y
(2.22)
we have uk+1 = Ayk+1 − b, and the iteration (2.19) can be equivalently written as the multiplier iteration
λk+1 = λk + αk (Ayk+1 − b).
15
(2.23)
This is precisely the (first order) augmented Lagrangian method. It is equivalent to the proximal iteration
1
λk+1 ∈ arg maxr Q(λ) −
kλ − λk k2 ,
λ∈ℜ
2αk
[cf. Eq. (2.13)]. In view of Eqs. (2.20) and (2.23), it can also be written in the gradient-like form
˜
λk+1 = λk + αk ∇Q(λ
k+1 ),
(2.24)
˜
where ∇Q(λ
k+1 ), the special subgradient of Q at λk+1 , is given by
˜
∇Q(λ
k+1 ) = Ayk+1 − b.
(2.25)
Note that the minimizing yk+1 in Eq. (2.22) need not exist or be unique. Its existence must be assumed
in some way, e.g., by assuming that H has compact level sets. As an example, it can be verified that for
1
the two-dimensional/single constraint problem of minimizing H(y) = ey , subject to y 1 + y 2 = 0, y 1 ∈ ℜ,
y 2 ≥ 0, the dual optimal solution is λ∗ = 0, but there is no primal optimal solution. For this problem, the
augmented Lagrangian algorithm will generate sequences {λk } and {yk } such that λk → 0 and yk → −∞.
Incremental Augmented Lagrangian Methods
The duality between the proximal and augmented Lagrangian minimizations outlined above is generic, and
holds in other related contexts, based on a similar use of the Fenchel duality theorem. In the context of the
separable problem (2.1), it holds in an incremental form where Q(λ) is replaced by
qik (λ),
as in the IP iteration (2.9), or is replaced by
qik (λ) +
X
i6=ik
˜ i (λℓ )′ (λ − λk ),
∇q
i
as in the IAP iteration (2.10). We refer to these two methods as the incremental augmented Lagrangian
method (abbreviated IAL), and the incremental aggregated augmented Lagrangian method (abbreviated
IAAL).
Based on the discussion of the algorithm (2.22)-(2.24), the IAL method,
1
kλ − λk k2 ,
λk+1 ∈ arg maxr qik (λ) −
λ∈ℜ
2αk
can be implemented as follows, as already noted in [Ber15], Section 6.4.3.
16
Incremental Augmented Lagrangian Iteration (IAL)
Select a component index ik , and update the single component y ik according to
i
k
yk+1
∈ arg min
y ik ∈Yi
k
n
o
αk
hik (y ik ) + λ′k (Aik y ik − bik ) +
kAik y ik − bik k2 ,
2
(2.26)
i
while keeping the others unchanged, yk+1
= yki for all i 6= ik . Then update λ according to
i
k
λk+1 = λk + αk (Aik yk+1
− bik ).
(2.27)
As in the IP method, all component indexes should be selected for iteration in Eq. (2.26) with equal
long-term frequency. Note that the augmented Lagrangian minimization is decoupled with respect to the
components y i , thus overcoming the major limitation of the augmented Lagrangian approach for separable
problems.
To derive the IAAL method, we use the equivalent form (1.11)-(1.12) of the IAP algorithm. We see
then that the method has similar form to the IAL method, except that λk is first translated by a multiple
of the sum of the delayed subgradients. In particular, the IAAL iteration takes the form
1
kλ − νk k2 ,
λk+1 ∈ arg maxr qik (λ) −
λ∈ℜ
2αk
where
νk = λk + αk
X
i6=ik
˜ i (λℓ ).
∇q
i
(2.28)
Applying the relations (2.22)-(2.24), it follows that we can write the IAAL iteration in two steps: Select a
component index ik , and update the single component y ik according to
i
k
yk+1
∈ arg min
y ik ∈Yi
k
n
hik (y ik ) + νk′ (Aik y ik − bik ) +
o
αk
kAik y ik − bik k2 ,
2
(2.29)
i
while keeping the others unchanged, yk+1
= yki for all i 6= ik . Then update λ according to
i
k
λk+1 = νk + αk (Aik yk+1
− bik ).
(2.30)
˜ i (λℓ ), needed for the computation of νk in Eq. (2.28), are generated by
Note that the subgradients ∇q
i
˜ i (λℓ ) = Ai y i − bi ,
∇q
i
ℓi
17
∀ i 6= ik ,
[cf. Eq. (2.25)]. Thus by streamlining the preceding relations, we see that the IAAL updates are written as
2
X
αk
ik
yk+1
∈ arg min
hik (y ik ) + λ′k (Aik y ik − bik ) +
Aik y ik − bik +
(Ai yℓii − bi )
,
2
y ik ∈Yi
i6=ik
k
If we denote b =
Pm
i=1 bi ,
i
k
− bik +
λk+1 = λk + αk Aik yk+1
X
i6=ik
(Ai yℓii − bi ) .
and neglect the constant term −λ′k bik from the augmented Lagrangian, we can
write the iteration in a way that it depends on the scalars bi only through their sum b.
Incremental Aggregated Augmented Lagrangian (IAAL) Iteration
Select a component index ik , and update the single component y ik according to
X
αk
ik
yk+1
∈ arg min
hik (y ik ) + λ′k Aik y ik +
Aik y ik +
Ai yℓii − b
2
y ik ∈Yi
i6=ik
k
2
,
(2.31)
i
while keeping the others unchanged, yk+1
= yki for all i 6= ik . Then update λ according to
i
k
+
λk+1 = λk + αk Aik yk+1
X
i6=ik
Ai yℓii − b .
(2.32)
By comparing the IAL method (2.26)-(2.27) with the IAAL method (2.31)-(2.32), we see that they
require comparable computations per iteration. While the IAL method requires a diminishing stepsize αk
for convergence, the IAAL method can converge with a constant stepsize, assuming that the dual function
components have Lipschitz continuous gradients, and the dual function is strongly concave (cf. Prop. 1.1).
Intuitively, if it can use a constant stepsize, the IAAL method should be asymptotically more effective than
the IAL method. Of course, if Q is not strongly convex (as for example in the important case where Q
is polyhedral, which arises in integer programming), our analysis guarantees the convergence of the IAAL
method only if the stepsize αk is diminishing. In this case it is unclear which of the IAL and IAAL methods
is more effective on a given problem.
Both the IAL and IAAL algorithms require an initial multiplier λ0 . Regarding the delayed indexes ℓi
in the IAAL algorithm, if the iteration is executed at a single processor, it is most appropriate to choose ℓi
to be the iteration index at which the component y i was last changed prior to the current index k, so ℓi ≤ k
(if a component y i has not yet been updated prior to k, we take ℓi = 0 and let y0i be some initial choice
18
for y i ). In this case, the formal statement of the IAAL method is again given by Eqs. (2.29)-(2.30), with
ℓi replaced by k for all i 6= ik . However, a different value of ℓi may apply if the iteration is executed in a
distributed asynchronous computing environment, as in the corresponding IAS method of [NBB01].
Note that the multiplier λk is updated each time a component y i is updated, which suggests that
the stepsize αk should be chosen carefully, possibly through some experimentation. Moreover, the strong
convexity assumption of Q is essential for the convergence of the method with a constant stepsize. Indeed a
three-dimensional example by Chen, He, Ye, and Yuan [CHH14] can be used to show that the IAAL algorithm
need not converge for any value of constant stepsize if the strong convexity assumption is violated.† An
alternative possibility is to perform a batch of component updates y i of the form (2.29) between multiplier
updates of the form (2.30). For example, one may restructure the IAAL iteration so that it consists of a full
i
cycle of updates of y 1 , . . . , y m , sequentially according to Eq. (2.31), to obtain yk+1
, i = 1, . . . , m, and only
then to update λ according to
λk+1 = λk + αk
m
X
i=1
Note that this sequential update of
y1, . . . , ym
!
i
Ai yk+1
−b .
according to Eq. (2.31) amounts to a cycle of coordinate
descent iterations for minimizing the augmented Lagrangian. Therefore, this variant of the IAAL iteration
may be viewed as an implementation of the augmented Lagrangian method with approximate minimization
of the augmented Lagrangian using coordinate descent. An algorithm of this type may be interesting and
has been suggested in the past (see Bertsekas and Tsitsiklis [BeT89], Example 4.4, and Eckstein [Eck12]).
Its linear convergence has been shown under certain assumptions by Hong and Luo [HoL13]. The algorithm
is worthy of further investigation, particularly in view of favorable computational results given by Wang,
Hong, Ma, and Luo [WHM13]. Let us also note that the work by Hong, Chang, Wang, Razaviyayn, Ma,
and Luo [HCW14] derives an algorithm for the separable problem (2.1) that is quite similar to the IAAL
algorithm, using different assumptions and line of development. The paper [HCW14] proves convergence but
not a linear convergence rate result.
Comparison with ADMM
We will now compare the IAAL iteration with the ADMM. We note that there is a well-known connection
of the ADMM and augmented Lagrangian methods, which was clarified long ago through a series of papers.
In particular, Lions and Mercier [LiM79] proposed a splitting algorithm for finding a zero of the sum of two
maximal monotone operators, known as the Douglas-Ratchford algorithm. It turns out that this algorithm
† While the paper [CHH14] is entitled “The Direct Extension of ADMM for Multi-Block Convex Minimization
Problems ...,” it considers an algorithm that is not a special case of ADMM, so a convergence counterexample is
possible. A correct specialization of ADMM for separable problems (dating from 1989 but unknown to the authors
of [CHH14]) will be given shortly, and is convergent under the same broadly applicable conditions as ADMM.
19
contains as a special case the ADMM, as shown in [Gab83]. The paper by Eckstein and Bertsekas [EcB92]
showed that the general form of the proximal algorithm for finding a zero of maximal monotone operator,
proposed by Rockafellar [Roc76a], [Roc76b], contains as a special case the Douglas-Ratchford algorithm and
hence also the ADMM. Thus the ADMM and the augmented Lagrangian method have a common ancestry:
they are both special cases of the general form of the proximal algorithm for finding a zero of a maximal
monotone operator. The common underlying structure of the two methods is reflected in similar formulas,
but ADMM has the advantage of flexibility to allow decomposition, at the expense of a typically slower
practical convergence rate.
A convenient decomposition-based form of ADMM for the separable problem (2.1) was derived (together
with the corresponding coordinate descent version of the augmented Lagrangian method) in [BeT89], Section
3.4 and Example 4.4 (see also [Ber15], Section 5.4.2). Wang, Hong, Ma, and Luo [WHM13], apparently
unaware of this form of ADMM, give related algorithms (referred to as Algorithms 2 and 3 in their paper),
which, however, involve updating m multiplier vectors in place of the single multiplier update of the following
algorithm. At iteration k, and given λk , the ADMM algorithm of [BeT89] generates λk+1 as follows.
ADMM Iteration for Separable Problems
Perform a separate augmented Lagrangian minimization over y i , for each i = 1, . . . , m,
2
m
X
1
α
j
i
i
′
i
i
i
,
i = 1, . . . , m,
Aj yk − b
Ai y − Ai yk +
yk+1 ∈ arg min hi (y ) + λk Ai y +
2
m j=1
y i ∈Yi
(2.33)
and then update λk according to
λk+1
α
= λk +
m
m
X
i=1
i
Ai yk+1
!
−b .
(2.34)
Note that contrary to the augmented Lagrangian method, where the best strategy for adjusting α is
usually clear, see e.g., [Ber82], there is no clear way to adjust the parameter α to improve performance
in ADMM. As a result for efficiency α is often determined by trial and error. A closely related but more
refined form of ADMM, also derived in [BeT89a], Section 3.4, Example 4.4, aims to improve the parameter
selection by exploiting the structure of the matrices Ai . It uses a coordinate-dependent parameter
α
mj
in
iteration (2.34), in place of α/m, where mj is the number of submatrices Ai that have nonzero jth row. In
this version, the multiplier update essentially involves diagonal scaling. The iteration maintains additional
vectors zki ∈ ℜr , i = 1, . . . , m, which represent estimates of Ai y i at the optimum, and has the following form,
20
where Aji denotes the jth row of the matrix Ai .
Diagonally Scaled ADMM Iteration for Separable Problems
Perform a separate augmented Lagrangian minimization over y i , for each i = 1, . . . , m,
n
α
hi (y i ) + λ′k Ai y i +
Ai y i − zki
i
2
y ∈Yi
i
yk+1
∈ arg min
2
o
,
i = 1, . . . , m,
(2.35)
and then update λk and zk according to
λjk+1
=
λjk
α
+
mj
m
X
i=1
i
i
zk+1
= Ai yk+1
+
i
Aji yk+1
− bj
!
λk − λk+1
,
α
,
j = 1, . . . , r,
i = 1, . . . , m.
(2.36)
(2.37)
Note that the preceding two ADMM iterations coincide when there is no nonzero row in any of the
matrices Ai , i.e., mj = m for all j. In comparing the IAAL iteration (2.31)-(2.32), and the ADMM iterations
(2.33)-(2.34) and (2.35)-(2.37), we note that they involve fairly similar operations. In particular, the ADMM
mutiplier update (2.34) approximates an average (over a full cycle of m components) of the IAAL multiplier
updates (2.32), and is executed m times less frequently; this is reminiscent of the difference between the
proximal and incremental proximal iterations. The different multiplier update frequencies of IAAL and
ADMM suggests that assuming IAAL converges, its stepsize αk should be chosen much smaller than the
stepsize α in ADMM, say
αk ∈
α i
,
m m2
hα
,
as a crude approximation, for comparable performance. There are also two other major differences:
(a) The ADMM iterations have guaranteed convergence for any constant stepsize α, and under weaker
conditions (differentiability of qi and strong convexity of Q are not required). On the other hand the
IAAL method requires a diminishing stepsize in general, or (under Lipschitz continuity of ∇qi and
strong convexity of Q) a constant stepsize that is not arbitrary, but must be sufficiently small.
(b) In the IAAL method a single component y i is updated at each iteration, while in the ADMM all
components y i are updated. For some problems, this may work in favor of IAAL, particularly for large
m, a case that generally seems to favor incremental methods.
Thus for the separable problems of this section, one may roughly view the IAAL method as an incremental
variant of ADMM, where the advantage of incrementalism may be offset by less solid convergence properties.
21
A computational comparison of the two methods will be helpful in clarifying their relative merits.
The diagonally scaled ADMM iteration (2.35)-(2.37) suggests also a similar diagonal scaling for the
IAAL iteration. The simplest way to accomplish this is to use the IAAL method (2.31)-(2.32) after scaling
the constraints, i.e., after multiplying the r constraint equations with different scaling factors, which in turn
will introduce diagonal scaling for the dual variables. Proposition 1.1 will still apply under this form of
scaling, assuming Lipschitz continuity of ∇qi and strong convexity of Q.
Comparison with the Methods of Tadjewski [Tad89] and Ruszczynski [Rus95]
The methods of [Tad89] and [Rus95] are motivated by the earlier algorithm of [StW75], and apply to the
separable constrained optimization problem of this section. They are similar to each other, but use different
assumptions. The method of [Tad89] requires differentiability and second order sufficiency assumptions,
but applies to nonconvex separable problems that may have a duality gap, while the method of [Rus95]
applies to separable problems with convex, possibly nondifferentiable cost function. These methods are also
similar to our IAAL method (2.29)-(2.30), but they use different approximations of the quadratic penalty
terms. In particular, instead of the vectors yℓii that appear in Eqs. (2.29) and (2.30), they use other terms
that are iteratively adjusted, with the aim to improve the approximation of the quadratic penalty terms
of the standard augmented Lagrangian. Both papers [Tad89] and [Rus95] provide a convergence analysis,
involving suitable choices of various parameters, although the convergence results obtained are not as strong
as the ones for ADMM. A major difference of the methods of [Tad89] and [Rus95] from our IAAL method is
that, like the ADMM, they update all the components y i simultaneously at each iteration, so they are not
incremental in character.
3.
PROOF OF PROPOSITION 1.1
Similar to other convergence proofs of incremental gradient methods, including the one of [GOP15] for the
IAG method that we follow, the proof of Prop. 1.1 is based on viewing the IAP iteration with constant
stepsize αk ≡ α,
xk+1 = xk − α ∇fik (xk+1 ) +
X
i6=ik
∇fi (xℓi ) ,
(3.1)
as a gradient method with errors in the calculation of the gradient [cf. Eqs. (1.15), (1.16)]. To deal with the
delays in the iterates, we use the following lemma, due to Feyzmahdavian, Aytekin, and Johansson [FAJ14]
(which is also used in the convergence proof of [GOP15]):
22
Lemma 3.1:
Let {βk } be a nonnegative sequence satisfying
βk+1 ≤ pβk + q
max
max{0,k−d}≤ℓ≤k
βℓ ,
∀ k = 0, 1, . . . ,
for some positive integer d and nonnegative scalars p and q such that p + q < 1. Then we have
βk ≤ ρk β0 ,
∀ k = 0, 1, . . . ,
1
where ρ = (p + q) 1+d .
In the following proof we take the stepsize α as small as is needed for the various calculations to be
valid. Also for convenience in expressing various formulas involving delays, we consider the algorithm for
large enough iteration indexes, so that all the delayed iteration indexes in the following calculations are larger
than 0 (for this it will be sufficient to consider the algorithm as starting at an iteration k ≥ 2b). Note that
the Lipschitz condition on ∇fi implies a Lipschitz condition and a bound on ∇F . In particular, denoting
L=
m
X
Li ,
i=1
we have for all x, z ∈ ℜn ,
∇F (x)−∇F (z) =
m
X
i=1
∇fi (x) −
m
X
i=1
∇fi (z) ≤
m
X
i=1
∇fi (x)−∇fi (z) ≤
m
X
i=1
Li kx−zk = Lkx−zk. (3.2)
As a special case, for z = x∗ , where x∗ is the unique minimum of F , we have
∇F (xℓ ) = ∇F (xℓ ) − ∇F (x∗ ) ≤ Lkxℓ − x∗ k,
∀ ℓ ≥ 0.
(3.3)
We break down the proof of Prop. 1.1 in steps, first writing the iteration (3.1) as a gradient iteration
with errors, then carrying along the errors in a standard line of linear convergence analysis of gradient
methods without errors, then bounding the errors, and finally using Lemma 3.1:
(a) We write the iteration (3.1) as a gradient method with errors
xk+1 = xk − α ∇F (xk ) + ek ,
(3.4)
where the error term ek is given by
ek = ∇fik (xk+1 ) − ∇fik (xk ) +
23
X
i6=ik
∇fi (xℓi ) − ∇fi (xk ) .
(3.5)
(b) We relate the gradient error ek to the distance kxk − x∗ k by verifying the relation
kxk+1 − x∗ k2 = kxk − x∗ k2 − 2α∇F (xk )′ (xk − x∗ ) + α2 ∇F (xk )
2
+ Ek ,
(3.6)
where
′
Ek = α2 kek k2 − 2α xk − α∇F (xk ) − x∗ ek .
(3.7)
This is done by subtracting x∗ from both sides of Eq. (3.4), norm-squaring both sides, and carrying out the
straightforward calculation.
(c) We use Eq. (3.7) to bound |Ek | according to
|Ek | ≤ α2 kek k2 + 2αkek k xk − x∗ ,
(3.8)
for all sufficiently small α. In particular, from Eq. (3.7), we have
|Ek | ≤ α2 kek k2 + 2αkek k xk − x∗ − α∇F (xk ) ,
and Eq. (3.8) is obtained from the preceding relation by using the inequality
xk − x∗ − α∇F (xk ) ≤ kxk − x∗ k.
which holds for α sufficiently small; this is a consequence of the fact that under the gradient Lipschitz
assumption, a gradient iteration (with no error) reduces the distance to x∗ for α ∈ (0, 1/L] (see e.g., [Ber15],
Prop. 6.1.6).
(d) We use the strong convexity assumption
′
∇F (x) − ∇F (y) (x − y) ≥ σkx − yk2 ,
∀ x, y ∈ ℜn ,
(3.9)
where σ is the coefficient of strong convexity and the Lipschitz condition (3.2), to invoke the relation
∇F (xk )′ (xk − x∗ ) ≥
σL
1
kxk − x∗ k2 +
∇F (xk )
σ+L
σ+L
2
;
(3.10)
see e.g., [Nes14], Th. 2.1.22, or [Ber15], Prop. 6.1.9(b). This will be used to bound the term ∇F (xk )′ (xk −x∗ )
of Eq. (3.6).
(e) We show that for α ≤
2
σ+L ,
we have
σL
∗
2
kxk − x∗ k2 + |Ek |.
kxk+1 − x k ≤ 1 − 2α
σ+L
In particular, using the relations (3.6) and (3.10), we have
σL
1
2
∗
2
∗
2
∗
2
+ α2 ∇F (xk )
kxk+1 − x k ≤ kxk − x k − 2α
∇F (xk )
kxk − x k +
σ+L
σ+L
σL
2
2
∗
2
≤ 1 − 2α
kxk − x k + α α −
∇F (xk ) + |Ek |,
σ+L
σ+L
24
(3.11)
2
+ |Ek |
from which Eq. (3.11) follows.
(f) We prove that the error ek is proportional to the stepsize α, and to the maximum distance of the
iterates from x∗ over the past 2b iterates:
kek k ≤ O(α)
max
k−2b≤ℓ≤k
kxℓ − x∗ k.
(3.12)
This is straightforward, using the Lipschitz assumption on ∇fi and the bound (3.3) on ∇F .
In particular, from Eq. (3.5), we have
kek k ≤ ∇fik (xk+1 ) − ∇fik (xk ) +
≤ Lik kxk+1 − xk k +
≤ Lik kxk+1 − xk k +
X
i6=ik
X
i6=ik
X
i6=ik
∇fi (xℓi ) − ∇fi (xk )
Li kxk − xℓi k
(3.13)
Li kxk − xk−1 k + · · · + kxℓi +1 − xℓi k .
Moreover from Eqs. (3.3) and (3.4),
kxℓ+1 − xℓ k = α ∇F (xℓ ) + αkeℓ k ≤ αLkxℓ − x∗ k + αkeℓ k,
∀ ℓ ≥ 0.
(3.14)
Using this relation for ℓ in the range [k − b, k] in Eq. (3.13), we obtain
k
X
(1 − αLik )kek k ≤ O(α)
ℓ=k−b
kxℓ −
x∗ k
+
k−1
X
ℓ=k−b
!
keℓ k ,
where for p ≥ 1, we generically use O(αp ) to denote any function of α such that for some scalar γ > 0, we
have O(αp ) ≤ γαp for all α in some bounded open interval containing the origin. Thus,
kek k ≤ O(α)
k
X
ℓ=k−b
kxℓ −
x∗ k
+
k−1
X
ℓ=k−b
!
keℓ k .
(3.15)
From Eq. (3.5), we also have
keℓ k ≤ Liℓ kxℓ+1 − xℓ k +
≤ L kxℓ+1 −
x∗ k
X
i6=iℓ
Li kxℓ − xℓi k
+ kxℓ −
x∗ k
+
X
i6=iℓ
Li kxℓ − x∗ k + kxℓi − x∗ k .
(3.16)
Since for ℓ in the range [k − b, k − 1], ℓi lies in the range [k − 2b, k − 1], it follows that
keℓ k ≤ c
max
k−2b≤ℓ≤k
kxℓ − x∗ k,
∀ ℓ ∈ [k − b, k − 1],
where c is some constant that is independent of k and ℓ. Combining this with Eq. (3.15), we obtain Eq.
(3.12).
25
(g) We use Eqs. (3.8), (3.11), and (3.12) to obtain
kxk+1 −
x∗ k2
σL
≤ 1 − 2α
kxk − x∗ k2 + O(α2 ) max kxℓ − x∗ k2 .
k−2b≤ℓ≤k
σ+L
(3.17)
In particular, the two terms bounding |Ek | in Eq. (3.8) are α2 kek k2 and αkek k xk − x∗ , which in view of
Eq. (3.12) are bounded by terms that are O(α4 ) and O(α2 ) times maxk−2b≤ℓ≤k kxℓ − x∗ k2 , respectively.
σL
(h) We use Eq. (3.17) and Lemma 3.1, with d = 2b, βk = kxk − x∗ k2 , p = 1 − 2α σ+L
, and q = O(α2 ),
√
so that p + q < 1 for sufficiently small α. This shows that βk = kxk − x∗ k converges linearly to 0, and
completes the proof.
Q.E.D.
Convergence Rate Comparison for Small Stepsizes
Note that Eq. (3.17) provides a more refined rate of convergence estimate. While this estimate is not very
precise, because of the second order term on the right in Eq. (3.17), it shows that the ratio
σL
L
=
σ+L
1 + L/σ
where L =
Pm
i=1
Li and σ is the coefficient of strong convexity, plays an important role, and in particular
the convergence rate is improved when the “condition number” L/σ is small. The role of the ratio L/σ
in determining the convergence rate of gradient methods (without error) is well-known; see e.g., [Nes04],
[Ber15].
Convergence rate estimates like the one of Eq. (3.17) can also be similarly derived for IAG (as shown in
[GOP15]), and for the standard nonincremental gradient method [for which the error term |Ek | in Eq. (3.11)
is equal to 0]. These estimates, to first order [i.e., after neglecting the second order term in the right-hand
side of Eq. (3.17)], are identical for IAP, IAG, and for the standard nonincremental gradient method. This
suggests that for very small values of α, IAP and IAG perform comparably, while the nonincremental gradient
method performs much worse because it requires m times as much overhead per iteration to calculate the
full gradient of the cost function.
4.
NONQUADRATIC INCREMENTAL PROXIMAL AND AUGMENTED LAGRANGIAN
METHODS
The augmented Lagrangian methods of Section 2 apply to linear equality constrained problems for which the
multiplier vector λ is unconstrained. This allows the application of the linear convergence result of Prop. 1.1.
We will now consider convex inequality constraints, whose multipliers must be nonnegative. As a result the
dual problem involves an orthant constraint, and the linear convergence result of Prop. 1.1 does not apply.
26
Unfortunately, when there is an orthant constraint [i.e., X = {x | x ≥ 0} instead of X = ℜn in Eq. (1.9)],
the proof of Prop. 1.1 breaks down because the critical inequality (3.3) fails. In fact, to our knowledge, a
linear convergence rate result for the IAG method (1.8) applied with an orthant constraint is not currently
available. Moreover, the convergence of the augmented Lagrangian-like methods discussed in Section 2 has
been analyzed only for the equality-constrained case. In this section we will try to address this difficulty by
using a different (nonquadratic) proximal approach.
In particular, we will introduce incremental augmented Lagrangian methods for convex inequality constraints, where the quadratic penalty in the augmented Lagrangian is replaced by a suitable nonquadratic
penalty. One of our objectives is to develop linearly convergent methods that can exploit separability, similar
to the ones of Section 2. A second objective is to develop corresponding dual linearly convergent incremental
aggregated gradient and proximal methods for differentiable minimization subject to nonnegativity constraints.
Nonquadratic Augmented Lagrangian Methods for Inequality Constraints
Consider the convex programming problem
minimize
H(y)
(4.1)
subject to y ∈ Y,
Gj (y) ≤ 0, j = 1, . . . , r,
where H : ℜn 7→ (−∞, ∞) and Gj : ℜn 7→ (−∞, ∞) are convex functiona, and Y is a convex set. The
corresponding dual problem is
maximize
Q(µ)
(4.2)
subject to µ ≥ 0,
where Q : ℜr 7→ [−∞, ∞) is the concave function of the multiplier vector µ = (µ1 , . . . , µr ), given by
Q(µ) = inf
y∈Y
H(y) +
r
X
j=1
µj Gj (y) ,
µ ∈ ℜr .
(4.3)
We will apply an augmented Lagrangian method, first proposed by Kort and Bertsekas [KoB72], and
further developed in a number of subsequent works, including the monograph [Ber82] (Chapter 5). The
method makes use of a nonquadratic penalty function ψ : ℜ 7→ ℜ with the following properties:
(i) ψ is twice differentiable and ∇2 ψ(t) > 0 for all t ∈ ℜ,
(ii) ψ(0) = 0, ∇ψ(0) = 1,
(iii) limt→−∞ ψ(t) > −∞,
(iv) limt→−∞ ∇ψ(t) = 0 and limt→∞ ∇ψ(t) = ∞.
27
The most common and interesting special case is the exponential
ψ(s) = exp(s) − 1,
s ∈ ℜ.
(4.4)
The corresponding exponential augmented Lagrangian method and its dual, a proximal algorithm known as
the entropy minimization algorithm, has been analyzed first in [KoB72] and [Ber82], and then by Tseng and
Bertsekas [TsB93]. Related classes of methods, which also contain the exponential and entropy methods as
special cases, were proposed and analyzed later by Iusem, Svaiter, and Teboulle [IST94]; see also the survey
by Iusem [Ius99], which contains followup work and many references.
The augmented Lagrangian algorithm corresponding to ψ and problem (4.2) maintains multipliers
µjk
> 0, j = 1, . . . , r, for the inequality constraints, and consists of finding
r
X
µjk
j
yk+1 ∈ arg min H(y) +
,
ψ
α
G
(y)
j
k
y∈Y
αjk
(4.5)
j=1
where αjk > 0, j = 1, . . . , r, are penalty parameters, followed by the multiplier iteration
µjk+1 = µjk ∇ψ ajk Gj (yk+1 ) ,
j = 1, . . . , r.
(4.6)
Alternatively and equivalently, based on the Fenchel duality theorem, one may show that the multiplier
iteration can be written in the proximal form
µk+1 ∈ arg maxr
µ∈ℜ
Q(µ) −
r
X
µjk
j=1
αjk
ψ⋆
µj
!
µjk
,
(4.7)
where Q is the dual function given by Eq. (4.3), and ψ ⋆ is the convex conjugate of ψ.
To see the equivalence of the expressions (4.6) and (4.7), let us write
ujk+1 = Gj (yk+1 ),
j = 1, . . . , r,
and note that the augmented Lagrangian minimization (4.5) yields
r
X
µjk
j j
uk+1 ∈ arg
min
P (u) +
ψ(α
u
)
,
k
u=(u1 ,...,ur )∈ℜr
αjk
(4.8)
j=1
where P is the primal function
P (u) =
inf
y∈Y, Gj (y)≤uj , j=1,...,r
H(y).
Then the minimization in Eq. (4.8) is the Fenchel dual to the maximization (4.7). By applying the Fenchel
duality theorem, we have that the maximizing vector in Eq. (4.7) is equal to the gradient
r
j
X
µk
j j
,
∇
j ψ(αk u )
j=1 αk
u=u
k+1
28
so it is given by the formula (4.6).
Note that while the dual problem is to maximize Q(µ) subject to µ ≥ 0, the proximal maximization
(4.7) is unconstrained. The reason is that the conjugate ψ ⋆ takes the value ∞ outside the nonnegative
orthant, and has the character of a barrier function within the nonnegative orthant. As an example, for the
exponential function (4.4) the conjugate is the entropy function
t ln(t) − 1 + 1
ψ ⋆ (t) = 1
∞
if t > 0,
(4.9)
if t = 0,
if t < 0.
An important advantage of the nonquadratic augmented Lagrangian method versus its quadratic counterpart, is that it leads to twice differentiable augmented Lagrangians. This advantage also carries over to
the incremental augmented Lagrangian methods to be presented next.
Nonquadratic Incremental Augmented Lagrangian Methods for Inequality Constraints
Consider now the separable constrained optimization problem
minimize
m
X
hi (y i )
i=1
subject to
yi
∈ Yi , i = 1, . . . , m,
m
X
(4.10)
gji
i=1
(y i )
≤ 0,
where hi and gji are convex real-valued functions, and Yi are convex sets. Similar to the development of
Section 2, the corresponding incremental aggregated augmented Lagrangian method, which parallels IAAL,
maintains a vector µk > 0 and operates as follows.
Incremental Aggregated Augmented Lagrangian Iteration for Inequalities (IAALI)
Select a component index ik , and update the single component y ik according to
r
j
X
X
µk j
ik
i )
ik ) +
ψ
α
g
(y
g
(y
,
yk+1
∈ arg min
hik (y ik ) +
jik
ji ℓi
k
αjk
y ik ∈Yi
(4.11)
i6=ik
j=1
k
i
while keeping the others unchanged, yk+1
= yki for all i 6= ik . Then update µ according to
µjk+1
=
µjk ∇ψ ajk
gji (y ik )
k
k+1
+
m
X
i6=ik
29
gji (yℓii ) ,
j = 1, . . . , r.
(4.12)
Note that the minimization (4.11) is of low dimension, but involves the nonquadratic penalty function
ψ. Thus even when the component y ik is one-dimensional, this minimization will likely require some form
of iterative line search. Note also that the update formula (4.12) can equivalently be written as
!
r
j
X
X
j
µ
µ
k ⋆
µk+1 ∈ arg maxr qik (µ) +
ψ
∇qi (µℓi )′ (µ − µk ) −
,
µ∈ℜ
αjk
µjk
i6=ik
(4.13)
j=1
where qi are the dual function components, given by
r
X
qi (µ) = inf
hi (y i ) +
µj gij (y i ) ,
y i ∈Y i
µ ∈ ℜr ,
j=1
i = 1, . . . , m.
The form (4.13) of the method can be viewed as an incremental aggregated proximal method for maximizing
Pm
Q(µ) = i=1 qi (µ) over µ ≥ 0, where
r
X
µj gji (y i ) ,
i = 1, . . . , m;
(4.14)
qi (µ) = inf
hi (y i ) +
y i ∈Yi
j=1
cf. Eq. (2.3). The convergence properties of the IAALI and the corresponding incremental aggregated
proximal method (4.13) for solving the dual problem
maximize
m
X
qi (µ)
i=1
subject to µ ≥ 0,
are interesting research subjects, as we will now discuss.
Nonquadratic Incremental Aggregated Proximal Algorithm for Nonnegativity Constraints
Consider the minimization problem
minimize
def
F (x) =
m
X
fi (x)
(4.15)
i=1
subject to x ≥ 0,
where fi : ℜn 7→ ℜ, i = 1, . . . , m, are convex real-valued functions. When translated to this minimization
context, the algorithm (4.13) maintains a vector xk > 0 that is updated as follows.
Nonquadratic Incremental Aggregated Proximal Iteration for X = {x | x ≥ 0}
Select a component index ik , and obtain xk+1 as
xk+1 ∈ arg minn
x∈ℜ
fik (x) +
X
i6=ik
∇fi (xℓi )′ (x − xk ) +
30
n
X
xjk
j=1
αjk
ψ⋆
xj
!
xjk
.
(4.16)
The analysis of the convergence properties of this algorithm is beyond the scope of this paper, and will
be the subject of a separate publication. In particular, it is interesting to investigate the linear convergence
of the method (4.16) when the parameters ajk are constant (but sufficiently small), under the appropriate
Lipschitz continuity and strong convexity assumptions, similar to Prop. 1.1. Note that by differentiating the
cost function in the minimization of Eq. (4.16), we obtain the optimality condition, which can be written as
1
1
xk+1
⋆
∇ψ
1
α1
x
k
k
X
..
∇fik (xk+1 ) +
∇fi (xℓi ) +
(4.17)
= 0.
.
xm
i6=ik
1
k+1
⋆
αm ∇ψ
xm
k
k
This expression may be used in the line of proof of Section 3 in place of the corresponding formula (1.14)
for the unconstrained IAP algorithm (1.14), which can be written in the form
∇fik (xk+1 ) +
X
i6=ik
∇fi (xℓi ) +
xk+1 − xk
= 0.
α
(4.18)
When ψ (and hence also ψ ⋆ ) is quadratic and αjk ≡ α, the two preceding formulas coincide. However,
contrary to iteration (4.18), the iteration (4.17) preserves the strict positivity of the iterates (xk > 0 for all
k), and addresses the orthant-constrained problem (4.15).
Entropy-Based Incremental Aggregated Proximal Algorithm for Nonnegativity Constraints
For an illustration of the algorithm (4.16), consider the special case where ψ is the exponential function and
ψ ⋆ is the entropy function, so that
ψ(s) = exp(s) − 1,
t ln(t) − 1 + 1 if t > 0,
ψ ⋆ (t) = 1
if t = 0,
∞
if t < 0,
∇ψ ⋆ (t)
=
(
ln(t)
if t > 0,
does not exist if t ≤ 0,
[cf. Eqs. (4.4) and (4.9)]. Then by using a constant stepsize αj for each coordinate, Eq. (4.17) takes the form
ln
xjk+1
xjk
!
X
∂f
(x
)
∂f
(x
)
i ℓi
i
k+1
,
+
= −αj k j
∂x
∂xj
j = 1, . . . , n,
(4.19)
i6=ik
where ik is the component index selected for iteration k. We can write this iteration as
ln
xjk+1
xjk
!
=
−αj
∂F (xk )
j
+ ek ,
∂xj
(4.20)
where ek = (e1k , . . . , enk ) is the error vector
ek = ∇fik (xk+1 ) − ∇fik (xk ) +
31
X
i6=ik
∇fi (xℓi ) − ∇fi (xk ) ,
(4.21)
that played an important role in the proof of Prop. 1.1 [cf. Eq. (3.5)].
We will use the line of analysis of Section 3 to speculate about the linear convergence of iteration
(4.19) and its equivalent form (4.20)-(4.21). Assume that the minimum x∗ satisfies the strict complementary
slackness condition
∂F (x∗ )
> 0,
∂xj
∀ j ∈ J 0,
(4.22)
where J 0 = j | (xj )∗ = 0 , and speculate on the behavior of {xk } in a small neighborhood around x∗ .
Consider first the iterates xjk , j ∈ J 0 , in a small neighborhood around x∗ . We note that the errors ejk
∂F (xk )
, for all
∂xj
j
j
the logarithms ln(xk+1 /xk ), j ∈ J 0 ,
so the sequences {xjk }, j ∈ J 0 , are
of Eq. (4.21) are near 0 and by Eq. (4.22), are negligible relative to the gradient components
j ∈ J 0 . In view of the form of iteration (4.20) and the condition (4.22),
are negative, and hence the ratios xjk+1 /xjk , j ∈ J 0 , are within [0, 1),
linearly decreasing towards 0.
Consider next the iterates xjk , j ∈
/ J 0 , in a small neighborhood around x∗ . They are close to the
corresponding positive numbers (xj )∗ , j ∈
/ J 0 , and they are iterated according to
ln(xjk+1 ) = ln(xjk ) − αj
∂F (xk )
j
+
e
k ,
∂xj
j∈
/ J 0,
(4.23)
[cf. Eq. (4.20)]. This looks like an incremental aggregated gradient iteration in the logarithms ln(xj ), j ∈
/ J 0.
Indeed by making the transformation of variables z j = ln(xj ), j = 1, . . . , n, for xj > 0, and introducing the
function
H(z 1 , . . . , z n ) = F exp(z 1 ), . . . , exp(z n ) ,
and its gradient, which is related to the gradient of F through the relation
∂F exp(z 1 ), . . . , exp(z n )
∂F (x)
∂H(z)
j)
=
exp(z
= xj
,
∂z j
∂xj
∂xj
j = 1, . . . , n,
we see that the iteration (4.23) can be written as
j
zk+1
=
zkj
−
αj
xjk
∂H(zk )
∂z j
+ αj ejk ,
j∈
/ J 0,
where xjk = exp(zkj ). Thus, neglecting the effect of the coordinates xj , j ∈ J 0 , that are fast diminishing to 0,
the iteration behaves like the IAP method restricted to the space of the coordinate logarithms z j = ln(xj ),
j∈
/ J 0 , with coordinate-dependent stepsizes
near x∗ .
αj
j
xk
that are close to the positive constants
αj
,
(xj )∗
j ∈ J 0 , for xk
By combining the preceding argument with the proof of Prop. 1.1, we can show that the method
converges to x∗ locally, i.e., when started sufficiently close to x∗ , assuming the strict complementarity
condition (4.22), and the appropriate stepsize, Lipschitz continuity, and strong convexity conditions. The
proof is long and will be deferred to a future publication. Moreover, for j ∈
/ J 0 , ln(xjk ) converges to
32
ln (xj )∗ linearly, while for j ∈ J 0 , {xjk } also converges to (xj )∗ linearly. However, a more sophisticated
argument is needed to show global and linear convergence of {xk } to x∗ , by combining the line of proof
of Prop. 1.1 with the existing convergence proofs of the entropy minimization algorithm and its dual, the
exponential method of multipliers.
Entropy-Based Incremental Aggregated Gradient Algorithm for Nonnegativity Constraints
Finally let us note the analog of the IAG method for nonnegativity constraints. In analogy with Eq. (4.19)
it has the form
ln
xjk+1
xjk
!
= −αj
m
X
∂fi (xℓi )
,
∂xj
i=1
j = 1, . . . , n,
or equivalently
xjk+1
=
xjk
exp
−αj
[the difference from Eq. (4.19) is the use of
m
X
∂fi (xℓi )
∂xj
i=1
∂fi (xℓ
k
∂xj
ik
)
!
,
j = 1, . . . , n,
(4.24)
∂fi (xk+1 )
k
].
∂xj
This iteration should be
in place of
compared with the IAS method (1.5), for the case where the functions fi are differentiable, and the stepise
αk is a constant α:
"
xk+1 = xk − α
m
X
i=1
#+
∇fi (xℓi )
,
(4.25)
where [·]+ denotes projection onto the nonnegative orthant. We may view the method (4.25) as the constrained version of the IAG method (1.8) with constant stepsize for which, however, no linear convergence
proof is presently available.†
The iteration (4.24) may also be viewed as an incremental version of the mirror descent method;
see Beck and Teboulle [BeT03], the surveys by Juditsky and Nemirovski [JuN11a], [JunN11b], and the
references quoted there, and the author’s presentation in [Ber15], Section 6.6. Using similar arguments to
the case of iteration (4.19), we can show that the iteration (4.24) converges linearly to x∗ , when started
sufficiently close to x∗ , assuming the strict complementarity condition (4.22), and the appropriate constant
stepsize, and other conditions. Note that the iteration (4.24) may be implemented more conveniently than
the proximal iteration (4.16), as it does not require a proximal minimization. However, the iteration (4.24)
is not suitable as the basis for the development of an incremental augmented Lagrangian method, such as
IAALI [cf. Eqs. (4.11)-(4.12)].
† A local linear convergence result for the constrained IAG method (4.25) is possible, assuming the strict complementarity condition (4.22). In particular, it can be shown that there is a sphere centered at x∗ such that if x0 belongs
to that sphere, then the sequence generated by iteration (4.25) stays within that sphere and converges linearly to
x∗ . The idea of the proof is that after the first iteration, all the iterates satisfy xjk = 0 for all indices j ∈ J 0 , so the
method essentially reduces to the IAG method in the space of variables xj , j ∈
/ J 0.
33
A final comment relates to the choice of the stepsizes αj in iteration (4.24). For the coordinates that
P
∂fi (xℓ )
i
are bounded away from 0 (i.e., for j ∈
/ J 0 ) we have asymptotically m
≈ 0, so from a Taylor
i=1
∂xj
expansion of the exponential in Eq. (4.24), we obtain
!
m
X
∂f
(x
)
1
i
ℓ
j
j
i
+
xk+1 = xk 1 + −αj
j
∂x
2
i=1
m
X
∂fi (xℓi )
−αj
∂xj
i=1
!2
+ · · · .
By discarding the second and higher order terms for j ∈
/ J 0 , we see that approximately,
xjk+1 ≈ xjk − αj xjk
m
X
∂fi (xℓi )
,
∂xj
i=1
j∈
/ J 0.
This suggests scaling the stepsizes αj for j ∈
/ J 0 , so that αj is inversely proportional to the optimal value
(xj )∗ . On the other hand, for j ∈ J 0 , it makes sense to choose αj large (subject to a positive lower bound)
in order to accelerate the convergence of xjk to (xj )∗ = 0. Thus a reasonable heuristic is to set
αj =
α
,
max{x̄j , δ}
j = 1, . . . , n,
where x̄j is an estimate for the optimal coordinate value (xj )∗ , α is some positive scalar, which corresponds
to the stepsize of the constrained IAG iteration (4.25), and δ is a small positive constant. One may also
consider updating the values αj in the course of the algorithm, as better estimates x̄j are obtained.
5.
CONCLUDING REMARKS
In this paper we have proposed IAP, an incremental aggregated proximal method, and we have shown that
under favorable assumptions, it attains a linear convergence rate, using a constant (but sufficiently small)
stepsize. The application of this method in a dual context, to separable constrained optimization problems,
yields the IAAL method, an incremental augmented Lagrangian method that preserves and exploits the
separable structure. The principal difference of our method relative to the several alternative augmented
Lagrangian-based proposals, is its incremental character and its high update frequency of the multiplier λk ;
the alternative methods, except Algorithm 1 of [WHM13] and the one of [RoT15], but including the proper
version of ADMM for separable problems, update all the primal variables y i , i = 1, . . . , m, simultaneously
rather than sequentially, so they are not incremental in nature. Moreover, the alternative methods update
the multipliers m times less frequently than IAAL. A systematic computational comparison of our methods
with the nonincremental alternatives will be helpful in clarifying what advantages our incremental approach
may hold.
There are several analytical issues relating to the IAAL method, which require further investigation.
For example a more refined convergence rate analysis may point the way to adaptive stepsize adjustment
schemes, and/or forms of scaling based on second derivatives of the cost function and the matrices Ai . There
34
are analyses of this type for ADMM; see the paper by Giselsson and Boyd [GiB15], and the references cited
there. Another possibility is to use a momentum term in the updating formula for the multiplier λ. A
third possibility is to control the degree of incrementalism by “batching” multiple augmented Lagrangian
iterations involving multiple components.
We have also proposed linearly converging extensions of IAAL for problems with convex inequality
constraints. These are based on a nonquadratic augmented Lagrangian approach such as the exponential,
and its dual version, which is an incremental aggregated entropy algorithm (4.19). The fuller investigation
of this method, as well as the method (4.24), which is the exponential analog of the IAG method for
nonnegativity constraints, are important subjects for investigation.
6.
REFERENCES
[AFB06] Ahn, S., Fessler, J., Blatt, D., and Hero, A. O., 2006. “Convergent Incremental Optimization Transfer
Algorithms: Application to Tomography,” IEEE Transactions on Medical Imaging, Vol. 25, pp. 283-296.
[BLY15] Bragin, M. A., Luh, P. B., Yan, J. H., Yu, N., and Stern, G. A., 2015. “Convergence of the Surrogate
Lagrangian Relaxation Method,” J. of Optimization Theory and Applications, Vol. 164, pp. 173-201.
[BNO03] Bertsekas, D. P., Nedić, A., and Ozdaglar, A. E., 2003. Convex Analysis and Optimization, Athena Scientific,
Belmont, MA.
[BPC11] Boyd, S., Parikh, N., Chu, E., Peleato, B., and Eckstein, J., 2011. Distributed Optimization and Statistical
Learning via the Alternating Direction Method of Multipliers, Now Publishers Inc, Boston, MA.
[BeT89] Bertsekas, D. P., and Tsitsiklis, J. N., 1989. Parallel and Distributed Computation: Numerical Methods,
Prentice-Hall, Englewood Cliffs, N. J.
[BeT03] Beck, A., and Teboulle, M., 2003. “Mirror Descent and Nonlinear Projected Subgradient Methods for Convex
Optimization,” Operations Research Letters, Vol. 31, pp. 167-175.
[Ber78] Bertsekas, D. P., 1978. “Local Convex Conjugacy and Fenchel Duality,” Preprints of 7th Triennial World
Congress of IFAC, Helsinki, Finland, Vol. 2, pp. 1079-1084.
[Ber79] Bertsekas, D. P., 1979. “Convexification Procedures and Decomposition Methods for Nonconvex Optimization
Problems,” J. of Optimization Theory and Applications, Vol. 29, pp. 169-197.
[Ber82] Bertsekas, D. P., 1982. Constrained Optimization and Lagrange Multiplier Methods, Academic Press, NY;
republished in 1996 by Athena Scientific, Belmont, MA. On line at http://web.mit.edu/dimitrib/www/lagrmult.html.
[Ber09] Bertsekas, D. P., 2009. Convex Optimization Theory, Athena Scientific, Belmont, MA.
[Ber10] Bertsekas, D. P., 2010. “Incremental Gradient, Subgradient, and Proximal Methods for Convex Optimization:
A Survey,” Lab. for Information and Decision Systems Report LIDS-P-2848, MIT; arXiv:1507.01030.
[Ber11] Bertsekas, D. P., 2011. “Incremental Proximal Methods for Large Scale Convex Optimization,” Math. Programming, Vol. 129, pp. 163-195.
[Ber15] Bertsekas, D. P., 2015. Convex Optimization Algorithms, Athena Scientific, Belmont, MA.
35
[CHH14] Chen, C., He, B., Ye, Y., and Yuan, X., 2014. “The Direct Extension of ADMM for Multi-Block Convex
Minimization Problems is not Necessarily Convergent,” Mathematical Programming, published on line.
[ChT94] Chen, G., and Teboulle, M., 1994. “A Proximal-Based Decomposition Method for Convex Minimization
Problems,” Mathematical Programming, Vol. 64, pp. 81-101.
[DLP14] Deng, W., Lai, M. J., Peng, Z., and Yin, W., 2014. “Parallel Multi-Block ADMM with O (1/k) Convergence,”
arXiv preprint arXiv:1312.3040v2.
[DaL15] Dang, C., and Lan, G., (2015). “Randomized First-order Methods for Saddle Point Optimization,” arXiv
preprint arXiv:1409.8625v3.
[EcB92] Eckstein, J., and Bertsekas, D. P., 1992. “On the Douglas-Rachford Splitting Method and the Proximal Point
Algorithm for Maximal Monotone Operators,” Math. Programming, Vol. 55, pp. 293-318.
[Eck12] Eckstein, J., 2012. “Augmented Lagrangian and Alternating Direction Methods for Convex Optimization: A
Tutorial and Some Illustrative Computational Results,” RUTCOR Research Report RRR 32-2012, Rutgers, Univ.
[Eve63] Everett, H., 1963. “Generalized Lagrange Multiplier Method for Solving Problems of Optimal Allocation of
Resources,” Operations Research, Vol. 11, pp. 399-417.
[FAJ14] Feyzmahdavian, H. R., Aytekin, A., and Johansson, M., 2014. “A Delayed Proximal Gradient Method with
Linear Convergence Rate,” in Prop. of 2014 IEEE International Workshop on Machine Learning for Signal Processing
(MLSP), pp. 1-6.
[GOP15] Gurbuzbalaban, M., Ozdaglar, A., and Parrilo, P., 2015. “On the Convergence Rate of Incremental Aggregated Gradient Algorithms,” arXiv preprint arXiv:1506.02081.
[GaM76] Gabay, D., and Mercier, B., 1976. “A Dual Algorithm for the Solution of Nonlinear Variational Problems
via Finite-Element Approximations,” Comp. Math. Appl., Vol. 2, pp. 17-40.
[Gab79] Gabay, D., 1979. Methodes Numeriques pour l’Optimization Non Lineaire, These de Doctorat d’Etat et
Sciences Mathematiques, Uni. Pierre at Marie Curie (Paris VI).
[Gab83] Gabay, D., 1983. “Applications of the Method of Multipliers to Variational Inequalities,” in M. Fortin and
R. Glowinski, eds., Augmented Lagrangian Methods: Applications to the Solution of Boundary-Value Problems,
North-Holland, Amsterdam.
[GiB15] Giselsson, P., and Boyd, S., 2015. “Metric Selection in Douglas-Rachford Splitting and ADMM,” arXiv
preprint arXiv:1410.8479v4.
[GlM75] Glowinski, R. and Marrocco, A., 1975. “Sur l’ Approximation par Elements Finis d’ Ordre un et la Resolution
par Penalisation-Dualite d’une Classe de Problemes de Dirichlet Non Lineaires” Revue Francaise d’Automatique
Informatique Recherche Operationnelle, Analyse Numerique, R-2, pp. 41-76.
[HCW14] Hong, M., Chang, T.-H., Wang, X., Razaviyayn, M., Ma, S., and Luo, Z.-Q., 2013. “A Block Successive
Upper Bound Minimization Method of Multipliers for Linearly Constrained Convex Optimization,? arXiv preprint
arXiv:1401.7079v1.
[HaM11] Hamdi, A., and Mishra, S. K., 2011. “Decomposition Methods Based on Augmented Lagrangians: A Survey,”
36
in Topics in Nonconvex Optimization, Springer, N. Y., pp. 175-203.
[HoL13] Hong, M., and Luo, Z. Q., 2013. “On the Linear Convergence of the Alternating Direction Method of
Multipliers,” arXiv preprint arXiv:1208.3922v3.
[IST94] Iusem, A. N., Svaiter, B. F., and Teboulle, M., 1994. “Entropy-Like Proximal Methods in Convex Programming,” Math. of Operations Research, Vol. 19, pp. 790-814.
[Ius99] Iusem, A. N., 1999. “Augmented Lagrangian Methods and Proximal Point Methods for Convex Minimization,”
Investigacion Operativa, Vol. 8, pp. 11-49.
[JuN11a] Juditsky, A., and Nemirovski, A., 2011. “First Order Methods for Nonsmooth Convex Large-Scale Optimization, I: General Purpose Methods,” in Optimization for Machine Learning, by Sra, S., Nowozin, S., and Wright,
S. J. (eds.), MIT Press, Cambridge, MA, pp. 121-148.
[JuN11b] Juditsky, A., and Nemirovski, A., 2011. “First Order Methods for Nonsmooth Convex Large-Scale Optimization, II: Utilizing Problem’s Structure,” in Optimization for Machine Learning, by Sra, S., Nowozin, S., and
Wright, S. J. (eds.), MIT Press, Cambridge, MA, pp. 149-183.
[KoB72] Kort, B. W., and Bertsekas, D. P., 1972. “A New Penalty Function Method for Constrained Minimization,”
Proc. 1972 IEEE Confer. Decision Control, New Orleans, LA, pp. 162-166.
[LiM79] Lions, P. L., and Mercier, B., 1979. “Splitting Algorithms for the Sum of Two Nonlinear Operators,” SIAM
J. on Numerical Analysis, Vol. 16, pp. 964-979.
[Mai13] Mairal, J., 2013. “Optimization with First-Order Surrogate Functions,” arXiv preprint arXiv:1305.3120.
[Mai14] Mairal, J., 2014. “Incremental Majorization-Minimization Optimization with Application to Large-Scale
Machine Learning,” arXiv preprint arXiv:1402.4419.
[Mar70] Martinet, B., 1970. “Regularisation d’ Inéquations Variationelles par Approximations Successives,” Revue
Fran. d’Automatique et Infomatique Rech. Opérationelle, Vol. 4, pp. 154-159.
[NBB01] Nedić, A., Bertsekas, D. P., and Borkar, V., 2001. “Distributed Asynchronous Incremental Subgradient
Methods,” Proc. of 2000 Haifa Workshop “Inherently Parallel Algorithms in Feasibility and Optimization and Their
Applications,” by D. Butnariu, Y. Censor, and S. Reich, Eds., Elsevier, Amsterdam.
[NeB01] Nedić, A., and Bertsekas, D. P., 2001. “Incremental Subgradient Methods for Nondifferentiable Optimization,” SIAM J. on Optimization, Vol. 12, 2001, pp. 109-138.
[NeB10] Nedić, A., and Bertsekas, D. P., 2010. “The Effect of Deterministic Noise in Subgradient Methods,” Math.
Programming, Ser. A, Vol. 125, pp. 75-99.
[Ned11] Nedić, A., 2011. “Random Algorithms for Convex Minimization Problems,” Math. Programming, Ser. B,
Vol. 129, pp. 225-253.
[Nes04] Nesterov, Y., 2004. Introductory Lectures on Convex Optimization, Kluwer Academic Publisher, Dordrecht,
The Netherlands.
[RoT15] Robinson, D. P., and Tappenden, R. E., 2015. “A Flexible ADMM Algorithm for Big Data Applications,”
arXiv preprint arXiv:1502.04391.
37
[Roc73] Rockafellar, R. T., 1973. “A Dual Approach to Solving Nonlinear Programming Problems by Unconstrained
Optimization,” Math. Programming, pp. 354-373.
[Roc76a] Rockafellar, R. T., 1976. “Monotone Operators and the Proximal Point Algorithm,” SIAM J. on Control
and Optimization, Vol. 14, pp. 877-898.
[Roc76b] Rockafellar, R. T., 1976. “Augmented Lagrangians and Applications of the Proximal Point Algorithm in
Convex Programming,” Math. of Operations Research, Vol. 1, pp. 97-116.
[Rus95] Ruszczynski, A., 1995. “On Convergence of an Augmented Lagrangian Decomposition Method for Sparse
Convex Optimization,” Math. of Operations Research, Vol. 20, pp. 634-656.
[SLB13] Schmidt, M., Le Roux, N., and Bach, F., 2013. “Minimizing Finite Sums with the Stochastic Average
Gradient,” arXiv preprint arXiv:1309.2388.
[TaM85] Tanikawa, A., and Mukai, M., 1985. “A New Technique for Nonconvex Primal-Dual Decomposition of a
Large-Scale Separable Optimization Problem,” IEEE Trans. Autom. Control, Vol. AC-30, pp. 133-143
[Tad89] Tatjewski, P., 1989. “New Dual-Type Decomposition Algorithm for Nonconvex Separable Optimization Problems,” Automatica, Vol. 25, pp. 233-242.
[TsB93] Tseng, P., and Bertsekas, D. P., 1993. “On the Convergence of the Exponential Multiplier Method for Convex
Programming,” Math. Programming, Vol. 60, pp. 1-19.
[WHM13] Wang, X., Hong, M., Ma, S., Luo, Z. Q., 2013. “Solving Multiple-Block Separable Convex Minimization
Problems Using Two-Block Alternating Direction Method of Multipliers,” arXiv preprint arXiv:1308.5294.
[WaB13] Wang, M., and Bertsekas, D. P., 2013. “Incremental Constraint Projection-Proximal Methods for Nonsmooth
Convex Optimization,” Lab. for Information and Decision Systems Report LIDS-P-2907, MIT, to appear in SIAM J.
on Optimization.
[WaB15] Wang, M., and Bertsekas, D. P., 2015. “Incremental Constraint Projection Methods for Variational Inequalities,” Mathematical Programming, Vol. 150, pp. 321-363.
38
| 3cs.SY
|
arXiv:1511.00954v1 [math.CO] 3 Nov 2015
EFFECTIVE INVARIANT THEORY OF PERMUTATION
GROUPS USING REPRESENTATION THEORY
NICOLAS BORIE
Abstract. Using the theory of representations of the symmetric group, we
propose an algorithm to compute the invariant ring of a permutation group.
Our approach have the goal to reduce the amount of linear algebra computations and exploit a thinner combinatorial description of the invariant ring.
Computational Invariant Theory, representation theory, permutation group
This is a drafty old version : full corrected text are available at :
http://www.springer.com/
1. Introduction
Invariant theory has been a rich and central area of algebra ever since the eighteenth theory, with practical applications [6, § 5] in the resolution of polynomial
systems with symmetries (see e.g. [4], [10], [20, § 2.6], [7]), in effective Galois theory
(see e.g. [5], [1], [11]), or in discrete mathematics (see e.g. [22, 16] for the original
motivation of the second author). The literature contains deep and explicit results
for special classes of groups, like complex reflection groups or the classical reductive
groups, as well as general results applicable to any group. Given the level of generality, one cannot hope for such results to be simultaneously explicit and tight in
general. Thus the subject was effective early on: given a group, one wants to calculate the properties of its invariant ring. Under the impulsion of modern computer
algebra, computational methods, and their implementations, have largely expanded
in the last twenty years [12, 20, 23, 6, 13, 14]. However much progress is still needed
to go beyond toy examples and enlarge the spectrum of applications.
Classical approaches solving the problem of computing invariant ring use elimination techniques in vector spaces of too high dimensions. Gröbner basis become
impracticable when the number of variables goes up (around 10 for modern computers). The evaluation approach proposed by the author in [3] required a permutation
group whose index in the symmetric group is relatively controlled (around 1000 for
modern computers). Each approaches localize the algebra reduction in vector spaces
still of too large dimensions. Gröbner basis approaches works with monomials of
degree d over n variables and linear reduction over a space spanned by these monomials is costly. The evaluation approach, proposed by the author in his thesis, does
the linear algebra in a free module spanned by the cosets of the symmetric group by
a permutation group (i.e. it have the index for dimension). In both case, as linear
reduction globally cost the cube of the dimension of the space, one cannot hope to
go so much further with classical approaches even with the progress of computer.
We propose in this article an approach following the idea that adding more
combinatorics in invariant theory help to produce more efficient algorithms whose
outputs could perhaps reveal some combinatorics also ; the long time goal being
1
2
NICOLAS BORIE
having a combinatorial description of invariant ring (generators or couple of primary
secondary invariants families). Since Hilbert, This problem have been solve only
in very restrictive and special cases (for example, [9] give secondary invariants for
Young subgroups of symmetric groups). We focus on the problem of computing secondary invariants of finite permutation groups in the non modular case. Assuming
that, we will shows how to localize computations inside selected irreducible representations of the symmetric group. These spaces are smaller than the ones used in
classical approaches and we can largely take advantages of the combinatorial results
coming from the theory of representations of the symmetric group.
2. Invariant ring and representations of the symmetric group
For any set S, we will denote |S| the cardinality of the set S.
2.1. Invariant ring of permutation group and application to combinatorics. Our approach start from a result in one key article of invariant theory
written by Stanley [18, proposition 4.9] mixing invariant of finite group and combinatorics. We recall this general result:
Proposition 2.1. Let θ1 , . . . , θn be an homogeneous set of parameters for K[x]G ,
where G is any finite subgroup of GL(Kn ) of order |G|. Set di = deg(θi ) and
t = d1 . . . dn /|G|. Then the action of G on the quotient ring S = K[x]/(θ1 , . . . , θn )
is isomorphic to t times the regular representation of G.
Applying this result to Sn the symmetric group of degree n with θi = ei (elementary symmetric polynomial)
Y
X
(
xi ),
(1)
ei :=
P ⊂{1,...,n} i∈P
we recover the well known result that the ring K[x]/(e1 , . . . , en ) is isomorphic the
n!
= 1). This
regular representation RR(Sn ) of the symmetric group (here t = n!
well known quotient K[x]/(e1 , . . . , en ) is called the coinvariant ring of the symmetric
group in the algebraic combinatorics world and several basis of this ring have been
explicitly built (Harmonic polynomials, Schubert Polynomials, Descents monomials
and more [2, 15, 9]).
(2)
K[x]/(e1 , . . . , en ) ∼Sn RR(Sn )
Let G a group of permutations, subgroup of Sn , we know reapply the result
of Stanley to G with the same homogeneous set of parameters formed with the
elementary symmetric polynomials. Then, the ring of coinvariant of the symmetric
group is now also isomorphic to t = n!/|G| time the regular representation of the
group G.
n!/|G|
(3)
K[x]/(e1 , . . . , en ) ∼G
M
RR(G)
i=1
We know that for any permutation group G, in the non modular case, the ring
of invariant under the action of G is a Cohen-Macaulay algebra. This imply that
there exist a family of generator ηi making the ring of invariant under the action
of G a free module of rank n!/|G| over the ring of symmetric polynomials.
EFFECTIVE INVARIANT THEORY USING REPRESENTATION THEORY
3
n!/|G|
(4)
K[x]G =
M
ηi K[e1 , . . . , en ]
i=1
Taking the quotient on both side by the ideal (e1 , . . . , en ) and keeping ηi as
representative of its equivalent class in the quotient, we have
n!/|G|
(5)
K[x]G /(e1 , . . . , en ) =
M
K · ηi
i=1
As K[x]G is, by definition, the subspace of K[x] on which the action of G is
trivial, the result of Stanley imply in particular that the polynomials ηi span the
subspace of the coinvariant of the symmetric group on which the action of G is
trivial. A way to construct the ηi is thus to search them as G-fixed point inside
the ring of coinvariant of the symmetric group and that could be done irreducible
representation by irreducible representation.
The theory of representations of the symmetric group has been largely studied,
this bring us to formulate the following problem:
Problem 2.2. Let n a positive integer and G a permutation group, subgroup of
Sn . Construct an explicit basis of the trivial representations of G appearing in
each irreducible subrepresentation of Sn inside the quotient K[x]/(e1 , . . . , en ).
A first step to solve this problem constitute in having a basis of the coinvariant
for the symmetric group respecting the action of Sn such that it can be partitioned
by irreducible representations. We expose this basis in the next section.
2.2. Representations of the symmetric group. We recall in this section some
results describing the irreducible representations of the symmetric group.
For a positive integer n, we will call λ partition of n (denoted λ ⊢ n) a non
increasing sequence of integers λ = (λ1 , . . . , λr ) whose entries sum up to n.
sage:
sage:
sage:
P = Partitions(4); P
Partitions of the integer 4
P.list()
[[4], [3, 1], [2, 2], [2, 1, 1], [1, 1, 1, 1]]
Partitions(8).random_element()
[3, 2, 2, 1]
For a positive integer n, irreducible representations of the symmetric group Sn
are indexed by the Partitions of n. Since we have a finite group, the multiplicity of
an irreducible representation inside the regular representation equal its dimension.
This information can be collected studying standard tableaux.
Let n a positive integer and λ = (λ1 , . . . , λr ) a partition of n. A tableau of shape
λ is a diagram of square boxes disposed in raw such that the first raw contains λ1
boxes, on top of its, a second raw contains λ2 boxes and so on. A standard tableau
of shape λ is a filled tableau of shape λ with integer from 1 to n such that integers
are increasing in column and raw. We will denote ST ab(λ) the set of standard
tableaux of shape λ.
We can ask Sage to display a such object.
4
NICOLAS BORIE
sage:
T = StandardTableaux([3,2,2,1]).random_element(); T
[[1, 2, 3], [4, 5], [6, 8], [7]]
sage:
latex(T)
7
6 8
4 5
1 2 3
We can also iterate or generate all tableaux of a given shape.
sage:
S = StandardTableaux([2,2,1]); S
sage:
for T in S: latex(T)
3
2 5
1 4
4
2 5
1 3
Standard tableaux of shape [2, 2, 1]
4
3 5
1 2
5
2 4
1 3
5
3 4
1 2
The number f λ of standard tableaux of a given shape λ can be easily computed
using the hook-length formula [2, formula 2.1]. As standard tableaux of shape λ
constitute a basis indexing of the vector space associated to the irreducible representation of the symmetric group indexed by λ, and because this same representation
must have for multiplicity its dimension inside the regular representation of Sn , we
have
X
(6)
(f λ )2 = |Sn | = n!
λ⊢n
The following computation illustrate this equality and check that hook-length
formula is well implemented in Sage.
sage:
....:
....:
....:
....:
sage:
def check_multiplicities(n):
dim = 0
for p in Partitions(n):
dim = dim + StandardTableaux(p).cardinality()**2
return dim
for i in range(1, 9) : print i, " : ", check_multiplicities(i)
1 : 1
2 : 2
3 : 6
4 : 24
5 : 120
6 : 720
7 : 5040
8 : 40320
We know describe a last useful object for our algorithmic to come which gather
information about all irreducible representations of the symmetric group : the
character table.
We recall that character of a representation is a map which associate the trace
of matrices for each group element. Such map are constant on conjugacy classes
and conjugacy classes of the symmetric group are indexed also by partitions. Any
permutation τ ∈ Sn have a single disjoint cycles representation and it belong to
EFFECTIVE INVARIANT THEORY USING REPRESENTATION THEORY
5
the conjugacy class indexed by the partition µ = (µ1 , . . . , µr ) if its disjoint cycles
representation contains r cycles of size respectively µ1 , µ2 , . . . , µr .
The character table of the symmetric group gather in a square matrices the
value of characters of irreducible representations on each conjugacy classes of the
symmetric group.
sage:
G = SymmetricGroup(3); G
sage:
G.character_table()
Symmetric group of order 3! as a permutation group
1
2
1
sage:
....:
−1
0
1
1
−1
1
for C in G.conjugacy_classes():
Permutation(C.representative()).cycle_type()
[ 1, 1, 1 ]
[ 2, 1 ]
[ 3 ]
3. Higher Specht polynomials for the symmetric group
Algorithmic in invariant theory must, at some point, construct invariant polynomials. Most current approaches use the Reynolds operator or an orbit sum over a
group of a special monomial. When the group become large, such invariant become
very large even they are stored in a sparse manner inside a computer, the number
of terms can easily fit with n! when G is a permutation group with a small index
in Sn .
For our approach, with focuses on the combinatorics of the quotient K[x]/(e1 , . . . , en ),
the higher specht polynomials will constitute the perfect family to get explicit answer to Problem 2.2.
The quotient K[x]/(e1 , . . . , en ) is isomorphic to the regular representation of Sn
in which we have several copies of irreducible representation following their dimension. The Specht polynomials, which are associated standard tableaux, allows to
construct explicit subspace of K[x] isomorphic to an irreducible representation of
the symmetric group. For λ ⊢ n a partition, the span of Specht polynomials associated to standard tableaux of shape λ realize explicitly the irreducible representation
of Sn indexed by λ. Now, we will see that the higher Specht polynomial take care
of multiplicities of irreducible representation inside the coinvariant. They are indexed by pair of standard tableaux of the same shape and they constitute a basis
of the Sn -module K[x]/(e1 , . . . , en ). Among all known basis of the coinvariants
for the symmetric group (Harmonic, Schubert, monomials under the staircase, descents monomials, ...), the higher Specht polynomial constitute a basis which can
be partitioned by irreducible Sn -module by construction.
let λ ⊢ n a partition and S, T two standard tableaux of shape λ. We define the
word w(S) by reading the tableau S from the top to the bottom in consecutive
columns, starting from the left. The number 1 in the word w(S) as for index 0.
Now, recursively, if the number k in the word has index p, then k + 1 has index
p + 1 if it lies to the left of k in the word, it has index p otherwise. For example,
6
NICOLAS BORIE
with the two tableaux
T = 2 4
S= 3 5
1 2 4
1 3 5
The reading of the Tableau S give 31524, now placing step by step the indices, we
get
3 10 5 2 4
initialization
3 10 5 20 4
right : 0 → 0
31 10 5 20 4
lef t : 0 → 1
31 10 5 20 41 right : 1 → 1
w(S) = 31 10 52 20 41 lef t : 1 → 2
Filling the index in corresponding cell of the tableau S, We obtain i(S), the index
tableau of S.
i(S) = 1 2
0 0 1
Now, using the tableaux T and i(S), cells of T giving variable index and correi(S)
sponding cell of i(S) giving exponent, we build monomials xT as follow.
T = 2 4
1 3 5
i(S)
xT
i(S) = 1 2
0 0 1
= x01 x12 x03 x24 x15
Here are all the monomials in three variables.
i(S)
S
T
i(S)
xT
1 2 3
1 2 3
0 0 0
1
3
1 2
3
1 2
1
0 0
x3
3
1 2
2
1 3
1
0 0
x2
2
1 3
3
1 2
1
0 1
x2 x3
2
1 3
2
1 3
1
0 1
x2 x3
3
2
1
3
2
1
2
1
0
x2 x23
For T a standard tableaux of shape λ, let R(T ) and C(T ) denote the row stabilizer and the column stabilizer of T respectively and consider the Young symmetrizer
X
X
sign(τ )τ σ
(7)
ǫT :=
σ∈R(T ) τ ∈C(T )
which is an element of the group algebra Q[Sn ]. We know define the polynomial
FTS by
(8)
i(S)
FTS (x1 , . . . , xn ) := ǫT (xT
).
EFFECTIVE INVARIANT THEORY USING REPRESENTATION THEORY
7
Theorem 3.1. Let n a positive integer, the family of n! polynomials FTS for S, T
running over standard tableaux of the same shape form a basis of the Sym(x)module K[x].
Terasoma and Yamada proved it using the usual bilinear form in its context : the
divided difference associated to the longest element of the symmetric group [21].
In Three variable, here is the basis of K[x1 , x2 , x3 ] as a Sym(x1 , x2 , x3 )-module.
i(S)
xT
FTS
S
T
1 2 3
1 2 3
1
6
3
1 2
3
1 2
x3
2(x3 − x1 )
3
1 2
2
1 3
x2
2(x2 − x1 )
2
1 3
3
1 2
x2 x3
x2 (x3 − x1 )
2
1 3
2
1 3
x2 x3
x3 (x2 − x1 )
3
2
1
3
2
1
x2 x23
(x3 − x1 )(x3 − x2 )(x2 − x1 )
We will know try to solve Problem 2.2 by searching linear combination of higher
Specht polynomials stabilized by the action of a permutation group.
4. Combinatorial description of the invariant ring
We now try to slice the invariant ring finer than degree by degree. As irreducible
representations of the symmetric group are homogeneous, we will build format series
mixing degree statistic and partitions.
4.1. A refinement of the Moliens series. Let G ⊂ Sn a permutation group.
Any module Sn -stable is also G-stable, thus any representation of Sn is also a
representation of G. Usually, An irreducible representation of Sn will not stay
irreducible when restricted to G. We are searching trivial representations of G
inside irreducible representation of Sn and that can be done with a scalar product
of character.
For G a permutation group, we will denote C(G) the set of conjugacy classes of
G. The usual scalar product between two characters χ and ψ of G is given by
1 X
(9)
hχ, ψi =
|C|χ(σ)ψ(σ)
(σ chosen arbitrary ∈ C)
|G|
C∈C(G)
Proposition 4.1. Let λ ⊢ n a partition of the positive integer n. Let G ⊂ Sn
a permutation group. The multiplicity of the trivial representation of G inside the
irreducible representation of Sn indexed by λ is given by mλ (G, Sn ) with
1 X
|C|M(λ,cycle type(σ)) ,
(σ chosen arbitrary ∈ C)
(10) mλ (G, Sn ) :=
|G|
C∈C(G)
8
NICOLAS BORIE
where M(λ,cycletype(σ)) is the coefficient of the character table of Sn indexed by the
partitions λ and cycle type(σ).
Proof. This just consist in using the usual scalar product of characters for G with
the trivial character of G. Thus we can remark that value of characters can be read
on the character table of Sn because conjugacy classes of G are subset of conjugacy
classes of Sn and traces of matrices do not change when a representation of Sn is
viewed as a representation of G.
Definition 4.2. Let G ⊂ Sn a permutation group. Using a formal set of variable
t = (tλ )λ⊢n indexed by partitions of n, we define the trivial multiplicities enumerator P (G, t) as follow
X
(11)
P (G, t) :=
mλ (G, Sn )tλ
λ⊢n
P (G, t) count the multiplicities of the trivial representation of G inside the irreducible representations of the symmetric group of degree n themselves indexed by
the partitions of the integer n.
For G = h(1, 2)(3, 4), (1, 4)(2, 3)i.
sage:
G = PermutationGroup([[(1,2),(3,4)],[(1,4),(2,3)]]); G
Permutation Group with generators [(1,2)(3,4), (1,4)(2,3)]
sage:
....:
....:
....:
for C in G.conjugacy_classes():
card = C.cardinality()
ct = Permutation(C.representative()).cycle_type()
print card, ct
1 [1, 1,
1 [2,
1 [2,
1 [2,
sage:
1, 1]
2]
2]
2]
trivial_representations_in_symmetric_representations(G)
{[1, 1, 1, 1]:
1, [2, 1, 1]:
0, [2, 2]:
2, [3, 1]:
0, [4]:
1}
For this group, we thus have
P (G, t) = t[1,1,1,1] + 2t[2,2] + t[4]
Definition 4.3. Let λ ⊢ n a partition of a positive integer n and z a formal
variable. The will denote φ(λ, z) the representation appearance polynomial defined
as follow
X
z cocharge(T )
(12)
φ(λ, z) :=
T ∈ST ab(λ)
Where the sum run over all standard tableaux T of shape λ.
φ(λ, z) make the link between the degree z and the irreducible representations
of Sn isomorphic the abstract one indexed by λ appearing inside the quotient
K[x]/Sym+ (x). φ(λ, 1) = |ST ab(λ)| give the multiplicities of the irreducible representation indexed by λ. Generally, if φ(λ, z) has coefficient an integer k for a term
in z d , this means that k Sym(x)-module isomorphic to the irreducible representation of Sn indexed by λ can be built inside the graded quotient K[x]/Sym+ (x)
at degree d. The higher Specht Polynomials realize explicitly these representations
EFFECTIVE INVARIANT THEORY USING REPRESENTATION THEORY
9
because the cocharge is exactly the sum of the entries of tableau i(S) (or the degree
of the corresponding Specht).
Proposition 4.4. Let G ⊂ Sn a permutation group. The trivial multiplicities
enumerator P (G, t) and the Hilbert series H(G, z) are related by
(13)
H(G, z) =
P (G, tλ → φ(λ, z))
(1 − z)(1 − z 2 ) · · · (1 − z n )
Proof. This result is a consequence of some statements about the combinatorics
of standard tableaux. As discussed previously, once we have the dimension of Gtrivial space inside each irreducible representation, it remains to know at which
degree copies of irreducible representation lies into the quotient K[x]/Sym+ (x).
The cocharge of standard tableaux is the right statistic partitioning the occurrences
of Sn -spaces along the degree.
Back with the example G = h(1, 2)(3, 4), (1, 4)(2, 3)i,
sage:
G = PermutationGroup([[(1,2),(3,4)],[(1,4),(2,3)]]); G
Permutation Group with generators [(1,2)(3,4), (1,4)(2,3)]
sage:
G.molien_series()
(x2 − x + 1)/(x6 − 2x5 − x4 + 4x3 − x2 − 2x + 1)
sage:
S4 = SymmetricGroup(4); S4
sage:
S4.molien_series()
Symmetric group of order 4!
as a permutation group
1/(x10 − x9 − x8 + 2x5 − x2 − x + 1)
sage:
G.molien_series() / S4.molien_series()
x6 + 2x4 + 2x2 + 1
On the other side, we had
(14)
P (G, t) = t[1,1,1,1] + 2t[2,2] + t[4]
Let us now list standard tableaux of shape [1, 1, 1, 1], [2, 2]and[4].
S
i(S)
cocharge(S)
1 2 3 4
0 0 0 0
0
3 4
1 2
1 1
0 0
2
2 4
1 3
1 2
0 1
4
4
3
2
1
3
2
1
0
6
We have φ([1, 1, 1, 1], z) = 1, φ([2, 2], z) = z 2 + z 4 and φ([4], z) = z 6 . Injecting these
evaluations inside the trivial multiplicities enumerator, we recover
(15)
P (G, tλ → φ(λ, z)) = 1 + 2z 2 + 2z 4 + z 6
10
NICOLAS BORIE
which give the number of secondary invariants degree by degree when elementary
symmetric polynomials are taken as primary invariants. This polynomials is also the
quotient of the two Hilbert series that can be computed in GAP [8] using Molien’s
Formula.
4.2. Secondary invariants built from higher Specht polynomial. Let G ∈
Sn a permutation group, λ ⊢ n a partition of n. Let us suppose that we have
calculated mλ (G, Sn ) and we now want to build explicitly the secondary invariant polynomials. We are in the case in which we have an homogeneous G-stable
space inside which we want to construct a finite and known number of independent
invariant polynomials under the action of G.
The usual way to dealt with this problem is to built an explicit family spanning
the concerned space by generating polynomials forming a basis. Then basis element
by basis element, we apply the Reynolds operator and some linear algebra to get a
free family of the wanted dimension. knowing this dimension give a stopping criteria
often very important since computations are extremely heavy even for small number
of variables.
In our context, even the usual approach would work, as permutation are often
given by a list of generators, we can even forget the Reynolds operator.
Proposition 4.5. Let G ∈ Sn a permutation group given by some generator : G =
hσ1 , . . . , σr i. Let λ ⊢ n a partition. The G-trivial abstract space inside the abstract
representation of Sn indexed by λ is given by the intersection of the eigenspace of
the representation matrices of σ1 , . . . , σr associated the eigenvalue 1.
Proof. We view here representation indexed by λ as the formal free module generated by the standard tableaux of shape λ. We know that a subspace of dimension
mλ (G, Sn ) have its elements invariant under the action of G. Being invariant under
G is equivalent to be stabilized by the Reynolds operator but is also equivalent (by
definition in fact) to be stabilized by the action of the generators of G. Since we are
working inside a representation of Sn , each permutation have an associated matrix
and the kernel of this matrix characterize the formal subspace stabilized
5. Algorithm building secondary invariants
We now present an effective algorithm exploiting the approach using the representation of the symmetric group.
Computation dependencies :
•
•
•
•
Character table of the symmetric group
Conjugacy classes of the group G (cardinalities and representatives)
Matrices of irreducible representation of the symmetric group
Some linear algebra
The returned set is composed by linear combinations of higher Specht polynomials. These polynomials can be easily evaluated but, as they contains a lot of
Vandermonde factors, there expansion on a set of n formal variables is an heavy
computation.
EFFECTIVE INVARIANT THEORY USING REPRESENTATION THEORY
11
Compute secondary invariant using Sn representations Input : σ1 , σ2 , . . . σn a set
of permutations of size n generating a group G.
def SecondaryInvariants(σ1, σ2 , . . . σn ) :
G ← P ermutationGroup(σ1 , σ2 , . . . σn )
Sn ← SymmetricGroup(n)
S ← {}
for λ ∈ P artition(n) :
if mλ (G, Sn ) 6= 0 :
G ← V ectorSpace(Q, fλ )
for i ∈ {1, 2, . . . , n} :
G ← G ∩ kernel(Mλ (σi ) − Id)
abstract secondary ← basis(G)
for µ ∈ StandardT ableaux(λ) :
for P ∈ abstract secondary :
S ← S ∪ HigherSpecthP olynomials(P, µ)
return S
5.1. A large trace of the algorithm. A teen years computational challenge
consist in computing a generating family of the ring of invariants of the group acting
on the edges of graphs over
5 nodes. This group is a subgroup of the symmetric
group of degree 10 = 52 and has for cardinality 5! = 120. As far as we know, no
computer algebra system has already handle a such computation. After 12 hours
of computation, Magma, Singular and the evaluation approach written in Sage did
not finish (in Sage, more precisely, around 5 percent of this computation(linear
algebra) was done after 24 hours). We tried our approach on this group and we got
the following verbose :
The trace have should be read has the following pattern
[partition] ambient dimension --> (number of standard tableaux for this shape)
rank in S_n repr : (dimension of the G-trivial space)
sage: load("invariants.py")
sage: G = TransitiveGroup(10,12)
sage: Specht_basis_of_trivial_representations(G, verbose=True)
[3, 2, 2, 1, 1, 1] ambient dimension --> 315
rank in S_n repr : 2
[6, 1, 1, 1, 1] ambient dimension --> 126
rank in S_n repr : 3
[6, 4] ambient dimension --> 90
rank in S_n repr : 3
[4, 3, 1, 1, 1] ambient dimension --> 525
rank in S_n repr : 5
[5, 2, 2, 1] ambient dimension --> 525
rank in S_n repr : 4
[5, 2, 1, 1, 1] ambient dimension --> 448
rank in S_n repr : 4
[3, 2, 1, 1, 1, 1, 1] ambient dimension --> 160
rank in S_n repr : 1
[6, 2, 1, 1] ambient dimension --> 350
rank in S_n repr : 2
[4, 4, 2] ambient dimension --> 252
rank in S_n repr : 5
12
NICOLAS BORIE
[2, 1, 1, 1, 1, 1, 1, 1, 1] ambient dimension --> 9
rank in S_n repr : 1
[3, 3, 2, 2] ambient dimension --> 252
rank in S_n repr : 2
[4, 4, 1, 1] ambient dimension --> 300
rank in S_n repr : 2
[4, 2, 2, 2] ambient dimension --> 300
rank in S_n repr : 5
[4, 2, 2, 1, 1] ambient dimension --> 567
rank in S_n repr : 3
[2, 2, 2, 1, 1, 1, 1] ambient dimension --> 75
rank in S_n repr : 2
[3, 2, 2, 2, 1] ambient dimension --> 288
rank in S_n repr : 3
[4, 2, 1, 1, 1, 1] ambient dimension --> 350
rank in S_n repr : 3
[3, 1, 1, 1, 1, 1, 1, 1] ambient dimension --> 36
rank in S_n repr : 1
[7, 1, 1, 1] ambient dimension --> 84
rank in S_n repr : 1
[5, 1, 1, 1, 1, 1] ambient dimension --> 126
rank in S_n repr : 3
[2, 2, 2, 2, 2] ambient dimension --> 42
rank in S_n repr : 3
[6, 3, 1] ambient dimension --> 315
rank in S_n repr : 1
[8, 2] ambient dimension --> 35
rank in S_n repr : 2
[3, 3, 3, 1] ambient dimension --> 210
rank in S_n repr : 2
[3, 3, 1, 1, 1, 1] ambient dimension --> 225
rank in S_n repr : 1
[5, 4, 1] ambient dimension --> 288
rank in S_n repr : 3
[5, 3, 2] ambient dimension --> 450
rank in S_n repr : 3
[10] ambient dimension --> 1
rank in S_n repr : 1
[4, 3, 2, 1] ambient dimension --> 768
rank in S_n repr : 6
[7, 2, 1] ambient dimension --> 160
rank in S_n repr : 1
[6, 2, 2] ambient dimension --> 225
rank in S_n repr : 3
[5, 3, 1, 1] ambient dimension --> 567
rank in S_n repr : 5
[3, 3, 2, 1, 1] ambient dimension --> 450
rank in S_n repr : 4
total : 30240
n! / |G| : 30240
TOTAL CPU TIME : 414.837207
Our algorithm took 414 seconds to generated the 30240 secondary invariants as
linear combinations of higher Specht polynomials. We still believe that the computation of a couple primary and secondary invariants for this group is unreachable
for Magma, Singular and the evaluation approach in less than 24 hours.
EFFECTIVE INVARIANT THEORY USING REPRESENTATION THEORY
13
For example, inside the symmetric representation associated with the partition
(4, 3, 2, 1). The algorithm making the Gauss reduction build a space of dimension 6
inside an ambient space of dimension 768. Using the Hook-length formula, on can
check that there exist 768 standard tableaux of shape (4, 3, 2, 1)
(16)
1
3 1
5 3 1
7 5 3 1
10!
= 768
7·5·5·3·3·3
5.2. Implementation details. Our approach involves a lot a technology around
representation theory and symbolic computation, most of the required prerequisites
are available in current computer algebras system. For our problem, dependencies
have been implemented in GAP [8] or Sage [19]. We implemented a first version in
Sage (400 lines of code with tests and documentation) to test the efficienty of this
approach.
Here are the main steps of computation.
• Compute the character table of the symmetric group.
This task is done by GAP [8]. The Murnaghan-Nakayama rule is an example
of working algorithm. This is a small part of computations and it can also
be precomputed and store for all small symmetric group (for degree at most
20).
• Enumerate cardinality and a representative of each conjugacy class of G.
Also handled by GAP [8] and interfaced in Sage [19]. The general problem
of enumerating conjugacy classes of finite group is not simple but there are
efficient algorithms for permutation group.
• Compute the trivial multiplicities enumerator of G.
We computed a short function which just iterate over all partitions and
compute a simple scalar product (see Formula 10).
• Calculate the matrices of σ1 , σ2 , . . . , σn inside the irreducible representations of Sn .
Once we have identified a G-trivial space inside an irreducible of Sn , we
build the matrices of the generators of G using the Sage [19] object : SymmetricGroupRepresentation which admits as argument a partition. The
returned object is able to build matrices of the abstract representation from
a permutation given in line notation.
• Compute the intersection of stabilized subspace of these matrices.
We use in a loop the method intersection of Sage VectorSpace and the
method kernel of Sage Matrix class. Our linear algebra is thus handle by
Sage [19].
• Transcript abstract combinations of standard tableaux in term of higher
Specht polynomials.
Sage and especially Sage-Combinat [17] contains the combinatorial family
of standard tableaux. A lot of combinatorial statistic are availlable on such
objects like the cocharge. We implemented the higher Specht polynomials
as Sage polynomials computed from a pair of Sage StandardTableaux of
the same shape.
Our code is completely not optimized. Since most of computations in invariants
(minimal generating set, secondary invariants, Hironaka decomposition) present
14
NICOLAS BORIE
large theoretical exponential complexity, we believe that, before code source refinement, we need new algorithms with better asymptotically behavior. However, a
huge factor of execution time can easily be wined from our code with cache remember, better interface between GAP and Sage, parralelisation (each computer
core can echelonize matrices for a special irreducible representation), we wanted to
benchmarks our approach using the current available (and open source) technology.
Therefore, anyone (even a student) can reproduce our experience rapidly and for
free.
5.3. Complexity. The rich literature about effective invariant theory does not
provide a lot a fin complexity bound for algorithms. Gröbner basis admit very
O(n)
general complexity bounds (in worst case 22
for n variables) which appears to
be overestimated compared to their effective behavior. The thesis [3] of the author
present an evaluation approach to compute the invariants inside a quotient of a
reduced dimension. The algorithm computing secondary with this technique has a
3
complexity in O((n!)2 + (n!)
|G|2 ). Using the representation of the symmetric group, it
is still very hard to establish a fin bound. However, we can produce better bounds.
Theorem 5.1. Let G a permutation group, subgroup of Sn , given by r generators.
The complexity of the linear algebra reduction in algorithm computing the secondary
invariants of G in Section 5 is bounded by
X
(17)
r·(
mλ (|G|, Sn )(fλ )2 )
λ⊢n
where fλ is the number of standard tableaux of shape λ and p(n) is the number of
partition of size n.
Proof. This is just a straightforward counting of the reductions the r matrices of
permutation for each irreducible representation of the symmetric group. For each
partition λ of n, we have to construct a free family of mλ (|G|, Sn ) vectors (i.e. the
rank) inside a space of dimension f λ (i.e. number of indeterminates) from r · f λ
equations (the vertical concatenation of the r matrices). Summing these operations
give the announced bound.
Corollary 5.2. Let G a permutation group, subgroup of Sn , given by r generators.
The complexity of the algorithm computing the secondary invariants in Section 5
3
has a complexity in O(r · (n!) 2 ).
Proof. Let us denote fmax := max{f λ }. We thus have
λ⊢n
(18)
r·(
X
mλ (|G|, Sn )(fλ )2 ) 6 r · (
λ⊢n
X
fmax (fλ )2 )
λ⊢n
since, at worse, the irreducible representation of Sn is composed only by element
stabilized point wise by G. The fmax term can go out the sum and using formula 6
we get the bound
(19)
r · fmax · n!
2
Using the triangular inequality, we roughly have fmax
6 n! and that give the
result.
EFFECTIVE INVARIANT THEORY USING REPRESENTATION THEORY
15
6. Further developments
Even the impressive algorithmic efficiently of this approach using the representation of the symmetric group, it has also the advantage of putting a lot of combinatrics inside this problem often classed inside effective algebraic geometry.
acknowledgements
This research was driven by computer exploration using the open-source mathematical software Sage [19]. In particular, we perused its algebraic combinatorics
features developed by the Sage-Combinat community [17], as well as its group
theoretical features provided by GAP [8].
References
[1] I. Abdeljaouad. Théorie des Invariants et Applications à la Théorie de Galois effective. PhD
thesis, Université Paris 6, 2000.
[2] F. Bergeron. Algebraic combinatorics and coinvariant spaces. CMS Treatises in Mathematics.
2009.
[3] N. Borie. Calcul des invariants des groupes de permutations par transformée de Fourier. PhD
thesis, Laboratoire de Mathématiques, Université Paris Sud, 2011.
[4] A. Colin. Solving a system of algebraic equations with symmetries. J. Pure Appl. Algebra,
117/118:195–215, 1997. Algorithms for algebra (Eindhoven, 1996).
[5] A. Colin. Théorie des invariants effective; Applications à la théorie de Galois et à la résolution de systèmes algébriques; Implantation en AXIOM. PhD thesis, École polytechnique,
1997.
[6] H. Derksen and G. Kemper. Computational invariant theory. Springer-Verlag, Berlin, 2002.
[7] J. Faugère and S. Rahmany. Solving systems of polynomial equations with symmetries using
SAGBI-Gröbner bases. In Proceedings of the 2009 international symposium on Symbolic and
algebraic computation, pages 151–158, 2009.
[8] The GAP Group, Lehrstuhl D für Mathematik, RWTH Aachen, Germany and SMCS, U. St.
Andrews, Scotland. GAP – Groups, Algorithms, and Programming, 1997.
[9] A. Garsia and D. Stanton. Group actions on stanley-reisner rings and invariants of permutation groups. Advances in Mathematics, 51(2):107 – 201, 1984.
[10] K. Gatermann. Symbolic solution of polynomial equation systems with symmetry. KonradZuse-Zentrum für Informationstechnik Berlin, 1990.
[11] K. Geissler and J. Klüners. Galois group computation for rational polynomials. J. Symbolic
Comput., 30(6):653–674, 2000. Algorithmic methods in Galois theory.
[12] G. Kemper. The invar package for calculating rings of invariants. IWR Preprint 93-94, University of Heidelberg, 1993.
[13] S. King. Fast Computation of Secondary Invariants. Arxiv preprint math/0701270, 2007.
[14] S. King. Minimal generating sets of non-modular invariant rings of finite groups. Arxiv
preprint math/0703035, 2007.
[15] A. Lascoux and M.-P. Schützenberger. Polynômes de Schubert. C. R. Acad. Sci. Paris Sér.
I Math., 294(13):447–450, 1982.
[16] M. Pouzet and N. M. Thiéry. Invariants algébriques de graphes et reconstruction. C. R. Acad.
Sci. Paris Sér. I Math., 333(9):821–826, 2001.
[17] T. Sage-Combinat community. Sage-Combinat: enhancing Sage as a toolbox for computer
exploration in algebraic combinatorics, 2008.
[18] R. P. Stanley. Invariants of finite groups and their applications to combinatorics. Bull. Amer.
Math. Soc. (N.S.), 1(3):475–511, 1979.
[19] W. Stein et al. Sage Mathematics Software (Version 3.3). The Sage Development Team,
2009. http://www.sagemath.org.
[20] B. Sturmfels. Algorithms in invariant theory. Springer-Verlag, Vienna, 1993.
[21] T. Terasoma and H. Yamada. Higher Specht polynomials for the symmetric group. Proc.
Japan Acad. Ser. A Math. Sci., 69(2):41–44, 1993.
[22] N. M. Thiéry. Algebraic invariants of graphs: a study based on computer exploration.
SIGSAM Bulletin, 34(3):9–20, Sept. 2000.
16
NICOLAS BORIE
[23] N. M. Thiéry. Computing minimal generating sets of invariant rings of permutation groups
with SAGBI-Gröbner basis. In Discrete models (Paris, 2001), pages 315–328 (electronic).
2001.
Univ. Paris Est Marne-La-Vallée, Laboratoire d’Informatique Gaspard Monge, Cité
Descartes, Bât Copernic – 5, bd Descartes Champs sur Marne 77454 Marne-la-Vallée
Cedex 2, France
| 0math.AC
|
Zero-Shot Deep Domain Adaptation
Kuan-Chuan Peng, Ziyan Wu, Jan Ernst
Siemens Corporate Technology
Princeton, NJ, USA
arXiv:1707.01922v2 [cs.CV] 16 Nov 2017
{kuanchuan.peng, ziyan.wu, jan.ernst}@siemens.com
Abstract
The existing methods of domain adaptation (DA) work
under the assumption that the task-relevant target-domain
training data is given. However, such assumption can be
violated, which is often ignored by the prior works. To
tackle this issue, we propose zero-shot deep domain adaptation (ZDDA), which uses the privileged information from the
task-irrelevant dual-domain pairs. ZDDA learns a sourcedomain representation which is not only tailored for the task
of interest (TOI) but also close to the target-domain representation. Therefore, the source-domain TOI solution (e.g.
the classifier for classification tasks) which is jointly trained
with the source-domain representation can be applicable
to both the source and target representations. Using the
MNIST, Fashion-MNIST, NIST, EMNIST, and SUN RGB-D
datasets, we show that ZDDA can perform DA in classification tasks without the access to the task-relevant targetdomain training data. We also extend ZDDA to perform
sensor fusion in the SUN RGB-D scene classification task
by simulating the task-relevant target-domain representations with the task-relevant source-domain data. To the best
of our knowledge, ZDDA is the first DA and sensor fusion
method which needs no task-relevant target-domain data.
Figure 1: We propose zero-shot deep domain adaptation (ZDDA)
for domain adaptation and sensor fusion. ZDDA learns from the
task-irrelevant dual-domain pairs when the task-relevant targetdomain training data is unavailable. In this example domain adaptation task (MNIST [25]→MNIST-M [12]), the task-irrelevant
gray-RGB pairs are from the Fashion-MNIST [41] dataset and
the Fashion-MNIST-M dataset (the colored version of the FashionMNIST [41] dataset with the details in Sec. 4.1).
The goal of a DA task is to derive solution(s) of the TOI for
both the source and target domains.
The state-of-the-art DA methods such as [5, 13–15, 23,
26, 31, 33, 35–39, 42, 45] are proposed to solve DA tasks
under the assumption that the task-relevant data, the data
directly applicable and related to the TOI (regardless of
whether it is labeled or not), in the target domain is available at training time, which is not always true in practice.
For instance, in real business use cases, acquiring the taskrelevant target-domain training data can be infeasible due
to the combination of the following reasons: 1) Unsuitable tools at the field. 2) Product development timeline. 3)
Budget limitation. 4) Data import/export regulations. Such
impractical assumption is also assumed true in the existing
works of sensor fusion such as [27, 43], where the goal is
to obtain a dual-domain (source and target) TOI solution
which is robust to noise in either domain. This unsolved
issue motivates us to propose zero-shot deep domain adap-
1. Introduction
The useful information to solve practical tasks often
exists in different domains captured by various sensors,
where a domain can be either a modality or a dataset. For
instance, the 3-D layout of a room can be either captured
by a depth sensor or inferred from the RGB images. In
real-world scenarios, it is highly likely that we can only
access limited amount of data in certain domain(s). The
performance of the solution (e.g. the classifier for classification tasks) we learn from one domain often degrades
when the same solution is applied to other domains, which
is caused by domain shift [16] in a typical domain adaptation (DA) task, where source-domain training data, targetdomain training data, and a task of interest (TOI) are given.
1
tation (ZDDA), a DA and sensor fusion approach which
learns from the task-irrelevant dual-domain training pairs
without using the task-relevant target-domain training data,
where we use the term task-irrelevant data to refer to the
data which is not task-relevant.
We illustrate what ZDDA is designed to achieve in
Figure 1 using an example DA task (MNIST [25]→MNISTM [12]). We recommend that the readers view all the
figures and tables in color. In Figure 1, the source and
target domains are gray scale and RGB images respectively, and the TOI is digit classification with both the
MNIST [25] and MNIST-M [12] testing data. We assume
that the MNIST-M [12] training data is unavailable. In this
example, ZDDA aims at using the MNIST [25] training data
and the task-irrelevant gray-RGB pairs from the FashionMNIST [41] dataset and the Fashion-MNIST-M dataset (the
colored version of the Fashion-MNIST [41] dataset with the
details in Sec. 4.1) to train digit classifiers for MNIST [25]
and MNIST-M [12] images. Specifically, ZDDA achieves
this by simulating the RGB representation using the gray
scale image and building a joint network with the supervision of the TOI in the gray scale domain. We present the
details of ZDDA in Sec. 3.
We make the following contributions: 1) To the best
of our knowledge, our proposed method, ZDDA, is
the first deep learning based method performing domain
adaptation between different image modalities (instead
of just different datasets in the same modality such as
the Office dataset [28]) without using the task-relevant
target-domain training data. We show ZDDA’s efficacy
using the MNIST [25], Fashion-MNIST [41], NIST [17],
EMNIST [10], and SUN RGB-D [32] datasets with cross
validation. 2) Given no task-relevant target-domain
training data, we show that ZDDA can perform sensor
fusion and that ZDDA is more robust to noisy testing data
in either source or target or both domains compared with a
naive fusion approach in the scene classification task from
the SUN RGB-D [32] dataset.
can be unavailable in reality. In contrast, we propose
ZDDA to learn from the task-irrelevant dual-domain pairs
without using the task-relevant target-domain training data.
One part of ZDDA includes simulating the target-domain
representation using the source-domain data, and similar
concepts have been mentioned in [18, 20]. However, both
of [18, 20] require the access to the task-relevant dualdomain training pairs, but ZDDA needs no access to the
task-relevant target-domain data.
The name “zero-shot domain adaptation” has already
been used by Blitzer et al. [7] and Yang et al. [44]. However,
both methods require additional prerequisites. Blitzer’s
method [7] needs access to the task-relevant target-domain
data. Requiring that the domain descriptor is always
observed and accurate, Yang’s method [44] shows its efficacy using only different datasets in the same modality
(specifically, the Office dataset [28]) instead of the data
in different modalities. ZDDA requires none of these two
prerequisites, and we show the efficacy of ZDDA with the
data in different modalities.
In terms of sensor fusion, Ngiam et al. [27] define
the three components for multimodal learning (multimodal
fusion, cross modality learning, and shared representation
learning) based on the modality used for feature learning,
supervised training, and testing, and experiment on audiovideo data with their proposed deep belief network and
autoencoder based method. Targeting on the temporal data,
Yang et al. [43] follow the setup of multimodal learning
in [27], and validate their proposed encoder-decoder architecture using video-sensor and audio-video data. Although
certain progress about sensor fusion is achieved in the
previous works [27, 43], we are unaware of any existing
sensor fusion method which overcomes the issue of lacking
task-relevant target-domain training data, which is the issue
that ZDDA is designed to solve.
3. Our proposed method — ZDDA
Given a task of interest (TOI), a source domain Ds ,
and a target domain Dt , our proposed method, zero-shot
deep domain adaptation (ZDDA), is designed to achieve the
following two goals: 1) Domain adaptation: Derive the
solutions of the TOI for both Ds and Dt when the taskrelevant training data in Dt is unavailable. We assume that
we have access to the task-relevant labeled training data in
Ds and the task-irrelevant dual-domain pairs in Ds and Dt .
2) Sensor fusion: Given the previous assumption, derive
the solution of TOI when the testing data in both Ds and
Dt is available. The testing data in either Ds or Dt can be
noisy. We assume that there is no prior knowledge available
about the type of noise and which domain gives noisy data
at testing time.
For convenience, we use a scene classification task in
RGB-D as an example TOI to explain ZDDA, but ZDDA
2. Related work
Domain adaptation (DA) has been extensively studied in
computer vision and applied to various applications such
as image classification [5, 13–15, 23, 26, 31, 33, 35–
39, 42, 45], semantic segmentation [40, 46], and image
captioning [9]. With the advance of deep neural networks
in recent years, the state-of-the-art methods successfully
perform DA with (fully or partially) labeled [9, 14, 23, 26,
35] or unlabeled [5, 13–15, 31, 33, 35–40, 42, 45] taskrelevant target-domain data. Although different strategies
such as the domain adversarial loss [36] and the domain
confusion loss [35] are proposed to improve the performance in the DA tasks, most of the existing methods
need the task-relevant target-domain training data, which
2
Figure 2: An overview of the ZDDA training procedure. We use the images from the SUN RGB-D [32] dataset for illustration. ZDDA
simulates the target-domain representation using the source-domain data, builds a joint network with the supervision from the source
domain, and trains a sensor fusion network. In step 1, we choose to train s1 and fix t, but we can also train t and fix s1 to simulate the
target-domain representation. In step 2, t can also be trainable instead of being fixed, but we choose to fix it to make the number of trainable
parameters manageable. The details are explained in Sec. 3.
can be applied to other TOIs/domains. In this example, Ds
and Dt are depth and RGB images respectively. According
to the our previous assumption, we have access to the taskrelevant labeled depth data and task-irrelevant RGB-D pairs
at training time. The training procedure of ZDDA is illustrated in Figure 2, where we simulate the RGB representation using the depth image, build a joint network with the
supervision of the TOI in depth images, and train a sensor
fusion network in step 1, step 2, and step 3 respectively.
We use the ID marked at the bottom of each convolutional
neural networks (CNN) in Figure 2 to refer to each CNN.
In step 1, we create two CNNs, s1 and t, to take the
depth and RGB images of the task-irrelevant RGB-D pairs
as input. The purpose of this step is to find s1 and t such
that feeding the RGB image into t can be approximated by
feeding the corresponding depth image into s1. We achieve
this by fixing t and enforcing the L2 loss on top of s1 and
t at training time. We choose to train s1 and fix t here, but
training t and fixing s1 can also achieve the same purpose.
The L2 loss can be replaced with any suitable loss functions
which encourage the similarity of the two input representations, and our selection is inspired by [18, 20]. The design
in step 1 is similar to the hallucination architecture [20]
and the supervision transfer [18], but we require no taskrelevant dual-domain training pairs. Instead, we use the
task-irrelevant dual-domain training pairs.
After step 1, we add another CNN, s2 (with the same
network architecture as that of s1), and a classifier to the
network (as shown in step 2) to learn from the label of the
training depth images. The classifier in our experiment is
a fully connected layer for simplicity, but other types of
classifiers can also be used. The newly added CNN takes
the task-relevant depth images as input, and shares all the
weights with the original source CNN, so we use s2 to
refer to both of them. t is the same as that in step 1. At
training time, we pre-train s2 from s1 and fix t. Our choice
of fixing t is inspired by the adversarial adaptation step in
ADDA [36]. t can also be trainable in step 2, but given
our limited amount of data, we choose to fix it to make
the number of trainable parameters manageable. s2 and the
source classifier are trained such that the weighted sum of
the softmax loss and L2 loss are minimized. The softmax
loss can be replaced with other losses suitable for the TOI.
After step 2, we expect to obtain a depth representation which is close to the RGB representation in the feature
space and performs reasonably well with the trained classifier in the scene classification. Step 1 and step 2 can
be done in one step with properly designed curriculum
learning, but we separate them not only because of clarity
but also because of the difficulty of designing the learning
curriculum before training. After step 2, we can form
the scene classifier in depth/RGB (denoted as CD /CRGB )
by concatenating s2/t and the trained source classifier (as
shown in Figure 3a), which meets our first goal, domain
adaptation. We use the notation ZDDA2 to refer to the
method using the training procedure in Figure 2 up to step
2 and the testing procedure in Figure 3a.
To perform sensor fusion, we propose step 3, where
we train a joint classifier for RGB-D input using only the
task-relevant depth training data. We create two CNNs, s3
and s4 (each with the same network architecture as that of
CN Ns1 ), and add a concatenation layer on top of them to
3
can form three scene classifiers in RGB, depth, and RGB-D
domains (one classifier per domain), and our trained RGBD classifier is expected to be able to handle noisy input with
reasonable performance degradation. We use the notation
ZDDA3 to refer to the method using the training procedure in Figure 2 up to step 3 and the testing procedure in
Figure 3b.
4. Experiment setup
(a) testing source-/target-domain data
4.1. Datasets
For domain adaptation (DA), we validate the efficacy
of ZDDA under classification tasks using the MNIST [25],
Fashion-MNIST [41], NIST [17], EMNIST [10], and SUN
RGB-D [32] datasets. For sensor fusion, we experiment on
the SUN RGB-D [32] dataset. We summarize the statistics of these datasets in Table 1, where we list the dataset
IDs which we use to refer to these datasets. For DM ,
DF , DN , and DE , we create the colored version of these
datasets (DM -M, DF -M, DN -M, and DE -M) according to
the procedure proposed in Ganin’s work [12] — blending
the gray scale images with the patches randomly extracted
from the BSDS500 dataset [6]. These colored datasets
and the original ones are used to construct four DA tasks
adapting from gray scale to RGB images. For each DA task,
we use one of the other three pairs of the datasets (original
and colored ones) as the task-irrelevant data. For example,
for the DA task DM → DM -M, DF and DF -M together
are one possible choice as the task-irrelevant data. The DA
task DM → DM -M is acknowledged as one of the standard
experiments to test the efficacy of the DA methods in recent
works [5, 8, 13, 19, 29, 30], so we adopt this experiment
and extend it to DF , DN , and DE .
DS contains 10335 RGB-D pairs belonging to 45
different scenes. For each RGB-D pair, both the raw
(noisy) depth image and post-processed clean depth image
are provided, and we choose to use the raw depth image to
simulate the real-world scenarios. Out of the 45 scenes, we
select the following 10 scenes: computer room (0), conference room (1), corridor (2), dining room (3), discussion area
(4), home office (5), idk (6), lab (7), lecture theatre (8), and
study space (9), where the number after each scene is the
scene ID we use to refer to each scene. The 8021 RGBD pairs belonging to the other scenes are used as the taskirrelevant training data. The 10 scenes are selected based
on the following two constraints: 1) Each scene contains at
least 150 RGB-D pairs in DS , which ensures a reasonable
amount of task-relevant data. 2) The total number of the
RGB-D pairs belonging to the selected 10 scenes is minimized, which maximizes the amount of the task-irrelevant
training data. We empirically find that the amount and
diversity of the task-irrelevant training data are important
for ZDDA. To avoid the bias toward the scene with more
(b) testing sensor fusion
Figure 3: An overview of the ZDDA testing procedure. We use
the SUN RGB-D [32] images for illustration.
concatenate their output representations. The concatenated
representation is connected to a joint classifier. At training
time, we pre-train s3 and s4 from s2 and s1 respectively and
fix s4. Both s3 and s4 take the task-relevant depth images
as the input. To train a more robust RGB-D scene classifier, we randomly select some inputs of s3 and s4, and
optionally add noise to them independently. We supervise
the entire network with the label of the depth training data
for the scene classification, which is done by the softmax
loss enforced on top of the joint classifier.
According to step 1, the output of s4 is expected to
simulate the RGB representation as if we feed the taskrelevant RGB image to t. This expectation is based on the
assumption that the relationship between the dual-domain
pairwise data is similar, regardless of whether the data is
task-relevant or task-irrelevant. Given the simulated RGB
representation, s3 is trained to learn a depth representation suitable for the RGB-D scene classification without
the constraint of the L2 loss in step 2. At testing time,
s4 is replaced with t which takes the task-relevant RGB
testing images as input with optional noise added to test the
ZDDA’s performance given noisy RGB-D testing data (as
shown in Figure 3b). In Figure 3b, we also test replacing
“RGB images and t” with “depth images and s4” to evaluate the performance of ZDDA in step 3 given only testing
depth images. After the training procedure in Figure 2, we
4
original dataset
dataset ID
classification subject
MNIST [25]
DM
digit
Fashion-MNIST [41]
DF
clothing
NIST [17]
DN
letter
EMNIST [10]
DE
letter
SUN RGB-D [32]
DS
scene
original image size
# classes
# training images
# testing images
class labels
balanced class?
example images
28×28
10
60000
10000
0-9
N
28×28
10
60000
10000
dress, coat, etc.
Y
128×128
52
387361
23941
A-Z, a-z
N
28×28
26
124800
20800
Aa-Zz
Y
∼VGA
45
see Sec. 4.1 for details
see Sec. 4.1 for details
corridor, lab, etc.
N
colored dataset [12]
example images
MNIST-M
Fashion-MNIST-M
NIST-M
EMNIST-M
N/A
N/A
Table 1: The statistics of the datasets we use. For NIST, we use the “by class” dataset, remove the digits, and treat uppercase and lowercase
letters as different classes. For EMNIST, we use the “EMNIST Letters” split which only contains the letters. We create the colored datasets
from the original ones using Ganin’s method [12] (see Sec. 4.1 for details). We refer to each dataset by the corresponding dataset ID (e.g.
DN and DN -M refer to the NIST and the NIST-M datasets, respectively).
wards. For example, when using DF and DF -M as the taskirrelevant data in the DA task DM → DM -M, we train a
CNN (denoted as CN Nref ) with the LeNet [3] architecture from scratch using the images and labels of DF -M,
and pre-train the target CNNs in Figure 2 from CN Nref .
We follow similar procedures for other DA tasks and taskirrelevant datasets involving DM , DF , DN , and DE .
For the experiment involving DS , we mostly use
GoogleNet [34] as the BNA, but we also use AlexNet [24]
and SqueezeNet v1.1 [21] in the cross validation experiment with respect to different BNAs. Since only limited
amount of RGB-D pairs are available in DS , we pre-train all
the CNNs in Figure 2 from the BVLC GoogleNet model [2],
BVLC AlexNet model [1], and the reference SqueezeNet
model [4] when the BNA is GoogleNet [34], AlexNet [24],
and SqueezeNet v1.1 [21], respectively. These pre-trained
models are trained for the ImageNet [11] classification task.
For the optionally added noise in ZDDA3 , we experiment on training/testing with noise-free data and noisy data.
In the latter case, given that no prior knowledge about the
noise is available, we use the black image as the noisy
image to model the extreme case where no information in
the noisy image is available. We train ZDDA3 step 3 with
the augmented training data formed by copying the original task-relevant source-domain training data 10 times and
replacing ptrain % of the images selected randomly with the
black images. We follow this procedure twice independently and use the two augmented training datasets as the
inputs of the two source CNNs in step 3. We empirically set
ptrain = 20. The testing data in Figure 3b is constructed
by replacing ptest % of the original testing images selected
randomly with the black images, and we evaluate ZDDA
under different ptest s. For all the experiments, the number
base network
source/target CNN architecture
architecture (BNA) (up to where in BNA (inclusive))
LeNet [3]
GoogleNet [34]
AlexNet [24]
SqueezeNet v1.1 [21]
ip1
pool5/7x7 s1
fc7
fire9/concat
Table 2: The base network architecture (BNA) we use in our
experiments. For each BNA, We specify the layer separating the
source/target CNN and the source classifier in Figure 2. The layer
name in the right column is based on the official Caffe [22] and
SqueezeNet v1.1 [21] implementation of each BNA.
data, for each of the selected 10 scenes, we randomly select
89/38 RGB-D pairs as the task-relevant training/testing
data. When experimenting on different scene classification
tasks using different selections of scenes, we only use the
training/testing data associated with those selected scenes
as the task-relevant data.
4.2. Training details
We use Caffe [22] to implement ZDDA. Table 2 lists
the base network architecture (BNA) we use and the layer
separating the source/target CNN and the source classifier
in Figure 2. For instance, in the case when the BNA is
LeNet [3], the architecture of each source/target CNN in
Figure 2 is the LeNet [3] architecture up to the “ip1” layer,
and the rest of the LeNet [3] architecture is used as the
source classifier. For the DA tasks involving DM , DF , DN ,
and DE , we use the LeNet [3] as the BNA and train all the
CNNs in Figure 2 from scratch except that the target CNN is
pre-trained from the task-irrelevant dataset and fixed after5
task-irrelevant data
DM → DM -M
→
DF → DF -M
→
DN → DN -M
→
DE → DE -M
→
none (source only)
39.04 / 39.31
33.77 / 33.77
8.59 / 8.79
33.70 / 33.70
DM , DM -M
DF , DF -M
DN , DN -M
DE , DE -M
N/A
73.15 / 72.96
91.99 / 92.00
94.84 / 94.82
51.55 / 51.55
N/A
43.87 / 43.87
65.30 / 65.30
34.25 / 33.35
21.93 / 21.24
N/A
N/A
71.20 / 71.20
46.93 / 46.93
N/A
N/A
target only
97.33 / 97.34
84.44 / 84.44
62.13 / 61.99
89.52 / 89.52
Table 3: The overall / average per class accuracy (%) of the domain adaptation tasks (gray scale images → RGB images) formed by
the datasets in Table 1, where we introduce the dataset IDs and use them to refer to the datasets here. The middle four rows show the
performance of ZDDA2 . The color of each cell reflects the performance ranking in each column, where darker is better.
of the output nodes of the source/joint classifiers is set to be
the number of classes in the TOI, and these classifiers are
trained from scratch. For the joint classifiers, we use two
fully connected layers unless otherwise specified, where the
first fully connected layer of the joint classifier has 1024
output nodes. The visualization of the network architectures
of the source CNN, source classifier, and joint classifier for
each BNA is shown in the supplementary material.
In terms of the training parameters used in Figure 2 for
the task involving DS when the BNA is GoogleNet [34],
we use a batch size of 32 and a fixed learning rate
10−5 /10−6 /10−3 for step 1/2/3. The learning rate is chosen
such that the trained network can converge under a reasonable amount of time. We set the weight of the softmax
loss and the L2 loss in step 2 to be 103 and 1 respectively such that both losses have comparable numerical
values. Step 1/2/3 are trained for 104 /103 /103 iterations.
For the other training parameters, we adopt the default ones
used in training the BVLC GoogleNet model [2] for the
ImageNet [11] classification task unless otherwise specified. The details of the training parameters for all the other
experiments are in the supplementary material. In general,
we adopt the default training parameters used in training
each BNA for either the MNIST [25] or ImageNet [11] classification tasks in the Caffe [22] and SqueezeNet v1.1 [21]
implementation unless otherwise specified.
method
[13]
[30]
[19]
[29]
[8]
accuracy (%) 76.66 86.70 89.53 94.20 98.20
ZDDA2
94.84
Table 4: The performance comparison of the domain adaptation
task MNIST→MNIST-M. For ZDDA2 , we report the best overall
accuracy from Table 3. All the listed methods except ZDDA2 use
the MNIST-M training data. Without the access to the MNISTM training data, ZDDA2 can still achieve the accuracy comparable to those of the competing methods (even outperform some
of them) in this task.
the performance without applying any DA method. For the
baseline of sensor fusion, we compare ZDDA3 with a naive
fusion method by predicting the label with the highest probability from CRGB and CD in Sec. 3.
5. Experimental result
We first compare ZDDA2 with the baseline in four
domain adaptation (DA) tasks (adapting from gray scale to
RGB images) involving DM , DF , DN , and DE , and the
result is summarized in Table 3, where the numbers represent the overall/average per class accuracy (%). Darker
cells in each column represent better classification accuracy in each task. In Table 3, the middle four rows represent the performance of ZDDA2 . {DN , DN -M} and {DE ,
DE -M} cannot be the task-irrelevant data for each other
because they are both directly related to the letter classification tasks. Table 3 shows that regardless of which
task-irrelevant data we use, ZDDA2 significantly outperforms the baseline (source only). We find that the similarity
between the DA task and the task-irrelevant data is related to
ZDDA2 ’s performance improvement over the baseline. For
example, in the DA task DM → DM -M, using {DE , DE M} as the task-irrelevant data outperforms using {DF , DF M}, which is consistent with our intuition that letters are
semantically more similar to digits compared with clothing.
Second, in Table 4, we compare ZDDA2 with the
4.3. Performance references and baselines
To obtain the performance references of the fully supervised methods, we train a classifier with the BNA in Table 2
in each domain using the task-relevant training data and
labels in that domain. When the BNA is LeNet [3], we train
the classifier from scratch. For the other BNAs, we pre-train
the classifier in the same way as that described in Sec. 4.2.
After training, for each DA task, we get two fully supervised
classifiers Cf s,s and Cf s,t in the source and target domains
respectively. For the baseline of the DA task, we directly
feed the target-domain testing images to Cf s,s to obtain
6
exp.
ID
method
training
modality
testing
modality
2
3
4
number of classes
5
6
7
1
2
3
GoogleNet
ZDDA2
ZDDA3
D
D+PRGB-D
D+PRGB-D
D
D
D
85.53
88.16
88.16
83.33
85.96
86.84
82.89
83.55
84.87
70.00
77.89
77.89
67.11
70.18
72.37
4
5
6
GoogleNet
ZDDA2
GoogleNet
D
D+PRGB-D
RGB
RGB
RGB
RGB
68.42
80.26
88.16
57.02
78.07
85.09
56.58
76.32
84.87
48.95
67.37
79.47
7
ZDDA3
D+PRGB-D
RGB-D
88.16
85.96
85.53
selected scene IDs (introduced in Sec. 4.1)
0∼1
0∼2
0∼3
8
9
10
59.02
66.92
66.92
54.28
64.80
64.47
50.88
62.28
64.33
51.84
59.74
63.16
42.11
57.89
78.07
45.11
53.76
68.80
40.46
47.37
70.07
34.50
45.03
69.88
31.58
43.16
63.68
76.32
72.81
68.42
65.13
63.16
63.16
0∼4
0∼5
0∼6
0∼7
0∼8
0∼9
Table 5: Performance comparison with different numbers of classes in scene classification. The reported numbers are classification
accuracy (%). The color of each cell reflects the performance ranking in each column, where darker color means better performance.
PRGB-D represents the task-irrelevant RGB-D pairs.
existing DA methods because the DA task DM → DM M is considered as one of the standard experiments in
recent works [8, 13, 19, 29, 30]. Although this is not a
fair comparison (because ZDDA2 has no access to the taskrelevant target-domain training data), we find that ZDDA2
can reach the accuracy comparable to those of the stateof-the-art methods (even outperform some of them), which
supports that ZDDA2 is a promising DA method when the
task-relevant target-domain training data is unavailable.
Third, we test the efficacy of ZDDA on the DA tasks
constructed from DS (adapting from depth to RGB images).
We compare ZDDA with the baseline under different scene
classification tasks by changing the number of scenes
involved. The result is summarized in Table 5, where we
list the training and testing modalities for each method. We
also list the scene IDs (introduced in Sec. 4.1) involved in
each task. Darker cells represent better accuracy in each
column. For simplicity, we use Ei to refer to the experiment
specified by exp. ID i in this section. For the fully supervised methods in depth domain, ZDDA (E2 , E3 ) outperforms the baseline (E1 ) due to the extra information brought
by the task-irrelevant RGB-D pairs. We find that for most
listed tasks, ZDDA3 (E3 ) outperforms ZDDA2 (E2 ), which
is consistent with our intuition because the source representation in ZDDA2 is constrained by the L2 loss, while the
counterpart in ZDDA3 is learned without the L2 constraint
given the simulated target representation. The fully supervised method in RGB domain (E6 ) outperforms the baseline
of the domain adaptation (E4 ) and ZDDA2 (E5 ) because E6
has access to the task-relevant RGB training data which is
unavailable for E4 and E5 . The performance improvement
from E4 to E5 is caused by ZDDA2 ’s training procedure as
well as the extra task-irrelevant RGB-D training pairs. E3
and E7 perform similarly, which supports that the simulated
target representation in ZDDA3 is similar to the real one.
To test the consistency of the performance of ZDDA
training testing validation on validation on
method modality modality train/test splits class choices
GN
D
ZDDA2 D+PRGB-D
ZDDA3 D+PRGB-D
D
D
D
52.63±1.76
56.89±2.13
58.37±3.08
53.98±1.68
62.05±1.97
62.49±1.74
GN
D
RGB
ZDDA2 D+PRGB-D RGB
GN
RGB
RGB
31.26±1.76
44.47±2.50
66.26±1.60
32.60±2.37
45.47±2.57
67.95±2.20
ZDDA3 D+PRGB-D RGB-D
58.68±3.10
62.13±1.50
10 / 5
9 / 10
# of classes / # of folds
Table 6: Validation of ZDDA’s performance (in mean classification accuracy (%)) with different training/testing splits and choices
of classes in scene classification. GN stands for GoogleNet [34].
The definition of PRGB-D and the representation of the cell color in
each column are the same as those in Table 5.
compared to that of the baseline, we perform the following
three experiments. First, we conduct 5-fold cross validation with different training/testing splits for the 10-scene
classification. Second, we perform 10-fold validation with
different selections of classes for the 9-scene classification (leave-one-class-out experiment out of the 10 selected
scenes introduced in Sec. 4.1). Third, we validate ZDDA’s
performance with different base network architectures. The
results of the first two experiments are presented in Table 6
(the extended version is in the supplementary material), and
the result of the third experiment is shown in Table 7. The
results of all the experiments are consistent with our observations from Table 5.
In Table 5, Table 6, and Table 7, the classification accuracy is reported under the condition of noise-free training
and testing data. To let ZDDA be more robust to noisy
7
(a) naive fusion
(b) ZDDA3
(c) accuracy improvement from (a) to (b)
Figure 4: Performance comparison between the two sensor fusion methods with black images as the noisy images. We compare the
classification accuracy (%) of (a) naive fusion and (b) ZDDA3 under different noise levels in both RGB and depth testing data. (c) shows
that ZDDA3 outperforms the naive fusion under most conditions.
method
training
modality
testing
modality
GN
BNA
AN
data in the RGB domain, we can still train an RGB-D fusion
model, and that the performance degrades smoothly when
the noise increases. In addition to using black images as the
noise model, we evaluate the same trained joint classifier
in ZDDA3 using another noise model (adding a black rectangle with a random location and size to the clean image)
at testing time, and the result (in the supplementary material) also supports that ZDDA3 outperforms the naive fusion
method. Although we only use black images as the noise
model for ZDDA3 at training time, we expect that adding
different noise models can improve the performance and
robustness of ZDDA3 .
SN
BNA
D
ZDDA2 D+PRGB-D
ZDDA3 D+PRGB-D
D
D
D
51.84 49.74 48.68
59.74 51.05 56.32
63.16 51.05 56.32
BNA
D
ZDDA2 D+PRGB-D
BNA
RGB
RGB
RGB
RGB
31.58 30.26 26.58
43.16 40.00 35.79
63.68 59.47 57.37
ZDDA3 D+PRGB-D
RGB-D
63.16 51.84 56.05
Table 7: Validation of ZDDA’s performance with different base
network architectures (BNA) in scene classification. The reported
numbers are classification accuracy (%). GN, AN, and SN stand
for GoogleNet [34], AlexNet [24], and SqueezeNet v1.1 [21],
respectively. The definition of PRGB-D and the representation of
the cell color in each column are the same as those in Table 5.
6. Conclusion and future work
We propose zero-shot deep domain adaptation (ZDDA),
a novel approach to perform domain adaptation (DA) and
sensor fusion with no need of the task-relevant targetdomain training data which can be inaccessible in reality.
Our key idea is to use the task-relevant source-domain
data to simulate the task-relevant target-domain representations by learning from the task-irrelevant dual-domain pairs.
Experimenting on the MNIST [25], Fashion-MNIST [41],
NIST [17], EMNIST [10], and SUN RGB-D [32] datasets,
we show that ZDDA outperforms the baselines in DA and
sensor fusion even without the task-relevant target-domain
training data. In the task adapting from MNIST [25] to
MNIST-M [12], ZDDA can even outperform several stateof-the-art DA methods which require access to the MNISTM [12] training data. We believe that ZDDA can be straightforwardly extended to handle other tasks of interest by
modifying the loss functions in Figure 2 step 2/3, and we
share a template of ZDDA’s training procedure and two
examples of such extensions in the supplementary material.
input, we train ZDDA3 step 3 with noisy training data (we
use ptrain = 20 as explained in Sec. 4.2), and evaluate the
classification accuracy under different noise conditions for
both RGB and depth testing data. The result is presented
in Figure 4, where ZDDA3 (Figure 4b) outperforms the
naive fusion method (Figure 4a) under most conditions,
and the performance improvement is shown in Figure 4c.
Both Figure 4a and Figure 4b show that the performance
degradation caused by the noisy depth testing data is larger
than that caused by the noisy RGB testing data, which
supports that the trained RGB-D classifier relies more on
the depth domain. Traditionally, training a fusion model
requires the task-relevant training data in both modalities.
However, we show that without the task-relevant training
8
References
[19] P. Haeusser, T. Frerix, A. Mordvintsev, and D. Cremers.
Associative domain adaptation. In ICCV, pages 2765–2773,
2017.
[20] J. Hoffman, S. Gupta, and T. Darrell. Learning with side
information through modality hallucination. In CVPR, pages
826–834, 2016.
[21] F. N. Iandola, S. Han, M. W. Moskewicz, K. Ashraf, W. J.
Dally, and K. Keutzer. SqueezeNet: AlexNet-level accuracy
with 50x fewer parameters and <0.5MB model size. arXiv
preprint arXiv: 1602.07360, 2016.
[22] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. B.
Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. arXiv preprint
arXiv: 1408.5093, 2014.
[23] P. Koniusz, Y. Tas, and F. Porikli. Domain adaptation by
mixture of alignments of second- or higher-order scatter
tensors. In CVPR, pages 4478–4487, 2017.
[24] A. Krizhevsky, I. Sutskever, and G. E. Hinton. ImageNet
classification with deep convolutional neural networks. In
NIPS, pages 1097–1105, 2012.
[25] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradientbased learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
[26] S. Motiian, M. Piccirilli, D. A. Adjeroh, and G. Doretto.
Unified deep supervised domain adaptation and generalization. In ICCV, pages 5715–5725, 2017.
[27] J. Ngiam, A. Khosla, M. Kim, J. Nam, H. Lee, and A. Y. Ng.
Multimodal deep learning. In ICML, pages 689–696, 2011.
[28] K. Saenko, B. Kulis, M. Fritz, and T. Darrell. Adapting
visual category models to new domains. In ECCV, pages
213–226, 2010.
[29] K. Saito, Y. Ushiku, and T. Harada. Asymmetric tri-training
for unsupervised domain adaptation. In ICML, 2017.
[30] O. Sener, H. O. Song, A. Saxena, and S. Savarese. Learning
transferrable representations for unsupervised domain adaptation. In NIPS, pages 2110–2118, 2016.
[31] K. Sohn, S. Liu, G. Zhong, X. Yu, M.-H. Yang, and M. Chandraker. Unsupervised domain adaptation for face recognition
in unlabeled videos. In ICCV, pages 3210–3218, 2017.
[32] S. Song, S. Lichtenberg, and J. Xiao. SUN RGB-D: A RGBD scene understanding benchmark suite. In CVPR, pages
567–576, 2015.
[33] B. Sun and K. Saenko. Deep CORAL: Correlation alignment
for deep domain adaptation. In ECCV Workshops, pages
443–450, 2016.
[34] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. E. Reed,
D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich.
Going deeper with convolutions. In CVPR, pages 1–9, 2015.
[35] E. Tzeng, J. Hoffman, T. Darrell, and K. Saenko. Simultaneous deep transfer across domains and tasks. In ICCV,
pages 4068–4076, 2015.
[36] E. Tzeng, J. Hoffman, K. Saenko, and T. Darrell. Adversarial
discriminative domain adaptation. In CVPR, pages 7167–
7176, 2017.
[37] H. Venkateswara, J. Eusebio, S. Chakraborty, and
S. Panchanathan. Deep hashing network for unsupervised
domain adaptation. In CVPR, pages 5018–5027, 2017.
[38] Y. Wang, W. Li, D. Dai, and L. V. Gool. Deep domain adaptation by geodesic distance minimization. In ICCV, pages
2651–2657, 2017.
[1] BAIR/BVLC AlexNet model.
http://dl.
caffe.berkeleyvision.org/bvlc_alexnet.
caffemodel. Accessed: 2017-03-02.
[2] BAIR/BVLC GoogleNet model. http://dl.caffe.
berkeleyvision.org/bvlc_googlenet.
caffemodel. Accessed: 2017-03-02.
[3] Lenet architecture in the Caffe tutorial.
https:
//github.com/BVLC/caffe/blob/master/
examples/mnist/lenet.prototxt.
[4] SqueezeNet v1.1 model.
https://github.
com/DeepScale/SqueezeNet/blob/master/
SqueezeNet_v1.1/squeezenet_v1.1.
caffemodel. Accessed: 2017-02-11.
[5] R. Aljundi and T. Tuytelaars. Lightweight unsupervised
domain adaptation by convolutional filter reconstruction. In
ECCV Workshops, pages 508–515, 2016.
[6] P. Arbelaez, M. Maire, C. Fowlkes, and J. Malik. Contour
detection and hierarchical image segmentation. TPAMI,
33:898–916, 2011.
[7] J. Blitzer, D. P. Foster, and S. M. Kakade. Zero-shot domain
adaptation: A multi-view approach. In Technical Report TTITR-2009-1. Technological institute Toyota, 2009.
[8] K. Bousmalis, N. Silberman, D. Dohan, D. Erhan, and
D. Krishnan. Unsupervised pixel-level domain adaptation
with generative adversarial networks. In CVPR, pages 3722–
3731, 2017.
[9] T.-H. Chen, Y.-H. Liao, C.-Y. Chuang, W.-T. Hsu, J. Fu, and
M. Sun. Show, adapt and tell: Adversarial training of crossdomain image captioner. In ICCV, pages 521–530, 2017.
[10] G. Cohen, S. Afshar, J. Tapson, and A. van Schaik. EMNIST:
An extension of MNIST to handwritten letters. arXiv
preprint arXiv: 1702.05373, 2017.
[11] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. FeiFei. ImageNet: A large-scale hierarchical image database.
In CVPR, pages 248–255, 2009.
[12] Y. Ganin and V. Lempitsky. Unsupervised domain adaptation
by backpropagation. In ICML, volume 37, pages 1180–1189,
2015.
[13] Y. Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle,
F. Laviolette, M. Marchand, and V. Lempitsky. Domainadversarial training of neural networks. Journal of Machine
Learning Research, 17(59):1–35, 2016.
[14] T. Gebru, J. Hoffman, and F.-F. Li. Fine-grained recognition
in the wild: A multi-task domain adaptation approach. In
ICCV, pages 1349–1358, 2017.
[15] M. Ghifary, W. B. Kleijn, M. J. Zhang, D. Balduzzi, and
W. Li. Deep reconstruction - classification networks for
unsupervised domain adaptation. In ECCV, pages 597–613,
2016.
[16] A. Gretton, A. J. Smola, J. Huang, M. Schmittfull, K. M.
Borgwardt, and B. Schölkopf. Covariate shift and local
learning by distribution matching, pages 131–160. MIT
Press, Cambridge, MA, USA, 2009.
[17] P. Grother and K. Hanaoka. NIST special database 19
handprinted forms and characters. U.S. Department of
Commerce, 2016.
[18] S. Gupta, J. Hoffman, and J. Malik. Cross modal distillation
for supervision transfer. In CVPR, pages 2827–2836, 2016.
9
[39] C. Wu, W. Wen, T. Afzal, Y. Zhang, Y. Chen, and H. Li. A
compact DNN: Approaching GoogLeNet-level accuracy of
classification and domain adaptation. In CVPR, pages 5668–
5677, 2017.
[40] M. Wulfmeier, A. Bewley, and I. Posner. Addressing appearance change in outdoor robotics with adversarial domain
adaptation. In IROS, 2017.
[41] H. Xiao, K. Rasul, and R. Vollgraf. Fashion-MNIST: A
novel image dataset for benchmarking machine learning
algorithms. arXiv preprint arXiv: 1702.05374, 2017.
[42] H. Yan, Y. Ding, P. Li, Q. Wang, Y. Xu, and W. Zuo. Mind
the class weight bias: Weighted maximum mean discrepancy
for unsupervised domain adaptation. In CVPR, pages 2272–
2281, 2017.
[43] X. Yang, P. Ramesh, R. Chitta, S. Madhvanath, E. A. Bernal,
and J. Luo. Deep multimodal representation learning from
temporal data. In CVPR, pages 5447–5455, 2017.
[44] Y. Yang and T. M. Hospedales. Zero-shot domain adaptation
via kernel regression on the grassmannian. In BMVC Workshop on Differential Geometry in Computer Vision, 2015.
[45] J. Zhang, W. Li, and P. Ogunbona. Joint geometrical and
statistical alignment for visual domain adaptation. In CVPR,
pages 1859–1867, 2017.
[46] Y. Zhang, P. David, and B. Gong. Curriculum domain adaptation for semantic segmentation of urban scenes. In ICCV,
pages 2020–2030, 2017.
10
| 1cs.CV
|
EVENT-TRIGGERED CONTROL OF INFINITE-DIMENSIONAL
SYSTEMS∗
arXiv:1801.05521v1 [math.OC] 17 Jan 2018
MASASHI WAKAIKI AND HIDEKI SANO†
Abstract. We consider the event-triggered control of continuous-time linear infinite-dimensional
systems. We show that an event trigger that measures the difference of a control input leads to the
positive minimum inter-event time if a feedback operator is compact. Moreover, under certain natural
assumptions on the infinite-dimensional system, we show that there exists an event trigger such that
the closed-loop system is exponentially stable.
Key words. event-triggered control, infinite-dimensional systems, stabilization
AMS subject classifications. 34G10, 93C62, 93C25, 93D15
1. Introduction. Event-triggered control is one of resource-aware control methods, and its intervals of data transmissions are determined by a predefined condition
on the data. As a result, network and energy resources are consumed only when the
data is necessary for control. In addition to such networked-control applications, the
analysis and synthesis of event-triggered control is interesting from a theoretical viewpoint, because event triggering interacts continuous-time and discrete-time dynamics
in a way different from the usual periodic sampled-data control schemes. Most of the
existing studies on event-triggered control have been developed for finite-dimensional
systems, but some researchers have recently extended to infinite-dimensional systems,
e.g., systems with output delays [12], first-order hyperbolic systems [6], and secondorder parabolic systems [11, 21].
In this paper, we consider the following system with state space X and input
space U :
ẋ(t) = Ax(t) + Bu(t),
where A is the generator of a strongly continuous semigroup T (t) on X and B is
a bounded linear operator from U into the extrapolation space X−1 . To infinitedimensional systems described by this abstract evolution equation, the results of periodic sampled-data control for finite-dimensional systems have been generalized in a
number of papers; see [13–15, 18–20]. Instead of periodic control updating, we here
generate the control input u(t) by an event-triggered scheme based on the difference
of u(t) such as
(1.1a)
tk+1 = inf t > tk : kF x(t) − u(tk )kU > εkx(tk )kX
∀k ∈ N0
(1.1b)
u(t) = F u(tk )
∀t ∈ [tk , tk+1 ), ∀k ∈ N0 ,
where ε is a positive constant and F is a bounded linear operator from X to U . Since a
small ε leads to frequent updates of the control input, we would expect that the eventtriggered feedback system is exponentially stable for a sufficiently small ε if A + BF
generates an exponentially stable semigroup. One of the fundamental problems we
consider is whether or not this intuition is correct.
In addition to closed-loop stability, the minimum inter-event time
inf (tk+1 − tk )
k∈N0
∗ Submitted
to the editors DATE.
Funding: This work was supported by JSPS KAKENHI Grant Numbers JP17K14699.
† Graduate School of System Informatics, Kobe University, Nada, Kobe, Hyogo 657-8501, Japan
([email protected], [email protected]).
1
2
MASASHI WAKAIKI AND HIDEKI SANO
should be guaranteed to be positive; otherwise, infinitely many events might occur in
finite time. This phenomenon is called the Zeno behavior (see [7]) and makes eventtriggered control schemes infeasible for practical implementation. To guarantee the
Zeno freeness is the unique theoretical specification of event-triggered control, and
for finite-dimensional systems, the minimum inter-event time has been extensively
investigated; see [1, 4, 22]. For example, in the finite-dimensional case, it was shown
in [22] that the event-trigger
tk+1 = inf t > tk : kx(t) − x(tk )kX > εkx(t)kX }
∀k ∈ N0
satisfies inf k∈N0 (tk+1 − tk ) > 0. However, this is not true for the infinite-dimensional
case, which is illustrated in Example 2.2. This is the reason why we use the eventtrigger (1.1a) based on the difference of the control input, and we see in Section 2
that the time sequence {tk }∞
k=0 constructed by (1.1a) satisfies inf k∈N0 (tk+1 − tk ) > 0
if the feedback operator F is compact.
After the discussion on the minimum inter-event time in Section 2, we analyze the
exponential stability of the event-triggered feedback system in Section 3 for the case
where the input operator B is bounded. Introducing a norm on the state space with
respect to which the semigroup generated by A + BF is a contraction, we provide a
sufficient condition on the event trigger parameter ε for the exponential stability of
the feedback system. Moreover, under a certain assumption on the semigroup T (t),
we obtain another sufficient condition for exponential stability. While we obtain the
former result via a trajectory-based approach, a key element in the latter result is the
application of the Lyapunov stability theorem.
In Section 4, we study the case of unbounded input operators. We first focus
on a system with a finite-dimensional unstable part and a feedback operator that
stabilizes the unstable part but does not act on the residual stable part. In this case,
the feedback operator has a specific structure, but we can achieve the exponential
stability of the feedback system by using less conservative event triggers for the finitedimensional part. Second we assume that the semigroups generated by A and A+BF
with suitable domains are analytic and exponentially stable, respectively. Under these
assumptions, we show that for every compact feedback operator, there exist periodic
event triggers [8, 9] achieving exponential stability.
Notation and terminology. We denote by N0 the set of nonnegative integers.
For α ∈ R, we define Cα := {s ∈ C : Res > α}. Let X and Y be Banach spaces. Let
us denote the space of all bounded linear operators from X to Y by B(X, Y ), and set
B(X) := B(X, X). Let A be an operator from X to Y . The domain of A is denoted
by D(A) and the spectrum of A by σ(A). For a subset S ⊂ X, let A|S denote the
restriction of A to S, namely,
A|S x = Ax
∀x ∈ D(A) ∩ S.
For a strongly continuous semigroup T (t) on X, the exponential growth bound of T (t)
is denoted by ω(T ), that is, ω(T ) := limt→∞ ln kT (t)k/t. We say that the strongly
continuous semigroup T (t) is exponentially stable if ω(T ) < 0.
2. Event-triggered feedback system.
2.1. Infinite-dimensional system. Let a time sequence {tk }∞
k=0 satisfy t0 = 0
and
(2.1)
inf (tk+1 − tk ) > 0.
k∈N0
EVENT-TRIGGERED CONTROL OF INFINITE-DIMENSIONAL SYSTEMS
3
Let us denote by X and U the state space and the input space, and both of them
are Hilbert spaces. Let k · k and h·, ·i denote the norm and the inner product of X,
respectively. The space X−1 stands for the completion of X in the norm kxk−1 =
k(λI − A)−1 xk, where λ ∈ C is any element in the resolvent set of A. Clearly,
X ,→ X−1 and B(U, X) ⊂ B(U, X−1 ).
Consider the following infinite-dimensional system:
(2.2a)
ẋ(t) = Ax(t) + Bu(t) ∀t ≥ 0;
(2.2b)
u(t) = F x(tk )
x(0) = x0 ∈ X
∀t ∈ [tk , tk+1 ), ∀k ∈ N0 .
where x(t) ∈ X, u(t) ∈ U , A is the generator of a strongly continuous semigroup T (t)
on X, and the input operator B and the feedback operator F satisfy B ∈ B(U, X−1 )
and F ∈ B(X, U ). We say that B is bounded if B ∈ B(U, X); otherwise B is unbounded.
The semigroup T (t) can be extended to a strongly continuous semigroup on X−1 ,
and its generator on X−1 is an extension of A to X. We shall use the same symbols
T (t) and A for the original ones and the associated extensions.
For the the abstract evolution equation (2.2), we define a function x recursively
by
(2.3a)
x(0) = x0
Z
(2.3b)
τ
T (s)BF x(tk )ds ∀τ ∈ (0, tk+1 − tk ], ∀k ∈ N0
x(tk + τ ) = T (τ )x(tk ) +
0
as in the periodic sampled-data case [14]. Considering T (t) as a semigroup on X−1 ,
we find from the standard theory of strongly continuous semigroups that x given by
(2.3) satisfies
(2.4)
x ∈ C(R+ , X)
and x|[tk ,tk+1 ] ∈ C 1 ([tk , tk+1 ], X−1 ) ∀k ∈ N0 .
Additionally, the following differential equation in X−1 holds:
(2.5)
ẋ(t) = Ax(t) + BF x(tk )
∀t ∈ (tk , tk+1 ), ∀k ∈ N0 .
Since only x defined by (2.3) satisfies the properties (2.4) and (2.5), we can say that
this x is the unique solution of the abstract evolution equation (2.2).
Definition 2.1 (Exponential stability). The system (2.2) is exponential stable if
there exist M ≥ 1 and γ > 0 such that x given by (2.3) satisfies
kx(t)k ≤ M e−γt kx0 k
∀x0 ∈ X, ∀t ≥ 0.
We call the constant γ (an upper bound on) the decay rate of the system (2.2).
2.2. Minimum inter-event times. We call inf k∈N0 (tk+1 − tk ) the minimum
inter-event time. If this value is zero, then the event trigger may update the input
infinitely fast, which cannot be realized in digital platform. In this section, we show
that the following event triggers guarantee the positive inter-event time under certain
assumptions:
(2.6)
tk+1 := inf t > tk : kF x(t) − u(tk )kU > εkx(tk )k
∀k ∈ N0
(2.7)
tk+1 := inf t > tk : kF x(t) − u(tk )kU > εkx(t)k
∀k ∈ N0 .
4
MASASHI WAKAIKI AND HIDEKI SANO
Whereas we here use the event triggers (2.6) and (2.7) that measure the difference
of the inputs, the commonly-used event triggers for finite-dimensional systems [22] are
based on the difference of the state such as
(2.8)
tk+1 := inf t > tk : kx(t) − x(tk )k > εkx(t)k}
∀k ∈ N0 .
If the semigroup T (t) is uniformly continuous, then one can show that the inter-event
time of the event trigger (2.8) is positive. However, systems described by uniformly
continuous semigroups do not contain practical infinite-dimensional systems. On the
other hand, in the following simple example, event triggers based on the difference of
the state lead to inf k∈N0 (tk+1 − tk ) = 0 in finite time.
Example 2.2. Let X = L2 (0, ∞), the Hilbert space of square integrable functions
on (0, ∞) with the usual inner product, and consider the shift operator
∀x ∈ L2 (0, ∞), ∀s > 0.
(T (t)x)(s) = x(t + s)
Then T (t) is a strongly continuous semigroup on L2 (0, ∞), and as discussed in Remark 7 in [2], T (t) is strongly stable but T ∗ (t) is not.
Set an initial state x0 ∈ L2 (0, ∞) to be
(
1 s≤1
x0 (s) =
0 s > 1,
and x(t) = T (t)x0 for all t > 0. Assume that ε < 1, and define a time sequence
{tk }∞
k=0 by t0 = 0 and
(2.9)
tk+1 := inf t > tk : kx(t) − x(tk )kL2 > εkx(tk )kL2 }.
If tk ∈ [0, 1), then kx(tk )k2L2 = 1 − tk and
Z 1−tk
2
kT (τ )x(tk ) − x(tk )kL2 =
1dτ = τ
∀τ ∈ [0, 1 − tk ].
1−(tk +τ )
It follows that
tk+1 = tk + ε(1 − tk )
Similarly, if we define
∀k ∈ N0 .
{tk }∞
k=0
tk+1
by t0 = 0 and (2.8), then
ε
(1 − tk )
∀k ∈ N0 .
= tk +
1+ε
Both of the time sequences {tk }∞
k=0 are monotonically increasing and converge to 1.
Thus the minimum inter-event time inf k∈N0 (tk+1 − tk ) is zero.
For τ ≥ 0, define the operator Sτ : U → X by
Z τ
Sτ u :=
T (s)Buds.
0
The following lemma is useful for the discussion on the minimum inter-event time:
Lemma 2.3 ([14]). For any τ ≥ 0, Sτ ∈ B(U, X), and for any θ > 0,
sup kSτ kB(U,X) < ∞.
0≤τ ≤θ
Moreover, if F ∈ B(X, U ) is compact, then
(2.10)
lim kSτ F kB(X) = 0.
τ →0
EVENT-TRIGGERED CONTROL OF INFINITE-DIMENSIONAL SYSTEMS
5
For compact operators, the next lemma is also known:
Lemma 2.4 ([14]). Let X, Y , and Z be Hilbert spaces and let Γ : [0, 1] → B(X, Y )
be given. If limt→0 Γ∗ (t)y = 0 for all y ∈ Y and if Λ ∈ B(Y, Z) is compact, then
lim kΛΓ(t)kB(X,Z) = 0.
t→0
From these lemmas, we obtain the following result:
Lemma 2.5. Assume that F ∈ B(X, U ) is compact. Set
x(τ ) = (T (τ ) + Sτ F )x0
(2.11)
∀τ > 0;
x0 ∈ X.
For every ε > 0, there exists θ > 0 such that
kF x(τ ) − F x0 kU ≤ εkx0 k
∀x0 ∈ X, ∀τ ∈ [0, θ).
Proof. Since (2.11) leads to
x(τ ) − x(0) = (T (τ ) − I)x(0) + Sτ F x(0)
∀τ ≥ 0,
it follows that
(2.12)
kF x(τ ) − F x(0)kU ≤ kF (T (τ ) − I) + F Sτ F kB(X,U ) · kx(0)k
∀τ ≥ 0.
From Lemma 2.3,
lim kF Sτ F kB(X,U ) = 0.
τ →0
Since T ∗ (t) is strongly continuous, we obtain
lim k(T ∗ (τ ) − I)xk = 0
τ →0
∀x ∈ X,
and hence it follows from Lemma 2.4 that
lim kF (T (τ ) − I)kB(X,U ) = 0.
τ →0
Thus, for every ε > 0, there exists θ > 0 such that
kF (T (τ ) − I) + F Sτ F kB(X,U ) ≤ kF (T (τ ) − I)kB(X,U ) + kF Sτ F kB(X,U )
(2.13)
<ε
∀τ ∈ [0, θ).
Combining (2.12) and (2.13), we obtain the desired result.
From Lemma 2.5, we see that the minimum inter-event time inf k∈N0 (tk+1 − tk ) is
positive under the event trigger (2.6).
Theorem 2.6. Assume that F ∈ B(X, U ) is compact. For the system (2.2), set
the time sequence {tk }∞
k=0 recursively satisfying t0 = 0 and (2.6). Then for every
ε > 0, the time sequence {tk }∞
k=0 satisfies inf k∈N0 (tk+1 − tk ) > 0.
Proof. From (2.4), it follows that x(tk ) ∈ X for every k ∈ N0 . Therefore, we see
from (2.3) and Lemma 2.5 that, for every ε > 0, there exists θ > 0 such that
kF x(tk + τ ) − F x(tk )kU ≤ εkx(tk )k
∀τ ∈ [0, θ),
∀k ∈ N0 .
Since u(tk ) = F x(tk ) for every k ∈ N0 , the time sequence {tk }∞
k=0 satisfies
inf (tk+1 − tk ) ≥ θ > 0.
k∈N0
This completes the proof.
6
MASASHI WAKAIKI AND HIDEKI SANO
We next investigate the minimum inter-event time of the other event trigger (2.7).
In addition to Lemma 2.5, we need the following estimate:
Lemma 2.7. Assume that F ∈ B(X, U ) is compact, and set x(t) as in (2.11).
There exist c1 > 0 and s1 > 0 such that the semigroup T (t) satisfies
kT (s1 )x0 k ≥ c1 kx0 k
(2.14)
∀x0 ∈ X
if and only if there exist c2 > 0 and θ > 0 such that
(2.15)
kx0 k ≤ c2 kx(τ )k
∀x0 ∈ X, ∀τ ∈ [0, θ).
Proof. Suppose first that (2.14) holds for some c1 > 0 and s1 > 0. Since there
exists M ≥ 1 such that
kT (t)kB(X) ≤ M
∀t ∈ [0, s1 ],
it follows from (2.14) that
c1 kx0 k ≤ kT (s1 )x0 k ≤ kT (s1 − t)T (t)x0 k ≤ M kT (t)x0 k
∀x0 ∈ X, ∀t ∈ [0, s1 ].
Therefore,
(2.16)
kT (t)x0 k ≥
c1 0
kx k
M
∀x0 ∈ X, ∀t ∈ [0, s1 ].
From (2.10), there exists s2 ∈ [0, s1 ] such that
(2.17)
kSt F x0 k ≤ kSt F kB(X) · kx0 k <
c1
kx0 k
2M
∀x0 ∈ X, ∀t ∈ [0, s2 ].
Combining (2.16) and (2.17), we obtain
(2.18)
k(T (t) + St F )x0 k ≥ kT (t)x0 k − kSt F x0 k ≥
c1
kx0 k
2M
∀x0 ∈ X, ∀t ∈ [0, s2 ].
Therefore, (2.15) holds with c2 := 2M/c1 and θ := s2 .
Conversely, if (2.15) holds for some θ > 0 and c2 > 0, then
kx0 k ≤ c2 kT (τ )x0 k + c2 kSτ F x0 k
∀x0 ∈ X, ∀τ ∈ [0, θ).
Using (2.10) as in the discussion above, we find that there exists s1 ∈ (0, θ) such that
kT (s)x0 k ≥
1
kx0 k
2c2
∀x0 ∈ X, ∀s ∈ [0, s1 ].
This completes the proof.
Remark 2.8. Suppose that B is bounded, i.e., B ∈ B(U, X). Then
lim kSτ k = 0,
τ →0
and hence the compactness of F is not required in Lemma 2.7.
Remark 2.9. The condition (2.14) appears also in Theorem 2 of [16] for the
applicability of the Lyapunov stability theorem, and it is shown in Corollary 1 of [16]
that T (t) satisfies (2.14) and its range is dense in X if and only if T (t) can be extended
to a strongly continuous group on X.
EVENT-TRIGGERED CONTROL OF INFINITE-DIMENSIONAL SYSTEMS
7
We see from Lemmas 2.5 and 2.7 that the event trigger (2.6) leads to the positive
minimum inter-event time for a strongly continuous semigroup T (t) satisfying (2.14).
Theorem 2.10. Assume that F ∈ B(X, U ) is compact and that the semigroup
T (t) satisfies (2.14) for some c1 > 0 and s1 > 0. For the system (2.2), set the time
sequence {tk }∞
k=0 recursively satisfying t0 = 0 and (2.7). Then for every ε > 0, the
time sequence {tk }∞
k=0 satisfies inf k∈N0 (tk+1 − tk ) > 0.
Proof. Combining Lemmas 2.5 and 2.7, we have that, for every ε > 0, there exists
θ > 0 such that
kF x(tk + τ ) − u(tk )kU ≤ εkx(t)k,
∀τ ∈ [0, θ),
∀k ∈ N0 .
Thus the time sequence {tk }∞
k=0 satisfies inf k∈N0 (tk+1 − tk ) ≥ θ > 0.
3. Stability analysis under bounded control. In this section, we analyze
closed-loop stability in the case in which the input operator B is bounded. If the
feedback operator F is compact and if the semigroup generated by A + BF is exponentially stable, then the event-triggered feedback system is exponentially stable
provided the event-triggered parameter ε is sufficiently small.
Fixing τM > 0, we first set a time sequence {tk }∞
k=0 recursively satisfying t0 = 0
and
n
o
(3.1) tk+1 := min tk + τM , inf t > tk : kF x(t) − u(tk )kU > εkx(tk )k
∀k ∈ N0 .
Theorem 3.1. Assume that A generates a strongly continuous semigroup T (t)
on X, B ∈ B(U, X), and F ∈ B(U, X) is compact. Assume that there exists M ≥ 1
and ω > 0 such that the semigroup TBF (t) generated by A + BF satisfies
kTBF (t)k ≤ M e−ωt
(3.2)
∀t ≥ 0.
If we set the time sequence {tk }∞
k=0 to be t0 = 0 and (3.1) with a parameter ε > 0
satisfying
(3.3)
ε<
ω
,
M kBkB(U,X)
then the system (2.2) is exponentially stable with decay rate γ defined by
log (1 − ε0 )e−ωτ + ε0
(3.4)
γ := min
,
τm ≤τ ≤τM
−τ
where
ε0 := ε
M kBkB(U,X)
,
ω
τm := inf (tk+1 − tk ).
k∈N0
Proof. As in Theorem 5.2 on p. 19 in [17] and Theorem 3.1 in [14], we introduce
a new norm | · | on X, which is defined by
|x| := sup keωt TBF (t)xk.
t≥0
It follows from (3.2) that
keωt TBF (t)xk ≤ eωt · M e−ωt kxk = M kxk
∀x ∈ X, ∀t ≥ 0.
8
MASASHI WAKAIKI AND HIDEKI SANO
On the other hand,
kxk = keω0 TBF (0)xk ≤ sup keωt TBF (t)xk
∀x ∈ X.
t≥0
Hence we have
kxk ≤ |x| ≤ M kxk
(3.5)
∀x ∈ X.
Moreover,
|TBF (t)x| = sup keωs TBF (s + t)xk = sup keω(s−t) TBF (s)xk
s≥0
(3.6)
=e
s≥t
−ωt
ωs
sup ke TBF (s)xk
s≥t
≤ e−ωt sup keωs TBF (s)xk = e−ωt |x|
∀x ∈ X, ∀t ≥ 0.
s≥0
Noting
(3.7)
ẋ(t) = (A + BF )x(t) − B(F x(t) − F x(tk ))
∀t ∈ (tk , tk+1 ), ∀k ∈ N0 ,
we see from a routine calculation (see, e.g., Exercise 3.3 in [3] and Appendix A) that
x(tk + τ ) in (2.3) can be written as
(3.8)
Z τ
x(tk +τ ) = TBF (τ )x(tk )−
TBF (τ −s)B(F x(tk +s)−F x(tk ))ds ∀τ ∈ (0, tk+1 −tk ].
0
Since the event-trigger condition in (3.1) leads to
sup
kF x(tk + s) − F x(tk )kU ≤ εkx(tk )k,
0≤s<tk+1 −tk
using the properties (3.5) and (3.6), we obtain
Z τ
|x(tk + τ )| ≤ |TBF (τ )x(tk )| +
|TBF (τ − s)B(F x(tk + s) − F x(tk ))|ds
Z τ0
≤ e−ωτ |x(tk )| +
e−ω(τ −s) |B(F x(tk + s) − F x(tk ))|ds
0
1 − e−ωτ
M kBkB(U,X) · sup kF x(tk + s) − F x(tk )kU
≤e
|x(tk )| +
ω
0≤s≤τ
−ωτ
≤ (1 − ε0 )e
+ ε0 · |x(tk )|
∀τ ∈ (0, tk+1 − tk ], ∀k ∈ N0
−ωτ
where ε0 is defined as in (3.4) and satisfies 0 < ε0 < 1 from (3.3).
Since f (τ ) := (1 − ε0 )e−ωτ + ε0 satisfies f (0) = 1 and f 0 (τ ) = (ε0 − 1)ωe−ωτ < 0
for all τ ≥ 0, it follows that
(1 − ε0 )e−ωτ + ε0 < 1
∀τ > 0.
Therefore, γ ∈ R defined by (3.4) satisfies γ > 0, and
|x(tk + τ )| ≤ e−γτ |x(tk )|
∀τ ∈ (0, tk+1 − tk ], ∀k ∈ N0 .
EVENT-TRIGGERED CONTROL OF INFINITE-DIMENSIONAL SYSTEMS
9
In particular,
|x(tk+1 )| ≤ e−γτ |x(tk )|
(3.9)
∀k ∈ N0 .
Using (3.9) recursively, we obtain
|x(t)| ≤ e−γt |x0 |
∀x0 ∈ X, ∀t ≥ 0.
Thus we see from (3.5) that
kx(t)k ≤ |x(t)| ≤ e−γt |x0 | ≤ M e−γt kx0 k
∀x0 ∈ X, ∀t ≥ 0.
This completes the proof.
Remark 3.2. In general, it is difficult to find a constant M ≥ 1 satisfying (3.2).
If TBF can be expanded in a Riesz basis, then we can characterize such a constant
M in the following way. Let A + BF be a Riesz spectral operator in Definition 2.3.4
on p. 41 in [3] and have simple eigenvalues {λn }n∈N and corresponding eigenvectors
{fn }n∈N . Then there exist positive constants α and β such that for every N ∈ N and
every scalars ρ1 , . . . , ρN such that
α
N
X
n=1
2
|ρn | ≤
N
X
2
ρn fn
≤β
n=1
N
X
|ρn |2 .
n=1
Let {gn }n∈N be the eigenvectors (A + BF )∗ such that {fn }n∈N and {gn }n∈N are
biorthogonal. As shown in the proof of Theorem 2.3.5 on p.p. 41–45 in [3], TBF (t)
satisfies
r
β −ω(TBF )t
kTBF (t)k ≤
e
∀t ≥ 0.
α
p
Thus a constant M ≥ 1 satisfying (3.2) is given by M = β/α. In particular, if
{fn }n∈N is an orthogonal basis, then α = β = 1, and hence M = 1.
The following example illustrates the result in Theorem 3.1:
Example 3.3. Consider a metal rod of length one that can be heated along its
length:
∂z
∂2z
(ξ, t) = 2 (ξ, t) + v(ξ, t)
∂t
∂ξ
z(0, t) = 0, z(1, t) = 0,
∀t ≥ 0
where z(ξ, t) and v(ξ, t) are the temperature of the rod and the addition of heat along
the bar at time t ≥ 0 and position ξ, respectively. We can reformulate this equation
as an abstract evolution equation (2.2a) with X = U = L2 (0, 1), the Hilbert space of
square integrable functions on (0, 1) with the usual inner product, and x(t) = z(·, t),
u(t) = v(·, t). We also introduce operators
Ah =
d2 h
dξ 2
with domain
D(A) = {h ∈ W 2,2 (0, 1) : h(0) = 0, h(1) = 0},
10
MASASHI WAKAIKI AND HIDEKI SANO
and B = I. Similarly to Example 6.2.8 on p. 299 in [3], the optimal control with cost
functional
Z ∞Z 1
kz(x, t)k2L2 + kv(x, t)k2L2 dxdt
0
0
is given by
u(t) =
∞
X
−n2 π 2 +
p
n4 π 4 + 1 hx(t), φn iL2 φn =: F x(t),
n=1
where φn (ξ) :=
√
2
2 sin(nπξ) and {φn }∞
n=1 is an orthonormal basis of L (0, 1). Since
p
∀n ∈ N,
−n2 π 2 + n4 π 4 + 1 ≤ 1
it follows that the feedback operator F is bounded. Moreover, we can show that F is
compact. Indeed, define the finite-rank operator FN ∈ B(L2 (0, 1)) by
FN x =
N
X
−n2 π 2 +
p
n4 π 4 + 1 hx, φn iL2 φn
∀x ∈ L2 (0, 1),
n=1
Then, we find from the Parseval equality that
kF x −
FN xk2L2
∞
p
X
−n2 π 2 + n4 π 4 + 1 hx, φn iL2 φn
=
n=N +1
∞
X
L2
p
−n2 π 2 + n4 π 4 + 1 hx, φn iL2
=
2
2
n=N +1
∞
2 X
p
2
|hx, φn iL2 |
≤ −(N + 1)2 π 2 + (N + 1)4 π 4 + 1
n=N +1
2
p
≤ −(N + 1)2 π 2 + (N + 1)4 π 4 + 1 kxk2L2
∀x ∈ L2 (0, 1).
We therefore obtain
kF − FN kB(L2 ) ≤ −(N + 1)2 π 2 +
and since
lim
N →∞
−(N + 1)2 π 2 +
p
p
(N + 1)4 π 4 + 1,
(N + 1)4 π 4 + 1 = 0,
it follows that the finite-rank operator FN uniformly converges to F . Hence the
feedback operator F is compact.
Using the expansion of A:
Ax =
∞
X
−n2 π 2 hx, φn iL2 φn
n=1
with domain
(
D(A) =
x ∈ L2 (0, 1) :
∞
X
n=1
)
|n2 π 2 |2 · |hx, φn iL2 |2 < ∞ ,
EVENT-TRIGGERED CONTROL OF INFINITE-DIMENSIONAL SYSTEMS
11
we see that
ω(TBF ) = −2π 2 +
p
π 4 + 1.
It follows from the discussion in Remark 3.2 that the semigroup TBF (t) satisfies
kTBF (t)kB(L2 ) ≤ e−(2π
2
√
− π 4 +1)t
∀t ≥ 0.
It follows from Theorem 3.1 that if the event-trigger parameter ε > 0 satisfies
p
ε < 2π 2 − π 4 + 1,
then the closed-loop system with the event trigger (3.1) is exponential stable.
{tk }∞
k=0
We next set a time sequence
recursively as in (2.7) with t0 = 0 and
show the exponential stability of the closed-loop system under such an event trigger.
Instead of the trajectory-based approach in Theorem 3.1, we here apply the Lyapunov
stability theorem.
Theorem 3.4. Assume that A generates a strongly continuous semigroup T (t)
on X, B ∈ B(U, X) and that F ∈ B(U, X) is compact. Assume further that the
semigroup T (t) satisfies (2.14) for some c1 > 0 and s1 > 0 and that the semigroup
TBF (t) generated by A+BF is exponential stable. If we set the time sequence {tk }∞
k=0
to be t0 = 0 and (2.7) with a parameter ε > 0 satisfying
Z ∞
1
∗
(3.11)
ε<
,
where P x :=
TBF
(t)TBF (t)xdt ∀x ∈ X,
2kP BkB(U,X)
0
then the infinite-dimensional system (2.2) is exponentially stable with decay rate γ
defined by
(3.12)
1 − 2εkP BkB(U,X)
R∞
.
2 0 kTBF (t)k2 dt
γ :=
Proof. For all s1 > 0, there exists M ≥ 1 such that
kT (t)kB(X) ≤ M,
(3.13)
kTBF (t)kB(X) ≤ M
∀t ∈ [0, s1 ],
and we obtain (2.16) from (2.14). Since TBF (t) satisfies the following identity (see
Theorem 3.2.1 on p. 110 in [3]):
Z
(3.14)
t
TBF (t − s)BF T (s)xds
TBF (t)x = T (t)x +
∀x ∈ X,
0
it follows from (2.14) and (3.13) that
Z
t
kTBF (t)xk ≥ kT (t)xk −
kTBF (t − s)BF T (s)xkds
0
c
1
− tM 2 kBF kB(X) kxk
∀x ∈ X, ∀t ∈ [0, s1 ].
≥
M
Therefore,
(3.15)
kTBF (t)xk ≥
c1
kxk
2M
∀x ∈ X, ∀t ∈ [0, s3 ],
12
MASASHI WAKAIKI AND HIDEKI SANO
where
c1
s3 := min s1 ,
3
2M kBF kB(X)
> 0.
Since TBF is exponential stable, it follows from Theorem 5.1.3 in [3] that there
exists a positive operator P ∈ B(X) such that the following Lyapunov inequality
holds:
(3.16)
h(A + BF )x, P xi + hP x, (A + BF )xi = −kxk2
∀x ∈ X,
and such an operator P is given as in (3.11). Using (3.15), we have from Theorem 2
in [16] that there exist α, β > 0 such that
(3.17)
αkxk2 ≤ hP x, xi ≤ βkxk2
Since
Z
hP x, xi =
∞
kTBF (t)xk2 dt ≤
Z
∀x ∈ X.
∞
kTBF (t)k2 dt · kxk2 ,
0
0
it follows that we can choose β > 0 so that
Z ∞
β≤
kTBF (t)k2 dt.
0
Assume that x0 ∈ D(A) := {x ∈ X : Ax ∈ X}. Then the mild solution of (2.3)
is also a classical solution, namely, x satisfies the differential equation (2.5) in X.
Moreover, if we define the implementation-induced error e(t) := u(t) − F x(t), then
ke(t)kU < εkx(t)k
under the event trigger (2.7). Therefore, using (3.7) and (3.16), we find that for every
t ∈ (tk , tk+1 ) and every k ∈ N0 , V (t) := hP x(t), x(t)i satisfies
dV
(t) = hP ẋ(t), x(t)i + hẋ(t), P x(t)i
dt
= hP ((A + BF )x(t) + Be(t)), x(t)i + hx(t), P ((A + BF )x(t) + Be(t))i
≤ −kx(t)k2 + hP Be(t), x(t)i + hx(t), P Be(t)i
≤ −kx(t)k2 + 2kP BkB(U,X) · kx(t)k · ke(t)kU
≤ −(1 − 2εkP BkB(U,X) )kx(t)k2
≤ −2γV (x(t)),
where γ > 0 is defined as in (3.12). We see from the positive definiteness (3.17) that
αkx(t)k2 ≤ V (t) ≤ e−2γt V (x0 ) ≤ βe−2γt kx0 k2
∀t ≥ 0,
and hence
r
(3.18)
kx(t)k ≤
β −γt 0
e kx k
α
∀x0 ∈ D(A), ∀t ≥ 0.
Finally, we show the exponential stability for all initial states in X. Fix x0 ∈ X
and te ≥ 0 arbitrarily, and let x(t) be the solution of the abstract evolution equation
(2.3) with the initial state x0 . Since Sτ ∈ B(U, X) for every τ ≥ 0 by Lemma 2.3, x(t)
EVENT-TRIGGERED CONTROL OF INFINITE-DIMENSIONAL SYSTEMS
13
depends continuously on the initial state x0 in the sense that there exists a constant
L = L(te ) ≥ 1 such that
kx(t)k ≤ Lkx0 k
∀x0 ∈ X, ∀t ∈ [0, te ].
Moreover, since D(A) is dense in X, it follows that for every ε > 0, there exists
ζ 0 ∈ D(A) such that the solution ζ(t) of the abstract evolution equation (2.3) with
the initial state ζ 0 satisfies
kx(t) − ζ(t)k ≤ ε
∀t ∈ [0, te ].
Therefore we have from (3.18) that
kx(t)k ≤ kx(t) − ζ(t)k + kζ(t)k ≤
r
r !
β
β −γt 0
ε+
e kx k
1+
α
α
Since ε > 0 was arbitrary, we obtain
r
β −γt 0
e kx k
kx(t)k ≤
α
∀t ∈ [0, te ].
∀t ∈ [0, te ].
Thus the feedback system (2.2) is exponentially stable with decay rate γ.
Remark 3.5. If M ≥ 1 and ω > 0 satisfy kTBF (t)kB(X) ≤ M e−ωt for all t ≥ 0,
then we obtain
M2
kP kB(X) ≤
.
2ω
Therefore, (3.11) and (3.12) can be rewritten as
ε<
ω
,
M 2 kBkB(U,X)
γ≥
ω − εM 2 kBkB(U,X)
.
M2
4. Stability analysis under unbounded control. Throughout this section,
we consider the unbounded input operator B ∈ B(U, X−1 ). We provide two eventtriggered control schemes for the exponential stabilization of the feedback system.
The first approach is based on system decomposition, and the second one employs a
periodic event trigger developed in [8, 9]
4.1. Event-triggered control based on system decomposition.
4.1.1. System decomposition. In what follows, we shall place a number of
assumptions on the infinite-dimensional system (2.2) and recall the decomposition of
infinite-dimensional systems under unbounded control used in [13–15].
Assumption 4.1. There exists α < 0 such that σ(A) ∩ Cα consists of finitely
many eigenvalues of A with finite algebraic multiplicities.
If Assumption 4.1 holds, then we can decompose X by a standard technique (see, e.g.,
Lemma 2.5.7 in [3]) as follows. There exists a rectifiable, closed, simple curve Γ in C
not intersecting σ(A) and containing σ(A) ∩ Cα in its interior and σ(A) ∩ (C \ Cα ) in
its exterior. The operator Π : X → X, defined by
Z
1
(4.1)
Π :=
(sI − A)−1 ds,
2πi Γ
14
MASASHI WAKAIKI AND HIDEKI SANO
where Γ is traversed once in the counterclockwise direction, is a projection operator,
and we can decompose X to be
(4.2)
X = X + ⊕ X −,
where X + := ΠX and X − := (I − Π)X.
This decomposition satisfies dim X + < ∞ and X + ⊂ D(A), and X + , X − are T (t)invariant for all t ≥ 0. Define
(4.3)
A+ := A|X + ,
A− := A|X − ,
T + (t) := T (t)|X + ,
T − (t) := T (t)|X − .
Then
σ(A− ) = σ(A) ∩ (C \ Cα ).
σ(A+ ) = σ(A) ∩ Cα ,
Note that A+ and A− generate the semigroups T + (t) on X + and T − (t) on X − , respectively. The semigroup T − (t) can be extended to a strongly continuous semigroup
on the extrapolation space (X − )−1 , and its generator on (X − )−1 is an extension of
A− to X − . The same symbols T − (t) and A− will be used to denote these extensions.
Since the spectrum of the operator A on X is equal to the spectrum of the operator
A on X−1 , the projection operator Π on X defined by (4.1) can be extended to a
projection Π−1 on X−1 . If λ ∈ ρ(A) and if λI − A is considered as an operator in
B(X, X−1 ), then Π−1 is similar to Π:
Π−1 = (λI − A)Π(λI − A)−1
and satisfies Π−1 X−1 = ΠX = X + . Using the extended projection operator Π−1 , we
can decompose the control operator B ∈ B(U, X−1 ):
B + := Π−1 B,
B − := (I − Π−1 )B.
Since (X − )−1 and (X−1 )− := (I − Π−1 )X−1 are both completions of X − endowed
with the norm k · k−1 , we can identify (X − )−1 and (X−1 )− (see, e.g., the footnote 3
of p. 1213 in [13]). We also decompose the feedback operator F ∈ B(X, U ):
F + := F |X + ,
F − := F |X − .
In addtion to Assumption 4.1, we impose the following assumptions:
Assumption 4.2. The exponential growth ω(T − ) satisfies ω(T − ) < 0.
Assumption 4.3. (A+ , B + ) is controllable.
Remark 4.4. Define the operator ABF : D(ABF ) ⊂ X → X by
(4.4)
ABF x = (A + BF )x
with domain
D(ABF ) := {x ∈ X : (A + BF )x ∈ X}.
It is shown in [14] that if A generates an analytic semigroup and if there exists
a compact operator F ∈ B(X, U ) such that the semigroup generated by ABF is
exponentially stable, then Assumptions 4.1–4.3 hold.
EVENT-TRIGGERED CONTROL OF INFINITE-DIMENSIONAL SYSTEMS
15
4.1.2. Event-triggered control. For every x ∈ X, define x+ := Πx and x− :=
(I − Π)x. We set a feedback operator F ∈ B(X, U ) to be F − = 0. Then the control
input is given by
(4.5)
u(t) = F x(tk ) = F + x+ (tk )
∀t ∈ [tk , tk+1 ), ∀k ∈ N0 .
In this section, we set the time sequence {tk }∞
k=0 so that the finite-dimensional system
(4.6a)
x+ (0) = (x0 )+ ∈ X +
(4.6b)
ẋ+ (t) = A+ x+ (t) + B + F + x+ (tk )
∀t ∈ (tk , tk+1 ), ∀k ∈ N0 .
is exponentially stable with decay rate β ≥ −ω(T − ). For example, as in Theorem
3.1, we can set a time sequence {tk }∞
k=0 recursively satisfying
(4.7)
n
o
∀k ∈ N0 .
tk+1 := min tk + τM , inf t > tk : kF + x+ (t) − u(tk )kU > εkx+ (tk )k
Similarly to Theorem 3.4, we can use the following event trigger:
(4.8)
n
o
tk+1 := min tk + τM , inf t > tk : kF + x+ (t) − u(tk )kU > εkx+ (t)k
∀k ∈ N0 .
Since X + is finite dimensional, we obtain less conservative conditions on the
event-trigger parameter ε for the feedback system (4.6) to be stable. In particular,
we obtain the following result on the event trigger (4.8):
Proposition 4.5. Consider the finite-dimensional system (4.6) and the event
trigger (4.8) with t0 = 0. Assume that the input space U is finite dimensional and
set β > 0. If there exist positive matrices P , Q and a positive scalar κ such that the
following linear matrix inequalities are feasible:
Q − ε2 κI −P B +
(4.9a)
0
−(B + )∗ P
κI
(4.9b)
(A+ + B + F + )∗ P + P (A+ + B + F + ) −β 2 P − Q,
then the finite-dimensional system (4.6) is exponential stable with decay rate β for all
τM > 0.
Similarly to Theorem III.3 in [5], we can prove Proposition 4.5, of which proof can be
found in the Appendix B.
Theorem 4.6. Let Assumptions 4.1–4.3 hold, and assume F − = 0. If the feedback gain F + and the time sequence {tk }∞
k=0 are chosen so that the finite-dimensional
system (4.6) is exponentially stable with decay rate γ0 ≥ −ω(T − ) and there exist
τM ≥ τm > 0 such that τm ≤ tk+1 − tk ≤ τM for all k ∈ N0 , then the infinitedimensional system (2.2) is exponentially stable with decay rate γ ≥ ω(T − ) − ε for
every ε > 0.
Rτ
Rτ
Proof. Setting Sτ+ = 0 T + (s)B + ds and Sτ− = 0 T − (s)B − ds, we have from
(2.3) that for every τ ∈ (0, tk+1 − tk ] and every k ∈ N0 ,
+
+
+
+
x+ (tk )
x (tk + τ )
T (τ )
0
x (tk )
Sτ +
F
0
=
+
(4.10)
.
x− (tk + τ )
0
T − (τ ) x− (tk )
Sτ−
x− (tk )
16
MASASHI WAKAIKI AND HIDEKI SANO
The above x+ is a unique solution of (4.6), and hence there exists M0 ≥ 1 such that
kx+ (t)k ≤ M0 e−γ0 t kx+ (0)k
(4.11)
∀t ≥ 0.
On the other hand, it follows from (4.10) that for every τ ∈ (0, tk+1 − tk ] and every
k ∈ N0 ,
x− (tk + τ ) = T − (τ )x− (tk ) + Sτ− F + x+ (tk )
= T − (tk + τ )x− (0) + Sτ F + x+ (tk )
+
k
X
T − (tk + τ − t` )St−` −t`−1 F + x+ (t`−1 ).
`=1
Since B − ∈ B(U, (X−1 )− ) and since T − (t) is a strongly continuous semigroup on
(X − )−1 = (X−1 )− , it follows from Lemma 2.3 that Sτ− ∈ B(U, X − ) for every τ > 0
and that there exists L ≥ 0 such that
(4.12)
Sτ− F +
sup
τm ≤τ ≤τM
B(X + ,X − )
≤ L.
Therefore, for all τ ∈ (0, tk+1 − tk ] and all k ∈ N0 ,
kx− (tk + τ )k ≤ kT − (tk + τ )kB(X − ) · kx− (0)k + Lkx+ (tk )k
(4.13)
+
k
X
kT − (tk + τ − t` )kB(X − ) · Lkx+ (t`−1 )k
`=1
Note that k is the number of the events that occur during [0, tk + τ ) for all τ ∈
(0, tk+1 − tk ] and hence satisfies
k≤
(4.14)
tk + τ
τm
∀τ ∈ (0, tk+1 − tk ].
Moreover, for every γ1 ∈ (0, −ω(T − )), there exists M1 ≥ 1 such that
kT − (t)kB(X − ) ≤ M1 e−γ1 t
(4.15)
∀t ≥ 0.
From (4.11)–(4.15), for all τ ∈ (0, tk+1 − tk ] and all k ∈ N0 ,
kx− (tk + τ )k ≤ M1 e−γ1 (tk +τ ) kx− (0)k + LM0 e−γ1 tk kx+ (0)k
+
k
X
M1 e−γ1 (tk +τ −t` ) · LM0 e−γ1 t`−1 kx+ (0)k
`=1
−γ1 (tk +τ )
kx− (0)k + (k + 1)LM0 M1 e−γ1 (tk +τ −τM ) kx+ (0)k
tk + τ
+ 1 LM0 M1 e−γ1 (tk +τ −τM ) kx+ (0)k
≤ M1 e−γ1 (tk +τ ) kx− (0)k +
τm
≤ M1 e
Note that, for every γ2 ∈ (0, γ1 ), there exists M2 ≥ 1 such that
t
∀t ≥ 0.
+ 1 e−γ1 t ≤ M2 e−γ2 t
τm
Then for every τ ∈ (0, tk+1 − tk ] and k ∈ N0 ,
(4.16)
kx− (tk + τ )k ≤ M1 e−γ1 (tk +τ ) kx− (0)k + M3 e−γ2 (tk +τ ) kx+ (0)k,
EVENT-TRIGGERED CONTROL OF INFINITE-DIMENSIONAL SYSTEMS
17
where M3 := LM0 M1 M2 eγ1 τM . Since
kx+ (0)k ≤ kΠkB(X) · kx0 k,
kx− (0)k ≤ 1 + kΠkB(X) · kx0 k,
it follows from (4.11) and (4.16) that
kx+ (t)k ≤ M + e−γ0 t kx0 k,
kx− (t)k ≤ M − e−γ2 t kx0 k
∀t ≥ 0,
where
M + := M0 kΠkB(X) ,
M − := M1 + (M1 + M3 )kΠkB(X) .
Thus we obtain
kx(t)k = kx+ + x− k ≤ kx+ k + kx− k ≤ (M + + M − )e−γ2 t kx0 k
∀t ≥ 0.
Since γ1 ∈ (0, −ω(T − )) and γ2 ∈ (0, γ1 ) were arbitrary, the infinite-dimensional system (2.2) is exponentially stable with decay rate γ ≥ ω(T − ) − ε for every ε > 0.
4.2. Stability under periodic event triggers. In Theorem 4.6, the feedback
operator F ∈ B(X, U ) has a specific structure F − = F |X − = 0. In contrast, we here
assume that A generates an analytic semigroup, and use a periodic event trigger proposed in [8, 9]. Then we see that for every compact feedback operator F ∈ B(U, X),
there exists a periodic event trigger such that the closed-loop system (2.2) is exponential stability.
Fixing h > 0, ε > 0, and `M ∈ N, we set the time sequence {tk }∞
k=0 to be t0 = 0
and
(4.17a)
ψk := min t > tk : kF x(t) − u(tk )kU > εkx(tk )k ∧ t = `h with ` ∈ N
(4.17b) tk+1 := min{tk + `M h, ψk }
∀k ∈ N0 .
We call this event trigger a periodic event trigger. The above time sequence {tk }∞
k=0
satisfies tk ∈ {`h : ` ∈ N0 } and τ ≤ tk+1 − tk ≤ `M h for every k ∈ N0 .
Theorem 4.7. Assume that A generates an analytic semigroup T (t) on X, B ∈
B(U, X−1 ), and F ∈ B(X, U ) is compact. If the semigroup generated by ABF in
(4.4) is exponentially stable, then there exist h∗ > 0 and ε∗ > 0 such that for every
h ∈ (0, h∗ ) and every ε ∈ (0, ε∗ ), the infinite-dimensional system (2.2) with periodic
event trigger (4.17a) is exponentially stable.
Proof. By Theorem 4.8 in [14], there exists h∗ > 0 such that for every h ∈ (0, h∗ ),
the operator ∆h ∈ B(X) defined by
∆h := T (h) + Sh F
is power stable, i.e., there exist M ≥ 1 and δ < 1 such that
k∆kh kB(X) ≤ M δ k
∀k ∈ N0 .
Fix h ∈ (0, h∗ ) and let tk = `k h for every k ∈ N0 . Then under the periodic event
trigger (4.17a), we have from (2.3) that for all p = 0, . . . , `k+1 − `k and all k ∈ N0 ,
x((`k + p + 1)h) = T (h)x((`k + p)h) + Sh u((`k + p)h)
From Lemma 2.3, it follows that Sh ∈ B(U, X), and hence the proof of Theorem 4.7
is the discrete-time counterpart of Theorem 3.1.
18
MASASHI WAKAIKI AND HIDEKI SANO
Define the implementation-induced error e by
(4.18) e((`k + p)h)) := u(`k h) − F x((`k + p)h) ∀p = 0, . . . , `k+1 − `k − 1, ∀k ∈ N0 .
For every p = 0, . . . , `k+1 − `k − 1 and every k ∈ N0 ,
x((`k + p + 1)h) = T (h)x((`k + p)h) + Sh F x((`k + p)h) + Sh e((`k + p)h)
= ∆h x((`k + p)h) + Sh e((`k + p)h).
Hence, for all k ∈ N0 ,
`k+1 −`k −1
(4.19)
`
x(`k+1 h) = ∆hk+1
−`k
x(`k h) +
X
`
∆hk+1
−`k −p−1
Sh e((`k + p)h).
p=0
We introduce a new norm | · |d on X defined by
|x|d := sup kδ −` ∆`h xk,
`≥0
and as (3.5) and (3.6), this norm has the following properties:
(4.20)
|∆kh x|d ≤ δ k |x|d
kxk ≤ |x|d ≤ M kxk,
∀x ∈ X.
Under the periodic event trigger (4.17a), the implementation-induced error e in (4.18)
satisfies
(4.21)
ke((`k + p)h)k ≤ εkx(`k h)k
∀p = 0, . . . , `k+1 − `k − 1, ∀k ∈ N0 .
Combining (4.19)–(4.21), we obtain
`k+1 −`k −1
|x(`k+1 h)|d ≤ δ `k+1 −`k |x(`k h)|d + M kSh kB(U,X)
X
δ `k+1 −`k −p−1 |x(`k h)|d
p=0
= δ
`k+1 −`k
(1 − ε0 ) + ε0 · |x(`k h)|d
where
ε0 := ε
∀k ∈ N0 ,
M kSh kB(U,X)
.
1−δ
Choose the event-trigger parameter ε > 0 so that δ(1 − ε0 ) + ε0 < 1, namely,
ε<
1−δ
.
M kSh kB(U,X)
If we define γ ∈ R by
γ :=
min
`=1,...,`M
− log(δ ` (1 − ε0 ) + ε0 )
,
`h
then γ > 0, and applying (4.20), we obtain
kx(`k+1 h)k ≤ |x(`k+1 h)|d ≤ e−γ(`k+1 −`k )h |x(`k h)|d
(4.22)
≤ e−γ`k+1 h |x(0)|d ≤ M e−γ`k+1 h kx0 k
∀k ∈ N0 .
EVENT-TRIGGERED CONTROL OF INFINITE-DIMENSIONAL SYSTEMS
19
From Lemma 2.3, there exists L ≥ 1 such that
kT (τ )kB(X) + kSτ F kB(X) ≤ L
∀τ ∈ [0, `M h].
Therefore, we see from (2.3) and (4.22) that
kx(`k h + τ )k ≤ kT (τ )kB(X) · kx(`k h)k + kSτ F kB(X) · kx(`k h)k
≤ Lkx(`k h)k
≤ eγ`M h LM · e−γ(`k h+τ ) kx0 k
∀τ ∈ (0, (`k+1 − `k )h], ∀k ∈ N0 .
Thus, the infinite-dimensional system (2.2) with periodic event trigger (4.17a) is exponentially stable with decay rate γ > 0.
Remark 4.8. As easily seen from the proof of Theorem 4.7, the same conclusion
holds for the following event trigger based on the difference of the state:
ψk := min t > tk : kx(t) − x(tk )k > εkx(tk )k ∧ t = `h with ` ∈ N
tk+1 := min{tk + `M h, ψk }
∀k ∈ N0 .
Remark 4.9. In Theorem 4.7, we assume that A generates an analytic semigroup
and that F is compact. These assumptions are used for the existence of sampling periods with respect to which the periodic sampled-data system is exponentially stable,
and we can replace it with different assumptions such as those of Corollary 2.3 in [20].
4.3. Example. We illustrate the event-triggered control method in Theorem
4.6 through an Euler-Bernoulli beam with structural damping [15]. Let ξ ∈ [0, 1] and
t ≥ 0 denote the space and time variables. We assume that the Euler-Bernoulli beam
is hinged at the one end of the beam ξ = 0 and has a freely sliding clamped end at the
other end ξ = 1. Suppose that the shear force u(t) is applied at ξ = 1. The dynamics
of the Euler-Bernoulli beam is given by
(4.24a)
(4.24b)
∂2z
∂3z
∂4z
(ξ,
t)
−
2γ
(ξ,
t)
+
(ξ, t) = 0
∂t2
∂ξ 2 ∂t
∂ξ 4
∂2z
∂z
z(0, t) = 0,
(0, t) = 0,
(1, t) = 0,
∂ξ 2
∂ξ
∀t ≥ 0
−
∂3z
(1, t) = u(t),
∂ξ 3
where z(ξ, t) is the lateral deflection of the beam at time t > 0 and location ξ along
the beam and γ ∈ (0, 1) is the damping constant.
4.3.1. Abstract evolution equation of Euler-Bernoulli beams. We here
recall the results developed in Sec. 5 of [15]: an abstract evolution equation of the
form (2.2a) for the Euler-Bernoulli beam (4.24) and the expansions of the generator
A and the input operator B with respect to a certain basis.
We introduce the operator A0 : D(A0 ) ⊂ L2 (0, 1) → L2 (0, 1),
A0 h =
d4 h
dξ 4
with domain
D(A0 ) = {h ∈ W 4,2 (0, 1) : h(0) = 0, h00 (0) = 0, h0 (1) = 0, h000 (1) = 0}.
20
MASASHI WAKAIKI AND HIDEKI SANO
1
We consider the state space X = D(A02 ) × L2 (0, 1), which is a Hilbert space with the
inner product
1
1
x1
y1
,
= hA02 x1 , A02 y1 iL2 + hx2 , y2 iL2 .
x2
y2
Define the operator
"
0
A=
−A0
#
I
1
with domain D(A) = D(A0 ) × D(A02 )
1
2
−2γA0
and introduce the state vector
"
#
z(·, t)
x(t) := ∂z
.
(·, t)
∂t
Then we can rewrite (4.24) in the form ẋ = Ax for the case u = 0. Using the technique
of [10], we can set the input operator B ∈ B(R, X−1 ) to be
0
Bu = u
, where δ1 denotes the Dirac distribution with support at ξ = 1.
δ1
Let us next obtain the expansions of A with respect to a Riesz basis of X. The
eigenvalues of A are given by
2
π
p
∀n ∈ N,
(4.25)
λ±n = −γ ± i 1 − γ 2 · − + nπ
2
and the associate eigenvectors are
√
f±n =
2
en /λ±n
2
p
en
1 − −γ ∓ i 1 − γ 2
where en ∈ L2 (0, 1) is defined by
π
√
en (ξ) := 2 sin − + nπ ξ
2
∀n ∈ N,
∀ξ ∈ [0, 1], ∀n ∈ N
and is the eigenvector of A0 that associated with the eigenvalue (−π/2 + nπ)4 . Define
Z∗ := Z \ {0}. Then we see that the eigenvectors {fn }n∈Z∗ is a Riesz basis of X.
Furthermore,
1 −en /λ∓n
g±n = √
∀n ∈ N,
en
2
are the eigenvectors of A∗ with associated eigenvalues λ±n such that {fn }n∈Z∗ and
{gn }n∈Z∗ are biorthogonal. Thus A is a Riesz spectral operator in Definition 2.3.4 on
p. 41 in [3], and it follows from Theorem 2.3.5 on p. 41 in [3] that
X
(4.26)
Ax =
λn hx, gn ifn
∀x ∈ D(A).
n∈Z∗
and
(
D(A) =
x∈X:
)
X
n∈Z∗
2
2
|λn | · |hx, gn i| < ∞ .
EVENT-TRIGGERED CONTROL OF INFINITE-DIMENSIONAL SYSTEMS
21
Moreover, A is the infinitesimal generator of an analytic semigroup T (t) given by
X
T (t)x =
eλn t hx, gn ifn
∀x ∈ X.
n∈Z∗
Noting that {fn }n∈Z∗ is also a Schauder basis of X−1 , we can expand B in X−1 in
the following way:
X
(4.27)
Bu = u
(−1)|n|+1 fn
∀u ∈ R.
n∈Z∗
4.3.2. Numerical simulation. Since σ(A) = {λn : n ∈ Z∗ }, we see that
Assumption 4.1 holds for every α < 0. We here set α ∈ (−9γπ 2 /4, −γπ 2 /4) in the
system decomposition of Sec. 5.1 and check where or not Assumptions 4.2 and 4.3.
In this case, σ(A) ∩ Cα = {λ−1 , λ1 } and ω(T − ) = −9γπ 2 /4 < 0. The subspace X +
is spanned by {f−1 , f1 }, and using this basis and the expansions (4.26), (4.27), we
can rewrite A+ and B + as
λ
0
1
A+ = −1
, B+ =
.
0
λ1
1
Clearly, (A+ , B + ) is controllable. Thus Assumptions 4.2 and 4.3 are satisfied.
Let γ = 1/15 and set the feedback gain F to be
F + x = F |X + x = −
13 2
γπ (hx, f−1 i + hx, f1 i) ,
4
F − = F |X − = 0.
Similarly to A+ and B + , we can rewrite F + with respect to the basis {f−1 , f1 } in the
following way:
13
F + = − γπ 2 1 1 .
4
We see from Proposition 4.5 that if the event-trigger parameter ε > 0 satisfies ε ≤ 0.54,
then this finite-dimensional system (4.6) with the event trigger (4.8) is exponential
stable with decay rate β ≥ −ω(T − ) = 9γπ 2 /4.
Fig. 1 illustrates the time responses of the Euler-Bernoulli beam with initial states
z(ξ, 0) = 1 − cos(πξ),
∂z
(ξ, 0) = 0
∂t
ξ ∈ [0, 1].
We apply the event trigger (4.8) with parameters ε = 0.54 and τM = 1 and approximate the state space X by the linear span of {fn : n ∈ Z∗ , |n| ≤ 15} for the
computation of the time responses. Fig. 1a shows that the time responses of kz(·, t)kL2
are close between the event-triggered control and the continuous-time control, and we
observe from Fig. 1b that the event trigger (4.8) can reduce the number of the control
updates.
Appendix A. Derivation of (3.8). Assume that u ∈ L1 (0, t; U ), the Hilbert
space of integrable vector-valued functions on (0, t) with the usual norm, and consider
the equation: ẋ(t) = Ax(t) + Bu(t) with bounded input operator B ∈ B(U, X). Since
TBF (t) satisfies
Z
(A.1)
t
TBF (t − s)BF T (s)xds
TBF (t)x = T (t)x +
0
∀x ∈ X,
22
MASASHI WAKAIKI AND HIDEKI SANO
kz(·, t)kL2
77
66
Event triggered control
55
Continuous-time control
No control
44
33
22
11
0 00
0
0.5
1
1.5
1
2
2
2.5
3
3
Time t
(a) Position norm kz(·, t)kL2 .
54
Event triggered control
u(t)
43
Continuous-time control
32
21
10
0-1 0
0
0.5
1
1.5
1
2
2
2.5
3
3
Time t
(b) Input u(t).
Fig. 1: Time response.
as shown, e.g., in Theorem 3.2.1 on p. 110 in [3], we obtain
Z
0
t
TBF (t − s)BF x(s)ds
Z t
Z
=
TBF (t − s)BF T (s)x0 +
0
s
T (s − α)Bu(α)dα ds
0
Z tZ s
TBF (t − s)BF T (s)x0 ds +
TBF (t − s)BF T (s − α)Bu(α)dαds
0
0
0
Z tZ s
= TBF (t)x0 − T (t)x0 +
TBF (t − s)BF T (s − α)Bu(α)dαds.
Z
t
=
0
0
EVENT-TRIGGERED CONTROL OF INFINITE-DIMENSIONAL SYSTEMS
23
On the other hand, from (A.1) and Fubini’s theorem, we also have
Z
0
t
TBF (t − s)Bu(s)ds
Z t
Z
=
T (t − s)Bu(s) +
0
Z
t−s
TBF (t − s − α)BF T (α)Bu(s)dα ds
0
t
Z tZ
t
TBF (t − β)BF T (β − s)Bu(s)dβds
T (t − s)Bu(s)ds +
=
Z
s
0
0
t
β
Z tZ
TBF (t − β)BF T (β − s)Bu(s)dsdβ.
T (t − s)Bu(s)ds +
=
0
0
0
Therefore,
TBF (t)x0 +
t
Z
TBF (t − s)B(u(s) − F x(s))ds
0
0
= T (t)x0 +
Z tZ
β
T (t − s)Bu(s)ds +
= T (t)x +
Z tZ
−
0
t
Z
0
s
TBF (t − β)BF T (β − s)Bu(s)dsdβ
0
0
TBF (t − s)BF T (s − α)Bu(α)dαds
0
Z
t
T (t − s)Bu(s)ds.
0
Thus, we obtain (3.8).
Appendix B. Proof of Proposition 4.5. For simplicity of notation, we omit
the superscript +. Define the implementation-induced error e(t) by
e(t) := F x(tk ) − F x(t)
∀t ∈ [tk , tk+1 ), ∀k ∈ N0 .
Then we obtain
∀t ∈ (tk , tk+1 ), ∀k ∈ N0 .
ẋ(t) = (A + BF )x(t) + Be(t)
If we define the Lyapunov function V (t) := x(t)∗ P x(t), then using the Lyapunov
inequality (4.9b), we obtain
dV
(t) ≤ −β 2 V (t) − x∗ (t)Qx(t) + x∗ (t)P Be(t) + e∗ (t)B ∗ P x(t)
dt
∗
x(t)
−Q P B x(t)
= −β 2 V (t) +
∀t ∈ (tk , tk+1 ), ∀k ∈ N0 .
e(t)
B∗P
0
e(t)
Moreover, we see from (4.9a) that if x(t) and e(t) satisfies
(B.1)
then
∗ 2
x(t)
−ε I
e(t)
0
∗
x(t)
−Q
e(t)
B∗P
0
I
x(t)
≥ 0,
e(t)
PB
0
x(t)
≥ 0.
e(t)
24
MASASHI WAKAIKI AND HIDEKI SANO
Since the condition (B.1) holds for all t ≥ 0 under the event-triggered condition (4.8)
with arbitrary τM > 0, it follows that
dV
(t) ≤ −β 2 V (t)
dt
∀t ∈ (tk , tk+1 ), ∀k ∈ N0 .
Since P is a positive definite matrix, we see that the finite-dimensional feedback
system (4.6) is exponential stable with decay rate β. This completes the proof.
REFERENCES
[1] D. P. Borgers and W. P. M. H. Heemels, Event-separation properties of event-triggered
control systems, IEEE Trans. Automat. Control, 59 (2014), pp. 2644–2656.
[2] R. F. Curtain and J. C. Oostveen, Necessary and sucient conditions for strong stability of
distributed parameter systems, Systems Control Lett., 37 (1999), pp. 11–18.
[3] R. F. Curtain and H. J. Zwart, An Introduction to Infinite-Dimensional Linear Systems
Theory, New York: Springer, 1995.
[4] V. S. Dolk, D. P. Borgers, and W. P. M. H. Heemels, Output-based and decentralized
dynamic event-triggered control with guaranteed Lp -gain performance and Zeno-freeness,
IEEE Trans. Automat. Control, 62 (2017), pp. 34–49.
[5] M. C. F. Donkers and W. P. M. H. Heemels, Output-based event-triggered control with guaranteed L∞ -gain and improved and decentralized event-triggering, IEEE Trans. Automat.
Control, 57 (2012), pp. 1362–1376.
[6] N. Espitia, A. Girard, N. Marchand, and C. Prieur, Event-based control of linear hyperbolic
systems of conservation laws, Automatica, 70 (2016), pp. 275–287.
[7] R. Goebel, R. Sanfelice, and A. Teel, Hybrid dynamical systems, IEEE Control Syst. Mag.,
29 (2009), pp. 28–93.
[8] W. P. M. H. Heemels and M. C. F. Donkers, Model-based periodic event-triggered control
for linear systems, Automatica, 49 (2013), pp. 698–711.
[9] W. P. M. H. Heemels, M. C. F. Donkers, and A. R. Teel, Periodic event-triggered control
for linear systems, IEEE Trans. Automat. Control, 58 (2013), pp. 847–861.
[10] L. F. Ho and D. L. Russell, Admissible input elements for systems in Hilbert space and a
Carleson measure criterion, SIAM J. Control Optim., 21 (1983), pp. 614–640.
[11] Z. Jiang, B. Cui, W. Wu, and B. Zhuang, Event-driven observer-based control for distributed
parameter systems using mobile sensor and actuator, Comput. Math. Appl., 72 (2016),
pp. 2854–2864.
[12] D. Lehmann and J. Lunze, Event-based control with communication delays and packet losses,
Int. J. Control, 85 (2012), pp. 563–577.
[13] H. Logemann, Stabilization of well-posed infinite-dimensional systems by dynamic sampleddata feedback, SIAM J. Control Optim., 51 (2013), pp. 1203–1231.
[14] H. Logemann, R. Rebarber, and S. Townley, Stability of infinite-dimensional sampled-data
systems, Trans. Amer. Math. Soc., 355 (2003), pp. 3301–3328.
[15] H. Logemann, R. Rebarber, and S. Townley, Generalized sampled-data stabilization of wellposed linear infinite-dimensional systems, SIAM J. Control Optim., 44 (2005), pp. 1345–
1369.
[16] A. Pazy, On the applicability of Lyapunov’s theorem in Hiilbert space, SIAM J. Math. Anal.,
3 (1972), pp. 291–294.
[17] A. Pazy, Semigroups of Linear Operators and Applications to Partial Differential Equations,
New York: Springer, 1983.
[18] R. Rebarber and S. Townley, Generalized sampled data feedback control of distributed parameter systems, Systems & Control Letters, 34 (1998), pp. 229–240.
[19] R. Rebarber and S. Townley, Nonrobustness of closed-loop stability for infinite-dimensional
systems under sample and hold, IEEE Trans. Automat. Control, 47 (2002), pp. 1381–1385.
[20] R. Rebarber and S. Townley, Robustness with respect to sampling for stabilization of Riesz
spectral systems, IEEE Trans. Automat. Control, 51 (2006), pp. 1519–1522.
[21] A. Selivanov and E. Fridman, Distributed event-triggered control of diffusion semilinear
PDEs, Automatica, 68 (2016), pp. 344–351.
[22] P. Tabuada, Event-triggered real-time scheduling of stabilizing control tasks, IEEE Trans.
Automat. Control, 52 (2007), pp. 1680–1685.
| 3cs.SY
|
arXiv:1711.07068v1 [cs.CV] 19 Nov 2017
Diverse and Accurate Image Description Using a
Variational Auto-Encoder with an Additive Gaussian
Encoding Space
Liwei Wang
Alexander G. Schwing
Svetlana Lazebnik
{lwang97, aschwing, slazebni}@illinois.edu
University of Illinois at Urbana-Champaign
Abstract
This paper explores image caption generation using conditional variational autoencoders (CVAEs). Standard CVAEs with a fixed Gaussian prior yield descriptions
with too little variability. Instead, we propose two models that explicitly structure
the latent space around K components corresponding to different types of image
content, and combine components to create priors for images that contain multiple
types of content simultaneously (e.g., several kinds of objects). Our first model
uses a Gaussian Mixture model (GMM) prior, while the second one defines a novel
Additive Gaussian (AG) prior that linearly combines component means. We show
that both models produce captions that are more diverse and more accurate than
a strong LSTM baseline or a “vanilla” CVAE with a fixed Gaussian prior, with
AG-CVAE showing particular promise.
1
Introduction
Automatic image captioning [9, 11, 18–20, 24] is a challenging open-ended conditional generation
task. State-of-the-art captioning techniques [23, 32, 36, 1] are based on recurrent neural nets with
long-short term memory (LSTM) units [13], which take as input a feature representation of a provided
image, and are trained to maximize the likelihood of reference human descriptions. Such methods are
good at producing relatively short, generic captions that roughly fit the image content, but they are
unsuited for sampling multiple diverse candidate captions given the image. The ability to generate
such candidates is valuable because captioning is profoundly ambiguous: not only can the same image
be described in many different ways, but also, images can be hard to interpret even for humans, let
alone machines relying on imperfect visual features. In short, we would like the posterior distribution
of captions given the image, as estimated by our model, to accurately capture both the open-ended
nature of language and any uncertainty about what is depicted in the image.
Achieving more diverse image description is a major theme in several recent works [6, 14, 27, 31, 35].
Deep generative models are a natural fit for this goal, and to date, Generative Adversarial Models
(GANs) have attracted the most attention. Dai et al. [6] proposed jointly learning a generator to
produce descriptions and an evaluator to assess how well a description fits the image. Shetty et
al. [27] changed the training objective of the generator from reproducing ground-truth captions to
generating captions that are indistinguishable from those produced by humans.
In this paper, we also explore a generative model for image description, but unlike the GAN-style
training of [6, 27], we adopt the conditional variational auto-encoder (CVAE) formalism [17, 29].
Our starting point is the work of Jain et al. [14], who trained a “vanilla” CVAE to generate questions
given images. At training time, given an image and a sentence, the CVAE encoder samples a latent z
vector from a Gaussian distribution in the encoding space whose parameters (mean and variance)
come from a Gaussian prior with zero mean and unit variance. This z vector is then fed into a decoder
that uses it, together with the features of the input image, to generate a question. The encoder and the
decoder are jointly trained to maximize (an upper bound on) the likelihood of the reference questions
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
Predicted Object Labels:
‘person’ ‘cup’ ‘donut’ ‘dining table’
Predicted Object Labels:
‘cup’ ‘fork’ ‘knife’ ‘sandwich’ ‘dining table’ ‘mouse’
AG-CVAE:
a woman sitting at a table with a cup of coffee
a person sitting at a table with a cup of coffee
a table with two plates of donuts and a cup of coffee
a woman sitting at a table with a plate of coffee
a man sitting at a table with a plate of food
AG-CVAE:
a close up of a plate of food on a table
a table with a plate of food on it
a plate of food with a sandwich on it
a white plate topped with a plate of food
a plate of food on a table next to a cup of coffee
LSTM Baseline:
a close up of a table with two plates of coffee
a close up of a table with a plate of food
a close up of a plate of food on a table
a close up of a table with two plates of food
a close up of a table with plates of food
LSTM Baseline:
a close up of a plate of food on a table
a close up of a plate of food with a sandwich
a close up of a plate of food
a close up of a plate of food on a white plate
a close up of a plate of food with a sandwich on it
Figure 1: Example output of our proposed AG-CVAE approach compared to an LSTM baseline
(see Section 4 for details). For each method, we show top five sentences following consensus
re-ranking [10]. The captions produced by our method are both more diverse and more accurate.
Object Labels: ‘person’
AG-CVAE sentences:
a man and a woman standing in a room
a man and a woman are playing a game
a man standing next to a woman in a room
a man standing next to a woman in a field
a man standing next to a woman in a suit
Object Labels: ‘person’,‘bus’
AG-CVAE sentences:
a man and a woman sitting on a bus
a man and a woman sitting on a train
a man and woman sitting on a bus
a man and a woman sitting on a bench
a man and a woman are sitting on a bus
Object Labels: ‘person’, ‘remote’
AG-CVAE sentences:
a man and a woman playing a video game
a man and a woman are playing a video game
a man and woman are playing a video game
a man and a woman playing a game with a remote
a woman holding a nintendo wii game controller
Object Labels: ‘person’, ‘train’
AG-CVAE sentences:
a man and a woman sitting on a train
a woman and a woman sitting on a train
a woman sitting on a train next to a train
a woman sitting on a bench in a train
a man and a woman sitting on a bench
Figure 2: Illustration of how our additive latent space structure controls the image description process.
Modifying the object labels changes the weight vectors associated with semantic components in
the latent space. In turn, this shifts the mean from which the z vectors are drawn and modifies the
resulting descriptions in an intuitive way.
given the images. At test time, the decoder is seeded with an image feature and different z samples,
so that multiple z’s result in multiple questions.
While Jain et al. [14] obtained promising question generation performance with the above CVAE
model equipped with a fixed Gaussian prior, for the task of image captioning, we observed a tendency
for the learned conditional posteriors to collapse to a single mode, yielding little diversity in candidate
captions sampled given an image. To improve the behavior of the CVAE, we propose using a set of K
Gaussian priors in the latent z space with different means and standard deviations, corresponding to
different “modes” or types of image content. For concreteness, we identify these modes with specific
object categories, such as ‘dog’ or ‘cat.’ If ‘dog’ and ‘cat’ are detected in an image, we would like to
encourage the generated captions to capture both of them.
Starting with the idea of multiple Gaussian priors, we propose two different ways of structuring
the latent z space. The first is to represent the distribution of z vectors using a Gaussian Mixture
model (GMM). Due to the intractability of Gaussian mixtures in the VAE framework, we also
introduce a novel Additive Gaussian (AG) prior that directly adds multiple semantic aspects in the
z space. If an image contains several objects or aspects, each corresponding to means µk in the
latent space, then we require the mean of the encoder distribution to be close to a weighted linear
combination of the respective means. Our CVAE formulation with this additive Gaussian prior
(AG-CVAE) is able to model a richer, more flexible encoding space, resulting in more diverse and
accurate captions, as illustrated in Figure 1. As an additional advantage, the additive prior gives us an
interpretable mechanism for controlling the captions based on the image content, as shown in Figure
2. Experiments of Section 4 will show that both GMM-CVAE and AG-CVAE outperform LSTMs
and “vanilla” CVAE baselines on the challenging MSCOCO dataset [5], with AG-CVAE showing
marginally higher accuracy and by far the best diversity and controllability.
2
Background
Our proposed framework for image captioning extends the standard variational auto-encoder [17]
and its conditional variant [29]. We briefly set up the necessary background here.
Variational auto-encoder (VAE): Given samples x from a dataset, VAEs aim at modeling the data
likelihood p(x). To this end, VAEs assume that the data points x cluster around a low-dimensional
manifold parameterized by embeddings or encodings z. To obtain the sample x corresponding to an
embedding z, we employ the decoder p(x|z) which is often based on deep nets. Since the decoder’s
posterior p(z|x) is not tractably computable we approximate it with a distribution q(z|x) which is
2
referred to as the encoder. Taking together all those ingredients, VAEs are based on the identity
log p(x) − DKL [q(z|x), p(z|x)] = Eq(z|x) [log p(x|z)] − DKL [q(z|x), p(z)],
(1)
which relates the likelihood p(x) and the conditional p(z|x). It is hard to compute the KL-divergence
DKL [q(z|x), p(z|x)] because the posterior p(z|x) is not readily available from the decoder distribution p(x|z) if we use deep nets. However, by choosing an encoder distribution q(z|x) with sufficient
capacity, we can assume that the non-negative KL-divergence DKL [q(z|x), p(z|x)] is small. Thus,
we know that the right-hand-side is a lower bound on the log-likelihood log p(x), which can be
maximized w.r.t. both encoder and decoder parameters.
Conditional variational auto-encoders (CVAE): In tasks like image captioning, we are interested
in modeling the conditional distribution p(x|c), where x are the desired descriptions and c is some
representation of content of the input image. The VAE identity can be straightforwardly extended by
conditioning both the encoder and decoder distributions on c. Training of the encoder and decoder
proceeds by maximizing the lower bound on the conditional data-log-likelihood p(x|c), i.e.,
log pθ (x|c) ≥ Eqφ (z|x,c) [log pθ (x|z, c)] − DKL [qφ (z|x, c), p(z|c)] ,
(2)
where θ and φ, the parameters for the decoder distribution pθ (x|z, c) and the encoder distribution
qφ (z|x, c) respectively. In practice, the following stochastic objective is typically used:
max
θ,φ
N
1 X
log pθ (xi |z i , ci ) − DKL [qφ (z|x, c), p(z|c)],
N i=1
s.t. ∀i z i ∼ qφ (z|x, c).
It approximates the expectation Eqφ (z|x,c) [log pθ (x|z, c)] using N samples z i drawn from the approximate posterior qφ (z|x, c) (typically, just a single sample is used). Backpropagation through
the encoder that produces samples z i is achieved via the reparameterization trick [17], which is
applicable if we restrict the encoder distribution qφ (z|x, c) to be, e.g., a Gaussian with mean and
standard deviation output by a deep net.
3
Gaussian Mixture Prior and Additive Gaussian Prior
Our key observation is that the behavior of the trained CVAE crucially depends on the choice of
the prior p(z|c). The prior determines how the learned latent space is structured, because the KLdivergence term in Eq. (2) encourages qφ (z|x, c), the encoder distribution over z given a particular
description x and image content c, to be close to this prior distribution. In the vanilla CVAE
formulation, such as the one adopted in [14], the prior is not dependent on c and is fixed to a
zero-mean unit-variance Gaussian. While this choice is the most computationally convenient, our
experiments in Sec. 4 will demonstrate that for the task of image captioning, the resulting model has
poor diversity and worse accuracy than the standard maximum-likelihood-trained LSTM. Clearly, the
prior has to change based on the content of the image. However, because of the need to efficiently
compute the KL-divergence in closed form, it still needs to have a simple structure, ideally a Gaussian
or a mixture of Gaussians.
Motivated by the above considerations, we encourage the latent z space to have a multi-modal
structure composed of K modes or clusters, each corresponding to different types of image content.
Given an image I, we assume that we can obtain a distribution c(I) = (c1 (I), . . . , cK (I)), where the
entries ck are nonnegative and sum to one. In our current work, for concreteness, we identify these
with a set of object categories that can be reliably detected automatically, such as ‘car,’ ‘person,’ or
‘cat.’ The MSCOCO dataset, on which we conduct our experiments, has direct supervision for 80
such categories. Note, however, our formulation is general and can be applied to other definitions of
modes or clusters, including latent topics automatically obtained in an unsupervised fashion.
GMM-CVAE: We can model p(z|c) as a Gaussian mixture with weights ck and components with
means µk and standard deviations σk :
p(z|c) =
K
X
ck N z |µk , σk2 I ,
(3)
k=1
where ck is defined as the weights above and µk represents the mean vector of the k-th component.
In practice, for all components, we use the same standard deviation σ.
3
Decoder
Decoder
C3
C1
Cz
switch
C1 C2 C3
C2
Z
Z
Cluster
Vector
…
Cluster
Vector
C1 C2 C3
…
Encoder
Encoder
(a) GMM-CVAE
(b) AG-CVAE
Figure 3: Overview of GMM-CVAE and AG-CVAE models. To sample z vectors given an image,
GMM-CVAE (a) switches from one cluster center to another, while AG-CVAE (b) encourages the
embedding z for an image to be close to the average of its objects’ means.
It is not directly tractable to optimize Eq. (2) with the above GMM prior. We therefore approximate
the KL divergence stochastically [12]. In each step during training, we first draw a discrete component
k according to the cluster probability c(I), and then sample z from the resulting Gaussian component.
Then we have
1
1
σk
+ 2 Eqφ (z|x,ck ) kz − µk k22 −
DKL [qφ (z|x, ck ), p(z|ck )] = log
σφ
2σ
2
(4)
2
2
σφ + kµφ − µk k2
σk
1
= log
+
− , ∀k ck ∼ c(I).
σφ
2σk2
2
We plug the above KL term into Eq. (2) to obtain an objective function, which we optimize w.r.t. the
encoder and decoder parameters φ and θ using stochastic gradient descent (SGD). In principle, the
prior parameters µk and σk can also be trained, but we obtained good results by keeping them fixed
(the means are drawn randomly and all standard deviations are set to the same constant, as will be
further explained in Section 4).
At test time, in order to generate a description given an image I, we first sample a component index k
from c(I), and then sample z from the corresponding component distribution. One limitation of this
procedure is that, if an image contains multiple objects, each individual description is still conditioned
on just a single object.
AG-CVAE: We would like to structure the z space in a way that can directly reflect object cooccurrence. To this end, we propose a simple novel conditioning mechanism with an additive
Gaussian prior. If an image contains several objects with weights ck , each corresponding to means µk
in the latent space, we want the mean of the encoder distribution to be close to the linear combination
of the respective means with the same weights:
!
K
X
2
p(z|c) = N z
ck µk , σ I ,
(5)
k=1
PK
where σ 2 I is a spherical covariance matrix with σ 2 = k=1 c2k σk2 . Figure 3 illustrates the difference
between this AG-CVAE model and the GMM-CVAE model introduced above.
In order to train the AG-CVAE model using the objective of Eq. (2), we need to compute the
KL-divergence DKL [qφ (z|x, c), p(z|c)] where qφ (z|x, c) = N (z | µφ (x, c), σφ2 (x, c)I) and the prior
p(z|c) is given by Eq. (5). Its analytic expression can be derived to be
2
K
X
1
1
DKL [qφ (z|x, c), p(z|c)] = log
+ 2 Eqφ z −
ck µk −
2σ
2
k=1
P
K
σφ2 + kµφ − k=1 ck µk k2
1
σ
+
− .
= log
2
σφ
2σ
2
σ
σφ
We plug the above KL-divergence term into Eq. (2) to obtain the stochastic objective function for
training the encoder and decoder parameters. We initialize the mean and variance parameters µk and
σk in the same way as for GMM-CVAE and keep them fixed throughout training.
4
2
µ 1 , log
2
µ 2 , log(
)
Wc1
)
Wc2
2
2
Reconstruction
Loss
……
µ , log(
1
µ
K
, log(
2
K
)
WcK
hT
LSTM
LSTM
LSTM
LSTM
WI
Wc
We
We
……
Image
Feature
Cluster
Vector
w1
w2
……
LSTM
p0
p1
h0
h1
P(Real/Fake)
……
hT
LSTM
LSTM
LSTM
LSTM
LSTM
We
WI
Wc
Wz
We
We
……
We
wT
Image
Feature
Cluster
Vector
z
ws
w1
……
wT
LSTM
Figure 4: Illustration of our encoder (left) and decoder (right). See text for details.
Next, we need to specify our architectures for the encoder and decoder, which are shown in Fig. 4.
The encoder uses an LSTM to map an image I, its vector c(I), and a caption into a point in the
latent space. More specifically, the LSTM receives the image feature in the first step, the cluster
vector in the second step, and then the caption word by word. The hidden state hT after the last step
2
is transformed into K mean vectors, µφk , and K log variances, log σφk
, using a linear layer for each.
2
For AG-CVAE, the µφk and σφk are then summed with weights ck and c2k respectively to generate
the desired µφ and σφ2 encoder outputs. Note that the encoder is used at training time only, and the
input cluster vectors are produced from ground truth object annotations.
The decoder uses a different LSTM that receives as input first the image feature, then the cluster
vector, then a z vector sampled from the conditional distribution of Eq. (5). Next, it receives a ‘start’
symbol and proceeds to output a sentence word by word until it produces an ‘end’ symbol. During
training, its c(I) inputs are derived from the ground truth, same as for the encoder, and the log-loss is
used to encourage reconstruction of the provided ground-truth caption. At test time, ground truth
object vectors are not available, so we rely on automatic object detection, as explained in Section 4.
4
Experiments
4.1 Implementation Details
We test our methods on the MSCOCO dataset [5], which is the largest “clean” image captioning
dataset available to date. The current (2014) release contains 82,783 training and 40,504 validation
images with five reference captions each, but many captioning works re-partition this data to enlarge
the training set. We follow the train/val/test split released by [23]. It allocates 118, 287 images for
training, 4, 000 for validation, and 1, 000 for testing.
Features. As image features, we use 4,096-dimensional activations from the VGG-16 network [28].
The cluster or object vectors c(I) are 80-dimensional, corresponding to the 80 MSCOCO object
categories. At training time, c(I) consist of binary indicators corresponding to ground truth object
labels, rescaled to sum to one. For example, an image with labels ‘person,’ ‘car,’ and ‘dog’ results in a
cluster vector with weights of 1/3 for the corresponding objects and zeros elsewhere. For test images
I, c(I) are obtained automatically through object detection. We train a Faster R-CNN detector [26]
for the MSCOCO categories using our train/val split by fine-tuning the VGG-16 net [28]. At test
time, we use a threshold of 0.5 on the per-class confidence scores output by this detector to determine
whether the image contains a given object (i.e., all the weights are once again equal).
Baselines. Our LSTM baseline is obtained by deleting the z vector input from the decoder architecture shown in Fig. 4. This gives a strong baseline comparable to NeuralTalk2 [1] or Google Show
and Tell [33]. To generate different candidate sentences using the LSTM, we use beam search with
a width of 10. Our second baseline is given by the “vanilla” CVAE with a fixed Gaussian prior
following [14]. For completeness, we report the performance of our method as well as all baselines
both with and without the cluster vector input c(I).
Parameter settings and training. For all the LSTMs, we use a one-hot encoding with vocabulary
size of 11,488, which is the number of words in the training set. This input gets projected into a word
embedding layer of dimension 256, and the LSTM hidden space dimension is 512. We found that
the same LSTM settings worked well for all models. For our three models (CVAE, GMM-CVAE,
and AG-CVAE), we use a dimension of 150 for the z space. We wanted it to be at least equal to the
number of categories to make sure that each z vector corresponds to a unique set of cluster weights.
The means µk of clusters for GMM-CVAE and AG-CVAE are randomly initialized on the unit ball
5
obj #z std beam
B4
B3
B2
B1
C
R
M
S
10
0.413 0.515 0.643 0.790 1.157 0.597 0.285 0.218
LSTM
X
10
0.428 0.529 0.654 0.797 1.202 0.607 0.290 0.223
20 0.1
0.261 0.381 0.538 0.742 0.860 0.531 0.246 0.184
CVAE X 20
2
0.312 0.421 0.565 0.733 0.910 0.541 0.244 0.176
20 0.1
0.371 0.481 0.619 0.778 1.080 0.582 0.274 0.209
GMM- X 20
2
0.423 0.533 0.666 0.813 1.216 0.617 0.298 0.233
CVAE X 20
2
2
0.449 0.553 0.680 0.821 1.251 0.624 0.299 0.232
X 100 2
0.494 0.597 0.719 0.856 1.378 0.659 0.325 0.261
X 100 2
2
0.527 0.625 0.740 0.865 1.430 0.670 0.329 0.263
20 0.1
0.431 0.537 0.668 0.814 1.230 0.622 0.300 0.235
X 20
2
0.451 0.557 0.686 0.829 1.259 0.630 0.305 0.243
AGX 20
2
2
0.471 0.573 0.698 0.834 1.308 0.638 0.309 0.244
0.532 0.631 0.749 0.876 1.478 0.682 0.342 0.278
CVAE X 100 2
X 100 2
2
0.557 0.654 0.767 0.883 1.517 0.690 0.345 0.277
Table 1: Oracle (upper bound) performance according to each metric. Obj indicates whether the
object (cluster) vector is used; #z is the number of z samples; std is the test-time standard deviation;
beam is the beam width if beam search is used. For the caption quality metrics, C is short for Cider,
R for ROUGE, M for METEOR, S for SPICE.
obj #z std beam
B4
B3
B2
B1
C
R
M
10
0.286 0.388 0.529 0.702 0.915 0.510 0.235
LSTM
X
10
0.292 0.395 0.536 0.711 0.947 0.516 0.238
20 0.1
0.245 0.347 0.495 0.674 0.775 0.491 0.217
CVAE X 20
2
0.265 0.372 0.521 0.698 0.834 0.506 0.225
20 0.1
0.271 0.376 0.522 0.702 0.890 0.507 0.231
GMM- X 20
2
0.278 0.388 0.538 0.718 0.932 0.516 0.238
CVAE X 20
2
2
0.289 0.394 0.538 0.715 0.941 0.513 0.235
X 100 2
0.292 0.402 0.552 0.728 0.972 0.520 0.241
X 100 2
2
0.307 0.413 0.557 0.729 0.986 0.525 0.242
20 0.1
0.287 0.394 0.540 0.715 0.942 0.518 0.238
AGX 20
2
0.286 0.391 0.537 0.716 0.953 0.517 0.239
CVAE
X 20
2
2
0.299 0.402 0.544 0.716 0.963 0.518 0.237
X 100 2
0.301 0.410 0.557 0.732 0.991 0.527 0.243
X 100 2
2
0.311 0.417 0.559 0.732 1.001 0.528 0.245
Table 2: Consensus re-ranking using CIDEr. See caption of Table 1 for legend.
S
0.165
0.170
0.147
0.158
0.166
0.170
0.169
0.174
0.177
0.168
0.172
0.173
0.177
0.179
and are not changed throughout training. The standard deviations σk are set to 0.1 at training time and
tuned on the validation set at test time (the values used for our results are reported in the tables). All
networks are trained with SGD with a learning rate that is 0.01 for the first 5 epochs, and is reduced
by half every 5 epochs. On average all models converge within 50 epochs.
4.2 Results
A big part of the motivation for generating diverse candidate captions is the prospect of being able to
re-rank them using some discriminative method. Because the performance of any re-ranking method
is upper-bounded by the quality of the best candidate caption in the set, we will first evaluate different
methods assuming an oracle that can choose the best sentence among all the candidates. Next, for a
more realistic evaluation, we will use a consensus re-ranking approach [10] to automatically select a
single top candidate per image. Finally, we will assess the diversity of the generated captions using
uniqueness and novelty metrics.
Oracle evaluation. Table 1 reports caption evaluation metrics in the oracle setting, i.e., taking the
maximum of each relevant metric over all the candidates. We compare caption quality using five
metrics: BLEU [25], METEOR [7], CIDEr [30], SPICE [2], and ROUGE [21]. These are calculated
using the MSCOCO caption evaluation tool [5] augmented by the author of SPICE [2]. For the
LSTM baseline, we report the scores attained among 10 candidates generated using beam search (as
suggested in [23]). For CVAE, GMM-CVAE and AG-CVAE, we sample a fixed number of z vectors
from the corresponding prior distributions (the numbers of samples are given in the table).
The high-level trend is that “vanilla” CVAE falls short even of the LSTM baseline, while the upperbound performance for GMM-CVAE and AG-CVAE considerably exceeds that of the LSTM given
6
beam % unique % novel
size
per image sentences
LSTM X
10
0.656
CVAE X 20 2
0.118
0.820
X 20 2
0.594
0.809
GMM- X 20 2
2
0.539
0.716
CVAE X 100 2
0.376
0.767
X 100 2
2
0.326
0.688
X 20 2
0.764
0.795
AGX 20 2
2
0.698
0.707
CVAE X 100 2
0.550
0.745
X 100 2
2
0.474
0.667
Table 3: Diversity evaluation. For each method, we report the percentage of unique candidates
generated per image by sampling different numbers of z vectors. We also report the percentage of
novel sentences (i.e., sentences not seen in the training set) out of (at most) top 10 sentences following
consensus re-ranking. It should be noted that for CVAE, there are 2,466 novel sentences out of 3,006.
For GMM-CVAE and AG-CVAE, we get roughly 6,200-7,800 novel sentences.
obj
#z
std
Predicted Object Labels:
'bottle' 'refrigerator'
Predicted Object Labels:
'person' 'horse' 'bear'
AG-CVAE:
an open refrigerator filled with lots of food
a refrigerator filled with lots of food and drinks
a refrigerator filled with lots of food
a large open refrigerator filled with lots of food
a refrigerator filled with lots of food and other items
AG-CVAE:
a man standing next to a brown horse
a man is standing next to a horse
a person standing next to a brown and white horse
a man standing next to a horse and a man
a man holding a brown and white horse
LSTM Baseline:
a refrigerator filled with lots of food
a refrigerator filled with lots of food on top
a refrigerator filled with lots of food inside
a refrigerator filled with lots of food inside of it
a refrigerator filled with lots of food and other items
LSTM Baseline:
a close up of a person with a horse
a close up of a horse with a horse
a black and white photo of a man wearing a hat
a black and white photo of a person wearing a hat
a black and white photo of a man in a hat
(a)
(b)
Predicted Object Labels:
'person' 'backpack' 'umbrella'
Predicted Object Labels:
'person' ‘bed’
AG-CVAE:
a person holding an umbrella in front of a building
a woman holding a red umbrella in front of a building
a person holding an umbrella in the rain
a man and woman holding an umbrella in the rain
a man holding a red umbrella in front of a building
AG-CVAE:
a baby laying on a bed with a blanket
a woman laying on a bed with a baby
a man laying on a bed with a baby
a baby laying in a bed with a blanket
a baby is laying in bed with a cat
LSTM Baseline:
a man holding an umbrella on a city street
a man holding an umbrella in the rain
a man is holding an umbrella in the rain
a person holding an umbrella in the rain
a man holding an umbrella in the rain with an umbrella
LSTM Baseline:
a baby is laying on a bed with a blanket
a baby is laying on a bed with a stuffed animal
a little girl laying in a bed with a blanket
a little girl laying on a bed with a blanket
a man laying in a bed with a blanket
(c)
(d)
Figure 5: Comparison of captions produced by our AG-CVAE method and the LSTM baseline. For
each method, top five captions following consensus re-ranking are shown.
the right choice of standard deviation and a large enough number of z samples. AG-CVAE obtains the
highest upper bound. A big advantage of the CVAE variants over the LSTM is that they can be easily
used to generate more candidate sentences simply by increasing the number of z samples, while the
only way to do so for the LSTM is to increase the beam width, which is computationally prohibitive.
In more detail, the top two lines of Table 1 compare performance of the LSTM with and without the
additional object (cluster) vector input, and show that it does not make a dramatic difference. That is,
improving over the LSTM baseline is not just a matter of adding stronger conditioning information
as input. Similarly, for CVAE, GMM-CVAE, and AG-CVAE, using the object vector as additional
conditioning information in the encoder and decoder can increase accuracy somewhat, but does not
account for all the improvements that we see. One thing we noticed about the models without the
object vector is that they are more sensitive to the standard deviation parameter and require more
careful tuning (to demonstrate this, the table includes results for several values of σ for the CVAE
models).
Consensus re-ranking evaluation. For a more realistic evaluation we next compare the same models
after consensus re-ranking [10, 23]. Specifically, for a given test image, we first find its nearest
neighbors in the training set in the cross-modal embedding space learned by a two-branch network
proposed in [34]. Then we take all the ground-truth reference captions of those neighbors and
calculate the consensus re-ranking scores between them and the candidate captions. For this, we
use the CIDEr metric, based on the observation of [22, 30] that it can give more human-consistent
evaluations than BLEU.
7
Object Labels: ‘cat’ ‘suitcase’
GMM-CVAE:
a black and white cat is sitting in a suitcase
a cat that is sitting in a suitcase
a cat that is sitting on a suitcase
a cat sitting on top of a suitcase
a black and white cat sitting on a suitcase
a cat sitting in a suitcase on a table
AG-CVAE:
a small white and black cat sitting on top of a suitcase
a cat sitting on a piece of luggage
a small gray and white cat sitting in a suitcase
a white cat sitting on top of a suitcase
a black and white cat sitting in a suitcase
a black and white cat sitting on top of a suitcase
a cat that is sitting on a table
a black and white cat sitting next to a suitcase
a cat sitting in front of a suitcase
a cat sitting on a wooden bench in the sun
a close up of a cat sitting on a suitcase
a cat sitting on top of a blue suitcase
a large brown and white cat sitting on top of a suitcase
a cat sitting on top of a suitcase
a white cat with a suitcase
Object Labels: ‘cup’ ‘dining table’
‘teddy bear’
GMM-CVAE:
a teddy bear sitting next to a teddy bear
a teddy bear sitting on a table next to a table
a teddy bear sitting on top of a table
a teddy bear sitting on a table next to a cup of coffee
a stuffed teddy bear sitting next to a table
a stuffed teddy bear sitting on a table
a teddy bear sitting next to a table filled with stuffed animals
a teddy bear is sitting on a table
ateddy bear sitting on a table next to a teddy bear
AG-CVAE:
a white teddy bear sitting next to a table
a couple of stuffed animals sitting on a table
a teddy bear sitting next to a bunch of flowers
a couple of teddy bears sitting on a table
a large teddy bear sitting on a table
a bunch of stuffed animals sitting on a table
a group of teddy bears sitting on a table
a large teddy bear sitting on a table next to a table
a teddy bear sitting next to a pile of books
a group of teddy bears sitting next to each other
a white teddy bear sitting on a wooden table
two teddy bears sitting next to each other
a couple of teddy bears sitting next to each other
a white teddy bear sitting next to a table
a teddy bear sitting next to a wooden table
a large stuffed animal sitting on top of a table
Object Labels: ‘cat’ ‘suitcase’ ‘chair’
GMM-CVAE:
a cat that is sitting in a suitcase
a cat sitting on top of a suitcase
a cat sitting in a suitcase on the floor
a black and white cat is sitting in a suitcase
a close up of a cat on a suitcase
AG-CVAE:
a white and black cat sitting in a suitcase
a cat that is sitting on a chair
a white and black cat sitting on top of a suitcase
a black and white cat sitting on a chair
a cat sitting on a chair in a room
a large brown and white cat sitting on top of a desk
a cat sitting on a wooden bench in the sun
a close up of a cat sitting on a suitcase
a black and white cat sitting next to a piece of luggage
a small white and black cat sitting in a chair
a black and white cat sitting on top of a suitcase
a cat sitting on top of a blue chair
a cat sitting on top of a suitcase
Object Labels: ‘cup’ ‘dining table’
‘teddy bear’ ‘sandwich’ ‘cake’
GMM-CVAE:
a teddy bear sitting next to a teddy bear
a teddy bear sitting on a table next to a cup of coffee
a teddy bear sitting on a table with a teddy bear
a teddy bear with a teddy bear sitting on top of it
a teddy bear sitting on top of a table
a teddy bear sitting next to a cup of coffee
a table with a teddy bear and a teddy bear
a teddy bear sitting on a table next to a glass of coffee
two teddy bears sitting on a table next to each other
AG-CVAE:
a table topped with a cake
a couple of cake sitting on top of a table
a table with a cake and a bunch of stuffed animals
a cake with a bunch of coffee on it
a white teddy bear sitting next to a glass of coffee
a table with a cake and a bear on it
a table with a bunch of teddy bears
a table with two plates of food on it
a table topped with a variety of food
a table with two teddy bears
a table with a cake and a plate of food
a couple of sandwiches sitting on top of a table
a table topped with a cake and two plates of food
a table with a bunch of cakes on it
a table with a cake and a cup of coffee
a white plate of food next to a table
a white table topped with lots of food
Figure 6: Comparison of captions produced by GMM-CVAE and AG-CVAE for two different versions
of input object vectors for the same images. For both models, we draw 20 z samples and show the
resulting unique captions.
Table 2 shows the evaluation based on the single top-ranked sentence for each test image. While the
re-ranked performance cannot get near the upper bounds of Table 1, the numbers follow a similar
trend, with GMM-CVAE and AG-CVAE achieving better performance than the baselines in almost
all metrics. It should also be noted that, while it is not our goal to outperform the state of the art in
absolute terms, our performance is actually better than some of the best methods to date [23, 37],
although [37] was trained on a different split. AG-CVAE tends to get slightly higher numbers than
GMM-CVAE, although the advantage is smaller than for the upper-bound results in Table 1. One
of the most important take-aways for us is that there is still a big gap between upper-bound and
re-ranking performance and that improving re-ranking of candidate sentences is an important future
direction.
Diversity evaluation. To compare the generative capabilities of our different methods we report
two indicative numbers in Table 3. One is the average percentage of unique captions in the set of
candidates generated for each image. This number is only meaningful for the CVAE models, where
we sample candidates by drawing different z samples, and multiple z’s can result in the same caption.
For LSTM, the candidates are obtained using beam search and are by definition distinct. From Table
3, we observe that CVAE has very little diversity, GMM-CVAE is much better, but AG-CVAE has the
decisive advantage.
Similarly to [27], we also report the percentage of all generated sentences for the test set that have
not been seen in the training set. It only really makes sense to assess novelty for sentences that
are plausible, so we compute this percentage based on (at most) top 10 sentences per image after
consensus re-ranking. Based on the novelty ratio, CVAE does well. However, since it generates
fewer distinct candidates per image, the absolute numbers of novel sentences are much lower than for
GMM-CVAE and AG-CVAE (see table caption for details).
8
Qualitative results. Figure 5 compares captions generated by AG-CVAE and the LSTM baseline on
four example images. The AG-CVAE captions tend to exhibit a more diverse sentence structure with
a wider variety of nouns and verbs used to describe the same image. Often this yields captions that
are more accurate (‘open refrigerator’ vs. ‘refrigerator’ in (a)) and better reflective of the cardinality
and types of entities in the image (in (b), our captions mention both the person and the horse while the
LSTM tends to mention only one). Even when AG-CVAE does not manage to generate any correct
candidates, as in (d), it still gets the right number of people in some candidates. A shortcoming of
AG-CVAE is that detected objects frequently end up omitted from the candidate sentences if the
LSTM language model cannot accommodate them (‘bear’ in (b) and ‘backpack’ in (c)). On the
one hand, this shows that the capacity of the LSTM decoder to generate combinatorially complex
sentences is still limited, but on the other hand, it provides robustness against false positive detections.
Controllable sentence generation. Figure 6 illustrates how the output of our GMM-CVAE and
AG-CVAE models changes when we change the input object vectors in an attempt to control the
generation process. Consistent with Table 3, we observe that for the same number of z samples,
AG-CVAE produces more unique candidates than GMM-CVAE. Further, AG-CVAE is more flexible
than GMM-CVAE and more responsive to the content of the object vectors. For the first image
showing a cat, when we add the additional object label ‘chair,’ AG-CVAE is able to generate some
captions mentioning a chair, but GMM-CVAE is not. Similarly, in the second example, when we add
the concepts of ‘sandwich’ and ‘cake,’ only AG-CVAE can generate some sentences that capture
them. Still, the controllability of AG-CVAE leaves something to be desired, since, as observed above,
it has trouble mentioning more than two or three objects in the same sentence, especially in unusual
combinations.
5
Discussion
Our experiments have shown that both our proposed GMM-CVAE and AG-CVAE approaches
generate image captions that are more diverse and more accurate than standard LSTM baselines.
While GMM-CVAE and AG-CVAE have very similar bottom-line accuracies according to Table 2,
AG-CVAE has a clear edge in terms of diversity (unique captions per image) and controllability, both
quantitatively (Table 3) and qualitatively (Figure 6).
Related work. To date, CVAEs have been used for image question generation [14], but as far as we
know, our work is the first to apply them to captioning. In [8], a mixture of Gaussian prior is used in
CVAEs for colorization. Their approach is essentially similar to our GMM-CVAE, though it is based
on mixture density networks [4] and uses a different approximation scheme during training.
Our CVAE formulation has some advantages over the CGAN approach adopted by other recent
works aimed at the same general goals [6, 27]. GANs do not expose control over the structure of
the latent space, while our additive prior results in an interpretable way to control the sampling
process. GANs are also notoriously tricky to train, in particular for discrete sampling problems like
sentence generation (Dai et al. [6] have to resort to reinforcement learning and Shetty et al. [27] to an
approximate Gumbel sampler [15]). Our CVAE training is much more straightforward.
While we represent the z space as a simple vector space with multiple modes, it is possible to impose
on it a more general graphical model structure [16], though this incurs a much greater level of
complexity. Finally, from the viewpoint of inference, our work is also related to general approaches
to diverse structured prediction, which focus on extracting multiple modes from a single energy
function [3]. This is a hard problem necessitating sophisticated approximations, and we prefer to
circumvent it by cheaply generating a large number of diverse and plausible candidates, so that “good
enough” ones can be identified using simple re-ranking mechanisms.
Future work. We would like to investigate more general formulations for the conditioning information c(I), not necessarily relying on object labels whose supervisory information must be provided
separately from the sentences. These can be obtained, for example, by automatically clustering nouns
or noun phrases extracted from reference sentences, or even clustering vector representations of entire
sentences. We are also interested in other tasks, such as question generation, where the cluster vectors
can represent the question type (‘what is,’ ‘where is,’ ‘how many,’ etc.) as well as the image content.
Control of the output by modifying the c vector would in this case be particularly natural.
Acknowledgments: This material is based upon work supported in part by the National Science
Foundation under Grants No. 1563727 and 1718221, and by the Sloan Foundation. We would like to
thank Jian Peng and Yang Liu for helpful discussions.
9
References
[1] Neuraltalk2. https://github.com/karpathy/neuraltalk2.
[2] P. Anderson, B. Fernando, M. Johnson, and S. Gould. Spice: Semantic propositional image caption
evaluation. In ECCV, 2016.
[3] D. Batra, P. Yadollahpour, A. Guzman-Rivera, and G. Shakhnarovich. Diverse M-Best Solutions in Markov
Random Fields. In ECCV, 2012.
[4] C. M. Bishop. Mixture density networks. 1994.
[5] X. Chen, H. Fang, T.-Y. Lin, R. Vedantam, S. Gupta, P. Dollár, and C. L. Zitnick. Microsoft coco captions:
Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015.
[6] B. Dai, D. Lin, R. Urtasun, and S. Fidler. Towards diverse and natural image descriptions via a conditional
gan. ICCV, 2017.
[7] M. Denkowski and A. Lavie. Meteor universal: Language specific translation evaluation for any target
language. In Proceedings of the EACL 2014 Workshop on Statistical Machine Translation, 2014.
[8] A. Deshpande, J. Lu, M.-C. Yeh, and D. Forsyth. Learning diverse image colorization. CVPR, 2017.
[9] J. Devlin, H. Cheng, H. Fang, S. Gupta, L. Deng, X. He, G. Zweig, and M. Mitchell. Language models for
image captioning: The quirks and what works. arXiv preprint arXiv:1505.01809, 2015.
[10] J. Devlin, S. Gupta, R. Girshick, M. Mitchell, and C. L. Zitnick. Exploring nearest neighbor approaches
for image captioning. arXiv preprint arXiv:1505.04467, 2015.
[11] A. Farhadi, M. Hejrati, M. Sadeghi, P. Young, C. Rashtchian, J. Hockenmaier, and D. Forsyth. Every
picture tells a story: Generating sentences from images. In ECCV, 2010.
[12] J. R. Hershey and P. A. Olsen. Approximating the kullback leibler divergence between gaussian mixture
models. In ICASSP, 2007.
[13] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780, 1997.
[14] U. Jain, Z. Zhang, and A. Schwing. Creativity: Generating diverse questions using variational autoencoders.
CVPR, 2017.
[15] E. Jang, S. Gu, and B. Poole. Categorical reparameterization with gumbel-softmax. ICLR, 2017.
[16] M. J. Johnson, D. Duvenaud, A. Wiltschko, S. Datta, and R. Adams. Structured vaes: Composing
probabilistic graphical models and variational autoencoders. NIPS, 2016.
[17] D. P. Kingma and M. Welling. Auto-encoding variational bayes. ICLR, 2014.
[18] R. Kiros, R. Salakhutdinov, and R. Zemel. Multimodal neural language models. In ICML, 2014.
[19] G. Kulkarni, V. Premraj, V. Ordonez, S. Dhar, S. Li, Y. Choi, A. C. Berg, and T. L. Berg. Babytalk:
Understanding and generating simple image descriptions. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 35(12):2891–2903, 2013.
[20] P. Kuznetsova, V. Ordonez, A. C. Berg, T. L. Berg, and Y. Choi. Generalizing image captions for image-text
parallel corpus. In ACL, 2013.
[21] C.-Y. Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out:
Proceedings of the ACL-04 workshop, volume 8. Barcelona, Spain, 2004.
[22] S. Liu, Z. Zhu, N. Ye, S. Guadarrama, and K. Murphy. Improved image captioning via policy gradient
optimization of spider. ICCV, 2017.
[23] J. Mao, W. Xu, Y. Yang, J. Wang, Z. Huang, and A. Yuille. Deep captioning with multimodal recurrent
neural networks (m-rnn). ICLR, 2015.
[24] M. Mitchell, X. Han, J. Dodge, A. Mensch, A. Goyal, A. Berg, K. Yamaguchi, T. Berg, K. Stratos, and
H. Daumé III. Midge: Generating image descriptions from computer vision detections. In Proceedings
of the 13th Conference of the European Chapter of the Association for Computational Linguistics, pages
747–756. Association for Computational Linguistics, 2012.
[25] K. Papineni, S. Roukos, T. Ward, and W.-J. Zhu. Bleu: a method for automatic evaluation of machine
translation. In ACL. Association for Computational Linguistics, 2002.
[26] S. Ren, K. He, R. Girshick, and J. Sun. Faster R-CNN: Towards real-time object detection with region
proposal networks. In NIPS, 2015.
[27] R. Shetty, M. Rohrbach, L. A. Hendricks, M. Fritz, and B. Schiele. Speaking the same language: Matching
machine to human captions by adversarial training. ICCV, 2017.
[28] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition.
arXiv preprint arXiv:1409.1556, 2014.
[29] K. Sohn, H. Lee, and X. Yan. Learning structured output representation using deep conditional generative
models. In NIPS, 2015.
[30] R. Vedantam, C. Lawrence Zitnick, and D. Parikh. Cider: Consensus-based image description evaluation.
In CVPR, 2015.
[31] A. K. Vijayakumar, M. Cogswell, R. R. Selvaraju, Q. Sun, S. Lee, D. Crandall, and D. Batra. Diverse beam
search: Decoding diverse solutions from neural sequence models. arXiv preprint arXiv:1610.02424, 2016.
[32] O. Vinyals, A. Toshev, S. Bengio, and D. Erhan. Show and tell: A neural image caption generator. In
CVPR, 2015.
[33] O. Vinyals, A. Toshev, S. Bengio, and D. Erhan. Show and tell: Lessons learned from the 2015 mscoco
image captioning challenge. IEEE transactions on pattern analysis and machine intelligence, 2016.
[34] L. Wang, Y. Li, and S. Lazebnik. Learning deep structure-preserving image-text embeddings. In CVPR,
2016.
[35] Z. Wang, F. Wu, W. Lu, J. Xiao, X. Li, Z. Zhang, and Y. Zhuang. Diverse image captioning via grouptalk.
In IJCAI, 2016.
10
[36] K. Xu, J. Ba, R. Kiros, K. Cho, A. Courville, R. Salakhudinov, R. Zemel, and Y. Bengio. Show, attend and
tell: Neural image caption generation with visual attention. In ICML, 2015.
[37] Q. You, H. Jin, Z. Wang, C. Fang, and J. Luo. Image captioning with semantic attention. In CVPR, 2016.
11
| 1cs.CV
|
Happy Travelers Take Big Pictures:
A Psychological Study with Machine Learning and Big Data
Xuefeng Liang, Lixin Fan† , Yuen Peng Loh‡ , Yang Liu, Song Tong
‡
IST, Graduate School of Informatics, Kyoto University, 606-8501 Kyoto, Japan
†
Nokia Technologies, Tampere 33100, Finland
Centre of Image and Signal Processing, University of Malaya, Kuala Lumpur 50603, Malaysia
arXiv:1709.07584v1 [cs.CV] 22 Sep 2017
[email protected], [email protected]
[email protected], liu.yang.24m, [email protected]
(a)
(b)
(c)
(d)
(e)
(f)
Figure 1: Wide view angle (a-c) vs. narrow view angle (d-f). The view angle of a travel photo is influenced by the photographer’s
subconscious emotional state, which is a living example of the psychological broaden-and-build theory.
Abstract
In psychology, theory-driven researches are usually conducted with extensive laboratory experiments, yet rarely
tested or disproved with big data. In this paper, we make
use of 418K travel photos with traveler ratings to test the
influential “broaden-and-build” theory, that suggests positive
emotions broaden one’s visual attention. The core hypothesis
examined in this study is that positive emotion is associated
with a wider attention, hence highly-rated sites would trigger
wide-angle photographs. By analyzing travel photos, we find
a strong correlation between a preference for wide-angle photos and the high rating of tourist sites on TripAdvisor. We are
able to carry out this analysis through the use of deep learning algorithms to classify the photos into wide and narrow
angles, and present this study as an exemplar of how big data
and deep learning can be used to test laboratory findings in
the wild.
Introduction
Recent advances in AI technologies (especially Deep Learning) in conjunction with big data offer psychologists an
unprecedented opportunity to test theories outside the laboratory. Cognitive scientists and psychologists have been
increasingly embracing big data and machine learning to
significantly further theory-driven understanding of human
behavior and cognition. For example, the sequential dependence functions in higher-order cognition were investigated on millions online reviews posted on Yelp (Vinson,
Dale, and Jones 2016), a machine learning model trained
on a standard corpus of online text resulted in human-like
semantic biases (Caliskan, Bryson, and Narayanan 2017),
emerging studies demonstrated that big data or naturally
occurring data sets (BONDS) can be used as a complement to traditional laboratory paradigms and refine theories
(Griffiths 2015; Goldstone and Lupyan 2016; Jones 2016;
Paxton and Griffiths 2017). Following in the footsteps of earlier calls to action, we present here an example of leveraging
state-of-the-art machine learning techniques and BONDS
as a complement to test psychological theories. Concretely
we investigate a real world scenario in which the travelers’
photo taking behavior is influenced by a hypothesized psychological mechanism, namely the broaden-and-build theory of positive emotions (Fredrickson 2004; Fredrickson and
Branigan 2005).
According to Fredrickson’s influential theory, positive
emotions broaden (globalize) the attentional scope of the
observer and result in processing of a global picture, while
negative emotions correlate with a narrowed (localized) attentional focus and induce the processing of local elements.
This psychological hypothesis was supported by extensive
laboratory experiments (Rowe, Hirsh, and Anderson 2007;
Tamir and Robinson 2007; Pourtois, Schettino, and Vuilleumier 2013; Vanlessen et al. 2013). They widely employed a
flanker task that required participants to respond to a globallocal visual processing task, in which the visual stimuli were
either compatible geometric figures / letters or incompatible
ones (see supplementary 1 for details). However, to the best
of our knowledge, this theory has not been tested with realworld big data. Moreover, it is imprudent to embrace any of
these theories blindly since traditional psychological experiments are often conducted in a restricted laboratory environment with limited number of subjects that may result in
a considerable bias.
In order to scrutinize the broaden-and-build theory in the
travel photo taking scenario, we first develop a deep learning
algorithm with a performance in sync with a human, subsequently, cross-check photographers’ behaviors by analyzing
big data, and address the confounding factors with a set of
carefully designed experiments. The results demonstrate that
travel photographers’ inclination to specific camera viewpoint, e.g. wide-angle (figures 1a-1c) vs. narrow-angle (figures 1d-1f) is largely influenced by photographers’ emotion
at the time of photo taking. Such kinds of influence, which
might be subconscious to photographers themselves, nevertheless, is statistically consistent and significant. Roughly
speaking, photographers seem to prefer wider-angle photos to narrow-angle ones at high rating tourist sites, while
for lower rating sites, the preference appears to be moderate or even going in the reverse direction (see Fig. 7 and
Experiments for details). This finding is in accord with the
notion of “positive emotions broaden attention and trigger
wide-angle photographs”. Moreover, our study demonstrates
a substantial boost of the numbers and diversity of experimental subjects by taking advantages of machine learning
techniques and the vast amount of behavior data already
available on the internet, which is challenging for traditional
laboratory paradigms.
It is our hope that the set up of experiments as well as
the proposed deep learning algorithm can be a new method
added into the psychologists toolbox. In addition, the methods adopted in this work have potential significance to realworld applications, such as discovering obscure but highvalue tourist sites (Zhuang et al. 2014), preventing mental
illness of special populations through mining their social
media data (Stewart and Davis 2016) and so on.
Materials and Methods
Here we discuss the data and methods employed to investigate our hypothesis. Specifically, we detail our criteria and
procedures in tourist sites selection and photo collection, followed by our proposed machine learning algorithms.
Tourist Sites Selection
To test the “broaden-and-build” theory using BONDS, we
studied the travel photos from selected tourist sites that are
hosted on TripAdvisor (https://www.tripadvisor.com). The
selection is based on five criteria: (1) Popularity: Recommended by top search engines - TripAdvisor, National Geographic and Travel + Leisure; (2) Objectivity: Having at
least 1.5K votes for each site regardless of language, age,
gender, nationality, etc; (3) Generality: Located across in
Asia, Europe, and Americas; (4) Diversity: Keeping site
types as diverse as possible, but avoid religious places; (5)
Independence: Having an appropriate distance from other
sites to avoid cross-rating. Based on their available locations, 70 sites were selected and travel photos associated
with these sites were used as our study targets (see supplementary 2 for more details). Figure 3 illustrates the positions
of these 70 sites on the distribution of 12K suitable candidates (the green curve), while Fig. 2 shows geo-locations of
30 samples of these sites.
Photo Datasets
Table 1: Datasets for the estimation of machine learning
methods and psychological experiments.
Name
1
Dtrain
1
Dtest
D2
D3
# of photos
43,906
8,722
418K
10,000
# of wide-angle
25,776
4,141
–
–
# of narrow-angle
18,130
4,581
–
–
Source
Flickr
TripAdvisor
TripAdvisor
YFCC100m
We used three newly collected datasets in this study, as
shown in Table 1. D1 is for the estimation of our proposed
1
machine learning methods, where the training data in Dtrain
were collected from Flickr according to the geo-locations of
1
aforementioned sites. Whereas, the testing data, Dtest
were
made up of evenly distributed amount of photos from 10
tourist sites, that were randomly collected from TripAdvisor. The reason of choosing data from different sources is
twofold: (1) to avoid the overlap between the training and
testing datasets; (2) photos hosted on TripAdvisor were uploaded by travelers who rated the tourist sites, thus the siteratings and photo contents would be closely related. Due to
the second consideration, we created the dataset D2 , which
consists of 418K travel photos taken at the 70 tourist sites
1
and collected from TripAdvisor without overlapping Dtest
,
to test the hypothesized correlation between tourists’ positive emotion and the choice of wide-angle photos. The third
set, called D3 , consists of 10K random photos collected
from YFCC100m dataset (Thomee et al. 2016) without using geo-tag or any other keywords. These photos were used
to test the preferences of photo-taking behaviors in a completely random (or neutral emotion) mode.
Nevertheless, raw data collected in a completely uncontrolled manner as such are error-prone. The following rectification procedures have been applied to these photos. Firstly,
we scrutinized all photos where the erroneously tagged,
meaningless, and duplicated photos, i.e. noises, were filtered
from the dataset. Secondly, selfies were eliminated from our
datasets due to their intrinsic ambiguities, i.e. the attention
of such photos are on both the narrow-angle of one or more
persons as well as the wide-angle of the background1 .
After rectifying the data, the D1 dataset was then labeled for the estimation of our machine learning algorithms.
1
To build up the training dataset Dtrain
and testing dataset
1
Dtest , about 55K photos were manually labeled as either
wide-angle or narrow-angle. We recruited five subjects (5
male, mean age = 28) and designed a binary classification
task for them. Before the task, 10 wide-angle and 10 narrowangle photos were demonstrated to let all subjects have a
correct understanding. In the task, 20 photos (4 rows and 5
columns) were simultaneously shown on the screen to give a
better visual comparison, and each subject classified photos
into two categories. This procedure was iteratively carried
out until all photos were checked. After collecting the batch
results, we removed those ambiguous photos that had less
1
The trend of selfies is a relatively recent cultural phenomenon
and fast becoming an integral trend in everyday people and also
travelers. Though different from the intuition of the current study,
it is in our interest to look into it in the future.
(a)
(b)
Figure 2: (a) The 20 samples of tourist sites across the world, and (b) A zoom-in view of 10 samples in California (USA).
Figure 3: The 70 selected sites are superimposed on the
number distribution of 12K suitable sites w.r.t. site-ratings
(the green curve), where the sites with similar ratings are
aligned vertically.
than 4 consistent votes. By this way, D1 consists of 52,628
photos in total with an almost perfect agreement (Fleiss’
kappa k = 0.873) among five subjects, and was used as the
ground truth for the estimation of our proposed methods.
Methods
In order to effectively test our hypothesis on such large
dataset D2 , we developed two task-optimized machine
learning models for wide-angle and narrow-angle classification. This section gives a detailed account of our designs,
evaluations, and analysis of said models.
(a)
(b)
(c)
(d)
Figure 4: Example photos of HVS cues to determine view
angle. (a) Narrow-angle of focus lens model, (b) Narrowangle of spatially large but conceptually small object, (c)
Wide-angle of spatially large and conceptually large object,
(d) Wide-angle of spatially small and conceptually small object.
HVS Model The first model mimics the basics of human
visual system (HVS) in determining viewpoints, and is formulated by two cues: a focus cue and a scale cue. The
focus cue is based on the finding that a large number of
professionally shot close-up view photos adhere to the focus lens model of HVS (Tsotsos 2011) where it focuses on
the center object (focus) while the surrounding background
is blurred (fringe), as shown in Fig. 4a. To model it, we
transform images into the frequency domain by using the
Non-subsampled Contourlet Transform (NSCT) (Da Cunha,
Zhou, and Do 2006), in which SURF features (Bay, Tuytelaars, and Van Gool 2006) are extracted and quantized using
Fisher Vector (Perronnin, Sánchez, and Mensink 2010). Afterwards, the classification is implemented by a trained support vector machine (SVM). However, many narrow-angle
photos shot by low-cost cameras (e.g. smart phones) do not
follow the focus model where entire scene appears sharp,
such as Fig. 4b. Therefore, the scale cue is derived from observers’ ability to differentiate the views by measuring the
size of objects, namely the spatial size (the object size measured in the photo indicated by the boxes in Fig. 4b and 4c
are bigger than the one in Fig. 4d) and the conceptual size
(the realistic proportion of the object; a person in Fig. 4b
is a small object but a building in Fig. 4c is a big object).
Referring to Fig. 4b - 4d, a narrow-angle can be determined
if the object is spatially large but conceptually small, otherwise, the photo is a wide-angle. We measure the spatial
size by an object bounding box proposal method, namely
Adobe refined BING boxes (Fang et al. 2016). Whereas, the
conceptual size is measured by a fine-tuned convolutional
neural network (CNN) (Krizhevsky, Sutskever, and Hinton
2012). Hence, this HVS model built following two specific
visual cues of human vision can address distinct photo characteristics.
CF-CNN Model We secondly looked into a deep learning
technique, by using a single CNN to perform this view angle
classification as opposed to the hand-designed HVS model.
This is on account of the success shown by CNN at discovering high level features for a variety of tasks (Donahue
et al. 2014; Zeiler and Fergus 2014; Yosinski et al. 2015;
Lee et al. 2017). However, conventional CNNs only utilize
single high level feature after multiple layers of convolution. According to our pilot investigation, features that are
crucial for view angle classification may vanish after multiple convolution and pooling operations in the conventional
CNNs. Therefore, we designed a cumulative feature CNN
(CF-CNN) that extracts features from each stage and ac-
Figure 5: The architecture of our proposed cumulative feature CNN (CF-CNN). The features from each convolution layer are
cumulated into one representation.
cumulates them into one representation, hence incorporating both low and high level features for the classification
task. Figure 5 illustrates the architecture of our model, where
travel photos are the inputs and the outputs are their respective narrow and wide angle categorization. Specifically, we
introduced additional convolution paths (convx) on each existing convolution (conv) layer, to produce 1024-dimension
features. The new convx layers are placed after pooling layers (pool), and if the conv is not followed by a pool (conv3conv4), pooling layers are added for them (poolx) before the
convx, as shown by the path illustration in Fig. 5. We use
the pooling size [3 × 3] for all pool and poolx. The kernel
sizes of conv are transfered from the AlexNet architecture
(Krizhevsky, Sutskever, and Hinton 2012), while the convx
kernel follows the size of the feature map that is to be convolved, e.g. the feature map after pool1 is 27 × 27 × 96,
hence the kernel size of convx1 is 27 × 27 × 96 and mapped
to 1024 neurons. These convx layers are to be trained endto-end with all the other conv layers, thus the kernels are expected to focus on significant features from different levels.
Hence, they are directly summed up to obtain the cumulative
feature and proceed to the subsequent fully connected layers
for classification.
1
We used 3/4 of wide and narrow angle photos in Dtrain
for training while the remaining for validation. During the
training process, each training image was augmented by resizing the shorter side to 256 dimensions while maintaining
aspect ratio, and then a random cropping and flipping is performed, followed by normalization by subtracting with the
average image of the dataset. Finally, a 227 × 227 × 3 dimension image was fed to the network.
This model is trained end-to-end using the stochastic gradient descend approach with training batch size of 230,
weight decay of 0.0005, and the learning rate that logarithmically reduces from η = 10−5 after every training epoch.
Additionally, we transferred ImageNet pre-trained weights
from the AlexNet for conv1 - conv5 to improve the generalization of the main feature extraction layers of our model
(Yosinski et al. 2014). In order to prevent over-fitting, the
training was stopped at 200 epochs where there was no significant reduction in the trend of the validation error. The dif-
ference between the validation and training errors was 0.047,
an acceptable range of over-fitting as the validation performance achieved over 80%. Thus, we proceed to perform
later classification experiments using this CF-CNN model.
Performance Evaluation and Analysis The performances of HVS and CF-CNN models were evaluated on
1
dataset Dtest
, the travel photos of 10 sites collected from
TripAdvisor. The CF-CNN achieved 88.12% overall classification accuracy, a major improvement in comparison to
the HVS model that only reached 64.06%. Table 2 shows
that CF-CNN outperforms HVS model at all sites. Additionally, we show in Table 3, the ratio of the wide-angle
against narrow-angle photos based on CF-CNN’s classification (disregarding accuracy) closely matches the ratio of
ground truth. This is an indication that the trained CF-CNN
is better than the hand-designed HVS approach with a considerable likeness to a human, therefore, the CF-CNN was
used for testing the “broaden-and-build” theory with real
world data.
Table 2: Classification results of HVS and CNN models on
narrow and wide angle photos from the 10 tourist sites.
Site
1
2
3
4
5
6
7
8
9
10
Total
HVS model
Correct/Total
Narrow
Wide
74/87
237/832
172/240
342/617
246/305
196/518
255/339
233/437
345/439
231/429
343/405
201/425
420/504
145/366
555/689
109/201
502/648
210/269
717/925
18/47
3629/4581
1958/4141
Accuracy
37.76%
59.98%
53.71%
62.89%
66.36%
65.54%
64.94%
74.61%
77.64%
75.62%
64.06%
CF-CNN model
Correct/Total
Narrow
Wide
46/87
819/832
129/240
590/617
265/305
465/518
303/339
333/437
353/439
396/429
385/405
313/425
480/504
207/366
654/689
176/201
592/648
244/269
903/925
33/47
4110/4581
3576/4141
Accuracy
94.12%
83.90%
88.70%
81.96%
86.29%
84.10%
78.97%
93.26%
91.17%
96.30%
88.12%
Table 3: Comparison of wide-angle and narrow-angle ratio
between the ground truth, HVS model, and CF-CNN model.
Site
GT
HVS
CF-CNN
1
9:1
3:7
9:1
2
7:3
5:5
8:2
3
6:4
3:7
6:4
4
6:4
4:6
5:5
5
5:5
4:6
6:4
6
5:5
3:7
4:6
7
4:6
3:7
3:7
8
2:8
3:7
2:8
9
3:7
4:6
3:7
10
1:9
2:8
1:9
model as follows:
W = βr R + βs S + β0 ,
Figure 6: Examples of narrow (top) and wide (bottom) angle
photos with their high activation regions (bright areas).
We also take a further look into the CF-CNN to explore
the contributing factors to its performance, by visualizing
the last activation maps (highest level features) of the network to find out the spatial location in the photos that is
responsible for the classification. The features from lower
level layers through convx are not visualized as it has been
known that they are less abstract features like edges and
high frequency details. Specifically, we extract the activation
maps produced by the last pooling operation (pool5) for each
test image, where the dimension of the maps are 6 × 6 × 256
as shown in Fig. 5. Max pooling is again performed on the
extracted maps in the third dimension to obtain an aggregated map with 6 × 6 dimension, where it is then resized
to the size of the original image. This final map is used to
mask the luminance channel of the original image to obtain
a visualization of the area in which the features are used for
classification.
This operation has given us an interesting insight of the
wide and narrow view angle classification task. Mainly, the
border of an image is a major contributor to the classification as opposed to objects only, as we had initially thought.
Figure 6 shows several examples where the activations are
on the fringe of the image, even though the objects within
the image are clearly shown irrespective of the viewpoints.
This interesting finding suggests that the strong classification can be achieved by “looking” at the image fringe instead of the objects, which goes beyond the focus cue and
scale cue designed in the HVS model. We believe that this
is one of the component missing from the HVS model that
caused its under-performance.
Experiments
We hypothesize that prominent tourist sites induces positive
emotions to travelers, and subsequently prompt them to capture more wide-angle photos than narrow-angle ones. To test
the broaden-and-build theory in photo-taking behaviors, we
structure the analysis to lay out a simple linear regression
(1)
in which W is the proportion of wide-angle photos, R and S
are the rating score and the (approximated) size of the tourist
site, and βr , βs are respective parameters to be estimated,
and β0 is the offset.
The model (1) is derived from the “broaden-and-build”
theory based on two assumptions: (1) emotions in the experiments reported here are considered to be represented by
traveler ratings on TripAdvisor; (2) the scope of attention is
naturally or unconsciously manifested by the choice of the
view-angle of tourist-taken photos. Note that a competing
factor, the site-size S, is included in the model because it
might also affect the choice of the view-angle. In this paper
we adopt the Pearson correlation coefficients (PCC) to quantify and compare the influences of R and S with respect to
W (see Tables 4 and 5 below).
First of all, the model (1) is fitted to W , R and S of 70
tourist sites, which are elaborated in the supplementary 2.
The optimal fitting is reached with parameters (βr = 0.124,
βs = 1.23e−4, β0 = −0.443). Note that the relative low
R-squared (R2 = 0.646) of the model indicates a certain
amount of data cannot be explained by the model. In order to
look for the most influential predictor, we conduct following
two experiments.
Experiment 1: The test with respect to the site-rating
The aim of this experiment is to assess how emotions induced by different tourist sites would affect the choice of
the view-angle of travel photos. All 418K travel photos in
dataset D2 from 70 tourist sites were classified by the CFCNN. Figure 7 plots the site-rating against the proportion of
wide-angle photos at each site.
Table 4: The Pearson correlation coefficients (PCCs) between the site-rating and the proportion of wide-angle photos with the P-values.
# of sites
PCC
P-value
World-wide
70
0.78
3.0e−15
Asia
18
0.75
5.8e−4
Europe
25
0.75
1.9e−5
Americas
27
0.81
2.7e−7
Results: Figure 7a shows a notable correlation between
the site-ratings and the proportions of wide-angle photos
across the world. The Pearson correlation coefficient (PCC
= 0.78 with P = 3.0e−15) indicates a strong site-ratings
dependent preference. Thus, we deem the site-rating R the
principal predictor of the model (1). Not surprisingly, this
observation is consistent with the broaden-and-build theory.
To further investigate the influence of local region and
culture, these 70 sites are classified into three subgroups according to their geo-locations, Asia, Europe, and Americas.
As shown in Table 4 and Fig. 7b, both the trends of the proportion of wide-angle photos and PCCs of three subgroups
are very similar to those of the joint group. Conceivably, the
influence of local regions and cultures is negligible.
Experiment 2: The test with respect to the site-size The
choice of view angle of travel photos may also be affected by
(a)
(b)
(c)
Figure 7: (a) A strong correlation between the site-ratings and the proportions of wide-angle photos in the 70 tourist sites from
across the world. (b) The analogous correlations appear across Asia, Europe, and Americas. (c) A modest correlation between
the site-sizes and the proportions of wide-angle photos.
the size of the tourist site, because people are naturally inclined to take wide-angle photos at a location which has an
open space or large object of interest and vice versa. Hence,
the size of the site could be a confounding factor as shown
in the model (1). The aim of this experiment is to assess the
relation between the preferences of photo-taking behaviors
and site-sizes. To this end, we define the site-size according
to the size of the object/region of interest at the location. For
the sites that have the obvious object of interest, we refer
to their physical sizes in meters (e.g. statues and buildings).
If no such object available, we estimate the size of the region of interest in meters. For the sites with extremely open
space (e.g. mountains, canyons, and seashores), their sizes
are capped to 1km (see supplementary 2 for details).
Table 5: The Pearson correlation coefficients (PCCs) between the site-size and the proportion of wide-angle photos
with the P-values.
# of sites
PCC
P-value
All sites
70
0.62
6.6e−9
Small sites
25
0.44
0.026
Medium sites
14
0.33
0.024
other hand, positive emotions reinforce the tendency for a
happy or excited photographer to take wide-angle photos regardless of site-sizes. This “modulation” effect is not only
in line with the broaden-and-build theory tested in the laboratory, but also suggests that the visual attention is the result
of multiple factors.
Experiment 3: The test on random photos While the linear regression model (1) discloses the influences exerted by
emotions and site-sizes on photo taking behaviors, the aim
of this experiment is to assess the “default” behavior in case
of a completely random mode (neutral emotion). Therefore,
a site independent dataset D3 was randomly collected from
YFCC100m dataset (Thomee et al. 2016) without using geotag or any other keywords, in which 10K photos were classified by the CF-CNN for the statistical analysis. YFCC100m
is a subset of Flickr containing 100 million data, which has
not only travel photos but a vast diversity (see supplementary 3 for example photos).
Large sites
31
0.61
0.002
Results: Figure 7c and Table 5 illustrate a modest correlation (PCC = 0.62) between the size of the site and the
proportion of wide-angle photos. This correlation is noticeably weaker, with a margin of 0.16, than the PCC (= 0.78)
between the site-rating and the proportion of the wide-angle
photos. Since Fig. 7c shows that these sites are unevenly
distributed according to the site-size, we further look into
this factor and separate them into three subgroups, namely
small sites, medium sites, and large sites. Specifically, the
sites with size ≤ 10 meters are in the small group, the sites
with size > 100 meters are in the large group, and others
make up the medium group. We calculate the PCC of each
group and list it in Table 5. These results show even weaker
correlations in the three subgroups. In our view, the dwarf influence of the site-size with respect to that of the site-rating
reinforces our hypothesis under examination.
Note that the interplay between the site-size, human emotion and photo-taking behavior is twofold: on the one hand,
open spaces or large objects make it easy to take wide-angle
photos and vice-versa, but are a secondary factor. On the
Figure 8: The proportions of wide-angle photos in lower rating sites (on average), 10K random data, and high rating
sites (on average).
In order to have a comparison with other emotional states,
we choose the sites, whose ratings are higher than 9.0, as
high rating sites according to the site distribution in Fig. 3
in line with the three-sigma rule in statistics. Whereas, sites
with rating lower than 7.4 are termed lower rating sites.
Results: It turns out that, the proportion of the wide-angle
photos (the green bar in Fig. 8) in the 10K random data
reaches approximately 40%. A close investigation of those
random photos revealed that a vast majority of narrow-angle
photos are cliché photos of everyday life. The proportion of
40%, although slightly in favour of narrow-angle photos, reveals a statistically “normal behavior” in composing wide
vs. narrow angle photos. Since random photos are, supposedly, taken under neural emotion, this particular ratio serves
as a reference and is compared against ratios estimated in
other mood states.
For the high rating sites, the average proportion approximates to 76%, which is not apparent in lower rating sites.
Moreover, the average proportion of 41% for the lower rating sites closely resembles the ratio of random photos. We
conjecture that this similarity can be ascribed to the neutral emotion associated with lower rating sites, i.e. such sites
are unable to induce positive emotions, subsequently, travelers’ photo-taking behaviors are not influenced in a positive manner. On the other hand, the significant high-ratio
associated with high rating sites induces photo-taking behavior via broadened visual attention. Another finding is the
greater deviation of view proportions in lower rating sites
(15.5%) than high rating ones (6%). This signifies that the
good sites share a consistent ability to induce positive emotions to tourists, which is lacking for lower rating sites.
Conclusion and Discussion
In this work, we tested the psychological broaden-and-build
theory outside the laboratory by leveraging recent machine
learning methods and big data from the internet. Our study
revealed a strong correlation between the preference for
wide-angle photos and the high rating of tourist sites. This
preference is ascribed to the notion that positive emotions
broaden visual attention and trigger wide-angle photo compositions. Alternatively, neutral emotion induces a slight favor of narrow-angle photos, which is likely associated with
those lower rating sites. In addition, by controlling the condition of site-size, our result suggests that the visual attention is the result of multiple factors. We are able to carry out
this analysis through the development of a deep learning algorithm for photo view angle classification, which achieves
a performance in sync with a human. We hope that the set
up of experiments as well as the proposed algorithm can be
a new method added into the psychologists toolbox.
Moreover, the methods adopted in this work have potential significance to real-world applications. For example,
recent researches have been focusing on discovering new
tourism resources through mining text or evaluating picture
quality in SNS. However, few of them tried to link tourists’
experiences and mood states with these data, particularly the
image data. The broaden-and-build theory with support of
real world big data in this study can add a new measure for
such task and boost tourism economics. In mental & welfare
heathcare filed, researchers are also reviewing big data resources and their use to characterise applications to address
mental illness, e.g. suicide prevention. The other side of the
broaden-and-build theory (i.e. negative emotions induce a
narrowed attention) with our machine learning method can
help such special populations to have better lives through
mining their SNS data.
References
[Bay, Tuytelaars, and Van Gool 2006] Bay, H.; Tuytelaars,
T.; and Van Gool, L. 2006. Surf: Speeded up robust features. In European Conference on Computer Vision, 404–
417. Springer.
[Caliskan, Bryson, and Narayanan 2017] Caliskan,
A.;
Bryson, J. J.; and Narayanan, A. 2017. Semantics derived
automatically from language corpora contain human-like
biases. Science 356(6334):183–186.
[Da Cunha, Zhou, and Do 2006] Da Cunha, A. L.; Zhou, J.;
and Do, M. N. 2006. The nonsubsampled contourlet transform: theory, design, and applications. IEEE Transactions
on Image Processing 15(10):3089–3101.
[Donahue et al. 2014] Donahue, J.; Jia, Y.; Vinyals, O.; Hoffman, J.; Zhang, N.; Tzeng, E.; and Darrell, T. 2014. Decaf:
A deep convolutional activation feature for generic visual
recognition. In International Conference on Machine Learning, 647–655.
[Fang et al. 2016] Fang, Z.; Cao, Z.; Xiao, Y.; Zhu, L.; and
Yuan, J. 2016. Adobe boxes: Locating object proposals using object adobes. IEEE Transactions on Image Processing
25(9):4116–4128.
[Fredrickson and Branigan 2005] Fredrickson, B. L., and
Branigan, C. 2005. Positive emotions broaden the scope of
attention and thought-action repertoires. Cognition & emotion 19(3):313–332.
[Fredrickson 2004] Fredrickson, B. L. 2004. The broadenand-build theory of positive emotions.
Philosophical
Transactions of the Royal Society B: Biological Sciences
359(1449):1367.
[Goldstone and Lupyan 2016] Goldstone, R. L., and Lupyan,
G. 2016. Discovering psychological principles by mining
naturally occurring data sets. Topics in Cognitive Science
8(3):548–568.
[Griffiths 2015] Griffiths, T. L. 2015. Manifesto for a new
(computational) cognitive revolution. Cognition 135:21–23.
[Jones 2016] Jones, M. N. 2016. Developing cognitive theory by mining large-scale naturalistic data. Big Data in Cognitive Science 1–12.
[Krizhevsky, Sutskever, and Hinton 2012] Krizhevsky, A.;
Sutskever, I.; and Hinton, G. E. 2012. Imagenet classification with deep convolutional neural networks. In Advances
in Neural Information Processing Systems, 1097–1105.
[Lee et al. 2017] Lee, S. H.; Chan, C. S.; Mayo, S. J.; and
Remagnino, P. 2017. How deep learning extracts and learns
leaf features for plant classification. Pattern Recognition
71:1–13.
[Paxton and Griffiths 2017] Paxton, A., and Griffiths, T. L.
2017. Finding the traces of behavioral and cognitive processes in big data and naturally occurring datasets. Behavior
Research Methods 1–9.
[Perronnin, Sánchez, and Mensink 2010] Perronnin,
F.;
Sánchez, J.; and Mensink, T. 2010. Improving the fisher
kernel for large-scale image classification. In European
Conference on Computer Vision, 143–156. Springer.
[Pourtois, Schettino, and Vuilleumier 2013] Pourtois,
G.;
Schettino, A.; and Vuilleumier, P. 2013. Brain mechanisms
for emotional influences on perception and attention: what
is magic and what is not. Biological Psychology.
[Rowe, Hirsh, and Anderson 2007] Rowe, G.; Hirsh, J. B.;
and Anderson, A. K. 2007. Positive affect increases the
breadth of attentional selection. Proceedings of the National
Academy of Sciences 104(1):383–388.
[Stewart and Davis 2016] Stewart, R., and Davis, K. 2016.
big datain mental health research: current status and emerging possibilities. Social psychiatry and psychiatric epidemiology 51(8):1055–1072.
[Tamir and Robinson 2007] Tamir, M., and Robinson, M. D.
2007. The happy spotlight: Positive mood and selective attention to rewarding information. Personality and Social
Psychology Bulletin 33(8):1124–1136.
[Thomee et al. 2016] Thomee, B.; Shamma, D. A.; Friedland, G.; Elizalde, B.; Ni, K.; Poland, D.; Borth, D.; and
Li, L.-J. 2016. Yfcc100m: The new data in multimedia research. Communications of the ACM 59(2):64–73.
[Tsotsos 2011] Tsotsos, J. K. 2011. A computational perspective on visual attention. MIT Press.
[Vanlessen et al. 2013] Vanlessen, N.; Rossi, V.; De Raedt,
R.; and Pourtois, G. 2013. Positive emotion broadens attention focus through decreased position-specific spatial encoding in early visual cortex: Evidence from erps. Cognitive,
Affective, & Behavioral Neuroscience 13(1):60–79.
[Vinson, Dale, and Jones 2016] Vinson, D. W.; Dale, R.; and
Jones, M. N. 2016. Decision contamination in the wild: Sequential dependencies in yelp review ratings. In Proceedings
of the 38th Annual Meeting of the Cognitive Science Society,
1433–1438.
[Yosinski et al. 2014] Yosinski, J.; Clune, J.; Bengio, Y.; and
Lipson, H. 2014. How transferable are features in deep neural networks? In Advances in Neural Information Processing
Systems, 3320–3328.
[Yosinski et al. 2015] Yosinski, J.; Clune, J.; Nguyen, A.;
Fuchs, T.; and Lipson, H. 2015. Understanding neural networks through deep visualization. arXiv preprint
arXiv:1506.06579.
[Zeiler and Fergus 2014] Zeiler, M. D., and Fergus, R. 2014.
Visualizing and understanding convolutional networks.
In European Conference on Computer Vision, 818–833.
Springer.
[Zhuang et al. 2014] Zhuang, C.; Ma, Q.; Liang, X.; and
Yoshikawa, M. 2014. Anaba: An obscure sightseeing spots
discovering system. In International Conference on Multimedia and Expo.
| 1cs.CV
|
Practical combinations of repetition-aware data
structures∗
Djamal Belazzougui1 , Fabio Cunial2 , Travis Gagie3,4 , Nicola
Prezza5 , and Mathieu Raffinot6
1
2
CERIST, Algeria.
Max Planck Institute of Molecular Cell Biology and Genetics, Dresden,
Germany.
Department of Computer Science, University of Helsinki, Finland.
Helsinki Institute for Information Technology, Finland.
Department of Mathematics and Computer Science, University of Udine, Italy.
Laboratoire Bordelais de Recherche en Informatique, CNRS, Bordeaux,
France.
arXiv:1604.06002v2 [cs.DS] 21 Apr 2016
3
4
5
6
Abstract
Highly-repetitive collections of strings are increasingly being amassed by genome sequencing and
genetic variation experiments, as well as by storing all versions of human-generated files, like
webpages and source code. Existing indexes for locating all the exact occurrences of a pattern
in a highly-repetitive string take advantage of a single measure of repetition. However, multiple,
distinct measures of repetition all grow sublinearly in the length of a highly-repetitive string. In
this paper we explore the practical advantages of combining data structures whose size depends on
distinct measures of repetition. The main ingredient of our structures is the run-length encoded
BWT (RLBWT), which takes space proportional to the number of runs in the Burrows-Wheeler
transform of a string. We describe a range of practical variants that combine RLBWT with the
set of boundaries of the Lempel-Ziv 77 factors of a string, which take space proportional to the
number of factors. Such variants use, respectively, the RLBWT of a string and the RLBWT of
its reverse, or just one RLBWT inside a bidirectional index, or just one RLBWT with support for
unidirectional extraction. We also study the practical advantages of combining RLBWT with the
compact directed acyclic word graph of a string, a data structure that takes space proportional
to the number of one-character extensions of maximal repeats. Our approaches are easy to
implement, and provide competitive tradeoffs on significant datasets.
1998 ACM Subject Classification E.1 Data structures, F.2.2 Pattern matching.
Keywords and phrases repetitive strings, locate, count, run-length encoded BWT, Lempel-Ziv
factorization, CDAWG.
Digital Object Identifier 10.4230/LIPIcs.CVIT.2016.23
1
Introduction
Locating all the exact occurrences of a string in a massive, highly-repetitive collection of
similar texts is a fundamental primitive in the post-genome era, in which genomes from
multiple related species, from multiple strains of the same species, and from multiple
individuals, are being sequenced at an increasing pace. Most data structures designed for
∗
This work was partially supported by Academy of Finland under grant 284598 (Center of Excellence in
Cancer Genetics Research).
© D. Belazzougui et al.;
licensed under Creative Commons License CC-BY
42nd Conference on Very Important Topics (CVIT 2016).
Editors: John Q. Open and Joan R. Acces; Article No. 23; pp. 23:1–23:18
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
23:2
Practical combinations of repetition-aware data structures
such repetitive collections take space proportional to a specific measure of repetition, for
example the number z of factors in a Lempel-Ziv parsing [1, 18], or the number r of runs in
a Burrows-Wheeler transform [21]. In previous work we showed how to achieve competitive
theoretical tradeoffs between space and time in locate queries, by combining data structures
that depend on distinct measures of repetition, where all such measures grow sublinearly in
the length of a highly-repetitive string [3]. Specifically, we described a data structure that
takes approximately O(z + r) words of space, and that reports all the occurrences of a pattern
of length m in a string of length n in O(m(log log n + log z) + pocc · log z + socc · log log n)
time, where pocc and socc are the number of primary and of secondary occurrences,
respectively (see Section 2.2). This compares favorably to the O(m2 h + (m + occ) log z)
reporting time of Lempel-Ziv 77 (LZ77) indexes [18], where h is the height of the parse
tree. It also compares favorably in space to solutions based on the run-length encoded BWT
(RLBWT) and on suffix array samples [21], which take O(n/k + r) words of space to achieve
O(m log log n + k · occ · log log n) reporting time, where k is the sampling rate. We also
introduced a data structure whose size depends on the number of right-extensions of maximal
repeats, and that reports all the occ occurrences of a pattern in O(m log log n + occ) time.
The main component of our constructions is the RLBWT, which we use for counting the
number of occurrences of a pattern, and which we combine with the CDAWG and with data
structures from Lempel-Ziv indexes, rather than with suffix array samples, for answering
locate queries.
In this paper we engineer a range of practical variants of such approaches, and we
compare their space-time tradeoffs to a representative set of state-of-the-art indexes for
repetitive collections, including RLCSA [21], a number of LZ77 implementations [17], and a
recent implementation of the hybrid index [32]. One of our indexes based on RLBWT and
LZ77 factors uses an amount of memory comparable to LZ77 indexes, but it answers count
queries between two and four orders of magnitude faster than all LZ77 and hybrid index
implementations. For long patterns, our index uses less space than the hybrid index, and it
answers locate queries between one and two orders of magnitude faster than a number of
LZ77 implementations, and as fast as the fastest LZ77 implementation. With short patterns,
our index based on RLBWT and CDAWG answers locate queries between four and ten times
faster than a version of RLCSA that uses comparable memory, and with extremely short
patterns our index achieves speedups even greater than ten with respect to RLCSA.
2
2.1
Preliminaries
Strings
Let Σ = [1..σ] be an integer alphabet, let # = 0 ∈
/ Σ be a separator, and let T ∈ [1..σ]n−1 be
a string. We denote by T the reverse of T , and by PT # (W ) the set of all starting positions
of a string W ∈ [0..σ]+ in the circular version of T #. We set ΣrT # (W ) = {a ∈ [0..σ] :
|PT # (W a)| > 0} and Σ`T # (W ) = {a ∈ [0..σ] : |PT # (aW )| > 0}. A repeat W ∈ Σ+ is a
string that satisfies |PT # (W )| > 1. A repeat W is right-maximal (respectively, left-maximal)
iff |ΣrT # (W )| > 1 (respectively, iff |Σ`T # (W )| > 1). A maximal repeat is a repeat that is
both left- and right-maximal. We say that a maximal repeat W is rightmost (respectively,
leftmost) if no string W V with V ∈ [0..σ]+ is left-maximal (respectively, if no string V W
with V ∈ [0..σ]+ is right-maximal).
For reasons of space we assume the reader to be familiar with the notion of suffix tree
STT # = (V, E) and of suffix-link tree of T #, which we do not define here. We denote by
`(γ), or equivalently by `(u, v), the label of edge γ = (u, v) ∈ E, and we denote by `(v) the
D. Belazzougui et al.
23:3
string label of node v ∈ V . It is well known that a string W is right-maximal (respectively,
left-maximal) in T # iff W = `(v) for some internal node v of STT # (respectively, iff W = `(v)
for some internal node v of STT # ). Since left-maximality is closed under prefix operation,
there is a bijection between the set of all maximal repeats of T # and the set of all nodes of
the suffix tree of T # that lie on paths that start from the root and that end at nodes labelled
by rightmost maximal repeats. Symmetrically, since right-maximality is closed under suffix
operation, there is a bijection between the set of all maximal repeats and the set of all nodes
of the suffix tree of T # that lie on paths that start from the root and that end at nodes
labelled by leftmost maximal repeats.
The compact directed acyclic word graph of T # (denoted by CDAWGT # in what follows)
is the minimal compact automaton that recognizes the set of suffixes of T # [4, 7]. It can
be seen as the minimization of STT # in which all leaves are merged to the same node (the
sink) that represents T # itself, and in which all nodes except the sink are in one-to-one
correspondence with the maximal repeats of T # [29] (the source corresponds to the empty
string). The set of accepting nodes consists of the sink and of all maximal repeats that also
occur as a suffix of T #. Like in the suffix tree, transitions are labelled by substrings of T #.
Since a maximal repeat corresponds to a subset of its right-maximal suffixes, CDAWGT # can
be built by putting in the same equivalence class all nodes of STT # that belong to the same
maximal unary path of explicit Weiner links. Note also that the subgraph of STT # induced
by maximal repeats is isomorphic to a spanning tree of CDAWGT # .
For reasons of space we assume the reader to be familiar with the notion and uses of
the Burrows-Wheeler transform of T , including the C array, LF mapping, and backward
search. In this paper we use BWTT # to denote the BWT of T #, and we use range(W ) =
[sp(W )..ep(W )] to denote the lexicographic interval of a string W in a BWT that is implicit
from the context. We say that BWTT # [i..j] is a run iff BWTT # [k] = c ∈ [0..σ] for all
k ∈ [i..j], and moreover if any substring BWTT # [i0 ..j 0 ] such that i0 ≤ i, j 0 ≥ j, and either
i0 6= i or j 0 =
6 j, contains at least two distinct characters. It is well known that repetitions
in T # induce runs in BWTT # : for example, the BWT of W k # consists of r ≤ |W | runs
of length at least k and of a run of length one. We denote by rT # the number of runs in
BWTT # , and we call run-length encoded BWT (denoted by RLBWTT # ) any representation
of BWTT # that takes O(rT # ) words of space, and that supports rank and select operations
(see e.g. [20, 21, 30]). Since the difference between rT # and rT # is negligible in practice, to
simplify notation we denote both of them by r when T is implicit from the context.
The Lempel-Ziv 77 factorization of T [34], abbreviated with LZ77 in the rest of the paper,
is the greedy decomposition T1 T2 · · · Tz of T defined as follows. Assume that T is virtually
preceded by the set of distinct characters in its alphabet, and assume that T1 T2 · · · Ti has
already been computed for some prefix of length k of T : then, Ti+1 is the longest prefix of
T [k + 1..n] such that there is a j ≤ k that satisfies T [j..j + |Ti+1 | − 1] = Ti+1 .
In the rest of the paper we drop subscripts whenever they are clear from the context.
2.2
String indexes
The run-length compressed suffix array of T #, denoted by RLCSAT # in what follows, consists
of a run-length compressed rank data structure for BWTT # , and of a sampled suffix array,
denoted by SSAT # [21]. Given a pattern P ∈ [1..σ]m , we use the rank data structure to
find the interval of BWTT # that contains all characters that precede the occurrences of P
in T #: the length of this interval, uncompressed, is the number of such occurrences. To
locate a specific occurrence, we start at the character that precedes it in BWTT # and we use
rank queries to move backward, until we reach a character whose position has been sampled.
CVIT 2016
23:4
Practical combinations of repetition-aware data structures
Thus, the average time for locating an occurrence is inversely proportional to the size of
SSAT # , and fast locating needs a large SSA regardless of the compressibility of the dataset.
Mäkinen et al. suggested ways to reduce the size of the SSA [21], but they did not perform
well enough in real repetitive datasets for the authors to include them in the software they
released.
For reasons of space we assume the reader to be familiar with LZ77 indexes (see e.g.
[13, 16]). Here we just recall that a primary occurrence of a pattern P in T is one that crosses
or ends at a phrase boundary in the LZ77 factorization T1 T2 · · · Tz of T . All other occurrences
are called secondary. Once we have computed primary occurrences, locating all socc secondary
occurrences reduces to two-sided range reporting, and it takes O(socc · log log n) time with
a data structure of O(z) words of space [16]. To locate primary occurrences, we use a
data structure for four-sided range reporting on a z × z grid, with a marker at (x, y) if the
x-th LZ factor in lexicographic order is preceded in the text by the lexicographically y-th
reversed prefix ending at a phrase boundary. This data structure takes O(z) words of space,
and it returns all the phrase boundaries that are immediately followed by a factor in the
specified range, and immediately preceded by a reversed prefix in the specified range, in
O((1 + k) log z) time, where k is the number of phrase boundaries reported [5]. Kärkkäinen
and Ukkonen used two PATRICIA trees [22], one for the factors and the other for the
reversed prefixes ending at phrase boundaries [16]. To locate primary occurrences, we query
the first tree for the range of distinct factors in left-to-right lexicographic order that start
with P [i + 1..m], and we query the second tree for the range of reversed prefixes T [1..pi − 1]
starting with P [1..i], for all i ∈ [1..m]. The ranges returned by the trees are correct iff any
factor starts with P [i + 1..m] and any reversed prefix at a phrase boundary starts with
P [1..i]. To check the first range, we choose any factor in the range and compare its first
m − i characters to P [i + 1..m]. To check the second range, we choose any reversed prefix in
the range and compare its first i characters to P [1..i]. This takes O(m) time for every i, thus
O(m2 ) time in total, assuming that T is not compressed. Replacing the uncompressed text
by an augmented compressed representation, we can store T in O(z log n) space such that
later, given P , we can find all occ occurrences of P in O(m log m + occ · log log n) time [13].
If we know in advance that all patterns will be of length at most M , then we can store in
a FM-index the substrings of T consisting of characters within distance M of the nearest
phrase boundary, and use that to find primary occurrences. This approach, called hybrid
indexing, has been proposed several times recently: see e.g. [32] and references therein for
more details.
2.3
Composite repetition-aware string indexes
It is possible to combine RLBWTT # with the set of all starting positions p1 , p2 , . . . , pz of LZ
factors of T , building a data structure that takes O(z + r) words of space, and that reports all
the pocc primary occurrences of a pattern P ∈ [1..σ]m in O(m(log log n+log z)+pocc·log z)
time [3]. Since such data structure is at the core of the paper, we summarize how it works in
what follows.
The same primary occurrence of P in T can cover up to m boundaries between two LZ
factors. Thus, we consider every possible way of placing, inside P , the rightmost boundary
between two factors, i.e. every possible split of P in two parts P [1..k − 1] and P [k..m] for
k ∈ [2..m], such that P [k..m] is either a factor or a proper prefix of a factor. For every such
k, we use four-sided range reporting queries to list all the occurrences of P in T that conform
to the split, as described in Section 2.2. We encode the sequence p1 , p2 , . . . , pz implicitly,
as follows: we use a bitvector last[1..n] such that last[i] = 1 iff SAT # [i] = n − pj + 2 for
D. Belazzougui et al.
23:5
some j ∈ [1..z], i.e. iff SAT # [i] is the last position of a factor. We represent such bitvector
as a predecessor data structure with partial ranks, using O(z) words of space [33]. Let
STT # = (V, E) be the suffix tree of T #, and let V 0 = {v1 , v2 , . . . , vz } ⊆ V be the set of loci
in STT # of all LZ factors of T . Consider the list of node labels L = `(v1 ), `(v2 ), . . . , `(vz ),
sorted in lexicographic order. It is easy to build a data structure that takes O(z) words of
space, and that implements in O(log z) time function I(W, V 0 ), which returns the (possibly
empty) interval of W in L (see e.g. [3]). Together with last, RLBWTT # and RLBWTT # ,
this data structure is the output of our construction.
Given P , we first perform a backward search in RLBWTT # to determine the number of
occurrences of P in T #: if this number is zero, we stop. During backward search, we store
in a table the interval [ik ..jk ] of P [k..m] in BWTT # for every k ∈ [2..m]. Then, we compute
0
the interval [i0k−1 ..jk−1
] of P [1..k − 1] in BWTT # for every k ∈ [2..m], using backward search
0
in RLBWTT # : if rank1 (last, jk−1
) − rank1 (last, i0k−1 − 1) = 0, then P [1..k − 1] never ends
at the last position of a factor, and we can discard this value of k. Otherwise, we convert
0
0
[i0k−1 ..jk−1
] to the interval [rank1 (last, i0k−1 − 1) + 1..rank1 (last, jk−1
)] of all the reversed
prefixes of T that end at the last position of a factor. Rank operations on last can be
implemented in O(log log n) time using predecessor queries. We get the lexicographic interval
of P [k..m] in the list of all distinct factors of T using operation I(P [k..m], V 0 ), in O(log z)
time. We use such intervals to query the four-sided range reporting data structure.
It is also possible to combine RLBWTT # with CDAWGT # , building a data structure that
takes O(eT # ) words of space, and that reports all the occ occurrences of P in O(m log log n +
occ) time, where eT # is the number of right-extensions of maximal repeats of T # [3].
Specifically, for every node v in the CDAWG, we store |`(v)| in a variable v.length. Recall
that an arc (v, w) in the CDAWG means that maximal repeat `(w) can be obtained by
extending maximal repeat `(v) to the right and to the left. Thus, for every arc γ = (v, w)
of the CDAWG, we store the first character of `(γ) in a variable γ.char, and we store the
length of the right extension implied by γ in a variable γ.right. The length γ.left of the
left extension implied by γ can be computed by w.length − v.length − γ.right. For every
arc of the CDAWG that connects a maximal repeat W to the sink, we store just γ.char and
the starting position γ.pos of string W · γ.char in T . The total space used by the CDAWG
is O(eT # ) words, and the number of runs in BWTT # can be shown to be O(eT # ) as well [3].
An alternative construction could use CDAWGT # and RLBWTT # .
We use the RLBWT to count the number of occurrences of P in T in O(m log log n)
time: if this number is not zero, we use the CDAWG to report all the occ occurrences of P
in O(occ) time, using a technique already sketched in [6]. Specifically, since we know that
P occurs in T , we perform a blind search for P in the CDAWG, as is typically done with
PATRICIA trees. We keep a variable i, initialized to zero, that stores the length of the prefix
of P that we have matched so far, and we keep a variable j, initialized to one, that stores
the starting position of P inside the last maximal repeat encountered during the search. For
every node v in the CDAWG, we choose the arc γ such that γ.char = P [i + 1] in constant
time using hashing, we increment i by γ.right, and we increment j by γ.left. If the search
leads to the sink by an arc γ, we report γ.pos + j and we stop. If the search ends in a node
v that is associated with the maximal repeat W , we determine all the occurrences of W in T
by performing a depth-first traversal of all nodes reachable from v in the CDAWG , updating
variables i and j as described above, and reporting γ.pos + j for every arc γ that leads to
the sink. The total number of nodes and arcs reachable from v is O(occ).
CVIT 2016
23:6
Practical combinations of repetition-aware data structures
3
Combining RLBWT and LZ factors in practice
We implement1 the combination of RLBWT and LZ factorization described in Section 2.3,
exploring a range of practical variants of decreasing size. Specifically, in addition to the
version described in Section 2.3 (which we call full in what follows), we implement a variant
in which we drop RLBWTT # , simulating it with a bidirectional index (we call this variant
bidirectional in what follows), a variant in which we drop RLBWTT # , the four-sided range
reporting data structure, and the subset of suffix tree nodes (we call this variant light in
what follows), and another variant in which we use a sparse version of the LZ parsing (we
call this sparse in what follows). Moreover, we design a number of practical optimization to
speed up locate queries: see Appendix C.
We implement all variants using a representation of the RLBWT that is more spaceefficient than the one described in [30]. Recall that the latter is encoded as follows: they
store one character per run in a string H ∈ Σr , they mark with a one the beginning of each
run in a bitvector Vall [0..n − 1], and for every c ∈ Σ they store the lengths of all runs of
character c consecutively in a bit-vector Vc : specifically, every c-run of length k is represented
in Vc as 10k−1 . This representation allows one to map rank and access queries on BWTT #
to rank, select and access queries on H, Vall , and Vc . By gap-encoding the bitvectors, this
representation takes r(2 log(n/r) + log σ)(1 + o(1)) bits of space. We reduce the multiplicative
factor of term log(n/r) by storing in Vall just one out of 1/ ones, where 0 < ≤ 1 is
an arbitrary constant. It is easy to see that we are still able to answer all queries on the
RLBWT by using the vectors Vc to reconstruct the positions of the missing ones in Vall ,
using r (1 + ) log(n/r) + log σ (1 + o(1)) bits of space, but query times are multiplied by a
factor 1/. In all our experiments we set to 1/8. We represent H as a Huffman-encoded
string (wt_huff<> in SDSL), and gap-encoded bitvectors with Elias-Fano (sd_vector<> in
SDSL).
3.1
Full index
The first variant is an engineered version of the data structure described in Section 2.3. We
store both RLBWTT # and RLBWTT # . A gap-encoded bitvector end[0..n−1] of z log(n/z)(1+
o(1)) bits marks the rank, among all the suffixes of T #, of every suffix T [i..n − 1]# such that
n − i − 2 is the last position of an LZ factor of T . Symmetrically, a gap-encoded bitvector
begin[0..n − 1] of z log(n/z)(1 + o(1)) bits marks the rank, among all the suffixes of T #, of
every suffix T [i..n − 1]# such that i is the first position of an LZ factor of T .
Geometric range data structures (4-sided and 2-sided) are implemented as wavelet trees
(wt_int<> in SDSL). The 4-sided range data structure supports locating primary occurrences,
by storing the permutation of the z LZ factors of T , sorted lexicographically, in the order
induced by the corresponding ones in end: in other words, every character of the wavelet
tree is the lexicographic rank of an LZ factor, among all the LZ factors of T . For locating
primary occurrences, we need to label every point in the 4-sided range data structure with
a text position. We allocate log z bits rather than log n bits to every such label, by using
as label the rank of the corresponding one in array begin, thus the data structure takes
2z log z(1 + o(1)) bits of space. The 2-sided range data structure stores z two-dimensional
points whose coordinates are both in [1..|T |]. For locating secondary occurrences, every
1
The source code of all our implementations is available at [26, 27], and it is based on the SDSL
library [15].
D. Belazzougui et al.
23:7
such point is again labeled with the rank of the corresponding one in array begin. The
wavelet tree that implements the 2-sided range data structure can only store points whose
set of x coordinates is [1..z]: we map text coordinates to this domain using a gap-encoded
bitvector, which takes z log(n/z)(1 + o(1)) bits of space. Some coordinates could be repeated,
since two LZ factors could share the same source start point: we keep track of duplicates
using a succinct bitvector that takes z + o(z) bits of space, in which a coordinate that is
repeated k > 0 times is encoded as 10k−1 . Overall, the 2-sided range data structure takes
z(2 log n + 1)(1 + o(1)) bits of space.
Finally, we need a way to compute the lexicographic range of a string among all the
LZ factors of T . We implement a simpler and more space-efficient strategy than the one
proposed in [3]. Specifically, recall that LZ factors are right-maximal substrings of T #, or
equivalently they are nodes of the suffix tree of T #. Recall also that the BWT intervals
of two nodes of the suffix tree of T # are either disjoint or contained in one another. We
sort the BWT intervals of all LZ factors by the order induced by the pre-order traversal
of the suffix tree of T #: two distinct nonempty intervals [i0 ..j 0 ] and [i00 ..j 00 ] are such that
[i0 ..j 0 ] < [i00 ..j 00 ] iff j 0 < i00 , or iff [i00 ..j 00 ] is contained in [i0 ..j 0 ]. The data structure is just the
sorted array V of such intervals, and it takes 2z log n bits of space2 . Given the BWT interval
[i..j] of a string W , we find its lexicographic range among all sorted distinct LZ factors, in
O(log z) time, as follows: (1) we binary-search V using the order described above, finding all
intervals that are strictly smaller than [i..j]; (2) starting from the first interval in V that is
greater than or equal to [i..j], we find all intervals in V that equal [i..j] or are contained in
it, i.e. all intervals of factors that are either W itself or a right-extension of W . This requires
just one binary search, since all such intervals are contiguous in V .
In summary, the full index takes 6z log n + 2(1 + )r log(n/r) + 2r log σ · (1 + o(1)) bits
of space, and it supports count queries in O(m · (log(n/r) + log σ)) time and locate queries
in O((m + occ) · log n) time.
3.2
Bidirectional index
We can drop RLBWTT # and simulate it using just RLBWTT # , by applying the synchronization step performed in bidirectional BWT indexes (see e.g. [2] and references therein). This
strategy penalizes the time complexity of locate queries, which becomes quadratic in the
length of the pattern. Moreover, since in our implementation we store run-lengths separately
for each character, a synchronization step requires σ rank queries to find the number of
characters smaller than a given character inside a BWT interval. This operation could be
performed in O(log σ) time if the string were represented as a wavelet tree. In summary, the
bidirectional variant of the index takes 6z log n + (1 + )r log(n/r) + r log σ · (1 + o(1)) bits
of space, it supports count queries in O(m · (log(n/r) + log σ)) time, and it supports locate
queries in O(m2 σ log(n/r) + (m + occ) · log n) time.
3.3
Light index with LZ sparsification
Once we have computed the interval of the pattern in BWTT # , we can locate all its primary
occurrences by just forward-extracting at most m characters for each occurrence inside the
2
The sequence of first positions of all intervals in the sorted array is non-decreasing, thus we use gap
encoding to save z log z bits of space. This adds a multiplicative factor of O(log(n/z)) to all query times.
For clarity we describe just the simpler version in which intervals are encoded as 2z integers of log n
bits each.
CVIT 2016
23:8
Practical combinations of repetition-aware data structures
range: this is because every primary occurrence of the pattern overlaps with the last position
of an LZ factor. We implement forward extraction with select queries on RLBWTT # . This
approach requires just RLBWTT # , the 2-sided range data structure, a gap-encoded bitvector
endT that marks the last position of every LZ factor in the text, a gap-encoded bitvector
endBW T that marks the last position of every LZ factor in BWTT # , and z integers of log z
bits each, connecting corresponding ones in endBW T and in endT : this array plays the role
of the sparse suffix array sampling used in RLCSA.
We can reduce space even further by sparsifying the LZ factorization. Intuitively, the
factorization of a highly-repetitive collection of strings T = T1 T2 · · · Tk , where T2 , . . . , Tk
are similar to T1 , is much denser inside T1 than it is inside T2 · · · Tk , thus excluding long
enough contiguous regions from the factorization (i.e. not outputting factors inside such
regions) could reduce the number of factors in dense regions. Formally, let d > 0, and
consider the following generalization3 of LZ77, denoted here by LZ77-d: we factor T as
X1 Y1 X2 Y2 · · · Xzd Yzd , where zd is the size of the factorization, Yi ∈ Σd for all i ∈ [1..zd ], and
Xi is the longest prefix of Xi Yi · · · Xzd Yzd that appears at least twice in X1 Y1 X2 Y2 · · · Xi To
make the index described in this section work with LZ77-d, we need to sample the suffix
array of T # at the lexicographic ranks that correspond to the last position of every Xi ,
and we need to redefine primary occurrences as those that are not fully contained inside an
Xi . During locate we now need to extract d additional characters before each occurrence
of the pattern, in order to locate primary occurrences that start inside a Yi . The 2-sided
range data structure must also be built on the (sources of the) factors X1 , . . . , Xzd . This
implementation of the index takes zd (3 log n + log(n/zd )) + (1 + )r log(n/r) · (1 + o(1))
bits of space, it answers locate queries in O((occ + 1) · (m + d) · log n) time and count queries
in O(m(log(n/r) + log σ)) time.
4
Combining RLBWT and CDAWG in practice
We implement4 the combination of RLBWT and CDAWG described in Section 2.3, and we
study the effect of two representations of the CDAWG in memory. In the first representation,
the graph is encoded as a sequence of variable-length integers: every integer is represented as
a sequence of bytes, in which the seven least significant bits of every byte are used to encode
the integer, and the most significant bit flags the last byte of the integer. Nodes are stored
in the sequence according to their topological order in the graph obtained from the CDAWG
by inverting the direction of all arcs: to encode a pointer from a node v to its successor w
in the CDAWG, we store the difference between the first byte of v and the first byte of w
in the sequence. If w is the sink, such difference is replaced by a shorter code. We choose
to store the length of the maximal repeat that corresponds to each node, rather than the
offset of `(v) inside `(w) for every arc (v, w), since such lengths are short and smaller than
the number of arcs in practice.
In the second encoding we exploit the fact that the subgraph of the suffix tree of T #
induced by maximal repeats is a spanning tree of CDAWGT # (see Section 2.1). Specifically,
we encode such spanning tree with the balanced parenthesis scheme described in [23], and we
resolve the arcs of the CDAWG that belong to the tree using corresponding tree operations.
Such operations work on node identifiers, thus we need to convert node identifiers to the
3
4
The version of LZ77 considered in this paper is obtained by setting d = 0, and by requiring the text to
be virtually preceded by all its distinct characters.
The source code of all our implementations is available at [28].
D. Belazzougui et al.
23:9
first byte in the byte sequence of the CDAWG, and vice versa. For this, we encode the
monotone sequence of the first byte of all n nodes in the byte sequence using the quasi-succinct
representation by Elias and Fano, which uses at most 2 + log(N/n) bits per starting position,
where N is the number of bytes in the byte sequence [9].
Finally, we observe that classical CDAWG construction algorithms (e.g. the online
algorithm described in [7]) are not space-efficient, and we design linear-time algorithms that
build a representation of the CDAWG from BWTT # or BWTT # , using optimal additional
space. Specifically, let enumerateLeft(i, j) be a function that returns the set of distinct
characters that appear in BWTT # [i..j], not necessarily in lexicographic order. We prove the
following lemmas in Appendix A and B:
I Lemma 1. Let T ∈ [1..σ]n−1 #, where # = 0. Given a representation of BWTT # that
answers enumerateLeft in time tEL per element in its output, and LF in time tLF , we can
build the topology of CDAWGT # , as well as the first character and the length of the label of
each arc, in randomized O(n(tEL + tLF )) time and zero space in addition to the input and
the output.
I Lemma 2. Let T ∈ [1..σ]n−1 #, where # = 0. Given a representation of BWTT # that
answers enumerateLeft in time tEL per element in its output, and LF in time tLF , we can
build the topology of CDAWGT # , as well as the first character and the length of the label of
each arc, in randomized O(n(tEL + tLF )) time and O(σ 2 log2 n) bits of space in addition to
the input and the output.
5
Experimental results
We test our implementations on five DNA datasets from the Pizza&Chili repetitive corpus [11],
which include the whole genomes of approximately 36 strains of the same eukaryotic species
(“Saccharomyces cerevisiae” and “Saccharomyces paradoxus” in the plots), a collection
of 23 and approximately 78 thousand substrings of the genome of the same bacterium
(respectively “Escherichia coli” and “Haemophilus influenzae”), and an artificially repetitive
string obtained by concatenating a hundred mutated copies of the same substring of the
human genome (denoted by “pseudo-real” in the plots)5 . We compare our results to the
FM index implementation in SDSL [15] with sampling rate 2i for i ∈ [5..10] (represented
by black circles in all plots), to an implementation of RLCSA6 [24] with the same sampling
rates (triangles in the plots), to the five variants in the implementation of the LZ77 index
described in [17] (squares), and to a recent implementation of the compressed hybrid index
[32] (diamonds). The FM index uses RRR bitvectors in its wavelet tree. For brevity, we call
LZ1 the implementation of the LZ77 index that uses suffix trie and reverse trie. For each
process, and for each pattern length 2i for i ∈ [3..10], we measure the maximum resident
set size and the number of CPU seconds that the process spends in user mode7 , both for
5
6
7
Compressing such files with p7zip [25], an implementation of LZ77 with large window, makes the
uncompressed files between 21 and 370 times bigger than the corresponding compressed files [11].
We compile the sequential version with PSI_FLAGS and SA_FLAGS turned off (thus, a gap-encoded
bitvector rather than a succinct bitvector is used to mark sampled positions in the suffix array). The
block size of psi vectors (RLCSA_BLOCK_SIZE) is 32 bytes.
We perform all experiments on a single core of a 6-core, 2.50 GHz, Intel Xeon E5-2640 processor, with
access to 128GiB of RAM and running CentOS 6.3. We measure resources with GNU Time 1.7, and we
compile with GCC 5.3.0.
CVIT 2016
23:10
Practical combinations of repetition-aware data structures
locate and for count queries, discarding the time for loading the indexes and averaging our
measurements over one thousand patterns8 .
We observe two distinct regimes for locate queries, corresponding to short patterns
(shorter than approximately 64) and to long patterns, respectively (Figures 1 and 3). Our full,
bidirectional and light index implementations (red circles in all plots) do not achieve any new
useful tradeoff, in any dataset, neither with short nor with long patterns. As expected, the
running time per pattern of the bidirectional index depends quadratically on pattern length,
but we observe a superlinear growth for the light index as well. The optimizations described
in Appendix C (red dots) are effective only for the bidirectional index, their effectiveness
increases with pattern length, and they manage to shave up to 80% of running time with
patterns of length 1024. The size of the bidirectional index on disk is on average 20% smaller
than the size of the full index on disk, and the size of the light index on disk is approximately
20% smaller than the size of the bidirectional index on disk.
We experiment with skipping 2i characters before opening a new phrase in the light index
with LZ sparsification (green in all plots), where i ∈ [5..10]. The size of the sparse index with
skip rate 25 on disk is approximately 35% smaller than the size of the light index on disk.
With short patterns, the memory used by the sparse index becomes smaller than RLCSA
and comparable to the LZ index, but its running time per occurrence is between one and two
orders of magnitude greater than the LZ index and comparable to RLCSA with sampling
rates equal to or greater than 2048 (Figure 1, top). With long patterns, however, the sparse
index becomes between one and two orders of magnitude faster than all variants of the LZ
index, except variant LZ1, while using comparable memory. As a function of pattern length,
the running time per occurrence of the sparse index grows more slowly than the running time
of LZ1, suggesting that the sparse index becomes as fast as LZ1 for patterns of length between
1024 and 2048 (Figure 3, top). The sparse index is approximately 1.5 orders of magnitude
slower than the hybrid index, but since the size of the hybrid index depends on maximum
pattern length, the sparse index becomes smaller than the hybrid index for patterns of length
between 64 and 128, and possibly even shorter (Figure 4, top). As expected, the sparse index
is faster than both the LZ index and the hybrid index in count queries, especially for short
patterns: specifically, the sparse index is between two and four orders of magnitude faster
than all variants of the LZ index, with the largest difference for patterns of length 8 (Figure
3, bottom). The difference between the sparse index and variant LZ1 shrinks as pattern
length increases. Similar trends hold for the hybrid index. The full, bidirectional and light
indexes show similar count times as the sparse index.
The disk size of the CDAWG is comparable to the disk size of RLCSA with sampling
rate between 4 and 8 (Figure 4, bottom). Using the succinct representation of the CDAWG
(blue dots in all plots) shaves between 20% and 30% of the disk size and resident set of the
non-succinct representation (blue circles). However, using the non-succinct representation
shaves between 20% and 80% of the time of the succinct representation, depending on dataset
and pattern length. Using the CDAWG to answer locate queries does not achieve any new
tradeoff with long patterns (Figure 1, bottom). However, with short patterns the running
time per occurrence of the CDAWG is between 4 and 10 times smaller than the running time
per occurrence of a version of RLCSA that uses comparable memory, and with patterns of
length two the CDAWG achieves speedups even greater than 10.
8
We generate random patterns that contain just characters in {a, c, g, t} using the genpatterns tool
from the Pizza&Chili corpus [11].
D. Belazzougui et al.
23:11
Figure 1 Space-time tradeoffs of our indexes (color) and of the state of the art (black). Top row:
patterns of length 16. Bottom row: patterns of length 512. For more clarity, RLCSA and sparse
index are tested also on additional configurations not mentioned in Section 5.
Figure 2 Space-time traeoffs of the CDAWG (blue) compared to RLCSA (triangles) with sampling
rate 2i , i ∈ [3..5]. Patterns of length 8, 6, 4, 2 (from left to right).
CVIT 2016
23:12
Practical combinations of repetition-aware data structures
Figure 3 Locate time per occurrence (top) and count time per pattern (bottom), as a function
of pattern length, for the sparse index with skip rate 2i , i ∈ [5..10], the LZ77 index, and the hybrid
index. Count plots show also the FM index and RLCSA.
Figure 4 (Top) Disk size of the sparse index with skip rate 2i , i ∈ [10..15], compared to the hybrid
index with maximum pattern length 2i , i ∈ [3..10], the LZ77 index, and RLCSA with sampling rate
2i , i ∈ [10..15]. (Bottom) Disk size of the CDAWG compared to RLCSA with sampling rate 2i ,
i ∈ [2..5].
D. Belazzougui et al.
23:13
Acknowledgements. We thank Miguel Ángel Martínez for providing implementations of the
variants described in [17], and Daniel Valenzuela for providing the implementation described
in [32].
6
Future work
Designing indexes for repetitive texts is an increasingly active field and it is beyond the scope
of this paper to review all the recent proposals — see, e.g., [10, 14, 31, 32] — especially since
some have not been implemented. We would like to draw attention to an index described
simultaneously and independently by Do et al. [8] and Gagie et al. [12], however, because we
think it can be improved and made competitive in practice.
Their index is intended for collections of many similar strings, such as databases of
genomes from the same species. The main idea is to choose one of the strings as a reference
and build an FM-index for the Relative Lempel-Ziv parse [19] of the entire dataset with
respect to that reference, treating the phrases as meta-characters. Using FM-indexes for
the reference and its reverse and some auxiliary data structures, we can apply dynamic
programming to quickly compute all the ways any given pattern can be decomposed into
the suffix of a phrase (possibly empty), a sequence of complete phrases, and the prefix of a
phrase (possibly empty). This is possible because the dictionary of potential phrases — i.e.,
all the substrings of the reference — is fixed, albeit very large, and does not change while we
parse (in contrast to the dictionaries for LZ77 and LZ78).
Using the FM-index for the parse and some more auxiliary data structures, we can quickly
find whether and where any of the possible decompositions of the pattern occur in the the
parse of the dataset. These occurrences correspond to occurrences of the pattern that cross
phrase boundaries. We can then quickly find all the other occurrences of the pattern in the
dataset. Unfortunately, if there are many distinct phrases in the parse, the FM-index for it
may not compress well. This raises the question of how we can reduce the number of distinct
phrases without increasing the number of phrases too much.
Suppose we build a compressed kth-order de Bruijn graph for the collection of strings (i.e.,
a kth-order de Bruijn graph in which we have collapsed every maximal path whose internal
nodes have in- and out-degree 1); assign each of its edges with a distinct meta-character;
consider each string as a walk on the graph and, for each edge the walk crosses, replacing
the substring causing us to cross that edge by the edge’s meta-character. This results in a
parse in which the number of distinct phrases is the number of edges in the graph (assuming
each of the strings in the collection has length at least k). Notice also that, for any pattern
of length at least k, the pattern corresponds to at most one walk on the uncompressed graph
(which may start and finish in the middle of edges in the compressed graph); if there is no
such walk, then the pattern does not occur in the collection. Using a de Bruijn graph also
removes the need for choosing a reference and using range reporting.
We may be able to improve the compression by, first, removing low-frequency edges in
the original de Bruijn graph before compressing it and, second, replacing substrings only
when they cause us to cross high-frequency edges with sufficiently long edge labels in the the
compressed de Bruijn graph. With this modification, a pattern may appear in the collection
even if it does not correspond to a walk on the uncompressed graph. However, if a substring
causes us to cross a high-frequency edge with a long edge label in the compressed graph,
then we can certainly replace that substring by the edge’s meta-character. It follows that,
for any pattern of length at least k, we need perform at most four searches in the FM-index
for the parse to determine whether and where that pattern occurs in the collection.
CVIT 2016
23:14
Practical combinations of repetition-aware data structures
We plan to implement and test this modification of Do et al.’s and Gagie et al.’s index
soon and report the results in a future paper.
Acknowledgements. We thank Miguel Ángel Martínez for providing implementations of the
variants described in [17], and Daniel Valenzuela for providing the implementation described
in [32].
References
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Diego Arroyuelo, Gonzalo Navarro, and Kunihiko Sadakane. Stronger Lempel-Ziv based
compressed text indexing. Algorithmica, 62(1-2):54–101, 2012.
Djamal Belazzougui. Linear time construction of compressed text indices in compact space.
In Proceedings of the 46th Annual ACM Symposium on Theory of Computing, pages 148–
193. ACM, 2014.
Djamal Belazzougui, Fabio Cunial, Travis Gagie, Nicola Prezza, and Mathieu Raffinot.
Composite repetition-aware data structures. In Combinatorial Pattern Matching, pages
26–39. Springer, 2015.
Anselm Blumer, Janet Blumer, David Haussler, Ross McConnell, and Andrzej Ehrenfeucht. Complete inverted files for efficient text retrieval and analysis. Journal of the
ACM, 34(3):578–595, 1987.
Timothy M Chan, Kasper Green Larsen, and Mihai Pătraşcu. Orthogonal range searching on the RAM, revisited. In Proceedings of the twenty-seventh annual symposium on
computational geometry, pages 1–10. ACM, 2011.
Maxime Crochemore and Christophe Hancart. Automata for matching patterns. In Handbook of formal languages, pages 399–462. Springer, 1997.
Maxime Crochemore and Renaud Vérin. Direct construction of compact directed acyclic
word graphs. In Alberto Apostolico and Jotun Hein, editors, CPM, volume 1264 of Lecture
Notes in Computer Science, pages 116–129. Springer, 1997.
Huy Hoang Do, Jesper Jansson, Kunihiko Sadakane, and Wing-Kin Sung. Fast relative
lempel-ziv self-index for similar sequences. Theoretical Computer Science, 532:14–30, 2014.
Peter Elias and Richard A Flower. The complexity of some simple retrieval problems.
Journal of the ACM (JACM), 22(3):367–379, 1975.
Hector Ferrada, Travis Gagie, Tommi Hirvola, and Simon J. Puglisi. Hybrid indexes for
repetitive datasets. Philosophical Transactions of the Royal Society of London A, 372(2016),
2014.
Paolo Ferragina and Gonzalo Navarro. Pizza&Chili repetitive corpus. http://pizzachili.
dcc.uchile.cl/repcorpus.html. Accessed: 2016-04-10.
Travis Gagie, Pawel Gawrychowski, Juha Kärkkäinen, Yakov Nekrich, and Simon J. Puglisi.
A faster grammar-based self-index. In Proceedings of the 6th Conference on Language and
Automata Theory and Applications, pages 240–251, 2012.
Travis Gagie, Paweł Gawrychowski, Juha Kärkkäinen, Yakov Nekrich, and Simon J Puglisi. LZ77-based self-indexing with faster pattern matching. In LATIN 2014: Theoretical
Informatics, pages 731–742. Springer, 2014.
Travis Gagie, Pawel Gawrychowski, Juha Kärkkäinen, Yakov Nekrich, and Simon J. Puglisi.
Lz77-based self-indexing with faster pattern matching. In Proceedings of the 11th Latin
American Symposium on Theoretical Informatics, pages 731–742, 2014.
Simon Gog, Timo Beller, Alistair Moffat, and Matthias Petri. From theory to practice: Plug
and play with succinct data structures. In 13th International Symposium on Experimental
Algorithms, (SEA 2014), pages 326–337, 2014.
D. Belazzougui et al.
23:15
16
Juha Kärkkäinen and Esko Ukkonen. Lempel-Ziv parsing and sublinear-size index structures for string matching. In Proc. 3rd South American Workshop on String Processing
(WSP’96, pages 141–155, 1996.
17
Sebastian Kreft. Self-index based on lz77. Master’s thesis, Department of Computer Science,
University of Chile, 2010.
18
Sebastian Kreft and Gonzalo Navarro. On compressing and indexing repetitive sequences.
Theoretical Computer Science, 483:115–133, 2013.
19
Shanika Kuruppu, Simon J. Puglisi, and Justin Zobel. Relative lempel-ziv compression
of genomes for large-scale storage and retrieval. In Proceedings of the 17th Symposium on
String Processing and Information Retrieval, pages 201–206, 2010.
20
Veli Mäkinen and Gonzalo Navarro. Succinct suffix arrays based on run-length encoding.
In Combinatorial Pattern Matching, pages 45–56. Springer, 2005.
21
Veli Mäkinen, Gonzalo Navarro, Jouni Sirén, and Niko Välimäki. Storage and retrieval of
highly repetitive sequence collections. Journal of Computational Biology, 17(3):281–308,
2010.
22
Donald R Morrison. Patricia—practical algorithm to retrieve information coded in alphanumeric. Journal of the ACM (JACM), 15(4):514–534, 1968.
23
J. Ian Munro and Venkatesh Raman. Succinct representation of balanced parentheses and
static trees. SIAM J. Comput., 31(3):762–776, March 2002. URL: http://dx.doi.org/
10.1137/S0097539799364092, doi:10.1137/S0097539799364092.
24
Adam Novak. Convenient repository for/fork of the RLCSA library. https://github.com/
adamnovak/rlcsa. Accessed: 2016-04-10.
25
Igor Pavlov. P7ZIP home. http://p7zip.sourceforge.net. Accessed: 2016-04-10.
26
Nicola Prezza. lz-rlbwt: Run-length compressed Burrows-Wheeler transform with LZ77
suffix array sampling. https://github.com/nicolaprezza/lz-rlbwt. Accessed: 2016-0410.
27
Nicola Prezza.
lz-rlbwt-sparse: Run-length compressed Burrows-Wheeler transform with sparse LZ77 suffix array sampling. https://github.com/nicolaprezza/
lz-rlbwt-sparse. Accessed: 2016-04-10.
28
Mathieu Raffinot. locate-cdawg: Replacing sampling by CDAWG localisation in BWT
indexing approaches. https://github.com/mathieuraffinot/locate-cdawg. Accessed:
2016-04-10.
29
Mathieu Raffinot. On maximal repeats in strings.
80(3):165–169, 2001.
30
Jouni Sirén, Niko Välimäki, Veli Mäkinen, and Gonzalo Navarro. Run-length compressed
indexes are superior for highly repetitive sequence collections. In String Processing and
Information Retrieval, 15th International Symposium, SPIRE 2008, Melbourne, Australia,
November 10-12, 2008., pages 164–175, 2008.
31
Yoshimasa Takabatake, Yasuo Tabei, and Hiroshi Sakamoto. Improved esp-index: A practical self-index for highly repetitive texts. In Proceedings of the 13th Symposium on Experimental Algorithms, pages 338–350, 2014.
32
Daniel Valenzuela. CHICO: A compressed hybrid index for repetitive collections. In Proceedings of the fifteenth International Symposium on Experimental Algorithms (SEA 2016),
Lecture Notes in Computer Science. Springer, June 2016.
33
Dan E Willard. Log-logarithmic worst-case range queries are possible in space Theta(n).
Information Processing Letters, 17(2):81–84, 1983.
34
Jacob Ziv and Abraham Lempel. A universal algorithm for sequential data compression.
IEEE Transactions on information theory, 23(3):337–343, 1977.
Information Processing Letters,
CVIT 2016
23:16
Practical combinations of repetition-aware data structures
A
Proof of Lemma 1
We use the algorithm described in [2] to enumerate a representation of every node of STT #
by performing a depth-first traversal of the suffix link tree of T #. Such algorithm works in
O(n · tEL ) time and in O(σ 2 log2 n) bits of working space9 , and it provides, for each node v,
its interval in BWTT # and the length of its label, as well as the list of all its children, and
for every such child w, the interval of w in BWTT # and the first character of the label of
edge (v, w) in STT # .
Since the label of every node of the CDAWG is a maximal repeat of T #, the set of
nodes of the CDAWG (excluding the sink Ω) is in one-to-one correspondence with a subset
of the nodes of the suffix tree. Specifically, a node v of STT # corresponds to a node v 0
of the CDAWG if and only if `(v) is a left-maximal substring of T #. We can check the
left-maximality of `(v) by counting the number of distinct characters in the BWT interval
of v. Every time this number is greater than one, we have discovered a new node v 0 of the
CDAWG, and we assign to it a unique identifier by incrementing a global counter. Then, we
scan every child w of v in STT # , and we store in a hash table a tuple (i, j, v 0 .id, |`(v)|, c),
where [i..j] is the interval of w and is used as key, and v 0 .id is the unique number assigned
to v 0 . Note that every quadruplet we insert in the hash table has a unique key. If the BWT
interval of w is of length one, then w is a leaf, an arc connects v 0 in the CDAWG to the
sink Ω with character c, and c is the first character of the label of edge (v, w) in STT # . We
compute the starting position of `(v) · c for of all such arcs (v 0 , Ω) in the CDAWG in batch,
by inverting BWTT # and querying the hash table. The hash table can be implemented
to support both insertion and querying in O(1) randomized time. BWT inversion takes
O(n · tLF ) time and an amount of memory that can be charged to the output.
To build all arcs of the CDAWG that are not directed to the sink, we perform another
traversal of the suffix-link tree, in the same order as the first traversal. Assume that, during
this second traversal, we enumerate a node w whose BWT interval is present in the hash
table: then, w is the child of a node v of the suffix tree that corresponds to a node v 0 of the
CDAWG. If w corresponds to a node w0 of the CDAWG as well, i.e. if `(w) if a left-maximal
substring of T #, we add arc (v 0 , w0 ) to the CDAWG. Otherwise w is not left-maximal, thus
the CDAWG must contain arc (v 0 , u0 ) where `(u0 ) = W · `(w) is the shortest left-extension
of `(w) to be left-maximal, and the corresponding node u in the suffix tree can be reached
from w by a unary path in the suffix-link tree. Thus, we keep an auxiliary buffer, initially
empty. Every time we encounter a node w of the suffix tree whose interval [i..j] is such
that a tuple (i, j, v 0 .id, |`(v 0 )|, c) exists in the hash table, we append to the buffer tuple
(v 0 .id, |`(w)|−|`(v 0 )|, c). Moreover, if w is left-maximal, we empty the buffer and we transform
every tuple (v 0 .id, k, c) in the buffer into an arc (v 0 .id, w0 .id, k, c) of the CDAWG. The size
of such buffer can be charged to the output.
B
Proof of Lemma 2
We proceed as in Appendix A, traversing the suffix-link tree of T #, and enumerating the
intervals in BWTT # of every node of the suffix tree of T #, i.e. of every right-maximal
substring of T #. Once we detect that a node v is also left-maximal in T #, we create a
9
The enumeration algorithm described in [2] uses the stack trick to fit the working space in O(σ 2 log2 n)
bits. Without such trick, the working space would be proportional to the largest number of left-extensions
of maximal repeats that lie in the same path of the suffix-link tree of T #, which can be charged to the
output of Lemma 1.
D. Belazzougui et al.
23:17
new node v 0 of the CDAWG, we assign a new unique identifier to it, we enumerate all the
left-extensions c · `(v) in T #, and we push tuple (i, j, v 0 .id, |`(v)|, c) in a hash table, where
[i..j] is the interval of c · `(v) in BWTT # and is used as key, and v 0 .id is the unique number
assigned to v 0 . If the BWT interval of c · `(v) is of length one, an arc connects v 0 to the sink
Ω with character c in the CDAWG. We compute the starting position in T # of c · `(v) for of
all such arcs (v 0 , Ω) in the CDAWG in batch, by inverting BWTT # and querying the hash
table. Note that, since we are possibly pushing the intervals in BWTT # of the destinations
of implicit Weiner links in the suffix tree of T #, the hash table has to allow the presence of
distinct tuples with the same key.
To build all arcs of the CDAWG that are not directed to the sink, we perform another
traversal of the suffix-link tree of T #, in the same order as the first traversal. Assume that,
during this second traversal, we enumerate a node w of the suffix tree of T # whose interval
in BWTT # is present in the hash table in a set of tuples T = {(i, j, v 0 .id, |`(v)|, c)}. If w
corresponds to a node w0 of the CDAWG as well, i.e. if `(w) if a left-maximal substring of T #,
we add arc (v 0 , w0 ) to the CDAWG for every tuple in T . Otherwise w is not left-maximal in
T #, thus the CDAWG must contain arc (v 0 , u0 ) for every tuple in T , where `(u0 ) = W · `(w)
is the shortest left-extension of `(w) to be left-maximal in T #, and the corresponding node
u in the suffix tree of T # can be reached from w by a unary path in the suffix-link tree of
T #. Thus, we keep an auxiliary buffer, initially empty. Every time we encounter a node w
of the suffix tree of T # whose interval [i..j] is such that a set of tuples T exists in the hash
table, we append to the buffer a corresponding set of tuples {(v 0 .id, |`(w)|, c)}. Moreover, if
w is left-maximal in T #, we empty the buffer and we transform every tuple (v 0 .id, k, c) in
the buffer into an arc (v 0 .id, w0 .id, |`(w)| − k + 1, c) of the CDAWG. The size of such buffer
can be charged to the output.
C
Speeding up locate queries on indexes based on RLBWT and LZ
factors
On indexes based on RLBWT and LZ factors, locate queries can be further engineered in a
number of ways:
1. Thanks to the RLBWT, we know the total number of occurrences of P in T before starting
to locate them: thus, we can stop locating as soon as we have found all occurrences.
2. We could add a compressed bitvector first[1..n] that flags a position i iff SAT # [i] = pj
for some j ∈ [1..z]. As we backward-search P in BWTT # , we could mark the positions
k ∈ [2..m] such that the interval of P [k..m] in BWTT # contains only zeros in first, and
discard them in the following steps. For every discarded suffix, this strategy saves a O(m)
backward search in the bidirectional index.
3. The four-sided range reporting data structure could use the reversed prefixes T [1..pi ]
rather than the reversed prefixes T [1..pi − 1] for all i ∈ [1..z]. This would allow checking,
at position k in P , whether P [1..k − 1] ends at a position j of T such that j is the last
position of a factor, and such that T [j + 1] = P [k]. To implement an ever more stringent
filter, one could store an additional four-sided range reporting data structure that uses
the reversed prefixes T [1..pi − 1 + h] for h > 1, reverting to the four-sided data structure
for h = 1 when k > m − h.
4. In the bidirectional index, we could quit the synchronized backward search for P [1..k − 1]
in BWTT # as soon as we find a suffix P [i..k − 1] which is not right-maximal, and which
is not followed by P [k] in T #, or which does not contain a factor as a suffix. Moreover,
we could quit the backward search as soon as we detect that P [i..k − 1] is right-maximal,
CVIT 2016
23:18
Practical combinations of repetition-aware data structures
but it is neither a factor nor the suffix of a factor, and it does not contain a factor as a
suffix. Such tests can be implemented e.g. using variations of function I, and by replacing
the interval of P [1..k − 1] with the σ + 1 subintervals of P [1..k − 1] · a for all a ∈ [0..σ],
as described in [2]. Every backward step would have a O(σ) overhead in this case.
5. In the bidirectional index, we could speed up the backward search that we have to perform
for every P [1..k − 1], by precomputing a table of intervals in BWTT # and BWTT # for
all strings of length h. This would take at most 4σ h log n bits of additional space, or at
most z(h log σ + 4 log n) bits if we just store the h-mers that suffix a factor.
Due to lack of space, in the main paper we study just the effects of the first two
optimizations.
| 8cs.DS
|
The Power of Data Reduction for Matching
George B. Mertzios1 , André Nichterlein∗1,2 , and Rolf Niedermeier2
arXiv:1609.08879v2 [cs.DS] 16 Mar 2017
1
School of Engineering and Computing Sciences, Durham University, UK,
[email protected]
2
Institut für Softwaretechnik und Theoretische Informatik, TU Berlin, Germany,
{andre.nichterlein,rolf.niedermeier}@tu-berlin.de
Abstract
Finding maximum-cardinality matchings in undirected graphs is arguably one of the most
central √
graph primitives. For m-edge and n-vertex graphs, it is well-known to be solvable
in O(m n) time; however, for several applications this running time is still too slow. We
investigate how linear-time (and almost linear-time) data reduction (used as preprocessing)
can alleviate the situation. More specifically, we focus on (almost) linear-time kernelization.
We start a deeper and systematic study both for general graphs and for bipartite graphs.
Our data reduction algorithms easily comply (in form of preprocessing) with every solution
strategy (exact, approximate, heuristic), thus making them attractive in various settings.
1
Introduction
“Matching is a powerful piece of algorithmic magic” [18]. In Matching, given a graph, one has to
compute a maximum-cardinality set of nonoverlapping edges. Matching is arguably among the
most fundamental graph-algorithmic primitives allowing for a polynomial-time algorithm. More
√
specifically, on an n-vertex and m-edge graph a maximum matching can be found in O(m n)
time [17]. Improving this upper time bound, even for bipartite graphs, resisted decades of research.
Recently, however, Duan and Pettie [8] presented a linear-time algorithm that computes a (1 − ǫ)approximate maximum-weight matching,
where the running time dependency on ǫ is ǫ−1 log(ǫ−1 ).
√
For the unweighted case, the O(m n) algorithm of Micali and Vazirani [17] implies a lineartime (1 − ǫ)-approximation, where in this case the running time dependency on ǫ is ǫ−1 [8]. We
take a different route: First, we do not give up the quest for optimal solutions. Second, we
focus on efficient data reduction rules—not solving an instance but significantly shrinking its
size before actually solving the problem1 . In the context of decision problems, in parameterized
algorithmics this is known as kernelization, a particularly active area of algorithmic research on
NP-hard problems.
The spirit behind our approach is thus closer to the identification of efficiently (i.e. linearly)
solvable special cases of Matching. There is quite some body of work in this direction. For
instance, since an augmenting path can be found in linear time [10], the standard augmenting
path-based algorithm runs in O(s(n + m)) time, where s is the number of edges in the maximum
matching. Yuster [20] developed an O(rn2 log n)-time algorithm, where r is the difference between
maximum and minimum vertex degree of the input graph. Moreover, there are linear-time algorithms for computing maximum matchings in special graph classes, including convex bipartite [19],
strongly chordal [7], and chordal bipartite graphs [6].
All this and the more general spirit of “parameterization for polynomial-time solvable problems” [12] (also referred to as “FPT in P” or “FPTP” for short) forms the starting point of our
∗ Supported by a postdoc fellowship of the German Academic Exchange Service (DAAD) while at Durham
University.
1 Doing so, however, we focus here on the unweighted case.
Parameter k
Table 1: Our kernelization results.
running time
kernel size
Results for Matching
Feedback edge number
Feedback vertex number
O(n + m) time
O(kn) time
Results for Bipartite Matching
Distance to chain graphs O(n + m) time
O(k) vertices and edges
2O(k) vertices and edges
(Theorem 2.3)
(Theorem 2.19)
O(k 3 ) vertices
(Theorem 3.10)
research. Remarkably, Fomin et al. [9] recently developed an algorithm to compute a maximum
matching in graphs of treewidth k in O(k 4 n log n) randomized time.
Following the paradigm of kernelization, that is, provably effective and efficient data reduction,
we provide a systematic exploration of the power of polynomial-time data reduction for Matching.
Thus, our aim (fitting within FPTP) is to devise problem kernels that are computable in (almost)
linear time. A particular motivation for this is that with such very efficient kernelization algorithms
it is possible to transform multiplicative (O(f (k)(n+m)) into additive (O(f ′ (k)+n+m)) “(almost)
linear-time FPTP” algorithms. Furthermore, kernelization algorithms (typically based on data
reduction rules) can be used as a preprocessing to heuristics or approximation algorithms with the
goal of getting larger matchings.
As kernelization is usually defined for decision problems, we use in the remainder of the paper
the decision version of Matching. In a nutshell, a kernelization of a decision problem instance
is an algorithm that produces an equivalent instance whose size can solely be upper-bounded by
a function in the parameter (preferably a polynomial). The focus on decision problems is justified
by the fact that all our results, although formulated for the decision version, in a straightforward
way extend to the corresponding optimization version.
(Maximum-Cardinality) Matching
Input:
An undirected graph G = (V, E) and a nonnegative integer s.
Question: Is there a size s subset MG ⊆ E of nonoverlapping (i.e. disjoint) edges?
Since solving the given instance and returning a trivial yes- or no-instance always produces a
constant-size kernel in polynomial time, we are looking for kernelization algorithms that are faster
than the algorithms solving the problem. For NP-hard problems, each kernelization algorithm,
since running in polynomial time, is (presumably) faster than any solution algorithm. This is,
of course, no longer true when applying kernelization to a polynomial-time solvable problem like
Matching. While the focus of classical kernelization for NP-hard problems is mostly on improving
the size of the kernel, we particularly emphasize that for polynomially solvable problems it now
becomes crucial to also focus on the running time of the kernelization algorithm. Moreover, the
parameterized complexity analysis framework can also be applied to the kernelization algorithm
itself. For example, a kernelization algorithm running in O(k 5 n) time (k is the “problem specific”
parameter) might be preferable to another one running in O(n3 ) time. In this paper, we present
kernelization algorithms for Matching which run in linear time (see Sections 2.1 and 3) or in
almost linear time (i.e. in O(kn) time, see Section 2.2).
Our contributions. In this paper we present three efficiently computable kernels for
Matching (see Table 1 for an overview). All our parameterizations can be categorized as “distance to triviality” [13]. They are motivated as follows. First, note that maximum-cardinality
matchings be can trivially found in linear time on trees (or forests). So we consider the corresponding edge deletion distance (feedback edge number) and vertex deletion distance (feedback
vertex number). Notably, there is a trivial linear-time algorithm for computing the feedback edge
number and there is a linear-time factor-4 approximation algorithm for the feedback vertex number [1]. We mention in passing that the parameter vertex cover number, which is lower-bounded
by the feedback vertex number, has been frequently studied for kernelization [3, 4]. In particular,
2
Gupta and Peng [14] (implicitly) provided a quadratic-size kernel for Matching with respect to
the parameter vertex cover number. Coming to bipartite graphs, note that our parameterization
by vertex deletion distance to chain graphs is motivated as follows. First, chain graphs form one
the most obvious easy cases for bipartite graphs where Matching can be solved in linear time [19].
Second, we show that the vertex deletion distance of any bipartite graph to a chain graph can be
2-approximated in linear time. Moreover, vertex deletion distance to chain graphs lower-bounds
the vertex cover number of a bipartite graph.
An overview of our main results is given in Table 1. We study kernelization for Matching
parameterized by the feedback vertex number, that is, the vertex deletion distance to a forest (see
Section 2). As a warm-up we first show that a subset of our data reduction rules for the “feedback
vertex set kernel” also yields a linear-time computable linear-size kernel for the typically much
larger parameter feedback edge number (see Section 2.1). As for Bipartite Matching no faster
algorithm is known than on general graphs, we kernelize Bipartite Matching with respect to
the vertex deletion distance to chain graphs (see Section 3).
Seen from a high level, our two main results employ the same algorithmic strategy, namely
upper-bounding (as a function of the parameter) the number of neighbors in the appropriate
vertex deletion set X; that is, in the feedback vertex set or in the deletion set to chain graphs,
respectively. To achieve this we develop new “irrelevant edge techniques” tailored to these two
kernelization problems. More specifically, whenever a vertex v of the deletion set X has large
degree, we efficiently detect edges incident to v whose removal does not change the size of the
maximum matching. Then the remaining graph can be further shrunk by scenario-specific data
reduction rules. While this approach of removing irrelevant edges is natural, the technical details
and the proofs of correctness can become quite technical and combinatorially challenging. In
particular, for the case of feedback vertex number k we could only upper-bound the number of
neighbors of each vertex in X by 2O(k) .
As a technical side remark, we emphasize that in order to achieve an (almost) linear-time
kernelization algorithm, we often need to use suitable data structures and to carefully design the
appropriate data reduction rules to be exhaustively applicable in linear time, making this form of
“algorithm engineering” much more relevant than in the classical setting of mere polynomial-time
data reduction rules.
Notation and Observations. We use standard notation from graph theory. In particular
all paths we consider are simple paths. Two paths in a graph are called internally vertex-disjoint if
they are either completely vertex-disjoint or they overlap only in their endpoints. A matching in a
graph is a set of pairwise disjoint edges. Let G = (V, E) be a graph and let M ⊆ E be a matching
in G. The degree of a vertex is denoted by deg(v). A vertex v ∈ V is called matched with respect
to M if there is an edge in M containing v, otherwise v is called free with respect to M . If the
matching M is clear from the context, then we omit “with respect to M ”. An alternating path
with respect to M is a path in G such that every second edge of the path is in M . An augmenting
path is an alternating path whose endpoints are free. It is well known that a matching M is
maximum if and only if there is no augmenting path for it. Let M ⊆ E and M ′ ⊆ E be two
matchings in G. We denote by G(M, M ′ ) := (V, M △ M ′ ) the graph containing only the edges in
the symmetric difference of M and M ′ , that is, M △ M ′ := M ∪ M ′ \ (M ∩ M ′ ). Observe that
every vertex in G(M, M ′ ) has degree at most two.
max
For a matching M ⊆ E for G we denote by MG
(M ) a maximum matching in G with the
max
largest possible overlap (in number of edges) with M . That is, MG
(M ) is a maximum matching
′
max
in G such that for each maximum matching M for G it holds that |M △ M ′ | ≥ |M △ MG
(M )|.
max
Observe that, if M is a maximum matching for G, then MG (M ) = M . Furthermore observe
max
that G(M, MG
(M )) consists of only odd-length paths and isolated vertices, and each of these
max
paths is an augmenting path for M . Moreover the paths in G(M, MG
(M )) are as short as
possible:
max
Observation 1.1. For any path v1 , v2 , . . . , vp in G(M, MG
(M )) it holds that {v2i−1 , v2j } ∈
/E
for every 1 ≤ i < j ≤ p/2.
3
Proof. Assume that {v2i−1 , v2j } ∈ E. Then v1 , v2 , . . . , v2i−2 , v2i−1 , v2j , v2j+1 , . . . , vp is a shorter
path which is also an augmenting path for M in G. The corresponding maximum matching M ′
max
max
satisfies |M △ MG
(M )| > |M △ M ′ |, a contradiction to the definition of MG
(M ).
Observation 1.2. Let G = (V, E) be a graph with a maximum matching MG , let X ⊆ V be a
vertex subset of size k, and let MG−X be a maximum matching for G − X. Then, |MG−X | ≤
|MG | ≤ |MG−X | + k.
Kernelization. A parameterized problem is a set of instances (I, k) where I ∈ Σ∗ for a
finite alphabet Σ, and k ∈ N is the parameter. We say that two instances (I, k) and (I ′ , k ′ ) of
parameterized problems P and P ′ are equivalent if (I, k) is a yes-instance for P if and only if
(I ′ , k ′ ) is a yes-instance for P ′ . A kernelization is an algorithm that, given an instance (I, k)
of a parameterized problem P , computes in polynomial time an equivalent instance (I ′ , k ′ ) of P
(the kernel ) such that |I ′ | + k ′ ≤ f (k) for some computable function f . We say that f measures
the size of the kernel, and if f (k) ∈ k O(1) , we say that P admits a polynomial kernel. Often, a
kernel is achieved by applying polynomial-time executable data reduction rules. We call a data
reduction rule R correct if the new instance (I ′ , k ′ ) that results from applying R to (I, k) is
equivalent to (I, k). An instance is called reduced with respect to some data reduction rule if
further application of this rule has no effect on the instance.
2
Kernelization for Matching on General Graphs
In this section, we investigate the possibility of efficient and effective preprocessing for Matching.
As a warm-up, we first present in Section 2.1 a simple, linear-size kernel for Matching with respect
to the parameter “feedback edge set”. Exploiting the data reduction rules and ideas used for this
kernel, we then present in Section 2.2 the main result of this section: an exponential-size kernel
for the smaller parameter “feedback vertex number”.
2.1
Warm-up: Parameter feedback edge number
We provide a linear-time computable linear-size kernel for Matching parameterized by the feedback edge number, that is, the size of a minimum feedback edge set. Observe that a minimum
feedback edge set can be computed in linear time via a simple depth-first search or breadth-first
search. The kernel is based on the next two simple data reduction rules due to Karp and Sipser
[16]. They deal with vertices of degree at most two.
Reduction Rule 2.1. Let v ∈ V . If deg(v) = 0, then delete v. If deg(v) = 1, then delete v and
its neighbor and decrease the solution size s by one (v is matched with its neighbor).
Reduction Rule 2.2. Let v be a vertex of degree two and let u, w be its neighbors. Then remove v,
merge u and w, and decrease the solution size s by one.
The correctness was stated by Karp and Sipser [16]. For completeness, we give a proof.
Lemma 2.1. Reduction Rules 2.1 and 2.2 are correct.
Proof. If v has degree zero, then clearly v cannot be in any matching and we can remove v.
If v has degree one, then let u be its single neighbor. Let M be a maximum matching of size
at least s for G. Then v is matched in M since otherwise adding the edge {u, v} would increase
the size of the matching. Thus, a maximum matching in G′ = G − u − v is of size at least s − 1.
Conversely, a maximum matching of size s − 1 in G′ can easily be extended by the edge {u, v} to
a maximum matching of size s in G.
If v has degree two, then let u and w be its two neighbors. Let M a maximum matching of
size at least s. If v is not matched in M , then u and w are matched since otherwise adding the
edge {u, v} resp. {v, w} would increase the size of the matching. Thus, deleting v and merging u
4
and w decreases the size of M by one (M looses either the edge incident to v or one of the edges
incident to u and w). Hence, the resulting graph G′ has a maximum matching of size at least s− 1.
Conversely, let M ′ be a matching of size at least s − 1 for G′ . If the merged vertex vw is free,
then M := M ′ ∪ {{u, v}} is a matching of size s in G. Otherwise, vw is matched to some vertex y.
Then matching y in G with either v or w (at least one of the two vertices is a neighbor of y) and
matching u with the other vertex yields a matching of size s for G.
Although Reduction Rules 2.1 and 2.2 are correct, it is not clear whether Reduction Rule 2.2
can be exhaustively applied in linear time. However, for our purpose it suffices to consider the
following restricted version which we can exhaustively apply in linear time.
Reduction Rule 2.3. Let v be a vertex of degree two and u, w be its neighbors with u and w
having degree at most two. Then remove v, merge u and w, and decrease s by one.
Lemma 2.2. Reduction Rules 2.1 and 2.3 can be exhaustively applied in O(n + m) time.
Proof. We give an algorithm which exhaustively applies Reduction Rules 2.1 and 2.3 in linear
time. First, using bucket sort, sort the vertices by degree and keep three lists containing all
degree-zero/one/two vertices. Then one applies Reduction Rules 2.1 and 2.3 in a straightforward
way. When a neighbor of a vertex is deleted, then check if the vertex has now degree zero, one, or
two. If yes, then add the vertex to the corresponding list.
We next show that this algorithm runs in linear time. First, observe that the deletion of
each each degree-zero vertex can be done in constant time as no further vertices are affected.
Second, consider a degree-one vertex v with a neighbor u and observe that deleting u and v can
be done O(deg(v)) time since one needs to update the degrees of all neighbors of v. Furthermore,
decreasing s by one can be done in constant time for each deleted degree-one vertex. Finally,
consider a degree-two vertex v with two neighbors u and w, each of degree at most two. Deleting v
takes constant time. To merge u and w iterate over all neighbors of u and add them to the
neighborhood of w. If a neighbor u′ of u is already a neighbor of w, then decrease the degree of u′
by one. Then, relabel w to be the new contracted vertex uw.
Overall, the worst-case running
time to apply Reduction Rules 2.1 and 2.3 exhaustively can
P
be upper-bounded by O(n + v∈V deg(v)) = O(n + m).
Theorem 2.3. Matching admits a linear-time computable linear-size kernel with respect to the
parameter “feedback edge number” k.
Proof. Apply Reduction Rules 2.1 and 2.3 exhaustively in linear time (see Lemma 2.2). We claim
that the reduced graph G = (V, E) has less than 12k vertices and 13k edges. Denote with X ⊆ E a
≥3
1
2
feedback edge set for G, |X| ≤ k. Furthermore, denote with VG−X
, VG−X
, and VG−X
the vertices
1
that have degree one, two, and more than two in the G−X. Thus, |VG−X | ≤ 2k as each leaf in G−X
≥3
1
has to be incident to an edge in X. Next, since G−X is a forest (or tree), we have |VG−X
| < |VG−X
|
≥3
and thus |VG−X | < 2k. Finally, each degree-two vertex in G needs at least one neighbor of degree
2
at least three since G is reduced with respect to Reduction Rule 2.3. Thus, the vertices in VG−X
≥3
are either incident to an edge in X or adjacent to one of the at most |VG−X | + 2k vertices
≥3
in G that have degree at least three. Since the sum over all degrees of vertices in VG−X
is at
P
≥3
1
2
most v∈V ≥3 degG−X (v) ≤ 2|VG−X | + |VG−X | < 6k, it follows that |VG−X | ≤ 8k. Thus, the
G−X
≥3
1
2
number of vertices in G is |VG−X
| + |VG−X
| + |VG−X
| ≤ 12k. Since G − X is a forest, it follows
that G has at most |V | + k ≤ 13k edges.
√
Applying the O(m n)-time algorithm for Matching [17] on the kernel yields:
Corollary 2.4. Matching can be solved in O(n + m + k 1.5 ) time, where k is the feedback vertex
number.
5
2.2
Parameter feedback vertex number
We next provide for Matching a kernel of size 2O(k) computable in O(kn) time where k is the
“feedback vertex number”. Using a known linear-time factor 4-approximation algorithm [1], we
can approximate feedback vertex set and use it in our kernelization algorithm.
Roughly speaking, our kernelization algorithm extends the linear-time computable kernel with
respect to the parameter “feedback edge set”. Thus, Reduction Rules 2.1 and 2.3 play an important role in the kernelization. Compared to the other kernels presented in this paper, the kernel
presented here comes at the price of higher running time O(kn) and bigger kernel size (exponential
size). It remains open whether Matching parameterized by the “feedback vertex number” admits
a linear-time computable kernel (possibly of exponential size), and whether it admits a polynomial
kernel computable in O(kn) time.
Subsequently, we describe our kernelization algorithm which keeps in the kernel all vertices in
the given feedback vertex set X and shrinks the size of G − X. Before doing so, we need some
further notation. In this section, we assume that each tree is rooted at some arbitrary (but fixed)
vertex such that we can refer to the parent and children of a vertex. A leaf in G − X is called a
bottommost leaf either if it has no siblings or if all its siblings are also leaves. (Here, bottommost
refers to the subtree with the root being the parent of the considered leaf.) The outline of the
algorithm is as follows (we assume throughout that k < log n since otherwise the input instance
is already a kernel of size O(2k )):
1. Reduce G wrt. Reduction Rules 2.1 and 2.3.
2. Compute a maximum matching MG−X in G − X.
3. Modify MG−X in linear time such that only the leaves of G − X are free(Section 2.2.1).
4. Bound the number of free leaves in G − X by k 2 (Section 2.2.2).
5. Bound the number of bottommost leaves in G − X by O(k 2 2k )(Section 2.2.3).
6. Bound the degree of each vertex in X by O(k 2 2k ). Then, use Reduction Rules 2.1 and 2.3
to provide the kernel of size 2O(k) (Section 2.2.4).
Whenever we reduce the graph at some step, we also show that the reduction is correct. That
is, the given instance is a yes-instance if and only if the reduced one is a yes-instance. The
correctness of our kernelization algorithm then follows by the correctness of each step. We discuss
in the following some details of each step.
2.2.1
Steps 1 to 3
By Lemma 2.2 we can perform Step 1 in linear time. By Lemma 2.1 this step is correct.
A maximum matching in Step 2 can be computed by repeatedly matching a free leaf to its neighbor and by removing both vertices from the graph (thus effectively applying Reduction Rule 2.1
to G − X). By Lemma 2.2, this can be done in linear time.
Step 3 can be done in O(n) time by traversing each tree in MG−X in a BFS manner starting
from the root: If a visited inner vertex v is free, then observe that all children are matched
since MG−X is maximum. Pick an arbitrary child u of v and match it with v. The vertex w that
was previously matched to u is now free and since it is a child of u, it will be visited in the future.
Observe that Steps 2 and 3 do not change the graph but only the auxiliary matching MG−X , and
thus these steps are correct.
2.2.2
Step 4.
Recall that our goal is to upper-bound the number edges between vertices of X and V \X, since we
can then use a simple analysis as for the parameter “feedback edge set”. Furthermore, recall that
by Observation 1.2 the size of any maximum matching in G is at most k plus the size of MG−X .
6
Now, the crucial observation is that if a vertex x ∈ X has at least k neighbors in V \ X that are
free wrt. MG−X , then there exists a maximum matching where x is matched to one of these k
vertices since at most k − 1 can be “blocked” by other matching edges. This means we can delete
all other edges incident to x. Formalizing this idea, we obtain the following reduction rule.
Reduction Rule 2.4. Let G = (V, E) be a graph, let X ⊆ V be a subset of size k,and let
MG−X be a maximum matching for G − X. If there is a vertex x ∈ X with at least k free
neighbors Vx = {v1 , . . . , vk } ⊆ V \ X, then delete all edges from x to vertices in V \ Vx .
Lemma 2.5. Reduction Rule 2.4 is correct and can be exhaustively applied in O(n + m) time.
Proof. We first discuss the correctness and then the running time. Denote by s the size of a
maximum matching in the input graph G = (V, E) and by s′ the size of a maximum matching
in the new graph G′ = (V ′ , E ′ ), where some edges incident to x are deleted. We need to show
that s = s′ . Since any matching in G′ is also a matching in G, we easily obtain s ≥ s′ . It remains
max
to show s ≤ s′ . To this end, let MG := MG
(MG−X ) be a maximum matching for G with the
maximum overlap with MG−X (see ??). If x is free wrt. MG or if x matched to a vertex v that
is also in G′ a neighbor of x, then MG is also a matching in G′ (MG ⊆ E ′ ) and thus we have
in this case s ≤ s′ . Hence, consider the remaining case where x is matched to some vertex v
such that {v, x} ∈
/ E ′ , that is, the edge {v, x} was deleted by Reduction Rule 2.4. Hence, x has k
neighbors v1 , . . . , vk in V \ X such that each of these neighbors is free wrt. MG−X and none of the
edges {vi , x}, i ∈ [k], was deleted. Observe that by the choice of MG , the graph G(MG−X , MG )
(the graph over vertex set V and the edges that are either in MG−X or in MG , see ??) contains
exactly s− |MG−X | paths (we do not consider isolated vertices as paths). Each of these paths is an
augmenting path for MG−X . By Observation 1.2, we have s − |MG−X | ≤ k. Observe that {v, x} is
an edge in one of these augmenting paths; denote this path with P . Thus, there are at most k − 1
paths G(MG−X , MG ) that do not contain x. Also, each of these paths contains exactly two vertices
that are free wrt. MG−X : the endpoints of the path. This means that no vertex in X is an inner
vertex on such a path. Furthermore, since MG−X is a maximum matching, it follows that for each
path at most one of these two endpoints is in V \ X. Hence, at most k − 1 vertices of v1 , . . . , vk are
contained in the k − 1 paths of G(MG−X , MG ) except P . Therefore, one of these vertices, say vi ,
is free wrt. MG and can be matched with x. Thus, by reversing the augmentation along P and
′
′
adding the edge {vi , x} we obtain another matching MG
of size s. Observe that MG
is a matching
′
′
for G and for G and thus we have s ≤ s . This completes the proof of correctness.
Now we come to the running time. We exhaustively apply the data reduction rule as follows.
First, initialize for each vertex x ∈ X a counter with zero. Second, iterate over all free vertices
in G − X in an arbitrary order. For each free vertex v ∈ V \ X iterate over its neighbors in X. For
each neighbor x ∈ X do the following: if the counter is less than k, then increase the counter by one
and mark the edge {v, x} (initially all edges are unmarked). Third, iterate over all vertices in X.
If the counter of the currently considered vertex x is k, then delete all unmarked edges incident
to x. This completes the algorithm. Clearly, it only deletes edges incident to a vertex x ∈ X only
if x has k free neighbors in V \ X and the edges to these k neighbors are kept. The running time
is O(n + m): When iterating over all free vertices in V \ X we consider each edge at most once.
Furthermore, when iterating over the vertices in X, we again consider each edge at most once.
To finish Step 4, we exhaustively apply Reduction Rule 2.4 in linear time. Afterwards, there
are at most k 2 free (wrt. to MG−X ) leaves in G − X that have at least one neighbor in X since
each of the k vertices in X is adjacent to at most k free leaves. Thus, applying Reduction Rule 2.1
we can remove the remaining free leaves that have no neighbor in X. However, since for each
degree-one vertex also its neighbor is removed, we might create new free leaves and need to again
apply Reduction Rule 2.4 and update the matching (see Step 3). This process of alternating
application of Reduction Rules 2.1 and 2.4 stops after at most k rounds since the neighborhood
of each vertex in X can be changed by Reduction Rule 2.4 at most once. This shows the running
time O(k(n + m)). We next show how to improve this to O(n + m) and arrive at the final lemma
of this subsection.
7
Algorithm 1: Reduce(G, MG−X ).
Input: A matching instance (G = (V, E), s) and a feedback vertex set X ⊆ V for G with |X| = k.
Output: An equivalent matching instance (G′ , s′ ) such that X is also a feedback vertex set for G′
and a maximum matching MG′ −X for G′ − X such that only at most k2 leaves in G′ − X
are free.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
Reduce G wrt. Reduction Rules 2.1 and 2.3
Compute a maximum matching MG−X as described in Step 3
foreach x ∈ X do c(x) ← 0
// c(x) will store the number of free neighbors for x
foreach e ∈ E do marked(e) ← False
L ← stack containing all free leaves in G − X
while L is not empty do
u ← pop(L)
foreach x ∈ NG (u) ∩ X do
// Check whether Reduction Rule 2.4 is applicable for x
c(x) ← c(x) + 1, marked({u, x}) ← True
// fix u as free neighbor of x
if c(x) = k then
// x has enough free neighbors: apply Reduction Rule 2.4
foreach y ∈ NG (x) ∩ X do delete {x, y}
foreach v ∈ NG (x) \ X do
if marked({x, v}) = False then
delete {x, v}
// Next deal with the case that v becomes a degree-one vertex
if degG (v) = 1 and v is free then push v on L
if degG (v) = 1 and v is matched then
delete v and its neighbor, s ← s − 1
if degG (u) = 1 then
// u has no neighbors in x
v ← neighbor of u in G − X; w ← matched neighbor of v
delete u and v from G
MG−X ← MG−X \ {{v, w}}, s ← s − 1
// update MG−X and s
if w is now a leaf in G − X then
add w to L
// w is free, so add w to the list of vertices to check
else
P ← arbitrary alternating path from w to a leaf u′ in the subtree rooted in w
augment MG−X along P
// ensure that the only free vertices are leaves
push u′ on L
// u′ is free, add u′ to the list of vertices to check
return (G, s) and MG−X .
Lemma 2.6. Given a matching instance (G, s) and a feedback vertex set X, Algorithm 1 computes
in linear time an instance (G′ , s′ ) with feedback vertex set X and a maximum matching MG′ −X
in G′ − X such that the following holds.
• There is a matching of size s in G if and only if there is a matching of size s′ in G′ .
• Each vertex that is free wrt. MG′ −X is a leaf in G′ − X.
• There are at most k 2 free leaves in G′ − X.
Proof. In the following, we explain Algorithm 1 which reduces the graph with respect to
Reduction Rules 2.1 and 2.4 and updates the matching MG−X as described in Step 3. The algorithm performs in (Lines 1 and 2) Steps 1 to 3. As described in the previous section, this can
be done in linear time. Next, Reduction Rule 2.4 is applied in Lines 8 to 17 using the approach
described in the proof of Lemma 2.5: For each vertex in x a counter c(x) is maintained. When
iterating over the free leaves in G − X, these counters will be updated. If a counter c(x) reaches k,
then the algorithm knows that x has k fixed free neighbors and according to Reduction Rule 2.4
8
the edges to all other vertices can be deleted (see Line 10). Observe that once the counter c(x)
reaches k, the vertex x will never be considered again by the algorithm since its only remaining
neighbors are free leaves in G − X that already have been popped from stack L. The only difference from the description in the proof of Lemma 2.5 is that the algorithm reacts if the degree of
some vertex v in G − X is decreased to one (see Lines 15 to 17). If v is matched, then simply remove v and its matched neighbor from G and MG−X . Otherwise, add v to the list L of unmatched
degree-one vertices and defer dealing with v to a latter stage of the algorithm.
Observe that the matching MG−X still satisfies the property that each free vertex in G − X is
a leaf since only matched vertex pairs were deleted so far. When deleting unmatched degree-one
vertices and their respective neighbor, the maximum matching MG−X needs to be updated to
satisfy this property. The algorithm does this from Lines 18 to 27: Let u be an entry in L such
that u has degree one in Line 18, that is, u is a free leaf in G − X and has no neighbors in X.
Then, following Reduction Rule 2.1, delete u and its neighbor v and decrease the solution size s
by one (see Lines 20 and 21). Let w denote the previously matched neighbor of v. Since v was
removed, w is now free. If w is a leaf in G − X, then we can simply add it to L and in this way
deal with it later. If w is not a leaf, then we need to update MG−X since only leaves are allowed to
be free. To this end, take an arbitrary alternating path P from w to a leaf u′ of the subtree with
root w and augment along P (see Lines 25 and 26). This can be done as follows: Pick an arbitrary
child w1 of w. Let w2 be the matched neighbor of w1 . Since w is the parent of w1 , it follows
that w2 is a child of w1 . Now, remove {w1 , w2 } from MG−X and add {w1 , w}. If w2 is a leaf,
then the alternating path P is found with u′ = w2 and augmented. Otherwise, repeat the above
procedure with w2 taking the role of w. This completes the algorithm. Its correctness follows
from the fact that it only deletes edges and vertices according to Reduction Rules 2.1 and 2.4.
It remains to show the running time of O(n + m). To this end, we prove that the algorithm
considers each edge in E only two times. First, consider the edges incident to a vertex x ∈ X.
These edges will be inspected at most twice by the algorithm: Once, when it is marked (see
Line 9). The second time is when it is deleted. This bounds the running time in the first part
(Lines 8 to 17).
Now consider the remaining edges within G − X. To this end, observe that the algorithm
performs two actions on the edges: deleting the edges (Line 20) and finding and augmenting
along an alternating path (Lines 25 and 26). Clearly, after deleting an edge it will no longer
be considered, so it remains to show that edge is part of at most one alternating path used in
Lines 25 and 26. Assume toward a contradiction that the algorithm augments along an edge twice
or more. From all the edges that are augmented twice or more let e ∈ E be one that is closest
to the root of the tree e is contained in, that is, there is no edge closer to a root. Let w1 and u′1
be the endpoints of the first augmenting path P1 containing e and w2 and u′2 the endpoints of
the second augmenting path P2 containing e. Observe that for each augmenting path chosen in
Line 25 it holds that one endpoint is a leaf and the other endpoint is an ancestor of this leaf.
Assume without loss of generality that u′1 and u′2 are the leaves and w1 and w2 are their respective
ancestors. Let u1 and v1 (u2 and v2 ) be the vertices deleted in Line 20 which in turn made w1
(w2 ) free. Observe that e does not contain any of these four vertices u1 , v1 , u2 , v2 since before
augmenting P1 (P2 ) the vertices u1 and v1 (u2 and v2 ) are deleted. Since e is contained in both
paths, either w1 is an ancestor of w2 or vice versa: the case w1 = w2 cannot happen since for the
second augmenting path the endpoint w2 = w1 would not be matched to v2 ; a contradiction (see
Line 19).
We next consider the case that w1 is an ancestor of w2 (the other case will be handled subsequently). Denote with w2′ the neighbor of w2 on P ′ . Observe that e = {w2 , w2′ } since e is chosen as
being closest to the root. We next distinguish the two cases whether or not e is initially matched.
If e is initially free, then e is matched after augmenting along P1 . Then, by choice of P2 , e is
not changed until the augmentation along P2 . This, however, is a contradiction since augmenting
along P2 only happens after the matched edge {w2 , v2 } is deleted. Since w2′ 6= v2 and e is matched
all the time until u2 and v2 are deleted, this means that w2 would be matched to two vertices.
Thus, consider the case that e is initially matched. Then, after augmenting along P1 , e is free
and w2 is matched to its parent pw2 . As a consequence, v2 is not matched to w2 , neither before
9
nor after the augmentation of P1 . Since the algorithm augments along P2 only after it deleted u2
and v2 where v2 is matched to w2 , it follows that the edge {w2 , v2 } is augmented before the algorithm augments along P2 . Denote with P3 the augmenting path containing the edge {w2 , v2 }.
Since w2 is apparently not a free leaf, it follows that P3 needs to contain the matched neighbor
of w2 , which is pw2 . This means that the edge {w2 , pw2 } is augmented at least twice (through P1
and P3 ). However, {w2 , pw2 } is closer to the root than e = {w2 , w2′ }, a contradiction to the choice
of e. This completes the case that w1 is an ancestor of w2 .
We now consider the remaining case where w2 is an ancestor of w1 . In this case we have e =
{w1 , w1′ } where w1′ is the neighbor of w1 on P1 . Observe that w1′ is a child of w1 . Furthermore,
observe that after the augmentation along P1 the leaf u′1 is free and can be reached by an alternating
path from w1 . Hence, before and after the augmentation along P1 it holds that w1 can reach exactly
one free leaf via an alternating path (u1 and u′1 ). Observe that this is true even if the algorithm
removes u′1 since then a new free leaf will be created. Thus, before deleting u2 and v2 (right before
the augmentation along P2 ), there is an augmenting path in G − X from u2 to w1 and to the free
leaf reachable from w1 . This is a contradiction to the fact that the matching MG−X is maximum.
We conclude that each edge in E will be augmented at most once. Thus, the algorithm considers
each edge at most twice (when augmenting it and when deleting it). Hence, the algorithm runs in
linear time.
Summarizing, in Step 4 we apply Algorithm 1 in order to obtain an instance with at most k 2
free vertices in G − X that are all leaves. By Lemma 2.6 this can be done in linear time. Furthermore, Lemma 2.6 also shows that the step is correct.
2.2.3
Step 5
In this step we reduce the graph in O(kn) time so that at most k 2 (2k + 1) bottommost leaves will
remain in the forest G − X. We will restrict ourselves to consider leaves that are matched with
their parent vertex in MG−X and that do not have a sibling. We call these bottommost leaves
interesting. Any sibling of a bottommost leaf is by definition also a leaf. Thus, at most one of
these leaves (the bottommost leaf or its siblings) is matched with respect to MG−X and all other
leaves are free. Recall that in the previous step we upper-bounded the number of free leaves with
respect to MG−X by k 2 . Hence there are at most k 2 bottommost leaves that are not interesting.
Our general strategy for this step is to extend the idea behind Reduction Rule 2.4: We want
to keep for each pair of vertices x, y ∈ X at most k different internally vertex-disjoint augmenting
paths from x to y. (For ease of notation we keep k paths although keeping k/2 is sufficient.) In
this step, we only consider augmenting paths of the form x, u, v, y where v is a bottommost leaf
and u is v’s parent in G − X. Assume that the parent u of v is adjacent to some vertex x ∈ X.
Observe that in this case any augmenting path starting with the two vertices x and u has to
continue to v and end in a neighbor of v. Thus, the edge {x, u} can be only used in augmenting
paths of length three. Furthermore, all these length-three augmenting paths are clearly internally
vertex-disjoint. If we do not need the edge {x, u} because we kept k augmenting paths from x
already, then we can delete {x, u}. Furthermore, if we deleted the last edge from u to X (or u
had no neighbors in X in the beginning), then u is a degree-two vertex in G and can be removed
by applying Reduction Rule 2.2. As the child v of u is a leaf in G − X, it follows that v has at
most k+1 neighbors in G. We show below (Lemma 2.7) that the application of Reduction Rule 2.2
to remove u takes O(k) time. As we remove at most n vertices, at most O(kn) time is spent on
Reduction Rule 2.2 in this step.
We now show that after a simple preprocessing one application of Reduction Rule 2.2 in the
algorithm above can indeed be performed in O(k) time.
Lemma 2.7. Let u be a leaf in the tree G − X, v be its parent, and let w be the parent of v. If v
has degree two in G, then applying Reduction Rule 2.2 to v (deleting v, contracting u and v, and
setting s := s − 1) can be done in O(k) time plus O(kn) time for an initial preprocessing.
10
Algorithm 2: FVS-Kernel Step 5.
Input: A matching instance (G = (V, E), s), a feedback vertex set X ⊆ V of size k for G
with k < log n, and a maximum matching MG−X for G − X with at most k2 free vertices
in G − X that are all leaves.
Output: An equivalent matching instance (G′ , s′ ) such that X is also a feedback vertex set for G′
and G − X is a tree with at most k2 (2k + 1) bottommost leaves, and a maximum
matching MG′ −X for G′ − X with at most k2 free vertices in G′ − X that are all leaves.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
Fix an arbitrary bijection f : 2X → {1, . . . , 2k }
foreach v ∈ V \ X do
Set fX (v) ← f (N (v) ∩ X)
// The number fX (v) < n can be read in constant time.
Initialize a table Tab of size k · 2k with Tab[x, f (Y )] ← 0 for all x ∈ X, ∅ ( Y ⊆ X
P ← List containing all parents of interesting bottommost leaves
while P is not empty do
u ← pop(P )
v ← child vertex of u in G − X
foreach x ∈ N (u) ∩ X do
if Tab[x, fX (v)] < k then
Tab[x, fX (v)] ← Tab[x, fX (v)] + 1
else
delete {x, u}
if u has now degree two in G then
Apply Reduction Rule 2.2 to u
// This decreases s by one.
vw ← vertex resulting from the merge of v and the parent w of u
if vw is now an interesting bottommost leaf then
add the parent of vw to P
return (G, s) and MG−X .
Proof. The preprocessing is to simply create a partial adjacency matrix for G with the vertices
in X in one dimension and V in the other dimension. This adjacency matrix has size O(kn) and
can clearly be computed in O(kn) time.
Now apply Reduction Rule 2.2 to v. Deleting v takes constant time. To merge u and w iterate
over all neighbors of u. If a neighbor u′ of u is already a neighbor of w, then decrease the degree
of u′ by one, otherwise add u′ to the neighborhood of w. Then, relabel w to be the new merged
vertex uw.
Since u is a leaf in G − X and its only neighbor in G − X, namely v, is deleted, it follows that
all remaining neighbors of u are in X. Thus, using the above adjacency matrix, one can check in
constant time whether u′ is a neighbor of w. Hence, the above algorithm runs in O(deg(u)) = O(k)
time.
The above ideas are used in Algorithm 2 which we use for this step (Step 5). The algorithm
is explained in the proof of the following lemma stating the correctness and the running time of
Algorithm 2.
Lemma 2.8. Let (G = (V, E), s) be a matching instance, let X ⊆ V be a feedback vertex set, and
let MG−X be a maximum matching for G − X with at most k 2 free vertices in G − X that are all
leaves. Then, Algorithm 2 computes in O(kn) time an instance (G′ , s′ ) with feedback vertex set X
and a maximum matching MG′ −X in G′ − X such that the following holds.
• There is a matching of size s in G if and only if there is a matching of size s′ in G′ .
• There are at most k 2 (2k + 1) bottommost leaves in G′ − X.
• There are at most k 2 free vertices in G′ − X and they are all leaves.
11
Proof. We start with describing the basic idea of the algorithm. To this end, let {u, v} ∈ E be
an edge such that v is an interesting bottommost leaf, that is, without siblings and matched to
its parent u by MG−X . Counting for each pair x ∈ N (u) ∩ X and y ∈ N (v) ∩ X one augmenting
path gives in a simple worst-case analysis O(k 2 ) time per edge, which is too slow for our purposes.
Instead, we count for each pair consisting of a vertex x ∈ N (u) ∩ X and a set Y = N (v) ∩ X one
augmenting path. In this way, we know that for each y ∈ Y there is one augmenting path from x
to y without iterating through all y ∈ Y . This comes at the price of considering up to k2k such
pairs. However, we will show that we can do the computations in O(k) time per considered edge
in G − X. The main reason for this improved running time is a simple preprocessing that allows
for a bottommost vertex v to determine N (v) ∩ X in constant time.
The preprocessing is as follows (see Lines 1 to 3): First, fix an arbitrary bijection f between the
set of all subsets of X to the numbers {1, 2, . . . , 2k }. This can be done for example by representing
a set Y ⊆ X = {x1 , . . . , xk } by a length-k binary string (a number) where the ith position is 1
if and only if xi ∈ Y . Given a set Y ⊆ X such a number can be computed in O(k) time in a
straightforward way. Thus, Lines 1 to 3 can be performed in O(kn) time. Furthermore, since we
assume that k < log n (otherwise the input instance is already an exponential kernel), we have
that f (Y ) < n for each Y ⊆ X. Thus, reading and comparing these numbers can be done in
constant time. Furthermore, in Line 3 the algorithm precomputes for each vertex the number
corresponding to its neighborhood in X.
After the preprocessing, the algorithm uses a table Tab where it counts an augmenting path
from a vertex x ∈ X to a set Y ⊆ X whenever a bottommost leaf v has exactly Y as neighborhood
in X and the parent of v is adjacent to x (see Lines 4 to 18). To do this in O(kn) time, the
algorithm proceeds as follows: First, it computes in Line 5 the set P which contains all parents
of interesting bottommost leaves. Clearly, this can be done in linear time. Next, the algorithm
processes the vertices in P . Observe that further vertices might be added to P (see Line 18) during
this processing. Let u be the currently processed vertex of P , let v be its child vertex, and let Y
be the neighborhood of v in X. For each neighbor x ∈ N (u) ∩ X, the algorithm checks whether
there are already k augmenting paths between x and Y with a table lookup in Tab (see Line 10).
If not, then the table entry is incremented by one (see Line 11) since u and v provide another
augmenting path. If yes, then the edge {x, u} is deleted in Line 13 (we show below that this does
not change the maximum matching size). If u has degree two after processing all neighbors of u
in X, then by applying Reduction Rule 2.2, we can remove u and contract its two neighbors v
and w. It follows from Lemma 2.7 that this application of Reduction Rule 2.2 can be done in O(k)
time. Hence, Algorithm 2 runs in O(kn) time.
Recall that all vertices in G−X that are free wrt. MG−X are leaves. Thus, the changes to MG−X
by applying Reduction Rule 2.2 in Line 15 are as follows: First, the edge {u, v} is removed and
second the edge {w, q} is replaced by {vw, q} for some q ∈ V . Hence, the matching MG−X after
running Algorithm 2 has still at most k 2 free vertices and all of them are leaves.
It remains to prove that (a) the deletion of the edge {x, u} in Line 13 results in an equivalent
instance and (b) that the resulting instance has at most O(k2k ) bottommost leaves. First, we
show (a). To this end, assume towards a contradiction that the new graph G′ := G − {x, u}
has a smaller maximum matching than G (clearly, G′ cannot have a larger maximum matching).
Thus, any maximum matching MG for G has to contain the edge {x, u}. This implies that
the child v of u in G − X is matched in MG with one of its neighbors (except u): If v is free
wrt. MG , then deleting {x, u} from MG and adding {v, u} yields another maximum matching not
containing {x, u}, a contradiction. Recall that N (v) = {u} ∪ Y where Y ⊆ X since v is a leaf
in G − X. Thus, each maximum matching MG for G contains for some y ∈ Y the edge {v, y}.
Observe that Algorithm 2 deletes {x, u} only if there are at least k other interesting bottommost
leaves v1 , . . . , vk in G − X such that their respective parent is adjacent to x and N (vi ) ∩ X = Y
(see Lines 9 to 13). Since |Y | ≤ k, it follows by the pigeon hole principle that at least one of these
vertices, say vi , is not matched to any vertex in Y . Thus, since vi is an interesting bottommost
leaf, it is matched to its only remaining neighbor: its parent ui in G − X. This implies that there
12
is another maximum matching
′
MG
:= (MG \ {{v, y}, {x, u}, {ui, vi }}) ∪ {{vi , y}, {x, ui }, {u, v}},
a contradiction to the assumption that all maximum matchings for G have to contain {x, u}.
We next show (b) that the resulting instance has at most k 2 (2k + 1) bottommost leaves. To
this end, recall that there are at most k 2 bottommost leaves that are not interesting (see discussion
at the beginning of this subsection). Hence, it remains to upper-bound the number of interesting
bottommost leaves. Observe that each parent u of an interesting bottommost leaf has to be
adjacent to a vertex in X since otherwise u would have been deleted in Line 15. Furthermore,
after running Algorithm 2, each vertex x ∈ X is adjacent to at most k2k parents of interesting
bottommost leaves (see Lines 10 to 13). Thus, the number of interesting bottommost leaves is at
most k 2 2k . Therefore the number of bottommost leaves is upper-bounded by k 2 (2k + 1).
2.2.4
Step 6
In this subsection, we provide the final step of our kernelization algorithm. Recall that in the
previous steps we have upper-bounded the number of bottommost leaves in G − X by O(k 2 2k ),
we computed a maximum matching MG−X for G − X such that at most k 2 vertices are free
wrt. MG−X and all free vertices are leaves in G − X. Using this, we next show how to reduce G
to a graph of size O(k 3 2k ). To this end we need some further notation. A leaf in G − X that is
not bottommost is called a pendant. We define T to be the pendant-free tree (forest) of G − X,
that is, the tree (forest) obtained from G − X by removing all pendants. The next observation
shows that G − X is not much larger than T . This allows us to restrict ourselves in the following
on giving an upper bound on the size of T .
Observation 2.9. Let G − X be as described above with vertex set V \ X and let T be the
pendant-free tree (forest) of G − X with vertex set VT . Then, |V \ X| ≤ 2|VT | + k 2 .
Proof. Observe that V \ X is the union of all pendants in G − X and VT . Thus, it suffices to show
that G − X contains at most |VT | + k 2 pendants. To this end, recall that we have a maximum
matching for G − X with at most k 2 free leaves. Thus, there are at most k 2 leaves in G − X
that have a sibling which is also a leaf since from two leaves with the same parent at most one
can be matched. Hence, all but at most k 2 pendants in G − X have pairwise different parent
vertices. Since all these parent vertices are in VT , it follows that the number of pendants in G − X
is |VT | + k 2 .
We use the following observation to provide an upper bound on the number of leaves of T .
Observation 2.10. Let F be a forest, let F ′ be the pendant-free forest of F , and let B be the set
of all bottommost leaves in F . Then, the set of leaves in F ′ is exactly B.
Proof. First observe that each bottommost leaf of F is a leaf of F ′ since we only remove vertices
to obtain F ′ from F . Thus, it remains to show that each leaf v in F ′ is a bottommost leaf in F .
We distinguish two cases of whether or not v is a leaf in F : First, assume that v is not a
leaf in F . Thus, all of it child vertices have been removed. Since we only remove pendants to
obtain F ′ from F and since each pendant is a leaf, it follows that v is in F the parent of one or
more leaves u1 , . . . , uℓ . Thus, by definition, all these leaves u1 , . . . , uℓ are bottommost leaves, a
contradiction to the fact that they were deleted when creating F ′ .
Second, assume that v is a leaf in F . If v is a bottommost leaf, then we are done. Thus, assume
that v is not a bottommost leaf and therefore a pendant. However, since we remove all pendants
to obtain F ′ from F , it follows that v is not contained in F ′ , a contradiction.
From Observation 2.10 it follows that the set B of bottommost leaves in G − X is exactly the
set of leaves in T . In the previous step we reduced the graph such that |B| ≤ k 2 (2k + 1). Thus,
T has at most k 2 (2k + 1) vertices of degree one and, since T is a tree (a forest), T also has at
most k 2 (2k + 1) vertices of degree at least three. Let VT2 be the vertices of degree two in T and
13
w
u
vs1
vt1
vs2
vt2
v
y
x
Figure 1: The situation in the proof of Lemma 2.11. The augmenting path from u to v intersects
the two augmenting paths Pwx and Pwy from w to x and y, respectively. Bold edges indicate
edges in the matching, dashed edges indicate odd-length alternating paths starting with the first
and last edge not being in the matching. The gray paths in the background highlight the different
augmenting paths: the initial paths from u to v, w to x, and x to y as well as the new paths
from u to x and w to v as postulated by Lemma 2.11.
let VT6=2 be the remaining vertices in T . From the above it follows that |VT6=2 | ≤ 2k 2 (2k + 1).
Hence, it remains to bound the size of VT2 . To this end, we will upper-bound the degree of
each vertex in X by O(k 2 2k ) and then use Reduction Rules 2.1 and 2.3. We will check for each
edge {x, v} ∈ E with x ∈ X and V \ X whether we “need” it. This check will use the idea from
the previous subsection where each vertex in X needs to reach each subset Y ∈ X at most k
times via an augmenting path. Similarly as in the previous section, we want to keep “enough” of
these augmenting paths. However, this time the augmenting paths might be long, while different
augmenting paths might overlap. To still use the basic approach, we use the following lemma
stating that we can still somehow replace augmenting paths.
Lemma 2.11. Let MG−X be a maximum matching in the forest G−X. Let Puv be an augmenting
path for MG−X in G from u to v. Let Pwx , Pwy , and Pwz be three internally vertex-disjoint
augmenting paths from w to x, y, and z, respectively, such that Puv intersects all of them. Then,
there exist two vertex-disjoint augmenting paths with endpoints u, v, w, and one of the three
vertices x, y, and z.
Proof. Label the vertices in Puv alternating as odd or even with respect to Puv so that no two
consecutive vertices have the same label, u is odd, and v is even. Analogously, label the vertices
in Pwx , Pwy , and Pwz as odd and even with respect to Pwx , Pwy , and Pwz respectively so that w
is always odd. Since all these paths are augmenting, it follows that each edge from an even vertex
to its succeeding odd vertex is in the matching MG−X and each edge from an odd vertex to its
succeeding even vertex is not in the matching. Observe that Puv intersects each of the other
paths at least at two consecutive vertices, since every second edge must be an edge in MG−X .
Since G − X is a forest and all vertices in X are free with respect to MG−X , it follows that the
intersection of two augmenting paths is connected and thus a path. Since Puv intersects the three
augmenting paths from w, it follows that at least two of these paths, say Pwx and Pwy , have a
“fitting parity”, that is, in the intersections of Puv with Pwx and with Pwy the even vertices with
respect to Puv are either even or odd with respect to both Pwx and Pwy .
Assume without loss of generality that in the intersections of the paths the vertices have
the same label with respect to the three paths (if the labels differ, then revert the ordering of
the vertices in Puv , that is, exchange the names of u and v and change all labels on Puv to its
opposite). Denote with vs1 and vt1 the first and the last vertex in the intersection of Puv and Pwx .
Analogously, denote with vs2 and vt2 the first and the last vertex in the intersection of Puv and Pwy .
Assume without loss of generality that Puv intersects first with Pwx and then with Pwx . Observe
that vs1 and vs2 are even vertices and vt1 and vt2 are odd vertices since the intersections have to
start and end with edges in MG−X (see Fig. 1 for an illustration). For an arbitrary path P and for
two arbitrary vertices p1 , p2 of P , denote by p1 − P − p2 the subpath of P from p1 to p2 . Observe
that u − Puv − vt1 − Pwx − x and w − Pwy − vt2 − Puv − v are two vertex-disjoint augmenting
14
Algorithm 3: Algorithm for computing Step 6 of our kernel wrt. the parameter “feedback
vertex number”.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
Input: A matching instance (G = (V, E), s), a feedback vertex set X ⊆ V of size k for G
with k < log n and at most k2 (2k + 1) bottommost leaves in G − X, and a maximum
matching MG−X for G − X with at most k2 free vertices in G − X that are all leaves.
Output: An equivalent matching instance (G′ , s′ ) such that G′ contains at most O(k3 2k ) vertices
and edges.
Fix an arbitrary bijection f : 2X → {1, . . . , 2k }
foreach v ∈ V \ X do
Set fX (v) ← f (N (v) ∩ X)
// The number fX (v) < n can be read in constant time.
Initialize a table Tab of size k · 2k with Tab[x, f (Y )] ← 0 for x ∈ X, ∅ ( Y ⊆ X
T ← pendant-free tree (forest) of G − X
VT≥3 ← vertices in T with degree ≥ 3
foreach x ∈ X do
foreach v ∈ N (x) \ X do
if Keep-Edge(x, v) = false then
// Is {x, v} needed for an augmenting path?
delete {x, v}
exhaustively apply Reduction Rules 2.1 and 2.3
return (G, s).
Function Keep-Edge(x ∈ X, v ∈ V \ X)
if v is free wrt. MG−X or v ∈ VT≥3 then return true
w ← matched neighbor of v in MG−X
if w ∈ VT≥3 or w is adjacent to free leaf in G − X then return true
if w has at least one neighbor in X and Tab[x, fX (w)] < 6k2 then
Tab[x, fX (w)] ← Tab[x, fX (w)] + 1
return true
21
foreach neighbor u 6= v of w that is matched wrt. MG−X and fulfills {u, x} ∈
/ E do
if Keep-Edge(u, x) = true then return true
22
return false
20
paths.
Algorithm description. We now provide the algorithm for Step 6 (see Algorithm 3 for a
pseudocode). The algorithm uses the same preprocessing (see Lines 1 to 3) as Algorithm 2. Thus,
the algorithm can determine whether two vertices have the same neighborhood in X in constant
time. As in Algorithm 2, Algorithm 3 uses a table Tab which has an entry for each vertex x ∈ X
and each set Y ⊆ X. The table is filled in such a way that the algorithm detected for each y ∈ Y
at least Tab[x, Y ] internally vertex-disjoint augmenting paths from x to y.
The main part of the algorithm is the boolean function ‘Keep-Edge’ in Lines 13 to 22 which
makes the decision on whether or not to delete an edge {x, v} for v ∈ V \ X and x ∈ X. The
function works as follows for edge {x, v}: Starting at v the graph will be explored along possible
augmenting paths until a “reason” for keeping the edge {x, v} is found or further exploration is
possible.
If the vertex v is free wrt. MG−X , then {x, v} is an augmenting path and we keep {x, v} (see
Line 14). Observe that in Step 4 we upper-bounded the number of free vertices by k 2 and all
these vertices are leaves. Thus, we keep a bounded number of edges incident to x because the
corresponding augmenting paths can end at a free leaf. We provide the exact bound below when
discussing the size of the graph returned by Algorithm 3. In Line 14, the algorithm stops exploring
the graph and keeps the edge {x, v} if v has degree at least three in T . The reason is to keep the
graph exploration simple by following only paths in T . This ensures that the running time for
exploring the graph from x does not exceed O(n). Since the number of vertices in T with degree
15
at least three is bounded (see discussion after Observation 2.10), it follows that only a bounded
number of such edges {x, v} are kept.
If v is not free wrt. MG−X , then it is matched with some vertex w. If w is adjacent to some
leaf u in G − X that is free wrt. MG−X , then the path x, v, w, u is an augmenting path. Thus, the
algorithm keeps in this case the edge {x, v}, see Line 16. Again, since the number of free leaves
is bounded, only a bounded number of edges incident to x will be kept. If w has degree at least
three in T , then the algorithm stops the graph exploration here and keeps the edge {x, v}, see
Line 16. Again, this is to keep the running time at O(kn) overall.
Let Y ⊆ X denote the neighborhood of w in X. Thus the partial augmenting path x, v, w
can be extended to each vertex in Y . Thus, if the algorithm did not yet find 6k 2 paths from x
to vertices whose neighborhood in X is also Y , then the table entry Tab[x, fX (w)] (where fX (w)
encodes the set Y = N (w) ∩ X) is increased by one and the edge {x, v} will be kept (see Lines 18
and 19). (Here we need 6k 2 paths since these paths might be long and intersect with many other
augmenting paths, see proof of Lemma 2.15 for the details of why 6k 2 is enough.) If the algorithm
already found 6k 2 “augmenting paths” from x to Y , then the neighborhood of w in X is irrelevant
for x and the algorithm continues.
In Line 20, all above discussed cases to keep the edge {x, v} do not apply and the algorithm
extends the partial augmenting part x, v, w by considering the neighbors of w except v. Since the
algorithm dealt with possible extensions to vertices in X in Lines 17 to 19 and with extensions to
free vertices in G − X in Line 14, it follows that the next vertex on this path has to be a vertex u
that is matched wrt. MG−X . Furthermore, since we want to extend a partial augmenting path
from x, we require that u is not adjacent to x as otherwise x, u would be another, shorter partial
augmenting path from x to u and we do not need the currently stored partial augmenting path.
Statements on Algorithm 3. For each edge {x, z} with x ∈ X and z ∈ V \ X we denote
by P (x, z) the induced subgraph of G − X on the vertices that are explored in the function KeepEdge when called in Line 9 with x and z. More precisely, we initialize P (x, z) := ∅. Whenever
the algorithm reaches Line 14, we add v to P (x, z). Furthermore, whenever the algorithm reaches
Line 17, we add w to P (x, z).
We next show that P (x, z) is a path or a path with one additional pendant.
Lemma 2.12. Let x ∈ X and z ∈ V \ X be two vertices such that {x, z} ∈ E. Then, P (x, z)
is either a path or a tree with exactly one vertex z ′ having more than two neighbors in P (x, z).
Furthermore, z ′ has degree exactly three and z is a neighbor of z ′ .
Proof. We first show that all vertices in P (x, z) except z and its neighbor z ′ have degree at most
two in P (x, z). Observe that having more vertices than z and z ′ in P (x, z) requires Algorithm 3
to reach Line 20.
Let w be the currently last vertex when Algorithm 3 continues the graph exploration in Line 20.
Observe that the algorithm therefore dealt with the case that w has degree at least three in T in
Line 16. Thus, w is either a pendant leaf in G − X or w ∈
/ VT≥3 (that is, w has degree at most two
in T ). In the first case, there is no candidate to continue and the graph exploration stops. In the
second case, w has degree at most two in T .
We next show that any candidate u for continuing the graph exploration in Line 21 is not a
leaf in G − X. Assume toward a contradiction, that u is a leaf in G − X. Since the parent w of u
is matched with some vertex v 6= u (this is how w is chosen, see Line 15), it follows that u is not
matched. This implies that the function ’Keep-Edge’ would have returned true in Line 16 and
would not have reached Line 20, a contradiction. Thus, the graph exploration follows only vertices
in T . Furthermore, the above argumentation implies that w is not adjacent to a leaf unless this
leaf is its predecessor v in the graph exploration.
We now have two cases: Either w is not adjacent to a leaf in G−X or v = z is a leaf and w = z ′
is its matched neighbor. In the first case, w has at most one neighbor u 6= v since w ∈
/ VT≥3 . Hence,
′
w has degree two in P (x, z). In the second case, w = z has at most two neighbors u 6= v and u′ 6= v.
Thus, z ′ has degree at most three.
16
We set for x ∈ X
Px := {P (x, v) | {x, v} ∈ E ∧ v ∈ V \ X}
to be the union of all induced subgraphs wrt. x.
Lemma 2.13. There exists a partition of Px into Px = PxA ∪ PxB such that all graphs within PxA
and within PxB are pairwise disjoint.
Proof. Since G − X is a tree (or forest), G − X is also bipartite. Let A and B be its two color
classes (so A ∪ B = V \ X). We define the two parts PxA and PxB as follows: A subgraph P ∈ Px
is in PxA if the neighbor v of x in P is contained in A, otherwise P is in PxB .
We show that all subgraphs in PxA and PxB are pairwise vertex-disjoint. To this end, assume
toward a contradiction that two graphs P, Q ∈ PxA share some vertex. (The case P, Q ∈ PxB is
completely analogous.) Let p1 and q1 be the first vertex in P and Q respectively, that is, p1 and q1
are adjacent to x in G. Observe that p1 6= q1 . Let u 6= x be the first vertex that is in P and in Q.
By Lemma 2.12, P and Q are paths or trees with at most one vertex of degree more than two and
this vertex has degree three and is the neighbor of p1 or q1 , respectively. This implies together
with q1 , p1 ∈ A that either u = p1 or u = q1 . Assume without loss of generality that u = p1 .
Since p1 ∈ A and q1 ∈ A and u is a vertex in Q, it follows that Algorithm 3 followed u in the
graph exploration from q1 in Line 21. However, this is a contradiction since the algorithm checks
in Line 20 whether the new vertex u in the path is not adjacent to x. Thus, all subgraphs in PxA
and PxB are pairwise vertex-disjoint.
We next show that if Tab[x, f (Y )] = 6k 2 for some x ∈ X and Y ∈ X (recall that f maps Y
to a number, see Line 1), then there exist at least 3k internally vertex-disjoint augmenting paths
from x to Y .
Lemma 2.14. If in Line 17 of Algorithm 3 it holds for x ∈ X and Y ⊆ X that Tab[x, f (Y )] = 6k 2 ,
then there exist in G wrt. MG−X at least 3k alternating paths from x to vertices v1 , . . . , v3k2
such that all these paths are pairwise vertex-disjoint (except x) and N (vi ) ∩ X = N (w) ∩ X for
all i ∈ [3k 2 ].
Proof. Note that each time Tab[x, f (Y )] is increased by one (see Line 18), the algorithm found a
vertex w such that there is an alternating path P from x to w and N (w) ∩ X = Y . Furthermore,
since the function Keep-Edge returns true in this case, the edge from x to its neighbor on P is
not deleted in Line 10. Thus, there exist at least 6k 2 alternating paths from x to vertices whose
neighborhood in X is exactly Y . By Lemma 2.13, it follows that at least half of these 6k 2 paths
are vertex-disjoint.
The next lemma shows that Algorithm 3 is correct and runs in O(kn) time.
Lemma 2.15. Let (G = (V, E), s) be a matching instance, let X ⊆ V be a feedback vertex set of
size k with k < log n and at most k 2 (2k + 1) bottommost leaves in G − X, and let MG−X be a
maximum matching for G − X with at most k 2 free vertices in G − X that are all leaves. Then,
Algorithm 3 computes in O(kn) time an equivalent instance (G′ , s′ ) of size O(k 3 2k ).
Proof. We split the proof into three claims, one for the correctness of the algorithm, one for the
returned kernel size, and one for the running time.
Claim 2.16. The input instance (G, s) is a yes-instance if and only if the instance (G′ , s′ ) produced
by Algorithm 3 is a yes-instance.
Proof. Observe that the algorithm changes the input graph only in two lines: Lines 10 and 11.
By Lemma 2.1, applying Reduction Rules 2.1 and 2.3 yields an equivalent instance. Thus, it
remains to show that deleting the edges in Line 10 is correct, that is, it does not change the size
of a maximum matching. To this end, observe that deleting edges does not increase the size of
a maximum matching. Thus, we need to show that the size of the maximum matching does not
decrease. Assume toward a contradiction that it does.
17
Let {x, v} be the edge whose deletion decreased the maximum matching size. Redefine G to be
the graph before the deletion of {x, v} and G′ to be the graph after the deletion of {x, v}. Recall
that Algorithm 3 gets as additional input a maximum matching MG−X for G − X. Let MG :=
max
MG
(MG−X ) be a maximum matching for G with the largest possible overlap with M and
let GM := G(MG−X , MG ) = (V, MG−X △ MG ) (cf. ??). Since {x, v} ∈ MG \ MG−X and x is free
wrt. MG−X it follows that there is a path P in GM with one endpoint being x.
Recall that since P is a path in GM it follows that P is an augmenting path for MG−X . Since
all vertices in X are free wrt. MG−X , it follows that all vertices in P except the endpoints are
in V \ X. Let z be the second endpoint of this path P . We call a vertex on P an even (odd) vertex
if it has an even (odd) distance to x on P . (So x is an even vertex and v and z are odd vertices).
Observe that v is the only odd vertex in P adjacent to x since otherwise there would be another
augmenting path from x to z which only uses vertices from P implying the existence of another
maximum matching that does not use {x, v}, a contradiction. Let u be the neighbor of z in P .
Since no odd vertex on P except v is adjacent to x, it follows that the graph exploration in
the function Keep-Edge starting from x and v in Line 9 either reached u or returned true before.
If z ∈ V \ X, then in both cases, the function Keep-Edge would have returned true in Line 9 and
Algorithm 3 would not have deleted {x, v}, a contradiction. Thus, assume that z ∈ X. Therefore,
the function Keep-Edge considered the vertex u in Line 17 but did not keep the edge {x, v}.
Thus, when considering u, it holds that Tab[x, fX (u)] = 6k 2 , where fX (u) encodes Y := N (u) ∩ X
and z ∈ Y .
By Lemma 2.14, it follows that there are 3k 2 pairwise vertex-disjoint (except x) alternating
paths from x to vertices u1 , . . . , uk with N (ui ) ∩ X = Y . Thus, there are 3k 2 internally vertexdisjoint paths Q from x to y in G. If one of the paths Q ∈ Q does not intersect any path in GM , then
reverting the augmentation along P and augmenting along Q would results in another maximum
matching not containing {x, v}, a contradiction. Thus, assume that each path in Q intersects at
least one path in GM .
For each two paths Q1 , Q2 ∈ Q that are intersected by the same path P ′ in GM it holds
that each further path P ′′ in GM can intersect at most one of Q1 and Q2 : Assume toward a
contradiction that P ′′ does. Since no path in GM except P contains x and z it follows that all
intersections between the paths are within G − X. Since P ′ and P ′′ are vertex-disjoint and Q1
and Q2 are internally vertex-disjoint, it follows that there is a cycle in G − X, a contradiction to
the fact that X is a feedback vertex set.
Since 3k 2 > 3k + k 2 , it follows from the pigeon hole principle that there is a path P ′ ∈ GM
that intersects at least three paths Q1 , Q2 , Q3 ∈ Q such that no further path in GM intersects
them. We can now apply Lemma 2.11 and obtain two vertex-disjoint augmenting paths Q and Q′ .
Thus, reverting the augmentation along P and P ′′ and augment along Q and Q′′ yields another
maximum matching for G which does not contain {x, v}, a contradiction.
Claim 2.17. The graph G′ returned by Algorithm 3 has at most O(k 3 2k ) vertices and edges.
Proof. We first show that each vertex x ∈ X has degree at most O(k 2 2k ) in G′ . To this end,
we need to count the number of neighbors v ∈ N (x) \ X where the function Keep-Edge returns
true in Line 9. By Lemma 2.12, the function Keep-Edge explores the graph along one or two
paths (essentially growing from one starting point into two directions). Recall that Px denotes the
subgraphs induced by the graph exploration of Keep-Edge for the neighbors of x. By Lemma 2.13
there is a partition of Px into PxA and PxB such that within each part the subgraphs are pairwise
vertex-disjoint. We consider the two parts independently. We start with bounding the number
of graphs in PxA where the function ’Keep-Edge’ returned true (the analysis is completely analog
for PxB ).
Since all explored subgraphs are disjoint and all free vertices in G − X wrt. MG−X are leaves,
it follows that Algorithm 3 returned at most k 2 times true in Line 16 due to w being adjacent to
a free leaf in G − X. Also, the algorithm returns at most k 2 times true in Line 14 due to v being
free. Furthmore, the algorithm returns at most 6k 2 · 2k times true in Line 19. Finally, we show
that the algorithm returns at most k 2 · (2k − 1) times true in Lines 14 and 16, respectively. It
follows from the discussion below Observation 2.10 that T , the pendent-free tree of G − X, has at
18
most k 2 (2k + 1) leaves (denoted by VT1 ) and k 2 (2k + 1) vertices of degree at least three (denoted
by VT≥3 ). Let VT be the vertices of T . Since T is a tree (or forest), it has more vertices than edges
and hence
X
degT (v) < 2|VT |
v∈VT
which implies
X
v∈VT≥3
degT (v) < 2 · |VT≥3 | + |VT1 |.
Thus, Algorithm 3 returns at most 2 · |VT≥3 | + |VT1 | < 3k 2 (2k + 1) times true in Line 16 due to w
being a vertex in VT≥3 . Also, Algorithm 3 returns at most |VT≥3 | ≤ k 2 (2k + 1) times true in Line 14
due to v being a vertex in VT≥3 .
Summarizing, considering the graph explorations in PxA , Algorithm 3 returned at most
k 2 + k 2 + 6k 2 · 2k + 2k 2 (2k + 1) ∈ O(k 2 2k )
times true in the function Keep-Edge. Analogously, considering the graph explorations in PxA ,
Algorithm 3 also returned at most O(k 2 2k ) times true. Hence, each vertex x ∈ X has degree at
most O(k 2 2k ) in G′ .
We now show that the exhaustive application of Reduction Rules 2.1 and 2.3 indeed results in
a kernel of the claimed size. To this end, denote with VG1′ −X , VG2′ −X , and VG≥3
′ −X the vertices that
have degree one, two, and at least three in G′ − X. We have |VG1′ −X | ∈ O(k 3 2k ) since each vertex
in X has degree at most O(k 2 2k ) and G′ is reduced wrt. Reduction Rule 2.1. Next, since G′ − X
≥3
1
3 k
is a forest (or tree), we have |VG≥3
Finally, each
′ −X | < |VG′ −X | and thus |VG′ −X | ∈ O(k 2 ).
′
degree-two vertex in G needs at least one neighbor of degree at least three since G′ is reduced
with respect to Reduction Rule 2.3. Thus, each vertex in VG2′ −X is either incident to a vertex
in X or adjacent to one of the at most O(k 2 2k ) vertices in G′ − X that have degree at least three.
Thus, |VG2′ −X | ∈ O(k 3 2k ). Summarizing, G′ contains at most O(k 3 2k ) vertices and edges.
Claim 2.18. Algorithm 3 runs in O(kn) time.
Proof. First, observe that Lines 1 to 6 can be done in O(kn) time: The preprocessing and
table initialization can be done in O(kn) time as discussed in Section 2.2.3. Furthermore, T
and VT≥3 can clearly be computed in O(n + m) ≤ O(kn) time. Second, by Lemma 2.2, applying
Reduction Rules 2.1 and 2.3 can be done in O(n + m) time. Thus, it remains to show that each
iteration of the foreach-loop in Line 7 can be done in O(n) time.
By Lemma 2.13, the explored graphs Px from x can be partitioned into two parts such that
within each part all subgraphs are vertex-disjoint. Thus, each vertex in G − X is visited only twice
during the execution of the function Keep-Edge. Furthermore, observe that in Lines 17 and 18 the
table can be accessed in constant time. Thus, the function Keep-Edge only checks once whether a
vertex in V \ X has a neighbor in X, namely in Line 20. This single check can be done in constant
time. Since the rest of the computation is done on G − X which has less than |V \ X| edges, it
follows that each iteration of the foreach-loop in Line 7 can indeed be done in O(n) time.
This completes the proof of Lemma 2.15.
Our kernelization algorithm for the parameter “feedback vertex number” essentially calls
Steps 1 to 6.
Theorem 2.19. Matching parameterized by the feedback vertex number k admits a kernel of
size 2O(k) . It can be computed in O(kn) time.
Proof. First, using the linear-time factor-four approximation of Bar-Yehuda et al. [1], we compute
an approximate feedback vertex set X with |X| ≤ 4k. Then, we apply Steps 1 to 6. Applying the
first three steps is rather straightforward, see Section 2.2.1. For the remaining three steps, we use
Algorithms 1 to 3. By Lemmas 2.6, 2.8 and 2.15, this can be done in O(kn) time and results in a
kernel of size O((4k)3 24k ) = 2O(k) .
19
a1
a2
a3
a4
a5
a6
a7
A
b7
b6
b5
b4
b3
b2
b1
B
Figure 2: A chain graph. Note that the ordering of the vertices in A is going from left to right
while the ordering of the vertices in B is going from right to left. The reason for these two orderings
being drawn in different directions is that a maximum matching can be drawn as parallel edges,
see e. g. the bold edges. In fact, Algorithm 4 computes such matchings with the matched edges
being parallel to each other.
√
Applying the O(m n)-time algorithm for Matching [17] on the kernel yields:
Corollary 2.20. Matching can be solved in O(kn + 2O(k) ) time, where k is the feedback vertex
number.
3
Kernelization for Matching on Bipartite Graphs
In this section, we investigate the possibility of efficient and effective preprocessing for Bipartite
Matching. In particular, we show a linear-time computable polynomial-size kernel with respect
to the parameter “distance k to chain graphs”. In the first part of this section, we provide the
definition of chain graphs and describe how to compute the parameter. In the second part, we
discuss the kernelization algorithm.
Definition and computation of the parameter. We first define chain graphs which are
a subclass of bipartite graphs with special monotonicity properties.
Definition 1 ([5]). Let G = (A, B, E) be a bipartite graph. Then G is a chain graph if each of its
two color classes A, B admits a linear order w.r.t. neighborhood inclusion, that is, A = {a1 , . . . , aα }
and B = {b1 , . . . , bβ } where N (ai ) ⊆ N (aj ) and N (bi ) ⊆ N (bj ) whenever i < j.
Observe that if the graph G contains twins, then there is more than one linear order w.r.t.
neighborhood inclusion. To avoid ambiguities, we fix for the vertices of the color class A (resp. B)
in a chain graph G = (A, B, E) one linear order ≺A (resp. ≺B ) such that, for two vertices u, v ∈ A
(resp. u, v ∈ B), if u ≺A v (resp. if u ≺B v) then N (u) ⊆ N (v). In the remainder of the section
we consider a bipartite representation of a given chain graph G = (A, B, E) where the vertices of
A (resp. B) are ordered according to ≺A (resp. ≺B ) from left to right (resp. from right to left), as
illustrated in Figure 2. For simplicity of notation we use in the following ≺ to denote the orderings
≺A and ≺B whenever the color class is clear from the context.
We next show, that we can approximate the parameter and the corresponding vertex subset
in linear time. To this end, we use the following characterization of chain graphs.
Lemma 3.1 ([5]). A bipartite graph is a chain graph if and only if it does not contain an induced 2K2 .
Lemma 3.2. There is a linear-time factor-4 approximation for the problem of deleting a minimum
number of vertices in a bipartite graph in order to obtain a chain graph.
Proof. Let G = (A, B, E) be a bipartite graph. We compute a set S ⊆ A ∪ B such that G − S is a
chain graph and S is at most four times larger than a minimum size of such a set. The algorithm
iteratively tries to find a 2K2 and deletes the four corresponding vertices until no further 2K2 is
20
Algorithm 4: An algorithm that computes a maximum matching M in the chain graph G
such that all edges in M are parallel (see Fig. 2 for a visualization.)
Input: A chain graph G = (V, E), V = A ∪ B, A = {a1 , . . . , aα } and B = {b1 , . . . , bβ }
with N (ai ) ⊆ N (aj ) and N (bi ) ⊆ N (bj ) for i < j.
Output: A maximum matching of G where all matched edges are parallel.
1
2
3
Compute the size s of a maximum matching in G using an algorithm of Steiner and Yeomans [19]
M ← {{aα−s+1 , bβ }, {aα−s+2 , bβ−1 }, . . . , {aα , bβ−s+1 }}
return M .
found. Since in each 2K2 , by Lemma 3.1, at least one vertex needs to be removed, the algorithm
yields the claimed factor-4 approximation.
The details of the algorithm are as follows: First, it initializes S = ∅ and sorts the vertices in A
and in B by their degree; the vertices in A = {a1 , . . . , aα } in increasing order and the vertices B =
{b1 , . . . , bβ } in decreasing order, that is, deg(a1 ) ≤ . . . ≤ deg(aα ) and deg(b1 ) ≤ . . . ≤ deg(bβ ).
Since the degree of each vertex is at most max{α, β}, this can be done in linear time with e. g.
Bucket Sort. At any stage the algorithm deletes all vertices of degree zero and all vertices which are
adjacent to all vertices in the other partition. The deleted vertices are not added to S since these
vertices can not participate in a 2K2 . Next, the algorithm recursively processes the vertices in A
in a nondecreasing order of their degrees. Let a ∈ A be a minimum-degree vertex and let b ∈ B
be a neighbor of a. Since b is not adjacent to all vertices in A (otherwise b would be deleted),
there is a vertex a′ ∈ A that is not adjacent to b. Since deg(a) ≤ deg(a′ ) it follows that a′ has
a neighbor b′ that is not adjacent to a. Hence, the four vertices a, a′ , b, b′ induce only two edges:
{a, b} and {a′ , b′ } and thus form a 2K2 . Thus, the algorithm adds the four vertices to S, deletes
them from the graph, and continues with a vertex in A that has minimum degree.
As to the running time, we now show that, after the initial sorting, the algorithm considers
each edge only twice: Selecting a and b as described above can be done in O(1) time. To select a′ ,
the algorithm simply iterates over all vertices in A until it finds a vertex that is not adjacent to b.
In this way at most deg(b) + 1 vertices are considered. Similarly, by iterating over the neighbors
of a′ , one finds b′ . Hence, the edges incident to a, a′ , b, and b′ are used once to find the vertices
and a second time when these vertices are deleted. Thus, using appropriate data structures, the
algorithm runs in O(n + m) time.
Kernelization. In the rest of this section, we provide a linear-time computable kernel for
Bipartite Matching with respect to the parameter vertex deletion distance k to chain graphs.
The intuitive description of the kernelization is as follows. First we upper bound by O(k) the
number of neighbors of each vertex in the deletion set. Then we mark O(k 2 ) special vertices and
we use the monotonicity properties of chain graphs to upper bound the number of vertices that
lie between any two consecutive marked edges, thus bounding the total size of the reduced graph
to O(k 3 ) vertices.
Let G = (A, B, E) be the bipartite input graph, where V = A ∪ B, and let X ⊆ V be a vertex
subset such that G − X is a chain graph. By Lemma 3.2, we can compute an approximate X
in linear time. The kernelization algorithm is as follows: First, compute a specific maximum
matching MG−X ⊆ E in G − X with Algorithm 4 where all edges in MG−X are “parallel” and
all matched vertices are consecutive in the ordering ≺A and ≺B , see also Fig. 2. Since in convex
graphs matching is linear-time solvable [19] and convex graphs are a super class of chain graphs,
this can be done in O(n + m) time. We use MG−X in our kernelization algorithm to obtain some
local information about possible augmenting paths. For example, each augmenting path has at
least one endpoint in X. Forming this into a data reduction rule, with s denoting the size of a
maximum matching, yields the following.
Reduction Rule 3.1. If |MG−X | ≥ s, then return a trivial yes-instance; if s > |MG−X | + k,
then return a trivial no-instance.
21
The correctness of the above reduction rule follows from Observation 1.2.
Without loss of generality, we will be able to assume that each vertex in X is either matched
in M with another vertex in X or with a “small-degree vertex” in G − X. This means that an
augmenting path starting at some vertex in X will “enter” the chain graph G−X in a small-degree
V \X
vertex. We now formalize this concept. For a vertex x ∈ X we define Nsmall (x) to be the set of
the k neighbors of x in V \ X with the smallest degree, formally,
V \X
Nsmall (x) := {w ∈ N (x) \ X | k > |{u ∈ N (x) \ X | u ≺ w}|}.
Lemma 3.3. Let G = (V, E) be a bipartite graph and let X ⊆ V be a vertex set such that G − X
is a chain graph. Then, there exists a maximum matching MG for G such that every matched
V \X
vertex x ∈ X is matched to a vertex in Nsmall (x) ∪ X.
′
Proof. Assume, towards a contradiction, that there is no such matching MG . Let MG
be a
maximum matching for G that maximizes the number of vertices x ∈ X that are matched to a
V \X
V \X
′
′
vertex in Nsmall (x) ∪ X, that is, let MG
maximize |{x ∈ X | {u, x} ∈ MG
∧ u ∈ Nsmall (x) ∪ X}|.
V \X
Let x ∈ X be a vertex that is not matched with any vertex in Nsmall (x) ∪ X, that is, x is
V \X
V \X
matched to a vertex u ∈ V \ (Nsmall (x) ∪ X). If there is an unmatched vertex w ∈ Nsmall (x)
′
′′
′
in MG , then the matching MG := MG ∪ {{x, w}} \ {{u, x}} is a maximum matching with more
V \X
′
vertices x ∈ X (compared to MG
) that are matched to a vertex in Nsmall (x) ∪ X, a contradiction.
V \X
V \X
Hence, assume that there is no free vertex in Nsmall (x). Since |Nsmall (x)| = |X| = k, it follows
V \X
that at least one vertex w ∈ Nsmall (x) is matched to a vertex v ∈ V \ X. Observe that, by
V \X
definition of Nsmall (x), we have NG−X (w) ⊆ NG−X (u). Thus, we have {u, v} ∈ E and thus,
′′
′
MG := MG ∪ {{x, w}, {u, v}} \ {{u, x}, {w, v}} is a maximum matching with more vertices in X
′
(compared to MG
) fulfilling the condition of the lemma, a contradiction.
Based on Lemma 3.3, we can provide our next data-reduction rule.
Reduction Rule 3.2. Let (G, s) be an instance reduced with respect to Reduction Rule 3.1 and
V \X
let x ∈ X. Then delete all edges between x and V \ Nsmall (x).
Clearly, Reduction Rule 3.2 can be exhaustively applied in O(n + m) time by one iteration
over A and B in the order ≺.
The high-level idea of our kernel is as follows: Keep in the kernel all vertices of X. For each
V \X
vertex x ∈ X keep all vertices in Nsmall (x) and if a kept vertex is matched, then keep also the
vertex with whom it is matched. Denote with K the set of the vertices kept so far in V \ X.
Consider an augmenting path P = x, a1 , b1 , . . . , aℓ , bℓ , y from a vertex x ∈ B ∩X to a vertex y ∈
A ∩ X. Observe that if a1 ≺ aℓ , then also {b1 , aℓ } ∈ E and thus P ′ = x, a1 , b1 , aℓ , bℓ , y is an
augmenting path. Furthermore, the vertices in the augmenting path P ′ are a subset of K ∪ X
and, thus, by keeping these vertices (and the edges between them), we also keep the augmenting
path P ′ in our kernel. Hence, it remains to consider the more complicated case that aℓ ≺ a1 . To
this end, we next show that in certain “areas” of the chain graph G− X the number of augmenting
paths “passing through” such an area is bounded.
Definition 2. Let G = (A, B, E) be a chain graph and let M be a matching in G. Furthermore
let a ∈ A, b ∈ B with {a, b} ∈ M . Then #lmv(b, M ) (resp. #rmv(a, M )) is the number of
neighbors of b (resp. of a) that are to the left of a (resp. to the right of b); formally,
#lmv(b, M ) := |{a′ ∈ N (b) | a′ ≺ a}|,
#rmv(a, M ) := |{b′ ∈ N (a) | b′ ≺ b}|.
In Definition 2 the terms “left” and “right” refer to the ordering of the vertices of A and B in the
bipartite representation of G, as illustrated in Figure 2. The abbreviation #rmv (#lmv) stands for
“number of vertices right (left) of the matched vertex”. We set #lmv(a, M ) := #lmv(b, M ) and
#rmv(b, M ) := #rmv(a, M ). Finally, we define #rmv(a1 , a2 , M ) := mina1 ≺a′ ≺a2 {#rmv(a′ , M )}
for a1 , a2 ∈ A and #lmv(b1 , b2 , M ) := minb2 ≺b′ ≺b1 {#lmv(b′ , M )} for b1 , b2 ∈ B.
22
Lemma 3.4. Let G = (A, B, E) be a chain graph and M be a maximum matching for G computed
by Algorithm 4. Let a, b ∈ V with {a, b} ∈ M . Then the number of vertex-disjoint alternating paths
that (1) start and end with edges not in M and that (2) have endpoints left of a and right of b is
at most min{#lmv(b, M ), #rmv(a, M )}.
Proof. We prove the case #lmv(b, M ) ≤ #rmv(a, M ), that is, min{#lmv(b, M ), #rmv(a, M )} =
#lmv(b, M ). The case #lmv(b, M ) > #rmv(a, M ) follows by symmetry (with switched roles
of a and b). Let #aug denote the number of vertex-disjoint alternating paths from {a′ ∈ A |
a′ ≺ a} to {b′ ∈ B | b′ ≺ b} such that the first and last edge are not in M . Furthermore,
let ab1 , . . . , ab#lmv(b,M) be the neighbors of b that are to the left of a, that is, ab1 ≺ ab2 ≺ . . . ≺
ab#lmv(b,M) ≺ a. Since G is a chain graph it follows that no vertex a′ ∈ A with a′ ≺ ab1 is adjacent
to any vertex b′ ∈ B with b′ b. Furthermore, for any edge {a′ , b′ } ∈ E with a ≺ a′ and b ≺ b′
it follows from the construction of M (see Algorithm 4) that {a′ , b′ } ∈
/ M . Hence, any of these
alternating paths has to contain at least one vertex from ab1 , . . . , ab#lmv(b,M) . Since the alternating
paths are vertex-disjoint it follows that #aug ≤ #lmv(b, M ).
From the previous lemma, we directly obtain the following.
Lemma 3.5. Let G = (A, B, E) be a chain graph and let M be the maximum matching for G
computed by Algorithm 4. Let a1 , a2 ∈ A and b1 , b2 ∈ B with {a1 , b1 }, {a2 , b2 } ∈ M with a1 ≺ a2 .
Then there are at most #lmv(b1 , b2 , M ) vertex-disjoint alternating paths that (1) start and end
with edges not in M and that (2) have endpoints left of a1 and right of b2 .
Lemma 3.5 states that the number of augmenting paths passing through the area between a1
and a2 is bounded. Using this, we want to replace this area by a gadget with O(k) vertices. To
this end, we need further notation. For each kept vertex v ∈ K, we may also keep some vertices
to the right and to the left of v. We call these k vertices the left buffer (right buffer ) of v.
Definition 3. Let G = (A, B, E) be a chain graph and let M be the maximum matching for G
computed by Algorithm 4. Let a1 , a2 ∈ A and b1 , b2 ∈ B with {a1 , b1 }, {a2 , b2 } ∈ M and a1 ≺ a2 .
Then the (at most) #lmv(b1 , b2 , M ) vertices to the right of a1 form the right buffer B r (a1 , M )
of a1 ; formally,
B r (a1 , M ) := {a ∈ A | a1 ≺ a ∧ |{a′ ∈ A | a1 ≺ a′ ≺ a}| ≤ min{#lmv(b1 , b2 , M ), k}}.
Analogously,
B ℓ (a2 , M ) := {a ∈ A | a ≺ a2 ∧ |{a′ ∈ A | a ≺ a′ ≺ a2 }| ≤ min{#lmv(b1 , b2 , M ), k}},
B r (b1 , M ) := {b ∈ B | b ≺ b1 ∧ |{b′ ∈ B | b ≺ b′ ≺ b1 }| ≤ min{#lmv(b1 , b2 , M ), k}},
B ℓ (b2 , M ) := {b ∈ B | b2 ≺ b ∧ |{b′ ∈ B | b2 ≺ b′ ≺ b}| ≤ min{#lmv(b1 , b2 , M ), k}}.
Note that in Definition 3 each of the sets B r (a1 , M ), B ℓ (a2 , M ), B r (b1 , M ), and B ℓ (b2 , M )
depends on all four vertices a1 , a2 , b1 , b2 ; we omit these dependencies from the names for the sake
of presentation.
Reduction Rule 3.3. Let (G, s) be an instance reduced with respect to Reduction Rule 3.1.
Let a1 , a2 ∈ K ∩ A with a1 ≺ a2 and {a1 , b1 }, {a2 , b2 } ∈ MG−X such that A′ := {a ∈ A |
a1 ≺ a ≺ a2 } is of size at least 2 min{#lmv(b1 , b2 , M ), k} + 1 and A′ ∩ K = ∅. Then delete all
vertices in A′ \ (B r (a1 , MG−X ) ∪ B ℓ (a2 , MG−X )) and their matched neighbors in B, add all edges
between the vertices in the right buffer of a1 and the vertices in the left buffer of b2 , and decrease s
by the number of removed matched vertex pairs.
Lemma 3.6. Reduction Rule 3.3 is correct and can be exhaustively applied in O(n + m) time.
Proof. We first introduce some notation and provide some general observations. Let a1 , a2 , b1 ,
and b2 be as stated in Reduction Rule 3.3. Denote by A′ (resp. B ′ ) the set of vertices between a1
23
′
and a2 (resp. between b1 and b2 ). Further denote by A′D ⊆ A′ and BD
⊆ B ′ the sets of deleted ver′
′
′
′
tices. Note that |A | = |B | and |AD | = |BD | since MG−X was produced by Algorithm 4. Denote
x
the vertices in the buffers of a1 , a2 , b1 , and b2 by B x (yz , MG−X ) = {y1x , . . . , ymin{#lmv(b
}
1 ,b2 ,M),k}
for x ∈ {r, ℓ}, y ∈ {a, b}, z ∈ [2], and x = r ⇐⇒ z = 1.
Since the input instance is reduced with respect to Reduction Rule 3.1, it follows that s − k ≤
|MG−X | < s. Denote by MG′ −X := MG−X ∩ E ′ the matching obtained from MG−X by deleting
all edges not in the reduced graph G′ . Recall that s is reduced by the number of matched edges
that are removed. We next show in Claims 3.7 and 3.8 that the input instance (G = (V, E), s)
is a yes-instance if and only if the produced instance (G′ = (V ′ , E ′ ), s′ ) is a yes-instance. Before
we present these two claims, observe that there is a perfect matching between the vertices in A′D
′
and BD
, and thus
s − |MG−X | = s′ − |MG′ −X |.
(1)
Claim 3.7. If (G, s) is a yes-instance, then (G′ , s′ ) is a yes-instance.
max
max
Proof. Recall that MG
(MG−X ) for G minimizes the size of MG−X △ MG
(MG−X ).
max
Since (G, s) is a yes-instance it holds that |MG (MG−X )| ≥ s. For brevity we set GM :=
max
max
G(MG−X , MG
(MG−X )) = (V, MG−X △ MG
(MG−X )). Note that GM is a graph that only
contains odd-length paths. We will show that there are as many vertex-disjoint augmenting paths
for MG′ −X in G′ as there are paths in GM . This will show that G′ contains a matching of size
(1)
max
(MG−X )| − |MG−X | ≥ |MG′ −X | + s − |MG−X | = |MG′ −X | + s′ − |MG′ −X | = s′ .
|MG′ −X | + |MG
′
To this end, observe that all paths that do not use vertices in V \ V ′ = A′D ∪ BD
are also contained
′
M
′
in G . Thus, consider the paths in G that use vertices in V \ V . Denote by P M the set of all
paths in GM using vertices in V \ V ′ and set t := |P M |. Consider now an arbitrary i ∈ [t], and
let PiM ∈ P M . Denote by v1i , v2i , . . . , vpi i the vertices in PiM in the corresponding order, that is,
i
i
, v2j+1
} ∈ MG−X for all j ∈ [pi /2]. Observe
v1i and vpi i are the endpoints of PiM and we have {v2j
M
that exactly one endpoint of Pi is in A and the other endpoint is in B, since PiM is an odd-length
path. Assume without loss of generality that v1i ∈ A and vpi i ∈ B. Thus, the vertices in PiM with
odd (even) index are in A (B).
We next show that for any two vertices vji , vℓi of PiM with j < ℓ < pi − 1 and both being
in A \ X, it follows that vji ≺ vℓi . First, observe that if j = 1, then vji ∈ A \ X is a free vertex
wrt. MG−X . Since vℓi is matched wrt. MG−X and since MG−X is computed by Algorithm 4, it
follows that v1i ≺ vℓi . Thus, assume that j > 1 and ℓ > 1 (thus j ≥ 3 and j > 1). Assume toward
i
a contradiction that vℓi ≺ vji . Since ℓ < pi − 1, we have vℓ+1
∈ B \ X and since G − X is a chain
i
i
graph, it follows that {vj , vℓ+1 } ∈ E, a contradiction to Observation 1.1. Thus, vji ≺ vℓi .
We next show that the path PiM contains at least one vertex vji left of a1 and at least one
vertex vℓi right of b2 . Recall that MG−X was computed by Algorithm 4 and, thus, the free vertices
are the smallest wrt. the ordering ≺ (see also Fig. 2). Thus, if one endpoint of PiM is in (A∪B)\X,
then this vertex is either v1i and left of a1 or it is vpi 1 and right of b2 . Thus, assume that the
endpoints of PiM are in X. We showed in the previous paragraph that v3i ≺ v5i ≺ . . . ≺ vpi 1 −3 .
Thus, we also have vpi i −2 ≺ vpi i −4 ≺ . . . ≺ v2i since MG−X is computed by Algorithm 4. Since we
i
assumed that some vertices of PiM to be in V \ V ′ , it follows that for at least one vertex v2j+1
i
it holds that a1 ≺ v2j+1 ≺ a2 . Furthermore since by assumption no vertex between a1 and a2 or
between b1 and b2 is in K, it follows that v3i ≺ a1 (since v3i ∈ K) and vpi i −2 ≺ b2 (since vpi i −2 ∈ K).
M
For each i ∈ [t] denote by aPi the last vertex on the path PiM that is not right of a1 ,
M
that is, aPi is the vertex on PiM such that for each vertex a′ ∈ A \ X that is in PiM it holds
M
M
that a1 ≺ a′ or a′ aPi . It follows from the previous paragraph that aPi exists. Analogously
M
M
to aPi , for each i ∈ [t] denote by bPi the first vertex on the path PiM that is not left of b2 ,
M
that is, bPi is the vertex on PiM such that for each vertex b′ ∈ B that is in PiM it holds
M
that b2 ≺ b′ or b′ bPi . This means that in GM there is for each i ∈ [t] an alternating path
24
M
M
from aPi to bPi starting and ending with non-matched edges and all these paths are pairwise
vertex-disjoint. We show that also in G′ there are pairwise vertex-disjoint alternating paths
M
M
M
M
M
from aPi to bPi . Assume without loss of generality that aP1 ≺ aP2 ≺ . . . ≺ aPt . Since
M
M
in each path PiM , i ∈ [t], the successor of aPi is to the right of b1 , it follows that aPi has at
least i neighbors right of b1 . Since the right buffer of b1 contains the #lmv(b1 , b2 , MG−X ) ≥ t (see
M
M
Lemma 3.5) vertices to the right of b1 , we have {aPi , bri } ∈ E. By symmetry, we have {bPi , aℓi } ∈
E. Recall that MG−X forms a perfect matching between B r (b1 , MG−X ) and B r (a1 , MG−X ) as
well as between B ℓ (a2 , MG−X ) and B ℓ (b2 , MG−X ). Since Reduction Rule 3.3 added all edges
between B r (a1 , MG−X ) and B ℓ (b2 , MG−X ) to E ′ , it follows that each path PiM can be completed
M
M
as follows: aPi , bri , ari , bℓi , aℓi , bPi ; note that exactly the edges {bri , ari } and {bℓi , aℓi } are in MG−X .
Thus, each path in P M can be replaced by an augmenting path for MG′ −X in G′ and all these
augmenting paths are vertex-disjoint. Thus, there are as many augmenting paths for MG′ −x in G′
as there are paths in GM and therefore (G′ , s′ ) is a yes-instance.
Claim 3.8. If (G′ , s′ ) is a yes-instance, then (G, s) is a yes-instance.
Proof. Let MG′ a maximum matching for G′ . Observe that |MG′ | ≥ s′ . We construct a matching MG for G as follows. First, copy all edges from MG′ ∩ E into MG . Second, add all edges
′
′
from MG−X ∩ V \V
, that is, a perfect matching between A′D and BD
is added to MG . Observe
2
that if all edges in MG′ are also in E, then MG is a matching of size s in G. Thus, assume
that some edges in MG′ are not in E, that is, {ari1 , bℓj1 }, . . . , {arit , bℓjt } ∈ MG′ \ E for some t ∈
[#lmv(b1 , b2 , MG−X )]. Observe that s − |MG | ≤ t. Clearly, the vertices ari1 , . . . , arit , brj1 , . . . , brjt are
free with respect to MG .
We show that there are t pairwise vertex-disjoint augmenting paths from {ari1 , . . . , arit }
to {bℓj1 , . . . , bℓjt }; note, however, that these paths are not necessarily from arir to bℓjr , where r ∈ [t].
To this end, recall that by definition of #lmv(b1 , b2 , MG−X ) each vertex b ∈ B with b2 ≺ b ≺ b1
has at least #lmv(b1 , b2 , MG−X ) neighbors to the left of its matched neighbor. This allows us
to iteratively find augmenting paths as follows: To create the q th augmenting path Pq start with
some vertex bℓjq . Denote by v the last vertex added to Pq (in the beginning we have v = bℓjq ).
If v ∈ A, then add to Pq the neighbor matched to v. If v ∈ B, then do the following: if v is
adjacent to a vertex a ∈ {ari1 , . . . , arit }, then add a to Pq , otherwise add the leftmost neighbor
of v to Pq . Repeat this process until Pq contains a vertex from {ari1 , . . . , arit }. After we found Pq
remove all vertices of Pq from G. If q < t, then continue with Pq+1 . Observe that any two vertices
of Pq that are in A have at least #lmv(b1 , b2 , MG−X ) − 1 other vertices of A between them (in
the ordering of the vertices of A, see Figure 2). Thus, after a finite number of steps, Pq will
reach a vertex in {ari1 , . . . , arit }. Furthermore, it follows that after removing the vertices of Pq it
holds that #lmv(b1 , b2 , MG−X ) is decreased by exactly one: Pq contains for each vertex b ∈ B
at most one vertex among the #lmv(b1 , b2 , MG−X ) neighbors of b that are directly to the left of
its matched neighbor in MG−X . Thus, in each iteration we have #lmv(b1 , b2 , MG−X ) > 0. It
follows that the above procedure constructs t vertex-disjoint augmenting paths from {ari1 , . . . , arit }
to {bℓj1 , . . . , bℓjt }. Hence, G contains a matching of size s and thus (G, s) is a yes-instance.
The correctness of the data reduction rule follows from the previous two claims. It remains to
prove the running time. To this end, observe that the matching MG−X is given. Computing all
degrees of G can be done in O(n + m) time. Also #lmv(v, MG−X ) can be computed in linear time:
For each vertex b ∈ B one has to check for each neighbor of b whether it is to the left of b’s matched
neighbor and to adjust #lmv(b, MG−X ) accordingly. Furthermore,
#lmv(b1 , b2 , MG−X )
P computing P
′
and removing the vertices in A′D and BD
can be done in O( b∈B ′ deg(b) + a∈A′ deg(a)) time.
Thus, Reduction Rule 3.3 can be exhaustively applied in O(n + m) time.
We next upper-bound the number of free vertices with respect to MG−X . Let
Afree := {a ∈ A | a is free with respect to MG−X }
and
Akfree := {a ∈ Afree | |{a′ ∈ Afree | a a′ }| ≤ k},
25
that is, Akfree contains the k rightmost free vertices in A \ X. Observe that all vertices in Akfree
k
are left of MG−X . Analogously, denote by Bfree
the set containing the k leftmost free vertices
in B \ X.
Reduction Rule 3.4.
to Reduction Rule 3.1. Then
Let (G, s) be an instance
reduced with respect
k
k
delete all vertices in Afree \ (K ∪ Afree ) ∪ Bfree \ (K ∪ Bfree ) .
Lemma 3.9. Reduction Rule 3.4 is correct an can be applied in O(n + m) time.
Proof. The running time is clear. It remains to show the correctness. Let (G, s) be the input
instance reduced with respect to Reduction Rule 3.1 and let (G′ , s) be the instance produced by
Reduction Rule 3.4. We show that deleting the vertices in Afree \ (K ∪ Akfree ) yields an equivalent
k
instance. It then follows from symmetry that deleting the vertices in Bfree \ (K ∪ Bfree
) yields also
an equivalent instance.
We first show that if (G, s) is a yes-instance, then also the produced instance (G′ , s) is a yesinstance. Let (G, s) be a yes-instance and MG be a maximum matching for G. Clearly, |MG | ≥ s.
Observe that for each removed vertex a ∈ Afree \(K ∪Akfree) it holds that every vertex a′ ∈ Akfree is to
the right of a, that is, a ≺ a′ and thus NG−X (a) ⊆ NG−X (a′ ). Since (G, s) is reduced with respect
to Reduction Rule 3.1, it follows that |MG−X | ≥ |MG |−k. Thus, there exist at most k augmenting
paths for MG−X in G. If none of these augmenting paths ends in a vertex a ∈ Afree \ (K ∪ Akfree ),
then all augmenting paths exist also in G′ and thus (G′ , s) is a yes-instance. If one of these
augmenting paths, say P , ends in a, then at least one vertex a′ ∈ Akfree is not endpoint of any of
these augmenting paths. Since a ∈
/ K it follows from Lemma 3.3 that the neighbor b of a on P is
indeed in B \X. Since NG−X (a) ⊆ NG−X (a′ ), it follows that {a′ , b} ∈ E and thus we can replace a
by a′ in the augmenting path. By exhaustively applying the above exchange argument, it follows
that we can assume that none of the augmenting paths uses a vertex in Afree \ (K ∪ Akfree ). Thus,
all augmenting paths are also contained in G′ and hence the resulting instance (G′ , s) is still a
yes-instance.
Finally observe that if (G′ , s) is a yes-instance, then also (G, s) is a yes-instance, since G′ is
a subgraph of G. Thus any matching of size s in G′ is also a matching in G. Hence (G, s) is a
yes-instance.
Theorem 3.10. Matching on bipartite graphs admits a cubic-vertex kernel with respect to the
vertex deletion distance to chain graphs. The kernel can be computed in linear time.
Proof. Let (G, s) be the input instance with G = (V, E), the two partitions V = A ∪ B, and X ⊆
V such that G − X is a chain graph. If X is not given explicitly, then use the linear-time
factor-four approximation provided in Lemma 3.2 to compute X. The kernelization is as follows:
First, compute MG−X in linear time with Algorithm 4. Next compute the set K. Then, apply
Reduction Rules 3.1, 3.3 and 3.4. By Lemmas 3.6 and 3.9, this can be done in linear time. Let bK
ℓ
K
K
the leftmost vertex in K ∩ B and aK
r rightmost vertex in A ∩ K. Let aℓ and br be their matched
neighbors. Since |K| ≤ k 2 and we reduced the instance with respect to Reduction Rule 3.3, it
K
K
follows that the number of vertices between aK
ℓ and ar as well as the number of vertices between bℓ
K
3
and br is at most 2k , respectively. Furthermore, there are at most 2k free vertices left in V \ X
since we reduced the instance with respect to Reduction Rule 3.4. It remains to uper-bound the
K
number of matched vertices left of bK
ℓ and right of ar .
K
Observe that all vertices left of bℓ are matched with respect to MG−X . If there are more
a
b
than 2k vertices to the left of bK
ℓ , then do the following: Add four vertices aℓ , bℓ , xℓ , xℓ to V . The
idea is that {aℓ , bℓ } should be an edge in MG−X such that aℓ ∈ A and bℓ ∈ B are in K and there is
no vertex left of bℓ . This means we add these vertices to simulate the situation where the leftmost
vertex in B \ X is also in K. To ensure that aℓ and bℓ are in K and that they are not matched with
some vertices in G, we add xaℓ and xbℓ to X and make xaℓ respectively xbℓ to their sole neighbors.
In this way, we ensure that there is maximum matching in the new graph that is exactly two
edges larger than the maximum matching in the old graph. In this new graph we can then apply
Reduction Rule 3.3 to reduce the number of vertices between bℓ and bK
ℓ . Formally, we add the
26
following edges. Add {aℓ , xaℓ }, {bℓ , xbℓ } to E. Add all edges between bℓ and the vertices in B \ X.
Let a be the rightmost vertex in Akfree . Then, add edges between aℓ and NG−X (a). Set b′ ≺ bℓ
for each b′ ∈ B \ X, set a ≺ aℓ for each vertex a ∈ Afree , and set aℓ ≺ a′ for each matched vertex
in A \ X. Furthermore, add {aℓ , bℓ } to MG−X and add aℓ and bℓ to K. Finally, increase s by two.
Next, apply Reduction Rule 3.3 in linear time, then remove aℓ , bℓ , xaℓ , xbℓ and reduce s by two. After
this procedure, it follows that there are at most 2k vertices left of bK
ℓ . If there are more than 2k
vertices right of the rightmost vertex aK
in
A∩K,
then
use
the
same
procedure
as above. Thus, the
r
total number of vertices in the remaining graph is at most |X| + 2k + 4k 3 = O(k 3 ). Furthermore,
observe that adding and removing the four vertices as well as applying Reduction Rule 3.3 can be
done in linear time. Thus, the overall running time of the kernelization is O(n + m).
Applying the O(n2.5 )-time algorithm for Bipartite Matching [15] on the kernel yields:
Corollary 3.11. Matching can be solved in O(k 7.5 + n + m) time, where k is the vertex deletion
distance to chain graphs.
4
Conclusion
We focussed on kernelization results for Matching. There remain numerous challenges for future
research as discussed in the end of this concluding section. First, however, let us discuss the
closely connected issue of FPTP algorithms for Matching. There is a generic augmenting pathbased approach to provide FPTP algorithms for Matching: To begin with, note that one can
find an augmenting path in linear time [2, 11, 17]. Now the solving algorithm for Matching
parameterized by some vertex deletion distance k works as follows:
1. Use a constant-factor linear-time (approximation) algorithm to compute a vertex set X such
that G − X is a “trivial” graph (where Matching is linear-time solvable).
2. Compute in linear time an initial maximum matching M in G − X.
3. Start with M as a matching in G and increase the size at most |X| = k times to obtain
in O(k · (n + m)) time a maximum matching for G.
From this we can directly derive that Matching can be solved in O(k(n + m)) time, where k
is one of the following parameters: feedback vertex number, feedback edge number, vertex cover
number. Bipartite Matching can be solved in O(k(n + m)) time, where k is the vertex deletion
distance to chain graphs. Using our kernelization results, the multiplicative dependence of the
running time on parameter k can now be made an additive one. For instance, in this way the
running time for Bipartite Matching parameterized by vertex deletion distance to chain graphs
“improves” from O(k(n + m)) to O(k 7.5 + n + m).
We conclude with listing some questions and tasks for future research. Can the size or the
running time of the kernel with respect to feedback vertex set (see Section 2) be improved? Is there
a linear time computable kernel for Matching parameterized by the treewidth t (assuming that
t is given)? This would complement the recent randomized O(t4 n log n) time algorithm [9]. Can
one extend the kernel of Section 3 from Bipartite Matching to Matching parameterized by
the distance to chain graphs? These are only three very concrete questions. There are numerous
more, including further parameterizations or any form of kernel lower bound results. Finally, will
Matching become the “drosophila” of FPTP studies akin to Vertex Cover for classical FPT
studies was? We hope we gave some reasons for believing that.
References
[1] R. Bar-Yehuda, D. Geiger, J. Naor, and R. M. Roth. Approximation algorithms for the feedback vertex set problem with applications to constraint satisfaction and Bayesian inference.
SIAM Journal on Computing, 27(4):942–959, 1998.
27
[2] N. Blum. A new approach to maximum matching in general graphs. In Proceedings of the 17th
International Colloquium on Automata, Languages, and Programming (ICALP ’90), volume
443 of LNCS, pages 586–597. Springer, 1990.
[3] H. L. Bodlaender, B. M. P. Jansen, and S. Kratsch. Preprocessing for treewidth: A combinatorial analysis through kernelization. SIAM Journal on Discrete Mathematics, 27(4):
2108–2142, 2013.
[4] H. L. Bodlaender, B. M. P. Jansen, and S. Kratsch. Kernelization lower bounds by crosscomposition. SIAM Journal on Discrete Mathematics, 28(1):277–305, 2014.
[5] A. Brandstädt, V. B. Le, and J. P. Spinrad. Graph Classes: a Survey, volume 3 of SIAM
Monographs on Discrete Mathematics and Applications. SIAM, 1999.
[6] M. Chang. Algorithms for maximum matching and minimum fill-in on chordal bipartite
graphs. In Proceedings of the 7th International Symposium on Algorithms and Computation
(ISAAC ’96), volume 1178 of LNCS, pages 146–155. Springer, 1996.
[7] E. Dahlhaus and M. Karpinski. Matching and multidimensional matching in chordal and
strongly chordal graphs. Discrete Applied Mathematics, 84(1–3):79–91, 1998.
[8] R. Duan and S. Pettie. Linear-time approximation for maximum weight matching. Journal
of the ACM, 61(1):1:1–1:23, 2014.
[9] F. V. Fomin, D. Lokshtanov, M. Pilipczuk, S. Saurabh, and M. Wrochna. Fully polynomialtime parameterized computations for graphs and matrices of low treewidth. In Proceedings of
the 28th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA ’17), pages 1419–
1432. SIAM, 2017.
[10] H. N. Gabow and R. E. Tarjan. A linear-time algorithm for a special case of disjoint set
union. Journal of Computer and System Sciences, 30(2):209–221, 1985.
[11] H. N. Gabow and R. E. Tarjan. Faster scaling algorithms for general graph-matching problems. Journal of the ACM, 38(4):815–853, 1991.
[12] A. C. Giannopoulou, G. B. Mertzios, and R. Niedermeier. Polynomial fixed-parameter algorithms: A case study for longest path on interval graphs. In Proceedings of the 10th
International Symposium on Parameterized and Exact Computation (IPEC ’15), volume 43
of LIPIcs, pages 102–113. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2015.
[13] J. Guo, F. Hüffner, and R. Niedermeier. A structural view on parameterizing problems:
Distance from triviality. In Proceedings of the 1st International Workshop on Parameterized
and Exact Computation (IWPEC ’04), volume 3162 of LNCS, pages 162–173. Springer, 2004.
[14] M. Gupta and R. Peng. Fully dynamic (1+ e)-approximate matchings. In Proceedings of
the 54th Annual IEEE Symposium on Foundations of Computer Science (FOCS ’13), pages
548–557. IEEE Computer Society, 2013.
[15] J. E. Hopcroft and R. M. Karp. An n5/2 algorithm for maximum matchings in bipartite
graphs. SIAM Journal on Computing, 2(4):225–231, 1973.
[16] R. M. Karp and M. Sipser. Maximum matchings in sparse random graphs. In Proceedings of
the 22nd Annual IEEE Symposium on Foundations of Computer Science (FOCS ’81), pages
364–375. IEEE Computer Society, 1981.
p
[17] S. Micali and V. V. Vazirani. An O( |V ||E|) algorithm for finding maximum matching
in general graphs. In Proceedings of the 21st Annual IEEE Symposium on Foundations of
Computer Science (FOCS ’80), pages 17–27. IEEE, 1980.
[18] S. S. Skiena. The Algorithm Design Manual. Springer, 2010.
[19] G. Steiner and J. S. Yeomans. A linear time algorithm for maximum matchings in convex
bipartite graphs. Comput. Math. Appl., 31:91–96, 1996.
[20] R. Yuster. Maximum matching in regular and almost regular graphs. Algorithmica, 66(1):
87–92, 2013.
28
| 8cs.DS
|
Provable quantum state tomography via non-convex methods
Anastasios Kyrillidis,1, ∗ Amir Kalev,2, † Dohuyng Park,3, ‡ Srinadh
Bhojanapalli,4, § Constantine Caramanis,5, ¶ and Sujay Sanghavi5, ∗∗
1
arXiv:1711.02524v3 [quant-ph] 19 Nov 2017
IBM T. J. Watson Research Center
2
University of Maryland
3
Facebook
4
Toyota Technological Institute at Chicago
5
University of Texas at Austin
With nowadays steadily growing quantum processors, it is required to develop new quantum
tomography tools that are tailored for high-dimensional systems. In this work, we describe such
a computational tool, based on recent ideas from non-convex optimization. The algorithm excels
in the compressed-sensing-like setting, where only a few data points are measured from a lowrank or highly-pure quantum state of a high-dimensional system. We show that the algorithm can
practically be used in quantum tomography problems that are beyond the reach of convex solvers,
and, moreover, is faster than other state-of-the-art non-convex approaches. Crucially, we prove that,
despite being a non-convex program, under mild conditions, the algorithm is guaranteed to converge
to the global minimum of the problem; thus, it constitutes a provable quantum state tomography
protocol.
I.
INTRODUCTION
Like any other processor, the behavior of a quantum information processor must be characterized, verified, and
certified. Quantum state tomography (QST) is one of
the main tools for that purpose [1]. Yet, it is generally
an inefficient procedure, since the number of parameters
that specify quantum states, grows exponentially with
the number of sub-systems. This inefficiency has two
practical manifestations: (i) without any prior information, a vast number of data points needs to be collected
[1]; (ii) once the data is gathered, a numerical procedure
should be executed on an exponentially-high dimensional
space, in order to infer the quantum state that is most
consistent with the observations. Thus, to perform QST
on nowadays steadily growing quantum processors [2, 3],
we must introduce novel, more efficient, techniques for
its completion.
Since often the aim in quantum information processing
is to coherently manipulate pure quantum states (i.e.,
states that can be equivalently described with rank-1,
positive semi-definite (PSD) density matrices), the use
of such prior information is the modus operandi towards
making QST manageable, with respect to the amount of
data required [4–7]. Compressed sensing (CS) [8, 9] –and
its extension to guaranteed low-rank approximation [10–
12]– has been applied to QST [5, 13] within this context.
In particular, it has been proven [5, 13–15] that convex programming guarantees robust estimation of pure
∗
†
‡
§
¶
∗∗
Author to whom correspondence should be addressed; [email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
n-qubit states from much less information than common
wisdom dictates, with overwhelming probability.
These advances, however, leave open the question
of how efficiently one can estimate exponentially largesized quantum states, from a limited set of observations.
Since convex programming is susceptible of provable performance, typical QST protocols rely on convex programs [5, 7, 13]. Nevertheless, their Achilles’ heel remains
the high computational and storage complexity. In particular, due to the PSD nature of density matrices, a
key step is the repetitive application of Hermitian eigenproblem solvers. Such solvers include the well-established
family of Lanczos methods [16–19], the Jacobi-Davinson
SVD type of methods [20], as well as preconditioned hybrid schemes [21], among others; see also the recent article in [22] for a more complete overview. Since –at least
once per iteration– a full eigenvalue decomposition is required in such convex programs, these eigensolvers contribute a O((2n )3 ) computational complexity, where n is
the number of qubits of the quantum processor. It is obvious that the recurrent application of such eigensolvers
makes convex programs impractical, even for quantum
systems with a relatively small number n of qubits [5, 23].
Ergo, to improve the efficiency of QST, we need to
complement it with numerical algorithms that can efficiently handle large search spaces using limited amount
of data, while having rigorous performance guarantees.
This is the purpose of this work. Inspired by the recent
advances on finding the global minimum in non-convex
problems [24–38], we propose the application of alternating gradient descent in QST, that operates directly on
the assumed low-rank structure of the density matrix.
The algorithm –named Projected Factored Gradient Decent (ProjFGD) and described below in detail– is based
on the recently analyzed non-convex method in [29] for
PSD matrix factorization problems. The added twist is
the inclusion of further constraints in the optimization
program, that makes it applicable for tasks such as QST.
2
In general, finding the global minimum in non-convex
problems is a hard problem. However, our approach
assumes certain regularity conditions –that are, however, satisfied by common CS-inspired protocols in practice [5, 7, 13]– and a good initialization –which we make
explicit in the text; both lead to a fast and provable estimation of the state of the system, even with limited
amount of data. Our numerical experiments show that
our scheme outperforms in practice state-of-the-art approaches for QST.
Apart from the QST application, our aim is to broaden
the results on efficient, non-convex recovery within the
set of constrained low-rank matrix problems. Our developments maintain a connection with analogous results
in convex optimization, where standard assumptions are
made. However, this work goes beyond convexity, in an
attempt to justify recent findings that non-convex methods show significant acceleration, as compared to stateof-the-art convex analogs.
δr < 1; i.e.,
(1 − δr )kρk2F ≤ kM(ρ)k22 ≤ (1 + δr )kρk2F ,
where
k · kF denote the Frobenius norm, is satisfied ∀ρ ∈
n
n
C2 ×2 such that rank(ρ) ≤ r.
An accurate estimation of ρ? is obtained by solving,
essentially, a convex optimization problem constrained
to the set of quantum states [13], consistent with the
measured data. Two such convex program examples are:
minimize
n
n
ρ∈C2
For the above setting, we assume that the data is
given in the form of expectation values of n-qubit Pauli
observables. An n-qubit Pauli observable is given by
P = ⊗nj=1 sj where sj ∈ {1, σx , σy , σz }. There are
4n such observables in total. In general, one needs to
have the expectation values of all 4n Pauli observables to
uniquely reconstruct ρ? ; i.e., performing for many repetitions/shots the same experiment and then taking the
expectation of the results (e.g., counts of qubit registers).
Since ρ? is a highly-pure density matrix, we apply the
CS result on Pauli measurements [5, 15], that guarantees robust estimation, with high probability, from just
m = O(r2n n6 ) randomly chosen Pauli observables (in expectation). Key property to achieve this is the restricted
isometry property [15]:
Definition 1 (Restricted Isometry
Property
(RIP) for
n
n
Pauli measurements). Let M : C2 ×2 → Rm be a linear
n
map, such that (M(ρ))i = √2m Tr(Pi ·ρ), for i = 1, . . . , m.
Then, with high probability over the choice of m = δc2 ·
r
(r2n n6 ) Pauli observables Pi , where c > 0 is an absolute
constant, M satisfies the r-RIP with constant δr , 0 ≤
(1)
and,
minimize
n
n
×2
1
2
· ky − M(ρ)k22
subject to ρ 0,
Tr(ρ) ≤ 1,
QUANTUM STATE TOMOGRAPHY SETUP
We begin by describing the problem of QST. We are
focusing here on QST of a highly-pure n-qubit state from
Pauli measurements. In particular, let yn∈ Rm be the
measurement vector with elements yi = √2m Tr(Pi · ρ? ) +
ei , i = 1, . . . , m, for some measurement error ei . Here, ρ?
denotes the unknown n-qubit densitynmatrix,
associated
n
with the pure quantum state; Pi ∈ C2 ×2 is a randomly
n
chosen Pauli observable; and the normalization √2m is
chosen nto follow
the results of [15]. For brevity, we denote
n
m
M : C2 ×2 →
R
as the linear “sensing” map, such that
n
2
(M(ρ))i = √m Tr(Pi · ρ), for i = 1, . . . , m.
Tr(ρ)
subject to ρ 0,
ky − M(ρ)k2 ≤ ,
ρ∈C2
II.
×2
(2)
where ρ 0 captures the positive semi-definite assumption, k · k2 is the vector Euclidean `2 -norm, and > 0 is
a parameter related to the error level in the model. Key
in both programs is the combination of the PSD and the
trace constraints: combined, they constitute the tightest
convex relaxation to the low-rank, PSD structure of the
unknown ρ? ; see also [10].
As was discussed in the introduction, the problem with
convex programs, such as (1) and (2), is their inefficiency
when applied in high-dimensional systems: most practical solvers for (1)-(2) are iterative and handling PSD
constraints adds an immense complexity overhead per iteration, especially when n is large; see also Section V.
In this work, we propose to use non-convex programming for QST of low-rank density matrices, which leads
to higher efficiency than typical convex programs. We
achieve this by restricting the optimization over the intrinsic non-convex structure of rank-r PSD matrices.
This allow us to “describe” an 2n × 2n PSD matrix
with
only O(2n r) space, as opposed to the O (2n )2 ambient space. Even more substantially, our program has
theoretical guarantees of global convergence, similar to
the guarantees of convex programming, while maintaining faster preformace than the latter. These properties
make our scheme ideal to complement the CS methodology for QST in practice.
III.
PROJECTED FACTORED GRADIENT
DECENT ALGORITHM
Optimization criterion recast: At its basis, the Projected Factored Gradient Decent (ProjFGD) algorithm
transforms convex programs, such as in (1)-(2), by imposing the factorization of a d × d PSD matrix ρ such
3
that ρ = AA† . This factorization, popularized by Burer and Monteiro [39, 40] for solving semi-definite convex programming instances, naturally encodes the PSD
constraint, removing the expensive eigen-decomposition
projection step. For concreteness, we focus here on the
convex program (2), where d = 2n . In order to encode the
trace constraint, ProjFGD enforces additional constraints
on A. In particular, the requirement that Tr(ρ) ≤ 1 is
translated to the convex constraint kAk2F ≤ 1, where
k · kF is the Frobenius norm. The above recast the program (2) as a non-convex program:
minimize
A∈Cd×r
1
2
· ky − M(AA† )k22
subject to kAk2F ≤ 1.
(3)
Observe that, while the constraint set is convex, the
objective is no longer convex due to the bilinear transformation of the parameter space ρ = AA† . Such criteria have been studied recently in machine learning and
signal processing applications [24–38]. Here, the added
twist is the inclusion of further matrix norm constraints,
that makes it proper for tasks such as QST; as we show
in Appendices A and B, such addition complicates the
algorithmic analysis.
The prior knowledge that rank(ρ? ) ≤ r? is imposed in
the program by setting A ∈ Cd×r? . In real experiments,
the state of the system, ρ? , could be full rank, but often
is highly-pure with only few dominant eigenvalues. In
this case, ρ? is well-approximated by a low-rank matrix
of rank r, which can be much smaller than r? , similar to
the CS methodology. Therefore in the ProjFGD protocol
we set A ∈ Cd×r . In this form, A contains much less
variables to maintain and optimize than a d × d PSD
matrix, and thus it is easier to update and to store its
iterates.
An important issue in optimizing (3) over the factored
space is the existence of non-unique possible factorizations for a given ρ. To see this, if ρ = AA† , then for any
unitary matrix R ∈ Cr×r such that RR† = I, we have
bA
b† , where A
b = AR. Since we are interested in obρ=A
taining a low-rank solution in the original space, we need
a notion of distance to the low-rank solution ρ? over the
factors. We use the following unitary-invariant distance
metric:
Definition 2. Let matrices A, A? ∈ Cd×r . Define:
Dist (A, A? ) := min kA − A? RkF ,
R:R∈U
where U is the set of r × r unitary matrices.
The ProjFGD algorithm: At heart, ProjFGD is a projected gradient descent algorithm over the variable A.
The pseudocode is provided in Algorithm 1.
Algorithm 1 ProjFGD pseudocode for (3)
1: Input: Function f , target rank r, # iterations T .
†
2: Output: ρ = AT AT .
3:
4:
5:
6:
7:
8:
Initialize ρ0 randomly or set ρ0 := 2/Lb · ΠC 0 M† (y) .
Set A0 ∈ Cd×r such that ρ0 = A0 A†0 .
Set step size η as in (8).
for t = 0 to T− 1 do
At+1 = ΠC At − η∇f (At A†t ) · At .
end
First, some properties of the objective in (3). Denote
g(A) = 12 · ky − M(AA† )k22 and f (ρ) = 12 · ky − M(ρ)k22 .
Due to the symmetry of f , i.e., f (ρ) = f (ρ† ), the gradient
of g(A) with respect to A variable is given by
∇g(A) = ∇f (ρ) + ∇f (ρ)† · A = 2∇f (ρ) · A,
where ∇f (ρ) = −2M† (y − M(ρ)), and M† is the adjoint operator for M. For the Pauli measurements case
we consider in this paper, the adjoint
operator for an
n Pm
input vector b ∈ Rm is M† (b) = √2m i=1 bi Pi .
d×r
Let ΠC (B) denote
the projection a matrix B ∈ C
onto the set C = A : A ∈ Cd×r , kAk2F ≤ 1 , as in (3).
For this particular C, ΠC (B) = ξ(B) · B, where ξ(·) ∈
(0, 1]; in the case where ξ(·) = 1, B ∈ C already. As
an initialization, we compute ΠC 0 (·), which denotes the
projection onto the set of PSD matrices with trace bound
Tr(ρ) ≤ 1; we discuss later in the text how to complete
this step in practice.
The main iteration of ProjFGD is in line 7 of Algorithm
1, where it applies a simple update rule over the factors:
At+1 = ΠC At − η∇f (At A†t ) · At ,
Observe that the input argument in ΠC (·) is:
At − η∇f (At A†t ) · At = At − η∇g(At );
i.e., it performs gradient descent over A variable, with
step size η. Any constants are “absorbed” in the step
size selection, for clarity.
Two vital components of our algorithm are: (i) the
initialization step and, (ii) step size selection.
A.
Initialization ρ0
Due to the bilinear structure in (3), at first glance it is
not clear whether the factorization ρ = AA† introduces
spurious local minima, i.e., local minima that do not
exist in (1)-(2), but are “created” after the substitution
ρ = AA† . This necessitates careful initialization, in order
to obtain the global minimum.
Before describing our initialization procedure, we find
it helpful to first discuss an initialization procedure for
an altered version of (3) where trace constraints are excluded. In this case, (3) transforms to:
minimize
A∈Cd×r
1
2
· ky − M(AA† )k22 .
(4)
4
Under this setting, the following theory stems from [30]:
Theorem 3. Suppose the unknown ρ? is a rank-r density
matrix with a non-unique factorization ρ? = A? A†? , for
A? ∈ Cd×r . Under the noiseless model, the observations
satisfy y = M(ρ? ). Assuming that the linear map M
satisfies the restricted isometry property in Definition 1,
with constant δ4r / 0.0363, any critical point A satisfying
first- and second-order optimality conditions is a global
minimum.
Corollary 4. Suppose the unknown ρ? is a full rank density matrix and let ρ?,r denote its best rank-r approximation, in the Eckart-Young-Minsky-Steward sense [41, 42].
Let ρ?,r have a non-unique factorization ρ?,r = A? A†? , for
A? ∈ Cd×r . Under the noiseless model, the observations
satisfy y = M(ρ? ). Assuming that the linear map M
satisfies the restricted isometry property in Definition 1,
with constant δ4r ≤ 1/200, any critical point A satisfying
first- and second-order optimality conditions satisfy:
Dist(A, A? ) ≤
1250
3σr (ρ? )
· kM (ρ? − ρ?,r )k2 .
In plain words, under the noiseless model and with
high probability (that depends on the random structure
of the sensing map M), Theorem 3 states that the nonconvex change of variables ρ = AA† does not introduce
any spurious local minima in the low-rank ρ? case, and
random initialization is sufficient for Algorithm 1 to find
the global minimum, assuming a proper step size selection. Further, when ρ? is not low-rank, there are low-rank
solutions ρ?,r , close to ρ? ; how much close is a function
of the spectrum of ρ? and its best rank-r approximation
residual (Corollary 4). In these cases, Step 3 in Algorithm boils down to random initialization.
The above cases hold when M corresponds to a
POVM; then the explicit trace constraint is redundant.
In the more general case where the trace constraint is
present, a different approach is followed. In that case,
the initial point ρ 0 is set as ρ0 := 1/Lb · ΠC 0 (−∇f (0)) =
2/L
b · ΠC 0 M† (y) , where ΠC 0 (·) denotes the projection
onto the set of PSD matrices ρ that satisfy Tr(ρ) ≤ 1.
b represents an approximation of L, where L is
Here, L
such that for all rank-r matrices ρ, ζ:
k∇f (ρ) − ∇f (ζ)kF ≤ L · kρ − ζkF .
(5)
(This also means that f is restricted gradient Lipschitz
continuous with parameter L. We defer the reader to the
Appendix A for more information). In practice, we set
b ∈ (1, 2).
L
This is the only place in the algorithm where
eigenvalue-type calculation is required. The projection
ΠC 0 (·) is given in [43]. Given M† (y), it is described with
the following criterion:
minimize
ρ0 ∈Cd×d
1
2
· kρ0 − M† (y) k2F
subject to ρ0 0,
Tr(ρ0 ) ≤ 1,
(6)
To solve this problem, we first compute its eigendecomposition M† (y) := ΦΛΦ† , where Φ is a unitary matrix containing the eigenvectors of the input matrix. Due to the fact that the Frobenius norm is invariant under unitary transformations, [43] proves that
b > , where Λ
b is a diagonal matrix, computed
ρ0 = ΦΛΦ
via:
minimize
b
Λ
subject to
b − Λk2F
· kΛ
X
b ii ≤ 1,
Λ
1
2
(7)
i
b 0.
Λ
The last part can be easily solved using the projection
onto the unit simplex [44–46].
Alternatively, in practice, we could just use a standard projection onto the set of PSD matrices ρ0 :=
2/L
b · Π+ M† (y) ; our experiments show that it is sufficient and can be implemented by any off-the-shelf eigenvalue solver. In that case, the algorithm generates an initial matrix A0 ∈ Cd×r by truncating the computed eigendecomposition, followed by a projection onto the convex set, C, defined by set of constraints in the program,
e0 ). In our case C = {A ∈ Cd×r : kAk2 ≤ 1}.
A0 = ΠC (A
F
Note again that the projection operation is a simple
entry-wise scaling, for A ∈
/ C, ΠC (A) = ξ(A) · A, where
ξ(A) = kAk−1
.
F
Apart from the procedure mentioned above, we could
also use more specialized spectral methods for initialization [26, 47] or, alternatively, run convex algorithms, such
as (2) for only a few iterations. However, this choice often
leads to an excessive number of full or truncated eigenvalue decompositions [28], which constitutes it a nonpractical approach.
The discussion regarding the step size and what type
of guarantees we obtain is discussed next.
B.
Step size selection and theoretical guarantees
Focusing on (3), we provide theoretical guarantees for
ProjFGD. Our theory dictates a specific constant step size
selection that guarantees convergence to the global minimum, assuming a satisfactory initial point ρ0 is provided.
Let us first describe the local convergence rate guarantees of ProjFGD.
Theorem 5 (Local convergence rate for QST). Let ρ?
be a rank-r quantum state density matrix of an n-qubit
system with
a non-unique factorization ρ? = A? A†? , for
2n ×r
A? ∈ C
. Let y ∈ Rm be the measurement vector of m = O(rn6 2n ) random n-qubit Pauli observables,
and M be the corresponding sensing map, such that
yi = (M(ρ? ))i + ei , ∀i = 1, . . . , m. Let the step η in
ProjFGD satisfy:
η≤
1
b 1 (ρ0 )+σ1 (∇f (ρ0 ))) ,
128(Lσ
(8)
5
where σ1 (ρ) denotes the leading singular value of ρ. Here,
b ∈ (1, 2) and ρ0 = A0 A† is the initial point such that:
L
0
Dist(A0 , A? ) ≤ γ 0 σr (A? ),
σr (ρ? )
1
4r )
for γ 0 := c · (1−δ
(1+δ4r ) · σ1 (ρ? ) , c ≤ 200 , where δ4r is the RIP
constant. Let At be the estimate of ProjFGD at the t-th
iteration.; then, the new estimate At+1 satisfies
Dist(At+1 , A? )2 ≤ α · Dist(At , A? )2 ,
(9)
(1−δ4r )·σr (ρ? )
where α := 1 − 550((1+δ
< 1. Further, At+1
4r )σ1 (ρ? )+kek2 )
satisfies Dist(At+1 , A? ) ≤ γ 0 σr (A? ), ∀t.
The above theorem provides a local convergence guarantee: given an initialization point ρ0 = A0 A†0 close
enough to the optimal solution –in particular, where
Dist(A0 , A? ) ≤ γ 0 σr (A? ) is satisfied– our algorithm converges locally with linear rate. In particular, in or2
der to obtain Dist(A
T , A? ) ≤ ε, ProjFGD requires
0
T = O log γ ·σrε(A? ) number of iterations. We conjecture that this further translatespinto linear convergence
√
√
in the infidelity metric, 1 − Tr(
ρT ρ? ρT ) 2 .
The per-iteration complexity of ProjFGD is dominated by the application of the linear map M and by
matrix-matrix multiplications. We note that, while both
eigenvalue decomposition and matrix multiplication are
known to have O((2n )2 r) complexity in Big-Oh notation,
the latter is at least two-orders of magnitude faster than
the former on dense matrices [33].
The proof of Theorem 5 is provided in the Appendix A.
We believe that our result, as stated in its most generality, complements recent results from the machine learning
and optimization communities, where different assumptions were made [26], or where constraints on A cannot
be accommodated [29].
So far, we assumed ρ0 is provided such that
Dist(A0 , A? ) ≤ γ 0 σr (A? ). The next theorem shows that
our initialization could achieve this guarantee (under assumptions) and turn the above local convergence guarantees to convergence to the global minimum.
†
Lemma 6. Let
A0 be such that ρ0 = A0 A0 =
−1
ΠC 0 L · ∇f (0) . Consider the problem in (3) where
M satisfies the RIP property for some constant δ4r ∈
(0, 1). Further, assume the optimum point ρ? satisfies
rank(ρ? ) = r. Then, A0 computed as above satisfies:
where γ 0
srank(ρ) =
Dist(A0 , A? ) ≤ γ 0 · σr (A? ),
r
1−δ
p
1− 1+δ4r
√ 4r · τ (ρ? ) ·
=
srank(ρ? ) and
2( 2−1)
kρkF
σ1 (ρ) .
This initialization introduces further restrictions on
?)
the condition number of ρ? , τ (ρ? ) = σσr1 (ρ
(ρ? ) , and the condition number of the objective function, which is pro1+δ4r
portional to ∝ 1−δ
. In particular, the initialization
4r
assumptions in Theorem 5 are satisfied by Lemma 6 if
and only if M satisfies RIP with constant δ4r fulfilling
the following expression:
√ √
q
1 + δ4r
1
2( 2−1)
4r
≤
.
· 1 − 1−δ
·√
1+δ
200
4r
1 − δ4r
r · τ 2 (ρ? )
While such conditions are hard to check a priori, our experiments showed that our initialization, as well as the
random initialization, work well in practice, and this behavior has been observed repeatedly in all the experiments we conducted. Thus, the method returns the exact solution of the convex programming problem, while
being orders of magnitude faster.
IV.
RELATED WORK
We focus on efficient methods for QST; for a broader
set of citations that go beyond QST, we defer the reader
to [33] and references therein.
The use of non-convex algorithms in QST is not
new [48, 49], and dates before the introduction of the
CS protocol in QST settings [5]. Assuming a multinomial distribution, [48] focus on the normalized negative
log-likelihood objective (see Eq. (2) in [48]) and propose
an diluted non-convex iterative algorithm for its solution.
The suggested algorithm exhibits good convergence and
monotonic increase of the likelihood objective in practice;
despite its success, there is no argument that guarantees
its performance, neither a provable setup for its execution.
[50, 51] use the reparameterization ρ = AA† in a Lagrange augmented maximum log-likelihood (ML) objective (see Eq. (3) in [50] and Eq. (9) in [51]), under the
multinomial distribution assumption. The authors state
that such a problem can be solved by standard numerical procedures for searching the maximum of the ML
objective [50], and use the downhill simplex method for
its solution, over the parameters of the matrix A [52].
Albeit [50, 51] rely on the uniqueness of the ML solution
before the reformulation ρ = AA† (due to the convexity
of the original problem), there are no theoretical results
on the non-convex nature of the transformed objective
(e.g., the presence of spurious local minima).
[53] consider the case of maximum likelihood quantum state tomography, under additive Gaussian noise, in
the informationally complete case. Assuming the measurement operators are traceless, simple linear inversion
techniques are shown to work accurately to infer the constrained ML state in a single projection step, from the
unconstrained ML state. As an extension, [54] present
a GPU implementation of the algorithm that recovers
a simulated 14-qubit density matrix within four hours;
however, implementations of a linear system inversion
could increase dramatically the computational and storage complexity, as the dimension of the problem grows.
Based on the extremal equations for the multinomial ML objective, [49] propose a fixed-point iteration
6
steepest-ascent method on A (with user-defined hyperparameters, such as the step size of the ascent). How
many iterations required and how to set up initial conditions are heuristically defined. Typically these methods,
as discussed in [55], lead to ill-conditioned optimization
problems, resulting in slow convergence.
[55] propose a hybrid algorithm that (i) starts with a
conjugate-gradient (CG) algorithm in the A space, in order to get initial rapid descent, and (ii) switch over to
accelerated first-order methods in the original ρ space,
provided one can determine the switchover point cheaply.
Under the multinomial ML objective, in the initial CG
phase, the Hessian of the objective is computed per iteration (i.e., a d2 × d2 matrix), along with its eigenvalue
decomposition. Such an operation is costly, even for moderate values of d, and heuristics are proposed for its completion. In the later phase, the authors exploit “momentum” techniques from convex optimization, that lead to
provable acceleration when the objective is convex; as we
state in the Conclusions section, such acceleration techniques have not considered in the factored space A, and
constitute an interesting research direction. From a theoretical perspective, [55] provide no convergence or convergence rate guarantees.
[56] use in practice the general parameterization of
AA†
density matrices, ρ = Tr(AA
† ) , that ensures jointly positive definiteness and unity in the trace. There, in order to attain the maximum value of the log-likehoood
objective, a steepest ascent method is proposed over A
variables, where the step size η is an arbitrarily selected
but sufficient small parameter. There is no discussion
regarding convergence and convergence rate guarantees,
as well as any specific set up of the algorithm (step size,
initialization, etc.).
[43] study the QST problem in the original parameter
space, and propose a projected gradient descent algorithm. The proposed algorithm applies both in convex
and non-convex objectives, and convergence only to stationary points could be expected. [57] extend the work in
[43] with two first-order variants, using momentum motions, similar to the techniques proposed by Polyak and
Nesterov for faster convergence in convex optimization
[58]. The above algorithms operate in the informationally complete case. Similar ideas in the informationally
incomplete case can be found in [59, 60].
Very recently, [61] presented an experimental implementation of CS tomography of a n = 7 qubit system,
where only 127 Pauli basis measurements are available.
To achieve recovery in practice –within a reasonable time
frame over hundreds of problem instances– the authors
proposed a computationally efficient estimator, based on
the factorization ρ = AA† . The resulting method resembles the gradient descent on the factors A, as the one
presented in this paper. However, the authors focus only
on the experimental efficiency of the method and provide
no specific results on the optimization efficiency of the algorithm, what are its theoretical guarantees, and how its
components (such as initialization and step size) affect
its performance (e.g., the step size is set to a sufficiently
small constant).
One of the first provable algorithmic solutions for the
QST problem was through convex approximations [10]:
this includes nuclear norm minimization approaches [5],
as well as proximal variants, as the one that follows:
minimize
ρ0
kM(ρ) − yk2F + λTr(ρ).
(10)
See also [5] for the theoretical analysis. Within this
context, we mention the work of [62]: there, the
AccUniPDGrad algorithm is proposed –a universal primaldual convex framework with sharp operators, in lieu of
proximal low-rank operators– where QST is considered
as an application. AccUniPDGrad combines the flexibility of proximal primal-dual methods with the computational advantages of conditional gradient (Frank-Wolfelike) methods. We will use this algorithm for comparisons
in the experimental section.
[63] presents SparseApproxSDP algorithm that solves
the QST problem in (2), when the objective is a generic
gradient Lipschitz smooth function, by updating a putative low-rank solution with rank-1 refinements, coming
from the gradient. This way, SparseApproxSDP avoids
computationally expensive operations per iteration, such
as full eigen-decompositions. In theory, at the r-th iteration, SparseApproxSDP is guaranteed to compute a 1r approximate solution,
with rank at most r, i.e., achieves
a sublinear O 1ε convergence rate. However, depending on ε, SparseApproxSDP might not return a low rank
solution.
Finally, [60] propose Randomized Singular Value Projection (RSVP), a projected gradient descent algorithm
for QST, which merges gradient calculations with truncated eigen-decompositions, via randomized approximations for computational efficiency.
Overall, our program is tailored for tomography of
highly-pure quantum states, by incorporating this constraint into the structure of A. This has two advantages.
First, it results in a faster algorithm that enables us
to deal many-qubit state reconstruction in a reasonable
time; and, second, it allows us prove the accuracy of the
ProjFGD estimator under model errors and experimental
noise, similar to the CS results.
V.
NUMERICAL EXPERIMENTS
We conducted experiments in an Matlab environment,
installed in a Linux-based system with 256 GB of RAM,
and equipped with two Intel Xeon E5-2699 v3 2.3GHz,
45M Cache, 9.60GT/s. In all the experiments, the error
reported in the Frobenius metric, kb
ρ−ρ? kF /kρ? kF , where
ρb is the estimation of the true state ρ? . Note that for a
pure state ρ, kρkF = 1. For some experiments we also
p √ √ 2
report the infidelity metric 1−Tr
ρ? ρb ρ? . We will
also use Sd to denote the set of d × d density matrices
Sd = {% : % ∈ Cd×d , % 0, Tr(%) = 1}.
7
d = 27
σ=0
Algorithm Time [s]
(1)
(2)
ProjFGD
46.01
77.12
0.28
d = 213
σ = 0.05
kb
ρ−ρ? kF
kρ? kF
Time [s]
5.3538e-07
3.0645e-04
3.2224e-08
58.48
65.53
0.30
kb
ρ−ρ? kF
kρ? kF
σ=0
Infidelity
6.0405e-02 3.0394e-02
6.1407e-02 3.0559e-02
2.3540e-02 1.3820e-04
σ = 0.05
Time [s]
kb
ρ−ρ? kF
kρ? kF
Time [s]
1314.01
6.8469e-08
1487.22
kb
ρ−ρ? kF
kρ? kF
Infidelity
3.1104e-02 1.9831e-03
TABLE I. All values are median values over 10 independent Monte Carlo iterations.
As a first set of experiments, we compare the efficiency of ProjFGD with second-order cone convex programs. State of the art solvers within this class of solvers
are the SeDuMi [64] and SDPT3 [65] methods; for their
use, we rely on the off-the-shelf Matlab wrapper CVX [66].
In our experiments, we observed that SDPT3 was faster
and we select it for our comparison.
The setting is as described in Section II: we consider rank-1 normalized density matrices ρ? ∈ S2n ,
from wen which we obtain Pauli measurements such that
yi = √2m Tr(Pi · ρ? ) + ei , i = 1, . . . , m, for some i.i.d.
Gaussian measurement error ei , with variance σ, i.e.,
∼ CN (0, σ · I). We consider both convex formulations
(1)-(2) and compare it to the ProjFGD estimator with
r = 1; in figures we use the notation CVX 1 and CVX 2
for simplicity.
We consider two cases: (i) n = 7, and (ii) n = 13.
Table I shows median values of 10 independent experimental realizations for m = 37 rd log d; this selection of m
was made so that all algorithms return a solution close
to the optimum ρ? . Empirically, we have observed that
ProjFGD succeeds even for cases m = O(rd). We consider
both noiseless σ = 0 and noisy σ = 0.05 settings.
In order to accelerate the execution of convex programs, we set the solvers in CVX to low precision. From
Table I, we observe that our method is two orders of
magnitude faster than second-order methods for n = 7:
ProjFGD achieves better performance (in both error metrics, and in both noisy/noiseless cases), faster. For the
higher n = 13 qubit case, we could not complete the
experiments for (1)-(2) due to system crash (RAM overflow). Contrariwise, our method was able to complete
the task with success within about 22 minutes of CPU
time.
Figures 1-2 show graphically how second-order convex
vs. our first-order non-convex schemes scale, as a function of time. In Figure 1, we fix the dimension to d = 27
and study how increasing the number of observations m
affects the performance of the algorithms. We observe
that, while in the ProjFGD, more observations lead to
faster convergence [67], the same does not hold for the
second-order cone programs. In Figure 2, we fix the number data points to m = 37 rd log d, and we scale the dimension d. It is obvious that the convex solvers do not
scale easily beyond n = 7, whereas our method handles
cases up to n = 13, within reasonable time.
10 3
2.1e-04
1.7e-04 1.7e-04
10 2
Time (sec.)
Comparison of ProjFGD with second-order
methods
6.5e-04
2.6e-06
3.8e-04
2.8e-04 3.1e-04
2.9e-04
4.0e-07
3.7e-07
9.8e-07 7.3e-07
1.9e-04 1.8e-04
5.4e-07
4.3e-07
3.4e-07
2.6e-07
3.9e-07
CVX 1
CVX 2
ProjFGD
10 1
10 0
1.8e-07
7.2e-08
10 -1
500
5.0e-08
1000
3.2e-08 3.1e-08 2.2e-08 1.9e-08 2.1e-08
1.6e-08 1.5e-08
1500
2000
2500
3000
3500
Number of data points
FIG. 1. Dimension fixed to d = 27 with rank(ρ? ) = 1. The
figure depicts the noiseless setting, σ = 0. Numbers within
figure are the error in Frobenius norm achieved (median values).
10 4
10 3
Time (sec.)
A.
10 2
CVX 1
CVX 2
ProjFGD
10 1
10 0
10 -1
10 -2 1
10
10 2
10 3
10 4
Dimension d
FIG. 2. Number of data points set to m = 73 rd log d. Rank of
optimum point is set to rank(ρ? ) = 1. The figure depicts the
noiseless setting.
8
B.
Comparison of ProjFGD with first-order methods
Here, we compare our method with more efficient firstorder methods, both convex (AccUniPDGrad [62]) and
non-convex (SparseApproxSDP [63] and RSVP [60]).
We consider two settings: ρ? ∈ Sd is (i) a pure state
(i.e., rank(ρ? ) = 1) and, (ii) a nearly low-rank state.
In the latter case, we construct ρ? = ρ?,r + ζ, where
ρ?,r is a rank-deficient PSD satisfying rank(ρ?,r ) = r,
and ζ ∈ Cd×d is a full-rank PSD noise term with a fast
decaying eigen-spectrum, significantly smaller than the
leading eigenvalues of ρ?,r . In other words, we can wellapproximate ρ? with ρ?,r . For all cases, we model the
measurement vector as y = M(ρ? ) + e; here, the noise
is such that kek = 10−3 . The number of data points m
satisfy m = Csam · rd, for various values of Csam > 0.
For all algorithms, we assumed r = rank(ρ?,r ) is known
and use it to reconstruct a rank-r approximation of ρ? .
All methods that require an SVD routine use lansvd(·)
from the PROPACK software package. Experiments and
algorithms are implemented in a Matlab environment;
we used non-specialized and non-mexified code parts for
all algorithms. For initialization, we use the same starting point for all algorithms, which is either specific (Section III A) or random. As a stopping criterion, we use
kρt+1 −ρt kF
≤ tol; we set the tolerance parameter to
kρt+1 kF
tol := 5 · 10−6 .
Convergence plots. Figure 3 (two-leftmost plots) illustrates the iteration and timing complexities of each
algorithm under comparison, for a pure state recovery
setting (r = 1) of a highly-pure ρ? . Here, n = 12 which
corresponds to a d2 = 16, 777, 216 dimensional problem;
moreover, we assume Csam = 3 and thus the number of
data points are m = 12, 288. For initialization, we use
the proposed initialization in Section III A for all algorithms: we compute −M† (y), extract factor A0 as the
best-r PSD approximation of −M† (y), and project A0
onto C.
It is apparent that ProjFGD converges faster to a vicinity of ρ? , compared to the rest of the algorithms; observe
also the sublinear rate of SparseApproxSDP in the inner
plots, as reported in [63].
Table II contains recovery error and execution time
results for the case n = 13 (d = 8096); in this case,
we solve a d2 = 67, 108, 864 dimensional problem. For
this case, RSVP and SparseApproxSDP algorithms were
excluded from the comparison, due to excessive execution time. Appendix C provides extensive results, where
similar performance is observed for other values of d = 2n
and Csam .
Figure 3 (rightmost plot) considers the more general
case where ρ? is nearly low-rank: i.e., it can be wellapproximated by a density matrix ρ?,r where r = 20 (lowrank density matrix). In this case, n = 12, m = 245, 760
for Csam = 3. As the rank in the model, r, increases,
algorithms that utilize an SVD routine spend more CPU
time on singular value/vector calculations. Certainly, the
Algorithm
kb
ρ−ρ? kF
kρ? kF
Time [s]
AccUniPDGrad
ProjFGD
7.4151e-02
8.6309e-03
2354.4552
1214.0654
TABLE II. Comparison results for reconstruction and efficiency, for n = 13 qubits and Csam = 3.
same applies for matrix-matrix multiplications; however,
in the latter case, the complexity scale is milder than
that of the SVD calculations. Further metadata are also
provided in Table III.
Setting: r = 5.
Algorithm
kb
ρ−ρ?,r kF
kρ?,r kF
SparseApproxSDP
RSVP
AccUniPDGrad
ProjFGD
3.17e-02
5.15e-02
2.01e-02
1.20e-02
Setting: r = 20.
Time [s]
kb
ρ−ρ?,r kF
kρ?,r kF
Time [s]
3.74
0.78
0.36
0.06
5.49e-02
1.71e-02
1.54e-02
7.12e-03
4.38
0.38
0.33
0.04
TABLE III. Results for reconstruction and efficiency. Time
reported is in seconds. For all cases, Csam = 3 and n = 10.
For completeness, in Appendix C we provide results
that illustrate the effect of random initialization: Similar
to above, ProjFGD shows competitive behavior by finding a better solution faster, irrespective of initialization
point.
Timing evaluation (total and per iteration). Figure 4
highlights the efficiency of our algorithm in terms of time
complexity, for various problem configurations. Our algorithm has fairly low per iteration complexity (where
the most expensive operation for this problem is matrixmatrix and matrix-vector multiplications). Since our algorithm shows also fast convergence in terms of the number of iterations, this overall results into faster convergence towards a good approximation of ρ? , even as the
dimension increases. Figure 4 shows how the total execution time scales with parameters n and r.
Overall performance. ProjFGD shows a substantial improvement in performance, as compared to the stateof-the-art algorithms; we would like to emphasize also
that projected gradient descent schemes, such as in [60],
are also efficient in small- to medium-sized problems,
due to their fast convergence rate. Further, convex approaches might show better sampling complexity performance (i.e., as Csam decreases). Nevertheless, one can
perform accurate MLE reconstruction for larger systems
in the same amount of time using our methods for such
small- to medium-sized problems. We defer the reader to
Appendix C, due to space restrictions.
9
Rank r = 1 - Data points m = 3 " d " r
Rank r = 1 - Data points m = 3 " d " r
AccUniPDGrad
ProjFGD
0.2
200
10
10 -2
0
0.5
400
600
10
20
-1
10
SparseApproxSDP
0
10 -1
1
800
30
2000 4000 6000 8000 100001200014000
2000 4000 6000 8000 10000 12000
1000
Number of iterations
0
10
10 -1
;!;?;r kF
kb
k;?;r kF
0.4
0
RSVP
SparseApproxSDP
;!;?;r kF
kb
k;?;r kF
0.6
10 -2
-1
;!;? kF
kb
k;? kF
;!;? kF
kb
k;? kF
0.8
10 -1
10
RSVP
SparseApproxSDP
1
-2
AccUniPDGrad
ProjFGD
10 0
0
;!;? kF
kb
k;? kF
10
-1
;!;? kF
kb
k;? kF
10
Rank r = 20 - Data points m = 3 " d " r
AccUniPDGrad
ProjFGD
Cumulative time (sec)
10-2
0
40
200
400
Cumulative time (sec)
Number of iterations
Cumulative time (sec)
10-2
600
0
100
200
300
400
Cumulative time (sec)
Total time (log-scale)
Total time (log-scale)
FIG. 3. Left and middle panels: Convergence performance of algorithms under comparison w.r.t. error in Frobenius norm
vs. (i) the total number of iterations (left) and (ii) the total execution time. Both cases correspond to Csam = 3, r = 1 (pure
state) and n = 12 (i.e., d = 4096). Right panel: Nearly low-rank state case—we approximate ρ? with ρ?,r , with r = 20. In
this setting, n = 12 (i.e., d = 4096) and Csam = 3.
10
8
6
4
RSVP
SparseApproxSDP
AccUniPDGrad
ProjFGD
2
0
-2
-4
6
8
10
n-qubit (d = 2n )
12
10
8
6
4
RSVP
SparseApproxSDP
AccUniPDGrad
ProjFGD
2
0
1
5
20
Rank r
FIG. 4. Timing bar plot.y-axis shows total execution time
in (log10 -scale) while x-axis corresponds to different n values.
Top panel corresponds to r = 1 and Csam = 6; bottom panel
corresponds to n = 10 and Csam = 3. All cases are noiseless.
VI.
SUMMARY AND CONCLUSIONS
In this work, we propose a non-convex algorithm,
dubbed as ProjFGD, for estimating a highly-pure quantum state, in a high-dimensional Hilbert space, from relatively small number of data points. We showed empirically that ProjFGD is orders of magnitude faster than
state-of-the-art convex and non-convex programs, such
as [62],[63], and [60]. More importantly, we prove that
under proper initialization and step-size, the ProjFGD is
guaranteed to converge to the global minimum of the
problem, thus ensuring a provable tomography proce-
dure; see Theorem 5 and Lemma 6.
In our setting, we model the state as a low-rank PSD
matrix. This, in turn, means that the estimator is biased
towards low-rank states. However, such bias is inherent
to all CS-like QST protocols by the imposition of the
positivity constraint [13].
Our techniques and proofs can be applied –in some
cases under proper modifications– to scenaria beyond the
ones considered in this work. We restricted our discussions to a measurement model of random Pauli observables, that satisfies RIP. We conjecture that our results
apply for other “sensing” settings, that are informationally complete for low-rank states; see e.g., [7]. The results
presented here are independent of the noise model and
could be applied for non-Gaussian noise models, such as
those stemming from finite counting statistics. Lastly,
while here we focus on state tomography, it would be interesting to explore similar techniques for the problem of
process tomography.
We conclude with a short list of interesting future research directions. Our immediate goal is the application
of ProjFGD in real-world scenaria; this could be completed by utilizing the infrastructure at the IBM T.J.
Watson Research Center [3]. This could complement the
results found in [61] for a different quantum system.
Beyond this practical implementation, we identify the
following interesting open questions. First, the ML estimator is one of the most frequently-used methods for
QST experiments. Beyond its use as point estimator,
it is also used as a basis for inference around the point
estimate, via confidence intervals [68] and credible regions [69]. However, there is still no rigorous analysis
when the factorization ρ = AA† is used.
The work of [55] considers accelerated gradient descent methods for QST in the original parameter space
ρ: Based on the seminal work of Polyak and Nesterov
[58] on convex optimization first-order methods, one can
achieve orders of magnitude acceleration (both in theory
and practice), by exploiting the momentum from previous iterates. It remains an open question how our ap-
10
proach could exploit acceleration techniques that lead to
faster convergence in practice, along with rigorous approximation and convergence guarantees. Further, distributed/parallel implementations, like the one in [54],
remain widely open using our approach, in order to accelerate further the execution of the algorithm. Research
along these directions is very interesting and is left for
future work.
Finally, while we saw numerically that a random initialization under noisy and constrained settings works
well, a careful theoretical treatment for this case is an
open problem.
To shed some light on the notions of (restricted) strong convexity and smoothness and how they relate to the QST objective, consider the restricted isometry property, which holds
with high probability under Pauli measurements for low rank
ρ? [11, 15]; here, we present a simplified version of the definition in the main text:
Definition 9 (Restricted Isometry Property (RIP)). A linear
map M satisfies the r-RIP with constant δr , if
(1 − δr )kρk2F ≤ kM(ρ)k22 ≤ (1 + δr )kρk2F ,
is satisfied for all matrices ρ ∈ Cd×d such that rank(ρ) ≤ r.
According to the quadratic loss function in QST:
f (ρ) = 12 ky − M(ρ)k22 ,
ACKNOWLEDGMENTS
Anastasios Kyrillidis is supported by the IBM Goldstine Fellowship and Amir Kalev is supported by the Department of Defense.
Appendix A: Theory
Notation. For matrices ρ, ζ ∈ Cd×d , hρ, ζi = Tr ρ† ζ represents their inner product. We use kρkF and σ1 (ρ) for the
Frobenius and spectral norms of a matrix, respectively. We
denote as σi (ρ) the i-th singular value of ρ. ρr denotes the
best rank-r approximation of ρ.
1.
Problem generalization, notation and definitions
To expand the generality of our scheme, we re-state the
problem setting for a broader set of objectives. The following arguments hold for both real and complex matrices. We
consider criteria of the following form:
its Hessian is given by M† M(·). Then, (restricted) strong
convexity suggests that:
kM(ρ)k22 ≥ C · kρk2F ,
for a restricted set of directions ρ, where C > 0 is a small
constant. Then, the correspondence of restricted strong convexity and smoothness with the RIP is obvious: both lower
and upper bound the quantity kM(ρ)k22 , where ρ is drawn
from a restricted (low-rank) set. It turns out that linear maps
that satisfy the RIP for low rank matrices, also satisfy the restricted strong convexity; see Theorem 2 in [71].
By assuming RIP, the condition number κ of f depends
on the RIP constants of the linear map M; in particular,
1+δ
∝ 1−δ
, since the eigenvalues of
one can show that κ = L
µ
†
M M lie between 1−δ and 1+δ, when restricted to low-rank
matrices. Observe that for δ sufficiently small and dimension
d sufficiently large, κ ≈ 1, with high probability.
We assume the optimum ρ? satisfies rank(ρ? ) = r? . For
our analysis, we further assume we know r? and set r? ≡ r.
As suggested in the main text, we solve (A1) in the factored
space, as follows:
minimize
A∈Cd×r
minimize
ρ∈Cd×d
f (ρ)
subject to
ρ 0, ρ ∈ C 0 .
(A1)
To make connection with the QST objective, set f (ρ) =
1
· ky − M(ρ)k22 , and ρ ∈ C 0 ⇔ Tr(ρ) ≤ 1. Apart from
2
the least-squares objective in QST, our theory extends to applications that can be described by strongly convex functions
f with gradient Lipschitz continuity. Further, our ideas can be
applied in a similar fashion to the case of restricted smoothness and restricted strong convexity [70]. We state these standard definitions below for the square case.
Definition 7. Let f : Cd×d → R be convex and differentiable. f is restricted µ-strongly convex if for all rank-r matrices ρ, ζ ∈ Cd×d ,
f (ζ) ≥ f (ρ) + h∇f (ρ) , ζ − ρi +
µ
2
kζ − ρk2F .
(A2)
Definition 8. Let f : Cd×d → R be a convex differentiable
function. f is restricted gradient Lipschitz continuous with
parameter L (or L-smooth) if for all rank-r matrices ρ, ζ ∈
Cd×d ,
k∇f (ρ) − ∇f (ζ)kF ≤ L · kρ − ζkF .
(A3)
ρ ∈ Cd×d ,
f (AA† )
subject to
A ∈ C.
(A4)
In the QST setting, A ∈ C ⇔ kAk2F ≤ 1.
In our theory we mostly focus on sets C that satisfy the
following assumptions.
Assumption 1. For ρ 0, there is A ∈ Cd×r and r ≤ d such
that ρ = AA† . Then, C 0 ⊆ Cd×d is endowed with a constraint
set C ⊆ Cd×r that (i) for each ρ ∈ C 0 , there is an subset in
C where each A ∈ C satisfies ρ = AA† and (ii) its projection
operator, say ΠC (B) = argminA∈C 12 kA − Bk2F for B ∈ Cd×r ,
is an entrywise scaling operation on the input B.
We also require the following faithfulness assumption [26]:
Assumption 2. Let E denote the set of equivalent facd×d
torizations
that lead to a rank-r
; i.e.,
n
o matrix ρ? ∈ C
†
E := A? ∈ Cd×r : ρ? = A? A? . Then, we assume E ⊆ C,
i.e., the resulting convex set C in (A4) (from C 0 in (A1)) respects the structure of E.
Summarizing, by faithfulness of C (Assumption 2), we assume that E ⊆ C. This means that the feasible set C in
(A4) contains all matrices A? that lead to ρ? = A? A†? in
(A1). Moreover, we assume both C, C 0 are convex sets and
11
there exists a “mapping” from C 0 to C, such that the two constraints are “equivalent”: i.e., ∀A ∈ C, we are guaranteed that
ρ = AA† ∈ C 0 . We restrict our discussion on norm-based sets
for C such that Assumption 1 is satisfied. As a representative example, consider the QST case where, for any ρ = AA† ,
Tr(ρ) ≤ 1 ⇔ kAk2F ≤ 1.
For our analysis, we will use the following step sizes:
ηb =
η? =
1
,
128 Lσ1 (ρt ) + σ1 QAt Q†At ∇f (ρt )
1
.
128 (Lσ1 (ρ? ) + σ1 (∇f (ρ? )))
Here, L is the Lipschitz constant in (5) and QAt Q†At represents the projection onto the column space of At . In our
algorithm, as described in the main text, we use the following
step size:
η≤
1
128 (Lσ1 (ρ0 ) + σ1 (∇f (ρ0 )))
for given initial ρ0 .
While different, by Lemma A.5 in [29], we know that ηb ≥ 65 η
11
and 10
η ≤ η ≤ 10
η? . Thus, in our proof, we will work
11 ?
with step size ηb, which is equivalent –up to constants– to the
original step size η in the proposed algorithm.
For ease of exposition, we re-define the sequence of updates: At is the current estimate in the factored space,
et+1 = At − ηb∇f (ρt )At is the putative solution after the
A
et+1 might not belong in C),
gradient step (observe that A
e
and At+1 = ΠC (At+1 ) is the projection step onto C. Observe that for the constraint cases we consider in this paper,
et+1 ) = ξt (A
et+1 ) · A
et+1 , where ξt (·) ∈ (0, 1); in
At+1 = ΠC (A
the case ξt (·) = 1, the algorithm simplifies to the algorithm
in [29]. For simplicity, we drop the subscript and the parenthesis of the ξ parameter; these values are apparent from the
context.
An important issue in optimizing f over the factored space
is the existence of non-unique possible factorizations. We use
the following rotation invariant distance metric:
Definition 10. Let matrices A, B ∈ Cd×r . Define:
Dist (A, B) := min kA − BRkF ,
Theorem 11. Let C ⊆ Cd×r be a convex, compact, and faithful set, with projection operator satisfying the assumptions
described above. Let f be a convex function satisfying Definitions 7 and 8.
Let At ∈ C be the current estimate and ρt = At A†t . Assume current point At satisfies Dist(At , A? ) ≤ γ 0 σr (A? ),
(ρ? )
µ
1
· σσr1 (ρ
, c ≤ 200
, and given ξt (·) &
for γ 0 := c · L
?)
0.78 per iteration, the new
estimate
of ProjFGD, At+1
=
†
†
ΠC At − ηb∇f (At At ) · At = ξt · At − ηb∇f (At At ) · At satisfies
Dist(At+1 , A? )2 ≤ α · Dist(At , A? )2 ,
(A5)
r (ρ? )
< 1. Further, At+1
where α := 1 − 550(Lσ1 (ρµ·σ
? )+σ1 (∇f (ρ? )))
0
satisfies Dist(At+1 , A? ) ≤ γ σr (A? ).
When applied to the QST setting, we obtain the following
variation of the above theorem:
Theorem 12 (Local convergence rate for QST). Let ρ? be
the quantum state of an n-qubit system, y ∈ Rm be the measurement vector of m = O(rn6 2n ) random n-qubit Pauli observables, and M be the corresponding sensing map, such that
yi = (M(ρ? ))i + ei , ∀i = 1, . . . , m.
Let At be the current estimate of ProjFGD. Assume At satis4r ) σr (ρ? )
·
, c≤
fies Dist(At , A? ) ≤ γ 0 σr (A? ), for γ 0 := c · (1−δ
(1+δ4r ) σ1 (ρ? )
1
,
where
δ
is
the
RIP
constant.
Then,
the
new
estimate
4r
200
At+1 satisfies
Dist(At+1 , A? )2 ≤ α · Dist(At , A? )2 ,
(1−δ4r )·σr (ρ? )
where α := 1 − 550((1+δ
< 1. Further, At+1
4r )σ1 (ρ? )+kek2 )
satisfies Dist(At+1 , A? ) ≤ γ 0 σr (A? ).
3.
Proof of Theorem 11
For our analysis, we make use of the following lemma [72,
Chapter 3], which characterizes the effect of projections onto
convex sets w.r.t. to inner products, as well as provides a
type-of triangle inequality for such projections; see also Figure
5 for a simple illustration.
R:R∈U
where U is the set of r × r unitary matrices R.
We assume that ProjFGD is initialized with a “good” starting point ρ0 = A0 A†0 , such that:
(A1)
Lemma 13. Let U ∈ C ⊆ Cd×r and V ∈ Cd×r where V ∈
/ C.
Then,
hΠC (V ) − U, V − ΠC (V )i ≥ 0.
A0 ∈ C and Dist(A0 , A? ) ≤ γ 0 σr (A? ) for γ 0 :=
(ρ? )
µ
1
, where c ≤ 200
· σσr1 (ρ
.
c· L
?)
Here, σr (·) denotes the r-singular value of the input matrix,
in descending order. Later in the text, we present an initialization that, under assumptions, leads further to global
convergence results.
2.
Generalized theorem
Next, we present the full proof of the following generalization of Theorem 5:
C
U
⇧C (V )
V
FIG. 5. Illustration of Lemma 13
(A6)
12
We start with the following series of (in)equalities:
2
Dist (At+1 , A? ) = min kAt+1 −
R∈U
A? Rk2F
(i)
?
≤ kAt+1 − A? RA
k2
t F
(ii)
?
et+1 + A
et+1 − A? RA
= kAt+1 − A
k2
t F
?
et+1 k2F + kA
et+1 − A? RA
= kAt+1 − A
k2
t F
D
E
?
et+1 , A
et+1 − A? RA
+ 2 At+1 − A
,
t
?
where (i) is due to the fact RA
:= argminR∈U kAt − A? Rk2F ,
t
et+1 .
(ii) is obtained by adding and subtracting A
Focusing on the second term of the right hand side, we
et+1 to obtain:
substitute A
?
?
et+1 − A? RA
k2
kA
k2 = kAt − ηe∇f At A†t At − A? RA
t F
t F
?
= kAt − A? RA
k2 + ηe2 k∇f At A†t At k2F
t F
D
E
?
− 2e
η ∇f At A†t At , At − A? RA
t
Then, our initial equation transforms into:
et+1 k2F
Dist (At+1 , A? )2 ≤ kAt+1 − A
+ Dist (At , A? )2 + ηe2 k∇f At A†t At k2F
D
E
?
− 2e
η ∇f At A†t At , At − A? RA
t
D
E
?
et+1 , A
et+1 − A? RA
+ 2 At+1 − A
t
Focusing further on the last term of the expression above, we
obtain:
D
E
?
et+1 , A
et+1 − A? RA
At+1 − A
t
D
E
?
et+1 , A
et+1 − At+1 + At+1 − A? RA
= At+1 − A
t
D
E
et+1 , A
et+1 − At+1
= At+1 − A
E
D
?
et+1 , At+1 − A? RA
+ At+1 − A
t
et+1 ≡ At+1 for all
Observe that, in the special case where A
t, i.e., the iterates are always within C before the projection
step, the above equation equals to zero and the recursion is
identical to that of [29][Proof of Theorem 4.2]. Here, we are
et+1 6≡ At+1 for some t—
more interested in the case where A
et+1 6∈ C. By faithfulness (Assumption 2), observe that
thus A
†
?
?
?
A? RA
∈ C and ρ? = A? RA
A? RA
= A? A†? . Moreover,
t
t
t
et+1 ): Then, according to Lemma 13 and focusAt+1 = ΠC (A
?
et+1 , the last
ing on eq. (A6), for U := A? RA
and V := A
t
term in the above equation satisfies:
D
E
?
et+1 , At+1 − A? RA
At+1 − A
≤ 0,
t
and, thus, the expression above becomes:
D
E
?
et+1 , A
et+1 − A? RA
et+1 k2F .
At+1 − A
≤ −kAt+1 − A
t
Therefore, going back to the original recursive expression, we
obtain:
et+1 k2F
Dist (At+1 , A? )2 ≤ −kAt+1 − A
+ Dist (At , A? )2 + ηe2 k∇f At A†t At k2F
D
E
?
− 2e
η ∇f At A†t At , At − A? RA
t
For the last term, we use the following descent lemma 14;
its proof is provided in Section A 4.
et+1 = At − ηb∇f (ρt ) ·
Lemma 14 (Descent lemma). Let A
At . For f restricted L-smooth and µ-strongly convex, and
under the same assumptions with Theorem 11, the following
inequality holds true:
?
et+1 k2F
2e
η ∇f (At A†t ) · At , At − A? RA
+ kAt+1 − A
t
≥ ηe2 k∇f (At A†t )At k2F +
3e
ηµ
10
· σr (ρ? ) · Dist(At , A? )2 .
Using the above lemma in our expression, we get:
ηµ
Dist (At+1 , A? )2 ≤ 1 − 3e
· σr (ρ? ) · Dist(At , A? )2 .
10
The expression for α is obtained by observing ηe ≥ 56 η and
11
≤ η ≤ 10
η? , from Lemma 20 in [29]. Then, for η? ≤
C
and C = 1/128, we have:
Lσ1 (ρ? )+σ1 (∇f (X? ))
10
η
11 ?
1−
3 · 10
· 56 η? µ
3e
ηµ
11
· σr (ρ? ) ≤ 1 −
· σr (ρ? )
10
10
15
η? µ · σr (ρ? )
=1−
66
µ · σr (ρ? )
15
=1−
66 128(Lσ1 (ρ? ) + σ1 (∇f (ρ? )))
µ · σr (ρ? )
≤1−
=: α
550(Lσ1 (ρ? ) + σ1 (∇f (ρ? )))
where α < 1.
Concluding the proof, the condition Dist(At+1 , A? )2 ≤
γ 0 σr (A? ) is naturally satisfied, since α < 1.
4.
Proof of Lemma 14
?
et+1 = At −e
.
Recall A
η ∇f (ρt )At and define ∆ := At −A? RA
t
Before presenting the proof, we need the following lemma
that bounds one of the error terms arising in the proof of
Lemma 14. This is a variation of Lemma 6.3 in [29]. The
proof is presented in Section A 5.
Lemma 15. Let f be L-smooth and µ-restricted strongly convex. Then, under the assumptions of Theorem 11 and assum1
ing step size ηb =
, the following
†
128(Lσ1 (ρt )+σ1 (∇f (ρt )QAt QA ))
t
bound holds true:
D
E
∇f (ρt ), ∆∆† ≥ − η5e k∇f (ρt )At k2F −
µσr (ρ? )
10
· Dist(At , A? )2 .
(A7)
Now we are ready to present the proof of Lemma 14.
Proof of Lemma 14. First we rewrite the inner product as
shown below.
?
h∇f (ρt )At , At − A? RA
i
t
D
E
?
= ∇f (ρt ), ρt − A? RAt A†t
D
E
?
= 12 h∇f (ρt ), ρt − ρ? i + ∇f (ρt ), 12 (ρt + ρ? ) − A? RA
A†t
t
D
E
= 12 h∇f (ρt ), ρt − ρ? i + 12 ∇f (ρt ), ∆∆† ,
(A8)
which follows by adding and subtracting
1
ρ .
2 ?
13
Let us focus on bounding the first term on the right hand
side of (A8). Consider points ρt = At A†t and ρt+1 =
At+1 A†t+1 ; by assumption, both ρt and ρt+1 are feasible points
in (A4). By smoothness of f , we get:
f (ρt ) ≥ f (ρt+1 ) − h∇f (ρt ), ρt+1 − ρt i −
(i)
≥ f (ρ? ) − h∇f (ρt ), ρt+1 − ρt i −
L
2
L
2
µ
2
et+1 k2 = kξ · A
et+1 − A
et+1 k2
kAt+1 − A
F
F
= (1 − ξ)2 · kAt k2F + (1 − ξ)2 ηe2 · k∇f (ρt )At k2F
kρt+1 − ρt k2F
kρt+1 − ρt k2F ,
(A9)
where (i) follows from optimality of ρ? and since ρt+1 is a
feasible point (ρt+1 0, ΠC 0 (ρt+1 ) = ρt+1 ) for problem (A1).
Moreover, by the restricted strong convexity of f , we get,
f (ρ? ) ≥ f (ρt ) + h∇f (ρt ), ρ? − ρt i +
For
thelast term in the above expression and given At+1 =
et+1 = ξ · A
et+1 for some ξ ∈ (0, 1), we further observe:
ΠC A
kρ? − ρt k2F .
(A10)
− 2 (1 − ξ)2 · ηe · h∇f (ρt )At , At i
Combining the above equality with the first term on
the right hand side in (A13), we obtain the expression in
(A14). Focusing on the first term, let Θt := I + 2(1−ξ)
· ηe ·
3ξ−1
∇f (ρt )QAt Q†At ; then, σd (Θt ) ≥ 1 −
nition of ηe and the fact that ηe ≤
2(1−ξ)
1
· 128
, by the defi3ξ−1
1
. Then:
†
128σ1 ∇f (ρt )QAt QA
t
Combining equations (A9), and (A10), we obtain:
3ξ−1
2
h∇f (ρt ), ρt − ρ? i ≥ h∇f (ρt ), ρt − ρt+1 i
−
L
2
kρt+1 − ρt k2F +
µ
2
kρ? − ρt k2F
(A11)
By the nature of the projection ΠC (·) step, it is easy to verify
that
ρt+1 = ξ 2 · ρt − ηe∇f (ρt )ρt Λt − ηeΛ†t ρ†t ∇f (ρt )† ,
where Λt = I − η2e QAt Q†At ∇f (Xt ) ∈ Cd×d and QAt Q†At denoting the projection onto the column space of At . Notice that,
for step size ηe, we have
Λt 0,
σ1 (Λt ) ≤ 1 + 1/256
and
σd (Λt ) ≥ 1 − 1/256.
Using the above ρt+1 characterization in (A11), we obtain:
h∇f (ρt ), ρt − ρ? i −
(i)
≥ ∇f (ρt ), 1 − ξ
µ
2
2
kρ? − ρt k2F +
L
2
≥
(3ξ−1)2
4
≥
(3ξ−1)2
4
ρt + 2e
η · ξ 2 · h∇f (ρt ), ∇f (ρt )Xt Λt i
−
(ii)
≥
≥ (1 −
2
F
· kAt k2F · σd (Θt )2
2
1
· 1 − 2(1−ξ)
·
· kAt k2F
3ξ−1
128
η
eµσr (ρ? )
10
· Dist(Ut , A? )2
(3ξ−1)2
·
1− 1−
(1 − ξ)2 −
4
η
eL
2
·
1
128
2
· kAt k2F
(A15)
Let us focus on the term
bounded as follows:
η
eL
2
2(1−ξ)
3ξ−1
kρt − ρt+1 k2F
η
eL
2
kρt − ρt+1 k2F ; this can be
η
eL
kAt A†t − At+1 A†t+1 k2F
2
†
η
eL
kAt At − At A†t+1 + At A†t+1 − At+1 A†t+1 k2F
2
η
eL
kAt (At − At+1 )† + (At − At+1 ) A†t+1 k2F
2
kρt − ρt+1 k2F =
=
1/256) k∇f (ρt )At k2 .
F
· At
?
et+1 k2
2e
η ∇f (ρt ) · At , At − A? RA
+ kAt+1 − A
F
t
2
η
eµ
2
2
1
≥ 255·ξ
−
·
η
e
k∇f
(X
)U
k
+
kρ
−
ρt k2F
t
t
?
F
128
5
2
+
= Tr(∇f (ρt )∇f (ρt )At A†t ) − η2e Tr(∇f (ρt )∇f (ρt )At A†t ∇f (ρt ))
≥ 1 − η2e kQAt Q†At ∇f (ρt )k2 k∇f (ρt )At k2F
3ξ−1
Θt
2
=
The above transform (A13) as follows:
Tr(∇f (ρt )∇f (ρt )ρt Λt )
F
et+1 k2F
1 − ξ 2 · ηe h∇f (ρt )At , At i + kAt+1 − A
2
2
2(1−ξ)
1
· kAt k2F
·
1
−
1
−
·
≥ (1 − ξ)2 − (3ξ−1)
4
3ξ−1
128
−
where: (i) follows from symmetry of ∇f (ρt ) and ρt and, (ii)
follows from the sequence equalities and inequalites:
2
Combining the above, we obtain the following bound:
kρt − ρt+1 k2F
1 − ξ 2 · h∇f (ρt )At , At i + 2e
η · ξ 2 · Tr(∇f (ρt )∇f (ρt )ρt ) · σd (Λt )
η ·ξ2
k∇f (ρt )At k2F ,
(A12)
≥ 1 − ξ 2 · h∇f (ρt )At , At i + 255·e
128
· At + (1 − ξ) · ηe∇f (ρt ) · At
=
(i)
≤ ηeL · kAt (At − At+1 )† k2F + k (At − At+1 ) A†t+1 k2F
(ii)
≤ ηeL σ1 (At+1 )2 + σ1 (At )2 · kAt+1 − At k2F .
Combining the above in the expression we want to lower
?
et+1 k2F , we
bound: 2e
η ∇f (ρt ) · At , At − A? RA
+ kAt+1 − A
t
obtain:
?
et+1 k2F
2e
η h∇f (ρt ) · At , At − A? RA
i + kAt+1 − A
t
D
E
et+1 k2F
= ηe h∇f (ρt ), ρt − ρ? i + ηe ∇f (ρt ), ∆∆† + kAt+1 − A
η 2 ·ξ2
k∇f (ρt )At k2F
≥ 1 − ξ 2 · ηe h∇f (ρt )At , At i + 255·e
128
+
−
+
η
eµ
2
kρ? − ρt k2F −
η
e2
k∇f (ρt )At k2F
5
et+1 k2F
kAt+1 − A
η
eL
2
−
kρt − ρt+1 k2F
η
eµσr (ρ? )
10
where (i) is due to the identity kA + Bk2F ≤ 2kAk2F + 2kBk2F
and (ii) is due to the Cauchy-Schwarz inequality. By definition of At+1 , we observe that:
σ1 (At+1 )2 = σ1 (ξ · (At − ηe∇f (ρt )At ))2
2
(i)
≤ ξ 2 · σ1 (At )2 · σ1 I − ηe∇f (ρt )QAt Q†At
(ii)
≤ 1+
· Dist(At , A? )2
(A13)
2
1
128
· σ1 (At )2 .
where (i) is due to Cauchy-Schwarz and (ii) is obtained by
1
and since ξ ∈ (0, 1).
substituting ηe ≤
†
128σ1 ∇f (ρt )QAt QA
t
14
1 − ξ 2 · ηe h∇f (ρt )At , At i + (1 − ξ)2 · kAt k2F + (1 − ξ)2 ηe2 · k∇f (ρt )At k2F
− 2 (1 − ξ)2 · ηe · h∇f (ρt )At , At i =
1−ξ
2
− 2 (1 − ξ)
2
· ηe h∇f (ρt )At , At i + (1 − ξ)2 · kAt k2F + (1 − ξ)2 ηe2 · k∇f (ρt )At k2F =
(3ξ − 1) (1 − ξ) · ηe h∇f (ρt )At , At i + (1 − ξ)2 · kAt k2F + (1 − ξ)2 ηe2 · k∇f (ρt )At k2F =
2
2
3ξ−1
· At + (1 − ξ) · ηe∇f (ρt ) · At F + (1 − ξ)2 − (3ξ−1)
kAt k2F .
2
4
= ηeL
We conjecture that the lower bound on ξt (·) for more
generic cases of C could possibly be improved with a different
analysis.
2
+ 1 · σ1 (At )2 · kAt+1 − At k2F
1 2
+ 1 · σ1 (ρt ) · kAt+1 − At k2F
1 + 128
1
128
1+
1.8
2
1
1+
+1
128
· kAt+1 − At k2F
128
2
1
1+
+1
128
et+1 − At k2
· kξ · A
F
128
2
1
1+
+1
128
· k(ξ − 1)At − ξ · ηe∇f (ρt ) · At k2F
128
2
1
+1
1+
128
(1 − ξ)2 ·
· kAt k2F
64
2
1
+1
1+
128
+
· ξ 2 · ηe2 · k∇f (ρt ) · At k2F
64
=
≤
" 92
0.4
−
kρ? −
ρt k2F
(1 − ξ)2 ·
−
1−
· ξ2
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1 2
) )
128
"
· ηe2 k∇f (ρt )At k2F
η
eµσr (ρ? )
· Dist(At , A? )2
10
!
2
1
1+
+1
128
64
0.9
0.8
0.7
0.6
0.5
(i)
2
≥ ηe
·
1− 1−
k∇f (ρt )At k2F
(1 − ξ)2 ·
+
−
(3ξ−1)2
4
(ii)
·
+
η
eµ
2
1−
1
128
·
2
·
kAt k2F
η
eµσr (ρ? )
kρ? − ρt k2F −
10
!
2
1
1+
+1
128
64
1− 1−
≥ ηe2 k∇f (ρt )At k2F +
2(1−ξ)
3ξ−1
η
eµ
2
2(1−ξ)
3ξ−1
·
1
128
kρ? − ρt k2F −
2
· Dist(At , A? )
!
η
eµσr (ρ? )
10
1+
1
128
64
2
· Dist(At , A? )2
(A16)
255·ξ2
128
0.2
0.3
0.1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
9
η
eµ
2
Finally, we bound
Lemma from [28]:
kρ? − ρt k2F using the following
−
+1
−
· ξ 2 ≥ 1; see also Figure 6 (top panel), and
(ii) is due to the non-negativity of the constant in front of
kAt k2F ; see also Figure 6 (bottom panel).
In the special case where C = A ∈ Cd×r : kAk2F ≤ 1 , as
in QST, the assumption ξ & 0.78 is always satisfied, according
to the following Corollary; the proof is provided in Subsection
A 6:
Corollary 16. If C =
A ∈ Cd×r : kAkF ≤ 1 , then
128
ProjFGD inherently satisfies 129 ≤ ξt (·) ≤ 1, for every t. I.e.,
it guarantees (A5) without assumptions on ξt (·).
1
5
0.4
FIG. 6. Behavior of constants, depending on ξ, in expression
(A16).
· kAt k2F
where (i) is due to the assumption ξ & 0.78 and thus
2
1 2
) +1
(1+ 128
)
64
−
(3ξ−1)2
4
!
(1 ! 9)2 " (1 !
!
+
128
64
0.1
1
2(1!9)
39!1
1
5
0
9
" (1 ! (1 !
+
η
eµ
2
−
9 : 0:78 %
0.2
-0.2
(39!1)2
4
255·ξ
128
≥
+1
1
0.8
0
?
et+1 k2
2e
η ∇f (ρt ) · At , At − A? RA
+ kAt+1 − A
F
t
!
2
1
1+
1.2
0.6
where in the last inequality we substitute ηe; observe that
ηe ≤ 128Lσ11 (ρt ) . Combining this result with (A15), we obtain:
2
1.4
1 2
) +1
(1+ 128
64
=
1.6
!
≤
1
5
kρt − ρt+1 k2F ≤ ηeL
!
η
eL
2
kρt − ρt+1 k2F can be further bounded as follows:
255"9 2
128
η
eL
2
Thus,
(A14)
Lemma 17. For any A, B ∈ Cd×r , we have:
kAA† − BB † k2F ≥ 2 ·
√
2 − 1 · σr (A)2 · Dist(A, B)2 .
In our case, this translates to:
η
eµ
2
kρ? − ρt k2F ≥ ηeµ ·
√
2 − 1 · σr (ρ? ) · Dist(At , A? )2 ,
15
and we can thus conclude to:
?
A? RA
t
et+1 k2
2e
η ∇f (ρt ) · At , At −
+ kAt+1 − A
F
√
2
2
≥ ηe k∇f (ρt )At kF + ηeµ
2 − 1 σr (ρ? ) · Dist(At , A? )2
η
eµσr (ρ? )
10
· Dist(At , A? )2
√
1
2 − 1 − 10
ηeµσr (ρ? ) · Dist(At , A? )2
= ηe2 k∇f (ρt )At k2F +
−
= ηe2 k∇f (ρt )At k2F +
3e
ηµ
10
and
σ1 (∇f (ρt )At ) = σ1 ∇f (ρt )QAt Q†At A
≥ σ1 ∇f (ρt )QAt Q†At · σr (At )
1
≥ σ1 ∇f (ρt )QAt Q†At · 1 − 200
· σr (A? ),
· σr (ρ? ) · Dist(At , A? )2
This completes the proof.
5.
Finally, (v) is due to the facts:
σ1 (∇f (ρt )A? ) = σ1 ∇f (ρt )QA? Q†A? A?
≥ σ1 ∇f (ρt )QA? Q†A? · σr (A? )
Proof of Lemma 15
by the proofof (a variant of)Lemma A.3 in [29]. Thus, for
the term σ1 ∇f (ρt )QA? Q†A? , we have
Proof. We can lower bound ∇f (ρt ), ∆∆† as follows:
E (i) D
E
∇f (ρt ), ∆∆† = Q∆ Q†∆ ∇f (ρt ), ∆∆†
≥ − Tr Q∆ Q†∆ ∇f (ρt )∆∆†
D
σ1 ∇f (ρt )QA? Q†A? ≤
≤
1
σ
σr (A? ) 1
(∇f (ρt )A? )
1
201
σ
σr (A? ) 199 1
(∇f (ρt )At )
†
201σ1 (A? ) 201
≤ 200σ
σ
1 ∇f (ρt )QAt QAt .
(A
)
199
r
?
(ii)
≥ −σ1 Q∆ Q†∆ ∇f (ρt ) Tr(∆∆† )
(iii)
(A18)
≥ − σ1 QAt Q†At ∇f (ρt )
+ σ1 QA? Q†A? ∇f (ρt )
Using (A18) in (A17), we obtain:
!
Dist(At , A? )2 .
(A17)
Note that (i) follows from the fact ∆ = Q∆ Q†∆ ∆ and (ii) follows from |Tr(AB)| ≤ σ1 (A) Tr(B), for PSD matrix B (Von
Neumann’s trace inequality [73]). For the transformation in
(iii), we use that fact that the column space of ∆, Span(∆),
is a subset of Span(At ∪ A? ), as ∆ is a linear combination of
?
.
At and A? RA
t
For the second term in the parenthesis above, we first derive
the following inequalities; their use is apparent later on:
σ1 (∇f (ρt )A? )
(i)
E
≥ − σ1 QAt Q†At ∇f (ρt )
+
≥−
201σ1 (A? ) 201
σ
200σr (A? ) 199 1
QAt Q†At ∇f (ρt )
21·τ (A? )
kQAt Q†At ∇f (ρt )k2
10
?)
where τ (A? ) := σσr1 (A
.
(A? )
We remind that ηb =
!
· Dist(At , A? )2
· Dist(At , A? )2
1
.
†
128(Lσ1 (ρt )+σ1 (QAt QA ∇f (ρt )))
Then,
t
· σ1 QAt Q†At ∇f (ρt ) · Dist(At , A? )2
21·τ (A? )
· ηe · 128L · σ1 (ρt ) σ1 QAt Q†At ∇f (ρt ) · Dist(At , A? )2
≤
10
2
21·τ (A? )
+
· ηe · 128 · σ1 QAt Q†At ∇f (ρt ) · Dist(At , A? )2
10
21·τ (A? )
10
(iii)
≤ σ1 (∇f (ρt )At ) + ∇f (ρt )QAt Q†At
+ σ1 ∇f (ρt )QA? Q†A? σ1 (∆)
(iv)
(A19)
To bound the first term on the right hand side, we observe
that
1
σ (A? )
200 r
σ1 QAt Q†At ∇f (ρt ) ≤
µσr (ρt )
21·τ (A? )
·10
10
σ1 QAt Q†At ∇f (ρt ) ≥
µσr (ρt )
.
21·τ (A? )
·10
10
or
(v)
≤ σ1 (∇f (ρt )At ) +
≤
∇f (ρt ), ∆∆†
we have:
≤ σ1 (∇f (ρt )At ) + σ1 (∇f (ρt )∆)
≤ σ1 (∇f (ρt )At ) + σ1 ∇f (ρt )Q∆ Q†∆ σ1 (∆)
(ii)
≤ σ1 (∇f (ρt )At ) + ∇f (ρt )QAt Q†At
+ σ1 ∇f (ρt )QA? Q†A?
D
200
199
·
1
1
(1− 200
)
·
1
σ
200 1
(∇f (ρt )At )
1
σ1 (∇f (ρt )A? )
+ 200
1
σ1 (∇f (ρt )At ) + 200
σ1 (∇f (ρt )A? ) .
We use this trivial information to obtain (A20) where κ :=
?
where (i) is due to triangle inequality on A? RA
= At −
t
∆, (ii) is due to generalized Cauchy-Schwarz inequality, (iii)
is due to triangle inequality and the fact that the column
span of ∆ can be decomposed into the column span of At
and A? , by construction of ∆, (iv) is due to the assumption
Dist(At , A? ) ≤ γ 0 · σr (A? ) and
σ1 (∆) ≤ Dist(At , A? ) ≤
1 σr (ρ? )
200 σ1 (ρ? )
· σr (A? ) ≤
1
200
· σr (A? ).
L
µ
and τ (ρ) := σσr1 (ρ)
for a rank-r matrix ρ. Combining (A20)
(ρ)
with (A19), we obtain the expression in (A21) where (i) follows from ηe ≤ 128Lσ11 (ρt ) , (ii) is due to Lemma A.3 in [29] and
using the bound Dist(At , A? ) ≤ γ 0 σr (A? ) by the hypothesis
of the lemma, (iii) is due to σr (ρ? ) ≤ 1.1σr (ρt ) by Lemma
2
A.3 in [29], due to the facts σr (ρt )σ1 QAt Q†At ∇f (ρt )
≤
kA†t ∇f (Xt )k2F and (11κτ (ρ? ) ·
21·τ (A? )
10
+ 1) ≤ 12κτ (ρ? ) ·
16
21·τ (A? )
10
· ηe · 128L · σ1 (ρt ) σ1 QAt Q†At ∇f (ρt ) · Dist(At , A? )2
21·τ (A? )
10
≤ max
·128·e
η ·Lσ1 (ρt )·µσr (ρt )
21·τ (A? )
·10
10
128·e
η ·Lσ1 (ρt )·µσr (ρt )
10
≤
21·τ (A? )
10
· Dist(At , A? )2 , ηe
· Dist(At , A? )2 + ηe
21·τ (A? )
10
2
≤
µσr (ρt )
10
(ii)
≤
(iii)
≤
(iii)
≤
· Dist(At , A? )2 + 10κτ (ρt ) ·
µσr (ρt )
10
21·τ (A? )
10
· Dist(At , A? )2 + 11κτ (ρ? ) ·
µσr (ρt )
10
· Dist(At , A? )2 +
12·212
102
µσr (ρt )
10
· Dist(At , A? )2 +
η
e
σ
5 1
≥−
η
e
k∇f (ρt )At k2F
5
+
µσr (ρ? )
10
· Dist(At , A? )2 .
Proof of Corollary 16
et+1 kF ≤ kAt kF + ηb · k∇f (ρt )At kF
kA
≤ kAt kF + ηb · σ1 ∇f (ρt )QAt Q†At · kAt kF
≤ 1 + ηb · σ1 ∇f (ρt )QAt Q†At
≤ (1 +
(A20)
+1 ·
2
· 128 · ηeσ1 QAt Q†At ∇f (ρt ) · Dist(At , A? )2
21·τ (A? )
10
2
· 128 · ηeσ1 QAt Q†At ∇f (ρt ) · (ρ0 )2 σr (ρ? )
11·(ρ0 )2
10
(A21)
For our initialization, we restrict our attention to the full
rank (r = d) case; the case of r < d assumes that there is a
projection step that projects at the same time onto the PSD
cone and C at the same time. In the full rank case, C 0 is
a convex set and includes the full-dimensional PSD cone, as
well as other norm constraints, as described in the main text.
In the particular case of QST, we make no restrictions; [43]
provides an efficient projection procedure that satisfies the
constraints C 0 and holds for any r.
Let us denote ΠC 0 (·) the corresponding projection step,
where all constraints are satisfied simultaneously. Then, the
initialization we propose follows similar motions with that in
[29]: We consider the projection of the weighted negative gradient at 0, i.e., − L1 · ∇f (0), onto C 0 . I.e.,
· ∇f (0) .
(B3)
ρ0 = A0 A†0 = ΠC 0 −1
L
Assuming a first-order oracle model, where we access f only
though function evaluations and gradient calculations, (B3)
provides a cheap way to find an initial point with some approximation guarantees as follows [74]:
1
)
128
where the first inequality follows from the triangle inequality,
the second holds by the property kABkF ≤ σ1 (A) · kBkF ,
and the third follows because the step size is bounded above
1
et+1 ) =
by ηb ≤
. Hence, we get ξ(A
†
et+1 kF
kA
21·τ (A? )
10
(∇f (ρt )At )2
We have
1
2
· 128 · 10κτ (ρt )σ1 QAt Q†At ∇f (ρt ) · Dist(At , A? )2
· κ · τ (ρ? )2 · 128 · ηeσ1 (∇f (ρt )At )2 ·
This completes the proof.
6.
+1 ·
21·τ (A? )
10
and τ (A? )2 = τ (ρ? ). Finally, (iv) follows from sub1
and using Lemma
stituting γ 0 := c · κ1 · τ (ρ1? ) for c = 200
1
A.3 in [29] (due to the factor 200 , all constants above lead to
bounding the term with the constant 15 ).
Thus, we can conclude:
E
2
2
· 128 · 10κτ (ρt )σ1 QAt Q†At ∇f (ρt ) · Dist(At , A? )2
21·τ (A? )
,
10
∇f (ρt ), ∆∆†
21·τ (A? )
10
· σ1 QAt Q†At ∇f (ρt ) · Dist(At , A? )2
(i)
D
128σ1 (∇f (ρt )QAt QA )
t
128
.
129
d×r
Lemma 18. Let
be such that ρ0 = A0 A†0 =
A0 ∈ C
−1
ΠC 0 L · ∇f (0) . Consider the problem in (B2) where f is
assumed to be L-smooth and µ-strongly convex, with optimum
point ρ? such that rank(ρ? ) = r. We apply ProjFGD with A0
as the initial point. Then, in this generic case, A0 satisfies:
≥
Dist(A0 , A? ) ≤ γ 0 · σr (A? ),
where γ 0 =
Appendix B: Initialization
q
1−µ/L
√
τ 2 (A? )
2( 2−1)
p
srank(ρ? ), srank(ρ) =
kρkF
.
σ1 (ρ)
Proof. To show this, we start with:
In this section, we present a specific initialization strategy
for ProjFGD. For completeness, we repeat the definition of the
optimization problem at hand, both in the original space:
minimize
ρ∈Cd×d
f (ρ)
subject to
ρ ∈ C0.
(B1)
−1
∇f (0), ρ0
L
and the factored space:
minimize
A∈Cd×r
f (AA† )
kρ0 − ρ? k2F = kρ? k2F + kρ0 k2F − 2 hρ0 , ρ? i .
(B4)
Recall that ρ0 = A0 A†0 = ΠC 0 −1
· ∇f (0) by assumption,
L
where ΠC 0 (·) is a convex projection. Then, by Lemma 13:
subject to
A ∈ C.
(B2)
− ρ? ≥ hρ0 , ρ0 − ρ? i .
(B5)
Observe that 0 ∈ Cd×d is a feasible point, since it is PSD
and satisfy any common symmetric norm constraints, as the
17
ones considered in this paper. Hence, using strong convexity
of f around 0, we get,
f (ρ? ) −
µ
kρ? k2F ≥ f (0) + h∇f (0), ρ? i
2
(i)
= f (0) + h∇f (0), ρ0 i + h∇f (0), ρ? − ρ0 i
(ii)
≥ f (0) + h∇f (0), ρ0 i + hL · ρ0 , ρ0 − ρ? i .
(B6)
where (i) is by adding and subtracting h∇f (0), ρ0 i, and (ii)
is due to (B5). Further, using the smoothness of f around 0,
we get:
In the case of n = 12, for all algorithms, we use as initial point
f0 ) such that ρ0 = A
e0 A
e† where ρ0 = Π+ (−A∗ (y))
U0 = ΠC (A
0
and Π+ (·) is the projection onto the PSD cone. In the case
of n = 10, we use random initialization. Configurations are
described in the caption of each figure. Table IV contains information regarding total time required for convergence and
quality of solution for some of these cases. Results on almost
pure density states, i.e., r > 1, are provided in Figure 9.
Next, we also provide pseudocode for our approach. Input
arguments for ProjFGD is (i) f grad that specifies the gradient
operator; in our case, we set
1
f (ρ0 ) ≤ f (0) + h∇f (0), ρ0 i +
(i)
≤ f (ρ? ) −
≤ f (ρ0 ) −
µ
kρ? k2F
2
µ
kρ? k2F
2
L
kρ0 k2F
2
+ hL · ρ0 , ρ? i −
+ hL · ρ0 , ρ? i −
L
kρ0 k2F
2
L
kρ0 k2F .
2
where (i) follows from (B6) by upper bounding the quantity f (0) + h∇f (0), ρ0 i, (ii) follows from the assumption that
f (ρ? ) ≤ f (ρ0 ). Hence, rearranging the above terms, we get:
hρ0 , ρ? i ≥ 12 kρ0 k2F +
µ
kρ? k2F .
2L
Combining the above inequality with (B4), we obtain,
q
µ
kρ0 − ρ? kF ≤ 1 − L
· kρ? kF .
Given, A0 such that ρ0 = A0 A†0 and A? such that ρ? = A? A†? ,
we use Lemma 17 from [28] to obtain:
q √
kA0 A†0 − A? A†? kF ≥ 2( 2 − 1) · σr (A? ) · Dist(A0 , A? ).
0 −ρ? kF
· kρ? kF ≤ γ 0 · σr (A? )
Thus: Dist(A0 , A? ) ≤ √ kρ
√
2( 2−1)·σr (A? )
q
p
µ/L
√
where γ 0 = 2(1−
· τ 2 (A? ) · srank(ρ? ).
2−1)
Such initialization, while being simple, introduces further
restrictions on the condition number τ (ρ? ), and the condition number of function f . Finding such simple initializations
with weaker restrictions remains an open problem; however,
as shown in [26, 28, 29], one can devise specific deterministic
initialization for a given application.
In practice, the projection ΠC 0 (·) step might not be easy
to compute, due to the joint involvement of convex sets. A
practical solution would be to sequentially project − L1 ·∇f (0)
onto the individual constraint sets. Let Π+ (·) denote the projection onto the PSD cone. Then, we can consider the ape0 A
e† = Π+ (e
e
proximate point: ρe0 = A
0
ρ0 ). Given A0 , we can
e
perform an additional step: A0 = ΠC A0 , to guarantee that
A0 ∈ C. In the special case of QST, one can use the procedure
in [43].
Appendix C: Additional experiments and
pseudocode
Figures 7-8 show further results regarding the QST problem, where r = 1 and n = 10, 12, respectively. For each case,
we present both the performance in terms of number of iterations needed, as well as what is the cumulative time required.
f_grad = @ ( rho ) - Mt ( y - M ( rho ) ) ;
where M denotes the forward linear operator over Pauli observables, and Mt its adjoint. And, (ii) params, a Matlab structure
that contains several hyperparameters. Here, params contains
.d, the dimension d; .r, the rank of the density matrix; .init,
the choice between random or specific initialization; .Ainit,
the random initial A0 in case params.init = 0; .stepsize
that selects a conservative (theory) versus a practical step
size; .iter, the maximum number of iterations; .tol, the
tolerance for the stopping criterion.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
function [ rhohat , Ahat ] = ProjFGD ( f_grad , params )
options.tol = 10^ -3; d = params.d ; r = params.r ;
% Random initialization
if ( params.init == 0)
Acur = params.Ainit ;
Acur = Acur. / norm ( Acur , ' fro ') ;
rhocur = Acur * Acur ';
rhoprev = rhocur ;
% Use of PROPACK
[¬, S1 , ¬] = lansvd ( Acur , 1 , 'L ' , options ) ;
norm_grad = f_grad ( rhocur ) ;
[¬, S2 , ¬] = lansvd ( norm_grad , 1 , 'L ' , options ) ;
grad0 = f_grad ( zeros (d , d ) ) ;
grad1 = f_grad ( ones (d , d ) ) ;
L_est = 2* norm ( grad0 - grad1 , ' fro ') / d ;
% Our initialization
elseif ( params.init == 1)
% Compute 1/ L * gradf (0)
grad0 = f_grad ( zeros (d , d ) ) ;
grad1 = f_grad ( ones (d , d ) ) ;
L_est = 2* norm ( grad0 - grad1 , ' fro ') / d ;
rhocur = -(1/ L_est ) * grad0 ;
% P_r ( gradf (0) ) using PROPACK
[ Ar , Sr , ¬] = lansvd ( rhocur , r , 'L ' , options ) ;
Acur = Ar * sqrt ( Sr ) ;
Acur = Acur. / norm ( Acur , ' fro ') ;
rhocur = Acur * Acur ';
rhoprev = rhocur ;
S1 = Sr ;
norm_grad = f_grad ( rhocur ) ;
[¬, S2 , ¬] = svds ( norm_grad , 1) ;
end
if ( params.stepsize == 0) % Theory
eta = 1/( 128 * ( L_est * S1 (1 ,1) + S2 (1 ,1) ) ) ;
elseif ( params.stepsize == 1) % More practical
eta = 1/( 10 * L_est * S1 (1 ,1) + S2 (1 ,1) ) ;
end ;
i = 1;
while ( i ≤ params.iters )
f_gradrho = f_grad ( rhocur ) ;
gradA = f_gradrho * Acur ;
Acur = Acur - eta * gradA ;
norm_Acur = norm ( Acur , ' fro ')
if ( norm_Acur > 1)
Acur = Acur. / norm ( Acur , ' fro ') ;
end ;
rhocur = Acur * Acur ';
% Test stopping criterion
if (( i > 1) && ( norm ( rhocur - rhoprev , ' fro ') <
params.tol * norm ( rhocur , ' fro ') ) )
break ;
end
i = i + 1;
rhoprev = rhocur ;
end
rhohat = Acur * Acur ';
Ahat = Acur ;
...
18
[1] J. Altepeter, E. Jeffrey, and P. Kwiat, “Photonic state
tomography,” Advances in Atomic, Molecular, and Optical Physics 52, 105–159 (2005).
[2] J. Zhang, G. Pagano, Hess P. W., A. Kyprianidis,
P. Becker, H. Kaplan, A. V. Gorshkov, Z.-X. Gong,
and Monroe C., “Observation of a many-body dynamical phase transition with a 53-qubit quantum simulator,”
arXiv preprint arXiv:1708.01044 (2017).
[3] “IBM-Q:
Quantum
Computing
Research,”
https://www.research.ibm.com/ibm-q/.
[4] S. Flammia, A. Silberfarb, and C. Caves, “Minimal informationally complete measurements for pure states,”
Foundations of Physics 35, 1985–2006 (2005).
[5] D. Gross, Y.-K. Liu, S. Flammia, S. Becker, and J. Eisert, “Quantum state tomography via compressed sensing,” Physical review letters 105, 150401 (2010).
[6] T. Heinosaari, L. Mazzarella, and M. Wolf, “Quantum
tomography under prior information,” Communications
in Mathematical Physics 318, 355–374 (2013).
[7] C. Baldwin, I. Deutsch, and A. Kalev, “Strictly-complete
measurements for bounded-rank quantum-state tomography,” Physical Review A 93, 052105 (2016).
[8] D. Donoho, “Compressed sensing,” IEEE Transactions
on information theory 52, 1289–1306 (2006).
[9] R. Baraniuk, “Compressive sensing,” IEEE signal processing magazine 24, 118–121 (2007).
[10] B. Recht, M. Fazel,
and P. Parrilo, “Guaranteed
minimum-rank solutions of linear matrix equations via
nuclear norm minimization,” SIAM review 52, 471–501
(2010).
[11] E. Candès and Y. Plan, “Tight oracle inequalities for lowrank matrix recovery from a minimal number of noisy
random measurements,” IEEE Transactions on Information Theory 57, 2342–2359 (2011).
[12] E. Candès and B. Recht, “Exact matrix completion
via convex optimization,” Foundations of Computational
mathematics 9, 717 (2009).
[13] A. Kalev, R. Kosut, and I. Deutsch, “Quantum tomography protocols with positivity are compressed sensing
protocols,” NPJ Quantum Information 1, 15018 (2015).
[14] S. Flammia and Y.-K. Liu, “Direct fidelity estimation
from few Pauli measurements,” Physical Review Letters
106, 230501 (2011).
[15] Y.-K. Liu, “Universal low-rank matrix recovery from
Pauli measurements,” in Advances in Neural Information
Processing Systems (2011) pp. 1638–1646.
[16] E. Kokiopoulou, C. Bekas, and E. Gallopoulos, “Computing smallest singular triplets with implicitly restarted
Lanczos bidiagonalization,” Applied numerical mathematics 49, 39–61 (2004).
[17] J. Baglama and L. Reichel, “Augmented implicitly
restarted Lanczos bidiagonalization methods,” SIAM
Journal on Scientific Computing 27, 19–42 (2005).
[18] J. Baglama and L. Reichel, “Restarted block Lanczos
bidiagonalization methods,” Numerical Algorithms 43,
251–272 (2006).
[19] J. Cullum, R. Willoughby, and M. Lake, “A Lanczos
algorithm for computing singular values and vectors of
large matrices,” SIAM Journal on Scientific and Statistical Computing 4, 197–215 (1983).
[20] M. Hochstenbach, “A Jacobi–Davidson type SVD
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
method,” SIAM Journal on Scientific Computing 23,
606–628 (2001).
L. Wu and A. Stathopoulos, “A preconditioned hybrid
SVD method for accurately computing singular triplets of
large matrices,” SIAM Journal on Scientific Computing
37, S365–S388 (2015).
A. Stathopoulos, E. Romero, and L. Wu, “Extended
functionality and interfaces of the PRIMME eigensolver,”
SIAM News Blog (2017).
H. Haffner, M. Riebe, C. Becher, C. Roos, P. Schmidt,
J. Benhelm, T. Korber, R. Blatt, D. Chek-Al-kar,
and W. Dur, “Scalable multi-particle entanglement of
trapped ions,” Nature 438, 643–646 (2006).
R. Sun and Z.-Q. Luo, “Guaranteed matrix completion
via nonconvex factorization,” in IEEE 56th Annual Symposium on Foundations of Computer Science, FOCS 2015
(2015) pp. 270–289.
T. Zhao, Z. Wang, and H. Liu, “A nonconvex optimization framework for low rank matrix estimation,” in Advances in Neural Information Processing Systems (2015)
pp. 559–567.
Y. Chen and M. Wainwright, “Fast low-rank estimation
by projected gradient descent: General statistical and algorithmic guarantees,” arXiv preprint arXiv:1509.03025
(2015).
P. Jain, C. Jin, S. Kakade, and P. Netrapalli, “Computing matrix squareroot via non convex local search,”
arXiv preprint arXiv:1507.05854 (2015).
S. Tu, R. Boczar, M. Simchowitz, M. Soltanolkotabi, and
B. Recht, “Low-rank solutions of linear matrix equations
via Procrustes flow,” in Proceedings of the 33rd International Conference on International Conference on Machine Learning-Volume 48 (JMLR. org, 2016) pp. 964–
973.
S. Bhojanapalli, A. Kyrillidis, and S. Sanghavi, “Dropping convexity for faster semi-definite optimization,” in
29th Annual Conference on Learning Theory, Proceedings of Machine Learning Research, Vol. 49, edited by
Vitaly Feldman, Alexander Rakhlin, and Ohad Shamir
(PMLR, Columbia University, New York, New York,
USA, 2016) pp. 530–582.
D. Park, A. Kyrillidis, C. Carmanis, and S. Sanghavi,
“Non-square matrix sensing without spurious local minima via the Burer-Monteiro approach,” in Artificial Intelligence and Statistics (2017) pp. 65–74.
R. Ge, J. Lee, and T. Ma, “Matrix completion has no
spurious local minimum,” in Advances in Neural Information Processing Systems (2016) pp. 2973–2981.
D. Park, A. Kyrillidis, S. Bhojanapalli, C. Caramanis,
and S. Sanghavi, “Provable Burer-Monteiro factorization
for a class of norm-constrained matrix problems,” arXiv
preprint arXiv:1606.01316 (2016).
D. Park, A. Kyrillidis, C. Caramanis, and S. Sanghavi, “Finding low-rank solutions to matrix problems, efficiently and provably,” arXiv preprint arXiv:1606.03168
(2016).
Y. Li, Y. Liang, and A. Risteski, “Recovery guarantee of non-negative matrix factorization via alternating
updates,” in Advances in Neural Information Processing
Systems (2016) pp. 4987–4995.
X. Li, Z. Wang, J. Lu, R. Arora, J. Haupt, H. Liu, and
19
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
T. Zhao, “Symmetry, saddle points, and global geometry of nonconvex matrix factorization,” arXiv preprint
arXiv:1612.09296 (2016).
Q. Tran-Dinh and Z. Zhang, “Extended Gauss-Newton
and Gauss-Newton-ADMM algorithms for low-rank matrix optimization,” arXiv preprint arXiv:1606.03358
(2016).
L. Wang, X. Zhang, and Q. Gu, “A universal variance
reduction-based catalyst for nonconvex low-rank matrix
recovery,” arXiv preprint arXiv:1701.02301 (2017).
R. Ge, C. Jin, and Y. Zheng, “No spurious local minima
in nonconvex low rank problems: A unified geometric
analysis,” arXiv preprint arXiv:1704.00708 (2017).
Samuel Burer and Renato D. C. Monteiro, “A nonlinear programming algorithm for solving semidefinite programs via low-rank factorization,” Mathematical Programming 95, 329–357 (2003).
Samuel Burer and Renato D. C. Monteiro, “Local minima
and convergence in low-rank semidefinite programming,”
Mathematical Programming 103, 427–444 (2005).
C. Eckart and G. Young, “The approximation of one matrix by another of lower rank,” Psychometrika 1, 211–218
(1936).
G. Stewart, “On the early history of the singular value
decomposition,” SIAM review 35, 551–566 (1993).
D. Gonçalves, M. Gomes-Ruggiero, and C. Lavor, “A
projected gradient method for optimization over density
matrices,” Optimization Methods and Software 31, 328–
341 (2016).
C. Michelot, “A finite algorithm for finding the projection
of a point onto the canonical simplex of Rn ,” Journal
of Optimization Theory and Applications 50, 195–200
(1986).
J. Duchi, S. Shalev-Shwartz, Y. Singer, and T. Chandra, “Efficient projections onto the `1 -ball for learning
in high dimensions,” in Proceedings of the 25th international conference on Machine learning (ACM, 2008) pp.
272–279.
A. Kyrillidis, S. Becker, V. Cevher, and C. Koch, “Sparse
projections onto the simplex,” in International Conference on Machine Learning (2013) pp. 235–243.
Qinqing Zheng and John Lafferty, “A convergent gradient
descent algorithm for rank minimization and semidefinite
programming from random linear measurements,” in Advances in Neural Information Processing Systems (2015)
pp. 109–117.
J. Řeháček, Z. Hradil, E. Knill, and A. I. Lvovsky, “Diluted maximum-likelihood algorithm for quantum tomography,” Phys. Rev. A 75, 042108 (2007).
Y. S. Teo, J. Řeháček,
and Z. Hradil, “Informationally incomplete quantum tomography,” Quantum Measurements and Quantum Metrology 1 (2013),
10.2478/qmetro-2013-0006.
K. Banaszek, G. D’Ariano, M. Paris, and M. Sacchi,
“Maximum-likelihood estimation of the density matrix,”
Physical Review A 61, 010304 (1999).
M. Paris, G. D’Ariano, and M. Sacchi, “Maximumlikelihood method in quantum estimation,” in AIP Conference Proceedings, Vol. 568 (AIP, 2001) pp. 456–467.
J. Nelder and R. Mead, “A simplex method for function
minimization,” The computer journal 7, 308–313 (1965).
J. Smolin, J. Gambetta, and G. Smith, “Efficient method
for computing the maximum-likelihood quantum state
from measurements with additive Gaussian noise,” Phys-
ical review letters 108, 070502 (2012).
[54] Z. Hou, H.-S. Zhong, Y. Tian, D. Dong, B. Qi, L. Li,
Y. Wang, F. Nori, G.-Y. Xiang, and C.-F. Li, “Full reconstruction of a 14-qubit state within four hours,” New
Journal of Physics 18, 083036 (2016).
[55] Jiangwei Shang, Zhengyun Zhang, and Hui Khoon Ng,
“Superfast maximum-likelihood reconstruction for quantum tomography,” Phys. Rev. A 95, 062336 (2017).
[56] Y. S. Teo, J. Řeháček, and Z. Hradil, “Informationally
incomplete quantum tomography,” Quantum Measurements and Quantum Metrology 1, 57–83 (2013).
[57] E. Bolduc, G. Knee, E. Gauger, and J. Leach, “Projected
gradient descent algorithms for quantum state tomography,” npj Quantum Information 3, 44 (2017).
[58] Y. Nesterov, “A method of solving a convex programming problem with convergence rate O(1/k2 ),” in Soviet
Mathematics Doklady, Vol. 27 (1983) pp. 372–376.
[59] A. Kyrillidis and V. Cevher, “Matrix recipes for hard
thresholding methods,” Journal of mathematical imaging
and vision 48, 235–265 (2014).
[60] S. Becker, V. Cevher, and A. Kyrillidis, “Randomized
low-memory singular value projection,” in 10th International Conference on Sampling Theory and Applications
(Sampta) (2013).
[61] C. Riofrı́o, D. Gross, S.T. Flammia, T. Monz, D. Nigg,
R. Blatt, and J. Eisert, “Experimental quantum compressed sensing for a seven-qubit system,” Nature Communications 8 (2017).
[62] A. Yurtsever, Quoc T. Dinh, and V. Cevher, “A universal primal-dual convex optimization framework,” in Advances in Neural Information Processing Systems (2015)
pp. 3150–3158.
[63] E. Hazan, “Sparse approximate solutions to semidefinite
programs,” Lecture Notes in Computer Science 4957,
306–316 (2008).
[64] J. Sturm, “Using SeDuMi 1.02, a MATLAB toolbox for
optimization over symmetric cones,” Optimization methods and software 11, 625–653 (1999).
[65] R. Tütüncü, K.-C. Toh,
and M. Todd, “Solving
semidefinite-quadratic-linear programs using SDPT3,”
Mathematical programming 95, 189–217 (2003).
[66] Inc. CVX Research, “CVX: Matlab software for disciplined convex programming, version 2.0,” http://cvxr.
com/cvx (2012).
[67] V. Chandrasekaran and M. Jordan, “Computational and
statistical tradeoffs via convex relaxation,” Proceedings
of the National Academy of Sciences 110, E1181–E1190
(2013).
[68] Matthias Christandl and Renato Renner, “Reliable quantum state tomography,” Physical Review Letters 109,
120403 (2012).
[69] Jiangwei Shang, Hui Khoon Ng, Arun Sehrawat, Xikun
Li, and Berthold-Georg Englert, “Optimal error regions
for quantum state estimation,” New Journal of Physics
15, 123026 (2013).
[70] A. Agarwal, S. Negahban, and M. Wainwright, “Fast
global convergence rates of gradient methods for highdimensional statistical recovery,” in Advances in Neural
Information Processing Systems (2010) pp. 37–45.
[71] Y. Chen and S. Sanghavi, “A general framework for highdimensional estimation in the presence of incoherence,”
in Communication, Control, and Computing (Allerton),
2010 48th Annual Allerton Conference on (IEEE, 2010)
pp. 1570–1576.
20
[72] S. Bubeck, “Convex optimization: Algorithms and complexity,” Foundations and Trends® in Machine Learning
8, 231–357 (2015).
[73] L. Mirsky, “A trace inequality of John von Neumann,”
Monatshefte für mathematik 79, 303–306 (1975).
[74] As we show in the experiments section, a random initialization performs well in practice, without requiring
the additional calculations involved in (B3). However, a
random initialization for the constraint case provides no
guarantees whatsoever.
21
Rank r = 1 - Data points m = 6 " d " r
10
0
AccUniPDGrad
ProjFGD
AccUniPDGrad
ProjFGD
0
10
0.4
0.2
0
0
0.5
200
400
600
1
800
1000
20
30
-4
100
0
40
200
300
400
20
40
10
50
-4
60
0
1
250
300
20
40
60
0
AccUniPDGrad
ProjFGD
10 0
10
-1
;!;? kF
kb
k;? kF
10
200
Rank r = 1 - Data points m = 10 " d " r
10
RSVP
SparseApproxSDP
;!;? kF
kb
k;? kF
;!;? kF
kb
k;? kF
-1
0.5
150
Number of iterations
AccUniPDGrad
ProjFGD
10-2
0
100
Number of iterations
10 0
RSVP
SparseApproxSDP
0.2
10 -3
500
0
10 0
-1
0.4
0
1
Rank r = 1 - Data points m = 6 " d " r
10
AccUniPDGrad
ProjFGD
;!;? kF
kb
k;? kF
0.5
0.6
10 -2
Number of iterations
Rank r = 1 - Data points m = 3 " d " r
10
0
Number of iterations
Number of iterations
10
0.2
0.8
-1
RSVP
SparseApproxSDP
-2
10 -2
10 -1
;!;? kF
kb
k;? kF
10
0.4
10 -2
RSVP
SparseApproxSDP
1
10
;!;? kF
kb
k;? kF
0
10
10
0.6
0
Number of iterations
10-2
10 -1
0
-2
;!;? kF
kb
k;? kF
0.6
10 -2
10
RSVP
SparseApproxSDP
1
0.8
;!;? kF
kb
k;? kF
;!;? kF
kb
k;? kF
10 -1
AccUniPDGrad
ProjFGD
0
10-2
RSVP
SparseApproxSDP
1
0.8
;!;? kF
kb
k;? kF
10
-1
;!;? kF
kb
k;? kF
10
Rank r = 1 - Data points m = 10 " d " r
100
;!;? kF
kb
k;? kF
Rank r = 1 - Data points m = 3 " d " r
10 -2
10 -2
Cumulative time (sec)
10-2
0
200
400
10 -3
10-4
2000 4000 6000 8000 10000 12000
600
5000
10000
15000
Cumulative time (sec)
0
Cumulative time (sec)
500
1000
1500
10
5000
-4
2000
10000
15000
Cumulative time (sec)
0
Cumulative time (sec)
1000
2000
3000
Cumulative time (sec)
−ρ? kF
FIG. 7. Convergence performance of algorithms under comparison w.r.t. kbρkρ
vs. (i) the total number of iterations (top)
? kF
and (ii) the total execution time (bottom). First, second and third column corresponds to Csam = 3, 6 and 10, respectively. For
e† where ρ0 = Π+ (−A∗ (y)).
f0 ) such that ρ0 = A
e0 A
all cases, r = 1 (pure state setting) and n = 12. Initial point is U0 = ΠC (A
0
10 0
600
20
40
200
10
0
10 0
10
0
10
20
30
Cumulative time (sec)
60
80
400
AccUniPDGrad
ProjFGD
10 0
0
10
20
40
20
40
60
Cumulative time (sec)
40
50
10 0
80
RSVP
SparseApproxSDP
10 -2
10 -4
10
0
800 1000
AccUniPDGrad
ProjFGD
Cumulative time (sec)
-4
600
Rank r = 1 - Data points m = 10 " d " r
200 400 600 800 10001200
10
30
100
10-2
RSVP
SparseApproxSDP
10 -2
600
400
Number of iterations
Rank r = 1 - Data points m = 6 " d " r
100
;!;? kF
kb
k;? kF
;!;? kF
kb
k;? kF
;!;? kF
kb
k;? kF
-2
Cumulative time (sec)
-4
40
10-2
RSVP
SparseApproxSDP
200
20
200
Number of iterations
-4
Number of iterations
AccUniPDGrad
ProjFGD
10
10 -4
800 1000
10
60
Rank r = 1 - Data points m = 3 " d " r
10-2
600
RSVP
SparseApproxSDP
10 -2
Number of iterations
-4
Number of iterations
100
400
10 0
;!;? kF
kb
k;? kF
0
;!;? kF
kb
k;? kF
800 1000
AccUniPDGrad
ProjFGD
;!;? kF
kb
k;? kF
10
400
Number of iterations
Rank r = 1 - Data points m = 10 " d " r
100
10-2
RSVP
SparseApproxSDP
10 -2
;!;? kF
kb
k;? kF
200
10 0
;!;? kF
kb
k;? kF
10 -2
-4
AccUniPDGrad
ProjFGD
10-2
RSVP
SparseApproxSDP
;!;? kF
kb
k;? kF
;!;? kF
kb
k;? kF
10-2
Rank r = 1 - Data points m = 6 " d " r
100
;!;? kF
kb
k;? kF
AccUniPDGrad
ProjFGD
;!;? kF
kb
k;? kF
Rank r = 1 - Data points m = 3 " d " r
100
500
1000 1500 2000
Cumulative time (sec)
-4
0
20
40
60
80
100
Cumulative time (sec)
−ρ? kF
FIG. 8. Convergence performance of algorithms under comparison w.r.t. kbρkρ
vs. (i) the total number of iterations (top)
? kF
and (ii) the total execution time (bottom). We consider random initialization for all algorithms. First, second and third
column corresponds to Csam = 3, 6 and 10, respectively. For all cases, r = 1 (pure state setting) and n = 10.
22
n = 6, Csam = 3.
Algorithm
kb
ρ−ρ? kF
kρ? kF
RSVP
SparseApproxSDP
AccUniPDGrad
ProjFGD
5.1496e-05
4.6323e-03
4.0388e-05
2.4116e-05
n = 6, Csam = 6.
Total time
kb
ρ−ρ? kF
kρ? kF
0.7848
3.7404
0.3634
0.0599
1.8550e-05
2.2469e-03
2.4064e-05
1.6052e-05
n = 8, Csam = 3.
RSVP
SparseApproxSDP
AccUniPDGrad
ProjFGD
1.5774e-04
4.1639e-03
3.5122e-05
2.4388e-05
5.7347
16.1074
1.1006
0.6918
4.6056e-04
3.6310e-03
3.0456e-05
9.2352e-06
5.2470e-05
2.2011e-03
2.4634e-05
1.5431e-05
379.8635
658.7082
33.3585
13.9547
4.7811e-03
3.1717e-03
8.8050e-05
8.4761e-06
Total time
0.3791
4.3775
0.3311
0.0441
6.6328e-06
1.4776e-03
1.9032e-05
1.1419e-05
0.1203
3.8536
0.4911
0.0446
n = 8, Csam = 10.
3.8649
33.7608
1.8428
0.8994
2.9583e-05
1.7631e-03
1.7719e-05
1.0561e-05
n = 10, Csam = 6.
1.8017e-04
2.1911e-03
1.9931e-05
5.8515e-06
n = 12, Csam = 3.
RSVP
SparseApproxSDP
AccUniPDGrad
ProjFGD
Total time
n = 8, Csam = 6.
n = 10, Csam = 3.
RSVP
SparseApproxSDP
AccUniPDGrad
ProjFGD
n = 6, Csam = 10.
kb
ρ−ρ? kF
kρ? kF
n = 10, Csam = 10.
331.1315
1326.5374
56.9693
19.3982
9.7585e-05
1.7687e-03
4.5022e-05
1.0460e-05
n = 12, Csam = 6.
14029.1525
13635.4238
461.2084
266.8203
1.0843e-03
3.6954e-03
5.2367e-05
4.7399e-06
4.6548
85.0633
3.9440
1.8804
307.9554
2245.2301
88.2965
49.4528
n = 12, Csam = 10.
15028.2836
15041.6235
904.0507
440.7193
5.6169e-04
5.0197e-03
4.5660e-05
1.1871e-05
15067.7249
15051.4497
1759.6698
1159.2885
TABLE IV. Summary of comparison results for reconstruction and efficiency.
As a stopping criterion, we used
−6
kρt+1 − ρt k2/kρ
, where ρi is the estimate at the i-th iteration. Time reported is in seconds. Initial point is
t+1 k2 ≤ 5 · 10
e† where ρ0 = Π+ (−A∗ (y)).
f0 ) such that ρ0 = A
e0 A
U0 = ΠC (A
0
10 -2
400
600
0
100
200
300
400
Number of iterations
-2
1000
800 1000
500
2000
3000
Cumulative time (sec)
10-4
0
200
400
Cumulative time (sec)
600
AccUniPDGrad
ProjFGD
RSVP
10-2
10 0
200
400
600
800 1000
Number of iterations
0
50
100
150
200
Number of iterations
AccUniPDGrad
ProjFGD
RSVP
10-2
SparseApproxSDP
10 -1
10-4
Rank r = 20 - Data points m = 3 " d " r
100
250
10 0
SparseApproxSDP
;!;?;r kF
kb
k;?;r kF
SparseApproxSDP
10 -1
10
200
Number of iterations
10-4
10 0
;!;?;r kF
kb
k;?;r kF
10 -1
Rank r = 20 - Data points m = 3 " d " r
100
;!;?;r kF
kb
k;?;r kF
10-2
SparseApproxSDP
;!;?;r kF
kb
k;?;r kF
;!;?;r kF
kb
k;?;r kF
10 0
AccUniPDGrad
ProjFGD
RSVP
;!;?;r kF
kb
k;?;r kF
10-2
Rank r = 5 - Data points m = 3 " d " r
100
;!;?;r kF
kb
k;?;r kF
AccUniPDGrad
ProjFGD
RSVP
;!;?;r kF
kb
k;?;r kF
Rank r = 5 - Data points m = 3 " d " r
100
10 -1
10-4
200
400
600
800 1000
Number of iterations
0
50
100
150
200
250
Number of iterations
−ρ? kF
FIG. 9. Convergence performance of algorithms under comparison w.r.t. kbρkρ
vs. (i) the total number of iterations (left)
? kF
and (ii) the total execution time (right). The two left plots correspond to the case r = 5 and the two right plots to the case
f0 ) such that ρ0 = A
e0 A
e† where ρ0 = Π+ (−A∗ (y)).
r = 20. In all cases Csam = 3 and n = 10. Initial point is U0 = ΠC (A
0
| 8cs.DS
|
Fooling Views: A New Lower Bound Technique for Distributed
Computations under Congestion
Amir Abboud∗
Keren Censor-Hillel†
Seri Khoury†
Christoph Lenzen‡
arXiv:1711.01623v3 [cs.DS] 13 Dec 2017
December 14, 2017
Abstract
We introduce a novel lower bound technique for distributed graph algorithms under bandwidth limitations. We define the notion of fooling views and exemplify its strength by proving
two new lower bounds for triangle membership in the Congest(B) model:
1. Any 1-round algorithm requires B ≥ c∆ log n for a constant c > 0.
2. If B = 1, even in constant-degree graphs any algorithm must take Ω(log∗ n) rounds.
The implication of the former is the first proven separation between the Local and the Congest
models for deterministic triangle membership. The latter result is the first non-trivial lower
bound on the number of rounds required, even for triangle detection, under limited bandwidth.
All previous known techniques are provably incapable of giving these bounds. We hope that our
approach may pave the way for proving lower bounds for additional problems in various settings
of distributed computing for which previous techniques do not suffice.
∗
IBM Almaden Research Center. [email protected].
Technion, Department of Computer Science, {ckeren,serikhoury}@cs.technion.ac.il. Supported in part by
the Israel Science Foundation (grant 1696/14).
‡
MPI for Informatics, Saarland Informatics Campus, [email protected].
†
1
Introduction
In a group of n players with names of O(log n) bits, how many days does it take for someone to find
out if there are three players that are all friends with each other? All computation is free, and in
the beginning, all players know their friends. Then, on each day, each player can send B = O(log n)
bits privately to each of its friends.
This multi-party communication puzzle is known as the triangle detection problem in the popular Congest model of distributed computing. Its complexity is poorly understood. In the naive
protocol, each player (node) tells all its friends about all its other friends (sends its neighborhood
to every neighbor). This takes a single round in the Local model and O(n) rounds in Congest.
A clever randomized protocol of Izumi and Le Gall [30] provides a solution with O(n2/3 (log n)2/3 )
rounds in Congest, and this is essentially all we know about the problem. Before our work, it
could not be ruled out that the problem can be solved in O(1) rounds, or even a single round, even
with a bandwidth of B = 1!
Open Question 1. What is the round complexity of triangle detection in the Congest model of
distributed computing?
Triangle detection is an extensively studied problem in most models of computation. In the
centralized setting, the best known algorithm involves taking the cube of the adjacency matrix of the
graph. It runs in O(n2.3729 ) time and was found using a complex computer program [39, 62]. If one
wishes to avoid the impractical matrix multiplication, the problem can be solved in O(n3 / log4 n)
time [8, 19, 65]. Other works have designed algorithms for sparse graphs [6], for real-world graphs
(e.g. [60]), for listing all the triangles [12, 29], for approximately counting their number [34], for
weighted variants (e.g. [20]), and much more (an exhaustive list is infeasible). Moreover, conjectures
about the time complexity of triangle detection and of its variants [1–3, 28, 64] are among the
cornerstones of fine-grained complexity (see [63]). Other highly non-trivial algorithms were designed
for it in settings such as: distributed models [15, 16, 21, 22, 25, 30, 32], quantum computing [11, 14,
38,40–42,47,61], and Map-Reduce [4,58]. It is truly remarkable that such a basic problem has lead
to so much research.
From a technical perspective, Open Question 1 is one of the best illustrations of the lack and
necessity of new techniques for proving lower bounds in distributed computing. In this work, we
present a novel lower bound technique, providing a separation between the Local and Congest(B)
models for a problem for which previous techniques are provably incapable of doing so.1 This is the
first progress towards answering Open Question 1 from the lower bounds side.
1.1
Prior lower bound techniques and their limits
To date, there are essentially two techniques for deriving lower bounds for distributed graph algorithms. The first is the indistinguishability technique of Linial [45], which is the main source for
lower bounds in the Local model, where the message size is unrestricted. This technique argues
that any r-round algorithm, regardless of message size, can be seen as a function that maps the rhop neighborhood of a node to its output. Here, the topology of the graph is labeled by the unique
O(log n)-bit node identifiers and any other input provided to the nodes; for randomized algorithms,
we simply give each node an infinite string of unbiased random bits as part of the input.
1
Congest(B) stands for the synchronous model with a bandwidth of B bits. In particular, the standard Local
and Congest models correspond to Congest(∞) and Congest(log n), respectively.
1
This technique has resulted in a large number of locality lower bounds, e.g. [13, 36, 44, 45, 50].
For these problems, it is a long-standing open question whether higher lower bounds can be found
in the Congest(B) model, see e.g. [52]. Note that part of its appeal is that one can entirely
“forget” about the algorithm: for instance, a 2-round coloring algorithm is just interpreted as a
function assigning a color to each possible 2-neighborhood, and a correct algorithm must assign
distinct colors to any pair of neighborhoods that may belong to adjacent nodes in any feasible
input graph; more generally, this gives rise to the so-called r-neighborhood graph, and showing that
r rounds are insufficient for coloring with c colors equates to showing that the chromatic number
of the r-neighborhood graph is larger than c.
Unfortunately, as this technique does not take bandwidth restrictions into account, it cannot
show any separation between the Local and Congest(B) models. Triangle detection is possibly
one of the most extreme examples for this, as it can be solved in a single round in Local but
seems to require nΩ(1) rounds in Congest(log n). Additional examples are symmetry breaking
problems, such as Maximal Independent Set (MIS) and (∆ + 1)-Coloring, upon which we elaborate
in Section 5.
The second tool available for generating distributed lower bounds is the information bottleneck
technique, first introduced implicitly by Peleg and Rubinovich [56]. Here, the idea is to reduce a 2party communication complexity problem (typically set disjointness) to a distributed problem, and
argue that a fast distributed algorithm under limited bandwidth would imply a protocol exchanging
few bits. This approach yields a large number of strong lower bounds for a wide range of global
problems, in which there is no bound on the distance up to which a local change in the input may
affect the output. Examples are, e.g., [17, 26, 43, 49, 59], but a complete list would justify an entire
survey by its own.
Lower bounds based on information bottlenecks can also be proven for local problems. For
instance, Drucker et al. show an nΩ(1) lower bound for detecting k-cliques or k-cycles for any fixed
k > 3 [22]. However, this technique is inherently incapable of proving lower bounds for many
problems. In particular, it completely fails for the problem of triangle detection. As discussed by
Drucker et al., this is because no matter how we divide the nodes of the graph among the two
players, one of them will know about the triangle. One may, in principle, hope for lower bounds
based on multi-party communication complexity or information complexity, but to date no such
result is known.
Intuitively, a lower bound for triangle detection must combine the two techniques. We have
to argue that when a small number of bits is sent, the nodes do not have enough information to
distinguish between a distance-2 neighborhood in which there is a triangle and one in which there
isn’t. Interestingly, Drucker et al. prove that no such technique is possible without breakthroughs in
circuit complexity in the related Congested-Clique model [22], where nodes can send messages
to all other nodes, not only neighbors.2 Indeed, it is still open whether triangle detection can be
solved in no(1) rounds even in this powerful model (it is likely that it is [22]), but now we know that
the lack of super-constant lower bounds can be blamed on our inability to prove “computational
hardness” results in CS: in a similar vein, we do not know whether 3SAT can be solved in polynomial
or even linear time. In contrast, no such barrier is known in the standard Congest(B) model,
where communication is limited to the input graph. This is an embarrassing situation, as we have
a huge gap, but no well-known barrier to blame it on.
2
This result applies to small output, e.g. the triangle detection problem, which is a decision problem. For listing
all triangles in the graph, a tight bound of Ω(n1/3 / log n) holds [21, 30].
2
With a clever usage of Rusza-Szemerdi graphs, Drucker et al. were also able to prove a strong
lower bound for triangle detection [22], under the restriction that each node sends the same
message to all of its neighbors in each round (in a broadcast fashion). More specifically, they
show that a deterministic protocol
in the Clique-Broadcast model (and therefore also Congest√
log
n
Broadcast) requires Ω(n/e
) rounds, and that (essentially) the same lower bound holds for
randomized protocols under the Strong Exponential Time Hypothesis, a popular conjecture about
the time complexity of k-SAT. Unfortunately, even under such conjectures, we do not know how to
get any non-trivial lower bound in the standard Congest(B) model.
Finally, it is worth pointing out a subtlety about the statement that 2-player communication
complexity cannot provide any lower bound for triangle detection. This statement is fully accurate
only under the assumption that nodes initially know the identifiers of their neighbors, as this renders
it trivial to infer from the joint view of two neighbors whether they participate in a triangle. This
assumption is known as KT1 , where KTi means Knowledge of Topology up to distance i (excluding
edges with both endpoints in distance i), as was first defined in [7]. The difference between KT0 , in
which a node knows only its own identifier, and KT1 , in which a node knows also the identifiers of
its neighbors, has been a focus of abundant studies, in particular concerning the message complexity
of distributed algorithms (see, e.g., [7, 23, 33, 51, 54] and references therein).
Note that acquiring knowledge on the neighbors’ identifiers requires no more than sending
O(log n) bits over each edge, so the distinction between KT0 and KT1 is insubstantial for the round
complexity in Congest(B) for B ∈ Ω(log n); KT1 is therefore the default assumption throughout
n
wide parts of the literature. However, in KT0 a lower bound of Ω( log
B ) on the round complexity
of triangle detection follows from a simple counting argument. As, ultimately, the goal is to show
lower bounds of ω(log n), we consider KT1 in this work.
n1−o(1)
1.2
Our contribution
In this paper, we introduce fooling views, a technique for proving lower bounds for distributed
algorithms with congestion. We are able to show the first non-trivial round complexity lower
bounds on triangle detection in KT1 , separating the Local and Congest(B) models:
1. Triangle membership3 in one round requires B ≥ c∆ log n for a constant c > 0 (Section 3).
2. If B = 1, triangle detection requires Ω(log∗ n) rounds, even if ∆ = 2, and even if the size of
the network is constant and n is the size of the namespace (Section 4).
We stress that we do not view our main contribution as the bounds themselves: while the bandwidth
lower bound for single-round algorithms is tight, it hardly comes as a surprise that such algorithms
need to communicate the entire neighborhood. Additionally, we do not believe that with 1-bit
messages, extremely fast triangle detection is possible. Rather, we present a novel technique that
enables to separate the two models, which is infeasible with prior lower bound techniques. We hope
this to be a crucial step towards resolving the large gap between lower and upper bounds, which
in contrast to other models is not justified by, e.g., conditional hardness results.
The basic idea of fooling views is that they combine reasoning about locality with bandwidth
restrictions. Framing this in terms of neighborhood graphs, this would mean to label neighborhoods by the information nodes have initially and the communication the algorithm performs over
3
In the triangle membership problem, every node must indicate whether it participates in a triangle.
3
the edges incident to a node. However, this communication depends on the algorithm and the
communication received in earlier rounds, enforcing more challenging inductive reasoning to prove
multi-round lower bounds.
To capture the intuition for our technique for B = 1, think about a node that receives the same
messages from its neighbors regardless of whether it participates in a triangle or not as a fooled
node. Intuitively, in a triangle {u, w, v} of a given network, if one of the nodes u, v and w is able
to detect the triangle after t rounds of communication, then it may simply inform the other two
nodes about the triangle during round number t + 1. Thus, it is crucial to maintain a perpetual
state of confusion for all nodes involved in the triangle.
However, if the task is to detect whether a specific triple of IDs is connected by a triangle or not,
then the nodes can solve this by simply exchanging only one bit of communication. Accordingly,
our goal is to keep a large subset of the namespace fooled as long as possible. To this end, think of
a triangle {u, v, w} for which none of u, v and w is able to detect the triangle as a fooled triangle.
Our main idea is to show that if there are many fooled triangles after t rounds, then there are
many triangles among them that are fooled after t + 1 rounds as well. In order to express this
intuition, one of our ingredients in the proof is the following extremal combinatorics result by Paul
Erdös [24].
Theorem 1 ( [24], Theorem 1). Any k-uniform hypergraph of n nodes which contains at least
1−k
nk−`
edges, must contain a complete k-partite k-uniform hypergraph such that each part of it is
of size `.
Using this theorem of Erdös, we are able to show that if there are many fooled triangles after t
rounds, then there is a set of nodes such that each triple in the set is a fooled triangle after t + 1
rounds. Blending counting and indistinguishability arguments with this theorem, we can derive
our lower bound for multi-round algorithms.
Our Ω(log∗ n) bound serves as a proof of concept that our technique has the power to break
through the bounds of previous techniques by demonstrating that this is indeed possible. Note
that purely information-theoretic reasoning runs into the obstacle that in the KT1 model, Θ(log n)
bits have already crossed each edge “before the algorithm starts.” Accordingly, we argue that our
approach represents a qualitative improvement over existing techniques. Proving lower bounds
higher than log∗ n requires new ideas, but we are hopeful that combining our technique with a
more sophisticated analysis will lead to much higher lower bounds, both for triangle detection and
for other non-global problems discussed above.
As an additional indication that the proposed technique is of wider applicability, we apply
it to k-cycle detection for k > 3, for showing lower bounds on the bandwidth of optimal-round
algorithms. However, here the information bottleneck technique is applicable again, and we do not
obtain stronger bounds using our fooling views. The details of these constructions are given in
Appendix A; the main body of the paper focuses on triangle detection. We conclude with open
questions in Section 5.
1.3
Further related work
Edge-crossings. The basic topology components for our lower bound in Section 4 are triangles
and 6-cycles. The main hardness that we show for a node in deciding whether it participates in a
triangle or not comes from not knowing the neighbors of its neighbor. That is, a node is unable to
distinguish between two triangles and one 6-cycle, because only difference between these two cases
4
is a single edge crossing, which are two node-disjoint edges for which we swap the endpoints from
{w, x}, {w0 , x0 } to {w, x0 }, {w0 , x}. Edge crossings have previously aided the construction of lower
bounds, such as lower bounds for message complexity of broadcast [7] or of symmetry breaking [51],
as well as lower bounds for proof-labeling schemes [10].
The Congest(1) model. While we view our results for the Congest(1) model more as a proof
of concept for our technique rather than as a bound that attempts to capture the true complexity,
this model has been attracting interest by itself in previous work. It has been shown in [48] that
the cornerstone O(log n)-round algorithms for maximal independent set [5,46] can be made to work
with even with a bandwidth of B = 1, and this problem was studied also in, e.g., [9, 35]. Using the
√
standard framework of reduction from 2-party communication complexity, a Ω( n) lower bound
for the number of rounds required for 4-cycle detection can be directly deduced from [22]. In fact,
all lower bounds obtained using this framework are with respect to the bandwidth B, and therefore
imply lower bounds also for the case of B = 1.
2
Model and Definitions
Our model is a network of n nodes, each having an ID in [N ], for some polynomial N in n. Each
node starts with knowledge of its ID as well as the IDs of its neighbors. This is known as the KT1
model, and differs from the KT0 model in which each node starts only with knowledge of its own
ID. The nodes communicate in synchronous rounds, in which each node can send a B-bit messages
to each of its neighbors.
The model we consider is the Congest(B) model [55], where B is the bandwidth that is given
for each message4 . We will focus on the following problem.
Definition 1. [Triangle Membership]. In the triangle membership problem, each node needs to
detect whether it is a part of a triangle.
We remark that our second lower bound also applies to the triangle detection problem, where it is
sufficient that some node learns that there is a triangle, without being able to tell who participates.
This is also the guarantee given by the sublinear-round algorithm of Izumi and Le Gall [30].
3
A Bandwidth Lower Bound for 1-round Triangle Membership
In this section we show the following theorem.
Theorem 2. The triangle membership problem cannot be solved by a single-round algorithm in
n−2 ∆−1
the Congest(B) model unless B ≥ log(( 2(∆−1)
)
).
Observe that this implies an Ω(∆ log n) lower bound on B for ∆ = O(n1− ), and an Ω(n) lower
bound on B for ∆ = n−2
4 + 1. This lower bound is significantly easier to obtain than the one
in Section 4. This has the advantage of demonstrating the technique in a simpler context before
delving into the proof of Theorem 3.
4
For simplifying the exposition, we will assume that nodes send B bits in each round and cannot send less bits or
remain silent. It is easy to verify that this only affects the constants in our asymptotic notation.
5
n−2 ∆−1
The main line of proof is to show that if the size of the messages is less than log ( 2(∆−1)
)
,
then there are three nodes u, w, v ∈ V , such that if the only neighbors of u are v and w, then
u receives the same messages from v and w during the single communication round regardless of
whether v and w are connected (see Figure 1). This is the standard notion of indistinguishability,
given here under bandwidth restrictions for the first time.
𝑣
𝑚𝑣→𝑢
𝑤
𝑢
𝑣
𝑚𝑤→𝑢
𝑚𝑣→𝑢
𝑤
𝑢
𝑚𝑤→𝑢
Figure 1: The node u receives the same messages regardless of whether v and w are connected.
For simplicity, we assume that each node has exactly ∆ neighbors, except some special node u,
which is fixed in the rest of this section, which has only two neighbors. Denote by N (v) the set
of neighbors of a node v, and let mv→u (S) denote the message sent from v to u during the single
round, given that N (v) = S ∪ {u}.
The following two notions of fooling sets of nodes and fooling nodes are what allows us to
capture indistinguishability in this setting. These are the specific shapes that our notion of fooling
views takes for obtaining this result.
Definition 2. Let v ∈ V \ {u}. A set of nodes S is called a (v, u)-fooling set if there is another
set of nodes S 0 6= S such that mv→u (S) = mv→u (S 0 ).
A node w ∈ V \ {v, u} is called a (v, u)-fooling node if there are two sets of nodes S 6= S 0 , such
that w ∈ S and w ∈
/ S 0 , and mv→u (S) = mv→u (S 0 ).
We denote by Fnodes (v, u) the set of (v, u)-fooling nodes.
Our first step towards proving Theorem 2 is to show that for each v ∈ V \ {u} there are many
(v, u)-fooling nodes.
n−2 ∆−1
Lemma 1 (Many fooling nodes). If B < log(( 2(∆−1)
)
), then for each v ∈ V \ {u} it holds
n−2
that |Fnodes (v, u)| ≥ 2 .
n−2
0
Proof. Assume towards a contradiction that |Fnodes (v, u)| < n−2
2 . Thus, there are k ≥ 2 non
(v, u)-fooling nodes, denote this set of non (v, u)-fooling nodes by N Fnodes (v, u), and denote the
family of sets of size ∆ − 1 over nodes in N Fnodes (v, u) by N Fsets (v, u). It holds that
n−2
∆−1
|N Fnodes (v, u)|
n−2
2
|N Fsets (v, u)| =
≥
≥
∆−1
∆−1
2(∆ − 1)
Observe that v must send a unique message to u on each of these sets, since otherwise, there are
two sets S1 6= S2 ∈ N Fsets (v, u) such that mv→u (S1 ) = mv→u (S2 ), and therefore, by Definition 2,
6
n−2 ∆−1
at least one node in N Fnodes (v, u) is (v, u)-fooling. This implies that B ≥ log(( 2(∆−1)
)
), a
contradiction.
The strength of having many (v, u)-fooling nodes for every node v is that it implies that there
is a node w∗ that is a (v, u)-fooling node for many nodes v.
n−2 ∆−1
Lemma 2. If B < log(( 2(∆−1)
)
), then there is a node w∗ ∈ V and a set of nodes Pw∗ of size
n−2
∗
∗
2 , such that for each v ∈ Pw , it holds that w ∈ Fnodes (v, u).
Proof. For each w, v ∈ V , let Xw,v be an indicator defined as follows:
(
1 if w ∈ Fnodes (v, u)
Xw,v =
0 otherwise
Since
X
X
Xw,v =
w∈V \{u} v∈V \{u}
X
X
Xw,v ,
v∈V \{u} w∈V \{u}
Lemma 1 gives that:
X
X
Xw,v ≥ (n − 1) ·
w∈V \{u} v∈V \{u}
n−2
2
Therefore, there must be a node w∗ and a set of nodes Pw∗ of size
it holds that Xw,v = 1. 5
n−2
2 ,
such that for each v ∈ Pw∗ ,
Lemma 2 is what allows us to prove that u cannot solve triangle membership, as follows.
w∗ |
Proof of Theorem 2: Let w∗ be the node provided by Lemma 2. It holds that |P
≥
∆−1
n−2 ∆−1
n−2 ∆−1
∗
( 2(∆−1) )
. If w sends to u less than log ( 2(∆−1) )
bits, then there are two sets of nodes
0
0
}).
} ⊆ Pw∗ , such that mw∗ →u ({v1 , ..., v∆−1 }) = mw∗ →u ({v10 , ..., v∆−1
{v1 , ..., v∆−1 } =
6 {v10 , ..., v∆−1
∗
∗
∗
0
0
∗
Let v be a node such that v ∈ {v1 , ..., v∆−1 } and v ∈
/ {v1 , ..., v∆−1 }. Since v ∈ Pw∗ , by Definition 2, it holds that there are two sets of nodes S 6= S 0 , such that w∗ ∈ S and w∗ ∈
/ S 0 , and
mv∗ →u (S) = mv∗ →u (S 0 ).
n−2 ∆−1
To summarize, if the size of the messages is less than log ( 2(∆−1)
)
, then there are two nodes
w, v ∈ V \ {u}, such that if u is connected to both v and w, then u receives the same messages
from v and w during the single communication round regardless of whether v and w are connected
(see Figure 1).
4
A Round Lower Bound for Triangle Detection with B = 1
Our main goal in this section is to prove the following theorem.
Theorem 3. Any algorithm for triangle membership in the Congest(1) model requires Ω(log∗ n)
rounds, even for graphs of maximum degree 2.
5
In fact, one can show that there are many such nodes as w∗ , but we do not use this in this section.
7
The hard instances for this setting are as simple as tripartite graphs with degree 2, as follows.
˙ 2 ∪A
˙ 3 , E) | E ⊆ (A1 × A2 ) ∪ (A1 × A3 ) ∪ (A2 × A3 )} be the family of tripartite
Let G = {(V = A1 ∪A
graphs such that for each G ∈ G, it holds that |A1 | = |A2 | = |A3 | = n/3, and for every i ∈ {1, 2, 3},
each node in Ai has exactly one neighbor in each of Aj and Ak , where j 6= k 6= i ∈ {1, 2, 3}.
In a nutshell, we show that there must be 6 nodes that cannot detect whether they constitute a 6cycle or two triangles. Remarkably, this simple intuitive task requires us to develop novel machinery
and make use of known results in extremal graph theory. Our approach has the compelling feature
that – up to one round – in this setting triangle detection is equivalent to triangle membership,
as a node detecting a triangle can infer that it is part of a triangle itself and just needs to inform
its neighbors. Thus, we obtain the same lower bound for the, in general, possibly easier problem
of triangle detection. Moreover, the lower bound depends only on the set of possible IDs, not the
actual number of nodes in the graph.
Theorem 4. Any algorithm for triangle detection in the Congest(1) model requires Ω(log∗ N )
rounds, even for graphs of maximum degree 2 with only 6 nodes, given a namespace of size N .
The main challenge in proving a lower bound for more than one round is that for t > 1 the
communication between the nodes in round t does not depend only on their IDs and their sets of
neighbors (as is the case for the first round), but rather it depends also on their views after t − 1
rounds, i.e., on the set of messages that the nodes receive during each of the first t − 1 rounds. The
view of a node after t − 1 rounds may depend on all the nodes in its (t − 1)-hop neighborhood.
˙ 2 ∪A
˙ 3 , E) ∈ G,
Nevertheless, since the family of graphs G is defined such that, for each G = (A1 ∪A
it holds that each node in each part of G has exactly one neighbor in each of the other two
parts, any cycle in any G ∈ G must be a connected component. Therefore, if there is a triple
(u, w, x) ∈ A1 × A2 × A3 that is connected by a triangle (u, w, x) in G, then the communication
between the nodes u, w and x during any round in G depends only on the nodes u, w and x.
Similarly, if there is a 6-cycle (u1 , w1 , x1 , u2 , w2 , x2 ) in G, then the communication between the
nodes u1 , w1 , x1 , u2 , w2 , x2 during any round in G depends only on the nodes u1 , w1 , x1 , u2 , w2 , x2 .
Our main line of proof is to show that for any algorithm for triangle membership with B =
˙ 2 ∪A
˙ 3 such that u1 receives the same messages during
1, there exist u1 , w1 , x1 , u2 , w2 , x2 ∈ A1 ∪A
∗
the first Ω(log n) rounds in two different scenarios, the first of which is a scenario in which u1
participates in a triangle (u1 , x1 , w1 ), and the second is a scenario in which u1 participates in a
6-cycle (u1 , x1 , w2 , u2 , x2 , w1 ). See Figure 2 for an illustration.
˙ 2 ∪A
˙ 3 , we denote by mtw→u ((u, w, x)) the message sent from
Given three nodes u, w, x ∈ A1 ∪A
w to u during round t, given that u, w and x are connected by a triangle. Similarly, given six nodes
˙ 2 ∪A
˙ 3 , we denote by mtw1 →u1 ((u1 , w1 , x2 , u2 , w2 , x1 )) the message sent
u1 , w1 , x1 , u2 , w2 , x2 ∈ A1 ∪A
from w1 to u1 during round t, given that the nodes u1 , w1 , x2 , u2 , w2 , x1 are connected by a 6-cycle
(u1 , x1 , w2 , u2 , x2 , w1 ).
The structure of the proof for the lower bound on the number of rounds is as follows. In Section 4.1 we present the notion of fooling sets of triangles which shares a similar spirit to Definition 2
and is the basis of our fooling views technique in this section. Next, in section 4.2 we present the
connection to 6-cycles and deduce our main result.
4.1
Fooling sets of triangles
Definition 3. Fix a pair of nodes (u, w) ∈ (Ai × Aj ) ∩ E, for i 6= j, and let X = {x1 , ..., x|X| } ⊆ Ak
for k 6= i, j. A set of triangles {(u, w, x) | x ∈ X} is called a (w, u)t -fooling set of triangles
8
1
2
1
1
2
1
2
1
2
2
1
2
Figure 2: The structures we use are the triangle (u1 , x1 , w1 ) and the 6-cycle (u1 , x1 , w2 , u2 , x2 , w1 ).
if w sends the same message to u during each of the first t rounds, in each of the triangles in
{(u, w, x) | x ∈ X}. Formally, for each 1 ≤ i ≤ t,
miw→u ((u, w, x1 )) = miw→u ((u, w, x2 )) = ... = miw→u ((u, w, x|X| )).
We extend the notion of a fooling set of triangles to the notion of a fooling rectangle of triangles:
Definition 4. Fix a node u ∈ Ai . A rectangle W × X ⊆ Aj × Ak , for j 6= k 6= i, is called a
ut -fooling rectangle of triangles if it satisfies the following properties.
1. For each w ∈ W , it holds that {(u, w, x) | x ∈ X} is a (w, u)t -fooling set of triangles.
2. For each x ∈ X, it holds that {(u, w, x) | w ∈ W } is an (x, u)t -fooling set of triangles.
Finally, we extend the notion of a fooling rectangle of triangles to a fooling cube of triangles:
Definition 5. A cube U × W × X ⊆ A1 × A2 × A3 is called a t-fooling cube of triangles if:
1. For each u ∈ U , it holds that W × X is a ut -fooling rectangle of triangles.
2. For each w ∈ W , it holds that U × X is a wt -fooling rectangle of triangles.
3. For each x ∈ X, it holds that U × W is a xt -fooling rectangle of triangles.
The following observation follows immediately from Definition 5.
Observation 1. A1 × A2 × A3 is a 0-fooling cube of triangles.
Our next step towards proving Theorem 3 is to show
√ that any t-fooling cube of triangles of size
s contains a (t + 1)-fooling cube of triangles of size Ω( log log log s).
Lemma 3. If there is a t-fooling cube of triangles U × W × X ⊆ A1 × A2 × A3 of size |U | = |W | =
|X| = s, then there is a constant
β and a (t + 1)-fooling cube of triangles U 0 × W 0 × X 0 ⊆ U × W × X
√
of size |U 0 | = |W 0 | = |X 0 | = β log log log s.
The key combinatorial ingredient for proving Lemma 3 is the following corollary of Theorem 1.
Corollary 1 (of Theorem 1). For every constant α, for sufficiently large s, any s × s × s boolean
2
cube that contains at least e−1/α s3 entries
√ that are 1, contains a 1-monochromatic subcube, such
that each side of the subcube is of size α log s.
9
To prove Lemma 3, we first prove a weaker claim, which guarantees the existence of a sufficiently
large cube of triangles satisfying only the first property in Definition 5. That is, we prove the
existence of a cube U 0 × W 0 × X 0 ⊆ U × W × X such that for each u√∈ U 0 , it holds that W 0 × X 0
is a ut+1 -fooling rectangle of triangles and |U 0 | = |W 0 | = |X 0 | = Θ( log n). Then, applying this
claim to the 3 different sides of the cube U × W × X gives Lemma 3.
Claim 1. If there is a t-fooling cube of triangles U × W × X ⊆ A1 × A2 × A3 of size |U | =
|W | = |X| = s, then √
there is a constant γ and a cube U 0 × W 0 × X 0 ⊆ U × W × X of size
0
0
0
|U | = |W | = |X | = γ log s, such that for each u ∈ U 0 , it holds that W 0 × X 0 is a ut+1 fooling
rectangle of triangles.
Proof. Fix u ∈ U , and consider the messages that u receive from its neighbors in W ∪ X during
round t + 1. Since B = 1, for each w ∈ W , it holds that there exist s/2 nodes x1 , ..., xs/2 ∈ X
t+1
t+1
such that mt+1
w→u ((u, w, x1 )) = mw→u ((u, w, x2 )) = ... = mw→u ((u, w, xs/2 )). Denote this specific
t+1
t+1 for each (w, x) ∈ W × X as
message by m̃w→u . Therefore, defining an indicator variable Yu,w,x
follows
(
t+1
1 if mt+1
w→u ((u, w, x)) = m̃w→u
t+1
Yu,w,x
=
0 otherwise
gives that
X X
t+1
Yu,w,x
= s2 /2,
w∈W x∈X
which implies
X X
x∈X w∈W
t+1
Yu,w,x
=
X X
t+1
Yu,w,x
= s2 /2.
w∈W x∈X
It follows that there are at least s/4 nodes x1 , ..., xs/4 in X, such that each xi ∈ {x1 , ..., xs/4 }
t+1 = 1.
has a set of nodes P (xi ) ⊆ W of size s/4, such that for each w ∈ P (xi ), it holds that Yu,w,x
i
xi
xi
Furthermore, for each xi , it holds that there are |P (xi )|/2 nodes w1 , ..., w|P (xi )|/2 ∈ P (xi ) such
xi
xi
t+1
1
that mt+1
xi →u ((u, w1 , xi )) = ... = mxi →u ((u, w|P (xi )|/2 , xi )). Denote this specific message by m̃xi →u .
t+1 be an indicator defined as follows:
For each (u, w, x) ∈ U × W × X, let Zu,w,x
(
t+1
t+1
t+1
1 if (mt+1
w→u ((u, w, x)) = m̃w→u ) ∧ (mx→u ((u, w, x)) = m̃x→u )
t+1
Zu,w,x
=
0 otherwise
t+1 is a boolean cube of size s × s × s, and that the above argument implies that
Notice that Zu,w,x
it contains at least s3 /32 entries that are 1.
t+1 contains a 1-monochromatic subcube such that each
Therefore, by Corollary 1, the cube Zu,w,x
p
side of the subcube is of size log s/ log 32. Denote this subcube by U 0 × W 0 × X 0 , and the claim
follows.
Proof of Lemma 3: To finish the proof of Lemma 3, we apply Claim 1 on each of the three sides
of the t-fooling cube of triangles U × W × X, and we deduce that there is a constant
√ β and a (t + 1)fooling cube of triangles U 0 ×W 0 ×X 0 ⊆ U ×W ×X of size |U 0 | = |W 0 | = |X 0 | = β log log log s.
10
What we have so far, by Observation 1 and Lemma 3, is that there are sufficiently many triangles
in which the nodes receive the same messages from their neighbors. What remains is to actually
capture the two different scenarios of a node participating in a triangle or not, while having the
node keep receiving the same messages in both. For this we next discuss 6-cycles.
4.2
Triangles and 6-Cycles
Having defined fooling structures of triangles in the previous section, our final step towards proving
Theorem 3 is presenting a connection between fooling structures of triangles and 6-cycles. We start
with the following definitions.
Definition 6. Let (u1 , w1 , x1 ), (u2 , w2 , x2 ) ∈ Ai × Aj × Ak be two disjoint triples, for i 6= j 6= k ∈
{1, 2, 3}. The triangle (u1 , w1 , x1 ) and the 6-cycle (u1 , w1 , x2 , u2 , w2 , x1 ) are called (3 ↔ 6, u1 )t fooling, if u1 receives the same messages from w1 and x1 during each of the first t rounds, in the
triangle (u1 , w1 , x1 ) and in the 6-cycle (u1 , w1 , x2 , u2 , w2 , x1 ). Formally, for each 1 ≤ i ≤ t,
miw1 →u1 ((u1 , w1 , x1 )) = miw1 →u1 ((u1 , w1 , x2 , u2 , w2 , x1 ))
∧mix1 →u1 ((u1 , w1 , x1 )) = mix1 →u1 ((u1 , w1 , x2 , u2 , w2 , x1 )).
To capture the connection between triangles and 6-cycles, we extend the notion of a t-fooling
cube of triangles into the notion of a (3 ↔ 6)t -fooling cube.
Definition 7. A cube U × W × X ⊆ A1 × A2 × A3 is called a (3 ↔ 6)t -fooling cube, if:
1. The cube U × W × X is a t-fooling cube of triangles (see Definition 5).
2. For each pair of disjoint triples (u1 , w1 , x1 ), (u2 , w2 , x2 ) ⊆ U × W × X it holds that:
(a) The triangle (u1 , w1 , x1 ) and the 6-cycle (u1 , w1 , x2 , u2 , w2 , x1 ) are (3 ↔ 6, u1 )t -fooling
(see Definition 6).
(b) The triangle (u1 , w1 , x1 ) and the 6-cycle (u1 , w1 , x1 , u2 , w2 , x2 ) are (3 ↔ 6, w1 )t fooling.
(c) The triangle (u1 , w1 , x1 ) and the 6-cycle (u1 , w2 , x2 , u2 , w1 , x1 ) are (3 ↔ 6, x1 )t -fooling.
∗
Our goal next is to show that there is a (3 ↔ 6)Ω(log n) -fooling cube U × W × X of size
|U | = |W | = |X| ≥ 2. The following observation follows immediately from definition 7.
Observation 2. A1 × A2 × A3 is a (3 ↔ 6)0 -fooling cube of size n/3.
In general, in life, it is good to know that there is a (3 ↔ 6)0 -fooling cube somewhere in the
wild. However, as we are interested in a lower bound on the number of rounds, we need to show
that the amazing (3 ↔ 6)0 -fooling cube contains sufficiently large sets of fooling cubes during many
rounds. For this, we need to prove the following lemma.
Lemma 4. If there is a (3 ↔ 6)t -fooling cube U ×W ×X ⊆ A1 ×A2 ×A3 of size |U | = |W | = |X| = s,
t+1
0
0
0
0
0
0
then
√ there is a (3 ↔ 6) -fooling cube U × W × X ⊆ U × W × X of size |U | = |W | = |X | =
Ω( log log log s).
11
Proof. By the first property of Definition 7, the cube U × W × X is a t-fooling cube of triangles.
Therefore, by Lemma 3, there
U 0 × W 0 × X 0 ⊆ U × W × X of
√ is a (t + 1)-fooling cube of triangles
0
0
0
0
0
size |U | = |W | = |X | = Ω( log log log s). We show that U × W × X 0 is also a (3 ↔ 6)t+1 -fooling
cube. That is, we show that each pair of disjoint triples (u1 , w1 , x1 ), (u2 , w2 , x2 ) ∈ U 0 × W 0 × X 0
satisfies Properties 2a, 2b and 2c of Definition 7.
We start with Property 2a. That is, we first prove that for each pair of disjoint triples
(u1 , w1 , x1 ), (u2 , w2 , x2 ) ∈ U 0 × W 0 × X 0 it holds that the triangle (u1 , w1 , x1 ) and the 6-cycle
(u1 , w1 , x2 , u2 , w2 , x1 ) are (3 ↔ 6, u1 )t+1 -fooling. Observe that since U 0 × W 0 × X 0 is a (3 ↔ 6)t fooling cube, by Property 2a of Definition 7, it holds that for each 1 ≤ i ≤ t,
miw1 →u1 ((u1 , w1 , x1 )) = miw1 →u1 ((u1 , w1 , x2 , u2 , w2 , x1 ))
∧mix1 →u1 ((u1 , w1 , x1 )) = mix1 →u1 ((u1 , w1 , x2 , u2 , w2 , x1 )).
Therefore, in order to show that the triangle (u1 , w1 , x1 ) and the 6-cycle (u1 , w1 , x2 , u2 , w2 , x1 ) are
(3 ↔ 6, u1 )t+1 -fooling, it remains to show that the above holds also for t + 1. That is, we need to
show that
t+1
mt+1
w1 →u1 ((u1 , w1 , x1 )) = mw1 →u1 ((u1 , w1 , x2 , u2 , w2 , x1 ))
t+1
∧mt+1
x1 →u1 ((u1 , w1 , x1 )) = mx1 →u1 ((u1 , w1 , x2 , u2 , w2 , x1 )).
Observe that since U 0 × W 0 × X 0 is a (t + 1)-fooling cube of triangles, it holds that
t+1
mt+1
w1 →u1 ((u1 , w1 , x1 )) = mw1 →u1 ((u1 , w1 , x2 ))
∧mt+1
x1 →u1 ((u1 , w1 , x1 ))
=
mt+1
x1 →u1 ((u1 , w2 , x1 )).
(1)
(2)
Furthermore, Since U 0 × W 0 × X 0 is also a (3 ↔ 6)t -fooling cube, by Property 2b of Definition 7,
for the two triples (u1 , w1 , x2 ), (u2 , w2 , x1 ), it holds that for each 1 ≤ i ≤ t,
miu1 →w1 ((u1 , w1 , x2 )) = miu1 →w1 ((u1 , w1 , x2 , u2 , w2 , x1 ))
∧mix2 →w1 ((u1 , w1 , x2 )) = mix2 →w1 ((u1 , w1 , x2 , u2 , w2 , x1 )),
which means that w1 has the same view after t rounds both in the case that it participates in a
triangle ((u1 , w1 , x2 )), and in the case that it participates in a 6-cycle ((u1 , w1 , x2 , u2 , w2 , x1 )). This,
in turn, implies that it sends the same message to u1 during round t + 1, in these two scenarios,
that is:
t+1
mt+1
w1 →u1 ((u1 , w1 , x2 )) = mw1 →u1 ((u1 , w1 , x2 , u2 , w2 , x1 )).
Combining this with Equation (1), gives that
t+1
mt+1
w1 →u1 ((u1 , w1 , x1 )) = mw1 →u1 ((u1 , w1 , x2 , u2 , w2 , x1 )).
Similarly, by Property 2c of Definition 7, for the two triples (u1 , w2 , x1 ), (u2 , w1 , x2 ), it holds
that for each 1 ≤ i ≤ t,
miu1 →x1 ((u1 , w2 , x1 )) = miu1 →x1 ((u1 , w1 , x2 , u2 , w2 , x1 ))
∧miw2 →x1 ((u1 , w2 , x1 )) = miw2 →x1 ((u1 , w1 , x2 , u2 , w2 , x1 )),
12
which implies that
t+1
mt+1
x1 →u1 ((u1 , w2 , x1 )) = mx1 →u1 ((u1 , w1 , x2 , u2 , w2 , x1 )).
Combining this with Equation (2), gives that
t+1
mt+1
x1 →u1 ((u1 , w1 , x1 )) = mx1 →u1 ((u1 , w1 , x2 , u2 , w2 , x1 )),
which completes the proof that for each pair of disjoint triples (u1 , w1 , x1 ), (u2 , w2 , x2 ) ∈ U 0 ×W 0 ×X 0
it holds that the triangle (u1 , w1 , x1 ) and the 6-cycle (u1 , w1 , x2 , u2 , w2 , x1 ) are (3 ↔ 6, u1 )t+1 fooling, i.e., the cube U 0 × W 0 × X 0 satisfies Property 2a of Definition 7. By symmetric arguments,
U 0 × W 0 × X 0 also satisfies Properties 2b and 2c of Definition 7.
∗
Proof of Theorem 3: Observe that it is sufficient to prove that there is a (3 ↔ 6)Ω(log n) -fooling
cube U × W × X ⊆ A1 × A2 × A3 of size |U | = |W | = |X| ≥ 2. By Observation 2, A1 × A2 × A3 is a
(3 ↔ 6)0 -fooling cube of size n/3. Furthermore, by Lemma 4, for any t ≥ 0, if there is a (3 ↔ 6)t fooling cube U t × W t × X t of size |U t | = |W t | = |X t | = s, then there is √
a (3 ↔ 6)t+1 -fooling cube
U t+1 ×W t+1 ×X t+1 ⊆ U t ×W t ×X t of size |U t+1 | = |W t+1 | = |X t+1 | = Ω( log log log s). Therefore,
∗
applying Lemma 4 repeatedly Ω(log∗ n) times implies that there is a (3 ↔ 6)Ω(log n) -fooling cube
U × W × X ⊆ A1 × A2 × A3 of size |U | = |W | = |X| ≥ 2.
5
Discussion and Open Questions
Being a first-step type of contribution, this work would not be complete without pointing out many
additional open questions for which our fooling views technique is a possible candidate as the road
for making progress.
First, we raise the question of whether our specification of the triangle membership problem has
an inherently different complexity than that of triangle detection. The latter is the standard way
of phrasing decision problems in distributed computing (everyone outputs NO for a no instance,
someone outputs YES for a yes instance). We believe that the bound we give for single-round
algorithms in Section 3 should hold also for triangle detection, but this seems to require a deeper
technical analysis. Our lower bound for the number of rounds in Congest(1) does hold also for
triangle detection as explained in Section 4. Note that the sublinear algorithm of [30] solves triangle
detection but it is not clear how to make it solve triangle membership. A closely-related problem
is that of triangle listing, where all triangles need to be output. The work of [30] also gives the first
sublinear algorithm for triangle listing, completing in O(n3/4 log n) rounds in the Congest model,
as well as an Ω(n1/3 / log n) lower bound (see also [53]).
Open Question 2. Do the triangle membership, triangle detection, and triangle listing problems
have different complexities in the Congest model?
Our results in this paper are for deterministic algorithms, but we do not see a technical obstacle
in making them work for randomized algorithms as well.
Open Question 3. Is the deterministic complexity of triangle membership/detection strictly larger
than that of its randomized complexity in the Congest model?
13
Section 3 gives a tight bound on the bandwidth required for an optimal-round algorithm. In
Appendix A we address the bandwidth of optimal-round algorithms for k-cycles. This question can
be asked about further problems, even those for which we do not know yet what the exact round
complexity is, such as various symmetry breaking problems.
Open Question 4. What is the bandwidth complexity of optimal-round distributed algorithms for
various problems?
Whether gaps between the Local and Congest models occur for symmetry breaking problems
is a central open question, which we hope our technique can shed light upon. Prime examples are
MIS and (∆ + 1)-Coloring, but it is not known whether there is indeed a gap. The reason that
reductions from 2-party communication problems are provably incapable of proving lower bounds
for these problems is that any partial solution that is obtained by a greedy algorithm is extendable
into a valid solution for the entire graph, which means that one player can solve the problem for its
set of nodes and deliver only the state of nodes on the boundary to the other player, for completing
the task. Another way to see why arguing about communication only will not suffice here is to
notice that simulating the sequential greedy algorithm requires in fact very little communication
in total, despite taking many rounds.
Open Question 5. For various symmetry breaking problems, is the complexity in the Congest
model strictly higher than its counterpart in the Local model?
Acknowledgements: We are grateful to Michal Dory, Eyal Kushilevitz, and Merav Parter for
stimulating discussions. We are also grateful to Ivan Rapaport, Eric Remila, and Nicolas Schabanel,
a point that was made by them helped in significantly simplifying Section 3.
References
[1] A. Abboud, A. Backurs, and V. V. Williams. If the current clique algorithms are optimal, so
is Valiant’s parser. In FOCS, pages 98–117, 2015.
[2] A. Abboud and V. Vassilevska Williams. Popular conjectures imply strong lower bounds for
dynamic problems. In FOCS, pages 434–443, 2014.
[3] A. Abboud, V. Vassilevska Williams, and H. Yu. Matching triangles and basing hardness on
an extremely popular conjecture. In STOC, pages 41–50, 2015.
[4] F. N. Afrati, D. Fotakis, and J. D. Ullman. Enumerating subgraph instances using mapreduce. In Data Engineering (ICDE), 2013 IEEE 29th International Conference on, pages
62–73. IEEE, 2013.
[5] N. Alon, L. Babai, and A. Itai. A fast and simple randomized parallel algorithm for the
maximal independent set problem. J. Algorithms, 7(4):567–583, 1986.
[6] N. Alon, R. Yuster, and U. Zwick. Color-coding. Journal of the ACM (JACM), 42(4):844–856,
1995.
[7] B. Awerbuch, O. Goldreich, D. Peleg, and R. Vainish. A trade-off between information and
communication in broadcast protocols. J. ACM, 37(2):238–256, 1990.
14
[8] N. Bansal and R. Williams. Regularity lemmas and combinatorial algorithms. Theory of
Computing, 8(1):69–94, 2012.
[9] A. Bar-Noy, J. Naor, and M. Naor. One-bit algorithms. Distributed Computing, 4:3–8, 1990.
[10] M. Baruch, P. Fraigniaud, and B. Patt-Shamir. Randomized proof-labeling schemes. In Proceedings of the 2015 ACM Symposium on Principles of Distributed Computing (PODC), pages
315–324, 2015.
[11] A. Belovs. Span programs for functions with constant-sized 1-certificates. In STOC, pages
77–84. ACM, 2012.
[12] A. Björklund, R. Pagh, V. V. Williams, and U. Zwick. Listing triangles. In International
Colloquium on Automata, Languages, and Programming, pages 223–234. Springer, 2014.
[13] S. Brandt, O. Fischer, J. Hirvonen, B. Keller, T. Lempiäinen, J. Rybicki, J. Suomela, and
J. Uitto. A lower bound for the distributed lovász local lemma. In STOC, pages 479–488,
2016.
[14] H. Buhrman, C. Durr, M. Heiligman, P. Hoyer, F. Magniez, M. Santha, and R. De Wolf.
Quantum algorithms for element distinctness. In CCC, pages 131–137. IEEE, 2001.
[15] K. Censor-Hillel, E. Fischer, G. Schwartzman, and Y. Vasudev. Fast distributed algorithms
for testing graph properties. In DISC, pages 43–56. Springer, 2016.
[16] K. Censor-Hillel, P. Kaski, J. H. Korhonen, C. Lenzen, A. Paz, and J. Suomela. Algebraic
methods in the congested clique. In PODC, pages 143–152. ACM, 2015.
[17] K. Censor-Hillel, T. Kavitha, A. Paz, and A. Yehudayoff. Distributed construction of purely
additive spanners. In Proceedings of the 30th International Symposium on Distributed Computing (DISC), Paris, France, September 27-29, 2016, pages 129–142, 2016.
[18] K. Censor-Hillel, S. Khoury, and A. Paz. Quadratic and near-quadratic lower bounds for the
CONGEST model. DISC 2017. Also in CoRR, abs/1705.05646, 2017.
[19] T. M. Chan. Speeding up the four russians algorithm by about one more logarithmic factor.
In SODA, pages 212–217, 2015.
[20] A. Czumaj and A. Lingas. Finding a heaviest vertex-weighted triangle is not harder than
matrix multiplication. SIAM Journal on Computing, 39(2):431–444, 2009.
[21] D. Dolev, C. Lenzen, and S. Peled. ” tri, tri again”: Finding triangles and small subgraphs in
a distributed setting. In DISC, pages 195–209. Springer, 2012.
[22] A. Drucker, F. Kuhn, and R. Oshman. On the power of the congested clique model. In
Proceedings of the ACM Symposium on Principles of Distributed Computing (PODC), pages
367–376, 2014.
[23] M. Elkin. A simple deterministic distributed MST algorithm, with near-optimal time and
message complexities. In PODC, pages 157–163. ACM, 2017.
15
[24] P. Erdös. On extremal problems of graphs and generalized graphs. Israel Journal of Mathematics, 2(3):183–190, Sep 1964.
[25] P. Fraigniaud, I. Rapaport, V. Salo, and I. Todinca. Distributed testing of excluded subgraphs.
In DISC, pages 342–356. Springer, 2016.
[26] S. Frischknecht, S. Holzer, and R. Wattenhofer. Networks cannot compute their diameter
in sublinear time. In Proceedings of the Twenty-Third Annual ACM-SIAM Symposium on
Discrete Algorithms (SODA), pages 1150–1162, 2012.
[27] J. Håstad and A. Wigderson. The randomized communication complexity of set disjointness.
Theory of Computing, 3(1):211–219, 2007.
[28] M. Henzinger, S. Krinninger, D. Nanongkai, and T. Saranurak. Unifying and strengthening
hardness for dynamic problems via the online matrix-vector multiplication conjecture. In
STOC, pages 21–30. ACM, 2015.
[29] A. Itai and M. Rodeh. Finding a minimum circuit in a graph. SIAM Journal on Computing,
7(4):413–423, 1978.
[30] T. Izumi and F. L. Gall. Triangle finding and listing in CONGEST networks. In Proceedings of
the ACM Symposium on Principles of Distributed Computing, (PODC), pages 381–389, 2017.
[31] S. Jukna. Extremal Combinatorics - With Applications in Computer Science. Texts in Theoretical Computer Science. An EATCS Series. Springer, 2001.
[32] J. Kari, M. Matamala, I. Rapaport, and V. Salo. Solving the induced subgraph problem in the
randomized multiparty simultaneous messages model. In SIROCCO, pages 370–384. Springer,
2015.
[33] V. King, S. Kutten, and M. Thorup. Construction and impromptu repair of an MST in a
distributed network with o(m) communication. In PODC, pages 71–80. ACM, 2015.
[34] M. N. Kolountzakis, G. L. Miller, R. Peng, and C. E. Tsourakakis. Efficient triangle counting
in large graphs via degree-based vertex partitioning. Internet Mathematics, 8(1-2):161–185,
2012.
√
[35] K. Kothapalli, C. Scheideler, M. Onus, and C. Schindelhauer. Distributed coloring in Õ( log n)
bit rounds. In Proceedings of the 20th International Parallel and Distributed Processing Symposium (IPDPS), 2006.
[36] F. Kuhn, T. Moscibroda, and R. Wattenhofer. Local computation: Lower and upper bounds.
J. ACM, 63(2):17, 2016.
[37] E. Kushilevitz and N. Nisan. Communication complexity. In Cambridge University Press,
1997.
[38] F. Le Gall. Improved quantum algorithm for triangle finding via combinatorial arguments. In
FOCS, pages 216–225. IEEE, 2014.
16
[39] F. Le Gall. Powers of tensors and fast matrix multiplication. In Proceedings of the 39th
international symposium on symbolic and algebraic computation, pages 296–303. ACM, 2014.
[40] F. Le Gall and S. Nakajima. Multiparty quantum communication complexity of triangle
finding. In TQC, to appear, 2017.
[41] F. Le Gall and S. Nakajima. Quantum algorithm for triangle finding in sparse graphs. Algorithmica, 79(3):941–959, 2017.
[42] T. Lee, F. Magniez, and M. Santha. Improved quantum query algorithms for triangle finding
and associativity testing. In SODA, pages 1486–1502. SIAM, 2013.
[43] C. Lenzen and B. Patt-Shamir. Improved distributed steiner forest construction. In PODC,
pages 262–271, 2014.
[44] N. Linial. Distributive graph algorithms global solutions from local data. In Proceedings of the
28th Annual Symposium on Foundations of Computer Science, FOCS, pages 331–335, 1987.
[45] N. Linial. Locality in distributed graph algorithms. SIAM J. Comput., 21(1):193–201, 1992.
[46] M. Luby. A simple parallel algorithm for the maximal independent set problem. SIAM J.
Comput., 15(4):1036–1053, 1986.
[47] F. Magniez, M. Santha, and M. Szegedy. Quantum algorithms for the triangle problem. SIAM
Journal on Computing, 37(2):413–424, 2007.
[48] Y. Métivier, J. M. Robson, N. Saheb-Djahromi, and A. Zemmari. An optimal bit complexity
randomized distributed MIS algorithm. Distributed Computing, 23(5-6):331–340, 2011.
[49] D. Nanongkai, A. D. Sarma, and G. Pandurangan. A tight unconditional lower bound on
distributed randomwalk computation. In PODC, pages 257–266. ACM, 2011.
[50] M. Naor. A lower bound on probabilistic algorithms for distributive ring coloring. SIAM J.
Discrete Math., 4(3):409–412, 1991.
[51] S. Pai, G. Pandurangan, S. V. Pemmaraju, T. Riaz, and P. Robinson. Symmetry breaking in
the congest model: Time- and message-efficient algorithms for ruling sets. In DISC 2017. Also
in CoRR, abs/1705.07861, 2017.
[52] S. Pai, G. Pandurangan, S. V. Pemmaraju, T. Riaz, and P. Robinson. Symmetry breaking in
the congest model: Time-and message-efficient algorithms for ruling sets. In PODC, to appear,
2017.
[53] G. Pandurangan, P. Robinson, and M. Scquizzato. Tight bounds for distributed graph computations. CoRR, abs/1602.08481, 2016.
[54] G. Pandurangan, P. Robinson, and M. Scquizzato. A time- and message-optimal distributed
algorithm for minimum spanning trees. In STOC, pages 743–756. ACM, 2017.
[55] D. Peleg. Distributed Computing: A Locality-Sensitive Approach. Society for Industrial and
Applied Mathematics, 2000.
17
[56] D. Peleg and V. Rubinovich. A near-tight lower bound on the time complexity of distributed
minimum-weight spanning tree construction. SIAM J. Comput., 30(5):1427–1442, 2000.
[57] A. A. Razborov. On the distributional complexity of disjointness. Theor. Comput. Sci.,
106(2):385–390, 1992.
[58] A. D. Sarma, F. N. Afrati, S. Salihoglu, and J. D. Ullman. Upper and lower bounds on the
cost of a map-reduce computation. In VLDB, volume 6, pages 277–288. VLDB Endowment,
2013.
[59] A. D. Sarma, S. Holzer, L. Kor, A. Korman, D. Nanongkai, G. Pandurangan, D. Peleg, and
R. Wattenhofer. Distributed verification and hardness of distributed approximation. SIAM J.
Comput., 41(5):1235–1265, 2012.
[60] T. Schank and D. Wagner. Finding, counting and listing all triangles in large graphs, an
experimental study. In WEA, pages 606–609. Springer, 2005.
[61] M. Szegedy. On the quantum query complexity of detecting triangles in graphs. arXiv preprint
quant-ph/0310107, 2003.
[62] V. V. Williams. Multiplying matrices faster than coppersmith-winograd. In STOC, pages
887–898. ACM, 2012.
[63] V. V. Williams. Hardness of easy problems: basing hardness on popular conjectures such as the
strong exponential time hypothesis (invited talk). In LIPIcs-Leibniz International Proceedings
in Informatics, volume 43, 2015.
[64] V. V. Williams and R. Williams. Subcubic equivalences between path, matrix and triangle
problems. In FOCS, pages 645–654, 2010.
[65] H. Yu. An improved combinatorial algorithm for boolean matrix multiplication. In ICALP,
pages 1094–1105. Springer, 2015.
A
k-Cycle Membership for k ≥ 4
An immediate question is whether we can get lower bounds on the bandwidth for additional roundoptimal algorithms. We show here how to generalize our lower bound technique to apply for
detecting membership in larger cycles. However, curiously, as we show in Section A.1 for the sake
of comparison, for cycles that are larger than 3, the standard approach of reductions from 2-party
communication complexity problems allows for stronger lower bounds.
An optimal-round algorithm for solving k-cycle membership completes within exactly b(k−1)/2c
rounds, by a simple (standard) indistinguishability argument (this is even with unlimited bandwidth). This is because after t rounds of communication a node may have information only about
nodes at distance at most t + 1 from it. Therefore, after b(k − 1)/2c − 1 rounds of communication,
a node cannot distinguish whether it participates in a k-cycle or not.
To see how we generalize our technique, observe first that the lower bound given in Theorem 2
holds even when a specific node u is given as input to all the nodes, and only u needs to solve the
triangle membership problem. This is helpful in extending our lower bound to the case of k-cycles
when k ≥ 4, and hence we define this problem formally.
18
(
)/
(
)/
Figure 3: Constructing G0 out of G.
Definition 8. (Fixed-Node Triangle Membership). In the Fixed-Node Triangle Membership problem, all nodes are given the identity of a specific node u, and node u needs to detect whether it is
a part of a triangle.
The proof of Theorem 2 actually proves the following theorem.
Theorem 5. The fixed-node triangle membership problem cannot be solved by a single-round algon−2 ∆−1
rithm in the Congest(B) model unless B ≥ log(( 2(∆−1)
)
).
We formally extend the membership problem to larger cycles, as follows.
Definition 9. (k-Cycle Membership). In the k-Cycle Membership problem, each node needs to
detect whether it is a part of a k-cycle.
Now, Theorem 5 can be used to prove the following.
Theorem 6. Let k ∈ O(n1− ), for some constant 0 < < 1. The k-cycle membership problem
cannot be solved by an optimal b(k − 1)/2c-round algorithm in the Congest(B) model unless
n−2 ∆−1
B ≥ c log(( 2(∆−1)
)
), for some constant c ≥ 0.
We prove Theorem 6 by showing a reduction from the fixed-node triangle membership problem,
given in Definition 8. That is, we show that an algorithm for solving the k-cycle membership
problem can be used to solve the fixed-node triangle membership problem.
Proof of Theorem 6: First, we show how to construct an appropriate instance for the k-cycle
membership problem, given an instance for the fixed-node triangle membership problem. We start
with describing the construction for an odd value of k, and then show how to tweak it to handle
even values of k as well.
Let (G = (V, E), u) be an instance of the fixed-node triangle membership problem where |V | = n
e,
and let E(u) be the set of edges incident to the node u. For an odd k, we define an instance of
the k-cycle membership problem, G0 = (V 0 , E 0 ), where |V 0 | = n, as follows. We replace each edge
(u, v) ∈ E(u) with a path Puv of length (k − 3)/2 + 1, going through (k − 3)/2 new nodes, denoted
uv1 , . . . , uv(k−3)/2 , containing the edges (u, uv1 ), (uv1 , uv2 ), . . . , (uv(k−3)/2−1 , uv(k−3)/2 ), (uv(k−3)/2 , v).
For example, for k = 7, we replace each edge (u, v) by a path of three edges (u, uv1 ), (uv1 , uv2 ), (uv2 , v),
going through two new intermediate nodes uv1 and uv2 (see Figure 3). Formally, the new graph
19
G0 = (V 0 , E 0 ) is defined as:
V 0 = V ∪ {uvi | (u, v) ∈ E(u), 1 ≤ i ≤ (k − 3)/2}
E 0 = (E \ E(u)) ∪ {(u, uv1 )), (uv1 , uv2 ), . . . , (uv(k−3)/2−1 , uv(k−3)/2 ), (uv(k−3)/2 , v) | (u, v) ∈ E(u)}
The following observation follows directly from the construction above.
Observation 3. For an odd k, the node u participates in a k-cycle in G0 if and only if it participates
in a triangle in G.
Since an algorithm may use the IDs of the nodes, we need to also assign unique IDs to nodes
in G0 . We can do this in any consistent arbitrary manner, say, by assigning IDG0 (x) = IDG (x) if
x ∈ V , and ID(uvi ) = ID(v) ◦ bin(i) for nodes in V 0 \ V , where bin(i) is the binary representation
of i. Notice that ∆(G0 ) = ∆(G).
Next, assume towards a contradiction, that there is an algorithm A0 that solves the k-cycle
membership problem in an optimal number of (k−1)/2 rounds in the Congest(B) model with B <
n
e−2
log(( 2e(∆−1)
)∆−1 ). We show an algorithm A that solves the fixed-node triangle membership problem
n
e−2
on (G = (V, E), u) in a single round of the Congest(B) model with B < log(( 2e(∆−1)
)∆−1 ). As
this contradicts Theorem 5, and n = ke
n, this completes the proof for all values of k in O(n1− ) for
any constant 0 ≤ ≤ 1.
We construct A such that nodes in G simulate the nodes in G0 running algorithm A0 , as follows.
In A, each node in V , sends to its neighbors in G the message it sends in the first round of A0 on G0 .
If v ∈ V is a neighbor of u in G, then it sends to u the message it sends to uv(k−3)/2 in A0 . Then,
for each of its neighbors v, the node u has all the messages that the node uv(k−3)/2 receives in the
first round of A0 . Now, by backwards induction, for each i, 1 ≤ i ≤ (k − 3)/2, by local simulation at
the node u, it knows the view of the node uvi at the end of the first (k − 3)/2 − i + 1 rounds of A0 .
This implies that u knows the message sent to it by uv1 in round (k − 3)/2 + 1 = (k − 1)/2 of A0 .
Since in (k − 1)/2 rounds of A0 the node u knows whether it is in a k-cycle in G0 , by Observation 3,
it thus knows in a single round whether it is in a triangle in G.
To handle even value of k, we construct G0 in a similar manner of replacing each edge (u, v) ∈
E(u) by a path of (k − 4)/2 nodes. In addition, each edge (v, w) ∈ E \ E(u), is replaced by a
path of length two which consists of an additional node vw. Similarly to Observation 3, the node
u participates in a k-cycle in G0 if and only if it participates in a triangle in G. As in the case
for odd k, given an algorithm A0 for k-cycle membership in G0 , the simulation of (k − 4)/2 + 1 =
(k − 2)/2 = b(k − 1)/2c rounds of it in G requires only a single round of communication. Therefore,
the reduction carries over for even values of k as well. Observe that here n = ke
n+n
e2 , therefore, as
in the case of odd k, we achieve the same asymptotic lower bound as in the triangle membership
problem, for any value of k in O(n1− ).
A.1
k-Cycle membership for k ≥ 4 using communication complexity
Here we show a lower bound on the bandwidth needed for any optimal-round algorithm for solving
k-cycle membership for k ≥ 4 by using the standard framework of reduction from a 2-party communication complexity problem. For simplicity, we will show this for even values of k, but a similar
construction works for odd values as well. For k > 4, the bound is larger compared with our proof
of Section A.
20
Theorem 7. The k-cycle membership problem cannot be solved by a deterministic optimal-round
k−4
algorithm in the Congest(B) model unless B ≥ Ω(∆ 2 +1 log(n)), and it cannot be solved by a
randomized optimal-round algorithm, which succeeds with high probability6 , in the Congest(B)
k−4
model unless B ≥ Ω(∆ 2 +1 ), for any integers n, ∆, k such that
(∆ − 1)
k−4
+1
2
= O(n1− ),
for some constant 0 < < 1.
In order to prove Theorem 7, we show a reduction from the 2-party communication complexity
problem DISJK,S .
A 2-party communication complexity problem [37] consists of a function f : {0, 1}K ×{0, 1}K →
{TRUE, FALSE}, and two strings, x, y ∈ {0, 1}K , that are given as inputs for two players, Alice
and Bob, respectively. The players exchange bits of communication in order to compute f (x, y),
according to a protocol π. The communication complexity CC(π) of a protocol π for computing
f is the maximal number of bits, taken over all input pairs (x, y), exchanged between Alice and
Bob. The communication complexity CC(f ) of f is the minimum, taken over all protocols π that
compute f , of CC(π).
In the S-Disjointness problem (DISJK,S ), each of the players Alice and Bob receives a K-bits
input string containing exactly S ones, and the function f is DISJK,S (x, y), whose output is FALSE
if there is an index i ∈ {0, ..., K − 1} such that xi = yi = 1, and TRUE otherwise. Observe that
for S > K2 , the function is constant. For S ≤ K2 The deterministic communication complexity of
7
DISJK,S is known to be Ω(log K
S ) [31, 37] , while its randomized communication complexity is
known to be Ω(S) [27,57]. For the reduction, we adapt the formalization of Family of Lower Bound
Graphs given in [18] to our setting.
Definition 10 (Definition 1 (simplified) of [18]: Family of Lower Bound Graphs). Fix
an integer K, a function f : {0, 1}K × {0, 1}K → {TRUE, FALSE}. The family of graphs {Gx,y =
(V, Ex,y ) | x, y ∈ {0, 1}K }, is said to be a family of lower bound graphs w.r.t. f and k-cycle
membership if the following properties hold:
˙ A ∪V
˙ B ∪˙ Ve a fixed
(1) The set of nodes V is the same for all graphs, and we denote by V = {u}∪V
partition of it;
(2) Only the existence of edges in VA × Ve may depend on x;
(3) Only the existence of edges in VB × Ve may depend on y;
(4) The node u participates in a k-cycle in Gx,y iff f (x, y) = FALSE.
Observe that given a family of lower bound graphs {Gx,y = (V, Ex,y ) | x, y ∈ {0, 1}K } w.r.t.
to DISJK,S and k-cycle membership, if Alice and Bob can simulate an algorithm for k-cycle
membership on u, then by checking the output of u at the end of the algorithm they can solve
DISJK,S (x, y).
6
We say that an event occurs with high probability if it occurs with probability 1 − n1c , for some constant c ≥ 1.
It can be proved by the rank method for proving lower bounds for deterministic protocols in communication
complexity. To read more about the rank method, see for example [37] section 1.4. For the proof of the lower bound
on the rank of S-Disjointness, see [31], page 175.
7
21
...
...
...
...
...
...
...
...
Δ−1 −
ℎ
( − 4)/2
Figure 4: The reduction from Communication Complexity to k-cycle membership.
The proof of Theorem 7 is organized as follows. First, we construct a family of lower bound
graphs, and next, we show that given an algorithm ALG for k-cycle membership with messages of
size B, Alice and Bob can simulate ALG on Gx,y by exchanging only O(B) bits.
We now construct the following family of lower bound graphs, by describing a fixed graph
construction G = (V, E), which we then generalize to a family of graphs {Gx,y = (V, Ex,y ) |
x, y ∈ {0, 1}K }, which we show to be a family lower bound graphs w.r.t. to DISJK,S and k-cycle
membership.
The fixed graph construction: The fixed graph construction (Figure 4) consists of a tree T
and a path P of size n − |T |. The tree T is a tree in which the root node of T , denoted by u, is
connected to two nodes w and v, such that each of w and v is a root of a (∆ − 1)-regular tree of
depth k−4
2 . Denote by leaves(w) and leaves(v) the set of leaves of the tree rooted at w and the set
of leaves of the tree rooted at v, respectively. We define VA and VB to be leaves(w) and leaves(v)
respectively.
Adding edges corresponding to the inputs: Each of the players Alice and Bob receives as
k−4
the input a set of nodes in P of size (∆ − 1) 2 +1 . Alice connects the nodes in her input to the
k−4
(∆ − 1) 2 leaves of the tree rooted at v, such that each leaf is connected to ∆ − 1 nodes in P .
k−4
Similarly, Bob connects the nodes in his input to the (∆ − 1) 2 leaves of the tree rooted at w,
such that each leaf in connected to ∆ − 1 nodes in P . The following observation follows directly
from the construction.
Observation 4. The node u participates in a cycle of length k in Gx,y if and only if the two sets
of Alice and Bob are not disjoint.
Therefore, given an algorithm ALG for k-cycle membership, if Alice and Bob can simulate ALG
k−4
on u then they can solve S-Disjointness, where S = (∆ − 1) 2 +1 and the size of the input stings is
k−4
!
k−4
2
+1
X
2
(∆
−
1)
−
1
K = n − |T | ≥ n − 1 + 2
(∆ − 1)i ≥ n − 1 + 2
∆−2
i=0
Observe that for
(∆ − 1)
k−4
+1
2
22
= O(n1− )
k−4
+1
2
for some constant 0 < < 1, it holds that K = Θ(n), and log K
log(n)). It remains
S = Ω(∆
to show that given an algorithm for k-cycle membership with messages of size B, Alice and Bob
can simulate ALG on u by exchanging only O(B) bits.
Proof of Theorem 7: Let ALG be a (k/2 − 1)-round algorithm for solving the k-cycle memk/2−1
k/2−1
bership problem. Let {m1v→u , ..., mv→u } and {m1w→u , ..., mw→u } be the two sets of messages sent
from v and w to u during the k/2 − 1 rounds, where, e.g., miv→u is the message sent from v to u
in round i.
The crucial observation is that Alice and Bob can simulate the nodes v, w, and u during the
first k/2 − 2 rounds without any communication, because the k/2 − 2 neighborhoods of these nodes
are fixed. Therefore, in order for the players to compute the output of u after k/2 − 1 rounds, it
k/2−1
suffices for Alice to send to Bob the message mv→u , and for Bob to send to Alice the message
k/2−1
mw→u .
By Observation 4, this implies that 2B bits suffice for the players to correctly compute DISJK,S .
Therefore, by the lower bounds on DISJK,S , any deterministic optimal-round algorithm for k-cycle
k−4
+1
2
membership requires messages of size Ω(log K
log(n)), and any randomized optimalS ) = Ω(∆
round algorithm which succeeds with high probability requires Ω(S) = Ω(∆
n, ∆, k such that
k−4
(∆ − 1) 2 +1 = O(n1− )
k−4
+1
2
), for any integers
for some constant 0 < < 1.
We mention that one can use the set-disjointness function instead of DISJK,S to obtain a lower
k−4
bound of (∆ − 1) 2 , by simply connecting each leaf to a single node on the path, based on the
inputs. However, this gives a rather strong bound on ∆ with respect to n because then the number
of leaves must be also linear in n. Such a bound for ∆ does not occur in our given construction.
23
| 8cs.DS
|
TIME PARALLEL GRAVITATIONAL COLLAPSE SIMULATION
ANDREAS KREIENBUEHL
arXiv:1509.01572v3 [gr-qc] 28 Dec 2016
Center for Computational Sciences and Engineering, Lawrence Berkeley National Laboratory, 1 Cyclotron
Road, Berkeley CA 94720, United States of America
PIETRO BENEDUSI
Institute of Computational Science, Faculty of Informatics, Università della Svizzera italiana, Via
Giuseppe Buffi 13, 6904 Lugano, Switzerland
DANIEL RUPRECHT
School of Mechanical Engineering, University of Leeds, Woodhouse Lane, Leeds LS2 9JT, United Kingdom
ROLF KRAUSE
Institute of Computational Science, Faculty of Informatics, Università della Svizzera italiana, Via
Giuseppe Buffi 13, 6904 Lugano, Switzerland
Abstract. This article demonstrates the applicability of the parallel-in-time method Parareal to the
numerical solution of the Einstein gravity equations for the spherical collapse of a massless scalar field. To
account for the shrinking of the spatial domain in time, a tailored load balancing scheme is proposed and
compared to load balancing based on number of time steps alone. The performance of Parareal is studied
for both the sub-critical and black hole case; our experiments show that Parareal generates substantial
speedup and, in the super-critical regime, can reproduce Choptuik’s black hole mass scaling law.
1. Introduction
Einstein’s field equations of general relativity (GR) consist of ten coupled, non-linear, hyperbolic-elliptic
partial differential equations (PDEs). Because gravity couples to all forms of energy, there is an enormous
dynamic range of spatiotemporal scales in GR. Hence, usually only the application of advanced numerical
methods can provide solutions and in numerical relativity [2, 4] extensive use of high-performance computing
(HPC) is made [28, 34].
Today, almost all HPC architectures are massively parallel systems connecting large numbers of compute
nodes by a high-speed interconnect. In numerical simulations, the power of these systems can only be
harnessed by algorithms that feature a high degree of concurrency; every algorithm with strong serial
dependencies can only provide inferior performance on massively parallel computers. For the solution of
PDEs, parallelization strategies have been developed mainly for spatial solvers. However, in light of the
E-mail addresses: [email protected], [email protected], [email protected], [email protected].
Date: December 30, 2016.
2010 Mathematics Subject Classification. 35Q76, 65M25, 65Y05, 83C57.
Key words and phrases. Einstein-Klein-Gordon gravitational collapse, Choptuik scaling, Parareal, spatial coarsening, load
balancing, speedup.
1
2
TIME PARALLEL GRAVITATIONAL COLLAPSE SIMULATION
rapid increase in the number of cores in supercomputers, methods that offer additional concurrency along
the temporal axis have recently begun to receive more attention.
The idea of parallelization-in-time was introduced in 1964 [37]. In the 1980s and 1990s, time and
spacetime multigrid methods were studied [24–26]. More recently, the now widely used time parallel method
Parareal was proposed [33]. Other recently introduced parallel-in-time methods are PFASST [13, 35],
RIDC [10], or MGRIT [14]. A historical overview is offered in [19].
Given the demonstrated potential of parallel-in-time integration methods for large-scale parallel simulations [44], these methods could be beneficial for the numerical relativity community. However, their
application is not straightforward and often it is unclear a priori if good performance can be achieved. In
this article, we therefore investigate the principal applicability of the time parallel Parareal method to solving Einstein’s equations describing spherical, gravitational collapse of a massless scalar field. The system
is also referred to as an Einstein-Klein-Gordon system because it is equivalent to a Klein-Gordon equation
expressed in the context of GR, i.e. on a back-reacting, curved geometry. It defines a basic gravitational
field theory and is of interest therefore not only in numerical relativity but also in, e.g., quantum gravity [27, 31, 46]. A summary of numerically derived results is given in [23]; the work by Choptuik [8] brought
forward novel, physical results and is of particular interest here because we will show that Parareal correctly
reproduces the expected mass scaling law.
Mathematical theory shows that Parareal performs well for diffusive problems with constant coefficients [21]. For diffusive problems with space- or time-dependent coefficients, numerical experiments show
that Parareal can converge quickly too [32]. However, given the theory for basic constant-coefficient hyperbolic PDEs [21], it can be expected that Parareal applied to convection dominated problems converges
too slowly for meaningful speedup to be possible. Special cases with reasonable performance are discussed
in [17] and for certain hyperbolic PDEs it was found that some form of stabilization is required for Parareal
to provide speedup [7, 12, 20, 42]. Surprisingly, no stabilization is required for the equations describing
gravitational collapse; we demonstrate that plain Parareal can achieve significant speedup. A detailed analytical investigation of why this is the case would definitely be of interest but is left out for future work.
One reason could be that we solve in characteristic coordinates for which the discretization is aligned with
the directions of propagation [18, 30].
The article is structured as follows: In Section 2 we define the system of Einstein field equations that
we solve using Parareal. In addition, we give details on the numerical approach and discuss the interplay
between Parareal and the particular structure of the spatial mesh. In Section 3 we discuss the Parareal
method. Then, in Section 4 numerical results are presented. Finally, in Section 5 we conclude with a
summary and discussion.
2. Equations
2.1. Gravitational collapse. The Einstein field equations in Planck units normalized to 4πG/c4 = 1 are
(2.1.1)
Gµν = 2Tµν ,
where µ, ν ∈ {0, 1, 2, 3} index time (via 0) and space (via 1, 2, and 3).1 Once the non-gravitational matter
content is specified by a definition of the energy-momentum tensor Tµν , possibly along with equations of
state that together satisfy the continuity equations ∇µ Tµν = 0, Equation (2.1.1) defines a set of ten partial
differential equations for ten unknown metric tensor field components gµν .2 In all generality, the equations
are coupled, non-linear, and hyperbolic-elliptic in nature. Six of the ten equations are hyperbolic evolution
equations, while the remaining four are elliptic constraints on the initial data; they represent the freedom
to choose spacetime coordinates. For the matter content, we consider a minimally coupled massless scalar
field φ with energy-momentum tensor
1
(2.1.2)
Tµν = ∇µ φ∇ν φ − gµν g αβ ∇α φ∇β φ.
2
For the metric tensor field gµν in spherical symmetry it is natural to introduce a parametrization in terms of
Schwarzschild coordinates (t, r). Here, t is the time coordinate of a stationary observer at infinite radius r,
1We omit the addition of the cosmological constant term Λg
µν on the left-hand side in Equation (2.1.1) because observations
suggest 0 < Λ 1 (see, e.g., [29]); the term’s impact on black hole formation as studied here can be neglected.
2We use the Einstein summation convention.
TIME PARALLEL GRAVITATIONAL COLLAPSE SIMULATION
3
which measures the size of spheres centered at r = 0. In [8] the resulting Einstein field equations are analyzed
numerically. In particular, adaptive mesh refinement [5] is used to resolve the black hole formation physics.
In [22] the same investigation is carried out in double null or characteristic coordinates (τ, ρ) without mesh
refinement (see, however, [41, 45]). Finally, in [31] the effect of quantum gravity modifications on the
collapse is studied in adjusted characteristic coordinates. Here we use characteristic coordinates (τ, ρ) as
well but exclude quantum gravity modifications. Also, for simplicity, we will refer to τ as a time coordinate
and to ρ as a space coordinate.
Making the ansatz
(2.1.3)
gµν dxµ dxν = −2∂ρ rHdτ dρ + r2 (dϑ2 + [sin(ϑ)dϕ]2 )
for the metric tensor field and using an auxiliary field h for the spacetime geometry along with an auxiliary
field Φ for the matter content, the complete field equations are
(H − h)(Φ − φ)
h
∂τ Φ =
,
(2.1.4)
∂τ r = − ,
2
2r
for r and Φ, and
∂ρ r
∂ρ r
∂ρ r
(2.1.5)
∂ρ φ =
(Φ − φ),
∂ρ H =
H(Φ − φ)2 ,
(H − h),
∂ρ h =
r
r
r
for φ, H, and h (see [22]). Overall the system can be seen as a wave equation for the massless scalar field
φ on a back-reacting, curved geometry. Boundary conditions at (τ, ρ = τ ) are r = 0 and regularity of Φ, φ,
H, and h, which implies Φ = φ and H = h at the boundary [11, 30]. Consistent initial data at (τ = 0, ρ)
are
ρ
(2.1.6)
r= ,
Φ = (1 + ρ∂ρ )φ,
2
where we choose for φ the Gaussian wave packet
2 !
ρ − ρ0
ρ3
exp −
(2.1.7)
φ(0, ρ) = φ0
.
1 + ρ3
δ0
We also performed tests for initial data similar in shape to the hyperbolic tangent function much like
Choptuik did in [8] for purely serial time stepping. Since in this case we found Parareal’s performance to
resemble strongly that for the case of the Gaussian wave packet we do not include these results here. The
initial scalar field configuration is thus characterized by an amplitude φ0 , mean position ρ0 , and width δ0 .
Depending on the value of these parameters, the solution to Equations (2.1.4) and (2.1.5) can describe a
bounce of the wave packet or black hole formation near the boundary at r = 0. A black hole appears when
the outward null expansion
r
1 2h
+
(2.1.8)
Θ =
,
r H
which measures the relative rate of change of a cross-sectional area element of a congruence of out-going
null curves, approaches zero [38]. The black hole mass is
r
(2.1.9)
M= ,
2
evaluated at the point (τ + , ρ+ ) toward which Θ+ vanishes.
2.2. Numerical solution. The numerical grid is depicted in Figure 1a. It is parametrized by the characteristic coordinates τ and ρ, which are used for numerical integration; τ is used as coordinate representing
time and ρ as coordinate representing space. Integration thus takes place on a right triangle with initial
data defined along the lower right-hand leg. Clearly, the spatial domain becomes smaller as the solution
is advanced in τ . Note that the domain is not exactly a right triangle because at the upper-most corner a
small sub-triangle is missing. This “buffer” zone of extent λ is needed for the spatial part of the numerical
stencil to fit. The computational domain thus consists of all points (τ, ρ) ∈ [0, L − λ] × [0, L] with L = 80,
λ = 0.625, and ρ ≥ τ .
As a time stepping method for the solution of the equations in (2.1.4), we use a second-order LaxWendroff Richtmyer two-step method on a fine spacetime grid (see [30]). To employ the time parallel
method Parareal (see Section 3), we need a second, computationally cheap, time integration method. Here,
4
TIME PARALLEL GRAVITATIONAL COLLAPSE SIMULATION
t
0.2
ρ
τ
r
Field φ
0
−0.2
−0.4
τ
τ
τ
τ
−0.6
0
20
40
60
≈ 45.352
= 20
≈ 16.758
=0
80
Coordinate ρ
(a) The numerical domain. It is
parametrized by the
characteristic coordinates τ and
ρ.
(b) Scalar field solution snapshots for a black
hole-free setting. The peak of the Gaussian evolves
along the constant coordinate value ρ ≈ 20, which is
also when the bounce occurs in τ .
Figure 1. The computational domain (left) and sub-critical gravitational scalar field evolution (right).
we choose the explicit first-order Euler method on a coarse spacetime mesh. For Parareal to be efficient,
the cost of the coarse method has to be small compared to that of the fine one: by choosing a simple
first-order method on the coarse grid for C we obtain a good coarse-to-fine ratio (see Section 3.4). For
optimal speedup, the right balance between the difference in accuracy and difference in cost between C and
F has to be found.
For the integration in space of the equations in (2.1.5) we use a second-order Runge-Kutta method [30].
Snapshots of scalar field evolution resulting from the chosen fine grid discretization are shown in Figure 1b,
where φ evolves along constant lines of ρ until a bounce occurs at r = 0. The figure also shows how the
size of the domain decreases during the evolution: for τ = 0 the left boundary is at ρ = 0 while for τ = 20
it is at ρ = 20.
2.3. Mass scaling. In practice, the simulation terminates when a black hole forms because H grows
without bound in this case (see [11] for details). Figure 2a provides a simplified illustration of a black hole
region (dotted portion) and shows where the simulation comes to a halt (dashed line). Thus, to determine
the black hole mass M , we record minimal expansion values via the scalar (rΘ+ )mi = minρ {rΘ+ } derived
from Equation (2.1.8). The last such recorded minimal value before the termination of the simulation
defines a characteristic coordinate (τ + , ρ+ ) (see again Figure 2a), which we can use to define an r and M
via Equation (2.1.9). The scalar (rΘ+ )mi approaches 0 when (τ, ρ) nears (τ + , ρ+ ), as is shown in the lower
portion of Figure 2b.
Based on numerical experiments, Choptuik presents, among other things, a relation between the amplitude φ0 of the Gaussian in Equation (2.1.7) and the black hole mass M [8]. He shows that there is a critical
value φ?0 such that for φ0 < φ?0 there is a bounce (sub-critical case), while for φ0 > φ?0 there is a black hole
(super-critical case). Based thereon, he demonstrates that the black hole mass scales with φ0 − φ?0 > 0
according to the law M ∝ (φ0 − φ?0 )γ with γ being a positive constant of the same value for various initial
data profiles. We demonstrate that Parareal can correctly capture this black hole mass scaling law although
our coarse level Euler method alone cannot. Also, Parareal requires less wall-clock time than F, which can
be beneficial for the investigation of the high-accuracy demanding critical solution [8, 23] that requires the
simulation of numerous black holes [22]. This analysis however is omitted in this article and left for future
work.
TIME PARALLEL GRAVITATIONAL COLLAPSE SIMULATION
5
τ
r
(τ
+
,ρ +
)
•
Scalar (rΘ+ )mi
ρ
1.5
Scalar (rΘ+ )mi
t
1.5
1
0.5
0
1
0.5
0
0
20
40
60
80
Coordinate τ
(a) The simulation terminates
at τ + , when a black hole forms
at ρ+ .
(b) Minimal weighted outward null expansion
indicating a bounce (top) and black hole formation
(bottom) are shown.
Figure 2. Illustrations to clarify super-critical gravitational collapse.
3. Parareal
3.1. Algorithm. Parareal [33] is a method for the solution of initial value problems
(3.1.1)
∂τ u(τ ) = f (τ, u(τ )),
u(0) = u0 ,
0 ≤ τ ≤ T.
Here, as is outlined in the previous section, f comes from discretizing Equations (2.1.4) and (2.1.5), and
T = L − λ marks the end time. Parareal starts with a decomposition of the time domain into Npr temporal
subintervals (TSs) defined in terms of times τ p such that
(3.1.2)
[τ 1 , τ 2 ] ∪ ... ∪ [τ Npr −1 , τ Npr ] = [0, L − λ].
Now denote by F some serial time integration method of high accuracy and cost (in our case this is the
second-order Lax-Wendroff Richtmyer two-step method), and by C a cheap and possibly much less accurate
method (in our case this is the explicit first-order Euler method). Instead of running the fine method
subinterval by subinterval serially in time, Parareal performs the iteration
p
p
(3.1.3)
up+1
=
C
u
−
C
u
+
F
up[i] ,
[i+1]
[i+1]
[i]
where super-scripts index time or process number p ∈ {1, ..., Npr } and sub-scripts iterations i ∈ {1, ..., Nit }.
The advantage is that the expensive computation of the fine method can be performed in parallel over all
TSs at once. Here, we assume that the number of TSs is equal to the number Npr of cores (or processes)
used for the time direction. Good speedup can be obtained if C is fast in comparison to F but still accurate
enough for Parareal to converge rapidly. See Section 3.4 for a more detailed discussion of Parareal’s speedup.
In Section 2.2 we hinted at the interchangeability of the characteristic coordinates τ and ρ for the
numerical integration. Therefore, theoretically, Parareal could also be used for the spatial integration to
simultaneously parallelize both time and space. However, such an interweaving of two Parareal iterations
is not discussed in this article; it is put aside for future work.
3.2. Spatial coarsening in Parareal. In order to make C cheaper and improve speedup, we not only use
a less accurate time stepper for C but also employ a coarsened spatial discretization with a reduced number
of degrees-of-freedom. Therefore, we need a spatial interpolation I and restriction R operator. In this case
(see, e.g., [15]), the Parareal algorithm is given by
p
p
p
(3.2.1)
up+1
[i+1] = IC Ru[i+1] − IC Ru[i] + F u[i] .
6
TIME PARALLEL GRAVITATIONAL COLLAPSE SIMULATION
As restriction operator R we use point injection. For the interpolation operator I we use polynomial (i.e.
Lagrangian) interpolation of order 3, 5, and 7.3 It has been shown that, even for simple toy problems,
convergence of Parareal can deteriorate if spatial coarsening with low-order interpolation is used. As
demonstrated in Section 4.1, this also holds true for the here studied problem.
3.3. Implementation. We have implemented two different realizations of Parareal. In a “standard” version Pst (see Listing 3a), the Parareal correction is computed on each TS up to a uniformly prescribed
iteration number. In contrast, in the “modified” implementation Pmo (see Listing 3b), Parareal corrections
are only performed on TSs where the solution may not yet have converged. Because Parareal always converges at a rate of at least one TS per iteration, we only iterate on a TS if its assigned MPI rank is greater
than or equal to the current Parareal iteration number (see line 8 in Listing 3b). Otherwise, no further
iterations are needed and performed, and the process remains idle. Thus, as the iteration progresses, more
and more processes enter an idle state. In an implementation to be realized in future work, the criterion
for convergence used here will be replaced by a check for some residual tolerance [3]. This could negatively
affect the observed performance since it requires essentially one more iteration to compute the residual.4 It
also bears mentioning that it has very recently been demonstrated that parallel-in-time integration methods
are good candidates to provide algorithm-based fault tolerance [36, 43].
Another difference between the standard and modified implementation is that in the former, after each
time parallel fine evolution, a copy of the fine grid solution has to be created (see line 10 in Listing 3a). In
the modified Listing 3b this copying is circumvented by the use of two alternating indices “j” and “k” in
lines 9 and 10, respectively. The iteration number determines their value which, in turn, determines the
fine grid solution buffer that is used to send or receive data by means of the corresponding MPI routines
(see lines 14 and 22 in Listing 3b). The two implementations also have slightly different requirements in
terms of storage. As can be seen in line 15 in Listing 3a, in Pst on the first TS or, equivalently, for the
first MPI rank, the fine grid solution has to be assigned initial data at the beginning of each iteration. This
requires one additional buffer to be held in storage. Other than that both implementations need one coarse
grid solution buffer and three fine grid buffers for each TS.
3.4. Speedup. We denote by Rco the coarse and by Rfi the fine time stepper’s runtime. Recalling that
Nit denotes the number of iterations required for Parareal to converge given Npr processes, Parareal’s
theoretically achievable speedup is
−1
Nit
Npr Rfi
Nit Rco
+
≤ min
,
,
(3.4.1)
S=
1+
Npr Rfi
Npr
Nit Rco
as is discussed, e.g., in [35]. The estimate is valid only for the ideal case, where runtimes across subintervals
are perfectly balanced. In the presence of load imbalances in time however, i.e. differences in the runtimes of
C and F across TSs, maximum speedup is reduced [32]. Because the spatial domain we consider is shrinking
in time, a tailored decomposition of the time axis has to be used to provide well balanced computational
load, as is discussed in the next section.
3.5. Load balancing. Because we integrate over a triangular computational spacetime domain (see Figure 1a), a straight forward, uniform partitioning of the time axis results in imbalanced computational load
in time. The first load balancing (LB) strategy, which henceforth we will refer to as LB1, is based on this
straight forward, basic decomposition of the time axis. It assigns to each TS the same number of time steps
without regard to their computational cost. Because of the shrinking domain, TSs at later times carry fewer
p
spatial degrees-of-freedom so that the per-process runtimes Rco
and Rfip of the coarse and fine time stepper,
respectively, are larger for the earlier TSs than for the later ones. Figure 4a shows how this partition leads
to an imbalanced computational load in time because the portion extending across the “early-middle” TS
[e, m] covers a larger area and thus a larger number of grid points than the portion over the “middle-late”
TS [m, l].
3We also tested barycentric interpolation [6, 16] but found the performance in terms of runtimes and speedup (see Sections 3.4 and 4) to be inferior.
4In [3] a version of Parareal is discussed that can be used to proceed the integration beyond a given end time. It is based
on an optimized scheduling of those tasks which become idle in our implementation.
TIME PARALLEL GRAVITATIONAL COLLAPSE SIMULATION
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
i f p > 1 then // In itializa tion
Coarse ( co ; τ 1 → τ p )
I n t e r p ( co 7→ f i [ 0 ] )
i f p < Npr then // Prediction
Coarse ( co ; τ p → τ p+1 )
I n t e r p ( co 7→ f i [ 2 ] )
f o r i = 1 : Nit do // Iteration
i f p < Npr then
F i n e ( f i [ 0 ] ; τ p → τ p+1 )
fi [1] = fi [0]
f i [ 1 ] −= f i [ 2 ]
i f p > 1 then
MPI Recv ( f i [ 0 ] ; p ⇐ p − 1 )
else
Init ( fi [0])
R e s t r i c t ( f i [ 0 ] 7→ co )
i f p < Npr then
Coarse ( co ; τ p → τ p+1 )
I n t e r p ( co 7→ f i [ 2 ] )
f i [ 1 ] += f i [ 2 ]
i f p < Npr then
MPI Send ( f i [ 1 ] ; p ⇒ p + 1 )
i f p > 1 then // Ini tializa tion
Coarse ( co ; τ 1 → τ p )
I n t e r p ( co 7→ f i [ 0 ] )
i f p < Npr then // Prediction
Coarse ( co ; τ p → τ p+1 )
I n t e r p ( co 7→ f i [ 2 ] )
f o r i = 1 : Nit do // Iteration
i f p >= i then
j = ( i +1) % 2
k = i % 2
i f p < Npr then
F i n e ( f i [ j ] ; τ p → τ p+1 )
i f p > i then
MPI Recv ( f i [ k ] ; p ⇐ p − 1 )
f i [ j ] −= f i [ 2 ]
R e s t r i c t ( f i [ k ] 7→ co )
i f p < Npr then
Coarse ( co ; τ p → τ p+1 )
I n t e r p ( co 7→ f i [ 2 ] )
f i [ j ] += f i [ 2 ]
i f p < Npr then
MPI Send ( f i [ j ] ; p ⇒ p + 1 )
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
(a) The standard Parareal implementation Pst .
7
(b) The modified Parareal implementation Pmo .
Figure 3. Pseudo code for the standard and modified Parareal implementation. The
variable “co” denotes the coarse grid solution and “fi” an array of three fine grid buffers.
t
t
τ =l
τ =l
ρ
τ
r
r
τ
=
m
τ
=
m
τ
ρ
τ =e
(a) Imbalanced load in time
from load balancing LB1.
τ =e
(b) Balanced load in time from
load balancing LB2.
Figure 4. Illustration of two different approaches for the decomposition of the time domain, i.e. LB1 (left) and LB2 (right).
Figure 4 suggests that early in time TSs should have a shorter extent in time than later ones. Thus,
in the second strategy, which in the following we will refer to as LB2, we also consider the cost of time
p
steps in order to balance the runtime Rco
+ Rfip over all processes p. We use a decomposition of the time
axis in TSs such that the sum of the total coarse and fine runtime is balanced over all TSs, i.e. such that
p
Rco + Rfi = Npr (Rco
+ Rfip ) for any process p. This is done by a bisection approach, making use of the fact
the we use explicit rather than implicit time integrators (cf. the discussion in [32]), and thus that the cost
of a time step from τ to τ + ∆τ is directly proportional to the number of spatial degrees-of-freedom at τ .
Therefore, the total spacetime domain is first divided into two parts of roughly equal number of grid points
8
TIME PARALLEL GRAVITATIONAL COLLAPSE SIMULATION
as is sketched in Figure 4b. Then, each part is divided again and again until the required number of TSs
is reached. Note that this limits the possible numbers of TSs to powers of 2.
Figure 5 shows Vampir5 traces for one simulation featuring LB1 (Figure 5a) and one LB2 (Figure 5b).
The horizontal axes correspond to runtime, while the vertical axes depict MPI rank numbers from 1 (lower)
to 8 (upper). In each case, three Parareal iterations are performed. Green regions indicate the coarse and
fine integrators carrying out work. Time spent in MPI receives (including waiting time) is shown in red.
We observe how LB1 leads to load imbalance and incurs significant wait times in processes handling later
TS. In contrast, the processes’ idle times (shown in red) in MPI receives are almost invisible in the case of
LB2. Elimination of wait times leads to a significant reduction in runtime and increase in speedup, as will
be shown in Section 4.
(a) Vampir trace for LB1. The Parareal runtime is Rpa = 7.964 (s).
(b) Vampir trace for LB2. The Parareal runtime is Rpa = 5.436 (s).
Figure 5. Vampir traces for the implementation Pmo with (Npr , Nit ) = (8, 3) for two
different load balancing strategies.
4. Results
Speedup and runtime measurements were performed on the Cray XC40 supercomputer Piz Dora6 at the
Swiss National Supercomputing Centre (CSCS) in Lugano, Switzerland. It features 1,256 compute nodes,
which all hold two 12-core Intel Xeon E5-2690v3 processors. This results in a total of 30,144 compute cores
and a peak performance of 1.254 PFlops; it occupies position 56 in the Top500 November, 2014 list.7 On
Piz Dora, we used the GNU compiler collection8 version 4.9.2 and the runtimes we provide do not include
the cost of I/O operations. Some simulations measuring convergence were performed on a machine located
at the Università della Svizzera italiana that is maintained by members of the Institute of Computational
Science of the Faculty of Informatics.9
For the results presented in the following we use a coarse grid resolution of (∆τ )co = (∆ρ)co = ∆co =
L/2,048 ≈ 0.039 and a fine grid resolution of ∆fi = ∆co /8 ≈ 0.005. We have also determined a reference
solution to approximately measure the serial fine stepper’s discretization error. For this we have used again
the serial fine time stepper but with a step size of ∆re = ∆fi /4 ≈ 0.001.
4.1. Sub-critical. First we consider the sub-critical case, where no black holes form. Figure 6 shows for
Npr = 256 and two different sets of initial data parameters the relative defect
(4.1.1)
D[i] =
kr[i] − rfi k2
,
krfi k2
5https://www.vampir.eu/
6http://user.cscs.ch/computing systems/piz dora/
7http://www.top500.org/list/2014/11
8https://gcc.gnu.org
9https://www.ics.usi.ch/index.php/ics-research/resources
TIME PARALLEL GRAVITATIONAL COLLAPSE SIMULATION
9
which measures the difference between the Parareal solution r[i] after i iterations and the serial fine solution
rfi as a function of the characteristic coordinate τ .
In Figure 6a we use the initial data parameters (φ0 , ρ0 , δ0 ) = (0.035, 20, 1), which results in an “early”
bounce of the wave packet at about τ = 20. For the simulations in Figure 6b, the values are (φ0 , ρ0 , δ0 ) =
(0.01, 75, 1), which leads to a “late” bounce at about τ = 75. Defects are plotted for Nit ∈ {1, 2, 3, 4}
along with the serial coarse and fine solution’s estimated discretization error krco − rre k2 /krfi k2 and krfi −
rre k2 /krfi k2 labeled “Coarse” and “Fine”, respectively. We observe that in Figure 6a, the data for Nit = 3
is somewhat jagged because for LB2 there are various start and end times of TSs near the bounce region.
In any case, Parareal converges in two iterations: for Nit = 2, the defect is below the discretization error for
all τ . In fact, without the bounce region near τ = 20, only one iteration would be required for convergence.
For the late bounce scenario in Figure 6b, we also observe that the rate of convergence at the final time
τ = L − λ gives an indication of the convergence at all τ . In the following we thus focus on convergence at
the final time. Convergence for the other evolved field Φ is not shown but was found to be at least as good
as for r.10
10−2
10−6
10−9
Nit = 1
Nit = 2
Nit = 3
Nit = 4
Coarse
Fine
10−12
10−15
0
20
40
60
80
Coordinate τ
(a) Parareal’s defect over time for an early
bounce scenario.
Defect D[i] for Npr = 256
Defect D[i] for Npr = 256
10−3
Nit = 1
Nit = 2
Nit = 3
Nit = 4
Coarse
Fine
10−5
10−8
10−11
10−14
0
20
40
60
80
Coordinate τ
(b) Defect of Parareal over time for a late
bounce situation.
Figure 6. Defect in r between Parareal and the fine method over time for fixed Npr = 256.
Figures 7a and 7b illustrate the defect of Parareal at the end of the simulation at τ = L − λ for various
values of Npr with third-order interpolation (left) and fifth-order interpolation (right). For third-order
interpolation, Parareal does not converge at all. The configuration stalls at a defect of about 10−2 until the
iteration count equals Npr . There, Parareal converges by definition but cannot provide any speedup. In
contrast, Parareal shows good convergence behavior for fifth-order interpolation. For Npr less than 64, the
defect of Parareal falls below the approximate discretization error of the fine method after a single iteration.
Otherwise, for Npr ≥ 64 up to Npr = 512, two iterations are required.
The resulting speedups with correspondingly adjusted values for Nit are shown in Figure 7c for both load
balancing strategies (see the discussion in Section 3.5). In addition, the projected speedup according to
Equation (3.4.1) is shown. The fine-to-coarse ratio Rfi /Rco was determined experimentally and found to be
about 74. Up to Npr = 64, for the advanced load balancing, speedup closely mirrors the theoretical curve
while the basic load balancing performs significantly worse. For Npr ≥ 64, measured speedups fall short
of the theoretical values, peak at Npr = 256, and then start to decrease. Note that the theoretical model
(blue line in Figure 7c) does take into account the scaling limit from the serial correction step according to
Amdahl’s law. The difference between theory and measured speedup is therefore due to other overheads
(communication and transfer between meshes) as analysed below.
10Convergence seems to be unaffected by the load balancing. In tests not documented here we found that for LB1 it takes
two iterations for Parareal to converge as well.
10
TIME PARALLEL GRAVITATIONAL COLLAPSE SIMULATION
Although the load balancing strategy LB2 results in significantly better speedup than the basic approach
LB1, the peak value provided by both schemes is essentially the same. This is because for increasingly large
numbers of cores, the computational load per TS eventually becomes small and imbalances in computational
load insignificant. Instead, runtime is dominated by overhead from, e.g., communication in time. The
communication load is independent of the chosen load balancing and depends solely on the number of TSs;
for every TS one message has to be sent and received once per iteration (save for the first and last TS).
Therefore, it can be expected that ultimately both approaches to load balancing lead to comparable peak
values. Below we demonstrate that the saturation in speedup is related to a significant increase in time
spent in MPI routines; eventually, communication cost starts to dominate over the computational cost left
on each time slice and the time parallelization saturates just as spatial parallelization does.
Defect D[i] at τ = L − λ
Defect D[i] at τ = L − λ
10−3
10−5
10−8
Npr = 2
Npr = 8
Npr = 32
Npr = 128
Npr = 512
Coarse
Fine
10−11
10−14
1
8
10−6
10−9
10−12
Theory
LB2
LB1
90
75
Speedup S
Npr = 2
Npr = 8
Npr = 32
Npr = 128
Npr = 512
Coarse
Fine
10−2
60
45
30
15
0
64
512
10−15
1
Iterations Nit
(a) Defect for late bounce and
interpolation order 3.
8
64
512
Iterations Nit
(b) Defect for late bounce and
interpolation order 5.
2
8
32 128
Cores Npr
512
(c) Parareal speedup for fifth-order
interpolation.
Figure 7. Parareal’s performance for the sub-critical case in terms of convergence for
polynomial interpolation orders 3 and 5, and in terms of speedup.
Figure 8 illustrates the reason behind the drop-off in speedup beyond Npr = 256. First, define
X p
p
p
(4.1.2)
Rpa
= Rco
+ Rfip +
Rst ,
st
p
where Rst
denotes runtime spent in stages that are different from coarse and fine integration on the TS
assigned to process p. For now, we consider only overhead from sending and receiving data as well as from
interpolation; other overheads are not further analyzed here. Next, we introduce the total overhead on a
TS as the sum of all stage-runtimes or
X p
p
(4.1.3)
Oto
=
Rst ,
st
which is also the runtime spent neither in the coarse nor fine integrator for a given p. The average overhead
p
is now defined as the geometric mean value of Oto
over all TSs, which is
PNpr p
p=1 Oto
.
(4.1.4)
Oav =
Npr
Finally, we define the relative overhead for individual stages on a TS as
p
Rst
p ,
Rpa
p
where Rpa
is the runtime of Parareal at processor p. Ideally, as is assumed for the derivation of the speedup
p
p
p
model given in Equation (3.4.1), Rco
and Rfip are the dominant costs. In this case, Rco
+ Rfip ≈ Rpa
so that
p
according to Equation (4.1.2) we have Oto ≈ 0 and therefore Oav ≈ 0 by definition. However, as can be seen
in Figure 8a, Oav is small only for small values of Npr . For Npr ≥ 32 it increases rapidly, which indicates
(4.1.5)
p
Ost
=
TIME PARALLEL GRAVITATIONAL COLLAPSE SIMULATION
11
that the overhead from communication and other sources starts to play a more dominant role when Npr is
increased.
Figure 8b shows the relative overhead from Equation (4.1.5) for Npr ∈ {32, 512} and p ∈ {1, ..., Npr }
for the three different stages st ∈ {Interpolation, Send, Receive}; “Send” and “Receive” refer to the corresponding MPI routines. There is a significant increase in relative overhead in all three stages as the number
of cores grows, causing the eventual drop-off in speedup for increasing Npr .
101
p
Overhead Ost
for Npr ∈ {32, 512}
Overhead Oav
0.9
0.6
0.3
0
2
8
32
128
Cores Npr
(a) Average overhead.
512
10−2
Receive, Npr = 512
Receive, Npr = 32
Interpolate, Npr = 512
Interpolate, Npr = 32
Send, Npr = 512
Send, Npr = 32
10−5
10−8
2
8
32
128
512
Core p
(b) Overhead caused by three different
Parareal stages.
Figure 8. Overhead from communication and other sources increases with Npr , which
leads to Parareal’s speedup decay.
4.2. Super-critical. We consider now the more complex case in which a black hole forms at some time
during the simulation. The goal is to compute the black hole’s position via Equation (2.1.8) so that its
mass can be determined from Equation (2.1.9) (see Section 2.3). Because the characteristic coordinates
(τ, ρ) do not allow us to continue the simulation past the black hole formation event, we need a way to keep
the simulation from terminating when Θ+ approaches 0 (see Figure 2b).
To avoid the need to adaptively modify the decomposition of the time domain, we carry out the supercritical case study using initial data parameter values near (φ0 , ρ0 , δ0 ) = (0.01, 75, 1), which we have also
used for the results in Figure 6b. With these parameters and in particular for φ0 ≥ 0.01, for all investigated
partitions of the time axis with Npr ≤ 256, the black hole generated by the fine time integrator forms in
the last TS unless φ0 becomes too large (ρ0 and δ0 are fix). Thus, Parareal can be used over all TSs except
for the last one, where only the fine method is executed to compute the black hole’s position. The C++
implementation uses a try-throw-catch approach to prevent complete termination of the simulation; if the
radicand in the definition of Θ+ in Equation (2.1.8) fails to be non-negative, an exception is thrown such
that the Parareal iteration can continue. As the Parareal iteration converges and better and better starting
values are provided for F on the last TS, the accuracy of the computed black hole position improves. A
more general implementation aiming at production runs would need to allow for black hole formation in
TSs before the last one but this is left for future work. In this article, the focus lies on investigating the
principal applicability of Parareal to the simulation of gravitational collapse.
Figure 9a depicts the Choptuik scaling that results from solutions computed with Parareal for Npr = 256
after the first three iterations. Table 1 lists the generated values of φ?0 and γ (see Section 2.3), and errors
compared to the value provided by the fine integrator, which agrees with the result in [22]. As can be
seen in Figure 9a, the coarse integrator C alone cannot adequately resolve black holes with φ0 − φ?0 . 10−9
(they are too small for C to be “visible”) and its γ is wrong by about 20%. This means that the coarse
method is too “coarse” in the sense that, on its own, it cannot correctly capture the physics underlying the
investigated problem. Nonetheless, Parareal is not only capable of generating the correct black hole physics
but can do so after only one iteration.
12
TIME PARALLEL GRAVITATIONAL COLLAPSE SIMULATION
Theory
LB2
90
10−1
75
Mass M
Speedup S
10−2
10−3
Coarse
Nit = 1
Nit = 2
Nit = 3
Fine
10−4
60
45
30
15
0
2
10−12 10−10 10−8 10−6 10−4
8
Criticality φ0 − φ?0
(a) Choptuik scaling from Parareal.
32
128
Cores Npr
512
(b) Parareal speedup.
Figure 9. Parareal’s performance for the super-critical case.
φ?0
Value
Coarse
Nit = 1
Nit = 2
Nit = 3
Fine
γ
Error (%)
Value
Error (%)
−1
0.458
0.377
0.370
0.381
0.381
20.21
1.05
2.89
0
0
0.01057748 7.25 · 10
0.01055915 5.51 · 10−1
0.01050240 1.01 · 10−2
0.01050135 9.52 · 10−5
0.01050134
0
Table 1. Approximate values and relative errors for the critical amplitude φ?0 and resulting
straight line slope γ.
Figure 9b visualizes the speedup achieved in the super-critical case including the theoretical estimate
according to Equation (3.4.1). The numbers of iterations required for Parareal to converge are derived
from an analysis just like the one plotted in Figure 7b for the sub-critical case and basically the values are
identical. Up to 64 processes, good speedup close to the theoretical bound is observed. For larger core
numbers however, speedup reaches a plateau and performance is no longer increasing. As in the sub-critical
case, as Npr increases, the computing times per TS eventually become too small and Parareal’s runtime
becomes dominated by, e.g., communication (see Figure 8). Even though the temporal parallelization
eventually saturates, substantial acceleration of almost a factor of 30 using 128 cores in time is possible,
corresponding to a parallel efficiency of about 23%.
5. Conclusion
The article assesses the performance of the parallel-in-time integration method Parareal for the numerical
simulation of gravitational collapse of a massless scalar field in spherical symmetry. It gives an overview
of the dynamics and physics described by the corresponding Einstein field equations and presents the
employed numerical methods to solve them. Because the system is formulated and solved in characteristic
coordinates, the computational spacetime domain is triangular so that later time steps carry fewer spatial
degrees-of-freedom. A strategy for balancing computational cost per subinterval instead of just number of
steps is discussed and its benefits are demonstrated by traces using the Vampir tool. Numerical experiments
are presented for both the sub- and super-critical case. Parareal converges rapidly for both and, for the
latter, correctly reproduces Choptuik’s mass scaling law after only one iteration despite the fact that the
used coarse integrator alone generates a strongly flawed mass scaling law. This underlines the capability of
Parareal to quickly correct a coarse method that does not resolve the dynamics of the problem. The results
TIME PARALLEL GRAVITATIONAL COLLAPSE SIMULATION
13
given here illustrate that Parareal and presumably other parallel-in-time methods as well can be used to
improve utilization of parallel computers for numerical studies of black hole formation.
Multiple directions for future research emerge from the presented results. Evaluating performance gains
for computing the critical solution [8, 23] would be valuable. Next, complexer collapse scenarios such
as in the Einstein-Yang-Mills system [9], axial symmetry [39], or binary black hole spacetimes [40] could
be addressed. An extended implementation of Parareal could utilize a more sophisticated convergence
criterion [3], a more flexible black hole detection, and parallelism in space via, e.g., again Parareal. The
latter would be possible because the integration along the characteristic we took to represent space is for
the solution of initial value problems just like in the temporal direction. Another topic of interest is that
of adaptive mesh refinement [1]: how it can be used efficiently in connection with Parareal or other time
parallel methods seems to be an open problem. As discussed in the introduction, a mathematical analysis of
the convergence behavior of Parareal for Einstein’s equations would be of great interest as well, particularly
since the good performance is unexpected in view of the negative theoretical results for basic hyperbolic
problems. Finally, incorporating a parallel-in-time integration method into a software library widely used
for black hole or other numerical relativity simulations would be the ideal way to make this new approach
available to a large group of domain scientists.11
Acknowledgments
We would like to thank Matthew Choptuik from the University of British Columbia in Vancouver, Canada
and Jonathan Thornburg from the Indiana University in Bloomington (IN), United States of America for
providing feedback and suggestions on an earlier version of the manuscript. We would also like to thank
Jean-Guillaume Piccinali and Gilles Fourestey from the Swiss National Supercomputing Center (CSCS) in
Lugano, Switzerland and Andrea Arteaga from the Swiss Federal Institute of Technology Zurich (ETHZ),
Switzerland for discussions concerning the hardware at CSCS.
This research is funded by the Deutsche Forschungsgemeinschaft (DFG) as part of the “ExaSolvers”
project in the Priority Programme 1648 “Software for Exascale Computing” (SPPEXA) and by the Swiss
National Science Foundation (SNSF) under the lead agency agreement as grant SNSF-145271. The research
of A.K., D.R., and R.K. is also funded through the ”FUtuRe SwIss Electrical InfraStructure” (FURIES)
project of the Swiss Competence Centers for Energy Research (SCCER) at the Commission for Technology
and Innovation (CTI) in Switzerland.
References
[1] Private communication with Jonathan Thornburg from the Indiana University in Bloomington (IN),
United States of America.
[2] M. Alcubierre. Introduction to 3+1 Numerical Relativity, volume 140 of International Series of
Monographs on Physics. Oxford University Press, Oxford, UK, 1st edition, Jun 2008. ISBN
9780199205677.
[3] E. Aubanel. Scheduling of tasks in the parareal algorithm. Parallel Computing, 37(3):172–182, Mar
2011. DOI 10.1016/j.parco.2010.10.004.
[4] T. W. Baumgarte and S. L. Shapiro. Numerical Relativity. Solving Einstein’s Equations on the
Computer. Cambridge University Press, Cambridge, UK, 1st edition, Aug 2010. ISBN 9780521514071.
[5] M. J. Berger and J. Oliger. Adaptive mesh refinement for hyperbolic partial differential equations.
Journal of Computational Physics, 53(3):484–512, Mar 1984. DOI 10.1016/0021-9991(84)90073-1.
URL http://www.sciencedirect.com/science/article/pii/0021999184900731.
[6] J.-P. Berrut and L. N. Trefethen. Barycentric Lagrange Interpolation. SIAM Review, 46(3):501–517,
Sep 2004. DOI 10.1137/S0036144502417715.
[7] F. Chen, J. S. Hesthaven, and X. Zhu. On the Use of Reduced Basis Methods to Accelerate and
Stabilize the Parareal Method. In A. Quarteroni and G. Rozza, editors, Reduced Order Methods for
Modeling and Computational Reduction, volume 9 of MS&A - Modeling, Simulation and Applications,
pages 187–214. Springer International Publishing, 2014. URL
http://dx.doi.org/10.1007/978-3-319-02090-7_7.
11A copy of the library Lib4PrM for the Parareal method can be obtained by cloning the Git repository https://scm.tiedu.ch/repogit/lib4prm.
14
TIME PARALLEL GRAVITATIONAL COLLAPSE SIMULATION
[8] M. W. Choptuik. Universality and scaling in gravitational collapse of a massless scalar field. Physical
Review Letters, 70(1):9–12, Jan 1993. DOI 10.1103/PhysRevLett.70.9.
[9] M. W. Choptuik, E. W. Hirschmann, and R. L. Marsa. New critical behavior in Einstein-Yang-Mills
collapse. Physical Review D, 60(12):124011, Nov 1999. DOI 10.1103/PhysRevD.60.124011. URL
http://arxiv.org/abs/gr-qc/9903081.
[10] A. J. Christlieb, C. B. Macdonald, and B. W. Ong. Parallel high-order integrators. SIAM Journal on
Scientific Computing, 32(2):818–835, 2010. URL http://dx.doi.org/10.1137/09075740X.
[11] D. Christodoulou. Bounded variation solutions of the spherically symmetric Einstein-scalar field
equations. Communications on Pure and Applied Mathematics, 46(8):1131–1220, 1993. DOI
10.1002/cpa.3160460803.
[12] X. Dai and Y. Maday. Stable Parareal in Time Method for First- and Second-Order Hyperbolic
Systems. SIAM Journal on Scientific Computing, 35(1):A52–A78, 2013. URL
http://dx.doi.org/10.1137/110861002.
[13] M. Emmett and M. L. Minion. Toward an Efficient Parallel in Time Method for Partial Differential
Equations. Communications in Applied Mathematics and Computational Science, 7:105–132, 2012.
URL http://dx.doi.org/10.2140/camcos.2012.7.105.
[14] R. D. Falgout, S. Friedhoff, T. V. Kolev, S. P. MacLachlan, and J. B. Schroder. Parallel time
integration with multigrid. SIAM Journal on Scientific Computing, 36:C635–C661, 2014. URL
http://dx.doi.org/10.1137/130944230.
[15] P. F. Fischer, F. Hecht, and Y. Maday. A parareal in time semi-implicit approximation of the
Navier-Stokes equations. In R. Kornhuber and et al., editors, Domain Decomposition Methods in
Science and Engineering, volume 40 of Lecture Notes in Computational Science and Engineering,
pages 433–440, Berlin, 2005. Springer. URL http://dx.doi.org/10.1007/3-540-26825-1_44.
[16] M. S. Floater and K. Hormann. Barycentric Rational Interpolation with no Poles and High Rates of
Approximation. Numerische Mathematik, 107(2):315–331, Aug 2007. DOI
10.1007/s00211-007-0093-y.
[17] M. J. Gander. Analysis of the Parareal Algorithm Applied to Hyperbolic Problems using
Characteristics. Bol. Soc. Esp. Mat. Apl., 42:21–35, 2008.
[18] M. J. Gander. Analysis of the Parareal algorithm applied to hyperbolic problems using charateristics.
Boletı́n de la Sociedad Española de Matemática Aplicada, (42):21–35, Mar 2008. URL
http://www.sema.org.es/ojs/index.php?journal=sema&page=article&op=view&path%5B%5D=19.
[19] M. J. Gander. 50 years of Time Parallel Time Integration. In Multiple Shooting and Time Domain
Decomposition. Springer, 2015. URL http://dx.doi.org/10.1007/978-3-319-23321-5_3.
[20] M. J. Gander and M. Petcu. Analysis of a Krylov Subspace Enhanced Parareal Algorithm for Linear
Problem. ESAIM: Proc., 25:114–129, 2008. URL http://dx.doi.org/10.1051/proc:082508.
[21] M. J. Gander and S. Vandewalle. Analysis of the Parareal Time-Parallel Time-Integration Method.
SIAM Journal on Scientific Computing, 29(2):556–578, 2007. URL
http://dx.doi.org/10.1137/05064607X.
[22] D. Garfinkle. Choptuik scaling in null coordinates. Physical Review D, 51(10):5558–5561, May 1994.
DOI 10.1103/PhysRevD.51.5558. URL http://arxiv.org/abs/gr-qc/9412008.
[23] C. Gundlach and J. M. Martı́n-Garcı́a. Critical phenomena in gravitational collapse. Living Reviews
in Relativity, 10(5), Dec 2007. DOI 10.12942/lrr-2007-5. URL
http://arxiv.org/abs/0711.4620.
[24] W. Hackbusch. Parabolic multi-grid methods. Computing Methods in Applied Sciences and
Engineering, VI, pages 189–197, 1984. URL http://dl.acm.org/citation.cfm?id=4673.4714.
[25] G. Horton. The time-parallel multigrid method. Communications in Applied Numerical Methods, 8
(9):585–595, 1992. URL http://dx.doi.org/10.1002/cnm.1630080906.
[26] G. Horton, S. Vandewalle, and P. Worley. An Algorithm with Polylog Parallel Complexity for Solving
Parabolic Partial Differential Equations. SIAM Journal on Scientific Computing, 16(3):531–541,
1995. URL http://dx.doi.org/10.1137/0916034.
[27] V. Husain. Critical Behavior in Quantum Gravitational Collapse. Advanced Science Letters, 2(2):
214–220(7), Jun 2009. DOI 10.1166/asl.2009.1028. URL http://arxiv.org/abs/0808.0949.
TIME PARALLEL GRAVITATIONAL COLLAPSE SIMULATION
15
[28] L. E. Kidder, M. A. Scheel, S. A. Teukolsky, E. D. Carlson, and G. B. Cook. Black hole evolution by
spectral methods. Physical Review D, 62(8):084032, Sep 2000. DOI 10.1103/PhysRevD.62.084032.
URL http://arxiv.org/abs/gr-qc/0005056.
[29] E. Komatsu, J. Dunkley, M. R. Nolta, C. L. Bennett, B. Gold, G. Hinshaw, N. Jarosik, D. Larson,
M. Limon, L. Page, D. N. Spergel, M. Halpern, R. S. Hill, A. Kogut, S. S. Meyer, G. S. Tucker, J. L.
Weiland, E. Wollack, and E. L. Wright. Five-Year Wilkinson Microwave Anisotropy Probe
Observations: Cosmological Interpretation. The Astrophysical Journal Supplement Series, 180(2):330,
Feb 2009. DOI 10.1088/0067-0049/180/2/330. URL http://arxiv.org/abs/0803.0547.
[30] A. Kreienbuehl. Quantum Cosmology, Polymer Matter, and Modified Collapse. PhD thesis,
University of New Brunswick, Fredericton Campus, Department of Mathematics and Statistics, Aug
2011.
[31] A. Kreienbuehl, V. Husain, and S. S. Seahra. Modified general relativity as a model for quantum
gravitational collapse. Classical and Quantum Gravity, 29(9):095008, May 2012. DOI
10.1088/0264-9381/29/9/095008. URL http://arxiv.org/abs/1011.2381.
[32] A. Kreienbuehl, A. Naegel, D. Ruprecht, R. Speck, G. Wittum, and R. Krause. Numerical simulation
of skin transport using Parareal. Computing and Visualization in Science, 17:99 – 108, 2015. URL
http://dx.doi.org/10.1007/s00791-015-0246-y.
[33] J.-L. Lions, Y. Maday, and G. Turinici. A ”parareal” in time discretization of PDE’s. Comptes
Rendus de l’Acadmie des Sciences - Series I - Mathematics, 332:661–668, 2001. URL
http://dx.doi.org/10.1016/S0764-4442(00)01793-6.
[34] F. Loeffler, J. Faber, E. Bentivegna, T. Bode, P. Diener, R. Haas, I. Hinder, B. C. Mundim, C. D.
Ott, E. Schnetter, G. Allen, M. Campanelli, and P. Laguna. The Einstein Toolkit: a community
computational infrastructure for relativistic astrophysics. Classical and Quantum Gravity, 29(11):
115001, May 2012. DOI 10.1088/0264-9381/29/11/115001. URL
http://arxiv.org/abs/1111.3344.
[35] M. L. Minion. A Hybrid Parareal Spectral Deferred Corrections Method. Communications in Applied
Mathematics and Computational Science, 5(2):265–301, 2010. URL
http://dx.doi.org/10.2140/camcos.2010.5.265.
[36] A. S. Nielsen and J. S. Hesthaven. Fault tolerance in the parareal method. In Proceedings of the
ACM Workshop on Fault-Tolerance for HPC at Extreme Scale, FTXS ’16, pages 1–8, New York, NY,
USA, 2016. ACM. ISBN 978-1-4503-4349-7. DOI 10.1145/2909428.2909431. URL
http://dx.doi.org/10.1145/2909428.2909431.
[37] J. Nievergelt. Parallel methods for integrating ordinary differential equations. Commun. ACM, 7(12):
731–733, 1964. URL http://dx.doi.org/10.1145/355588.365137.
[38] E. Poisson. A Relativist’s Toolkit. The Mathematics of Black-Hole Mechanics. Cambridge University
Press, Cambridge, UK, 2004.
[39] F. Pretorius. Numerical Simulations of Gravitational Collapse. PhD thesis, The University of British
Columbia, 2002.
[40] F. Pretorius. Evolution of Binary Black-Hole Spacetimes. Physical Review Letters, 95(12):121101,
Sep 2005. DOI 10.1103/PhysRevLett.95.121101. URL http://arxiv.org/abs/gr-qc/0507014.
[41] F. Pretorius and L. Lehner. Adaptive mesh refinement for characteristic codes. Journal of
Computational Physics, 198(1):10–34, 2004. DOI 10.1016/j.jcp.2004.01.001. URL
http://arxiv.org/abs/gr-qc/0302003.
[42] D. Ruprecht and R. Krause. Explicit parallel-in-time integration of a linear acoustic-advection
system. Computers & Fluids, 59(0):72–83, 2012. URL
http://dx.doi.org/10.1016/j.compfluid.2012.02.015.
[43] R. Speck and D. Ruprecht. Toward fault-tolerant parallel-in-time integration with {PFASST}.
Parallel Computing, pages –, 2016. DOI 10.1016/j.parco.2016.12.001. URL
http://dx.doi.org/10.1016/j.parco.2016.12.001.
[44] R. Speck, D. Ruprecht, R. Krause, M. Emmett, M. L. Minion, M. Winkel, and P. Gibbon. A
massively space-time parallel N-body solver. In Proceedings of the International Conference on High
Performance Computing, Networking, Storage and Analysis, SC ’12, pages 92:1–92:11, Los Alamitos,
CA, USA, 2012. IEEE Computer Society Press. URL http://dx.doi.org/10.1109/SC.2012.6.
16
TIME PARALLEL GRAVITATIONAL COLLAPSE SIMULATION
[45] J. Thornburg. Adaptive mesh refinement for characteristic grids. General Relativity and Gravitation,
43(5):1211–1251, May 2011. DOI 10.1007/s10714-010-1096-z. URL
http://arxiv.org/abs/0909.0036.
[46] J. Ziprick and G. Kunstatter. Dynamical singularity resolution in spherically symmetric black hole
formation. Physical Review D, 80(2):024032, Jul 2009. DOI 10.1103/PhysRevD.80.024032. URL
http://arxiv.org/abs/0902.3224.
| 5cs.CE
|
Toward Depth Estimation Using Mask-Based
Lensless Cameras
M. Salman Asif
arXiv:1711.03527v1 [cs.CV] 9 Nov 2017
Department of Electrical and Computer Engineering
University of California, Riverside
Email: [email protected]
Abstract—Recently, coded masks have been used to demonstrate a thin form-factor lensless camera, FlatCam, in which a
mask is placed immediately on top of a bare image sensor. In
this paper, we present an imaging model and algorithm to jointly
estimate depth and intensity information in the scene from a
single or multiple FlatCams. We use a light field representation
to model the mapping of 3D scene onto the sensor in which light
rays from different depths yield different modulation patterns.
We present a greedy depth pursuit algorithm to search the 3D
volume and estimate the depth and intensity of each pixel within
the camera field-of-view. We present simulation results to analyze
the performance of our proposed model and algorithm with
different FlatCam settings.
I. I NTRODUCTION
Lens-based cameras are standard vision sensors in system
that records visual information. However, lens-based cameras
are bulky, heavy, and rigid—partly because of the size and
material of a lens. The shape of a lens-based camera is also
fixed in a cube-like form because of the physical constraints
on placing the lens at a certain distance from the sensor. A
lensless camera can potentially be very thin and lightweight,
can operate over a large spectral range, can provide an
extremely wide field of view, and can have curved or flexible
shape.
Recently, a new lensless imaging system, called FlatCam,
was proposed in [1]. FlatCam consists of a coded binary mask
placed at a small distance from a bare sensor. FlatCam can be
viewed as an example of a coded aperture system in which the
mask was placed extremely close to the sensor [2]. The mask
pattern was selected in a way that the image formation model
takes a linear separable form. Image reconstruction from the
sensor measurements requires solving a linear inverse problem.
One limitation of the imaging model in [1] is that it assumes
a 2D scene that consists of a single plane at a fixed distance
from the camera. In this paper, we present a new imaging
model for FlatCam in which the scene consists of multiple
planes at different (unknown) depths. We use a light field
representation in which light rays from different depths yield
different modulation patterns. We use the lightfield representation to analyze the sensitivity of FlatCam to the sampling
pattern and depth mismatch. We present a greedy algorithm
that jointly estimates the depths and intensity of each pixel.
We present simulation results to demonstrate the performance
of our algorithm under different settings.
Fig. 1: 3D imaging with a mask-based lensless camera that consists of a bare
sensor with a fixed, binary mask on top of it. Every light source from within
the camera field-of-view casts a shadow of the mask on sensor, resulting in a
multiplexed image on the sensor. The shadow of any light source depends on
its 3D location with respect to the mask-sensor assembly. A depth-selective
pursuit algorithm reconstructs the 3D image of the scene.
II. BACKGROUND AND RELATED WORK
A pinhole camera is a classical example of a lensless camera
in which an opaque mask with a single pinhole is placed
in front of a light-sensitive surface. A pinhole camera can
potentially take an arbitrary shape. However, a major drawback
of a pinhole camera is that it only allows a tiny fraction of the
ambient light to pass through the single pinhole; therefore, it
typically requires very long exposure times. Coded aperture
imaging systems extend the idea of a pinhole camera by
using a mask with multiple pinholes [3], [2], [4], [5], [6],
[7]. However, the image formed on the sensor is a linear
superposition of images from multiple pinholes. We need to
solve an inverse problem to recover the underlying scene
image from sensor measurements. The primary purpose of the
coded aperture is to increase the amount of light recorded at
the sensor.
A coded aperture system offers another advantage by virtue
of encoding light from different directions and depths differently. Note that a bare sensor can provide the intensity of a
light source but not its spatial location. A mask in front of the
sensor encodes directional information of the source in the
sensor measurements. In a coded aperture system, every light
source in the scene casts a unique shadow of the mask onto the
sensor. Therefore, sensor measurements encode information
about locations and intensities of all the light sources in the
scene. Consider a single light source with a dark background;
the image formed on the sensor will be a shadow of the mask.
If we change the angle of the light source, the mask shadow
on the sensor will shift. Furthermore, if we change the depth
of the light source, the size of the shadow will change (see
Figure 2). Thus, we can represent the relationship between
(a)
(c)
Pinhole
mask
Sensor
Coded
mask
Sensor
(b)
(d)
Pinhole
mask
Sensor
Coded
mask
Sensor
Fig. 2: Examples of imaging with pinhole and coded mask-based cameras.
Light rays from all direction hit the mask; rays can only pass through transparent regions (holes). (a,b) Pinhole cameras preserve angular information
but lose depth information as points along the same angle yield identical
images, irrespective of their depths. (c,d) Coded aperture-based cameras
record coded combination of light from different directions and better preserve
depth information.
all the points in the scene and the sensor measurements as a
linear system, which depends on the pattern and the placement
of the mask. We can solve this system using an appropriate
computational algorithm to recover the image of the scene.
The depth-dependent imaging capability in coded aperture
systems is known since the pioneering work in this domain [8],
[2]. The following excerpt in [2] summarizes it well: ”One
can reconstruct a particular depth in the object by treating the
picture as if it was formed by an aperture scaled to the size
of the shadow produced by the depth under consideration.”
However, the classical methods assume that the scene consists
of a single plane at known depth. In this paper, we assume
that the depth scene consists of multiple depth planes and the
true depth map is unknown at the time of reconstruction.
Coded-aperture cameras have traditionally been used for
imaging wavelengths beyond the visible spectrum (e.g., Xray and gamma-ray imaging), for which lenses or mirrors
are expensive or infeasible [3], [2], [4], [5], [6], [7]. Maskbased lensless designs have been proposed for flexible field-ofview selection in [9], compressive single-pixel imaging using
a transmissive LCD panel [10], and separable coded masks
[11]. In recent years, coded masks and light modulators have
been added to lens-based cameras in different configurations
to build novel imaging devices that can capture image and
depth [12] or 4D light field [13], [14] from a single coded
image. Light field imaging by moving a lensless cameras has
been demonstrated in [15] and 3D imaging with a single-shot
diffuser-based lensless camera was recently demonstrated in
[16].
III. F LAT C AM : R EPLACING LENSES WITH CODED MASKS
AND COMPUTATIONS
FlatCam is a coded aperture system that consists of a bare,
planar sensor and a binary mask [1]. Coded-aperture cameras
have traditionally been used for imaging wavelengths beyond
the visible spectrum (e.g., X-ray and gamma-ray imaging), for
which lenses or mirrors are expensive or infeasible [3], [2],
[6]. A bare sensor can provide information about the intensity
of a light source but not its spatial location. By adding a mask
in front of the sensor, we can encode directional information
of the source in the sensor measurements.
The imaging model in [1] assumes that the scene consists of
a single plane parallel to mask-sensor planes. Let us consider
a 1D imaging system, shown in Fig. 3a, in which a 1D mask
is placed at a distance d in front of a 1D sensor array with M
pixels, and the scene consists of a single plane at distance D
from the sensor with N scene pixels. Let us denote a scene
pixel as l(θ), where θ is uniformly distributed along an angular
interval [θ− , θ+ ]. We can represent the measurement at sensor
pixel s as
I(s) =
θ+
X
ϕ(s, θ)l(θ),
(1)
θ=θ−
where ϕ(θ, s) denotes the modulation coefficient of the mask
for a light ray between scene pixel l(θ) and the sensor pixel
at location s. We can write (1) in a compact form as
I = Φl,
(2)
where Φ denotes an M × N system matrix that maps N scene
pixels to M sensor pixels. This is a linear system that we can
solve using an appropriate computational algorithm to recover
the image of the scene (more details can be found in [1]).
IV. D EPTH ESTIMATION USING F LAT C AMS
A. Imaging model
The model in (1) assumes a 2D scene that consists of a
single plane at some fixed depth. The system matrix Φ encodes
mapping of scene points from that plane to the sensor pixels.
In our new model, we consider a 3D scene that consists
of multiple planes, each of which contribute to the sensor
measurements. Without loss of generality, we consider a 1D
imaging system in Fig. 3a and assume that the sensor plane
is centered at the origin and the mask plane is placed in front
of it at distance d. The scene consists of K planes at depths
[D1 , . . . , DK ] and the scene pixels are distributed uniformly
along angles in interval [θ− , θ+ ], as before. We can describe
the measurement at sensor pixel s as
I(s) =
θ+
DK X
X
ϕ(s, θ, D)l(θ, D),
(3)
D=D1 θ=θ−
where ϕ(s, θ, D) denotes modulation coefficient of the mask
for a light ray between light source l(θ, D) and the sensor
pixel at location s. We can write (3) in a compact form as
I=
DK
X
ΦD lD ,
(4)
D=D1
where each lD denotes intensity of N pixels in a plane at depth
D and ΦD is an M × N matrix that represents the mapping
of lD onto the sensor.
For the case of 3D imaging, let us represent the light
distribution as L(θx , θy , z) for z > d. Measurements for a
sensor pixel located at (su , sv ) can then be described using
(a) Imaging system geometry. Mask
and sensor planes are separated by
distance d. Each point in the scene
(at any angle θ and distance D) contributes to the sensor measurements.
Our goal is to jointly estimate depth
and intensity of each pixel within the
field-of-view.
(b) Lightfield representation of the
system. Angle and depth of a scene
point encode intercept and depth
of the respective line in lightfield.
Horizontal lines denote mask patten.
Lightfield is first modulated by the
mask pattern and then integrated at
the sensor plane.
Fig. 3: Geometry of the imaging system in 1D and the corresponding light
field representation.
the following system of linear equations:
I(su , sv ) =
X
L(θx , θy , z)×
θx ,θy ,z
mask
1−
d
z
d
su + d sin θx , 1 −
sv + d sin θy ,
z
(5)
where mask[u, v] denotes the transparency value of the mask
at location (u, v) within the mask plane. If the mask pattern is
symmetric and separable in (θx , θy ) space, then we can write
the 2D measurements in (5) as
I=
DK
X
ΦD LD ΦTD ,
(6)
D=D1
where ΦD is an M × N matrix and LD denotes N × N light
distribution corresponding to plane at depth D.
Furthermore, we can include multiple cameras at different
locations and orientations with respect to a reference frame.
Such a system will provide us multiple sensor measurements
of the form
DK
X
Ic =
ΦD,c LD ΦTD,c ,
(7)
D=D1
where Ic denotes sensor measurements at camera c and ΦD,c
is matrix that represents mapping of LD onto camera c.
B. Joint image and depth reconstruction
We estimate the depth and intensity of each pixel within the
field-of-view of our cameras using a greedy depth-selective
algorithm, in which we assume a sparse prior that L(θx , θy , D)
has nonzero value only for one depth. Our proposed algorithm
is inspired by structured sparse recovery algorithms in modelbased compressive sensing [17].
To simplify the presentation, let us represent (6) or (7) as
the following general linear system:
I = A(L),
(8)
where L is an N × N × K light distribution with N × N
spatial resolution and K depth planes, A denotes the linear
measurement operator in (6) or (7), and I denotes the sensor
measurements.
Suppose our current estimate of L is L̃ with exactly one
depth assigned to each pixel. Let us denote the initial depth
map as Ω. In our experiments, we initialize the depth estimate
with the farthest plane in the scene. Our proposed depthselective pursuit algorithm is an iterative method that performs
the following three main steps at every iteration:
Compute proxy depth estimate. We first select new candidate
depth for each pixel by picking the maximum magnitude
in the following proxy map corresponding to each pixel:
P = AT [I − A(L̃)]. Let us denote the new depth map as
Ω̃.
Merge depths and estimate image. We first merge the
original depth estimate with the proxy depth estimate. Let us
denote the merged depth support as T = {Ω ∪ Ω̃}. Then we
solve a least-squares problem over the merged depth support
as L̂ = arg minL kIT − AT Lk22 .
Prune depth and threshold image. We prune the depth
estimate at every spatial location by picking the depth corresponding to higher pixel intensity in L̂. Finally, we threshold
L̂ to L̃ that has only one nonzero depth per spatial location.
C. Depth sampling and sensitivity
For a single camera, sensor measurements for a single point
source at location (θ, D) can be described as
d
s + d sin θ .
(9)
I(s) ∝ mask 1 −
D
From this lightfield expression we note that the slope of a line
corresponding to any light source is inversely proportional to
its depth. As a light source moves farther from the mask-sensor
assembly, its line would rotate around the center. As a light
source moves along a plane at a fixed depth, its line would
shift with the same slope. Therefore, in our imaging model, we
select depth planes in a given range by sampling lightfield at
uniform angles, which results in planes at non-uniform depths.
V. E XPERIMENTAL RESULTS
To validate the performance of our proposed imaging model
and reconstruction algorithm, we performed extensive simulations under different settings of FlatCam parameters.
First, we show results of a simple simulation in which
our scene consists of three depth planes as shown in Fig. 4.
We simulated an imaging system in which a binary mask
is placed at 1mm distance from the sensor. We simulated
a 3D voxel space with ten depth planes within a depth
range of 100mm and 3m. We chose the ten depth planes by
uniformly sampling the lightfield representation. We simulated
a scene with 128 × 128 spatial resolution, a mask with a
binary random sequence, and a sensor with 256 × 256 pixels.
To generate sensor measurements, we assumed that the 3D
scene consists of three planes chosen at random out of ten
fixed planes. We added a small amount of Gaussian noise
Number of imaging depth planes (K)
5
10
15
20
25
Single camera
33.83
33.58
31.27
30.5
30.99
Three cameras
39.07
39.09
40
39.54
39.58
TABLE I: PSNR (in dB) comparison of reconstructed images for different
number of depth planes and different number of cameras.
(a) Test scene with three cards
placed at three different depths
picked out of K = 10 depth planes
at random.
(b) Image reconstructed by assuming
that the scene consists of a single
plane at a fixed depth.
PSNR = 15 dB
number of tilted planes. The simulation results of three camera
system are also summarized in Table I. An example of image
reconstruction for this case is shown in Fig. 4(d).
R EFERENCES
(c) Image reconstructed using
depth-selective algorithm that
jointly estimates the depth and
intensity of each pixel.
PSNR = 33.5 dB
(d) Image reconstructed with three
cameras via depth-selective algorithm that jointly estimates the
depth and intensity of each pixel.
PSNR = 39 dB
Fig. 4: Simulation results to demonstrate the effect of depth sensitivity and
a result for our joint depth and intensity reconstruction algorithm.
in the true measurements. To reconstruct the 3D scene, we
solved the depth-selective pursuit algorithm described in the
previous section. The results are presented in Fig. 4, where
(a) denotes pixel intensities for three planes in a 3D scene,
(b) denotes image reconstructed by assuming that all pixels
belong to a plane at fixed depth, and (c) denotes images
reconstructed by solving the depth-pursuit algorithm on the
same measurements.
Next we discuss an experiment that demonstrates the robustness of our proposed model and method against mismatch
in the locations of the original depth planes and those used
for reconstruction. In this experiment, we simulated imaging
system with K = {5, 10, 15, 20, 25} depth planes chosen
uniformly in the lightfield representation. We selected the three
depth planes for the scene at random and calculated the peak
signal to noise ratio (PSNR) for the recovered images. We
present PSNR for each test (averaged over ten instances) in
Table I. We see that the quality of reconstruction remains
almost the same as we increase the number of depth planes in
our model. The computational complexity, however, slightly
increases as we increase the number of depth planes.
Finally, we present an experiment in which we simulated
a system with three lensless cameras in a convex geometry,
where one camera is used as a reference to generate depth
planes in the scene. The other two cameras see tilted planes
in their field of view. An advantage of such a configuration
is that the depth information of the pixels is converted into
angular information. However, this configuration also makes
a strong assumption that the scene only consists of the finite
[1] M. S. Asif, A. Ayremlou, A. Sankaranarayanan, A. Veerarghavan, and
R. Baraniuk, “Flatcam: Thin, bare-sensor cameras using coded aperture
and computation,” IEEE Transactions on Computational Imaging, vol. 3,
no. 3, pp. 384–397, Sept 2017. 1, 2
[2] E. Fenimore and T. Cannon, “Coded aperture imaging with uniformly
redundant arrays,” Applied optics, vol. 17, no. 3, pp. 337–347, 1978. 1,
2
[3] R. Dicke, “Scatter-hole cameras for x-rays and gamma rays,” The
Astrophysical Journal, vol. 153, p. L101, 1968. 1, 2
[4] T. Cannon and E. Fenimore, “Coded aperture imaging: Many holes make
light work,” Optical Engineering, vol. 19, no. 3, pp. 193–283, 1980. 1,
2
[5] P. Durrant, M. Dallimore, I. Jupp, and D. Ramsden, “The application
of pinhole and coded aperture imaging in the nuclear environment,”
Nuclear Instruments and Methods in Physics Research Section A:
Accelerators, Spectrometers, Detectors and Associated Equipment, vol.
422, no. 1, pp. 667–671, 1999. 1, 2
[6] D. J. Brady, Optical imaging and spectroscopy. John Wiley & Sons,
2009. 1, 2
[7] A. Busboom, H. Elders-Boll, and H. Schotten, “Uniformly redundant
arrays,” Experimental Astronomy, vol. 8, no. 2, pp. 97–123, 1998. 1, 2
[8] H. H. Barrett and F. A. Horrigan, “Fresnel zone plate imaging of gamma
rays; theory,” Appl. Opt., vol. 12, no. 11, pp. 2686–2702, Nov 1973.
[Online]. Available: http://ao.osa.org/abstract.cfm?URI=ao-12-11-2686
2
[9] A. Zomet and S. K. Nayar, “Lensless imaging with a controllable
aperture,” in IEEE Computer Society Conference on Computer Vision
and Pattern Recognition, vol. 1, 2006, pp. 339–346. 2
[10] G. Huang, H. Jiang, K. Matthews, and P. Wilford, “Lensless imaging by
compressive sensing,” in 20th IEEE International Conference on Image
Processing, 2013, pp. 2101–2105. 2
[11] M. J. DeWeert and B. P. Farm, “Lensless coded-aperture imaging with
separable doubly-Toeplitz masks,” Optical Engineering, vol. 54, no. 2,
pp. 023 102–023 102, 2015. 2
[12] A. Levin, R. Fergus, F. Durand, and W. T. Freeman, “Image and depth
from a conventional camera with a coded aperture,” in ACM Transactions
on Graphics (TOG), vol. 26, no. 3. ACM, 2007, p. 70. 2
[13] A. Veeraraghavan, R. Raskar, A. Agrawal, A. Mohan, and J. Tumblin,
“Dappled photography: Mask enhanced cameras for heterodyned light
fields and coded aperture refocusing,” ACM Transactions on Graphics
(TOG), vol. 26, no. 3, p. 69, 2007. 2
[14] K. Marwah, G. Wetzstein, Y. Bando, and R. Raskar, “Compressive
light field photography using overcomplete dictionaries and optimized
projections,” ACM Transactions on Graphics (TOG), vol. 32, no. 4, p. 46,
2013. 2
[15] C. Zhang and T. Chen, “Light field capturing with lensless cameras,”
in IEEE International Conference on Image Processing (ICIP), vol. 3.
IEEE, 2005, pp. III–792. 2
[16] N. Antipa, G. Kuo, R. Heckel, B. Mildenhall, E. Bostan, R. Ng, and
L. Waller, “Diffusercam: Lensless single-exposure 3d imaging,” arXiv
preprint arXiv:1710.02134, 2017. 2
[17] R. G. Baraniuk, V. Cevher, M. F. Duarte, and C. Hegde, “Modelbased compressive sensing,” IEEE Transactions on Information Theory,
vol. 56, no. 4, pp. 1982–2001, 2010. 3
| 1cs.CV
|
M ODULAR C ONTINUAL L EARNING IN A U NIFIED
V ISUAL E NVIRONMENT
Blue Sheffer
Department of Psychology
Stanford University
Stanford, CA 94305
[email protected]
arXiv:1711.07425v2 [cs.LG] 12 Dec 2017
Kevin T. Feigelis
Department of Physics
Stanford Neurosciences Institute
Stanford University
Stanford, CA 94305
[email protected]
Daniel L. K. Yamins
Departments of Psychology and Computer Science
Stanford Neurosciences Institute
Stanford University
Stanford, CA 94305
[email protected]
A BSTRACT
A core aspect of human intelligence is the ability to learn new tasks quickly and
switch between them flexibly. Here, we describe a modular continual reinforcement
learning paradigm inspired by these abilities. We first introduce a visual interaction
environment that allows many types of tasks to be unified in a single framework.
We then describe a reward map prediction scheme that learns new tasks robustly
in the very large state and action spaces required by such an environment. We
investigate how properties of module architecture influence efficiency of task
learning, showing that a module motif incorporating specific design principles (e.g.
early bottlenecks, low-order polynomial nonlinearities, and symmetry) significantly
outperforms more standard neural network motifs, needing fewer training examples
and fewer neurons to achieve high levels of performance. Finally, we present a
meta-controller architecture for task switching based on a dynamic neural voting
scheme, which allows new modules to use information learned from previouslyseen tasks to substantially improve their own learning efficiency.
I NTRODUCTION
In the course of everyday functioning, people are constantly faced with real-world environments in which they are required to shift unpredictably between multiple, sometimes unfamiliar,
tasks (Botvinick & Cohen, 2014). They are nonetheless able to flexibly adapt existing decision
schemas or build new ones in response to these challenges (Arbib, 1992). How humans support
such flexible learning and task switching is largely unknown, both neuroscientifically and algorithmically (Wagner et al., 1998; Cole et al., 2013).
We investigate solving this problem with a neural module approach in which simple, task-specialized
decision modules are dynamically allocated on top of a largely-fixed underlying sensory system (Andreas et al., 2015; Hu et al., 2017). The sensory system computes a general-purpose visual representation from which the decision modules read. While this sensory backbone can be large, complex, and
learned comparatively slowly with significant amounts of training data, the task modules that deploy
information from the base representation must, in contrast, be lightweight, quick to be learned, and
easy to switch between. In the case of visually-driven tasks, results from neuroscience and computer
vision suggest the role of the fixed general purpose visual representation may be played by the ventral
visual stream, modeled as a deep convolutional neural network (Yamins & DiCarlo, 2016; Razavian
et al., 2014). However, the algorithmic basis for how to efficiently learn and dynamically deploy
visual decision modules remains far from obvious.
1
controller
agent
ReMaP
module
A
fixed visual backbone
xt
W
H
rt
TouchStream
environment
W
H
policy
at
Figure 1: Modular continual learning in the TouchStream environment The TouchStream environment is a
touchscreen-like GUI for continual learning agents, in which a spectrum of visual reasoning tasks can be posed
in a large but unified action space. On each timestep, the environment (cyan box) emits a visual image (xt ) and a
reward (rt ). The agent recieves xt and rt as input and emits an action at . The action represents a “touch” at
some location on a two-dimensional screen e.g. at ∈ {0, . . . , H − 1} × {0, . . . , W − 1}, where H and W are
the screen height and width. The environment’s policy is a program computing xt and rt as a function of the
agent’s action history. The agent’s goal is to learn how to choose optimal actions to maximize the amount of
reward it recieves over time. The agent consists of several component neural networks including a fixed visual
backbone (yellow inset), a set of learned neural modules (grey inset), and a meta-controller (red inset) which
mediates the deployment of these learned modules for task solving. The modules use the ReMaP algorithm
§ 2 to learn how to estimate reward as a function of action (heatmap), conditional on the agent’s recent history.
Using a sampling policy on this reward map, the agent chooses an optimal action to maximize its aggregate
reward.
In standard supervised learning, it is often assumed that the output space of a problem is prespecified
in a manner that just happens to fit the task at hand – e.g. for a classification task, a discrete output
with a fixed number of classes might be determined ahead of time, while for a continuous estimation
problem, a one-dimensional real-valued target might be chosen instead. This is a very convenient
simplification in supervised learning or single-task reinforcement learning contexts, but if one is
interested in the learning and deployment of decision structures in a rich environment defining tasks
with many different natural output types, this simplification becomes cumbersome.
To go beyond this limitation, we build a unified environment in which many different tasks are
naturally embodied. Specifically, we model an agent interacting with a two-dimensional touchscreenlike GUI that we call the TouchStream, in which all tasks (discrete categorization tasks, continuous
estimation problems, and many other combinations and variants thereof) can be encoded using a
single common and intuitive – albeit large – output space. This choice frees us from having to
hand-design or programmatically choose between different output domain spaces, but forces us
to confront the core challenge of how a naive agent can quickly and emergently learn the implicit
“interfaces” required to solve different tasks.
We then introduce Reward Map Prediction (ReMaP) networks, an algorithm for continual reinforcement learning that is able to discover implicit task-specific interfaces in large action spaces like those
of the TouchStream environment. We address two major algorithmic challenges associated with
learning ReMaP modules. First, what module architectural motifs allow for efficient task interface
learning? We compare several candidate architectures and show that those incorporating certain
intuitive design principles (e.g. early visual bottlenecks, low-order polynomial nonlinearities and
symmetry-inducing concatenations) significantly outperform more standard neural network motifs,
needing fewer training examples and fewer neurons to achieve high levels of performance. Second,
what system architectures are effective for switching between tasks? We present a meta-controller
architecture based on a dynamic neural voting scheme, allowing new modules to use information
learned from previously-seen tasks to substantially improve their own learning efficiency.
In § 1 we formalize the TouchStream environment. In § 2, we introduce the ReMaP algorithm. In § 3,
we describe and evaluate comparative performance of multiple ReMaP module architectures on a
variety of TouchStream tasks. In § 4, we describe the Dynamic Neural Voting meta-controller, and
evaluate its ability to efficiently transfer knowledge between ReMaP modules on task switches.
2
a
stimulus-response
b
c
match-to-sample
MS-COCO MTS
d
localization
image:
or
sample
then
match
sample
then
match
then
reward
map:
Figure 2: Exemplar TouchStream tasks. Illustration of several task paradigms explored in this work using the
TouchStream Environment. The top row depicts observation xt and the bottom shows the ground truth reward
maps (with red indicating high reward and blue indicating low reward). a. Binary Stimulus-Response task. b.
stereotyped Match-To-Sample task. c. The Match-To-Sample task using the MS-COCO dataset. d. Object
localization.
R ELATED W ORK
Modern deep convolutional neural networks have had significant impact on computer vision and
artificial intelligence (Krizhevsky et al., 2012), as well as in the computational neuroscience of vision
(Yamins & DiCarlo (2016)). There is a recent but growing literature on convnet-based neural modules,
where they have been used for solving compositional visual reasoning tasks (Andreas et al., 2015;
Hu et al., 2017). In this work we apply the idea of modules to solving visual learning challenges
in a continual learning context. Existing works rely on choosing between a menu of pre-specified
module primitives, using different module types to solve subproblems involving specific input-output
datatypes, without addressing how these modules’ forms are to be discovered in the first place. In this
paper, we show a single generic module architecture is capable of automatically learning to solve a
wide variety of different tasks in a unified action/state space, and a simple controller scheme is able
to switch between such modules.
Our results are also closely connected with the literature on lifelong (or continual) learning (Kirkpatrick et al., 2016; Rusu et al., 2016). A part of this literature is concerned with learning to solve new
tasks without catastrophically forgetting how to solve old ones (Zenke et al., 2017; Kirkpatrick et al.,
2016). The use of modules obviates this problem, but instead shifts the hard question to one of how
newly-allocated modules can be learned effectively. The continual learning literature also directly
addresses knowlege transfer to newly allocated structures (Chen et al., 2015; Rusu et al., 2016;
Fernando et al., 2017), but largely addresses how transfer learning can lead to higher performance,
rather than addressing how it can improve learning speed. Aside from reward performance, we focus
on issues of speed in learning and task switching, motivated by the remarkably efficient adaptability
of humans in new task contexts. Existing work in continual learning also largely does not address
which specific architecture types learn tasks efficiently, independent of transfer. By focusing first
on identifying architectures that achieve high performance quickly on individual tasks (§ 3), our
transfer-learning investigation then naturally focuses more on how to efficiently identify when and
how to re-use components of these architectures (§ 4). Most of these works also make explicit a
priori assumptions about the structure of the tasks to be encoded into the models (e.g. output type,
number of classes), rather than address the more general question of emergence of solutions in an
embodied case, as we do.
Meta-reinforcement learning approaches such as Wang et al. (2016); Duan et al. (2016), as well as
the schema learning ideas of e.g. Arbib (1992); McClelland (2013) typically seek to address the
issue of continual learning by having a complex meta-learner extract correlations between tasks
over a long timescale. In our context most of the burden of environment learning is placed on the
individual modules, so our meta-controller can thus be comparatively light-weight compared to
typical meta-reinforcement approaches. Unlike our case, meta-learning has mostly been limited to
small state or action spaces. Some recent work in general reinforcement learning (e.g. Ostrovski et al.
(2017); Dulac-Arnold et al. (2015)) has addressed the issue of large action spaces, but has not sought
to address multitask transfer learning in these large action spaces.
1
T HE T OUCH S TREAM E NVIRONMENT
Agents in a real-world environment are exposed to many different implicit tasks, arising without
predefined decision structures, and must learn on the fly what the appropriate decision interfaces are
3
for each situation. Because we are interested in modeling how agents can do this on-the-fly learning,
our task environment should mimic the unconstrained nature of the real world. Here, we describe the
TouchStream environment, which attempts to do this in a simplified two-dimensional domain.
Our problem setup consists of two components, an “environment” and an “agent,” interacting over an
extended temporal sequence (Fig. 1). At each timestep t, the environment emits an RGB image xt of
height H and width W , and a scalar reward rt . Conversely, the agent accepts images and rewards as
input and chooses an action at in response. The action space A available to the agent consists of a
two-dimensional pixel grid {0, . . . , H − 1} × {0, . . . , W − 1} ⊂ Z2 , of the same height and width as
its input image. The environment is equipped with a policy (unknown to the agent) that on each time
step computes image xt and reward rt as a function of the history of agent actions {a0 , . . . , at−1 },
images {x0 , . . . , xt−1 } and rewards {r0 , . . . , rt−1 }.
In this work, the agent is a neural network, composed of a visual backbone with fixed weights, together
with a meta-controller module whose parameters are learned by interaction with the environment.
The agent’s goal is to learn to enact a policy that maximizes its reward obtained over time. Unlike an
episodic reinforcement learning context, the TouchStream environment is continuous: throughout
the course of learning the agent is never signaled when it should reset to some “initial” internal
state. However, unlike the traditional continuous learning context of e.g. Sutton & Barto (1998), a
TouchStream may implicitly define many different tasks, each of which is associated with its own
characteristic reward schedule. The agent experiences a continual stream of tasks, and any implicit
association between reward schedule and state reset must be discovered by the agent.
By framing the action space A of the agent as all possible pixel locations and the state space as any
arbitrary image, a very wide range of possible tasks are unified in this single framework, at the cost
of requiring the agents’ action space to be congruent to its input state space, and thus be quite large.
This presents two core efficiency challenges for the agent: on any given task, it must be able to both
quickly recognize what the “interface” for the task is, and transfer such knowledge across tasks in a
smart way. Both of these goals are complicated by the fact that both the large size of agent’s state and
action spaces.
Although we work with modern large-scale computer vision-style datasets and tasks in this work,
e.g. ImageNet (Deng et al. (2009)) and MS-COCO (Lin et al. (2014)), we are also inspired by visual
psychology and neuroscience, which have pioneered techniques for how controlled visual tasks can
be embodied in real reinforcement learning paradigms (Horner et al., 2013; Rajalingham et al., 2015).
Especially useful are three classes of task paradigms that span a range of the ways discrete and
continuous estimation tasks can be formulated – including Stimulus-Response, Match-To-Sample,
and Localization tasks (Fig. 2).
Stimulus-Response Tasks: The Stimulus-Response (SR) paradigm is a common approach to physically embodying discrete categorization tasks (Gaffan & Harrison, 1988). For example, in the simple
two-way SR discrimination task shown in Fig. 2a, the agent is rewarded if it touches the left half
of the screen after being shown an image of a dog, and the right half after being shown a butterfly.
SR tasks can be made more difficult by increasing the number of image classes or the complexity of
the reward boundary regions. In our SR experiments, we use images and classes from the ImageNet
dataset (Deng et al., 2009).
Match-To-Sample Tasks: The Match-to-Sample (MTS) paradigm is another common approach to
assessing visual categorization abilities (Murray & Mishkin, 1998). In the MTS task shown in Fig.
2b, trials consist of a sequence of two image frames – the “sample” screen followed by the “match”
screen – in which the agent is expected to remember the object category seen on the sample frame,
and then select an onscreen “button” (really, a patch of pixels) on the match screen corresponding to
the sample screen category. Unlike SR tasks, MTS tasks require some working memory and more
localized spatial control. More complex MTS tasks involve more sophisticated relationships between
the sample and match screen. In Fig. 2c, using the MS-COCO object detection challenge dataset (Lin
et al., 2014), the sample screen shows an isolated template image indicating one of the 80 MS-COCO
classes, while the match screen shows a randomly-drawn scene from the dataset containing at least
one instance of the sample-image class. The agent is rewarded if its chosen action is located inside
the boundary of an instance (e.g. the agent “pokes inside”) of the correct class. This MS-COCO
MTS task is a “hybrid” of categorical and continuous elements, meaning that if phrased as a standard
4
supervised learning problem, both categorical readout (i.e. class identity) and a continous readout
(i.e. object location) would be required.
Localization: Fig. 2d shows a two-step continuous localization task in which the agent is supposed
to mark out the bounding box of an object by touching opposite corners on two successive timesteps,
with reward proportionate to the Intersection over Union (IoU) value of the predicted bounding box
GT ∩B̂)
relative to the ground truth bounding box IoU = Area(B
. In localization, unlike the SR and
Area(BGT ∪B̂)
MTS paradigms, the choice made at one timestep constrains the agent’s optimal choice on a future
timestep (e.g. picking the upper left corner of the bounding box on the first step contrains the lower
right opposite corner to be chosen on the second).
Although these tasks can become arbitrarily complex along certain axes, the tasks presented here
require only fixed-length memory and future prediction. That is, each task requires only knowledge
of the past kb timesteps, and a perfect solution always exists within kf timesteps from any point. The
minimal required values of kb and kf are different across the various tasks in this work. However, in
the investigations below, we set these to the maximum required values across tasks, i.e. kb = 1 and kf
= 2. Thus, the agent is required to learn for itself when it is safe to ignore information from the past
and when it is irrelevant to predict past a certain point in the future.
We will begin by considering a restricted case where the environment runs one semantic task indefinitely, showing how different architectures learn to solve such individual tasks with dramatically
different levels of efficiency (§ 2-3). We will then expand to considering the case where the environment’s policy consists of a sequence of tasks with unpredictable transitions between tasks, and
exhibit a meta-controller that can cope effectively with this expanded domain (§ 4).
2
R EWARD M AP P REDICTION
The TouchStream environment necessarily involves working with large action and state spaces.
Methods for handling this situation often focus on reducing the effective size of action/state spaces,
either via estimating pseudo-counts of state-action pairs, or by clustering actions (Ostrovski et al.,
2017; Dulac-Arnold et al., 2015). Here we take another approach, using a neural network to directly
approximate the (image-state modulated) mapping between the action space and reward space,
allowing learnable regularities in the state-action interaction to implicitly reduce the large spaces into
something manageable by simple choice policies. We introduce an off-policy algorithm for efficient
multitask reinforcement learning in large action and state spaces: Reward Map Prediction, or ReMaP.
2.1
R E M A P N ETWORK A LGORITHM
As with any standard reinforcement learning situation, the agent seeks to learn an optimal policy
π = p(at | xt ) defining the probability density p over actions given image state xt . The ReMaP
algorithm is off-policy, in that π is calculated as a simple fixed function of the estimated reward.
A ReMaP network MΘ is a neural network with parameters Θ, whose inputs are a history over
previous timesteps of (i) the agent’s own actions, and (ii) an activation encoding of the agent’s state
space; and which explicitly approximates the expected reward map across its action space for some
number of future timesteps. Mathematically:
h
i
k
MΘ : [Ψt−kb :t , ht−kb :t−1 ] 7−→ m1t , m2t , . . . , mt f
where kb is the number of previous timesteps considered; kf is the length of future horizon to be
considered; Ψt−kb :t is the history [ψ(xt−kb ), . . . , ψ(xt )] of state space encodings produced by fixed
backbone network ψ(·), ht−kb :t−1 is the history [at−kb . . . , at−1 ] of previously chosen actions, and
each mi ∈ map(A, R) – that is, a map from action space to reward space. The predicted reward
maps are constructed by computing the expected reward obtained for a subsample of actions drawn
randomly from A:
Z
j
mt : at 7→ E [rt+j | at , ht−kb :t−1 , Ψt−kb :t ] =
rt+j p(rt+j | at , ht−kb :t−1 , Ψt−kb :t ).
(1)
R
where rt+j is the predicted reward j steps into the future horizon. Having produced kf reward
prediction maps, one for each timestep of its future horizon, the agent needs to determine what
5
h
i
k
it believes will be the single best action over all the expected reward maps m1t , m2t , . . . , mt f .
The ReMaP algorithm formulates doing so by normalizing the predictions across each of these kf
maps into separate probability distributions, and sampling an action from the distribution which has
maximum variance. That is, the agent computes its policy π as follows:
k
f
π = VarArgmaxj=1
{Dist[Norm[mjt ]]},
(2)
N orm[m] = m − min m(x)
(3)
where
x∈A
is a normalization that removes the minimum of the map,
f (m)
f
(m(x))
A
Dist[m] = R
(4)
ensures it is a probability distribution parameterized by functional family f (·), and VarArgmax is
an operator which chooses the input with largest variance.
The sampling procedure described in equation (2) uses two complementary ideas to exploit spatial
and temporal structure to efficiently explore a large action space. Since rewards in real physical
tasks are spatially correlated, the distribution-based sampler in Equation (4) allows for more effective
exploration of potentially informative actions than would the single-point estimate of an apparent
optimum (e.g. an -greedy policy). Further, in order to reduce uncertainty, the ReMaP algorithm
explores timesteps with greatest reward map variance. The VarArgmax function nonlinearly upweights
the timeframe with highest variance to exploit the fact that some points in time carry disproportianate
relevance for reward outcome, somewhat analagously to how max-pooling operates in convolutional
networks. Although any standard action selection strategy can be used in place of the one in (2)
(e.g. pseudo -greedy over all kf maps), we have empirically found that this policy is effective at
efficiently exploring our large action space.
The parameters Θ of a ReMaP network are learned by gradient descent on the loss of the reward
prediction error Θ∗ = argminΘ L [mt (at ), rt , ; Θ] with map mjt compared to the true reward rt+j .
Only the reward prediction in mt corresponding to the action chosen at timestep t participates in
loss calculation and backpropagation of error signals. A minibatch of maps, rewards, and actions is
collected over several consecutive inference passes before performing a parameter update.
The ReMaP algorithm is summarized in 1.
Algorithm 1: ReMaP – Reward Map Prediction
Initialize ReMaP network M
Initialize state and action memory buffers Ψt−kb :t and ht−kb :t−1
for timestep t = 1,T do
Observe xt , encode with state space network ψ(·), and append to state buffer
Subsample set of potential action choices at uniformly from A
Produce kf expected reward maps of at from eq. (1)
Select action according to policy π as in (2)
Execute action at in environment, store in action buffer, and receive reward rt
Calculate loss for this and previous kf − 1 timesteps
if t ≡ 0 mod batch size then
Perform parameter update
Throughout this work, we take our fixed backbone state space encoder to be the VGG-16 convnet,
pretrained on ImageNet (Simonyan & Zisserman, 2014). Because the resolution of the input to this
network is 224x224 pixels, our action space A = {0, . . . , 223} × {0, . . . , 223}. By default, the
functional family f used in the action selection scheme in Eq. (4) is the identity, although on tasks
benefiting from high action precision (e.g. Localization or MS-COCO MTS), it is often optimal to
sample a low-temperature Boltzmann distribution with f (x) = e−x/T . Reward prediction errors are
calculated using the cross-entropy loss (where logits are smooth approximations to the Heaviside
function in analogy to eq. (5)).
6
3
E FFICIENT N EURAL M ODULES FOR TASK L EARNING
The main question we seek to address in this section is: what specific neural network structure(s)
should be used in ReMaP modules? The key considerations are that such modules (i) should be easy
to learn, requiring comparatively few training examples to discover optimal parameters Θ∗ , and (ii)
easy to learn from, meaning that an agent can quickly build a new module by reusing components of
old ones.
Intuitive Example: As an intuition-building example, consider the case of a simple binary StimulusResponse task, as in Fig. 2a (“if you see a dog touch on the right, if a butterfly touch on the left").
One decision module that is a “perfect” reward predictor on this task is expressed analytically as:
M [Ψt ](ax , ay ) = H(ReLU(W Ψt ) · ReLU(ax ) + ReLU(−W Ψt ) · ReLU(−ax ))
(5)
where H is the Heaviside function, ax and ay are the x and y components of the action a ∈ A relative
to the center of the screen, and W is a 1 × |Ψt | matrix expressing the class boundary (bias term
omitted for clarity). If W Ψt is positive (i.e. the image is of a dog) then ax must also be positive (i.e.
touch is on the right) to predict positive reward; conversly, if W Ψt is negative (i.e. butterfly), ax
must be negative (i.e. left touch) to predict reward. If neither of these conditions hold, both terms are
equal to zero, so the formula predicts no reward. Since vertical location of the action does not affect
reward, ay is not involved in reward calculation on this task.
Equation (5) has three basic ideas embedded in its structure:
• there is an early visual bottleneck, in which the high-dimensional general purpose feature representation Ψt is greatly reduced in dimension (in this case, from the 4096 features of VGG’s FC6
layer, to 1) prior to combination with action space,
• there is a multiplicative interaction between the action vector and (bottlenecked) visual features,
and
• there is symmetry, e.g. the first term of the formula is the sign-antisymmetric partner of the second
term, reflecting something about the spatial structure of the task.
In the next sections, we show these three principles can be generalized into a parameterized family of
networks from which the visual bottleneck (the W parameters), and decision structure (the form of
equation (5)) can emerge naturally and efficienty via learning for any given task of interest.
3.1
T HE EMS M ODULE
In this section we define a generic ReMaP module which is lightweight, encodes all three generic
design principles from the “perfect” formula, and uses only a small number of learnable parameters.
Define the concatenated square nonlinearity as
Sq : x 7−→ x ⊕ x2
and the concatenated ReLU nonlinearity (Shang et al. (2016)) as
CReLU : x 7−→ ReLU(x) ⊕ ReLU(−x)
where ⊕ denotes vector concatenation. The CReS nonlinearity is then defined as the composition of
CReLU and Sq, e.g.
CReS(x) : x 7−→ ReLU(x) ⊕ ReLU(−x) ⊕ ReLU2 (x) ⊕ ReLU2 (−x).
The CReS nonlinearity introduces multiplicative interactions between its arguments via its Sq
component and symmetry via its use of CReLU.
Definition. The (n0 , n1 , . . . , nk )-Early Bottleneck-Multiplicative-Symmetric (EMS) module is the
ReMaP module given by
B = CReLU(W0 Ψ + b0 )
l1 = CReS(W1 (B ⊕ a) + b1 )
li = CReS(Wi li−1 + bi ) for i > 1
where Wi and bi are learnable parameters, Ψ are features from the fixed visual encoding network,
and a is the action vector in A.
7
Sample
Match
Reward Maps
a
4
6
18
83
125
96
115
140
1536
b
64
Training episode (in thousands)
Figure 3: Decision interfaces emerge naturally over the course of training. The ReMaP modules allow the
agent to discover the implicit interfaces for each task. We observe that learning generally first captures the
emergence of natural physical constructs before learning task-specific decision rules. Examples of this include:
a. onscreen “buttons” appearing on the match screen of an MTS task before the specific semantic meaning
of each button is learned (arrows indicate random motion), and b. the general discovery of objects and their
boundaries before the task-specific category rule is applied. This image is best viewed in color.
The EMS structure builds in each of the three principles described above. The B stage represents the
early bottleneck in which visual encoding inputs are bottlenecked to size n0 before being combined
with actions, and then performs k CReS stages, introducing multiplicative symmetric interactions
between visual features and actions. From this, the “perfect” module definition for the binary SR
task in eq. (5) then becomes a special case of a two-layer EMS module. Note that the visual features
to be bottlenecked can be from any encoder; in practice, we work with both fully connected and
convolutional features of the VGG-16 backbone.
In the experiments that follow, we compare the EMS module to a wide variety of alternative control
motifs, in which the early bottleneck, multiplicative, and symmetric features are ablated. Multiplicative nonlinearity and bottleneck ablations use a spectrum of more standard activation functions,
including ReLU, tanh, sigmoid, elu (Clevert et al., 2015), and CReLU forms. In late bottleneck
(fully-ablated) architectures – which are, effectively, “standard” multi-layer perceptrons (MLPs)
– action vectors are concatenated directly to the output of the visual encoder before being passed
through subsequent stages. In all, we test 24 distinct architectures. Detailed information on each can
be found in the Supplement.
3.2
E XPERIMENTS
We compared each architecture across 12 variants of visual SR, MTS, and localization tasks, using
fixed visual encoding features from layer FC6 of VGG-16. Task variants ranged in complexity from
simple (e.g. a binary SR task with ImageNet categories) to more challenging (e.g. a many-way
ImageNet MTS task with result buttons appearing in varying positions on each trial). The most
complex tasks are two variants of localization, either with a single main salient object placed on
a complex background (similar to images used in Yamins & DiCarlo (2016)), or complex scenes
from MS-COCO (see Fig. 3b). Details of the tasks used in these experiments can be found in the
Supplement. Module weights were initialized using a normal distribution with µ = 0.0, σ = 0.01, and
optimized using the ADAM algorithm (Kingma & Ba (2014)) with parameters β1 = 0.9, β2 = 0.999
and = 1e−8. Learning rates were optimized on a per-task, per-architecture basis in a cross-validated
fashion. For each architecture and task, we ran optimizations from five different initialization seeds to
obtain mean and standard error due to initial condition variability. For fully-ablated “late-bottleneck”
modules, we measured the performance of modules of three different sizes (small, medium, and
large), where the smallest version is equivalent in size to the EMS module, and the medium and large
versions are much larger (Table S1).
Emergence of Decision Structures: A key feature of ReMaP modules is that they are able to
discover de novo the underlying output domain spaces for a variety of qualitatively distinct tasks (Fig.
3; more examples in Fig. S2). The emergent decision structures are highly interpretable and reflect
the true interfaces that the environment implicitly defines. The spatiotemporal patterns of learning are
robust across tasks and replicable across initial seedings, and thus might serve as a candidate model
of interface use and learning in humans. In general, we observe that the modules typically discover
8
Reward
a
b
1.0
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
Reward
c
1.0
0.9
e TA-N-AUC
0.2
20
40
60
80
100
120
140
d
1.0
50
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
100
200
300
400
500
150
200
1.0
0.9
Training episodes (in thousands)
100
200
400
600
800
Module
0.997 ± 0.006
EMS
0.944 ± 0.059
No symm
0.818 ± 0.122
No mult
0.717 ± 0.127 None (large)
0.566 ± 0.184 No mult/symm
0.522 ± 0.213 None (medium)
0.403 ± 0.170 None (small)
1000
Training episodes (in thousands)
Figure 4: EMS modules as components of an efficient visual learning system. Validation reward obtained
over the course of training for modules on a. 4-way stimulus-response with a reward map split into four
quadrants, b. 2-way MTS with randomly moving match templates, c. 4-way MTS with two randomly moving
class templates shown at a time, and d. 4-way MTS with four randomly positioned images shown at a time.
Lines indicate mean reward over five different weight initializations. For clarity, seven of the total 24 tested
architectures are displayed (see results for remaining architectures in Supplement). e. The TA-N-AUC metric
is the area under the learning curve, normalized to the highest performing module within a task (over all 24
modules), averaged across all 12 tasks. Error values are standard deviations from the mean TA-N-AUC.
the underlying “physical structures” needed to operate the task interface before learning the specific
decision rules needed to solve the task.
For example, in the case of a discrete MTS categorization task (Fig. 3a), this involves the quick
discovery of onscreen “buttons” corresponding to discrete action choices before these buttons are
mapped to their semantic meaning. In the case of in the MS-COCO MTS task (Fig. 3b), we observe
the initial discovery of high salience object boundaries, and followed by category-specific refinement.
It is important to note that the visual backbone was trained on a categorization task, quite distinct
from the localization task in MS-COCO MTS. Thus, the module had to learn this very different
decision structure, as well as the class boundaries of MS-COCO, from scratch during training.
Efficiency of the EMS module: The efficiency of learning was measured by computing the taskaveraged, normalized area under the learning curve (TA-N-AUC) for each of the 24 modules tested,
across all 12 task variants. Fig. 4a-d shows characteristic learning curves for several tasks, summarized in the table in Fig. 4e. Results for all architectures for all tasks are shown in Supplement
Figure S1. We find that the EMS module is the most efficient across tasks (0.997 TA-N-AUC).
Moreover, the EMS architecture always achieves the highest final reward level on each task.
Increasing ablations of the EMS structure lead to increasingly poor performance, both in terms of
learning efficiency and final performance. Ablating the low-order polynomial interaction (replacing
Sq with CReLU) had the largest negative effect on performance (0.818 TA-N-AUC), followed in
importance by the symmetric structure (0.944 TA-N-AUC). Large fully-ablated models (no bottleneck,
using only ReLU activations) performed significantly worse than the smaller EMS module and
the single ablations (0.717 TA-N-AUC), but better than the module with neither symmetry nor
multiplicative interactions (0.566 TA-N-AUC). Small fully-ablated modules with the same number of
parameters as EMS were by far the least efficient (0.403 TA-N-AUC) and oftentimes achieved much
lower final reward. In summary, the main conceptual features by which the special-case architecture
in eq. (5) solves the binary SR task are both individually helpful, combine usefully, and can be
parameterized and efficiently learned for a variety of visual tasks. These properties are critical to
achieving effective task learning compared to standard MLP structures.
In a second experiment focusing on localization tasks, we tested an EMS module using convolutional
features from the fixed VGG-16 feature encoder, reasoning that localization tasks could benefit from
finer spatial feature resolution. We find that using visual features with explicit spatial information
9
a
b
0.6
0.5
0.7
0.6
Convolutional EMS
EMS
No symmetry
No multiplication
No multiplication/symmetry
None (large)
None (medium)
None (small)
SVM baseline
Reward
IOU
0.4
0.3
0.2
0.5
0.4
0.3
0.1
1000
2000
3000
4000
5000
6000
Training episodes (in thousands)
10000
20000
30000
40000
50000
Training episodes (in thousands)
Figure 5: Convolutional bottlenecks allow for fine resolution localization and detection in complex scenes.
a. Mean Intersection over Union (IoU) obtained on the localization task. b. Reward obtained on the MS-COCO
match-to-sample variant. Both of these require their visual systems to accomodate for finer spatial resolution
understanding of the scene, and more precise action placement than the SR or (non-COCO) MTS tasks. The
convolutional variant of the EMS module uses skip connections from the conv5 and FC6 layers of VGG-16 as
input, whereas the standard EMS uses only the FC6 layer as input.
substantially improves task performance and learning efficiency on these tasks (Fig. 5). To our
knowledge, our results on MS-COCO are the first demonstrated use of reinforcement learning to
achieve instance-level object segmentations. Reward curves (measuring bounding box IoU) in Fig.
5a show little difference between any of the late bottleneck modules at any size. The only models
to consistently achieve an IoU above 0.4 are the EMS-like variants, especially with convolutional
features. For context, a baseline SVR trained using supervised methods to directly regress bounding
boxes using the same VGG features results in an IoU of 0.369.
4
DYNAMIC N EURAL VOTING FOR TASK S WITCHING
So far, we’ve considered the case where the TouchStream consists of only one task. However, agents
in real environments are often faced with having to switch between tasks, many of which they may
be encountering for the first time. Ideally, such agents would repurpose knowledge from previously
learned tasks when it is relevant to a new task.
Formally, we now consider environment policies consisting of sequences of tasks T = {τ1 , τ2 , ..., τΩ },
each of which may last for an indeterminate period of time. Consider also a set of modules M, where
each module corresponds to a task-specific policy πω (a | x) = p (at | xt , τω ). When a new task
begins, we cue the agent to allocate a new module MΩ+1 which is added to the set of modules M. In
the learning that follows allocation, the weights in old modules are held fixed while the parameters in
the new module MΩ+1 are trained. However, the output of the system is not merely the output of the
new module, but instead is a dynamically allocated mixture of pathways through the computation
graphs of the old and new modules. This mixture is determined by a meta-controller (Fig. 6). The
meta-controller is itself a neural network which learns a dynamic distribution over (parts of) modules
to be used in building the composite execution graph. Intuitively, this composite graph is composed
of a small number of relevant pathways that mix and match parts of existing modules to solve the
new task, potentially in combination with new module components that need to be learned.
4.1
DYNAMIC N EURAL VOTING
We define a meta-controller that assigns weights to each layer in each module in M. Let piω be the
weight associatedP
with the ith layer in module ω. These weights are probabilistic on a per layer basis,
e.g. piω ≥ 0 and ω piω = 1 and can be interpreted as the probability of the controller selecting the
ith layer lωi for use in the execution graph, with distribution πi = {piω }. For such an assignment of
weights, the composite execution graph defined by the meta-controller is generated by computing the
sum of the activations of all the components at layer i weighted by the probabilities piω . These values
are then passed on to the next layer where this process repeats. Mathematically, the composite layer
at stage i can be expressed as
h
i
X
i
i ˜i−1
i ˜i−1
˜li =
p
M
(
l
)
=
E
M
(
l
)
.
(6)
π
ω ω M
ω M
M
i
ω
10
controller
p1ω Mω1
p2ω Mω2
p1ω Mω1
p2ω Mω2
+
+
visual
input
˜l0
M
composite
layer
˜l1
M
A
= fixed parameters
˜l2
M
Figure 6: The Dynamic Neural Voting Controller. Dynamic Neural Voting solves new tasks by computing a
composite execution graph through previously learned and newly allocated modules. Shown here is an agent
with two existing modules (yellow and green), as one newly allocated module is being learned (blue). For each
layer, the controller takes as input the activations of all three modules and outputs a set of “voting results” —
probabilistic weights to be used to scale activations within the corresponding components. Voting can be done
on either a per-layer basis or a per-unit basis, for clarity only the layer voting method is depicted. The weighted
sum of these three scaled outputs is used as input to the next stage in the computation graph. If the new task
can be solved through a combination of existing module components, these will be weighted highly, while the
new module will be effectively unused e.g. is assigned low weights. If however the task is quite different than
previously solved tasks, the new module will play a larger role in the execution graph as it learns to solve the
task.
where Mωi (·) is the operator that computes the ith layer of module ω, and ˜l0 := ψ(xt ) is the original
encoded input state.
The question now is, where do these probabilistic weights come from? The core of our procedure
is a dynamic neural voting process in which the controller network learns a Boltzmann distribution
over module activations to maximize reward prediction accuracy. This process is performed at each
module layer, where the module weightings for a given layer are conditioned on the results of voting
at the previous layer. That is,
"
!
#
M
i
i
i ˜i−1
i
p = softmax W
M (l ) + b
(7)
ω
ω
M
where p i = (pi0 , pi1 , ..., piΩ ) are the module weights at layer i, ⊕ is concatenation,
and W i ∈
1
1
R(Ω·L)×Ω is a learnable weight matrix of the controller.
This voting procedure operates in an online fashion,
such that the controller is continously learning
1
1
its meta-policy while the agent is taking actions. As defined, the meta-controller constitutes a
fully-differentiable neural network and is learned by gradient descent online.
1 mechanism involves voting across the units of M. Specifically, the
A useful refinement of the above
i,j 1
meta-controller now assigns probabilistic weights pi,j
ω to neuron nω (the jth unit in layer i of module
ω). In contrast to the layer-voting scheme, the dynamically generated execution graph computed 1by
the meta controller now becomes composite neurons with activations:
h
i
X
i,j ˜i−1
i,j ˜i−1
ñi,j
pi,j
(8)
ω Mω (lM ) = Eπi ,j Mω (lM ) .
M =
ω
i
which are concatenated to form the composite layer ˜lM
. The generalization of equation (7) to the
single-unit voting scheme then becomes:
"
!
#
M
i,j
i,j
i,j ˜i−1
i,j
p = softmax W
Mω (lM ) + b
(9)
ω
i,j
i,j
i,j
where p i,j = (pi,j
∈ RΩ×Ω .
0 , p1 , ..., pΩ ) are the unit-level weights across modules, and W
11
Empirically, we find that the initialization schemes of the learnable controller parameters are an
important consideration in the design, and that two specialized transformations also contribute slightly
to its overall efficiency. For details on these, please refer to the Supplement.
The dynamic neural voting mechanism achieves meta-control through a neural network optimized
online via gradient descent while the modules are solving tasks, rather than a genetic algorithm that
operates over a longer timescale as in the work of Fernando et al. (2017). Moreover, in contrast to the
work of Rusu et al. (2016) the voting mechanism eliminates the need for fully-connected adaptation
layers between modules, thus substantially reducing the number of parameters required for transfer.
Reward
4.2
S WITCHING E XPERIMENTS
1.0
0.8
0.6
0.4
From Scratch
After Switch
Reuse
Fraction
0.2
1.0
0.8
0.6
20
30
40
Batch Updates
Figure 7: Dyanmic Neural Voting quickly corrects
for “no-switch” switches. Although a new module
is allocated for each task transition, if the new task
is identitcal to the original task, the controller quickly
learns to reuse the old module components. Top: postswitching learning curve for the EMS module on a binary
stimulus-response task, after being trained on the same
task. For clarity, only the Layer Voting method is compared against a baseline module trained from scratch.
Bottom: fraction of the original module reused over the
course of post-switch learning, calculated by averaging
the voting weights of each layer in the original module.
0
10
“No-switch” switches: Our first experiments
tested how the dynamic neural voting mechanism would respond to “no-switch” switches,
i.e. ones in which although a switch cue was
given and a new module allocated, the environment policy’s task did not actually change (Fig
7). We find that in such cases, performance
almost instantly approaches pre-switch levels
(e.g. there is very little penalty in attempting
an uneccessary switch). Moreover, we find that
the weightings the controller applies to the new
module is low: in other words, the system recognizes that no new module is needed and acts
accordingly by concentrating its weights on the
existing module. These results show that, while
we formally assume that the agent is cued as
when task switches occurs, in theory it could
implement a completely autonomous monitoring policy, in which the agent simply runs the
allocation procedure if a performance “anomoly”
occurs (e.g. a sustained drop in reward). If
the system determines that the new module was
unneeded, it could simply reallocate the new
module for a later task switch. In future work,
we plan to implement this policy explicitly.
“Real” switches: We next tested how the dynamic voting controller handled switches in which the
environment policy substantially changed after the switching cue. Using both the EMS module and
(for control) the large fully-ablated module as described in § 3.2, the dynamic neural voting controller
was evaluated on 15 switching experiments using multiple variants of SR and MTS tasks. Specifically,
these 15 switches cover a variety of distinct (but not mutually exclusive) switching types including:
•
•
•
•
•
•
addition of new classes to the dataset (switch indexes 2, 7, 11 in the table of Fig. 8)
replacing the current class set entirely with a new non-overlapping class set (switch ids. 1, 3)
addition of visual variability to a previously less variable task (switch id. 6)
addition of visual interface elements e.g. new buttons (switch id. 8)
transformation of interface elements e.g. screen rotation (switch ids. 12, 13, 14, 15)
transitions between different task paradigms e.g. SR to MTS tasks and vice-versa (switch ids. 4, 5,
9, 10).
Controller hyperparameters were optimized in a cross-validated fashion (see Appendix G.1), and
optimizations for three different initialization seeds were run to obtain mean and standard error.
Figures 8a and b show characteristic post-switch learning curves for the EMS module for both the
Layer Voting and Single-Unit Voting methods. Additional switching curves can be found in the
Supplement. Cumulative reward gains relative to learning from scratch were quantified by Relative
switch
)−AU C(M )
Gain in AUC: RGain = AU C(M AU C(M
, where M is the module trained from scratch on
)
12
a
b
Reward
0.8
0.6
EMS
Single-Unit Voting
Layer Voting
0.4
0.2
10
20
30
40
50
10
20
EMS
c
d
Layer Voting
Single-Unit Voting
Relative AUC Gain
0.5
40
50
Layer Voting
Single-Unit Voting
1.4
1.2
0.4
1.0
0.3
0.8
0.2
0.6
0.1
0.4
0.0
e
30
None (Large)
0.2
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
0.5
Transfer Gain
0.4
0.3
0.2
0.1
0.0
1
1.
2.
3.
4.
5.
6.
7.
8.
2
3
Base Task
2-way SR
2-way SR
2-way stationary MTS
2-way SR
2-way stationary MTS
2-way stationary MTS
2-way vert-motion horiz-flip MTS
4-way 2-shown vert-motion MTS
4
5
6
7
8
Switch Task
2-way SR new classes
4-way double-binary SR
2-way stationary MTS new classes
2-way stationary MTS
2-way SR
2-way vert-motion horiz-flip MTS
4-way 2-shown vert-motion MTS
4-way 4-shown permuted MTS
9
9.
10.
11.
12.
13.
14.
15.
10
11
12
Base Task
4-way double-binary SR
4-way 4-shown stationary MTS t
2-way SR
4-way double-binary SR
2-way SR
2-way SR
2-way SR
13
14
15
Switch Task
4-way 4-shown stationary MTS
4-way quadrant SR
4-way quadrant SR
4-way quadrant SR
class reversal
squeezed map
90◦ map rotation
Task switching with Dynamic Neural Voting. Post-Switching learning curves for the EMS module on the
4-way Quadrant SR task after learning a. 2-way SR task and b. a 4-way MTS task with 4 match screen class
templates. Both the Layer Voting method and Single-Unit Voting method are compared against a baseline
module trained on the second task from scratch. Across all twelve task switches, we evaluate the Relative
Gain in AUC over baseline (RGain) using both voting methods for c. the EMS module and d. the large-sized
fully-ablated late bottleneck MLP. e. Transfer Gain (TGain) metrics are compared for both module types for
each of the voting mechanisms. Colors are as in c. (EMS module) and d. (fully-ablated module).
Figure 8
13
the second task, and M switch is the module transferred from an initial task using the dynamic voting
controller. We find that the dynamic voting controller allows for rapid positive transfer of both module
types across all 15 task switches, and the general Single-Unit voting method is a somewhat better
transfer mechanism than the Layer Voting method (Fig. 8c). Both the EMS module and the large
fully-ablated module, which was shown to be inefficient on single-task performance in § 3.2, benefit
from dynamic neural voting (Fig. 8 d).
EMS modules are more “switchable”: To quantify how fast switching gains are realized, we use
Transfer Gain: T Gain = T∆∆max , where T∆max = argmax(∆t ) is the time where the maximum
max
amount of reward difference between M switch and M occurs, and ∆max is the reward difference
at that time. Qualitatively, a high score on the Transfer Gain metric indicates that a large amount
of relative reward improvement has been achieved in a short amount of time (see Figure S7 for a
graphical illustration of the relationship between the RGain and T Gain metrics). While both the
EMS and large fully-ablated modules have positive Transfer Gain, EMS scores significantly higher
on this metric, i.e. is significantly more “switchable” than the large fully-ablated module (Fig. 8e).
We hypothesize that this is due to the EMS module being able to achieve high task performance
with significantly fewer units than the larger fully-ablated module, making the former easier for the
dynamic neural voting controller to operate on.
5
C ONCLUSION AND F UTURE D IRECTIONS
In this work, we introduce the TouchStream environment, a continual reinforcement learning framework that unifies a wide variety of spatial decision-making tasks within a single context. We describe
a general algorithm (ReMaP) for learning light-weight neural modules that discover implicit task
interfaces within this large-action/state-space environment. We show that a particular module architecture (EMS) is able to remain compact while retaining high task performance, and thus is especially
suitable for flexible task learning and switching. We also describe a simple but general dynamic
task-switching architecture that shows substantial ability to transfer knowledge when modules for
new tasks are learned.
A crucial future direction will be to expand insights from the current work into a more complete
continual-learning agent. We will need to show that our approach scales to handle dozens or hundreds
of task switches in sequence. We will also need to address issues of how the agent determines when
to build a new module and how to consolidate modules when appropriate (e.g. when a series of tasks
previously understood as separate can be solved by a single smaller structure). It will also be critical
to extend our approach to handle visual tasks with longer horizons, such as navigation or game play
with extended strategic planning, which will likely require the use of recurrent memory stores as part
of the feature encoder.
From an application point of view, we are particularly interested in using techniques like those
described here to produce agents that can autonomously discover and operate the interfaces present
in many important real-world two-dimensional problem domains, such as on smartphones or the
internet (Grossman, 2007). We also expect many of the same spatially-informed techniques that
enable our ReMaP/EMS modules to perform well in the 2-D TouchStream environment will also
transfer naturally to a three-dimensional context, where autonomous robotics applications (Devin
et al., 2016) are very compelling.
R EFERENCES
Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Deep compositional question
answering with neural module networks. CoRR, abs/1511.02799, 2015. URL http://arxiv.
org/abs/1511.02799.
Michael A Arbib. Schema theory. The Encyclopedia of Artificial Intelligence, 2:1427–1443, 1992.
Matthew M Botvinick and Jonathan D Cohen. The computational and neural basis of cognitive
control: charted territory and new frontiers. Cognitive science, 38(6):1249–1285, 2014.
Tianqi Chen, Ian Goodfellow, and Jonathon Shlens. Net2net: Accelerating learning via knowledge
transfer. arXiv preprint arXiv:1511.05641, 2015.
14
Djork-Arné Clevert, Thomas Unterthiner, and Sepp Hochreiter. Fast and accurate deep network
learning by exponential linear units (elus). CoRR, abs/1511.07289, 2015. URL http://arxiv.
org/abs/1511.07289.
Michael W Cole, Jeremy R Reynolds, Jonathan D Power, Grega Repovs, Alan Anticevic, and Todd S
Braver. Multi-task connectivity reveals flexible hubs for adaptive task control. Nat. Neurosci., 16
(9):1348–1355, sep 2013.
J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical
Image Database. In IEEE CVPR, 2009.
Coline Devin, Abhishek Gupta, Trevor Darrell, Pieter Abbeel, and Sergey Levine. Learning modular
neural network policies for multi-task and multi-robot transfer. CoRR, abs/1609.07088, 2016. URL
http://arxiv.org/abs/1609.07088.
Yan Duan, John Schulman, Xi Chen, Peter L. Bartlett, Ilya Sutskever, and Pieter Abbeel. Rl$ˆ2$:
Fast reinforcement learning via slow reinforcement learning. CoRR, abs/1611.02779, 2016. URL
http://arxiv.org/abs/1611.02779.
Gabriel Dulac-Arnold, Richard Evans, Peter Sunehag, and Ben Coppin. Reinforcement learning in
large discrete action spaces. CoRR, abs/1512.07679, 2015. URL http://arxiv.org/abs/
1512.07679.
Chrisantha Fernando, Dylan Banarse, Charles Blundell, Yori Zwols, David Ha, Andrei A. Rusu,
Alexander Pritzel, and Daan Wierstra. Pathnet: Evolution channels gradient descent in super neural
networks. CoRR, abs/1701.08734, 2017. URL http://arxiv.org/abs/1701.08734.
David Gaffan and Susan Harrison. Inferotemporal-frontal disconnection and fornix transection in
visuomotor conditional learning by monkeys. Behavioural Brain Research, 31(2):149 – 163,
1988. ISSN 0166-4328. doi: https://doi.org/10.1016/0166-4328(88)90018-6. URL http:
//www.sciencedirect.com/science/article/pii/0166432888900186.
Lev Grossman. Invention of the year: The iphone. Time Magazine Online, 1, 2007.
Alexa E Horner, Christopher J Heath, Martha Hvoslef-Eide, Brianne A Kent, Chi Hun Kim, Simon RO
Nilsson, Johan Alsiö, Charlotte A Oomen, Andrew Holmes, Lisa M Saksida, et al. The touchscreen
operant platform for testing learning and memory in rats and mice. Nature protocols, 8(10):
1961–1984, 2013.
R. Hu, J. Andreas, M. Rohrbach, T. Darrell, and K. Saenko. Learning to Reason: End-to-End Module
Networks for Visual Question Answering. ArXiv e-prints, April 2017.
Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization.
abs/1412.6980, 2014. URL http://arxiv.org/abs/1412.6980.
CoRR,
James Kirkpatrick, Razvan Pascanu, Neil C. Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis
Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic
forgetting in neural networks. CoRR, abs/1612.00796, 2016. URL http://arxiv.org/abs/
1612.00796.
A Krizhevsky, I Sutskever, and G Hinton. ImageNet classification with deep convolutional neural
networks. Advances in Neural Information Processing Systems, 2012.
Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James
Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO:
common objects in context. CoRR, abs/1405.0312, 2014. URL http://arxiv.org/abs/
1405.0312.
James L McClelland. Incorporating rapid neocortical learning of new schema-consistent information
into complementary learning systems theory. Journal of Experimental Psychology: General, 142
(4):1190, 2013.
15
Elisabeth A. Murray and Mortimer Mishkin. Object recognition and location memory in monkeys with
excitotoxic lesions of the amygdala and hippocampus. Journal of Neuroscience, 18(16):6568–6582,
1998. ISSN 0270-6474. URL http://www.jneurosci.org/content/18/16/6568.
Georg Ostrovski, Marc G. Bellemare, Aäron van den Oord, and Rémi Munos. Count-based exploration
with neural density models. CoRR, abs/1703.01310, 2017. URL http://arxiv.org/abs/
1703.01310.
R. Rajalingham, K. Schmidt, and J. J. DiCarlo. Comparison of object recognition behavior in human
and monkey. J. Neurosci., 35(35), 2015.
Ali S Razavian, Hossein Azizpour, Josephine Sullivan, and Stefan Carlsson. Cnn features offthe-shelf: an astounding baseline for recognition. In Computer Vision and Pattern Recognition
Workshops (CVPRW), 2014 IEEE Conference on, pp. 512–519. IEEE, 2014.
Andrei A. Rusu, Neil C. Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick,
Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. CoRR,
abs/1606.04671, 2016. URL http://arxiv.org/abs/1606.04671.
Wenling Shang, Kihyuk Sohn, Diogo Almeida, and Honglak Lee. Understanding and improving
convolutional neural networks via concatenated rectified linear units. CoRR, abs/1603.05201, 2016.
URL http://arxiv.org/abs/1603.05201.
Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image
recognition. arXiv preprint arXiv:1409.1556, 2014.
Richard S. Sutton and Andrew G. Barto. Introduction to Reinforcement Learning. MIT Press,
Cambridge, MA, USA, 1st edition, 1998. ISBN 0262193981.
Anthony D Wagner, Daniel L Schacter, Michael Rotte, Wilma Koutstaal, Anat Maril, Anders M Dale,
Bruce R Rosen, and Randy L Buckner. Building memories: remembering and forgetting of verbal
experiences as predicted by brain activity. Science, 281(5380):1188–1191, 1998.
Jane X. Wang, Zeb Kurth-Nelson, Dhruva Tirumala, Hubert Soyer, Joel Z. Leibo, Rémi Munos,
Charles Blundell, Dharshan Kumaran, and Matt Botvinick. Learning to reinforcement learn. CoRR,
abs/1611.05763, 2016. URL http://arxiv.org/abs/1611.05763.
D Yamins, H Hong, C F Cadieu, E A Solomon, D Seibert, and J J DiCarlo. Performance-optimized
hierarchical models predict neural responses in higher visual cortex. Proceedings of the National
Academy of Sciences, 2014.
Daniel LK Yamins and James J DiCarlo. Using goal-driven deep learning models to understand
sensory cortex. Nature neuroscience, 19(3):356–365, 2016.
Friedemann Zenke, Ben Poole, and Surya Ganguli. Improved multitask learning through synaptic
intelligence. arXiv preprint arXiv:1703.04200, 2017.
S UPPLEMENTARY M ATERIAL
A
TASK VARIANTS
The EMS module and all ablation controls were evaluated on a suite of 13 stimulus-response,
match-to-sample, localization, and MS-COCO MTS variants:
1. 2-way SR - standard binary SR task
2. 4-way double binary SR - four class variant of SR, where each class is assigned either to the right
or left half of the action space
3. 4-way quadrant SR - four class variant of SR, where each class is assigned to only a quadrant of
the action space
4. 2-way stationary MTS - standard binary MTS task with stereotyped and non-moving match screens
16
5. 2-way stationary horiz-flip MTS - two class variant MTS task where the match templates’ horizontal placement is randomly chosen, but confined within the same vertical plane
6. 2-way stationary vert-motion MTS - two class variant MTS task where the match templates’
vertical position is randomly chosen, but each class is confined to a specific side
7. 2-way stationary vert-motion horiz-flip MTS - two class variant MTS task where the match
templates’ positions are completely random
8. 4-way 2-shown MTS - four class variant MTS task where only two class templates are shown on
the match screen (appearing with random horizontal location as well)
9. 4-way 2-shown vert-motion MTS - same as above, but with random vertical motion for the
templates
10. 4-way 4-shown stationary MTS - four class variant MTS task where all four class templates are
shown on the match screen, but with fixed positions.
11. 4-way 4-shown permuted MTS - same as above, but with randomly permuted locations of all
match templates
12. Localization - Localization task
13. MS-COCO MTS - 80-way MTS task using the MS-COCO detection challenge dataset, where
match screens are randomly samples scenes from the dataset
B
E XPERIMENT DETAILS AND DATASETS
Stimulus-Response Experiment Details: Image categories used are drawn from the Image-Net
2012 ILSVR classification challenge dataset Deng et al. (2009). Four unique object classes are taken
from the dataset: Boston Terrier, Monarch Butterfly, Race Car, and Panda Bear. Each class has 1300
unique training instances, and 50 unique validation instances.
Match-To-Sample Experiment Details: Sample screen images drawn from the same Image-Net
class set as the Stimulus-Response tasks. One face-centered, unobstructed class instance is also drawn
from the Image-Net classification challenge set and used as a match screen template image for that
class. Class template images for the match screen were held fixed at 100x100 pixels. For all variants
of the MTS task, we keep a six pixel buffer between the edges of the screen and the match images,
and a twelve pixel buffer between the adjascent edges of the match images themselves. Variants
without vertical motion have the match images vertically centered on the screen.
Localization Experiment Details: The Localization task uses synthetic images containing a single
main salient object placed on a complex background (similar to images used in Yamins & DiCarlo
(2016); Yamins et al. (2014)). There are a total of 59 unique classes in this dataset. In contrast
to other single-class localization datasets (e.g. Image-Net) which are designed to have one large,
face-centered, and centrally-focused object instance and for which a trivial policy of “always poke
in image corners” could be learned, this synthetic image set offers larger variance in instance scale,
position, and rotation so the agent is forced into learning non-trivial policies requiring larger precision
in action selection.
MS-COCO MTS Experiment Details This task uses the entire MS-COCO detection challenge
dataset Lin et al. (2014). On every timestep, a sample screen chosen from one of the 80 MS-COCO
classes. These are constructed to be large, unobstructed, face centered representations of the class. For
the match screen, we sample a random scene from MS-COCO containing any number of objects, but
containing at least a single instance of the sample class. The agent is rewarded if its action is located
inside any instance of the correct class. Both modules use sample actions from a low-temperature
Boltzmann policy from eq. (4), which was empirically found to result in more precise reward map
prediction.
C
C.1
M ODULES
U NITS P ER L AYER
Table S1 aggregates the number of units per layer for the EMS and ablated modules which was used
when conducting single-task and task-switching experiments. Only fully-connected modules’ layer
sizes are shown here. For details on the convolutional bottleneck EMS module, please refer to C.2.
17
Table S1: Number of units per layer for investigated modules
Base-task
EMS
SR
MTS
LOC
8
32
128
No
symm
8
32
128
No
Mult
8
32
128
No
None- None- Nonemult/symm
Small Med
Large
8
8
128
512
32
32
128
512
128
128
512
1024
Normalized Validation AUC
2-way SR
1.0
4-way double binary SR
4-way quadrant SR
0.9
2-way stationary MTS
0.8
2-way vert. motion MTS
0.7
2-way horiz. flip MTS
2-way motion/flip MTS
0.6
4-way 2-shown MTS
0.5
4-way 2-shown vert-motion MTS
0.4
4-way 4-shown stationary MTS
0.3
4-way 4-shown permuted MTS
Pa
rtia
EM
S
l sy
mm
No
sym
No
m
sym
m/p
atia
l mu
lt
No
Non
mult
eC
ReL
u (la
rge)
Non
eR
eLu
(larg
e)
Non
e eL
u (la
Non
rge)
eC
ReL
u (m
ediu
No
m)
mult
/sym
me
No
Lu
mult
/sym
mR
Non
e
Lu
e eL
u (m
ediu
Non
m)
eR
eLu
(me
dium
Non
eC
)
ReL
u (s
mall
Non
)
e ta
nh (
larg
No
e)
mult
/sym
m ta
n
h
Non
e sig
(larg
No
e
)
mult
/sym
ms
Non
ig
eR
eLu
(sm
all)
Non
e eL
u (s
mall
Non
)
e sig
(me
dium
Non
e ta
)
nh (
med
ium
Non
)
e ta
nh (
sma
ll)
Non
e sig
(sm
all)
Localization
Figure S1: Exhaustive module performance study of the EMS module and 23 ablation control modules,
measured as the Area Under the Curve for all SR, MTS, and LOC task variants. Shown is the AUC normalized
to the highest performing module in a task. Results in fig. 4 have further averaged this over the vertical task axis,
and report only a salient subset of the ablations.
C.2
T HE C ONVOLUTIONAL -EMS M ODULE
This is a "Convolutional Bottleneck" extension of the EMS module shown in the paper, where skip
connections link the conv5 and the FC6 representation of the visual backbone. Here, the "scenelevel" representation stored in the FC6 ReMaP memory buffer is tiled spatially to match the present
convolution dimensions (here 14x14), and concatenated onto its channel dimension. A series of 1x1
convolutions plays the role of a shallow visual bottleneck, before the activations are vectorized and
concatenated with A as input to the CReS layers of the standard EMS module.
The results in the paper are shown for a bottleneck consisting of a single tanh and two CReS
convolutions, with 128 units each. The Downstream layers use 128 units each as well.
The motivation for the convolutional bottleneck is that lower-level features are useful for complex
spatial tasks such as Localization and Object Detection, and hence may result in a more precise policy.
By tiling the entire scene-level representation along the convolution layer’s channel dimension, a
form of multiplicative template-matching is possible between objects that must be memorized (e.g.
MS-COCO MTS templates) and what is inside the present scene.
D
E XHAUSTIVE A BLATION S TUDY
In all, we investigated 23 distinct ablations on the EMS module, across all twelve task variants outlined
in sec A (Fig. S1). Symmetry ablations replace CReS with the activation x 7→ ReLU(x) ⊕ x2
Multiplicative ablations are denoted by specifying the nonlinearity used in place of CReS (where this
is one of ReLU, tanh, sigmoid, elu Clevert et al. (2015), or CReLU Shang et al. (2016)). This
additionally includes one partial symmetry ablation (denoted “partial symm”) where only the visual
bottleneck is symmetric, and one which ablates the ReLU from the “no symm” module (denoted “no
symm/partial-mult”).
18
Table S2: Module learning rates
EMS
Partial symm
No symm
No symm/partial mult
No mult/symm ReLU
No mult/symm tanh
No mult/symm sig
No mult/symm eLU
No mult/symm CReLU
None ReLU(small)
None ReLU(medium)
None ReLU(large)
None tanh(small)
None tanh(medium)
None tanh(large)
None sig(small)
None sig(medium)
None sig(large)
None eLU(small)
None eLU(medium)
None eLU(large)
None CReLU(small)
None CReLU(medium)
None CReLU(large)
D.1
2-way
SR
4-way
double
binary
SR
4-way
stationary
SR
2-way
stationary
MTS
2-way
vertmotion
MTS
2-way
horiz flip
MTS
2-way
motion/flip
MTS
4-way
2-shown
MTS
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−4
10−4
10−4
10−4
10−4
10−4
10−3
10−3
10−4
10−3
10−3
10−4
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−4
10−4
10−4
10−4
10−4
10−4
10−3
10−3
10−4
10−3
10−3
10−4
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−4
10−3
10−3
10−3
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−3
10−3
10−3
10−3
10−3
10−4
5·10−4
5·10−4
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−4
10−4
10−3
10−3
10−4
10−3
10−3
10−4
10−3
10−4
10−4
10−3
10−4
10−4
5·10−4
5·10−4
10−3
10−3
10−3
10−4
10−4
10−3
10−3
10−3
10−4
10−4
10−3
10−3
10−4
10−3
10−3
10−4
10−3
10−4
10−4
10−3
10−4
10−4
5·10−4
5·10−4
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−4
10−4
10−3
10−3
10−4
10−3
10−3
10−4
10−3
10−4
10−4
10−4
10−4
10−4
5·10−4
5·10−4
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−4
10−4
10−3
10−3
10−4
10−3
10−3
10−4
10−3
10−4
10−4
10−3
10−4
10−4
5·10−4
5·10−4
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−4
10−4
10−3
10−3
10−4
10−3
10−3
10−4
10−3
10−4
10−4
10−3
10−4
10−4
4-way
2-shown
vertmotion
MTS
5·10−4
5·10−4
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−3
10−4
10−4
10−3
10−3
10−4
10−3
10−3
10−4
10−3
10−4
10−4
10−3
10−4
10−4
4-way
4-shown
stationary
MTS
5·10−4
5·10−4
10−3
10−3
10−3
10−4
10−4
10−3
10−3
10−3
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−3
10−3
10−4
10−3
10−3
10−4
4-way
4-shown
permuted
MTS
LOC
5·10−4
5·10−4
2·10−4
2·10−4
10−3
10−4
10−4
10−3
10−3
10−3
10−4
10−4
10−4
10−4
10−4
10−3
10−4
10−4
10−3
10−4
10−4
10−3
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
10−4
H YPERPARAMETERS
Learning rates for the ADAM optimizer were chosen on a per-task basis through cross-validation on
a grid between [10−4 ,10−3 ] for each architecture. Values used in the present study may be seen in
Table S2.
E
A DDITIONAL MS-COCO R EWARD M APS
Five additional reward map examples for the MS-COCO MTS task are provided in in Figure S2.
Examples are plotted over the course of learning.
F
F.1
A DDITIONAL L EARNING C URVES
S INGLE -TASK A BLATION E XPERIMENTS
Learning trajectories for seven additional tasks are provided in Figure S3. Modules capable of
convergence on a task were run until this was acheived, but AUC values for a given task are calculated
at the point in time when the majority of models converge.
F.2
DYNAMIC VOTING C ONTROLLER AND EMS MODULE TASK -S WITCHING E XPERIMENTS
Additional trajectories for ten unshown switching curves are provided in Figure S4.
G
G.1
DYNAMIC VOTING C ONTROLLER AUGMENTATIONS
L EARNABLE PARAMETER I NITIALIZATIONS
Here we describe the weight initialization scheme that was found to be optimal for use with the
dynamic voting controller. For simplicity, consider the layer-voting mechanism, with learnable
19
Sample
Match
Reward Maps
64
128
704
1344
1984
64
128
704
1344
1984
64
704
1984
2624
3904
64
704
1344
2176
3904
64
704
1344
6016
Training episode (in thousands)
8384
Figure S2: Examples of the emergence of decision interfaces in MSCOCO MTS Reward map predictions
over the course of training for 5 different object classes.
20
Reward
2-way SR
4-way double binary SR
2-way stationary MTS
2-way horiz-flip MTS
2-way vert-motion MTS
4-way 2-shown MTS
4-way 4-shown stationary MTS
EMS
No symm
No mult
None (large)
No mult/symm
None (medium)
None (small)
Training Episodes
Figure S3: Additional Single-task performance ablation Learning curves. Seven learning curves shown for
task variants not seen in the main text body. Shown are the same ablations as the main text.
21
2-way SR to 4-way double binary SR
2-way stationary MTS to 2-way stationary MTS new classes
2-way SR to 2-way stationary MTS
2-way stationary MTS to 2-way SR
2-way stationary MTS to 2-way vert-motion horiz-flip MTS
2-way vert-motion horiz-flip MTS to 4-way 2-shown vert-motion MTS
4-way 2-shown vert-motion MTS to 4-way 4-shown permuted MTS
4-way double binary SR to 4-way 4-shown stationary MTS
4-way double binary SR to 4-way quadrant SR
Reward
2-way SR to 2-way SR new classes
EMS
Single-Unit Voting
Layer Voting
22
Training Episodes
weight matricies W i and biases bi . The intended biasing scheme is achieved through initializing the
elements of these parameters to:
0
if i = 1, ω < Ω
|N (µ , 0.001)|
i
1
Wω ∼ |N (µ , 0.001)|
if i > 1, ω < Ω
|N (0.01, 0.001)| if ω = Ω
0
b
biω = b1
0.1
(10)
if i = 1, ω < Ω
if i > 1, ω < Ω
if ω = Ω
(11)
This initialization technique was also generalized for use with the single-unit voting mechanism.
For the switching experiments presented in section § 4.2, we sweep the hyperparameters on a narrow
band around the default scheme. The ranges for these are: µ0 ∈ [0.01, 0.005] , b0 ∈ [0.1, 0.01] , µ1 ∈
[0.01, 0.02] , and b1 ∈ [0.1, 0.2, 0.5, 1.0].
G.2
TARGETED T RANSFORMATIONS
Two additional switching mechanisms were added to the controller to augment its ability to switch
between taks which are remappings of the action space or reward policy of a preexisting module.
G.2.1
ACTION T RANSFORMATIONS
we note that efficient modules are those which can effectively produce a minimal representation of
the interaction between action space A and observation xt . If the agent’s optimal action space shifts
to A0 while the remainder of the task context remains fixed, the controller should allow for rapid
targeted remapping A 7→ A0 . Since we formulate the modules as ReMaP Networks, and A is an
input feature basis, we can achieve remappings of this form through a fully-connected transformation:
a0τ = f (Wa aτ + b)
(12)
where aτ = [ht−kb:t−1 , at ] is the vector of action histories, and Wa and b embed aτ into new action
space A0 using only a small number of learnable parameters.
Pseudo Identity-Preserving Transformation In practice, we initialize the parameters in eq. (12)
such that the transformation is pseudo identity-preseving, meaning that the representation learned at
this level in the original module is not destroyed prior to transfer.
This is done by initializing Wa to be an identity matrix I|aτ | with a small amount of Gaussian noise
∼ N (0.0, σ 2 ) added to break symmetry. b is initialized to be a vector of ones of size |aτ |.
G.2.2
R EWARD M AP T RANSFORMATIONS
Each of the kf maps mt (x) reflects the agent’s uncertainty in the environment’s reward policy. If the
task context remains stationary, but the environment transitions to new reward schedule R0 that no
longer aligns with the module’s policy π, the controller could to this transition by e.g. containing a
mechanism allowing for targeted transformation of m(x) and hence also π.
One complication that arises under ReMaP is that since each task-module learns its optimal action
space internally, m(x) are in the basis of R rather than A. Therefore, transformations on the map
distribution must also re-encode A before mapping to R0 .
In this work, we investigate a shallow “adapter” neural network that lives on top of the existing
module and maps R 7→ R0 . Its first and second layers are defined by
l1 (x) = f (W1 [m(x)
g(aτ ), aτ ] + b1
m(x)0 ∝ W2 l1 + b2
23
(13)
(14)
where g(aτ ) is a similar transformation on A as above, denotes elementwise multiplication, W1 is a
0
learnable matrix embedding into a hidden state, and W2 ∈ R|l1 |×|R | is a learnable matrix embedding
into R0
Pseudo Identity-Preserving Transformation Similar to the transformation on the action space,
we modify the reward-map transformation to be pseudo identity-preserving as well. This is done
by modifying eq. (13) such that the original maps are concatenated on to the beginning of the
transformation input vector:
l1 (x) = f (W1 [m(x), m(x)
g(aτ ), aτ ] + b1
The intended map-preserving transformation is accomplished via initializing W1 and W2 as:
1.0 + N (0.0, ) if i = j, i < R
W (i,j) ∼
N (0.0, )
otherwise
G.3
(15)
(16)
TARGETED T RANSFORMATION H YPERPARAMETERS
Both of the targeted transformations have several hyperparameters. We conducted a grid search to
optimize these in a cross-validated fashion, on a set of test task switches designed to be solved by
one of the targeted transformations (Fig. S5). Each was conducted independently of the dynamic
voting controller, and independently of the other transformation. Optimal hyperparameters found in
these experiments were fixed for use in the integrated dynamic voting controller, and were not further
optimized afterwards.
Action Transformation Hyperparameters We conducted three tests using the stimulus-response
paradigm: class reversal (in which the left class becomes the right class and vice-versa), a horizontal
rotation of the reward boundaries (such that right becomes up and left becomes down), and a “switch”
to the original task (intended to test the identity-preserving component).
In this work, we find that a single, non-activated linear transformation (f in (12)) is optimal for
this new state-space embedding, using kb ∗ 2 units, and initialized such that the idendity-preserving
transformation weights have σ = 0.01. The learning rate for this transformation was found to be
optimal at 0.1.
Reward Map Transformation Hyperparameters We conducted two tests using the stimulusresponse paradigm: a “squeezing” task (where there is no longer any reward dispensed on the
lower half of the screen), and a “switch” to the original task (intended to test the identity-preserving
component).
In this work, we find the optimal activations in eq. (15) to be f (·) = CReS and g(·) = ReLU, with
4 units in the hidden layer. in the weight initialization scheme was found optimal at 0.001, and an
initial bias of 0.01. The optimal learning rate for this transformation was found to be 0.01.
G.4
T RANSFORM A BLATION
A study was conducted to determine the relative benefit of the targeted transformations (Fig. S6),
where it was determined that the primary contribution of the dynamic neural controller was in fact
the voting mechanism (although the transformations did supplement this as well).
G.5
D EPLOYMENT SCHEME OF TASK MODULES
When cued into task transition, the controller freezes the learnable parameters of the old task-module,
and deploys a new unitialized task-module. The controller then initializes the action and reward map
transformation networks as described in G.2 on top of the old module. These transformations are also
voted on inside the dynamic neural controller at every timestep.
H
S WITCHING M ETRICS
Figure S7 graphically illustrates the metrics used inside the paper to quantify switching performance:
RGain and T Gain.
24
a
Environment Reward
Policy Reversal
Stimulus-Response
image:
Task
Switch
reward
map:
Training
Trials
b
Environment Reward
Policy Rotation
Stimulus-Response
image:
Task
Switch
reward
map:
Training
Trials
c
Environment Reward
Policy Squeeze
Stimulus-Response
image:
Task
Switch
reward
map:
Training
Trials
Figure S5: Action and reward map transformation switch examples. Three task switching experiments were
performed to optimize the hyperparameters of the targeted transformations that augment the dynamic neural
voting controller. These switches are also retested in the fully-integrated meta-controller and shown in the
original switching result figure. a. Binary stimulus-response class reversals, where the left class becomes the
right class, and vice-versa. b. Rotations of the binary stimulus-response reward boundaries. c. A “squeezing” of
the binary stimulus-response reward boundaries, where no reward is given on the new task on the bottom half of
the screen, regardless of class shown.
25
Relative AUC Gain
0.5
0.4
0.3
0.2
0.1
0.0
−0.1
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Figure S6: Targeted controller transform ablation. Relative AUC Gain for the EMS module over the same
switching snearios in the paper, but with the targeted transformations ablated.
1.0
0.9
Reward
0.8
TGain =
Δmax
Δmax
tΔ
max
0.7
RGain =
0.6
0.5
tΔ
AUC( t)
AUC(Bt )
max
5
10
15
20
25
30
40
Training episodes
Figure S7: Illustration of switching performance metrics. We quantify the switching performance of the
dynamic neural controller and task-modules by two metrics: “relative gain in AUC” (ratio of green to purple
shaded regions), and “transfer” gain (difference of reward at T∆max ). Relative AUC measures the overall gain
relative to scratch, and the transfer gain measures the speed of transfer. Curve shown is the EMS module with
Single-Unit voting method evaluated on a switch from a 4-way MTS task with two randomly moving class
templates to a 4-way MTS task with four randomly moving templates.
26
| 2cs.AI
|
Smooth backfitting of proportional hazards – A new approach
projecting survival data
Munir Hiabu
Cass Business School, City University London, U.K.
arXiv:1707.04622v1 [math.ST] 14 Jul 2017
E-mail: [email protected]
Enno Mammen
Heidelberg University, Germany
E-mail: [email protected]
Marı́a Dolores Martı́nez-Miranda
Cass Business School, City University London, U.K.
E-mail: [email protected]
Jens Perch Nielsen
Cass Business School, City University London, U.K.
E-mail: [email protected]
Summary. Smooth backfitting has proven to have a number of theoretical and practical advantages in structured regression. Smooth backfitting projects the data down onto the structured space of interest providing a direct link between data and estimator. This paper introduces the ideas of smooth backfitting to survival analysis in a proportional hazard model,
where we assume an underlying conditional hazard with multiplicative components. We develop asymptotic theory for the estimator and we use the smooth backfitter in a practical
application, where we extend recent advances of in-sample forecasting methodology by allowing more information to be incorporated, while still obeying the structured requirements of
in-sample forecasting.
Keywords: Aalen’s multiplicative model; Local linear kernel estimation; Survival
data; In-sample forecasting.
1.
Introduction
Purely unconstrained nonparametric models suffer from the curse of dimensionality in high
dimensional data spaces. Most often, some structure has to be introduced to stabilize the
system and to allow to visualize, interpret, extrapolate and forecast the properties of the
underlying data. The smooth backfitting algorithm of Mammen et al. (1999) considered the
simplest nonparametric structure in the regression context - the additive structure - and
it has been a successful update of kernel smoothing regression backfitting algorithms with
many theoretical and practical advantages to earlier approaches of regression backfitting.
The still to this day popular regression backfitting algorithms of Hastie and Tibshirani
(1990) are numerical iterating-procedures estimating one component given the estimates of
2
Hiabu et al.
the rest. In contrast the smooth backfitter is a direct projection of the data down onto
the structured space of interest. This direct relationship between data and estimates gives
a more solid grip on what is being estimated and the theoretical properties underlying it,
see also Nielsen and Sperlich (2005). The purpose of this paper is to introduce smooth
backfitting to the field of survival analysis and nonparametric smooth hazard estimation.
While the additive structure is the most natural and most widely used in regression, the
multiplicative structure seems more natural in hazard estimation. The omnipresent Cox
regression model is a proportional hazard model and many extensions and alternatives to the
Cox hazard model have been formulated in a multiplicative framework. We have therefore
chosen the multiplicative hazard structure as the natural place to start when introducing
smooth backfitting to survival analysis. Smooth multiplicative backfitting is theoretically
more challenging than additive smooth backfitting. The smooth backfitting multiplicative
regression structure was analysed in detail by Yu et al. (2008) as a special case of general
additive models. Yu et al. (2008) proved that the multiplicative structure - in contrast
to the simpler additive regression models - provides asymptotic theory with a number of
non-trivial interactions between exposure available in different directions. Naturally, the
asymptotics provided here for smooth backfitting of multiplicative hazards contain similar
interactive components in the asymptotic theory. We are able to provide a simple algorithm
first projecting the data down onto an unconstrained estimator, and then further projecting
the unconstrained estimator down onto the multiplicative space of interest. Our numerical
algorithms are greatly simplified by a new principle of weighting the projection according
to the final estimates.
Let Z = (Z1 , . . . , Zd ) be a d-dimensional covariate process observed as long as our
observed object is under exposure. We are interested in the conditional hazard of a nonnegative random variable T
α(t|Z) = lim h−1 Pr [T ∈ [t, t + h)| T ≥ t, {Z(s), s ≤ t}] .
(1)
α(t|Z) = α(t, Z(t)),
(2)
h↓0
We assume that
where α is some unknown smooth function depending on the time t and the value of the
covariate at only the time point t. In many cases, T might be subject to some filtering.
Filtered observations are present in a vast variety of topics including right censoring in
experimental studies like clinical trials or left truncation in insurance loss data. A first
version of the non-parametric model (2) was introduced in Beran (1981) where the author
only considered time independent covariates and a filtering scheme of only right censoring.
Dabrowska (1987) showed weak convergence of the estimator presented there. The more
general model with time dependent covariates and also more general filtering patterns were
analysed in McKeague and Utikal (1990) and Nielsen and Linton (1995) as part of a counting
process model. The estimator of Nielsen and Linton (1995) was identified as the natural
local constant estimator in this context and also generalized to the local linear estimator in
Nielsen (1998). We observe n independent and identically distributed copies of the process
(N, Y, Z), where Y is a predictable process and N a counting process with intensity
λ(t) = α(t, Z(t))Y (t),
(3)
Smooth backfitting of proportional hazards
3
The multiplicative intensity assumption (3) of the counting process is known as Aalen’s
multiplicative intensity model. Andersen et al. (1993) give a comprehensive overview of
how to embed various survival data models into this counting process formulation. In Section 2.1, we show how to embed left truncation and right censoring.
Non-parametric approaches like (2) and (3) are attractive because of the minimal assumptions on the underlying model compared to for example a fully parametric approach. However, estimation accuracy decreases rapidly with the number of dimensions - also known
as curse of dimensionality - a weakness that can be overcome by introducing assumptions
of separable structures on the underlying hazard, see also Stone (1985). In this paper we
assume that the conditional hazard is multiplicative i.e.
α(t, z) = α0 (t)α1 (z1 ) · · · αd (zd ).
(4)
Algorithms for kernel smoothing of (4) have been provided in Hastie and Tibshirani (1990)
and Fan et al. (1997) and most recently in Lin et al. (2016). Kernel smoothing within
the model framework (4) was first analysed in the filtered survival context in Linton et al.
(2003) who based their approach on the principle of marginal integration Linton and Nielsen
(1995). Marginal integration, however, requires a rectangular support of the data. The data
example of this paper is taken from the non-life insurance challenge of estimating outstanding liabilities. This data has triangular support and the approach of Linton et al. (2003)
is therefore not feasible. We show in Section 4 that our smooth projection approach - that
directly links the data to the underlying multiplicative structure - also works on triangle
support as our data example requires. In Section 2 the underlying survival model is set
up. In Section 3 it is first pointed out that unconstrained multidimensional hazard estimators can be considered to be ratios of smooth occurrence and smooth exposure. This is
unlike regression, where only the local constant Nadaraya-Watson estimator exhibits this
simple structure. Secondly, the smooth backfitting estimator is defined as a projection of
any unconstrained hazard estimator that enjoys the simple ratio structure. In section 4
the asymptotic properties are given for the smooth backfitting estimator defined in Section
3 and general sufficient conditions are given for the asymptotic properties of the unconstrained smooth occurrence and unconstrained smooth exposure that our smooth backfitter
is based on. In Section 5 we consider a sophisticated version of in-sample forecasting enabled
by our new smooth backfitter. We introduce a smooth extension of the popular actuarial
chain ladder model. In-sample forecasting is possible because of the imposed multiplicative
structure. In the concluding Section 6 we point out that while multiplicative hazard estimation is a natural place to start, other hazard structures might be interesting to consider
in the future. Martinussen and Scheike (2006) for example consider a rich class of additive
and combined additive and multiplicative structures that could be interesting to explore in
future work. Most of the proofs are deferred to the Appendix.
2.
Aalen’s multiplicative intensity model
We consider Aalen’s multiplicative intensity model that allows for very general observations
schemes. It covers filtered observations arising from left truncation and right censoring but
also more complicated patterns of occurrence and exposure. In the next section we describe
4
Hiabu et al.
how to embed left truncation and right censoring into this framework. In contrast to
Linton et al. (2003) we will hereby allow the filtering to be correlated to the survival time
and be represented in the covariate process. We briefly summarize the general model we
are assuming.
We observe n iid copies of the stochastic processes (N (t), Y (t), Z(t)), t ∈ [0, R0 ], R0 > 0.
Here, N denotes a right-continuous counting process which is zero at time zero and has
jumps of size one. The process Y is left-continuous and takes values in {0, 1} where the value
1 indicates that the individual is under risk. Finally,
Z is a d-dimensional left-continuous
Q
covariate process with values in a rectangle dj=1 [0, Rj ] ⊂ Rd .The multivariate process
((N1 , Y1 , Z1 ), . . . , (Nn , Yn , Zn )), i = 1, . . . , n, is adapted to the filtration Ft which satisfies
the usual conditions. Now we assume that Ni satisfies Aalen’s multiplicative intensity
model, that is
λi (t) = lim h−1 E[Ni ((t + h)−) − Ni (t−)| Ft− ] = α(t, Zi (t))Yi (t).
h↓0
(5)
The deterministic function α(t, z) is called hazard function and it is the failure rate of an
individual at time t given the covariate Z(t) = z.
2.1. Left truncation and right censoring time as covariates
The most prominent example for Aalen’s multiplicative intensity model is filtered observation due to left truncation and right censoring. We now show how to embed model
(2) with covariate, Z, possibly carrying truncation and censoring information into Aalen’s
multiplicative intensity model. Every covariate coordinate can carry individual truncation
information as long as it corresponds to left truncation. That is, we observe (T, Z) if and
only if (T, Z(T )) ∈ I, where the set I is compact and it holds that if (t1 , Z(t1 )) ∈ I
and t2 ≥ t1 , then (t2 , Z(t2 )) ∈ I, a.s. The set I is allowed to be random but is independent of T given the covariate process Z. Furthermore, T can be subject to right censoring with censoring time C. We assume that also T and C are conditional independent
given the covariate process Z. This includes the case where the censoring time equals
one covariate coordinate. In conclusion, we observe n iid copies of (Te, Z ∗ , I, δ), where
δ = 1(T ∗ < C), Te = min(T ∗ , C), and (T ∗ , Z ∗ ) is the truncated version of (T, Z), i.e,
(T ∗ , Z ∗ ) arises from (T, Z) by conditioning on the event (T, Z(T )) ∈ I.
Then, for each subject, i = 1, . . . , n, we can define a counting process Ni as
n
o
Ni (t) = 1 Tei ≤ t, δi = 1 ,
∗
e
with respect to the filtration Fi,t = σ
Ti ≤ s, Zi (s), Ii , δi : s ≤ t ∪ N , where N is
a class of null-sets that completes the filtration. After straightforward computations one
can conclude that under the setting above, including (2), Aalen’s multiplicative intensity
model is satisfied with
αz (t) = α(t, z) = lim h−1 Pr{Ti ∈ [t, t + h)| Ti ≥ t, Zi (t) = z},
h↓0
∗
Yi (t) = 1 (t, Zi (t)) ∈ Ii , t ≤ Tei .
Smooth backfitting of proportional hazards
3.
5
The smooth backfitting estimator of multiplicative hazards
We describe the smooth backfitting problem in two steps. First the data is projected down
onto an unconstrained space resulting in an unconstrained estimator. Secondly this unconstrained estimator is projected down onto the multiplicative space of interest. We first
show that both a local constant and a local linear projection in the first step lead to estimators that are simple ratios of smoothed occurrence and smoothed exposure. This resembles
the simple structure known from the local constant Nadaraya-Watson estimator in regression, however, local linear regression does not satisfy this simple structure. This is the
background for our approach of being able to derive general underlying conditions for our
smooth backfitter to work the encompass both local constant and local linear estimators.
In fact it encompasses all estimators with the simple ratio structure including all local polynomial kernel hazard estimators and other situations, where the unconstrained estimator
can be expressed as such simple ratio.
3.1.
First step: projecting the data down onto the unstructured space resulting in an
unconstrained estimator
In this section we concentrate on the local constant and local linear estimators defined
inNielsen (1998) as a projection of the data down onto the unconstrained space. We
notice that both these two estimators can be expressed as a ratio of a smoothed occurrence and a smoothed exposure. This will be important in the next section where the
unconstrained estimator is projected further down to the multiplicative space of interest.
We introduce the notation Xi (t) = (t, Zi (t)). We also set x = (t, z), with coordinates
x0 = t, x1 = z1 , . . . , xd = zd , and write the hazard as α(t, z) = α(x).
To estimate the components of the structured hazard in (7) below, we will need an unstructured pilot estimator of the hazard α first. We propose the local linear kernel estimator,
α̂LL (x), based on least squares (cf. Nielsen (1998)). Its value in x is defined as the solution
θb0 in the equation
θb0
θb1
!
n Z
X
" Z
2
1 s+ε
T
dNi (u) − θ0 − θ1 (x − Xi (s))
= arg
min
lim
ε s
θ0 ∈R,θ1 ∈Rd+1 ε→0
i=1
Z s+ε
2 #
1
−
Kb (x − Xi (s))Yi (s) ds.
dNi (u)
ε s
(6)
Q
In the following, we restrict ourselves to a multiplicative kernel K(u0 , . . . , ud ) = dj=0 k(uj ),
Q
and a one-dimensional bandwidth b, with Kb (u) = dj=0 b−1 k(b−1 uj ), where for simplicity
of notation the bandwidth b does not depend on j. More general choices would have
been possible with the cost of extra notation. The local linear estimator includes boundary
corrections so that the bias is of same order at the boundary as in the interior of the support,
namely O(b2 ), or for the more general case of varying bandwidths we do not consider here,
O(max1≤j≤d+1 b2j ). The local constant estimator achieves only slower rates at the boundary
region and local polynomial estimators of higher order, like in regression, have the usual
6
Hiabu et al.
drawback known from higher order kernels, that they perform poorly as long as sample
sizes are not very large.
The solution of the least square minimisation (6) can be rewritten as the ratio of smooth
estimators of the number of occurrence and the exposure, see Gámiz et al. (2013) for details.
b LL (x) = n−1
O
n Z
X
1 − (x − Xi (s))D(x)−1 c1 (x) Kb (x − Xi (s))dNi (s),
i=1
b LL (x) = n
E
n Z
X
−1
1 − (x − Xi (s))D(x)−1 c1 (x) Kb (x − Xi (s))Yi (s)ds,
i=1
where the components of the (d + 1)−dimensional vector c1 are
c1j (x) = n
−1
n Z
X
Kb (x − Xi (s))(xj − Xij (s))Yi (s)ds,
j = 0, . . . , d,
i=1
and the entries (djk ) of the (d + 1) × (d + 1)−dimensional matrix D(x) are given by
djk (x) = n
−1
n Z
X
Kb (x − Xi (s))(xj − Xij (s))(xk − Xik (s))Yi (s)ds.
i=1
b LL (x)/E
b LL (x). Compare this
The local linear estimator is then defined as α
bLL (x) = O
estimator with the local constant version that can be defined as a similar ratio, α
bLC (x) =
b LC (x)/E
b LC (x), but with smoothed occurrence and smoothed exposure given by
O
b LC (x) = κn (x)
O
b LC (x) = κn (x)
E
n Z
X
i=1
n Z
X
Kb (x − Xi (s))dNi (s),
Kb (x − Xi (s))Yi (s)ds,
i=1
Z
κn (x) =
−1
Kb (x − u) du
.
Under standard smoothing conditions, if b is chosen of order n−1/(4+d+1) , then the bias of
and α
bLC (x) is of order n−2/(4+d+1) and the variance is of order n−4/(4+d+1) , which is
the optimal rate of convergence in the corresponding regression problem, see Stone (1982).
For an asymptotic theory of these estimators see Linton et al. (2003).
α
bLL (x)
3.2. Structured smooth backfitting estimator via solution-weighted minimization
In this section we will project the unconstrained estimator of the previous section down
onto the multiplicative space of interest. Due to filtering, observations
are assumed to
Q
be only available on a subset of the full support, X ⊆ R = dj=0 [0, Rj ]. Our estimators
are restricted to this set and assumptions on X and the data generating functions are
Smooth backfitting of proportional hazards
7
given in the next section. Our calculations simplify via a new principle we call solutionweighted minimization. We assume that we have the solution and use it strategically in
the least squares weighting. While the procedure is not directly feasible to compute, it
is made feasible by defining it as an iterative procedure. In the sequel we will assume a
multiplicative structure of the hazard α, i.e.,
α(x) = α∗
d
Y
αj (xj ),
(7)
j=0
where αj , j = 0, . . . , d, are some functions and α∗ is a constant. For identifiability of the
components, we make the following further assumption:
Z
αj (xj )wj (xj ) dxj = 1, j = 0, . . . , d,
where wj is some weight function.
We also need the following notation:
Ft (z) = P r (Z1 (t) ≤ z| Y1 (t) = 1) ,
y(t) = E[Y1 (t)].
By denoting ft (z) the density corresponding to Ft (z) with respect to the Lebesgue measure,
we also define
E(x) = ft (z)y(t)
and O(x) = E(x)α(x).
We define the estimators α
b∗ and α
b = (α
b0 , . . . , α
bd ) of the hazard components in (7) as
solution of the following system of equations:
R
b
Xxk O(x)dx−k
k = 0, . . . , d,
(8)
α
bk (xk ) = R
Q
b
b∗ j6=k α
bj (xj )E(x)dx
−k
Xxk α
Z
α
bk (xk )wk (xk ) dxk = 1, k = 0, . . . , d.
(9)
Here Xxk denotes the set {(x0 , . . . , xk−1 , xk+1 , . . . , xd )| (x0 , . . . , xd ) ∈ X }, and
b and O
b are some full-dimensional esx−k = (x0 , . . . , xk−1 , xk+1 , . . . , xd ). Furthermore, E
timators of E and O. We will discuss below that the system above has a solution with
probability tending to one. In the next section we will show asymptotic properties of the
b and
estimator. We will see that we do not require that the full-dimensional estimators E
b
O are consistent. We will only need asymptotic consistency of marginal averages of the
estimators, see below. This already highlights that our estimator efficiently circumvents the
curse of dimensionality.
In practice, system (8) can be solved by the following iterative procedure:
R
b
Xxk O(x)dx−k
(r+1)
, k = 0, . . . , d
(10)
α
bk
(xk ) = R
Qk−1 (r+1)
Q
(r)
b
α
b
(xj ) d+1 α
b (xj )E(x)dx
−k
Xxk
j=0
j
j=k+1
j
8
Hiabu et al.
After a finite number of cycles or after a termination criterion applies, the last values of
(r+1)
α
bk
(xk ), k = 0, ..., d, are multiplied by a factor such that the constraint (9) is fulfilled with
the above choice of wk (xk ). This can always be achieved by multiplication with constants.
This gives the backfitting approximations of α
bk (xk ) for k = 0, ..., d.
The estimator α
b can be motivated as a weighted least squares estimator with random
data-adaptive weights. To see this consider the estimator α = (α∗ , α0 , . . . , αd ) that minimizes
2
Z
d
Y
min
αj (xj ) w(x)dx,
α
e(x) − α∗
(11)
α
X
j=0
b
b
where w(x) is some weighting and α
e(x) = O(x)/
E(x)
is an unconstrained full-dimensional
estimator of α. This gives
R
Q
e(x) dj=0 αj (xj )w(x)dx
X α
∗
α = R nQ
,
o2
d
w(x)dx
j=0 αj (xj )
X
and (α0 , . . . , αd ) can be described via the backfitting equation
R
Q
e(x) j6=k αj (xj )w(x)dx−k
Xxk α
, k = 0, . . . , d.
αk (xk ) = R
nQ
o2
∗
w(x)dx−k
j6=k αj (xj )
Xx α
(12)
k
The asymptotic variance of kernel estimators of α is proportional to α(x)/E(x), see e.g.
Linton and Nielsen (1995). This motivates the choice w(x) = E(x)/α(x). However, this
choice is not possible because E(x) and α(x) are unknown. One could use w(x) = Ě(x)/α̌(x)
where Ě(x) and α̌(x) are some pilot estimators of E and α. We follow another idea and we
propose to weight the minimization (11) with its solution. We choose
b
E(x)
w(x) = Q
,
bi (x)
iα
(13)
and heuristically, by putting αj = α
bj and by plugging (13) into (12), we get (8). The next
section discusses existence and asymptotic properties of the solution α
bj of (8).
4.
Asymptotic properties of the smooth backfitter of multiplicative hazards
The estimator α
bj is defined as solution of a nonlinear operator equation. We are going to
approximate this equation by a linear equation that can be interpreted as equation that
arises in nonparametric additive regression models, and then show that the solution of the
linear equation approximates α
bj . The linear equation and its solution is well understood
from the theory of additive models. This will be our essential step to arrive at an asymptotic
understanding of our estimator α
bj . Assumptions [A1]-[A7] below and [B1]-[B7] are of standard nature in marker dependent hazard papers and can be verified for the local constant
Smooth backfitting of proportional hazards
9
and local linear estimators we are interested, see in particular Nielsen and Linton (1995),
Nielsen (1998) and Linton et al. (2003) for related calculations. However, one should notice
that the conditions are not restricted to the local constant or local linear smoothers. They
are not even tight to kernel smoothers. Any smoother could be used as long as it obeys the
structure of being a ratio of a smoothed occurrence and a smoothed exposure.
For our main theorem we make the following assumptions. We hereby do not make
assumptions on the full support R but only on a subset X ⊆ R. We will make use
of the following notations: Xxj has been defined after equation (8) and Xxj ,xk denotes
the set {(xl : l ∈ {0, ..., d}\{j, k})| (x0 , . . . , xd ) ∈ X }. Furthermore, we define Xk =
{xk | (x0 , . . . , xd ) ∈ X for some values of (xl : l 6= k)}, Xj,k = {(xj , xk )| (x0 , . . . , xd ) ∈
X for some values of (xl : l 6∈ {j, k})}, and Xj(xk ) = {xj | (x0 , . . . , xd ) ∈ X for some values
of (xl : l 6∈ {j, k})}.
A1 The function E(x) is two times continuously differentiable for x ∈ X and inf x∈X E(x) >
0.
A2 The hazard α is two times continuously differentiable for x ∈ X and inf x∈X α(x) > 0.
A3 The kernel K has compact support which is without loss of generality supposed to be
[−1, 1]. Furthermore it is symmetric and continuous.
A4 It holds that nb5 → cb for a constant 0 < cb < ∞ as n → ∞.
A5 It holds that
Z
1
dxj dxk < ∞
Xj,k Oj (xj )Ok (xk )
R
Q
for j, k = 0, ..., d, j 6= k, where Oj (xj ) = O(x) dx−j and O(x) = α∗ dj=0 αj (xj )E(x).
R
A6 It holds that the two-dimensional marginal occurences Oj,k (xj , xk ) = O(x) dx−(j,k)
are bounded from above and bounded away from 0.
A7 For some δ > 0 it holds that for j, k = 0, ..., d, j 6= k
Z
1
dxj dxk < ∞,
1+δ
(xj )Ok (xk )
Xj,k Oj
Z
1
sup
dxj < ∞,
1−δ
(xj )Ok (xk )
xk ∈Xk Xj(xk ) Oj
Z
1
sup
dxj < ∞.
1/2
1/2
xk ∈Xk Xj(xk ) O
j (xj )Ok (xk )
Note that assumptions [A1]-[A4] are standard in kernel smoothing theory. In Assumptions
[A5] and [A6] we only assume that the two-dimensional marginal occurrences of O are
bounded from above and bounded away from 0, but we do not make the assumption that the
one-dimensional marginal occurrences have this property. This allows that the support of a
two-dimensional marginal density Ojk has a triangular shape {(xj , xk ) : xj +xk ≤ c; xj , xk ≥
0} for some constant c > 0. This can be easily seen. Suppose for simplicity that Ojk is the
10
Hiabu et al.
uniform density on the triangle. Then Oj (xj ) = 2c−2 (c − xj )+ and Ok (xk ) = 2c−2 (c − xk )+
and we have
Z
Z
1
1
π2
2
dx
dx
=
< ∞.
dxj dxk =
j
k
2
Oj (xj )Ok (xk )
3c2
xj +xk ≤c; xj ,xk ≥0 c (c − xj )(c − xk )
Thus, our assumption [A5] on one-dimensional marginals is fulfilled. One can easily verify
that also [A7] holds for this example. This discussion can be extended to other shapes of
two-dimensional marginals that differ from rectangular supports. Note also that [A5] and
[A7] trivially hold if the one-dimensional marginal Oj are bounded away from zero.
The solutions α
b0 , . . . , α
bd of (8) can be rewritten as solutions of
Z
Z
b
O(x)dx
−k −
Xxk
Xxk
α
b
∗
d
Y
b
α
bj (xj )E(x)dx
−k = 0,
k = 0, . . . , d.
j=0
R
R
Q
Since, Xx O(x)dx−k − Xx α∗ dj=0 αj (xj )E(x)dx−k = 0, the difference of those two zerok
k
terms is zero as well, and we have
Z
d
d
Y
Y
b
b k (xk ) −
αj (xj ) E(x)dx
α
bj (xj ) − α∗
α
b∗
0=∆
−k
Xxk
j=0
j=0
Z
d
d
Y
Y
∗
∗
b
b
b
b
αj (xj )E(x)dx
(14)
= ∆k (xk ) −
(1 + δ )
{1 + δj (xj )} − 1 α
−k ,
Xxk
j=0
j=0
where
Z
Z
n
o
b
O(x) − O(x) dx−k −
b k (xk ) =
∆
Xxk
Xxk
α
∗
d
Y
b
αj (xj ){E(x)
− E(x)}dx−k ,
j=0
α
bj (xj ) − αj (xj )
δbj (xj ) =
,
αj (xj )
α
b ∗ − α∗
δb∗ =
.
α∗
Note that δb is defined as a root of a non-linear operator. Motivated by (14), we define an
∗
approximation, δ and δ j (xj ) (0 ≤ j ≤ d), as solution of the linear equation
Z
d
d
X
Y
∗
b
b
δ +
δ j (xj ) α∗
αj (xj )E(x)dx
(15)
−k = ∆k (xk )
Xxk
under the constraint
R
j=0
j=0
δ k (xk )ωk (xk )dxk = 0 where
ωk (xk ) =
Z Y
d
j=0
b
αj (xj )E(x)dx
−k .
Smooth backfitting of proportional hazards
11
R Qd
b
This is equal to the constraint (9) for the choice wk (xk ) =
j=0 αj (xj )E(x)dx−k /(αk (xk )
R Qd
b
j=0 αj (xj )E(x)dx). Under this constraint one has that
o
R n
R
Q
b
b
− E(x)}dx
O(x)
−
O(x)
dx − X α∗ dj=0 αj (xj ){E(x)
X
∗
.
δ =
R
Q
b
α∗ d αj (xj )E(x)dx
j=0
X
Note that the norming of the constraint cannot be used in practice because α is unknown
but it will simplify the theoretical discussion and the results can be carried over to feasible
weighting.
Equation (15) can be rewritten as an integral equation of the second kind
XZ
∗
δ k (xk ) +
π
bk,j (xk , xj )δ j (xj )dxj = µ
bk (xk ) − δ ,
j6=k
Xj(xk )
with
e
O(x)
= α∗
d
Y
αj (xj )Ê(x),
j=0
Z
e
e
Oj,k (xj , xk ) = O(x)
dx−(j,k) ,
Z
ek (xk ) = O(x)
e
O
dx−k ,
π
bk,j (xk , xj ) =
µ
bk (xk ) =
ej,k (xj , xk )
O
,
ek (xk )
O
b k (xk )
∆
.
ek (xk )
O
Note that all these functions depend on n. The integral equation can also be simply written
∗
bδ = µ
b − δ with δ = (δ 0 , ..., δ d )| , where π
b is the integral operator with kernel π
bk,j .
as δ + π
b Before we come to this point we state a proposition
We will show that δ approximates δ.
that gives the asymptotics for δ .
b and O.
b We decompose
For the next results we need some conditions on the estimators E
A
B
C
µ
bk into three terms, µ
bk = µ
bk + µ
bk + µ
bk , which depend on n. For some deterministic
functions O∗ (x) and E ∗ (x) these terms are defined as:
n
o
n
o
R
R
Q
b
b
− Xx α∗ dj=0 αj (xj ) E(x)
− E ∗ (x) dx−k + Xx O(x)
− O∗ (x) dx−k
k
k
,
µ
bA
k (xk ) =
ek (xk )
O
R
R
Q
− Xx α∗ dj=0 αj (xj ) {E ∗ (x) − E(x)} dx−k + Xx {O∗ (x) − O(x)} dx−k
k
k
µ
bB
,
k (xk ) =
Ok (xk )
#
"
O
(x
)
k
k
µ
bC
−1 µ
bB
k (xk ) =
k (xk ).
ek (xk )
O
12
Hiabu et al.
∗
∗
Note that µ
bB
k are deterministic functions. Typical choices of O (x) and E (x) are the
A
b
b
expectations of E(x)
and O(x).
Then µ
bk (xk ) is the stochastic part of a one-dimensional
B
smoother and µ
bk (xk ) is its bias. Both terms are well understood and can be easily treated
by standard smoothing theory. We now want to develop an asymptotic theory for the
estimators α̂j where their asymptotic properties are described by the properties
of µ
bA
k (xk )
R r
∗,r
B
bk (xk )Ok (xk ) dxk
and µ
bk (xk ). We use the following normings of these quantities: µk = µ
R r
ek (xk ) dxk for r ∈ {A, B, C}.
for r ∈ {A, B, C} and µ∗,r,n
= µ
bk (xk )O
k
We assume that with
R Qd
l=0 αl (xl )E(x)dx−(k,j)
πk,j (xk , xj ) = R Q
d
l=0 αl (xl )E(x)dx−k
the following assumptions hold:
R
e 2 dx = OP (1) and
B1 It holds that O(x)
ej,k (xj , xk ) − Oj,k (xj , xk ) = oP ((log n)−1 )
O
R
uniformly over 0 ≤ j < k ≤ d and xj , xk , where Oj,k (xj , xk ) = O(x) dx−(j,k) .
B2
1/2
−1/5
sup |Oj (xj )µ
bA
)
j (xj )| = oP (n
xj
and
1/2
−1/5
sup |Oj (xj )µ
bB
)
j (xj )| = oP (n
xj
for 0 ≤ j ≤ d, where Oj (xj ) =
R
O(x) dx−j .
B3 For xj with Oj (xj ) > 0 it holds that
2
n2/5 µ
bA
j (xj ) → N(0, σj (xj ))
for 0 ≤ j ≤ d with some function σj2 (xj ) > 0.
B4
Z
and
Z
2
−4/5
µ
bA
)
j (xj ) Oj (xj ) dxj = OP (n
2
−4/5
µ
bB
)
j (xj ) Oj (xj ) dxj = O(n
for 0 ≤ j ≤ d.
B5 It holds that
sup
xj ∈Xj
1/2
Oj (xj )
Z
Xk(xj )
Oj,k (xj , xk ) A
µ
bk (xk )dxk = oP (n−2/5 (log n)−1 ).
Oj (xj )
Smooth backfitting of proportional hazards
B6 It holds for 0 ≤ j ≤ d that
Z
sup
xj ∈Xj
x∈X
1
Q
k∈{0,...,d}\{j} Ok (xk )
Z
sup
xj ∈Xj
x∈X
1/2
e
O(x)dx
−j = OP (1),
1/2
O(x)dx−j = O(1).
1
Q
k∈{0,...,d}\{j} Ok (xk )
13
B7 It holds that µ∗,B,n − µ∗,B = op (n−2/5 ).
We shortly discuss these assumptions. Condition [B1] is a mild consistency assumption
for a two-dimensional smoother. Also [B2] is a weak condition because in our setting
one-dimensional smoothers are typically OP (n−2/5 )-consistent. [B3] is a standard limit
result for many one-dimensional smoothers and [B4] assumes rates for L2 -norms of the
stochastic part and the bias part of a nonparametric one-dimensional smoother that are
standard under our smoothness assumptions. For the interpretation of [B5] note that the
integral on the left hand side in the formula is a global average of µ
bA
bA
k . Because µ
k is a
local average this is a global weighted average of mean zero random varibales. Thus one
expects a OP (n−1/2 ) rate for the integral. For the supremum of the integrals one expects
a OP ((log n)1/2 n−1/2 ) rate which is faster than the required rate. [B6] states a bound for
the total number of occurrences. It can be easily verified under the assumption that the
one dimensional marginals Ol are bounded from below. Furthermore, one can easily check
it if the one-dimensional marginals Ol are bounded from below for l 6∈ {j, k} and if the twodimensional marginal Oj,k has the properties discussed in the example after assumption
[A7]. The following proposition states a stochastic expansion for δ.
Proposition 1. Make the assumptions [A1]–[A7], [B1]–[B7]. Then the function δ =
(δ 0 , ..., δ d ), introduced in (15), exists and is uniquely defined, with probability tending to
one. Moreover, it has the following expansion:
δ−µ
bA − (I − π)−1 (µ
bB − µ∗,B )
O,∞
= op (n−2/5 ),
where, for a function f (x) = (f0 (x0 ), ..., fd (xd ))| , we define
1/2
kf kO,∞ = sup max |Oj (xj )fj (xj )|.
x∈X 0≤j≤d
R
P
Furthermore, the function π : L → L is defined as πk (f )(xk ) = j6=k πkj (xk , xj )fj (xj )dxj
R
R
for f ∈ L = L1 ×...×Ld with Lj = {δj : Xj → R : Xj δj2 (xj )Oj (xj ) dxj < ∞, Xj δj (xj )Oj (xj )
dxj = 0}.
From the proposition we get as a corollary the asymptotic distribution of δ j (xj ).
Proposition 2. Make the assumptions [A1]–[A7], [B1]–[B7]. Then for xj (0 ≤ j ≤ d)
with Oj (xj ) > 0 it holds that
n2/5 {δ j (xj ) − [(I − π)−1 (µ
bB − µ∗,B )]j (xj )} → N(0, σj2 (xj )),
−2/5 ) we have that the bias
in distribution. Under the additional assumption µ
bB
j (xj ) = O(n
[(I − π)−1 (µ
bB − δ
B,∗
)]j (xj ) is of order O(n−2/5 ).
14
Hiabu et al.
The following theorem states that δ is indeed a good approximation of the relative
b
estimation error δ.
Theorem 3. Under assumptions [A1]–[A7], [B1]–[B7] it holds that with probability tending to one there exists a solution δb∗ and δb = (δb0 , . . . , δbd ) that solves equation (14) with
δb − δ
O,∞
= op (n−2/5 ),
∗
δb∗ − δ = op (n−2/5 ).
For this solution we get that
n2/5 {(α
bj − αj )(xj ) − αj (xj )[(I − π)−1 (µ
bB − µ∗,B )]j (xj )} → N(0, αj2 (xj )σj2 (xj )),
in distribution, for xj (0 ≤ j ≤ d) with Oj (xj ) > 0.
5.
In-sample forecasting of outstanding loss liabilities
The so-called chain ladder method is a popular approach to estimate outstanding liabilities. It started off as a deterministic algorithm, and it is used today for almost every single
insurance policy over the world in the business of non-life insurance. In many developed
countries, the non-life insurance industry has revenues amounting to around 5%. It is therefore comparable to - but smaller than - the banking industry. In every single product sold,
the chain ladder method (because actuaries hardly use other methods) comes in, estimating
the outstanding liabilities that eventually aggregate to the reserve - the single biggest number of most non-life insurers balance sheets. The insurers liabilities often amount to many
times the underlying value of the company. In Europe alone those outstanding liabilities
are estimated to accumulate to around e 1trn. It is therefore of obvious importance that
this estimate is not too far from the best possible estimate. We describe in this section
how the methodology introduced in this paper can be applied to provide a solution to this
challenging problem.
We analyze reported claims from a motor business line in Cyprus. The same data set has
been used by Hiabu et al. (2016) and it consists of the number of claims reported between
2004 and 2013. During these 10 years (3654 days), n = 58180 claims were reported. The
data are given as {(T1 , Z1 ), . . . , (Tn , Zn )}, where Zi denotes the underwriting date of claim
i, and Ti the time between underwriting date and the date of report of a claim in days,
also called reporting delay. Hence, in the notation of the previous sections, the covariate
underwriting date, Z(t) = Z, does not depend on time and has dimension d = 1. The data
exist on a triangle, with Ti + Zi ≤ 31 December 2013 = R0 , which is a subset of the full
support R = [0, R0 ]2 (0 = 1 January 2004). The aim is to forecast the number of future
claims from contracts written in the past which have not been reported yet. Figure 1 shows
the observed data that lie on a triangle, while the forecasts are required on the triangle
that added to the first completes a square. Here it is implicitly assumed that the maximum
reporting delay of a claim is 10 years. Actuaries call this assumption that the triangle is
fully run off. In our data set, this is a reasonable assumption looking at Figure 1.
Smooth backfitting of proportional hazards
15
70
60
50
60
40
40
20
0
120
30
100
120
100
80
80
60
t
20
60
40
40
20
z
20
10
0
Fig. 1. Histogram of claim numbers of a motor business line between 2004 and 2013. Axis z
represents the underwriting time (in months) and axis t the reporting delay (in months).
16
Hiabu et al.
The classical chain ladder method is able to provide a simple solution to the above
problem. Recently, Martı́nez-Miranda et al. (2013) have pointed out that this method can
be viewed as a multiplicative density method: the original, un-truncated random variable
(T, Z) having density f (t, z) = f1 (t)f2 (z); and the authors suggested to embed the method
in a more standard mathematical statistical vocabulary to engage mathematical statisticians
in future developments. In particular, Martı́nez-Miranda et al. (2013) showed that one could
consider the traditional chain ladder estimator as a multiplicative histogram in a continuous framework, and presented an alternative by projecting an unconstrained local linear
density down onto a multiplicative subspace. This approach was called continuous chain
ladder and it has been further analyzed by Mammen et al. (2015); Lee et al. (2015, 2017),
providing full asymptotic theory of the underlying density components. A related approach
by Hiabu et al. (2016) (see also Hiabu (2017)) proposes to transform the two-dimensional
multiplicative continuous chain ladder problem to two one-dimensional continuous hazard
estimation problems via an elegant time-reverting trick. The application considered in
this paper generalizes the most important of these reversed hazards to a two-dimensional
multiplicatively structured hazard. In this way the continuous chain ladder is improved
and generalized allowing more flexibility for the estimation of outstanding liabilities in the
insurance business.
In Hiabu et al. (2016) it is assumed that T and Z are independent, which means that
the underwriting date of a claim has no effect on the reporting delay. We are not going to
impose such a strong restriction. In order to discuss the independence assumption, consider
Figure 2. The points in the plots are derived by first transforming the data into a triangle
with dimension 3654 × 3654,
n
X
Nt,z =
I Ti = t, Zi = z , (t, z) ∈ {1, 2, . . . , 3654}2 , t + z ≤ 3654
i=1
Q
), with dimension 40 × 40, see
and then aggregating the data into a quarterly triangle, (Nt,z
also Hiabu (2017). Then, for t = 2, . . . , 5, one derives the quarterly hazard rate as ratio
Q Pz
Q
/ l=1 Nt,l
. These values are then scaled by an
of occurrence and exposure, α(t, z) = Nt,z
eye-picked norming factor, α0 (t), t = 2, . . . , 5, letting α(t, z) start at around 1 as a function
of z with fixed t. The final values, α1 (t, z) = α0 (t)α(t, z), are displayed in Figure 2. We
only show plots for t ≤ 5 since almost all claims are reported after five quarters.
If the independence assumption of Hiabu et al. (2016) is satisfied, the points should
lie around a horizontal line in each plot. If the multiplicative hazard assumption of this
paper is satisfied, then any smooth shape is allowed, but all four graphs must be equal
after correction for noise. This is because under the model which will be defined below, the
graphs, α1 (·, ·), with the first component fixed, mimic a quarterly version of α1 .
Inspecting the four plots, one can argue to see a negative drift of similar magnitude in
each graph, the values decaying from around 1 to 0.8. This indicates that the approach of
this paper should give a better fit to the data compared to the model of Hiabu et al. (2016).
From this discussion we now continue with embedding our observations in the proportional hazard framework. Afterwards we will show how the hazard estimate can be used
to forecast the number of outstanding claims. First note that we cannot apply the approach of this paper directly, since in this application we only observe T if T ≤ R0 − Z,
Smooth backfitting of proportional hazards
α1(2,z)
1.2
1.2
1.1
1.1
1.0
α1(3,z)
0.9
0.8
1.0
0.9
0.8
0.7
0.7
0
5
10 15 20 25 30
39
0
z; 1=Q1 2004, 39=Q3 2013
α1(4,z)
17
5
10 15 20 25 30
38
z; 1=Q1 2004, 38=Q2 2013
1.2
1.2
1.1
1.1
1.0
α1(5,z)
0.9
0.8
1.0
0.9
0.8
0.7
0.7
0
5
10
15
20
25
30
z; 1=Q1 2004, 37=Q1 2013
37
0
5
10
15
20
25
30
36
z; 1=Q1 2004, 36=Q4 2012
Fig. 2. Scaled quarterly hazard rates of the first four development quarters.
which is a right truncation. Analogue to Hiabu et al. (2016), we transform the random variable T to T R = R0 − T . This has the result that the right truncation truncation becomes left truncation, T R ≥ Z. Thus we consider the random variable T R
as our variable of interest. With the notation considered in Section 2.1, we now have
T = T R , d = 1, Z(t) = Z, δ = 1, I = {(tR , z) ∈ R|0 ≤ z ≤ tR }. We conclude that the
counting process Ni (tR ) = 1 TiR ≤ tR , satisfies Aalen’s multiplicative intensity model
with respect to the filtration given in Section 2.1 and
αz (tR ) = α(tR , z) = lim h−1 Pr{T R ∈ [tR , tR + h)| T R ≥ tR , Z = z},
h↓0
R
R
Yi (t ) = 1 (t , Zi ) ∈ I, tR ≤ TiR,∗ .
We can therefore estimate the unstructured hazard using the local linear estimator estimator described in Section 3.1. Note that we have X = I. The components of the
multiplicative conditional hazard are then computed as in (10). These estimators require
the choice of the bandwidth parameter, which was assumed to be scalar in order to simplify the notation in this paper. In this application we generalize this restriction allowing
for different smoothing levels in each dimension, namely reporting delay and underwriting
time. The bandwidth parameter is then a vector b = (b0 , b1 ) and we estimate it using
cross-validation (see further details in Appendix A). To alleviate the computational burden
of cross-validation we aggregated the data triangle Nt,z considering bins of two days when
applying a discrete version of the estimators described in Appendix C. After several trials
we run the cross-validation minimization over b0 ∈ {1300, 1400, 1500, 1600, 1700, 1800} and
Hiabu et al.
1500
1000
500
0
2000
1500
1000
500
0
differences between hazard estimates
−0.05
2000
−0.10
underwriting
0.00
0.05
0.10
18
development
Fig. 3. Difference between structured and unstructured hazard estimator, α
e(x) − α
b(x), on a 2-day
grid.
b1 ∈ {2, 3, 4, 5}. The cross-validated bandwidth components were b0 = 1600 and b1 = 3
(unit=2 days).
The results of the estimation procedure are given in Figure 4 and Figure 3. The first
figure shows the estimated components of the multiplicatively structured hazard estimator,
and the latter shows the difference, α
e(x)−α
b(x), between the structured and the unstructured
estimators.
Finally the total number of outstanding claims, i.e. the reserve, can be estimated as
R R0
Z R0 −t
n
b
X
R0 −Zi fZi (t)dt
b
Reserve =
, fz (t) = α
b0 (R0 − t)α
b1 (z) exp −
α
b0 (s)α
b1 (z)ds .
R R0 −Zi
fbZi (t)dt
0
i=1 0
Note that fbz (t) is an estimator of the conditional density of the survival time T . The reserve
can be also decomposed further to provide the ’cash flow’ of the next periods. If the future
is divided into M periods, each of them with length δ = R0 /M , then the amount of claims
19
50
α1(t)
150
0.95
0
0.85
α2(z)
Smooth backfitting of proportional hazards
0
2
4
6
8
Underwriting year z: 0=01/Jan/2004, 10=31/Dec/2013
10
0
2
4
6
8
10
t; T−t=Reporting delay in years
Fig. 4. The estimated multiplicative hazard components .
Table 1. Number of outstanding claims for future quarters; 1 = 2014 Q1, . . . , 39 = 2022 Q3.
The backfitting approach in this paper (PH) is compared with the chain ladder method (CLM)
and the approach in Hiabu et al. (2016).
Future quarter
1
2
3
4
5 6 7 8 9 10 11 12 – 39 Tot.
Hiabu et al. 2016 970 684 422 166 14 5 3 2 1 1
1
0
2270
CLM 948 651 387 148 12 5 3 2 1 1
1
0
2160
PH 872 621 400 130 53 7 4 3 2 1
1
1
2193
forthcoming in the ath (a = 1, 2, . . . , M ) period can be estimated by
R (R0 −Zi +aδ)∧R0
n
b
X
(R0 −Zi +aδ−1)∧R0 fZi (t)dt
ReserveP (a) =
.
R R0 −Zi
bZ (t)dt
f
i
i=1
0
Table 1 shows the estimated number of of outstanding claims for future quarters. We
compare the approach of this paper with the results derived by Hiabu et al. (2016) and
the traditional chain ladder method. The two latter approaches have in common that
they assume independence between underwriting date, Z, and reporting delay, T . We see
that while all approaches estimate a similar total number of outstanding claims (reserve),
those two approaches have distributions over the quarters that are very different from the
results obtained by the method proposed in this paper. It seems that the violation of
the independence assumption has not a big influence on the reserve, since it balances the
different development patterns arising from different periods out. However, the problem
becomes quite serious if one is interested in more detailed estimates like the cash flow.
6.
Conclusion
This paper provided a first introduction of smooth backfitting into survival analysis and
hazard estimation. The starting point has been the popular proportional hazard model
with fully nonparametric components. One could imagine that smooth backfitting could
play a role in a long list of structured problems in semiparametric and nonparametric
survival analysis. One could for example imagine that smooth backfitting can provide
useful extensions of some of the practical dynamic survival models of Martinussen and
Scheike (2006) and one can also think of applications to many of known extensions of the
Cox regression model where the understanding of the link between data and estimators is
improved via the direct projection approach of smooth backfitting.
20
A.
Hiabu et al.
Bandwidth selection
A crucial problem in practice is finding the right amount of smoothing when using nonparametric approaches. In the application described in this paper we have considered the
maybe most straightforward way to estimate the optimal bandwidth – the data-driven
cross-validation method.
The data-driven cross-validation method in density estimation goes back to Rudemo
(1982) and Bowman (1984). Nowadays, a slightly modified version (see Hall (1983)) is used
which aims to minimize the integrated squared error. In our framework, the cross-validation
bandwidth has been proposed in Nielsen and Linton (1995). Cross-validation arises from
the idea to minimize the integrated squared error
n
−1
n Z
X
i=1
R0
0
[α
b{Xi (s)} − α{Xi (s)}]2 Yi (s)ds.
By expanding the square,
only two of the three terms depend on P
theRbandwidth and are thus
R
considered. While α
b(Xi (s))2 ds is feasible, we have to estimate i α
b(Xi (s))α(Xi (s))Yi (s)ds.
In cross-validation this is done by the unbiased leave-one-out estimator
Z
α
b[i] {Xi (s)}dNi (s),
where α
b[i] is the leave-one-out version, which arises from the definition of structured estimator α
b by setting Ni = 0. Finally we define the cross-validated bandwidth, bCV , as
bCV = arg min
b
n Z
X
i=1
α
b(Xi (s))2 ds − 2
n Z
X
α
b[i] {Xi (s)}dNi (s).
(16)
i=1
Theoretical properties of cross-validation in hazard estimation in the one dimensional case
are derived in Mammen et al. (2015). To our knowledge there is no theoretical analysis
of cross-validation in the multivariate hazard case of this paper. An extensive simulation
study of the multivariate case can be found in Gámiz et al. (2013).
B.
Proofs
B.1. Proof of Proposition 1
The proof of this proposition follows the lines of the proof of Theorem 1 in Mammen et al.
(1999) but it needs some major modifications in the last steps of the proof because we have
weaker assumptions than the ones assumed in the latter theorem. We outline that the first
part of the proof in Mammen et al. (1999) also goes through under our weaker assumptions
and we show how additional arguments can be used in theRlast part.
Note that under our assumptions [A5], [A6] we get that Ojk (xj , xk )2 Oj (xj )−1 Ok (xk )−1
dxj dxk < ∞. As in Lemma 1 in Mammen et al. (1999) this implies that for some constants
c, C > 0
c max kδj k ≤ kδ0 + ... + δd k ≤ C max kδj k
(17)
0≤j≤d
0≤j≤d
Smooth backfitting of proportional hazards
21
for δj ∈ Lj where k...k denotes the norm km(x)k2 = m(x)2 O(x) dx. Furthermore, one
gets that kT k = sup{kT (δ0 + ... + δd )k : δj ∈ Lj with kδ0 + ... + δd k < 1} < 1, where here
T is the operator T = Ψd · ... · Ψ0 with
R
Ψj (δ ∗ + δ0 + ... + δd )(x) = δ ∗ + δ0 (x0 ) + ... + δj−1 (xj−1 )
+δj∗ (xj ) + δj+1 (xj+1 ) + ... + δd (xd ),
XZ
∗
δj (xj ) = −
δk (xk )πj,k (xj , xk ) dxk
k6=j
for (δ ∗ , δ0 , ..., δd ) ∈ L. Furthermore, note that for j 6= k it holds that
ej (xj ) − Oj (xj )
O
= oP (1),
Oj (xj )
Z
Z
ej,k (xj , xk )
Oj,k (xj , xk )
O
−
Oj (xj )Ok (xk ) Oj (xj )Ok (xk )
ej,k (xj , xk )
Oj,k (xj , xk )
O
−
ej (xj )Ok (xk ) Oj (xj )Ok (xk )
O
!2
Oj (xj )Ok (xk )dxj dxk = oP (1),
!2
Oj (xj )Ok (xk )dxj dxk = oP (1).
These equations follow from [A5], [A6] and [B1]. To see this note that [A6] and [B1] imply
ej,k (xj , xk )−Oj,k (xj , xk ) = oP ((log n)−1 )Oj,k (xj , xk ).
that, uniformly for xj , xk it holds that O
This gives that
ej (xj )/Oj (xj )] − 1 = oP ((log n)−1 ),
[O
(18)
uniformly for xj ∈ Xj and 0 ≤ j ≤ d. Together with [A5] and [B1], this implies the three
equations. As in Lemma 2 in Mammen et al. (1999) we conclude from these equations that
kT̂ kn < γ
b j , Lej and
for some γ < 1 with probability tending to one. Here, we define T̂ , k...kn , Xn,j , Ψ
ej , π
Le as T , k...k, Xj , Ψj , Lj and L but with Oj , πjk replaced by O
bjk (0 ≤ j, k ≤ d; j 6= k). In
R
R
2
e
e
ej (xj ) dxj =
particular, we put Lj = {δj : Xj → R : Xj δj (xj )Oj (xj ) dxj < ∞, Xj δj (xj )O
0}, Le = Le1 × ... × Led , and kT kn = sup{kT (δ0 + ... + δd )kn : δj ∈ Lej with kδ0 + ... + δd kn < 1}.
Arguing as in the first part of Lemma 3 in Mammen et al. (1999) this gives that δ k (x) =
B
C
r
+ δ k (x) + δ k (x), where for r ∈ {A, B, C} the functions δ k ∈ Lek are defined by
A
δ k (x)
r
r
δ 0 (x0 ) + ... + δ d (xd ) =
s
X
br,[s] (x)
Tbl τbr (x) + R
l=0
with
br,[s] k
kR
≤
Cγ s
with probability tending to one for some constant C > 0. Here we put
b d (µ
b d · ... · Ψ
b 1 (µ
) + ... + Ψ
brd−1 − µ∗,r,n
brd − µ∗,r,n
),
τbr = Ψ
br0 − µ∗,r,n
0
d−1 ) + (µ
d
∞
X
br,[s] (x) =
R
Tbl τbr (x).
l=s+1
22
Hiabu et al.
Up to this point we followed closely the arguments in the proof of Theorem 1 in Mammen
et al. (1999). The arguments of the further parts of the proof of the latter theorem would
need that, in our notation,
Z
e 2 (xj , xk )
O
j,k
sup
dxk
(19)
2
e
xj ∈Xj Xk(xj ) Oj (xj )Ok (xk )
is bounded by a constant, with probability tending to one. This would imply that with
probability tending to one for some constant C > 0 for all functions g : Xk(xj ) → R
Z
ej,k (xj , xk )
O
sup
g(xk )dxk ≤ Ckgk,
(20)
ej (xj )
xj ∈Xj
O
Xk(xj )
as can be seen by application of the Cauchy-Schwarz inequality. The proof of Theorem 1 in
r,[s]
Mammen et al. (1999) shows that this can be used to show that supx∈X ,0≤j≤d |Rj (x)| ≤
Cγ s with probability tending to one for some constant C > 0. Unfortunately in our setting
(19) does not hold and thus we cannot follow that (20) holds in our setting. Indeed, one can
check that in general (20) does not hold under our assumptions. Consider e.g. the set-up
discussed after the statement of assumption [A7]. Thus we do not have that T and T̂ map
a function with bounded L2 -norm into a function with bounded L∞ -norm. This also does
not hold if we replace the L∞ -norm by our weighted norm k..kO,∞ . We now argue that
after twice application of T or T̂ a function with bounded k..k-norm is transformed into
a function with bounded k..kO,∞ -norm. This follows from the following two estimates for
functions g : Xk → R with some constant C > 0
!2
Z
Z
Z
Oj,k (xj , xk )
1−δ
Ok (xk )g 2 (xk )dxk , (21)
g(xk )dxk Oj (xj )dxj ≤ C
Oj (xj )
Xk
Xj
Xk(xj )
Z
1/2
Z
Oj,k (xj , xk )
1/2
1−δ
2
sup Oj (xj )
g(xk )dxk ≤ C
Ok (xk )g (xk )dxk
.(22)
Oj (xj )
xj ∈Xj
Xk(xj )
Xk
Furthermore, it holds with probability tending to one, that
!2
Z
Z
Z
ej,k (xj , xk )
O
Ok (xk )g 2 (xk )dxk , (23)
g(xk )dxk Oj1−δ (xj )dxj ≤ C
e
Oj (xj )
Xk
Xj
Xk(xj )
1/2
Z
Z
ej,k (xj , xk )
O
1/2
1−δ
2
sup Oj (xj )
g(xk )dxk ≤ C
Ok (xk )g (xk )dxk
.(24)
ej (xj )
xj ∈Xj
O
Xk(xj )
Xk
Below we will also use that a function with bounded k..kO,∞ -norm is mapped by T and T̂
into a function with bounded k..kO,∞ -norm. This follows from
Z
Oj,k (xj , xk )
1/2
1/2
sup Oj (xj )
g(xk )dxk ≤ C ∗ sup Ok (xk )|g(xk )|,
(25)
Oj (xj )
xj ∈Xj
xk ∈Xk
Xk(xj )
Z
ej,k (xj , xk )
O
1/2
1/2
sup Oj (xj )
g(xk )dxk ≤ C ∗ sup Ok (xk )|g(xk )|
(26)
ej (xj )
xj ∈Xj
xk ∈Xk
O
Xk(xj )
Smooth backfitting of proportional hazards
23
with probability to one. We now show (21)–(26). The bound (25) follows directly from the
last inequality in Condition A7. For the proof of (21) note that the left hand side of (21)
can be bounded by a constant times
Z
Xxj ,xk
1
dxj dxk
1+δ
Oj (xj )Ok (xk )
Z
Ok (xk )g 2 (xk )dxk .
Xk
Thus, (21) follows by application of the first inequality in Condition A7. For the proof of
(22) note that the left hand side of (22) can be bounded by a constant times
Z
sup
xk ∈Xk
Xj(xk )
1
dxj
1−δ
Oj (xj )Ok (xk )
Z
Xk
!1/2
Ok1−δ (xk )g 2 (xk )dxk
.
Here, (22) follows by application of the second inequality in Condition A7. For the proof
of (23), (24) and (26) one uses (18) to show that the left hand sides of the equations in
ej and O
ek , respectively.
Condition A7 are of order OP (1) if one replaces Oj and Ok by O
Thus, one can show (23), (24) and (26) by using the same arguments as in the proofs of
(21), (22) and (25).
We now want to show that
r
r
kδ 0 (x0 ) + ... + δ d (xd ) −
∞
X
T l τ r (x)kO,∞ = oP (n−2/5 ),
(27)
l=0
where
brd−1 − µ∗,r
brd − µ∗,r
τ r = Ψd · ... · Ψ1 (µ
br0 − µ∗,r
0 ) + ... + Ψd (µ
d−1 ) + (µ
d )
and where for δ = (δ ∗ , δ0 , ..., δd )| ∈ R×L we define kδ ∗ +δ0 +...+δd kO,∞ as k(δ0 , ..., δd )| kO,∞ ∨
|δ ∗ |.
Using (21)–(26), kT k < 1 and the fact that kT̂ kn < γ for some γ < 1 with probability
tending to one, one gets that for (27) it suffices to show that for all choices of c > 0
k
cX
log n
Tbl τbr (x) − T l τ r (x)kO,∞ = oP (n−2/5 ).
(28)
l=0
For the proof of this claim it suffices to show that the norm of each summand is of order
oP (n−2/5 (log n)−1 ). This can be shown by using condition B1, (21)–(26), and
24
Hiabu et al.
#
!2
ej,k (xj , xk ) O
ej,k (xj , xk )
O
−
g(xk )dxk Oj1−δ (xj )dxj
ej (xj )
ej (xj )
O
O
Xk(xj )
Xj
Z
−1
= oP ((log n) )
Ok (xk )g 2 (xk )dxk ,
Xk
#
"
Z
ej,k (xj , xk )
ej,k (xj , xk ) O
O
1/2
g(xk )dxk
sup Oj (xj )
−
ej (xj )
ej (xj )
xj ∈Xj
O
O
Xk(xj )
Z
1/2
Ok1−δ (xk )g 2 (xk )dxk
= oP ((log n)−1 )
,
Xk
#
"
Z
e
e
O
(x
,
x
)
O
(x
,
x
)
j
j
j,k
k
j,k
k
1/2
−
g(xk )dxk
sup Oj (xj )
ej (xj )
ej (xj )
xj ∈Xj
O
O
Xk(x )
Z
Z
"
(29)
(30)
(31)
j
1/2
= oP ((log n)−1 ) sup Ok (xk )|g(xk )|.
xk ∈Xk
Claims (29)–(31) can be shown similarly as (21)–(26) by using additionally condition B1.
−2/5 ) because of [B1] and [B4], see also
− µ∗,r
For r = B we note that µ∗,r,n
k
k = oP (n
P
l B
(18) and that the sum of the elements of (I − π)−1 (µ
bB − µB,∗ ) is equal to ∞
l=0 T τ (x).
P∞ l C
−2/5
For r = C one checks easily that k l=0 T τ kO,∞ = oP (n
). For the statement of the
P∞ l A
−2/5
proposition it remains to show that k l=1 T τ kO,∞ = oP (n
) and that kτ r − (µ
br0 −
∗,r
∗,r
r
−2/5
µ0 + ... + µ
bd − µd )kO,∞ = oP (n
). For the proof of these two claims one applies
condition B5.
B.2. Proof of Proposition 2
The statement of Proposition 2 follows immediately from [B3] and Proposition 1.
B.3. Proof of Theorem 3
The main tool to prove this theorem is the Newton-Kantorovich theorem, see for example
Deimling (1985). Since this theorem is central in our considerations we will state it here.
Theorem 4 (Newton-Kantorovich theorem). Consider Banach spaces X , Y and
a map F : Br (x0 ) = {x : kx − x0 k ≤ r} ⊂ X 7→ Y for x0 ∈ X and r > 0. We assume that
the Fréchet derivative F 0 exists for x ∈ Br (x0 ), that it is invertible and that the following
conditions are satisfied
(a) kF 0 (x0 )−1 F (x0 )k ≤ γ,
(b) kF 0 (x0 )−1 k ≤ β,
(c) kF 0 (x) − F 0 (x∗ )k ≤ lkx − x∗ k for all x, x∗ ∈ Br (x0 ),
(d) 2γβl < 1 and 2γ < r.
Smooth backfitting of proportional hazards
25
Then the equation
F (x) = 0
has a unique solution x∗ in B 2γ (x0 ) and furthermore, x∗ can be approximated by Newtons
iterative method
xk+1 = xk − F 0 (xk )−1 F (xk ),
and it holds that
kxk − x∗ k ≤
γ
2k−1
q2
k
−1
,
with q = 2γβl < 1.
We come now to the proof of Theorem 3.
Proof (of Theorem 3). Equation (14) can be rewritten as
b δb∗ , δb0 , . . . , δbd ) = 0,
F(
where
b ∗ , f0 , . . . , fd )(x) = Fbk (f ∗ , f0 , . . . , fd )(x)
F(f
k=∗,0,...,d
.
with
Fb∗ (f ∗ , f0 , . . . , fd )(x) =
Z
d
Y
∗
(1 + f )
{1 + fj (xj )} − 1
X
×
d
Y
j=0
Z
b
αj (xj )E(x)dx
−
j=0
Fbk (f ∗ , f0 , . . . , fd )(x) =
Z
Xxk
×
d
Y
b k (xk )dxk ,
∆
Xk
d
Y
∗
(1 + f )
{1 + fj (xj )} − 1
j=0
b
b
b ∗
αj (xj )E(x)dx
−k − ∆k (xk ) − F∗ (f , f0 , . . . , fd )(x)
j=0
R
b k (xk )dxk does not depend on k.
for k = 0, . . . , d. Note that Xk ∆
We define an additional operator F by the following equations
F(f ∗ , f0 , . . . , fd )(x) = (Fk (f ∗ , f0 , . . . , fd )(x))k=∗,0,...,d
26
Hiabu et al.
with
F∗ (f ∗ , f0 , . . . , fd )(x) =
Z
X
×
d
Y
{1 + f ∗ }
d
Y
{1 + fj (xj )} − 1
j=0
αj (xj )E(x)dx,
j=0
Fk (f ∗ , f0 , . . . , fd )(x) =
Z
Xxk
×
d
Y
{1 + f ∗ }
d
Y
{1 + fj (xj )} − 1
j=0
αj (xj )E(x)dx−k − F∗ (f ∗ , f0 , . . . , fd )(x)
j=0
for k = 0, . . . , d.
Note that F(0) = 0. The Fréchet derivatives of Fb and F in 0 are
Z
d
X
b
f ∗ +
fj (xj ) α(x)E(x)dx,
Fb∗0 (0)(f ) =
X
j=0
F∗0 (0)(f ) =
Z
X
f ∗ +
d
X
Fbk0 (0)(f ) =
Xxk
−
f ∗ +
Fk0 (0)(f ) =
Xxk
−
d
X
b
fj (xj ) α(x)E(x)dx
−k
j=0
Fb∗0 (0)(f ),
Z
fj (xj ) α(x)E(x)dx,
j=0
Z
f ∗ +
d
X
fj (xj ) α(x)E(x)dx−k
j=0
F∗0 (0)(f )
for k = 0, . . . , d.
The main idea of our proof is to apply the Newton-Kantorovich theorem, Theorem 4,
with the mapping F = Fb and norm k(f0 , ..., fd )kO,∞ ∨|f ∗ | which in abuse of notation we also
∗
denote by k(f ∗ , f0 , ..., fd )kO,∞ . As starting point x0 we choose x0 = (δ , δ). In our application of Theorem 4, the spaces X and Y are equal to R × {(f0 , ..., fd )| : k(f0 , ..., fd )| kO,∞ <
R
ej (xj )dxj = 0 for j = 0, ..., d}. We consider F and Fb as operators from X to
∞, fj (xj )O
X . Note that FX ⊂ X because of [B6] and the last assumption of [A7]. Note that we get
from [B6] and the last assumption of [A7] that
Z
1
O(x)dx = O(1).
Qd
1/2
x∈X
k=0 Ok (xk )
Smooth backfitting of proportional hazards
27
b ⊂ X , with
Similarly, one uses [B6] and the last assumption of [A7] to show that FX
probability tending to one.
We will show that
∗
Fb (δ , δ)
= op (n−2/5 ),
(32)
O,∞
and that Fb0 is locally Lipschitz around 0, i.e., that there exist constants r∗ , C such that
with probability tending to one
Fb0 (g)(f ) − Fb0 (g ∗ )(f )
O,∞
≤ C kg − g ∗ kO,∞ kf kO,∞
for all
g, g ∗ ∈ Br∗ (0).
(33)
Furthermore, we will show, that
F 0 (0) is invertible, with
F 0 (0)−1
O,∞
< C ∗,
for some C ∗ > 0.
(34)
We now argue that by application of the Newton-Kantorovich theorem (32)-(34) imply that
∗
b
(δ , δ) − (δb∗ , δ)
O,∞
= op (n−2/5 ).
(35)
This implies the statement of the theorem.
∗
= oP (1), the inequality (33)
We now show that (32)-(34) imply (35). Since (δ , δ)
O,∞
∗
also holds with a constant r for all g, g ∗ ∈ Br (δ , δ) with probability tending to one.
This gives condition (c) of the Newton-Kantorovich theorem.
Furthermore, by application of (18) we get that Fb0 (0) − F 0 (0)
= oP (1). This
O,∞
∗
together with (δ , δ)
O,∞
= oP (1) and (33) gives
∗
Fb0 (δ , δ) − F 0 (0)
O,∞
= op (1).
Therefore
with probability tending to one, condition (34) also holds if F 0 (0) is replaced by
∗
Fb0 (δ , δ) . Thus, we get from (32)-(34) that conditions (a)–(d) of the Newton-Kantorovich
theorem are fulfilled with probability tending to one, with γ = oP (n−2/5 ). This shows (35).
∗
It remains to show (32), (33) and (34). For the proof of (32) note that (δ , δ)
=
O,∞
op
(n−1/5 )
and that Fb0 is Lipschitz. A first order Taylor expansion yields
∗
b
b
F(δ)
= F(0)
+ Fb0 (0) (δ , δ) + op (n−2/5 ).
∗
0
b
b
b
b
Equation (32) follows from F(0) + F (0) (δ , δ) = F(0)
− F(0)
= 0.
Claim (33) follows directly from assumption [B6].
For the proof of (34) we have to show that F 0 (0) is invertible. For the proof of this
claim we start by showing that it is bijective. For the proof of injectivity, assume that
28
Hiabu et al.
F 0 (0)(f ) = 0 for some f = (f ∗ , f0 , ..., fd )| ∈ X . We will show that this implies that f = 0.
It holds that
Z
d
X
f ∗ +
fj (xj ) α(x)E(x)dx = 0,
X
j=0
Z
Xxk
f ∗ +
d
X
fj (xj ) α(x)E(x)dx−k = 0,
for all k = 0, . . . , d.
j=0
R
R
P
With f¯j (xj ) = fj (xj ) − fj (uj )v(u)du and f¯∗ = f ∗ + dj=0 fj (uj )v(u)du for v(u) =
R
α(u)E(u)/ α(s)E(s)ds this implies that
f¯∗
Z
α(x)E(x)dx = 0
X
and thus it holds that f¯∗ = 0. Furthermore, we get that for k = 0, . . . , d
Z
Z
Z
d
d
X
X
f¯j (xj ) α(x)E(x)dx−k dxk =
f¯k (xk )
f¯j (xj ) α(x)E(x)dx.
0=
f¯k (xk )
Xk
Xxk
X
j=0
j=0
By summing these terms up over k, we get that
Z X
d
X
j=0
2
f¯j (xj ) α(x)E(x)dx = 0,
which implies that
d
X
f¯j (xj ) = 0,
a.e. on X .
j=0
By application of (17) this implies that f¯j and fj are constant functions. Because of
R
b
fj (uj )α(u)E(u)du
= 0 this implies f = 0.
Now we check that F 0 (0) is surjective. Consider g = (g ∗ , g0 , ..., gd )| ∈ X such that
0
hF (0)(f ), gi =
F∗0 (0)(f )g ∗
+
d Z
X
k=0
Xk
Fk0 (0)(f )(xk )gk (xk )dxk = 0
(36)
for all f = (f ∗ , f0 , ..., fd )| ∈ X . We will show that then g = 0. This implies that g = 0
is the only element in Y that is perpendicular to the range space of F 0 (0). Since F 0 (0) is
linear, this shows that F 0 (0) is surjective.
Smooth backfitting of proportional hazards
29
From (36) one gets with the choice fk = gk and f ∗ = g ∗ that
Z
Z
d Z
d
X
X
0 = (g ∗ )2
α(x)E(x)dx +
gk (xk )
gj (xj ) α(x)E(x)dx−k dxk
X
k=0
Xk
Xxk
j=0
2
Z
Z
d
X
= (g ∗ )2
α(x)E(x)dx +
gj (xj ) α(x)E(x)dx.
X
X
j=0
With exactly the same arguments as for the injectivity we conclude that g = 0 and that
g ∗ = 0. Thus, we have shown that F 0 (0) is invertible.
It remains to show that F 0 (0)−1 is bounded. By the bounded inverse theorem for this
claim it suffices to show that F 0 (0) is bounded. Boundedness of F 0 (0) can be shown by
application of (26). This concludes the proof of Theorem 3.
C.
Discrete data
Data is given as Nr0 ,r , with (r0 , r) ∈ Idisc , Idisc = {(r0 , r)| r0 = 1, . . . , R0 ; r = 0, . . . , R0 − 1
and r0 ≤ r}. We define occurrence Or0 ,r and exposure Er0 ,r .
nr0 Z r+1
X
0
dNr0 ,j (s) = Nr0 ,(R0 −r) ,
Or ,r =
Er0 ,r =
j=1 r
nr0 Z r+0.5
X
j=1
Yr0 ,j (s)ds = Yr0 ,j (r + 0.5) =
r−0.5
X
Nr0 ,k .
k≤(R0 −r)
Then the unstructured local linear hazard estimator, α
e, becomes
P
0
−1
0
r0 ,r∈Idisc 1 − (x − (r + 0.5, r ))Ddisc (x) c1,disc (x) Kb (x − (r + 0.5, r ))Or0 ,r
,
α
e(x) = P
0
−1
0
r0 ,r∈Idisc {1 − (x − (r + 0.5, r ))Ddisc (x) c1,disc (x)} Kb (x − (r + 0, 5, r ))Er0 ,r
where Ddisc and c1,disc are the discrete versions of D and c1 , respectively:
X
c11,disc (x) = n−1
Kb (x − (r + 0.5, r0 ))(t − r + 0.5)Er0 ,r ,
r0 ,r∈Idisc
X
c12,disc (x) = n−1
Kb (x − (r + 0.5, r0 ))(t − r0 )Er0 ,r ,
r0 ,r∈Idisc
d00,disc (x) =
X
Kb (x − (r + 0.5, r0 ))(t − r + 0.5)2 Er0 ,r ,
r0 ,r∈Idisc
d01,disc (x) =
X
Kb (x − (r + 0.5, r0 ))(t − r + 0.5)(z − r0 )Er0 ,r ,
r0 ,r∈Idisc
d11,disc (x) =
X
r0 ,r∈Idisc
Kb (x − (r + 0.5, r0 ))(z − r0 )2 Er0 ,r .
30
Hiabu et al.
The cross-validation criterion can be written as
X
Q(b) = n−1
α
b(r, r0 ) − α(r, r0 )
2
Er0 ,r ,
0
r ,r∈Idisc
and thus
b b = n−1
Q
X
α
b(r, r0 )
2
Er0 ,r − 2
r0 ,r∈Idisc
X
0
α
b[r,r ] (r, r0 )Or0 ,r .
r0 ,r∈Idisc
Finally,
Z
b
fz (t) = α
b0 (R0 − t)α
b1 (z) exp −
R0 −t
0
α
b0 (s)α
b1 (z)ds .
Acknowledgements
The research of the second author was supported by the Deutsche Forschungsgemeinschaft
through the Research Training Group RTG 1953. Third author acknowledges the support from the Spanish Ministry of Economy and Competitiveness, through grant number
MTM2013- 41383P, which includes support from the European Regional Development Fund
(ERDF).
References
Andersen, P., O. Borgan, R. Gill, and N. Keiding (1993). Statistical Models Based on
Counting Processes. New York: Springer.
Beran, R. (1981). Nonparametric regression with randomly censored survival data. Technical report, Dept. Statist., Univ. California, Berkeley.
Bowman, A. W. (1984). An alternative method of cross-validation for the smoothing of
density estimates. Biometrika 71, 353–360.
Dabrowska, D. M. (1987). Non-parametric regression with censored survival time data.
Scand. Actuar. J. 14, 181–197.
Deimling, K. (1985). Nonlinear functional analysis. Berlin: Springer.
Fan, J., I. Gijbels, and M. King (1997). Local likelihood and local partial likelihood in
hazard regression. Ann. Stat. 25, 1661–1690.
Gámiz, M. L., L. Janys, M. D. Martı́nez-Miranda, and J. P. Nielsen (2013). Bandwidth
selection in marker dependent kernel hazard estimation. Comput. Stat. Data An. 68,
155–169.
Hall, P. (1983). Large sample optimality of least squares cross-validation in density estimation. Ann. Stat. 11, 1156–1174.
Hastie, T. J. and R. J. Tibshirani (1990). Generalized Additive Models. London: Chapman
and Hall.
Smooth backfitting of proportional hazards
31
Hiabu, M. (2017). On the relationship between classical chain ladder and granular reserving.
Scand. Actuar. J. To appear.
Hiabu, M., E. Mammen, M. D. Martı́nez-Miranda, and J. P. Nielsen (2016). In-sample
forecasting with local linear survival densities. Biometrika 103, 843–859.
Lee, Y. K., E. Mammen, J. P. Nielsen, and B. U. Park (2015). Asymptotics for in-sample
density forecasting. Ann. Stat. 43, 620–651.
Lee, Y. K., E. Mammen, J. P. Nielsen, and B. U. Park (2017). Operational time and
in-sample density forecasting. Ann. Stat. To appear.
Lin, H., Y. He, and J. Huang (2016). A global partial likelihood estimation in the additive
cox proportional hazards model. J. Statist. Plann. Inference 169, 71–87.
Linton, O. B. and J. P. Nielsen (1995). A kernel method of estimating structured nonparametric regression based on marginal integration. Biometrika 82, 93–100.
Linton, O. B., J. P. Nielsen, and S. Van de Geer (2003). Estimating multiplicative and
additive hazard functions by kernel methods. Ann. Stat. 31, 464–492.
Mammen, E., O. B. Linton, and J. P. Nielsen (1999). The existence and asymptotic properties of a backfitting projection algorithm under weak conditions. Ann. Stat 27, 1443–1490.
Mammen, E., M. D. Martı́nez-Miranda, and J. P. Nielsen (2015). In-sample forecasting
applied to reserving and mesothelioma. Insurance Math. Econom. 61, 76–86.
Martı́nez-Miranda, M. D., J. P. Nielsen, S. Sperlich, and R. Verrall (2013). Continuous
chain ladder: Reformulating and generalising a classical insurance problem. Expert. Syst.
Appl. 40, 5588–5603.
Martinussen, T. and T. H. Scheike (2006). Dynamic regression models for survival data.
New York: Springer.
McKeague, I. W. and K. J. Utikal (1990). Inference for a nonlinear counting process regression model. Ann. Stat. 18, 1172–1187.
Nielsen, J. P. (1998). Marker dependent kernel hazard estimation from local linear estimation. Scand. Actuar. J. 1998, 113–124.
Nielsen, J. P. and O. B. Linton (1995). Kernel estimation in a non-parametric marker
dependent hazard model. Ann. Stat. 23, 1735–1748.
Nielsen, J. P. and S. Sperlich (2005). Smooth backfitting in practice. J. Roy. Statist. Soc.
Ser. B 67, 43–61.
Rudemo, M. (1982). Empirical choice of histograms and kernel density estimators. Scand.
J. Stat. 9, 65–78.
Stone, C. J. (1982). Optimal global rates of convergence for nonparametric regression. Ann.
Stat. 10, 1040–1053.
32
Hiabu et al.
Stone, C. J. (1985). Additive regression and other nonparametric models. Ann. Stat. 13,
689–705.
Yu, K., B. U. Park, and E. Mammen (2008). Smooth backfitting in generalized additive
models. Ann. Stat. 36, 228–260.
| 10math.ST
|
To appear in Theory and Practice of Logic Programming.
1
arXiv:1210.2297v2 [cs.PL] 9 Oct 2012
Diagrammatic confluence for
Constraint Handling Rules∗
Rémy Haemmerlé
Technical University of Madrid
Abstract
Confluence is a fundamental property of Constraint Handling Rules (CHR) since, as in
other rewriting formalisms, it guarantees that the computations are not dependent on
rule application order, and also because it implies the logical consistency of the program
declarative view. In this paper we are concerned with proving the confluence of nonterminating CHR programs. For this purpose, we derive from van Oostrom’s decreasing
diagrams method a novel criterion on CHR critical pairs that generalizes all preexisting
criteria.We subsequently improve on a result on the modularity of CHR confluence, which
permits modular combinations of possibly non-terminating confluent programs, without
loss of confluence.
KEYWORDS: CHR, confluence, decreasing diagrams, modularity of confluence.
1 Introduction
Constraint Handling Rules (CHR) is a committed-choice constraint logic programming language, introduced by Frühwirth (1998) for the easy development of constraint solvers. It has matured into a general-purpose concurrent programming language. Operationally, a CHR program consists of a set of guarded rules that rewrite
multisets of constrained atoms. Declaratively, a CHR program can be viewed as a
set of logical implications executed on a deduction principle.
Confluence is a basic property of rewriting systems. It refers to the fact that
any two finite computations starting from a common state can be prolonged so as
to eventually meet in a common state again. Confluence is an important property
for any rule-based language, because it is desirable for computations to not be dependent on a particular rule application order. In the particular case of CHR, this
property is even more desirable, as it guarantees the correctness of a program (Abdennadher et al. 1999; Haemmerlé et al. 2011): any program confluent has a consistent logical reading. Confluence of a CHR program is also a fundamental prerequisite for logical completeness results (Abdennadher et al. 1999; Haemmerlé 2011a),
∗ The research leading to these results has received funding from the Programme for Attracting
Talent / young PHD of the MONTEGANCEDO Campus of International Excellence (PICD),
the Madrid Regional Government under the CM project P2009/TIC/1465 (PROMETIDOS),
and the Spanish Ministry of Science under the MEC project TIN-2008-05624 (DOVES).
2
Rémy Haemmerlé
makes possible program parallelization (Frühwirth 2005; Meister 2006), and may
simplify program equivalence analyses (Abdennadher and Frühwirth 1999; Haemmerlé 2011b).
Following the pioneering research of Abdennadher et al. (1996), most existing
work dealing with the confluence of CHR limits itself to terminating programs (see
for instance the works by Abdennadher (1997) and Duck et al. (2007)). Nonetheless,
proving confluence without global termination assumptions is still a worthwhile
objective.
From a theoretical point of view, this is an interesting topic, because, as illustrated by the following example typical CHR programs fail to terminate on the level
of abstract semantics, even if they do terminate on more concrete levels. Indeed,
number of analytical results for the language rest on the notion of confluence, but
only when programs are considered with respect to abstract semantics. For instance,
in the current state of knowledge, even a result as important as the guarantee of
correction by confluence only holds when programs are considered with respect to
the most general operation semantics for CHR, namely the very abstract semantics.
Example 1 (Partial order constraint)
Let P1 be the classic CHR introductory example, namely the constraint solver for
partial order. This consists of the following four rules, which define the meaning of
the user-defined symbol ≤ using the built-in equality constraint = :
duplicate
reflexivity
antisymmetry
transitivity
@
@
@
@
x ≤ y \ x ≤ y ⇐⇒ >
x ≤ x ⇐⇒ >
x ≤ y, y ≤ x ⇐⇒ x = y
x ≤ y, y ≤ z =⇒ x ≤ z
The duplicate rule implements so-called duplicate removal. In other words, it
states that if two copies of the same user-defined atom are present, then one of
them can be removed. The reflexivity and transitivity rules respectively state that
any atom of the form x ≤ x can be removed, and that two atoms x ≤ y and x ≤ y
can be substituted with the built-in constraint x = y. Finally, the transitivity rule
is a propagation rule. It states that if x ≤ y and y ≤ z are present, then the atom
x ≤ z may be added.
It is well know that this program, like any other program using propagation rules,
faces the so-called trivial non-termination problem when considered with respect to
the very abstract semantics. Indeed, for these semantics, a propagation rule applies
to any state it produces, leading to trivial loops. In order to solve this problem,
Abdennadher (1997) proposed a token-based semantics in which propagation rules
may be applied only once to the same combination of atoms. Nonetheless, such a
proposal does not solve all the problems of termination. Indeed the transitivity rule
may loop on queries containing a cycle in a chain of inequalities when considered
against Abdennadher’s semantics. Consider, for instance, the query x ≤ y, y ≤ x.
In fact, in order for P1 to be terminating, the rules of reflexivity, antisymmetry,
and transitivity must have priority over the transitivity rule. This behaviour can
be achieved by considering concrete semantics, such as the refined semantics of
Diagrammatic confluence for Constraint Handling Rules
3
Duck et al. (2005). These semantics reduce the non-determinism of the CHR execution model by applying the rules in textual order.
In exchange for gaining termination, the most concrete semantics lose a number
of analytical results. For instance, as explained by Frühwirth (2009), although any
CHR program can be run in parallel in abstract semantics, one can obtain incorrect
results for programs written with the refined semantics in mind. Indeed, if the
result of a program relies on a particular rule application order, parallel execution
will garble this order, leading to unexpected results. Interestingly, confluence on
an abstract (but possibly non-terminating) level may come to the rescue of the
most concrete semantics: If a program is confluent on a semantic level where the
rule application order is not specified, then the result will not be dependent on
the particular application order. Similar considerations have been discussed for
equivalences of CHR programs (Haemmerlé 2011b).
From a more practical point of view, proving confluence without the assumption
of termination is important, because it may be desirable to prove the confluence
of a program for which termination cannot be inferred. Indeed, there exist very
simple programs, such as the Collatz function, for which termination is only a conjecture (Guy 2004). Furthermore, since CHR is now a general-purpose language,
analytical tools for the language must handle programs that do not terminate on any
semantic level—for instance, interpreters for a Turing-complete language (Sneyers
et al. 2009), or typical concurrent programs (see the numerous examples of concurrent systems given by Milner (1999)). We have also recently demonstrated that
non-terminating execution models for CHR yield elegant frameworks for programming with coinductive reasoning (Haemmerlé 2011a). As a motivating example for
the class of intrinsically non-terminating programs, we will use the following solution for the seminal dining philosophers problem.
Example 2 (Dining philosophers)
Consider the following CHR program P2 that implements a solution to the dining
philosophers problem extended to count the number of times a philosopher eats:
eat
thk
@
@
t(x, y, i), f(x), f(y)
e(x, y, i)
⇐⇒
⇐⇒
e(x, y, i + 1)
f(x), f(y), t(x, y, i)
The atom f(x) represents the fork x, the atom e(x, y, i) (resp. t(x, y, i)) represents
an eating (thinking) philosopher seated between forks x and y, who has already
eaten i times. On the one hand, the rule eat, states that if a thinking philosopher
is seated between two forks lying on the table, then he may start eating once he
has picked up both forks. On the other hand, the rule thk states that a philosopher
may stop eating if he puts down the forks he has been using. The initial state
corresponding to n dining philosophers seated around a table can be encoded by
the set of atoms f(1), t(1, 2, 0), f(2), t(2, 3, 0), · · · f(n), t(n, 1, 0).
Despite the fact that this program is intrinsically non-terminating, we may be
interested in its confluence, for example, so that we may make use of one of the
previously mentioned applications (e.g. confluence simplifies observational equivalence (Haemmerlé 2011b)). Confluence of P2 may also simplify the proofs of funda-
4
Rémy Haemmerlé
mental properties of concurrent systems, such as, for instance, the absence of deadlock: Starting from the initial state, one can easily construct a derivation where the
ith philosopher (i ∈ 1, . . . , n) has eaten an arbitrary number of times. Hence if P2 is
confluent, we can then infer that it is possible to extend any finite derivation such
that the ith philosopher eats strictly more, i.e. no derivation leads to a deadlock.
To the best of our knowledge, the only existing principle for proving confluence of
non-terminating programs is the so-called strong confluence criterion (Haemmerlé
and Fages 2007; Raiser and Tacchella 2007). However this criterion appears to be too
weak to apply to common CHR programs, such as Examples 1 and 2. In this paper,
we are concerned with extending CHR confluence theory to be able to capture
a large class of possibly non-terminating programs. For this purpose we derive
from the so-called decreasing diagrams technique a novel criterion that generalizes
all existing confluence criteria for CHR. The decreasing diagrams technique is a
method developed by van Oostrom (1994) which subsumes all sufficient conditions
for confluence. Applying this method requires that all local rewrite peaks (i.e. points
where the rewriting relation diverges because of non-determinism) can be completed
into so-called decreasing diagrams.
The present paper presents two main contributions. In Sect. 4, we present a
particular instantiation of the decreasing diagrams technique to CHR, and show
that in the context of this particular instantiation, the verification of decreasingness can be restricted to the standard notion of critical pairs. Then in Sect. 5, we
extend the so-called modularity of confluence (Frühwirth 2009) so as to be able to
combine programs which have independently been proven confluent, without losing
confluence.
2 Preliminaries on abstract confluence
In this section, we gather some required notations, definitions, and results on the
confluence of abstract rewriting systems. Terese’s compendium (2003) can be referred to for a more detailed presentation.
A rewrite relation (or rewrite for short) is a binary relation on a set of objects E.
For any rewrite →, the symbol ← will denote its converse, →≡ its reflexive closure,
→+ its transitive closure, and its transitive-reflexive closure. We will use →α ·→β
to denote the left-composition of all rewrites →α and →β . A family of rewrites is a
set (→α )α∈I of rewrites indexed by a set I of labels. For such a family and any set
S
K, →K will denote the union α∈(K∩I) (→α ).
P
P
A reduction is a finite sequence of rewriting steps of the form (e0 −
→
→
α 1 e1 −
α2
P
P
··· −
→
→
αn en ). Such a reduction would be abreviated as e0 −
ᾱ en with ᾱ =
α1 , α2 , . . . , αn when the intermediary states e1 , . . . , en−1 are not relevant. A peak
P
P
is a pair of reductions el ←
−ᾱ e −
→
β̄ er from a common element e. A local peak
is a peak formed by two one-step reductions. A valley is a pair of reductions
P
P
P
0 P
el −
→
−β̄ er ending in a common element e0 . A peak el ←
−ᾱ e −
→
ᾱ e ←
ᾱ0 er
P
P
P
P
is joinable by −
→
−β 0 if it is true that el −
→
− β 0 er .
β ·←
β ·←
≡
Fig. 1. Confluence
Fig. 2. Local Confluence
5
≡
Diagrammatic confluence for Constraint Handling Rules
Fig. 3. Strong confluence
A rewrite → is terminating if there is no infinite sequence of the form e0 → e1 →
e2 . . . Furthermore, we will say that → is confluent if (·) ⊆ (·) holds, locally
confluent if (←·→) ⊆ (·) holds, and strongly confluent1 if (←·→) ⊆ (→≡ ·←≡ )
holds. Figures 1, 2 and 3 graphically represent these definitions. Following standard
diagrammatic notation, solid edges stand for universally quantified rewrites, while
dashed edges represent existentially quantified rewrites.
By the seminal lemma of Newman (1942), we know that a terminating and locally
confluent rewrite is confluent. Another famous result due to Huet (1980) ensures
that strong confluence implies confluence.
We now present a slight variation due to Hirokawa and Middeldorp (2010) of
the so-called decreasing diagrams technique, which is more suitable for our purposes. The interest of the decreasing diagrams method (van Oostrom 1994) is that
it reduces problems of general confluence to problems of local confluence. In exchange, the method requires the confluence diagrams (i.e. the way peaks close)
to be decreasing with respect to a labeling provided with a wellfounded preorder.
The method is complete in the sense that any countable confluence rewrite can be
equipped with such a labeling. But because confluence is an undecidable property,
finding such labeling may be difficult.
In the rest of this paper, we will say that a preorder < is wellfounded, if the strict
preorder associated with < (i.e. α β iff α < β but not β < α) is a terminating
relation. Let (→α )α∈I be a family of rewrites and < be a wellfounded preorder on
I. A local peak el ←α e →β er (α, β ∈ I) is decreasing with respect to < if the
following holds:
· ←≡ {α} ·
{β}
{α,β}
<
e0
{α,β}
· →≡ {β} ·
<
{α}
el
er
(?)
1
<
where for any set K of labels, K stands for {γ ∈ I | ∃δ ∈ K.δ < γ} and K for
{γ ∈ I | ∃δ ∈ K.δ γ}. A family (→α )α∈I of rewrites is (locally) decreasing if all
local peaks of the form u ←α · →β v (α, β ∈ I) are decreasing with respect to a
common wellfounded preorder on I. A rewrite is (locally) decreasing if it is the union
of some decreasing families of rewrites. Property (?) is graphically represented in
Figure 4.
For the sake of simplicity, we use a definition weaker than the one of Huet (1980). It is worth
noting, that the counterexamples given in introduction stay relevant for the general definition.
Rémy Haemmerlé
β
α
6
{β
}
{α
}
<
{β
{α
≡
≡
}
}
<
{α
,β
}
,β
{α
}
Fig. 4. Local decreasingness
Theorem 3 (Decreasing Diagram (van Oostrom 1994))
A countable rewrite is confluent if and only if it is locally decreasing.
We recall now some other state-of-the-art results which will be used later.
Lemma 4 (Terese 2003)
(i) For all rewrites →1 , →2 if (←1 · 2 ) ⊆ (2 · 1 ), then (1 · 2 ) ⊆ (2 · 1 ).
(ii) For all rewrites →1 ,→2 s.t. →1 ⊆ →2 ⊆ 1 , →2 is confluent iff →1 is confluent.
3 Preliminaries on Constraint Handling Rules
In this section, we recall the syntax and the semantics of CHR. Frühwirth’s book
(2009) can be referred to for a more general overview of the language.
3.1 Syntax
The formalization of CHR assumes a language of (built-in) constraints containing
equality over some theory C, and defines (user-defined) atoms using a different set of
predicate symbols. In the following, R will denote an arbitrary set of identifiers. By
a slight abuse of notation, we allow confusion of conjunctions and multiset unions,
omit braces around multisets, and use the comma for multiset union. We use fv(φ)
to denote the set of free variables of a formula φ. The notation ∃ -ψ φ denotes the
existential closure of φ with the exception of free variables of ψ.
A (CHR) program is a finite set of eponymous rules of the form:
(r @ K\H ⇐⇒ G | B; C)
where K (the kept head), H (the removed head), and B (the user body) are multisets
of atoms, G (the guard) and C (the built-in body) are conjunctions of constraints
and, r ∈ R (the rule name) is an identifier assumed unique in the program. Rules
in which both heads are empty are prohibited. An empty guard > (resp. an empty
kept head) can be omitted with the symbol | (resp. with the symbol \). Rules are
Diagrammatic confluence for Constraint Handling Rules
7
divided into two classes: simplification rules2 if the removed head is non-empty and
propagation rules otherwise. Propagation rules can be written using the alternative
syntax:
(r @ K =⇒ G | B; C)
3.2 Operational semantics
In this section, we recall the equivalence-based operational semantics ωe of Raiser
et al. (2009). It is equivalent to the very abstract semantics ωva of Frühwirth (1998),
which is the most general operational semantics of CHR. We prefer the former because it includes an rigorous notion of equivalence, which is an essential component
of confluence analysis.
A (CHR) state is a tuple hC; E; x̄i, where C (the user store) is a multiset of atoms,
E (the built-in store) is a conjunction of constraints, and x̄ (the global variables)
is a finite set of variables. Unsurprisingly, the local variables of a state are those
variables of the state which are not global. When no confusion can occur, we will
syntactically merge user and built-in stores. We may futhermore omit the global
variables component when states have no local variables. In the following, we use Σ
to denote the set of states. Following Raiser et al., we will always implicitly consider
states modulo a structural equivalence. Formally, this state equivalence is the least
equivalence relation ≡ over states satisfying the following rules:
•
•
•
•
hE; C; x̄i ≡ hE; D; x̄i if C ∃ -(E,x̄) C ↔ ∃ -(E,x̄) D
hE; ⊥; x̄i ≡ hF; ⊥; ȳi
hA, c; C, c=d; x̄i ≡ hA, d; C, c=d; x̄i
hA; C; x̄i ≡ hA; C; {y} ∪ x̄i if y ∈
/ fv(A, C).
Once states are considered modulo equivalence, the operation semantics of CHR
can be expressed by a single rule. Formally the operational semantcs of a program
P
P is given by the least relation −
→
on states satisfying the rule:
(r @ K\H ⇐⇒ G|B; C) ∈ Pρ lv(r) ∩ fv(E, D, x̄) = ∅
P
hK, H, E; G, D; x̄i −
→
hK, B, E; G, C, D; x̄i
P
where ρ is a renaming. A program P is confluent (resp. terminating) if −
→
is confluent (resp. terminating).
Before going further, we recall an important property of CHR semantics. This
property, monotonicity, means that if a transition is possible in a state, then the
same transition is possible in any larger state. To help reduce the level of verbosity
we introduce the notion of the quantified conjunction of states (Haemmerlé and
Fages 2007). This operator allows the composition of states with disjoint local variables while quantifying some of their global variables (i.e. changing global variables
2
Unlike standard presentations, our definition does not distinguish between simplification rules
form the so-called simpagation rules.
8
Rémy Haemmerlé
into local ones). Formally, the quantified conjunction is a binary operator on states
parametrized by a set of variables z̄ satisfying:
hE; C; x̄i ⊕z̄ hF; D; ȳi = hE, F; C, D; (x̄ȳ) \ z̄i if (fv(E, C) ∩ fv(F, D)) ⊆ (x̄ ∩ ȳ)
Note the side condition is not restrictive, as local variables can always be renamed
using the implicit state equivalence.
Proposition 5 (Monotonicity of CHR)
Let P be a CHR program, S, S1 , S2 be CHR states, and x̄ be a set of variables.
P
P
If S1 −
→
S2 , then S1 ⊕x̄ S −
→
S2 ⊕x̄ S
3.3 Declarative semantics
Owing to its origins in the tradition of CLP, the CHR language features declarative
semantics through direct interprestation in first-order logic. Formally, the logical
reading of a rule of the form:
K\H ⇐⇒ G | B; C
is the guarded equivalence:
∀ (K ∧ G) → H ↔ ∃ -(K,H) (G ∧ C ∧ B)
The logical reading of a program P within a theory C is the conjunction of the
logical readings of its rules with the constraint theory C. It is denoted by CP.
Operational semantics is sound and complete with respect to this declarative
semantics (Frühwirth 1998; Abdennadher et al. 1999). Furthermore, any program
confluent with respect to ωe has a consistent logical reading (Abdennadher et al.
1999; Haemmerlé et al. 2011).
4 Diagrammatic confluence for Constraint Handling Rules
In this section, we are concerned with proving confluence of a large class of CHR
programs. Indeed, as explained in the introduction, existing criteria are not sufficiently powerful to infer confluence of common non-terminating programs. (See
Examples 13 and 14 for concrete examples). To avoid this limitation, we will derive
from the decreasing diagrams technique a novel csriterion on CHR critical pairs
that generalizes both local and strong confluence criteria. An analogue criterion
has been developed for linear Term Rewriting Systems (TRS) (Jouannaud and van
Oostrom 2009).
4.1 Labels for Constraint Handling Rules
In order to apply the decreasing diagram technique to CHR, we will need first to
label CHR transitions. In this work, we will use two labelings proposed by van
Oostrom (2008) for TRS. The first one is the so-called rule-labeling. It consists of
P
labeling each transition a −
→
b with the name of the applied rule. This labeling
Diagrammatic confluence for Constraint Handling Rules
9
is ideal for capturing strong confluence-like properties for linear TRS. Within the
proof of our main result, we will also use the so-called self-labeling which consists
P
of labeling each transition a −
→
b with its source a. This second labeling captures
the confluence of terminating rewrites.
In practice, we will assume that the set R of rule identifiers is defined as a disjoint
union Ri ] Rc . For a given program P, we denote by P i (resp. P c ) the set of rules
form P built with Ri (resp. Rc ). We call P i the inductive part of P, because we will
subsequently assume that P i is terminating, while P c will be called coinductive, as
it will be typically non-terminating.
Definition 6 (Rule-labeling)
P
The rule-labeling of a CHR program P is the family of rewrites (−
→
r )r∈R indexed
{r}
P
by rule identifiers, where −
→
−→. A preorder < on rule identifiers is admissible,
r =−
if any inductive rule identifier is strictly smaller than any coinductive one (i.e. for
any ri ∈ Ri and any rc ∈ Rc , rc ri holds).
4.2 Critical peaks
In TRS, the basic techniques used to prove confluence consist of showing various
confluence criteria on a finite set of special cases, called critical pairs. Critical pairs
are generated by a superposition algorithm, in which one attempts to capture the
most general way the left-hand sides of the two rules of the system may overlap.
The notion of critical pairs has been successfully adapted to CHR by Abdennadher
et al. (1996). Here, we introduce a slight extension of the notion that takes into
account the rule-labeling we have just defined.
Definition 7 (Critical peak )
Let us assume that r1 and r2 are CHR rules renamed apart:
(r1 @ K1 \H1 ⇐⇒ G1 | B1 ; C1 ) ∈ P1
(r2 @ K2 \H2 ⇐⇒ G2 | B2 ; C2 ) ∈ P2
A critical ancestor (state) Sc for the rules r1 and r2 is a state of the form:
∩
∆
Sc = hH∆
1 , H1 , H2 ; D; x̄i
satisfying the following properties:
•
•
•
•
∩
∩
∩
∩
(K1 , H1 )=(H
˙ ∆
˙ ∆
1 , H1 ), (K2 , H2 )=(H
2 , H2 ), H1 6= ∅, and H2 6= ∅;
x¯1 = fv(K1 , H1 ), x¯2 = fv(K2 , H2 ) and x̄ = x¯1 ∪ x¯2 ;
D = (H∩
˙ ∩
1 =H
2 , G1 , G2 ) and ∃D is C-satisfiable;
∩
H1 6⊆ K1 or H∩
2 6⊆ K2 .
Then the following tuple is called a critical peak between r1 and r2 at Sc :
P
P
∆
hK1 , B1 , H∆
−r1 Sc −
→
r2 hK2 , B2 , H1 ; D, C2 ; x̄i
2 ; D, C1 ; x̄i ←
A critical peak between a program P and a program Q is a critical peak between a
rule of P and a rule of Q. A critical peak of a program P is a critical peak between
P and itself. A critical peak is inductive if it involves only inductive rules (i.e. a
critical peak of P i ), or coinductive if it involves at least one coinductive rule (i.e. a
critical peak between P c and P).
10
Rémy Haemmerlé
Example 8
Consider the solver partial order P1 , given in Example 1. The following ciritial peak
stems from overlapping the heads of the rules antisymmetry and transitivity:
P
P
1
1
hx = yi ←−−
−→
anti. hx ≤ y, y ≤ xi −
trans. hx ≤ y, y ≤ x, x ≤ xi
4.3 Rule-decreasingness
We now come to our main result, showing that the study of decreasingness with
respect to the rule-labeling can be restricted to critical peaks without loss of generality.
Definition 9 (Critical rule-decreasingness)
A program P is (critically) rule-decreasing w.r.t. an admissible preorder < if:
• the inductive part of P is terminating,
Pi
Pi
• all inductive critical peaks of P are joinable by − · −, and
• all coinducitve critical peaks of P are decreasing w.r.t. <.
A program is rule-decreasing if it is rule-decreasing with respect to some admissible preorder. A rule-decreasing program is strongly rule-decreasing if it is purely
coinductive (i.e. without inductive rules).
Theorem 10
Rule-decreasing programs are confluent.
Proof
Let us assume that P is a rule-decreasing program w.r.t. a given preorder <R . Now
P
let (−
→
α )α∈(Σ∪Rc ) , the family of rewrites indexed by rule or state, be defined as
P
−
→α =
( i
−P−→ ∩({α} × Σ)
{α}
−−→
if α ∈ Σ
(self-labeling on inductive part)
if α ∈ Rc
(rule-labeling on coinductive part)
i
P
Let < be the union of <R , −
→ + , and {(r, α) | r ∈ R & α ∈ I}. By assuming
without loss of generality that R is finite (i.e. <R is trivially wellfounded), we
obtain that < is wellfounded. With the help of Theorem 3, it suffices to prove that
P
P
each peak Sα ←
−α S −
→
β Sβ (α, β ∈ (Rc ∪ Σ)) is decreasing w.r.t. <. We distinguish
two cases:
1 The rules rα and rβ used to respectively produce Sα and Sβ apply to different parts
{rβ }
{rα }
of S. By monotonicity of CHR transitions, we infer Sα −−−→ S 0 ←−−− Sβ . We have
to show this valley respects property (?) within the definition of the decreasing
diagrams. We proceed by cases on the types of the rules rα and rβ :
i
P +
P
0
1.1 rα is inductive. We have α = S, α −
→ Sα , and Sβ −
→
Sβ S .
i
P
P
0
i
1.1.1 rβ is inductive. We have β = S, β −
→ Sβ , and Sα −
→
Sα S . Since P is
P
0 P
terminating, we infer α Sα and β Sβ . We conclude Sα −
→Sα S ←
−Sβ Sβ ,
i.e. the peak is decreasing w.r.t. <.
Diagrammatic confluence for Constraint Handling Rules
11
P
0
1.1.2 rβ is coinductive. We have β ∈ Rc , Sα −
→
β S , and β Sα . We conclude
P
P
0
−β Sβ , i.e. the peak is decreasing w.r.t. <.
Sα −
→Sα S ←
P
0
1.2 rα is coinductive. We have α ∈ Rc and Sβ −
→
α S .
1.2.1 rβ is inductive. The case is symmetric with case 1.1.2.
P
P
0
1.2.2 rβ is coinductive. We have β ∈ Rc and Sα −
→
→
β S . We conclude Sα −
α
P
−β Sβ , i.e. the peak is decreasing w.r.t. <.
S0 ←
2 The applications of the rules rα and rβ used to respectively produce Sα and Sβ
P
P
overlap. There should exist a critical peak Rα ←
−rα Sc −
→
rβ Rβ , a state R, and a
set of variables ȳ, such that S ≡ Sc ⊕x̄ R, Sα ≡ Rα ⊕x̄ R, and Rβ ≡ Rβ ⊕x̄ R. We
proceed by cases on the types of rules rα and rβ :
2.1 Both rules are inductive: We have β = α = S, and by hypothesis we have
i
i
i
i
P
P
0 P
1 P
Rα ≡ Rα
−→ Rα
−→ · · · Sαm ≡ S 0 ≡ Rβn · · · ←
− Rβ1 ←
− Rβ0 ≡ Rβ
By monotony of CHR we infer:
i
i
i
i
P
P
P
P
Sα ≡ Sα0 −
→ Sα1 −
→ · · · Sαm ≡ S ≡ Sβn · · · ←
− Sβ1 ←
− Sβ0 ≡ Sβ
i
where Sαi = Rα
⊕x̄ R (for i ∈ 0, . . . m), Sβi = Rβi ⊕x̄ R (for i ∈ 0, . . . n), and
P
S = S 0 ⊕ R. By construction of (−
→
α )α∈Σ×Rc we get:
P
P
P
1 P
0 S
Sα −
→
→ Sα1 · · · Sαm ≡ S ≡ Sβn · · · ←
−Sβ1 Sβ1 ←
−Sβ0 Sβ
Sα
α −
To conclude about the discussion of the decreasingness of the peak, it is just
P +
necessary to notice that for any i ∈ 0, . . . m and any j ∈ 0, . . . n, both S −
→
Sαi
j
j
P +
i
and S −
→ Sβ hold, i.e. Sα , Sβ ∈ {α, β}.
2.2 One of the rules is coinductive. By hypothesis we have
· ←≡ {r1 } ·
{r1 ,r2 }
· ←≡ {r1 } ·
{r2 }
Rβ
{r2 }
Sβ
{r1 ,r2 }
<
·
{r1 ,r2 }
· →≡ {r2 } ·
<
{r1 }
Rα
or equivalently by monotony of CHR:
<
·
{r1 ,r2 }
· →≡ {r2 } ·
<
{r1 }
Sα
Theorem 10 strictly subsumes all the criteria for proving confluence of CHR
programs we are aware of, namely the local confluence (Abdennadher et al. 1999)
and the strong confluence (Haemmerlé and Fages 2007) criteria.
Corollary 11 (Local confluence)
P
P
A terminating program P is confluent if its critical peaks are joinable by −
·
−.
Corollary 12 (Strong confluence)
P ≡
P ≡
A program P is confluent if its critical peaks are joinable by −
→
·←
− .
The following examples show that the rule-decreasingness criterion is more powerful than both local and strong confluence criteria.
12
Rémy Haemmerlé
hx ≤ x, x ≤ yi
trans.
anti.
hx ≤ xi
hx ≤ y, y ≤ z, z ≤ yi
hx ≤ x, x ≤ y, x ≤ xi
hx ≤ y, y = zi
ref lex.
anti.
trans.
anti.
hx ≤ y, y ≤ z, z ≤ y, x ≤ zi
dupl.
hx ≤ y, y ≤ xi
anti.
hx ≤ y, x ≤ z, y = zi
Fig. 5. Some rule-decreasing critical peaks for P1
Example 13
Consider the solver P1 for partial order given in Example 1. Since P1 is trivially
non-terminating one cannot apply local confluence criterion. Strong confluence does
not apply either, because of some non-strongly joinable critical peaks. For instance,
considere the peak given at Example 8:
P
P
1
1
hx = yi ←−−
−→
anti. hx ≤ y, y ≤ xi −
trans. hx ≤ y, y ≤ x, x ≤ xi
It can be seen that hx = yi may not be reduced, and that the right-hand side cannot
be rewritten into the left-hand side in less than two steps (e.g. by using reflexivity
and antisymmetry rules).
Nonetheless, confluence of P1 can be deduced using the full generality of Theorem 10. For this purpose, assume that all rules except transitivity are inductive
and take any admissible preorder. Clearly the inductive part of P1 is terminating. Indeed the application of any one of the three first rules strictly reduces the
number of atoms in a state. Then by a systematic analysis of all critical peaks of
P1 , we prove that each peak can be closed while respecting the hypothesis of ruledecreasingness. In fact all critical peaks can be closed without using transitivity.
Some rule-decreasing diagrams involving the transitivity rule are given as examples
in Figure 5.
Example 14
Consider the program P2 implementing the dining philosophers problem, as given
in Example 2. The confluence of P2 cannot be inferred by either local or strong
confluence. On the one hand, P2 is obviously non-terminating, and hence prevents
the application of the local confluence criterion. On the other hand, P2 has critical
P2
P2
peaks which are not in (−− · −−). Consider as an example the peak given in
Figure 6. It is critical for the rule eating with itself, but it is not joinable by
P2
≡ P2 ≡
(−−→
· ←−− ). However, the figure shows that it is joinable by
P
P
P
P
P
P
2 ≡
2 ≡
2 ≡
2 ≡
2 ≡
2 ≡
−−→
−→
−→
−−
−−
−−
thk · −
eat · −
thk · ←
thk · ←
eat · ←
thk
i.e. the peak is decreasing. In fact, all the critical peaks of P2 involve only the rule
eat and may be closed in a similar manner. Thus, by assuming that the eat rule is
coinductive and strictly greater than thk, we can infer, using Theorem 10, that P2
is confluent.
Diagrammatic confluence for Constraint Handling Rules
13
hf(x), f(y), f(z), t(x, y, i), t(y, z, j)i
eat
hf(z), e(x, y, i + 1), t(y, z, j)i
eat
hf(x), t(x, y, i), e(y, z, j + 1)i
thk
thk
hf(x), f(y), f(z), t(x, y, i + 1), t(y, z, j)i
hf(x), f(y), f(z), t(x, y, i), t(y, z, j + 1)i
eat
eat
hf(x), t(x, y, i + 1), e(y, z, j + 1)i
thk
hf(z), e(x, y, i + 1), t(y, z, j + 1)i
thk
hf(x), f(y), f(z), t(x, y, i + 1), t(y, z, j + 1)i
Fig. 6. A rule-decreasing critical peak of P2
4.4 On program partitioning
The rule-decreasingness criterion is based on the division of the program into a
terminating part and a possibly non-terminating one. Since a program can be partitioned in multiple ways, it may be the case that the rule-decreasingness of a
program depends on the splitting used (see Example 16). From a purely theoretical
point of view, this is not a particular drawback, since the property we aim at proving (i.e. the confluence of program) is undecidable. From a more pragmatical point
of view, it appears that the classic examples of CHR programs can be proved to
be rule-decreasing without any assumption of termination. In particular, we were
unable to find a counterexample of a confluent but non-strongly rule-decreasing
program in Frühwirth’s book (2009).
Example 15
Consider the CHR solver for partial order given in Example 1. Assuming that any
rule is coinductive, P can be shown strongly rule-decreasing with respect the order
< satisfying:
transitivity duplicate antisymmetry reflexivity
As illustrated by Figure 5, critical peaks involving transitivity rules may be closed
using only rules that are strictly smaller. Similarly, one can verify that any critical
peak between a given rule α and a smaller (or equal) one can be closed using only
rules strictly smaller than α (i.e. all the peaks are trivialy decreasing).
The choice of a good partition may simplify proofs of rule-decreasingness: by
maximizing the inductive part of a program, the number of peaks which must be
proved decreasing (i.e. the coinductive critical peaks) is reduced. Indeed, while the
joinability of a peak with respect to the inductive part of program – which must
14
Rémy Haemmerlé
be terminating – is a decidable problem and can be efficiently automatized,3 the
rule-decreasingness of a peak with respect to a possibly non-terminating program
is likely to be undecidable.4 Consequently, a good partition will limit the use of
heuristics or human interactions necessary to infer a rule-decreasing diagram for
each coinductive critical peak.
Since termination is also an undecidable property, we cannot expect to fully
automatize the search for the optimal partition, and we must content ourselves
with heuristic procedures. Despite the fact that the formal development of such
procedures is beyond the scope of this paper, our practical experience suggests that
a trivial partitioning may be interesting. This partition consists of considering as
inductive only those rules that strictly reduce the number of atoms in a state. Even
if this choice is not necessarily optimal and may even produce bad partitions, it
does seem to produce relevant partitions for typical CHR solvers, as illustrated by
Example 13.
We now give two counterexamples. The first shows that rule-decreasingness can
be dependent on particular splittings, while the second presents a confluent program
which is not rule-decreasing.
Example 16
Consider the following CHR rules:
duplicate @ p(x)\p(x) ⇐⇒ >
s− @ p(s(x)) ⇐⇒ p(x)
s+ @ p(x) ⇐⇒ p(s(x))
−
+
We denote by P16
the program built from the duplicate and s− rules, and by P16
+
the program built from the duplicate and s rules.
−
P16
is clearly terminating: the duplicate rule strictly reduces the number of atoms
in a state, while s− leaves the number of atoms unchanged, but strictly reduces the
−
size of the argument of one of them. We can also verify that P16
has a single critical
peak. Figure 7 shows the only way this peak may be closed. Thus, by assuming that
all rules are inductive, we can infer that the program is rule-decreasing. However if
−
s− is assumed to be coinductive, we can verify that the sole critical peak of P16
is
decreasing with respect to no admissible order.
−
+
As in the case of P16
, P16
yields only one critical peak which is decreasing with
respect to no admissible order (see Figure 8). However, this time s+ is not terminat+
ing, and so cannot been assumed inductive. Consequently P16
cannot be inferred
to be confluent using Theorem 10.
5 Modularity of CHR confluence
In this section, we are concerned with proving the confluence of union of confluent programs in a modular way (in particular of those programs proved confluent using the rule-decreasing criterion). In practice, we improve on a result of
3
4
See the works about CHR local confluence (Abdennadher et al. 1999; Abdennadher 1997).
Decreasingness of a peak for a given order seems a more difficult problem than joinability
without termination assumption—which is itself undecidable.
Diagrammatic confluence for Constraint Handling Rules
hp(s(x)), p(s(x))i
p
s−
hp(s(x))i
hp(x)i
hp(x), p(x)i
p
hp(s(x)), p(x)i
s−
s−
p
hp(x), p(x)i
−
Fig. 7. Critical peak of P16
15
s+
hp(x)i
hp(s(x)), p(x)i
s+
hp(s(x))i
s+
p
hp(s(x)), p(s(x))i
−
Fig. 8. Critical peak of P16
Frühwirth (2009) which states that a terminating union of confluent programs which
do not overlap (i.e. which do not have a critical peak) is confluent. In particular,
we allow some overlapping and we drop the termination hypotheses.
Theorem 17 (Modularity of confluence)
Let P and Q be two confluent CHR programs. If any critical peak between P and
Q
P ≡
Q is joinable by −
·←
− , then PQ is confluent.
Before formally proving the theorem, it is worth noting that, despite the fact that
modularity of confluence and the rule-decreasing theorem have similar flavors, both
results have different scopes. Indeed, on the one hand modularity of confluence does
not assume anything about the way in which P and Q are confluent. For instance, if
P and Q are two rule-decreasing programs, Theorem 17 does not require the union
of the inductive parts of P and Q to be terminating, while Theorem 10 does. This
is important since, termination is not a modular property: even if two terminating
programs do not share any user-defined atoms, one cannot be sure that their union
is terminating. (See Section 5.4 of Frühwirth’s book (2009) for more details.) On
the other hand, the rule-decreasing criterion allows the critical peaks to be closed
in a more complex way than Theorem 17 permits.
The proof of the theorem rests on the following lemma, which states that under
Q
P
the hypotheses of Theorem 17, −
→
“strongly commutes” with −
→.
Lemma 18
Q
Q
Q
P ≡
P
P ≡
If critical peaks between P and Q are in −
· ←
− , then (←
−· −
) ⊆ (−
· ←
− ).
Proof
Q
We prove by induction on the length of the derivation Sc −
S 0 that for any peak
Q
Q
P
P ≡
0
0
S ←
− Sc −
S , the property S −
· ←
− S holds. The base case Sc ≡ S 0 is
Q
Q
P
immediate. For the inductive case S ←
− Sc −
S 00 −
→ S 0 , we know by the induction
Q
P ≡
hypothesis that there exists a state R, such that S −
R ←
− S 00 . From here,
Q
P ≡
it is sufficient to prove that R −
· ←
− S 0 and to use the definition of relation
Q
Q
P
composition in order to conclude. We assume that S 00 −
→ R, otherwise R −
·←
−
≡ 0
S holds trivially. We distinguish two cases: either the rules involved in the local
Q
P
peak R ←
− S 00 −
→ S 0 apply to different parts of S 00 , or else their applications
16
Rémy Haemmerlé
Q
P
overlap. In the first case, we use CHR monotonicity to infer R −
→·←
− S 00 . In the
Q
P
000
00
−·−
→ S , a state R0 , and a set
second case, there must exist a critical peak R ←
00
0
000
0
of variables x̄, such that R ⊕x̄ R ≡ R, S ⊕x̄ R ≡ S 0 . Then by the hypotheses
Q
P ≡
and CHR monotonicity, we obtain the results that R −
·←
− S0.
Proof of Theorem 17
Q
P
Let →1 = −
, →2 = −
. On one hand, by the confluence of P and Q, we have
(←1 · →1 ) ⊆ (→1 · ←1 ) and (←2 · →2 ) ⊆ (→2 · ←2 ). (Note that 1 = →1 and
2 = →2 .) On the other hand, by combining Lemma 18 and case (i) of Lemma 4,
we infer (←1 · →2 ) ⊆ (→2 · ←1 ). By a trivial application of Theorem 3, we find that
PQ
PQ
→{1,2} is confluent. We conclude by noting −−→ ⊆ →{1,2} ⊆ −−, and apply case
PQ
PQ
(ii) of Lemma 4. (It is worth noting that →{1,2} equals neither −−→ nor −−.)
6 Conclusion
By employing the decreasing diagrams technique in CHR, we have established a
new criterion for CHR confluence that generalizes local and strong confluence criteria. The crux of this novel criterion rests on the distinction between the terminating part (the so-called inductive part) and non-terminating part (the so-called
coinductive part) of a program, together with the labeling of transitions by rules.
Importantly, we demonstrate that in the particular case of the proposed application
of the decreasing diagrams, the check on decreasingness can be restricted to the sole
critical pairs, hence making it possible to automatize the process. We also improve
on a result about the so-called modularity of confluence, which allows a modular
combination of rule-decreasing programs, without loss of confluence.
It is worth saying that all the diagrammatic proofs sketched in the paper have
been systematically verified by a prototype of a diagrammatic confluence checker.
In practice, this checker automatically generates all the critical pairs of a program
provided with an admissible order, then using user-defined tactics (finit sets of
reductions) tries to join these while respecting rule-decreasingness.
Current work involves investigating the development of heuristics to automatically infer rule-decreasingness without human interaction. We also plan to develop
a new completion procedure based on the criterion presented here. Because duplicate removal is an important programming idiom of CHR, the development of
new confluence-proof techniques capable of dealing with confluent but non-ruledecreasing programs, like those given in Example 16, is also worth investigating.
Diagrammatic confluence for Constraint Handling Rules
17
References
Abdennadher, S. 1997. Operational semantics and confluence of constraint propagation
rules. In Proceedings of the International Conference on Principles and Practice of
Constraint Programming (CP). LNCS, vol. 1330. Springer, Berlin, Germany, 252–266.
Abdennadher, S. and Frühwirth, T. 1999. Operational equivalence of CHR programs
and constraints. In Proceedings of the International Conference on Principles and Practice of Constraint Programming (CP). LNCS, vol. 1713. Springer, Berlin, Germany,
43–57.
Abdennadher, S., Frühwirth, T., and Meuss, H. 1996. On confluence of Constraint
Handling Rules. In Proceedings of the International Conference on Principles and Practice of Constraint Programming (CP). LNCS, vol. 1118. Springer, Berlin, Germany,
1–15.
Abdennadher, S., Frühwirth, T. W., and Meuss, H. 1999. Confluence and semantics
of constraint simplification rules. Constraints 4, 2, 133–165.
Duck, G. J., Stuckey, P. J., Garcı́a de la Banda, M., and Holzbaur, C. 2005.
The refined operational semantics of Constraint Handling Rules. In Proceedings of the
International Conference on Logic Programming (ICLP). LNCS, vol. 3668. Springer,
Berlin, Germany, 90–104.
Duck, G. J., Stuckey, P. J., and Sulzmann, M. 2007. Observable confluence for
constraint handling rules. In Proceedings of the International Conference on Logic Programming (ICLP). LNCS, vol. 4670. Springer, Berlin, Germany, 224–239.
Frühwirth, T. 1998. Theory and practice of Constraint Handling Rules. J. Logic Programming, Special Issue on Constraint Logic Programming 37, 1–3, 95–138.
Frühwirth, T. 2005. Parallelizing union-find in Constraint Handling Rules using confluence. In Proceedings of the International Conference on Logic Programming (ICLP).
LNCS, vol. 3668. Springer, Berlin, Germany, 113–127.
Frühwirth, T. 2009. Constraint Handling Rules. Cambrige University Press, Cambrige,
UK.
Guy, R. 2004. Unsolved Problems in Number Theory. Problem Books in Mathematics.
Springer, Berlin, Germany.
Haemmerlé, R. 2011a. (Co)-Inductive semantics for Constraint Handling Rules. Theory and Practice of Logic Programming, 27th Int’l. Conference on Logic Programming
(ICLP’11) Special Issue 11, 4–5, 593–609.
Haemmerlé, R. 2011b. Observational equivalences for linear logic concurrent constraint
languages. Theory and Practice of Logic Programming, 27th Int’l. Conference on Logic
Programming (ICLP’11) Special Issue 11, 4–5, 469–485.
Haemmerlé, R. and Fages, F. 2007. Abstract critical pairs and confluence of arbitrary
binary relations. In Proceedings of the International Conference on Rewriting Techniques
and Applications (RTA). Number 4533 in LNCS. Springer, Berlin, Germany, 214–228.
Haemmerlé, R., López, P., and Hermenegildo, M. 2011. CLP projection for Constraint Handling Rules. In International ACM SIGPLAN Conference on Principles
and Practice of Declarative Programming (PPDP). ACM Press, New York, NY, USA,
137–148.
Hirokawa, N. and Middeldorp, A. 2010. Decreasing diagrams and relative termination.
In IJCAR. LNCS, vol. 6173. Springer, Berlin, Germany, 487–501.
Huet, G. 1980. Confluent reductions: Abstract properties and applications to term rewriting systems: Abstract properties and applications to term rewriting systems. Journal
of the ACM 27, 4, 797–821.
Jouannaud, J.-P. and van Oostrom, V. 2009. Diagrammatic confluence and completion. In Proceedings of 36th Internatilonal Collogquium on Automata, Languages and
Programming: ICALP 2009. LNCS, vol. 5556. Springer, Berlin, Germany, 212–222.
18
Rémy Haemmerlé
Meister, M. 2006. Fine-grained parallel implementation of the preflow-push algorithm
in CHR. In Workshop on Logic Programming (WLP). INFSYS Research report 184306-02. T.U.Wien, Vienna, Austria, 172–181.
Milner, R. 1999. Communicating and mobile systems - the Pi-calculus. Cambrige University Press, Cambrige, UK.
Newman, M. H. A. 1942. On theories with a combinatorial definition of ”equivalence”.
Annals of mathematics 43, 2.
Raiser, F., Betz, H., and Frühwirth, T. 2009. Equivalence of CHR states revisited.
In Proceedings of the International Workshop on Constraint Handling Rules (CHR).
Report CW 555. Kath. Univ. Leuven, Leuven, Belgium, 34–48.
Raiser, F. and Tacchella, P. 2007. On confluence of non-terminating CHR programs.
In Proceedings of the International Workshop on Constraint Handling Rules (CHR).
63–76.
Sneyers, J., Schrijvers, T., and Demoen, B. 2009. The computational power and
complexity of Constraint Handling Rules. ACM Trans. Program. Lang. Syst. 31, 2.
Terese. 2003. Term Rewriting Systems. Cambrige University Press, Cambrige, UK.
van Oostrom, V. 1994. Confluence by decreasing diagrams. Theor. Comput. Sci. 126, 2,
259–280.
van Oostrom, V. 2008. Confluence by decreasing diagrams converted. In Proceedings of
the International Conference on Rewriting Techniques and Applications (RTA). LNCS.
Springer, Berlin, Germany, 306–320.
| 6cs.PL
|
1
Design Exploration of Hybrid CMOS-OxRAM
Deep Generative Architectures
arXiv:1801.02003v1 [cs.ET] 6 Jan 2018
Vivek Parmar, Member, IEEE, and Manan Suri, Member, IEEE
Indian Institute of Technology-Delhi, Hauz Khas, New Delhi-110016, email: [email protected]
Abstract—Deep Learning and its applications have gained
tremendous interest recently in both academia and industry.
Restricted Boltzmann Machines (RBMs) offer a key methodology
to implement deep learning paradigms. This paper presents
a novel approach for realizing hybrid CMOS-OxRAM based
deep generative models (DGM). In our proposed hybrid DGM
architectures, HfOx based (filamentary-type switching) OxRAM
devices are extensively used for realizing multiple computational
and non-computational functions such as: (i) Synapses (weights),
(ii) internal neuron-state storage, (iii) stochastic neuron activation
and (iv) programmable signal normalization. To validate the
proposed scheme we have simulated two different architectures:
(i) Deep Belief Network (DBN) and (ii) Stacked Denoising Autoencoder for classification and reconstruction of hand-written digits
from a reduced MNIST dataset of 6000 images. Contrastivedivergence (CD) specially optimized for OxRAM devices was
used to drive the synaptic weight update mechanism of each
layer in the network. Overall learning rule was based on greedylayer wise learning with no back propagation which allows the
network to be trained to a good pre-training stage. Performance
of the simulated hybrid CMOS-RRAM DGM model matches
closely with software based model for a 2-layers deep network.
Top-3 test accuracy achieved by the DBN was 95.5%. MSE of the
SDA network was 0.003, lower than software based approach.
Endurance analysis of the simulated architectures show that for
200 epochs of training (single RBM layer), maximum switching
events/per OxRAM device was ∼ 7000 cycles.
Index Terms—Deep Learning, Stacked Denoising Autoencoder
(SDA), Deep Belief Network (DBN), RRAM, Deep Generative
Models, Restricted Boltzmann Machine (RBM)
I. I NTRODUCTION
HE neurons in our brain have a capacity to process a large
amount of high dimensional data from various sensory
inputs while still focusing on the most relevant components
for decision making [1] [2]. This implies that the biological
neural networks have a capacity to perform dimensionality
reduction to facilitate decision making. In the field of machine
learning, artificial neural networks also require a similar capability because of the availability of massive amounts of high
dimensional data being generated everyday through various
sources for digital information. Thus it becomes imperative
to derive an efficient method for dimensionality reduction to
facilitate tasks like classification, feature learning, storage, etc.
Deep generative networks such as Autoencoders have been
shown to perform better than many commonly used statistical
techniques such as PCA (principal component analysis), ICA
(Independent Component Analysis) for encoding and decoding
of high dimensional data [3]. These networks are traditionally
trained using gradient descent based on back-propagation.
However it is observed that for deep networks, gradient
T
descent doesn’t converge and gets stuck in a local minima in
case of purely randomized initialization [4]. A solution to this
problem is, weight initialization by utilizing a generative layerby-layer training procedure based on Contrastive Divergence
(CD) algorithm [5].
To maximize the performance of this algorithm, a dedicated
hardware implementation is required to accelerate computation
speed. Traditionally CMOS based designs have been used for
this by utilizing commonly available accelerator like GPUs [6],
FPGAs [7], ASICs [8] [9], etc. Recently with the introduction
of the emerging non-volatile memory devices such as PCM,
CBRAM, OxRAM, MRAM, etc, there is further optimization
possible in design of a dedicated hardware accelerators given
the fact that they allow replacement of certain large CMOS
blocks while simultaneously emulating storage and compute
functionalities [10] [11] [12] [13] [14] [15], [16] [17] [18].
Recent works that present designs of Contrastive Divergence
based learning using resistive memory devices are [19], [20].
In [19] the authors propose the use of a two-memristor model
as a synapse to store one synaptic weight. In [20] the authors
have experimentally demonstrated a 45-synapse RBM realized
with 90 resistive phase change memory (PCM) elements
trained with a bio-inspired variant of the contrastive divergence
algorithm, implementing Hebbian and anti-Hebbian weight
update. Both these designs justify the use of RRAM devices
as dense non-volatile synaptic arrays. Also both make use of
a spike based programming mechanism for gradually tuning
the weights. Negative weights have been implemented by
using two devices in place of a single device per synapse.
It is apparent that in order to implement more complex
learning rules with larger and deeper networks the hardware
complexity and area footprint increases considerably while
using this simplistic design strategy. As a result, there is a need
to increase further increase the functionality of the RRAM
devices in the design beyond simple synaptic weight storage.
In [21] we have described a design exploiting the intrinsic
device-to-device variability as a substitute for the randomly
distributed hidden layer weights in order to gain both area
and power savings. In [22], we have made use of another
property of the RRAM devices by exploiting the cycle-to-cycle
variability in device switching to create a stochastic neuron
as a basic building block for a hybrid CMOS-OxRAM based
Restricted Boltzmann Machines (RBM) circuit.
In this paper we build upon our previous work on hybrid
CMOS-OxRAM RBM with the following novel contributions:
• Design of deep generative models (DGM) that utilize the
hybrid CMOS-RRAM RBM as a building block.
2
Design of programmable output normalization block for
stacking multiple hybrid RBMs.
• Simulation and performance analysis of two types of
DGM architectures at 8-bit synaptic weight resolution: (i)
Deep Belief Networks (DBN) and (ii) Stacked Denoising
Autoencoders (SDA)
• Analysis of learning performance (accuracy, MSE) while
using only greedy layer-wise training (without backprop).
• Analysis of learning impact on RRAM device endurance.
In our hybrid CMOS-OxRAM DGM implementation the
OxRAM devices have been exploited for four different storage
and compute functions: (i) Synaptic weight matrix, (ii) neuron
internal state storage, (iii) stochastic neuron firing and (iv)
programmable gain control block. Section II discusses the
basics of OxRAM and deep generative networks. Section III
describes the implementation details of our proposed hybrid
CMOS-OxRAM DGM architectures. Section IV discusses
simulation results and Section V gives the conclusions.
•
II. BASICS OF OX RAM AND DGM A RCHITECTURES
A. OxRAM Working
Fig. 2. Cycle-to-Cycle ON/OFF-state resistance distribution for HfOx device
presented in [22].
by using an optional selector device (i.e. 1R-1T/1D configuration). OxRAM devices are known to demonstrate cycle-tocycle (C2C) (shown in Fig. 2), and device-to-device (D2D)
variability [25] [26], [27]. In our proposed architecture, we
exploit OxRAM (a) C2C switching variability for realization
of stochastic neuron circuit, (b) binary resistive switching for
realization of synaptic weight arrays/neuron internal state storage and (c) SET-state resistance modulation for normalization
block.
B. Restricted Boltzmann Machines (RBM)
Fig. 1. Basic IV characteristics for HfOx OxRAM device with switching
principle indicated. Experimental data corresponding to device presented in
[22].
OxRAM devices are two-terminal MIM-type structures
(metal-insulator-metal) sandwiching an active metal-oxide
based insulator layer, between metallic electrodes (see Fig.
1). The active layer exhibits reversible non-volatile switching behavior on application of appropriate programming
current/voltage across the device terminals. In the case of
filamentary- OxRAM devices, formation of a conductive filament in the active layer, leads the device to a low-resistance
(LRS/On) SET-state, while dissolution of the filament puts
the device in a high-resistance (HRS/Off) RESET-state. The
conductive filament is composed of oxygen vacancies and
defects [23]. SET-state resistance (LRS) level can be defined
by controlling the dimensions of the conductive filament [24]
[23], which depends on the amount of current flowing through
the active layer. Current flowing through the active layer is
controlled either by externally imposed current compliance or
Unsupervised learning based on generative models has
gained importance with use of deep neural networks. Besides
being useful for pre-training a supervised predictor, unsupervised learning in deep architectures can be of interest to learn
a distribution and generate samples from it [28]. RBMs in particular, are widely used as building blocks for deep generative
models such as DBN and SDA. Both these models are made
by stacking RBM blocks on top of each other. Training of such
models using traditional back-propagation based approaches is
a computationally intensive problem. Hinton et.al. [5] showed
that such models can be trained very fast through greedy layerwise training making the task of training deep networks based
on stacking of RBMs more feasible. Each RBM block consists
of two layers of fully connected stochastic sigmoid neurons
as shown in Fig. 3. The input or the first layer of the RBM
is called the visible layer and the second (feature detector)
layer is called the hidden layer. Each RBM is trained using
Fig. 3. Graphical representation of RBM hidden/visible nodes.
3
CD algorithm as described in [29]. The output layer of the
bottom RBM acts as the visible layer for the next RBM.
C. Stacked Denoising Autoencoder (SDA)
An autoencoder network is a deep learning framework
mostly used for denoising corrupted data [30], dimensionality
reduction [3] and weight initialization applications. In recent
years random weight initialization techniques have been preferred over use of generative training networks [31], however
DGMs continue to be the ideal candidate for dimensionality
reduction and denoising applications. Autoencoder network is
basically realized using two networks:
1) An ’encoder’ network which has layers of RBMs stacked
on the top of one another.
2) A mirrored ’decoder’ network with same weights as that
of the encoder layer for data reconstruction.
The stack of RBMs in autoencoder are trained layer-wise
one after the other. An ’unrolled’ autoencoder network with
the encoder and decoder is shown in Fig. 4.
D. Deep Belief Network (DBN)
DBNs are probabilistic generative models that are composed
of multiple layers of stochastic, latent variables [32]. The latent
variables typically have binary values and are often called
hidden units or feature detectors. The top two layers have
undirected, symmetric connections between them and form
an associative memory. The lower layers receive top-down,
directed connections from the layer above. The states of the
units in the lowest layer represent a data vector. A typical
DBN is shown in Fig. 5 which uses a single RBM as the
first two layers followed by a sigmoid belief network (logistic
regression layer) for the final classification output.
The two most significant DBN properties are:
1) There is an efficient, layer-by-layer procedure for learning the top-down, generative weights that determine how
the variables in one layer depend on the variables in the
layer above.
2) After learning, the values of the latent variables in every
layer can be inferred by a single, bottom-up pass that
starts with an observed data vector in the bottom layer
and uses the generative weights in the reverse direction.
DBNs have been used for generating and recognizing images
[5], [33], [4], video sequences [34], and motion-capture data
[35]. With low number of units in the highest layer, DBNs
perform non-linear dimensionality reduction and can learn
short binary codes, allowing very fast retrieval of documents
or images [36].
III. I MPLEMENTATION OF P ROPOSED A RCHITECTURES
Basic building block of both SDA and DBN is the RBM.
In our simulated architectures, within a single RBM block
OxRAM devices are used for multiple functionalities. The
basic RBM block (shown in Fig. 6(a)) is replicated, with
the hidden layer memory states of the first RBM acting as
visible layer memory for the next RBM block and so on
(Fig. 6(b)). All RBM blocks have a common weight update
module described in Section III-C. Post training, the learned
synaptic weights along with the sigmoid block can be used for
reconstructing the test data. Architecture sub-blocks consist of:
A. Synaptic Network
Synaptic network of each RBM block was simulated using a
1T-1R HfOx OxRAM matrix. Each synaptic weight is digitally
encoded in a group of binary switching OxRAM devices,
where the number of devices used per synapse depends on
the required weight resolution. For all architectures simulated
in this work we have used 8-bit resolution (i.e. 8 OxRAM
devices/per synapse).
B. Stochastic Neuron Block
Fig 6(d), shows the stochastic sigmoid neuron block. Each
neuron (hidden or visible) has a sigmoid response, which was
Fig. 4. (a) Basic RBM blocks stacked to form a deep autoencoder. (b)
Denoising noisy image using autoencoder.
Fig. 5. DBN architecture comprising of stacked RBMs
4
Fig. 6. (a) Individual RBM training layer architecture. RBM training block symbols, ’H’, ’V’ and ’S’ represent hidden layer memory, visible layer memory,
and synaptic network respectively. (b) Cascaded RBM blocks for realizing the proposed deep autoencoder with shared weight update module (c) Fully digital
CD based weight update module. (d) Block level design of single stochastic sigmoid neuron
implemented using a low-power 6-T sigmoid circuit ( [37]).
Gain of the sigmoid circuit can be tuned by optimizing the
scaling of the six transistors. Voltage output of the sigmoid
circuit is compared with the voltage drop across the OxRAM
device, with the help of a comparator. The HfOx based
device is repeatedly cycled ON/OFF. C2C intrinsic RON and
ROF F variability of the OxRAM device leads to a variable
reference voltage for the comparator. This helps to translate
the deterministic sigmoid output to a neuron output, which
is effectively stochastic in nature. At any given moment, a
specific neuron’s output determines it’s internal state, which
needs to be stored for RBM driven learning. Neuron internal
state is stored using individual OxRAM devices placed after
the comparator. Single OxRAM/per neuron is sufficient for
state storage, since RBM requires each neuron to only have a
binary activation state.
C. CD Weight Update Block
The weight update module is a purely digital circuit that
reads the synaptic weights and internal neuron states. It
updates the synaptic weights during learning based on the
CD RBM algorithm ( [29]). The block consists of an array
of weight update circuits, one of which is shown in Fig 6(c).
Synaptic weight is updated by 4Wij 1, based on the previous
(v, h) and current (v, h) internal neuron states of the mutually
connected neurons in the hidden and visible layers. CD is
is realized using two AND gates and a comparator (having
outputs -1, 0, +1). Input to the first AND gate is previous
internal neuron states, while the input to second AND gate
is the current internal neuron states. Based on the comparator
output, (learning rate) will either be added, subtracted, or
not applied to the current synaptic weight (Wij).
0
0
4 Wij = (vhT − v h T )
(1)
D. Output Normalization block
In order to chain the mixed-signal design of RBM we
need to ensure the signal output at each layer is having
an enhancement in the dynamic range so that the signal
doesn’t deteriorate as the network depth increases. For this
purpose we proposed a hybrid CMOS-OxRAM programmable
normalization circuit (see Fig. 7) whose gain and bias can be
tuned based on OxRAM resistance programming.
The circuit schematic of the programmable normalization
block is shown in Fig. 7(a)). In order to check variation in
gain, we have considered programming the OxRAM in three
different SET states (∼ 3.2 kΩ, 6.6 kΩ, and 22.6 kΩ).
The differential amplifier consisting of a DC gain control
circuit and a biasing circuit is used to implement the normalization function. A two stage amplifier consisting of transistors
N3, N4, N5, N6, P3, P4, P5, P6, and P7 is used. DC gain of
the circuit is controlled using a constant gm circuit whose
output is fed into N3. The constant gm circuit consists of
transistors N1, N2, P1, P2 and one OxRAM. Based on the
OxRAM resistance, gm of the circuit can be changed thereby
changing the output potential. This affects Vgs of N3 thereby
controlling the gain of the circuit.
To validate the design, we performed simulation of the circuit using an OxRAM device compact model ( [26]) and 90 nm
CMOS design kit. The simulated variation in the gain of the
circuit based on the resistance state of the OxRAM is shown
in Fig. 7(b). Gain control through OxRAM programming was
found to be more prominent at higher operating frequencies.
Bias control is implemented by a potential divider circuit
(Rf and the OxRAM). The potential divider circuit determines
the potential across Vg of P6. Input V2 is swept from 0 V to 1
V. If the potential across P6 increases for a fixed V2 the output
switching voltage also increases thereby controlling the bias
5
Fig. 7. Programmable normalization circuit: (a) Circuit Schematic, (b) Gain variation w.r.to variation in OxRAM resistance state
of the output.
IV. D EEP L EARNING S IMULATIONS AND R ESULTS
Simulations of the proposed architectures (DBN, SDA) were
performed in MATLAB. Both generative networks with CD
algorithm and behavioral model of all blocks described in section III were simulated. Stochastic sigmoid neuron activation
and normalization circuits were simulated in Cadence Virtuoso using 90 nm CMOS design kit and Verilog-A OxRAM
compact model [26].
A. Stacked Denoising Autoencoder performance analysis
We trained two autoencoder networks each having the same
number of neurons in the final encoding layer, but varying
levels of depth, and compared their denoising performance
(see Fig. 8). In each network a single synaptic weight was
realized using 8 OxRAM devices (8-bit resolution). All neurons have a logistic activation except for the last ten units in
the classification layer, which are linear. The networks were
trained on a reduced MNIST dataset of 5000 images and
tested for denoising 1000 new salt-and-pepper noise corrupted
images (see Fig. 8).
TABLE I
P ROPOSED AUTOENCODER PERFORMANCE FOR REDUCED MNIST
Network
784x100x784
784x100x40x100x784
Implementation
MSE
Software
0.010
Hybrid OxRAM SDA
0.003
Software
0.049
Hybrid OxRAM SDA
1.095
Table I presents the learning performance of the proposed
SDA-1 and SDA-2. Increasing depth in the network was
not useful with the current learning algorithm and tuning
parameters.
Fig. 8. (a) 3-layer deep SDA-1, (b) 5-layer deep SDA-2. Denoising results
of 100 corrupted MNIST images for: (c) SDA1 and (d) SDA2.
B. Deep Belief Network performance analysis
We simulated two deep belief network architectures shown
in Fig. 9. (4 and 5 layer variants) Performance of the network
was measured by testing on 1000 samples from the reduced
MNIST dataset. The results for the same are shown in Table
II. We measured test accuracy using 3 parameters :
1) Top 1 accuracy : correct class corresponds to output
neuron with highest response.
2) Top 3 accuracy : correct class corresponds to the top 3
output neurons with highest response.
3) Top 5 accuracy : correct class corresponds to the top 5
output neurons with highest response.
From Table II, the performance of simulated Hybrid CMOSOxRAM DBN matches closely with software based accuracy
(2-3% lower) for a DBN formed with 2 RBMs. There is a
significant drop in test accuracy for the DBN with 3 RBMs.
This is acceptable as the goal of the greedy layer-wise training
is to pre-train the network to a good state before using
back-propagation to allow faster convergence. Thus lower
6
TABLE III
M AXIMUM OX RAM SWITCHING ACTIVITY FOR 5 LAYER SDA
( TRAINING )
Device placement
L1-784
L2-100
L3-40
W1
W2
Max Switching activity
596
3074
542
6808
5000
TABLE IV
M AXIMUM OX RAM SWITCHING ACTIVITY FOR 5 LAYER DBN
( TRAINING )
Fig. 9. Simulated 4 and 5 layer DBN architecture.
accuracy after layer-wise training for a deeper network is
acceptable as the weights would be further optimized using
back-propagation.
TABLE II
P ROPOSED DBN P ERFORMANCE FOR R EDUCED MNIST
Device placement
L1-784
L1-784
L2-160
L3-80
L4-40
W1
W2
W3
W4
Max Switching activity
596
428
2069
3026
420
6798
5000
2500
2500
Test accuracy
Network
784x100x40x10
784x160x80x40x10
Implementation
Top-1
Top-3
Top-5
Software
93.10%
98.70%
99.40%
Hybrid OxRAM DBN
78.70%
95.50%
98.80%
Software
93.70%
98.50%
99.40%
Hybrid OxRAM DBN
21.30%
61.40%
79.60%
C. Tuning VOxRAM amplifying gain
The sigmoid activation circuits in the network use a gain
factor in order to balance for the low current values obtained
as a result of the OxRAM device resistance values. If the
amplification is low it will lead to saturation and the network will not learn a proper reconstruction of the data. This
necessitates proper tuning of the amplifier gain for effective
learning. In our architecture, amplifier gain for VOxRAM is
an important hyper-parameter along with the standard ones
(momentum, decay rate, learning rate, etc.) and is different for
each consecutive pair of layers. A higher dimensional input to
a layer will require a lower amplifying gain for VOxRAM and
vice-versa.
D. Switching activity analysis for the Proposed architecture
Resistive switching of OxRAM devices is observed in
following sections of the architecture:
1) Synaptic matrix
2) Stochastic neuron activation
3) Internal neuron state storage.
RRAM devices suffer from limited cycling endurance (∼
0.1 million cycles) [38].For stochastic neuron activation, the
OxRAM device is repeatedly cycled to OFF state and the
voltage drop across the device is used to generate the stochastic
signal fed to one of the comparator inputs. Thus the neuron
activation block related switching activity depends on the
number of data samples as well as number of epochs. The
maximum switching per device for any layer can be estimated
by using (2):
Nevents = Nepochs ∗ Nsamples ∗ Nbatch
(2)
Another part of the architecture where the OxRAM device
may observe a significant number of switching events is the
synaptic matrix. Since we are interested in device endurance,
we consider the worst case, i.e. the of maximum number of
hits a particular OxRAM device will take during the entire
weight update procedure. For worst case analysis we make
the following assumptions• While bit encoding the synaptic weight (4 or 8 or 16),
there exists an OxRAM device that is switched every
single time.
Thus the maximum possible number of hits a device would
take during the synaptic weight update procedure can be
estimated using (3):
Nswitchevents = Nbatch ∗ Nepochs
(3)
Simulated switching activity for reduced MNIST training for
each neuron layer and synaptic matrix is shown in Table III and
Table IV corresponding to both SDA and DBN architectures
respectively. Key observations can be summarized as:
• Increasing depth of the network increases amount of
switching for hidden layers.
• Increasing depth of the network doesn’t have significant
impact on the the switching events in the synaptic matrix.
7
V. C ONCLUSION
In this paper we proposed a novel methodology to realize
DGM architectures using mixed-signal type hybrid CMOSRRAM design framework. We achieve deep generative models
by proposing a strategy to stack multiple RBM blocks. Overall
learning rule used in this study is based on greedy-layer wise
learning with no back propagation which allows the network
to be trained to a good pre-training stage. RRAM devices
are used extensively in the proposed architecture for multiple
computing and storage actions. Total RRAM requirement for
the largest simulated network was 139 kB for DBN and 169 kB
for SDA. Simulated architectures show that the performance
of the proposed DGM models matches closely with software
based models for 2 layers deep network. The top-3 test accuracy achieved by the DBN for reduced MNIST was ∼ 95.5%.
MSE of SDA network was 0.003. Endurance analysis shows
resonable maximum switching activity. Future work would
focus on realizing an optimal strategy to implement backpropagation with the proposed architecture to enable complete
training of the DGM on the hybrid DGM architecture.
ACKNOWLEDGEMENT
This research activity under the PI Prof. M. Suri is partially
supported by the Department of Science & Technology (DST),
Government of India and IIT-D FIRP Grant. Authors would
like to express gratitude to S. Chakraborty. The authors would
like to thank F. Alibart and D. Querlioz for the HfOx device
data.
A PPENDIX
Code for all simulations discussed in the paper is available
at https://gitlab.com/vivekp312/oxram-sda-sim.git. Any interested researchers can contact the authors for access to the code
repository.
R EFERENCES
[1] J. W. Pillow and E. P. Simoncelli, “Dimensionality reduction in neural
models: an information-theoretic generalization of spike-triggered average and covariance analysis,” Journal of vision, vol. 6, no. 4, pp. 9–9,
2006.
[2] J. P. Cunningham and M. Y. Byron, “Dimensionality reduction for largescale neural recordings,” Nature neuroscience, vol. 17, no. 11, pp. 1500–
1509, 2014.
[3] G. E. Hinton and R. S. Zemel, “Autoencoders, minimum description
length, and helmholtz free energy,” Advances in neural information
processing systems, pp. 3–3, 1994.
[4] Y. Bengio, P. Lamblin, D. Popovici, and H. Larochelle, “Greedy layerwise training of deep networks,” in Advances in neural information
processing systems, 2007, pp. 153–160.
[5] G. E. Hinton, S. Osindero, and Y. W. Teh, “A fast learning algorithm for
deep belief nets,” Neural computation, vol. 18, no. 7, pp. 1527–1554,
2006.
[6] R. Raina, A. Madhavan, and A. Y. Ng, “Large-scale deep unsupervised
learning using graphics processors,” in Proceedings of the 26th annual
international conference on machine learning. ACM, 2009, pp. 873–
880.
[7] S. K. Kim, L. C. McAfee, P. L. McMahon, and K. Olukotun, “A highly
scalable restricted boltzmann machine fpga implementation,” in Field
Programmable Logic and Applications, 2009. FPL 2009. International
Conference on. IEEE, 2009, pp. 367–372.
[8] P. Merolla, J. Arthur, F. Akopyan, N. Imam, R. Manohar, and D. S. .
Modha, “September),” A digital neurosynaptic core using embedded
crossbar memory with, vol. 45, pp. 1–4, 2011.
[9] E. Stromatias, D. Neil, M. Pfeiffer, F. Galluppi, S. B. Furber, and
S.-C. Liu, “Robustness of spiking deep belief networks to noise and
reduced bit precision of neuro-inspired hardware platforms,” Frontiers
in neuroscience, vol. 9, 2015.
[10] M. Suri, O. Bichler, D. Querlioz, O. Cueto, L. Perniola, V. Sousa,
D. Vuillaume, C. Gamrat, and B. DeSalvo, “Phase change memory as
synapse for ultra-dense neuromorphic systems: Application to complex
visual pattern extraction, electron devices meeting (iedm), 2011 ieee
international, vol., no,” pp., vol. 4, no. 4, pp. 5–7, December 2011.
[11] F. Alibart, E. Zamanidoost, and D. B. Strukov, “Pattern classification by
memristive crossbar circuits using ex situ and in situ training,” Nature
communications, vol. 4, p. 2072, 2013.
[12] J. J. Yang, D. B. Strukov, and D. R. Stewart, “Memristive devices for
computing,” Nature nanotechnology, vol. 8, no. 1, pp. 13–24, 2013.
[13] B. De Salvo, Silicon non-volatile memories: paths of innovation. John
Wiley & Sons, 2013.
[14] B. L. Jackson, B. Rajendran, G. S. Corrado, M. Breitwisch, G. W.
Burr, R. Cheek, K. Gopalakrishnan, S. Raoux, C. T. Rettner, A. Padilla
et al., “Nanoscale electronic synapses using phase change devices,” ACM
Journal on Emerging Technologies in Computing Systems (JETC), vol. 9,
no. 2, p. 12, 2013.
[15] A. F. Vincent, J. Larroque, W. S. Zhao, N. B. Romdhane, O. Bichler,
C. Gamrat, J. o. Klein, S. Galdin-Retailleau, and D. Querlioz, “Spintransfer torque magnetic memory as a stochastic memristive synapse,”
In Circuits and Systems (ISCAS), vol. 2014, pp. 1074–1077, 2014.
[16] H.-S. P. Wong and S. Salahuddin, “Memory leads the way to better
computing,” Nature nanotechnology, vol. 10, no. 3, pp. 191–194, 2015.
[17] G. W. Burr, R. M. Shelby, S. Sidler, C. Di Nolfo, J. Jang, I. Boybat,
R. S. Shenoy, P. Narayanan, K. Virwani, E. U. Giacometti et al.,
“Experimental demonstration and tolerancing of a large-scale neural
network (165 000 synapses) using phase-change memory as the synaptic
weight element,” IEEE Transactions on Electron Devices, vol. 62, no. 11,
pp. 3498–3507, 2015.
[18] V. Milo, G. Pedretti, R. Carboni, A. Calderoni, N. Ramaswamy, S. Ambrogio, and D. Ielmini, “Demonstration of hybrid cmos/rram neural
networks with spike time/rate-dependent plasticity,” in Electron Devices
Meeting (IEDM), 2016 IEEE International. IEEE, 2016, pp. 16–8.
[19] A. M. Sheri, A. Rafique, W. Pedrycz, and M. Jeon, “Contrastive divergence for memristor-based restricted boltzmann machine,” Engineering
Applications of Artificial Intelligence, vol. 37, pp. 336–342, 2015.
[20] S. B. Eryilmaz, E. Neftci, S. Joshi, S. Kim, M. BrightSky, H. L. Lung,
C. Lam, G. Cauwenberghs, and H. S. P. Wong, “Training a probabilistic
graphical model with resistive switching electronic synapses,” IEEE
Transactions on Electron Devices, vol. 63, no. 12, pp. 5004–5011, Dec
2016.
[21] M. Suri and V. Parmar, “Exploiting intrinsic variability of filamentary
resistive memory for extreme learning machine architectures,” IEEE
Transactions on Nanotechnology, vol. 14, no. 6, pp. 963–968, Nov 2015.
[22] M. Suri, V. Parmar, A. Kumar, D. Querlioz, and F. Alibart, “Neuromorphic hybrid rram-cmos rbm architecture,” in 2015 15th Non-Volatile
Memory Technology Symposium (NVMTS), Oct 2015, pp. 1–6.
[23] H.-S. P. Wong, H.-Y. Lee, S. Yu, Y.-S. Chen, Y. Wu, P.-S. Chen, B. Lee,
F. T. Chen, and M.-J. Tsai, “Metal–oxide rram,” vol. 100, no. 6. IEEE,
2012, pp. 1951–1970.
[24] M. Suri, D. Querlioz, O. Bichler, G. Palma, E. Vianello, D. Vuillaume,
C. Gamrat, and B. DeSalvo, “Bio-inspired stochastic computing using binary cbram synapses,” Electron Devices, IEEE Transactions on,
vol. 60, no. 7, pp. 2402–2409, 2013.
[25] C. Baeumer, R. Valenta, C. Schmitz, A. Locatelli, T. O. Mente, S. P.
Rogers, A. Sala, N. Raab, S. Nemsak, M. Shim et al., “Subfilamentary
networks cause cycle-to-cycle variability in memristive devices,” ACS
nano, vol. 11, no. 7, pp. 6921–6929, 2017.
[26] H. Li, Z. Jiang, P. Huang, Y. Wu, H.-Y. Chen, B. Gao, X. Liu, J. Kang,
and H.-S. Wong, “Variation-aware, reliability-emphasized design and
optimization of rram using spice model,” in Design, Automation & Test
in Europe Conference & Exhibition (DATE), 2015. IEEE, 2015, pp.
1425–1430.
[27] D. Ielmini, “Resistive switching memories based on metal oxides: mechanisms, reliability and scaling,” Semiconductor Science and Technology,
vol. 31, no. 6, p. 063002, 2016.
[28] Y. Bengio et al., “Learning deep architectures for ai,” Foundations and
trends R in Machine Learning, vol. 2, no. 1, pp. 1–127, 2009.
[29] G. Hinton, “A practical guide to training restricted boltzmann machines,”
Momentum, vol. 9, no. 1, p. 926, 2010.
[30] H. Vincent, P.and Larochelle, I. Lajoie, Y. Bengio, and P. Manzago,
“Stacked denoising autoencoders: Learning useful representations in a
8
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
deep network with a local denoising criterion,” Journal of Machine
Learning Research, no. 11, pp. 3371–3408, 2010.
X. Glorot and Y. Bengio, “Understanding the difficulty of training deep
feedforward neural networks,” In Proc. AISTATS, vol. 9, pp. 249–256,
2010.
G. E. Hinton, “Deep belief networks,” Scholarpedia, vol. 4, no. 5, p.
5947, 2009.
F. J. Huang, Y.-L. Boureau, Y. LeCun et al., “Unsupervised learning
of invariant feature hierarchies with applications to object recognition,”
in Computer Vision and Pattern Recognition, 2007. CVPR’07. IEEE
Conference on. IEEE, 2007, pp. 1–8.
I. Sutskever and G. Hinton, “Learning multilevel distributed representations for high-dimensional sequences,” in Artificial Intelligence and
Statistics, 2007, pp. 548–555.
G. W. Taylor, G. E. Hinton, and S. T. Roweis, “Modeling human
motion using binary latent variables,” in Advances in neural information
processing systems, 2007, pp. 1345–1352.
G. E. Hinton and R. R. Salakhutdinov, “Reducing the dimensionality of
data with neural networks,” science, vol. 313, no. 5786, pp. 504–507,
2006.
D. Pan and B. M. Wilamowski, “A vlsi implementation of mixed-signal
mode bipolar neuron circuitry,” In Neural Networks, Proceedings of the
International Joint Conference on, vol. 2, pp. 971–976, 2003.
S. Balatti, S. Ambrogio, Z.-Q. Wang, S. Sills, A. Calderoni, N. Ramaswamy, and D. Ielmini, “Pulsed cycling operation and endurance
failure of metal-oxide resistive (rram),” in Electron Devices Meeting
(IEDM), 2014 IEEE International. IEEE, 2014, pp. 14–3.
| 9cs.NE
|
Sympiler: Transforming Sparse Matrix Codes by Decoupling
Symbolic Analysis
arXiv:1705.06575v1 [cs.PL] 18 May 2017
Kazem Cheshmi
Adobe Research
Cambridge, MA, US
[email protected]
Michelle Mills Strout
Maryam Mehri Dehnavi
University of Arizona
Tucson, AZ, US
[email protected]
ABSTRACT
Sympiler is a domain-specific code generator that optimizes sparse
matrix computations by decoupling the symbolic analysis phase
from the numerical manipulation stage in sparse codes. The computation patterns in sparse numerical methods are guided by the
input sparsity structure and the sparse algorithm itself. In many
real-world simulations, the sparsity pattern changes little or not at
all. Sympiler takes advantage of these properties to symbolically
analyze sparse codes at compile-time and to apply inspector-guided
transformations that enable applying low-level transformations to
sparse codes. As a result, the Sympiler-generated code outperforms
highly-optimized matrix factorization codes from commonly-used
specialized libraries, obtaining average speedups over Eigen and
CHOLMOD of 3.8× and 1.5× respectively.
KEYWORDS
Matrix computations, sparse methods, loop transformations, domainspecific compilation
1
Shoaib Kamil
Rutgers University
Piscataway, NJ, US
[email protected]
INTRODUCTION
Sparse matrix computations are at the heart of many scientific applications and data analytics codes. The performance and efficient
memory usage of these codes depends heavily on their use of specialized sparse matrix data structures that only store the nonzero
entries. However, such compaction is done using index arrays that
result in indirect array accesses. Due to these indirect array accesses,
it is difficult to apply conventional compiler optimizations such as
tiling and vectorization even for static index array operations like
sparse matrix vector multiply. A static index array does not change
during the algorithm; for more complex operations with dynamic
index arrays such as matrix factorization and decomposition, the
nonzero structure is modified during the computation, making conventional compiler optimization approaches even more difficult to
apply.
The most common approach to accelerating sparse matrix computations is to identify a specialized library that provides a manuallytuned implementation of the specific sparse matrix routine. A large
number of sparse libraries are available (e.g., SuperLU [22], MUMPS
[2], CHOLMOD [11], KLU [20], UMFPACK [15]) for different numerical kernels, supported architectures, and specific kinds of matrices.
Rutgers University
Piscataway, NJ, US
[email protected]
While hand-written specialized libraries can provide high performance, they must be manually ported to new architectures and may
stagnate as architectural advances continue. Alternatively, compilers can be used to optimize code while providing architecture
portability. However, indirect accesses and the resulting complex
dependence structure run into compile-time loop transformation
framework limitations.
Compiler loop transformation frameworks such as those based
on the polyhedral model use algebraic representations of loop nests
to transform code and successfully generate highly-efficient dense
matrix kernels [5, 10, 41, 54, 65, 67]. However, such frameworks
are limited when dealing with non-affine loop bounds and/or array
subscripts, both of which arise in sparse codes. Recent work has
extended polyhedral methods to effectively operate on kernels with
static index arrays by building run-time inspectors that examine the
nonzero structure and executors that use this knowledge to transform code execution [63, 66, 68–70]. However, these techniques
are limited to transforming sparse kernels with static index arrays.
Sympiler addresses these limitations by performing symbolic analysis at compile-time to compute fill-in structure and to remove
dynamic index arrays from sparse matrix computations. Symbolic
analysis is a term from the numerical computing community. It
refers to phases that determine the computational patterns that
only depend on the nonzero pattern and not on numerical values.
Information from symbolic analysis can be used to make subsequent numeric manipulation faster, and the information can be
reused as long as the matrix nonzero structure remains constant.
For a number of sparse matrix methods such as LU and Cholesky,
it is well known that viewing their computations as a graph (e.g.,
elimination tree, dependence graph, and quotient graph) and applying a method-dependent graph algorithm yields information about
dependences that can then be used to more efficiently compute
the numerical method [14]. Most high-performance sparse matrix
computation libraries utilize symbolic information, but couple this
symbolic analysis with numeric computation, further making it
difficult for compilers to optimize such codes.
This work presents Sympiler, which generates high-performance
sparse matrix code by fully decoupling the symbolic analysis from
numeric computation and transforming code to utilize the symbolic
information. After obtaining symbolic information by running a
symbolic inspector, Sympiler applies inspector-guided transformations, such as variable-sized blocking, resulting in performance
,,
1
•
•
L: {n,Lp,Li,Lx}
b
x
x1
2
x2
• 3
x3
x4
4
• 5
x5
* x6 =
•
•
•
• 6
x7
7
x8
•
• 8
x9
•
•
•
• 9
x10
•
• 10
1
•
Kazem Cheshmi, Shoaib Kamil, Michelle Mills Strout, and Maryam Mehri Dehnavi
x=b; // copy RHS to x
for(j=0;j<n;j++){
x[j]/=Lx[Lp[j]];
Dependence Graph (DGL )
for(p=Lp[j]+1;p<Lp[j+1];p++){
β = {i|bi 6= 0} = {1, 6}
x[Li[p]]-=Lx[p]*x[j];}}
ReachL (β) = {1, 6, 7, 8, 9, 10} 1
2
•
9
3
4
7
10
8
5
6
(a)
x=b;
for(j=0;j<n;j++){
if(x[j] != 0){
x[j]/=Lx[Lp[j]];
for(p=Lp[j]+1;p<Lp[j+1];p++)
x[Li[p]]-=Lx[p]*x[j];}}
x=b;
for(px=0;px<reachSetSize;px++){
j=reachSet[px];
x[j]/=Lx[Lp[j]];
for(p=Lp[j]+1;p<Lp[j+1];p++){
x[Li[p]]-=Lx[p]*x[j];}}
(c) Library implementation
(d) Decoupled code
(b) Forward substitution
x=b;
x[0] /= Lx[0]; // Peel col 0
for(p = 1; p < 3; p++)
x[Li[p]] -= Lx[p] * x[0];
for(px=1;px<3;px++){
j=reachSet[px];x[j]/=Lx[Lp[j]];
for(p=Lp[j]+1;p<Lp[j+1];p++)
x[Li[p]]-=Lx[p]*x[j];}
x[7] /= Lx[20]; // Peel col 7
for(p = 21; p < 23; p++)
x[Li[p]] -= Lx[p] * x[7];
for(px=4;px<reachSetSize;px++){
j=reachSet[px];x[j]/=Lx[Lp[j]];
for(p=Lp[j]+1;p<Lp[j+1];p++)
x[Li[p]]-=Lx[p]*x[j];}
(e) Sympiler-generated
Figure 1: Four different codes for solving the linear system in (a). In all four code variants, the matrix L is stored in compressed
sparse column (CSC) format, with {n,Lp,Li,Lx} representing {matrix order, column pointer, row index, nonzeros} respectively.
The dependence graph DG L is the adjacency graph of matrix L; the vertices of DG L correspond to columns of L and its edges
show dependencies between columns in triangular solve. Vertices corresponding to nonzero columns are colored blue, and
columns that must participate in the computation due to dependence structure are colored red; the white vertices can be
skipped during computation. The boxes around columns show supernodes of different sizes. (b) is a forward substitution algorithm. (c) is a library implementation that skips iterations when the corresponding entry in the x is zero. (d) is the decoupled
code that uses symbolic information given in reachSet, which is computed by performing a depth-first search on DG L . (e) is the
Sympiler-generated code which peels iterations corresponding to columns within the reach-set with more than 2 nonzeros.
equivalent to hand-tuned libraries. But Sympiler goes further than
existing numerical libraries by generating code for a specific matrix nonzero structure. Because matrix structure often arises from
properties of the underlying physical system that the matrix represents, in many cases the same structure reoccurs multiple times,
with different values of nonzeros. Thus, Sympiler-generated code
can combine inspector-guided and low-level transformations to
produce even more efficient code. The transformations applied by
Sympiler improves the performance of sparse matrix codes through
applying optimizations for a single-core such as vectorization and
increased data locality which should extend to improve performance on shared and distributed memory systems.
1.1
Motivating Scenario
Sparse triangular solve takes a lower triangular matrix L and a righthand side (RHS) vector b and solves the linear equation Lx = b for
x. It is a fundamental building block in many numerical algorithms
such as factorization [14, 44], direct system solvers [13], and rank
update methods [18], where the RHS vector is often sparse. A naïve
implementation visits every column of matrix L to propagate the
contributions of its corresponding x value to the rest of x (see
Figure 1b). However, with a sparse b, the solution vector is also
sparse, reducing the required iteration space of sparse triangular
solve to be proportional to the number of nonzero values in x.
Taking advantage of this property requires first determining the
nonzero pattern of x. Based on a theorem from Gilbert and Peierls
[34], the dependence graph DG L = (V , E) for matrix L with nodes
V = {1, ..., n} and edges E = {(j, i)|Li j , 0} can be used to compute
the nonzero pattern of x, where n is the matrix rank and numerical
cancellation is neglected. The nonzero indices in x are given by
Reach L (β) which is the set of all nodes reachable from any node in
β = {i |bi , 0}, and can be computed with a depth-first search of
the directed graph DG L staring with β. An example dependence
graph is illustrated in Figure 1a. The blue colored nodes correspond
to set β and the final reach-set Reach L (β) contains all the colored
nodes in the graph.
Figure 1 shows four different implementations of sparse triangular solve. Most solvers assume the input matrix L is stored in a
compressed sparse column (CSC) storage format. While the naïve
implementation in Figure 1b traverses all columns, the typical library implementation shown in Figure 1c skips iterations when the
corresponding value in x is zero.
The implementation in Figure 1d shows a decoupled code that
uses the symbolic information provided by the pre-computed reachset. This decoupling simplifies numerical manipulation and reduces
the run-time complexity from O(|b | +n+ f ) in Figure 1c to O(|b | + f )
in Figure 1d, where f is the number of floating point operations
and |b| is the number of nonzeros in b. Sympiler goes further by
Sympiler: Transforming Sparse Codes
building the reach-set at compile-time and leveraging it to generate
code specialized for the specific matrix structure and RHS. The
Sympiler-generated code is shown in Figure 1e, where the code
only iterates over reached columns and peels iterations where the
number of nonzeros in a column is greater than some threshold (in
the case of the figure, this threshold is 2). These peeled loops can
be further transformed with vectorization to speed up execution.
This shows the power of fully decoupling the symbolic analysis
phase from the code that manipulates numeric values: the compiler
can aggressively apply conventional optimizations, using the reachset to guide the transformation. On matrices from the SuiteSparse
Matrix Collection, the Sympiler-generated code shows speedups
between 8.4× to 19× with an average of 13.6× compared to the forward solve code (Figure 1b) and from 1.2× to 1.7× with an average
of 1.3× compared to the library-equivalent code (Figure 1c).
1.2
Static Sparsity Patterns
A fundamental concept that Sympiler is built on is that the structure
of sparse matrices in scientific codes is dictated by the physical
domain and as such does not change in many applications. For
example, in power system modeling and circuit simulation problems the sparse matrix used in the matrix computations is often
a Jacobian matrix, where the structure is derived from interconnections among the power system and circuit components such
as generation, transmission, and distribution resources. While the
numerical values in the sparse input matrix change often, a change
in the sparsity structure occurs on rare occasions with a change
in circuit breakers, transmission lines, or one of the physical components. The sparse systems in simulations in domains such as
electromagentics [24, 28, 29, 47], computer graphics [33], and fluid
mechanics [6] are assembled by discretizing a physical domain and
approximating a partial differential equation on the mesh elements.
A sparse matrix method is then used to solve the assembled systems.
The sparse structure originates from the physical discretization and
therefore the sparsity pattern remains the same except where there
are deformations or if adaptive mesh refinement is used. Sparse
matrices in many other physical domains exhibit the same behavior
and benefit from Sympiler.
1.3
Contributions
This work describes Sympiler, a sparsity-aware code generator
for sparse matrix algorithms that leverages symbolic information
to generate fast code for a specific matrix structure. The major
contributions of this paper are:
• A novel approach for building compile-time symbolic inspectors that obtain information about a sparse matrix, to
be used during compilation.
• Inspector-guided transformations that leverage compile-time
information to transform sparse matrix code for specific
algorithms.
• Implementations of symbolic inspectors and inspectorguided transformations for two algorithms: sparse triangular solve and sparse Cholesky factorization.
,,
• A demonstration of the performance impact of our code
generator, showing that Sympiler-generated code can outperform state-of-the-art libraries for triangular solve and
Cholesky factorization by up to 1.7× and 6.3× respectively.
2
SYMPILER: A SYMBOLIC-ENABLED CODE
GENERATOR
Sympiler generates efficient sparse kernels by tailoring sparse code
to specific matrix sparsity structures. By decoupling the symbolic
analysis phase, Sympiler uses information from symbolic analysis
to guide code generation for the numerical manipulation phase
of the kernel. In this section, we describe the overall structure of
the Sympiler code generator, as well as the domain-specific transformations enabled by leveraging information from the symbolic
inspector.
2.1
Sympiler Overview
Sparse triangular solve and Cholesky factorization are currently
implemented in Sympiler. Given one of these numerical methods
and an input matrix stored using compressed sparse column (CSC)
format, Sympiler utilizes a method-specific symbolic inspector to
obtain information about the matrix. This information is used to
apply domain-specific optimizations while lowering the code for
the numerical method. In addition, the lowered code is annotated
with additional low-level transformations (such as unrolling) when
applicable based on domain- and matrix-specific information. Finally, the annotated code is further lowered to apply low-level
optimizations and output to C source code.
Code implementing the numerical solver is represented in a
domain-specific abstract syntax tree (AST). Sympiler produces the
final code by applying a series of phases to this AST, transforming
the code in each phase. An overview of the process is shown in
Figure 2. The initial AST for triangular solve is shown in Figure 2a
prior to any transformations.
2.2
Symbolic Inspector
Different numerical algorithms can make use of symbolic information in different ways, and prior work has described run-time graph
traversal strategies for various numerical methods [12, 14, 45, 52].
The compile-time inspectors in Sympiler are based on these strategies. For each class of numerical algorithms with the same symbolic
analysis approach, Sympiler uses a specific symbolic inspector to
obtain information about the sparsity structure of the input matrix and stores it in an algorithm-specific way for use during later
transformation stages.
We classify the used symbolic inspectors based on the numerical
method as well as the transformations enabled by the obtained
information. For each combination of algorithm and transformation,
the symbolic inspector creates an inspection graph from the given
sparsity pattern and traverses it during inspection using a specific
inspection strategy. The result of the inspection is the inspection set,
which contains the result of running the inspector on the inspection
graph. Inspection sets are used to guide the transformations in
Sympiler. Additional numerical algorithms and transformations
can be added to Sympiler, as long as the required inspectors can be
described in this manner as well.
,,
Kazem Cheshmi, Shoaib Kamil, Michelle Mills Strout, and Maryam Mehri Dehnavi
Sparsity
Pattern
Symbolic Inspector
Numerical
Method
VS-Block
VI-Prune
Low-Level
Transformations
Code
Generation
Inspector-Guided Transformations
(a) Initial AST
(b) After VI-Prune
VI-Prune
for sol.j0 in 0..Lsp.n
VS-Block
x[bspj0 ]/=Lx[Lsp.diag(j0 )];
VS-Block
for sol.j1 in Lsp.colj0 ..Lsp.colj0 +1
x[Lsp.rowj1 ]-=Lx[j1 ]*x[bspj0 ];
peel(0,3)
for sol.p0 in 0..pruneSetSize
j0 =pruneSetp0 ;
x[bspj0 ]/=Lx[Lsp.diag(j0 )];
vec(0)
for sol.j1 in Lsp.colj0 ..Lsp.colj0 +1
x[Lsp.rowj1 ]-=Lx[j1 ]*x[bspj0 ];
(c) After Low-Level Transformations
s0 =pruneSet0 ;
x[bsps0 ]/=Lx[Lsp.diag(s0 )];
for sol.j1 in Lsp.cols0 ..Lsp.cols0 +1
x[Lsp.rowj1 ]-=Lx[j1 ]*x[bsps0 ];
for sol.p0 in 0..pruneSetSize
j0 =pruneSetp0 ;
x[bspj0 ]/=Lx[Lsp.diag(j0 )];
for sol.j1 in Lsp.colj0 ..Lsp.colj0 +1
x[Lsp.rowj1 ]-=Lx[j1 ]*x[bspj0 ];
1
Figure 2: Sympiler lowers a functional representation of a sparse kernel to imperative code using the inspection sets. Sympiler
constructs a set of loop nests and annotates them with some domain-specific information that is later used in inspectorguided transformations. The inspector-guided transformations use the lowered code and inspection sets as input and apply
transformations. Inspector-guided transformations also provide hints for further low-level transformations by annotating the
code. For instance, the transformation steps for the code in Figure 1 are: (a) Initial AST with annotated information showing
where the VI-Prune and VS-Block transformations apply. (b) The symbolic inspector sends the reach-set as pruneSet, which
VI-Prune uses to add hints to further steps— in this case, peeling iterations 0 and 3. (c). The hinted low-level transformations
are applied and final code generated (peeling is only shown for the iteration zero).
1
2
3
4
5
6
7
8
9
10
11
1
for(I 1 ){
.
.
for(I k < m) {
.
.
for(I n (I k , ..., I n−1 )) {
a[idx(I 1 ,...,I k ,...,I n )];
}
}
}
2
3
4
5
6
7
8
9
10
11
12
(a) Before
for(I 1 ) {
.
.
for(Ip < pruneSetSize) {
I k0 = pruneSet[Ip ];
.
.
for(I n (I k0 , ..., I n−1 )) {
a[idx(I 1 ,...,I k0 ,...,I n )];
}
}
}
(b) After
Variable Iteration Space Pruning, loop[k].VI-Prune(pruneSet,pruneSetSize)
1
1
2
3
4
5
for(I) {
for(J) {
B[idx1(I,J)] op1= a[idx2(I,J)];
}
}
2
3
4
5
6
7
(c) Before
for(b < blockSetSize) {
for( J1 < blockSet[b].x) {
for( J2 < blockSet [b].y ) {
B[idx1(b, J1 , J2 )] op1 = A[idx2(b, J1 , J2 )];
}
}
}
(d) After
2D Variable-Sized Blocking, loop[I].VS-Block(blockSet,blockSetSize)
Figure 3: The inspector-guided transformations. Top: The loop over Ik with iteration space m in (a) transforms to a loop over
Ip with iteration space pruneSetSize in (b). Any use of the original loop index Ik is replaced with its corresponding value from
pruneSet i.e., Ik0 . Bottom: The two nested loops in (c) are transformed into loops over variable-sized blocks in (d).
Sympiler: Transforming Sparse Codes
For our motivating example, triangular solve, the reach-set can be
used to prune loop iterations that perform work that is unnecessary
due to the sparseness of matrix or the right hand side. In this case,
the inspection set is the reach-set, and the inspection strategy is to
perform a depth-first search over the inspection graph, which is the
directed dependency graph DG L of the triangular matrix. For the
example linear system shown in Figure 1, the symbolic inspector
generates the reach-set {6, 1, 7, 8, 9, 10}.
2.3
,,
inputs in sparse kernels make blocking optimizations challenging.
The symbolic inspector identifies sub-kernels with similar structure
in the sparse matrix methods and the sparse inputs to provide the
VS-Block stage with “blockable” sets that are not necessarily of the
same size or consecutively located. These blocks are similar to the
concept of supernodes [44] in sparse libraries. VS-Block must deal
with a number of challenges:
• The block sizes are variable in a sparse kernel.
• Due to using compressed storage formats, the block elements may not be in consecutive memory locations.
• The type of numerical method used may need to change after applying this transformation. For example, applying VSBlock to Cholesky factorization requires dense Cholesky
factorization on the diagonal segment of the blocks, and
the off-diagonal segments of the blocks must be updated
using a set of dense triangular solves.
Inspector-guided Transformations
The initial lowered code along with the inspection sets obtained by
the symbolic inspector are passed to a series of passes that further
transform the code. Sympiler currently supports two transformations guided by the inspection sets: Variable Iteration Space Pruning
and 2D Variable-Sized Blocking, which can be applied independently
or jointly depending on the input sparsity. As shown in Figure 2a,
the code is annotated with information showing where inspectorguided transformations may be applied. The symbolic inspector
provides the required information to the transformation phases,
which decide whether to transform the code based on the inspection
sets. Given the inspection set and annotated code, transformations
occur as illustrated in Figure 3.
2.3.1 Variable Iteration Space Pruning. Variable Iteration Space
Pruning (VI-Prune) prunes the iteration space of a loop using information about the sparse computation. The iteration space for sparse
codes can be considerably smaller than that for dense codes, since
the computation needs to only consider iterations with nonzeros.
The inspection stage of Sympiler generates an inspection set that
enables transforming the unoptimized sparse code to a code with a
reduced iteration space.
Given this inspection set, the VI-Prune transformation can be
applied at a particular loop-level to the sparse code to transform
it from Figure 3a to Figure 3b. In the figure, the transformation is
applied to the k t h loop nest in line 4. In the transformed code the
iteration space is pruned to pruneSetSize, which is the inspection set size. In addition to the new loop, all references to Ik (the
loop index before transformation) are replaced by its corresponding value from the inspection set, pruneSet[Ip ]. Furthermore, the
transformation phase utilizes inspection set information to annotate specific loops with further low-level optimizations to be applied
by later stages of code generation. These annotations are guided by
thresholds that decide when specific low-level optimizations result
in faster code.
In our running example of triangular solve, the generated inspection set from the symbolic inspector enables reducing the iteration
space of the code. The VI-Prune transformation elides unnecessary
iterations due to zeros in the right hand side. In addition, depending
on the number of iterations the loops will run (which is known
thanks to the symbolic inspector), loops are annotated with directives to unroll and/or vectorize during code generation.
2.3.2 2D Variable-Sized Blocking. 2D Variable-Sized Blocking
(VS-Block) converts a sparse code to a set of non-uniform dense subkernels. In contrast to the conventional approach of blocking/tiling
dense codes, where the input and computations are blocked into
smaller uniform sub-kernels, the unstructured computations and
To address the first challenge, the symbolic inspector uses an
inspection strategy that provides an inspection set specifying the
size of each block. For the second challenge, the transformed code
allocates temporary block storage and copies data as needed prior
to operating on the block. Finally, to deal with the last challenge,
the synthesized loops/instructions in the lowering phase contain
information about the block location in the matrix, and when applying this transformation, the correct operation is chosen for each
loop/instruction. As with the VI-Prune transformation, VS-Block
also annotates loops with further low-level transformations such as
tiling to be applied during code generation. By leveraging specific
information about the matrix when applying the transformation,
Sympiler is able to mitigate all of the difficulties of applying VSBlock to sparse numerical methods.
An off-diagonal version of the VS-Block transformation is shown
in Figures 3c and 3d. As shown, a new outer loop is made that provides the block information to the inner loops using the given
blockSet. The inner loop in Figure 3c transforms to two nested
loops (lines 2–6) that iterate over the block specified by the outer
loop. The diagonal version VS-Block heavily depends on domain
information. More detailed examples of applying this transformation to triangular solve and Cholesky factorization is described in
Section 3.
2.4
Enabled Conventional Low-level
Transformations
While applying inspector-guided transformations, the original loop
nests are transformed into new loops with potentially different
iteration spaces, enabling the application of conventional low-level
transformations. Based on the applied inspector-guided transformations as well as the properties of the input matrix and right-hand
side vectors, the code is annotated with some transformation directives. An example of these annotations are shown in Figure 2b
where loop peeling is annotated within the VI-Pruned code. To
decide when to add these annotations, the inspector-guided transformations use sparsity-related parameters such as the average
block size. The main sources of enabling low-level transformations
are:
,,
Kazem Cheshmi, Shoaib Kamil, Michelle Mills Strout, and Maryam Mehri Dehnavi
(1) Symbolic information provides dependency information
at compile-time, allowing Sympiler to apply more transformations such as peeling based on the reach-set in Figure 1;
(2) Inspector-guided transformations remove some of the indirect memory accesses and annotate the code with potential
conventional transformations;
(3) Sparsity-specific code generation enables Sympiler to know
details such as loop boundaries at compile-time. Thus, several customized transformations are applied such vectorization of loops with iteration counts greater than a threshold;
Figure 1e shows how some of the iterations in the triangular
solve code after VI-Prune can be peeled. In this example, the inspection set used for VI-Prune is the reach-set {1, 6, 8, 9, 10}. Because
the reach-set is created in topological order, iteration ordering dependencies are met and thus code correctness is guaranteed after
loop peeling. As shown in Figure 2b, the transformed code after VIPrune is annotated with the enabled peeling transformation based
on the number of nonzeros in the columns (the column count).
The two selected iterations with column count greater than 2 are
peeled to replace them with a specialized kernel or to apply another
transformation such as vectorization.
3
CASE STUDIES
1
2
3
4
5
6
7
8
9
10
11
for(column j = 0 to n){
f = A(:,j)
PruneSet = The sparsity pattern of row j
for(every row r in PruneSet){ // Update
f -= L(j:n,r) * L(j,r);
}
L(k,k) = sqrt(f(k)); // Diagonal
for(off-diagonal elements in f){ // Off-diagonal
L(k+1:n,k) = f(k+1:n) / L(k,k);
}
}
Figure 4: The pseudo-code of the left-looking Cholesky.
Sympiler currently supports two important sparse matrix computations: triangular solve and Cholesky factorization. This section discusses some of the graph theory and algorithms used in
Sympiler’s symbolic inspector to extract inspections sets for these
two matrix methods. The run-time complexity of the Symbolic inspector is also presented to evaluate inspection overheads. Finally,
we demonstrate how the VI-Prune and VS-Block transformations
are applied using the inspection sets. Table 1 shows a classification
of the inspection graphs, strategies, and resulting inspection sets
for the two studied numerical algorithms in Sympiler. As shown in
Table 1, the symbolic inspector performs a set of known inspection
methods and generates some sets which includes symbolic information. The last column of Table 1 shows the list of transformations
enabled by each inspector-guided transformation. We also discuss
extending Sympiler to other matrix methods.
3.1
Sparse Triangular Solve
Theory: In the symbolic inspector, the dependency graph DG L is
traversed using depth first search (DFS) to determine the inspection
A
1 •
• 2
•
•
3
4 • •
• 5
•
6
•
• •
• • •
•
•
•
•
T
L
• •
•
• •
•
•
•
−→
•
•
7
•
8
•
9 •
• • • 10
1
10
2
• 3
9
4
• 5
• • • • 6
•
• • 7
• • •
• • 8
• • • • • • • 9
•
• • • 10
8
7
1
6
3
5
2
4
1
Figure 5: An example matrix A and its L factor from
Cholesky factorization. The corresponding elimination tree
(T ) of A is also shown. Nodes in T and columns in L highlighted with the same color belong to the same supernode.
The red nonzeros in L are fill-ins.
set for the VI-Prune transformation, which in this case is the reachset from DG L and the right-hand side vector. The graph DG L can
also be used to detect blocks with similar sparsity patterns, also
known as supernodes, in sparse triangular solve. The block-set,
which contains columns of L grouped into supernodes, are identified
by inspecting DG L using a node equivalence method. The node
equivalence algorithm first assumes nodes vi and v j are equivalent
and then compares their outgoing edges. If the outgoing edges go
to the same destination nodes then the two nodes are equal and are
merged.
Inspector-guided Transformations: Using the reach-set,VIPrune limits the iteration spaces of the loops in triangular solve to
only those that operate on the necessary nonzeros. The VS-Block
transformation changes the loops to apply blocking as shown in
Figure 2a in triangular solve. The diagonal block of each columnblock, which is a small triangular solve, is solved first. The solution
of the diagonal components is then substituted in the off-diagonal
segment of the matrix.
Symbolic Inspection: The time complexity of DFS on graph
DG L is proportional to the number of edges traversed and the
number of nonzeros in the RHS of the system. The time complexity
for the node equivalence algorithm is proportional to the number
of nonzeros in L. We provide overheads for these methods for the
tested matrices in Section 4.3.
3.2
Cholesky Factorization
Cholesky factorization is commonly used in direct solvers and
is used to precondition iterative solvers. The algorithm factors a
Hermitian positive definite matrix A into LLT , where matrix L is a
lower triangular matrix. Figure 5 shows an example matrix A and
the corresponding L matrix after factorization.
Theory: The elimination tree (etree) [17] is one of the most
important graph structures used in the symbolic analysis of sparse
factorization algorithms. Figure 5 shows the corresponding elimination tree for factorizing the example matrix A. The etree of A is a
spanning tree of G + (A) satisfying parent[j] = min{i > j : Li j , 0}
where G + (A) is the graph of L + LT . The filled graph or G + (A)
results at the end of the elimination process and includes all edges
of the original matrix A as well as fill-in edges. In-depth discussions
of the theory behind the elimination tree, the elimination process,
and the filled graph can be found in [14, 52].
Sympiler: Transforming Sparse Codes
,,
Table 1: Inspection and transformation elements in Sympiler for triangular solve and Cholesky. DG: dependency graph, SP
(RHS): sparsity patterns of the right-hand side vector, DFS: depth-first search, SP(A): sparsity patterns of the coefficient A, SP
(L j ): sparsity patterns of the j t h row of L, unroll: loop unrolling, peel: loop peeling, dist: loop distribution, tile: loop tiling.
Transformations
VI-Prune
VS-Block
Inspection
Graph
DG +
SP(RHS)
DG
Triangular Solve
Inspection
Inspection Set
Strategy
DFS
Prune-set
(reach-set)
Node
Block-set
equivalence (supernodes)
Figure 4 shows the pseudo-code of the left-looking sparse Cholesky,
which is performed in two phases of update (lines 3–6) and column
factorization (lines 7–10). The update phase gathers the contributions from the already factorized columns on the left. The column
factorization phase calculates the square root of the diagonal element and applies it to the off-diagonal elements.
To find the prune-set that enables the VI-Prune transformation,
the row sparsity pattern of L has to be computed; Figure 4 shows
how this information is used to prune the iteration space of the
update phase in the Cholesky algorithm. Since L is stored in column
compressed format, the etree and the sparsity pattern of A are used
to determine the L row sparsity pattern. A non-optimal method
for finding the row sparsity pattern of row i in L is that for each
nonzero Ai j the etree of A is traversed upwards from node j until
node i is reached or a marked node is found. The row-count of i is
the visited nodes in this subtree. Sympiler uses a similar but more
optimized approach from [14] to find row sparsity patterns.
Supernodes used in VS-Block for Cholesky are found with the L
sparsity pattern and the etree. The sparsity pattern of L is different
from A because of fill-ins created during factorization. However,
the elimination tree T along with the sparsity pattern of A are used
to find the sparsity pattern of L prior to factorization. As a result,
memory for L can be allocated ahead of time to eliminate the need
for dynamic memory allocation. To create the supernodes, the fillin pattern should be first determined. Equation (1) is based on a
theorem from [31] and computes the sparsity pattern of column
j in L, L j , where T (s) is the parent of node s in T and “\" means
exclusion. The theorem states that the nonzero pattern of L j is the
union of the nonzero patterns of the children of j in the etree and
the nonzero pattern of column j in A.
Ø© Ø
ª
Ls \{s}®
(1)
«j=T (s)
¬
When the sparsity pattern of L is obtained, the following rule is used
to merge columns to create basic supernodes: when the number
of nonzeros in two adjacent columns j and j − 1, regardless of the
diagonal entry in j − 1, is equal, and j − 1 is the only child of j in T ,
the two columns can be merged.
Inspector-guided transformations: The VI-Prune transformation applies to the update phase of Cholesky. With the row sparsity
pattern information, when factorizing column i Sympiler only iterates over dependent columns instead of all columns smaller than i.
The VS-Block transformation applies to both update and column
L j = Aj
Ø
{j}
Inspection
Graph
etree +
SP(A)
etree
+
ColCount(A)
Cholesky
Inspection
Strategy
Single-node
up-traversal
Up-traversal
Inspection Set
Prune-set
(SP(L j ))
Block-set
(supernodes)
Enabled
Low-level
dist, unroll, peel,
vectorization
tile, unroll, peel,
vectorization
factorization phases. Therefore, the outer loop in the Cholesky algorithm in Figure 4 is converted to a new loop that iterates over
the provided block-set. All references to the columns j in the inner
loops will be changed to the blockSet[j]. For the diagonal part of
the column factorization, a dense Cholesky needs to be computed
instead of the square root in the non-supernodal version. The resulting factor from the diagonal elements applies to the off-diagonal
rows through a sequence of dense triangular solves. VS-Block also
converts the update phase from vector operations to matrix operations.
Symbolic Inspection: The computational complexity for building the etree in sympiler is nearly O(|A|). The run-time complexity
for finding the sparsity pattern of row i is proportional to the number of nonzeros in row i of A. The method is executed for all columns
which results in a run-time of nearly O(|A|).The inspection overhead for finding the block-set for VS-Block includes the sparsity
detection which is done in nearly O(|A| + 2n) and the supernode
detection which has a run-time complexity of O(n) [14].
3.3
Other Matrix Methods
The inspection graphs and inspection strategies supported in the
current version of Sympiler can support a large class of commonlyused sparse matrix computations. The applications of the elimination tree go beyond the Cholesky factorization method and extend
to some of the most commonly used sparse matrix routines in
scientific applications such as LU, QR, orthogonal factorization
methods [46], and incomplete and factorized sparse approximate
inverse preconditioner computations [40]. Inspection of the dependency graph and proposed inspection strategies that extract
reach-sets and supernodes from the dependency graph are the fundamental symbolic analyses required to optimize algorithms such
as rank update and rank increase methods [18], incomplete LU(0)
[49], incomplete Cholesky preconditioners, and up-looking implementations of factorization algorithms. Thus, Sympiler with the
current set of symbolic inspectors can be made to support many of
these matrix methods. We plan to extend to an even larger class of
matrix methods and to support more optimization methods.
4
EXPERIMENTAL RESULTS
We evaluate Sympiler by comparing the performance to two stateof-the-art libraries, namely Eigen [36] and CHOLMOD [11], for
the Cholesky factorization method and the sparse triangular solve
algorithm. Section 4.1 discusses the experimental setup and experimental methodology. In Section 4.2 we demonstrate that the
,,
Kazem Cheshmi, Shoaib Kamil, Michelle Mills Strout, and Maryam Mehri Dehnavi
Table 2: Matrix set: The matrices are sorted based on the
number of nonzeros in the original matrix; nnz refers to
number of nonzeros, n is the rank of the matrix.
Sympiler: VS-Block
Sympiler: VS-Block+VI-Prune
Sympiler: VS-Block+VI-Prune+Low-Level
Eigen
Problem
Name
ID
1
cbuckle
2
Pres_Poisson
3
gyro
4
gyro_k
5
Dubcova2
6
msc23052
7
thermomech_dM
8
Dubcova3
9
parabolic_fem
10
ecology2
11
tmt_sym
n
13.7
14.8
17.4
17.4
65.0
23.1
204
147
526
1000
727
nnz (A)
(106 )
0.677
0.716
1.02
1.02
1.03
1.14
1.42
3.64
3.67
5.00
5.08
transformations enabled by Sympiler generate highly-optimized
codes for sparse matrix algorithms compared to state-of-the-art
libraries. Although symbolic analysis is performed only once at
compile-time for a fixed sparsity pattern in Sympiler, we analyze
the cost of the symbolic inspector in Section 4.3 and compare it
with symbolic costs in Eigen and CHOLMOD.
4.1
Methodology
We selected a set of symmetric positive definite matrices from [19],
which are listed in Table 2. The matrices originate from different
domains and vary in size. All matrices have real numbers and
are in double precision. The testbed architecture is a 3.30GHz
Intel®Core™i7-5820K processor with L1, L2, and L3 cache sizes
of 32KB, 256KB, and 15MB respectively and turbo-boost disabled.
We use OpenBLAS.0.2.19 [71] for dense BLAS (Basic Linear Algebra Subprogram) routines when needed. All Sympiler-generated
codes are compiled with GCC v.5.4.0 using the -O3 option. Each
experiment is executed 5 times and the median is reported.
We compare the performance of the Sympiler-generated code
with CHOLMOD [11] as a specialized library for Cholesky factorization and with Eigen [36] as a general numerical library. CHOLMOD
provides one of the fastest implementations of Cholesky factorization on single-core architectures [35]. Eigen supports a wide range
of sparse and dense operations including sparse triangular solve
and Cholesky. Thus, for Cholesky factorization we compare with
both Eigen and CHOLMOD while results for triangular solve are
compared to Eigen. Both libraries are installed and executed using
the recommended default configuration. Since Sympiler’s current
version does not support node amalgamation [26], this setting is
not enabled in CHOLMOD. For the Cholesky factorization both
libraries support the more commonly used left-looking (supernodal)
algorithm which is also the algorithm used by Sympiler. Sympiler
applies either both or one of the inspector-guided transformations
as well as some of the enabled low-level transformations; currently,
Sympiler implements unrolling, scalar replacement, and loop distibution from among the possible low-level transformations.
Triangular Solve GFLOP/s
2
(103 )
1
1
2
3
4
5
6
7
8
9
10
11
Figure 6: Sympiler’s performance compared to Eigen for triangular solve. The stacked-bars show the performance of
the Sympiler (numeric) code with VS-Block and VI-Prune.
The effects of VS-Block, VI-Prune, and low-level transformations on Sympiler performance are shown separately.
4.2
Performance of Generated Code
This section shows how the combination of the introduced transformations and the decoupling strategy enable Sympiler to outperform
two state-of-the-art libraries for sparse Cholesky and sparse triangular solve.
Triangular solve: Figure 6 shows the performance of Sympilergenerated code compared to the Eigen library for a sparse triangular
solve with a sparse RHS. The nonzero fill-in of the RHS in our
experiments is selected to be less than 5%. The sparse triangular
system solver is often used as a sub-kernel in algorithms such as
left-looking LU [14] and Cholesky rank update methods [18] or as
a solver after matrix factorizations. Thus, typically the sparsity of
the RHS in sparse triangular systems is close to the sparsity of the
columns of a sparse matrix. For the tested problems, the number of
nonzeros for all columns of L was less than 5%.
The average improvement of Sympiler-generated code, which we
refer to as Sympiler (numeric), over the Eigen library is 1.49×. Eigen
implements the approach demonstrated in Figure 1c, where symbolic analysis is not decoupled from the numerical code. However,
the Sympiler-generated code only manipulates numerical values
which leads to higher performance. Figure 6 also shows the effect
of each transformation on the overall performance of the Sympilergenerated code. In the current version of Sympiler the symbolic
inspector is designed to generate sets so that VS-Block can be applied before VI-Prune. Our experiments show that this ordering
often leads to better performance mainly because Sympiler supports
supernodes with a full diagonal block. As support for more transformations are added to Sympiler, we will enable it to automatically
decide the best transformation ordering. Whenever applicable, the
vectorization and peeling low-level transformations are also applied
after VS-Block and VI-Prune. Peeling leads to higher performance if
applied after VS-Block where iterations related to single-column supernodes are peeled. Vectorization is always applied after VS-Block
and does not lead to performance if only VI-Prune is applied.
Sympiler: Transforming Sparse Codes
2.0
Sympiler: VS-Block
Sympiler: VS-Block+Low-Level
Eigen (Numeric)
CHOLMOD (Numeric)
Triangular Solve Time / (Eigen Time)
Cholesky GFLOP/s
10
,,
5
1
2
3
4
5
6
7
8
9
10
1.5
1.0
0.5
11
Figure 7: The performance of Sympiler (numeric) for
Cholesky compared to CHOLMOD (numeric) and Eigen
(numeric). The stacked-bar shows the performance of the
Sympiler-generated code. The effect of VS-Block and lowlevel transformations are shown separately. The VI-Prune
transformation is already applied to the baseline code so it
is not shown here.
Matrices 3, 4, 5, and 7 do not benefit from the VS-Block transformation so their Sympiler run-times in Figure 6 are only for VI-Prune.
Since small supernodes often do not lead to better performance,
Sympiler does not apply the VS-Block transformation if the average
size of the participating supernodes is smaller than a threshold.This
parameter is currently hand-tuned and is set to 160. VS-Block is not
applied to matrices 3, 4, 5, and 7 since the average supernode size is
too small and thus does not improve performance. Also, since these
matrices have a small column count vectorization does not payoff.
Cholesky: We compare the numerical manipulation code of Eigen
and CHOLMOD for Cholesky factorization with the Sympilergenerated code. The results for CHOLMOD and Eigen in Figure 7
refer to the numerical code performance in floating point operations
per second (FLOP/s). Eigen and CHOLMOD both execute parts of
the symbolic analysis only once if the user explicitly indicates that
the same sparse matrix is used for subsequent executions. However,
even with such an input from the user, none of the libraries fully
decouple the symbolic information from the numerical code. This is
because they can not afford to have a separate implementation for
each sparsity pattern and also do not implement sparsity-specific
optimizations. For fairness, when using Eigen and CHOLMOD we
explicitly tell the library that the sparsity is fixed and thus report
only the time related to the the library’s numerical code (which
still contains some symbolic analysis).
As shown in Figure 7, for Cholesky factorization Sympiler performs up to 2.4× and 6.3× better than CHOLMOD and Eigen respectively. Eigen uses the left-looking non-supernodal approach
therefore, its performance does not scale well for large matrices.
CHOLMOD benefits from supernodes and thus performs well for
large matrices with large supernodes. However, CHOLMOD does
not perform well for some small matrices and large matrices with
small supernodes. Sympiler provides the highest performance for
almost all tested matrix types which demonstrates the power of
sparsity-specific code generation.
Sympiler (Numeric)
Sympiler (Symbolic)
Eigen
1
2
3
4
5
6
7
8
9
10
11
Figure 8: The figure shows the sparse triangular solve symbolic+numeric time for Sympiler and Eigen’s runtime normalized over the Eigen time (lower is better).
The application of kernel-specific and aggressive optimizations
when generating code for dense sub-kernels enables Sympiler to
generate fast code for any sparsity pattern. Since BLAS routines
are not well-optimized for small dense kernels they often do not
perform well for the small blocks produced when applying VSBlock to sparse codes [61]. Therefore, libraries such as CHOLMOD
do not perform well for matrices with small supernodes. Sympiler
has the luxury to generate code for its dense sub-kernels; instead
of being handicapped by the performance of BLAS routines, it
generates specialized and highly-efficient codes for small dense
sub-kernels. If the average column-count for a matrix is below a
tuned threshold, Sympiler will call BLAS routines [71] instead. Since
the column-count directly specifies the number of dense triangular
solves, which is the most important dense sub-kernel in Cholesky,
the average column-count is used to decide when to switch to BLAS
routines [71]. For example, the average column-count of matrices
3, 4, 6, and 8 are less than the column-count threshold.
Decoupling the prune-set calculation from the numerical manipulation phase also improves the performance of the Sympilergenerated code. As discussed in subsection 3.2, the sparse Cholesky
implementation needs to obtain the row sparsity pattern of L. The
elimination tree of A and the upper triangular part of A are both
used in CHOLMOD and Eigen to find the row sparsity pattern.
Since A is symmetric and only its lower part is stored, both libraries
compute the transpose of A in the numerical code to access its upper
triangular elements. Through fully decoupling symbolic analysis
from the numerical code, Sympiler has the L row sparsity information in the prune-set ahead of time and therefore, both the reach
function and the matrix transpose operations are removed from
the numeric code.
4.3
Symbolic Analysis Time
All symbolic analysis is performed at compile-time in Sympiler
and its generated code only manipulates numerical values. Since
symbolic analysis is performed once for a specific sparsity pattern,
its overheads amortize with repeat executions of the numerical
code. However, as demonstrated in Figures 8 and 9 even if the
numerical code is executed only once, which is not common in
Kazem Cheshmi, Shoaib Kamil, Michelle Mills Strout, and Maryam Mehri Dehnavi
scientific applications, the accumulated symbolic+numeric time of
Sympiler is close to Eigen for the triangular solve and faster than
both Eigen and CHOLMOD for Cholesky.
Triangular solve: Figure 8 shows the time Sympiler spends
to do symbolic analysis at compile-time, Sympiler (symbolic), for
the sparse triangular solve. No symbolic time is available for Eigen
since as discussed, Eigen uses the code in Figure 1c for its triangular
solve implementation. Figure 8 shows the symbolic analysis and
numerical manipulation time of Sympiler normalized over Eigen’s
run-time. Sympiler’s numeric plus symbolic time is on average
1.27× slower than the Eigen code. In addition, code generation and
compilation in Sympiler costs between 6–197× the cost of the numeric solve, depending on the matrix. It is important to note that
since the sparsity structure of the matrix in triangular solve does
not change in many applications, the overhead of the symbolic
inspector and compilation is only paid once. For example, in preconditioned iterative solvers a triangular system must be solved per
iteration, and often the iterative solver must execute thousands of iterations [8, 42, 50] until convergence since the systems in scientific
applications are not necessarily well-conditioned.
Cholesky: Sparse libraries perform symbolic analysis ahead of
time which can be re-used for same sparsity patterns and improves
the performance of their numerical executions. We compare the
analysis time of the libraries with Sympiler’s symbolic inspection
time. Figure 9 provides the symbolic analysis and numeric manipulation times for both libraries normalized to Eigen time. The time
spent by Sympiler to perform symbolic analysis is referred to as
Sympiler symbolic. CHOLMOD (symbolic) and Eigen (symbolic)
refer to the partially decoupled symbolic code that is only run once
if the user indicates that sparsity remains static. In nearly all cases
Sympiler’s accumulated time is better than the other two libraries.
Code generation and compilation, which are not shown in the chart,
add a very small amount of time, costing at most 0.3× the cost of
numeric factorization. Also, like the triangular solve example, the
matrix with a fixed sparsity pattern must be factorized many times
in scientific applications. For example, in Newton-Raphson (NR)
solvers for nonlinear systems of equations, a Jacobian matrix is factorized in each iteration and the NR solvers require tens or hundreds
of iterations to converge [21, 51].
5
RELATED WORK
Compilers for general languages are hampered by optimization
methods that either give up on optimizing sparse codes or only
apply conservative transformations that do not lead to high performance. This is due to the indirection required to index and loop
over the nonzero elements of sparse data structures. Polyhedral
methods are limited when dealing with non-affine loop nests or
subscripts [5, 10, 41, 54, 65, 67] which are common in sparse computations.
To make it possible for compilers to apply more aggressive loop
and data transformations to sparse codes, recent work [63, 66, 68–
70] has developed compile-time techniques for automatically creating inspectors and executors for use at run-time. These techniques
use an inspector to analyze index arrays in sparse codes at run-time
and an executor that uses this run-time information to execute code
with specific optimizations. These inspector-executor techniques
2.0
Cholesky Time / (Eigen Time)
,,
Sympiler (Numeric)
Sympiler (Symbolic)
Eigen (Numeric)
Eigen (Symbolic)
CHOLMOD (Numeric)
CHOLMOD (Symbolic)
1.5
1.0
0.5
1
2
3
4
5
6
7
8
9
10
11
Figure 9: The figure shows the symbolic+numeric time for
Sympiler, CHOLMOD, and Eigen for the Cholesky algorithm. All times are normalized over the Eigen’s accumulated symbolic+numeric time (lower is better).
are limited in that they only apply to sparse codes with static index
arrays; such codes require the matrix structure to not change during
the computation. The aforementioned approach performs well for
methods such as sparse incomplete LU(0) and Gauss-Seidel methods where additional nonzeros/fill-ins are not introduced during
computation. However, in a large class of sparse matrix methods,
such as direct solvers including Cholesky, LU, and QR decompositions, index arrays dynamically change during computation since
the algorithm itself introduces fill-ins. In addition, the indirections
and dependencies in sparse direct solvers are tightly coupled with
the algorithm, making it difficult to apply inspector-executor techniques.
Domain-specific compilers integrate domain knowledge into
the compilation process, improving the compiler’s ability to transform and optimize specific kinds of computations. Such an approach
has been used successfully for stencil computations [39, 55, 64], signal processing [53], dense linear algebra [37, 62], matrix assembly
and mesh analysis [1, 48], simulation [9, 43], and sparse operations [16, 56]. Though the simulations and sparse compilers use
some knowledge of matrix structure to optimize operations, they
do not build specialized matrix solvers.
Specialized Libraries are the typical approach for sparse direct
solvers. These libraries differ in (1) which numerical methods are
implemented, (2) the implementation strategy or variant of the
solver, (3) the type of the platform supported, and (4) whether the
algorithm is specialized for specific applications.
Each numerical method is suitable for different classes of matrices; for example, Cholesky factorization requires the matrix
be symmetric (or Hermitian) positive definite. Libraries such as
SuperLU [22], KLU [20], UMFPACK [12], and Eigen [36] provide
optimized implementations for LU decomposition methods. The
Cholesky factorization is available through libraries such as Eigen [36],
CSparse [14], CHOLMOD [11], MUMPS [2–4], and PARDISO [57,
58]. QR factorization is implemented in SPARSPAK[30, 32], SPLOOES
[7], Eigen [36], and CSparse [14]. The optimizations and algorithm
variants used to implement sparse matrix methods differ between
libraries. For example LU decomposition can be implemented using
Sympiler: Transforming Sparse Codes
multifrontal methods [12, 15, 38], left-looking [20, 22, 27, 30], rightlooking [25, 45, 59], and up-looking [13, 60] methods. Libraries are
developed to support different platforms such as sequential implementations [11, 14, 20], shared memory [15, 23, 57], and distributed
memory [3, 23]. Finally, some libraries are designed to perform well
on matrices arising from a specific domain. For example, KLU [20]
works best for circuit simulation problems. In contrast, SuperLUMT applies optimizations with the assumption that the input matrix
structure leads to large supernodes; such a strategy is a poor fit for
circuit simulation problems.
6
CONCLUSION
In this paper we demonstrated how decoupling symbolic analysis
from numerical manipulation can enable the generation of domainspecific highly-optimized sparse codes with static sparsity patterns.
Sympiler, the proposed domain-specific code generator, takes the
sparse matrix pattern and the sparse matrix algorithm as inputs to
perform symbolic analysis at compile-time. It then uses the information from symbolic analysis to apply a number of inspector-guided
and low-level transformations to the sparse code. The Sympilergenerated code outperforms two state-of-the-art sparse libraries,
Eigen and CHOLMOD, for the sparse Cholesky and the sparse
triangular solve algorithms.
REFERENCES
[1] Martin S Alnæs, Anders Logg, Kristian B Ølgaard, Marie E Rognes, and Garth N
Wells. 2014. Unified form language: A domain-specific language for weak formulations of partial differential equations. ACM Transactions on Mathematical
Software (TOMS) 40, 2 (2014), 9.
[2] Patrick R Amestoy, Iain S Duff, and J-Y L’Excellent. 2000. Multifrontal parallel
distributed symmetric and unsymmetric solvers. Computer methods in applied
mechanics and engineering 184, 2 (2000), 501–520.
[3] Patrick R Amestoy, Iain S Duff, Jean-Yves L’Excellent, and Jacko Koster. 2001. A
fully asynchronous multifrontal solver using distributed dynamic scheduling.
SIAM J. Matrix Anal. Appl. 23, 1 (2001), 15–41.
[4] Patrick R Amestoy, Abdou Guermouche, Jean-Yves LâĂŹExcellent, and Stéphane
Pralet. 2006. Hybrid scheduling for the parallel solution of linear systems. Parallel
computing 32, 2 (2006), 136–156.
[5] Corinne Ancourt and François Irigoin. 1991. Scanning polyhedra with DO loops.
In ACM Sigplan Notices, Vol. 26. ACM, 39–50.
[6] Dale Arden Anderson, John C Tannehill, and Richard H Pletcher. 1984. Computational fluid mechanics and heat transfer. (1984).
[7] Cleve Ashcraft and Roger G Grimes. 1999. SPOOLES: An Object-Oriented Sparse
Matrix Library.. In PPSC.
[8] Michele Benzi, Jane K Cullum, and Miroslav Tuma. 2000. Robust approximate
inverse preconditioning for the conjugate gradient method. SIAM Journal on
Scientific Computing 22, 4 (2000), 1318–1332.
[9] Gilbert Louis Bernstein, Chinmayee Shah, Crystal Lemire, Zachary Devito,
Matthew Fisher, Philip Levis, and Pat Hanrahan. 2016. Ebb: A DSL for Physical
Simulation on CPUs and GPUs. ACM Trans. Graph. 35, 2, Article 21 (May 2016),
12 pages. DOI:https://doi.org/10.1145/2892632
[10] Chun Chen. 2012. Polyhedra scanning revisited. ACM SIGPLAN Notices 47, 6
(2012), 499–508.
[11] Yanqing Chen, Timothy A Davis, William W Hager, and Sivasankaran Rajamanickam. 2008. Algorithm 887: CHOLMOD, supernodal sparse Cholesky factorization and update/downdate. ACM Transactions on Mathematical Software (TOMS)
35, 3 (2008), 22.
[12] Timothy A Davis. 2004. Algorithm 832: UMFPACK V4. 3—an unsymmetricpattern multifrontal method. ACM Transactions on Mathematical Software (TOMS)
30, 2 (2004), 196–199.
[13] Timothy A Davis. 2005. Algorithm 849: A concise sparse Cholesky factorization
package. ACM Transactions on Mathematical Software (TOMS) 31, 4 (2005), 587–
591.
[14] Timothy A Davis. 2006. Direct methods for sparse linear systems. Vol. 2. Siam.
[15] Timothy A Davis. 2011. Algorithm 915, SuiteSparseQR: Multifrontal multithreaded rank-revealing sparse QR factorization. ACM Transactions on Mathematical Software (TOMS) 38, 1 (2011), 8.
,,
[16] Timothy A Davis. 2013. Algorithm 930: FACTORIZE: An object-oriented linear
system solver for MATLAB. ACM Transactions on Mathematical Software (TOMS)
39, 4 (2013), 28.
[17] Timothy A Davis and William W Hager. 2005. Row modifications of a sparse
Cholesky factorization. SIAM J. Matrix Anal. Appl. 26, 3 (2005), 621–639.
[18] Timothy A Davis and William W Hager. 2009. Dynamic supernodes in sparse
Cholesky update/downdate and triangular solves. ACM Transactions on Mathematical Software (TOMS) 35, 4 (2009), 27.
[19] Timothy A Davis and Yifan Hu. 2011. The University of Florida sparse matrix
collection. ACM Transactions on Mathematical Software (TOMS) 38, 1 (2011), 1.
[20] Timothy A Davis and Ekanathan Palamadai Natarajan. 2010. Algorithm 907:
KLU, a direct sparse solver for circuit simulation problems. ACM Transactions
on Mathematical Software (TOMS) 37, 3 (2010), 36.
[21] Ailson P de Moura and Adriano Aron F de Moura. 2013. Newton–Raphson power
flow with constant matrices: a comparison with decoupled power flow methods.
International Journal of Electrical Power & Energy Systems 46 (2013), 108–114.
[22] James W Demmel, Stanley C Eisenstat, John R Gilbert, Xiaoye S Li, and
Joseph WH Liu. 1999. A supernodal approach to sparse partial pivoting. SIAM J.
Matrix Anal. Appl. 20, 3 (1999), 720–755.
[23] James W Demmel, John R Gilbert, and Xiaoye S Li. 1999. An asynchronous
parallel supernodal algorithm for sparse gaussian elimination. SIAM J. Matrix
Anal. Appl. 20, 4 (1999), 915–952.
[24] Richard C Dorf. 2006. Electronics, power electronics, optoelectronics, microwaves,
electromagnetics, and radar. CRC press.
[25] Iain S Duff, Nick IM Gould, John K Reid, Jennifer A Scott, and Kathryn Turner.
1991. The factorization of sparse symmetric indefinite matrices. IMA J. Numer.
Anal. 11, 2 (1991), 181–204.
[26] Iain S Duff and John K Reid. 1983. The multifrontal solution of indefinite sparse
symmetric linear. ACM Transactions on Mathematical Software (TOMS) 9, 3 (1983),
302–325.
[27] Iain S Duff and John Ker Reid. 1996. The design of MA48: a code for the direct
solution of sparse unsymmetric linear systems of equations. ACM Transactions
on Mathematical Software (TOMS) 22, 2 (1996), 187–226.
[28] Yousef El-Kurdi, Maryam Mehri Dehnavi, Warren J. Gross, and Dennis Giannacopoulos. 2015. Parallel finite element technique using Gaussian belief propagation. Computer Physics Communications 193, Complete (2015), 38–48. DOI:
https://doi.org/10.1016/j.cpc.2015.03.019
[29] Gonzalo Exposito-Dominguez, Jose-Manuel Fernandez Gonzalez, Pablo Padilla de
La Torre, and Manuel Sierra-Castaner. 2012. Dual circular polarized steering
antenna for satellite communications in X band. Progress In Electromagnetics
Research 122 (2012), 61–76.
[30] Alan George and Joseph WH Liu. 1979. The design of a user interface for a
sparse matrix package. ACM Transactions on Mathematical Software (TOMS) 5, 2
(1979), 139–162.
[31] Alan George and Joseph W. Liu. 1981. Computer Solution of Large Sparse Positive
Definite. Prentice Hall Professional Technical Reference.
[32] Alan George and Joseph W Liu. 1981. Computer solution of large sparse positive
definite. (1981).
[33] Sarah FF Gibson and Brian Mirtich. 1997. A survey of deformable modeling in
computer graphics. Technical Report. Citeseer.
[34] John R Gilbert and Tim Peierls. 1988. Sparse partial pivoting in time proportional
to arithmetic operations. SIAM J. Sci. Statist. Comput. 9, 5 (1988), 862–874.
[35] Nicholas IM Gould, Jennifer A Scott, and Yifan Hu. 2007. A numerical evaluation
of sparse direct solvers for the solution of large sparse symmetric linear systems
of equations. ACM Transactions on Mathematical Software (TOMS) 33, 2 (2007),
10.
[36] Gaël Guennebaud and Benoit Jacob. 2010. Eigen. URl: http://eigen. tuxfamily. org
(2010).
[37] John A Gunnels, Fred G Gustavson, Greg M Henry, and Robert A Van De Geijn.
2001. FLAME: Formal linear algebra methods environment. ACM Transactions
on Mathematical Software (TOMS) 27, 4 (2001), 422–455.
[38] Anshul Gupta, George Karypis, and Vipin Kumar. 1997. Highly scalable parallel
algorithms for sparse matrix factorization. IEEE Transactions on Parallel and
Distributed Systems 8, 5 (1997), 502–520.
[39] Justin Holewinski, Louis-Noël Pouchet, and P. Sadayappan. 2012. Highperformance Code Generation for Stencil Computations on GPU Architectures.
In Proceedings of the 26th ACM International Conference on Supercomputing (ICS
’12). ACM, New York, NY, USA, 311–320. DOI:https://doi.org/10.1145/2304576.
2304619
[40] Carlo Janna, Massimiliano Ferronato, and Giuseppe Gambolati. 2015. The use
of supernodes in factored sparse approximate inverse preconditioning. SIAM
Journal on Scientific Computing 37, 1 (2015), C72–C94.
[41] Wayne Kelly. 1998. Optimization within a unified transformation framework.
(1998).
[42] David S Kershaw. 1978. The incomplete CholeskyâĂŤconjugate gradient method
for the iterative solution of systems of linear equations. J. Comput. Phys. 26, 1
(1978), 43–65.
,,
Kazem Cheshmi, Shoaib Kamil, Michelle Mills Strout, and Maryam Mehri Dehnavi
[43] Fredrik Kjolstad, Shoaib Kamil, Jonathan Ragan-Kelley, David IW Levin, Shinjiro
Sueda, Desai Chen, Etienne Vouga, Danny M Kaufman, Gurtej Kanwar, Wojciech Matusik, and Saman Amarasinghe. 2016. Simit: A language for physical
simulation. ACM Transactions on Graphics (TOG) 35, 2 (2016), 20.
[44] Xiaoye S Li. 2005. An overview of SuperLU: Algorithms, implementation, and
user interface. ACM Transactions on Mathematical Software (TOMS) 31, 3 (2005),
302–325.
[45] Xiaoye S Li and James W Demmel. 2003. SuperLU_DIST: A scalable distributedmemory sparse direct solver for unsymmetric linear systems. ACM Transactions
on Mathematical Software (TOMS) 29, 2 (2003), 110–140.
[46] Joseph W. H. Liu. 1990. The Role of Elimination Trees in Sparse Factorization.
SIAM J. Matrix Anal. Appl. 11, 1 (Jan. 1990), 134–172. DOI:https://doi.org/10.
1137/0611010
[47] José A Zevallos Luna, Alexandre Siligaris, Cédric Pujol, and Laurent Dussopt.
2013. A packaged 60 GHz low-power transceiver with integrated antennas for
short-range communications.. In Radio and Wireless Symposium. 355–357.
[48] Fabio Luporini, David A Ham, and Paul HJ Kelly. 2016. An algorithm for the
optimization of finite element integration loops. arXiv preprint arXiv:1604.05872
(2016).
[49] Maxim Naumov. 2012. Parallel incomplete-LU and Cholesky factorization in
the preconditioned iterative methods on the GPU. NVIDIA Technical Report
NVR-2012-003 (2012).
[50] M Papadrakakis and N Bitoulas. 1993. Accuracy and effectiveness of preconditioned conjugate gradient algorithms for large and ill-conditioned problems.
Computer methods in applied mechanics and engineering 109, 3-4 (1993), 219–232.
[51] Roger P Pawlowski, John N Shadid, Joseph P Simonis, and Homer F Walker. 2006.
Globalization techniques for Newton–Krylov methods and applications to the
fully coupled solution of the Navier–Stokes equations. SIAM review 48, 4 (2006),
700–721.
[52] Alex Pothen and Sivan Toledo. 2004. Elimination Structures in Scientific Computing. (2004).
[53] Markus Püschel, José M. F. Moura, Jeremy Johnson, David Padua, Manuela Veloso,
Bryan Singer, Jianxin Xiong, Franz Franchetti, Aca Gacic, Yevgen Voronenko,
Kang Chen, Robert W. Johnson, and Nicholas Rizzolo. 2005. SPIRAL: Code
Generation for DSP Transforms. Proceedings of the IEEE, special issue on “Program
Generation, Optimization, and Adaptation” 93, 2 (2005), 232– 275.
[54] Fabien Quilleré, Sanjay Rajopadhye, and Doran Wilde. 2000. Generation of efficient nested loops from polyhedra. International Journal of Parallel Programming
28, 5 (2000), 469–498.
[55] Jonathan Ragan-Kelley, Connelly Barnes, Andrew Adams, Sylvain Paris, Frédo
Durand, and Saman Amarasinghe. 2013. Halide: a language and compiler for
optimizing parallelism, locality, and recomputation in image processing pipelines.
ACM SIGPLAN Notices 48, 6 (2013), 519–530.
[56] Hongbo Rong, Jongsoo Park, Lingxiang Xiang, Todd A Anderson, and Mikhail
Smelyanskiy. 2016. Sparso: Context-driven optimizations of sparse linear algebra.
In Proceedings of the 2016 International Conference on Parallel Architectures and
Compilation. ACM, 247–259.
[57] Olaf Schenk and Klaus Gärtner. 2004. Solving unsymmetric sparse systems of
linear equations with PARDISO. Future Generation Computer Systems 20, 3 (2004),
475–487.
[58] Olaf Schenk, Klaus Gärtner, and Wolfgang Fichtner. 2000. Efficient sparse LU
factorization with left-right looking strategy on shared memory multiprocessors.
BIT Numerical Mathematics 40, 1 (2000), 158–176.
[59] Kai Shen, Tao Yang, and Xiangmin Jiao. 2000. S+: Efficient 2D sparse LU factorization on parallel machines. SIAM J. Matrix Anal. Appl. 22, 1 (2000), 282–305.
[60] Andrew H Sherman. 1978. Algorithms for sparse Gaussian elimination with
partial pivoting. ACM Transactions on Mathematical Software (TOMS) 4, 4 (1978),
330–338.
[61] Jaewook Shin, Mary W Hall, Jacqueline Chame, Chun Chen, Paul F Fischer, and
Paul D Hovland. 2010. Speeding up Nek5000 with autotuning and specialization.
In Proceedings of the 24th ACM International Conference on Supercomputing. ACM,
253–262.
[62] Daniele G Spampinato and Markus Püschel. 2014. A basic linear algebra compiler.
In Proceedings of Annual IEEE/ACM International Symposium on Code Generation
and Optimization. ACM, 23.
[63] Michelle Mills Strout, Alan LaMielle, Larry Carter, Jeanne Ferrante, Barbara
Kreaseck, and Catherine Olschanowsky. 2016. An approach for code generation
in the sparse polyhedral framework. Parallel Comput. 53 (2016), 32–57.
[64] Yuan Tang, Rezaul Alam Chowdhury, Bradley C Kuszmaul, Chi-Keung Luk,
and Charles E Leiserson. 2011. The pochoir stencil compiler. In Proceedings
of the twenty-third annual ACM symposium on Parallelism in algorithms and
architectures. ACM, 117–128.
[65] Ananta Tiwari, Chun Chen, Jacqueline Chame, Mary Hall, and Jeffrey K
Hollingsworth. 2009. A scalable auto-tuning framework for compiler optimization. In Parallel & Distributed Processing, 2009. IPDPS 2009. IEEE International
Symposium on. IEEE, 1–12.
[66] Harmen LA Van Der Spek and Harry AG Wijshoff. 2010. Sublimation: expanding
data structures to enable data instance specific optimizations. In International
Workshop on Languages and Compilers for Parallel Computing. Springer, 106–120.
[67] Nicolas Vasilache, Cédric Bastoul, and Albert Cohen. 2006. Polyhedral code
generation in the real world. In International Conference on Compiler Construction.
Springer, 185–201.
[68] Anand Venkat, Mary Hall, and Michelle Strout. 2015. Loop and data transformations for sparse matrix code. In Proceedings of the 36th ACM SIGPLAN Conference
on Programming Language Design and Implementation. ACM, 521–532.
[69] Anand Venkat, Mahdi Soltan Mohammadi, Jongsoo Park, Hongbo Rong, Rajkishore Barik, Michelle Mills Strout, and Mary Hall. 2016. Automating wavefront
parallelization for sparse matrix computations. In Proceedings of the International
Conference for High Performance Computing, Networking, Storage and Analysis.
IEEE Press, 41.
[70] Anand Venkat, Manu Shantharam, Mary Hall, and Michelle Mills Strout. 2014.
Non-affine extensions to polyhedral code generation. In Proceedings of Annual
IEEE/ACM International Symposium on Code Generation and Optimization. ACM,
185.
[71] Z Xianyi. 2016. OpenBLAS: an optimized BLAS library. (2016).
| 6cs.PL
|
arXiv:1506.00716v1 [cs.CE] 2 Jun 2015
Tackling Exascale Software Challenges in
Molecular Dynamics Simulations
with GROMACS
Szilárd Páll1,2,†
Carsten Kutzner3
Mark James Abraham1,2,†
Berk Hess1,2
Erik Lindahl1,2,4
Feb 12, 2015
1 Science
for Life Laboratory, Stockholm and Uppsala, 171 21 Stockholm, Swe-
den,
2 Dept. Theoretical Physics, KTH Royal Institute of Technology, 10691 Stockholm,
Sweden,
3 Theoretical and Computational Biophysics Dept., Max Planck Institute for Biophysical Chemistry, Am Fassberg 11, 37077 Göttingen, Germany
4 Center for Biomembrane Research, Dept. Biochemistry & Biophysics, Stockholm University, SE-10691 Stockholm, Sweden
Address correspondence to: Erik Lindahl, [email protected]
The final publication is available at
http://link.springer.com/chapter/10.1007%2F978-3-319-15976-8_1
†
The authors contributed equally.
1
Abstract
GROMACS is a widely used package for biomolecular simulation, and over the
last two decades it has evolved from small-scale efficiency to advanced heterogeneous acceleration and multi-level parallelism targeting some of the largest
supercomputers in the world. Here, we describe some of the ways we have been
able to realize this through the use of parallelization on all levels, combined
with a constant focus on absolute performance. Release 4.6 of GROMACS uses
SIMD acceleration on a wide range of architectures, GPU offloading acceleration,
and both OpenMP and MPI parallelism within and between nodes, respectively.
The recent work on acceleration made it necessary to revisit the fundamental
algorithms of molecular simulation, including the concept of neighborsearching,
and we discuss the present and future challenges we see for exascale simulation
- in particular a very fine-grained task parallelism. We also discuss the software
management, code peer review and continuous integration testing required for
a project of this complexity.
2
Introduction
Molecular Dynamics simulation of biological macromolecules has evolved from
a narrow statistical-mechanics method into a widely applied biophysical research tool that is used outside theoretical chemistry. Supercomputers are now
as important as centrifuges or test tubes in chemistry. However, this success also
considerably raises the bar for molecular simulation implementations - it is no
longer sufficient to reproduce experimental results or e.g. show proof-of-concept
relative scaling. To justify the substantial supercomputing resources required by
many computational chemistry projects the most important focus today is simply absolute simulation performance and the scientific results achieved. Exascale
computing has potential to take simulation to new heights, but the combination
of challenges that face software preparing for deployment at the exascale to deliver these results are unique in the history of software. The days of simply
buying new hardware with a faster clock rate and getting shorter times to solution with old software are gone. The days of running applications on a single
core are gone. The days of heterogeneous processor design to suit floating-point
computation are back again. The days of performance being bounded by the
time taken for floating-point computations are ending fast. The need to design
with multi-core and multi-node parallelization in mind at all points is here to
stay, which also means Amdahl’s law 3 is more relevant than ever.1
A particular challenge for biomolecular simulations is that the computational
1 Amdahl’s law gives a model for the expected (and maximum) speedup of a program
when parallelized over multiple processors with respect to the serial version. It states
that the achievable speedup is limited by the sequential part of the program.
2
problem size is fixed by the geometric size of the protein and the atomic-scale
resolution of the model physics. Most life science problems can be reduced to
this size (or smaller). It is possible to simulate much larger systems, but it is
typically not relevant. Second, the timescale of dynamics involving the entire
system increases much faster than the length scale, due to the requirement of
sampling the exponentially larger number of ensemble microstates. This means
that weak scaling is largely irrelevant for life science; to make use of increasing
amounts of computational resources to simulate these systems, we have to rely
either on strong-scaling software engineering techniques, or ensemble simulation techniques.
The fundamental algorithm of molecular dynamics assigns positions and
velocities to every particle in the simulation system, and specifies the model
physics that governs the interactions between particles. The forces can then be
computed, which can be used to update the positions and velocities via Newton’s second law, using a given finite time step. This numerical integration
scheme is iterated a large number of times, and it generates a series of samples from the thermodynamic ensemble defined by the model physics. From
these samples, observations can be made that confirm or predict experiment.
Typical model physics have many components to describe the different kinds
of bonded and non-bonded interactions that exist. The non-bonded interactions
between particles model behaviour like van der Waals forces, or Coulomb’s law.
The non-bonded interactions are the most expensive aspects of computing the
forces, and the subject of a very large amount of research, computation and
optimization.
Historically, the GROMACS molecular dynamics simulation suite has aimed
at being a general-purpose tool for studying biomolecular systems, such as
shown in Fig. 1. The development of the simulation engine focused heavily
on maximizing single-core floating-point performance of its innermost compute
kernels for non-bonded interactions. These kernels typically compute the electrostatic and van der Waals forces acting on each simulation particle from its
interactions with all other inside a given spherical boundary. These kernels were
first written in C, then FORTRAN, and later optimized in assembly language,
mostly for commodity x86-family processors, because the data dependencies
of the computations in the kernels were too challenging for C or FORTRAN
compilers (then or now). The kernels were also specialized for interactions
within and between water molecules, because of the prevalence of such interactions in biomolecular simulations. From one point-of-view, this extensive use
of interaction-specific kernels can be seen as a software equivalent of applicationspecific integrated circuits.
Recognizing the need to build upon this good work by coupling multiple
processors, GROMACS 4.0 14 introduced a minimal-communication neutral territory domain-decomposition (DD) algorithm, 7,8 with fully dynamic load balancing. This spatial decomposition of the simulation volume created high-level
3
Figure 1: A typical GROMACS simulation system, featuring the ligand-gated ion-channel membrane protein
GLIC (colored), embedded in a lipid membrane (grey). The whole system is solvated in water (not shown),
giving a total of around 145,000 atoms. Image created with VMD. 15
data parallelism that was effective for near-linear scaling of the computation at
around 400 atoms per core. The DD implementation required the use of MPI
for message-passing parallel constructs. However, the needs of many simulation
users can be met within a single node, 23 and in that context the implementation
overhead of MPI libraries was too high, not to mention it is difficult to employ in distributed computing. In GROMACS 4.5, 20 we implemented a multithreaded MPI library with the necessary subset of the MPI API. The library has
both POSIX and Windows threads back-ends (hence called thread-MPI) and uses
highly efficient hardware-supported atomic and lock-free synchronization primitives. This allows the existing DD implementation to work across multiple cores
of a single node without depending on any external MPI library.
However, the fundamental limitation remained of a one-to-one mapping of
MPI ranks to cores, and to domains. On the one hand, there is always a limit to
how small a spatial domain can be, which will limit the number of domains the
simulation box can be decomposed into, which in turn limits the number of cores
that a parallelization with such a mapping can utilize. On the other hand, the
one-to-one domains to cores mapping is cache-friendly as it creates independent
data sets so that cores sharing caches can act without conflict, but the size of
the volume of data that must be communicated so that neighboring domains act
coherently grows rapidly with the number of domains. This approach is only
scalable for a fixed problem size if the latency of communication between all
cores is comparable and the communication book-keeping overhead grows only
linearly with the number of cores. Neither is true, because network latencies
4
are orders of magnitude higher than shared-cache latencies. This is clearly a
major problem for designing for the exascale, where many cores, many nodes
and non-uniform memory and communication latencies will be key attributes.
The other important aspect of the target simulations for designing for strong
scaling is treating the long-range components of the atomic interactions. Many
systems of interest are spatially heterogeneous on the nanometer scale (e.g. proteins embedded in membranes and solvated in water), and the simulation artefacts caused by failing to treat the long-range effects are well known. The de facto
standard for treating the long-range electrostatic interactions has become the
smooth particle-mesh Ewald (PME) method, 12 whose cost for N atoms scales
as N log( N ). A straightforward implementation where each rank of a parallel
computation participates in an equivalent way leads to a 3D Fast Fourier Transform (FFT) that communicates globally. This communication quickly limits the
strong scaling. To mitigate this, GROMACS 4.0 introduced a multiple-program
multiple-data- (MPMD) implementation that dedicates some ranks to the FFT
part; now only those ranks do all-to-all FFT communication. GROMACS 4.5
improved further by using a 2D pencil decomposition 11,16 in reciprocal space,
within the same MPMD implementation. This coarse-grained task parallelism
works well on machines with homogeneous hardware, but it is harder to port to
accelerators or combine with RDMA constructs.
The transformation of GROMACS needed to perform well on exascale-level
parallel hardware began after GROMACS 4.5. This requires radical algorithm
changes, and better use of parallelization constructs from the ground up, not as
an afterthought. More hands are required to steer the project, and yet the old
functionality written before their time must generally be preserved. Computer
architectures are evolving rapidly, and no single developer can know the details
of all of them. In the following sections we describe how we are addressing some
of these challenges, and our ongoing plans for addressing others.
3
3.1
Handling exascale software challenges: algorithms and parallelization schemes
Multi-level parallelism
Modern computer hardware is not only parallel, but exposes multiple levels
of parallelism depending on the type and speed of data access and communication capabilities across different compute elements. For a modern superscalar CPU such as Intel Haswell, even a single core is equipped with 8 different execution ports, and it is not even possible to buy a single-core chip. Add
hardware threads, complex communication crossbars, memory hierarchies, and
caches larger than hard disks from the 1990s. This results in a complex hierarchical organization of compute and communication/network elements from SIMD
5
units and caches to network topologies, each level in the hierarchy requiring a
different type of software parallelization for efficient use. HPC codes have traditionally focused on only two levels of parallelism: intra-node and inter-node.
Such codes typically rely solely on MPI parallelization to target parallelism on
multiple levels: both intra-socket, intra-node, and inter-node. This approach
had obvious advantages before the multi-core and heterogeneous computing
era when improvements came from CPU frequency scaling and evolution of
interconnect. However, nowadays most scientific problems require complex parallel software architecture to be able use petaflop hardware efficiently and going
toward exascale this is becoming a necessity. This is particularly true for molecular dynamics which requires reducing the wall-time per iteration to improve
simulation performance.
On the lowest level, processors typically contain SIMD (single instruction
multiple data) units which offer fine-grained data-parallelism through silicon
dedicated to executing a limited set of instructions on multiple, currently typically 4-16, data elements simultaneously. Exploiting this low-level and finegrained parallelism has become crucial for achieving high performance, especially with new architectures like AVX and Intel MIC supporting wide SIMD.
One level higher, multi-core CPUs have become the standard and several architectures support multiple hardware threads per core. Hence, typical multi-socket
SMP machines come with dozens of cores capable of running 2-4 threads each
(through simultaneous multi-threading, SMT, support). Simply running multiple processes (MPI ranks) on each core or hardware thread is typically less
efficient than multi-threading. Achieving strong scaling in molecular dynamics
requires efficient use of the cache hierarchy, which makes the picture even more
complex. On the other hand, a chip cannot be considered a homogeneous cluster either. Accelerator coprocessors like GPUs or Intel MIC, often referred to as
many-core, add another layer of complexity to the intra-node parallelism. These
require fine-grained parallelism and carefully tuned data access patterns, as well
as special programming models. Current accelerator architectures like GPUs
also add another layer of interconnect in form of PCIe bus (Peripheral Component Interconnect Express) as well as a separate main memory. This means that
data movement across the PCIe link often limits overall throughput. Integration
of traditional latency-oriented CPU cores with throughput-oriented cores like
those in GPUs or MIC accelerators is ongoing, but the cost of data movement
between the different units will at least for the foreseeable future be a factor that
needs to be optimized for.
Typical HPC hardware exhibits non-uniform memory access (NUMA) behavior on the node level: accessing data from different CPUs or cores of CPUs has
a non-uniform cost. We started multithreading trials quite early with the idea
of easily achieving load balancing, but the simultaneous introduction of NUMA
suddenly meant a processor resembled a cluster internally. Indiscriminately accessing memory across NUMA nodes will frequently lead to performance that
6
is lower than for MPI. Moreover, the NUMA behavior extends to other compute
and communication components: the cost of communicating with an accelerator
or through a network interface typically depends on the intra-node bus topology
and requires special attention. On the top level, the interconnect links together
compute nodes into a network topology. A side-effect of the multi-core evolution
is that, while the network capacity (latency and bandwidth) per compute node
has improved, the typical number of CPU cores they serve has increased faster;
the capacity available per core has decreased substantially.
In order to exploit the capabilities of each level of hardware parallelism,
a performance-oriented application needs to consider multiple levels of parallelism: SIMD parallelism for maximizing single-core/thread performance; multithreading to exploit advantages of multi-core and SMT (e.g. fast data sharing); inter-node communication-based parallelism (e.g. message passing with
MPI); and heterogeneous parallelism by utilizing both CPUs and accelerators
like GPUs.
Driven by this evolution of hardware, we have initiated a re-redesign of
the parallelization in GROMACS. In particular, recent efforts have focused on
improvements targeting all levels of parallelization: new algorithms for wide
SIMD and accelerator architectures, a portable and extensible SIMD parallelization framework, efficient multi-threading throughout the entire code, and an
asynchronous offload-model for accelerators. The resulting multi-level parallelization scheme implemented in GROMACS 4.6 is illustrated in Fig. 2. In the
following sections, we will give an overview of these improvements, highlighting
the advances they provide in terms of making efficient use of current petascale
hardware, as well as in paving the road towards exascale computing.
3.2
SIMD parallelism
All modern CPU and GPU architectures use SIMD-like instructions to achieve
high flop rates. Any computational code that aims for high performance will
have to make use of SIMD. For very regular work, such as matrix-vector multiplications, the compiler can generate good SIMD code, although manually tuned
vendor libraries typically do even better. But for irregular work, such as shortrange particle-particle non-bonded interactions, the compiler usually fails since
it cannot control data structures. If you think your compiler is really good at
optimizing, it can be an eye-opening experience to look at the raw assembly
instructions actually generated. In GROMACS, this was reluctantly recognized
a decade ago and SSE and Altivec SIMD kernels were written manually in assembly. These kernels were, and still are, extremely efficient for interactions
involving water molecules, but other interactions do not parallelize well with
SIMD using the standard approach of unrolling a particle-based Verlet-list. 25
It is clear that a different approach is needed in order to use wide SIMD execution units like AVX or GPUs. We developed a novel approach, where particles
7
short-ranged
interactions
• cluster-pair search
• bonded interactions
• integration of forces
SIMD
floating-point
operations
OpenMP threads
CPU
cores
serial
FFT
library
Nvidia GPUs
PP
nodes
• spread/gather
• 3D-FFT
• solve
MPI
ranks
PME
nodes
OpenMP threads
LCD-Pro
READER
AUDIO
PHONE
MIC
LINE-IN
CARD
SATA
USB
EJECT
DVD-RW
POWER
MENU
F7 F8
Esc
`
!
F4
F3
F1 F2
~
1
@
Tab
2
#
+
SELECT
F11 F12
F9 F10
+
_
-
ScrollLock
PrintScrn
SysRq
Pause
NumLock
Break
Insert
=
|
\
}
Home
ScrollLock
CapsLock
NumLock
PageUp
Delete
End
PageDown
/
*
7
Shift
8
Home
]
) 0
{ [
( 9
F5 F6
" '
* 8
O P
& 7
L : ;
?
^ 6
U I
% 5
J K M < , > . /
$
T Y
3 4 E R
G H
N
D F C V B
Q W
A S
CapsLock
Z X
9
4
PgUp
+
6
5
1
End
2
3
0
PgDn
Ins
.
Enter
Del
Ctrl
Alt Gr
Alt
Shift
Ctrl
workstations
ensembles
of simulations
compute
clusters
simulations
Figure 2: Illustration of multi-level parallelism in GROMACS 4.6. This exploits several kinds of finegrained data parallelism, a multiple-program multiple-data (MPMD) decomposition separating the shortrange particle-particle (PP) and long-range Particle Mesh Ewald (PME) force calculation algorithms, coarsegrained data parallelism with domain-decomposition (DD) over MPI ranks (implemented either on singlenode workstations or compute clusters), and ensembles of related simulations scheduled e.g. by a distributed
computing controller.
are grouped into spatial clusters containing fixed number of particles. 17 First,
the particles are placed on a grid in the x and y dimensions, and then binned
in the z dimension. This efficiently groups particles that are close in space, and
permits the construction of a list of clusters, each containing exactly M particles.
A list is then constructed of all those cluster pairs containing particles that may
be close enough to interact. This list of pairs of interacting clusters is reused over
multiple successive evaluations of the non-bonded forces. The list is constructed
with a buffer to prevent particle diffusion corrupting the implementation of the
model physics.
The kernels that implement the computation of the interactions between two
clusters i and j use SIMD load instructions to fill vector registers with copies
of the positions of all M particles in i. The loop over the N particles in j is
unrolled according to the SIMD width of the CPU. Inside this loop, SIMD load
instructions fill vector registers with positions of all N particles from the j cluster.
This permits the computation of N interactions between an i and all j particles
simultaneously, and the computation of M × N interactions in the inner loop
without needing to load particle data. With wide SIMD units it is efficient to
process more than one j cluster at a time.
M, N and the number of j clusters to process can be adjusted to suit the
underlying characteristics of the hardware. Using M=1 and N=1 recovers the
original Verlet-list algorithm. On CPUs, GROMACS uses M=4 and N=2, 4 or
8
8, depending on the SIMD width. On NVIDIA GPUs, we use M=8 and N=4
to calculate 32 interactions at once with 32 hardware threads executing in lockstep. To further improve the ratio of arithmetic to memory operations when
using GPUs, we add another level of hierarchy by grouping 8 clusters together.
Thus we store 64 particles in shared memory and calculate interactions with
about half of these for every particle in the cluster-pair list.
The kernel implementations reach about 50% of the peak flop rate on all
supported hardware, which is very high for MD. This comes at the cost of calculating about twice as many interactions as required; not all particle pairs in all
cluster pairs will be within the cut-off at each time step, so many interactions are
computed that are known to produce a zero result. The extra zero interactions
can actually be put to use as an effective additional pair list buffer additionally to the standard Verlet list buffer. As we have shown here, this scheme is
flexible, since N and M can be adapted to current and future hardware. Most
algorithms and optimization tricks that have been developed for particle-based
pair lists can be reused for the cluster-pair list, although many will not improve
the performance.
The current implementation of the cluster-based non-bonded algorithm already supports a wide range of SIMD instruction sets and accelerator architectures: SSE2, SSE4.1, AVX (256-bit and 128-bit with FMA), AVX2, BG/Q QPX, Intel MIC (LRBni), NVIDIA CUDA. An implementation on a field-programmable
gate array (FPGA) architecture is in progress.
3.2.1
Multi-threaded parallelism
Before GROMACS 4.6, we relied mostly on MPI for both inter-node and intranode parallelization over CPU cores. For MD this has worked well, since there is
little data to communicate and at medium to high parallelization all data fits in
L2 cache. Our initial plans were to only support OpenMP parallelization in the
separate Particle Mesh Ewald (PME) MPI ranks. The reason for using OpenMP
in PME was to reduce the number of MPI ranks involved in the costly collective
communication for the FFT grid transpose. This 3D-FFT is the only part of
the code that involves global data dependencies. Although this indeed greatly
reduced the MPI communication cost, it also introduced significant overhead.
GROMACS 4.6 was designed to use OpenMP in all compute-intensive parts
of the MD algorithm.2 Most of the algorithms are straightforward to parallelize
using OpenMP. These scale very well, as Fig. 3 shows. Cache-intensive parts of
the code like performing domain decomposition, or integrating the forces and
velocities show slightly worse scaling. Moreover, the scaling in these parts tends
to deteriorate with increasing number of threads in an MPI rank – especially with
2 At
the time of that decision, sharing a GPU among multiple MPI ranks was inefficient, so the only efficient way to use multiple cores in a node was with OpenMP within
a rank. This constraint has since been relaxed.
9
large number of threads in a rank, and when teams of OpenMP threads cross
NUMA boundaries. When simulating at high ratios of cores/particles, each MD
step can take as little as a microsecond. There are many OpenMP barriers used
in the many code paths that are parallelized with OpenMP, each of which takes
a few microseconds, which can be costly.
80
OpenMP
MPI
70
MPI+OpenMP (two ranks)
performance (ns/day)
60
50
40
30
20
10
0
0
2
4
6
8
10
12
14
16
#cores
Figure 3: Comparison of single-node simulation performance using MPI, OpenMP, and combined
MPI+OpenMP parallelization. The OpenMP multi-threading (blue) achieves the highest performance and
near linear scaling up to 8 threads. It only deteriorates when threads on OpenMP regions need to communicate across the system bus. In contrast, the MPI-only runs (red) that require less communication scale
well across sockets. Combining MPI and OpenMP parallelization with two ranks and varying the number of
threads (green) results in worse performance due to the added overhead of the two parallelization schemes.
Simulations were carried out on a dual-socket node with 8-core Intel Xeon E5-2690 (2.9 GHz Sandy Bridge).
Input system: RNAse protein solvated in a rectangular box, 24k atoms, PME electrostatics, 0.9 nm cut-off.
Accordingly, the hybrid MPI + OpenMP parallelization is often slower than
an MPI-only scheme as Fig. 3 illustrates. Since PP (particle-particle) ranks only
do low-volume local communication, the reduction in MPI communication from
using the hybrid scheme is apparent only at high parallelization. There, MPIonly parallelization (e.g. as in GROMACS 4.5) puts a hard upper limit on the
number of cores that can be used, due to algorithmic limits on the spatial domain
size, or the need to communicate with more than one nearest neighbor. With the
hybrid scheme, more cores can operate on the same spatial domain assigned to
an MPI rank, and there is no longer a hard limit on the parallelization. Strong
scaling curves now extend much further, with a more gradual loss of parallel
efficiency. An example is given in Fig. 4, which shows a membrane protein
system scaling to twice as many cores with hybrid parallelization and reach
double the peak performance of GROMACS 4.5. In some cases, OpenMP-only
parallelization can be much faster than MPI-only parallelization if the load for
each stage of the force computation can be balanced individually. A typical
example is a solute in solvent, where the solute has bonded interactions but
the solvent does not. With OpenMP, the bonded interactions can be distributed
equally over all threads in a straightforward manner.
10
ns/day
300
200
MPI only
MPI + 2 OpenMP
MPI + 4 OpenMP
GMX4.5, MPI only
100
0
0
192
384
576
#cores
768
960
Figure 4: Improvements in strong scaling performance since GROMACS 4.5, using the M × N kernels and
OpenMP parallelization in GROMACS 4.6. The plot shows simulation performance in ns/day for different
software versions and parallelization schemes. Performance with one core per MPI rank is shown for GROMACS 4.5 (purple) and 4.6 (black). Performance with GROMACS 4.6 is shown using two (red) and four
(green) cores per MPI rank using OpenMP threading within each MPI rank. Simulations were carried out on
the Triolith cluster at NSC, using two 8-core Intel E5-2660 (2.2 GHz Sandy Bridge) processors per node and
FDR Infiniband network. The test system is the GLIC membrane protein shown in Fig. 1 (144,000 atoms, PME
electrostatics.)
Heterogeneous parallelization
Heterogeneous architectures combine multiple types of processing units, typically latency- and throughput-oriented cores – most often CPUs and accelerators
like GPUs, Intel MIC, or FPGAs. Many-core accelerator architectures have been
become increasingly popular in technical and scientific computing mainly due
to their impressive raw floating point performance. However, in order to efficiently utilize these architectures, a very high level of fine-grained parallelism
is required. The massively parallel nature of accelerators, in particular GPUs, is
both an advantage as well as a burden on the programmer. Since not all tasks
are well suited for execution on the accelerators this often leads to additional
challenges for workload distribution and load balancing. Moreover, current heterogeneous architectures typically use a slow PCIe bus to connect the hardware
elements like CPUs and GPUs and move data between the separate global memory of each. This means that explicit data management is required. This adds a
further latency overhead to challenge algorithms like MD that already face this
as a parallelization bottle-neck.
GPU accelerators were first supported experimentally in GROMACS with
the OpenMM library, 10 which was used as a black box to execute the entire
simulation on the GPU. This meant that only a fraction of the diverse set of
GROMACS algorithms were supported and simulations were limited to singleGPU use. Additionally, while OpenMM offered good performance for implicitsolvent models, the more common type of runs showed little speedup (and in
some cases slowdown) over the fast performance on multi-core CPUs, thanks to
the highly tuned SIMD assembly kernels.
With this experience, we set out to provide native GPU support in GRO-
11
MACS 4.6 with a few important design principles in mind. Building on the
observation that highly optimized CPU code is hard to beat, our goal was to ensure that all compute resources available, both CPU and accelerators, are utilized
to the greatest extent possible. We also wanted to ensure that our heterogeneous
GPU acceleration supported most existing features of GROMACS in a single
code base to avoid having to reimplement major parts of the code for GPU-only
execution. This means that the most suitable parallelization is the offload model,
which other MD codes have also employed successfully. 9,18 As Fig. 5 illustrates,
we aim to execute the compute-intensive short-range non-bonded force calculation on GPU accelerators, while the CPU computes bonded and long-range
electrostatics forces, because the latter are communication intensive.
The newly designed future-proof SIMD-oriented algorithm for evaluating
non-bonded interactions with accelerator architectures in mind has been discussed already. It is highly efficient at expressing the fine-grained parallelism
present in the pair-force calculation. Additionally, the atom cluster-based approach is designed for data reuse which is further emphasized by the supercluster grouping. As a result, our CUDA implementation is characterized by a
high ratio of arithmetic to memory operations which allows avoiding memory
bottlenecks. These algorithmic design choices and the extensive performance
tuning led to strongly instruction-latency bound CUDA non-bonded kernels, in
contrast to most traditional particle-based GPU algorithms which are reported
to be memory bound 4,9 . Our CUDA GPU kernels also scale well, reaching peak
pair-force throughput already around 20,000 particles per GPU.
In contrast to typical data-parallel programming for homogeneous CPU-only
machines, heterogeneous architectures require additional code to manage task
scheduling and concurrent execution on the different compute elements, CPU
cores and GPUs in the present case. This is a complex component of our heterogeneous parallelization which implements the data- and control-flow with
the main goal of maximizing the utilization of both CPU and GPU by ensuring
optimal CPU-GPU execution overlap.
We combine a set of CPU cores running OpenMP threads with a GPU. As
shown in Fig. 5, the pair-lists required for the non-bonded computation are prepared on the CPU and transferred to the GPU where a pruning step is carried
out after which the lists are reused for up to 100 iterations. The extreme floatingpoint power of GPUs makes it feasible to use the much larger buffers required
for this. The transfer of coordinates, charges, and forces as well as compute
kernels are launched asynchronously as soon as data becomes available on the
CPU. This ensures overlap of CPU and GPU computation. Additional effort
has gone into maximizing overlap by reducing the wall-time of CPU-side nonoverlapping program parts through SIMD parallelization (in pair search and constraints) and efficient multi-threaded algorithms allowing GROMACS to achieve
a typical CPU-GPU overlap of 60-80%.
This scheme naturally extends to multiple GPUs by using the existing ef-
12
CPU
GPU
CUDA
Pair search,
transfer list
Domain decomp.
transfer x, q
Bonded F
MD step
Pair search & domain-decomposition: every 20-100 iterations
OpenMP
threads
offload
Non-bonded F
Non-bonded F
&
Pair-list
pruning
PME
Integration
transfer F, E
Constraints
Figure 5: GROMACS heterogeneous parallelization using both CPU and GPU resources during each simulation time-step. The compute-heavy non-bonded interactions are offloaded to the GPU, while the CPU is
responsible for domain-decomposition book-keeping, bonded force calculation, and lattice summation algorithms. The diagram shows tasks carried out during a GPU-accelerated normal MD step (black arrows) as
well as a step which includes the additional pair-search and domain-decomposition tasks are carried out (blue
arrows). The latter, as shown above in blue, also includes an additional transfer, and the subsequent pruning
of the pair list as part of the non-bonded kernel.
Source: http://dx.doi.org/10.6084/m9.figshare.971161. Reused under CC-BY; retrieved 22:15, March 23, 2014 (GMT).
13
ficient neutral-territory do main-decomposition implemented using MPI parallelization. By default, we assign computation on each domain to a single GPU
and a set of CPU cores. This typically means decomposing the system into as
many domains as GPUs used, and running as many MPI ranks per node as
GPUs in the node. However, this will often require to run a large number of
OpenMP threads in a rank (8-16 or even more with a single GPU per node),
potentially spanning across multiple NUMA domains. As explained in the previous section, this will lead to suboptimal multi-threaded scaling – especially
affecting cache-intensive algorithms outside the CPU-GPU overlap region. To
avoid this, multiple MPI ranks can share a GPU, which reduces the number of
OpenMP threads per rank.
The heterogeneous acceleration in GROMACS delivers 3-4x speedup when
comparing CPU only with CPU-GPU runs. Moreover, advanced features like
arbitrary simulation box shapes and virtual interaction sites are all supported
(Fig. 6). Even though the overhead of managing an accelerator is non-negligible,
GROMACS 4.6 shows great strong scaling in GPU accelerated runs reaching 126
atoms/core (1260 atoms/GPU) on common simulation systems (Fig. 7).
Based on a similar parallelization design, the upcoming GROMACS version
will also support the Intel MIC accelerator architecture. Intel MIC supports native execution of standard MPI codes using the so-called symmetric mode, where
the card is essentially treated as a general-purpose multi-core node. However,
as MIC is a highly parallel architecture requiring fine-grained parallelism, many
parts of typical MPI codes will be inefficient on these processors. Hence, efficient
utilization of Xeon Phi devices in molecular dynamics – especially with typical
bio-molecular simulations and strong-scaling in mind - is only possible by treating them as accelerators. Similarly to GPUs, this means a parallelization scheme
based on offloading only those tasks that are suitable for wide SIMD and highly
thread-parallel execution to MIC.
3.3
Ensemble simulations
The performance and scaling advances in GROMACS (and many other programs) have made it efficient to run simulations that simply were too large only
a few years ago. However, infrastructures such as the European PRACE provide
access only to problems that scale to thousands of cores. This used to be an impossible barrier for biomolecular dynamics on anything but ridiculously large
systems when an implementation could only run well with hundreds of particles per core. Scaling has improved, but the number of computational units in
supercomputers is growing even faster. There are now multiple machines in the
world that reach roughly a million cores. Under ideal conditions, GROMACS
can scale to levels where each PP rank handles 40 atoms, but there are few if any
concrete biological problems that require 40 million atoms without corresponding increases in the number of samples generated. Even in the theoretical case
14
Xeon X5650
Xeon X5650+Tesla C2075
2x Xeon X5650
2x (Xeon X5650+Tesla C2075)
21.2
53.5
cubic box
38.0
84.8
20.5
56.3
cubic box, NPT
35.9
88.6
28.4
77.3
dodec box
49.3
100.7
63.7
169.1
dodec box + vsites
110.0
202.4
20.8
cubic RF 1.1
40.7
38.9
70.0
30.4
55.4
55.6
cubic RF 0.9
91.2
0
50
100
150
200
250
ns/day
Figure 6: An important feature of the current heterogeneous GROMACS GPU implementation is that it works,
and works efficiently, in combination with most other features of the software. GPU simulations can employ
domain decomposition, non-standard boxes, pressure scaling, and virtual interaction sites to significantly
improve the absolute simulation performance compared to the baseline. Simulation system: RNAse protein
solvated in rectangular (24 K atoms) and rhombic dodecahedron (16.8 k atoms) box, PME electrostatics, cutoff 0.9 nm. Hardware: 2x Intel Xeon E5650 (2.67 GHz Westmere), 2x NVIDIA Tesla C2070 (Fermi) GPU
accelerators.
Figure 7: Strong scaling of GROMACS 4.6 on the HYDRA heterogeneous GPU-equipped machine in Garching,
Germany. Grey lines indicate linear scaling. The hybrid version of GROMACS scales very well and achieves
impressive absolute performance for both small and large systems. For the smaller systems, peak performance
is achieved with 150 atoms per core, and the larger systems achieve sustained effective flop rate of 0.2 petaflops
(only counting the number of useful floating-point operations, not the total).
Simulation systems: typical production systems of 81k atoms (circles), 2M atoms (stars), and 12M atoms
(triangles) in size. Hardware (per node): 2 10-core Xeon E5-2680v2 (2.8 GHz Ivy Bridge), 2 NVIDIA K20X,
InfiniBand FDR14 (4 x 14 Gb/s) network.
15
where we could improve scaling to the point where each core only contains a
single atom, the simulation system would still be almost an order of magnitude
larger than the example in Fig. 1.
To adapt to this reality, researchers are increasingly using large ensembles of
simulations, either to simply sample better, or new algorithms such as replica
exchange simulation, 24 Markov state models, 22 or milestoning 13 that analyze
and exchange data between multiple simulations to improve overall sampling.
In many cases, this achieves as much as two-fold superscaling, i.e., an ensemble of 100 simulations running on 10 nodes each might provide the same sampling efficiency as a single simulation running on 2000 cores. To automate this,
GROMACS has been co-developed with a new framework for Parallel Adaptive
Molecular Dynamics called Copernicus. 19 Given a set of input structures and
sampling settings, this framework automatically starts a first batch of sampling
runs, makes sure all simulations complete (with extensive support for checkpointing and restarting of failed runs), and automatically performs the adaptive
step data analysis to decide what new simulations to start in a second generation. The current ensemble sampling algorithms scale to hundreds or thousands
of parallel simulations (each using up to thousands of cores even for small systems). For the first time in many years, molecular dynamics might actually be
able to use all the cores available on next-generation supercomputers rather than
constantly being a generation or two behind.
3.4
Multi-level load balancing
Achieving strong scaling to a higher core count for a fixed-size problem requires
careful consideration of load balance. The advantage provided by spatial DD
is one of data locality and reuse, but if the distribution of computational work
is not homogeneous then more care is needed. A typical membrane protein
simulation is dominated by
• water, which is usually treated with a rigid 3-point model,
• a lipid membrane, whose alkyl tails are modeled by particles with zero
partial charge and bonds of constrained length, and
• a protein, which is modeled with a backbone of fixed-length bonds that
require a lengthy series of constraint calculations, as well as partial charge
on all particles.
These problems are well known, and are addressed in the GROMACS DD
scheme via automatic dynamic load balancing that distributes the spatial volumes unevenly according to the observed imbalance in compute load. This approach has limitations because it works at the level of DD domains that must
map to MPI ranks, so cores within the same node or socket have unnecessary
16
copies of the same data. We have not yet succeeded in developing a highly effective intra-rank decomposition of work to multiple cores. We hope to address
this via intra-node or intra-socket task parallelism.
One advantage of the PME algorithm as implemented in GROMACS is that it
is possible to shift the computational workload between the real- and reciprocalspace parts of the algorithm at will. This makes it possible to write code that
can run optimally at different settings on different kinds of hardware. The performance of the compute, communication and bookkeeping parts of the overall
algorithm vary greatly with the characteristics of the hardware that implements
it, and with the properties of the simulation system studied. For example, shifting compute work from reciprocal to real space to make better use of an idle
GPU increases the volume that must be communicated during DD, while lowering the required communication volume during the 3D FFTs. Evaluating how
best to manage these compromises can only happen at runtime.
The MPMD version of PME is intended to reduce the overall communication
cost on typical switched networks by minimizing the number of ranks participating in the 3D FFTs. This requires generating a mapping between PME and
non-PME ranks and scheduling data transfer to and from them. However, on
hardware with relatively efficient implementations of global communication, it
can be advantageous to prefer the SPMD implementation because it has more
regular communication patterns. 2 The same may be true on architectures with
accelerators, because the MPMD implementation makes no use of the accelerators on the PME ranks. The performance of both implementations is limited by
lack of overlap of communication and computation.
Attempts to use low-latency partitioned global address space (PGAS) methods that require single-program multiple-data (SPMD) approaches are particularly challenged, because the gain from any decrease in communication latency
must also overcome the overall increase in communication that accompanies the
MPMD-to-SPMD transition. 21 The advent of implementations of non-blocking
collective (NBC) MPI routines is promising if computation can be found to overlap with the background communication. The most straightforward approach
would be to revert to SPMD and hope that the increase in total communication
cost is offset by the gain in available compute time, however, the available performance is still bounded by the overall cost of the global communication. Finding
compute to overlap with the NBC on the MPMD PME ranks is likely to deliver
better results. Permitting PME ranks to execute kernels for bonded and/or nonbonded interactions from their associated non-PME ranks is the most straightforward way to achieve this overlap. This is particularly true at the scaling limit,
where the presence of bonded interactions is one of the primary problems in
balancing the compute load between the non-PME ranks.
The introduction of automatic ensemble computing introduces another layer
of decomposition, by which we essentially achieve MSMPMD parallelism: Multiplesimulation (ensemble), multiple-program (direct/lattice space), and multiple-
17
data (domain decomposition).
3.5
Managing the long-range contributions at exascale
A promising candidate for exascale-level biomolecular simulations is the use
of suitable implementations of fast-multipole methods such as ExaFMM. 6,27 At
least one implementation of FMM-based molecular dynamics running on 100,000
cores has been reported, 5 but so far the throughput on problems of comparable
size is only equivalent to the best PME-based implementations. FMM-based
algorithms can deliver linear scaling of communication and computation with
both the number of MPI ranks and the number of particles. This linear scaling
is expected to be an advantage when increasing the number of processing units
in the exascale era. Early tests showed that the iteration times of ExaFMM doing only long-range work and GROMACS 4.6 doing only short-range work on
homogeneous systems of the same size were comparable, so we hope we can
deploy a working version in the future.
3.6
Fine-grained task parallelism for exascale
We plan to address some of the exascale-level strong-scaling problems mentioned above through the use of a more fine-grained task parallelism than what
is currently possible in GROMACS. Considerable technical challenges remain
to convert OpenMP-based data-parallel loop constructs into series of tasks that
are coarse enough to avoid spending lots of time scheduling work, and yet fine
enough to balance the overall load. Our initial plan is to experiment with the
cross-platform Thread Building Blocks (TBB) library, 1 which can coexist with
OpenMP and deploy equivalent loop constructs in the early phases of development. Many alternatives exist; those that require the use of custom compilers,
runtime environments, or language extensions are unattractive because that increases the number of combinations of algorithm implementations that must be
maintained and tested, and compromises the high portability enjoyed by GROMACS.
One particular problem that might be alleviated with fine-grained task parallelism is reducing the cost of the communication required during the integration phase. Polymers such as protein backbones are modeled with fixed-length
bonds, with at least two bonds per particle, which leads to coupled constraints
that domain decomposition spreads over multiple ranks. Iterating to satisfy
those constraints can be a costly part of the algorithm at high parallelism. Because the spatial regions that contain bonded interactions are distributed over
many ranks, and the constraint computations cannot begin until after all the
forces for their atoms have been computed, the current implementation waits
for all forces on all ranks to be computed before starting the integration phase.
The performance of post-integration constraint-satisfaction phase is bounded by
18
the latency for the multiple communication stages required. This means that
ranks that lack atoms with coupled bonded interactions, such as all those with
only water molecules,literally have nothing to do at this stage. In an ideal implementation, such ranks could contribute very early in each iteration to complete all the tasks needed for the forces for the atoms involved in coupled bond
constraints. Integration for those atoms could take place while forces for interactions between unrelated atoms are being computed, so that there is computation
to do on all nodes while the communication for the constraint iteration takes
place. This kind of implementation would require considerably more flexibility
in the book-keeping and execution model, which is simply not present today.
4
4.1
Handling exascale software challenges: process and infrastructure
Transition from C to C++98
The major part of the GROMACS code base has been around 1-1.5 million lines
of C code since version 4.0 (http://www.ohloh.net/p/gromacs). Ideally, software engineering on such moderately large multi-purpose code bases would
take place within the context of effective abstractions. 26 For example, someone
developing a new integration algorithm should not need to pay any attention
to whether the parallelization is implemented by constructs from a threading library (like POSIX threads), a compiler-provided threading layer (like OpenMP),
an external message-passing library (like MPI), or remote direct memory access
(like SHMEM). Equally, she/he should not need to know whether the kernels
that compute the forces they are using as inputs are running on any particular kind of accelerator or CPU. Implementing such abstractions generally costs
some developer time, and some compute time. These are necessary evils if the
software is to be able to change as new hardware, new algorithms or new implementations emerge.
Considerable progress has been made in modularizing some aspects of the
code base to provide effective abstraction layers. For example, once the main MD
iteration loop has begun, the programmer does not need to know whether the
MPI layer is provided by an external library because the computation is taking
place on multiple nodes, or the internal thread-based implementation is working to parallelize the computation on a single node. Portable abstract atomic
operations have been available as a side-effect of the thread-MPI development.
Integrators receive vectors of positions, velocities and forces without needing
to know the details of the kernels that computed the forces. The dozens of
non-bonded kernels can make portable SIMD function calls that compile to the
correct hardware operations automatically.
However, the size of the top-level function that implements the loop over time
19
steps has remained at about 1800 code and comment lines since 4.0. It remains
riddled with special-case conditions, comments, and function calls for different parallelization conditions, integration algorithms, optimization constructs,
housekeeping for communication and output, and ensemble algorithms. The
function that computes the forces is even worse, now that both the old and new
non-bonded kernel infrastructures are supported! The code complexity is necessary for a general-purpose multi-architecture tool like GROMACS. However,
needing to be aware of dozens of irrelevant possibilities is a heavy barrier to participation in the project, because it is very difficult to understand all side effects
of a change.
To address this, we are in the process of a transition from C99 to C++98 for
much of this high-level control code. While we remain alert to the possibility
that HPC compilers will not be as effective at compiling C++98 as they are for
C99, the impact on execution time of most of this code is negligible and the
impact on developer time is considerable.
Our expectation is that the use of virtual function dispatch will eliminate
much of the complexity of understanding conditional code (including switch
statements over enumerations that must be updated in widely scattered parts
of the code), despite a slightly slower implementation of the actual function
call. After all, GROMACS has long used a custom vtable-like implementation
for run-time dispatch of the non-bonded interaction kernels. Objects managing
resources via RAII exploiting compiler-generated destructor calls for doing the
right thing will lead to shorter development times and fewer problems because
developers have to manage fewer things. Templated container types will help
alleviate the burden of manual memory allocation and deallocation. Existing
C++ testing and mocking libraries will simplify the process of developing adequate testing infrastructure, and existing task-parallelism support libraries such
as Intel TBB 1 will be beneficial.
It is true that some of these objectives could be met by re-writing in more
objected-oriented C, but the prospect of off-loading some tedious tasks to the
compiler is attractive.
4.2
Best practices in open-source scientific software development
Version control is widely considered necessary for successful software development. GROMACS used CVS in its early days and now uses Git (git clone
git://git.gromacs.org/gromacs.git). The ability to trace when behavior changed
and find some metadata about why it might have changed is supremely valuable.
Coordinating the information about desires of users and developers, known
problems, and progress with current work is an ongoing task that is difficult
with a development team scattered around the world and thousands of users
20
who rarely meet. GROMACS uses the Redmine issue-tracking system3 to discuss
feature development, report and discuss bugs, and to monitor intended and actual progress towards milestones. Commits in the git repository are expected to
reference Redmine issues where appropriate, which generates automatic HTML
cross-references to save people time finding information.
Peer review of scientific research is the accepted gold standard of quality because of the need for specialist understanding to fully appreciate, value, criticize
and improve the work. Software development on projects like GROMACS is
comparably complex, and our experience has been that peer review has worked
well there. Specifically, all proposed changes to GROMACS – even from the core
authors – must go through our Gerrit code-review website4 , and receive positive
reviews from at least two other developers of suitable experience, before they
can be merged. User- and developer-level documentation must be part of the
same change. Requiring this review to happen before acceptance has eliminated
many problems before they could be felt. It also creates social pressure for people to be active in reviewing others’ code, lest they have no karma with which to
get their own proposals reviewed. As features are implemented or bugs fixed,
corresponding Redmine issues are automatically updated. Gerrit also provides
a common venue for developers to share work in progress, either privately or
publicly.
Testing is one of the least favourite activities of programmers, who would
much rather continue being creative in solving new problems. The standard
procedure in software engineering is to deploy continuous integration, where
each new or proposed change is subjected to a range of automatic tests. In the
GROMACS project, we use Jenkins5 to build the project on a wide range of operating systems (MacOS, Windows, flavours of Linux), compilers (GNU, Intel,
Microsoft, clang; and several versions of each), and build configurations (MPI,
thread-MPI, OpenMP, different kinds of SIMD), and then automatically test the
results for correctness. This immediately finds problems such as programmers
using POSIX constructs that are not implemented on Windows. Most of our tests
detect regressions, where a change in the code leads to an unintended change
in behavior. Unfortunately, many of these tests are still structured around executing a whole MD process, which makes it difficult to track down where a
problem has occurred, unless the code change is tightly focused. This motivates the discipline of proposing changes that only have one logical effect, and
working towards adding module-level testing. New behaviors are expected to
be integrated alongside tests of that behavior, so that we continue to build upon
the test infrastructure for the future. All tests are required to pass before code
changes can be merged.
Testing regularly for changes in execution speed is an unsolved problem that
3 http://redmine.gromacs.org
4 http://gerrit.gromacs.org
5 http://jenkins.gromacs.org
21
is particularly important for monitoring our exascale software developments. It
is less suited for deployment via continuous integration, because of the quantity
of computation required to test the throughput of code like GROMACS with
proper load-balancing, at-scale, and on a range of hardware and input conditions. It would be good to be able to execute a weekly end-to-end test run that
shows that unplanned performance regressions have not emerged, but we have
not prioritized it yet. Waiting to do these tests until after feature stability is
achieved in the software-development life cycle is not appropriate, because that
requires extra work in identifying the point in time (ie. the git commit) where
the problem was introduced, and the same work identifying the correct way to
manage the situation. This is much better done while the change is fresh in developers’ minds, so long as the testing procedure is reasonably automatic. Also,
in the gap between commit and testing, a regression may be masked by some
other improvement. More extensive human-based testing before releases should
still be done; but avoiding protracted bug hunts just before releases makes for a
much happier team.
Cross-platform software requires extensive configuration before it can be
built. The system administrator or end user needs to be able to guide what kind
of GROMACS build takes place, and the configuration system needs to verify
that the compiler and machine can satisfy that request. This requires searching
for ways to resolve dependencies, and disclosing to the user what is being done
with what is available. It is important that compilation should not fail when
configuration succeeded, because the end user is generally incapable of diagnosing what the problem was. A biochemist attempting to install GROMACS
on their laptop generally does not know that scrolling back through 100 lines of
output from recursive make calls is needed to find the original compilation error,
and even then they will generally need to ask someone else what the problem
is and how to resolve it. It is far more efficient for both users and developers to
detect during configuration that compilation will fail, and to provide suggested
solutions and guidance at that time. Accordingly, GROMACS uses the CMake
build system (http://www.cmake.org), primarily for its cross-platform support,
but makes extensive use of its high-level constructs, including sub-projects and
scoped variables.
4.3
Profiling
Experience has shown that it is hard to optimize software, especially an HPC
code, based on simple measurements of total execution speed. It is often necessary to have a more fine-grained view of the performance of individual parts of
the code, details of execution on the individual compute units, as well as communication patterns. There is no value in measuring the improvement in execution
time of a non-bonded kernel if the execution time of the FFTs is dominant!
Standard practice is to use a profiling/tracing tool to explore which func-
22
tions or code lines consume important quantities of time, and to focus effort
on those. However, if the measurement is to provide useful information, the
profiler should perturb the execution time by a very small amount. This is particularly challenging with GROMACS because in our case an MD iteration is
typically in the range of a millisecond or less wall clock time around the current
scaling limit, and the functions that are interesting to profile might execute only
for microseconds. Overhead introduced by performance measurement that is
acceptable in other kinds of applications often leads to incorrect conclusions for
GROMACS. Statistical sampling from periodically interrupting the execution to
observe which core is doing which task could work in principle, but (for example) Intel’s VTune 3 Amplifier defaults to a 10 ms interval, which does not create
confidence that use of the tool would lead to accurate observations of events
whose duration is a thousand times shorter. Reducing the profiling overhead to
an acceptable level while still capturing enough information to be able to easily
interpret the performance measurements has proved challenging. Additionally,
this often required expert knowledge, assistance of the developers of the respective performance measurement tool. This makes it exceptionally hard to use
in-depth or large-scale profiling as part of the regular GROMACS development
workflow.
However, we have not been optimizing in the dark; the main mdrun simulation tool has included a built-in tracing-like functionality for many years. This
functionality relies on manual instrumentation of the entire source code-base
(through inlined start/stop timing functions) as well as low-overhead timing
measurements based on processor cycle counters. The great benefit is that the
log output of every GROMACS simulation contains a breakdown of detailed
timing measurements of the different code parts. However, this internal tracing functionality does not reach its full potential because the collected data is
typically displayed and analyzed through time-averages across MPI ranks and
time-steps, often hiding useful details.
To realize more of this potential, we have explored the possibility of more
detailed MPI rank-based statistics, including minimum and maximum execution
times across ranks as well as averages. However, this information is still less
detailed than that from a classical trace and profile visualizer. We are exploring
combining our internal instrumentation with a tracing library. By adding API
calls to various tracing libraries to our instrumentation calls, we can provide
native support for detailed trace-generation in GROMACS just by linking against
a tracing library like Extrae6 . This will make it considerably easier to carry
out performance analysis without the need for expert knowledge on collecting
performance data while avoiding influencing the program behavior by overhead.
6 http://www.bsc.es/computer-sciences/extrae
23
5
Future directions
GROMACS has grown from an in-house simulation code into a large international software project, which now also has highly professional developer, testing and profiling environments to match it. We believe the code is quite unique
in the extent to which it interacts with the underlying hardware, and while there
are many significant challenges remaining this provides a very strong base for
further extreme-scale computing development. However, scientific software is
rapidly becoming very dependent on deep technical computing expertise: Many
amazingly smart algorithms are becoming irrelevant since they cannot be implemented efficiently on modern hardware, and the inherent complexity of this
hardware makes it very difficult even for highly skilled physicists and chemists
to predict what will work. It is similarly not realistic to expect every research
group to afford a resident computer expert, which will likely require both research groups and computing centers to increasingly join efforts to create large
open source community codes where it is realistic to fund multiple full time developers. In closing, the high performance and extreme-scale computing landscape is currently changing faster than it has ever done before. It is a formidable
challenge for software to keep up with this pace, but the potential rewards of
exascale computing are equally large.
Acknowledgments
This work was supported by the European research Council (258980, BH), the
Swedish e-Science research center, and the EU FP7 CRESTA project (287703).
Computational resources were provided by the Swedish National Infrastructure
for computing (grants SNIC 025/12-32 & 2013-26/24) and the Leibniz Supercomputing Center.
References
[1] Intel Thread Building Blocks, https://www.threadingbuildingblocks.
org
[2] Abraham, M.J., Gready, J.E.: Optimization of parameters for molecular dynamics simulation using Smooth Particle-Mesh Ewald in GROMACS 4.5. J.
Comput. Chem. 32(9), 2031–2040 (JUL 15 2011)
[3] Amdahl, G.M.: Validity of the single processor approach to achieving large scale computing capabilities. In: Proceedings of the April 1820, 1967, Spring Joint Computer Conference. pp. 483–485. AFIPS ’67
(Spring), ACM, New York, NY, USA (1967), http://doi.acm.org/10.1145/
1465482.1465560
24
[4] Anderson, J.A., Lorenz, C.D., Travesset, A.: General purpose molecular
dynamics simulations fully implemented on graphics processing units. J.
Comput. Phys. 227, 5324–5329 (2008)
[5] Andoh, Y., Yoshii, N., Fujimoto, K., Mizutani, K., Kojima, H., Yamada, A.,
Okazaki, S., Kawaguchi, K., Nagao, H., Iwahashi, K., Mizutani, F., Minami,
K., Ichikawa, S.i., Komatsu, H., Ishizuki, S., Takeda, Y., Fukushima, M.:
MODYLAS: A Highly Parallelized General-Purpose Molecular Dynamics
Simulation Program for Large-Scale Systems with Long-Range Forces Calculated by Fast Multipole Method (FMM) and Highly Scalable Fine-Grained
New Parallel Processing Algorithms. Journal of Chemical Theory and Computation 9(7), 3201–3209 (2013), http://pubs.acs.org/doi/abs/10.1021/
ct400203a
[6] Arnold, A., Fahrenberger, F., Holm, C., Lenz, O., Bolten, M., Dachsel, H.,
Halver, R., Kabadshow, I., Gähler, F., Heber, F., Iseringhausen, J., Hofmann,
M., Pippig, M., Potts, D., Sutmann, G.: Comparison of scalable fast methods
for long-range interactions. Phys. Rev. E 88, 063308 (Dec 2013), http://
link.aps.org/doi/10.1103/PhysRevE.88.063308
[7] Bowers, K.J., Dror, R.O., Shaw, D.E.: Overview of neutral territory methods for the parallel evaluation of pairwise particle interactions. Journal
of Physics: Conference Series 16(1), 300 (2005), http://stacks.iop.org/
1742-6596/16/i=1/a=041
[8] Bowers, K.J., Dror, R.O., Shaw, D.E.: Zonal methods for the parallel execution of range-limited n-body simulations. J. Comput. Phys. 221(1), 303–329
(Jan 2007), http://dx.doi.org/10.1016/j.jcp.2006.06.014
[9] Brown, W.M., Wang, P., Plimpton, S.J., Tharrington, A.N.: Implementing
molecular dynamics on hybrid high performance computers - short range
forces. Comp. Phys. Comm. 182, 898–911 (2011)
[10] Eastman, P., Pande, V.S.: Efficient nonbonded interactions for molecular dynamics on a graphics processing unit. J. Comput. Chem. 31, 1268–72 (2010)
[11] Eleftheriou, M., Moreira, J.E., Fitch, B.G., Germain, R.S.: A volumetric FFT
for BlueGene/L. In: Pinkston, T.M., Prasanna, V.K. (eds.) High Performance
Computing - HiPC 2003, Lecture Notes in Computer Science, vol. 2913, pp.
194–203. Springer Berlin Heidelberg (2003)
[12] Essmann, U., Perera, L., Berkowitz, M.L., Darden, T., Lee, H., Pedersen,
L.G.: A smooth particle mesh Ewald method. J. Chem. Phys. 103(19), 8577–
8593 (1995)
25
[13] Faradjian, A., Elber, R.: Computing time scales from reaction coordinates
by milestoning. J. Chem. Phys. 120, 10880–10889 (2004)
[14] Hess, B., Kutzner, C., van der Spoel, D., Lindahl, E.: GROMACS 4: Algorithms for highly efficient, load-balanced, and scalable molecular simulation. J. Chem. Theory Comput. 4(3), 435–447 (2008)
[15] Humphrey, W., Dalke, A., Schulten, K.: VMD: Visual molecular dynamics.
J. Mol. Graph. 14(1), 33–38 (1996)
[16] Jagode, H.: Fourier transforms for the BlueGene/L communication network. Ph.D. thesis, The University of Edinburgh, Edinburgh, UK (2005)
[17] Páll, S., Hess, B.: A flexible algorithm for calculating pair interactions on SIMD architectures. Computer Physics Communications 184(12),
2641 – 2650 (2013), http://www.sciencedirect.com/science/article/
pii/S0010465513001975
[18] Phillips, J.C., Braun, R., Wang, W., Gumbart, J., Tajkhorshid, E., Villa, E.,
Chipot, C., Skeel, R.D., Kale, L., , Schulten, K.: Scalable molecular dynamics
with NAMD 26, 1781–1802 (2005)
[19] Pronk, S., Larsson, P., Pouya, I., Bowman, G.R., Haque, I.S., Beauchamp,
K., Hess, B., Pande, V.S., Kasson, P.M., Lindahl, E.: Copernicus: A new
paradigm for parallel adaptive molecular dynamics. In: Proceedings of 2011
International Conference for High Performance Computing, Networking,
Storage and Analysis. pp. 60:1–60:10. SC ’11, ACM, New York, NY, USA
(2011), http://doi.acm.org/10.1145/2063384.2063465
[20] Pronk, S., Páll, S., Schulz, R., Larsson, P., Bjelkmar, P., Apostolov, R.,
Shirts, M.R., Smith, J.C., Kasson, P.M., van der Spoel, D., Hess, B., Lindahl,
E.: GROMACS 4.5: a high-throughput and highly parallel open source
molecular simulation toolkit. Bioinformatics 29(7), 845–854 (2013), http:
//bioinformatics.oxfordjournals.org/content/29/7/845.abstract
[21] Reyes, R., Turner, A., Hess, B.: Introducing SHMEM into the GROMACS molecular dynamics application: experience and results. In:
M Weiland, A.J., Johnson, N. (eds.) Proceedings of the 7th International Conference on PGAS Programming Models. The University of Edinburgh (October 2013), http://www.pgas2013.org.uk/sites/default/
files/pgas2013proceedings.pdf
[22] Schütte, C., Winkelmann, S., Hartmann, C.: Optimal control of molecular
dynamics using Markov state models. Math. Program. (Series B) 134, 259–
282 (2012)
26
[23] Shirts, M., Pande, V.S.: Screen savers of the world unite! Science 290(5498),
1903–1904 (2000), http://www.sciencemag.org/content/290/5498/1903.
short
[24] Sugita, Y., Okamoto, Y.: Replica-exchange molecular dynamics method for
protein folding. Chem. Phys. Lett. 314, 141–151 (1999)
[25] Verlet, L.: Computer ”Experiments” on Classical Fluids. I. Thermodynamical Properties of Lennard-Jones Molecules. Phys. Rev. 159, 98–103 (Jul 1967),
http://link.aps.org/doi/10.1103/PhysRev.159.98
[26] Wilson, G., Aruliah, D.A., Brown, C.T., Chue Hong, N.P., Davis, M.,
Guy, R.T., Haddock, S.H.D., Huff, K.D., Mitchell, I.M., Plumbley, M.D.,
Waugh, B., White, E.P., Wilson, P.: Best Practices for Scientific Computing. PLoS Biol 12(1), e1001745 (01 2014), http://dx.doi.org/10.1371/
journal.pbio.1001745
[27] Yokota, R., Barba, L.A.: A tuned and scalable fast multipole method as a
preeminent algorithm for exascale systems. International Journal of High
Performance Computing Applications 26(4), 337–346 (2012), http://hpc.
sagepub.com/content/26/4/337.abstract
27
| 5cs.CE
|
Feeding Hand-Crafted Features for Enhancing the Performance of
Convolutional Neural Networks
arXiv:1801.07848v1 [cs.CV] 24 Jan 2018
Sepidehsadat Hosseini
Seoul National University
[email protected]
Seok Hee Lee
Seoul Nat’l Univ.
[email protected]
Abstract
Nam Ik Cho
Seoul National University
[email protected]
and extract the features from the training data without human intervention. In this paper, we show that feeding some
effective hand-crafted features to the CNN, along with the
input images, can enhance the performance of CNN at least
in the case of some face related tasks that we focus on. In
other words, enforcing the CNN to use the domain knowledge can increase the performance or can save the computations by reducing the depth. To be specific with the
age/gender estimation problem, since the most important
features are the angle and depth of the wrinkles in our faces,
we believe that the bio-inspired multi-scale Gabor filter responses [7] are the right features for this problem. Hence,
we propose a method to get the benefits of BIF, together
with the features that are learned by the CNN with the input
images. Precisely, we extract several Gabor filter responses
and concatenate them with the input image, which forms
a tensor input like a multi-channel image. The tensor input can be directly fed to the CNN, like we feed the multichannel image to the CNN. In addition to this scheme, we
let the first layer of the CNN to be a 1 × 1 convolution such
that a matrix is obtained at the first layer, which is actually a
weighted sum of the input image and Gabor responses. This
can also be considered a fusion of input image and filter
bank responses, which looks like an image with enhanced
trextures, and the fused image is fed to the CNN.
Since the convolutional neural network (CNN) is believed to find right features for a given problem, the study
of hand-crafted features is somewhat neglected these days.
In this paper, we show that finding an appropriate feature
for the given problem may be still important as they can enhance the performance of CNN-based algorithms. Specifically, we show that feeding an appropriate feature to the
CNN enhances its performance in some face related works
such as age/gender estimation, face detection and emotion
recognition. We use Gabor filter bank responses for these
tasks, feeding them to the CNN along with the input image. The stack of image and Gabor responses can be fed to
the CNN as a tensor input, or as a fused image which is a
weighted sum of image and Gabor responses. The Gabor
filter parameters can also be tuned depending on the given
problem, for increasing the performance. From the extensive experiments, it is shown that the proposed methods provide better performance than the conventional CNN-based
methods that use only the input images.
1. Introduction
The CNNs are gaining more and more attention as they
are successfully applied to many image processing and
computer vision tasks, providing better performance than
the non-CNN approaches. Face related tasks are not the exceptions, for example, the CNNs in [18, 33, 38] provide better face detection performance than the conventional methods such as Haar-like feature based face detector [30], local
binary pattern (LBP) based method [1] and deformable part
model based ones [?, 25]. In the case of age/gender classification, the CNN estimators [16, 32] give more accurate
results than the method based on the bio-inspired features
(BIF) [7], which is one of the best methods among the nonCNN approaches.
Most of CNNs from low to high-level vision problems
use the image (not the features) as the input, and they learn
Analysis of feature maps from some of convolution layers shows that the wrinkle features and face shapes are more
enhanced in our CNN than the conventional one that uses
only the pixel values as the input. As a result, the accuracy
of age/gender estimation is much improved compared to the
state-of-the-art image-domain CNNs [16, 32]. Moreover,
we test our approach on face detection and emotion recognition and also obtain some gains over the existing CNN
based methods [18, 33, 38]. In other tasks where some of
the hand-crafted features are apparently effective, we hope
that feeding such features along with the image may bring
better results.
1
2. Related work
the influence of race and gender by proposing a multi-task
network.
Gaobr filters. Nobel prize winners Hubel and Wiesel discovered that there are simple cells in the primary visual
cortex, where its receptive field is divided into subregions
that are layers covering the whole field [11]. Also in [21],
Petkov proposed the Gabor filter, as a suitable approximation of mammal’s visual cortex receptive field. The 2D
Gabor filter is a Gaussian kernel function adjusted by a sinusoidal wave, consisting of both imaginary and real parts
where the real part can be described as:
0
x0
x + γy 02
cos
2π
+
φ
gλ,θ,σ,γ (x, y) = exp −
2σ 2
λ
(1)
where x0 = x cos θ + y sin θ, y 0 = −x sin θ + y cos θ, and
λ, θ, φ , γ and σ are the wavelength of the real part of Gabor filter kernel, the orientation of the normal to the stripes
of function, phase offset, spatial ratio and standard deviation of the Gaussian envelope representatives respectively.
Fig. 1 is an example of Gabor filter response to a face image, which shows that they find the textures that correspond
to the given θ very well. Hence the Gabor filter responses
have been used in the applications where the (orientational)
textures play an important role such as fingerprint recognition [6], face detection [10], facial expression recognition
[14], age/gender estimation [7], text segmentation [26], super resolution [24], and texture description.
Face detection. There are a large number of face detection
methods, as it is also a very important topic. For details, refer to a complete survey on face detection done by Zafeiriou
et al.[37]. Like other computer vision problems, the CNNs
are now effectively used for face detection [33, 34, ?].
Facial Expression Recognition Emotion classification is a
relatively young and complicated task among many facerelated tasks. Since the facial expression recognition (FER)
plays an important role in human-machine interaction, recently more researches are being performed on this subject.
For some examples of conventional methods, Tang used
support vector machine (SVM) for this problem [29]. Ionwscu et al. also used SVM to improve Bag of Visual words
(BOW) approach [12]. Hassani et al. used the advantage of
facial landmarks along with CNNs [8]. More recent studies
are focused on using the CNNs for the FER [36, 17, 22].
3. Preparation of Input
We attempt our approach to several face related works
such as age/gender estimation, face detection, and emotion recognition. Each of them needs different CNN architecture, but they are all fed with the Gabor filter responses as the input along with the image. As can be seen
from the eq. (1), there are several parameters which induce different filter responses. In all the applications, we
prepare eight filter banks by combining the cases of four
θ = 0, π/4, π/2, 3π/4 and two φ = 0, π/2. The rest of
parameters σ, λ and γ are changed depending on the application. For the age and gender estimation problem, we set
σ = 2, λ = 2.5, and γ = 0.3.
Let Nf (=8 in all the experiments in this paper as stated
above) be the number of Gabor filters, and let Fgk be the response of k-th Gabor filter. Normally, we may just concatenate the input image and Nf responses as W ×H ×(Nf +1)
tensor input to a CNN as illustrated in Fig. 2(a). On the
other hand, we may consider fusing the input and Gabor responses as a single input and feed the matrix to the CNN
as shown in Fig. 2(b). The figure also shows that fusing
the input image and Gabor responses can be interpreted
as convolving the W × H × (Nf + 1) tensor input with
1 × 1 × (Nf + 1) filter. If we denote the coefficients of
this filter as [wi , w1 , w2 , · · · , wNf ], where wi is multiplied to the input image and the rest are multiplied to Gabor
responses, then the fused input is represented as
Figure 1: Demonstration of Gabor filter bank responses
with kernel size = 5 applied to an image. Responses for
four orientations (θ = 0, π/4, π/2, 3π/4) are shown.
Age/Gender Estimation. Predicting the age of a person
from a single image is one of the hardest tasks, which even
humans sometimes have difficulties in doing that. The reason is that aging depends on several factors such as living
habits, races, genetics, etc. The studies without using the
CNN are well summarized in Fu et al.’s survey [?]. Recent
works are mostly based on the CNN, for some examples,
Levi and Hassner’s work [16] was the first to adopt the CNN
for age/gender estimation, and Xing et al. [32] considered
F in = wi I +
Nf
X
wk Fgk
(2)
k=1
which is similar to the weighted fusion method in [28],[32].
Fig. 2(c) is an example of fused input, which can be con2
sidered a “wrinkle-enhanced” image. Both of concatenation and fusion approaches inject the Gabor responses as
the input to the CNN. From the extensive experiments, the
fusion approach in Fig. 2(b) shows slightly better performance (about 1%p increase in the case of gender estimation
and similarly to other tasks) while requiring slightly less
number of parameters.
as stated previously, and the GF CNN outperforms the existing methods by at least 3.1 %p on Aidence dataset and
1.3%p on Gallagher dataset.
For gender estimation, our method outperforms all the
other ones on Adience as shown in Table 2. The Table also
shows that the proposed network shows almost the same
performance as VGG hybrid on Webface dataset, while it
has ten times less number of parameters than the VGG.
For the analysis of the effects of feeding the Gabor responses, we compare some feature maps in Fig. 3. Specifically, Fig. 3(a) shows the feature maps from our GF CNN
and Fig. 3(b) from the CNN with only image input at
the same layer. It can be seen that the features from the
GF CNN contain more strong facial features and wrinkle
textures than the original network, which is believed to be
the cause of better performance.
4. Networks for face related problems
We apply the Gabor responses to the CNNs for the
age/gender estimation, face detection, and emotion recognition problems in the following subsections. At each subsection, we show that the performance is improved by feeding the Gabor responses as compared to the case of feeding
only the image input.
4.1. Age/Gender classification
A
Table 1: Age estimation (classification) results on Adience
& Gallagher datasets.
Network
The gender estimation is just a binary classification,
while the age estimation is implemented as a classification or regression problem. In the case of age estimation
as a classification problem (segmenting the age into several
ranges), the network shown in Fig. 2(a) or (b) is used. Each
convolution block consists of convolution layer, Relu, and
Max pooling, and each fully connected block consists of
fully connected layer, Relu and drop-out with the drop ratio
0.5.
B
Method
LBP [31]
LBP+FPLBP+Droupout 0.8 [31]
Eidinger [3]
Best from Levi [16]
Resnet[9]
PTP [13]
DAPP [13]
GT CNN[Ours]
GF CNN[Ours]
Gallagher
58.0
66.6
N.A.
N.A.
68.1
68.6
69.91
69.1
71.4
Dataset description
We perform age classification on two popular datasets,
Adience [3] and Gallagher dataset [4]. Both are from
flickr.com, including the pictures with large variations in
poses, appearances, lighting condition, unusual facial expressions, etc. Adience has approximatively 26K images of
2k subjects in 8 classes (0-2, 4-6, 8-13, 15-20, 25-32, 3842, 48-53, 60+), Gallagher dataset has 5K images with 28K
labeled faces, being divided into 7 classes (0-2, 3-7,8-12,
13-19, 20-36, 37-65, 66+). For gender estimation, we used
Adience and CASIA Webface [35]. It has 450K images of
10K subjects, which is obtained from the pictures on IMDB
and most of the pictures in the dataset are celebrities.
C
Adience
41.1
45.1
45.1
50.7
52.2
53.27
54.9
57.2
59.3
Table 2: Gender estimation results on Adience & Webface
datasets.
Method
BIF [7]
Eidinger [3]
Best from Levi [16]
Resnet[9]
GG
N etVHybrid
[32]
GT CNN[Ours]
GF CNN[Ours]
Adience
N.A
77.8
86.8
88.5
N.A
89.2
90.1
Webface
79.3
N.A.
N.A.
89.2
92.3
91.0
92.1
Test and result
We perform the experiments based on the standard fivefold, subject-exclusive cross-validation protocol for fair
comparison. Table 1 shows the results for age estimation, where GT CNN means our method that use Gabor responses as tensor input and GF CNN as fused input. It can
be observed that GF CNN is slightly better than GT CNN
4.2. Age regression
A
Network
Age estimation can also be implemented as a regression
problem when we wish to tell a person’s exact age, rather
3
Figure 2: Illustration of two input feeding methods. (a) The tensor input is directly fed to the CNN, (b) The tensor input
is fused to be an image and fed to the CNN. (c) Example of a fused image which is the weighted sum of image and Gabor
responses.
defined as
(a)
L(x) = −
M
1 X
|yˆi − yi |
M i=1
(4)
where M is the maximum age that we set, and yˆi is the
estimate of true age yi .
(b)
Figure 3: Comparison of feature maps after the first convolution layer in two networks. Features from (a) GF CNN
and (b) original CNN with image input.
Figure 4: Age regression network architecture.
than as a classification problem which tells the range (class)
of ages. We use the network shown in Fig. 4 for this problem. One of the main differences between the age classification and regression problem is that they need different
loss functions. For the classification problem above, we use
the Softmax loss defined as:
L(x) = −
N
1 X
Yiyi log piyi
N i=1
B
Dataset description
For age regression task, we perform the experiments on
two widely used datasets for age estimation in literature.
We choose CASIA-Webface dataset [35] as it consists of a
large amount of pictures, and also we used FG-Net database
which contains 1002 images of 82 subjects, where subjects’
ages range from 0 to 69.
(3)
where N is the number of classes, Yiyi is the one-hot encoding of sample’s age label, and piyi is the yi -th element
of predicted probability vector for xi . For the regression,
we use Mean Squared Error (MSE) or Mean Absolute Error (MAE) as the loss function. To be precise, the MAE is
C
Test and result
We used four-fold cross-validation protocol for Webface
dataset and the Leave-One-Person-Out (LOPO) test strategy
while working on FG-Net because the number of pictures in
4
FG-Net is small. Table 3 shows the result of age estimation.
It can be seen that our network shows better performance
than the state of the art method.
Table 3: Age estimation error on Adience and Gallagher
datasets. GF CNNresent means that we use residual learning.
Method
BIF[7]
RF[19]
EBIF[2]
GG
N etVHybrid
GF CNN[ours]
GT CNNresent [ours]
GF CNNresent [ours]
Casia Webface Dataset
10.65
9.38
N.A.
5.75
5.83
5.66
5.61
FG-net
4.77
4.21
3.17
N.A.
3.13
3.15
3.08
Figure 5: Illustration of three stages of face detection network architecture (GP-Net, GR-Net ,GO-Net).
C
For the bounding box regression and face classification,
we use the same loss as [38]. Specifically, we use crossentropy loss:
4.3. Face detection
A
Network
det
det
Ldet
i (xi ) = −(yi log(pi ) + (1 − yi )(1 − log(pi ))) (5)
Our face detector is a three-stage cascaded CNN which
is the same as Zhang et al.’s network [38], except that we
use the fusion of input and Gabor responses as shown in
Fig. 5. At stage 1, which is called P-Net, possible facial windows along with their bonding box regression vectors are obtained. Then the bounding boxes are calibrated,
and the highly overlapped ones are merged to others using non-maximum suppression (NMS). In the second and
third stages (called R-Net and O-Net respectively) the candidates are refined again using the calibration and NMS. For
all these three step networks we feed our Gabor fusion image.
About the Gabor filter parameters, it is noted that finding the facial components such as nose, mouse, eyes, etc.
are more important than the relatively straight and sometimes long wrinkles that were important in the previous
age/gender estimation. Hence we reduce the kernel size of
Gabor filter and also the parameters σ, λ and γ to 0.75, 2,
and 0.05 respectively.
B
Test and result
where pi is the probability of xi being a face and yidet is the
ground truth. For the bonding box we use:
LB.Box
(xi ) = ||ŷiB.box − yiB.box ||22
i
(6)
where ŷiB.box andyiB.box are the network output and ground
truth respectively. Table 4 shows that we can get better performance with almost same number parameters as
MTCNN. Figs. 6(a)-(c) show in all three stages using hand
crafted features can improve the performance and help increase the network convergence speed. To evaluate our face
detection method we compare our method with other six
sate- of-the-art methods on FDDB and our method outperform all of them as shown in Fig. 6(d). At last, we compare
our method’s run time with other CNN based methods and
results are in 5 as it can be seen while purposed method
has better performance than MTCNN and cascade CNN it
is almost as fast as them.
4.4. Facial expression recognition
Dataset description
A
In this section, we evaluate our network on Face Detection Dataset and Benchmark (FDDB) [15] which contains
2,845 images with 5K annotated faces taken in the wild.
There are two types of evaluation available on FDDB: discontinuous score which counts the number of detected faces
versus the number of false positives, and continuous score
which evaluates how much is the overlap of bounding boxes
on the faces between the ground truth and detected.
Network
The baseline network for FER is VGG-19 [27], and we just
add one more drop out after the last fully connected layer to
decrease the overlapping, as shown in Fig. 7. For the FER,
we think that the wrinkles again play an important role here,
and hence that we set the bandwith larger than the previous
case, specifically set σ = 1.4. Also, λ becomes large to 2.5,
and set γ = 0.1.
5
(a)
(b)
(c)
(d)
Figure 6: (a)-(c) Comparision between the three stages of MTCNN [38] (in orange) and our method (in Green). (d) Comparison of our performance with MTCNN [38], DP2MDF[25], cascade CNN [18], Faceness[33], Joint fasterRCNN [23] and
head hunter [20], where the numbers in the parentheses are the area under curve.
Table 4: Comparison of Validation Accuracy of Ours, CascadeCNN and MTCNN.
Group
Group1
Group2
Group3
CNN
12-Net[18]
P-Net[38]
GP-Net[ours].
24-Net[18]
R-Net[38]
GR-Net[ours]
48-Net[18]
O-Net[38]
GO-Net[ours]
Table 5: Runtime Comparison on the same GPU.
Method
Faceness [33]
MTCNN [38]
Cascade CNN [18]
GF-MTCNN[Ours]
Validation Accuracy
94.4%
94.6%
94.83%
95.1%
95.4%
95.61%
93.2%
95.4%
95.72%
B
Speed
20 FPS
99 FPS
100 FPS
99 FPS
Dataset description
We evaluate our network on FER 2013 dataset [5], which is
being labeled in seven classes(0=Angry, 1=Disgust, 2=Fear,
3=Happy, 4=Sad, 5=Surprise, and 6=Neutral). It contains
about 32K images, 28.5K for training and 3.5K for the test.
C
Test and result
Table 6 shows our result, where we compare our results with
the FER 2013 competition winners and other state of the art
methods. It can be seen that our network shows better performance than others. While VGGNet can also reach to
Figure 7: Illustration GF-VGG network for the FER.
6
69.8 %, adding our fusion module at the input of the network can increase the performance by 2.098%p.
Table 6: Results of FER.
Method
Radu + Marius + Cristi [12]
Unsupervised[5]
Maxim Milakov [5]
SVM [29]
VGGNet [27]
GF-VGGNet[Ours]
Accuracy on FER 2013
67.484%
69.267%
68.821%
71.162%
69.08%
72.198%
[6]
[7]
5. Conclusion
[8]
Most of CNNs for image understanding use the image as
the input, with the belief that the CNN will automatically
find the appropriate features from the data. In this paper,
we have shown that feeding appropriate hand-crafted features can lead to the improved results. Hence the domain
knowledge and study of appropriate features are important
for improving the CNN-based algorithms. Specifically, we
have shown that feeding the Gabor filter response to the
CNN leads to better performances in face related problems
such as age/gender estimation, face detection, and emotion
recognition. We hope there can be more applications that
can be benefited by our approach, i.e., there can be more
CNN-based image processing and vision algorithms that
can have gains by taking the appropriate features as the input.
[9]
[10]
[11]
[12]
[13]
References
[1] O. Bilaniuk, E. Fazl-Ersi, R. Laganire, C. Xu,
D. Laroche, and C. Moulder. Fast lbp face detection
on low-power simd architectures. In Proceedings of
the IEEE Conference on Computer Vision and Pattern
Recognition Workshops, pages 616–622, 2014.
[2] M. E. Deeb and M. El-Saban. Human age estimation
using enhanced bio-inspired features (ebif). In ICIP,
2010.
[3] E. Eidinger, R. Enbar, and T. Hassner. Age and gender estimation of unfiltered faces. IEEE Transactions
on Information Forensics and Security, 9(12):2170–
2179, 2014.
[4] A. C. Gallagher and T. Chen. Understanding images
of groups of people. In Computer Vision and Pattern
Recognition, 2009. CVPR 2009. IEEE Conference on,
pages 256–263. IEEE, 2009.
[5] I. Goodfellow, D. Erhan, P.-L. Carrier, A. Courville,
M. Mirza, B. Hamner, W. Cukierski, Y. Tang,
[14]
[15]
[16]
[17]
7
D. Thaler, D.-H. Lee, Y. Zhou, C. Ramaiah, F. Feng,
R. Li, X. Wang, D. Athanasakis, J. Shawe-Taylor,
M. Milakov, J. Park, R. Ionescu, M. Popescu,
C. Grozea, J. Bergstra, J. Xie, L. Romaszko, B. Xu,
Z. Chuang, and Y. Bengio. Challenges in representation learning: A report on three machine learning
contests, 2013.
C. Gottschlich. Curved-region-based ridge frequency
estimation and curved gabor filters for fingerprint image enhancement. IEEE Transactions on Image Processing, 21(4):2220–2227, 2011.
G. Guo, G. Mu, Y. Fu, and T. S. Huang. Human age
estimation using bio-inspired features. In Computer
Vision and Pattern Recognition, 2009. CVPR 2009.
IEEE Conference on. IEEE, 2009.
B. Hassani and M. H. Mahoor. Facial expression
recognition using enhanced deep 3d convolutional
neural networks. CoRR, abs/1705.07871, 2017.
K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. arXiv preprint
arXiv:1512.03385, 2015.
L.-L. Huang, A. Shimizu, and H. Kobatake. Robust face detection using gabor filter features. Pattern
Recognition Letters, 26(11):1641–1649, 2005.
D. H. Hubel and T. Wiese. Receptive fields, binocular interaction and functional architecture in the cat’s
visual cortex. Journal of Physiology, 160(1):106,154,
1962.
R. T. Ionescu and C. Grozea. Local learning to improve bag of visual words model for facial expression
recognition. 2013.
M. T. B. Iqbal, M. Shoyaib, B. Ryu, M. Abdullah-AlWadud, and O. Chae. Directional age-primitive pattern (dapp) for human age group recognition and age
estimation. IEEE Transactions on Information Forensics and Security, 12:2505–2517, 2017.
M. J. Lyons, S. Akamatsu, M. G. Kamachi, and J. Gyoba. Coding facial expressions with gabor wavelets. In
Automatic Face and Gesture Recognition, 1998. Proceedings. Third IEEE International Conference on.
IEEE, 1998.
V. Jain and E. Learned-Miller. Fddb: A benchmark for
face detection in unconstrained settings. Technical Report UM-CS-2010-009, University of Massachusetts,
Amherst, 2010.
G. Levi and T. Hassner. Age and gender classification using convolutional neural network. In Computer
Vision and Pattern Recognition Workshops (CVPRW),
2015 IEEE Conference on, pages 34–42. IEEE, 2015.
G. Levi and T. Hassner. Emotion recognition in the
wild via convolutional neural networks and mapped
binary patterns. In Proc. ACM International Conference on Multimodal Interaction (ICMI), November
2015.
[30] P. Viola and M. J. Jones. Robust real-time face detection. International journal of computer vision,
57(2):137–154, 2004.
[18] H. Li, Z. Lin, X. Shen, J. Brandt, and G. Hua. A convolutional neural network cascade for face detection.
In Computer Vision and Pattern Recognition (CVPR),
2015 IEEE Conference on, pages 5325–5334, 2015.
[31] T. Wu and R. Chellappa. Age invariant face verification with relative craniofacial growth model. In European Conference on Computer Vision ECCV 2012:
Computer Vision, ECCV 2012, volume 7577, pages
58–71, 2012.
[19] S. Li, S. Shan, and X. Chen. Relative Forest for Attribute Prediction, pages 316–327. Springer Berlin
Heidelberg, Berlin, Heidelberg, 2013.
[32] J. Xing, K. Li, W. Hu, C. Yuan, and H. Ling. Diagnosing deep learning models for high accuracy age
estimation from a single image. Pattern Recognit,
66:106–116, 2017.
[20] M. Mathias, R. Benenson, M. Pedersoli, and L. Van
Gool. Face detection without bells and whistles. In
ECCV, 2014.
[33] S. Yang, P. Luo, and X. Change Loy, Chenand Tang.
From facial parts responses to face detection: A deep
learning approach. In IEEE International Conference
on Computer Vision, pages 3676–3684. IEEE, 2015.
[21] N. Petkov. Biologically motivated computationally intensive approaches to image pattern recognition. Future Generation Computer Systems, 11(4–5):451,465,
1995.
[34] S. Yang, P. Luo, and X. Change Loy, Chenand Tang.
Wider face: A face detection benchmark. In Computer
Vision and Pattern Recognition (CVPR), 2016 IEEE
Conference on. IEEE, 2016.
[22] C. Pramerdorfer and M. Kampel. Facial expression recognition using convolutional neural networks:
State of the art. CoRR, abs/1612.02903, 2016.
[35] D. Yi, Z. Lei, S. Liao, and S. Z. Li. Learning face representation from scratch. CoRR, abs/1411.7923, 2014.
[23] H. Qin, J. Yan, X. Li, and X. Hu. Joint training of cascaded cnn for face detection. In Computer Vision and
Pattern Recognition (CVPR), 2016 IEEE Conference
on, 2016.
[36] Z. Yu and C. Zhang. Image based static facial expression recognition with multiple deep network learning.
November 2015.
[24] S. Ram Dogiwal, Y. Shishodia, and A. Upadhyaya.
Super resolution image reconstruction using wavelet
lifting schemes and gabor filters. In Confluence
The Next Generation Information Technology Summit (Confluence), 2014 5th International Conference.
IEEE, 2014.
[37] S. Zafeiriou, C. Zhang, and Z. Zhang. A survey on
face detection in the wild: Past, present and future.
Computer Vision and Image Understandings, 138:1–
24, 2015.
[38] K. Zhang, Z. Zhang, Z. Li, and Y. Qiao. Joint face detection and alignment using multi-task cascaded convolutional networks. IEEE Signal Processing Letters,
23(10):1499–1503, 2016.
[25] R. Ranjan and R. Patel, Vishal M.and Chellappa. A
deep pyramid deformable part model for face detection. In Biometrics Theory, Applications and Systems
(BTAS), 2015 IEEE 7th Int. Conf, pages 1–8, 2015.
[26] S. Sabari Raju, P. Basa Pati, and A. Ramakrishnan.
Gabor filter based block energy analysis for text extraction from digital document images. In Document
Image Analysis for Libraries, 2004. Proceedings. First
International Workshop on. IEEE, 2004.
[27] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition.
CoRR, abs/1409.1556, 2014.
[28] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and
Z. Wojna. Rethinking the inception architecture for
computer vision. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June
2016.
[29] Y. Tang. Deep learning using support vector machines.
CoRR, abs/1306.0239, 2013.
8
| 1cs.CV
|
arXiv:1508.01290v1 [math.AC] 6 Aug 2015
TORIC RINGS, INSEPARABILITY AND RIGIDITY
MINA BIGDELI, JÜRGEN HERZOG AND DANCHENG LU
Abstract. Let K be a field, H an affine semigroup and R = K[H] its toric ring
over K. In this paper we give an explicit description of the ZH-graded components
of the cotangent module T 1 (R) which classifies the infinitesimal deformations of
R. In particular, we are interested in unobstructed deformations which preserve
the toric structure. Such deformations we call separations. Toric rings which
do not admit any separation are called inseparable. We apply the theory to the
edge ring of a finite graph. The coordinate ring of a convex polyomino may be
viewed as the edge ring of a special class of bipartite graphs. It is shown that the
coordinate ring of any convex polyomino is inseparable. We call a bipartite graph
G semi-rigid if T 1 (R)a = 0 for all a with −a ∈ H. Here R is the edge ring of G.
A combinatorial description of semi-rigid graphs is given. The results are applied
to show that for n = 3, Gn is semi-rigid but not rigid, while Gn is rigid for all
n ≥ 4. Here Gn is the complete bipartite graph Kn,n with one edge removed.
Introduction
In this paper we study infinitesimal deformations and unobstructed deformations
of toric rings which preserve the toric structure, and apply this theory to edge
ideals of bipartite graphs. Already in [1] and [2], infinitesimal and homogeneous
deformations of toric varieties have been considered from a geometric point of view.
The view point of this paper is more algebraic and does not exclude non-normal toric
rings, having in mind toric rings which naturally appear in combinatorial contexts.
This aspect of deformation theory has also been pursued in the papers [4],[5] and [3]
where deformations of Stanley-Reisner rings attached to simplicial complexes were
studied.
Let K be a field. The infinitesimal deformations of a finitely generated K-algebra
R are parameterized by the elements of the cotangent module T 1 (R) which in the
case that R is a domain is isomorphic to Ext1R (ΩR/K , R), where ΩR/K denotes the
module of differentials of R over K. The ring R is called rigid if T 1 (R) = 0. We
refer the reader to [12] for the theory of deformation.
Let H be an affine semigroup and K[H] its affine semigroup ring. We are interested in the module T 1 (K[H]). This module is naturally ZH-graded. Here ZH
denotes the associated group of H which for an affine semigroup is a free group of
2010 Mathematics Subject Classification. Primary 13D10, 05E40; Secondary 13C13.
Key words and phrases. Deformation, Toric ring, Rigid, Inseparable, Bipartite graph, Convex
Polyomino.
The paper was written while the third author was visiting the Department of Mathematics of
University Duisburg-Essen. He wants to express his thanks for the hospitality.
1
finite rank. For each a ∈ ZH, the a-graded component T 1 (K[H])a of T 1 (K[H]) is a
finite dimensional K-vector space.
In Section 1 we describe the vector space T 1 (K[H])a and provide a method how
to compute its dimension. Let H ⊂ Zm with generators h1 , . . . , hn . Then the
associated group ZH of H is a subgroup of Zm , and K[H] is the K-subalgebra of the
±1
h1
hn
ring K[t±1
1 , . . . , tm ] of Laurent polynomials generated by the monomials t , . . . , t .
a(1)
a(m)
Here ta = t1 · · · tm for a = (a(1), . . . , a(m)) ∈ Zm . Let S = K[x1 , . . . , xn ] be the
polynomial ring over K in the indeterminates x1 , . . . , xn . Then S may be viewed as
a ZH-graded ring with deg xi = hi , and the K-algebra homomorphism S → K[H]
with xi 7→ thi is a homomorphism of ZH-graded K-algebras. We denote by IH
the kernel of this homomorphism. The ideal IH is called the toric ideal associated
with H. It is generated by homogeneous binomials. To describe these binomials,
consider the group homomorphism Zn → Zm with εi 7→ hi , where ε1 , . . . , εn is the
canonical basis of Zn . The kernel L of this group homomorphism is a lattice of Zn
and is called the relation lattice of H. For v = (v(1), . . . , v(n)) ∈ Zn we define the
Q
Q
v(i)
−v(i)
binomial fv = fv+ − fv− with fv+ = i, v(i)≥0 xi and fv− = i, v(i)≤0 xi
, and
let IL be the ideal generated by the binomials fv with v ∈ L.
P It is well known
that IH = IL . Each fv ∈ IH is homogeneous of degree h(v) = i, v(i)≥0 v(i)hi . Let
fv1 , . . . , fvs be a system of generators of IH . We consider the (s × n)-matrix
v1 (1) v1 (2) . . . v1 (n)
v2 (1) v2 (2) . . . v2 (n)
AH =
.
..
..
...
.
.
vs (1) vs (2) . . . vs (n)
Summarizing the results of Section 1, dimK T 1 (K[H])a can be computed as follows:
let l = rank AH , la be the rank of the submatrix of AH whose rows are the ith rows
of AH for which a + h(vi ) 6∈ H, and let da be the rank of the submatrix of AH whose
columns are the jth columns of AH for which a + hj ∈ H. Then
dimK T 1 (K[H])a = l − la − da .
In Section 2 we introduce the concept of separation for a torsionfree lattice L ⊂ Zn .
Note that a lattice L ⊂ Zn is torsionfree if and only if it is the relation lattice of some
affine semigroup. Given an integer i ∈ [n] = {1, 2, . . . , n}, we say that L admits an
i-separation if there exists a torsionfree lattice L′ ⊂ Zn+1 of the same rank as L such
that πi (IL′ ) = IL , where πi : S[xn+1 ] → S is the K-algebra homomorphism which
identifies xn+1 with xi . An additional condition makes sure that this deformation
which induces an element in T 1 (K(H))−hi is non-trivial, see 2.1 for the precise definition. We say that L is inseparable, if for all i, the lattice L admits no i-separation,
and we call H and its toric ring inseparable if its relation lattice is inseparable. In
particular, if the generators of H belong to a hyperplane of Zm , so that K[H] also
admits a natural standard grading, then H is inseparable if T 1 (K[H])−1 = 0. In
general, the converse is not true since the infinitesimal deformation given by nonzero element of T 1 (K[H])−1 may be obstructed. We demonstrate this theory and
2
show that a numerical semigroup generated by three elements which is not a complete intersection is i-separable for i = 1, 2, 3, while if it is a complete intersection
it is i-separable for at least two i ∈ {1, 2, 3}. For the proof of this fact we use the
structure theorem of such semigroups given in [8].
The last Section 3 is devoted to the study of T 1 (R) when R is the edge ring
of a bipartite graph. This class of rings has been well studied in combinatorial
commutative algebra, see e.g. [11] and [13]. For a given simple graph G of the vertex
set [n] one considers the edge ring R = K[G] which is the toric ring generated over
K by the monomials ti tj for which {i, j} is an edge of G. Viewing the edge ring a
semigroup ring K[H], the edges ei of G correspond the generators hi of the semigroup
H. We say that G is inseparable if the corresponding semigroup is inseparable. The
main result of the first part of this section is a combinatorial criterion for G being
inseparable. Let C be a cycle of G and e a chord of G. Then e splits C into two
disjoint connected components C1 and C2 which are obtained by restricting C to
the complement of e. A path P of G is called a crossing path of C with respect
to e if one end of P belongs to C1 and the other end to C2 . Now the criterion
(Corollary 3.5) says that a bipartite graph G is inseparable if and only if for any
cycle C which has a unique chord e, there exists a crossing path of C with respect
to e. In particular, if no cycle has a chord, then G is inseparable. By using this
criterion we show in Theorem 3.6 that the coordinate ring of any convex polyomino,
which may be interpreted as a special class of edge rings, is inseparable.
For the rest of this section we consider the semi-rigidity and rigidity of bipartite
graphs. We call H semi-rigid if T 1 (K(H))−a = 0 for all a ∈ H, and characterize
in Theorem 3.10 semi-rigidity of bipartite graphs in terms of the non-existence of
certain constellations of edges and cycles of the graph. To classify rigidity of bipartite
graphs is much more complicated, and we do not have a general combinatorial
criterion for when a bipartite graph is rigid. Instead we consider for each n ≥ 3 the
graph Gn which is obtained by removing an edge from the complete bipartite graph
Kn,n . It is shown in Proposition 3.12 that for n = 3, Gn is not rigid while for n ≥ 4,
Gn is rigid. It remains a challenging problem to classify all rigid bipartite graphs.
1. T 1 for toric rings
Let H be an affine semigroup, that is, a finitely generated subsemigroup of Zm for
some m > 0. Let h1 , . . . , hn be the minimal generators of H, and fix a field K. The
±1
toric ring K[H] associated with H is the K-subalgebra of the ring K[t±1
1 , . . . , tm ] of
a(1)
a(m)
Laurent polynomials generated by the monomials th1 , . . . , thn . Here ta = t1 · · · tm
for a = (a(1), . . . , a(m)) ∈ Zm .
Let S = K[x1 , . . . , xn ] be the polynomial ring over K in the variables x1 , . . . , xn .
The K-algebra R = K[H] has a presentation S → R with xi 7→ thi for i = 1, . . . , n.
The kernel IH ⊂ S of this map is called the toric ideal attached to H. Corresponding
to this presentation of K[H] there is a presentation Nn → H of H which can be
extended to the group homomorphism Zn → Zm with εi 7→ hi for i = 1, . . . , n,
where ε1 , . . . , εn denotes the canonical basis of Zn . Let L ⊂ Zn be the kernel of this
3
group homomorphism. The lattice L is called the relation lattice of H. Note that L
is a free abelian group and Zn /L is torsion-free.
For a vector v ∈ Zn with v = (v(1), . . . , v(n)), we set
X
X
v+ =
v(i)εi and v− =
−v(i)εi .
i, v(i)≥0
i, v(i)≤0
Then v = v+ −v− . It is a basic fact and well-known (see e.g. [7]) that IH is generated
by the binomials fv with v ∈ L, where fv = xv+ − xv− .
We define an H-grading on S by setting deg xi = hi . Then IH is a graded ideal
with deg fv = h(v), where
X
X
(1)
h(v) =
v(i)hi (=
−v(i)hi ).
i, v(i)≥0
i, v(i)≤0
Let v1 , . . . , vr be a basis of L. Since IH is a prime ideal we may localize S with
respect to this prime ideal and obtain
IH SIH = (fv1 , . . . , fvr )SIH .
In particular, we see that
(2)
height IH = rank L.
Let ΩR/K be the module of differentials of R over K. Since R is a domain, the
cotangent module T 1 (R) is isomorphic to Ext1R (ΩR/K , R), and since R is H-graded it
follows that ΩR/K is H-graded as well, and hence Ext1R (ΩR/K , R) and T 1 (R) are ZHgraded. Here ZH denotes the associated group of H, that is, the smallest subgroup
of Zm containing H. It is our goal to compute the graded components T 1 (R)a of
T 1 (R) for a ∈ ZH.
The module of differentials has a presentation
ΩR/K = (
n
M
Rdxi )/U,
i=1
L
where U is the submodule of the free R-module ni=1 Rdxi generated by the elements
dfv with v ∈ L, where
n
X
dfv =
(∂fv /∂xi )dxi .
i=1
Here ∂fv /∂xi stands for partial derivative of fv with respect to xi , evaluated modulo
IH .
One verifies at once that
n
X
dfv =
v(i)th(v)−hi dxi .
(3)
i=1
L
For i ∈ [n], the basis element dxi of ΩS/K ⊗S R = ni=1 Rdxi is given the degree
hi . Then U is an H-graded submodule of ΩS/K ⊗S R, and deg dfv = deg fv = h(v).
4
For any ZH-graded R-module M we denote by M ∗ the graded R-dual HomR (M, R).
Then the exact sequence of H-graded R-modules
0 → U → ΩS/K ⊗S R → ΩR/K → 0
gives rise to the exact sequence
(ΩS/K ⊗S R)∗ → U ∗ → T 1 (R) → 0
of ZH-graded modules. This exact sequence may serve as the definition of T 1 (R),
namely, to be the cokernel of (ΩS/K ⊗S R)∗ → U ∗ .
Let fv1 , . . . , fvs be a system of generators of IH , where we may assume that for
r ≤ s, the elements v1 , . . . , vr form a basis of L. In general s is much larger than r.
Observe that the elements dfv1 , . . . , dfvs form a system of generators of U.
We let F be a free graded R-module with basis g1 , . . . , gs such that deg gi =
deg dfvi for i = 1, . . . , s, and define the R-module epimorphism F → U by gi 7→ dfvi
for i = 1, . . . , s. The kernel of F → U we denote by C. The composition F →
ΩS/K ⊗S R of the epimorphism F → U with the inclusion map U → ΩS/K ⊗S R will
be denoted by δ. We identify U ∗ ⊂ F ∗ with its image in F ∗ . Then T 1 (R) = U ∗ / Im δ ∗
and U ∗ is the submodule of F ∗ consisting of all ϕ ∈ F ∗ with ϕ(C) = 0.
We first describe the ZH-graded components of U ∗ . Let a ∈ ZH. We denote by
KL the K-subspace of K n spanned by v1 , . . . , vs and by KLa the K-subspace of
KL spanned by the vectors vi with i ∈
/ Fa . Here the set Fa is defined to be
Fa = {i ∈ [s] : a + h(vi ) ∈ H}.
Then we have
Theorem 1.1. For all a ∈ ZH, we have
dimK (U ∗ )a = dimK KL − dimK KLa .
Proof. Let σ1 , . . . , σs be the canonical basis of K s and W ⊂ K s be the kernel of the
K-linear map K s → KL with σi 7→ vi for i = 1, . . . , s.
We will show that
(4)
(U ∗ )a ∼
= {µ ∈ K s : µ(i) = 0 for i ∈ [s] \ Fa and hµ, λi = 0 for all λ ∈ W },
as K-vector space.
Assuming this isomorphism has been proved,
let Xa be the image of W ⊂ K s
L
under the canonical projection K s → Va = i∈Fa Kσi . Then (4) implies that (U ∗ )a
is isomorphic to the orthogonal complement of Xa in Va . Thus,
(5)
dimK (U ∗ )a = |Fa | − dimK Xa .
5
L
Let Za = i6∈Fa Kσi and Ya the cokernel of Xa → Va . Then we obtain a commutative diagram with exact rows and columns
0
y
0 −−−→ W ∩ Za −−−→
y
0 −−−→
0 −−−→
W
y
Xa
y
0
y
0
y
Za −−−→ KLa −−−→ 0
y
y
−−−→ K s −−−→ KL −−−→ 0
y
y
−−−→ Va −−−→ Ya
y
y
−−−→ 0
0
0
0
∗
Now (5) implies that dimK (U )a = dimK Ya , and the diagram shows that dimK Ya =
dimK KL − dimK KLa .
It remains to prove the isomorphism (4). Observe that (U ∗ )a =P
{ϕ ∈ (F ∗ )a : ϕ(C) =
∗
0}, where C is the kernel of F → U. Let ϕ ∈ (F )a . Then ϕ = si=1 ϕ(gi )gi∗ , where
g1∗ , . . . , gs∗ is the basis of F ∗ dual to g1 , . . . , gs .
Since deg gi∗ = − deg dfvi = −h(vi ), it follows that ϕ ∈ (F ∗ )a if and only if
ϕ(gi ) = µ(i)ta+h(vi ) with µ(i) ∈ K and µ(i) = 0 if a + h(vi ) 6∈ H. Hence
X
µ(i)ta+h(vi ) gi∗ )(C) = 0}.
(U ∗ )a ∼
= {µ ∈ K s : µ(i) = 0 for i ∈ [s] \ Fa and (
i∈Fa
In order to complete the proof of (4) we only need to prove the following statement:
(6)
(
s
X
µ(i)ta+h(vi ) gi∗ )(C) = 0 if and only if hµ, λi = 0 for all λ ∈ W.
i=1
P
Let z ∈ Cb for some b ∈ H. Then z = i∈[s] λ(i)tb−h(vi ) gi with λ(i) ∈ K for i =
1, . . . , s and λ(i) = 0 if b−h(vi ) ∈
/ H since z ∈ Fa . Moreover, since z ∈ Ker(F → U)
b−h(v1 )
it follows that λ(1)t
dfv1 + · · · + λ(s)tb−h(vs ) dfvs = 0. This implies that
X X
X X
λ(i)vi (j)tb−hj )dxj = 0.
λ(i)tb−h(vi ) vi (j)th(vi )−hj dxj =
(
i∈[s]
b−h(vi )∈H
j∈[n]
j∈[n]
i∈[s]
b−h(vi )∈H
Note that if b − hj ∈
/ H, then for all i ∈ [s] with b − h(vi ) ∈ H, one has h(vi ) −
h
∈
/
H
and
so
v
(j)
= 0. Here we use the definition of P
h(vi ), see (1). Therefore,
i
Pj
i∈[s],b−h(vi )∈H λ(i)vi (j) = 0 for j = 1, . . . , n. This implies
i∈[s],b−h(vi )∈H λ(i)vi = 0.
In conclusion we see that
X
X
λ(i)tb−h(vi ) gi ∈ Cb if and only if
λ(i)vi = 0.
i∈[s],b−h(vi )∈H
i∈[s],b−h(vi )∈H
6
P
This particularly implies that if z = i∈[s] λ(i)tb−h(vi ) gi ∈ Cb , then λ = (λ(1), . . . , λ(s)) ∈
W.
Since
s
s
X
X
X
a+h(vi ) ∗
b−h(vi )
(
µ(i)t
gi )(
λ(i)t
gi ) = (
µ(i)λ(i))ta+b ,
i=1
i=1
i∈[s],b−h(vi )∈H
Ps
/ H or hµ, λi =
it
that ( i=1 µ(i)ta+h(vi ) gi∗)(Cb ) = 0 if and only if either a+b ∈
Pfollows
s
all i with b − h(vi ) ∈
/ H. In
i=1 µ(i)λ(i) = 0 for all λ ∈ W satisfying λ(i) = 0 forP
particular, we have if hµ, λi = 0 for all λ ∈ W , then ( si=1 µ(i)ta+h(vi ) gi∗ )(C) = 0.
P
a+h(vi ) ∗
gi )(C) = 0. Write a = a+ −a−
For the converse, we assume that ( si=1 µ(i)t
Ps
with a+ ∈ H and a− ∈ H, and set b0 = Pi=1 h(vi ) + a− . Since a + b0 ∈ H and
b0 − h(vi ) ∈ H for all i ∈ [s], and since ( si=1 µ(i)ta+h(vi ) gi∗ )(Cb0 ) = 0, it follows
that (µ, λ) = 0 for all λ ∈ W . Therefore the statement (6) has been proved and this
completes the proof.
Now for any a ∈ ZH we want to determine the dimension of (Im δ ∗ )a . We observe
that the ZH-graded R-module Im δ ∗ is generated by the elements
s
s
X
X
∗
∗
∗
δ ((dxi ) ) =
(∂fvj /∂xi )gj =
vj (i)th(vj )−hi gj∗ .
j=1
∗
j=1
∗
Note that deg δ ((dxi ) ) = −hi for i = 1, . . . , n.
For i = 1, . . . , n we set wi = (v1 (i), . . . , vs (i)), and for a ∈ ZH we let KDa be the
K-subspace of K s spanned by the vectors wi for which i ∈ Ga . Here the set Ga is
defined to be
Ga = {i ∈ [n] : a + hi ∈ H}.
Proposition 1.2. Let a ∈ ZH. Then
dimK (Im δ ∗ )a = dimK KDa .
Proof. The K-subspace (Im δ ∗ )a ⊂ (F ∗ )a is spanned by the vectors
s
X
a+hi ∗
∗
t
δ ((dxi ) ) =
vj (i)ta+h(vj ) gj∗
j=1
with i ∈ Ga .
The desired formula dimK (Im δ ∗ )a follows once we have shown that
X
X
µ(i)wi = 0.
µ(i)ta+hi δ ∗ ((dxi )∗ ) = 0 if and only if
i∈Ga
i∈Ga
Here µ(i) ∈ K for any i ∈ Ga . To prove it we notice that
s
s X
X
X
X
X
a+h(vj ) ∗
a+hi ∗
∗
µ(i)vj (i)ta+h(vj ) gj∗).
µ(i)(
vj (i)t
gj ) =
(
µ(i)t
δ ((dxi ) ) =
i∈Ga
P
i∈Ga
j=1
j=1 i∈Ga
P
Thus i∈Ga µ(i)ta+hi δ ∗ ((dxi )∗ ) = 0 if and only if P
i∈Ga µ(i)vj (i) = 0 for j = 1, . . . , s.
Since vj (i) = wi (j), this is the case if and only if i∈Ga µ(i)wi = 0.
7
Corollary 1.3. Let a ∈ ZH. Then dimK KDa + dimK KLa ≤ dimK KL. Equality
holds if and only if T 1 (R)a = 0.
Summarizing our discussions of this section we observe that all information which
is needed to compute dimK T 1 (R)a can be obtained from the (s × n)-matrix
v1 (1) v1 (2) . . . v1 (n)
v2 (1) v2 (2) . . . v2 (n)
.
AH =
..
..
...
.
.
vs (1) vs (2) . . . vs (n)
Indeed, dimK T 1 (K[H])a can be computed as follows: let l = rank AH , ra the rank
of the submatrix of AH whose rows are the ith rows of AH for which a + h(vi ) 6∈ H,
and let ca be the rank of the submatrix of AH whose columns are the jth columns
of AH for which a + hj ∈ H. Then
dimK T 1 (K[H])a = l − la − da .
(7)
Corollary 1.4. Suppose a ∈ H. Then T 1 (R)a = 0.
Proof. Since a ∈ H, it follows that G(a) = [n] and dimK Da = dimK KL = rank AH .
Thus the assertion follows from Corollary 1.3.
The inequality of Corollary 1.3 can also be deduced from the following lemma.
Lemma 1.5. Fix a ∈ ZH. Then vi (j) = 0 for every pair i, j with i ∈
/ Fa and
j ∈ Ga .
Proof. Assume on the contrary that vi (j) 6= 0, say vi (j) < 0, for some i ∈
/ Fa and
j ∈ Ga . Then
X
X
−vi (k)hk + (−vi (j) − 1)hj ∈ H.
vi (k)hk = hj + b, where b =
h(vi ) = −
k6=j
vi (k)<0
k
vi (k)<0
Since j ∈ Ga , we have a + hj ∈ H and so a + h(vi ) = (a + hj ) + b ∈ H. Consequently,
i ∈ Fa , a contradiction.
2. Separable and inseparable saturated lattices
In this section we study conditions under which an affine semigroup ring K[H]
is obtained from another affine semigroup ring K[H ′ ] by specialization, that is, by
reduction modulo a regular element. Of course we can always choose H ′ = H × N
in which case K[H ′ ] is isomorphic to the polynomial ring K[H][y] over K[H] in
the variable y, and K[H] is obtained from K[H ′ ] by reduction modulo the regular
element y. This trivial case we do not consider as a proper solution of finding an
K[H ′ ] that specializes to K[H]. If no non-trivial K[H ′ ] exists, which specializes to
K[H], then H will be called inseparable and otherwise separable. It turns out that
the separability of H is naturally phrased in terms of the relation lattice L of H.
8
Let L ⊂ Zn be a subgroup of Zn . Such a subgroup is often called a lattice. The
ideal IL generated by all binomials fv with v ∈ L is called the lattice ideal of L. The
following properties are known to be equivalent:
(i) Zn /L is torsionfree;
(ii) IL is a prime ideal;
(iii) there exists a semigroup H such that IL = IH .
A proof of these facts can be found for example in [7]. A lattice L for which Zn /L
is torsionfree is called a saturated lattice.
Let ε1 , . . . , εn be the canonical basis of Zn and ε1 , . . . , εn , εn+1 the canonical basis
of Zn+1 . Let i ∈ [n]. We denote by πi : Zn+1 → Zn the group homomorphism with
πi (εj ) = εj for j = 1, . . . , n and πi (εn+1 ) = εi . For convenience we denote again by
πi the K-algebra homomorphism S[xn+1 ] → S with πi (xj ) = xj for j = 1, . . . , n and
πi (xn+1 ) = xi .
Definition 2.1. Let L ⊂ Zn be a saturated lattice. We say that L is i-separable for
some i ∈ [n], if there exists a saturated lattice L′ ⊂ Zn+1 such that
(i) rank L′ = rank L;
(ii) πi (IL′ ) = IL ;
(iii) there exists a minimal system of generators fw1 , . . . , fws of IL′ such that the
vectors (w1 (n + 1), . . . , ws (n + 1)) and (w1 (i), . . . , ws (i)) are linearly independent.
The lattice L is called i-inseparable if it is not i-separable, and L is called inseparable
if it is i-inseparable for all i. Moreover, the lattice L′ satisfying (i)-(iii) is called an
i-separation lattice for L. We also call a semigroup H and its toric ring inseparable
if the relation lattice of H is inseparable.
Remark 2.2. Suppose that L′ ⊂ Zn+1 is an i-separation lattice for L. Let IL′ ⊂
S[xn+1 ] be the lattice ideal of L′ . It is easily seen that xn+1 − xi 6∈ IL′ because
rank L = rank L′ . Indeed, if xn+1 − xi ∈ IL′ , then S[xn+1 ]/IL′ ∼
= S/IL , and so
rank L′ = height IL′ = height IL + 1 = rank L + 1, contradicting Definition 2.1(i).
Moreover, xn+1 − xi is a non-zerodivisor of S[xn+1 ]/IL′ since S[xn+1 ]/IL′ is a domain. In particular, if fw1 , . . . , fws is a minimal system of generators of IL′ , then
πi (fw1 ), . . . , πi (fws ) is a minimal system of generators of IL , (see Lemma 2.5 for the
details). This implies that
wj (i)wj (n + 1) ≥ 0 for j = 1, . . . , s.
Indeed, xi divides πi (fwj ) if wj (i)wj (n + 1) < 0. Since πi (fwj ) is a minimal generator
of IL and since IL is a prime ideal, the polynomial πi (fwj ) must be irreducible. So,
wj (i)wj (n + 1) < 0 is not possible.
Let vj = πi (wj ) for j = 1, . . . , s. Since wj (i)wj (n + 1) ≥ 0 for j = 1, . . . , s, for all
j we have πi (fwj ) = fvj . Hence fv1 . . . , fvs is a minimal system of generators of IL .
For an affine semigroup H ⊂ Zm the semigroup ring K[H] is standard graded, if
and only if there exists a linear form ℓ = a1 z1 + a2 z2 + · · · + am zm in the polynomial
ring Q[z1 , . . . , zm ] such that ℓ(hi ) = 1 for all minimal generators hi of H.
9
The following result provides a necessary condition of i-inseparability. Recall from
[6] that an affine semigroup H is called positive if H0 = {0}, where H0 is the set of
invertible elements of H.
Theorem 2.3. Let H be a positive affine semigroup which is minimally generated
by h1 , . . . , hn , L ⊂ Zn the relation lattice of H. Suppose that L is i-separable. Then
T 1 (K[H])−hi 6= 0. In particular, if K[H] is standard graded, then L is inseparable,
if T 1 (K[H])−1 = 0.
Proof. Since L is i-separable, there exists a saturated lattice L′ satisfying the conditions (i) and (ii) as given in Definition 2.1. Since xn+1 − xi is a non-zerodivisor on
R′ = S[xn+1 ]/IL′ it follows that R′′ = R′ /(xn+1 − xi )2 R′ is an infinitesimal deformation of R (which is isomorphic to R′′ /((xn+1 − xi )R′′ ).
Let vj = πi (wj ) for j = 1, . . . , s. By Remark 2.2, we have πi (fwj ) = fvj for
j = 1, . . . , s and fv1 , . . . , fvs is a minimal system of generators of IL .
Note that S[xn+1 ] = S[xn+1 − xi ]. We set ε to be the residue class of xn+1 − xi in
S[xn+1 − xi ]/(xn+1 − xi )2 . Then S[xn+1 − xi ]/(xn+1 − xi )2 = S[ε]. Let σ : S[xn+1 ] →
S[ε] the canonical epimorphism and let J be the image of IL′ in S[ε]. Then R′′ =
S[ε]/J.
In order to determine the generators of J, we fix a j with 1 ≤ j ≤ s, and may
assume that w(n + 1) ≥ 0 and w(i) ≥ 0. Then modulo (xn+1 − xi )2 , we obtain
Y w (k) w (n+1)
Y w (k)
j
fwj =
−
xk j xn+1
xk j
1≤k≤n
wj (k)≥0
=
Y
1≤k≤n
wj (k)<0
w (k)
xk j
w (n+1)
(xi j
1≤k≤n
wj (k)≥0
= fvj + [wj (n + 1)(
Y
w (n+1)−1
+ wj (n + 1)xi j
ε) −
Y
w (k)
xk j
1≤k≤n
wj (k)<0
v (k)
xkj
)/xi ]ε.
1≤k≤n
vj (k)≥0
For the second equality we used that xn+1 = ε+xi and ε2 = 0, and the third equality
is due to the fact vj (i) = wj (i) + wj (n + 1).
The homomorphism ϕ : IL /IL2 → R corresponding to the infinitesimal deformation
S[ε]/J is given by
Y v (k)
xkj )/xi + IL = wj (n + 1)th(vj )−hi for j = 1, . . . , s,
ϕ(fvj + IL2 ) = wj (n + 1)(
1≤k≤n
vj (k)≥0
which induces the element
α=
X
wj (n + 1)th(vj )−hi gj∗ ∈ (U ∗ )−hi .
1≤j≤s
Since
it follows that G−hi = {i}, and this implies that (Imδ ∗ )−hi =
P H is positive
K 1≤j≤s vj (i)th(vj )−hi gj∗ , see Proposition 1.2. Assume α ∈ (Imδ ∗ )−hi . Then there
exists λ ∈ K such that
(w1 (n + 1), . . . , ws (n + 1)) = λ(v1 (i), . . . , vs (i)).
10
Since vj (i) = wj (i) + wj (n + 1) for j = 1, . . . , s, and since by condition (iii) of
Definition 2.1 the vectors (w1 (n+1), . . . , ws (n+1) and (w1 (i), . . . , ws (i)) are linearly
independent, we obtain a contradiction. Hence T 1 (R)−hi 6= 0, as required.
As a first example of a separable lattice we consider the relation lattice of a
numerical semigroup.
Discussion 2.4. Let H ⊂ N be the numerical semigroup minimally generated by
h1 , h2 , h3 with gcd(h1 , h2 , h3 ) = 1. Recall some facts from [8]. For i = 1, 2, 3 let ci
be the smallest integer such that ci hi ∈ Nhk + Nhℓ , where {i, k, ℓ} = [3], and let
rik and riℓ be nonnegative integers such that ci hi = rik hk + riℓ hℓ . Denote by L the
relation lattice of H. Then the three vectors
v1 = (c1 , −r12 , −r13 ),
v2 = (−r21 , c2 , −r23 ),
v3 = (−r31 , −r32 , c3 )
generate L. We have v1 + v2 + v3 = 0 if
(1) all rij 6= 0, or
(2) v1 = (c1 , −c2 , 0), v2 = (0, c2 , −c3 ) and v3 = (−c1 , 0, c3).
In case (1), fv1 , fv2 , fv3 is the unique minimal system of generators of IL . In case
(2), fv1 + fv2 + fv3 = 0, so that any two of the fvi ’s minimally generate IL .
An example for (1) is the semigroup with generators 3, 4 and 5, and example for
(2) is the semigroup with generators 6, 10 and 15. (3) If v1 + v2 + v3 6= 0, then there
exist distinct integers k, ℓ ∈ [3] such that vk + vℓ = 0 and rij 6= 0 for i ∈ [3] \ {k, ℓ}
and j ∈ {k, ℓ}. In this case IL is minimally generated by xci i − xrkik xrl il and xckk − xcℓℓ .
An example for (3) is the semigroup with generators 4, 5 and 6.
It is known and easy to prove that R P
= K[H] is not rigid. Indeed, since R
is quasi-homogeneous, the Euler relations ni=1 (∂f /∂xi )xi = (deg f )f imply that
there is an epimorphism χ : ΩR/K → m with χ(dxi ) 7→ thi where m = (th1 , th2 , th3 )
is the graded maximal ideal of R. Since rank ΩR/K = rank m = 1, it follows that
C = Ker χ is a torsion module. Thus we obtain the following exact sequence
0 → C → ΩR/K → m → 0,
which induces the long exact sequence
HomR (C, R) → Ext1R (m, R) → Ext1R (ΩR/K , R).
Since R is a 1-dimensional domain, R is Cohen-Macaulay, HomR (C, R) = 0 and
Ext1R (m, R) ∼
= m−1 /R 6= 0. It follows that Ext1R (ΩR/K , R) 6= 0. In other words, R is
not rigid.
Of course the same argument can be applied to any numerical semigroup generated
by more than 1 element.
We have seen that K[H] is not rigid. The next result shows that the relation
lattice of H is even i-separable for i ∈ [3] with T 1 (R)−hi 6= 0. To prove this we need
Lemma 2.5. Let L ⊂ Zn and L′ ⊂ Zn+1 be saturated lattices which satisfy the
conditions (i) and (ii) as given in Definition 2.1. Then
(a) IL and IL′ have the same number of minimal generators;
11
(b) fw1 , . . . , fws is a minimal system of generators of IL′ if and only if
πi (fw1 ), . . . , πi (fws ) is a minimal system of generators of IL .
Proof. (a) For any S[xn+1 ]/IL′ -module M we denote by M its reduction modulo
xn+1 − xi . The conditions (i) and (ii) of Definition 2.1 guarantee that xn+1 − xi is
a non-zerodivisor on S[xn+1 ]/IL′ and that S/IL ∼
= S[xn+1 ]/IL′ . From these facts (a)
follows.
(b) Suppose that fw1 , . . . , fws is a minimal system of generators of IL′ . Then IL
is generated by πi (fw1 ), . . . , πi (fws ) since πi (IL′ ) = IL . By (a), πi (fw1 ), . . . , πi (fws ) is
a minimal system of generators of IL .
Conversely, assume that πi (fw1 ), . . . , πi (fws ) is a minimal system of generators of
IL . We want to show that IL′ = (fw1 , . . . , fws ). Set J = (fw1 , . . . , fws ). Then we
obtain the following short exact sequence:
α
0 → IL′ /J → S[xn+1 ]/J → S[xn+1 ]/IL′ → 0
Here α is the natural epimorphism. By [6, Proposition 1.1.4], we obtain the exact
sequence
α
0 → IL′ /J → S[xn+1 ]/J −→ S[xn+1 ]/IL′ → 0.
Since πi (J) = πi (IL′ ) = IL it follows that α is an isomorphism, and so IL′ /J = 0.
Nakayama’s Lemma implies that IL′ /J = 0. Hence J = IL′ , as desired.
Proposition 2.6. Let H be a numerical semigroup as above and set R = K[H]. Let
L ⊂ Z3 be the relation lattice of H. With the notation of Discussion 2.4 we have:
(a) If v1 +v2 +v3 = 0, then dimK T 1 (R)−hi = 1 and L is i-separable for i = 1, 2, 3.
(b) If v1 + v2 + v3 6= 0, then there exists i ∈ [3] such that IL = (xci i − xrkik xrl il , xckℓ −
xcℓk ) with {i, k, l} = [3] and rik , ril 6= 0. In this case, T 1 (R)−hi = 0, and for
j 6= i we have that T 1 (R)−hj 6= 0 and that L is j-separable.
Proof. (a) We consider the case (1), where rij > 0 for all i and j, see Discussion 2.4.
Fix i ∈ [3]. Since all rij > 0 it follows that F−hi = {1, 2, 3}, and since H is a positive
semigroup we have G−hi = {i}. It follows from Corollary 1.1 and Proposition 1.2
∗
that dimK (U ∗ )−hi = 2 and dimK (Im(δ−h
) = 1. Hence dimK T 1 (R)−hi = 1.
i
Consider the vectors
w1 = (c1 − 1, −r12 , −r13 , 1),
w2 = (−r21 + 1, c2 , −r23 , −1),
w3 = (−r31 , −r32 , c3 , 0)
in Z4 , and set L′ = Zw1 + Zw2 + Zw3 . We will prove that L′ is a 1-separation of
L. First we show that L′ is saturated. Indeed, if aw ∈ L′ for some 0 6= a ∈ Z and
some w ∈ Z4 , then aw = a1 w1 + a2 w2 + a3 ws for some ai ∈ Z, and it follows that
av = a1 v1 +a2 v2 +a3 v3 , where v = π1 (w). This implies that v = k1 v1 +k2 v2 +k3 v3 for
some ki ∈ Z, since L is saturated. Thus (a1 −ak1 )v1 + (a2 −ak2 )v2 + (a3 −ak3 )v3 = 0
and so a1 − ak1 = a2 − ak2 = a3 − ak3 . It follows that (a1 − ak1 )w1 + (a2 − ak2 )w2 +
(a3 − ak3 )w3 = 0. Thus w = k1 w1 + k2 w2 + k3 w3 . Hence L′ is saturated. Next we
show π1 (IL′ ) = IL . It is clear that IL ⊆ π1 (IL′ ) since π1 (fwi ) = fvi for i = 1, 2, 3.
12
For the converse direction, we only need to note that π1 (L′ ) = L and that fπ1 (w)
divides π1 (fw ) for all w ∈ L′ .
Now, applying Lemma 2.5 we conclude that fw1 , fw2 , fw3 is a minimal system of
generators of IL′ satisfying condition (iii) of Definition 2.1. Consequently, L′ is a
1-separation of L. Similar arguments work for i = 2, 3.
In case (2), L is generated by any two of the vectors v1 = (c1 , −c2 , 0), v2 =
(0, c2 , −c3 ) and v3 = (−c1 , 0, c3 ). Let L′ ⊂ Z4 be a lattice generated by w1 =
(c1 − 1, c2 , 0, 1) and w2 = (−c1 , 0, c3 , 0). We claim that L′ is a 1-separation of L.
Indeed, the ideal of 2-minors I2 (W ) of the matrix W whose row vectors are w1
and w2 contains the elements c1 and c3 . By the choice of the ci ’s it follows that
gcd(c1 , c3 ) = 1. Thus, I2 (W ) = Z. This shows that L′ is saturated. Since π1 (fwi ) =
fvi for i = 1, 2 and since IL = (fv1 , fv2 ), Lemma 2.5 implies that IL′ = (fw1 , fw2 ) and
π1 (IL′ ) = IL . Since L′ satisfies also condition (iii) of Definition 2.1, it follows that
L is 1-separable. In the same way it is shown that L is i-separable for i = 2, 3.
(b) This is case (3) of Discussion 2.4 and we have IH = (xci i − xrkik xrl kl , xckk − xcℓℓ )
with {i, k, l} = [3]. Thus IH is a complete intersection and the exponents c1 , c2 and
c3 are all > 1. Without loss of generality we may assume that i = 2, k = 1 and
l = 3. Since the lattice L ⊂ Z3 with basis v1 = (−r21 , c2 , −r23 ), v2 = (−c3 , 0, c1 ) is
saturated, it follows that the ideal of 2-minors (c1 c2 , c2 c3 , c1 r21 + c3 r23 ) of
−r21 c2 −r23
−c3 0
c1
is equal to Z.
Consider the lattice L′ ⊂ Z4 whose basis w1 , w2 consists of the row vectors of
−r21 + 1 c2 r23 −1
.
−c3
0 c1 0
The ideal of 2-minors of this matrix contains (c1 c2 , c2 c3 , c1 r21 + c3 r23 ), and hence
is again equal to Z. Thus L′ is saturated. Furthermore we have π2 (fw1 ) = fv1 ,
π2 (fw2 ) = fv2 and π2 (L′ ) = L. This implies that π2 (IL′ ) = IL . Since rank L′ =
rank L = 2, the conditions (i) and (ii) of Definition 2.1 are satisfied. Applying
Lemma 2.5 we obtain IL′ = (fw1 , fw2 ). Since the condition (iii) of Definition 2.1 is
also satisfied we see that L is 1-separable. Similarly, one shows that L is 3-separable.
3. Edge rings of bipartite graphs
Let G be a finite simple graph on the vertex set [m], and let K a field. The Kalgebra R = K[G] = K[ti tj : {i, j} ∈ E(G)] is called the edge ring of G. Here E(G)
denotes the set of edges of G. We let n = |E(G)|, and denote by S the polynomial
ring over K in the indeterminates xe with e ∈ E(G). Let ϕ : S → K[G] be the
K-algebra homomorphism with xe 7→ ti tj for e = {i, j}. The toric ideal Ker ϕ will
be denoted by IG .
In this section we will discuss inseparability, semi-rigidity and rigidity of the edge
ring of a bipartite graph, which may as well be considered as the toric ring associated
with the affine semigroup H generated by the elements δi + δj with {i, j} ∈ E(G),
13
where δ1 , . . . , δm is a canonical basis of Zm . Let G be a bipartite graph. The
generators of IG are given in terms of even cycles of G. Recall that a walk in G is a
sequence C : i0 , i1 , . . . , iq such that {ik , ik+1 } is an edge of G for k = 0, 1, . . . , k − 1.
C is called a closed walk, if iq = i0 . The closed walk C is called a cycle if ij 6= ik
for all j 6= k with j, k < q, and it is called an even closed walk if q is even. Observe
that any cycle of bipartite graph is an even cycle.
Given any even cycle (more generally an even closed walk) C : i0 , i1 , . . . , i2q . The
edges of C are ekj = {ij , ij+1 } for j = 0, 1, . . . , 2q − 1 together with the edge
ek2q−1 = {i2q−1 , i0 }. We associate to C the vector v(C) ∈ Zn which defined as
(8)
v(C) =
q−1
X
εk2i −
q−1
X
εk2i+1
i=0
i=0
Here ε1 , . . . , εn denotes the canonical basis of Zn . Note that v(C) is determined by
C only up to sign. We call v(C) as well as −v(C) the vector corresponding to C.
For simplicity we write fC for fv(C) . Recall from [11] that the toric ideal IG of a
finite bipartite graph is minimally generated by indispensable binomials, that is, by
binomials, which up to sign, belong to any system of generators of IG . Furthermore,
a binomial f ∈ IG is indispensable if and only if f = fC , where C is an induced
cycle, that is, a cycle without a chord. In particular, if G′ is the graph obtained
from G by deleting all edges which do not belong to any cycle, then IG = IG′ S.
Therefore we may assume throughout this section that each edge of G belongs to
some cycle.
Now for the rest of this section we let G be a bipartite graph on the vertex set [m]
with edge set E(G) = {e1 , . . . , en }. With the edge ek = {i, j} we associate the vector
hk = δi + δj . Here δ1 , . . . , δm is the canonical basis of Zm . The semigroup generated
by h1 , . . . , hn we denote by H(G) or simply by H. Note that K[H(G)] = K[G].
Let {C1 , . . . , Cs } be the set of cycles of G and vi = v(Ci ) the vector corresponding
to Ci . We may assume that for i = 1, . . . , s1 ≤ s, the cycles Ci are the induced
cycles of G. Then IG is minimally generated by fv1 , . . . , fvs1 , see [11]. Of course, IG
is also generated by fv1 , . . . , fvs . In particular, if L is the relation lattice of H, then
KL is the vector space spanned by v1 , . . . , vs .
Let a ∈ ZH. As in Section 1 we set
Fa = {1 ≤ i ≤ s : a + h(vi ) ∈ H}, and KLa = SpanK {vi : i ∈ [s] \ Fa }.
In addition we now also set
Fa′ = {1 ≤ i ≤ s1 : a + h(vi ) ∈ H}, and KL′a = SpanK {vi : i ∈ [s1 ] \ Fa′ }.
In general, Fa′ is a proper subset of Fa . However, we have
Lemma 3.1. KL′a = KLa for all a ∈ ZH.
Proof. Since [s1 ] \ Fa′ ⊆ [s] \ Fa , we have KL′a ⊆ KLa . Let i ∈ ([s] \ Fa ) \ ([s1] \ Fa′ ).
Then a + h(vi ) ∈
/ H and Ci is a cycle with chords. In the following we describe a
process to obtain the induced cycles with vertex set contained in V (Ci ). Choose a
chord of Ci and note that this chord divides Ci into two cycles. If both cycles are
14
induced, then the process stops. Otherwise we divide as before, those cycles which
are not induced. Proceeding in this way, we obtain induced cycles of G, denoted
by Ci1 , . . . , Cik , such that E(Cij ) consists of at least one chord of Ci . Moreover, the
edges of Cij which are not chords of Ci , are edges of Ci .
In general, if C is a cycle and v = v(C), then
X
h(v) =
δj .
j∈V (C)
Hence it follows from the construction of the induced cycles Cij that h(vi ) − h(vij )
is the sum of certain terms δk1 + δk2 , where {k1 , k2 } is an edge of Ci , and hence
h(vi ) − h(vij ) ∈ H. Since a + h(vi ) ∈
/ H it follows that a + h(vij ) ∈
/ H for all j. This
′
′
implies that vij ∈ KLa for all j, and so vi ∈ KLa since vi is a linear combination of
the vij .
For the discussion on separability we need to know when T 1 (K[G])−hj vanishes,
see Theorem 2.3. For that we need to have the interpretation of F−hj for edge rings
which is given by the following formula:
(9)
F−hj = {i ∈ [n] : V (ej ) ⊂ V (Ci )}.
For the proof of this equation note that if V (ej ) ⊆ V (Ci ), then without loss of
generality we assume that Ci : 1, 2, . . . , 2t and that ej = {1, k} with k ∈ [2t].
Note that k is even, since G contains no odd cycle. It follows that −hj + h(vi ) =
(δ2 + δ3 ) + · · · + (δk−2 + δk−1 ) + (δk+1 + δk+2 ) + · · · + (δ2t−1 + δ2t ) ∈ H, and so i ∈ F−hj
by definition. Conversely, assume that V (ej ) * V (Ci ) and let k ∈ V (ej ) \ V (Ci ).
Then −hj + h(vi ) is a vector in Zm with the kth entry negative and thus it does not
belong to H. Therefore i ∈
/ F−hj .
Later we also shall need
Lemma 3.2. Let W : i1 , i2 , . . . , i2k , i1 be an even closed walk in G and let ej be an
edge of G with the property that ej 6= {ia , ib } with 1 ≤ a < b ≤ 2k. Then the vector
w = v(W ) ∈ KL belongs to KL−hj .
Proof. We may view W as a bipartite graph with bipartition {i1 , i3 , . . . , i2k−1 } and
{i2 , i4 , . . . , i2k }. Then we see that w belongs to the space spanned by the vectors
corresponding to the induced cycles of G with edges in W . This vector space is a
subspace of KL−hj , since ej is not an edge of any cycle with edges in W , as follows
from (9).
We call the space KL which is spanned by the vectors v1 , . . . , vs the cycle space of
G (with respect to K). Usually the cycle space is only defined over Z2 . For bipartite
graphs the dimension of the cycle space does not depend on K and is known to be
(10)
|E(G)| − |V (G)| + c(G).
where c(G) is the number of connected components of G, see [13, Corollary 8.2.13].
Inseparability. In this subsection we present some characterizations of bipartite
graphs G for which K[G] is inseparable.
15
Note that (9) says that i ∈ F−hj if and only if ej is an edge or a chord of Ci .
Accordingly, we split the set F−hj into the two subsets
(11)
Aj = {i ∈ [s] : ej is an edge of Ci },
and
(12)
Bj = {i ∈ [s] : ej is a chord of Ci }.
We also set V−hj = SpanK {vi : i ∈ [s] \ Aj }. Then, since by assumption all edges
of G belong to a cycle, we obtain
(13)
dimK V−hj = dimK KL − 1 for j = 1, . . . , n.
Indeed, let G\{ej } be the graph obtained from G by deleting the edge ej and leaving
vertices unchanged. Then V−hj is the cycle space of G \ {ej }.
Lemma 3.3. T 1 (R)−hj = 0 if and only if for all i ∈ Bj , one has vi ∈ KL−hj .
Proof. Since −hj + hi ∈ H if and only if i = j it follows that dim(Im δ ∗ )−hj = 1, see
Proposition 1.2. Thus, since KL−hj ⊆ V−hj , it follows from (13) that T 1 (R)−hj = 0
if and only if V−hj = KL−hj . Since V−hj = KL−hj + SpanK {vi : i ∈ Bj }, the
assertion follows.
For stating the next result we have first to introduce some concepts. Let C be
a cycle. Then the path P : i1 , i2 , i3 , . . . , ir−1 , ir (with r ≥ 2 and with ij 6= ik for all
j 6= k) is called a path chord of C if i1 , ir ∈ V (C) and ij 6∈ V (C) for all j 6= i1 , ir .
The vertices i1 and ir are called the ends of P . Note that any chord of C is a path
chord.
Let P be a path chord of C. We may assume that {i, i + 1} for i = 1, . . . , t
together with {1, 2t} are the edges of C and that i1 = 1 and ir = k with k 6= 1.
Let P ′ be another path chord of C. Then we say that P and P ′ cross each other
if one end of P ′ belongs to the interval [2, k − 1] and the other end of P ′ belongs
to [k + 1, 2t]. In particular, if P is a chord and P ′ crosses P , we say that P ′ is a
crossing path chord of C with respect to the chord P .
Theorem 3.4. Let G be a bipartite graph with edge set {e1 , . . . , en }, and let R =
K[G] be the edge ring of G. Then the following conditions are equivalent:
(a) T 1 (R)−hj 6= 0.
(b) There exists a cycle C of G for which ej is a chord, and there is no crossing
path chord P of C with respect to ej .
(c) The relation lattice of H(G) is j-separable.
Proof. (a) ⇒ (b): Assume that (b) does not hold. Let i ∈ Bj with Bj as defined in
12. By our assumption, Ci admits a path chord, denoted by P , which crosses ej .
Denote by i1 , i2 the two ends of P . Then C is the union of two paths P1 and P2
which both have ends i1 , i2 . Since P1 ∪ P and P2 ∪ P are cycles and ej is neither an
edge nor a chord of them, it follows from Lemma 3.2 that the vectors w1 = v(P1 ∪P )
and w2 = v(P2 ∪ P ) belong to KL−hj . Therefore, vi ∈ KL−hj because it is a linear
combination of w1 and w2 . Now applying Lemma 3.3, we obtain T 1 (R)−hj = 0, a
contradiction.
16
(b) ⇒ (c): We may assume that the cycle C as given in (b) has the edge set
E(C) = {e1 = {1, 2}, . . . , eℓ = {ℓ, ℓ + 1}, . . . , e2t = {2t, 1}},
and that ej = {1, k} with 2 < k < 2t − 1.
We let X be the set of all a ∈ [m] \ V (C) for which there is a path P from a to
some vertex of [2, k − 1], and we set Y = [m] \ (V (C) ∪ X).
We now define a graph G′ = G1 ∪ G2 , where G1 and G2 are disjoint graphs, that
is, V (G1 ) ∩ V (G2 ) = ∅. The graph G2 is the subgraph of G induced on X ∪ [k]. Next
e1 as the subgraph of G induced on Y ∪ [k + 1, 2t] ∪ {1, k}. Then
we first define G
e 1 by renaming 1 as m + 1 and k as m + 2. We claim that G1
G1 is obtained from G
and G2 are disjoint. Indeed, V (G1 ) ∩ V (G2 ) ⊆ [k + 1, 2t] ∩ X. Condition (b) implies
that [k + 1, 2t] ∩ X = ∅.
Now we claim that if we identify in G′ the vertex m + 1 with 1 and the vertex
m + 2 with k, then we obtain G. Indeed, let G′′ be the graph which is obtained
from G′ after this identification. We have to show that G′′ = G. Obviously, we have
V (G′′ ) = V (G) and E(G′′ ) ⊆ E(G). Let e ∈ E(G) \ E(G′′ ). Then e = {k1 , k2} with
k1 ∈ [k + 1, 2t] ∪ Y and k2 ∈ X ∪ [2, k − 1]. If k2 ∈ [2, k − 1], then k1 ∈ [1, k] ∩ X
by the definition of X. This is impossible since (X ∪ [1, k]) ∩ ([k + 1, 2t] ∪ Y ) = ∅;
If k2 ∈ X, then again by the definition of X it follows that k1 ∈ X ∪ [1, k], which is
impossible again in the same reason. Thus we have proved the claim.
Now the edge ring of G′ is of the form R′ = S ′ /IG′ = S ′ /(IG1 + IG2 )S ′ , where
S ′ = S[xn+1 ] and where the variable xn+1 corresponds to the edge en+1 = {m +
1, m + 2}. The variable xj corresponds to the edge ej if ej ∈ G, and to e ∈ E(G1 )
if e ∈ E(G1 ) \ {m + 1, m + 2} and e is mapped to ej by the identification map
G′ → G′′ = G. Let L be the relation lattice of H(G) and L′ be the relation lattice of
H(G′ ). Then L ⊂ Zn and L′ ⊂ Zn+1 are saturated lattices. We claim that L and L′
satisfy the conditions (i), (ii) and (iii) with respect to πj , see Definition 2.1. We first
show that πj (IL′ ) = IL . Let f be a minimal generator of IL . Then there exists an
e1 )
induced cycle D of G such that f = fD . Since G = G′′ it follows that V (D) ⊂ V (G
′
′
or V (D) ⊂ V (G2 ). Hence there is an induced cycle D in G whose image under the
identification map is D. Therefore, πj (fD′ ) = fD . This proves the condition (ii).
Since (ii) is satisfied, it follows that R′ /(xn+1 − xj )R′ ∼
= R. Moreover, xn+1 − xj is a
′
′
non-zerodivisor on R , since R is a domain. This implies that height IL′ = height IL .
In particular, rank L′ = rank L. Thus the condition (i) is also satisfied. Finally, by
the definition of G1 and G2 , there exist an induced cycle of G1 with en+1 as an edge,
say C1 , and an induced cycle of G2 with ej as an edge, say C2 . Let w1 = v(C1 ) and
w2 = v(C2 ). Then w1 (n + 1) 6= 0, w1 (j) = 0, w2 (n + 1) = 0 and w2 (j) 6= 0. This
implies the condition (iii).
The implication (c) ⇒ (a) follows from Theorem 2.3.
Corollary 3.5. Let G be a bipartite graph. Then K[G] is inseparable if and only
for any cycle C which has a unique chord e, there exists a crossing path chord of C
with respect to e. In particular, if no cycle of G has chord, then G is inseparable.
17
Proof. By Theorem 3.4, G is inseparable if and only if for any cycle C with a chord
e, there exists a crossing path chord of C with respect to e. Assume first that K[G]
is inseparable. Then, by what we just said, each cycle of G with a chord has the
desired property. Conversely, assume that each cycle with a unique chord has a
crossing path chord, and let C be a cycle with a chord e. Suppose C has another
chord, say e′ . If e′ crosses e, then we are done. Otherwise e′ divides C into two
smaller cycles C1 and C2 , and we may assume that e is a chord of C1 . Since C1 has
less chords than C, we may apply induction on the number of chords of a cycle and
deduce that there exists a crossing path chord of C1 with respect to e. Then this
path chord is also a crossing path chord of C with respect to e.
As an example of the theory which we developed so far we consider coordinate
rings of convex polyominoes. First we recall from [9] the definitions and some facts
about convex polyominoes.
Let R2+ = {(x, y) ∈ R : x, y ≥ 0}. We consider (R+ , ≤) as a partially ordered set
with (x, y) ≤ (z, w) if x ≤ z and y ≤ w. Let a, b ∈ Z2+ . Then the set [a, b] = {c ∈
Z2+ : a ≤ c ≤ b} is called an interval.
A cell C is an interval of the form [a, b], where b = a + (1, 1). The elements of C
are called vertices of C. We denote the set of vertices of C by V (C). The intervals
[a, a + (1, 0)], [a + (1, 0), a + (1, 1)], [a + (0, 1), a + (1, 1)] and [a, a + (0, 1)] are called
edges of C. The set of edges of C is denoted by E(C).
Let P be a finite collection of cells of Z2+ . Then two cells C and D are called
connected if there exists a sequence C : C = C1 , C2 , . . . , Ct = D of cells of P such
that for all i = 1, . . . , t − 1 the cells Ci and Ci+1 intersect in an edge. If the cells in C
are pairwise distinct, then C is called a path between C and D. A finite collection of
cells P is called a polyomino if every
S two cells of P are connected. The vertex set of
P, denoted V (P),
S is defined to be C∈P V (C) and the edge set of P, denoted E(P),
is defined to be C∈P E(C). A polyomino is said to be vertically or column convex
if its intersection with any vertical line is convex. Similarly, a polyomino is said to
be horizontally or row convex if its intersection with any horizontal line is convex. A
polyomino is said to be convex if it is row and column convex. Figure 1 shows two
polyominos whose cells are marked by gray color The right hand side polyomino is
convex while the left one is not.
Figure 1.
18
Let P be a polyomino, and let K be a field. We denote by S the polynomial over
K with variables xij with (i, j) ∈ V (P). A 2-minor xij xkl − xil xkj ∈ S with i < k
and j < l is called an inner minor of P if all the cells [(r, s), (r + 1, s + 1)] with
i ≤ r ≤ k − 1 and j ≤ s ≤ l − 1 belong to P. The ideal IP ⊂ S generated by all
inner minors of P is called the polyomino ideal of P. We also set K[P] = S/IP . It
has been shown in [9] that K[P] is a domain, and hence a toric ring, if P is convex.
A toric parametrization of K[P] will be given in the following proof.
Theorem 3.6. Let P be a convex polyomino. Then k[P] is inseparable.
Proof. Set AP = {hi : (i, j) ∈ V (P) for some j ∈ Z+ } and BP = {vj : (i, j) ∈
V (P) for some i ∈ Z+ }. We associate with P a bipartite graph G(P) such that
V (G(P)) = AP ∪ BP and E(G(P)) = {{hi , vj } : (i, j) ∈ V (P)}. Figure 2 shows a
polyomino and its associated bipartite graph.
h1
(4,4) (5,4)
(2,3) L(3,3) (4,3)
h2
h3
h4
h5
(5,3)
(1,2) (2,2)
J
(1,1) (2,1)
(3,2) (4,2) (5,2)
v1
(3,1)
v2
v3
v4
Figure 2.
We let K[G(P)] be the subring of the polynomial ring T = K[AP ∪ BP ] generated
by the monomials hi vj with {hi , vj } ∈ E(G(P)). In other words, K[G(P)] is the
edge ring of the bipartite graph G(P). Let, as above, S = K[xij : (i, j) ∈ V (P)]. As
shown in [9], IP is the kernel of the K-algebra homomorphism S → K[G(P)] with
xij 7→ hi vj . Thus K[P] ∼
= K[G(P)], and K[G(P)] is the desired toric parametrization. It is known from [11] that IP is generated by the binomials corresponding to
the cycles in G(P).
By using Corollary 3.5 it is enough to show that for any cycle C of G(P) which
has a unique chord, say e = {hi , vj }, there is a crossing path chord of C with respect
to e. Since G(P) is a bipartite graph, C is an even cycle, and also |C| > 4 because C
has a chord. Since every induced cycle of G(P) is a 4-cycle and since C has only one
chord, C must be a 6-cycle. Assume that the vertices of C are hi , vk1 , hℓ1 , vj , hℓ2 , vk2 ,
listed counterclockwise, and the chord of C is e = (hi , vj ) as above. With the
notation introduced, it follows that
(i, j), (i, k2 ), (ℓ2 , k2 ), (ℓ2 , j), (ℓ1, j), (ℓ1 , k1 , ), (i, k1)
are vertices of P. We consider the following cases.
Suppose first that (ℓ1 − i)(ℓ2 − i) > 0. Without loss of generality, we may assume
ℓ2 > ℓ1 > i. Then, since P is convex and (i, k2 ) and (ℓ2 , k2 ) are both vertices of P,
we have (ℓ1 , k2 ) is a vertex of P. It follows that {hℓ1 , vk2 } is an edge of G(P) which
19
is a chord of C, contradicting our assumption that C has a unique chord. Similarly
the case that (k1 − j)(k2 − j) > 0 is also not possible.
It remains to consider the case when (ℓ1 − i)(ℓ2 − i) < 0 and (k1 − j)(k2 − j) < 0.
Without loss of generality we may assume that ℓ1 < i < ℓ2 and k1 < j < k2 . Then
either (i−1, j +1) or (i+1, j −1) is a vertex of P by the connectedness and convexity
of P.
We may assume that (i − 1, j + 1) ∈ V (P). Note that (i − 1, k1 ) and (ℓ2 , j + 1)
belong to V (P). Thus we obtain the path vk1 , hi−1 , vj+1 , hℓ2 in G(P) which is a
crossing path chord of C with respect to e.
Semi-rigidity. We say that R is semi-rigid if T 1 (R)a = 0 for all a ∈ ZH with
−a ∈ H. In this subsection we consider this weak form of rigidity which however is
stronger than inseparability.
We again let G be a finite bipartite graph on the vertex set [m] with edge set
E(G) = {e1 , e2 , . . . , enP
}. The edge ring of G is the toric ring K[H] whose generators
are the elements hi = j∈V (ei ) δj , i = 1, . . . , n. Here δ1 , . . . , δm is the canonical basis
of Zm . As above we may assume that each edge of G belongs to a cycle and that
C1 , C2 , . . . Cs is the set of cycles of G and where C1 , . . . , Cs1 is the set of induced
cycles of G.
Let Ci be one of these cycles with edges ei1 , ei2 , . . . , ei2t labeled counterclockwise.
Two distinct edges e and e′ of Ci are said to be of the same parity in Ci if e = eij
and e′ = eik with j − k an even number.
Lemma 3.7. Let a = −hj − hk , and let i ∈ [s1 ]. Then i ∈ Fa′ , if and only if ej and
ek have the same parity in Ci . Moreover, if Fa′ 6= ∅, then KLa = KL−hj + KL−hk .
Proof. Since i ∈ [s1 ], the cycle Ci is an induced cycle.
, . . . , ei2t be the
P Let ei1 , ei2P
edges of Ci labeled counterclockwise. Then h(vi ) = tk=1 hi2k−1 = tk=1 hi2k . Thus
if ej and ek have the same parity in Ci , it follows that hj and hk belong to either one
of the above summands, so that a + h(vi ) ∈ H. This shows that i ∈ Fa′ . Conversely,
suppose that i ∈ Fa′ . Let hj = δj1 + δj2 and hk = δk1 + δk2 . For simplicity, we
may assume that δ1 , . . . , δ2t correspond to the vertices of Ci and that the edges of
Ci correspond to the elements δ2t + δ1 and δi + δi+1 for i = 1, . . . , 2t − 1. Then
h(vi ) = δ1 + · · · + δ2t and
a + h(vi ) = −δj1 − δj2 − δk1 − δk2 + δ1 + . . . + δ2t ∈ H.
P
In general, let h ∈ H, h = m
i=1 zi δi with zi ∈ Z. Then it follows that zi ≥ 0 for all
i. Hence it follows from (14) that ej and ek are edges of Ci with V (ej ) ∩ V (ek ) = ∅
(that is, the vertices j1 , j2 , k1 , k2 are pairwise different), and that a + h(vi ) is the
sum of all δi , i = 1, . . . , 2t with i 6= j1 , j2 , k1, k2 . Suppose the edges ej and ek do not
have the same parity in Ci . Then a + h(vi ) is the sum of S1 and S2 , where each of
S1 and S2 consists of an odd sum of δi . Hence none of these summands belongs to
H. Since S1 + S2 ∈ H, there exists a summand δr1 in S1 and a summand δr2 in S2
such that δr1 + δr2 ∈ H. This implies that {r1 , r2 } ∈ E(Ci ) because Ci has no chord.
However this is not possible. Indeed, if {r1 , r2 } ∈ E(Ci ), then r2 ≡ r1 + 1 mod 2t.
But this is not the case.
(14)
20
′
′
,
Next we show that KLa = KL−hj +KL−hk if Fa′ 6= ∅. Note that Fa′ ⊆ F−h
∩F−h
j
k
we have KL−hj + KL−hk ⊆ KLa by Lemma 3.1. In order to obtain the desired
′
equality, we only need to show that vi ∈ KL−hj + KL−hk for each i ∈ (F−h
∩
j
′
′
F−hk ) \ Fa .
′
′
) \ Fa′ . Since Fa′ 6= ∅, there exists an induced cycle, say C,
Let i ∈ (F−h
∩ F−h
j
k
such that ej and ek have the same parity in C. We may assume that V (C) = [2t]
and E(C) = {{1, 2}, {2, 3}, . . . , {2t − 1, 2t}, {2t, 1}}, and that ej = {1, 2} and ek =
{2k − 1, 2k} with 1 < k ≤ t. Since ej , ek do not have the same parity in Ci , we can
assume without loss of generality that E(Ci ) is
{{1, 2}, {2, i1}, {i1 , i2 }, . . . , {i2h , i2h+1 }, {i2h+1 , 2k}, {2k, 2k − 1}}
∪{{2k − 1, i2h+2 }, . . . , {i2ℓ , i2ℓ+1 }, {i2ℓ+1 , 1}}.
Then we have even closed walks
W1 : 2, 3, . . . , (2k − 1), 2k, i2h+1 , i2h . . . , i1 , 2
and
W2 : 1, 2, 3, . . . , (2k − 1), i2h+2 , . . . , i2ℓ+1 , 1.
Let w1 = v(W1 ) and w2 = v(W2 ). Since the vertex 1 belongs to ej but is not a vertex
of W1 , Lemma 3.2 implies that w1 ∈ KL−hj . Similarly it follows that w2 ∈ KL−hk .
Since vi differs at most by a sign from either w1 − w2 or w1 + w2 , it follows that
vi ∈ KL−hj + KL−hk , as required.
′
′
. Then KL−hj 6= KL−hk .
6= F−h
Lemma 3.8. Suppose that F−h
j
k
′
′
Proof. Let i ∈ F−h
\ F−h
. Then vi ∈ KL−hk and vi (j) 6= 0, since ej is an edge of
j
k
Ci . However the vectors v which belong to KL−hj have the property that v(j) = 0.
Hence vi ∈ KL−hk \ KL−hj , and this implies KL−hj 6= KL−hk .
Corollary 3.9. Assume that K[G] is inseparable. Let a = −hj − hk . Then
′
′
′
dimK KLa = dimK KL − 1 if F−a
6= ∅ and F−h
= F−h
.
j
k
Otherwise, dimK KLa = dimK KL.
Proof. Since we assume that G is inseparable, it follows from Corollary 1.1 and
Proposition 1.2 that dimK KL − dimK KL−hj = dimK (Im δ ∗ )−hj . Since by assumption each edge of G belongs to a cycle, it follows that dimK (Im δ ∗ )−hj = 1.
Thus dimK KL−hj = dimK KL − 1. Similarly, dimK KL−hk = dimK KL − 1. If
F ′ −hj = F ′ −hk , then KL−hj = KL−hk , and if moreover, F ′−a 6= ∅, then together
with Lemma 3.7 we have dimK KLa = dimK KL − 1, as desired.
Otherwise, there are two cases to consider. If F ′ −a = ∅, then KLa = KL, by
the definition of KLa and by Lemma 3.1. If F ′ −a 6= ∅ and F ′ −hj 6= F ′ −hk , then
KLa = KL−hj + KL−hk = KL, using Lemma 3.7 together with Lemma 3.8.
Theorem 3.10. Let G be a bipartite graph such that R = K[G] is inseparable. Then
the following statements are equivalent:
(a) K[G] is not semi-rigid;
21
(b) there exist edges e, f and an induced cycle C such that e, f have the same
parity in C and for any other induced cycle C ′ , e ∈ E(C ′ ) if and only if
f ∈ E(C ′ ).
Proof. (b) ⇒ (a): Let a = −g − h, where g and h are vectors in H corresponding to
the edges e and f respectively. Then dimK KLa = dimK KL − 1 by Corollary 3.9.
Note that Ga = ∅, we have (Im δ ∗ )a = 0. Therefore T 1 (R)a 6= 0 by Corollary 1.3,
and in particular, R is not semirigid.
P
(a) ⇒ (b): By assumption, there exists a = i∈[n] −ai hi ∈ ZH with ai ≥ 0 for
i = 1, · · · , n such that T 1 (R)a 6= 0. Note that ai ∈ {0, 1}, for otherwise, Fa′ = ∅ and
so KLa = KL. In particular T 1 (R)a = 0, a contradiction. Since R is inseparable,
it follows that |{i : ai 6= 0}| ≥ 2. If |{i : ai 6= 0}| = 2, then a = −hk − hj for
′
′
some 1 ≤ i 6= j ≤ n. Therefore, Fa′ 6= ∅ and F−h
= F−h
by Corollary 1.1 and
j
k
Corollary 3.9.
Let e and f be the edges corresponding to the vectors hj and hk , respectively.
Then, since Fa′ 6= ∅, there exists an induced cycle C of G such that e and f have
′
′
implies that for any
the same parity in C, by Lemma 3.7. Moreover, F−h
= F−h
j
k
′
′
′
induced cycle C of G, e ∈ E(C ) if and only if f ∈ E(C ).
Now suppose that |{i : ai 6= 0}| ≥ 3. Then there exists j and k with aj 6= 0 and
ak 6= 0, and we set b = −hj −hk . Note that Fa′ ⊆ Fb′ . This implies that KLb ⊆ KLa .
Therefore, since (Im δ ∗ )a = (Im δ ∗ )b = 0, we have T 1 (R)b 6= 0, and we are in the
previous case.
Corollary 3.11. Let P be a convex polyomino. Then K[P] is semi-rigid if and only
if P contains more than one cell.
Proof. Assume that P contains a unique cell. Then G(P) is a square and it is not
semi-rigid by Theorem 3.10.
Conversely, assume that K[P] is not semi-rigid. Then there exist two edges e, f
and an induced cycle C of G(P) satisfying the condition (b) in Theorem 3.10. Let
(i, j) and (k, ℓ) be vertices of P corresponding to the edge e and f , respectively.
Then the two edges of C other than e and f correspond to the vertices (i, ℓ) and
(k, j) of P. It follows that k 6= i and ℓ 6= j. Without loss of generality, we may
assume that k > i and ℓ > j. Then (i + 1, j + 1) ∈ V (P). Let C ′ be the induced
cycle of G(P) corresponding to the cell [(i, j), (i + 1, j + 1)] of P. Since C ′ contains
the edge e, C ′ must contain f by the condition (b) and thus k = i + 1 and ℓ = j + 1.
We claim that [(i, j), (i + 1, j + 1)] is the only cell of P. Suppose that this is not the
case. Then we let Ct , t = 1, 2, 3, 4 be four cells which share a common edge with
the cell [(i, j), (i + 1, j + 1)]. Note that P contains at least one of the Ct . Indeed,
since P is connected and since by assumption P contains a cell C different from
[(i, j), (i + 1, j + 1)], there exists a path in P between the cell [(i, j), (i + 1, j + 1)] and
C. This path must contain one of the Ct . However V (Ct ) contains exactly one of
the two vertices (i, j) and (i + 1, j + 1) for t = 1, . . . , 4. In other words, there exists
an induced cycle of G(P) which contains exactly one of the edges e and f . This is
contradicted to the condition (b) and thus our claim has been proved.
22
Classes of bipartite graphs which are semi-rigid or rigid. Let Gn be the bipartite
graph on vertex set [2n] with edge set
E(Gn ) = {{i, j} : i, j ∈ [2n], i − j is odd and {i, j} =
6 {1, 2n}}.
Thus Gn is obtained from the complete bipartite graph Kn,n by deleting one of its
edges. We observe that for n ≥ 3, Gn = G(Pn ) where Pn is the polyomino with
V (Pn ) = {(i, j) : 1 ≤ i, j ≤ n, (i, j) 6= (n, n)}.
For any edge e = {i, j} ∈ E(G) we use h(e) to denote the vector δi + δj ∈ Z2n .
Let C be a cycle. Then, as before, v(C) stands for the vector corresponding to C
and V (C) denotes the set of vertices of C. Our main result of this subsection is the
following:
Proposition 3.12. Let R be the edge ring of Gn .
(a) If n = 3, then R is semirigid, but not rigid.
(b) If n ≥ 4, then R is rigid.
We need some preparations. First, we introduce some notation.
P
Let a = i∈[2n] ai δi ∈ Z2n with ai ∈ Z for i = 1, . . . , 2n. We set
X
X
ai
and
ao =
ai .
ae =
i is even
i is odd
We also set
X
ℓ(a) = a1 + a2n
and
r(a) =
ai .
i∈{1,2n}
/
Lemma 3.13. Let a =
P
i∈[2n]
ai δi and H = H(Gn ). Then
(a) a ∈ ZH if and only if ae = ao .
(b) The following conditions are equivalent:
(i) a ∈ H;
(ii) ae = ao , ℓ(a) ≤ r(a) and ai ≥ 0 for all i = 1, . . . , 2n.
Proof. (a) First note that if i is even and j is odd, then δi + δj ∈ ZH. Indeed, if
{i, j} =
6 {1, 2n}, then δi + δj ∈ H; if {i, j} = {1, 2n}, then δ1 + δ2n = (δ1 + δ2 ) +
(δ3 + δ2n ) − (δ2 + δ3 ) ∈ ZH.
Suppose that ae = ao . We prove that a ∈ ZH by induction on |ae |. If ae = 0
then a = 0, and the assertion is trivial. Suppose that |ae | > 0. We only consider the
case when ae > 0 since the other case is similar. In this case there exist i, j ∈ [2n]
such that i is even, j is odd, ai > 0 and aj > 0. Let b = a − (δi + δj ). Then
be = ae − 1 = ao − 1 = bo , and so b ∈ ZH by induction. This implies that
a = b + (δi + δj ) ∈ ZH. Conversely, it is obvious that ae = ao if a ∈ ZH.
(b) (i) ⇒ (ii): Note that ℓ(h(e))
/ E(G).
P ≤ r(h(e)) for any e ∈ E(G) since {1, 2n} ∈
Now given a ∈ H. Then a = e∈E(G) ce h(e), where ce is a non-negative integer for
P
P
each e ∈ E(G). It follows that ℓ(a) = e∈E(G) ce ℓ(h(e)) ≤ e∈E(G) ce r(h(e)) = r(a),
as required.
23
(ii) ⇒ (i): We use induction on ℓ(a). If ℓ(a) = 0, we see that a ∈ H by induction
on ae as in the proof of (a). Assume that ℓ(a) > 0. Without restriction we may
further assume that a1 ≥ a2n . Then
X
ae − a2n =
ai > 0,
i∈even,
i6=2n
for otherwise ae = a2n ≤ a1 < ao , a contradiction. (For the inequality a1 < ao we
used that 0 < ℓ(a) ≤ r(a) = a3 + a5 + · · · + a2n−1 .) Hence there exists an even
number j 6= 2n with aj > 0. Set b = a − (δ1 + δj ). Then b ∈ H by induction and so
a = b + (δ1 + δj ) ∈ H.
Corollary 3.14. Let a ∈ ZH with ai ≥ 0 for all i ∈ [2n]. Then either a ∈ H or
a = b + k(δ1 + δ2n ), where k ≥ 1 and b ∈ H with ℓ(b) = r(b).
Proof. Suppose that a ∈
/ H. Then ℓ(a) > r(a), by Lemma 3.13. Note that by
Lemma 3.13 we have ae = ao , and hence ℓ(a) −r(a) = a1 + a2nP
−(ae −a2n ) −(ao −a1 )
is an even number, say 2k. Set b = (a1 − k)δ1 + (a2n − k)δ2n + i∈{1,2n}
ai δi . Suppose
/
that a1 < k. Then
X
X
a2n = 2k − a1 +
ai > a1 +
ai ≥ ao ,
i∈{1,2n}
/
i∈{1,2n}
/
a contradiction, because ae = ao . Therefore a1 ≥ k. Similarly, a2n ≥ k. Since
ℓ(b) = r(b), Lemma 3.13 implies that b ∈ H. Moreover, a = b + k(δ1 + δ2n ) by the
choices of k and b.
Lemma 3.15. Let a ∈ ZH such that E(Gn ) \ {e : a + h(e) ∈ H} contains no cycle.
Then dimK KL = dimK Da . In particular, T 1 (R)a = 0.
Proof. Let Γ′ be the graph with E(Γ′ ) = E(Gn ) \ {e : a + h(e) ∈ H}, and let Γ
be the graph obtained from Γ′ by adding all vertices of Gn which do not belong to
V (Γ′ ). Then Γ is a graph with no cycle and V (Γ) = [2n]. If Γ is a tree, then Γ is a
spanning tree of Gn . If Γ is not a tree, we choose a connected component Γ0 of Γ
and let Γ1 be the induced graph of Γ on the set [2n] \ V (Γ0 ). Since Gn is connected
there exists an edge e of Gn with one end in V (Γ0 ) and the other end in V (Γ1 ). We
let Γ′′ be the graph obtained from Γ by adding the edge e. Since neither Γ0 nor
Γ1 contains a cycle, it follows that Γ′′ has also no cycle, but one more edge than
Γ. Proceeding in this way we obtain after finitely many steps a spanning tree T of
Gn which contains E(Gn ) \ {e : a + h(e) ∈ H}. In particular, there exists a subset
of {e : a + h(e) ∈ H}, say {e1 , . . . , ek } such that E(T ) = E(Gn ) \ {e1 , e2 , . . . , ek }.
Since all spanning trees have the same number of edges, namely, 2n − 1, it follows
that k = n2 − 2n.
For each i = 1, . . . , k, T + ei contains a unique induced cycle, say Ci . Then for
all i = 1, . . . , k we have vi (i) ∈ {±1} and vi (j) = 0 if j 6= i and 1 ≤ j ≤ k. Here
vi = v(Ci ), the vector corresponding to the cycle Ci , for i = 1, . . . , k. It follows
that dimK Da ≥ k since (v1 (i), . . . , vk (i), . . . , vs1 (i)) ∈ Da for i = 1, . . . , k. Here s1 is
the number of induced cycles of Gn . On the other hand, dimK Da ≤ dimK KL and
24
dimk KL = |E(Gn )| − |V (Gn )| + 1 which is equal to k = n2 − 2n. It follows that
dimK KL = dimK Da , completing the proof by Proposition 1.2.
Proof of Proposition 3.12. (a) Since G3 is associated with a polyomino containing 8
vertices, we have R(= K[G3 ]) is semi-rigid by Corollary 3.11.
Let a = δ6 −δ1 −δ2 −δ4 . Then KLa is spanned by the vectors corresponding to the
cycles C1 : 3, 2, 5, 4, C2 : 3, 4, 5, 6 and C3 : 3, 2, 5, 6. This implies that dimK KLa = 2.
Since dimK Da = 0 and dimK KL = 3, we have T 1 (R)a 6= 0. In particular, R is not
rigid, as required.
P
(b) Let a = i∈[2n] ai δi ∈ ZH. We want to prove that T 1 (R)a = 0. There are
following cases to consider.
Case 1 : ai ≥ 0 for all i. By Corollary 3.14, either a ∈ H or a = b+k(δ1 +δ2n ), where
k ≥ 1 and b ∈ H with ℓ(b) = r(b). If a ∈ H, then T 1 (R)a = 0, see Corollary 1.4. If
a = b + k(δ1 + δ2n ) with k = 1, then for any edge e = {i, j} with e ∩ {1, 2n} = ∅,
we have a + δi + δj ∈ H by Lemma 3.13. It follows that E(G) \ {e : a + h(e) ∈ H}
contains no cycle, and so T 1 (R)a = 0 by Lemma 3.15. If k = 2, then for any
induced cycle C, a + h(v(C)) ∈ H if and only if V (C) ∩ {1, 2n} = ∅. This follows
from Lemma 3.13 and the fact that any induced cycle of Gn is a 4-cycle. To prove
KL = KLa , we have to show if V (C) ∩ {1, 2n} = ∅, then v(C) ∈ KLa . Given
an induced cycle C : i1 , i2 , i3 , i4 with V (C) ∩ {1, 2n} = ∅, where i1 , i3 are even and
i2 , i4 are odd. Then we obtain two cycles C1 : i1 , i2 , i3 , 1 and C2 : i3 , i4 , i1 , 1. Note
that v(C1 ), v(C2 ) ∈ KLa and v(C) is a linear combination of v(C1 ), v(C2), we have
KL = KLa and so T 1 (R)a = 0. If k ≥ 3, then for any induced cycle C, one has
a + h(v(C)) ∈
/ H by Lemma 3.13 and so KLa = KL. In particular, T 1 (R)a = 0.
Case 2: There exists a unique i ∈ [2n] with ai < 0. If ai ≤ −2, then Fa′ = ∅ and so
KLa = KL. In particular T 1 (R)a = 0. Hence we assume ai = −1. By symmetry,
we only need to consider the cases when i = 1 and when i = 3.
We first assume that i = 1. Since ae = ao , there exists an odd integer j 6= 1
such that aj > 0, and so a = b + δj − δ1 , where be = bo and bℓ ≥ 0 for each
ℓ ∈ [2n]. By Corollary 3.14, either b ∈ H or b = c + k(δ1 + δ2n ) with c ∈ H and
k > 0. The second case cannot happen because a1 = −1. Hence for any e ∈ E(G),
a + h(e) ∈ H if and only if 1 ∈ e. In other words, a + h(e) ∈ H if and only if
e ∈ {{1, 2}, {1, 4}, . . . , {1, 2n − 2}}. Denote {1, 2i} by ei for i = 1, . . . , n − 1. Let
Ci be the cycle 1, 2i, 3, 2n − 2 and let vi = v(Ci ) for i = 1, . . . , n − 2. Then for
i = 1, . . . , n − 2, we have vi (i) ∈ {±1} and vi (j) = 0 for j 6= i and j = 1, . . . , n − 2.
This implies that dimK Da ≥ n − 2. To compute dimK KLa , we notice that if C is
an induced cycle with 1 ∈
/ V (C), then a + h(v(C)) ∈
/ H and thus KLa contains the
cycle space of the complete bipartite graph with bipartition {3, 5, . . . , 2n − 1} and
{2, . . . , 2n}, which has the dimension (n − 1)n − n − (n − 1) + 1 = n2 − 3n + 2, see
(10). Thus T 1 (R)a = 0 because dimK Da = n2 − 2n.
Next we assume that i = 3. Then a = b + δj − δ3 , where bℓ ≥ 0 for all ℓ ∈ [2n]
and be = bo , j 6= 3 and j is odd. Moreover by Corollary 3.14, we have either
b ∈ H or b = c + k(δ1 + δ2n ) for some k ≥ 1 and with c ∈ H and ℓ(c) = r(c).
Suppose first that b ∈
/ H and k ≥ 2. Then for any cycle C, a + h(v(C)) ∈ H
25
implies V (C) ∩ {1, 2n} = ∅. Thus, similarly as in Case 1 we see that KLa = KL
and T 1 (R)a = 0. Suppose next that j 6= 1 and that b ∈ H or b ∈
/ H and k = 1.
Then a + h(e) ∈ H for any e ∈ {{3, 2}, . . . , {3, 2n − 2}}. Denote {3, 2t} by et for
t = 1, . . . , n−1. For t = 1, . . . , n−1, let Ct be the cycle 3, 2t, 5, 2n and let vt = v(Ct ),
the vector corresponding to Ct . Then vt (t) ∈ {±1} for t = 1, . . . , n − 1 and vt (k) = 0
for k 6= t. This implies that dimK Da = n − 1. On the other hand, KLa contains
the cycle space of the subgraph of Gn induced on {1, 5, . . . , 2n − 1} ∪ {2, 4, . . . , 2n},
which has the dimension n2 − 3n + 1. Thus T 1 (R)a = 0.
Finally suppose that j = 1 and that and also k = 1 if b ∈
/ H. If b ∈ H,
then we check that a + h(e) ∈ H for any e ∈ {{3, 2}, . . . , {3, 2n − 2}} and deduce
that T 1 (R)a = 0, in the same process as in the last case. If b ∈
/ H and k = 1,
then for any induced cycle C, we have a + h(v(C)) ∈ H if and only if 3 ∈ V (C)
and {1, 2n} ∩ V (C) = ∅. We claim that KLa = KL. Given an induced cycle
C : 3, i1 , i2 , i3 with a + h(v(C)) ∈ H. Here i1 and i3 are even and i2 is odd. We
let C1 : 3, i1 , i2 , 2n and C2 : i2 , i3 , 3, 2n. Then v(C1) and v(C2 ) belong to KLa and
v(C) is a linear combination of v(C1 ) and v(C2 ). Thus KLa = KL, as claimed. In
particular, T 1 (R)a = 0.
Case 3: |{k : ak < 0}| = 2. Without restriction we may assume ai = aj = −1
for some i 6= j. Indeed, if some ak ≤ −2, then Fa′ = ∅ by Lemma 3.13 and so
T 1 (R)a = 0. Assume first that both i and j are even. Then for any induced cycle
C such that {i, j} * V (C), we have v(C) ∈ KLa . Let C : k, i, ℓ, j be a cycle with
{i, j} ⊆ V (C). We choose an even number d ∈ [2n] \ {i, j, 2n}. Then we obtain
two cycles C1 :k, i, ℓ, d and C2 :ℓ, j, k, d. Since v(C) is a linear combination of v(C1)
and v(C2 ) and since v(Ct ) ∈ KLa for t = 1, 2, we have v(C) ∈ KLa and thus
KLa = KL. In particular T 1 (R)a = 0.
Next assume that i is even and j is odd and {i, j} =
6 {1, 2n}. Notice that we can
write a as a = b + k(δ1 + δ2n ) − (δi + δj ), where b ∈ H and k ≥ 0. Moreover, if k > 0
then ℓ(b) = r(b).
If k = 0, then dimK Da = 1, and KLa contains the cycle space of the graph which
is obtained from Gn by deleting the edge {i, j}. Hence dimK KLa ≥ dimK KL − 1,
and so T 1 (R)a = 0.
If k = 1, then for any induced cycle C, we have a + h(v(C)) ∈ H if and only if
{i, j} ⊆ V (C) and V (C) ∩ {1, 2n} = ∅. Let C : i, j, k, ℓ be an induced cycle such
that a + h(v(C)) ∈ H. Then the vectors v1 , v2 which correspond to cycles j, k, ℓ, 2n
and ℓ, i, j, 2n respectively belong to KLa and v(C) is a linear combination of v1 , v2 .
It follows that KL = KLa and T 1 (R)a = 0.
If k ≥ 2, we first note that {i, j} ∩ {1, 2n} = ∅. Indeed, if {i, j} ∩ {1, 2n} =
6 ∅,
′
then either ai ≥ 1 or aj ≥ 1, a contradiction. Thus Fa = ∅ by Lemma 3.13 and it
follows that KL = KLa . In particular T 1 (R)a = 0.
Finally assume that {i, j} = {1, 2n}. Then a = b − δ1 − δ2n with b ∈ H and so
Fa′ = ∅. It follows that KLa = KL and T 1 (R)a = 0.
Case 4: |{k : ak < 0}| = 3. We may assume that ai = aj = ak = −1. We
only need to consider the case when Fa′ 6= ∅. So we may assume i, k are even
and j is odd, and {1, 2n} * {i, j, k}. Let C : i, j, k, ℓ be an induced cycle such that
26
a+h(v(C)) ∈ H. We choose an even number d ∈ [2n]\{i, k, 2n} and let C1 : j, k, ℓ, d
and C2 : ℓ, i, j, d be two cycles in Gn . Then v(C1 ) and v(C2 ) belong to KLa and
v(C) is a linear combination of v(C1 ) and v(C2 ). This implies KLa = KL, and in
particular, T 1 (R)a = 0.
Case 5: |{k : ak < 0}| ≥ 4. If |{k : ak < 0}| = 4, we may assume that ai =
aj = ak = aℓ = −1. Then for any induced cycle C, a + h(v(C)) ∈ H implies that
V (C) = {i, j, k, ℓ}. We may assume that i and k are even numbers. Let t be an
odd number in [2n] \ {j, ℓ}, and let C1 : i, j, k, t and C2 : k, l, i, t be 4-cycles of Gn .
Since v(C) is a linear combination of v(C1 ) and v(C2 ), we have KLa = KL, and
consequently, T 1 (R)a = 0. If |{k : ak < 0}| > 4, then Fa′ = ∅ and so T 1 (R)a = 0.
Thus we have shown that T 1 (R)a = for all a, and this shows that R is rigid, as
desired.
References
[1] K. Altmann, Computation of the vector space T 1 for affine toric varieties, J. Pure Appl. Algebra
95 (1994) 239–259
[2] K. Altmann, Minkowski sums and homogeneous deformations of toric varieties, Tóhoku Math.
J. 47 (1995) 151–184
[3] K. Altmann, M. Bigdeli, J. Herzog, D. Lu, Algebraically rigid simplicial complexes and graphs,
arXiv:1503.08080
[4] K. Altmann, J. A. Christophersen, Deforming Stanley-Reisner schemes, Math. Ann. 348
(2010), 513–537.
[5] K. Altmann, J. A. Christophersen, Cotangent cohomology of Stanley-Reisner rings.
Manuscripta Math. 115 (2004), 361–378.
[6] W. Bruns, J. Herzog, Cohen-Macaulay rings, Cambridges Studies in Advanced Mathematics,
39, 1993.
[7] V. Ene, J. Herzog, Gröbner bases in Commutative Algebra, Graduate Studies in Mathematics
130, AMS 2012.
[8] J. Herzog, Generators and relations of abelian semigroups and semigroup rings, Manuscripta
Math. 3 (1970), 175–193.
[9] A. A. Qureshi, Ideals generated by 2-minors, collections of cells and stack polyominoes, J.
Algebra, 357 (2012), 279–303.
[10] H. Ohsugi, T. Hibi, Koszul bipartite graphs, Advances in Applied Mathematics, 22 (1999)
25–28.
[11] H. Ohsugi, T. Hibi, Toric ideals generated by quadratic binomials, J. Algebra. 218 (1999),
509–527.
[12] J. Stevens, Deformations of singularities, Lecture Notes in Mathematics (1811), Berlin:
Springer, 2003.
[13] R. Villarreal, Monomial Algebras, Pure and Applied Mathematics, Marcel Dekker, 2001.
Mina Bigdeli, Faculty of Mathematics, Institute for Advanced Studies in Basic
Sciences (IASBS), 45195-1159 Zanjan, Iran
E-mail address: [email protected]
Jürgen Herzog, Fakultät für Mathematik, Universität Duisburg-Essen, 45117
Essen, Germany
E-mail address: [email protected]
27
Dancheng Lu, Department of Mathematics, Soochow University, 215006 Suzhou,
P.R.China
E-mail address: [email protected]
28
| 0math.AC
|
A New Approach to Probabilistic Programming Inference
arXiv:1507.00996v2 [stat.ML] 9 Jul 2015
Frank Wood
Department of Engineering
University of Oxford
Jan Willem van de Meent
Vikash Mansinghka
Department of Statistics
Computer Science & AI Lab
Columbia University
Massachusetts Institute of Technology
Abstract
We introduce and demonstrate a new approach to inference in expressive probabilistic programming languages based on particle
Markov chain Monte Carlo. Our approach
is simple to implement and easy to parallelize. It applies to Turing-complete probabilistic programming languages and supports
accurate inference in models that make use
of complex control flow, including stochastic recursion. It also includes primitives
from Bayesian nonparametric statistics. Our
experiments show that this approach can
be more efficient than previously introduced
single-site Metropolis-Hastings methods.
1
Introduction
Probabilistic programming differs substantially from
traditional programming. In particular, probabilistic
programs are written with parts not fixed in advance
that instead take values generated at runtime by random sampling procedures. Inference in probabilistic
programming characterizes the conditional distribution of such variables given observed data assumed
to have been generated by executing the probabilistic
program. Exploring the joint distribution over program execution traces that could have generated the
observed data using Markov chain sampling techniques
is one way to produce such a characterization.
We propose a novel combination of ideas from probabilistic programming [3] and particle Markov chain
Monte Carlo (PMCMC) [1] that yields a new scheme
for exploring and characterizing the set of probable execution traces. As our approach is based on repeated
simulation of the probabilistic program, it is easy to
implement and parallelize. We show that our approach
supports accurate inference in models that make use of
complex control flow, including stochastic recursion, as
well as primitives from nonparametric Bayesian statistics. Our experiments also show that this approach can
be more efficient than previously introduced single-site
Metropolis-Hastings (MH) samplers [12].
2
Language
The probabilistic programming system Anglican1 exists in two versions. The results in this paper were obtained using what is now called interpreted Anglican2
which employed an interpreted execution model and a
language syntax derived from the Venture3 modeling
language. Since the time of original publication, both
the syntax and execution model have been changed.
The Venture style syntax is now deprecated, with the
new version using a syntax much closer to that of
Anglican’s host language Clojure. This new Anglican version4 , simply called Anglican, is a compiled
language; a continuation-passing style (CPS) transformation compiles Anglican programs into Clojure programs that are then subsequently compiled to Java
Virtual Machine (JVM) bytecode by the Clojure justin-time compiler.
The change in language execution model and syntax
affect neither the substance or the claims of this paper, however, readers wishing to experiment with the
language starting from the code examples appearing in
Section 4 would do well to note this evolution. What
does change is the absolute time required to perform
inference in the given models. In general, the latest
compiled version of Anglican is ten to one hundred
times faster than its interpreted ancestor.
1
This paper, which originally appeared in the Proceedings of
the 17th International Conference on Artificial Intelligence
and Statistics (AISTATS) 2014, has been updated to reflect
changes in the latest version of the Anglican language.
http://www.robots.ox.ac.uk/~fwood/anglican/
http://bitbucket.org/probprog/
interpreted-anglican
3
http://probcomp.csail.mit.edu/venture/
4
http://bitbucket.org/probprog/anglican
2
A New Approach to Probabilistic Programming Inference
2.1
Original Venture-Style Syntax
The original, deprecated Anglican syntax is a
Scheme/Lisp-dialect that is extended with three toplevel special forms that we refer to as directives
[ assume symbol < expr >]
[ observe < expr > < const >]
[ predict < expr >]
Here, each <expr> is a Scheme/Lisp-syntax expression,
symbol is a unique symbol, and <const> is a constantvalued deterministic expression.
Semantically assume’s are (random) variable (generative) declarations, observe’s condition the distribution
of assume’d variables by (noisily) constraining the output values of (random) functions of assume’d variables
to match observed data, and predict’s are “watches”
which report on (via print out) the values of variables
in program traces as they are explored. In Anglican,
probabilistic program interpretation is taken to be a
forever-continuing exploration of the space of execution traces that obey (where hard) or reflect (where
soft) the observe’d constraints in order to report functions (via predict’s) of the conditional distribution of
subsets of the assume’d variables.
Like Scheme/Lisp, Anglican eagerly and exchangeably recursively “evaluates” subexpressions, for instance of <expr> = (<proc> <arg> . . . <arg>), before
“applying” the procedure (which may be a random
procedure or special form) resulting from evaluating
the first expression <proc> to the value of its arguments. Anglican counts applications, reported as a
computational cost proxy in Sec 4. Anglican supports several special forms, notably (lambda (<arg>
. . . <arg>) <body>)) which allows creation of new procedures and (if <pred> <cons> <alt>) which supplies
branching control flow; also begin, let, define, quote,
and cond. Anglican exposes eval and apply. Built-in
deterministic procedures include list, car, cdr, cons,
mem, etc., and arithmetic procedures +, -, \, etc.
All randomness in the language originates from builtin “random primitives” of which there are two types.
Elementary random primitives, such as poisson, gamma,
flip, discrete, categorical, and normal, generate
independent and identically distributed (i.i.d.) samples when called repeatedly with the same arguments. Exchangeable random primitives, such as crp
or beta-bernoulli, return a random procedure with
internal state that generates an exchangeably distributed samples when called repeatedly.
In the interpreted version of Anglican the outer <proc>
in all observe <expr>’s must be a built-in random primitive, which guarantees that likelihood of output given
arguments can be computed exactly.
2.2
New Clojure-Style Syntax
The compiled version of Anglican now supports a syntax that more closely integrates with the host language
Clojure. The macro defquery defines an Anglican
program from within Clojure
( defquery symbol [ arg1 arg2 ...] < body >)
Here symbol is the name of the program, and the arguments may be used to pass values of parameters or
observed variables to the program. The set of allowable body expressions <body> is a subset of the Clojure
language, in which all basic Clojure language forms
and first order primitives are supported. Macros and
higher-order functions are not inherited from Clojure,
although a subset of macros5 and higher-order functions6 has been implemented.
In the CPS version of Anglican, random primitives
such as normal and discrete return first class distribution objects, rather than sampled values, and, furthermore, are user programmable without requiring modifications to the Anglican compiler. The special forms
sample and observe associate values with random variables drawn from a distribution
( sample < dist >)
( observe < dist > < value >)
The language additionally provides a data type for sequences of random variables that are not i.i.d., which
we refer to as a random process. A random process
implements two operations
( produce < process >)
( absorb < process > < value >)
The produce primitive returns a distribution object
for the next random variable in the sequence. The
absorb primitive returns an updated random process
instance, in which a value has been associated with
the next random variable in the sequence. A random
process is most commonly used to represent an exchangeably distributed sequence of random variables,
but it may be used to represent any sequence of random variables for which it is possible to construct a
distribution on the next variable given preceding variables. Random process constructors are customarily
identified with uppercase names (e.g. CRP) and, in the
same manner as before, are user programmable.
Finally, the predict form may be used at any point in
the program to generate labeled output values
( predict < label > < expr >)
5
when, cond
map, reduce, filter, some, repeatedly, comp,
partial
6
Frank Wood, Jan Willem van de Meent, Vikash Mansinghka
Given a previously defined query, inference may be
performed using the doquery macro
( doquery algorithm < symbol >
[ < arg1 > < arg2 > ...]
< opt1 > < opt2 > ...)
The inference algorithm may be specified using an
algorithm keyword, and any options to the inference
algorithm can be supplied as arguments as well. The
doquery macro constructs a lazy sequence of program
execution states that contain predicted values and optionally an importance weight, which may then be consumed and analyzed by, for instance, an outer Clojure
or Java program.
3
Inference
An execution trace is the sequence of memory states
resulting from the sequence of function applications
performed during the interpretation of a program. In
probabilistic programming systems like Anglican any
variable may be declared as being the output of a
random procedure. Such variables can take different
values in independent interpretations of the program.
This leads to a “many-worlds” computational trace
tree in which, at interpretation time, there is a branch
at every random procedure application.
To define the probability of a single execution trace,
first fix an ordering of the exchangeable lines of
the program and index the observe lines by n. Let
p(yn |θtn , xn ) be the likelihood of the observe’d output
yn where tn is a random procedure type (i.e. gamma,
poisson, etc.), θtn is its argument (possibly multidimensional), and xn is the set of all random procedure
application results computed before the likelihood of
observation yn is evaluated. Both the type tn and the
parameter θtn can be functions of any in-scope subset of xn . We can then define the probability of an
execution trace to be
p̃(y, x) ≡
N
Y
p(yn |θtn , xn )p̃(xn |xn−1 )
(1)
n=1
where y is the set of all observe’d quantities, x is the
set of all random procedure application results, and ∼
marks distributions which we can only sample.
The number and type of all random procedure applications performed before the nth observe may vary in
one program trace to the next. We define the probability of the sequence of their outputs xn to be
p̃(xn |xn−1 )
(2)
|xn \xn−1 |
=
Y
k=1
p(xn,k |θtn,k , xn,1:(k−1) , xn−1 ).
Here x0 = ∅ is the empty set and xn,j:k are the j
to kth values generated by random procedure applications in the trace up to observation likelihood computation n. The cardinality of the set xn \ xn−1 , notated
| · |, arises implicitly as the total number of random
procedure applications in a given execution trace. As
before, θtn,k are the arguments to, and tn,k the type
of, the (n, k)th random procedure – both of which may
be functions of subsets of in-scope subsets of variables
xn:1,(k−1) ∪xn−1 . Note that xn−1 ⊆ xn . Also note that
variable referencing defines a directed conditional dependency structure for the probability model encoded
by the program, i.e. xn,k need not (and often cannot
due to variable scoping) depend on the outputs of all
previous random procedure applications.
We use sampling to explore and characterize the distribution p̃(x|y) ∝ p̃(y, x), i.e. the distribution of all random procedure outputs that lead to different program
execution traces, conditioned on observed data. Related approaches include rejection sampling [3], singlesite MH [3, 12].
All members of the set of all directed probabilistic
models with fixed-structure joint distributions can be
expressed as probabilistic programs that “unroll” in all
possible execution traces into an equivalent joint distribution. As Church-like probabilistic programming
frameworks, Anglican included, support recursive procedures and branching on the values returned by random procedures, the corresponding set of models is
a superset of the set of all directed graphical models. Other related efforts eschew Turing-completeness
and operate on a restricted set of models [11, 9, 7, 10]
where inference techniques other than sampling can
more readily be employed.
3.1
A New Approach
Towards our new approach to probabilistic programming inference, first consider a standard sequential
Monte Carlo (SMC) recursion for sampling from a sequence of intermediate distributions that terminates
in p̃(x|y) ∝ p̃(y, x) where x and y are as before,
and the joint is given by Eq.’s 1 and 2. Note that a
sequence of intermediate approximating distributions
can be constructed from any syntactically allowed reordering of
p̃(x1 )p̃(x2 |x1 ) · · · p̃(xn |xn−1 )p(y1 |x1 ) · · · p(yn |xn ).
Assume that observation likelihoods are pushed as far
left in this sequence of approximating distributions as
possible; however it is clear how to proceed if this is
not the case. Assume we have 1 ≤ ` ≤ L unweighted
(`)
samples xn−1 ∼ p̃(xn−1 |y1:(n−1) ) and that from these
we will produce approximate samples from p̃(xn |y1:n ).
A New Approach to Probabilistic Programming Inference
To do so via importance sampling we may choose
any proposal distribution q(xn |xn−1 , y1:n ). Sampling
from this and weighting by the discrepancy between
it and the distribution of interest, p̃(xn , y1:n ), we ar(`)
rive at samples with unnormalized weights w̃n =
(`)
(`) (`)
p̃(x̂n , y1:n )/q(x̂n |xn−1 , y1:n ). Here hats notate the
difference between weighted and unweighted samples,
those with being weighted and vice versa.
This expression simplifies substantially in the “propose
from the prior” case where the proposal distribution
is defined to be the continued interpretation of the
program from observation likelihood evaluation n − 1
(`) (`)
(`) (`)
to n, i.e. q(x̂n |xn−1 , y1:n ) ≡ p(x̂n |xn−1 ). In this case
(`)
(`)
the weight simplifies to w̃n = p(yn |x̂n ). Sampling
P
(`)
(`)
an unweighted particle set xn ∼ ` wn δx̂(`) , where
n
P
(`)
(`)
(j)
wn = w̃n / j w̃n , completely describes SMC for
probabilistic program inference.
The SMC procedure described is, to first approximation, the inner loop of PMCMC. It corresponds to a
procedure whereby the probabilistic program is interpreted in parallel (possibly each particle in its own
thread or process) between observation likelihood calculations. Unfortunately, SMC with a finite set of particles is not itself directly viable for probabilistic programming inference for all the familiar reasons: particle degeneracy, inefficiency in models with global, continuous parameters, etc.
PMCMC, on the other hand, is directly viable. PMCMC for probabilistic programming inference is a MH
algorithm for exploring the space of execution traces
that uses SMC proposals internally. This, unlike prior
art, allows sampling of execution traces with changes
to potentially many more than one variable at a time.
The particular variant of PMCMC we discuss in this
paper is Particle-Gibbs (PG) although we have developed engines based on other PMCMC variants including particle independent Metropolis Hastings and conditional sequential Monte Carlo too. PG works by iteratively re-running SMC, with, on all but the first
sweep, reinsertion of a “retained” particle trace into
the set of particles at every stage of SMC. PG is theoretically justified as an MH transition operator that,
like the Gibbs operator, always accepts [1, 4]. In this
paper we describe PMCMC for probabilistic programming inference algorithmically in Alg.1 and experimentally demonstrate its relative efficacy for probabilistic
programming inference.
In Alg. 1 the function r(N, S) stands for multinomial sample N items from the set of pairs S =
{{w1 , θ1 }, . . . , {wM , θM }}, where each element of S
consists of an unnormalized weight wm and interpreter
memory states θm . For each sample value θm returned,
the function r also returns the original, corresponding
unnormalized weight wm . This kind of weight bookkeeping retains, for all particles, the results of the outermost observe likelihood function applications so that
the unnormalized weights are available in the retained
particle {w∗ , x∗ } in the next sweep.
Algorithm 1 PMCMC for Prob. Prog. Inference
L ← number of particles
S ← number of sweeps
(`)
(`)
{w̃N , xN } ← Run SMC
for s < S do
(`)
{·, x∗N } ← r(1, {1/L, xN })
(`)
{·, x0 } ← initialize L − 1 interpreters
for d ∈ ordered lines of program do
for ` < L − 1 do
(`)
(`)
x̄n−1 ← fork(xn−1 )
end for
if directive(d) == “assume” then
for ` < L − 1 do
(`)
(`)
x̄n ← interpret(d, x̄n−1 )
end for
(`)
(`)
{xn } ← {x̄n } ∪ x∗n
else if directive(d) == “predict” then
for ` < L − 1 do
(`)
interpret(d, x̄n−1 )
end for
interpret(d, x∗n−1 )
else if directive(d) == “observe” then
for ` < L − 1 do
(`)
(`)
(`)
{w̄n , x̄n } ← interpret(d, x̄n−1 )
end for
(`)
(`)
T ← r(L − 1, {w̄n , x̄n } ∪ {w̃n∗ , x∗n })
(`)
(`)
{w̃n , xn } ← T ∪ {w̃n∗ , x∗n }
end if
end for
end for
In Alg. 1 “Run SMC” means running one sweep of
the s loop with L particles and no retained particle,
d is a program line, and fork(·) means to copy the
entire interpreter memory datastructure (efficient implementations have characteristics similar to POSIX
fork() [8]). The command interpret(d, ·) means execute line d in the given interpreter. Only when interpreting an observe must the interpreter return a
weight, that being result of the outermost apply of
the observe statement. Bars indicate temporary data
structures, not averages. All sets are ordered with
unions implemented by append operations.
Note that there are more efficient PMCMC algorithms
for probabilistic programming inference. In particular,
there is no reason to fork unless an observe has just
Frank Wood, Jan Willem van de Meent, Vikash Mansinghka
been interpreted. Alg. 1 is presented in this form for
expositional purposes.
3.2
proposal density can now be expressed in a similar
fashion in terms of x\x0 and x ∩ x0 = x0 ∩ x, allowing
the full acceptance probability to be written as
Random Database
We refer to the MH approach to sampling over
the space of all traces proposed in [12] as “random
database” (RDB). A RDB sampler is a MH sampler
where a single variable drawn in the course of a particular interpretation of a probabilistic program is modified via a standard MH proposal, and this modification
is accepted by comparing the value of the joint distribution of old and new program traces. For completeness we review RDB here, noting a subtle correction to
the acceptance ratio proposed in the original reference
which is proper for a larger family of models.
The RDB sampler employs a data structure that holds
all random variables x associated with an execution
trace, along with the parameters and log probability
of each draw. Note that interpretation of a program
is deterministic conditioned on x. A new proposal
trace is initialized by picking a single variable xm,j
from the |x| random draws, and resampling its value
using a reversible kernel κ(x0m,j |xm,j ). Starting from
this initialization, the program is rerun to generate a
new set of variables x0 that correspond to a new valid
execution trace. In each instance where the random
procedure type remains the same, we reuse the existing value from the set x, rescoring its log probability
conditioned on the preceding variables where necessary. When the random procedure type has changed,
or a new random variable is encountered, its value is
sampled in the usual manner. Finally, we compute
the probability p(y|x0 ) by rescoring each observe as
needed, and accept with probability
p(y|x0 )p(x0 )q(x|x0 )
.
(3)
min 1,
p(y|x)p(x)q(x0 |x)
In order to calculate the ratio of the proposal probabilities q(x0 |x) and q(x|x0 ), we need to account for the
variables that were resampled in the course of constructing the proposal, as well as the fact that the
sets x0 and x may have different cardinalities |x0 | and
|x|. We will use the (slightly imprecise) notation x0 \x
to refer to the set of variables that were resampled,
and let x0 ∩ x represent the set of variables common
to both execution traces. The proposal probability is
now given by
q(x0 |x) =
κ(x0m,j |xm,j ) p(x0 \x | x0 ∩ x)
.
|x|
p(x0m,j |x0 ∩ x)
(4)
In our implementation, the initialization x0m,j is simply resampled conditioned on the preceding variables,
such that κ(x0m,j |xm,j ) = p(x0m,j |x0 ∩ x). The reverse
p(y|x0 ) p(x0 ) |x| p(x\x0 | x ∩ x0 )
.
p(y|x) p(x) |x0 | p(x0 \x | x0 ∩ x)
4
(5)
Testing
Programming probabilistic program interpreters is a
non-trivial software development effort, involving both
the correct implementation of an interpreter and the
correct implementation of a general purpose sampler.
The methodology we employ to ensure correctness of
both involves three levels of testing; 1) unit tests, 2)
measure tests, and 3) conditional measure tests.
4.1
Unit and Measure Tests
In the context of probabilistic programming, unit testing includes verifying that the interpreter correctly
interprets a comprehensive set of small deterministic
programs. Measure testing involves interpreting short
revealing programs consisting of assume and predict
statements (producing a sequence of ancestral, unconditioned samples, i.e. no observe’s). Interpreter output is tested relative to ground truth, where ground
truth is computed via exhaustive enumeration, analytic derivation, or some combination, and always in
a different, well-tested independent computational system like Matlab. Various comparisons of the empirical
distribution constructed from the accumulating stream
of output predicts’s and ground truth are computed;
Kulback-Leibler (KL) divergences for discrete sample
spaces and Kolmogorov Smirnov (KS) test statistics
for continuous sample spaces. While it is possible
to construct distribution equality hypothesis tests for
some combinations of test statistic and program we
generally are content to accept interpreters for which
there is clear evidence of convergence towards zero of
all test statistics for all measure tests. Anglican passed
all unit and measure tests.
4.2
Conditional Measure Tests
Measure tests involving conditioning provide additional information beyond that provided by measure
and unit tests. Conditioning involves endowing programs with observe statements which constrain or
weight the set of possible execution traces. Interpreting observe statements engages the full inference machinery. Conditional measure test performance is measured in the same way as measure test performance.
They are also how we compare different probabilistic
programming inference engines.
A New Approach to Probabilistic Programming Inference
24
2
4
22
2
2
2
0
-2
2
2-4
2
6
2
4
2
2
2
0
101
102
Simulation
2
-4
2
104
103
-2
2
2
-4
2
-6
5
10
6
10
randomdb
-6
10
1
10
2
10
2-10 0
10
3
pmcmc
randomdb
1
10
2
3
10
10
Simulation
24
2-4
2-6
pmcmc
randomdb
2-8
105
104
Apply
10
3
10
Simulation
4
10
2
2
0
2
-2
2
-4
2
-6
2
-8
2
-10
2
-12
pmcmc
randomdb
104
105
Apply
106
10
1
10
2
10
3
0.5
marginal
0.4
randomdb
pmcmc
0.3
0.2
1
10
10
1
100
10
0
pmcmc
randomdb
pmcmc
randomdb
10
1
10
2
10
10-2 0
10
3
Time [s]
1
10
101
2
4
6
8
10
12
2
4
3
10
10
10
Simulation
100
150
pmcmc
10-1
10
15
r
20
10-2
randomdb
10
0
102
103
10
1
Time [s]
posterior
randomdb
pmcmc
0.6
0.4
0.2
pmcmc
randomdb
randomdb
5
10
-2
0.8
200
enumerated
-1
1.0
KS
50
10
pmcmc
100
0
randomdb
-10
Number of Components
10-1
0
pmcmc
KS
2
-8
0
CDF
10
2
4
2
0.0
107
106
KS
KL
1
2
Prior
26
24
22
20
2-2
2-4
2-6
2-8
2-10
2-12
pmcmc
randomdb
2
-6
(d) Marsaglia
Posterior
KL
KL
2
24
22
20
2-2
2-4
2-6
2-8
2-10
2-12 0
10
2
0.1
(c) Branching
6
-4
Time [s]
2
2
2
0.6
-2
Apply
6
10
20
-10
2
-2
2
4
22
2
1
0
7
0
2
pmcmc
Random DB
10
2
-8
2
1
0
4
20
2
2
1
0
10
2
2-4
Time [s]
randomdb
2
2
Forward-Backward
pmcmc
2
-6
PMCMC
-2
4
-2
KL
KL*
2-6 0
10
pmcmc
randomdb
2
2
2
KL
20
2
6
KL
2
(b) DP Mixture
4
Frequency
26
KL*
KL*
(a) HMM
104
Apply
105
0.0
2
4
6
8
10
12
μ
Figure 1: Comparative conditional measure test performance: PMCMC with 100 particles vs. RDB.
5
Inference Engine Comparison
We compare PMCMC to RDB measuring convergence
rates for an illustrative set of conditional measure test
programs. Results from four such tests are shown in
Figure 1 where the same program is interpreted using
both inference engines. PMCMC is found to converge
faster for conditional measure test programs that cor-
respond to expressive probabilistic graphical models
with rich conditional dependencies.
The four test programs are: 1) a program that corresponds to state estimation in a hidden Markov model
(HMM) with continuous observations (HMM: Program 5.1, Figure 1a), 2) a program that corresponds
to learning an uncollapsed Dirichlet process (DP) mix-
Frank Wood, Jan Willem van de Meent, Vikash Mansinghka
ture of Gaussians with fixed hyperparameters (DP
Mixture: Program 5.2, Figure 1b), 3) a multimodal
branching with deterministic recursion program that
cannot be represented as a graphical model in which all
possible execution paths can be enumerated (Branching: Program 5.3, Figure 1c), and 4) a program that
corresponds to inferring the mean of a univariate normal generated via an Anglican-coded Marsaglia [6]
rejection-sampling algorithm that halts with probability one and generates an unknown number of internal random variables (Marsaglia: Program 5.4, Figure 1d). We refer to 1) and 2) as “expressive” models because they have complex conditional dependency
structures and 3) and 4) as simple models because the
programs encode models with very few free parameters. 1) and 2) illustrate our claims; 3) and 4) are included to document the correctness and completeness
of the Anglican implementation while also demonstrating that the gains illustrated in 1) and 2) do not come
at too great a cost even for simple programs for which,
a priori, PMCMC might be reasonably be expected to
underperform.
In Figures 1a-d there are three panels that report similar style findings across test programs and a fourth
that is specific to the individual test program. In all,
PMCMC results are reported for a single-threaded interpreter with 100 particles. The choice of 100 particles is largely arbitrary; our results are stable for a
large range of values. PMCMC is dark blue while RDB
is light orange. We report the 25% (lower dashed)
median (solid) and 75% (upper dashed) percentiles
over 25 runs with differing random number seeds. We
refer to KL/KL∗ /KS as distances and compute each
via a running average of the empirical distribution of
predict statement outputs to ground truth starting at
the first predict output. Note that lower is better.
We define the number of simulations to be the number of times the program is interpreted in its entirety.
For RDB this means that the number of simulations
is exactly the number of sampler sweeps; for PMCMC
it is the number of particles multiplied by the number
of sampler sweeps. The time horizontal axes report
wall clock time; the apply axes report the number of
function applications performed by the interpreter. In
the distance vs. time plots, observed single-threaded
PMCMC wall-clock times are reported via filled circles; the left-ward dotted lines illustrate hypothetically
what should be achievable via parallelism. Carefully
note that PMCMC requires completing a number of
simulations equal to the number of particles (here 100)
before emitting batched predict outputs. This means
that single-threaded implementations of PMCMC suffer from latency that RDB does not. Still, for some
programs both the quality of PMCMC’s predict outputs and PMCMC’s convergence rate is faster even
in direct wall clock time comparison to RDB. PMCMC appears to converge faster for some programs
than RDB even relative to the number of function applications. Equivalent results were obtained relative
to eval counts.
5.1
HMM
New
( defquery hmm
[ observations init-dist trans-dists obs-dists ]
( predict
: states
( reduce
( fn [ states obs ]
( let [ state ( sample ( get trans-dists
( peek states )))]
( observe ( get obs-dists state ) obs )
( conj states state )))
[( sample init-dist )]
observations )))
Original (deprecated)
[ assume in it i al -s t at e -d is t ( list (/ 1 3) (/ 1 3) (/ 1 3))]
[ assume g e t - s t a t e - t r a n s i t i o n - d i s t ( lambda ( s )
( cond ((= s 0) ( list .1 .5 .4)) ((= s 1) ( list .2 .2 .6))
((= s 2) ( list .15 .15 .7))))]
[ assume transition ( lambda ( prev-state )
( discrete ( g e t - s t a t e - t r a n s i t i o n - d i s t prev-state )))]
[ assume get-state ( mem ( lambda ( index )
( if ( <= index 0) ( discrete in i ti a l- st a te - di st )
( transition ( get-state (- index 1))))))]
[ assume g e t - s t a t e - o b s e r v a t i o n - m e a n ( lambda ( s )
( cond ((= s 0) -1) ((= s 1) 1) ((= s 2) 0)))]
[ observe ( normal ( g et -s t at e -o bs - me a n ( get-state 1)) 1) .9]
[ observe ( normal ( g et -s t at e -o bs - me a n ( get-state 2)) 1) .8]
.
.
.
[ observe ( normal ( g et -s t at e -o bs - me a n ( get-state 16)) 1) -1]
[ predict ( get-state 0)]
[ predict ( get-state 1)]
.
.
.
[ predict ( get-state 16)]
The HMM program corresponds to a latent state inference problem in an HMM with three states, onedimensional Gaussian observations (.9, .8, .7, 0, -.025,
5, 2, 0.1, 0, .13, .45, 6, .2, .3, -1, -1), with known means
and variances, transition matrix, and initial state distribution. The lines of the program were organized
with the observe’s in “time” sequence.
The KL∗ axis reports the sum of the Kulback-Leibler
divergences between the running sample average state
occupancy across all states of a HMM including the
initial state and one trailing predictive state KL∗S =
P
PS
1
i DKL ( S
s=1 δzi(s) ||γi ). Here δzi(s) (k) returns one
if simulation s has latent state z at time step i equal
to k and γi (k) is the true marginal probability of the
latent state indicator zi taking value k at time step
i. The vertical red line in the apply plot indicates the
time and number of applies it takes to run forwardbackward in the Anglican interpreter.
The fourth plot shows the learned posterior distribution over the latent state value for all time steps including both the initial state and a trailing predictive
time step. While RDB produces a reasonable approximation to the true posterior, it does so more slowly
and with greater residual error.
A New Approach to Probabilistic Programming Inference
5.2
DP Mixture
New
( defquery crp-mixture
[ observations alpha mu beta a b ]
( let [ precision-prior ( gamma a b )]
( loop [ observations observations
state-proc ( CRP alpha )
obs-dists {}
states []]
( if ( empty ? observations )
( do
( predict : states states )
( predict : num-clusters ( count obs-dists )))
( let [ state ( sample ( produce state-proc ))
obs-dist ( get obs-dists
state
( let [ l ( sample precision-prior )
s ( sqrt (/ (* beta l )))
m ( sample ( normal mu s ))]
( normal m ( sqrt (/ l )))))]
( observe obs-dist ( first observations ))
( recur ( rest observations )
( absorb state-proc state )
( assoc obs-dists state obs-dist )
( conj states state )))))))
Original (deprecated)
[ assume class-generator ( crp 1.72)]
[ assume class ( mem ( lambda ( n ) ( class-generator )))]
[ assume var ( mem ( lambda ( c ) (* 10 (/ 1 ( gamma 1 10)))))]
[ assume mean ( mem ( lambda ( c ) ( normal 0 ( var c ))))]
[ assume u ( lambda () ( list ( class 1) ( class 2) . . .
( class 9) ( class 10)))]
[ assume K ( lambda () ( count ( unique ( u ))))]
[ assume means ( lambda ( i c )
( if (= i c ) ( list ( mean c ))
( cons ( mean i ) ( means (+ i 1) c ) )))]
[ assume stds ( lambda ( i c )
( if (= i c ) ( list ( sqrt (* 10 ( var c ))))
( cons ( var i ) ( stds (+ i 1) c ) )))]
[ observe ( normal ( mean ( class 1)) ( var ( class 1))) 1.0]
[ observe ( normal ( mean ( class 2)) ( var ( class 2))) 1.1]
.
.
.
[ observe ( normal ( mean ( class 10)) ( var ( class 10))) 0]
[ predict ( u )]
[ predict ( K )]
[ predict ( means 1 ( K ))]
[ predict ( stds 1 ( K ))]
The DP mixture program corresponds to a clustering
with unknown mean and variance problem modelled
via a Dirichlet process mixture of one-dimensional
Gaussians with unknown mean and variance (normalgamma priors). The KL divergence reported is between the running sample estimate of the distribution
over the number of clusters in the data and the ground
truth distribution over the same. The ground truth
distribution over the number of clusters was computed
for this model and data by exhaustively enumerating
all partitions of the data (1.0, 1.1, 1.2, -10, -15, -20,
.01, .1, .05, 0), analytically computing evidence terms
by exploiting conjugacy, and conditioning on partition
cardinality. The fourth plot shows the posterior distribution over the number of classes in the data computed
by both methods relative to the ground truth.
This program was written in a way that was intentionally antagonistic to PMCMC. The continuous likelihood parameters were not marginalized out and the
observe statements were not organized in an optimal
ordering. Despite this, PMCMC outperforms RDB per
simulation, wall clock time, and apply count.
5.3
Branching
New
( defn fib [ n ]
( loop [ a 0 b 1 m 0]
( if (= m n )
a
( recur b (+ a b ) ( inc m )))))
( w i t h - p r i m i t i v e - p r o c e d u r e s [ fib ]
( defquery branching []
( let [ count-prior ( poisson 4)
r ( sample count-prior )
l ( if ( < 4 r )
6
(+ ( fib (* 3 r ))
( sample count-prior )))]
( observe ( poisson l ) 6)
( predict : r r ))))
Original (deprecated)
[ assume fib ( lambda ( n )
( cond ((= n 0) 1) ((= n 1) 1)
( else (+ ( fib (- n 1)) ( fib (- n 2))))))]
[ assume r ( poisson 4)]
[ assume l ( if ( < 4 r ) 6 (+ ( fib (* 3 r )) ( poisson 4)))]
[ observe ( poisson l ) 6]
[ predict r ]
The branching program has no corresponding graphical model. It was designed to test for correctness of
inference in programs with control logic and execution
paths that can vary in the number of sampled values.
It also illustrates mixing in a model where, as shown
in the fourth plot, there is a large mismatch between
the prior and the posterior, so rejection and importance sampling are likely to be ineffective. Because
there is only one observation and just a single named
random variable PMCMC and RDB should and does
achieve essentially indistinguishable performance normalized to simulation, time and apply count.
5.4
Marsaglia
New
( defm marsaglia-normal [ mu std ]
( let [ u ( u ni f or m- c on t in uo u s -1.0 1.0)]
( loop [ x ( sample u )
y ( sample u )]
( let [ s (+ (* x x ) (* y y ))]
( if ( < s 1.0)
(+ mu (* std (* x ( sqrt (* -2.0 (/ ( log s ) s ))))))
( recur ( sample u ) ( sample u )))))))
( defquery g au ss i an - ma rs a gl i a
[ observations sigma mu0 sigma0 ]
( let [ mu ( marsaglia-normal mu0 sigma0 )
likelihood ( normal mu sigma )]
( reduce ( fn [ _ obs ]
( observe likelihood obs ))
nil
observations )
( predict : mu mu )))
Original (deprecated)
[ assume marsaglia-normal
( lambda ( mu std )
( define x ( u n if o rm -c o nt i nu ou s -1.0 1.0))
( define y ( u n if o rm -c o nt i nu ou s -1.0 1.0))
( define s (+ (* x x ) (* y y )))
( if ( < s 1)
(+ mu (* std (* x ( sqrt (* -2.0 (/ ( log s ) s ))))))
( marsaglia-normal mu std )))]
[ assume std ( sqrt 2)]
[ assume mu ( marsaglia-normal 1 ( sqrt 5))]
[ observe ( normal mu std ) (+ 8 1)]
[ observe ( normal mu std ) 8]
[ predict mu ]
Frank Wood, Jan Willem van de Meent, Vikash Mansinghka
Marsaglia is a test program included here for completeness. It is an example of a type of program for
which PMCMC sometimes may not be more efficient.
Marsaglia is the name given to the rejection form of the
Box-Muller algorithm [2] for sampling from a Gaussian [6]. The Marsaglia test program corresponds to
an inference problem in which observed quantities are
drawn from a Gaussian with unknown mean and this
unknown mean is generated by an Anglican implementation of the Marsaglia algorithm for sampling from a
Gaussian. The KS axis is a Kolmogorov-Smirnov test
statistic [5] computed by finding the maximum deviation between the accumulating sample and analytically
derived ground truth cumulative distribution functions
(CDF). Equal-cost PMCMC, RDB, and ground truth
CDFs are shown in the fourth plot.
Because Marsaglia is a recursive rejection sampler it
may require many recursive calls to itself. We conjecture that RDB may be faster than PMCMC here because, while PMCMC pays no statistical cost, it does
pay a computational cost for exploring program traces
that include many random procedure calls that lead to
rejections whereas RDB, due to the implicit geometric prior on program trace length, effectively avoids
paying excess computational costs deriving from unnecessarily long traces.
26
24
4
22
2
20
20
KL
KL*
22
2-2
2-4
2-6
2-4
2-8
-6
2
2-2
100
101
102
2-10 0
10
104
103
Simulation
(a) HMM
101
102
104
103
Simulation
(b) DP Mixture
2
6
2
4
2
4
2
2
2
0
2
2
2
0
2
-2
2
-4
2
-6
10
KL
KL*
Figure 2: Effect of program line permutations
0
10
1
10
2
10
3
Simulation
(a) HMM
10
4
2
-2
2
-4
2
-6
2
-8
2
5.5
Line Permutation
Syntactically and semantically observe and predict’s
are mutually exchangeable (so too are assume’s up to
syntactic constraints). Given this and the nature of
PMCMC it is reasonable to expect that line permutations could effect the efficiency of inference. We explored this by randomly permuting the lines of the
HMM and DP Mixture programs. The results are
shown in Fig. 2 where blue lines correspond to median
(out of twenty five) runs of PMCMC for each of twenty
five program line permutations (including unmodified
(dark) and reversed (light)) and orange are the same
for RDB. For the HMM we found that the natural,
time-sequence ordering of the lines of the program resulted in the best performance for PMCMC relative
to RDB. This is because in this ordering the observe’s
cause re-weighting to happen as soon as possible in
each SMC phase of PMCMC. The effect of permuting code lines interpolates inference performance between optimal where PMCMC is best and adversarial
orderings where RDB is instead. RDB performance is
demonstrated here to be independent of the program
line ordering.
The DP Mixture results show PMCMC outperforming
RDB on all program reorderings. Further, it can be
seen that the original program ordering was not optimal with respect to PMCMC inference.
While PMCMC presents the opportunity for significant gains in inference efficiency, it does not prevent
programmers from seeking to further optimize performance manually. Programmers can influence inference
performance by reordering program lines, in particular pushing observe statements as near to the front
of the program as syntactically allowed, or restructuring programs to lazily rather than eagerly generate
latent variables. Efficiency gains via automatic transformations or online adaptation of the ordering may
be possible.
5.6
Number of Particles
Fig. 3 shows the number of particles in the PMCMC
inference engine affects performance. Performance improves as a function of the number of particles. In this
plot the red line indicates 100 particles. Increasing
from dark to light the number of particles plotted is 2,
5, 10, 20, 50, (red, 100), 200, and 500.
-10
10
0
10
1
10
2
10
3
10
4
Simulation
(b) DP Mixture
Figure 3: Effect of particle count on performance
6
Discussion
The PMCMC approach to probabilistic program interpretation appears to converge faster to the true
conditiona distribution over program execution traces
A New Approach to Probabilistic Programming Inference
for programs that correspond to expressive models
with dense conditional dependencies than single-site
Metropolis-Hastings methods. That PMCMC converges faster in our tests even after normalizing for
computational time is, to us, both surprising and striking. To reiterate, this includes all the computation
done in all the particle executions.
Current versions of Anglican include a multi-threaded
inference core but work remains to achieve optimal
parallelism due, in particular, to memory organization sub-optimality leading to excessive locking overhead. Using just a single thread, PMCMC surprisingly
sometimes outperforms RDB per simulation anyway in
terms of wall clock time.
Our specific choice of syntactically forcing observe’s
to be noisy requires language and interpreter level restrictions and checks. Hard constraint observe’s can
be supported in Anglican by exposing a Dirac likelihood to programmers. Persisting in not doing so
should help programmers avoid writing probabilistic
programs where finding even a single satisfying execution trace is NP-hard or not-computable, but also
could be perceived as requiring a non-intuitive programming style.
As explored in concurrent work on Venture, there may
be opportunities to improve inference performance by
partitioning program variables, either automatically or
via syntax constructs, and treating them differently
during inference. Then, for instance, some could be
sampled via plain MH, some via conditional SMC. One
simple way to do this would be to combine conditional
SMC and RDB MH.
Acknowledgments
We thank Xerox and Google for their generous support. We also thank Arnaud Doucet, Brooks Paige,
and Yura Perov for helpful discussions about PMCMC
and probabilistic programming in general.
References
[1] Christophe Andrieu, Arnaud Doucet, and Roman
Holenstein. Particle Markov chain Monte Carlo methods. Journal of the Royal Statistical Society: Series B
(Statistical Methodology), 72(3):269–342, 2010.
[2] George EP Box and Mervin E Muller. A note on the
generation of random normal deviates. The Annals of
Mathematical Statistics, 29(2):610–611, 1958.
[3] Noah Goodman, Vikash Mansinghka, Daniel Roy,
Keith Bonawitz, and Daniel Tarlow. Church: a
language for generative models.
arXiv preprint
arXiv:1206.3255, 2012.
[4] Roman Holenstein. Particle Markov Chain Monte
Carlo.
PhD thesis, The University of British
Columbia, 2009.
[5] Hubert W Lilliefors. On the Kolmogorov-Smirnov test
for normality with mean and variance unknown. Journal of the American Statistical Association, 62(318):
399–402, 1967.
[6] George Marsaglia and Thomas A Bray. A convenient
method for generating normal variables. Siam Review,
6(3):260–264, 1964.
[7] Tom Minka, J Winn, J Guiver, and D Knowles. Infer.NET 2.4, 2010. Microsoft Research Cambridge.
[8] The Open Group. IEEE Std 1003.1, 2004 Edition, 2004.
URL http://pubs.opengroup.org/
onlinepubs/009695399/functions/fork.html.
[9] Avi Pfeffer. IBAL: A probabilistic rational programming language. In IJCAI, pages 733–740. Citeseer,
2001.
[10] David Spiegelhalter, Andrew Thomas, Nicky Best,
and Wally Gilks. Bugs 0.5: Bayesian inference using
Gibbs sampling manual (version ii). MRC Biostatistics Unit, Institute of Public Health, Cambridge, UK,
1996.
[11] Stan Development Team. Stan modeling language
user’s guide and reference manual.
http://mcstan.org/, 2013.
[12] David Wingate, Andreas Stuhlmueller, and Noah D
Goodman. Lightweight implementations of probabilistic programming languages via transformational compilation. In Proceedings of the 14th international conference on Artificial Intelligence and Statistics, page
131, 2011.
| 6cs.PL
|
From Clarity to Efficiency for Distributed Algorithms ∗
Yanhong A. Liu
Scott D. Stoller
Bo Lin
arXiv:1412.8461v4 [cs.PL] 12 Mar 2017
Computer Science Department, Stony Brook University, Stony Brook, NY 11794, USA
{liu,stoller,bolin}@cs.stonybrook.edu
Abstract
1. Introduction
This article describes a very high-level language for clear
description of distributed algorithms and optimizations necessary for generating efficient implementations. The language supports high-level control flows where complex synchronization conditions can be expressed using high-level
queries, especially logic quantifications, over message history sequences. Unfortunately, the programs would be extremely inefficient, including consuming unbounded memory, if executed straightforwardly.
We present new optimizations that automatically transform complex synchronization conditions into incremental updates of necessary auxiliary values as messages are
sent and received. The core of the optimizations is the first
general method for efficient implementation of logic quantifications. We have developed an operational semantics of
the language, implemented a prototype of the compiler and
the optimizations, and successfully used the language and
implementation on a variety of important distributed algorithms.
Distributed algorithms are at the core of distributed systems.
Yet, developing practical implementations of distributed algorithms with correctness and efficiency assurances remains
a challenging, recurring task.
• Study of distributed algorithms has relied on either pseu-
docode with English, which is high-level but imprecise,
or formal specification languages, which are precise but
harder to understand, lacking mechanisms for building
real distributed systems, or not executable at all.
• At the same time, programming of distributed systems
has mainly been concerned with program efficiency and
has relied mostly on the use of low-level or complex
libraries and to a lesser extent on built-in mechanisms in
restricted programming models.
What’s lacking is (1) a simple and powerful language that
can express distributed algorithms at a high level and yet
has a clear semantics for precise execution as well as for
verification, and is fully integrated into widely used programming languages for building real distributed systems,
together with (2) powerful optimizations that can transform
high-level algorithm descriptions into efficient implementations.
This article describes a very high-level language, DistAlgo,
for clear description of distributed algorithms, combining
advantages of pseudocode, formal specification languages,
and programming languages.
Categories and Subject Descriptors D.1.3 [Programming
Techniques]: Concurrent Programming—Distributed programming; D.3.2 [Programming Languages]: Language
Classifications—Very high-level languages; D.3.4 [Programming Languages]: Processors—Code generation, Compilers, Optimization; F.3.1 [Logics and Meanings of Programs]: Specifying and Verifying and Reasoning about
Programs—Specification techniques; F.3.2 [Logics and
Meanings of Programs]: Semantics of Programming Languages—Operational semantics; I.2.4 [Computing Methodologies]: Knowledge Representation Formalisms and Methods—Predicate logic
General Terms
mance
• The main control flow of a process, including sending
messages and waiting on conditions about received messages, can be stated directly as in sequential programs;
yield points where message handlers execute can be specified explicitly and declaratively.
Algorithms, Design, Languages, Perfor-
• Complex synchronization conditions can be expressed
using high-level queries, especially quantifications, over
message history sequences, without manually writing
message handlers that perform low-level incremental updates and obscure control flows.
Keywords distributed algorithms, high-level queries and
updates, incrementalization, logic quantifications, message
histories, synchronization conditions, yield points
DistAlgo supports these features by building on an objectoriented programming language. We also developed an operational semantics for the language. The result is that distributed algorithms can be expressed in DistAlgo clearly at
a high level, like in pseudocode, but also precisely, like in
∗ This
work was supported in part by NSF under grants CCF1414078, CCF-1248184, CCF-0964196, CNS-0831298, and CCF0613913; and ONR under grants N000141512208, N000140910651 and
N000140710928.
1
2017/3/14
formal specification languages, facilitating formal verification, and can be executed as part of real applications, as in
programming languages.
Unfortunately, programs containing control flows with
synchronization conditions expressed at such a high level
are extremely inefficient if executed straightforwardly: each
quantifier can introduce a linear factor in running time, and
any use of the history of messages sent and received may
cause space usage to be unbounded.
We present new optimizations that allow efficient implementations to be generated automatically, extending previous optimizations to distributed programs and to the most
challenging quantifications.
We have implemented a prototype of the compiler and
the optimizations and experimented with a variety of important distributed algorithms, including Paxos, Byzantine
Paxos, and multi-Paxos. Our experiments strongly confirm
the benefits of the language and the effectiveness of the optimizations.
This article is a revised version of Liu et al. [56]. The
main changes are revised and extended descriptions of the
language and the optimization method, a new formal operational semantics, an abridged and updated description of the
implementation, and a new description of our experience of
using DistAlgo in teaching.
2. Expressing distributed algorithms
• Our method transforms sending and receiving of mes-
sages into updates to message history sequences, incrementally maintains the truth values of synchronization conditions and necessary auxiliary values as those
sequences are updated, and finally removes those sequences as dead code when appropriate.
Even when a distributed algorithm appears simple at a high
level, it can be subtle when necessary details are considered,
making it difficult to understand how the algorithm works
precisely. The difficulty comes from the fact that multiple
processes must coordinate and synchronize to achieve global
goals, but at the same time, delays, failures, and attacks can
occur. Even determining the ordering of events is nontrivial,
which is why Lamport’s logical clock [43] is so fundamental
for distributed systems.
• To incrementally maintain the truth values of general
quantifications, our method first transforms them into
aggregations, also called aggregate queries. In general,
however, translating nested quantifications simply into
nested aggregations can incur asymptotically more space
and time overhead than necessary. Our transformations
minimize the nesting of the resulting queries.
Running example. We use Lamport’s distributed mutual
exclusion algorithm [43] as a running example. Lamport developed it to illustrate the logical clock he invented. The
problem is that n processes access a shared resource, and
need to access it mutually exclusively, in what is called a
critical section (CS), i.e., there can be at most one process
in a critical section at a time. The processes have no shared
memory, so they must communicate by sending and receiving messages. Lamport’s algorithm assumes that communication channels are reliable and first-in-first-out (FIFO).
Figure 1 contains Lamport’s original description of the algorithm, except with the notation < instead of −→ in rule 5
(for comparing pairs of timestamps and process ids using
lexical ordering: (a,b) < (a2,b2) iff a < a2 or a = a2 and
b < b2) and with the word “acknowledgment” added in rule 5
(for simplicity when omitting a commonly omitted [29, 59]
small optimization mentioned in a footnote). This description is the most authoritative, is at a high level, and uses the
most precise English we found.
The algorithm satisfies safety, liveness, and fairness, and
has a message complexity of 3(n − 1). It is safe in that at
most one process can be in a critical section at a time. It
is live in that some process will be in a critical section if
there are requests. It is fair in that requests are served in the
order of the logical timestamps of the request messages. Its
message complexity is 3(n − 1) in that 3(n − 1) messages
are required to serve each request.
• Quantified order comparisons are used extensively in
nontrivial distributed algorithms. They can be incrementalized easily when not mixed with other conditions or
with each other. We systematically extract single quantified order comparisons and transform them into efficient
incremental operations.
Overall, our method significantly improves time complexities and reduces the unbounded space used for message history sequences to the auxiliary space needed for incremental
computation. Systematic incrementalization also allows the
time and space complexity of the generated programs to be
analyzed easily.
There has been a significant amount of related research,
as discussed in Section 7. Our work contains three main
contributions:
• A simple and powerful language for expressing dis-
tributed algorithms with high-level control flows and synchronization conditions, an operational semantics, and
full integration into an object-oriented language.
• A systematic method for incrementalizing complex syn-
chronization conditions with respect to all sending and
receiving of messages in distributed programs.
• A general and systematic method for generating effi-
Challenges. To understand how this algorithm is carried out
precisely, one must understand how each of the n processes
acts as both Pi and Pj in interactions with all other pro-
cient implementations of arbitrary logic quantifications
together with general high-level queries.
2
2017/3/14
in Figure 1; the simplified algorithm can be expressed using
basically two send statements, a receive definition, and an
await statement.
The results on the running example are shown in Figures 2–5, with details explained later. Figure 2 shows Lamport’s original algorithm expressed in DistAlgo; it also includes configuration and setup for running 50 processes each
trying to enter critical section at some point during its execution. Figures 3 and 4 show two alternative optimized programs after incrementalization; all lines with comments are
new except that the await statement is simplified. Figure 5
shows the simplified algorithm.
The algorithm is then defined by the following five rules. For
convenience, the actions defined by each rule are assumed to
form a single event.
1. To request the resource, process Pi sends the message
Tm :Pi requests resource to every other process, and puts that
message on its request queue, where Tm is the timestamp of the
message.
2. When process Pj receives the message Tm :Pi requests
resource, it places it on its request queue and sends a (timestamped) acknowledgment message to Pi .
3. To release the resource, process Pi removes any Tm :Pi
requests resource message from its request queue and sends
a (timestamped) Pi releases resource message to every other
process.
4. When process Pj receives a Pi releases resource message,
it removes any Tm :Pi requests resource message from its request queue.
5. Process Pi is granted the resource when the following two
conditions are satisfied: (i) There is a Tm :Pi requests resource
message in its request queue which is ordered before any other
request in its queue by the relation <. (To define the relation <
for messages, we identify a message with the event of sending
it.) (ii) Pi has received an acknowledgment message from every
other process timestamped later than Tm .
Note that conditions (i) and (ii) of rule 5 are tested locally by
Pi .
3. DistAlgo Language
To support distributed programming at a high level, four
main concepts can be added to commonly used high-level
programming languages, especially object-oriented languages, such as Python and Java: (1) distributed processes,
and sending messages, (2) control flows with yield points
and waits, and receiving messages, (3) synchronization
conditions using high-level queries of message history sequences, and (4) configuration of processes and communication mechanisms. DistAlgo supports these concepts, with
options and generalizations for ease of programming, as described below. A formal operational semantics for DistAlgo
is presented in Appendix A.
Figure 1. Original description in English.
Processes and sending of messages. Distributed processes
are concurrent executions of programmed instructions, like
threads in Java and Python, except that each process has its
private memory, not shared with other processes, and processes communicate by message passing. Three main constructs are used, for defining processes, creating processes,
and sending messages.
A process definition is of form (1) below. It defines a
type p of processes, by defining a class p that extends class
process. The process_body is a set of method definitions and
handler definitions, to be described.
cesses. Each process must have an order of handling all the
events according to the five rules, trying to reach its own goal
of entering and exiting a critical section while also responding to messages from other processes. It must also keep testing the complex condition in rule 5 as events happen.
State machine based formal specifications have been used
to fill in such details precisely, but at the same time, they are
lower-level and harder to understand. For example, a formal
specification of Lamport’s algorithm in I/O automata [59,
pages 647-648] occupies about one and a fifth pages, most
of which is double-column.
To actually implement distributed algorithms, details
for many additional aspects must be added, for example,
creating processes, letting them establish communication
channels with each other, incorporating appropriate logical
clocks (e.g., Lamport clock or vector clock [60]) if needed,
guaranteeing the specified channel properties (e.g., reliable,
FIFO), and integrating the algorithm with the application
(e.g., specifying critical section tasks and invoking the code
for the algorithm as part of the overall application). Furthermore, how to do all of these in an easy and modular fashion?
class p extends process:
process_body
(1)
A special method setup may be defined in process_body for
initially setting up data in the process before the process’s
execution starts. A special method run() may be defined in
process_body for carrying out the main flow of execution. A
special variable self refers to the process itself.
A process creation statement is of form (2) below. It
creates n new processes of type p at each node in the value
of expression node_exp, and returns the resulting process or
set of processes. A node is a running DistAlgo program on a
machine, and is identified by the host name of the machine
plus the name of the running DistAlgo program that can be
specified when starting the program.
Our approach. We address these challenges with the
DistAlgo language, compilation to executable programs, and
especially optimization by incrementalization of expensive
synchronizations, described in Sections 3, 4, and 5, respectively. An unexpected result is that incrementalization led us
to discover simplifications of Lamport’s original algorithm
n new p at node_exp
3
(2)
2017/3/14
The number n and the at clause are optional; the defaults
are 1 and the local node, respectively. A new process can be
set up by calling its setup method. A call start() on the
process then starts the execution of its run() method.
A statement for sending messages is of form (3) below.
It sends the message that is the value of expression mexp to
the process or set of processes that is the value of expression
pexp.
send mexp to pexp
multiple yield points without using method definition and
invocations. As syntactic sugar, a receive that is handled at
only one yield point can be written at that point.
Synchronization and associated actions can be expressed
using general, nondeterministic await statements. A simple
await statement is one of the two forms in (6) below. It waits
for the value of Boolean-valued expression bexp to become
true, for the first form, or waits for a timeout after time period
t, for the second form.
(3)
await bexp
await timeout t
A message can be any value but is by convention a tuple
whose first component is a string, called a tag, indicating the
kind of the message.
A general, nondeterministic await statement is of form (7)
below. It waits for any of the values of expressions bexp1 ,
..., bexpk to become true or a timeout after time period
t, and then nondeterministically selects one of statements
stmt1 , ..., stmtk , stmt whose corresponding conditions are
satisfied to execute. The or and timeout clauses are optional.
Control flows and handling of received messages. The
key idea is to use labels to specify program points where
control flow can yield to handling of messages and resume
afterwards. Three main constructs are used, for specifying
yield points, handling of received messages, and synchronization.
A yield point preceding a statement is of form (4) below,
where identifier l is a label. It specifies that point in the
program as a place where control yields to handling of unhandled messages, if any, and resumes afterwards.
-- l
(6)
await bexp1 : stmt1
or ...
or bexpk : stmtk
timeout t: stmt
(7)
An await statement must be preceded by a yield point, for
handling messages while waiting; if a yield point is not
specified explicitly, the default is that all message handlers
can be executed at this point.
These few constructs make it easy to specify any process that has its own flow of control while also responding to messages. It is also easy to specify any process that
only responds to messages, for example, by writing just
receive definitions and a run() method containing only
await false.
(4)
The label l is optional; it can be omitted when this yield
point is not explicitly referred to in any handler definitions,
defined next.
A handler definition, also called a receive definition, is
of form (5) below. It handles, at yield points labeled l1 ,
..., lj , un-handled messages that match some mexpi sent
from pexpi , where mexpi and pexpi are parts of a tuple
pattern; previously unbound variables in a pattern are bound
to the corresponding components in the value matched. The
handler_body is a sequence of statements to be executed for
the matched messages.
Synchronization conditions using high-level queries. Synchronization conditions and other conditions can be expressed using high-level queries—quantifications, comprehensions, and aggregations—over sets of processes and sequences of messages. High-level queries are used commonly
in distributed algorithms because (1) they make complex
synchronization conditions clearer and easier to write, and
(2) the complexity of distributed algorithms is measured by
round complexity and message complexity, not time complexity of local processing.
Quantifications are especially common because they directly capture the truth values of synchronization conditions.
We discovered a number of errors in our initial programs that
were written using aggregations in place of quantifications
before we developed the method to systematically optimize
quantifications. For example, we regularly expressed “v is
larger than all elements of s” as v > max s and either forgot to handle the case that s is empty or handled it in an
ad hoc fashion. Naive use of aggregation operators like max
may also hinder generation of more efficient implementations.
receive mexp1 from pexp1 , ..., mexpk from pexpk
at l1 , ..., lj :
(5)
handler_body
The from and at clauses are optional; the defaults are any
process and all yield points, respectively. If the from clause is
used, each message is automatically extended with the process id of the sender. A tuple pattern is a tuple in which each
component is a non-variable expression, a variable possibly
prefixed with "=", a wildcard, or recursively a tuple pattern.
A non-variable expression or a variable prefixed with “=”
means that the corresponding component of the tuple being
matched must equal the value of the non-variable expression or the variable, respectively, for pattern matching to succeed. A variable not prefixed with “=” matches any value and
becomes bound to the corresponding component of the tuple being matched. A wildcard, written as “_”, matches any
value. Support for receive mimics common usage in pseudocode, allowing a message handler to be associated with
4
2017/3/14
We define operations on sets; operations on sequences are
the same except that elements are processed in order, and
square brackets are used in place of curly braces.
agg sexp
(10)
agg { exp: v1 in sexp1 , ..., vk in sexpk | bexp }
• In the query forms above, each vi can also be a tuple
• A quantification is a query of one of the two forms in
pattern ti . Variables in ti are bound to the corresponding
components in the matched elements of the value of
sexpi . We omit | bexp when bexp is true.
(8) below, called existential and universal quantifications,
respectively, plus a set of parameters—variables whose
values are bound before the query. For a query to be
well-formed, every variable in it must be reachable from
a parameter—be a parameter or recursively be the leftside variable of a membership clause whose right-side
variables are reachable. Given values of parameters, the
query returns true iff for some or all, respectively, combinations of values of variables that satisfy all membership clauses vi in sexpi , expression bexp evaluates to
true. When an existential quantification returns true, all
variables in the query are also bound to a combination of
values, called a witness, that satisfy all the membership
clauses and condition bexp.
We use {} for empty set; use s.add(x) and s.del(x) for
element addition and deletion, respectively; and use x in s
and x not in s for membership test and its negation, respectively. We assume that hashing is used in implementing
sets, and the expected time of set initialization, element addition and removal, and membership test is O(1). We consider
operations that involve iterations over sets and sequences to
be expensive; each iteration over a set or sequence incurs a
cost that is linear in the size of the set or sequence. All quantifications, comprehensions, and aggregations are considered
expensive.
DistAlgo has built-in sequences received and sent, containing all messages received and sent, respectively, by a
process.
some v1 in sexp1 , ..., vk in sexpk | bexp
(8)
each v1 in sexp1 , ..., vk in sexpk | bexp
For example, the following query returns true iff each
element in s is greater than each element in s2.
• Sequence received is updated only at yield points; after
a message arrives, it will be handled when execution
reaches the next yield point, by adding the message to
received and running matching receive definitions, if
any, associated with the yield point. We use received
m from p interchangeably with m from p in received
to mean that message m from process p is in received;
from p is optional, but when specified, each message in
received is automatically extended with the process id
of the sender.
each x in s, x2 in s2 | x > x2
For another example, the following query, containing a
nested quantification, returns true iff some element in s
is greater than each element in s2. Additionally, when the
query returns true, variable x is bound to a witness—an
element in s that is greater than each element in s2.
some x in s | each x2 in s2 | x > x2
• A comprehension is a query of form (9) below. Given
values of parameters, the query returns the set of values
of exp for all combinations of values of variables that
satisfy all membership clauses vi in sexpi and condition
bexp.
{ exp: v1 in sexp1 , ..., vk in sexpk | bexp }
(9)
• Sequence sent is updated at each send statement; each
message sent to a process is added to sent. We use sent
m to p interchangeably with m to p in sent to mean
that message m to process p is in sent; to p is optional,
but when specified, p is the process to which m was sent
as specified in the send statement.
If implemented straightforwardly, received and sent can
create a huge memory leak, because they can grow unboundedly, preventing their use in practical programming. Our
method can remove them by maintaining only auxiliary values that are needed for incremental computation.
For example, the following query returns the set of products of x in s and x2 in s2 where x is greater than x2.
{x*y: x in s, x2 in s2 | x > x2}
We abbreviate {v: v in sexp | bexp} as {v in sexp |
Configuration. One can specify channel types, handling of
messages, and other configuration items. Such specifications
are declarative, so that algorithms can be expressed without
unnecessary implementation details. We describe a few basic
kinds of configuration items.
First, one can specify the types of channels for passing
messages. For example, the following statement configures
all channels to be FIFO.
bexp}.
• An aggregation, also called an aggregate query, is a query
of one of the two forms in (10) below, where agg is an
aggregation operator, including count, sum, min, and max.
Given values of parameters, the query returns the value of
applying agg to the set value of sexp, for the first form,
or to the multiset of values of exp for all combinations of
values of variables that satisfy all membership clauses vi
in sexpi and condition bexp, for the second form.
configure channel = fifo
5
2017/3/14
1 class P extends process:
2
def setup(s):
3
self.s := s
4
self.q := {}
Other
options for channel include reliable and
{reliable, fifo}. When either fifo or reliable is included, TCP is used for process communication; otherwise,
UDP is used. In general, channels can also be configured
separately for messages from any set of processes to any set
of processes.
One can specify how much effort is spent processing
messages at yield points. For example,
5
6
7
8
9
12
13
14
15
def mutex(task):
# run task with mutual exclusion
-- request
self.t := logical_time()
# 1 in Fig 1
send (’request’, t, self) to s
#
q.add((’request’, t, self))
#
# wait for own req < others in q
# and for acks from all in s
await each (’request’, t2, p2) in q |
# 5 in Fig 1
(t2,p2) != (t,self) implies (t,self) < (t2,p2)
and each p2 in s |
#
some received(’ack’, t2, =p2) | t2 > t
task()
# critical section
-- release
q.del((’request’, t, self))
# 3 in Fig 1
send (’release’, logical_time(), self) to s
#
16
17
18
receive (’request’, t2, p2):
q.add((’request’, t2, p2))
send (’ack’, logical_time(), self) to p2
# 2 in Fig 1
#
#
19
20
21
receive (’release’, _, p2):
for (’request’, t2, =p2) in q:
q.del((’request’, t2, p2))
# 4 in Fig 1
#
#
22
def run():
...
def task(): ...
mutex(task)
...
10
configure handling = all
11
configures the system to handle all un-handled messages at
each yield point; this is the default. For another example,
one can specify a time limit. One can also specify different
handling effort for different yield points.
Logical clocks [27, 43, 60] are used in many distributed
algorithms. One can specify the logical clock, e.g., Lamport
clock, that is used:
configure clock = Lamport
23
24
It configures sending and receiving of messages to update
the clock appropriately. A call logical_time() returns the
current value of the logical clock.
Overall, a DistAlgo program consists of a set of process definitions, a method main, and possibly other, conventional program parts. Method main specifies the configurations and creates, sets up, and starts a set of processes.
DistAlgo language constructs can be used in process definitions and method main and are implemented according to
the semantics described; other, conventional program parts
are implemented according to their conventional semantics.
# set of all other processes
# set of pending requests
#
#
#
#
#
25 def main():
#
...
#
26
configure channel = {reliable,
#
27
configure clock = Lamport
#
28
ps := 50 new P
#
29
for p in ps: p.setup(ps-{p}) #
30
for p in ps: p.start()
#
...
#
main method for the process
do non-CS tasks of the process
define critical section task
run task with mutual exclusion
do non-CS tasks of the process
main method for the application
do other tasks of the application
fifo}
use reliable and FIFO channel
use Lamport clock
create 50 processes of P class
pass to each process other processes
start the run method of each process
do other tasks of the application
Figure 2. Original algorithm (lines 6-21) in a complete program in DistAlgo.
Other language constructs. For other constructs, we use
those in high-level object-oriented languages. We mostly use
Python syntax (indentation for scoping, ’:’ for elaboration,
’#’ for comments, etc.) for succinctness, except with v := exp
for assignment and with a few conventions from Java (keyword extends for subclass, keyword new for object creation,
and omission of self, the equivalent of this in Java, when
there is no ambiguity) for ease of reading.
4. Compiling to executable programs
Compilation generates code to create processes on the specified machine, take care of sending and receiving messages,
and realize the specified configuration. In particular, it inserts appropriate message handlers at each yield point.
Processes and sending of messages. Process creation is
compiled to creating a process on the specified or default
machine and that has a private memory space for its fields.
Each process is implemented using two threads: a main
thread that executes the main flow of control of the process,
and a helper thread that receives and enqueues messages sent
to this process. Constructs involving a set of processes, such
as n new P, can easily be compiled into loops.
Sending a message m to a process p is compiled into calls
to a standard message passing API. If the sequence sent
is used in the program, we also insert sent.add(m to p).
Calling a method on a remote process object is compiled
into a remote method call.
Example. Figure 2 shows Lamport’s algorithm expressed
in DistAlgo. The algorithm in Figure 1 corresponds to the
body of mutex and the two receive definitions, 16 lines
total; the rest of the program, 14 lines total, shows how
the algorithm is used in an application. The execution of
the application starts with method main, which configures
the system to run (lines 25-30). Method mutex and the two
receive definitions are executed when needed and follow
the five rules in Figure 1 (lines 5-21). Recall that there is an
implicit yield point before the await statement.
Note that Figure 2 is not meant to replace Figure 1,
but to realize Figure 1 in a precisely executable manner.
Figure 2 is meant to be high-level, compared with lowerlevel specifications and programs.
Control flows and handling of received messages. Each
yield point l is compiled into a call to a message han6
2017/3/14
dler method l() that updates the sequence received, if
received is used in the program, and executes the bodies
of the receive definitions whose at clause includes l. Pre-
and quantifications are dominantly used in writing synchronization conditions and assertions in specifications and very
high-level programs. Unfortunately, if implemented straightforwardly, each quantification introduces a cost factor that is
linear in the size of the collection quantified over.
Optimizing expensive quantifications in general is difficult, which is a main reason that they are not used in practical programs, not even logic programs, and programmers
manually write more complex and error-prone code. The difficulty comes from expensive enumerations over collections
and complex combinations of join conditions. We address
this challenge by converting quantifications into aggregations that can be optimized systematically using previously
studied methods. However, a quantification can be converted
into multiple forms of aggregations. Which one to use depends on what kinds of updates must be handled, and on how
the query can be incrementalized under those updates. Direct
conversion of nested quantifications into nested aggregations
can lead to much more complex incremental computation
code and asymptotically worse time and space complexities
for maintaining the intermediate query results.
Note that, for an existential quantification, we convert it
to a more efficient aggregation if a witness is not needed;
if a witness is needed, we incrementally compute the set of
witnesses.
cisely:
• Each receive definition is compiled into a method that
takes a message m as argument, matches m against the
message patterns in the receive clause, and if the match-
ing succeeds, binds the variables in the matched pattern
appropriately, and executes the statement in the body of
this receive definition.
• Method l() compiled for yield point l does the following: for each un-handled message m from p to be handled,
(1) execute received.add(m from p) if received is used
in the program, (2) call the methods generated from the
receive definitions whose at clause includes l, and (3)
remove m from the message queue.
An await statement can be compiled into a synchronization using busy-waiting or blocking. We use blocking to wait
until a new message arrives or the timeout specified in await
is reached.
Configuration. Configuration options are taken into account during compilation in a straightforward way. Libraries
and modules are used as much as possible. For example,
when fifo or reliable channel is specified, the compiler
can generate code that uses TCP sockets.
Converting quantifications to aggregations.We present all
converted forms here and describe which forms to use after
we discuss the updates that must be handled. The correctness
of all rules presented have been proved, manually, using
first-order logic and set theory. These rules ensure that the
value of a resulting query expression equals the value of the
original quantified expression.
Table 1 shows general rules for converting single quantifications into equivalent aggregations that use aggregation
operator count. For converting universal quantifications, either rule 2 or 3 could be used. The choice does not affect the
asymptotic cost, but only small constant factors: rule 2 requires maintaining count s, and rule 3 requires computing
not; the latter is generally faster unless count s is already
needed for other purposes, and is certainly faster when not
bexp can be simplified, e.g., when bexp is a negation. The
rules in Table 1 are general because bexp can be any Boolean
expression, but they are for converting single quantifications.
Nested quantifications can be converted one at a time from
inside out, but the results may be much more complicated
than necessary. For example,
5. Incrementalizing expensive synchronizations
Incrementalization transforms expensive computations into
efficient incremental computations with respect to updates
to the values on which the computations depend. It (1) identifies all expensive queries, (2) determines all updates that
may affect the query result, and (3) transforms the queries
and updates into efficient incremental computations. Much
of incrementalization has been studied previously, as discussed in Section 7.
The new method here is for (1) systematic handling of
quantifications for synchronization as expensive queries, especially nested alternating universal and existential quantifications and quantifications containing complex order comparisons and (2) systematic handling of updates caused by
all sending, receiving, and handling of messages in the same
way as other updates in the program. The result is a drastic
reduction of both time and space complexities.
Expensive computations using quantifications. Expensive computations in general involve repetition, including
loops, recursive functions, comprehensions, aggregations,
and quantifications over collections. Optimizations were
studied most for loops, less for recursive functions, comprehensions, and aggregations, and least for quantifications,
basically corresponding to how frequently these constructs
have traditionally been used in programming. However,
high-level queries are increasingly used in programming,
each x in s | some x2 in s2 | bexp
would be converted using rule 1 to
each x in s | count {x2 in s2 | bexp} != 0
and then using rule 2 to
count {x in s | count {x2 in s2 | bexp} != 0}
= count s
7
2017/3/14
A simpler conversion is possible for this example, using a
rule in Table 2, described next.
from the parameter values. The most basic updates are assignments to query parameters, v := exp, where v is a query
parameter. Other updates are to objects and collections used
in the query. For objects, all updates can be expressed as field
assignments, o.f := exp. For collections, all updates can be
expressed as initialization to empty and element additions
and removals, s.add(x) and s.del(x).
For distributed algorithms, a distinct class of important
updates are caused by message passing. Updates are caused
in two ways:
Table 1. Rules for converting single quantifications.
Quantification
Aggregation
1 some x in s | bexp count {x in s | bexp} != 0
2
count {x in s | bexp} = count s
each x in s | bexp
3
count {x in s | not bexp} = 0
Table 2 shows general rules for converting nested quantifications into equivalent, but non-nested, aggregations that
use aggregation operator count. These rules yield much simpler results than repeated use of the rules in Table 1. For example, rule 2 in this table yields a much simpler result than
using two rules in Table 1 in the previous example. More
significantly, rules 1, 4, and 5 generalize to any number of
the same quantifier, and rules 2 and 3 generalize to any number of quantifiers with one alternation. We have not encountered more complicated quantifications than these in the algorithms we found. It is well known that more than one alternation is rarely used, so commonly used quantifications can
all be converted to non-nested aggregations. For example,
in twelve different algorithms expressed in DistAlgo [56],
there are a total of 50 quantifications but no occurrence of
more than one alternation.
Table 3 shows general rules for converting single quantifications with a single order comparison, for any linear order,
into equivalent queries that use aggregation operators max
and min. These rules are useful because max and min can in
general be maintained incrementally in O(log n) time with
O(n) space overhead. Additionally, when there are only element additions, max and min can be maintained most efficiently in O(1) time and space.
Table 4 shows general rules for decomposing Boolean
combinations of conditions in quantifications, to obtain
quantifications with simpler conditions. In particular, Boolean
combinations of order comparisons and other conditions can
be transformed to extract quantifications each with a single order comparison, so the rules in Table 3 can be applied, and Boolean combinations of inner quantifications
and other conditions can be transformed to extract directly
nested quantifications, so the rules in Table 2 can be applied.
For example,
1. Sending and receiving messages updates the sequences
sent and received, respectively. Before incrementalization, code is generated, as described in Section 4, to explicitly perform these updates.
2. Handling of messages by code in receive definitions
updates variables that are parameters of the queries for
computing synchronization conditions, or that are used
to compute the values of these parameters.
Once these are established, updates can be determined using
previously studied analysis methods, e.g., [33, 51].
Incremental computation. Given expensive queries and
updates to the query parameters, efficient incremental computations can be derived for large classes of queries and updates based on the language constructs used in them or by
using a library of rules built on existing data structures [51,
53, 54, 65].
For aggregations converted from quantifications, algebraic properties of the aggregation operators are exploited
to efficiently handle possible updates. In particular, each resulting aggregate query result can be obtained in O(1) time
and incrementally maintained in O(1) time per update to
the sets maintained and affected plus the time for evaluating
the conditions in the aggregation once per update. The total
maintenance time at each element addition or deletion to a
query parameter is at least a linear factor smaller than computing the query result from scratch. Additionally, if aggregation operators max and min are used and there are only element additions, the space overhead is O(1). Note that if max
and min are used naively when there are element deletions,
there may be an unnecessary overhead of O(n) space and
O(log n) maintenance time per update from using more sophisticated data structures to maintain the max or min under
element deletion [21, 84, 85].
Incremental computation improves time complexity only
if the total time of repeated expensive queries is larger than
that of repeated incremental maintenance. This is generally
true for incrementalizing expensive synchronization conditions because (1) expensive queries in the synchronization
conditions need to be evaluated repeatedly at each relevant
update to the message history, until the condition becomes
true, and (2) incremental maintenance at each such update is
at least a linear factor faster for single message updates and
no slower generally than computing from scratch.
each x in s | bexp implies y < x
can be converted using rule 8 in Table 4 to
each x in {x in s | bexp} | y < x
which can then be converted using rule 13 of Table 3 to
{x in s | bexp} = {} or y < min {x in s | bexp}
Updates caused by message passing. Recall that the parameters of a query are variables in the query whose values are bound before the query. Updates that may affect the
query result include not only updates to the query parameters but also updates to the objects and collections reachable
8
2017/3/14
Table 2. Rules for converting nested quantifications.
Nested Quantifications
1 some x in s | some x2 in s2 | bexp count
2 each x in s | some x2 in s2 | bexp count
3 some x in s | each x2 in s2 | bexp count
count
4
each x in s | each x2 in s2 | bexp count
5
count
{(x,x2):
{x: x in
{x: x in
{(x,x2):
{(x,x2):
{(x,x2):
9
10
11
12
13
14
15
16
some
some
some
some
some
some
some
some
each
each
each
each
each
each
each
each
Existential
x in s |
x in s |
x in s |
x in s |
x in s |
x in s |
x in s |
x in s |
Universal
x in s |
x in s |
x in s |
x in s |
x in s |
x in s |
x in s |
x in s |
Aggregation
y
x
y
x
y
x
y
x
<= x
>= y
>= x
<= y
< x
> y
> x
< y
• For non-nested quantifications, if the conditions contain
s != {} and y <= max s
no order comparisons or there are deletions from the sets
or sequences whose elements are compared, the rules in
Table 1 are used. The space overhead is linear in the sizes
of the sets maintained and being aggregated over.
s != {} and y >= min s
s != {} and y < max s
• For non-nested quantifications, if the conditions contain
order comparisons and there are only additions to the
sets or sequences whose elements are compared, the rules
in Table 4 are used to extract single quantified order
comparisons, and then the rules in Table 3 are used to
convert the extracted quantifications. In this case, the
space overhead is reduced to constant.
s != {} and y > min s
Aggregation
y
x
y
x
y
x
y
x
<= x
>= y
>= x
<= y
< x
> y
> x
< y
s = {} or y <= min s
s = {} or y >= max s
• For nested quantifications with one level of nesting, the
rules in Table 4 are used to extract directly nested quantifications, and then the rules in Table 2 are used. If the resulting incremental maintenance has constant-time overhead maintaining a linear-space structure, we are done. If
it is linear-time overhead maintaining a quadratic-space
structure, and if the conditions contain order comparisons, then the rules in Table 4 are used to extract single quantified order comparisons, and then the rules in
Table 3 are used. This can reduce the overhead to logarithmic time and linear space.
s = {} or y < min s
s = {} or y > max s
Table 4. Rules for decomposing conditions to extract quantified comparisons.
1
2
3
4
5
6
7
8
Quantification
some x in s
| not e
some x in s
| e1 and e2
some x in s
| e1 or e2
some x in s
| e1 implies e2
each x in s
| not e
each x in s
| e1 and e2
each x in s
| e1 or e2
each x in s
| e1 implies e2
Decomposed Quantifications
• In general, multiple ways of conversion may be possible,
besides small constant-factor differences between rules 2
and 3 in Table 1 and rules 4 and 5 in Table 2. In particular,
for nested quantifications with two or more alternations,
one must choose which two alternating quantifiers to
transform first, using rule 2 or 3 in Table 2. We have not
encountered such queries and have not studied this aspect
further. Our general method is to transform in all ways
possible, obtain the time and space complexities for each
result, and choose one with the best time and then space.
Complexities are calculated using the cost model of the
set operations given in Section 3. The number of possible
ways is exponential in the worst case in the size of the
query, but the query size is usually a small constant.
not each x in s | e
some x in {x
| e2
(some x in s
(some x in s
(some x in s
(some x in s
in s | e1}
|
|
|
|
e1) or
e2)
not e1) or
e2)
not some x in s | e
(each x in s
(each x in s
each x in {x
| e2
each x in {x
| e2
s2 | bexp} != 0
bexp} = count s
not bexp} != count s
s2 | bexp} =
s2}
s2 | not bexp} = 0
To allow the most efficient incremental computation under all given updates, our method transforms each top-level
quantification as follows:
Table 3. Rules for single quantified order comparison.
1
2
3
4
5
6
7
8
Aggregation
x in s, x2 in
s, x2 in s2 |
s, x2 in s2 |
x in s, x2 in
x in s, x2 in
x in s, x2 in
| e1) and
| e2)
in s | not e1}
in s | e1}
Table 5 summarizes well-known incremental computation
methods for these aggregate queries. The methods are expressed as incrementalization rules: if a query in the program
9
2017/3/14
matches the query form in the table, and each update to a parameter of the query in the program matches an update form
in the table, then transform the query into the corresponding replacement and insert at each update the corresponding
maintenance; fresh variables are introduced for each different query to hold the query results or auxiliary data structures. In the third rule, data structure ds stores the argument
set s of max and supports priority queue operations.
Example. In the program in Figure 2, three quantifications
are used in the synchronization condition in the await statement, and two of them are nested. The condition is copied
below, except that (’ack’,t2,=p2) in received is used in
place of received(’ack’,t2,=p2).
Table 5. Incrementalization rules for count and for max.
Converting quantifications into aggregations as described
using Tables 1 through 4 proceeds as follows. In the first conjunct, the universal quantification is converted using rule 2 or
3 in Table 1, because it contains an order comparison with
elements of q and there are element deletions from q; rule 3
is used here because it is slightly simpler after the negated
condition is simplified. In the second conjunct, the nested
quantification is converted using rule 2 in Table 2. The resulting expression is:
Query
count s
Updates
s := {}
s.add(x)
s.del(x)
Replacement
number
Inserted Maintenance
number := 0
if x not in s: number +:= 1
if x in s: number -:= 1
Cost
O(1)
Cost
O(1)
O(1)
O(1)
Query
max s
Updates
s := {x}
s.add(x)
Replacement
maximum
Inserted Maintenance
maximum := x
if x > maximum: maximum := x
Cost
O(1)
Cost
O(1)
O(1)
each (’request’, t2, p2) in q |
(t2,p2) != (t,self) implies (t,self) < (t2,p2)
and each p2 in s |
some (’ack’, t2, =p2) in received | t2 > t
count {(’request’, t2, p2) in q |
(t,self) > (t2,p2)} = 0
and
count {p2: p2 in s, (’ack’, t2, p2) in received |
t2 > t} = count s
Query
Replacement
Cost
max s
ds.max()
O(1)
Updates
Inserted Maintenance
Cost
s := {}
ds := new DS()
O(1)
s := {x} ds := new DS(); ds.add(x)
O(1)
s.add(x) if x not in s: ds.add(x)
O(log |s|)
s.del(x) if x in s: ds.del(x)
O(log |s|)
The rule for min is similar to the rule for max.
Updates to parameters of the first conjunct are additions
and removals of requests to and from q, and also assignment
to t. Updates to parameters of the second conjunct are additions of ack messages to received, and assignment to t,
after the initial assignment to s.
Incremental computation [51, 53, 54, 65] introduces variables to store the values of all three aggregations in the converted query, transforms the aggregations to use the introduced variables, and incrementally maintains the stored values at each of the updates, as follows, yielding Figure 3.
The overall incrementalization algorithm [51, 53, 65]
introduces new variables to store the results of expensive
queries and subqueries, as well as appropriate additional
values, forming a set of invariants, transforms the queries
and subqueries to use the stored query results and additional
values, and transforms updates to query parameters to also
do incremental maintenance of the stored query results and
additional values.
In particular, if queries are nested, inner queries are transformed before outer queries. Note that a comprehension
such as {x in s | bexp} is incrementalized with respect to
changes to parameters of Boolean expression bexp as well
as addition and removal of elements of s; if bexp contains
nested subqueries, then after the subqueries are transformed,
incremental maintenance of their query results become additional updates to the enclosing query.
At the end, variables and computations that are dead in
the transformed program are eliminated. In particular, sequences received and sent will be eliminated when appropriate, because queries using them have been compiled
into message handlers that only store and maintain values
needed for incremental evaluation of the synchronization
conditions.
• For the first conjunct, store the set value and the count
value in two variables, say earlier and number1, respectively, so first conjunct becomes number1 = 0; when t is
assigned a new value, let earlier be q and let number1
be its size, taking O(|earlier |) time, amortized to O(1)
time when each request in earlier is served; when a
request is added to q, if t is defined and (t,self) >
(t2,p2) holds, add the request to earlier and increment
number1 by 1, taking O(1) time; similarly for deletion
from q. A test of definedness, here t != undefined, is
inserted for any variable that might not be defined in the
scope of the maintenance code.
Note that when (’request’,t,self) in particular is
added to or removed from q, earlier and number1 are
not updated, because (t,self) > (t,self) is trivially
false.
• For the second conjunct, store the set value and the
two count values in three variables, say responded,
number2, and total, respectively, so the conjunct becomes number2 = total; when s is initialized in setup,
10
2017/3/14
1 class P extends process:
2
def setup(s):
3
self.s := s
4
self.total := count s
5
self.q := {}
assign total the size of s, taking O(|s|) time, done only
once for each process; when t is assigned a new value,
let responded be {}, and let number2 be 0, taking O(1)
time; when an ack message is added to received, if the
associated conditions hold, increment number2 by 1, taking O(1) time. A test of definedness of t is omitted
in the maintenance for receiving ack messages, because
t is always defined there; this small optimization is incorporated in an incrementalization rule, but it could be
done with a data-flow analysis that covers distributed data
flows.
6
7
8
9
10
11
12
13
14
15
Note that incrementalization uses basic properties about
primitives and libraries. These properties are incorporated
in incrementalization rules. For the running example, the
property used is that a call to logical_time() returns a
timestamp larger than all existing timestamp values, and
thus at the assignment to t in method mutex, we have that
earlier is q and responded is {}. So, an incrementalization rule for maintaining earlier specifies that at update
t := logical_time(), the maintenance is earlier := q;
similarly for maintaining responded. These simplifications
could be facilitated with data-flow analyses that determine
variables holding logical times and sets holding certain element types. Incrementalization rules can use any program
analysis results as conditions [54].
Figure 3 shows the optimized program after incrementalization of the synchronization condition on lines 10-11 in
Figure 2. All lines with comments are new except that the
synchronization condition in the await statement is simplified. The synchronization condition now takes O(1) time,
compared with O(|s|2 ) if computed from scratch. The tradeoff is the amortized O(1) time overhead at updates to t
and q and on receiving of ack messages. Using based representation for sets [17, 34, 64], maintaining earlier and
responded can each be done using one bit for each process.
Note that the sequence received used in the synchronization condition in Figure 2 is no longer used after incrementalization. All values needed for evaluating the synchronization condition are stored in new variables introduced:
earlier, number1, responded, number2, and total, a drastic
space improvement from unbounded for received to linear
in the number of processes.
# total num of other processes
16
17
18
19
def mutex(task):
-- request
self.t := logical_time()
self.earlier := q
#
self.number1 := count earlier #
self.responded := {}
#
self.number2 := 0
#
send (’request’, t, self) to s
q.add((’request’, t, self))
await number1 = 0
and number2 = total
#
task()
-- release
q.del((’request’, t, self))
send (’release’, logical_time(),
20
21
22
23
24
25
26
27
receive (’request’, t2, p2):
if t != undefined:
# if t is defined
if (t,self) > (t2,p2):
# comparison in conjunct 1
if (’request’,t2,p2) not in earlier:
# if not in earlier
earlier.add((’request’, t2, p2))
# add to earlier
number1 +:= 1
# increment number1
q.add((’request’, t2, p2))
send (’ack’, logical_time(), self) to p2
28
29
30
31
32
33
receive (’ack’, t2, p2):
if t2 > t:
if p2 in s:
if p2 not in responded:
responded.add(p2)
number2 +:= 1
34
35
36
37
38
39
40
41
receive (’release’, _, p2):
for (’request’, t2, =p2) in q:
if t != undefined:
# if t is defined
if (t,self) > (t2,p2):
# comparison in conjunct 1
if (’request’,t2,p2) in earlier:
# if in earlier
earlier.del((’request’, t2, p2))
# delete from earlier
number1 -:= 1
# decrement number1
q.del((’request’, t2, p2))
#
#
#
#
#
#
set
num
set
num
of
of
of
of
pending earlier requests
pending earlier requests
responded processes
responded processes
use maintained results
self) to s
new message handler
comparison in conjunct 2
membership in conjunct 2
if not responded already
add to responded
increment number2
Figure 3. Optimized program after incrementalization. Definitions of run and main are as in Figure 2.
Consider the first conjunct in the synchronization condition in the await statement in Figure 2, copied below:
each (’request’, t2, p2) in q |
(t2,p2) != (t,self) implies (t,self) < (t2,p2)
One might have written the following instead, because it
seems natural, especially if universal quantification is not
supported:
Example with naive use of aggregation operator min. Note
that the resulting program in Figure 3 does not need to use
a queue at all, even though a queue is used in the original
description in Figure 1; the variable q is simply a set, and
thus element addition and removal takes O(1) time.
We show that if min is used naively, a more sophisticated data structure [21, 84, 85] supporting priority queue
is needed, incurring an O(log n) time update instead of the
O(1) time in Figure 3. Additionally, for a query using min to
be correct, special care must be taken to deal with the case
when the argument to min is empty, because then min is undefined.
(t,self) < min {(t2,p2): (’request’, t2, p2) in q
| (t2,p2) != (t,self)}
However, that is incorrect, because the argument of min may
be empty, in which case min is undefined.
Instead of resorting to commonly used special values,
such as maxint, which is ad hoc and error prone in general,
the empty case can be added as the first disjunct of a disjunction:
11
2017/3/14
1 class P extends process:
2
def setup(s):
3
self.s := s
4
self.total := count s
5
self.q := {}
6
self.ds := new DS()
{(t2,p2): (’request’, t2, p2) in q
| (t2,p2) != (t,self)} = {}
or
(t,self) < min {(t2,p2): (’request’, t2, p2) in q
| (t2,p2) != (t,self)}
15
16
17
18
#
#
def mutex(task):
-- request
self.t := logical_time()
self.responded := {}
#
self.number := 0
#
send (’request’, t, self) to s
q.add((’request’, t, self))
await (ds.is_empty() or (t,self)
and number = total
#
task()
-- release
q.del((’request’, t, self))
send (’release’, logical_time(),
19
20
21
22
receive (’request’, t2, p2):
ds.add((t2,p2))
# add to data structure
q.add((’request’, t2, p2))
send (’ack’, logical_time(), self) to p2
23
24
25
26
27
28
receive (’ack’, t2, p2):
if t2 > t:
if p2 in s:
if p2 not in responded:
responded.add(p2)
number +:= 1
29
30
31
32
receive (’release’, _, p2):
for (’request’, t2, =p2) in q:
ds.del((t2,p2))
# delete from data structure
q.del((’request’, t2, p2))
7
8
9
10
11
12
13
14
In fact, the original universal quantification in the first conjunct in the await statement can be converted exactly to this
disjunction by using rule 8 in Table 4 and then rule 13 in Table 3. Our method does not consider this conversion because
it leads to a worse resulting program.
Figure 4 shows the resulting program after incrementalization of the synchronization condition that uses the disjunction above, where ds stores the argument set of min and
supports priority queue operations. All commented lines are
new compared to Figure 2 except that the synchronization
condition in the await statement is simplified. The program
appears shorter than Figure 3 because the long complex code
for maintaining the data structure ds is not included; it is in
fact similar to Figure 3 except that ds is used and maintained
instead of earlier and number1.
The program in Figure 4 is still a drastic improvement
over the original program in Figure 2, with the synchronization condition reduced to O(1) time and with received removed, just as in Figure 3. The difference is that maintaining
ds for incrementalizing min under element addition to and
deletion from q takes O(log |s|) time, as opposed to O(1)
time for maintaining earlier and number1 in Figure 3.
# total num of other processes
#
#
#
#
#
#
data structure for maintaining
requests by other processes
set of responded processes
num of responded processes
< ds.min())
use maintained results
self) to s
new message handler
comparison in conjunct 2
membership in conjunct 2
if not responded already
add to responded
increment number
Figure 4. Optimized program with use of min after incrementalization. Definitions of run and main are as in Figure 2.
Simplifications to the original algorithm. Consider the
original algorithm in Figure 2. Note that incrementalization
determined that there is no need for a process to update
auxiliary values for its own request, in both Figures 3 and 4.
Based on this, we discovered, manually, that updates to q for
a process’s own request do not affect the two uses of q, on
lines 9 and 35, in Figure 3 and the only use of q, on line 30,
in Figure 4. So we can remove them in Figures 3 and 4. In
addition, we can remove them on lines 9 and 14 in Figure 2
and remove the test (t2,p2) != (t,self), which becomes
always true, in the synchronization condition, yielding a
simplified original algorithm.
Furthermore, note that the remaining updates to q in Figure 2 merely maintain pending requests by others, so we can
remove lines 4, 17, 20, 21, and the entire receive definition
for release messages, by using, for the first conjunct in the
await statement,
clock and FIFO channels. The incrementalization rules for
maintaining the result of the new condition incorporate this
property in a similar way as described for Figure 3, except
it could be facilitated with also a data-flow analysis that determines the component of a received message holding the
sender of the message.
1 class P extends process:
2
def setup(s):
3
self.s := s
4
5
6
7
8
10
11
12
def mutex(task):
-- request
self.t := logical_time()
send (’request’, t, self) to s
await each received(’request’, t2, p2) |
not (some received(’release’, t3, =p2) | t3 > t2)
implies (t,self) < (t2,p2)
and each p2 in s |
some received(’ack’, t2, =p2) | t2 > t
task()
-- release
send (’release’, logical_time(), self) to s
13
14
receive (’request’, _, p2):
send (’ack’, logical_time(), self) to p2
9
each received(’request’, t2, p2) |
not (some received(’release’, t3, =p2) | t3 > t2)
implies (t,self) < (t2,p2)
Figure 5. Simplified algorithm. Definitions of run and main
are as in Figure 2.
Figure 5 shows the resulting simplified algorithm. Incrementalizing this program yields essentially the same programs as in Figures 3 and 4, except that it needs to use the
property that when a message is added to received, all messages from the same process in received have a smaller
timestamp. This property follows from the use of logical
6. Implementation and experiments
We have developed a prototype implementation of the compiler and optimizations for DistAlgo and evaluated it in im12
2017/3/14
plementing a set of well-known distributed algorithms, as
described previously [56]. We have also used DistAlgo in
teaching distributed algorithms and distributed systems, and
students used the language and system in programming assignments and course projects. We summarize results from
the former and describe experience with the latter, after an
overview and update about the implementation.
Our DistAlgo implementation takes DistAlgo programs
written in extended Python, applies analyses and optimizations, especially to the high-level queries, and generates executable Python code. It optionally interfaces with an incrementalizer to apply incrementalization before generating
code. Applying incrementalization uses the methods and implementation from previous work: a library of incrementalization rules was developed, manually but mostly following
a systematic method [51, 53], and applied automatically using InvTS [33, 54]. A set of heuristics are currently used to
select the best program generated from incrementalizing differently converted aggregations.
A more extensive implementation of DistAlgo than the
first prototype [56] has been released and is being gradually
improved [25]. Improved methods and implementation for
incrementalization are also being developed [57], to replace
manually written incrementalization rules, and to better select the best transformed programs.
for the incremental program, but grows linearly with the
number of requests for the original program.
• Compared with running times of our best, manually writ-
ten programs in programming languages, all running on a
single machine, our generated DistAlgo takes about twice
as long as our Python version, which takes about twice as
long as our Java version, which takes about twice as long
as our C version, which takes about four times as long as
our Erlang version.
Python is well known to be slow compared Java and C, and
we have not focused on optimizing constant factors. Erlang
is significantly faster than C and the rest because of its use
of light-weight threads to implement processes that is facilitated by its being a functional language. However, among
all our programs for Lamport’s distributed mutual exclusion,
Erlang is the only one besides un-incrementalized DistAlgo
whose memory usage for a fixed number of processes grows
linearly with the number of requests.
Programming distributed algorithms at a high level has
also allowed us to discover several improvements to correctness and efficiency aspects of some of the algorithms [55].
For example, in the pseudocode for multi-Paxos [82], in process Commander, waiting for p2b messages containing ballot b from a majority of acceptors is expressed by starting with a waitfor set initialized to acceptors and then, in
a for ever loop, repeatedly updating waitfor and testing
|waitfor| < |acceptors|/2 as each p2b message containing ballot b arrives. The test is incorrect if implemented directly in commonly used languages such as Java, and even
Python until Python 3, because / is integer division, which
discards any fractional part; for example, test 1 < 3/2 becomes false but should be true. In DistAlgo, the entire code
can simply be written as
Evaluation in implementing distributed algorithms. We
have used DistAlgo to implement a variety of well-known
distributed algorithms, including twelve different algorithms
for distributed mutual exclusion, leader election, and atomic
commit, as well as Paxos, Byzantine Paxos, and multiPaxos, as summarized previously [56]; results of evaluation
using these programs are as follows:
• DistAlgo programs are consistently small, ranging from
22 to 160 lines, and are much smaller than specifications
or programs written in other languages, mostly 1/2 to
1/5 of the size; also we were able to find only a few of
these algorithms written in other languages. Our own best
effort to write Lamport’s distributed mutual exclusion in
programming languages resulted in 272 lines in C, 216
lines in Java, 122 lines in Python, and 99 lines in Erlang,
compared with 32 lines in DistAlgo.
await count {a: received (’p2b’,=b) from a} >
(count acceptors)/2
using the standard majority test, and it is correct whether /
is for integer or float.
Experience in teaching distributed algorithms. DistAlgo
has also helped us tremendously in teaching distributed
algorithms, because it makes complex algorithms completely clear, precise, and directly executable. Students learn
DistAlgo quickly through even a small programming assignment, despite that most did not know Python before, thanks
to the power and clarity of Python.
In particular, students in distributed systems courses have
used DistAlgo in dozens of course projects, implementing
the core of network protocols and distributed graph algorithms [59]; distributed coordination services Chubby [16]
and Zookeeper [38]; distributed hash tables Kademlia [61],
Chord [79], Pastry [74], Tapestry [87], and Dynamo [24];
distributed file systems GFS [32] and HDFS [78]; distributed databases Bigtable [19], Cassandra [42], and Mega-
• Compilation times without incrementalization are all un-
der 0.05 seconds on an Intel Core-i7 2600K CPU with
16GB of memory; and incrementalization times are all
under 30 seconds. Generated code size ranges from 1395
to 1606 lines of Python, including 1300 lines of fixed library code.
• Execution time and space confirm the analyzed asymp-
totic time and space complexities. For example, for Lamport’s distributed mutual exclusion, total CPU time is linear in the number of processes for the incrementalized
program, but superlinear for the original program; for a
fixed number of processes, the memory usage is constant
13
2017/3/14
of the code is devoted to distributed systems aspects. The
numbers of teams that chose various other languages are:
Java 15, C++ 3, Go 3, Erlang 2, Node.js 2, Elixir (a variant
of Erlang) 1, JavaScript 1.
store [12]; distributed processing platform MapReduce [23];
and others.
All distributed programming features were used extensively in students’ programs—easy process creation and
setup and sending of messages, high-level control flows
with receive definitions as well as await for synchronization, and declarative configurations—with the exception of queries over message histories, because students had
been trained in many courses to handle events imperatively;
we have not evaluated incrementalization on students’ programs, because execution efficiency has not been a problem.
Overall, students’ experience helps confirm that DistAlgo
allows complex distributed algorithms and services to be
implemented much more easily than commonly used languages such as C++ and Java. We summarize two specific
instances below.
In a graduate class in Fall 2012, most of the 28 students initially planned to use Java or C++ for their course
projects, because they were familiar with those and wanted
to strengthen their experience of using them instead of using
DistAlgo in implementing distributed systems. However, after doing one programming assignment using DistAlgo, all
those students switched to DistAlgo for their course projects,
except for one student, who had extensive experience with
C++, including several years of internship at Microsoft Research programming distributed systems.
• In the last assignment, teams implemented an extension
to the banking service in one language of their choice.
59% of the teams chose DistAlgo for this, even though
most students (about 80%) did not know Python, and
none knew DistAlgo, at the beginning of the class. In
other words, a majority of students decided that implementation of this type of system is better in DistAlgo,
even compared to languages with which they had more
experience and that are more widely used.
• We asked each team to compare their experiences with
the two languages. Teams consistently reported that development in DistAlgo was faster and easier than development in the other language (even though most students
did not know Python before the project), and that the
DistAlgo code was significantly shorter. It is no surprise
that Java and C++ require more code, even when students
used existing networking libraries, which they were encouraged to do. Comparison with Erlang and Go is more
interesting, because they are high-level languages designed to support distributed programming. For the teams
that chose Erlang, the average DistAlgo and Erlang code
sizes, measured as non-empty non-comment line of code,
are 586 and 1303, respectively. For the teams that chose
Go, the average DistAlgo and Go code sizes are 465 and
1695, respectively.
• This student wrote about 3000 lines of C++, compared
to about 300 lines of DistAlgo written by several other
students who chose the same project of implementing
multi-Paxos and several optimizations. Furthermore, his
C++ program was incomplete, lacking some optimizations that other students’ DistAgo programs included.
7. Related work and conclusion
• The student did a re-implementation in DistAlgo quickly
A wide spectrum of languages and notations have been used
to describe distributed algorithms, e.g., [7, 29, 41, 44, 45,
59, 70–72, 81]. At one end, pseudocode with English is
used, e.g., [41], which gives a high-level flow of the algorithms, but lacks the details and precision needed for a complete understanding. At the other end, state machine based
specification languages are used, e.g., I/O automata [39, 59],
which is completely precise, but uses low-level control flows
that make it harder to write and understand the algorithms.
There are also many notations in between these extremes,
some being much more precise or completely precise while
also giving a high-level control flow, e.g., Raynal’s pseudocode [70–72] and Lamport’s PlusCal [45]. However, all
of these languages and notations lack concepts and mechanisms for building real distributed applications, and most of
the languages are not executable.
Many programming languages support programming of
distributed algorithms and applications. Most support distributed programming through messaging libraries, ranging
from relatively simple socket libraries to complex libraries
such as MPI [62]. Many support Remote Procedure Call
(RPC) or Remote Method Invocation (RMI), which allows
after the course1 , confirming that it took about 300 lines.
His biggest surprise was that his C++ program was an
order of magnitude slower than his DistAlgo program.
After several weeks of debugging, he found that it was
due to an improper use of some C++ library function.
The main contrast that the student concluded was the huge
advantage of DistAlgo over C++ in ease of programming
and program understanding, not to mention the unexpected
performance advantage.
In a graduate class in Fall 2014, each team of two students
first implemented a fault-tolerant banking service in two languages: DistAlgo and another language of their choice other
than Python. We excluded Python as the other language, because implementing the same service in such closely related
languages would be less educational. The service uses chain
replication [83] to tolerate crash failures. The service offers
only a few simple banking operations (get balance, deposit,
withdrawal, intra-bank transfer, inter-bank transfer), so most
1 The student wanted to do research on DistAlgo and so was asked to reimplement his project in DistAlgo.
14
2017/3/14
a process to call a subroutine in another process without
the programmer coding the details for this. Many also support asynchronous method invocation (AMI), which allows
the caller to not block and get the reply later. Some programming languages, such as Erlang [26, 46], which has
an actor-like model [2], have support for message passing
and process management built into the language. There are
also other well-studied languages for distributed programming, e.g., Argus [47], Lynx [76], SR [5], Concert/C [8],
and Emerald [15]. These languages all lack constructs for
expressing control flows and complex synchronization conditions at a much higher level; such high-level constructs are
extremely difficult to implement efficiently. DistAlgo’s construct for declaratively and precisely specifying yield points
for handling received messages is a new feature that we have
not seen in other languages. So is DistAlgo’s support of
history variables in high-level synchronization conditions in
non-deterministic await with timeout in a programming language. Our simple combination of synchronous await and
asynchronous receive allows distributed algorithms to be
expressed easily and clearly.
There has been much work on producing executable implementations from formal specifications, e.g., from process
algebras [37], I/O automata [31], Unity [35], and Seuss [40],
as well as from more recently proposed high-level languages for distributed algorithms, e.g., Datalog-based languages Meld [6], Overlog [4], and Bloom [13], a Prologbased language DAHL [58], and a logic-based language
EventML [14, 67]. An operational semantics was studied
recently for a variant of Meld, called Linear Meld, that allows updates to be encoded more conveniently than Meld
by using linear logic [22]. Compilation of DistAlgo to executable implementations is easy because it is designed to
be so and DistAlgo is given an operational semantics. Highlevel queries and quantifications used for synchronization
conditions can be compiled into loops straightforwardly, but
they may be extremely inefficient. None of these prior works
study powerful optimizations of quantifications. Efficiency
concern is a main reason that similar high-level language
constructs, whether for queries or assertions, are rarely used,
if supported at all, in commonly used languages.
Incrementalization has been studied extensively, e.g., [48,
69], both for doing it systematically based on languages, and
in applying it in an ad hoc fashion to specific problems.
However, all systematic incrementalization methods based
on languages have been for centralized sequential programs,
e.g., for loops [3, 30, 52], set languages [36, 53, 65], recursive functions [1, 49, 68], logic rules [50, 75], and objectoriented languages [51, 57, 63, 73]. This work is the first to
extend incrementalization to distributed programs to support
high-level synchronization conditions. This allows the large
body of previous work on incrementalization, especially on
sets and sequences, to be used for optimizing distributed programs.
Quantifications are the centerpiece of first-order logic,
and are dominantly used in writing synchronization conditions and assertions in specifications, but there are few results on generating efficient implementations of them. In
the database area, despite extensive work on efficient implementation of high-level queries, efficient implementation
of quantification has only been studied in limited scope or
for extremely restricted query forms, e.g., [9–11, 20]. In
logic programming, handling of universal quantification is
based on variants of brute-force Lloyd-Topor transformations, e.g., [28, 66]; even state-of-the-art logic programming
systems, e.g., [80], do not support universal quantification.
Our method is the first general and systematic method for incrementalizing arbitrary quantifications. Although they are
much more challenging to optimize than set queries, our
method combines a set of general transformations to transform them into aggregations that can be most efficiently incrementalized using the best previous methods.
To conclude, this article presents a powerful language and
method for programming and optimizing distributed algorithms. There are many directions for future work, from formal verification on the theoretical side, to generating code in
lower-level languages on the practical side, with many additional analyses and optimizations in between. In particular, a language with a high level of abstraction also faciliates formal verification, of not only the high-level programs,
but also the generated efficient implementations when they
are generated through systematic optimizations. Besides developing systematic optimizations, we have started to study
formal verification of distributed algorithms [18] and their
implementations by starting with their high-level, concise
descriptions in DistAlgo.
APPENDIX
A. Semantics of DistAlgo
We give an abstract syntax and operational semantics for a
core language for DistAlgo. The operational semantics is a
reduction semantics with evaluation contexts [77, 86].
A.1 Abstract Syntax
The abstract syntax is defined in Figures 6 and 7. We use
some syntactic sugar in sample code, e.g., we use infix notation for some binary operators, such as and and is.
Notation.
• A symbol in the grammar is a terminal symbol if it starts
with a lower-case letter.
• A symbol in the grammar is a non-terminal symbol if it
starts with an upper-case letter.
• In each production, alternatives are separated by a line-
break.
• * after a non-terminal means “0 or more occurrences”.
• + after a non-terminal means “1 or more occurrences”.
15
2017/3/14
Program ::= Configuration ProcessClass* Method
ProcessClass ::= class ClassName extends ClassName: Method * ReceiveDef *
ReceiveDef ::= receive ReceivePattern+ at Label + : Statement
receive ReceivePattern+ : Statement
ReceivePattern ::= Pattern from InstanceVariable
Method ::= def MethodName(Parameter *) Statement
defun MethodName(Parameter *) Expression
Statement ::=
InstanceVariable := Expression
InstanceVariable := new ClassName
InstanceVariable := { Pattern : Iterator * | Expression }
Statement ; Statement
if Expression: Statement else: Statement
for Iterator : Statement
while Expression: Statement
Expression.MethodName(Expression*)
send Tuple to Expression
Label await Expression : Statement AnotherAwaitClause*
Label await Expression : Statement AnotherAwaitClause* timeout Expression
skip
Expression ::= Literal
Parameter
InstanceVariable
Tuple
Expression.MethodName(Expression*)
UnaryOp(Expression)
BinaryOp(Expression,Expression)
isinstance(Expression,ClassName)
and(Expression,Expression)
/ / conjunction (short-circuiting)
or(Expression,Expression)
/ / disjunction (short-circuiting)
each Iterator | Expression
some Iterator | Expression
Tuple ::= (Expression*)
Figure 6. Abstract syntax, Part 1.
• tθ denotes the result of applying substitution θ to t. We
Constructs whose semantics is given by translation.
represent substitutions as functions from variables to expressions.
1. Constructors for all classes, and setup() methods for
process classes, are eliminated by translation into ordinary methods that assign to the fields of the objects.
Well-formedness requirements on programs.
2. A method call or field assignment that does not explicitly
specify the target object is translated into a method call
or field assignment, respectively, on self.
1. The top-level method in a program must be named
main. It gets executed in an instance of the pre-defined
process class when the program starts.
3. An await statement without an explicitly specified
label—in other words, the associated label is the empty
string—is translated into an await statement with an explicitly specified label, by generating a fresh label name
ℓ, replacing the empty label in that await statement with
ℓ, and inserting ℓ in every at clause in the class containing the await statement.
2. Each label used in a receive definition must be the label
of some statement that appears in the same class as the
receive definition.
3. Invocations of methods defined using def appear only in
method call statements. Invocations of methods defined
using defun appear only in method call expressions.
16
2017/3/14
UnaryOp ::= not
isTuple
len
BinaryOp ::= is
plus
select
/ / Boolean negation
/ / test whether a value is a tuple
/ / length of a tuple
/ / identity-based equality
/ / sum
/ / select(t,i) returns the i’th component of tuple t
Pattern ::= InstanceVariable
TuplePattern
TuplePattern ::= (PatternElement*)
PatternElement ::= Literal
InstanceVariable
=InstanceVariable
Iterator ::= Pattern in Expression
AnotherAwaitClause ::= or Expression : Statement
Configuration ::= configuration ChannelOrder ChannelReliability ...
ChannelOrder ::= fifo
unordered
ChannelReliability ::= reliable
unreliable
ClassName ::= ...
MethodName ::= ...
Parameter ::= ...
InstanceVariable ::= Expression.Field
Field ::= ...
Label ::= ...
Literal ::= BooleanLiteral
IntegerLiteral
...
BooleanLiteral ::= true
false
IntegerLiteral ::= ...
Figure 7. Abstract syntax, Part 2. Ellipses (“...”) are for common syntactic categories whose details are unimportant.
4. The Boolean operators and and each are eliminated as
follows: e1 and e2 is replaced with not(not(e1) or
not(e2)), and each iter | e is replaced with not(some
iter | not(e)).
constants and the variables prefixed with “=” in
(e1 , . . . , en ). Let ēj denote ej after removing the “=”
prefix, if any. The quantification is rewritten as some
x in s | isTuple(x) and len(x) is n and
(select(x,j1), . . ., select(x,jm)) is (ēj1 ,
. . ., ējm ) and bθ.
5. An aggregation is eliminated by translation into a comprehension followed by a for loop that iterates over the
set returned by the comprehension. The for loop updates
an accumulator variable using the aggregation operator.
• Consider the loop for (e1 , . . . , en ) in e : s. Let
6. Iterators containing tuple patterns are rewritten as iterators without tuple patterns, as follows.
• Consider the existential quantification some (e1 , . . . , en )
in s | b. Let x be a fresh variable. Let θ be the
substitution that replaces ei with select(x,i) for
each i such that ei is a variable not prefixed with
“=”. Let {j1 , . . . , jm } contain the indices of the
17
x and S be fresh variables. Let {i1 , . . . , ik } contain
the indices in (e1 , . . . , en ) of variables not prefixed
with “=”. Let θ be the substitution that replaces ei
with select(x,i) for each i in {i1 , . . . , ik }. Let
{j1 , . . . , jm } contain the indices in (e1 , . . . , en ) of
the constants and the variables prefixed with “=”. Let
ēj denote ej after removing the “=” prefix, if any.
Note that e may denote a set or sequence, and duplicate bindings for the tuple of variables (ei1 , . . . , eik )
2017/3/14
are filtered out if e is a set but not if e is a sequence.
The loop is rewritten as the code in Figure 8.
ables to a witness. Such existential quantifications could
be added as a new form of statement.
7. Comprehensions in which some variables are prefixed
with = are translated into comprehensions without such
prefixing. Specifically, for a variable x prefixed with =
in a comprehension, replace occurrences of =x in the
comprehension with occurrences of a fresh variable y,
and add the conjunct y is x to the Boolean condition.
7. Object creation and comprehension are statements, not
expressions, because they have side-effects. Comprehension has the side-effect of creating a new set.
8. Parameter must include self. The values of method parameters cannot be updated (e.g., using assignment statements). For brevity, local variables of methods are omitted from the core language. Consequently, assignment is
allowed only for instance variables.
8. Comprehensions are statically eliminated as follows. The
comprehension x := { e | x1 in e1 , . . ., xn in
en | b }, where each xi is a pattern, is replaced with
9. Semantically, the for loop copies the contents of a (mutable) sequence or set into an (immutable) tuple before
iterating over it, to ensure that changes to the sequence
or set by the loop body do not affect the iteration. An
implementation could use optimizations to achieve this
semantics without copying when possible.
x := new set
for x1 in e1 :
...
for xn in en :
if b:
x.add(e)
10. For brevity, among the standard arithmetic operations (+,
-, *, etc.), we include only one representative operation
in the abstract syntax and semantics; others are handled
similarly.
9. Wildcards are eliminated from tuple patterns by replacing
each occurrence of wildcard with a fresh variable.
10. Remote method invocation, i.e., invocation of a method
on another process after that process has been started, is
translated into message communication.
11. The semantics below does not model real-time, so timeouts in await statements are simply allowed to occur
non-deterministically.
12. We omit the concept of node (process location) from
the semantics, and we omit the node argument of the
constructor when creating instances of process classes,
because process location does not affect other aspects of
the semantics.
Notes.
1. ClassName must include process. process is a predefined class; it should not be defined explicitly. process
has fields sent and received, and it has a method
start.
13. We omit configure handling statements from the
syntax. The semantics is for configure handling =
all. Semantics for other configure handling options
can easily be added.
2. The grammar allows receive definitions to appear in
classes that do not extend process, but such receive
definitions are useless, so it would be reasonable to make
them illegal.
14. To support initialization of a process by its parent, a
process can access fields of another process and invoke
methods on another process before the latter process is
started.
3. The grammar does not allow labels on statements other
than await. A label ℓ on a statement s other than await
is treated as syntactic sugar for label ℓ on await true
: skip followed by statement s.
15. We require that all messages be tuples. This is an inessential restriction; it slightly simplifies the specification of
pattern matching for matching messages against patterns.
4. ClassName must include set and sequence. Sets and
sequences are treated as objects, because they are mutable. These are predefined classes that should not be
defined explicitly. Methods of set include add, del,
contains, min, max, and size. Methods of sequence
include add (which adds an element at the end of the sequence), contains, and length. We give the semantics
explicitly for a few of these methods; the others are handled similarly.
16. A process’s sent sequence contains pairs of the form
(m, d), where m is a message sent by the process to
destination d. A process’s received sequence contains
pairs of the form (m, s), where m is a message received
by the process from sender s.
A.2 Semantic Domains
5. Tuples are treated as immutable values, not as mutable
objects.
The semantic domains are defined in Figure 9.
Notation.
6. All expressions are side-effect free. For simplicity, we
treat quantifications as expressions, so existential quantifications do not have the side-effect of binding vari-
• D ∗ contains finite sequences of values from domain D.
• Set(D) contains finite sets of values from domain D.
18
2017/3/14
S := e
if isinstance(S,set):
S := { x : x in S | isTuple(x) and len(x) is n
and (select(x,j1), . . ., select(x,jm)) is (ēj1 , . . ., ējm ) }
for x in S:
sθ
else: / / S is a sequence
for x in S:
if (isTuple(x) and len(x) is n
and (select(x,j1), . . ., select(x,jm)) is (ēj1 , . . ., ējm ):
sθ
else:
skip
Figure 8. Translation of for loop to eliminate tuple pattern.
• D1 ⇀ D2 contains partial functions from D1 to D2 .
HeapType in the semantics, even though the information
in that function is distributed in the same way as the heap
itself in an implementation.
dom(f ) is the domain of a partial function f .
• The MsgQueue associated with a process by the last
{true, false}
component of a state contains messages, paired with the
sender, that have arrived at the process but have not yet
...
been handled by matching receive definitions.
...
...
A.3 Extended Abstract Syntax
ProcessAddress ∪ NonProcessAddress
Section A.1 defines the abstract syntax of programs that can
Val ∗
be written by the user. Figure 10 extends the abstract syntax
Bool ∪ Int ∪ Address ∪ Tuple
to include additional forms into which programs may evolve
Set(Val )
during evaluation. Only the new productions are shown here;
all of the productions given above carry over unchanged.
Val ∗
(Field ⇀ Val ) ∪ SetOfVal ∪ SeqOfVal
Expression ::= Address
Address ⇀ ClassName
Address.Field
Address ⇀ Object
Statement ::= for Variable intuple Tuple: Statement
ProcessAddress ⇀ LocalHeap
ProcessAddress × ProcessAddress
Figure 10. Extensions to the abstract syntax.
⇀ Tuple ∗
∗
MsgQueue = (Tuple × ProcessAddress)
The statement for v intuple t: s iterates over the
State = (ProcessAddress ⇀ Statement)
elements of tuple t, in the obvious way.
×HeapType × Heap × ChannelStates
A.4 Evaluation Contexts
×(ProcessAddress ⇀ MsgQueue)
Evaluation contexts, also called reduction contexts, are used
to identify the next part of an expression or statement to be
evaluated. An evaluation context is an expression or stateFigure 9. Semantic domains. Ellipses are used for semantic
ment with a hole, denoted [], in place of the next subdomains of primitive values whose details are standard or
expression or sub-statement to be evaluated. Evaluation conunimportant.
texts are defined in Figure 11.
Bool
Int
ProcessAddress
NonProcessAddress
Address
Tuple
Val
SetOfVal
SeqOfVal
Object
HeapType
LocalHeap
Heap
ChannelStates
=
=
=
=
=
=
=
=
=
=
=
=
=
=
Notes.
A.5 Transition Relations
• We require that ProcessAddress and NonProcessAddress
The transition relation for expressions has the form ht : h ⊢
e → e′ , where e and e′ are expressions, ht ∈ HeapType,
and h ∈ LocalHeap. The transition relation for statements
has the form σ → σ ′ where σ ∈ State and σ ′ ∈ State.
Both transition relations are implicitly parameterized by
the program, which is needed to look up method definitions
be disjoint.
• For a ∈ ProcessAddress and h ∈ Heap, h(a) is the
local heap of process a. For a ∈ Address and ht ∈
HeapType, ht(a) is the type of the object with address
a. For convenience, we use a single (global) function for
19
2017/3/14
Val ::= Literal
Address
(Val*)
C ::= []
(Val*,C ,Expression*)
C .MethodName(Expression*)
Address.MethodName(Val *,C ,Expression*)
UnaryOp(C )
BinaryOp(C ,Expression)
BinaryOp(Val,C )
isinstance(C,ClassName)
or(C ,Expression)
some Pattern in C | Expression
C .Field := Expression
Address.Field := C
InstanceVariable := C
C ; Statement
if C : Statement else: Statement
for InstanceVariable in C : Statement
for InstanceVariable intuple Tuple: C
send C to Expression
send Val to C
await Expression : Statement AnotherAwaitClause*
timeout C
Figure 11. Evaluation contexts.
and configuration information. The transition relation for
expressions is defined in Figure 12. The transition relation
for statements is defined in Figures 13–14.
Notation and auxiliary functions.
• extends(c1 , c2 ) holds iff class c1 is a descendant of class
c2 in the inheritance hierarchy.
• For c ∈ ClassName, new (c) returns a new instance of c.
new
(c) =
{} if c = set
hi if c = sequence
f0 otherwise
• For m ∈ MethodName and c ∈ ClassName, the rela-
tion methodDef (c, m, def ) holds iff (1) class c defines
method m, and def is the definition of m in c, or (2) c
does not define m, and def is the definition of m in the
nearest ancestor of c in the inheritance hierarchy that defines m.
• For h, h̄, h̄′ ∈ LocalHeap and ht, ht′ ∈ HeapType and
v, v̄ ∈ Val , the relation isCopy (v, h, h̄, ht, v̄, h̄′ , ht′ )
holds iff (1) v is a value in a process with local heap h,
i.e., addresses in v are evaluated with respect to h, (2)
v̄ is a copy of v for a process whose local heap was h̄
before v was copied into it and whose local heap is h̄′
after v is copied into it, i.e., v̄ is the same as v except that,
instead of referencing objects in h, it references newly
created copies of those objects in h̄′ , and (3) h̄′ and ht′
are versions of h̄ and ht updated to reflect the creation of
those objects. As an exception, because process addresses
are used as global identifiers, process addresses in v are
copied unchanged into v̄, and new copies of process
objects are not created. We give auxiliary definitions and
then a formal definition of isCopy .
For v ∈ Val , let addrs(v, h) denote the set of addresses
that appear in v or in any objects or values reachable from
v with respect to local heap h; formally,
• In the transition rules, a matches an address; v matches a
value (i.e., an element of Val); and ℓ matches a label.
• For an expression or statement e, e[x := y] denotes e
a ∈ addrs(v, h) ⇔
(v ∈ Address ∧ v = a)
∨ (v ∈ dom(h) ∧ h(v) ∈ Field ⇀ Val
∧ (∃f ∈ dom(h(v)). a ∈ addrs(h(v)(f ), h)))
∨ (v ∈ dom(h) ∧ h(v) ∈ SetOfVal ∪ SeqOfVal
∧ (∃v ′ ∈ h(v). a ∈ addrs(v ′ , h)))
∨ (∃v1 , . . . , vn ∈ Val. v = (v1 , . . . , vn )
∧ ∃i ∈ [1..n]. a ∈ addrs(vi , h))
with all occurrences of x replaced with y.
• A function matches the pattern f [x → y] iff f (x) equals
y. For example, in transition rules for statements, a
function P in ProcessAddress ⇀ Statement matches
P [a → s] if P maps process address a to statement s.
• For a function f , f [x := y] denotes the function that is
the same as f except that it maps x to y.
• f0 denotes the empty partial function, i.e., the partial
For v, v̄ ∈ Val and f ∈ Address ⇀ Address, the
relation subst(v, v̄, f ) holds iff v is obtained from v̄ by
replacing each occurrence of an address a in dom(f )
with f (a) (informally, f maps addresses of new objects
in v̄ to addresses of corresponding old objects in v);
formally,
function whose domain is the empty set.
• For a (partial) function f , f ⊖ a denotes the function that
is the same as f except that it has no mapping for a.
• Sequences are denoted with angle brackets, e.g., h0, 1, 2i ∈
Int ∗ .
• s@t is the concatenation of sequences s and t.
subst(v, v̄, f ) =
(v ∈ Bool ∪ Int ∪ (Address \ dom(f )) ∧ v̄ = v)
∨ (v ∈ dom(f ) ∧ f (v̄) = v)
∨ (∃v1 , . . . , vn , v̄1 , . . . , v̄n .
v = (v1 , . . . , vn ) ∧ v̄ = (v̄1 , . . . , v̄n )
∧ (∀i ∈ [1..n]. subst(vi , v̄i , f )))
• first(s) is the first element of sequence s.
• rest(s) is the sequence obtained by removing the first
element of s.
• length(s) is the length of sequence s.
20
2017/3/14
Similarly, for o, ō ∈ Object and f ∈ Address ⇀
Address, the relation subst (o, ō, f ) holds iff o is obtained from ō by replacing each occurrence of an address
1−1
a in dom(f ) with f (a). For sets S and S ′ , let S → S ′
′
be the set of bijections between S and S .
returns {s1 , . . . , sn }. Otherwise, receiveAtLabel ((m, a), ℓ,
c, h) returns ∅.
A.6 Executions
An execution is a sequence of transitions σ0 → σ1 → σ2 →
· · · such that σ0 is an initial state. The set of initial states
is defined in Figure 15. Intuitively, ap is the address of the
initial process, ar is the address of the received sequence
of the initial process, and as is the address of the sent
sequence of the initial process.
Informally, execution of the statement initially associated
with a process may eventually (1) terminate (i.e., the statement associated with the process becomes skip, indicating
that there is nothing left for the process to do), (2) get stuck
(i.e., the statement associated with the process is not skip,
and the process has no enabled transitions) due to an unsatisfied await statement or an error (e.g., the statement contains
an expression that tries to select a component from a value
that is not a tuple, or the statement contains an expression
that tries to read the value of a non-existent field), or (3) run
forever due to an infinite loop or infinite recursion.
Finally, isCopy is defined as follows (intuitively, A contains the addresses of the newly allocated objects):
isCopy (v, h, h̄, ht, v̄, h̄′ , ht′ ) ⇔
∃A ⊂ NonProcessAddress.
1−1
∃f ∈ A → (addrs(v, h) \ ProcessAddress).
A ∩ dom(ht) = ∅
∧ dom(ht′ ) = dom(ht) ∪ A
∧ dom(h̄′ ) = dom(h̄) ∪ A
∧ (∀a ∈ dom(ht). ht′ (a) = ht(a))
∧ (∀a ∈ dom(h̄). h̄′ (a) = h̄(a))
∧ (∀a ∈ A. ht′ (a) = ht(f (a))
∧ subst(h(a), h̄′ (a), f ))
• For m ∈ Val, a ∈ ProcessAddress, ℓ ∈ Label ,
h ∈ LocalHeap, and a receive definition d, if message m can be received from a at label ℓ by a process with local heap h using receive definition d, then
matchRcvDef (m, a, ℓ, h, d) returns the appropriately instantiated body of d.
Acknowledgments
We thank Michael Gorbovitski for supporting the use of InvTS for automatic incrementalization of DistAlgo programs.
We are grateful to the following people for their helpful
comments and discussions: Ken Birman, Andrew Black, Jon
Brandvein, Wei Chen, Ernie Cohen, Mike Ferdman, John
Field, Georges Gonthier, Leslie Lamport, Nancy Lynch,
Lambert Meertens, Stephan Merz, Don Porter, Michel Raynal, John Reppy, Emin Gün Sirer, Doug Smith, Gene Stark,
and Robbert van Renesse. We thank the anonymous reviewers for their detailed and helpful comments.
We first define some auxiliary relations and functions.
The relation matchesDefLbl (d, ℓ) holds iff receive
definition d either lacks an at clause or has an at
clause that includes ℓ. bound(P ) returns the set of
variables that appear in pattern P prefixed with “=”.
vars(P ) returns the set of variables that appear in P .
findSubstPat(m, a, h, P from x) returns the substitution θ with domain vars(P ) ∪ {x} such that m =
P θ ∧ θ(x) = a ∧ (∀y ∈ bound (P ). θ(y) = h(y)), if
any, otherwise it returns ⊥. findSubst(m, a, h, d) returns
findSubstPat(m, a, h, P from x) for the first receive
pattern P from x in d such that findSubstPat(m, a, h,
P from x) 6= ⊥, if any, otherwise it returns ⊥.
References
[1] U. A. Acar, G. E. Blelloch, and R. Harper. Adaptive functional programming. ACM Transactions on Programming
Languages and Systems, 28(6):990–1034, 2006.
[2] G. Agha. Actors: A Model of Concurrent Computation in
Distributed Systems. MIT Press, 1986.
If matchesDefLbl (d, ℓ) ∧ findSubst(m, a, h, d) 6= ⊥,
then matchRcvDef (m, a, ℓ, h, d) returns sθ, where s is
the body of d (i.e., the statement that appears in d) and
θ = findSubst(m, a, h, d), otherwise it returns ⊥.
[3] F. E. Allen, J. Cocke, and K. Kennedy. Reduction of operator
strength. In S. S. Muchnick and N. D. Jones, editors, Program
Flow Analysis, pages 79–101. Prentice-Hall, 1981.
• For m ∈ Val , a ∈ ProcessAddress, ℓ ∈ Label , c ∈
[4] P. Alvaro, T. Condie, N. Conway, J. Hellerstein, and R. Sears.
ClassName, and h ∈ LocalHeap, if message m can be
I do declare: Consensus in a logic language. ACM SIGOPS
received from a at label ℓ in class c by a process with local
Operating Systems Review, 43(4):25–30, 2010.
heap h, then receiveAtLabel ((m, a), ℓ, c, h) returns a set
[5] G. R. Andrews and R. A. Olsson. The SR Programming
of statements that should be executed when receiving m
Language: Concurrency in Practice. Benjamin Cummings,
in that context.
1993.
Specifically, if class c contains a receive definition d
[6] M. P. Ashley-Rollman, P. Lee, S. C. Goldstein, P. Pillai, and
such that matchRcvDef (m, a, ℓ, h, d) is not ⊥, then, letJ. D. Campbell. A language for large ensembles of indepenting d1 , . . . , dn be the receive definitions in c such that
dently executing nodes. In Proceedings of the 25th IntermatchRcvDef (m, a, ℓ, h, di ) is not ⊥, and letting si =
national Conference on Logic Programming, pages 265–280.
Springer, 2009.
matchRcvDef (m, a, ℓ, h, di ), receiveAtLabel ((m, a), ℓ, c, h)
21
2017/3/14
/ / field access
ht : h ⊢ a.f → h(a)(f )
if a ∈ dom(h) ∧ f ∈ dom(h(a))
/ / invoke method in user-defined class
ht : h ⊢ a.m(v1 , . . . , vn ) → e[self := a, x1 := v1 , . . . , xn := vn ]
if a ∈ dom(h) ∧ methodDef (ht(a), m, defun m(x1 , . . . , xn ) e)
/ / invoke method in pre-defined class (representative examples)
ht : h ⊢ a.contains(v1 ) → true if a ∈ dom(h) ∧ ht(a) = set ∧ v1 ∈ h(a)
ht : h ⊢ a.contains(v1 ) → false if a ∈ dom(h) ∧ ht(a) = set ∧ v1 6∈ h(a)
/ / unary operations
ht : h ⊢ not(true) → false
ht : h ⊢ not(false) → true
ht : h ⊢ isTuple(v) → true if v is a tuple
ht : h ⊢ isTuple(v) → false if v is not a tuple
ht : h ⊢ len(v) → n
if v is a tuple with n components
/ / binary operations
ht : h ⊢ is(v1 , v2 ) → true
if v1 and v2 are the same (identical) value
ht : h ⊢ plus(v1 , v2 ) → v3
if v1 ∈ Int ∧ v2 ∈ Int ∧ v3 = v1 + v2
ht : h ⊢ select(v1 , v2 ) → v3
if v2 ∈ Int ∧ v2 > 0 ∧ (v1 is a tuple with at least v2 components) ∧ (v3 is the v2 ’th component of v1 )
/ / isinstance
ht : h ⊢ isinstance(a, c) → true if ht(a) = c
ht : h ⊢ isinstance(a, c) → false if ht(a) 6= c
/ / disjunction
ht : h ⊢ or(true, e) → true
ht : h ⊢ or(false, e) → e
/ / existential quantification
ht : h ⊢ some x in a | e → e[x := v1 ] or · · · or e[x := vn ]
if (ht(a) = sequence ∧ h(a) = hv1 , . . . , vn i) ∨ (ht(a) = set ∧ hv1 , . . . , vn i is a linearization of h(a))
Figure 12. Transition relation for expressions.
[12] J. Baker, C. Bond, J. C. Corbett, J. Furman, A. Khorlin, J. Larson, J.-M. Léon, Y. Li, A. Lloyd, and V. Yushprakh. Megastore: Providing scalable, highly available storage for interactive services. In Proceedings of the Conference on Innovative
Database Research, pages 223–234, 2011.
[7] H. Attiya and J. Welch. Distributed Computing: Fundamentals, Simulations, and Advanced Topics. Wiley, 2nd edition,
2004.
[8] J. S. Auerbach, A. P. Goldberg, G. S. Goldszmidt, A. S. Gopal,
M. T. Kennedy, J. R. Rao, and J. R. Russell. Concert/C:
A language for distributed programming. In Proceedings
of the USENIX Winter 1994 Technical Conference. USENIX
Association, 1994.
[13] Berkeley Orders of Magnitude. Bloom Programming Language. http://www.bloom-lang.net, 2013. Lastest release April 23, 2013. Accessed January 14, 2017.
[14] M. Bickford. Component specification using event classes.
In Proceedings of the 12th International Symposium on
Component-Based Software Engineering, pages 140–155.
Springer, 2009.
[9] A. Badia. Question answering and database querying: Bridging the gap with generalized quantification. Journal of Applied
Logic, 5(1):3–19, 2007.
[10] A. Badia, D. Van Gucht, and M. Gyssens. Query languages
with generalized quantifiers. In R. Ramakrishnan, editor,
Applications of Logic in Databases. Kluwer Academic, 1995.
[15] A. P. Black, N. C. Hutchinson, E. Jul, and H. M. Levy. The
development of the Emerald programming language. In Proceedings of the 3rd ACM SIGPLAN Conference on History of
Programming Languages, pages 11–1–11–51, 2007.
[11] A. Badia, B. Debes, and B. Cao. An implementation of a
query language with generalized quantifiers. In Proceedings
of the 27th International Conference on Conceptual Modeling, pages 547–548. Springer, 2008.
[16] M. Burrows. The Chubby lock service for loosely-coupled
distributed systems. In Proceedings of the 7th USENIX Sym-
22
2017/3/14
/ / field assignment
(P [a → a′ .f := v], ht, h[a → ha[a′ → o]], ch, mq)
→ (P [a := skip], ht, h[a := ha[a′ := o[f := v]]], ch, mq)
/ / object creation
(P [a → a′ .f := new c], ht, h[a → ha[a′ → o]], ch, mq)
→ (P [a := skip], ht[a′ := c], h[a := ha[a′ := o[f := ac ], ac := new (c)]], ch, mq)
if ac 6∈ dom(ht) ∧ ac ∈ Address ∧ (ac ∈ ProcessAddress ⇔ extends(c, process))
/ / sequential composition
(P [a → skip; s], ht, h, ch, mq) → (P [a := s], ht, h, ch, mq)
/ / conditional statement
(P [a → if true : s1 else : s2 ], ht, h, ch, mq) → (P [a := s1 ], ht, h, ch, mq)
(P [a → if false : s1 else : s2 ], ht, h, ch, mq) → (P [a := s2 ], ht, h, ch, mq)
/ / for loop
(P [a → for x in a′ : s], ht, h, ch, mq) → (P [a := for x intuple (v1 , . . . , vn ) : s], ht, h, ch, mq)
if ((ht(a) = sequence ∧ h(a)(a′ ) = hv1 , . . . , vn i) ∨ (ht(a) = set ∧ hv1 , . . . , vn i is a linearization of h(a)(a′ )))
(P [a → for x intuple (v1 , . . . , vn ) : s], ht, h, ch, mq)
→ (P [a := s[x := v1 ]; for x intuple (v2 , . . . , vn ) : s], ht, h, ch, mq)
(P [a → for x intuple () : s], ht, h, ch, mq) → (P [a := skip], ht, h, ch, mq)
/ / while loop
(P [a → while e: s], ht, h, ch, mq) → (P [a := if e: (s; while e: s) else : skip], ht, h, ch, mq)
/ / invoke method in user-defined class
(P [a → a′ .m(v1 , . . . , vn )], ht, h, ch, mq)
→ (P [a := s[self := a, x1 := v1 , . . . , xn := vn ]], ht, h, ch, mq)
if a′ ∈ dom(h(a))
∧ ht(a′ ) 6∈ {process, set, sequence} ∧ methodDef (ht(a′ ), m, def m(x1 , . . . , xn ) s)
/ / invoke method in pre-defined class (representative examples)
/ / process.start allocates a local heap and sent and received sequences for the new process,
/ / and moves the started process to the new local heap.
(P [a → a′ .start()], ht, h[a → ha[a′ → o], ch, mq)
→ (P [a := skip, a′ := a′ .run()], ht[as := sequence, ar := sequence],
h[a := ha ⊖ a′ , a′ := f0 [a′ → o[sent := as , received := ar ], ar := hi, as := hi]], ch, mq)
if extends(ht(a′ ), process) ∧ (ht(a′ ) inherits start from process) ∧ ar 6∈ dom(ht) ∧ as 6∈ dom(ht)
∧ ar ∈ NonProcessAddress ∧ as ∈ NonProcessAddress
(P [a → a′ .add(v1 )], ht, h[a → ha], ch, mq)
→ (P [a := skip], ht, h[a := ha[a′ := ha(a′ ) ∪ {v1 }]], ch, mq)
if a′ ∈ dom(ha) ∧ ht(a′ ) = set
(P [a → a′ .add(v1 )], ht, h[a → ha], ch, mq)
→ (P [a := skip], ht, h[a := ha[a′ := ha(a′ )@hv1 i]], ch, mq)
if a′ ∈ dom(ha) ∧ ht(a′ ) = sequence
Figure 13. Transition relation for statements, Part 1.
23
2017/3/14
/ / send a message to one process. create copies of the message for the sender’s sent sequence
/ / and the receiver.
(P [a → send v to a2 ], ht, h[a → ha, a2 → ha2 ], ch, mq)
→ (P [a := skip], ht′′ , h[a := ha′ [as := ha(as )@h(v1 , a2 )i], a2 := ha′2 ],
ch[(a, a2 ) := ch((a, a2 ))@hv2 i], mq)
if a2 ∈ ProcessAddress ∧ as = ha(a)(sent) ∧ isCopy (v, ha, ha, ht, v1 , ha′ , ht′ )
∧ isCopy (v, ha′ , ha2 , ht′ , v2 , ha′2 , ht′′ )
/ / send to a set of processes
(P [a → send v to a′ ], ht, h[a → ha], ch, mq)
→ (P [a := for x in a′ : send v to x], ht, h[a := ha[as := ha(as )@h(v, a′ )i]], ch, mq)
if ht(a′ ) = set ∧ as = ha(a)(sent) ∧ (x is a fresh variable)
/ / message reordering
(P, ht, h, ch[(a, a′ ) → q], mq) → (P, ht, h, ch[(a, a′ ) := q ′ ], mq)
if (channel order is unordered in the program configuration) ∧ (q ′ is a permutation of q)
/ / message loss
(P, ht, h, ch[(a, a′ ) → q], mq) → (P, ht, h, ch[(a, a′ ) := q ′ ], mq)
if (channel reliability is unreliable in the program configuration) ∧ (q ′ is a subsequence of q)
/ / arrival of a message from process a at process a′ . remove message from channel, and append
/ / (message, sender) pair to message queue.
(P, ht, h, ch[(a, a′ ) → q], mq)
→ (P, ht, h, ch[(a, a′ ) := rest(q)], mq[a′ := mq(a′ )@h(first (q), a)i])
if length(q) > 0
/ / handle a message at a yield point. remove the (message, sender) pair from the message
/ / queue, append a copy to the received sequence, and prepare to run matching receive
/ / handlers associated with ℓ, if any. s has a label hence must be await.
(P [a → ℓ s], ht, h[a → ha], ch, mq[a → q])
→ (P [a := s′ [self := a]; ℓ s], ht′ , h[a → ha′ [ar → ha(ar )@hcopyi]], ch, mq[a := rest(q)])
if length(q) > 0 ∧ ar = ha(a)(received) ∧ isCopy (first (q), ha, ha, ht, copy, ha′ , ht′ )
∧ receiveAtLabel (first(q), ℓ, ht(a), ha′ ) = S ∧ s′ is a linearization of S
/ / await without timeout clause
(P [a → ℓ await e1 :s1 or · · · or en :sn ], ht, h, ch, mq) → (P [a := si ], ht, h, ch, mq)
if length(mq(a)) = 0 ∧ i ∈ [1..n] ∧ h(a) : ht ⊢ ei → true
/ / await with timeout clause, terminated by true condition
(P [a → ℓ await e1 :s1 or · · · or en :sn timeout v:s], ht, h, ch, mq) → (P [a := si ], ht, h, ch, mq)
if length(mq(a)) = 0 ∧ i ∈ [1..n] ∧ h(a) : ht ⊢ ei → true
/ / await with timeout clause, terminated by timeout (occurs non-deterministically)
(P [a → ℓ await e1 :s1 or · · · or en :sn timeout v:s], ht, h, ch, mq) → (P [a := s], ht, h, ch, mq)
if length(mq(a)) = 0 ∧ h(a) : ht ⊢ e1 → false ∧ · · · ∧ h(a) : ht ⊢ en → false
/ / context rule for expressions
h(a) : ht ⊢ e → e′
(P [a → C[e]], ht, h, ch, mq) → (P [a := C[e′ ]], ht, h, ch, mq)
/ / context rule for statements
(P [a → s], ht, h, ch, mq) → (P [a := s′ ], ht′ , h′ , ch′ , mq ′ )
(P [a → C[s]], ht, h, ch, mq) → (P [a := C[s′ ]], ht′ , h′ , ch′ , mq ′ )
Figure 14. Transition relation for statements, Part 2.
24
2017/3/14
Init =
{(P, ht, h, ch, mq) ∈ State |
∃ ap ∈ ProcessAddress,
ar ∈ NonProcessAddress,
as ∈ NonProcessAddress.
ar 6= as
∧ P = f0 [ap := ap .main()]
∧ht = f0 [ap := process, ar := sequence, as := sequence]
∧ h = f0 [ap := ha]
∧ ch = (λ(a1 , a2 ) ∈ ProcessAddress ×ProcessAddress. hi)
∧ mq = (λa ∈ ProcessAddress. hi)
where ha = f0 [ap := op , ar := hi, as := hi]
op = f0 [received := ar , sent := as ]}
Figure 15. Initial states.
[27] C. J. Fidge. Timestamps in message-passing systems that
preserve the partial ordering. In Proceedings of the 11th
Australian Computer Science Conference, pages 56–66, 1988.
[28] F. Fioravanti, A. Pettorossi, M. Proietti, and V. Senni. Program
transformation for development, verification, and synthesis of
programs. Intelligenza Artificiale, 5(1):119–125, 2011.
[29] V. K. Garg. Elements of Distributed Computing. Wiley, 2002.
[30] Gautam and S. Rajopadhye. Simplifying reductions. In Conference Record of the 33rd ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages, pages 30–41,
2006. ISBN 1-59593-027-2.
[31] C. Georgiou, N. A. Lynch, and P. M. andJoshua A. Tauber.
Automated implementation of complex distributed algorithms
specified in the IOA language. International Journal on Software Tools for Technology Transfer, 11(2):153–171, 2009.
[32] S. Ghemawat, H. Gobioff, and S.-T. Leung. The Google file
system. ACM SIGOPS Operating Systems Review, 37(5):29–
43, 2003.
posium on Operating Systems Design and Implementation,
pages 335–350, 2006.
[33] M. Gorbovitski, Y. A. Liu, S. D. Stoller, T. Rothamel, and
T. Tekle. Alias analysis for optimization of dynamic languages. In Proceedings of the 6th Symposium on Dynamic
Languages, pages 27–42. ACM Press, 2010.
[17] J. Cai, P. Facon, F. Henglein, R. Paige, and E. Schonberg. Type
analysis and data structure selection. In B. Möller, editor,
Constructing Programs from Specifications, pages 126–164.
North-Holland, 1991.
[34] D. Goyal. A Language Theoretic Approach to Algorithms.
PhD thesis, Department of Computer Science, New York University, 2000.
[18] S. Chand, Y. A. Liu, and S. D. Stoller. Formal verification
of multi-Paxos for distributed consensus. In Proceedings of
the 21st International Symposium on Formal Methods, pages
119–136. Springer, 2016.
[35] A. Granicz, D. M. Zimmerman, and J. Hickey. Rewriting
UNITY. In Proceedings of the 14th International Conference
on Rewriting Techniques and Applications, pages 138–147.
Springer, 2003.
[19] F. Chang, J. Dean, S. Ghemawat, W. C. Hsieh, D. A. Wallach,
M. Burrows, T. Chandra, A. Fikes, and R. E. Gruber. Bigtable:
A distributed storage system for structured data. ACM Transactions on Computer Systems, 26(2):4, 2008.
[36] A. Gupta, I. S. Mumick, and V. S. Subrahmanian. Maintaining views incrementally. In Proceedings of the 1993 ACM
SIGMOD International Conference on Management of Data,
pages 157–166, 1993.
[20] J. Claußen, A. Kemper, G. Moerkotte, and K. Peithner. Optimizing queries with universal quantification in object-oriented
and object-relational databases. In Proceedings of the 23rd
International Conference on Very Large Data Bases, pages
286–295. Morgan Kaufman, 1997.
[37] D. Hansel, R. Cleaveland, and S. A. Smolka. Distributed
prototyping from validated specifications. Journal of Systems
and Software, 70(3):275–298, 2004.
[21] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein.
Introduction to Algorithms. MIT Press, 3rd edition, 2009.
[38] P. Hunt, M. Konar, F. P. Junqueira, and B. Reed. ZooKeeper:
Wait-free coordination for internet-scale systems. In USENIX
Annual Technical Conference, page 9, 2010.
[22] F. Cruz, R. Rocha, S. C. Goldstein, and F. Pfenning. A linear logic programming language for concurrent programming
over graph structures. Theory and Practice of Logic Programming, 14:493–507, 7 2014.
[39] D. Kaynar, N. Lynch, R. Segala, and F. Vaandrager. The
Theory of Timed I/O Automata. Morgan & Claypool, 2nd
edition, 2010.
[23] J. Dean and S. Ghemawat. MapReduce: Simplified data processing on large clusters. Communications of the ACM, 51(1):
107–113, 2008.
[40] I. H. Krüger. An experiment in compiler design for a concurrent object-based programming language. Master’s thesis,
The University of Texas at Austin, 1996.
[24] G. DeCandia, D. Hastorun, M. Jampani, G. Kakulapati,
A. Lakshman, A. Pilchin, S. Sivasubramanian, P. Vosshall,
and W. Vogels. Dynamo: Amazon’s highly available keyvalue store. ACM SIGOPS Operating Systems Review, 41(6):
205–220, 2007.
[41] A. Kshemkalyani and M. Singhal. Distributed Computing:
Principles, Algorithms, and Systems. Cambridge University
Press, 2008.
[42] A. Lakshman and P. Malik. Cassandra: A decentralized structured storage system. ACM SIGOPS Operating Systems Review, 44(2):35–40, 2010.
[25] DistAlgo. DistAlgo: A Language for Distributed Algorithms. http://github.com/DistAlgo, 2016. Beta release
September 27, 2014. 1.0 release November 13, 2016.
[43] L. Lamport. Time, clocks, and the ordering of events in a
distributed system. Communications of the ACM, 21:558–565,
1978.
[26] Erlang. Erlang Programming Language. http://www.
erlang.org/, 2015. Last released December 18, 2015.
25
2017/3/14
[60] F. Mattern. Virtual time and global states of distributed
systems. In Proceedings of the International Workshop on
Parallel and Distributed Algorithms, pages 120–131. NorthHolland, 1989.
[44] L. Lamport. Specifying Systems: The TLA+ Language and
Tools for Hardware and Software Engineers. AddisonWesley, 2002.
[45] L. Lamport. The PlusCal algorithm language. In Proceedings
of the 6th International Colloquium on Theoretical Aspects of
Computing, pages 36–60. Springer, 2009.
[61] P. Maymounkov and D. Mazières. Kademlia: A peer-to-peer
information system based on the XOR metric. In Peer-to-Peer
Systems, pages 53–65, 2002.
[46] J. Larson. Erlang for concurrent programming. Communications of the ACM, 52(3):48–56, 2009.
[62] MPI. Message Passing Interface Forum. http://www.
mpi-forum.org/, Last released June 4, 2015.
[47] B. Liskov. Distributed programming in Argus. Communications of the ACM, 31(3):300–312, Mar. 1988.
[63] H. Nakamura. Incremental computation of complex object
queries. In Proceedings of the 16th ACM SIGPLAN Conference on Object-Oriented Programming, Systems, Languages,
and Applications, pages 156–165, 2001.
[48] Y. A. Liu. Systematic Program Design: From Clarity To
Efficiency. Cambridge University Press, 2013.
[49] Y. A. Liu and S. D. Stoller. Dynamic programming via static
incrementalization. Higher-Order and Symbolic Computation, 16(1–2):37–62, 2003.
[64] R. Paige. Real-time simulation of a set machine on a RAM.
In Proceedings of the International Conference on Computing and Information, pages 69–73. Canadian Scholars Press,
1989.
[50] Y. A. Liu and S. D. Stoller. From Datalog rules to efficient
programs with time and space guarantees. ACM Transactions
on Programming Languages and Systems, 31(6):1–38, 2009.
[65] R. Paige and S. Koenig. Finite differencing of computable
expressions. ACM Transactions on Programming Languages
and Systems, 4(3):402–454, 1982.
[51] Y. A. Liu, S. D. Stoller, M. Gorbovitski, T. Rothamel, and
Y. E. Liu. Incrementalization across object abstraction. In
Proceedings of the 20th ACM Conference on Object-Oriented
Programming, Systems, Languages, and Applications, pages
473–486, 2005.
[66] V. Petukhin. Programs with universally quantified embedded
implications. In Proceedings of the 4th International Conference on Logic Programming and Nonmonotonic Reasoning,
pages 310–324. Springer, 1997.
[52] Y. A. Liu, S. D. Stoller, N. Li, and T. Rothamel. Optimizing
aggregate array computations in loops. ACM Transactions on
Programming Languages and Systems, 27(1):91–125, 2005.
[67] PRL Project.
EventML.
http://www.nuprl.org/
software/\#WhatisEventML, 2013.
Lastest release
September 21, 2012. Accessed January 14, 2017.
[53] Y. A. Liu, C. Wang, M. Gorbovitski, T. Rothamel, Y. Cheng,
Y. Zhao, and J. Zhang. Core role-based access control: Efficient implementations by transformations. In Proceedings of
the ACM SIGPLAN 2006 Workshop on Partial Evaluation and
Program Manipulation, pages 112–120, 2006.
[68] W. Pugh and T. Teitelbaum. Incremental computation via
function caching. In Conference Record of the 16th Annual
ACM Symposium on Principles of Programming Languages,
pages 315–328, 1989.
[69] G. Ramalingam and T. Reps. A categorized bibliography
on incremental computation. In Conference Record of the
20th Annual ACM Symposium on Principles of Programming
Languages, pages 502–510, 1993.
[54] Y. A. Liu, M. Gorbovitski, and S. D. Stoller. A language and
framework for invariant-driven transformations. In Proceedings of the 8th International Conference on Generative Programming and Component Engineering, pages 55–64. ACM
Press, 2009.
[70] M. Raynal. Distributed Algorithms and Protocols. Wiley,
1988.
[55] Y. A. Liu, S. D. Stoller, and B. Lin. High-level executable
specifications of distributed algorithms. In Proceedings of
the 14th International Symposium on Stabilization, Safety,
and Security of Distributed Systems, pages 95–110. Springer,
2012.
[71] M. Raynal. Communication and Agreement Abstractions for
Fault-Tolerant Asynchronous Distributed Systems. Morgan &
Claypool, 2010.
[72] M. Raynal. Distributed Algorithms for Message-Passing Systems. Springer, 2013.
[56] Y. A. Liu, S. D. Stoller, B. Lin, and M. Gorbovitski. From
clarity to efficiency for distributed algorithms. In Proceedings
of the 27th ACM SIGPLAN Conference on Object-Oriented
Programming, Systems, Languages and Applications, pages
395–410, 2012.
[73] T. Rothamel and Y. A. Liu. Generating incremental implementations of object-set queries. In Proceedings of the 7th
International Conference on Generative Programming and
Component Engineering, pages 55–66. ACM Press, 2008.
[57] Y. A. Liu, J. Brandvein, S. D. Stoller, and B. Lin. Demanddriven incremental object queries. In Proceedings of the
18th International Symposium on Principles and Practice of
Declarative Programming, pages 228–241. ACM Press, 2016.
[74] A. Rowstron and P. Druschel. Pastry: Scalable, decentralized
object location, and routing for large-scale peer-to-peer systems. In Proceedings of the IFIP/ACM International Conference on Distributed Systems Platforms, Middleware 2001,
pages 329–350. Springer, 2001.
[58] N. P. Lopes, J. A. Navarro, A. Rybalchenko, and A. Singh.
Applying Prolog to develop distributed systems. Theory and
Practice of Logic Programming, 10(4-6):691–707, July 2010.
[75] D. Saha and C. R. Ramakrishnan. Incremental evaluation
of tabled logic programs. In Proceedings of the 19th International Conference on Logic Programming, pages 392–406.
Springer, 2003.
[59] N. A. Lynch. Distributed Algorithms. Morgan Kaufman,
1996.
26
2017/3/14
[76] M. L. Scott. The Lynx distributed programming language:
Motivation, design, and experience. Computer Languages, 16
(3):209–233, 1991.
[77] T. F. Serbanuta, G. Rosu, and J. Meseguer. A rewriting
logic approach to operational semantics. Information and
Computation, 207:305–340, 2009.
[78] K. Shvachko, H. Kuang, S. Radia, and R. Chansler. The
Hadoop distributed file system. In Proceedings of the 2010
IEEE 26th Symposium on Mass Storage Systems and Technologies, pages 1–10. IEEE CS Press, 2010.
[79] I. Stoica, R. Morris, D. Liben-Nowell, D. Karger,
M. Kaashoek, F. Dabek, and H. Balakrishnan. Chord:
A scalable peer-to-peer lookup protocol for internet applications. IEEE/ACM Transactions on Networking, 11(1):17–32,
2003.
[80] T. Swift, D. S. Warren, et al. The XSB System Version 3.7,x,
2016. http://xsb.sourceforge.net. Latest release July
6, 2016.
[81] G. Tel. Introduction to Distributed Algorithms. Cambridge
University Press, 2nd edition, 2000.
[82] R. van Renesse and D. Altinbuken. Paxos made moderately
complex. ACM Computing Surveys, 47(3):42:1–42:36, Feb.
2015.
[83] R. van Renesse and F. B. Schneider. Chain replication for
supporting high throughput and availability. In Proceedings
of the 6th USENIX Symposium on Operating Systems Design
and Implementation, pages 91–104. USENIX Association,
2004.
[84] D. E. Willard. Efficient processing of relational calculus expressions using range query theory. In Proceedings of the
1984 ACM SIGMOD International Conference on Management of Data, pages 164–175, 1984.
[85] D. E. Willard. An algorithm for handling many relational
calculus queries efficiently. Journal of Computer and System
Sciences, 65:295–331, 2002.
[86] A. K. Wright and M. Felleisen. A syntactic approach to type
soundness. Information and Computation, 115:38–94, 1994.
[87] B. Y. Zhao, L. Huang, J. Stribling, S. C. Rhea, A. D. Joseph,
and J. D. Kubiatowicz. Tapestry: A resilient global-scale
overlay for service deployment. IEEE Journal on Selected
Areas in Communications, 22(1):41–53, 2004.
27
2017/3/14
| 6cs.PL
|
Ternary Neural Networks with Fine-Grained
Quantization
arXiv:1705.01462v3 [cs.LG] 30 May 2017
Naveen Mellempudi1 , Abhisek Kundu1 , Dheevatsa Mudigere1 , Dipankar Das1 , Bharat Kaul1 , and
Pradeep Dubey2
1
Parallel Computing Lab, Intel Labs, Bangalore, India
Parallel Computing Lab, Intel Labs, Santa Clara, CA
2
Abstract
We propose a novel fine-grained quantization (FGQ) method to ternarize pretrained full precision models, while also constraining activations to 8 and 4-bits.
Using this method, we demonstrate minimal loss in classification accuracy on
state-of-the-art topologies without additional training. We provide an improved
theoretical formulation that forms the basis for a higher quality solution using
FGQ. Our method involves ternarizing the original weight tensor in groups of N
weights. Using N = 4, we achieve Top-1 accuracy within 3.7% and 4.2% of the
baseline full precision result for Resnet-101 and Resnet-50 respectively, while
eliminating 75% of all multiplications. These results enable a full 8/4-bit inference
pipeline, with best reported accuracy using ternary weights on ImageNet dataset,
with a potential of 9× improvement in performance. Also, for smaller networks
like AlexNet, FGQ achieves state-of-the-art results. We further study the impact of
group size on both performance and accuracy. With a group size of N = 64, we
eliminate ≈ 99% of the multiplications; however, this introduces a noticeable drop
in accuracy, which necessitates fine tuning the parameters at lower precision. We
address this by fine-tuning Resnet-50 with 8-bit activations and ternary weights at
N = 64, improving the Top-1 accuracy to within 4% of the full precision result
with < 30% additional training overhead. Our final quantized model can run on a
full 8-bit compute pipeline using 2-bit weights and has the potential of up to 15×
improvement in performance compared to baseline full-precision models.
1
Introduction
Today’s deep learning models achieve state-of-the-art results on a wide variety of tasks including
Computer Vision, Natural Language Processing, Automatic Speech Recognition and Reinforcement
Learning [1]. Mathematically, this involves solving a non-convex optimization problem with order
of millions or more parameters. Solving this optimization problem - also referred to as training the
neural network - is a compute-intensive process that, for current state-of-the-art networks, requires
days to weeks. Once trained, the network evaluates a function on specific input data - referred to as
inference. While the compute intensity for inference is much lower than that of training, owing to
the fact that inference is done on a large number of input data, the total computing resources spent
on inference is likely to dwarf those spent on training. The large and somewhat unique compute
requirements for both deep learning training and inference operations motivate the use of customized
low precision arithmetic [9, 2, 8, 24, 15, 13] and specialized hardware to run these computations
as efficiently as possible [5, 25, 21, 19, 11]. Most of these cases requires partial or full training of
network in low precision. Training at low-precision allows for the network to implicitly learn the low
precision representation (along with the inherent noise); however, it introduces significant resource
overheads which can be prohibitive for many resource-constrained applications, specifically those
involving edge devices.
Reducing precision for both weights and activations has significant power-performance implication
on system design. Low-precision not only allows increasing compute density, but also reduce
pressure on the memory sub-system. Most of the current solutions are focused on compressing the
model [14, 16], going as low as binary weights, which allows storing the model on the limited on-chip
local memory. However, activations (input) need to be fetched from external memory or I/O-device
(camera). Fetching data contributes to majority of the system power consumption. Hence reducing the
size of activations is essential for more efficient utilization of the available computational resources.
There have been a few solutions [8, 9] using lower precision representation for activations, however
they necessitate specialized hardware for efficient implementation. Further, with widespread adoption
of deep learning across various applications, such as autonomous driving, augmented reality etc, there
is an increased demand for inference tasks to be done on edge devices efficiently. To address both
the aforementioned system and application requirements, there is a general trend to move towards a
full lower precision inference pipeline [11]. This is evident from the advent of 8-bit and sub 8-bit
hardware such as Google’s TPU [11] and other main stream GPU1 , CPU offerings. Further, there is
also software support for 8-bit inference through popular frameworks such as TensorFlow, Theano
and compute libraries like NVidia’s TensorRT1 .
In this paper, we focus on enabling a sub 8-bit inference pipeline by using ternary weights and 8/4-bit
activations, with minimal or no re-training, and yet achieving near state-of-art accuracy. The rationale
behind our approach is to carefully convert the full-precision weights to low-precision, such that the
element-wise distance between full-precision and low-precision weights are small. Consequently,
the low-precision weights remain in the neighborhood of pre-trained full-precision weights in the
search space of network parameters, and we expect them to generalize in a similar manner, despite
no re-training.
We summarize our contributions below:
1. Based on an improved theoretical formulation, propose a novel fine-grained quantization
(FGQ) method to convert pre-trained models to a ternary representation with minimal loss
in test accuracy, without re-training.
2. With ternary weights, achieve classification accuracy (Top-1) of 73.85% with 8-bit activations (2w-8a) and 70.69% with 4-bit activations (2w-4a), on the ImageNet dataset[3] using
a pre-trained Resnet-101 model (no re-training). To the best of our knowledge, these are the
highest reported accuracies in this category on ImageNet dataset[17].
3. Demonstrate the general applicability of FGQ, with state-of-art results (2w-8a, 2w-4a) on
smaller models such as Resnet-50 and Alexnet[12]. And also show the efficacy of using
FGQ for (re)training at low precision.
4. Study the performance-accuracy trade-off using different group sizes. For a group of N
filters, we reduce the number of multiplications to one in every N additions, thus significantly
reducing computation complexity with a potential for up to 16× improvement baseline full
precision models.
The rest of the paper is organized as follows, Section2 discusses related work on ternary weights, low
precision inference and contrast them with FGQ. Section3 describes the FGQ formulation and the
theoretical basis for this method. This is followed by Section4, which includes experimental results
and related discussion. Finally, in Section5 we conclude with summarizing the implications of of
FGQ, our results and the future research directions.
2
Related Work
Deep learning inference using low-precision weights and activations is a well-researched topic. Many
researchers have experimented with custom data representations to perform deep learning tasks
and have shown significant benefits over the general purpose floating point representation. [20]
have show that 8-bit dynamically scaled fixed point representation [22] can be used to speed up
convolution neural networks using general purpose CPU hardware by carefully choosing right data
layout, compute batching and using implementation optimized for target hardware. With this they
show up to 4× improvement over an aggressively tuned floating point implementation. [5] have
1
https://devblogs.nvidia.com/parallelforall/new-pascal-gpus-accelerate-inference-in-the-data-center/
2
done a comprehensive study on the effect of low precision fixed point computation for deep learning
and have successfully trained smaller networks using 16-bit fixed point on specialized hardware.
This suggests that fixed point representations are better suited for low(er) precision deep learning.
There have also been recent efforts exploring 8-bit floating point representation [4], however such
schemes have the additional overhead of reduced precision since the exponent is replicated for each
value. Whereas with fixed point representation, using a single shared exponent improves capacity for
precision. Typically for deep neural networks, with reduced bit-widths it is desired to preserve the
numerical precision, since the loss in range can be augmented by the dynamic scaling of the shared
exponent.
Commonly, low precision networks are designed to be trained from scratch, leveraging the inherent
ability of network to learn the approximations introduced by low precision computations [14, 16, 9, 2,
8, 25, 6]. This can be prohibitive in applications which rely on using previously trained models. Such
use cases are typical in many edge device deployments. To address such cases, FGQ is developed
with motivation to be able to achieve state-of-art accuracies without any training and hence enabling
direct use of pre-trained models. This requirement results in the quantization scheme being quite
complex but making it more widely applicable and making it also easily usable for the former case with training from scratch.
Many of the recent reduced precision work, look at the low precision only for the weights while
retaining the activations in full precision [23, 21, 6, 15, 16, 14, 4]. Using low precision also for
activations is essential to realize the full power-performance benefits with using 2-bit (ternary) weights.
The hardware needs to operate at the throughput that is close to the precision of the weights (i.e. 16×
better throughput compared to using 32-bit weights). This cannot be achieved (or would be very hard
to achieve) when the activations are at full precision because streaming 32-bit activations from main
memory at that rate requires much higher(16×) bandwidth and compute engine needs to be much
wider to deliver the desired throughput. All of which will increase the area and power budget which
is not desirable when designing for low-power edge devices. Hence, reducing the size of activations
is essential for reducing the compute requirements at the edge. Using 8-bits and below for activations
dramatically reduces the design requirements on the edge and opens up the possibility of achieving
16× throughput improvement. [14, 16] propose low precision networks with binary weights, while
retaining the activations in full precision. [14] use a stochastic binarization scheme, achieving state-ofart (SOTA) accuracies on smaller data-sets (MNIST, CIFAR10, SVHN). [16] demonstrate near-SOTA
accuracies on the large ImageNet data-set using AlexNet topology. Further, they also demonstrate a
variant with binary weights and activations, with all computations are simplified bit-count operations
but with significant loss in accuracy. Lower precision for activations have also been used, [8] use
1-bit for both weights and activations for smaller networks. For larger Imagenet-class networks [9],
use 2-bit activations and binary weights showing reasonable accuracies. However, both these [8, 9]
use specialized data representation requiring custom hardware for efficient implementation. Other
solutions such as [24], employ a more tailored approach with different precision for each - weights
(1-bit), activations (2-bits) and gradients (6-bits); implemented with special-purpose hardware.
[13] introduces a theoretical formulation for ternary weight network using a threshold based approach
(symmetric threshold ±∆) with one scaling factor for each layer. They provide an approximation
to the optimal ternary representation assuming weights follow a Gaussian distribution. However,
one scaling factor per layer may not approximate the full-precision network well as the model
capacity is very limited in this case. To increase model capacity [25] modify this solution to use two
symmetric thresholds (±∆) and two scaling factors (separately for positive and negative weights).
However, despite improving the accuracy this approach typically makes the inferencing inefficient
by requiring multiple passes over the positive and negative values, hence increasing the bandwidth
requirements. [23] have proposed a post-facto incremental quantization approach, which aims to find
the optimal representation using an iterative method, constraining weights to either 0 or powers of 2,
using a 5-bit representation. and re-training activations with full precision. All the aforementioned
implementation require partial or full training of the network in low precision. Alternatively, [15]
used log quantization method on pre-trained models and achieved good accuracy by tuning the bit
length for each layer without re-training.
Achieving near-SOTA accuracy on the Imagenet dataset with deeper networks [7], without any
training in low precision (for both weights and activations) is still a challenge. Our work is an attempt
to address this problem and improve over existing approaches.
3
3
Ternary Conversion of Trained Network
Our goal is to convert the full-precision trained weights W to ternary values {−α, 0, +α}, α ≥ 0,
without re-training. We use a threshold (∆ > 0) based approach similar to [13]: i-th element Ŵi =
sign(Wi ), if |Wi | > ∆, and 0 otherwise. Then, the element-wise error is E(α, ∆) = kW −αŴk2F
and an optimal ternary representation α∗ Ŵ∗ ≈ W is as follows:
α ∗ , ∆∗
=
argmin E(α, ∆),
α ≥ 0, Ŵi ∈ {−1, 0, +1}, i = 1, 2, ..., n
s.t.
(1)
α≥0,∆>0
where n is the size of W (W ∈ Rn ). We hypothesize that weights that learn different types of
features may follow different distributions. Combining all the weights together represents a mixture
of various distributions, and ternarizing them using a single threshold (∆) and magnitude (α) may
not preserve the distributions of individual weights. Consequently, many weights are approximated
poorly (if not totally pruned out) leading to loss of valuable information that they learn. We may not
be able to compensate for this loss of information as we do not train the network in low precision.
This motivates us to use a fine-grained quantization technique involving multiple scaling factors in
order to increase model capacity that can lead to better preservation of distributions learned by filters.
Moreover, we hypothesize that positive and negative weight distributions are not always symmetric
around the mean, further refinement of this solution maybe possible using two separate thresholds,
∆p and ∆n > 0, for positive and negative weights, respectively, along with a scaling factor α to
ternarize the weights.
3.1
Our Formulation
Computing separate ∆ and α for each weight compensates for information loss and better preserves
the underlying distributions. However, such solution, while showing significant improvement in
accuracy, does not reduce the number of multiplications leading to a less efficient implementation.
Therefore, we seek to find a trade-off between achieving higher accuracy and reducing the total number
of multiplications. We propose a fine-grained quantization approach that creates groups of weights,
and ternarizes each group independently. Let us consider the weights represented as a vector W ∈ Rn .
We partition the set I of n indices into k disjoint subsets, c1 , c2 , ..., ck , with cardinality |ci | = ni , such
P
that, ci ∩ cj = ∅, ∪i ci = I, i ni = n. We can decompose W into k orthogonal vectors W(i) ∈ Rn ,
P
(i)
i = 1, ..., k, where j-th component Wj = Wj if j ∈ ci , otherwise 0. Clearly, i W(i) = W;
then we ternarize each orthogonal component W(i) as αi Ŵ(i) , where components of Ŵ(i) are in
{−1, 0, +1}. Threshold-based pruning never turns 0 to non-zero, and the following orthogonality
holds. W(i) ⊥ W(j) , Ŵ(i) ⊥ Ŵ(j) , W(i) ⊥ Ŵ(j) , for i 6= j. It follows that, (W(i) − αi Ŵ(i) ) ⊥
P
P
(W(j) − αj Ŵ(j) ), for i 6= j, and we have kW − i αi Ŵ(i) k2F = i kW(i) − αi Ŵ(i) k2F . Then,
(i)
for a given group of k filters {W(i) }, i = 1, ..., k, and Ŵj ∈ {−1, 0, +1}, ∀j,
X
X
α1∗ , .., αk∗ , Ŵ(1)∗ , .., Ŵ(k)∗ = argminkW −
αi Ŵ(i) k2F , =
argminkW(i) − αi Ŵ(i) k2F (2)
αi ,Ŵ(i)
i
i
αi ,Ŵ(i)
Therefore, we need to solve k independent sub-problems. This formulation allows a better ternary
approximation to the original full-precision weights, ensuring that they remain within a neighborhood
of the original solution in the complex search space of parameters, despite no re-training. Consequently, we expect the full-precision solution and the ternary counterpart to generalize in a similar
manner. From model capacity point of view, we can have only three distinct values, {−α, 0, +α},
for a ternary weight vector without such grouping. With k such groups, however, we can represent
2k + 1 distinct values, thus increasing model capacity linearly with number of groups.
We can solve each sub-problem using a threshold-based approach the following way. We are given a
vector of elements W ∈ Rn , and we can use separate thresholds for positive and negative weights,
+
−
along with one scaling factor α to ternarize W. Let I∆
= {i : Wi > ∆p }, I∆
= {i : Wi < −∆n },
for some ∆p , ∆n > 0. We want to solve
α∗ , ∆∗p , ∆∗n =
argmin
α≥0,∆p >0,∆n >0
kW − αŴk2F ,
4
s.t.
Ŵi ∈ {−1, 0, +1}, i = 1, ..., n.
(3)
We have the following analytical solution.
P
P
( i∈I + |Wi | + i∈I − |Wi |)2
∗
∗
∆
∆
∆p , ∆n = argmax
,
+
−
|I∆
| + |I∆
|
∆p >0,∆n >0
P
∗
α =
+
i∈I∆
|Wi | +
+
|I∆
|
+
P
−
i∈I∆
−
|I∆ |
|Wi |
+
−
Note that for ∆p = ∆n = ∆, I∆ = I∆
∪ I∆
, and (4) reproduces the formulation in ([13]).
X
X
α∗ = (
|Wi |)/|I∆ |, ∆∗ = argmax(
|Wi |)2 /|I∆ |
∆>0
i∈I∆
(4)
(5)
i∈I∆
The advantage of our formulation (2) is that the smaller independent sub-problems can be solved
efficiently using brute-force methods to achieve better approximation. However, we also explore
analytical solutions to establish the theoretical veracity of our approach. Assuming that the magnitude
of the learned weights follow exponential distribution with parameter λ, we analytically derive the
optimal ∆∗ from the following lemma.
P
Lemma 1. Using above notations, if |Wi | ∼ exp(λ), then, ∆∗ in (5) is ∆∗ ≈ 1/λ = i |Wi |/n
1
5
9
13
17
21
25
29
33
37
41
45
49
53
57
61
65
69
73
77
81
85
89
93
97
101
105
Relative Error Improvement (%)
From this analysis, we see the need for a
higher threshold value to prune larger number
Improvement of Error over Gaussian Assumption
14
of smaller elements. This is intuitive from the
shape of the model distributions, which are typ12
ically heavy-tailed distributions. In reality, how10
ever, it may not be appropriate to use a single
8
distribution to model the weights of all the layers
6
of a neural network. We can apply Kolmogorov4
Smirnov (K-S) test as a goodness-of-fit measure
2
to identify an appropriate reference distribution
0
(here we choose between Gaussian and expo-2
nential), and find ∆∗ accordingly. We approxLayers
imate a heavy-tailed distribution by an exponential one by pruning out some of the smaller Figure 1: Layer-wise improvement of theoretical
elements. This gives us an exponential approx- ternary error (one α per layer) over Gaussian asimation with smaller λ. Further, we can use sumption by choosing an appropriate distribution
maximum likelihood functions to estimate the using K-S test for ResNet-101 on Imagenet dataset.
parameters of such distributions.
pPn For Gaussian
2
N (0, σ), estimated σ̂ =
i=1 |Wi | /n = rms(W), and for exponential case, estimated paramePn
ter λ̂ = i=1 |Wi |/n. Based on such refined analysis, we observe significant improvement in the
theoretical ternary error over Gaussian assumption of [13] (Figure 1). It is interesting to observe
that for earlier convolution layers of ResNet-101 trained on ImageNet, the magnitude of the weights
follow exponential distribution, and later layer weights are Gaussian.
3.2
Weight Grouping
Our method (2) is agnostic to how the (full-precision) weights are grouped , but leverages that
consequence of grouping - which allows for solving these as independent sub-problems more
efficiently. The specifics of the grouping mechanism and memory layout used for accessing these
groups of weights is an independent problem to explore. The primary objective of grouping is to
minimize the dynamic range within each group and split the weights in such a way that the smaller
groups have a uniform distribution. This helps in reducing the complexity of finding an optimal
solution (α) for each independent sub-problem using either analytical or brute-force techniques.
However, to realize the full performance potential of ternarization, it is essential to ensure that the
grouping mechanism itself does not introduce a significant overhead. Similarity based clustering
algorithms such as K-means, despite being better at finding optimal grouping of weights that may
even lead to better accuracy, are not friendly for efficient implementations (in both software hardware),
because of the random grouping of elements from non-contiguous memory locations. This leads
to irregular memory accesses with longer latencies, to gather arbitrarily grouped weights that use a
common α, for a partial accumulation of the output.
5
Based on our empirical observations, we conclude that using static groups of weights that
are partitioned along input channels C achieves
best accuracy. The same element from multiple filters along C have significantly less variance, since they correspond to similar input features. Hence grouping such elements results in
∝
reduced dynamic range within the group. Such
Figure 2: Static Grouping: (R × S) sub-groups of a grouping also easily lends itself for efficient
n elements from contiguous filters along C dimen- implementation using both existing hardware
sion, αk consists of scaling factors.
and in software with using [K][C/N ][R ∗ S][N ]
layout for the weight tensor, where groups of N
elements are accessed from contiguous memory locations. Since the elements along C accumulate
to the same output feature, this layout is also amenable to efficient vectorization along K. Figure 2
shows an example of this grouping scheme applied to 3 × 3 filters. Each group of N 3 × 3 ternary
filters, has 3 × 3 scaling factors (α) corresponding to each element of the filter.
=
4
=
,
,
,…
,∝ = ∝
= ∝ , … , = 1. . , = 1 …
×
Experimental Results
For experimental results, we focused on Resnet50 and Resnet-101[7] using ILSVRC-2012[3]
dataset, to demonstrate the efficacy of our
method on large, sophisticated models using
2-bit weights and 8-bit activations (2w-8a). We
extended our study by applying FGQ on activations to help further reduce the precision of activations to 4-bits (2w-4a) and show results comparable with 8-bit activations for all the tested
networks. Further, towards establishing the
broader applicability of FGQ we demonstrate
state-of-the-art accuracy also for Alexnet[12].
Convolution Data Flow
FP32 weight Tensor
Convert to Ternary ( α ×
DFP-8 Tensor
FP32 activation Tensor
)
Convert to DFP-8
Emulation Library
Convert DFP-8 to FP32
FP32 weight tensor emulating
range & precision of DFP-8
DFP-8 Tensor
Convert DFP-8 to FP32
FP32 activation tensor emulating
range & precision of DFP-8
FP32-convolutions-FP32 output
Our setup consists of a modified version of
Intel Confidential
Caffe[10] that emulates low-precision dynamic Figure 3: Schematic describing our low preci2
fixed point (DFP ) computations described in sion experimental setup in Caffe, to emulate fineFig. 3. We use 32-bit accumulator for all our low grained quantization (FGQ) with ternary weights
precision computations to minimize the chances and 8-bit activations
of an overflow. We split the pre-trained weights
into groups of N elements using the mechanism
described in section3.2, and use brute-force technique to compute the floating point values of the
threshold (∆) and scaling-factor (α) for each group. The scaling factors are then quantized to a
8/4-bit fixed point and the weights are stored in the memory format described in3.2. The activations
are quantized to 8/4-bits before performing convolution operation and the 32-bit outputs are down
converted to 8-bit and appropriately rounded before they are passed to the next layer.
Our experiments indicate that it is essential to use higher precision of the first layer (8w-8a), to
minimize the accumulation of quantization loss. We also observe that using pre-trained parameters
in batch normalization layers leads to a loss of accuracy due to shift in variance introduced by
the quantization. We prevent this loss by recomputing batch normalization parameters during the
inference phase to compensate for the shift in variance.
We explored the accuracy-performance trade-off using different group sizes of N , our experiments
show that FGQ with a group size of N=4 (FGQ-N4) achieves highest accuracy with no re-training
and a potential 9× performance benefit. FGQ-N4 applied to a pre-trained Resnet-101 model with
2w-8a achieves Top-1 accuracy of 73.9%, which is within 4% of the full-precision results. With
activations reduced to 4-bits (2w-4a), the Top-1 accuracy drops only marginally to 70.7%. FGQ-N4
performs equally well on Resnet-50, achieving with Top-1 accuracy of 70.8% with 2w-8a which is
4.2% off from full-precision result, and 68.4% with 2w-4a. To the best of our knowledge, these are
2
Please not that fixed point and dynamic fixed point are used interchangeably
6
Table 1: Classification accuracy (Top-1) on ImageNet dataset achieved using FGQ with N=4 without
any re-training. Results on Resnet-101, Resnet-50 and Alexnet using 2w-8a and 2w-4a compared
with best published results
Networks
Our Baseline
FGQ-N4
2w-8a
FGQ-N4
2w-4a
no low precision training
Resnet-101
Resnet-50
AlexNet
77.50%
75.05%
56.83%
73.85%
70.76%
49.04%
70.69%
68.38%
49.00%
INQ
5w-32a [23]
dLAC
2w-32a [21]
DoReFa
1w-4a-32g [24]
with low precision re-training
74.81%3
-
73.85%3
-
50.3%3
the highest reported accuracies using 2w-8a and 2w-4a on Imagenet dataset[3] using state-of-the-art
networks.
To understand the general applicability of our method to a wider range of networks, we apply FGQ
to the smaller Alexnet[12] model. FGQ-N4 applied to a pre-trained Alexnet model, achieves 49%
Top-1 accuracy with 2w-8a without any re-training, this is 8% away from the baseline full-precision
result( 56.8%). With 2w-4a we do not see any further reduction in accuracy. There are no previously
published results that can be directly compared to FGQ, which perform quantizaion on pre-trained
models and work with end-to-end low-precision. Hence, we compare with [21, 23, 24], which are the
closest in terms of the networks used and/or the target precision. Our Alexnet result using FGQ-N
is comparable to previously published result[24] which is 6% away from the baseline using 1w-4a
while also employing training in low-precision with full precision gradients. Table 1 has a comparison
with previous reported results from[23] using 5-bit weights and [21] using ternary weights. While
they report slightly better absolute numbers, our numbers are relatively better because both these
results use full-precision activations and train the network in low precision to achieve those numbers.
While without any low precision training and reduced precision for activation, results with FGQ is
still competitive with other similar (aforementioned) results.3 With additional low precision training
with FGQ we are able significantly improve accuracy and get closer to state-of-art full precision
results, as outlined in the next section along with associated performance implications.
4.1
Discussion
In order to realize the full performance potential of ternary networks, the inference platform needs to
operate at the throughput that is close to the precision of the weights. This would increase the amount
of memory bandwidth required to stream activations by 16× and a compute engine that is much wider
to deliver the desired compute throughput. Building such solution around full-precision activations
would be prohibitive in terms of areas and power requirements, whereas it is more amenable to build
such solution when the activations are 8 or 4-bits.
Figure4a shows the performance Vs accuracy trade-off for Resnet-50[18] for a FGQ based 8-bit
inference design. Our model projects the lower bound of the performance potential based on the
percentage of FMA operations that can be converted into ternary accumulations at each group size
N . In the ideal case, where N is equal to the total number of weights in the layer, the best case
performance potential is 16× compared to the baseline full-precision performance. For a group size
of N=4, 75% of all FMA operations can be performed in ternary, Using slightly larger sub-groups
of N=8 we can replace 87.5% of FMA operations with ternary while losing an additional 3.7%
Top-1 accuracy. At group size N = 64, ≈ 99% of all FMA operations can be replaced by ternary
accumulations, resulting in 15× potential improvement in performance. But the performance comes
at a cost of significant drop in accuracy. Using larger groups of weights results in a poor ternary
approximation to the full-precision model. Consequently, the ternary solution moves away from the
full-precision local optima and display different generalization behavior.
3
It should be noted that both these works use Resnet-50 with slight variations and hence have slightly
different baseline accuracies. For [23] the baseline full precision a Top-1 accuracy is 73.22%, for [21] it is
76% and for [24] it is 55.9%
7
Top-1 Accuracy
80%
Baseline (FP32)
75%
70.51%
68.59%
70.33%
71.10%
65.08%
2.0
N=2
N=8
60%
N=16
55%
N=32
50%
N=64
45%
0x
2x
4x
6x
8x
10x
12x
Speedup (relative to Baseline FP32)
14x
1.5
1.0
0.5
0.0
16x
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
Training Error
N=4
65%
0
70%
Top-1 Accuracy
2.5
After
fine-tuning
N=64*
Epochs
(a)
(b)
Figure 4: a). Performance vs accuracy trade-off for Resnet-50 for a FGQ based 8-bit inference design,
b). Fine-tuning Resnet-50, with pre-initialized weights on Imagenet dataset.
We have trained low-precision (2w-8a) ResNet-50[18] at group size N=64 on ImageNet[3] dataset
to recover the accuracy lost because of ternarization. We initialized the network with a pre-trained
full-precision model and fine-tuned the parameters of the low-precision network. We reduced the
learning rate to an order of 1e-4 to avoid exploding gradients, retaining all the other hyper parameters
from full-precision training and performed gradient updates in full precision. After training for
20-epochs, we recover most of the lost accuracy and achieved 71.1% Top-1 and 90.01% Top-5
bringing it to within 4% of the full-precision baseline accuracy. Figure4b shows the reduction of
training error and improvements in validation accuracy.
5
Conclusion
We propose a fine-grained ternarization method which exploits local correlations in dynamic range
of the parameters to minimize the impact of quantization on overall accuracy. We demonstrate near
SOTA accuracy on Imagenet data-set using pre-trained models with quantized networks without
re-training. Using ternary weights on Resnet-101 and Resnet-50 with 8-bit activations our results
are within ≈ 4% from the full precision (F P 32) accuracy. Using 4-bit activations we see a further
drop of ≈ 3% in accuracy. To the best of our knowledge, these are the highest reported accuracies
using ternary weights and low-precision activations.
Our weight grouping based approach allows us to obtain solutions that can be tailored for specific
hardware, as well as, can be used on general purpose hardware, based on the accuracy and performance
requirements. Smaller group sizes with N=4 achieve best accuracy, and use 75% of the computations
ternary operations (simple 8-bit additions) and this is better suited for implementation on specialized
hardware. Larger group sizes are more suitable for current general purpose hardware, with a larger
portion of computations as low precision operations (≈ 99% for N=64), although this comes with
the cost of reduced accuracy. This gap may be bridged with additional low precision training as
shown in Section4. Our final quantized model can be efficiently run on full 8-bit compute pipeline,
thus offering a potential 16X performance-power benefit.
We continue to actively work on closing the current accuracy gap, exploring both low precision
(re)training and extensions to the FGQ method itself. Also we are looking into a more theoretical
exploration to better understand the formal relationship between the weight grouping and final
accuracy, with an attempt to establish realistic bounds for given network-performance-accuracy
requirements.
8
References
[1] Yoshua Bengio, Ian Goodfellow, and Aaron Courville. Deep learning. Book in preparation for
MIT Press, 2016.
[2] Matthieu Courbariaux, Itay Hubara, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized
neural networks: Training deep neural networks with weights and activations constrained to+ 1
or-1. arXiv preprint arXiv:1602.02830, 2016.
[3] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale
hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009.
IEEE Conference on, pages 248–255. IEEE, 2009.
[4] Tim Dettmers. 8-bit approximations for parallelism in deep learning. arXiv preprint
arXiv:1511.04561, 2015.
[5] Suyog Gupta, Ankur Agrawal, Kailash Gopalakrishnan, and Pritish Narayanan. Deep learning
with limited numerical precision. In ICML, pages 1737–1746, 2015.
[6] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections
for efficient neural network. In Advances in Neural Information Processing Systems, pages
1135–1143, 2015.
[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 770–778, 2016.
[8] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized
neural networks. In Advances in Neural Information Processing Systems, pages 4107–4115,
2016.
[9] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Quantized neural networks: Training neural networks with low precision weights and activations.
arXiv preprint arXiv:1609.07061, 2016.
[10] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick,
Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature
embedding. arXiv preprint arXiv:1408.5093, 2014.
[11] N Jouppi. Google supercharges machine learning tasks with tpu custom chip. Google Blog,
May, 18, 2016.
[12] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep
convolutional neural networks. In Advances in neural information processing systems, pages
1097–1105, 2012.
[13] Fengfu Li, Bo Zhang, and Bin Liu. Ternary weight networks. arXiv preprint arXiv:1605.04711,
2016.
[14] Zhouhan Lin, Matthieu Courbariaux, Roland Memisevic, and Yoshua Bengio. Neural networks
with few multiplications. arXiv preprint arXiv:1609.07061, 2016.
[15] Daisuke Miyashita, Edward H Lee, and Boris Murmann. Convolutional neural networks using
logarithmic data representation. arXiv preprint arXiv:1603.01025, 2016.
[16] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet
classification using binary convolutional neural networks. In ECCV, 2016.
[17] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng
Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual
recognition challenge. International Journal of Computer Vision, 115(3):211–252, 2015.
[18] Marcel Simon, Erik Rodner, and Joachim Denzler. Imagenet pre-trained models with batch
normalization. arXiv preprint arXiv:1612.01452v2, 2016.
[19] Yaman Umuroglu, Nicholas J Fraser, Giulio Gambardella, Michaela Blott, Philip Leong,
Magnus Jahre, and Kees Vissers. Finn: A framework for fast, scalable binarized neural network
inference. arXiv preprint arXiv:1612.07119, 2016.
[20] Vincent Vanhoucke, Andrew Senior, and Mark Z Mao. Improving the speed of neural networks
on cpus. In Proc. Deep Learning and Unsupervised Feature Learning NIPS Workshop, volume 1,
page 4. Citeseer, 2011.
9
[21] Ganesh Venkatesh, Eriko Nurvitadhi, and Debbie Marr. Accelerating deep convolutional
networks using low-precision and sparsity. arXiv preprint arXiv:1610.00324, 2016.
[22] Darrell Williamson. Dynamically scaled fixed point arithmetic. In Communications, Computers
and Signal Processing, 1991., IEEE Pacific Rim Conference on, pages 315–318. IEEE, 1991.
[23] Aojun Zhou, Anbang Yao, Yiwen Guo, Lin Xu, and Yurong Chen. Incremental network quantization: Towards lossless cnns with low-precision weights. poster at International Conference
on Learning Representations, 2017.
[24] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. Dorefa-net:
Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint
arXiv:1606.06160, 2016.
[25] Chenzhuo Zhu, Song Han, Huizi Mao, and William J Dally. Trained ternary quantization. arXiv
preprint arXiv:1612.01064, 2016.
10
Appendix
5.1
Proof of Lemma 1
Let n denote the number of elements. Let f (x) = λe−λx be the pdf of exponential distribution
with parameter λ > 0, and F (x) = 1 − e−λx be the cdf. Then,
Z
|I∆ | ≈ n
f (x)dx = n(1 − F (∆)) = ne−λ∆
x>∆
Furthermore,
Z
X
|Wi | ≈ n
Z
x>∆
i∈I∆
(λe−λx )xdx ≈
xf (x)dx = n
x>∆
n
λ
(λ∆ + 1) e−λ∆
Then,
G(∆)
=
G0 (∆)
=
G0 (∆)
=
00
G (∆)
(
P
|Wi |)2
|I∆ |
n
λ2
=
n
λ2
(1 + λ∆)2 e−λ∆
(2λ(1 + λ∆)e−λ∆ − λ(1 + λ∆)2 e−λ∆ )
0⇒∆=
< 0
∆=1/λ
i∈I∆
1
λ
(maxima)
Therefore,
∆∗ =
1
λ
=
1X
n
11
i
|Wi |
| 9cs.NE
|
arXiv:1605.02306v1 [math.GT] 8 May 2016
CONJUGATION-INVARIANT NORMS ON THE COMMUTATOR
SUBGROUP OF THE INFINITE BRAID GROUP
MITSUAKI KIMURA
Abstract. In this paper, we give a proof of the result of Brandenbursky and
Kȩdra which says that the commutator subgroup of the infinite braid group
admits stably unbounded norms. Moreover, we observe the norms which we
constructed are equivalent to the biinvariant word norm studied by Brandenbursky and Kȩdra.
1. Introduction
In [2], Burago, Ivanov, and Polterovich introduced the notion of conjugationinvariant norms and asked several problems. One of them is as follows:
Problem 1.1 ([2]). Does there exists a perfect group G which satisfies the following
conditions?
(1) The commutator length is stably bounded on [G, G].
(2) G admits a stably unbounded norm.
For the definitions on conjugation-invariant norms, see [2]. It is known that such
groups exist. Brandenbursky and Kȩdra [1] proved the following theorem:
Theorem 1.2 ([1]). The commutator subgroup of the infinite braid group [B∞ , B∞ ]
admits a stably unbounded norm.
Kawasaki [4] also showed that the commutator subgroup of Sympc (R2n )0 is such
a group, where Sympc (R2n )0 is the group of symplectomorphisms with compact
support isotopic to the identity of the standard symplectic space. In this paper, we
give a proof of Theorem1.2 by using the idea of Kawasaki in [4]. Kawasaki introduced ν-quasimorphisms (or relative quasimorphisms) and the (ν, p, q)-commutator
length clν,p,q . He proved that the existence of non-trivial ν-quasimorphism implies
stably unboundedness of clν,p,q (see Proposition 2.3).
To give our proof of Theorem1.2, we construct stably unbounded norms clνn ,p,q
on [B∞ , B∞ ] by observing that the signature of braids σ : B∞ → R is a νn quasimorphism in §2.1. In §2.2, we study the property of the norms clνn ,p,q and
prove the following theorem:
Theorem 1.3. For any integer n ≥ 2 and any real numbers p ≥ 1 and q ≥ 1,
the norm clνn ,p,q is equivalent to the biinvariant word norm k · k (whose stably
unboundedness is observed by Brandenbursky and Kȩdra).
Acknowledgement. The author would like to thank Professor Takashi Tsuboi
for his guidance and helpful advice. He also thanks to Morimichi Kawasaki for his
useful advice.
1
2
MITSUAKI KIMURA
2. Proofs of main results
2.1. The construction of stably unbounded norms. First, we explain the idea
of Kawasaki. The definitions of (ν, p, q)-commutator length and ν-quasimorphisms
are following.
Definition 2.1 ([4]). Let G be a group with a conjugation-invariant norm ν and
p, q ∈ R>0 . We define the (ν, p, q)-commutator subgroup [G, G]ν,p,q to be a subgroup
of G generated by the elements [f, g] ∈ G such that ν(f ) ≤ p, ν(g) ≤ q.
We define (ν, p, q)-commutator length clν,p,q : [G, G]ν,p,q → R≥0 by
∃fi , ∃gi ∈ G, (i = 1, . . . , k)
clν,p,q (h) = min k
, h = [f1 , g1 ] · · · [fk , gk ] .
ν(fi ) ≤ p, ν(gi ) ≤ q
We note that clν,p,q is a conjugation-invariant norm.
Definition 2.2 ([4]). Let G be a group with a conjugation-invariant norm ν. A
function φ : G → R is called a ν-quasimorphism (or quasimorphism relative to ν)
if there exists a constant C > 0 such that for every f, g ∈ G,
|φ(f g) − φ(f ) − φ(g)| ≤ C min{ν(f ), ν(g)}.
The concept of ν-quasimorphisms appeared earlier in the paper of Entov and
Polterovich ([3], Theorem 7.1.), which is called “controlled quasi-additivity” in [3].
Kawasaki proved the following proposition.
Proposition 2.3 ([4], Proposition 3.4.). Let φ be a ν-quasimorphism on G and
n
p, q ∈ R>0 . If there exist an element h ∈ [G, G]ν,p,q such that limn→∞ φ(hn ) > 0,
then clν,p,q is stably unbounded on [G, G]ν,p,q .
Next, we give an useful sufficient condition to prove that a function is a qK quasimorphism, where qK is a conjugation-invariant norm on a group G which is
normally generated by a subset K ⊂ G defined by
qK (f ) = min{l | ∃ki ∈ K, ∃gi ∈ G, f = k1g1 · · · klgl },
where xy denotes the conjugation xyx−1 .
Lemma 2.4 ([3], Lemma 7.3.). Let G be a group which is normally generated by
a subset K ⊂ G, and φ : G → R a function on G. If there exists a constant C > 0
such that the inequality |φ(gh) − φ(g) − φ(h)| ≤ C holds for any g ∈ G and any
h ∈ G with qK (h) = 1, then φ is a qK -quasimorphism.
Finally, we construct the stably unbounded norms on [B∞ , B∞ ] by proving the
signature of braids is a νn -quasimorphism, where νn = qBn .
We denote the n-braid group by Bn and let σ1 , . . . , σn−1 be the standard Artin
generators. For a braid α ∈ Bn , we denote the closure of α by α
b. For a braid
α ∈ Bn , σ(α) is defined to be the signature σ(b
α) of the link α
b. We consider the
standard inclusion ιn : Bn → Bn+1 , σi 7→ σi (i.e. ιn is “adding a trivial string”)
and obtain the following sequence
S∞ : B1 ⊂ B2 ⊂ · · · ⊂ Bn ⊂ · · · . We define the
infinite braid group B∞ to be n=1 Bn . Since the signature of braids σ : Bn → R
and the inclusion ιn : Bn → Bn+1 are compatible, i.e. σ(ιn (α)) = σ(α) for α ∈ Bn ,
σ : B∞ → R is well-defined.
Theorem 2.5. The signature of braids σ : B∞ → R is a νn -quasimorphism.
CONJUGATION-INVARIANT NORMS ON [B∞ , B∞ ]
3
Proof. By Lemma 2.4, it is sufficient to prove that |σ(αβ) − σ(α) − σ(β)| ≤ n for
α, β ∈ B∞ , νn (β) = 1. The assumption νn (β) = 1 implies the existence of a braid
γ ∈ B∞ such that β γ ∈ Bn ⊂ B∞ . Let m be a natural number such that αγ ∈ Bm
(m > n). Since the signature of braids is conjugation-invariant,
|σ(αβ) − σ(α) − σ(β)| = |σ(αγ β γ ) − σ(αγ ) − σ(β γ )|.
γ β γ from α
cγ by taking saddle moves m times. Note
cγ ⊔ β
We obtain the link α[
γ
c
that the link β has m − n unknot components since β γ has trivial strings after
(n + 1)-th one (Figure 2). It is known that the signature changes at most ±1 by
one saddle move (see [5] for example). Since the signature does not changed by
taking connected sum to an unknot, the signature changes at most n by the m
times saddle moves. Hence
γ β γ ) − σ(α
cγ ⊔ βcγ )| ≤ n.
|σ(αγ β γ ) − σ(αγ ) − σ(β γ )| = |σ(α[
Figure 1. saddle move
Figure 2.
Now we prove Theorem 1.2.
Lemma 2.6. Let G be a group normally generated by K ⊂ G and ν a conjugationinvariant norm on G. We assume that ν is bounded on K (it is so when K is
finite). If p, q ≥ supk∈K ν(k), then [G, G]ν,p,q = [G, G].
4
MITSUAKI KIMURA
Proof. For a, b, r, s ∈ G, the following equalities hold:
[ar, b] = [r, b]a [a, b],
[a, bs] = [a, b][a, s]b .
By using them, we can represent a element f of [G, G] as a product of commutators
of the form [k ∗ , l∗ ] (k, l ∈ K). Therefore f ∈ [G, G]ν,p,q if p, q ≥ supk∈K ν(k).
Since B∞ is normally generated by {σ1±1 }, we have the following corollary.
Corollary 2.7. For n ≥ 2 and p, q ≥ 1, clνn ,p,q is well-defined on [B∞ , B∞ ].
(Proof of Theorem 1.2) We apply Proposition 2.3 to the signature σ : B∞ → R.
Since σ is a νn -quasimorphism by Theorem 2.5, it is sufficient to see that the stabilization of σ is non-trivial and it is already known (see [1] for example). Therefore,
for n ≥ 2 and p, q ≥ 1, clνn ,p,q is a stably unbounded norm on [B∞ , B∞ ]νn ,p,q =
[B∞ , B∞ ] by Corollary 2.7.
2.2. The extremal property. We study the properties of the norms clνn ,p,q . Two
norms on a group are called equivalent if their ratio is bounded away from 0 and
∞. First, we will show that the norms clνn ,p,q are equivalent each other. It follows
from the fact that the norms have an “extremal property”.
Lemma 2.8. For x, y, z ∈ G, assume that x and y z commute. Then [x, y] is written
as the products of 4 conjugates of z or z −1 .
Proof. By the assumption, [x, [y, z]] = xy(y −1 )z x−1 y z y −1 = xyx−1 y −1 = [x, y].
Thus [x, y] = [x, [y, z]] = xz y z −1 x−1 z(z −1)y = z xy (z −1 )x z(z −1 )y .
We call that a norm µ on [B∞ , B∞ ] has the extremal property if µ satisfies
the following condition: for any conjugation-invariant norm ν on [B∞ , B∞ ] which
satisfies ν(αβ ) = ν(α) for all α ∈ [B∞ , B∞ ] and β ∈ B∞ (not only [B∞ , B∞ ]),
there exists a positive number λ > 0 such that ν ≤ λµ.
By using above lemma, we observe that the norms clνn ,p,q have the property.
Proposition 2.9. For n ≥ 2 and p, q ≥ 1, clνn ,p,q has the extremal property.
Proof. First we prove that clνn ,1,1 has the extremal property. Let α, β ∈ B∞
satisfy νn (α) = νn (β) = 1. Then there exists γ, γ ′ ∈ B∞ such that α
e := αγ ∈ Bn ,
′
′ −1
′
βe := β γ ∈ Bn . If m is sufficiently large, α
eγ γ and βe∆m commute, where
−1
−1
∆′m = A′n,m−1 · · · A′n,2 A′n,1 A′n,2 · · · A′n,m−1
′
and A′n,i is a commutator argyle braid appeared in [1] . By Lemma 2.8, [α, β]γ =
′
′
′ −1
e is written as the products of 4 conjugates of ∆′±1 . Thus
[αγ , β γ ] = [e
αγ γ , β]
m
ν([α, β]) ≤ 4ν(∆′m ). Since ∆′m is conjugate to A′n,1 , ν(∆′m ) = ν(A′n,1 ). Then we
obtain ν ≤ 4ν(A′n,1 ) clνn ,1,1 . Since clνn ,1,1 ≤ pq clνn ,p,q , the proposition follows.
Since clνn ,p,q (αβ ) = clνn ,p,q (α) for all α ∈ [B∞ , B∞ ] and β ∈ B∞ , it follows from
Proposition 2.9 that the norms clνn ,p,q are equivalent to each other.
Next, we also consider the property of the biinvariant word norm k · k := q{σ±1 } .
1
Proposition 2.10. The biinvariant word norm k · k has the extremal property.
CONJUGATION-INVARIANT NORMS ON [B∞ , B∞ ]
5
Proof. Let α ∈ [B∞ , B∞ ] with kαk = k. Then α is written as follows:
α =(σ1ε1 )α1 (σ1ε2 )α2 · · · (σ1εk )αk
=[α1 , σ1ε1 ]σ1ε1 [α2 , σ1ε2 ]σ1ε2 · · · [αk , σ1εk ]σ1εk
ε1
=[α1 , σ1ε1 ][α2 , σ1ε2 ]σ1 σ1ε1 +ε2 · · · [αk , σ1εk ]σ1εk
=···
=[α1 , σ1ε1 ][α2 , σ1ε2 ]∗ · · · [αk , σ1εk ]∗ σ1ε1 +···+εk ,
where εi ∈ {±1} and αi ∈ B∞ . Since α ∈ [B∞ , B∞ ], ε1 + · · · + εk = 0. Thus α is
′
written as a product of k commutators of the form [∗, σ1±1 ]∗ . Since αi and (σ1εi )∆m
commute in B∞ for a sufficient large m, ν([αi , σ1εi ]∗ ) = ν([αi , σ1εi ]) ≤ 4ν(∆′m ) =
4ν(A′2,1 ) and it follows that ν(α) ≤ 4ν(A′2,1 )k = 4ν(A′2,1 )kαk.
As a corollary of Proposition 2.9 and 2.10, it follows that the norm k · k is also
equivalent to the norms clνn ,p,q we constructed (Theorem 1.3) .
References
[1] M. Brandenbursky and J. Kȩdra, Concordance group and stable commutator length in braid
groups, Algebr. Geom. Topol. (to appear)
[2] D. Burago, S. Ivanov, and L. Polterovich, Conjugation-invariant norms on groups of geometric origin, Adv. Stud. Pure Math. 52 (2008), 221-250.
[3] M. Entov and L. Polterovich, Quasi-states and symplectic intersections, Comment. Math.
Helv., 81 (1) (2006), 75-99.
[4] M. Kawasaki, Relative quasimorphisms and stably umbounded norms on the group of symplectomorphisms of the Euclidean spaces, J. Symplectic Geom. (to appear)
[5] K. Murasugi, On a certain numerical invariant of link types, Trans. Amer. Math. Soc. 117
(1965), 387-422.
Graduate School of Mathematical Sciences, the University of Tokyo 3-8-1 Komaba,
Meguro-ku, Tokyo, 153-8914, Japan
E-mail address: [email protected]
| 4math.GR
|
INITIALIZATION OF MULTILAYER FORECASTING
ARTIFICIAL NEURAL NETWORKS
V.V. Bochkarev and Yu.S. Maslennikova
Kazan (Volga Region) Federal University
[email protected], [email protected]
Abstract
In this paper, a new method was developed for initialising artificial neural networks
predicting dynamics of time series. Initial weighting coefficients were determined for neurons
analogously to the case of a linear prediction filter. Moreover, to improve the accuracy of the
initialization method for a multilayer neural network, some variants of decomposition of the
transformation matrix corresponding to the linear prediction filter were suggested. The efficiency
of the proposed neural network prediction method by forecasting solutions of the Lorentz chaotic
system is shown in this paper.
Keywords: Artificial neural networks; Forecasting; Neural network initialization; Linear
prediction filter
Introduction
Reliable forecasts are necessary nowadays both in scientific research and daily operations.
At present, this problem is usually being solved by the linear prediction method. The latter
reduces the solution of a forecasting problem to the search of coefficients of the linear prediction
filter of the pth order that gives the best prediction of the current value of a real-valued sequence
x̂ n from its previous values:
xˆ n a1 xn1 a2 xn2 a p xn p ;
(1)
here p is the filter order, and the letter а denotes coefficients of the filter. As a rule, one
minimizes the sum of squared prediction errors considered as a function of the filter coefficients
[1]. Since the graph of this function is a paraboloid, the minimum point is unique.
The linear prediction method is efficient for stationary systems, i.e., those systems whose
properties do not change in course of time, while the distribution of parameter fluctuations is
close to the Gaussian distribution [1, 2]. In more complicated cases one should better use
nonlinear methods, e.g., artificial neural networks [3, 4]. However, this method also has some
drawbacks. Thus, in the case of neural networks, owing to the nonlinearity of the neurons
activation function, the error functional is more complicated, and the calculated local minimum
point is not necessary global. As a result, in spite of the great potential of neural networks [5],
the error of a neural network prediction model may exceed that of the linear one.
As is mentioned in [3], when using neural networks, the choice of the rule for the
initialization of neuron weighting coefficients plays an important role. Algorithms for the
initialization of neural networks imply the choice of initial values of weighting coefficients. This
choice can be performed by some deterministic technique or random one (e.g., the widely used
Nguyen-Widrow algorithm [5]). In the latter case, as a rule, the network is being trained several
times in order to increase the probability of finding the global minimum of the error function.
This leads to a multiple growth of the computational burden, while the obtained result may
appear to be unsatisfactory.
In [6] one proposes an interesting algorithm for constructing a hybrid neural network [4]
for time series prediction. First, one trains the network consisting of a single linear neuron. The
obtained coefficients are used for initializing a hybrid network, which also contains only one
neuron. Then one sequentially adds neurons to the network until the decrease of the prediction
error stops. In this paper we consider another approach which is also based on the preliminary
solution of the optimal linear prediction problem.
Formula (1) demonstrates that the linear prediction filter is representable as the one-layer
linear neural direct transmission network. This analogy allows us to use the preliminarily
calculated coefficients of the linear prediction model as a base, including them in the
corresponding elements of the neural network. The idea is to construct a neural network which
before training transforms data in just the same way as expression (1). In order to make the full
use of abilities of neural network algorithms for approximating complex dependencies, the
network has to consist of 2 or more layers [3]. And then we meet a difficulty. It is connected
with the ambiguity of the decomposition of the initial linear transform into several sequential
ones which correspond to layers of the neural network. In what follows we consider two possible
ways to overcome this difficulty.
The simplest algorithm for the initialization of a neural network based on coefficients of
the linear prediction model
Consider the simplest technique for solving the stated problem in the case when functions
implemented by various network layers are evidently separated. The error that occurs due to the
nonlinearity of neurons activation functions prior to the network training should be small. This
fact imposes certain restrictions on the choice of activation functions for layers. For example, the
hyperbolic tangent (the sigmoid) has a domain with a weak nonlinearity at the origin of
coordinates. When using neurons with sigmoidal activation functions, by properly scaling the
training set we can make transforms of a signal by the network neurons close to linear. Let us
consider, as an example, the application of the proposed initialization algorithm to a three-layer
neural direct transmission network [4].
Assume that input values of all neurons correspond to linear parts of the activation
function. Represent the resulting transformation matrix of a three-layer neural network as
follows:
A A3 A2 A1 ,
(2)
where А1, А2, А3 are transformation matrices of the 1st, 2nd, and 3d layers, respectively (Fig. 1).
The input layer of the network scales the data so as to make all values of the series be located
within the linear domain of the hyperbolic tangent.
0
0
А= 0 0 0 ... 1 / · ...
0
а1
1
0
...
0
1
...
0
а2
0
а3
0 0 0 ... 0
0 0 0 ... 0
... · 0 0 ... 0
... 1 0 0 0 ... 0
... а р 0 0 0 ...
...
...
...
Fig. 1. The transformation matrix of a neural network after the initialization.
The intermediate layer performs a weakly nonlinear transform with the help of the weight
matrix that contains coefficients а of the linear prediction model. The output layer performs the
converse data transform. By a proper choice of the coefficient one can make the error
stipulated by the nonlinearity of the activation function arbitrarily small. In this case the neural
network with weights chosen in such a way will estimate the current value x̂ k of the input signal
from preceding p sequential ones xk-1, xk-2, … xk-p in accordance with expression (1).
Note that results of the prediction of the given neural network before training nearly
coincide with the output of the linear prediction filter. When training the network with the help
of an algorithm with a monotonic decrease of the error (e.g., gradient and quasi-Newton
algorithms, those based on the conjugate gradients method, etc), the result obtained on any
training step is the same or even better than that obtained with the help of the linear prediction
filter. In this paper we use the Levenberg-Marquardt algorithm; it belongs to the class of quasiNewton methods that guarantee a high convergence rate [7].
Prediction of the Lorenz chaotic system
The proposed technique for the neural network initialization was applied for the
prediction of trajectories of the Lorenz system, which is the simplest example of a determinate
chaotic system [8]. It was first observed by Lorenz in numerical experiments when studying the
trajectory of a system of three connected quadratic ordinary differential equations that define
three modes of Oberbeck-Boussinesq equations for the convection of a liquid in a 2D layer
heated from below. The mentioned equations take the form
x y x ,
y xr z y,
z xy bz.
(3)
According to results of numerical modeling, solutions of the system with many values of
parameters asymptotically tend to unstable cycles, but have two evident clusterization centers
and thus form the so-called strange attractor [8]. Fig. 2 illustrates the solution to the system with
10, r 28, b 8 / 3 . Another important property of solutions to the Lorenz system is their
essential dependence on the initial condition, which constitutes the main feature of the chaotic
dynamics. Therefore, the long-term behavior of the system is poorly predictable.
Fig. 2. Trajectories of the Lorenz attractor with r=28 (at the left); the variation of the Х coordinate (at the
right)
We use the Lorenz system as the test problem due to the fact that it is a rather adequate
model for many real systems. In [8, 9] one mentions that this system approximately describes
oscillations of parameters in turbulent flows and variations of parameters of the geomagnetic
field. One also encounters such system of equations in models of economic processes [10].
For testing the proposed initialization algorithm we have constructed the series that
correspond to solutions of system (3). We have applied the Runge–Kutta method with the
automatic choice of the step value. We have thinned out the obtained series in order to get a
sample of an acceptable size covering a long time interval with the time step of 0.01 sec.
Moreover, we have removed the initial part of the sample in order to make the data used for
testing correspond to the movement near the attractor and to make the results independent of the
choice of the initial point.
Below we represent results of the prediction of the behavior of the given system with the
help of the neural network, where initial values of weights are chosen in accordance with
coefficients of the linear prediction filter. First we have calculated coefficients of the linear
prediction model and then on its base constructed the neural network in the way described above.
We predicted each current value from 5 sequential previous ones. The first and the second layer
of the neural network contained 5 neurons each. We trained the neural network with the help of
the error backpropagation method, minimizing the error functional by the Levenberg-Marquardt
algorithm within 300 epochs of training. The prediction was performed for a time interval
varying from 0.01 to 1 second.
Fig. 3. A part of the predicted series for the increase of the Х coordinate by 0.15 seconds; the
interior of the rectangle shows a zoomed part of the same series: the initial series (solid line), the linear
prediction (dashed line), the neural network prediction (dotted line)
Fig. 3 shows a part of the predicted series; for comparison we also depict the linear
prediction of the same series. This example is typical for the whole range of the prediction
interval, namely, the proposed initialization algorithm provides an essential improvement in
comparison both to the linear prediction method and to the neural network prediction method
used in the Nguyen-Widrow initialization algorithm. See Table 1 below for mean square
prediction errors for various time intervals in comparison to results obtained by other methods.
The results of numerical experiments prove the efficiency of the proposed neural network
initialization method. We can also ascertain that for the considered problem the neural network
prediction is much more efficient than the linear prediction method.
An improved network initialization method
The main drawback of the considered simplest algorithm consists in the fact that layers of
the network (at least, on the initial training stage) participate to some extent in the data
transformation. In order to fully use abilities of multilayer networks, it is desirable to make the
computational load of layers uniform. As was mentioned above, the described method for the
initialization of a neural network is not unique. According to the considered example of a threelayer neural network, one can include real orthogonal matrices U1 and U2 in the resulting
transformation matrix A (see expression (2) and Fig. 1). Then U1 U1 E , U1 U1 E , and
T
A A 3 U 2T U 2 A 2 U 1T U 1 A1 .
T
(4)
Therefore, before training, the neural network, where initial values of weighting coefficients of
layers are given by matrices U1 A 1 , U 2 A 2U 1T , and A 3U 2T , respectively, still completely
corresponds to the initial linear prediction filter. However, during the training process, because
of the nonlinearity of the neurons transfer function, we can expect an improvement of the
prediction accuracy. In order to choose optimal matrices U1 and U2, it is convenient to represent
them in the exponential form
U i e Gi .
(5)
As is known, in this case, in order to make matrices Ui orthogonal, matrices Gi have to be skewsymmetric
GiT Gi .
Let the number of neurons in the 1st and 2nd layers equal N. In the linear space of skewsymmetric NN-matrices (the dimension of this space equals N N 1 2 ) we choose some
basis, for example,
G k , m , k m,
Gi,kj,m m,i k , j k ,i m, j .
Then
Gi
x
k , m, i G
k , m .
(6)
k,m
For example, with N 3 we have
a b
0
Gi a 0 c a i
b c 0
0 1 0
1 0 0 bi
0 0 0
0 0 1
0 0 0
0 0 0 ci 0 0 1 .
1 0 0
0 1 0
If the number of neurons in layers is small, then one can determine unknowns in
expression (4) by the gradient descent method [7]. The most appropriate choice in this case is to
define the objective function as the error of the neural network which is initialized in the
mentioned way and then subject to the full course of training. However, it is clear that this
approach requires more calculations. In this case, only the calculation of the gradient from
coefficients xk ,m,i that enter in (6) requires ~ N 2 network training courses.
One can solve this problem either by using stochastic search algorithms or by the random
initialization of coefficients xk ,m,i . In this case instead of the exponential parameterization of
orthogonal matrices (5) one can perform the parameterization
U i I Gi I Gi ,
1
which requires less computations.
The search of the global minimum in this case requires multiple training of the neural
network. Another approach (we describe it below) is based on the following ideas. Let us choose
matrices U1 and U2 so as to make the nonlinear transformation of data in the neural network
bring the gain as soon as possible. The rate of the neural network training depends on the
gradient calculated by differentiating the error function with respect to neurons weights. We vary
coefficients xk ,m,i in (6) so as to maximize the norm of the gradient. Note that this process is
quite analogous to the training of the neural network by the error backpropagation method,
because, in essence, we just impose some linear constraints on the variation of weighting
coefficients in accordance with expressions (4). In view of this fact we expect the total time
consumption to be approximately two times greater than that of the simplest initialization
algorithm described above. The implementation of this algorithm needs minimal additions to
available libraries of neural network computations. For example, the network error function after
the first epoch of training can be used as the objective function. In both cases, one can solve the
optimization problem, for example, by the gradient descent method.
We have performed numerical experiments with the proposed algorithm for the same
model data and the same structure of the neural network as in the previous case. For simplicity of
programming, we used the error of the neural network after the first epoch of training as the
objective function. We performed 10 iterations of the gradient descent method. Then for the
network with calculated optimal values of parameters of initial weights we, as above, performed
300 epochs of training. With fixed parameters the total time consumption increased less than 2
times in comparison to the simplest initialization algorithm considered above.
See Table 1 for values of mean square prediction errors obtained by the application of 4
methods (the linear prediction method, the neural network prediction method with the NguyenWidrow initialization algorithm, the simplest algorithm, and the improved one based on the use
of coefficients of the linear prediction model) under the same conditions for prediction intervals
of 0.01, 0.02, 0.05, 0.1, and 1 second. Evidently, the use of the improved initialization algorithm
in all cases results in the further essential decrease of the prediction error.
Table 1
Prediction method
Linear prediction
method
Neural network
prediction method
Neural network
prediction method based
on the linear one
Improved neural
network prediction
method
σ (0.01 sec)
σ (0.02sec)
σ (0.05 sec)
σ (0.1 sec)
σ (1 sec)
3.13810-4
1310-4
0.0108
0.0646
3.4031
1.18410-4
9.44110-4
0.0062
0.0377
0.7285
0.50510-4
1.99710-4
0.0016
0.0097
0.0862
0.27510-4
0,98510-4
0.0009
0.0048
0.0564
For comparison we have also determined values of coefficients xk ,m,i by minimizing the error at
the output of the fully trained network. In this case the error of the prediction for 0.05 sec equals
0.00072 against the value of 0.0009 obtained by using the greatest gradient criterion. Evidently,
switching to the method that requires much more computations, we only insignificantly improve
the prediction accuracy.
Conclusion
According to the predicted behavior of the Lorenz chaotic system, the proposed
initialization method for a neural network allows one to essentially decrease the prediction error
in comparison both to that in the linear prediction method and the neural network one under the
use of universal initialization algorithms. It is also true that in this problem, under the same
conditions, the neural network prediction method is much more efficient than the linear one. The
improved initialization algorithm allows us to additionally improve the prediction accuracy by a
slight increase of the training time. Therefore, the developed method proved to be consistent for
predicting the behavior of trajectories of the Lorenz chaotic system, which allows one to use it
for predicting the behavior of such complex series.
References
1.
T.W.Anderson, «The Statistical Analysis of Time Series». Wiley-Interscience, 1994. 720 p.
2.
M.Arato, Linear Stochastic Systems with Constant Coefficients. A Statistical Approach.
Springer-Verlag, Berlin-Heidelberg-New York, 1982.
3.
A.I. Galushkin, Neural Networks Theory. Springer, 2007, 396 p.
4.
V.V. Kruglov, M.I. Dli, R.Y. Golunov. Moscow: “Fizmalit” Publisher, 2001. 224 p. (in
Russian).
5.
V. S. Medvedev and V. G. Potemkin, Neural networks: Matlab 6. Moscow: Dialog-MIFI,
2002. 496 p. (In Russian).
6.
A.Gholipour, C.Lucas, B.N.Araabi, and M.Shafiee, Solar Activity Forecast: Spectral
Analysis and Neurofuzzy Prediction. // Journal of Atmospheric and Solar-Terrestrial
Physics. (2005). V.67, p. 595-603.
7.
Ph.E.Gill, W.Murray, and M.H.Wright, Practical Optimization. N.Y.: Springer, 1981. 401 p.
8.
Richard M. Crownover, Introduction to Fractals and Chaos, Jones and Bartlett, 1995.
9.
P.G. Frick, Turbulence: Approaches and Models. 2-nd Edition. М.; Izhevsk: RDC “Regular
and Chaotic Dynamics”, 2010. 332 p. (In Russian).
10. Tönu Puu, Nonlinear Economic Dynamics. Springer, 4th ed., 1997, 288 p.
| 9cs.NE
|
arXiv:1802.07810v1 [cs.AI] 21 Feb 2018
Manipulating and Measuring Model Interpretability
Forough Poursabzi-Sangdeh
Daniel G. Goldstein
Jake M. Hofman
[email protected]
[email protected]
[email protected]
University of Colorado Boulder
Microsoft Research
Microsoft Research
Jennifer Wortman Vaughan
Hanna Wallach
[email protected]
[email protected]
Microsoft Research
Microsoft Research
Abstract
Despite a growing body of research focused on creating interpretable machine learning methods, there have been
few empirical studies verifying whether interpretable methods achieve their intended effects on end users. We present
a framework for assessing the effects of model interpretability on users via pre-registered experiments in which
participants are shown functionally identical models that vary in factors thought to influence interpretability. Using this
framework, we ran a sequence of large-scale randomized experiments, varying two putative drivers of interpretability:
the number of features and the model transparency (clear or black-box). We measured how these factors impact
trust in model predictions, the ability to simulate a model, and the ability to detect a model’s mistakes. We found
that participants who were shown a clear model with a small number of features were better able to simulate the
model’s predictions. However, we found no difference in multiple measures of trust and found that clear models did
not improve the ability to correct mistakes. These findings suggest that interpretability research could benefit from
more emphasis on empirically verifying that interpretable models achieve all their intended effects.
1
Introduction
Machine learning is increasingly used to make decisions that affect people’s lives in critical domains like criminal justice,
credit, lending, and medicine. Machine learning models are often evaluated based on their predictive performance on
held-out data sets, measured, for example, in terms of accuracy, precision, or recall. However, good performance on
held-out data may not be sufficient to convince decision makers that a model is trustworthy or reliable in the wild.
To address this problem, a new line of research has emerged that focuses on developing interpretable machine learning
methods. There are two common approaches. The first is to employ models in which the impact of each feature on
the model’s prediction is easy to understand. Examples include generative additive models (Lou et al., 2012, 2013;
Caruana et al., 2015) and point systems (Jung et al., 2017; Ustun and Rudin, 2016). The second is to provide post-hoc
explanations for (potentially complex) models. One thread of research in this direction looks at how to explain individual
predictions by learning simple local approximations of the model around particular data points (Ribeiro et al., 2016;
Lundberg and Lee, 2017) or estimating the influence of training examples (Koh and Liang, 2017), while another focuses
1
on visualizing model output (Wattenberg et al., 2016).
Despite the flurry of activity and innovation in this area, there is still no consensus about how to define, quantify, or
measure the interpretability of a machine learning model (Doshi-Velez and Kim, 2017). Indeed, different notions of
interpretability, such as simulatability, trustworthiness, and simplicity, are often conflated (Lipton, 2016). This problem
is exacerbated by the fact that there are different types of users of machine learning systems and these users may have
different needs in different scenarios. For example, the approach that works best for a regulator who wants to understand
why a particular person was denied a loan may be different from the approach that works best for a data scientist trying
to debug a machine learning model.
We take the perspective that the difficulty of defining interpretability stems from the fact that interpretability is
not something that can be directly manipulated or measured. Rather, interpretability is a latent property that can
be influenced by different manipulable factors (such as the number of features, the complexity of the model, the
transparency of the model, or even the user interface) and that impacts different measurable outcomes (such as an end
user’s ability to simulate, trust, or debug the model). Different factors may influence these outcomes in different ways.
As such, we argue that to understand interpretability, it is necessary to directly manipulate and measure the influence
that different factors have on real people’s abilities to complete tasks.
This endeavor goes beyond the realm of typical machine learning research. While the factors that influence interpretability are properties of the system design, the outcomes that we would ultimately like to measure are properties of human
behavior. Because of this, building interpretable machine learning models is not a purely computational problem. In
other words, what is or is not “interpretable” is defined by people, not algorithms. We therefore take an interdisciplinary
approach, building on decades of psychology and social science research on human trust in models (e.g., Önkal
et al., 2009; Dietvorst et al., 2015; Logg, 2017). The general approach used in this literature is to run randomized
human-subject experiments in order to isolate and measure the influence of different manipulable factors on trust. Our
goal is to apply this approach in order to understand interpretability—i.e., the relationships between properties of the
system design and properties of human behavior.
We present a sequence of large-scale randomized human-subject experiments, in which we varied factors that are
thought to make models more or less interpretable (Glass et al., 2008; Lipton, 2016) and measured how these changes
impacted people’s decision making. We focus on two factors that are often assumed to influence interpretability, but
rarely studied formally: the number of features and the model transparency, i.e., whether the model internals are clear
or a black box. We focus on laypeople as opposed to domain experts, and ask which factors help them simulate a
model’s predictions, gain trust in a model, and understand when a model will make mistakes. While others have
used human-subject experiments to validate or evaluate particular machine learning innovations in the context of
interpretability (e.g., Ribeiro et al., 2016; Lim et al., 2009), we attempt to isolate and measure the influence of different
factors in a more systematic way by taking an experimental approach.
In each of our pre-registered experiments, participants were asked to predict the prices of apartments in a single
neighborhood in New York City with the help of a machine learning model. Each apartment was represented in terms of
eight features: number of bedrooms, number of bathrooms, square footage, total rooms, days on the market, maintenance
fee, distance from the subway, and distance from a school. All participants saw the same set of apartments (i.e., the same
feature values) and, crucially, the same model prediction for each apartment, which came from a linear regression model.
What varied between the experimental conditions was only the presentation of the model. As a result, any observed
differences in the participants’ behavior between the conditions could be attributed entirely to the model presentation.
In our first experiment (Section 2), we hypothesized that participants who were shown a clear model with a small number
2
of features would be better able to simulate the model’s predictions and more likely to trust (and thus follow) the model’s
predictions. We also hypothesized that participants in different conditions would exhibit varying abilities to correct the
model’s inaccurate predictions on unusual examples. As predicted, we found that participants who were shown a clear
model with a small number of features were better able to simulate the model’s predictions; however, we did not find that
they were more likely to trust the model’s predictions and instead found no difference in trust between the conditions.
We also found that participants who were shown a clear model were less able to correct inaccurate predictions.
In our second experiment (Section 3), we scaled down the apartment prices and maintenance fees to match median
housing prices in the U.S. in order to determine whether the findings from our first experiment were merely an artifact of
New York City’s high prices. Even with scaled-down prices and fees, the findings from our first experiment replicated.
In our third experiment (Section 4), we dug deeper into our finding that there was no difference in trust between the
conditions. To make sure that this finding was not simply due to our measures of trust, we instead used the weight of
advice measure frequently used in the literature on advice-taking (Yaniv, 2004; Gino and Moore, 2007) and subsequently
used in the context of algorithmic predictions by Logg (2017). We hypothesized that participants would give greater
weight to the predictions of a clear model with a small number of features than the predictions of a black-box model
with a large number of features, and update their own predictions accordingly. We also hypothesized that participants’
behavior might differ if they were told that the predictions were made by a “human expert” instead of a black-box model
with a large number of features. Even with the weight of advice measure, we again found no difference in trust between
the conditions. We also found no difference in the participants’ behavior when they were told that the predictions were
made by a human expert.
We view these experiments as a first step toward a larger agenda aimed at quantifying and measuring the impact of
different manipulable factors that influence interpretability.
2
Experiment 1: Predicting apartment prices
Our first experiment was designed to measure the influence of the number of features and the model transparency on
three properties of human behavior that are commonly associated with interpretability: laypeople’s abilities to simulate
a model’s predictions, gain trust in a model, and understand when a model will make mistakes. Before running the
experiment, we posited and pre-registered three hypotheses:1
H1. Simulation. A clear model with a small number of features will be easiest for participants to simulate.
H2. Trust. Participants will be more likely to trust (and thus follow) the predictions of a clear model with a small
number of features than the predictions of a black-box model with a large number of features.
H3. Detection of mistakes. Participants in different conditions will exhibit varying abilities to correct the model’s
inaccurate predictions on unusual examples.
For the unusual examples, we intentionally did not pre-register any hypotheses about which conditions would make
participants more or less able to correct inaccurate predictions. On the one hand, if a participant understands the model
1 Links
to the pre-registration documents on aspredicted.org for each experiment are omitted to preserve author anonymity.
3
(a) Clear, two-feature condition (CLEAR -2).
(b) Black-box, two-feature condition (BB -2).
(c) Clear, eight-feature condition (CLEAR -8).
(d) Black-box, eight-feature condition (BB -8).
Figure 1: The four primary experimental conditions. In the conditions on top, the model used two features; on the
bottom, it used eight. In the conditions on the left, participants saw the model internals; on the right, they were presented
with the model as a black box.
better, she may be better equipped to correct examples on which the model makes mistakes. On the other hand, a
participant may place greater trust in a model she understands well, leading her to closely follow its predictions.
Prediction error. Finally, we pre-registered our intent to analyze participants’ prediction error in each condition, but
intentionally did not pre-register any directional hypotheses.
2.1
Experimental design
As explained in the previous section, we asked participants to predict apartment prices with the help of a machine
learning model. We showed all participants the same set of apartments and the same model prediction for each apartment.
What varied between the experimental conditions was only the presentation of the model. We considered four primary
4
(a)
(b)
Figure 2: Part of the testing phase in the first experiment: (a) participants were asked to guess the model’s prediction and
state their confidence (step 1) and (b) participants were asked to make their own prediction and state their confidence
(step 3).
experimental conditions in a 2 × 2 design:
• Some participants saw a model that uses only two features (number of bathrooms and square footage—the two
most predictive features), while some saw a model that uses all eight features. (Note that all eight feature values
were visible to participants in all conditions.)
• Some participants saw the model internals (i.e., a linear regression model with visible coefficients), while some
were presented with the model as a black box.
Screenshots from each of the four primary experimental conditions are shown in Figure 1. We additionally considered a
baseline condition in which there was no model available.
We ran the experiment on Amazon Mechanical Turk using psiTurk (Gureckis et al., 2016), an open-source platform
for designing online experiments. The experiment was IRB-approved. We recruited 1,250 participants, all located in
the U.S., with Mechanical Turk approval ratings greater than 97%. The participants were randomly assigned to the
five conditions (CLEAR -2, n = 248 participants; CLEAR -8, n = 247; BB -2, n = 247; BB -8, n = 256; and NO - MODEL,
n = 252). Each participant received a flat payment of $2.50.
Participants were first shown detailed instructions (including, in the clear conditions, a simple English description of the
corresponding two- or eight-feature linear regression model), before proceeding with the experiment in two phases.
In the training phase, participants were shown ten apartments in a random order. In the four primary experimental
conditions, participants were shown the model’s prediction of each apartment’s price, asked to make their own prediction,
and then shown the apartment’s actual price. In the baseline condition, participants were asked to predict the price of
each apartment and then shown the actual price.
In the testing phase, participants were shown another twelve apartments. The order of the first ten was randomized,
while the remaining two always appeared last, for reasons described below. In the four primary experimental conditions,
participants were asked to guess what the model would predict for each apartment (i.e., simulate the model) and to
indicate how confident they were in this guess on a five-point scale (Figure 2a). They were then shown the model’s
prediction and asked to indicate how confident they were that the model was correct. Finally, they were asked to make
5
their own prediction of the apartment’s price and to indicate how confident they were in this prediction (Figure 2b). In
the baseline condition, participants were asked to predict the price of each apartment and to indicate their confidence.
The apartments shown to participants were selected from a data set of actual Upper West Side apartments taken from
StreetEasy,2 a popular and reliable New York City real estate website, between 2013 and 2015. To create the models
for the four primary experimental conditions, we first trained a two-feature linear regression model on our data set
using ordinary least squares with Python’s scikit-learn library (Pedregosa et al., 2011), rounding coefficients to “nice”
numbers within a safe range.3 To keep the models as similar as possible, we fixed the coefficients for number of
bathrooms and square footage and the intercept of the eight-feature model to match those of the two-feature model,
and then trained a linear regression model with the remaining six features, following the same rounding procedure to
obtain “nice” numbers. The resulting coefficients are shown in Figure 1. When presenting the model predictions to
participants, we rounded predictions to the nearest $100,000.
To enable comparisons across experimental conditions, the ten apartments used in the training phase and the first
ten apartments used in the testing phase were selected from those apartments in our data set for which the rounded
predictions of the two- and eight-feature models agreed and chosen to cover a wide range of deviations between the
models’ predictions and the apartments’ actual prices. By selecting only apartments for which the two- and eight-feature
models agreed, we were able to ensure that what varied between the experimental conditions was only the presentation
of the model. As a result, any observed differences in the participants’ behavior between the conditions could be
attributed entirely to the model presentation.
The last two apartments used in the testing phase were chosen to test our third hypothesis—i.e., that participants in
different conditions will exhibit varying abilities to correct the model’s inaccurate predictions on unusual examples.
To test this hypothesis, we would ideally have used an apartment with strange or misleading features that caused the
two- and eight-feature models to make the same bad prediction. Unfortunately, there was no such apartment in our data
set, so we chose two examples to test different aspects of our hypothesis. Both of these examples exploited the models’
large coefficient ($350,000) for number of bathrooms. The first (apartment 11) was a one-bedroom, two-bathroom
apartment from our data set for which both models made high, but different, predictions. Comparisons between the
two- and eight- feature conditions were therefore impossible, but we could examine differences in accuracy between the
clear and black-box conditions. The second (apartment 12) was a synthetically generated one-bedroom, three-bathroom
apartment for which both models made the same (high) prediction, allowing comparisons between all conditions, but
ruling out accuracy comparisons since there was no ground truth. These apartments were always shown last to avoid the
previously studied phenomenon in which people trust a model less after seeing it make a mistake (Dietvorst et al., 2015).
2.2
Results
Having run our experiment, we compared participants’ behavior across the conditions. Doing so required us to compare
multiple responses from multiple participants, which was complicated by possible correlations among any given
participant’s responses. For example, some people might consistently overestimate apartment prices regardless of
the condition they are assigned to, while others might consistently provide underestimates. We addressed this by
fitting a mixed-effects model for each measure of interest to capture differences across conditions while controlling for
2 https://streeteasy.com/
3 In particular, for each coefficient, we found a value that was divisible by the largest possible exponent of ten and was in the safe range, which is
the coefficient value plus or minus stderr/4.
6
$100k
$0k
CLEAR−2
CLEAR−8
(a)
BB−2
BB−8
Experiment 1: Deviation
Experiment 1: Prediction error
Mean prediction error
$200k
Mean deviation from the model
Mean simulation error
Experiment 1: Simulation error
$150k
$100k
$50k
$0k
$300k
$200k
$100k
$0k
CLEAR−2
CLEAR−8
(b)
BB−2
BB−8
CLEAR−2 CLEAR−8
BB−2
BB−8 NO−MODEL
(c)
Figure 3: Results from our first experiment: (a) mean simulation error, (b) mean deviation of participants’ predictions
from the model’s prediction (a smaller value indicates higher trust), and (c) mean prediction error. Error bars indicate
one standard error.
participant-level effects—a standard approach for analyzing repeated measures experimental designs (Bates et al., 2015).
We derived all plots and statistical tests from these models; plots show averages with one standard error by condition
from the fitted models, and statistical tests report degrees of freedom, test statistics, and p-values under the models.4
Unless otherwise noted, all plots and statistical tests correspond to just the first ten apartments from the testing phase.
H1. Simulation. We defined a participant’s simulation error to be the absolute deviation between the model’s prediction,
m, and the participant’s guess for that prediction, um —that is, |m − um |. Figure 3a shows the mean simulation error in
the testing phase. As hypothesized, participants in the CLEAR -2 condition had lower simulation error, on average, than
participants in the other conditions (t(996) = 11.91, p < .001 for the contrast of CLEAR -2 with the other three primary
conditions). This suggests that, on average, participants in this condition had some understanding of how the model
works. Participants in the CLEAR -8 condition appeared to have higher simulation error, on average, than participants in
the BB -8 condition who could not see the model’s internals (t(996) = 3.00, p = .002 for the contrast of CLEAR -8 with
BB -8), though we note that this comparison is not one we pre-registered and could be due to chance.
H2. Trust. To measure trust, we calculated the absolute deviation between the model’s prediction, m, and the
participant’s prediction of the apartment’s price, ua —that is, |m − ua |; a smaller value indicates higher trust. Figure 3b
shows that contrary to our second hypothesis, we found no significant difference in participants’ deviation from the
model between CLEAR -2 and BB -8. (We did find statistically but not practically significant differences in participants’
self-reported confidence in the models’ predictions.)
H3. Detection of mistakes. We used the last two apartments in the testing phase (apartment 11 and apartment 12) to
test our third hypothesis. The models made erroneously high predictions on these examples. For both apartments, we
found that participants in the four primary experimental conditions overestimated the apartments’ prices, compared to
participants in the baseline condition. We suspect that this is due to an anchoring effect around the models’ predictions.
For apartment 11, we found no significant difference in participants’ deviation from the model’s prediction between
the four primary conditions (see Figure 4a). For apartment 12, we found a significant difference between the clear and
black-box conditions (t(996) = 2.96, p = .003 for the contrast of CLEAR -2 and CLEAR -8 with BB -2 and BB -8). In
particular, participants in the clear conditions deviated from the model’s prediction less, on average, than participants in
the black-box conditions, resulting in even worse final predictions of the apartment’s price (see Figure 4b). This finding
4 We follow standard notation, where, e.g., the result of a t-test with n degrees of freedom is reported as t(n) = x, p = y, where x is the value
of the test statistic and y is the corresponding p-value.
7
Mean deviation from the model
for apartment 12
Mean deviation from the model
for apartment 11
Experiment 1: Deviation in apartment 11
$150k
$100k
$50k
$0k
CLEAR−2
CLEAR−8
BB−2
BB−8
Experiment 1: Deviation in apartment 12
$300k
$250k
$200k
$150k
$100k
$50k
$0k
CLEAR−2
$15k
$10k
$5k
$0k
CLEAR−2
CLEAR−8
BB−2
BB−8
(b)
Experiment 2: Deviation in apartment 11
Mean deviation from the model
for apartment 12
Mean deviation from the model
for apartment 11
(a)
CLEAR−8
BB−2
BB−8
(c)
Experiment 2: Deviation in apartment 12
$30k
$20k
$10k
$0k
CLEAR−2
CLEAR−8
BB−2
BB−8
(d)
Figure 4: Mean deviation from the model for apartments 11 and 12 in our first experiment (top) and in our second
experiment (bottom). Error bars indicate one standard error. (Note that for apartment 11, comparisons between the twoand eight-feature conditions are not possible because the models make different predictions.)
contradicts the common intuition that transparency enables users to understand when a model will make mistakes.
Prediction error. We defined prediction error to be the absolute deviation between the apartment’s actual price, a,
and the participant’s prediction of the apartment’s price, ua —that is, |a − ua |. Figure 3c shows that we did not find a
significant difference between the four primary experimental conditions. However, participants in the baseline condition
had significantly higher error than participants in the four primary conditions (t(1248) = 15.27, p < .001 for the
contrast of the baseline with the four primary conditions).
3
Experiment 2: Scaled-down prices
One potential explanation for participants’ poor abilities to correct inaccurate predictions might be their lack of
familiarity with New York City’s unusually high apartment prices. For example, if a participant finds Upper West Side
prices to be unreasonably high even when the model is correct, she may not notice that the model has placed too much
weight on number of bathrooms. Our second experiment was designed to address this issue by replicating our first
experiment with apartment prices and maintenance fees scaled down to match median housing prices in the U.S. Before
running this experiment we pre-registered three hypotheses. The first two hypotheses (H4 and H5) are identical to H1
and H2 from our first experiment. We made the third hypothesis more precise than H3 to reflect the results of our first
experiment and a small pilot with scaled-down prices:
8
$10k
$0k
CLEAR−2
CLEAR−8
BB−2
(a)
BB−8
Experiment 2: Deviation
Experiment 2: Prediction error
$15k
Mean prediction error
$20k
Mean deviation from the model
Mean simulation error
Experiment 2: Simulation error
$10k
$5k
$0k
$30k
$20k
$10k
$0k
CLEAR−2
CLEAR−8
(b)
BB−2
BB−8
CLEAR−2 CLEAR−8
BB−2
BB−8 NO−MODEL
(c)
Figure 5: Results from our second experiment: (a) mean simulation error, (b) mean deviation of participants’ predictions
from the model’s predictions (a smaller value indicates higher trust), and (c) mean prediction error. Error bars indicate
one standard error.
H6. Detection of mistakes. Participants will be less likely to correct inaccurate predictions on unusual examples of a
clear model compared to a black-box model.
3.1
Experimental design
We first scaled down the apartment prices and maintenance fees from our first experiment by a factor of ten. To account
for this change, we also scaled down all regression coefficients (except for the coefficient for maintenance fee) by a factor
of ten. Apart from the description of the neighborhood from which the apartments were selected, the experimental design
was unchanged. We again ran the experiment on Amazon Mechanical Turk. We excluded people who had participated
in our first experiment, and recruited 750 new participants all of whom satisfied the selection criteria from our first
experiment. The participants were randomly assigned to the five conditions (CLEAR -2, n = 150; CLEAR -8, n = 150;
BB -2, n = 147; BB -8, n = 151; and NO - MODEL , n = 152). Each participant received a flat payment of $2.50.
3.2
Results
H4. Simulation. As hypothesized, and shown in Figure 5a, participants in the CLEAR -2 condition had significantly
lower simulation error, on average, than participants in the other conditions (t(596) = 10.28, p < .001 for the contrast
of CLEAR -2 with the other three primary conditions). This is in line with the finding from our first experiment.
H5. Trust. Contrary to our second hypothesis, and in line with the finding from our first experiment, we found no
significant difference in participants’ trust, as indicated by their deviation from the model, between CLEAR -2 and BB -8.
H6. Detection of mistakes. In line with the finding from our first experiment, we found no significant difference in
participants’ deviation from the model’s prediction between the four primary conditions for apartment 11 (see Figure 4c).
As hypothesized, and in line with the finding from our first experiment, participants in the clear conditions deviated
from the model’s prediction less, on average, than participants in the black-box conditions for apartment 12, resulting in
even worse final predictions of the apartment’s price (see Figure 4d). These findings both suggest that New York City’s
9
unusually high apartment prices do not explain participants’ poor abilities to correct inaccurate predictions.
Prediction error. Participants in the CLEAR -2 condition had statistically but not practically (< $3,000) significantly
lower prediction error (t(596) = 3.17, p = .001 for the contrast of CLEAR -2 with the other three primary conditions).
4
Experiment 3: Alternative measure of trust
In our first two experiments, we found that participants were no more likely to trust the predictions of a clear model with
a small number of features than the predictions of a black-box model with a large number of features, as indicated by the
deviation of their own predictions from the model’s prediction. However, perhaps another measure of trust would reveal
differences between the conditions. In this section, we therefore present our third experiment, which was designed to
allow us to compare participants’ trust across the conditions using an alternative measure of trust: the weight of advice
measure frequently used in the literature on advice-taking (Yaniv, 2004; Gino and Moore, 2007; Logg, 2017).
Weight of advice quantifies the degree to which people update their beliefs (e.g., predictions made before seeing the
model’s predictions) toward advice they are given (e.g., the model’s predictions). In the context of our experiment, it
is defined as |u2 − u1 | / |m − u1 |, where m is the model’s prediction, u1 is the participant’s initial prediction of the
apartment’s price before seeing m, and u2 is the participant’s final prediction of the apartment’s price after seeing m.
It is equal to 1 if the participant’s final prediction matches the model’s prediction and equal to 0.5 if the participant
averages their initial prediction and the model’s prediction.
To understand the benefits of comparing weight of advice across the conditions, consider the scenario in which u2 is
close to m. There are different reasons why this might happen. On the one hand, it could be the case that u1 was far
from m and the participant made a significant update to their initial prediction based on the model. On the other hand, it
could be the case that u1 was already close to m and the participant did not update her prediction at all. These two
scenarios are indistinguishable in terms of the participant’s deviation from the model’s prediction. In contrast, weight of
advice would be high in the first case and low in the second.
We additionally used this experiment as a chance to see whether participants’ behavior would differ if they were told
that the predictions were made by a “human expert” instead of a model. Previous studies have examined this question
from different perspectives with differing results (e.g., Önkal et al., 2009; Dietvorst et al., 2015). Most closely related to
our experiment, Logg (2017) found that when people were presented with predictions from either an algorithm or a
human expert, they updated their own predictions toward predictions from an algorithm more than they did toward
predictions from a human expert in a variety of domains. We were interested to see whether this finding would replicate.
We pre-registered four hypotheses:
H7. Trust (deviation). Participants’ predictions will deviate less from the predictions of a clear model with a small
number of features than the predictions of a black-box model with a large number of features.
H8. Trust (weight of advice). Weight of advice will be higher for participants who see a clear model with a small
number of features than for those who see a black-box model with a large number of features.
H9. Humans vs. machines. Participants will trust a human expert and a black-box model to differing extents. As
a result, their deviation from the model’s predictions and their weight of advice will also differ.
10
H10. Detection of mistakes. Participants in different conditions will exhibit varying abilities to correct the model’s
inaccurate predictions on unusual examples.
The first two hypotheses are variations on H2 from our first experiment, while the last hypothesis is identical to H3.
4.1
Experimental design
We considered the same four primary experimental conditions as in the first two experiments plus a new condition,
EXPERT, in which participants saw the same information as in BB -8, but with the black-box model labeled as “Human
Expert” instead of “Model.” We did not include a baseline condition because the most natural baseline would have been
to simply ask participants to predict apartment prices (i.e., the first step of the testing phase described below).
We again ran the experiment on Amazon Mechanical Turk. We excluded people who had participated in our first
two experiments, and recruited 1,000 new participants all of whom satisfied the selection criteria from our first two
experiments. The participants were randomly assigned to the five conditions (CLEAR -2, n = 202; CLEAR -8, n = 200;
BB -2, n = 202; BB -8, n = 198; and EXPERT , n = 197). Each participant received a flat payment of $1.50. We
excluded data from one participant who reported technical difficulties.
We asked participants to predict apartment prices for the same set of apartments used in the first two experiments.
However, in order to calculate weight of advice, we modified the experiment design so that participants were asked
for two predictions for each apartment during the testing phase: an initial prediction before being shown the model’s
prediction and a final prediction after being shown the model’s prediction. To ensure that participants’ initial predictions
were the same across the conditions, we asked for their initial predictions for all twelve apartments before introducing
them to the model or human expert and before informing them that they would be able to update their predictions. This
design has the added benefit of potentially reducing the amount of anchoring on the model or expert’s predictions.
Participants were first shown detailed instructions (which intentionally did not include any information about the
corresponding model or human expert), before proceeding with the experiment in two phases. In the (short) training
phase, participants were shown three apartments, asked to predict each apartment’s price, and shown the apartment’s
actual price. The testing phase consisted of two steps. In the first step, participants were shown another twelve apartments.
The order of all twelve apartments was randomized. Participants were asked to predict the price of each apartment. In
the second step, participants were introduced to the model or human expert before revisiting the twelve apartments. As
in the first two experiments, the order of the first ten apartments was randomized, while the remaining two (apartments
11 and 12) always appeared last. For each apartment, participants were first reminded of their initial prediction, then
shown the model or expert’s prediction, and then asked to make their final prediction of the apartment’s price.5
11
$50k
$0k
CLEAR−2 CLEAR−8
BB−2
(a)
BB−8
EXPERT
Experiment 3: Weight of Advice
Experiment 3: Prediction error
0.6
Mean prediction error
$100k
Mean weight of advice (WOA)
Mean deviation from the model
Experiment 3: Deviation
$150k
0.4
0.2
0.0
$200k
$100k
$0k
CLEAR−2 CLEAR−8
BB−2
(b)
BB−8
EXPERT
CLEAR−2 CLEAR−8
BB−2
BB−8
EXPERT
(c)
Figure 6: Results from our third experiment: (a) mean deviation of participants’ predictions from the model’s prediction
(a smaller value indicates higher trust), (b) mean weight of advice, and (c) mean prediction error. Error bars indicate
one standard error.
4.2
Results
H7. Trust (deviation). Contrary to our first hypothesis, and in line with the findings from our first two experiments, we
found no significant difference in participants’ deviation from the model between CLEAR -2 and BB -8 (see Figure 6a).
H8. Trust (weight of advice). Weight of advice is not well defined when a participant’s initial prediction matches
the model’s prediction (i.e., u1 = m). For each condition, we therefore calculated the mean weight of advice over
all participant–apartment pairs for which the participant’s initial prediction did not match the model’s prediction.6
This calculation can be viewed as calculating the mean conditioned on there being initial disagreement between the
participant and the model. Contrary to our second hypothesis, and in line with the findings for the measures of trust
in our first two experiments, we did not find a significant difference in participants’ weight of advice between the
CLEAR -2 and BB -8 conditions (see Figure 6b).
H9. Humans vs. machines. Contrary to our third hypothesis, we did not find a significant difference in participants’
trust, as indicated by either the deviation of their predictions from the model or expert’s prediction or by their weight of
advice, between the BB -8 and EXPERT conditions.
H10. Detection of mistakes. In contrast to our first two experiments, we did not find that participants in the clear
conditions were less able to correct inaccurate predictions.
5 We initially considered an alternative design in which participants were asked to predict each apartment’s price, shown the model’s prediction,
and then asked to update their own prediction before moving on to the next apartment. During pilots, it appeared that participants changed their initial
predictions in response to the model. To verify this, we ran a larger version of this experiment, hypothesizing that participants’ initial predictions
would deviate less from the model’s predictions in the CLEAR -2 condition. As predicted, this was indeed the case (t(241) = −3.41, p < .001).
The amount by which participants’ initial predictions change based on the model they see could be viewed as another measure of trust.
6 We found no significant difference in the fraction of times that participants’ initial predictions matched the model’s predictions.
12
5
Discussion and future work
We investigated how two factors that are thought to influence model interpretability—the number of features and
the model transparency—impact laypeople’s abilities to simulate a model’s predictions, gain trust in a model, and
understand when a model will make mistakes. Although we found that a clear model with a small number of features
was easier for participants to simulate, we found no difference in trust. We also found that participants were less able
to correct inaccurate predictions when they were shown a clear model instead of a black box. These findings suggest
that one should not take for granted that a “simple” or “transparent” model always leads to higher trust. However, we
caution readers against jumping to the conclusion that interpretable models are not valuable. Our experiments focused
on just one model, presented to one specific subpopulation, for only a subset of the scenarios in which interpretability
might play an important role. Instead, we see this work as the first of many steps towards a larger goal of rigorously
quantifying and measuring when and why interpretability matters.
The general experimental approach that we introduced—i.e., presenting people with models that make identical predictions but varying the presentation of these model in order to isolate and measure the impact of different factors on
people’s abilities to perform well-defined tasks—could be applied in a wide range of different contexts and may lead
to different conclusions in each. For example, instead of a linear regression model, one could examine decision trees
or rule lists in a classification setting. Or our experiments could be repeated with participants who are domain experts,
data scientists, or researchers in lieu of laypeople recruited on Amazon Mechanical Turk. Likewise, there are many
other scenarios to be explored such as debugging a poorly performing model, assessing bias in a model’s predictions, or
explaining why an individual prediction was made. We hope that our work can serve as a useful template for examining
the importance of interpretability in these and other contexts.
References
Douglas Bates, Martin Mächler, Ben Bolker, and Steve Walker. Fitting linear mixed-effects models using lme4. Journal
of Statistical Software, 67(1):1–48, 2015.
Rich Caruana, Yin Lou, Johannes Gehrke, Paul Koch, Marc Sturm, and Noemie Elhadad. Intelligible models for
healthcare: Predicting pneumonia risk and hospital 30-day readmission. In Proceedings of the 21st ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining (KDD), 2015.
Berkeley J. Dietvorst, Joseph P. Simmons, and Cade Massey. Algorithm aversion: People erroneously avoid algorithms
after seeing them err. Journal of Experimental Psychology: General, 144(1):114–126, 2015.
Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint
arXiv:1702.08608, 2017.
Francesca Gino and Don A. Moore. Effects of task difficulty on use of advice. Journal of Behavioral Decision Making,
20(1):21–35, 2007.
Alyssa Glass, Deborah L McGuinness, and Michael Wolverton. Toward establishing trust in adaptive agents. In
Proceedings of the 13th International Conference on Intelligent User Interfaces (IUI), 2008.
Todd M Gureckis, Jay Martin, John McDonnell, Alexander S Rich, Doug Markant, Anna Coenen, David Halpern,
Jessica B Hamrick, and Patricia Chan. psiTurk: An open-source framework for conducting replicable behavioral
experiments online. Behavior Research Methods, 48(3):829–842, 2016.
13
Jongbin Jung, Connor Concannon, Ravi Shro, Sharad Goel, and Daniel G. Goldstein. Simple rules for complex
decisions. arXiv preprint arXiv:1702.04690, 2017.
Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. In Proceedings of the
34th International Conference on Machine Learning (ICML), 2017.
Brian Y Lim, Anind K Dey, and Daniel Avrahami. Why and why not explanations improve the intelligibility of
context-aware intelligent systems. In Proceedings of the SIGCHI Conference on Human Factors in Computing
Systems (CHI), 2009.
Zachary C Lipton. The mythos of model interpretability. arXiv preprint arXiv:1606.03490, 2016.
Jennifer M. Logg. Theory of machine: When do people rely on algorithms? Harvard Business School NOM Unit
Working Paper No. 17-086, 2017.
Yin Lou, Rich Caruana, and Johannes Gehrke. Intelligible models for classification and regression. In Proceedings of
the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), 2012.
Yin Lou, Rich Caruana, Johannes Gehrke, and Giles Hooker. Accurate intelligible models with pairwise interactions. In
Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD),
2013.
Scott Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Advances in Neural Information
Processing Systems 30 (NIPS), 2017.
Dilek Önkal, Paul Goodwin, Mary Thomson, and Sinan Gönül. The relative influence of advice from human experts
and statistical methods on forecast adjustments. Journal of Behavioral Decision Making, 22:390–409, 2009.
Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu
Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vanderplas, Alexandre Passos, David Cournapeau,
Matthieu Brucher, Matthieu Perrot, and Édouard Duchesnay. Scikit-learn: Machine learning in Python. Journal of
Machine Learning Research, 12:2825–2830, 2011.
Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. Why should I trust you?: Explaining the predictions of any
classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining (KDD), 2016.
Berk Ustun and Cynthia Rudin. Supersparse linear integer models for optimized medical scoring systems. Machine
Learning Journal, 102(3):349–391, 2016.
Martin Wattenberg, Fernanda Viégas, and Moritz Hardt. Attacking discrimination with smarter machine learning. Accessed at https://research.google.com/bigpicture/attacking-discrimination-in-ml/,
2016.
Ilan Yaniv. Receiving other people’s advice: Influence and benefit. Organizational Behavior and Human Decision
Processes, 93:1–13, 2004.
14
| 2cs.AI
|
1
Recent Advances in Zero-shot Recognition
arXiv:1710.04837v1 [cs.CV] 13 Oct 2017
Yanwei Fu, Tao Xiang, Yu-Gang Jiang, Xiangyang Xue, Leonid Sigal, and Shaogang Gong
Abstract—With the recent renaissance of deep convolution neural networks, encouraging breakthroughs have been achieved on
the supervised recognition tasks, where each class has sufficient
training data and fully annotated training data. However, to
scale the recognition to a large number of classes with few
or now training samples for each class remains an unsolved
problem. One approach to scaling up the recognition is to develop
models capable of recognizing unseen categories without any
training instances, or zero-shot recognition/ learning. This article
provides a comprehensive review of existing zero-shot recognition
techniques covering various aspects ranging from representations
of models, and from datasets and evaluation settings. We also
overview related recognition tasks including one-shot and open
set recognition which can be used as natural extensions of zeroshot recognition when limited number of class samples become
available or when zero-shot recognition is implemented in a
real-world setting. Importantly, we highlight the limitations of
existing approaches and point out future research directions in
this existing new research area.
Index Terms—life-long learning, zero-shot recognition, oneshot learning, open-set recognition.
I. I NTRODUCTION
Humans can distinguish at least 30,000 basic object categories [1] and many more subordinate ones (e.g., breeds
of dogs). They can also create new categories dynamically
from few examples or purely based on high-level description.
In contrast, most existing computer vision techniques require
hundreds, if not thousands, of labelled samples for each
object class in order to learn a recognition model. Inspired
by humans’ ability to recognize without seeing examples, the
research area of learning to learn or lifelong learning [2], [3],
[4] has received increasing interests.
These studies aim to intelligently apply previously learned
knowledge to help future recognition tasks. In particular, a
major topic in this research area is building recognition models
capable of recognizing novel visual categories that have no
associated labelled training samples (i.e., zero-shot learning),
few training examples (i.e. one-shot learning), and recognizing
the visual categories under an ‘open-set’ setting where the
testing instance could belong to either seen or unseen/novel
categories.
These problems can be solved under the setting of transfer
learning. Typically, transfer learning emphasizes the transfer
Yanwei Fu and Xiangyang Xue are with the School of Data
Science, Fudan University, Shanghai, 200433, China. E-mail: {yanweifu,xyxue}@fudan.edu.cn;
Yu-Gang Jiang is with the School of Computer Science, Shanghai
Key Lab of Intelligent Information Processing, Fudan University. Email:
[email protected]; Yu-Gang Jiang is the corresponding author.
Leonid Sigal is with the Department of Computer Science, University of
British Columbia, BC, Canada. Email: [email protected];
Tao Xiang and Shaogang Gong are with the School of Electronic Engineering and Computer Science, Queen Mary University of London, E1 4NS, UK.
Email: {t.xiang, s.gong}@qmul.ac.uk.
of knowledge across domains, tasks, and distributions that are
similar but not the same. Transfer learning [5] refers to the
problem of applying the knowledge learned in one or more
auxiliary tasks/domains/sources to develop an effective model
for a target task/domain.
To recognize zero-shot categories in the target domain, one
has to utilize the information learned from source domain.
Unfortunately, it may be difficult for existing methods of
domain adaptation [6] to be directly applied on these tasks,
since there are only few training instances available on target
domain. Thus the key challenge is to learn domain-invariant
and generalizable feature representation and/or recognition
models usable in the target domain.
The rest of this paper is organized as follows: We give an
overview of zero-shot recognition in Sec. II. The semantic
representations and common models of zero-shot recognition
have been reviewed in Sec. III and Sec. IV respectively. Next,
we discuss the recognition tasks beyond zero-shot recognition
in Sec. V including generalized zero-shot recognition, openset recognition and one-shot recognition. The commonly used
datasets are discussed in Sec. VI; and we also discuss the
problems of using these datasets to conduct zero-shot recognition. Finally, we suggest some future research directions in
Sec. VII and conclude the paper in Sec. VIII.
II. OVERVIEW OF Z ERO - SHOT R ECOGNITION
Zero-shot recognition can be used in a variety of research
areas, such as neural decoding from fMRI images [7], face
verification [8], object recognition [9], video understanding
[10], [11], [12], [13], and natural language processing [14].
The tasks of identifying classes without any observed data
is called zero-shot learning. Specifically, in the settings of
zero-shot recognition, the recognition model should leverage
training data from source/auxiliary dataset/domain to identify the unseen target/testing dataset/domain. Thus the main
challenge of zero-shot recognition is how to generalize the
recognition models to identify the novel object categories
without accessing any labelled instances of these categories.
The key idea underpinning zero-shot recognition is to explore and exploit the knowledge of how an unseen class (in
target domain) is semantically related to the seen classes (in
the source domain). We explore the relationship of seen and
unseen classes in Sec. III, through the use of intermediatelevel semantic representations. These semantic representation
are typically encoded in a high dimensional vector space. The
common semantic representations include semantic attributes
(Sec. III-A) and semantic word vectors (Sec. III-B), encoding
linguistic context. The semantic representation is assumed to
be shared between the auxiliary/source and target/test dataset.
Given a pre-defined semantic representation, each class name
2
can be represented by an attribute vector or a semantic word
vector – a representation termed class prototype.
III. S EMANTIC R EPRESENTATIONS IN Z ERO - SHOT
R ECOGNITION
Because the semantic representations are universal and
shared, they can be exploited for knowledge transfer between
the source and target datasets (Sec. IV), in order to enable
recognition novel unseen classes. A projection function mapping visual features to the semantic representations is typically
learned from the auxiliary data, using an embedding model
(Sec. IV-A). Each unlabelled target class is represented in
the same embedding space using a class ‘prototype’. Each
projected target instance is then classified, using the recognition model, by measuring similarity of projection to the class
prototypes in the embedding space (Sec. IV-B). Additionally,
under an open set setting where the test instances could belong
to either the source or target categories, the instances of
target sets can also be taken as outliers of the source data;
therefore novelty detection [15] needs to be employed first to
determine whether a testing instance is on the manifold of
source categories; and if it is not, it will be further classified
into one of the target categories.
In this section, we review the semantic representations
used for zero-shot recognition. These representations can be
categorized into two categories, namely, semantic attributes
and beyond. We briefly review relevant papers in Table I.
The zero-shot recognition can be considered a type of
life-long learning. For example, when reading a description
‘flightless birds living almost exclusively in Antarctica’, most
of us know and can recognize that it is referring to a penguin,
even though most people have never seen a penguin in their
life. In cognitive science [16], studies explain that humans are
able to learn new concepts by extracting intermediate semantic
representation or high-level descriptions (i.e., flightless, bird,
living in Antarctica) and transferring knowledge from known
sources (other bird classes, e.g., swan, canary, cockatoo and
so on) to the unknown target (penguin). That is the reason
why humans are able to understand new concepts with no
(zero-shot recognition) or only few training samples (few-shot
recognition). This ability is termed “learning to learn”.
More interestingly, humans can recognize newly created
categories from few examples or merely based on high-level
description, e.g., they are able to easily recognize the video
event named “Germany World Cup winner celebrations 2014”
which, by definition, did not exist before July 2014. To teach
machines to recognize the numerous visual concepts dynamically created by combining multitude of existing concepts,
one would require an exponential set of training instances
for a supervised learning approach. As such, the supervised
approach would struggle with the one-off and novel concepts
such as “Germany World Cup winner celebrations 2014”, because no positive video samples would be available before July
2014 when Germany finally beat Argentina to win the Cup.
Therefore, zero-shot recognition is crucial for recognizing dynamically created novel concepts which are composed of new
combinations of existing concepts. With zero-shot learning,
it is possible to construct a classifier for “Germany World
Cup winner celebrations 2014”by transferring knowledge from
related visual concepts with ample training samples, e.g., “ FC
Bayern Munich - Champions of Europe 2013” and “ Spain
World Cup winner celebrations 2010”.
A. Semantic Attributes
An attribute (e.g., has wings) refers to the intrinsic characteristic that is possessed by an instance or a class (e.g., bird) (Fu
et al. [11]), or indicates properties (e.g., spotted) or annotations
(e.g., has a head) of an image or an object (Lampert et al. [9]).
Attributes describe a class or an instance, in contrast to the
typical classification, which names an instance. Farhadi et al.
[17] learned a richer set of attributes including parts, shape,
materials and etc. Another commonly used methodology (e.g.,
in human action recognition (Liu et al. [12]), and in attribute
and object-based modeling (Wang et al. [18])) is to take the attribute labels as latent variables on the training dataset, e.g., in
the form of a structured latent SVM model with the objective
is to minimize prediction loss. The attribute description of an
instance or a category is useful as a semantically meaningful
intermediate representation bridging a gap between low level
features and high level class concepts (Palatucci et al. [7]).
The attribute learning approaches have emerged as a promising paradigm for bridging the semantic gap and addressing data sparsity through transferring attribute knowledge in
image and video understanding tasks. A key advantage of
attribute learning is to provide an intuitive mechanism for
multi-task learning (Salakhutdinov et al. [19]) and transfer
learning (Hwang et al. [20]). Particularly, attribute learning
enables the learning with few or zero instances of each class
via attribute sharing, i.e., zero-shot and one-shot learning.
Specifically, the challenge of zero-shot recognition is to recognize unseen visual object categories without any training
exemplars of the unseen class. This requires the knowledge
transfer of semantic information from auxiliary (seen) classes,
with example images, to unseen target classes.
Later works (Parikh et al.[21], Kovashka et al. [22] and Berg
et al. [23]) extended the unary/binary attributes to compound
attributes, which makes them extremely useful for information
retrieval (e.g., by allowing complex queries such as “Asian
women with short hair, big eyes and high cheekbones”) and
identification (e.g., finding an actor whose name you forgot,
or an image that you have misplaced in a large collection).
In a broader sense, the attribute can be taken as one special
type of “subjective visual property” [24], which indicates
the task of estimating continuous values representing visual
properties observed in an image/video. These properties are
also examples of attributes, including image/video interestingness [25], [26], memorability [27], [28], aesthetic [29], and
human-face age estimation [30], [31]. Image interestingness
was studied in Gygli et al. [25], which showed that three
cues contribute the most to interestingness: aesthetics, unusualness/novelty and general preferences; the last of which
refers to the fact that people, in general, find certain types
3
of scenes more interesting than others, for example, outdoornatural vs. indoor-manmade. Jiang et al. [26] evaluated different features for video interestingness prediction from crowdsourced pairwise comparisons. ACM International Conference
on Multimedia Retrieval (ICMR) 2017 published special issue
(“multimodal understanding of subjective properties”1 ) on the
applications of multimedia analysis for subjective property
understanding, detection and retrieval. These subjective visual
properties can be used as an intermediate representation for
zero-shot recognition as well as other visual recognition tasks,
e.g., people can be recognized by the description of how pale
their skin complexion is and/or how chubby their face looks
[21]. In the next subsections, we will briefly review different
types of attributes.
1) User-defined Attributes: User-defined attributes are defined by human experts [32], [9], or a concept ontology [11].
Different tasks may also necessitate and contain distinctive
attributes, such as facial and clothes attributes [18], [33], [34],
[35], [36], [37], attributes of biological traits (e.g., age and
gender) [38], [39], product attributes (e.g., size, color, price)
[40] and 3D shape attributes [41]. Such attributes transcend the
specific learning tasks and are, typically, pre-learned independently across different categories, thus allowing transference
of knowledge [22], [42], [43]. Essentially, these attributes can
either serve as the intermediate representations for knowledge
transfer in zero-shot, one-shot and multi-task learning [40],
or be directly employed for advanced applications, such as
clothes recommendation [18].
Ferrari et al. [44] studied some elementary properties such
as colour and/or geometric pattern. From human annotations,
they proposed a generative model for learning simple color and
texture attributes. The attribute can be either viewed as unary
(e.g., red colour, round texture), or binary (e.g., black/white
stripes). The ‘unary’ attributes are simple attributes, whose
characteristic properties are captured by individual image
segments (appearance for red, shape for round). In contrast, the
‘binary’ attributes are more complex attributes, whose basic
element is a pair of segments (e.g., black/white stripes).
2) Relative Attributes: Attributes discussed above use single value to represent the strength of an attribute being
possessed by one instance/class; they can indicate properties
(e.g., spotted) or annotations of images or objects. In contrast,
relative information, in the form of relative attributes, can be
used as a more informative way to express richer semantic
meaning and thus better represent visual information. The
relative attributes can be directly used for zero-shot recognition
[21].
Relative attributes (Parikh et al. [21]) were first proposed
in order to learn a ranking function capable of predicting the
relative semantic strength of a given attribute. The annotators
give pairwise comparisons on images and a ranking function
is then learned to estimate relative attribute values for unseen images as ranking scores. These relative attributes are
learned as a form of richer representation, corresponding to
the strength of visual properties, and used in a number of tasks
including visual recognition with sparse data, interactive image
1 http://www.icmr2017.ro/call-for-special-sessions-s1.php
search (Kovashka et al. [22]), semi-supervised (Shrivastava et
al. [45]) and active learning (Biswas et al. [46], [47]) of visual
categories. Kovashka et al. [22] proposed a novel model of
feedback for image search where users can interactively adjust
the properties of exemplar images by using relative attributes
in order to best match his/her ideal queries.
Fu et al. [24] extended the relative attributes to “subjective
visual properties” and proposed a learning-to-rank model of
pruning the annotation outliers/errors in crowdsourced pairwise comparisons. Given only weakly-supervised pairwise
image comparisons, Singh et al. [48] developed an end-toend deep convolutional network to simultaneously localize and
rank relative visual attributes. The localization branch in [48]
is adapted from the spatial transformer network [49].
3) Data-driven attributes: The attributes are usually defined by extra knowledge of either expert users or concept ontology. To better augment such user-defined attributes, Parikh
et al. [50] proposed a novel approach to actively augment
the vocabulary of attributes to both help resolve intra-class
confusions of new attributes and coordinate the “name-ability”
and “discriminativeness” of candidate attributes. However,
such user-defined attributes are far from enough to model the
complex visual data. The definition process can still be either
inefficient (costing substantial effort of user experts) and/or
insufficient (descriptive properties may not be discriminative).
To tackle such problems, it is necessary to automatically discover more discriminative intermediate representations from
visual data, i.e. data-driven attributes. The data-driven attributes can be used in zero-shot recognition tasks [12], [11].
Despite previous efforts, an exhaustive space of attributes
is unlikely to be available, due to the expense of ontology
creation, and a simple fact that semantically obvious attributes,
for humans, do not necessarily correspond to the space of
detectable and discriminative attributes. One method of collecting labels for large scale problems is to use Amazon
Mechanical Turk (AMT) [73]. However, even with excellent
quality assurance, the results collected still exhibit strong label
noise. Thus label-noise [51] is a serious issue in learning from
either AMT, or existing social meta-data. More subtly, even
with an exhaustive ontology, only a subset of concepts from
the ontology are likely to have sufficient annotated training
examples, so the portion of the ontology which is effectively
usable for learning, may be much smaller. This inspired the
works of automatically mining the attributes from data.
Data-driven attributes have only been explored in a few
previous works. Liu et al. [12] employed an information theoretic approach to infer the data-driven attributes from training
examples by building a framework based on a latent SVM
formulation. They directly extended the attribute concepts in
images to comparable “action attributes” in order to better
recognize human actions. Attributes are used to represent
human actions from videos and enable the construction of
more descriptive models for human action recognition. They
augmented user-defined attributes with data-driven attributes
to better differentiate existing classes. Farhadi et al. [17] also
learned user-defined and data-driven attributes.
The data-driven attribute works in [12], [17], [74] are
limited. First, they learn the user-defined and data-driven
4
Different Types of Attributes
User-defined attributes
Relative attributes
Data-driven attributes
Video attributes
Concept ontology
Semantic word embedding
Papers
[32], [9][11][42], [43][18], [33], [34], [35], [36], [37][40][18][38], [39][44]
[21][48][24][46], [47][22][45]
[50][51][12], [17][11], [13][17][52]
[53][54][53][55][56], [57][58][59]
[60][61], [62][63][64], [65]
[66], [67], [68], [69], [70], [15][15][71][69][70], [66][72], [10]
TABLE I
D IFFERENT TYPES OF SEMANTIC REPRESENTATIONS FOR ZERO - SHOT RECOGNITION .
attributes separately, rather than jointly in the same framework. Therefore data-driven attributes may re-discover the
patterns that exist in the user-defined attributes. Second, the
data-driven attributes are mined from data and we do not
know the corresponding semantic attribute names for the
discovered attributes. For those reasons, usually data-driven
attributes can not be directly used in zero-shot learning. These
limitations inspired the works of [11], [13]. Fu et al. [11],
[13] addressed the tasks of understanding multimedia data
with sparse and incomplete labels. Particularly, they studied
the videos of social group activities by proposing a novel
scalable probabilistic topic model for learning a semi-latent
attribute space. The learned multi-modal semi-latent attributes
can enable multi-task learning, one-shot learning and zero-shot
learning. Habibian et al. [52] proposed a new type of video
representation by learning the “VideoStory” embedding from
videos and corresponding descriptions. This representation can
also be interpreted as data-driven attributes. The work won the
best paper award in ACM Multimedia 2014.
4) Video Attributes: Most existing studies on attributes
focus on object classification from static images. Another line
of work instead investigates attributes defined in videos, i.e.,
video attributes, which are very important for corresponding
video related tasks such as action recognition and activity understanding. Video attributes can correspond to a wide range of
visual concepts such as objects (e.g., animal), indoor/outdoor
scenes (e.g., meeting, snow), actions (e.g. blowing candle)
and events (e.g., wedding ceremony), and so on. Compared
to static image attributes, many video attributes can only be
computed from image sequences and are more complex in that
they often involve multiple objects.
Video attributes are closely related to video concept detection in Multimedia community. The video concepts in a
video ontology can be taken as video attributes in zero-shot
recognition. Depending on the ontology and models used,
many approaches on video concept detection (Chang et al.
[75], [76], Snoek et al. [54], Hauptmann et al. [53], Gan et
al. [64] and Qin et al. [77]) can therefore be seen as addressing
a sub-task of video attribute learning to solve zero-shot video
event detection. Some works aim to automatically expand (e.g.,
Hauptmann et al. [53] and Tang et al. [58]) or enrich (Yang et
al. [78]) the set of video tags [79], [55], [78] given a search
query. In this case, the expanded/enriched tagging space has
to be constrained by a fixed concept ontology, which may be
very large and complex [55], [80], [81]. For example, there is
a vocabulary space of over 20, 000 tags in [55].
Zero-shot video event detection has also attracted large
research attention recently. The video event is a higher level
semantic entity and is typically composed of multiple concepts/video attributes. For example, a “birthday party” event
consists of multiple concepts, e.g., “blowing candle” and
“birthday cake”. The semantic correlation of video concepts
has also been utilized to help predict the video event of
interest, such as weakly supervised concepts [82], pairwise
relationships of concepts (Gan et al. [65]) and general video
understanding by object and scene semantics attributes [56],
[57]. Note, a full survey of recent works on zero-shot video
event detection is beyond the scope of this paper.
B. Semantic Representations Beyond Attributes
Besides the attributes, there are many other types of semantic representations, e.g. semantic word vector and concept
ontology. Representations that are directly learned from textual
descriptions of categories have also been investigated, such
as Wikipedia articles [83], [84], sentence descriptions [85] or
knowledge graphs [61], [62].
1) Concept ontology: Concept ontology is directly used
as the semantic representation alternative to attributes. For
example, WordNet [86] is one of the most widely studied
concept ontologies. It is a large-scale semantic ontology
built from a large lexical dataset of English. Nouns, verbs,
adjectives and adverbs are grouped into sets of cognitive
synonyms (synsets) which indicate distinct concepts. The
idea of semantic distance, defined by the WordNet ontology,
is also used by Rohrbach et al. [61], [62] for transferring
semantic information in zero-shot learning problems. They
thoroughly evaluated many alternatives of semantic links between auxiliary and target classes by exploring linguistic bases
such as WordNet, Wikipedia, Yahoo Web, Yahoo Image, and
Flickr Image. Additionally, WordNet has been used for many
vision problems. Fergus et al. [60] leveraged the WordNet
ontology hierarchy to define semantic distance between any
two categories for sharing labels in classification. The COSTA
[63] model exploits the co-occurrences of visual concepts in
images for knowledge transfer in zero-shot recognition.
2) Semantic word vectors: Recently, word vector approaches, based on distributed language representations, have
gained popularity in zero-shot recognition [66], [67], [68],
[69], [70], [15]. A user-defined semantic attribute space is
pre-defined and each dimension of the space has a specific semantic meaning according to either human experts or concept
ontology (e.g., one dimension could correspond to ‘has fur’,
and another ‘has four legs’)(Sec. III-A1). In contrast, the semantic word vector space is trained from linguistic knowledge
bases such as Wikipedia and UMBCWebBase using natural
language processing models [71], [87]. As a result, although
5
the relative positions of different visual concepts will have
semantic meaning, e.g., a cat would be closer to a dog than a
sofa, each dimension of the space does not have a specific
semantic meaning. The language model is used to project
each class’ textual name into this space. These projections
can be used as prototypes for zero-shot learning. Socher et
al. [15] learned a neural network model to embed each image
into a 50-dimensional word vector semantic space, which
was obtained using an unsupervised linguistic model [71]
trained on Wikipedia text. The images from either known or
unknown classes could be mapped into such word vectors and
classified by finding the closest prototypical linguistic word in
the semantic space.
Distributed semantic word vectors have been widely used
for zero-shot recognition. Skip-gram model and CBOW model
[87], [88] were trained from a large scale of text corpora to
construct semantic word space. Different from the unsupervised linguistic model [71], distributed word vector representations facilitate modeling of syntactic and semantic regularities
in language and enable vector-oriented reasoning and vector
arithmetics. For example, V ec(“Moscow”) should be much
closer to V ec(“Russia”)+V ec(“capital”) than V ec(“Russia”)
or V ec(“capital”) in the semantic space. One possible explanation and intuition underlying these syntactic and semantic
regularities is the distributional hypothesis [89], which states
that a word’s meaning is captured by other words that cooccur with it. Frome et al. [70] further scaled such ideas to
recognize large-scale datasets. They proposed a deep visualsemantic embedding model to map images into a rich semantic
embedding space for large-scale zero-shot recognition. Fu
et al. [69] showed that such a reasoning could be used to
synthesize all different label combination prototypes in the
semantic space and thus is crucial for multi-label zero-shot
learning. More recent work of using semantic word embedding
includes [66], [67], [68].
More interestingly, the vector arithmetics of semantic emotion word vectors is matching the psychological theories
of Emotion, such as Ekman’s six pan-cultural basic emotions or Plutchik’s emotion. For example, V ec(“Sur prise”) +
V ec(“Sadness”) is very close to V ec(“Disappointment”); and
V ec(“Joy”) + V ec(“Trust”) is very close to V ec(“Love”).
Since there are usually thousands of words that can describe
emotions, zero-shot emotion recognition has been also investigated in [72] and [10].
IV. M ODELS FOR Z ERO - SHOT R ECOGNITION
With the help of semantic representations, zero-shot recognition can usually be solved by first learning an embedding
model (Sec. IV-A) and then doing recognition (Sec. IV-B). To
the best of our knowledge, a general ‘embedding’ formulation
of zero-shot recognition was first introduced by Larochelle
et al. [90]. They embedded handwritten character with a
typed representation which further helped to recognize unseen
classes.
The embedding models aim to establish connections between seen classes and unseen classes by projecting the lowlevel features of images/videos close to their corresponding
semantic vectors (prototypes). Once the embedding is learned,
from known classes, novel classes can be recognized based on
the similarity of their prototype representations and predicted
representations of the instances in the embedding space. The
recognition model matches the projection of the image features against the unseen class prototypes (in the embedding
space). In addition to discussing these models and recognition
methods in Sec. IV-A and Sec. IV-B, respectively, we will
also discuss the potential problems encountered in zero-shot
recognition models in Sec. IV-C.
A. Embedding Models
1) Bayesian Models: The embedding models can be
learned using a Bayesian formulation, which enables easy
integration of prior knowledge of each type of attribute to
compensate for limited supervision of novel classes in image
and video understanding. A generative model is first proposed
in Ferrari and Zisserman in [44] for learning simple color and
texture attributes.
Lampert et al. [32], [9] is the first to study the problem of
object recognition of categories for which no training examples
are available. Direct Attribute Prediction (DAP) and Indirect
Attribute Prediction (IAP) are the first two models for zeroshot recognition [32], [9]. DAP and IAP algorithms use a
single model that first learns embedding using Support Vector
Machine (SVM) and then does recognition using Bayesian formulation. The DAP and IAP further inspired later works that
employ generative models to learn the embedding, including
with topic models [13], [11], [91] and random forests [92].
We briefly describe the DAP and IAP models as follows,
• DAP Model. Assume the relation between known
classes, yi, ..., yk , unseen classes, z1, ..., z L , and descriptive attributes a1, ..., a M is given by the matrix of binary
y
z . Such a matrix encodes the
associations values am and am
presence/absence of each attribute in a given class. Extra
knowledge is applied to define such an association matrix,
for instance, by leveraging human experts (Lampert et
al. [32], [9]), by consulting a concept ontology (Fu et
al. [13]), or by semantic relatedness measured between
class and attribute concepts (Rohrbach et al. [61]). In the
training stage, the attribute classifiers are trained from
the attribute annotations of known classes yi, ..., yk . At
the test stage, the posterior probability p(am |x) can be
inferred for an individual attribute am in an image x. To
predict the class label of object class z,
p(z|x) = Σa ∈ {0,1} M p(z|a)p(a|x)
=
•
p(z)
p(az )
M
Ö
z
p(am |x)am
(1)
(2)
m=1
IAP Model. The DAP model directly learns attribute
classifiers from the known classes, while the IAP model
builds attribute classifiers by combining the probabilities
of all associated known classes. It is also introduced as
direct similarity-based model in Rohrbach et al. [61]. In
the training step, we can learn the probabilistic multiclass classifier to estimate p(yk |x) for all training classes
6
yi, ..., yk . Once p(a|x) is estimated, we use it in the same
way as in for DAP in zero-shot learning classification
problems. In the testing step, we predict,
K
p(am |x) = Σk=1
p(am |yk )p(yk |x)
(3)
2) Semantic Embedding: Semantic embedding learns the
mapping from visual feature space to the semantic space which
has various semantic representations. As discussed in Sec.
III-A, the attributes are introduced to describe objects; and the
learned attributes may not be optimal for recognition tasks.
To this end, Akata et al. [93] proposed the idea of label
embedding that takes attribute-based image classification as
a label-embedding problem by minimising the compatibility
function between an image and a label embedding. In their
work, a modified ranking objective function was derived from
the WSABIE model [94]. As object-level attributes may suffer
from the problems of the partial occlusions, scale changes
of images, Li et al. [95] proposed learning and extracting
attributes on segments containing the entire object; and then
joint learning for simultaneous object classification and segment proposal ranking by attributes. They thus learned the
embedding by the max-margin empirical risk over both the
class label as well as the segmentation quality. Other semantic
embedding algorithms have also been investigated such as
semi-supervised max-margin learning framework [96], [97],
latent SVM [67] or multi-task learning [20], [98], [99].
3) Embedding into Common Spaces: Besides the semantic
embedding, the relationship of visual and semantic space can
be learned by jointly exploring and exploiting a common
intermediate space. Extensive efforts [84], [99], [100], [101],
[102], [103], [104] had been made towards this direction.
Akata et al. [100] learned a joint embedding semantic space
between attributes, text and hierarchical relationships. Ba et
al. [84] employed text features to predict the output weights
of both the convolutional and the fully connected layers in a
deep convolutional neural network (CNN).
On one dataset, there may exist many different types of
semantic representations. Each type of representation may contain complementary information. Fusing them can potentially
improve the recognition performance. Thus several recent
works studied different methods of multi-view embedding. Fu
et al. [105] employed the semantic class label graph to fuse the
scores of different semantic representations. Similarly label relation graphs have also been studied in [106] and significantly
improved large-scale object classification in supervised and
zero-shot recognition scenarios.
A number of successful approaches to learning a semantic
embedding space reply on Canonical Component Analysis
(CCA). Hardoon et al. [107] proposed a general, kernel CCA
method, for learning semantic embedding of web images
and their associated text. Such embedding enables a direct
comparison between text and images. Many more works [108],
[109], [110], [111] focused on modeling the images/videos
and associated text (e.g., tags on Flickr/YouTube). Multi-view
CCA is often exploited to provide unsupervised fusion of
different modalities. Gong et al. [109] also investigated the
problem of modeling Internet images and associated text or
tags and proposed a three-view CCA embedding framework
for retrieval tasks. Additional view allows their framework
to outperform a number of two-view baselines on retrieval
tasks. Qi et al. [112] proposed an embedding model for jointly
exploring the functional relationships between text and image
features for transferring inter-model and intra-model labels to
help annotate the images. The inter-modal label transfer can
be generalized to zero-shot recognition.
4) Deep Embedding: Most of recent zero-shot recognition
models have to rely the state-of-the-art deep convolutional
models to extract the image features. As one of the first
works, DeViSE [70] extended the deep architecture to learn
the visual and semantic embedding; and it can identify visual
objects using both labeled image data as well as semantic
information gleaned from unannotated text. ConSE [66] constructed the image embedding approach by mapping images
into the semantic embedding space via convex combination of
the class label embedding vectors. Both DeViSE and ConSE
are evaluated on large-scale datasets, – ImageNet (ILSVRC)
2012 1K and ImageNet 2011 21K dataset.
To combine the visual and textual branches in the deep embedding, different loss functions can be considered, including
margin-based losses [70], [103], or Euclidean distance loss
[113], or least square loss [84]. Zhang et al. [114] employed
the visual space as the embedding space and proposed an endto-end deep learning architecture for zero-shot recognition.
Their networks have two branches: visual encoding branch
which uses convolutional neural network to encode the input
image as a feature vector, and the semantic embedding branch
which encodes the input semantic representation vector of each
class which the corresponding image belonging to.
B. Recognition Models in the Embedding Space
Once the embedding model is learned, the testing instances
can be projected into this embedding space. The recognition
can be carried out by using different recognition models. The
most common used one is nearest neighbour classifier which
classify the testing instances by assigning the class label in
term of the nearest distances of the class prototypes against
the projections of testing instances in the embedding space. Fu
et al. [13] proposed semi-latent zero-shot learning algorithm
to update the class prototypes by one step self-training.
Manifold information can be used in the recognition models
in the embedding space. Fu et al. [115] proposed a hypergraph structure in their multi-view embedding space; and
zero-shot recognition can be addressed by label propagation
from unseen prototype instances to unseen testing instances.
Changpinyo et al. [116] synthesized classifiers in the embedding space for zero-shot recognition. For multi-label zeroshot learning, the recognition models have to consider the cooccurrence/correlations of different semantic labels [63], [69],
[117].
Latent SVM structure has also been used as the recognition
models [118], [20]. Wang et al. [118] treated the object
attributes as latent variables and learnt the correlations of
attributes through an undirected graphical model. Hwang
et al. [20] utilized a kernelized multi-task feature learning
framework to learn the sharing features between objects and
7
Zebra
Prototype
The same ‘hasTail’ attribute
different visual appearance
Pig
Prototype
Pig
Prototype
(a) visual space
(b) attribute space
(c) multi-view embedding space
Fig. 1. Illustrating projection domain shift problem. Zero-shot prototypes are
annotated as red stars and predicted semantic attribute projections shown in
blue. Both Pig and Zebra share the same ‘hasTail’ attribute yet with very
different visual appearance of ‘Tail’. The figure comes from [115].
their attributes. Additionally, Long et al. [119] employed
the attributes to synthesize unseen visual features at training
stage; and thus zero-shot recognition can be solved by the
conventional supervised classification models.
the presence of ‘universal’ neighbors, or hubs, in the space.
Radovanovic et al. [124] was the first to study the hubness
problem; in [124] a hypothesis is made that hubness is an
inherent property of data distributions in the high dimensional
vector space. Nevertheless, Low et al. [125] challenged this
hypothesis and showed the evidence that hubness is rather
a boundary effect or, more generally, an effect of a density
gradient in the process of data generation. Interestingly, their
experiments showed that the hubness phenomenon can also
occur in low-dimensional data.
While causes for hubness are still under investigation, recent
works [126], [127] noticed that the regression based zero-shot
learning methods do suffer from this problem. To alleviate this
problem, Dinu et al. [126] utilized the global distribution of
feature instances of unseen data, i.e., in a transductive manner.
In contrast, Yutaro et al. [127] addressed this problem in an
inductive way by embedding the class prototypes into a visual
feature space.
V. B EYOND Z ERO - SHOT R ECOGNITION
C. Problems in Zero-shot Recognition
A. Generalized Zero-shot Recognition and Open-set Recognition
There are two intrinsic problems in zero-shot recognition,
namely projection domain shift problem (Sec. IV-C1) and
hubness problem (Sec. IV-C2).
1) Projection Domain Shift Problems: Projection domain
shift problem in zero-shot recognition was first identified by
Fu et al. [115]. This problem can be explained as follows:
since the source and target datasets have different classes, the
underlying data distribution of these classes may also differ.
The projection functions learned on the source dataset, from
visual space to the embedding space, without any adaptation
to the target dataset, will cause an unknown shift/bias. Figure
1 from [115] gives a more intuitive illustration of this problem.
It plots the 85D attribute space representation spanned by
feature projections which is learned from source data, and
class prototypes which are 85D binary attribute vectors. Zebra
and Pig are one of auxiliary and target classes respectively;
and the same ’hasTail’ semantic attribute means very different
visual appearance for Pig and Zebra. In the attribute space,
directly using the projection functions learned from source
datasets (e.g., Zebra) on the target datasets (e.g., Pig) will lead
to a large discrepancy between the class prototype of the target
class and the predicted semantic attribute projections.
To alleviate this problem, the transductive learning based
approaches were proposed, to utilize the manifold information
of the instances from unseen classes [115], [120], [121], [122],
[123], [102]. Nevertheless, the transductive setting assumes
that all the testing data can be accessed at once, which
obviously is invalid if the new unseen classes appear dynamically and unavailable before learning models. Thus inductive
learning base approaches [120], [116], [92], [105], [122] have
also been studied and these methods usually enforce other
additional constraints or information from the training data.
2) Hubness problem: The hubness problem is another
interesting phenomenon that may be observed in zero-shot
recognition. Essentially, hubness problem can be described as
In conventional supervised learning tasks, it is taken for
granted that the algorithms should take the form of “closed
set” where all testing classes should are known at training
time. The zero-shot recognition, in contrast, assumes that the
source and target classes cannot be mixed; and that the testing
data only coming from the unseen classes. This assumption,
of course, greatly and unrealistically simplifies the recognition
tasks. To relax the settings of zero-shot recognition and
investigate recognition tasks in a more generic setting, there
are several tasks advocated beyond the conventional zero-shot
recognition. In particular, generalized zero-shot recognition
[128] and open set recognition tasks have been discussed
recently [129], [130], [131], [132].
The generalized zero-shot recognition proposed in [128]
broke the restricted nature of conventional zero-shot recognition and also included the training classes among the testing
data. Chao et al. [128] showed that it is nontrivial and
ineffective to directly extend the current zero-shot learning
approaches to solve the generalized zero-shot recognition.
Such a generalized setting, due to the more practical nature, is
recommended as the evaluation settings for zero-shot recognition tasks [133].
Open-set recognition, in contrast, has been developed independently of zero-shot recognition. Initially, open set recognition aimed at breaking the limitation of “closed set” recognition setup. Specifically, the task of open set recognition tries
to identify the class name of an image from a very large set of
classes, which includes but is not limited to training classes.
The open set recognition can be roughly divided into two subgroups.
1) Conventional open set recognition: First formulated in
[134], [135], [130], [129], the conventional open set recognition only identifies whether the testing images come from
the training classes or some unseen classes. This category of
methods do not explicitly predict from which out of unseen
8
classes the testing instance, from the unseen classes, belongs
to. In such a setting, the conventional open set recognition is
also known as incremental learning [136], [137], [138].
2) Generalized open set recognition: The key difference
from the conventional open set recognition is that the generalized open set recognition also needs to explicitly predict
the semantic meaning (class) of testing instances even from
the unseen novel classes. This task was first defined and
evaluated in [131], [132] on the tasks of object categorization.
The generalized open set recognition can be taken as a most
general version of zero-shot recognition, where the classifiers
are trained from training instances of limited training classes,
whilst the learned classifiers are required to classify the testing
instances from a very large set of open vocabulary, say, 310 K
class vocabulary in [131], [132]. Conceptually similar, there
are vast variants of generalized open-set recognition tasks
which have been studied in other research community such
as, open-vocabulary object retrieval [139], [140], open-world
person re-identification [141] or searching targets [135], open
vocabulary scene parsing [142].
B. One-shot recognition
A closely-related problem to zero-shot learning is one-shot
or few-shot learning problem – instead of/apart from having
only textual description of the new classes, one-shot learning
assumes that there are one or few training samples for each
class. Similar to zero-shot recognition, one-shot recognition
is inspired by fact that humans are able to learn new object
categories from one or very few examples [143], [144].
Existing one-shot learning approaches can be divided into two
groups: the direct supervised learning based approaches and
the transfer learning based approaches.
1) Direct Supervised Learning-based Approaches: Early
approaches do not assume that there exist a set of auxiliary
classes which are related and/or have ample training samples
whereby transferable knowledge can be extracted to compensate for the lack of training samples. Instead, the target classes
are used to trained a standard classifier using supervised learning. The simplest method is to employ nonparametric models
such as kNN which are not restricted by the number of training
samples. However, without any learning, the distance metric
used for kNN is often inaccurate. To overcome this problem,
metric embedding can be learned and then used for kNN classification [145]. Other approaches attempt to synthesize more
training samples to augment the small training dataset [146],
[147], [148], [144]. However, without knowledge transfer from
other classes, the performance of direct supervised learning
based approaches is typically weak. Importantly, these models
cannot meet the requirement of lifelong learning, that is, when
new unseen classes are added, the learned classifier should still
be able to recognize the seen existing classes.
2) Transfer Learning-based One-shot Recognition: This
category of approaches follow a similar setting to zero-shot
learning, that is, they assume that an auxiliary set of training
data from different classes exist. They explore the paradigm of
learning to learn [16] or meta-learning [149] and aim to transfer knowledge from the auxiliary dataset to the target dataset
with one or few examples per class. These approaches differ in
(i) what knowledge is transferred and (ii) how the knowledge
is represented. Specifically, the knowledge can be extracted
and shared in the form of model prior in a generative model
[150], [151], [152], features [153], [154], [155], [156], [157],
[158], semantic attributes [13], [9], [121], [62], or contextual
information [159]. Many of these approaches take a similar
strategy as the existing zero-shot learning approaches and
transfer knowledge via a shared embedding space. Embedding
space can typically be formulated using neural networks (e.g.,
siamese network [160], [161]), discriminative (e.g., Support
Vector Regressors (SVR) [17], [9], [162]), metric learning
[163], [164], or kernel embedding [165], [154] methods.
Particularly, one of most common embedding ways is semantic
embedding which is normally explored by projecting the visual
features and semantic entities into a common new space. Such
projections can take various forms with corresponding loss
functions, such as SJE [100], WSABIE [166], ALE [93],
DeViSE [70], and CCA [102].
More recently deep meta-learning has received increasing
attention for few-shot learning [167], [168], [161], [169], [52],
[141], [170], [171], [144]. Wang et al. [172], [173] proposed
the idea of one-shot adaptation by automatically learning a
generic, category agnostic transformation from models learned
from few samples to models learned from large enough sample
sets. A model-agnostic meta-learning framework is proposed
by Finn et al. [174] which trains a deep model from the
auxiliary dataset with the objective that the learned model
can be effectively updated/fine-tuned on the new classes with
one or few gradient steps. Note that similar to the generalised
zero-shot learning setting, recently the problem of adding new
classes to a deep neural network whilst keeping the ability to
recognise the old classes have been attempted [175]. However,
the problem of lifelong learning and progressively adding new
classes with few-shot remains an unsolved problem.
VI. DATASETS IN Z ERO - SHOT R ECOGNITION
This section summarizes the datasets used for zero-shot
recognition. Recently with the increasing number of proposed
zero-shot recognition algorithms, Xian et al. [133] compared
and analyzed a significant number of the state-of-the-art methods in depth and they defined a new benchmark by unifying
both the evaluation protocols and data splits. The details of
these datasets are listed in Tab. II.
A. Standard Datasets
1) Animal with Attribute (AwA) dataset [32]: AwA consists
of the 50 Osher-son/Kemp animal category images collected
online. There are 30, 475 images with at least 92 examples of
each class. Seven different feature types are provided: RGB
color histograms, SIFT [176], rgSIFT [177], PHOG [178],
SURF [179], local self-similarity histograms [180] and DeCaf [181]. The AwA dataset defines 50 classes of animals,
and 85 associated attributes (such as furry, and has claws). For
the consistent evaluation of attribute-based object classification
methods, the AwA dataset defined 10 test classes: chimpanzee,
giant panda, hippopotamus, humpback whale, leopard, pig,
9
raccoon, rat, seal. The 6, 180 images of those classes are
taken as the test data, whereas the 24, 295 images of the
remaining 40 classes can be used for training. Since the images
in AwA are not available under a public license, Xian et
al. [133] introduced another new zero-shot learning dataset
– Animals with Attributes 2 (AWA2) dataset with 37,322
publicly licensed and released images from the same 50 classes
and 85 attributes as AwA.
2) aPascal-aYahoo dataset [17]: aPascal-aYahoo has a
12,695-image subset of the PASCAL VOC 2008 data set
with 20 object classes (aPascal); and 2,644 images that were
collected using the Yahoo image search engine (aYahoo) of 12
object classes. Each image in this data set has been annotated
with 64 binary attributes that characterize the visible objects.
3) CUB-200-2011 dataset [182]: CUB-200-2011 contains
11, 788 images of 200 bird classes. This is a more challenging
dataset than AwA – it is designed for fine-grained recognition
and has more classes but fewer images. All images are
annotated with bounding boxes, part locations, and attribute
labels. Images and annotations were filtered by multiple users
of Amazon Mechanical Turk. CUB-200-2011 is used as the
benchmarks dataset for multi-class categorization and part
localization. Each class is annotated with 312 binary attributes
derived from the bird species ontology. A typical setting is to
use 150 classes as auxiliary data, holding out 50 as target data,
which is the setting adopted in Akata et al. [93].
4) Outdoor Scene Recognition (OSR) Dataset [183]: OSR
consists of 2, 688 images from 8 categories and 6 attributes
(‘openness’, ‘natrual’, etc.) and an average 426 labelled pairs
for each attribute from 240 training images. Graphs constructed are thus extremely sparse. Pairwise attribute annotation was collected by AMT (Kovashka et al. [22]). Each
pair was labelled by 5 workers to average the comparisons by
majority voting. Each image also belongs to a scene type.
5) Public Figure Face Database (PubFig) [8]: PubFig is
a large face dataset of 58,797 images of 200 people collected
from the internet. Parikh et al. [21] selected a subset of PubFig
consisting of 772 images from 8 people and 11 attributes
(‘smiling’, ‘round face’, etc.). We annotate this subset as
PubFig-sub. The pairwise attribute annotation was collected
by Amazon Mechanical Turk [22]. Each pair was labelled
by 5 workers. A total of 241 training images for PubFig-sub
respectively were labelled. The average number of compared
pairs per attribute were 418.
6) SUN attribute dataset [184]: This is a subset of the SUN
Database [185] for fine-grained scene categorization and it has
14, 340 images from 717 classes (20 images per class). Each
image is annotated with 102 binary attributes that describe
the scenes’ material and surface properties as well as lighting
conditions, functions, affordances, and general image layout.
7) Unstructured Social Activity Attribute (USAA) dataset
[11]: USAA is the first benchmark video attribute dataset for
social activity video classification and annotation. The groundtruth attributes are annotated for 8 semantic class videos of
Columbia Consumer Video (CCV) dataset [186], and select
100 videos per-class for training and testing respectively.
These classes were selected as the most complex social
group activities. By referring to the existing work on video
ontology [187], [186], the 69 attributes can be divided into
five broad classes: actions, objects, scenes, sounds, and camera
movement. Directly using the ground-truth attributes as input
to a SVM, the videos can come with 86.9% classification
accuracy. This illustrates the challenge of USAA dataset: while
the attributes are informative, there is sufficient intra-class
variability in the attribute-space, and even perfect knowledge
of the instance-level attributes is also insufficient for perfect
classification.
8) ImageNet datasets [62], [61], [131], [116]: ImageNet
has been used in several different papers with relatively
different settings. The original ImageNet dataset has been
proposed in [188]. The full set of ImageNet contains over 15
million labeled high-resolution images belonging to roughly
22,000 categories and labelled by human annotators using
Amazon’s Mechanical Turk (AMT) crowd-sourcing tool. Starting in 2010, as part of the Pascal Visual Object Challenge, an
annual competition called the ImageNet Large-Scale Visual
Recognition Challenge (ILSVRC) has been held. ILSVRC
uses a subset of ImageNet with roughly 1,000 images in each
of 1,000 categories. In [121], [61], Robhrbach et al. split the
ILSVRC 2010 data into 800/200 classes for source/target data.
In [131], Fu et al. employed the training data of ILSVRC 2012
as the source data; and the testing part of ILSVRC 2012 as
well as the data of ILSVRC 2010 as the target data. The full
sized ImageNet data has been used in [116], [70], [66].
9) Oxford 102 Flower dataset [189]: Oxford 102 is a
collection of 102 groups of flowers each with 40 to 256
flower images, and total 8,189 images in total. The flowers
were chosen from the common flower species in the United
Kingdom. Elhoseiny et al. [83] generated textual descriptions
for each class of this dataset.
10) UCF101 dataset [190]: UCF101 is another popular
benchmark for human action recognition in videos, which
consists of 13, 320 video clips (27 hours in total) with 101
annotated classes. More recently, the THUMOS-2014 Action
Recognition Challenge [191] created a benchmark by extending upon the UCF-101 dataset (used as the training set).
Additional videos were collected from the Internet, including
2, 500 background videos, 1, 000 validation and 1, 574 test
videos.
11) Fudan-Columbia Video Dataset (FCVID) [192]:
FCVID contains 91, 223 web videos annotated manually into
239 categories. Categories cover a wide range of topics (not
only activities), such as social events (e.g., tailgate party),
procedural events (e.g., making cake), object appearances (e.g.,
panda) and scenic videos (e.g., beach). Standard split consists
of 45, 611 videos for training and 45, 612 videos for testing.
12) ActivityNet dataset [193]: ActivityNet is another largescale video dataset for human activity recognition and understanding and released in 2015. It consisted of 27,801
video clips annotated into 203 activity classes, totaling 849
hours of video. Comparing with existing dataset, ActivityNet
has more fine-grained action categories (e.g., “drinking beer”
and “drinking coffee”). ActivityNet had the settings of both
trimmed and untrimmed videos of its classes.
10
A
B
C
Dataset
AwA
aPascal-aYahoo
PubFig
PubFig-sub
OSR
ImageNet
ILSVRC 2010
ILSVRC 2012
Oxford 102 Flower
CUB-200-2011
SUN-attribute
USAA
UCF101
ActivityNet
FCVID
# instances
#classes
30475
50
15339
32
58,797
200
772
8
2688
8
15 million
22000
1.2 million
1000
1.2 million
1000
8189
102
11788
200
14340
717
1600
8
13320
101
27801
203
91223
239
TABLE II
#attribute
85
64
–
11
6
–
–
–
–
312
102
69
–
–
–
Annotation Level
per class
per image
per image
per image pairs
per image pairs
per image
per image
per image
–
per class
per image
per video
per video
per video
per video
DATASETS IN ZERO - SHOT RECOGNITION . T HE DATASETS ARE DIVIDED INTO THREE GROUPS : GENERAL IMAGE CLASSIFICATION (A), FINE - GRAINED
IMAGE CLASSIFICATION (B) AND VIDEO CLASSIFICATION DATASETS (C).
B. Discussion of Datasets.
In Tab. II, we roughly divide all the datasets into three
groups: general image classification, fine-grained image classification and video classification datasets. These datasets have
been employed widely as the benchmark datasets in many
previous works. However, we believe that when making a
comparison with the other existing methods on these datasets,
there are several issues that should be discussed.
1) Features: With the renaissance of deep convolutional
neural networks, deep features of images/videos have been
used for zero-shot recognition. Note that different types of
deep features (e.g., Overfeat [194], VGG-19[195], or ResNet
[196]) have varying level of semantic abstraction and representation ability; and even the same type of deep features,
if fine-tuned on different dataset and with slightly different
parameters, will also have different representative ability. Thus
it should be obvious, without using the same type of features,
it is not possible to conduct a fair comparisons among different
methods and draw any meaningful conclusion. Importantly
it is possible that the improved performance of one zero
shot recognition could be largely attributed to the better deep
features used.
2) Auxiliary data: As mentioned, zero-shot recognition can
be formulated in a transfer learning setting. The size and
quality of auxiliary data can be very important for the overall
performance of zero-shot recognition. Note that these auxiliary
data do not only include the auxiliary source image/video
dataset, but also refer to the data to extract/train the concept
ontology, or semantic word vectors. For example, the semantic word vectors trained on large-scale linguistic articles, in
general, are better semantically distributed than those trained
on small sized linguistic corpus. Similarly, GloVe [197] is
reported to be better than the skip-gram and CBOW models
[88]. Therefore, to make a fair comparison with existing
works, another important factor is to use the same set of
auxiliary data.
3) Evaluation: For many datasets, there is no agreed
source/target splits for zero-shot evaluation. Xian et al. [133]
suggested a new benchmark by unifying both the evaluation
protocols and data splits.
VII. F UTURE R ESEARCH D IRECTIONS
1) More Generalized and Realistic Setting: From the detailed review of existing zero-shot learning methods, it is clear
that overall the existing efforts have been focused on a rather
restrictive and impractical setting: classification is required for
new object classes only and the new unseen classes, though
having no training sample present, are assumed to be known.
In reality, one wants to progressively add new classes to the
existing classes. Importantly, this needs to be achieved without
jeopardizing the ability of the model to recognize existing seen
classes. Furthermore, we cannot assume that the new samples
will only come from a set of known unseen classes. Rather,
they can only be assumed to belong to either existing seen
classes, known unseen classes, or unknown unseen classes. We
therefore foresee a more generalized setting will be adopted
by the future zero-shot learning work.
2) Combining Zero-shot with Few-shot Learning: As mentioned earlier, the problems of zero-shot and few-shot learning
are closely related and as a result, many existing methods
use the same or similar models. However, it is somewhat
surprising to note that no serious efforts have been taken to
address these two problems jointly. In particular, zero-shot
learning would typically not consider the possibility of having
few training samples, while few-shot learning ignores the fact
that the textual description/human knowledge about the new
class is always there to be exploited. A few existing zeroshot learning methods [131], [13], [198], [85] have included
few-shot learning experiments. However, they typically use a
naive kNN approach, that is, each class prototype is treated as
a training sample and together with the k-shot, this becomes a
k+1-shot recognition problem. However, as shown by existing
zero-shot learning methods [115], the prototype is worth far
more that one training sample; it thus should be treated
differently. We thus expect a future direction on extending
the existing few-shot learning methods by incorporating the
prototype as a ‘super’-shot to improve the model learning.
3) Beyond object categories: So far the current zero-shot
learning efforts are limited to recognizing object categories.
However, visual concepts can have far complicated relationships than object categories. In particular beyond ob-
11
jects/nouns, attributes/adjectives are important visual concepts.
When combined with objects, the same attribute often has
different meaning, e.g., the concept of ‘yellow’ in yellow
face and a yellow banana clearly differs. Zero-shot learning
attributes with associated objects is thus an interesting future
research direction.
4) Curriculum learning: In a lifelong learning setting, a
model will incrementally learn to recognise new classes whilst
keep the capacity for existing classes. A related problem is thus
how to select the more suitable new classes to learn given the
existing classes. It has been shown that [138], [199], [200] the
sequence of adding different classes have a clear impact on the
model performance. It is therefore useful to investigate how
to incorporate the curriculum learning principles in designing
a zero-shot learning strategy.
VIII. C ONCLUSION
In this paper, we have reviewed the recent advances in zero
shot recognition. Firstly different types of semantic representations are examined and compared; the models used in zero shot
learning have also been investigated. Next, beyond zero shot
recognition, one-shot and open set recognition are identified as
two very important related topics and thus reviewed. Finally,
the common used datasets in zero-shot recognition have been
reviewed with a number of issues in existing evaluations of
zero-shot recognition methods discussed. We also point out a
number of research direction which we believe will the focus
of the future zero-shot recognition studies.
Acknowledgments. This work is supported in part by
two grants from NSF China (#61702108, #61622204,
#61572134), and an European FP7 project (PIRSESGA2013 − 612652). Yanwei Fu is supported by The
Program for Professor of Special Appointment (Eastern
Scholar) at Shanghai Institutions of Higher Learning.
R EFERENCES
[1] I. Biederman, “Recognition by components - a theory of human image
understanding,” Psychological Review, 1987. I
[2] X. Chen, A. Shrivastava, and A. Gupta, “NEIL: Extracting Visual
Knowledge from Web Data,” in IEEE International Conference on
Computer Vision, 2013. I
[3] A. Pentina and C. H. Lampert, “A PAC-bayesian bound for lifelong
learning,” in International Conference on Machine Learning, 2014. I
[4] S. Thrun and T. M. Mitchell, “Lifelong robot learning,” Robotics and
Autonomous Systems, 1995. I
[5] S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE Transactions on Data and Knowledge Engineering, vol. 22, no. 10, pp. 1345–
1359, 2010. I
[6] V. Patel, R. Gopalan, R. Li, and R. Chellappa, “Visual domain
adaptation: A survey of recent advances,” IEEE Signal Processing
Magazine (SPM), 2015. I
[7] M. Palatucci, G. Hinton, D. Pomerleau, and T. M. Mitchell, “Zero-shot
learning with semantic output codes,” in NIPS, 2009. II, III-A
[8] N. Kumar, A. C. Berg, P. N. Belhumeur, and S. K. Nayar, “Attribute
and simile classifiers for face verification,” in ICCV, 2009. II, VI-A5
[9] C. H. Lampert, H. Nickisch, and S. Harmeling, “Attribute-based
classification for zero-shot visual object categorization,” IEEE TPAMI,
2013. II, III-A, III-A1, III-A2, IV-A1, V-B2
[10] B. Xu, Y. Fu, Y.-G. Jiang, B. Li, and L. Sigal, “Video emotion
recognition with transferred deep feature encodings,” in ICMR, 2016.
II, III-A2, III-B2
[11] Y. Fu, T. Hospedales, T. Xiang, and S. Gong, “Attribute learning for
understanding unstructured social activity,” in ECCV, 2012. II, III-A,
III-A1, III-A3, III-A2, IV-A1, VI-A7
[12] J. Liu, B. Kuipers, and S. Savarese, “Recognizing human actions
by attributes,” in IEEE Conference on Computer Vision and Pattern
Recognition, 2011. II, III-A, III-A3, III-A2
[13] Y. Fu, T. M. Hospedales, T. Xiang, and S. Gong, “Learning multimodal latent attributes,” IEEE TPAMI, 2013. II, III-A2, III-A3, IV-A1,
IV-B, V-B2, VII-2
[14] J. Blitzer, D. P. Foster, and S. M. Kakade, “Zero-shot domain adaptation: A multi-view approach,” TTI-TR-2009-1, Tech. Rep., 2009. II
[15] R. Socher, M. Ganjoo, H. Sridhar, O. Bastani, C. D. Manning, and
A. Y. Ng, “Zero-shot learning through cross-modal transfer,” in NIPS,
2013. II, III-A2, III-B2
[16] S. Thrun, Learning To Learn: Introduction.
Kluwer Academic
Publishers, 1996. II, V-B2
[17] A. Farhadi, I. Endres, D. Hoiem, and D. Forsyth, “Describing objects
by their attributes,” in CVPR, 2009. III-A, III-A3, III-A2, V-B2, VI-A2
[18] X. Wang and T. Zhang, “Clothes search in consumer photos via color
matching and attribute learning,” in ACM International Conference
on Multimedia, 2011. [Online]. Available: http://doi.acm.org/10.1145/
2072298.2072013 III-A, III-A1, III-A2
[19] R. Salakhutdinov, A. Torralba, and J. Tenenbaum, “Learning to share
visual appearance for multiclass object detection,” in IEEE Conference
on Computer Vision and Pattern Recognition, 2011. III-A
[20] S. J. Hwang, F. Sha, and K. Grauman, “Sharing features between
objects and their attributes,” in IEEE Conference on Computer Vision
and Pattern Recognition, 2011. III-A, IV-A2, IV-B
[21] D. Parikh and K. Grauman, “Relative attributes,” in ICCV, 2011. III-A,
III-A2, VI-A5
[22] A. Kovashka, D. Parikh, and K. Grauman, “WhittleSearch: Image
search with relative attribute feedback,” in IEEE Conference on Computer Vision and Pattern Recognition, 2012. III-A, III-A1, III-A2,
VI-A4, VI-A5
[23] T. L. Berg, A. C. Berg, and J. Shih, “Automatic attribute discovery
and characterization from noisy web data,” in European Conference
on Computer Vision, 2010. III-A
[24] Y. Fu, T. M. Hospedales, J. Xiong, T. Xiang, S. Gong, Y. Yao,
and Y. Wang, “Robust estimation of subjective visual properties from
crowdsourced pairwise labels,” IEEE TPAMI, 2016. III-A, III-A2
[25] M. Gygli, H. Grabner, H. Riemenschneider, F. Nater, and L. V. Gool,
“The interestingness of images,” in IEEE International Conference on
Computer Vision, 2013. III-A
[26] Y.-G. Jiang, YanranWang, R. Feng, X. Xue, Y. Zheng, and H. Yang,
“Understanding and predicting interestingness of videos,” in AAAI
Conference on Artificial Intelligence, 2013. III-A
[27] P. Isola, D. Parikh, A. Torralba, and A. Oliva, “Understanding the
intrinsic memorability of images,” in Neural Information Processing
Systems, 2011. III-A
[28] P. Isola, J. Xiao, A. Torralba, and A. Oliva, “What makes an image
memorable?” in IEEE Conference on Computer Vision and Pattern
Recognition, 2011. III-A
[29] S. Dhar, V. Ordonez, and T. L. Berg, “High level describable attributes
for predicting aesthetics and interestingness,” in IEEE Conference on
Computer Vision and Pattern Recognition, 2011. III-A
[30] Y. Fu, G. Guo, and T. Huang, “Age synthesis and estimation via
faces: A survey,” IEEE Transactions on Pattern Analysis and Machine
Intelligence, 2010. III-A
[31] K. Chen, S. Gong, T. Xiang, and C. C. Loy, “Cumulative attribute
space for age and crowd density estimation,” in IEEE Conference on
Computer Vision and Pattern Recognition, 2013. III-A
[32] C. H. Lampert, H. Nickisch, and S. Harmeling, “Learning to detect
unseen object classes by between-class attribute transfer,” in CVPR,
2009. III-A1, III-A2, IV-A1, VI-A1
[33] E. M. Rudd, M. Gunther, and T. E. Boult, “Moon:a mixed objective
optimization network for the recognition of facial attributes,” in ECCV,
2016. III-A1, III-A2
[34] E. Rudd, M. Günther, and T. Boult, “Moon: A mixed objective
optimization network for the recognition of facial attributes,” arXiv
preprint arXiv:1603.07027, 2016. III-A1, III-A2
[35] J. Wang, Y. Cheng, and R. S. Feris, “Walk and learn: Facial attribute
representation learning from egocentric video and contextual data,” in
CVPR, 2016. III-A1, III-A2
[36] A. Datta, R. Feris, and D. Vaquero, “Hierarchical ranking of facial
attributes,” in IEEE International Conference on Automatic Face &
Gesture Recognition, 2011. III-A1, III-A2
12
[37] M. Ehrlich, T. J. Shields, T. Almaev, and M. R. Amer, “Facial attributes
classification using multi-task representation learning,” in Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition
Workshops, 2016, pp. 47–55. III-A1, III-A2
[38] R. Jafri and H. R. Arabnia, “A survey of face recognition techniques,”
in Journal of Information Processing Systems, 2009. III-A1, III-A2
[39] Z. Wang, K. He, Y. Fu, R. Feng, Y.-G. Jiang, and X. Xue, “Multitask deep neural network for joint face recognition and facial attribute
prediction,” in ACM ICMR, 2017. III-A1, III-A2
[40] A. Argyriou, T. Evgeniou, and M. Pontil, “Convex multi-task feature
learning,” in ACM ICMR, 2007. III-A1, III-A2
[41] D. F. Fouhey, A. Gupta, and A. Zisserman, “Understanding higherorder shape via 3d shape attributes,” in IEEE TPAMI, 2017. III-A1
[42] D. Vaquero, R. Feris, D. Tran, L. Brown, A. Hampapur, and M. Turk,
“Attribute-based people search in surveillance environments,” in IEEE
Workshop on Applications of Computer Vision (WACV), dec. 2009, pp.
1 –8. III-A1, III-A2
[43] G. Wang and D. Forsyth, “Joint learning of visual attributes, object
classes and visual saliency,” in IEEE International Conference on
Computer Vision, 2009, pp. 537–544. III-A1, III-A2
[44] V. Ferrari and A. Zisserman, “Learning visual attributes,” in Neural
Information Processing Systems, Dec. 2007. III-A1, III-A2, IV-A1
[45] A. Shrivastava, S. Singh, and A. Gupta, “Constrained semi-supervised
learning via attributes and comparative attributes,” in European Conference on Computer Vision, 2012. III-A2
[46] A. Biswas and D. Parikh, “Simultaneous active learning of classifiers
and attributes via relative feedback,” in IEEE Conference on Computer
Vision and Pattern Recognition, 2013. III-A2
[47] A. Parkash and D. Parikh, “Attributes for classifier feedback,” in
European Conference on Computer Vision, 2012. III-A2
[48] K. K. Singh and Y. J. Lee, “End-to-end localization and ranking for
relative attributes,” in ECCV, 2016. III-A2
[49] M. Jaderberg, K. Simonyan, A. Zisserman et al., “Spatial transformer
networks,” in Advances in Neural Information Processing Systems,
2015, pp. 2017–2025. III-A2
[50] D. Parikh and K. Grauman, “Interactively building a discriminative
vocabulary of nameable attributes,” in IEEE Conference on Computer
Vision and Pattern Recognition, 2011. III-A3, III-A2
[51] J. Tang, S. Yan, R. Hong, G.-J. Qi, and T.-S. Chua, “Inferring semantic
concepts from community-contributed images and noisy tags,” in ACM
International Conference on Multimedia, 2009. [Online]. Available:
http://doi.acm.org/10.1145/1631272.1631305 III-A3, III-A2
[52] A. Habibian, T. Mensink, and C. Snoek, “Videostory: A new multimedia embedding for few-example recognition and translation of events,”
in ACM MM, 2014. III-A2, III-A3, V-B2
[53] A. Hauptmann, R. Yan, W.-H. Lin, M. Christel, and H. Wactlar, “Can
high-level concepts fill the semantic gap in video retrieval? a case study
with broadcast news,” IEEE Transactions on Multimedia, vol. 9, no. 5,
pp. 958 –966, aug. 2007. III-A2, III-A4
[54] C. G. M. Snoek, B. Huurnink, L. Hollink, M. de Rijke, G. Schreiber,
and M. Worring, “Adding semantics to detectors for video retrieval,”
IEEE Transactions on Multimedia, vol. 9, pp. 975–986, 2007. III-A2,
III-A4
[55] G. Toderici, H. Aradhye, M. Pasca, L. Sbaiz, and J. Yagnik, “Finding
meaning on youtube: Tag recommendation and category discovery,” in
IEEE Conference on Computer Vision and Pattern Recognition, 2010,
pp. 3447–3454. III-A2, III-A4
[56] Z. Wu, Y. Fu, Y.-G. Jiang, and L. Sigal, “Harnessing object and scene
semantics for large-scale video understanding,” in CVPR, 2016. III-A2,
III-A4
[57] M. Jain, J. C. van Gemert, T. Mensink, and C. G. M. Snoek,
“Objects2action: Classifying and localizing actions without any video
example,” in ICCV, 2015. III-A2, III-A4
[58] J. Tang, X.-S. Hua, M. Wang, Z. Gu, G.-J. Qi, and X. Wu, “Correlative
linear neighborhood propagation for video annotation,” IEEE Transactions on Systems, Man, and Cybernetics, Part B, vol. 39, no. 2, pp.
409–416, 2009. III-A2, III-A4
[59] G.-J. Qi, X.-S. Hua, Y. Rui, J. Tang, T. Mei, and H.-J. Zhang,
“Correlative multi-label video annotation,” in ACM International
Conference on Multimedia, 2007. [Online]. Available: http://doi.acm.
org/10.1145/1291233.1291245 III-A2
[60] R. Fergus, H. Bernal, Y. Weiss, and A. Torralba, “Semantic label
sharing for learning with many categories,” in European Conference
on Computer Vision, 2010. III-A2, III-B1
[61] M. Rohrbach, M. Stark, and B. Schiele, “Evaluating knowledge transfer
and zero-shot learning in a large-scale setting,” in CVPR, 2012. III-A2,
III-B, III-B1, IV-A1, IV-A1, VI-A8
[62] M. Rohrbach, M. Stark, G. Szarvas, I. Gurevych, and B. Schiele, “What
helps where – and why? semantic relatedness for knowledge transfer,”
in CVPR, 2010. III-A2, III-B, III-B1, V-B2, VI-A8
[63] T. Mensink, E. Gavves, and C. G. Snoek, “Costa: Co-occurrence
statistics for zero-shot classification,” in IEEE Conference on Computer
Vision and Pattern Recognition, 2014. III-A2, III-B1, IV-B
[64] C. Gan, Y. Yang, L. Zhu, and Y. Zhuang, “Recognizing an action using
its name: A knowledge-based approach,” IJCV, 2016. III-A2, III-A4
[65] C. Gan, M. Lin, Y. Yang, G. de Melo, and A. G. Hauptmann, “Concepts
not alone: Exploring pairwise relationships for zero-shot video activity
recognition,” in AAAI, 2016. III-A2, III-A4
[66] M. Norouzi, T. Mikolov, S. Bengio, Y. Singer, J. Shlens, A. Frome,
G. S. Corrado, and J. Dean, “Zero-shot learning by convex combination
of semantic embeddings,” ICLR, 2014. III-A2, III-B2, IV-A4, VI-A8
[67] Z. Zhang and V. Saligrama, “Zero-shot learning via joint latent similarity embedding,” in CVPR, 2016. III-A2, III-B2, IV-A2
[68] ——, “Zero-shot recognition via structured prediction,” in ECCV,
2016. III-A2, III-B2
[69] Y. Fu, Y. Yang, T. Hospedales, T. Xiang, and S. Gong, “Transductive
multi-label zero-shot learning,” in British Machine Vision Conference,
2014. III-A2, III-B2, IV-B
[70] A. Frome, G. S. Corrado, J. Shlens, S. Bengio, J. Dean, M. Ranzato,
and T. Mikolov, “DeViSE: A deep visual-semantic embedding model,”
in NIPS, 2013. III-A2, III-B2, IV-A4, V-B2, VI-A8
[71] E. H. Huang, R. Socher, C. D. Manning, and A. Y. Ng, “Improving
word representations via global context and multiple word prototypes,”
in Association for Computational Linguistics 2012 Conference, 2012.
III-A2, III-B2
[72] B. Xu, Y. Fu, Y.-G. Jiang, B. Li, and L. Sigal, “Heterogeneous
knowledge transfer in video emotion recognition, attribution and summarization,” IEEE TAC, 2016. III-A2, III-B2
[73] A. Sorokin and D. Forsyth, “Utility data annotation with amazon
mechanical turk,” in IEEE Conference on Computer Vision and Pattern
Recognition Workshops, 2008. III-A3
[74] J. Qin, Y. Wang, L. Liu, J. Chen, and L. Shao, “Beyond semantic
attributes: Discrete latent attributes learning for zero-shot recognition,”
IEEE Signal Processing Letters, vol. 23, no. 11, pp. 1667–1671, 2016.
III-A3
[75] X. Chang, Y. Yang, G. Long, C. Zhang, and A. Hauptmann, “Dynamic
concept composition for zero-example event detection,” in AAAI, 2016.
III-A4
[76] X. Chang, Y. Yang, A. Hauptmann, E. P. Xing, and Y. Yu, “Semantic
concept discovery for large-scale zero-shot event detection,” in IJCAI,
2015. III-A4
[77] J. Qin, L. Liu, L. Shao, F. Shen, B. Ni, J. Chen, , and YunhongWang,
“Zero-shot action recognition with error-correcting output codes,” in
CVPR, 2017. III-A4
[78] K. Yang, X.-S. Hua, M. Wang, and H.-J. Zhang, “Tag tagging: Towards
more descriptive keywords of image content,” IEEE Transactions on
Multimedia, vol. 13, pp. 662 –673, 2011. III-A4
[79] T. Hospedales, S. Gong, and T. Xiang, “Learning tags from unsegmented videos of multiple human actions,” in International Conference
on Data Mining, 2011. III-A4
[80] H. Aradhye, G. Toderici, and J. Yagnik, “Video2text: Learning to annotate video content,” in Proc. IEEE Int. Conf. Data Mining Workshops
ICDMW ’09, 2009, pp. 144–151. III-A4
[81] W. Yang and G. Toderici, “Discriminative tag learning on youtube
videos with latent sub-tags,” in IEEE Conference on Computer Vision
and Pattern Recognition, 2011. III-A4
[82] S. Wu, F. Luisier, and S. Bondugula, “Zero-shot event detection using
multi-modal fusion of weakly supervised concepts,” in CVPR, 2016.
III-A4
[83] M. Elhoseiny, B. Saleh, and A. Elgammal, “Write a classifier: Zeroshot learning using purely textual descriptions,” in IEEE International
Conference on Computer Vision, December 2013. III-B, VI-A9
[84] J. L. Ba, K. Swersky, S. Fidler, and R. Salakhutdinov, “Predicting deep
zero-shot convolutional neural networks using textual descriptions,” in
ICCV, 2015. III-B, IV-A3, IV-A4
[85] S. Reed, Z. Akata, B. Schiele, and H. Lee., “Learning deep representations of fine-grained visual descriptions,” in CVPR, 2016. III-B,
VII-2
[86] G. A. Miller, “Wordnet: A lexical database for english,” Commun.
ACM, vol. 38, no. 11, pp. 39–41, Nov. 1995. III-B1
[87] T. Mikolov, K. Chen, G. Corrado, and J. Dean, “Efficient estimation of
word representation in vector space,” in Proceedings of Workshop at
International Conference on Learning Representations, 2013. III-B2
13
[88] T. Mikolov, I. Sutskever, K. Chen, G. Corrado, and J. Dean, “Distributed representations of words and phrases and their compositionality,” in Neural Information Processing Systems, 2013. III-B2, VI-B2
[89] Z. S. Harris, Distributional Structure. Dordrecht: Springer Netherlands, 1981, pp. 3–22. III-B2
[90] H. Larochelle, D. Erhan, and Y. Bengio, “Zero-data learning of new
tasks,” in AAAI, 2008. IV
[91] X. Yu and Y. Aloimonos, “Attribute-based transfer learning for object
categorization with zero/one training example,” in European Conference on Computer Vision, 2010. IV-A1
[92] D. Jayaraman and K. Grauman, “Zero shot recognition with unreliable
attributes,” in NIPS, 2014. IV-A1, IV-C1
[93] Z. Akata, F. Perronnin, Z. Harchaoui, and C. Schmid, “Labelembedding for attribute-based classification,” in CVPR, 2013. IV-A2,
V-B2, VI-A3
[94] J. Weston, S. Bengio, and N. Usunier, “Large scale image annotation:
learning to rank with joint word-image embeddings,” Machine Learning, 2010. IV-A2
[95] Z. Li, E. Gavves, T. E. J. Mensink, and C. G. M. Snoek, “Attributes
make sense on segmented objects,” in European Conference on Computer Vision, 2014. IV-A2
[96] X. Li and Y. Guo, “Max-margin zero-shot learning for multiclass
classification,” in AISTATS, 2015. IV-A2
[97] X. Li, Y. Guo, and D. Schuurmans, “Semi-supervised zero-shot classification with label representation learning,” in ICCV, 2015. IV-A2
[98] D. Jayaraman, F. Sha, and K. Grauman, “Decorrelating semantic visual
attributes by resisting the urge to share,” in CVPR, 2014. IV-A2
[99] S. J. Hwang and L. Sigal, “A unified semantic embedding: relating
taxonomies and attributes,” in NIPS, 2014. IV-A2, IV-A3
[100] Z. Akata, S. Reed, D. Walter, H. Lee, and B. Schiele, “Evaluation
of output embeddings for fine-grained image classification,” in CVPR,
2015. IV-A3, V-B2
[101] B. Romera-Paredes and P. Torr, “An embarrassingly simple approach
to zero-shot learning,” in ICML, 2015. IV-A3
[102] Y. Fu, T. M. Hospedales, T. Xiang, Z. Fu, and S. Gong, “Transductive
multi-view embedding for zero-shot recognition and annotation,” in
ECCV, 2014. IV-A3, IV-C1, V-B2
[103] Y. Yang and T. M. Hospedales, “A unified perspective on multi-domain
and multi-task learning,” in ICLR, 2015. IV-A3, IV-A4
[104] D. Mahajan, S. Sellamanickam, and V. Nair, “A joint learning framework for attribute models and object descriptions,” in IEEE International Conference on Computer Vision, 2011, pp. 1227–1234. IV-A3
[105] Z. Fu, T. Xiang, E. Kodirov, and S. Gong, “zero-shot object recognition
by semantic manifold distance,” in CVPR, 2015. IV-A3, IV-C1
[106] J. Deng, N. Ding, Y. Jia, A. Frome, K. Murphy, S. Bengio, Y. Li,
H. Neven, and H. Adam, “Large-scale object classification using label
relation graphs,” in ECCV, 2014. IV-A3
[107] D. R. Hardoon, S. Szedmak, and J. Shawe-Taylor, “Canonical correlation analysis; an overview with application to learning methods,” in
Neural Computation, 2004. IV-A3
[108] R. Socher and L. Fei-Fei, “Connecting modalities: Semi-supervised
segmentation and annotation of images using unaligned text corpora,”
in IEEE Conference on Computer Vision and Pattern Recognition,
2010. IV-A3
[109] Y. Gong, Q. Ke, M. Isard, and S. Lazebnik, “A multi-view embedding space for modeling internet images, tags, and their semantics,”
International Journal of Computer Vision, 2013. IV-A3
[110] S. J. Hwang and K. Grauman, “Learning the relative importance of
objects from tagged images for retrieval and cross-modal search,”
International Journal of Computer Vision, 2011. IV-A3
[111] Y. Wang and S. Gong, “Translating topics to words for image annotation,” in ACM International Conference on Conference on Information
and Knowledge Management, 2007. IV-A3
[112] G.-J. Qi, W. Liu, C. Aggarwal, and T. Huang, “Joint intermodal and
intramodal label transfers for extremely rare or unseen classes,” IEEE
TPAMI, 2017. IV-A3
[113] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov,
D. Erhan, V. Vanhoucke, and A. Rabinovich, “Going deeper with
convolutions,” in CVPR, 2015. IV-A4
[114] L. Zhang, T. Xiang, and S. Gong, “Learning a deep embedding model
for zero-shot learning,” in CVPR, 2017. IV-A4
[115] Y. Fu, T. M. Hospedales, T. Xiang, and S. Gong, “Transductive multiview zero-shot learning,” IEEE TPAMI, 2015. IV-B, 1, IV-C1, VII-2
[116] S. Changpinyo, W.-L. Chao, B. Gong, and F. Sha, “Synthesized
classifiers for zero-shot learning,” in CVPR, 2016. IV-B, IV-C1, VI-A8
[117] Y. Zhang, B. Gong, and M. Shah, “Fast zero-shot image tagging,” in
CVPR, 2016. IV-B
[118] Y. Wang and G. Mori, “A discriminative latent model of image region
and object tag correspondence,” in Neural Information Processing
Systems, 2010. IV-B
[119] Y. Long, L. Liu, L. Shao, F. Shen, G. Ding, and J. Han, “From zeroshot learning to conventional supervised classification: Unseen visual
data synthesis,” in CVPR, 2017. IV-B
[120] E. Kodirov, T. Xiang, Z. Fu, and S. Gong, “Unsupervised domain
adaptation for zero-shot learning,” in ICCV, 2015. IV-C1
[121] M. Rohrbach, S. Ebert, and B. Schiele, “Transfer learning in a
transductive setting,” in NIPS, 2013. IV-C1, V-B2, VI-A8
[122] Y. Li, D. Wang, H. Hu, Y. Lin, and Y. Zhuang, “Zero-shot recognition
using dual visual-semantic mapping paths,” in CVPR, 2017. IV-C1
[123] X. Xu, T. Hospedales, and S. Gong, “Transductive zero-shot action
recognition by word-vector embedding,” in IJCV, 2016. IV-C1
[124] B. Marco, L. Angeliki, and D. Georgiana, “Hubness and pollution:
Delving into cross-space mapping for zero-shot learning,” in ACL,
2015. IV-C2
[125] T. Low, C. Borgelt, S. Stober, and A. Nürnberger, The Hubness
Phenomenon: Fact or Artifact?, 2013. IV-C2
[126] G. Dinu, A. Lazaridou, and M. Baroni, “Improving zero-shot learning
by mitigating the hubness problem,” in ICLR workshop, 2014. IV-C2
[127] Y. Shigeto, I. Suzuki, K. Hara, M. Shimbo, and Y. Matsumoto, “Ridge
regression, hubness, and zero-shot learning,” in ECML/PKDD, 2015.
IV-C2
[128] W.-L. Chao, S. Changpinyo, B. Gong, and F. Sha., “An empirical study
and analysis of generalized zero-shot learning for object recognition in
the wild,” in ECCV, 2016. V-A
[129] W. J. Scheirer, L. P. Jain, and T. E. Boult, “Probability models for open
set recognition,” IEEE TPAMI, 2014. V-A, V-A1
[130] W. J. Scheirer, A. Rocha, A. Sapkota, and T. E. Boult, “Towards open
set recognition,” IEEE TPAMI, 2013. V-A, V-A1
[131] Y. Fu and L. Sigal, “Semi-supervised vocabulary-informed learning,”
in CVPR, 2016. V-A, V-A2, VI-A8, VII-2
[132] Y. Fu, H. Dong, Y. feng Ma, Z. Zhang, and X. Xue, “Vocabularyinformed extreme value learning,” in arxiv, 2017. V-A, V-A2
[133] Y. Xian, B. Schiele, and Z. Akata, “Zero-shot learning - the good, the
bad and the ugly,” in CVPR, 2017. V-A, VI, VI-A1, VI-B3
[134] A. Bendale and T. Boult, “Towards open world recognition,” in CVPR,
2015. V-A1
[135] H. Sattar, S. Muller, M. Fritz, and A. Bulling, “Prediction of search
targets from fixations in open-world settings,” in CVPR, 2015. V-A1,
V-A2
[136] R. Gomes, M. Welling, and P. Perona, “Incremental learning of
nonparametric bayesian mixture models,” in IEEE Conference on
Computer Vision and Pattern Recognition, 2008. V-A1
[137] C. P. Diehl and G. Cauwenberghs, “Svm incremental learning, adaptation and optimization,” in IJCNN, vol. 4, 20–24 July 2003, pp. 2685–
2690. V-A1
[138] S. A. Rebuffi, A. Kolesnikov, G. Sperl, and C. H. Lampert, “icarl:
Incremental classifier and representation learning,” in CVPR, 2017.
V-A1, VII-4
[139] S. Guadarrama, E. Rodner, K. Saenko, N. Zhang, R. Farrell, J. Donahue, and T. Darrell, “Open-vocabulary object retrieval,” in Robotics
Science and Systems (RSS), 2014. V-A2
[140] S. Guadarrama, E. Rodner, K. Saenko, and T. Darrell, “Understanding
object descriptions in robotics by open-vocabulary object retrieval and
detection,” in Journal International Journal of Robotics Research,
2016. V-A2
[141] W. Zheng, S. Gong, and T. Xiang, “Towards open-world person reidentification by one-shot group-based verification,” in IEEE TPAMI,
2016. V-A2, V-B2
[142] H. Zhao, X. Puig, B. Zhou, S. Fidler, and A. Torralba, “Open
vocabulary scene parsing,” in CVPR, 2017. V-A2
[143] Jankowski, Norbert, Duch, Wodzislaw, Grabczewski, and Krzyszto,
“Meta-learning in computational intelligence,” in Springer Science &
Business Media, 2011. V-B
[144] B. M. Lake and R. Salakhutdinov, “One-shot learning by inverting a
compositional causal process,” in NIPS, 2013. V-B, V-B1, V-B2
[145] J. Goldberger, G. E. Hinton, S. T. Roweis, and R. R.
Salakhutdinov, “Neighbourhood components analysis,” in Advances
in Neural Information Processing Systems 17, L. K.
Saul, Y. Weiss, and L. Bottou, Eds.
MIT Press, 2005,
pp. 513–520. [Online]. Available: http://papers.nips.cc/paper/
2566-neighbourhood-components-analysis.pdf V-B1
[146] T. D. Kulkarni, W. F. Whitney, P. Kohli, and J. Tenenbaum, “Deep
convolutional inverse graphics network,” in NIPS, 2015. V-B1
14
[147] T. D. Kulkarni, V. K. Mansinghka, P. Kohli, and J. B. Tenenbaum,
“Inverse graphics with probabilistic cad models,” in arxiv:1407.1339,
2014. V-B1
[148] B. Lake, R. Salakhutdinov, and J. Tenenbaum, “Human-level concept
learning through probabilistic program induction,” in NIPS, 2013. V-B1
[149] R. JVilalta and Y. Drissi, “A perspective view and survey of metalearning,” Artificial intelligence review, 2002. V-B2
[150] L. Fei-Fei, R. Fergus, and P. Perona, “A bayesian approach to unsupervised one-shot learning of object categories,” in IEEE International
Conference on Computer Vision, 2003. V-B2
[151] ——, “One-shot learning of object categories,” IEEE TPAMI, 2006.
V-B2
[152] T. Tommasi and B. Caputo, “The more you know, the less you learn:
from knowledge transfer to one-shot learning of object categories,” in
British Machine Vision Conference, 2009. V-B2
[153] E. Bart and S. Ullman, “Cross-generalization: learning novel classes
from a single example by feature replacement,” in CVPR, 2005. V-B2
[154] T. Hertz, A. Hillel, and D. Weinshall, “Learning a kernel function for
classification with small training samples,” in ICML, 2016. V-B2
[155] F. Fleuret and G. Blanchard, “Pattern recognition from one example
by chopping,” in NIPS, 2005. V-B2
[156] Y. Amit, Fink, S. M., and U. N., “Uncovering shared structures in
multiclass classification,” in ICML, 2007. V-B2
[157] L. Wolf and I. Martin, “Robust boosting for learning from few
examples,” in CVPR, 2005. V-B2
[158] A. Torralba, K. Murphy, and W. Freeman, “sharing visual features
for multiclass and multiview object detection,” in IEEE TPAMI, 2007.
V-B2
[159] A. Torralba, K. P. Murphy, and W. T. Freeman, “Using the forest
to see the trees: Exploiting context for visual object detection and
localization,” Commun. ACM, 2010. V-B2
[160] J. Bromley, J. Bentz, L. Bottou, I. Guyon, Y. LeCun, C. Moore,
E. Sackinger, and R. Shah, “Signature verification using a siamese
time delay neural network,” in IJCAI, 1993. V-B2
[161] G. Koch, R. Zemel, and R. Salakhutdinov, “Siamese neural networks
for one-shot image recognition,” in ICML – Deep Learning Workshok,
2015. V-B2
[162] W. Kienzle and K. Chellapilla, “Personalized handwriting recognition
via biased reg- ularization,” in ICML, 2006. V-B2
[163] A. Quattoni, M. Collins, and T. Darrell, “Transfer learning for image
classification with sparse prototype representations,” in IEEE Conference on Computer Vision and Pattern Recognition, 2008, pp. 1–8. V-B2
[164] M. Fink, “Object classification from a single example utilizing class
relevance metrics,” in NIPS, 2005. V-B2
[165] L. Wolf, T. Hassner, and Y. Taigman, “The one-shot similarity kernel,”
in ICCV, 2009. V-B2
[166] J. Weston, S. Bengio, and N. Usunier, “Wsabie: Scaling up to large
vocabulary image annotation,” in IJCAI, 2011. V-B2
[167] Santoro, S. Bartunov, M. Botvinick, D. Wierstra, and T. Lillicrap, “Oneshot learning with memory-augmented neural networks,” in arx, 2016.
V-B2
[168] L. Bertinetto, J. F. Henriques, J. Valmadre, P. Torr, and A. Vedaldi,
“Learning feed-forward one-shot learners,” in NIPS, 2016. V-B2
[169] A. Habibian, T. Mensink, and C. Snoek, “Video2vec embeddings
recognize events when examples are scarce,” in IEEE TPAMI, 2014.
V-B2
[170] O. Vinyals, C. Blundell, T. Lillicrap, K. Kavukcuoglu, and D. Wierstra,
“Matching networks for one shot learning,” in NIPS, 2016. V-B2
[171] H. Zhang, K. Dana, and K. Nishino, “Friction from reflectance: Deep
reflectance codes for predicting physical surface properties from oneshot in-field reflectance„” in ECCV, 2016. V-B2
[172] Y. Wang and M. Hebert, “Learning from small sample sets by combining unsupervised meta-training with cnns,” in NIPS, 2016. V-B2
[173] ——, “Learning to learn: model regression networks for easy small
sample learning,” in ECCV, 2016. V-B2
[174] C. Finn, P. Abbeel, and S. Levine, “Model-agnostic meta-learning
for fast adaptation of deep networks,” in Proceedings of the 34th
International Conference on Machine Learning, ser. Proceedings
of Machine Learning Research, D. Precup and Y. W. Teh,
Eds., vol. 70. International Convention Centre, Sydney, Australia:
PMLR, 06–11 Aug 2017, pp. 1126–1135. [Online]. Available:
http://proceedings.mlr.press/v70/finn17a.html V-B2
[175] A. A. Rusu, N. C. Rabinowitz, G. Desjardins, H. Soyer, J. Kirkpatrick,
K. Kavukcuoglu, R. Pascanu, and R. Hadsell, “Progressive neural
networks,” arXiv preprint arXiv:1606.04671, 2016. V-B2
[176] D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” International Journal of Computer Vision, vol. 60, 2004. VI-A1
[177] K. E. A. van de Sande, T. Gevers, and C. G. M. Snoek, “Evaluation of
color descriptors for object and scene recognition,” in IEEE Conference
on Computer Vision and Pattern Recognition, 2008. VI-A1
[178] A. Bosch, A. Zisserman, and X. Munoz, “Representing shape with a
spatial pyramid kernel,” in ACM International Conference on Image
and Video Retrieval, 2007. VI-A1
[179] H. Bay, A. Ess, T. Tuytelaars, and L. V. Gool, “SURF: Speeded up
robust features,” Computer Vision and Image Understanding, vol. 110,
no. 3, pp. 346–359, 2008. VI-A1
[180] E. Shechtman and M. Irani, “Matching local self-similarities across
images and videos,” in IEEE Conference on Computer Vision and
Pattern Recognition, 2007. VI-A1
[181] J. Donahue, Y. Jia, O. Vinyals, J. Hoffman, N. Zhang, E. Tzeng, and
T. Darrell, “Decaf: A deep convolutional activation feature for generic
visual recognition,” in International Conference on Machine Learning,
2014. VI-A1
[182] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie, “The
Caltech-UCSD Birds-200-2011 Dataset,” California Institute of Technology, Tech. Rep. CNS-TR-2011-001, 2011. VI-A3
[183] A. Oliva and A. Torralba, “Modeling the shape of the scene: Aholistic
representation of the spatial envelope,” International Journal of Computer Vision, vol. 42, 2001. VI-A4
[184] G. Patterson and J. Hays, “Sun attribute database: Discovering, annotating, and recognizing scene attributes.” in IEEE Conference on
Computer Vision and Pattern Recognition, 2012. VI-A6
[185] J. Xiao, J. Hays, K. A. Ehinger, A. Oliva, and A. Torralba, “Sun
database: Large-scale scene recognition from abbey to zoo,” in IEEE
Conference on Computer Vision and Pattern Recognition, 2010, pp.
3485–3492. VI-A6
[186] Y.-G. Jiang, G. Ye, S.-F. Chang, D. Ellis, and A. C. Loui, “Consumer
video understanding: A benchmark database and an evaluation of
human and machine performance,” in ACM International Conference
on Multimedia Retrieval, 2011. VI-A7
[187] Z.-J. Zha, T. Mei, Z. Wang, and X.-S. Hua, “Building a comprehensive
ontology to refine video concept detection,” in Proceedings of the
International Workshop on Workshop on Multimedia Information
Retrieval, ser. MIR ’07. New York, NY, USA: ACM, 2007, pp. 227–
236. [Online]. Available: http://doi.acm.org/10.1145/1290082.1290114
VI-A7
[188] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet:
A large-scale hierarchical image database,” in CVPR, 2009. VI-A8
[189] M.-E. Nilsback and A. Zisserman, “Automated flower classification
over a large number of classes,” in Proceedings of the Indian Conference on Computer Vision, Graphics and Image Processing, 2008.
VI-A9
[190] K. Soomro, A. R. Zamir, and M. Shah, “Ucf101: A dataset of 101
human action classes from videos in the wild,” CRCV-TR-12-01, 2012.
VI-A10
[191] H. Idrees, A. R. Zamir, Y.-G. Jiang, A. Gorban, I. Laptev, R. Sukthankar, and M. Shah, “The thumos challenge on action recognition
for videos "in the wild",” Computer Vision and Image Understanding,
2017. VI-A10
[192] Y.-G. Jiang, Z. Wu, J. Wang, X. Xue, and S.-F. Chang, “Exploiting
feature and class relationships in video categorization with regularized
deep neural networks,” in IEEE TPAMI, 2017. VI-A11
[193] F. C. H. V. E. B. Ghanem and J. C. Niebles, “Activitynet: A large-scale
video benchmark for human activity understanding,” in CVPR, 2015.
VI-A12
[194] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and Y. LeCun,
“Overfeat: Integrated recognition, localization and detection using
convolutional networks,” in ICLR, 2014. VI-B1
[195] K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman, “Return
of the devil in the details: Delving deep into convolutional nets,” in
BMVC, 2014. VI-B1
[196] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” arXiv preprint arXiv:1512.03385, 2015. VI-B1
[197] J. Pennington, R. Socher, and C. D. Manning, “Glove: Global vectors
for word representation,” in EMNLP, 2014. VI-B2
[198] Y. Xian, Z. Akata, G. Sharma, Q. Nguyen, M. Hein, and B. Schiele,
“Latent embeddings for zero-shot classification,” in CVPR, 2016. VII-2
[199] A. Pentina and C. H. Lampert, “Lifelong learning with non-i.i.d. tasks,”
in NIPS, 2015. VII-4
[200] A. Pentina, V. Sharmanska, and C. H. Lampert, “Curriculum learning
of multiple tasks,” in CVPR, 2015. VII-4
15
PLACE
PHOTO
HERE
PLACE
PHOTO
HERE
Yanwei Fu received the BSc degree in information
and computing sciences from Nanjing University
of Technology in 2008; and the MEng degree in
the Department of Computer Science & Technology
at Nanjing University in 2011, China. He is now
pursuing his PhD in vision group of EECS, Queen
Mary University of London. His research interest
is attribute learning, topic model, learning to rank,
video summarization and image segmentation.
Tao Xiang received the Ph.D. degree in electrical
and computer engineering from the National University of Singapore in 2002. He is currently a reader
(associate professor) in the School of Electronic
Engineering and Computer Science, Queen Mary
University of London. His research interests include
computer vision, machine learning, and data mining.
He has published over 140 papers in international
journals and conferences.
Leonid Sigal is an Associate Professor at the University of British Columbia. Prior to this he was a
Senior Research Scientist at Disney Research. He
completed his Ph.D. at Brown University in 2008;
PLACE
received his M.A. from Boston University in 1999,
PHOTO
and M.Sc. from Brown University in 2003. Leonid’s
HERE
research interests lie in the areas of computer vision,
machine learning, and computer graphics. Leonid’s
research emphasis is on machine learning and statistical approaches for visual recognition, understanding and analytics. He has published more than 70
papers in venues and journals in these fields (including TPAMI, IJCV, CVPR,
ICCV and NIPS).
PLACE
PHOTO
HERE
PLACE
PHOTO
HERE
Yu-Gang Jiang a Professor in School of Computer
Science, Fudan University, China. His Lab for Big
Video Data Analytics conducts research on all aspects of extracting high-level information from big
video data, such as video event recognition, object/scene recognition and large-scale visual search.
His work has led to many awards, including the
inaugural ACM China Rising Star Award and the
2015 ACM SIGMM Rising Star Award.
Xiangyang Xue Xiangyang Xue received the B.S.,
M.S., and Ph.D. degrees in communication engineering from Xidian University, Xi’an, China, in
1989, 1992 and 1995, respectively. He is currently a
Professor of Computer Science at Fudan University,
Shanghai, China. His research interests include multimedia information processing and machine learning.
PLACE
PHOTO
HERE
Shaogang Gong received the DPhil degree in 1989
from Keble College, Oxford University. He has been
Professor of Visual Com- putation at Queen Mary
University of Lon- don since 2001, a fellow of the
Institution of Electrical Engineers and a fellow of the
British Computer Society. His research inter- ests
include computer vision, machine learn- ing, and
video analysis.
| 2cs.AI
|
1
On the reduction of differential inclusions and
Lyapunov stability
arXiv:1703.07071v4 [cs.SY] 23 Jan 2018
Rushikesh Kamalapurkar, Warren E. Dixon, and Andrew R. Teel
Abstract—In this paper, locally Lipschitz regular functions are
utilized to identify and remove infeasible directions from differential inclusions. The resulting reduced differential inclusion is
point-wise smaller (in the sense of set containment) than the
original differential inclusion. The reduced inclusion is utilized
to develop a generalized notion of a time derivative for locally
Lipschitz candidate Lyapunov functions. The developed generalized derivative yields less conservative statements of Lyapunov
stability results, invariance-like results, and Matrosov results
for differential inclusions. Illustrative examples are included to
demonstrate the utility of the developed stability theorems.
Index Terms—differential inclusions, stability of NL systems,
stability of hybrid systems, nonlinear systems
I. I NTRODUCTION
Differential inclusions can be used to model and analyze
a large variety of practical systems. For example, systems
that utilize discontinuous control architectures such as sliding
mode control, multiple model and sparse neural network
adaptive control, finite state machines, gain scheduling control,
etc., are analyzed using the theory of differential inclusions.
Differential inclusions are also used to analyze robustness to
bounded perturbations and modeling errors, to model physical
phenomena such as coulomb friction and impact, and to model
differential games [1], [2].
Asymptotic properties of trajectories of differential inclusions are typically analyzed using Lyapunov-like comparison functions. Several generalized notions of the directional
derivative are utilized to characterize the change in the value
of the candidate Lyapunov function along the trajectories of
differential inclusions. Early results on stability of differential
inclusions that utilize nonsmooth candidate Lyapunov functions are based on Dini directional derivatives [3], [4] and
contingent derivatives [5, Chapter 6]. For locally Lipschitz
regular candidate Lyapunov functions, stability results based
on Clarke’s notion of generalized directional derivatives have
been developed in results such as [6]–[8]. In [6], Paden and
Rushikesh Kamalapurkar is with the School of Mechanical and
Aerospace Engineering, Oklahoma State University, Stillwater, OK, USA.
[email protected].
Warren E. Dixon is with the Department of Mechanical and
Aerospace Engineering, University of Florida, Gainesville, FL, USA.
[email protected].
Andrew R. Teel is with the Department of Electrical and Computer Engineering, University of California, Santa Barbara, CA, USA.
[email protected].
This research is supported in part by NSF award numbers 1509516 and
1508757, ONR grant number N00014-13-1-0151, and AFOSR Award Number
FA9550-15-1-0155. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily
reflect the views of the sponsoring agency.
Sastry utilize the Clarke gradient to develop a set-valued generalized derivative along with several Lyapunov-based stability
theorems. In [7], Bacciotti and Ceragioli introduce another
set-valued generalized derivative that results in sets that are
smaller, pointwise, than those generated by the set-valued
derivative in [6]; hence, the Lyapunov theorems in [7] are
generally less conservative than their counterparts in [6]. The
Lyapunov theorems developed by Bacciotti and Ceragioli have
also been shown to be less conservative than those based on
Dini and contingent derivatives, provided locally Lipschitz
regular candidate Lyapunov functions are employed (cf. [9,
Proposition 7]).
In this paper, and in the preliminary work in [?], locally
Lipschitz regular functions are utilized to identify and remove
the infeasible directions from a differential inclusion to yield a
pointwise smaller (in the sense of set containment) equivalent
differential inclusion. Using the reduced differential inclusion,
a novel generalization of the set-valued derivative concepts
in [6] and [7] is introduced for locally Lipschitz Lyapunov
functions. The developed technique yields less conservative
statements of Lyapunov stability results (cf. [3], [4], [6], [7],
[10], [11]), invariance results (cf. [8], [12]–[14]), invariancelike results for nonautonomous systems (cf. [15], [16, Theorem
2.5]), and Matrosov results (cf. [17]–[21]) for differential
inclusions.
The paper is organized as follows. Section II introduces the
notation and Sections III and IV review differential inclusions
and Clarke-gradient-based set-valued derivatives from [6] and
[7]. In Section V, locally Lipschitz regular functions are used
to identify the infeasible directions in a differential inclusion.
Section VI includes a novel generalization of the notion
of a time-derivative with respect to a differential inclusion.
Sections VII and VIII state Lyapunov stability results that
utilize the novel definition of the generalized time-derivative
for autonomous and nonautonomous differential inclusions, respectively. Illustrative examples where the developed stability
theory is less conservative than results such as [6] and [7]
are presented. Section IX summarizes the article and includes
concluding remarks.
II. N OTATION
The n−dimensional Euclidean space is denoted by Rn and
µ denotes the Lebesgue measure on Rn . Elements of Rn are
T
interpreted as column vectors and (·) denotes the vector
transpose operator. The set of positive integers excluding 0
is denoted by N. For a ∈ R, R≥a denotes the interval
[a, ∞) and R>a denotes the interval (a, ∞). Unless otherwise specified, t0 denotes the initial time and an interval is
2
assumed to be right-open and of nonzero length. The notation
F : A ⇒ B is used to denote a set-valued map from A
to the subsets of B. The notations co A, coA, A, Å, and
∂A are used to denote the convex hull, the closed convex
hull, the closure, the interior, and the boundary of a set A,
m
n
respectively. If a ∈ R
and b ∈ R then [a; b] denotes the
a
concatenated vector
∈ Rm+n . For A ⊆ Rm , B ⊆ Rn
b
A
the notations
and A × B are interchangeably used to
B
n
T
denote the set {[a;
Tb] | a ∈ A, b ∈ B}. For A, B ⊆ R , A B
denotes the set a b | a ∈ A, b ∈ B , A ± B denotes the set
{a ± b ∈ Rn | a ∈ A, b ∈ B}, and A ≤ B implies kak ≤ kbk,
∀a ∈ A, and ∀b ∈ B. The notation B (x, r), B (x, r) and
D (r, l) for x ∈ Rn and r, l > 0 is used to denote the
sets {y ∈ Rn | kx − yk ≤ r}, {y ∈ Rn | kx − yk < r}, and
{y ∈ Rn | r ≤ kyk ≤ l}, respectively. The notation |(·)| denotes the absolute value if (·) ∈ R and the cardinality if (·) is
a set. For A ⊂ Rn and x ∈ Rn , dist (x, A) := inf y∈A kx − yk.
The notations L∞ (A, B), C n (A, B), and Lip (A, B) denote
essentially bounded, n−times continuously differentiable, and
locally Lipschitz functions with domain A and codomain B,
respectively. The notation 0n denotes the zero element of Rn .
Whenever clear from the context, the subscript n is suppressed.
III. D IFFERENTIAL INCLUSIONS
n
Let F : R × R≥t0 ⇒ Rn be a set-valued map. Consider
the differential inclusion
ẋ ∈ F (x, t) .
(1)
A locally absolutely continuous function x : I → Rn is called
a solution to (1) over the interval I ⊆ R if
ẋ (t) ∈ F (x (t) , t) ,
(2)
for almost all t ∈ I [1, p. 50]. A solution is called complete if
I = R≥t0 and maximal if it does not have a proper right
extension which is also a solution to (1). If a solution is
maximal and if the set {x (t) | t ∈ I} is compact, then the
solution is called precompact. Similar to [22, Proposition 1],
it can be shown that any solution to (1) can be extended to
a maximal solution; hence, if a solution exists, then it can be
assumed to be maximal without loss of generality.
To facilitate the discussion, let D ⊆ Rn be open and
connected and let Ω := D × I. Let ID := [t0 , TD ), where
TD denotes the first exit time of the solution x (·) from D
i.e., TD := min (sup I, inf {t ∈ I | x (t) ∈
/ D}) and inf ∅ is
assumed to be ∞. Since D is open and x (·) ∈ C 0 (I, Rn ),
if x (t0 ) ∈ D, then ID 6= ∅. Throughout the manuscript,
St (A) denotes the set of maximal solutions to (1) such that
(x (t0 ) , t0 ) ∈ A ⊂ Rn × R≥t0 and S (A) denotes the set of
maximal solutions to ẋ ∈ F (x) where x (t0 ) ∈ A ⊂ Rn . The
following notions of weak and strong forward invariance are
utilized in this paper.
Definition 1. A set A ⊆ Rn is called weakly forward invariant
with respect to (1) if ∃x (·) ∈ St (A × R≥t0 ) such that x (t) ∈
A, ∀t ∈ I. It is called strongly forward invariant with respect
to (1) if ∀x (·) ∈ St (A × R≥t0 ) and ∀t ∈ I, x (t) ∈ A. 4
The following development focuses on set-valued maps that
admit local solutions.
Definition 2. Let D ⊆ Rn and let J be an interval. The
set-valued map F : Rn × R≥t0 ⇒ Rn is said to admit local
solutions over D×J if for all (y, τ ) ∈ D×J , ∃T ∈ R>τ such
that a solution x : I → Rn to (1), starting from x (τ ) = y,
exists over the interval I := [τ, T ) ⊆ R≥t0 .
Sufficient conditions for the existence of local solutions can
be found in [1, §7, Theorem 1] and [1, §7, Theorem 5]. The
following Lemma, stated to facilitate the analysis, is a slight
generalization of [22, Proposition 2].
Lemma 1. Let D ⊆ Rn be open and connected and let
F : Rn × R≥t0 ⇒ Rn be a set-valued map that admits
solutions over D × R≥t0 . Let x : [t0 , T ) → Rn , T > t0 , be
a solution to (1) such that {x (t) | t ∈ [t0 , T )} ⊂ D. If the set
∪t∈J F (x (t) , t) is bounded for every subinterval J ⊆ [t0 , T )
of finite length, then x (·) is complete.
Proof: For the sake of contradiction, assume that
T < ∞. Since the set ∪t∈[t0 ,T ) F (x (t) , t) is bounded,
(2) implies that ẋ (·) ∈ L∞ ([t0 , T ) , Rn ). Since t 7→
x (t) is locally absolutely continuous, ∀t1 , t2 ∈ [t0 , T ),
´ t2
ẋ (τ ) dτ . Since ẋ (·) ∈
kx (t2 ) − x (t1 )k2 =
t1
´ t2
´ t2
n
L∞ ([t0 , T ) , R ), t1 ẋ (τ ) dτ ≤ t12 M dτ , where M is
2
a positive constant. Thus, kx (t2 ) − x (t1 )k2 ≤ M |t2 − t1 |,
and hence, t 7→ x (t) is uniformly continuous on [t0 , T ).
Therefore, x (·) can be extended into a continuous function
x0 : [t0 , T ] → Rn . Since x (·) is continuous, D is open,
and {x (t) | t ∈ [t0 , T )} ⊂ D, x0 (T ) ∈ D. Since F admits
solutions over D ×R≥t0 , x0 (·) can be extended into a solution
to (1) on the interval [t0 , T 0 ) for some T 0 > T , which
contradicts the maximality of x (·). Hence, x (·) is complete.
The hypothesis of Lemma 1 that the set ∪t∈J F (x (t) , t) is
bounded for every subinterval J ⊆ [t0 , T ) of finite length is
met if, e.g., (x, t) 7→ F (x, t) is locally bounded over D×R≥t0
and x (·) is precompact (cf. [23, Proposition 5.15]).
IV. S ET- VALUED DERIVATIVES
The focus of this article is on the development of a less
conservative Lyapunov method for the analysis of differential
inclusions using Clarke’s notion of generalized directional
derivatives and gradients [24, p. 39]. Clarke gradients are
utilized in [4] by Paden and Sastry to introduce the following
generalized notion of the time derivative of a locally Lipschitz
and regular (in the sense of [24, Definition 2.3.4]) candidate
Lyapunov function with respect to a differential inclusion.
Definition 3. [4] For a regular function V ∈ Lip (Ω, R), the
set-valued derivative of V with respect to (1) is defined as
Ṽ˙ (x, t) :=
\
p∈∂V (x,t)
pT (F (x, t) × {1}) ,
3
where ∂V denotes the Clarke gradient of V , defined as (see
also, [24, Theorem 2.5.1])
∂V (x, t) :=
co {lim ∇V (xi , ti ) | (xi , ti ) → (x, t) , (xi , ti ) ∈
/ ΩV } , (3)
where ΩV is the set of measure zero where the gradient of V
is not defined.
4
Lyapunov stability theorems developed using the set-valued
derivative Ṽ˙ exploit the property that every upper bound of
the set Ṽ˙ (x (t) , t) is also an upper bound of V̇ (x (t) , t), for
almost all t where V̇ (x (t) , t) exists. The aforementioned fact
is a consequence of the following proposition.
Proposition 1. [4] Let x : I → Rn be a solution to (1) such
that x (t0 ) ∈ D. If V ∈ Lip (Ω, R) is a regular function, then
V̇ (x (t) , t) exists for almost all t ∈ ID and V̇ (x (t) , t) ∈
Ṽ˙ (x (t) , t), for almost all t ∈ ID .
Proof: See [4, Theorem 3].
In [7], the notion of a set-valued derivative is further
generalized via the following definition.
Definition 4. [7] For a regular function V ∈ Lip (Ω, R), the
set-valued derivative of V with respect to (1) is defined as
V˙ (x, t) :=
a ∈ R | ∃q ∈ F (x, t) | pT [q; 1] = a, ∀p ∈ ∂V (x, t) . (4)
4
The set-valued derivative in Definition 4 results in less
conservative statements of Lyapunov stability than Definition
3 since it is contained within the set-valued derivative in Definition 3 and, as evidenced by [7, Example 1], the containment
can be strict. The Lyapunov stability theorems developed in
[7] exploit the property that Proposition 1 also holds for V˙
(see [7, Lemma 1]).
In the following, notions of Lyapunov stability for differential inclusions are introduced.1
Definition 5. The differential inclusion ẋ ∈ F (x) is said to
be (strongly)
(a) stable at x = 0 if ∀ > 0 ∃δ > 0 such that if x (·) ∈
S B (0, δ) , then x (·) is complete and x (t) ∈ B (0, ),
∀t ≥ 0.
(b) asymptotically stable at x = 0 if it is stable
at x = 0
and ∃c > 0 such that if x (·) ∈ S B (0, c) , then x (·)
is complete and limt→∞ kx (t)k = 0.
(c) globally asymptotically stable at x = 0 if it is stable at
x = 0 and if x (·) ∈ S (Rn ) implies that x (·) is complete
and limt→∞ kx (t)k = 0.
4
The following proposition is an example of a typical Lyapunov stability result for time-invariant differential inclusions
that utilizes set-valued derivatives of the candidate Lyapunov
function. The proposition combines [7, Theorem 2] and a
specialization of [6, Theorem 3.1].
1 While the results in this paper are stated in terms of stability at the origin,
they extend in a straightforward manner to stability of arbitrary compact sets.
Proposition 2. Let F : Rn ⇒ Rn be an upper semicontinuous map with compact, nonempty, and convex values.
If V ∈ Lip (Rn , R) is a positive definite and regular function
such that either2 max V˙ (x) ≤ 0 or max Ṽ˙ (x) ≤ 0, ∀x ∈ Rn ,
then ẋ ∈ F (x) is stable at x = 0.
Proof: See [7, Theorem 2] and [6, Theorem 3.1].
The following section details a novel generalization of the
notion of a time-derivative with respect to a differential inclusion that yields less conservative statements of results such as
Proposition 2. The generalization relies on the observation that
locally Lipschitz regular functions can be utilized to reduce
differential inclusions pointwise to sets of feasible directions
that are smaller than the corresponding sets in (1).
V. R EDUCED DIFFERENTIAL INCLUSIONS
By definition, V˙ ⊆ Ṽ˙ , which implies max V˙ (x) ≤
max Ṽ˙ (x). In some cases [7, Example 1], V˙ can be a proper
subset of Ṽ˙ i.e., max V˙ (x) < max Ṽ˙ (x), which implies that
Lyapunov theorems based on V˙ can be less conservative than
those based on Ṽ˙ . A tighter bound on the evolution of V as
x (·) moves along an orbit of (1) can be obtained by examining
the following alternative representation of max V˙ ,3
max V˙ (x, t) =
min
max
p∈∂V (x,t) q∈GF
V (x,t)
pT [q; 1] ,
(5)
where, for any regular function U ∈ Lip (Rn × R≥t0 , R), and
any set-valued map H : Rn × R≥t0 ⇒ Rn , the reduction
n
n
GH
U : R × R≥t0 ⇒ R is defined as
GH
U (x, t) :=
q ∈ H (x, t) | ∃a ∈ R | pT [q; 1] = a, ∀p ∈ ∂U (x, t) . (6)
Proposition 2 and (5) suggest that the only directions in F
that affect the stability properties of solutions to (1) are those
included in GF
V , that is, the directions that map the Clarke
gradient of V into a singleton. The key observation in this
paper is that the statement above remains true even if V is
replaced with any arbitrary locally Lipschitz regular function
U . The following proposition formalizes the aforementioned
observation.
Proposition 3. Let F : Rn ⇒ Rn be a map with compact
and nonempty values that admits solutions over Rn . Let V ∈
Lip (Rn , R) be a positive definite and regular function and let
U ∈ Lip (Rn , R) be any other regular function. If
min
max pT q ≤ 0,
p∈∂V (x) q∈GF
U (x)
∀x ∈ Rn ,
then ẋ ∈ F (x) is stable at x = 0.
Proof: The proposition follows from the more general
result stated in Theorem 3.
2 The definitions of V˙ and Ṽ˙ translate to time-invariant systems as
˙
F
V (x) = a ∈ R | ∃q ∈ F (x) | pT q = a, ∀p ∈ ∂V (x) and Ṽ˙ (x) :=
T
T
p∈∂V (x) p F (x), respectively.
3 The minimization in (5) serves to maintain consistency of notation, but is
in fact, redundant.
4
Proposition 3 indicates that locally Lipschitz regular functions help discover the admissible directions in F . That
is, from the point of view of Lyapunov stability, only the
directions in GF
U are relevant, where U can be different from
the candidate Lyapunov function V . If U = V , Proposition 3
reduces to Proposition 2.
In fact, the differential inclusion ẋ ∈ GF
U (x, t) is, in a sense,
equivalent to the differential inclusion ẋ ∈ F (x, t). To make
the equivalence precise, the following definition of a reduced
differential inclusion is introduced.
ẋ (t) ∈ F̃U (x (t) , t), ∀t ∈ T . Since µ (ID \ T ) = 0,
ẋ (t) ∈ F̃U (x (t) , t), for almost all t ∈ ID .
The following example demonstrates the utility of Theorem
1.
∞
Definition 6. Let U := {Ui }i=1 be a collection of realvalued locally Lipschitz regular functions defined on Ω. The
set-valued map F̃U : Rn × R≥t0 ⇒ Rn , defined as
F
F̃U (x, t) := F (x, t) ∩ ∩∞
i=1 GUi (x, t) ,
∀t ∈ R≥0 , where sgn (x) denotes the sign of x. The function
U : R × R≥0 → R, defined as
(
|x|
|x| ≤ 1,
U (x, t) :=
2 |x| − 1 |x| > 1,
∀t ∈ R≥0 , satisfies U ∈ Lip R2 , R . In addition, since U is
convex, it is also regular [24, Proposition 2.3.6]. The Clarke
gradient of U can be computed using (3) as
[1, 2] × {0}
x = 1,
[−2,
−1]
×
{0}
x = −1,
∂U (x, t) = {sgn (x)} × {0}
0 < |x| < 1,
{2 sgn (x)} × {0} |x| > 1,
[−1, 1] × {0}
|x| = 0,
is called the U-reduced differential inclusion for (1).
4
The key utility of the reduction developed in Definition 6
is that the reduced differential is sufficient to characterize the
solutions to (1). That is, as demonstrated by the following
theorem, the differential inclusion ẋ ∈ F̃U (x, t) admits local
solutions if (1) admits local solutions and for every solution
x : I → Rn to (1), there exists a subinterval J ⊆ I, such that
x (·), when restricted to J , is also a solution to ẋ ∈ F̃U (x, t)
Theorem 1. Let U be a collection of countably many realvalued locally Lipschitz regular functions defined on Ω. If x :
I → Rn is a solution to (1) such that x (t0 ) ∈ D, then ẋ (t) ∈
F̃U (x (t) , t) for almost all t ∈ ID .
Proof: The proof closely follows the proof of [7, Lemma
1]. Consider the set of times T ⊆ ID where ẋ (t) is defined,
U̇i (x (t) , t) is defined ∀i ≥ 0, and ẋ (t) ∈ F (x (t) , t).
Since x (·) is a solution to (1) and since Ui ∈ Lip (Ω, R),
t 7→ Ui (x (t) , t) is absolutely continuous, and hence,
µ (ID \ T ) = 0. The objective is to show that ẋ (t) ∈
F̃U (x (t) , t), not just F (x (t) , t).
Since each function Ui is locally Lipschitz, for t ∈ T the
time derivative of Ui can be expressed as
(Ui (x (t)+hẋ (t) , t + h) − Ui (x (t) , t))
U̇i (x(t) , t) = lim
.
h→0
h
Since each Ui is regular, for i ≥ 1,
0
U̇i (x (t) , t) = Ui+
([x (t) ; t] , [ẋ (t) ; 1]) ,
= Uio ([x (t) ; t] , [ẋ (t) ; 1]) ,
=
max
pT [ẋ (t) ; 1] ,
Example 1. Consider the differential inclusion in (1), where
x ∈ R, and F : R × R≥0 ⇒ R is defined as
(
2 sgn (x − 1) |x| =
6 1,
F (x, t) :=
[−2, 5]
|x| = 1,
∀t ∈ R≥0 . The set GF
U is then given
{0}
F
GU (x, t) = ∅
F (x, t)
by
|x| = 1,
|x| = 0,
otherwise,
∀t ∈ R≥0 . Theorem 1 can then be invoked to conclude that
every solution x : I → Rn to ẋ ∈ F (x, t) satisfies ẋ (t) ∈
F̃{U } (x (t) , t) = GF
4
U (x (t) , t), for almost all t ∈ I.
VI. G ENERALIZED TIME DERIVATIVES
Proposition 3 and Theorem 1 suggest the following notion
of a generalized time derivative of V with respect to (1).
Definition 7. The U−generalized time derivative of V ∈
Lip (Ω, R) with respect to (1), denoted by V˙ U : Ω → R is
defined, ∀ (x, t) ∈ Ω, as
V˙ U (x, t) :=
min
max
pT [q; 1] ,
(7)
pT [q; 1] ,
(8)
p∈∂V (x,t) q∈F̃U (x,t)
p∈∂Ui (x(t),t)
0
U̇i (x (t) , t) = Ui−
([x (t) ; t] , [ẋ (t) ; 1]) ,
=
=
Uio
([x (t) ; t] , [ẋ (t) ; 1]) ,
min
T
p [ẋ (t) ; 1] ,
p∈∂Ui (x(t),t)
0
U+
0
U−
o
where
and
denote the right and left directional
derivatives, and U denotes the Clarke-generalized derivative
of U .4 Thus, pT [ẋ (t) ; 1] = U̇i (x (t) , t) , ∀p ∈ ∂Ui (x (t) , t),
which implies ẋ (t) ∈ GF
Ui (x (t) , t), for each i. Therefore,
4 See [24, p. 39] for definitions of the right and left directional derivatives
and the Clarke-generalized derivative.
if V is regular, and
V˙ U (x, t) :=
max
max
p∈∂V (x,t) q∈F̃U (x,t)
if V is not regular. The U−generalized time derivative is
understood to be −∞ when F̃U (x, t) is empty.
4
Definition 7 also facilitates a unified treatment of Lyapunov
stability theory using regular as well as nonregular candidate
Lyapunov functions. The candidate Lyapunov function will
be called a Lyapunov function if the U−generalized time
derivative is negative.
5
Definition 8. If V ∈ Lip (Ω, R) is positive definite and if
V˙ U (x, t) ≤ 0, ∀ (x, t) ∈ Ω, then V is called a U−generalized
Lyapunov function for (1).
4
˙
If V ∈ U then F̃U ⊆ GF
V , and hence, V U (x, t) ≤
˙
max V (x, t) , ∀ (x, t) ∈ Ω. Thus, by judicious selection of
the functions in U, V˙ (x, t) can be constructed to be less
U
conservative than the set-valued derivatives in [6] and [7].
Naturally, if U = {V } then V˙ U = V˙ .
In general, the U−generalized time derivative does not
satisfy the chain rule as stated in Proposition 1. However,
it satisfies the following weak chain rule which turns out
to be sufficient for Lyapunov-based analysis of differential
inclusions.
Theorem 2. If V ∈ Lip (Ω, R), then ∀x (·) ∈ St (Ω),
T F̃U (x (t) , t)
V̇ (x (t) , t) ∈ (∂V (x (t) , t))
,
{1}
(9)
for almost all t ∈ ID . In addition, if there exists a function
W : Ω → R such that
V˙ U (x, t) ≤ W (x, t) , ∀ (x, t) ∈ Ω,
(10)
then
V̇ (x (t) , t) ≤ W (x (t) , t) ,
(11)
for almost all t ∈ ID .
Proof: Let x (·) ∈ St (Ω). Since x (t0 ) ∈ D, D is open,
and x (·) is continuous, ID is nonempty. Consider a set of
times T ⊆ ID where ẋ (t), V̇ (x (t) , t), and U̇i (x (t) , t) are
defined ∀i ≥ 0 and ẋ (t) ∈ F̃U (x (t) , t). Using Theorem 1 and
the facts that x (·) is absolutely continuous and V is locally
Lipschitz, µ (ID \ T ) = 0.
If V is regular, then arguments similar to the
proof of Theorem 1 can be used to conclude that
V̇ (x (t) , t) = pT [ẋ (t) ; 1] , ∀p ∈ ∂V (x (t) , t) , ∀t ∈ T .
Thus, (7), (10), and Theorem
1 imply that
T
V̇ (x (t) , t) ∈ (∂V (x (t) , t)) F̃U (x (t) , t) × {1} and
V̇ (x (t) , t) ≤ W (x (t) , t), for almost all t ∈ ID .
If V is not regular, then [9, Proposition 4] (see also,
[25, Theorem 2]) can be used to conclude that, for almost
every t ∈ ID , ∃p0 ∈ ∂V (x (t) , t) such that V̇ (x (t) , t) =
pT0 [ẋ (t) ; 1]. Thus, (8), (10), and
Theorem 1 imply
that
T
V̇ (x (t) , t) ∈ (∂V (x (t) , t)) F̃U (x (t) , t) × {1} and
V̇ (x (t) , t) ≤ W (x (t) , t) for almost all t ∈ ID .
The following sections develop relaxed Lyapunov-like stability theorems for differential inclusions based on the properties of the U−generalized time derivative hitherto established.
VII. S TABILITY OF AUTONOMOUS SYSTEMS
In this section, U−generalized Lyapunov functions are
utilized to formulate less conservative extensions to stability
and invariance results for autonomous differential inclusions
of the form
ẋ ∈ F (x) ,
(12)
where F : Rn ⇒ Rn is a set-valued map.
A. Lyapunov stability
The following Lyapunov stability theorem is a consequence
of Theorem 2.
Theorem 3. Let 0 ∈ D and let F : Rn ⇒ Rn be a locally
bounded set-valued map with nonempty and compact values
that admits solutions over D. If there exists a U−generalized
Lyapunov function V : D → R for (12), then (12) is stable
at x = 0. If in addition, V˙ U (x) ≤ −W (x) , ∀x ∈ D, for
some positive definite function W ∈ C 0 (Rn , R), then (12) is
asymptotically stable at x = 0. Furthermore, if D = Rn and
if the sublevel sets Ll := {x ∈ Rn | V (x) ≤ l} are compact
for all l ∈ R>0 , then (12) is globally asymptotically stable at
x = 0.
Proof: Given > 0, let r > 0 be such that
B (0, r) ⊂
D
and
r
∈
(0,
].
Let
β
∈
0,
min
V
(x)
and Lβ :=
kxk=r
x ∈ B (0, r) | V (x) ≤ β . Since V is continuous, ∃δ > 0
such that B (0, δ) ⊂ Lβ . Using Theorem 2, which implies that
t 7→ V (x (t)) is nonincreasing on ID , and standard arguments
(see, e.g., [26, Theorem 4.8]), it can be shown that Lβ is
compact, (strongly) forward invariant, and Lβ ⊂ D. Hence,
every solution x (·) ∈ S (Lβ ) is precompact, and
by Lemma
1, complete. Furthermore, if x (·) ∈ S B (0, δ) then x (t) ∈
B (0, ), ∀t ∈ R≥t0 .
In addition, if V˙ U (x) ≤ −W (x) , ∀x ∈ D, for some
positive definite function W ∈ C 0 (Rn , R), then Theorem 2
implies that t 7→ V (x (t)) is strictly decreasing on R≥t0
provided x (t0 ) ∈ B (0, δ). Asymptotic stability and global
asymptotic stability (in the case where the sublevel sets of
V are compact) of (12) at x = 0 then follow from standard
arguments (see, e.g., [27, Section 5.3.2]).
The following example presents a case where tests based
on V˙ and Ṽ˙ are inconclusive but Theorem 3 can be used to
establish asymptotic stability.
Example 2. Let H : R ⇒ R be defined as
(
{0}
|y| =
6 1,
H (y) :=
[−1, 1] |y| = 1.
Let F : R2 ⇒ R2 be defined as
{−x1 + x2 } + H (x2 )
F (x) :=
.
{−x1 − x2 } + H (x1 )
Consider the differential inclusion in (1) and the candidate
2
Lyapunov function V : R2 → R defined as V (x) := 21 kxk2 .
Since V ∈ C 1 R2 , the set-valued derivatives V˙ in [7] and
Ṽ˙ in [6] are bounded by
V˙ (x),Ṽ˙ (x) ≤ −x21 − x22 + x1 H (x2 ) + x2 H (x1 ) .
(13)
Since neither Ṽ˙ nor V˙ can be shown to be negative semidefinite everywhere, the inequality in (13) is insufficient to draw
conclusions regarding the stability of (12).
The function U : R2 → R, defined as (see Fig. 1)
U (x) := max ((x1 − 1) , 0) − min ((x1 + 1) , 0)
+ max ((x2 − 1) , 0) − min ((x2 + 1) , 0)
6
In applications where a negative definite bound on the
derivative of the candidate Lyapunov function cannot be found
easily, the invariance principle is invoked. The following
section develops invariance results using U−generalized timederivatives.
B. Invariance principle
Analogs of the Barbashin-Krasovskii-LaSalle invariance
principle for autonomous differential inclusions appear in results such as [7], [12], [28]. Estimates of the limiting invariant
set that are less conservative than those developed in [7], [12],
[28] can be obtained by using locally Lipschitz regular functions to reduce the admissible directions in F . For example, the
following theorem extends the invariance principle developed
by Bacciotti and Ceragioli (see [7, Theorem 3]).
Fig. 1. The function U : R2 → R.
satisfies U ∈ Lip R2 , R . In addition, since U is convex, it is
also regular [24, Proposition 2.3.6]. The Clarke gradient of
U is given by,
#)
("
sgn 1 (x1 )
|x1 | =
6 1 ∧ |x2 | =
6 1,
sgn 1 (x2 )
"
#
co {0, sgn (x1 )}
|x1 | = 1 ∧ |x2 | =
6 1,
{sgn 1 (x2 )}
∂U (x) = "
#
{sgn
1
(x
)}
1
|x1 | =
6 1 ∧ |x2 | = 1,
co {0, sgn (x2 )}
"
#
co
{0,
sgn
(x
)}
1
|x1 | = 1 ∧ |x2 | = 1,
co {0, sgn (x2 )}
where
(
0
−1 < y < 1,
sgn 1 (y) :=
sgn (y) otherwise .
In this case, the reduced inclusion GF
U is given by
(
F (x) |x1 | =
6 1 ∧ |x2 | =
6 1,
GF
U (x) =
∅
otherwise.
2
F
Since GF
U (x) ⊂ F (x) , ∀x ∈ R , F̃{U } = GU . Since
∂V
2
1
V ∈ C R , R , ∂V (x) =
∂x (x) , and hence, the
{U } −generalized time derivative of V with respect to (12)
is given by
T
∂V
˙
V {U } (x) = max
(x)
q,
∂x
q∈F̃{U } (x)
"
#
i −x + x
h
1
2
|x1 | =
6 1 ∧ |x2 | =
6 1,
x1 x2
=
−x1 − x2
−∞
otherwise,
≤ −x21 − x22 .
Global asymptotic stability of (12) at x = 0 is then follows
from Theorem 3.
4
Theorem 4. Let F : Rn ⇒ Rn be locally bounded and outer
semicontinuous [23, Definition 5.4] over D and let F (x) be
nonempty, convex, and compact, ∀x ∈ D. Let V ∈ Lip (D, R)
and V˙ U (x) ≤ 0, ∀x ∈ D. Let Cn ⊂ D be a compact
o
strongly forward invariant set. If E := x ∈ D | V˙ U (x) = 0
and if M is the largest weakly forward invariant set in
E ∩ C, then x (·) ∈ S (C) implies that x (·) is complete
and limt→∞ dist (x (t) , M ) = 0.
Proof: Existence of x (·) follows from [1, §7, Theorem
1] and its completeness follows from Lemma 1. The same
argument as [7, Theorem 3] indicates that t 7→ V (x (t)) is
constant on ω (x (·)), the ω−limit set of x (·) [7, Definition
3]. Note that ω (x (·)) is weakly invariant [12, Proposition 2.8]
and ω (x (·)) ⊂ C. Let y : I → ω (x (·)) be a solution to
(12) such that y (t0 ) ∈ ω (x). The existence of such a solution
follows from weak invariance of ω (x (·)). Since t 7→ V (x (t))
is constant on ω (x (·)), V̇ (y (t)) = 0, ∀t ∈ R≥t0 .
Let T be a set of time instances where ẏ (t) is defined and
ẏ (t) ∈ F̃U (y (t)). If V is regular, then arguments similar
to the proof of Theorem 1 can be used to conclude that
∀t ∈ T and ∀p ∈ ∂V (y (t)), 0 = V̇ (x (t)) = pT ẏ (t).
Since V˙ U (x) = minp∈∂V (x) maxq∈F̃U (x) pT q ≤ 0, ∀x ∈ D
and pT ẏ (t) = 0, ∀p ∈ ∂V (y (t)), it follows that V˙ (y (t)) =
U
0, ∀t ∈ T , which means that y (t) ∈ E, for almost all t ≥ t0 .
If V is not regular then [9, Proposition 4] (see also, [25,
Theorem 2]) can be used to conclude that for almost every
t ∈ I, ∃p0 ∈ ∂V (y (t)) such that 0 = V̇ (y (t)) = pT0 ẏ (t).
Since, V˙ U (x) = maxp∈∂V (x) maxq∈F̃U (x) pT q ≤ 0, ∀x ∈ D
and pT0 ẏ (t) = 0 for some p0 ∈ ∂V (y (t)), it follows that
V˙ U (y (t)) = 0, ∀t ∈ T , which means that y (t) ∈ E for
almost all t ≥ t0 .
Since y (·) ∈ C 0 (R≥t0 , Rn ), y (t) ∈ E, ∀t ≥ t0 .
That is, ω (x (·)) ⊂ E, and hence, ω (x (·)) ⊂ E ∩
C. Since ω (x (·)) is weakly invariant, ω (x (·)) ⊂ M .
As a result, limt→∞ dist (x (t) , ω (x (·))) = 0 implies
limt→∞ dist (x (t) , M ) = 0.
The following corollary illustrates one of the many alternative ways to establish the existence of a compact strongly
forward invariant set needed to apply Theorem 4.
7
Corollary 1. Let F : Rn ⇒ Rn be locally bounded and
outer semicontinuous [23, Definition 5.4] over D and let
F (x) be nonempty, convex, and compact, ∀x ∈ D. Let
V ∈ Lip (D, R) and V˙ U (x) ≤ 0, ∀x ∈ D. Let l > 0
be such that the level set Ll := {x ∈ D : V (x) ≤ l} is
closednand a connected component
Cl of Ll is bounded. If
o
E := x ∈ D | V˙ U (x) = 0 and if M is the largest weakly
forward invariant set contained in E ∩Cl , then x (·) ∈ S Cl
implies that x (·) is complete and limt→∞ dist (x (t) , M ) = 0.
Proof: Let x : I → R be a maximal solution to (12)
such that x (t0 ) ∈ Cl . From Theorem 2, t 7→ V (x (t)) is
nonincreasing on ID . If x (t1 ) ∈
/ Cl for some t1 ∈ ID , then
continuity of x (·), the fact that Cl is a connected component
of Ll , and the fact that Ll is closed imply that x (t1 ) ∈
/ Ll ,
which is impossible since t 7→ V (x (t)) is nonincreasing on
ID . Hence, Cl is strongly forward invariant on ID . Since Cl
is bounded by assumption, x (·) is precompact, and hence,
complete, by Lemma 1. The conclusion of the corollary then
follows from Theorem 4 with C = Cl .
The invariance principle is often applied to conclude asymptotic stability at the origin in the form of the following
corollary.
complete solutions converge to the origin, x (·) ∈ S B (0, δ)
=⇒ limt→∞ x (t) = 0; hence, (12) is asymptotically stable at
x = 0. If D = Rn and the sublevel sets {x ∈ Rn | V (x) ≤ l}
are compact for all l ∈ R≥0 , then r, and hence, δ, can be
selected arbitrarily large; therefore, (12) is globally asymptotically stable at x = 0.
The following example demonstrates the utility of the
developed invariance principle.
Example 3. Let H : R ⇒ R be defined as
(
{0}
|y| =
6 1,
H (y) := 1 1
−2, 2
|y| = 1,
and let F : R2 ⇒ R2 be defined as
{x2 } + H (x2 )
F (x) :=
,
{−x1 − x2 } + H (x1 )
(14)
and consider the differential inclusion in (1). The candidate
Lyapunov function V : R2 → R is defined as V (x) :=
1/2 kxk2 . Since V ∈ C 1 R2 , R , the set-valued derivatives
2
˙
V in [7] and Ṽ˙ in [6] are bounded by
V˙ (x) , Ṽ˙ (x) ≤ −x22 + x2 H (x1 ) + x1 H (x2 ) .
(15)
Corollary 2. Let 0 ∈ D and V ∈ Lip (D, R) be a positive
definite function. Let F : Rn ⇒ Rn be locally bounded and
outer semicontinuous [23, Definition 5.4] over D and let F (x)
be nonempty, convex, and compact, ∀x ∈ D. If V˙ U (x) ≤ 0,
∀x ∈ D, and if for each ν > 0, no complete solution to
(12) remains in the level set {x ∈ D | V (x) = ν}, then (12)
is asymptotically stable at x = 0. In addition, if D = Rn
and the sublevel sets {x ∈ Rn | V (x) ≤ l} are compact for
all l ∈ R≥0 , then (12) is globally asymptotically stable at
x = 0.
That is, neither Ṽ˙ (x) nor V˙ (x) are negative semidefinite
everywhere, and hence, the inequality in (15) is inconclusive.
Let U : R2 → R be defined as in Example 2. The
{U } −reduced differential inclusion corresponding to F is
given by
F (x)
|x1 | =
6 1 ∧ |x2 | =
6 1,
{0} × {−1} + − 1 , 1 x = 1 ∧ x = 0,
1
2
2
1 21
F̃{U } (x) =
{0}
×
{1}
+
−
,
x
=
−1
∧
x2 = 0,
1
2 2
∅
otherwise.
Proof: To prove the corollary, it is first established that
every complete solution converges to the origin, and then it is
shown that all solutions are complete. Let x : R≥t0 → Rn
be any complete solution to (12). Since t 7→ V (x (t)) is
decreasing and bounded below, limt→∞ V (x (t)) = c for
some c ≥ 0. For all y ∗ ∈ ω (x (·)), ∃ {ti }i∈N ⊂ R≥t0 ,
limi→∞ ti = ∞ such that limi→∞ x (ti ) = y ∗ . Since V
is continuous, V (ω (x (·))) = {c}. If it can be shown that
c = 0, then positive definiteness of V would imply that
ω (x (·)) = {0}, and hence, limt→∞ x (t) = 0.
To prove that c = 0 using contradiction, assume that there
exists a complete solution x : R≥t0 → Rn such that c > 0.
Let y : R≥t0 → Rn be a solution to (12) such that y (t) ∈
ω (x (·)) , ∀t ∈ R≥t0 . Such a solution exists since ω (x (·))
is weakly invariant. Along the solution y, V (y (t)) = c > 0,
∀t ∈ R≥t0 , which contradicts the hypothesis that no complete
solution to (12) remains in the level set {x ∈ D | V (x) = ν}
for any ν > 0. Therefore, c = 0
Select r > 0 such that B (0, r) ⊂ D and δ > 0 such
that B (0, δ) ∈ Lβ := x ∈ B (0, r) | V (x) ≤ β , where
β ∈ 0, minkxk=r V (x) . Lyapunov stability of (12) at x = 0
and the fact that Lβ is strongly forward invariant follows from
Theorem 3. Since Lβ is compact, all solutions starting in Lβ
are precompact, and hence, complete, by Lemma 1. Since all
The {U } −generalized time derivative of V with respect to
ẋ = F (x) is then given by
x1 x2 q,
V˙ {U } (x) := max
q∈F̃{U } (x)
"
#
h
i
x
2
x1 x2
−x1 − x2
#
"
h
i
{0}
max 1 0
− 23 , − 12
=
"
#
h
i {0}
max −1 0 1 3
2, 2
−∞
≤
|x1 | =
6 1 ∧ |x2 | =
6 1,
x1 = 1 ∧ x2 = 0,
x1 = −1 ∧ x2 = 0,
otherwise,
−x22 .
In
this 2case, the set E in Corollary 1 is given by E =
x ∈ R | x2 = 0 . Since the level sets Ll are bounded and
connected, ∀l ∈ R≥0 , and since the largest invariant set
contained within E ∩ Ll is {[0; 0]}, ∀l ∈ R≥0 , Corollary 1
can be invoked to conclude that all solutions to (12) converge
to the origin.
8
From Theorem 2, V̇ (x (t) , t) ≤ −x22 (t) for almost all t ∈
R≥0 ; hence, given any ν > 0, a trajectory of (12) can remain
on the level set {x √
∈ Rn | V (x) = ν} if and only if x2 (t) =
0 and x1 (t) = ± 2ν, for all t ∈ R≥0 . From
1,
√ Theorem
the state [x1 ; x2 ] can remain
constant
at
±
2ν;
0
for
all
√
t ∈ R≥0 only if [0; 0] ∈ F ± 2ν; 0 , which is not true for
the inclusion in (14). Therefore, Corollary 2 can be invoked
to conclude that the system is globally asymptotically stable
at x = 0.
4
if the sublevel sets {x ∈ Rn | W (x) ≤ c} are compact ∀c ∈
R≥0 , then (1) is globally uniformly asymptotically stable at
x = 0.
Proof: Select r > 0 such
that B (0, r) ⊂ D. Let x (·) ∈
St (Ωc × R ≥t0 ) where Ωc := x ∈ B (0, r) |W (x) ≤ c for
some c ∈ 0, minkxk2 =r W (x) . Using Theorem 2 and [16,
Lemma 2],
V (x (t0 ) , t0 ) ≥ V (x (t) , t) ,
∀t ∈ ID .
(18)
VIII. S TABILITY OF N ONAUTONOMOUS SYSTEMS
Before stating the stability results for nonautonomous systems, the following definitions are stated.
Definition 9. The differential inclusion in (1) is said to be
(strongly)
(a) uniformly stable at x = 0, if ∀ > 0 ∃δ > 0 such that if
x (·) ∈ St B (0, δ) × R≥0 , then x (·) is complete and
x (t) ∈ B (0, ), ∀t ≥ t0 .
(b) globally uniformly stable at x = 0, if it is uniformly
stable at x = 0 and ∀ > 0 ∃∆ > 0 such that if x (·) ∈
St B (0, ) × R≥0 , then x (·) is complete and x (t) ∈
B (0, ∆), ∀t ≥ t0 .
(c) uniformly asymptotically stable at x = 0 if it is uniformly
stable at x = 0 and ∃c > 0 such that
∀ > 0 ∃T ≥ 0 such
that if x (·) ∈ St B (0, c) × R≥0 , then x (·) is complete
and x (t) ∈ B (0, ), ∀t ≥ t0 + T .
(d) globally uniformly asymptotically stable at x = 0 if it is
uniformly stable at x = 0 and ∀c,
> 0 ∃T ≥ 0 such
that if x (·) ∈ St B (0, c) × R≥0 , then x (·) is complete
and x (t) ∈ B (0, ), ∀t ≥ t0 + T .
4
While the results in this section are stated in terms of
stability of the entire state at the origin and uniformity with
respect to time, they extend in a straightforward manner
to partial stability, uniformity with respect to a part of the
state (see, e.g., [15, Definition 4.1]) and stability of arbitrary
compact sets.
A. Lyapunov stability
In this section, a basic Lyapunov-based stability result is
stated for nonautonomous differential inclusions.
Theorem 5. Let 0 ∈ D and let Ω := D × R≥t0 . Let F :
Rn × R≥t0 ⇒ Rn be a locally bounded set-valued map with
nonempty and compact values that admits solutions over Ω.
Let V ∈ Lip (Ω, R) be a positive definite function. If there
exist positive definite functions W , W ∈ C 0 (D, R) such that
W (x) ≤ V (x, t) ≤ W (x) , ∀ (x, t) ∈ Ω,
˙
V U (x, t) ≤ 0, ∀x ∈ D, and almost all t ∈ R≥t0 ,
(16)
then (1) is uniformly stable at x = 0. In addition, if there
exists a positive definite function W ∈ C 0 (D, R) such that
V˙ U (x, t) ≤ −W (x) ,
Using (18) and arguments similar to [26, Theorem 4.8], it can
be shown that all solutions x (·) ∈ St (Ωc × R≥t0 ) satisfy
x (t) ∈ B (0, r) on every interval of their existence. Therefore,
all solutions x (·) ∈ St (Ωc × R≥t0 ) are precompact, and
hence, complete, by Lemma 1. Since W is continuous and
positive definite, ∃δ > 0 such that B (0, δ) ⊂ Ωc . Since δ
is independent of t0 , uniform stability of (1) at x = 0 is
established. The rest of the proof is identical to [27, Section
5.3.2], and is therefore omitted.
In the following example, tests based on V˙ and Ṽ˙ are
inconclusive, but Theorem 5 can be invoked to conclude global
uniform asymptotic stability of the origin.
Example 4. Let H : R ⇒ R be defined as in Example 3 and
let F : R2 × R≥t0 ⇒ R2 be defined as
{−x1 + x2 (1 + g (t))} + H (x2 )
F (x, t) =
,
{−x1 − x2 } + H (x1 )
where g ∈ C 1 (R≥t0 , R), 0 ≤ g (t) ≤ 1, ∀t ∈ R≥t0 and ġ (t) ≤
g (t) , ∀t ∈ R≥t0 . Consider the differential inclusion in (1) and
the candidate Lyapunov function V : R2 × R≥t0 → R defined
as V (x, t) := x21 + (1 + g (t)) x22 . the candidate Lyapunov
2
2
2
function satisfies kxk2 ≤ V (x, t) ≤ 2 kxk2 , ∀ (x,
t) ∈ R ×
1
2
R≥t0 . In this case, since V ∈ C R × R≥t0 , R , similar to
[26, Example 4.20], the set-valued derivatives V˙ in [7] and
Ṽ˙ in [6] satisfy the bound
V˙ (x, t) , Ṽ˙ (x, t) ≤ −2x21 − 2x22 + 2x1 H (x2 )
+ 2x2 h (t) H (x1 ) , (19)
where h (t) := 1 + g (t) and the inequality 2 + 2g (t) − ġ (t) ≥
2 is utilized. Therefore, neither Ṽ˙ (x, t) nor V˙ (x, t) can be
shown to be negative semidefinite everywhere.
The function U1 : R2 × R≥t0 → R, defined as (see Fig. 1)
U1 (x, t) = max ((x1 − 1) , 0) − min ((x1 + 1) , 0)
+ max ((x2 − 1) , 0) − min ((x2 + 1) , 0) , (20)
(17)
∀x ∈ D and for almost all t ∈ R≥t0 , then (1) is uniformly
asymptotically stable at x = 0. Furthermore, if D = Rn and
∀t ∈ R≥t0 , satisfies U1 ∈ Lip R2 × R≥t0 , R . In addition,
since U1 is convex, it is also regular [24, Proposition 2.3.6].
9
The Clarke gradient of U1 is given by,
#)
("
sgn 1 (x1 )
× {0}
|x1 | =
6 1 ∧ |x2 | =
6 1,
sgn 1 (x2 )
#
"
co {0, sgn (x1 )}
× {0} |x1 | = 1 ∧ |x2 | =
6 1,
{sgn 1 (x2 )}
∂U1 (x, t) = "
#
{sgn
1
(x
)}
1
× {0} |x1 | =
6 1 ∧ |x2 | = 1,
co {0, sgn (x2 )}
"
#
co
{0,
sgn
(x
)}
1
× {0} |x1 | = 1 ∧ |x2 | = 1,
co {0, sgn (x2 )}
where ‘sgn 1’
{U1 } −reduced
F is given by
was introduced in Example 2. The
differential inclusion corresponding to
(
F (x, t) |x1 | =
6 1 ∧ |x2 | =
6 1,
F̃{U1 } (x, t) =
∅
otherwise.
The {U1 } −generalized time derivative of V with respect to
ẋ = F (x, t) is then given by
T
∂V
V˙ {U1 } (x) :=
max
(x, t)
[q; 1] ,
q∈F̃{U1 } (x,t) ∂ (x, t)
T
2x1
−x1 + x2 h (t)
6 1∧
|x1 | =
2x2 h (t) −x1 − x2 |x | =
1,
=
2 6
ġ (t) x22
1
−∞
otherwise,
2
≤ −2 kxk2 .
Theorem 5 can then be invoked to conclude that ẋ ∈ F (x, t)
is globally uniformly asymptotically stable at x = 0.
4
B. Invariance-like results
In applications such as adaptive control, Lyapunov methods
commonly result in semidefinite Lyapunov functions (i.e.,
candidate Lyapunov functions with time derivatives bounded
by a negative semidefinite function of the state). The following
theorem establishes the fact that if the function W in (17)
is positive semidefinite then t 7→ W (x (t)) asymptotically
decays to zero.
Theorem 6. Let 0 ∈ D and let Ω := D × R≥t0 . Let F : Rn ×
R≥t0 ⇒ Rn be a set-valued map with nonempty and compact
values that admits solutions over Ω. Let V ∈ Lip (Ω, R) be
a positive definite function that satisfies (16) and (17) where
W ∈ C 0 (D, R) is positive semidefinite. Select r > 0 such that
B (0, r) ⊂ D. If F is locally bounded, uniformly in t, over
Ω,5 then every solution x (·) ∈ St (Ωc × R≥t0 ), where Ωc :=
x ∈ B (0, r) | W (x) ≤ c and c ∈ 0, minkxk2 =r W (x) , is
complete, bounded, and satisfies limt→∞ W (x (t)) = 0.
Proof: Similar to the proof of [16, Corollary 1], it is
established that the bounds on V˙ F in (7) and (8) imply
5 A set-valued map F : Rn × R
n
≥0 ⇒ R is locally bounded, uniformly
in t, over Ω, if for every compact K ⊂ D, there exists M > 0 such that
∀ (x, t, y) such that (x, t) ∈ K × R≥t0 , and y ∈ F (x, t), kyk2 ≤ M .
that the cLf is nonincreasing along all the solutions to (1).
The nonincreasing property of the cLf is used to establish
boundedness of x (·), which is used to prove the existence and
uniform continuity of complete solutions. Barbălat’s lemma
[26, Lemma 8.2] is then used to conclude the proof. Let
x (·) ∈ St (Ωc × R≥t0 ). Using Theorem 2 and [16, Lemma
2],
V (x (t0 ) , t0 ) ≥ V (x (t) , t) ,
∀t ∈ ID .
(21)
Using (21) and arguments similar to [26, Theorem 4.8], it can
be shown that all solutions x (·) ∈ St (Ωc × R≥t0 ) satisfy
x (t) ∈ B (0, r) on every interval of their existence. Therefore,
all solutions x (·) ∈ St (Ωc × R≥t0 ) are precompact, and
hence, complete, by Lemma 1.
To establish uniform continuity of the solutions, it is observed that since F is locally bounded, uniformly in t, over
Ω, and x (t) ∈ B (0, r) on R≥t0 , the map t 7→ F (x (t) , t)
is uniformly bounded on R≥t0 . Hence, (2) implies that ẋ ∈
L∞ (R≥t0 ). Since t 7→ x (t) is locally absolutely continuous,
´t
∀t1 , t2 ∈ R≥t0 , kx (t2 ) − x (t1 )k2 = t12 ẋ (τ ) dτ . Since
2
´t
´t
ẋ ∈ L∞ , t12 ẋ (τ ) dτ ≤ t12 M dτ , where M is a positive
2
constant. Thus, kx (t2 ) − x (t1 )k2 ≤ M |t2 − t1 |, and hence,
t 7→ x (t) is uniformly continuous on R≥t0 .
Since x 7→ W (x) is continuous and B (0, r) is compact,
x 7→ W (x) is uniformly continuous on B (0, r). Hence,
t 7→ W (x´(t)) is uniformly continuous on R≥t0 . Furthert
more, t 7→ t0 W (x (τ )) dτ is monotonically increasing and
´t
from (17), t0 W (x (τ )) dτ ≤ V (x (t0 ) , t0 ) − V (x (t) , t)
´t
≤ V (x (t0 ) , t0 ). Hence, limt→∞ t0 W (x (τ )) dτ exists
and is finite. By Barbălat’s Lemma [26, Lemma 8.2],
limt→∞ W (x (t)) = 0.
In the following example V˙ and Ṽ˙ do not have a negative
semidefinite upper bound, but Theorem 6 can be invoked to
conclude partial stability.
Example 5. Let H : R ⇒ R be defined as in Example 3 and
let F : R2 × R≥t0 ⇒ R2 be defined as
F (x, t) =
{x2 (1 + g (t))} + H (x2 )
,
{−x1 − x2 } + H (x1 )
where g ∈ C 1 (R≥t0 , R), 0 ≤ g (t) ≤ 1, ∀t ∈ R≥t0 and ġ (t) ≤
g (t) , ∀t ∈ R≥t0 . Consider the differential inclusion in (1) and
the candidate Lyapunov function V : R2 × R≥t0 → R defined
as V (x, t) := x21 + (1 + g (t)) x22 . The candidate Lyapunov
2
2
2
function satisfies kxk2 ≤ V (x, t) ≤ 2 kxk2 , ∀ (x, t)
∈R ×
1
2
R≥t0 . In this case, since V ∈ C R × R≥t0 , R , the setvalued derivatives V˙ in [7] and Ṽ˙ in [6] are bounded by
V˙ (x, t) , Ṽ˙ (x, t) ≤ −2x22 + 2x2 h (t) H (x1 ) + 2x1 H (x2 ) ,
(22)
where h (t) := 1+g (t) and the inequality 2+2g (t)− ġ (t) ≥ 2
is utilized. Thus, neither Ṽ˙ nor V˙ are negative semidefinite
everywhere.
10
Let U1 be defined as in (20). The {U1 } −reduced differential
inclusion corresponding to F is given by
F (x, t)
|x1 | 6= 1 ∧ |x2 | 6= 1,
{0}× {−1}+ − 1 , 1 x = 1 ∧ x = 0,
1
2
2
1 21
F̃{U1} (x, t) =
{0}×
{1}+
−
,
x
=
−1
∧
x2 = 0,
1
2 2
∅
otherwise.
The {U1 } −generalized time derivative of V with respect to
ẋ = F (x, t) is then given by
T
∂V
˙
V {U1 } (x, t) :=
max
(x, t)
[q; 1] ,
q∈F̃{U1 } (x,t) ∂ (x, t)
T
x2 h (t)
2x1
|x1 | =
6 1∧
2x2 h (t) −x1 − x2
|x
|
=
6 1,
2
ġ (t) x22
1
h
i {0}
x1 = 1∧
max 2 0 0
32 , − 12 x = 0,
2
=
1
{0}
h
i
x1 = −1∧
max −2 0 0 21 , 32
x2 = 0,
1
−∞
otherwise,
≤ −2x22 .
Theorem 6 can then be invoked to conclude that t 7→ x1 (t) ∈
L∞ (R≥t0 , R) and limt→∞ x2 (t) = 0.
4
Theorem 6 and its counterparts are widely used in applications such as adaptive control to establish boundedness of the
system and convergence of the state to the origin. However,
since Theorem 6 is a partial stability result, it generally cannot
be used to establish convergence of the parameter estimates to
their true values. Under certain excitation conditions, parameter convergence can be established using auxiliary functions
[18]. Theorems that rely on such auxiliary functions fall under
the umbrella of Matrosov theorems, named after the seminal
work in [17]. The following section focuses on Matrosov
theorems.
C. Matrosov theorems
In this section, a less conservative generalization of Matrosov results for uniform asymptotic stability of nonautonomous systems is developed. In particular, the nonsmooth
version [21, Theorem 1] of the nested Matrosov theorem
[18, Theorem 1] is generalized. The following definitions of
Matrosov functions are inspired by [21].
Definition 10. Let γ, δ, ∆ > 0 be constants. A finite
set
M
of functions {Yj }j=1 ⊂ C 0 B (0m , γ) × D (δ, ∆) , R is said
to have the Matrosov property relative to (γ, δ, ∆) if ∀j ∈
{0, · · · , M },
(z, x) ∈ B (0m , γ) × D (δ, ∆) ∧
(Yi (z, x) = 0, ∀i ∈ {0, · · · , j}) =⇒ Yj+1 (z, x) ≤ 0,
where Y0 (z, x) = 0 and YM +1 (z, x) = 1, ∀ (z, x) ∈
B (0m , γ) × D (δ, ∆).
4
Definition 11. Let Ω := D × R≥t0 and let δ, ∆ > 0 be
constants such that D (δ, ∆) ⊂ D. Let F : Rn × R≥t0 ⇒ Rn
be a set-valued map with nonempty and compact values. The
M
functions {Wj }j=1 ⊂ Lip (Ω, R) are said to be Matrosov
functions for (F, δ, ∆) if ∃φ : Ω → Rm , γ > 0, and
M
{Yj }j=1 ⊂ C 0 B (0m , γ) × D (δ, ∆) , R such that
M
(a) The set of functions {Yj }j=1 has the Matrosov property
relative to (γ, δ, ∆),
(b) ∀j ∈ {1, · · · , M } and ∀ (x, t) ∈ D (δ, ∆) × R≥t0 ,
max {|Wj (x, t)| , |φ (x, t)|} ≤ γ,
(c) ∀j ∈ {1, · · · , M } there exists a collection of regular functions Uj ⊂ Lip (D (δ, ∆) × R≥t0 , R) such
˙ (x, t) ≤
that ∀ (x, t) ∈ D (δ, ∆) × R≥t0 , W
Uj
Yj (φ (x, t) , x)
4
The following technical Lemmas aid the proof of the Matrosov
theorem.
Lemma 2. Given δ > 0, ∃ > 0 such that
(z, x) ∈ B (0m , γ) × D (δ, ∆) ∧
(Yj (z, x) = 0, ∀j ∈ {1, · · · , M − 1}) =⇒ YM (z, x) ≤ −.
Proof: See [18, Claim 1].
Lemma 3. Let l ∈ {2, · · · , M }, ˜ > 0, and Ỹl ∈
C 0 (Rm × Rn , R). If
(z, x) ∈ B (0m , γ) × D (δ, ∆) ∧
(Yj (z, x) = 0, ∀j ∈ {1, · · · , l − 1}) =⇒ Ỹl (z, x) ≤ −˜
,
then ∃Kl−1 > 0 such that
(z, x) ∈ B (0m , γ) × D (δ, ∆) ∧
(Yj (z, x) = 0, ∀j ∈ {1, · · · , l − 2}) =⇒
˜
Kl−1 Yl−1 (z, x) + Ỹl (z, x) ≤ − .
2
Proof: See [18, Claim 2]
The Matrosov theorem can now be stated as follows.
Theorem 7. Let 0 ∈ D and let F : Rn × R≥t0 ⇒ Rn be
a set-valued map with nonempty and compact values such
that (1) is uniformly stable at x = 0. If for each pair of
numbers δ, ∆ ∈ R, such that 0 ≤ δ ≤ ∆ and D (δ, ∆) ⊂
D, there exist Matrosov functions for (F, δ, ∆), then (1) is
uniformly asymptotically stable at x = 0. If D = Rn and if
(1) is uniformly globally stable at x = 0 then (1) is uniformly
globally asymptotically stable at x = 0.
Proof: Select ∆ > 0 such that B (0, ∆) ⊂ D and let
r > 0 be such that x (·) ∈ St B (0, r) × R≥0 =⇒ x (t) ∈
B (0, ∆), ∀t ≥ t0 . Let ∈ (0, r) and select δ > 0 such that
x (·) ∈ St B (0, δ) × R≥0 =⇒ x (t) ∈ B (0, ), ∀t ≥ t0 .
By repeated application of Lemmas 2 and 3 it can be shown
that ∀δ > 0, ∃ζ > 0 and K1 , · · · , KM −1 > 0 such that
∀ (z, x) ∈ B (0m , γ) × D (δ, ∆),
Z (z, x) :=
M
−1
X
j=1
Kj Yj (z, x) + YM (z, x) ≤ −
ζ
2M −1
. (23)
11
Let W ∈ Lip (Ω, R) be defined as
W (x, t) :=
M
−1
X
Kj Wj (x, t) + WM (x, t) .
j=1
From Definition 11.b,
|V (x, t)| ≤ γ 1 +
M
−1
X
Kj =: η.
(24)
j=1
Fix (x0 , t0 ) ∈ B (0, r) × R≥0 and x (·) ∈ St ({(x0 , t0 )}).
The solution x (·) satisfies x (t) ∈ B (0, ∆), ∀t ∈ R≥t0 .
From Definition 11.c, V˙ Uj (x, t) ≤ Z (φ (x, t) , x), ∀ (x, t) ∈
D (δ, ∆) × R≥t0 , and hence, from Theorem 2,
V̇ (x (t) , t) ≤ Z (φ (x (t) , t) , x (t)) ,
for almost all t ∈ x
(23),
−1
(25)
Fig. 2. The function U2 .
(D (δ, ∆)). Using Definition 11.b and
Z (φ (x (t) , t) , x (t)) ≤ −
ζ
,
2M −1
(26)
for almost all t ∈ x−1 (D (δ, ∆)).
M
Let T > 2 ζ η . The claim is that kx (t)k ≤ , ∀t ≥ t0 + T .
If not, then x (t) ∈ D (δ, ∆), ∀t ∈ [t0 , t0 + T ]. Hence, from
(25) and (26),
ζ
V̇ (x (t) , t) ≤ − M −1 ,
(27)
2
for almost all t ∈ [t0 , t0 + T ]. Integrating (27) and using
Tζ
the bound in (24), 2M
≤ 2η, which contradicts T >
−1
2M η
(0, r), ∃T > 0 such that x (·) ∈
ζ . Hence, ∀ ∈
St B (0, r) × R≥0 =⇒ kx (t)k < , ∀t ≥ t0 + T , i.e.,
(1) is uniformly asymptotically stable at x = 0.
If D = Rn and if (1) is uniformly globally stable at x = 0
then r can be selected arbitrarily large, and hence, the result
is global
Example 6. Let H : R ⇒ R be defined as in Example 3
and let F : R2 × R≥t0 ⇒ R2 be defined as in Example 5.
Let U1 be defined as in (20). Let W1 : R2 × R≥t0 → R
be defined as W1 (x, t) := x21 + (1 + g (t)) x22 . It follows that
W˙ 1{U1 } (x, t) ≤ −2x22 , ∀x ∈ R2 , and uniform global stability
of (1) at x = 0 can be concluded from Theorem 5.
Let φ (x, t) = 0, ∀ (x, t) ∈ R2 × R≥t0 and let Y1 (z, x) :=
−2x22 , ∀ (z, x) ∈ R × R2 . Let W2 (x, t) := x1 x2 . The function
U2 : R2 × R≥t0 → R, defined as (see Fig. 2)
|x1 | > |x2 | ∧ x ∈ Sq,
|x1 |
U2 (x, t) = |x2 |
|x1 | ≤ |x2 | ∧ x ∈ Sq,
1 + U ∗ (x, t) x ∈
/ Sq,
∀t ∈ R≥t0 , where U ∗ (x, t) = max ((2x1 − 2) , 0) −
min ((2x1 + 2) , 0)
+
max ((2x2 − 2) , 0)
−
min ((2x2 + 2) , 0), ∀t
∈
R≥t0 and ’Sq’ denotes
the open unit square centered at the origin, satisfies
U2 ∈ Lip R2 × R≥t0 , R . In addition, since U2 is convex, it
is also regular [24, Proposition 2.3.6].
of U2 is given by
("
#)
2
sgn
1
(x
)
1
× {0}
2 sgn 1 (x2 )
"
#
co {0, 2 sgn (x )}
1
∂U2 (x, t) =
×{0}
{2 sgn 1 (x2 )}
"
#
{2
sgn
1
(x
)}
1
×{0}
co {0, 2 sgn (x2 )}
The Clarke gradient
|x1 | =
6 1 ∧ |x2 | =
6 1,
|x1 | = 1 ∧ |x2 | =
6 1,
|x1 | =
6 1 ∧ |x2 | = 1,
if x ∈
/ Sq,
#
"
[−1,
1]
× {0}
|x1 | = 0 ∧ |x2 | = 0,
[−1, 1]
"
#
sgn
(x
)
1
× {0}
|x1 | > |x2 | ,
{0}
∂U2 (x, t) = "
#
{0}
× {0}
|x1 | < |x2 | ,
sgn (x2 )
#
"
co
{0,
sgn
(x
)}
1
× {0} |x1 | = |x2 | > 0,
co {0, sgn (x2 )}
if x ∈ Sq, and
(
)
sgn
1
(x
)
,
1
co
( 2 sgn 1 (x1 )) × {0} |x1 | =
6 |x2 | ,
sgn 1 (x2 ) ,
co
2 sgn 1 (x2 )
∂U2 (x, t) = (
)
sgn (x1 ) ,
co
2 sgn (x1 ) , 0
× {0} |x1 | = |x2 | ,
(
)
sgn (x2 ) ,
co 2 sgn (x ) , 0
2
12
if x ∈ ∂Sq. The {U2 } −reduced differential inclusion corresponding to F is given by
6 1 ∧ |x2 | =
6 1,
F (x, t) x ∈
/ Sq ∧ |x1 | =
F (x, t) x ∈ Sq ∧ |x | =
|x2 | ,
1 6
F̃{U2 } (x, t) =
{0}
x
∈
Sq
∧
|x
|
=
0 ∧ |x2 | = 0,
1
∅
otherwise.
The {U2 } −generalized derivative of W2 is
T
x2 h (t)
x2
x1 −x1 − x2
1
0
˙
W2{U2 } (x, t) =
0
−∞
then given by
x∈
/ Sq∧
|x1 | =
6 1∧
|x2 | =
6 1∨
x ∈ Sq∧
|x1 | =
6 |x2 | ,
x ∈ Sq∧
|x1 | = 0∧
|x2 | = 0,
otherwise.
˙
2
2
That is, W
2{U2 } (x, t) ≤ −x1 − x2 x1 + 2x2 . If Y2 (z, x) :=
2
2
2
−x1 − x2 x1 + 2x2 , ∀ (z, x) ∈ R × R , then the functions
{Y1 , Y2 } have the Matrosov
property. Furthermore, since
W1 , W2 ∈ C 0 R2 × R≥t0 , R , ∀0 < δ < ∆, ∃γ > 0 such that
|W (x, t)| ≤ γ, ∀ (x, t) ∈ D (δ, ∆) × R≥t0 . Hence, {W1 , W2 }
are Matrosov functions for (F, δ, ∆), ∀0 < δ < ∆. Hence, by
Theorem 7, ẋ ∈ F (x, t) is uniformly globally asymptotically
stable at x = 0.
IX. C ONCLUSION
This paper demonstrates that locally Lipschitz regular functions can be used to identify infeasible directions in differential
inclusions. The infeasible directions can then be removed to
yield a point-wise smaller (in the sense of set containment)
equivalent differential inclusion. The reduction process is utilized to develop a novel generalization of the set-valued derivative for locally Lipschitz candidate Lyapunov functions. Less
conservative statements of Lyapunov stability and invariancelike results for differential inclusions are developed based on
the reduction using locally Lipschitz regular functions.
The fact that arbitrary locally Lipschitz regular functions
can be used to reduce differential inclusions to smaller sets
of admissible directions indicates that there may be a smallest
set of admissible directions corresponding to each differential
inclusion. Further research is needed to establish the existence
of such a set and to find a representation of it that facilitates
computation.
R EFERENCES
[1] A. F. Filippov, Differential equations with discontinuous right-hand
sides. Kluwer Academic Publishers, 1988.
[2] N. N. Krasovskii and A. I. Subbotin, Game-theoretical control problems.
New York: Springer-Verlag, 1988.
[3] E. Roxin, “Stability in general control systems,” J. Differ. Equ., vol. 1,
no. 2, pp. 115–150, 1965.
[4] B. E. Paden and S. S. Sastry, “A calculus for computing Filippov’s
differential inclusion with application to the variable structure control
of robot manipulators,” IEEE Trans. Circuits Syst., vol. 34, no. 1, pp.
73–82, Jan. 1987.
[5] J. P. Aubin and A. Cellina, Differential inclusions. Springer, Berlin,
1984.
[6] D. Shevitz and B. Paden, “Lyapunov stability theory of nonsmooth
systems,” IEEE Trans. Autom. Control, vol. 39 no. 9, pp. 1910–1914,
1994.
[7] A. Bacciotti and F. Ceragioli, “Stability and stabilization of
discontinuous systems and nonsmooth Lyapunov functions,” ESAIM
Control Optim. Calc. Var., vol. 4, pp. 361–376, 1999.
[8] Q. Hui, W. Haddad, and S. Bhat, “Semistability, finite-time stability,
differential inclusions, and discontinuous dynamical systems having a
continuum of equilibria,” IEEE Trans. Autom. Control, vol. 54, no. 10,
pp. 2465–2470, Oct. 2009.
[9] F. M. Ceragioli, “Discontinuous ordinary differential equations and
stabilization,” Ph.D. dissertation, Universita di Firenze, Italy, 1999.
[10] A. N. Michel and K. Wang, Qualitative theory of dynamical systems,
the role of stability preserving mappings. New York: Marcel Dekker,
1995.
[11] E. Moulay and W. Perruquetti, “Finite time stability of differential
inclusions,” IMA J. Math. Control Info., vol. 22, pp. 465–275, 2005.
[12] E. Ryan, “An integral invariance principle for differential inclusions with
applications in adaptive control,” SIAM J. Control Optim., vol. 36, no. 3,
pp. 960–980, 1998.
[13] H. Logemann and E. Ryan, “Asymptotic behaviour of nonlinear systems,” Amer. Math. Month., vol. 111, pp. 864–889, 2004.
[14] A. Bacciotti and L. Mazzi, “An invariance principle for nonlinear
switched systems,” Syst. Control Lett., vol. 54, pp. 1109–1119, 2005.
[15] W. M. Haddad, V. Chellaboina, and S. G. Nersesov, Impulsive and hybrid
dynamical systems. Princeton Series in Applied Mathematics, 2006.
[16] N. Fischer, R. Kamalapurkar, and W. E. Dixon, “LaSalle-Yoshizawa
corollaries for nonsmooth systems,” IEEE Trans. Autom. Control,
vol. 58, no. 9, pp. 2333–2338, Sep. 2013.
[17] V. M. Matrosov, “On the stability of motion,” J. Appl. Math. Mech.,
vol. 26, pp. 1337–1353, 1962.
[18] A. Lorı́a, E. Panteley, D. Popovic, and A. R. Teel, “A nested Matrosov
theorem and persistency of excitation for uniform convergence in stable
nonautonomous systems,” IEEE Trans. Autom. Control, vol. 50, no. 2,
pp. 183–198, 2005.
[19] R. Sanfelice and A. R. Teel, “Asymptotic stability in hybrid systems via
nested Matrosov functions,” IEEE Trans. Autom. Control, vol. 54, no. 7,
pp. 1569–1574, Jul. 2009.
[20] B. Paden and R. Panja, “Globally asymptotically stable ’PD+’ controller
for robot manipulators,” Int. J. Control, vol. 47, no. 6, pp. 1697–1712,
1988.
[21] A. R. Teel, D. Nešić, T.-C. Lee, and Y. Tan, “A refinement of Matrosov’s
theorem for differential inclusions,” Automatica, vol. 68, pp. 378–383,
2016.
[22] E. Ryan, “Discontinuous feedback and universal adaptive stabilization,”
in Control of Uncertain systems. Springer, 1990, pp. 245–258.
[23] R. T. Rockafellar and R. J.-B. Wets, Variational analysis. Springer
Science & Business Media, 2009, vol. 317.
[24] F. H. Clarke, Optimization and nonsmooth analysis. SIAM, 1990.
[25] J. J. Moreau and M. Valadier, “A chain rule involving vector functions
of bounded variation,” J. Funct. Anal., vol. 74, no. 2, pp. 333–345,
1987.
[26] H. K. Khalil, Nonlinear systems, 3rd ed. Upper Saddle River, NJ:
Prentice Hall, 2002.
[27] M. Vidyasagar, Nonlinear systems analysis, 2nd ed. SIAM, 2002.
[28] J. Alvarez, Y. Orlov, and L. Acho, “An invariance principle for discontinuous dynamic systems with applications to a Coulomb friction
oscillator,” ASME J. Dyn. Syst. Meas. Control, vol. 74, pp. 190–198,
2000.
13
Rushikesh Kamalapurkar received his M.S. and
Ph.D. degrees in 2011 and 2014, respectively, from
the Mechanical and Aerospace Engineering Department at the University of Florida. After working
for a year as a postdoctoral research fellow with
Dr. Warren E. Dixon, he was selected as the 201516 MAE postdoctoral teaching fellow. In 2016 he
joined the School of Mechanical and Aerospace
Engineering at the Oklahoma State University as an
Assistant professor. His primary research interest has
been intelligent, learning-based optimal control of
uncertain nonlinear dynamical systems. He has published 3 book chapters, 18
peer reviewed journal papers and 19 peer reviewed conference papers. His
work has been recognized by the 2015 University of Florida Department of
Mechanical and Aerospace Engineering Best Dissertation Award, and the 2014
University of Florida Department of Mechanical and Aerospace Engineering
Outstanding Graduate Research Award.
Warren E. Dixon received his Ph.D. in 2000 from
the Department of Electrical and Computer Engineering from Clemson University. He worked as a
research staff member and Eugene P. Wigner Fellow
at Oak Ridge National Laboratory (ORNL) until
2004, when he joined the University of Florida in the
Mechanical and Aerospace Engineering Department.
His main research interest has been the development
and application of Lyapunov-based control techniques for uncertain nonlinear systems. His work
has been recognized by the 2015 & 2009 American
Automatic Control Council (AACC) O. Hugo Schuck (Best Paper) Award, the
2013 Fred Ellersick Award for Best Overall MILCOM Paper, a 2012-2013
University of Florida College of Engineering Doctoral Dissertation Mentoring
Award, the 2011 American Society of Mechanical Engineers (ASME) Dynamics Systems and Control Division Outstanding Young Investigator Award, the
2006 IEEE Robotics and Automation Society (RAS) Early Academic Career
Award, an NSF CAREER Award, the 2004 Department of Energy Outstanding
Mentor Award, and the 2001 ORNL Early Career Award for Engineering
Achievement. He is a Fellow of both ASME and IEEE, an IEEE Control
Systems Society (CSS) Distinguished Lecturer, and served as the Director of
Operations for the Executive Committee of the IEEE CSS Board of Governors
(2012-2015). He was awarded the Air Force Commander’s Public Service
Award (2016) for his contributions to the U.S. Air Force Science Advisory
Board. He is currently or formerly an associate editor for ASME Journal of
Journal of Dynamic Systems, Measurement and Control, Automatica, IEEE
Transactions on Systems Man and Cybernetics: Part B Cybernetics, and the
International Journal of Robust and Nonlinear Control.
Andrew R. Teel received his A.B. degree in
Engineering Sciences from Dartmouth College in
Hanover, New Hampshire, in 1987, and his M.S.
and Ph.D. degrees in Electrical Engineering from
the University of California, Berkeley, in 1989 and
1992, respectively. After receiving his Ph.D., he was
a postdoctoral fellow at the Ecole des Mines de
Paris in Fontainebleau, France. In 1992 he joined the
faculty of the Electrical Engineering Department at
the University of Minnesota, where he was an assistant professor until 1997. Subsequently, he joined
the faculty of the Electrical and Computer Engineering Department at the
University of California, Santa Barbara, where he is currently a Distinguished
Professor and director of the Center for Control, Dynamical systems, and
Computation. His research interests are in nonlinear and hybrid dynamical
systems, with a focus on stability analysis and control design. He has received
NSF Research Initiation and CAREER Awards, the 1998 IEEE Leon K.
Kirchmayer Prize Paper Award, the 1998 George S. Axelby Outstanding Paper
Award, and was the recipient of the first SIAM Control and Systems Theory
Prize in 1998. He was the recipient of the 1999 Donald P. Eckman Award
and the 2001 O. Hugo Schuck Best Paper Award, both given by the American
Automatic Control Council, and also received the 2010 IEEE Control Systems
Magazine Outstanding Paper Award. In 2016, he received the Certificate of
Excellent Achievements from the IFAC Technical Committee on Nonlinear
Control Systems. He is Editor-in-Chief for Automatica, and a Fellow of the
IEEE and of IFAC.
| 3cs.SY
|
An Algorithmic Framework for Labeling Network Maps?
Jan-Henrik Haunert1 and Benjamin Niedermann2
1
arXiv:1505.00164v2 [cs.CG] 30 Dec 2015
2
University of Osnabrück, Germany
Karlsruhe Institute of Technology, Germany
Abstract. Drawing network maps automatically comprises two challenging steps, namely laying
out the map and placing non-overlapping labels. In this paper we tackle the problem of labeling an
already existing network map considering the application of metro maps. We present a flexible
and versatile labeling model that subsumes different labeling styles. We show that labeling a
single line of the network is NP-hard, even if we make very restricting assumptions about the
labeling style that is used with this model. For a restricted variant of that model, we then
introduce an efficient algorithm that optimally labels a single line with respect to a given cost
function. Based on that algorithm, we present a general and sophisticated workflow for multiple
metro lines, which is experimentally evaluated on real-world metro maps.
1
Introduction
Label placement and geographic network visualization are classical problems in cartography, which
independently of each other have received the attention of computer scientists. Label placement usually
deals with annotating point, line or area features of interest in a map with text labels such that the
associations between the features and the labels are clear and the map is kept legible [8]. Geographic
network visualization, on the other hand, often aims at a geometrically distorted representation of
reality that allows information about connectivity, travel times, and required navigation actions to be
retrieved easily. Computing a good network visualization is thus related to finding a layout of a graph
with certain favorable properties [15]. For example, to avoid visual clutter in metro maps, an octilinear
graph layout is often chosen, in which the orientation of each edge is a multiple of 45◦ [11,12,14].
Alternatively, one may choose a curvilinear graph layout, that is, to display the metro lines as curves
[5,13].
Computing a graph layout for a metro map and labeling the stops have been considered as two different problems that can be solved in succession [14], but also integrated solutions have been suggested
[11,12]. Nevertheless, in practice, metro maps are often drawn manually by cartographers or designers,
as the existing algorithms do not achieve results of sufficient quality in adequate time. For example,
Nöllenburg and Wolff [11] report that their method needed 10 hours and 31 minutes to compute a
labeled metro map of Sydney that they present in their article, while an unlabeled map for the same
instance was obtained after 23 minutes—both results were obtained without proof of optimality but
with similar optimality gaps. On the other hand Wang and Chi [14] present an algorithm that creates
the graph layout and labeling within one second, but they cannot guarantee that labels do not overlap
each other or the metro lines.
An integrated approach to computing a graph layout and labeling the stops allows consideration
to be given to all quality criteria of the final visualization. On the other hand, treating both problems
separately will probably reduce computation time. Moreover, we consider the labeling of a metro map
as an interesting problem on its own, since, in some situations, the layout of the network is given as
part of the input and must not be changed. In a semi-automatic workflow, for example, a cartographer
may want to draw or alter a graph layout manually before using an automatic method to place labels,
?
A preliminary version of this paper has appeared in Proc. 21st Int. Conf. on Computing Combinatorics
(COCOON’15), volume 9198 of Lect. Notes Comput. Sci., pages 689–700, Springer-Verlag. This research
was initiated during Dagstuhl Seminar 13151 “Drawing Graphs and Maps with Curves” in April 2013
probably to test multiple different labeling styles with the drawing. Hence, a labeling algorithm is
needed that is rather flexible in dealing with different labeling styles.
In this paper, we are given the layout of a metro map consisting of several metro lines on which
stops (also called stations) are located. For each stop we are further given its name, which should
be placed close to its position. We first introduce a versatile and general model for labeling metro
maps; see Section 2. Like many labeling algorithms for point sets [1,3,6], our algorithm uses a discrete
set of candidate labels for each point. Often, each label is represented by a rectangle wrapping the
text. Since we also want to use curved labels, however, we represent a label by a simple polygon that
approximates a fat curve, that is, a curve of certain width reflecting the text height. We then prove
that even in that simple model labeling a single metro line is NP-hard considering different labeling
styles. Hence, we restrict the set of candidates satisfying certain properties, which allows us to solve the
problem on one metro line C in O(n2 ) time, where n is the number of stops of C; see Section 4. This
algorithm optimizes the labeling with respect to a cost function that is based on Imhof’s [8] classical
criteria of cartographic quality. Utilizing that algorithm, we present an efficient heuristic for labeling
a metro map consisting of multiple metro lines; see Section 6. Our method is similar to the heuristic
presented by Kakoulis and Tollis [9], in the sense that it discards some label candidates to establish
a set of preconditions that allow for an efficient exact solution. Our model of quality is more general
than the one of Kakoulis and Tollis, however, as it not only takes the quality of individual labels but
also the quality of pairs of labels for consecutive metro stations into account. Finally, we evaluate our
approach presenting experiments conducted on realistic metro maps; see Section 8. Note that “stops”
on “metro lines” can refer more generally to points of interest on the lines of any kind of a network
map. We address labeling styles for octilinear graph layouts and curvilinear graph layouts that use
Bézier curves. The more general model behind our method, however, subsumes but is not limited to
these particular styles.
2
Labeling Model
We assume that the metro lines are given by directed, non-self-intersecting curves in the plane described
by polylines, which for example have been derived by approximating Bézier curves. We denote that
set of metro lines by M. Further, the stops of each metro line C ∈ M are given by an ordered set SC
of points on C going from the beginning to the end of C. For two stops s, s0 ∈ SC we write s < s0 if s
lies before s0 . We denote the union of the stops among all metro lines by S and call the pair (M, S) a
metro map.
For each stop s ∈ S we are further given a name that should be placed close to it. In contrast to
previous work, we do not follow traditional map labeling abstracting from the given text by bounding
boxes. Instead we model a label ` of a stop s ∈ S as a simple polygon. For example, a label could have
been derived by approximating a fat curve prescribing the name of the stop;Ssee Fig. 1. For each stop s
we are given a set Ks of labels, which we also call candidates of s. The set s∈S Ks is denoted by K.
Since “names should disturb other map content as little as possible”[8], we strictly forbid overlaps
between labels and lines as well as label-label overlaps. Further, each stop must be labeled. Hence, a
set L ⊆ K is called a labeling if (1) no two labels of L intersect each other, (2) no label ` ∈ L intersects
any metro line C ∈ M, and (3) for each stop s ∈ S there is exactly one label ` ∈ L ∩ Ks .
Definition 1 (MetroMapLabeling).
Given: Metro map (M, S), candidates K and cost function w : 2K → R+ .
Find, if it exists: Optimal labeling L of (M, S, K, w), i.e., w(L) ≤ w(L0 ) for any labeling L0 ⊆ K.
The model allows us to create arbitrarily shaped label candidates for a metro map. In our evaluation
we have considered two different labeling styles. The first style, OctilinStyle, creates for each stop
a set of octilinear rectangles as label candidates; see Fig. 2. We use that style for octilinear maps. The
second style, CurvedStyle, creates for each stop a set of fat Bézier curves as label candidates, which
are then approximated by simple polygons; see Fig. 1. We use that style for curvilinear metro maps,
lm
n
L−1
{
1
2 v2
v1
{
1
2 v1
p2
`τ
p3 v2
bτ
L1
lm
2
lm
4
p4
s
C
n
p1
C
(a)
(b)
Fig. 1: Construction of curved candidates. (a) Construction of a single label. (b) Candidates Ks =
L1 ∪ L−1 for stop s.
in order to adapt the curvilinear style of the metro map. The basic idea is that a label perpendicularly
emanates from a stop with respect to its metro line and then becomes horizontal to sustain legibility.
In the following section we motivate our choice of candidates based on cartographic criteria and give
detailed technical descriptions for both labeling styles.
2.1
Two Examples of Labeling Styles
We extracted the rules for generating label candidates from Imhof’s “general principles and requirements” for map labeling [8]. For schematic network maps, the need for “legibility” implies that we must
not destroy the underlying design principle with clutter. To this end, we generate candidate labels that
adhere to the schematics of the network. That is, we use straight horizontal and diagonal labels with
octilinear layouts and curved labels with curvilinear layouts.
We now describe more precisely, how we defined the labeling styles CurvedStyle and OctilinStyle, which are used for curvilinear layouts and octilinear layouts, respectively.
Curvilinear Metro Maps. For CurvedStyle, assume that the given metro map is curvilinear.
In order to achieve a “clear graphic association” between a label and the corresponding point p, we
construct the simple polygon prescribing a candidate label based on a curve ` (possibly a straight-line
segment) that emanates from p. The candidate label itself is a continuous section of ` that does not
directly start in p but at a certain configurable distance from it. We define the end of the candidate
label on ` based on the text length and assign a non-zero width to the curve section to represent the text
height. In the case that p lies on a single curved line C, we require that ` and C are perpendicular in p
to enhance the angular resolution of the final drawing. By bending ` towards the horizontal direction,
we avoid steep labels. We approximate ` by a simple polygon consisting of a constant number of line
segments.
We now describe the construction of a single candidate more specifically. For each stop s of each
metro line C we create a constant number of curved labels adapting the curvilinear style of the metro
map. The basic idea is that a label perpendicularly emanates from s with respect to C and then
becomes horizontal to sustain legibility; see Fig. 1. Let n = (nx , ny ) be the normalized normal vector
of C at s. Further, let d ∈ {−1, 1} and c1 , c2 ∈ R+ be pre-defined constants. For τ = (c1 , c2 , d) we
define the fat cubic Bézier curve bτ by the following four control points; see Fig. 1. p1 = s, p2 =
s + 0.5 · v 1 , p3 = s + v 1 + 0.5 · v 2 , p4 = s + v 1 + v 2 , where v 1 = c1 · n, v 2 = sgn(n) · (d · c2 , 0),
and sgn(n) = 1 if nx > 0 and sgn(n) = −1 otherwise. We define the thickness of bτ to be the predefined height of a label. Let `τ be the sub-curve of bτ that starts at p1 and has the length of the name
of s and let `0τ be the curve when mirroring `τ at s. Further, let lm be the length of the longest name of
lm
a stop in S and let Ld = {`τ , `0τ | τ ∈ {(lm , lm , d), ( lm
2 , lm , d), ( 4 , lm , d)}}. If n has an orientation less
◦
than or equal to 60 , we set Ks = L1 and otherwise Ks = L1 ∪L−1 . Hence, if n is almost vertical and C
`5
`3
`03
`05
`05
`4
`2
p
`1
s l 0
`1
`02
(a)
`04
`2
`01 `0
2
p
s
(b)
`1
l
l
`1
s
`03
`3
`2
`02
(c)
l
l0
l0
s
l0
l
s
l
s
(d)
(e)
(f)
Fig. 2: Construction of octilinear candidates for a stop s. (a) s lies on a horizontal segment. (b) s lies on
a diagonal segment. (c) s lies on a vertical segment. (d) s lies on a crossing of two diagonal segments l
and l0 . (e) s lies on a crossing of a vertical and horizontal segment. (f) s lies on a crossing of a vertical
and diagonal segment.
is therefore almost horizontal at s, we also add the labels L−1 pointing into the opposite x-direction
than n. In our experiments we did not let the labels start at s, but with a certain offset to s, in order
to avoid intersections with C.
Octilinear Metro Maps. For OctilinStyle assume that the metro map is octilinear. We model
the labels as horizontal and diagonal rectangles. Let l be the line segment of C on which s lies and
let R be an axis-aligned rectangle that is the bounding box of the name of s. Further, let c be a circle
around s with a pre-defined radius. We place the labels such that they touch the border of c, but they
do not intersect the interior of c. Hence, the labels have a pre-defined offset to s.
If l is horizontal, we place five copies `1 , . . . , `5 of R above l as follows; see Fig. 2(a). We place `1 , `2
and `3 such that the left-bottom corner of `1 , the midpoint of `2 ’s bottom edge and the right-bottom
corner of `3 coincides with the topmost point of c. We rotate `4 by 45◦ counterclockwise and place it at c
such that the midpoint of its left side touches c, i.e., that midpoint lies on a diagonal through s. Finally,
`5 is obtained by mirroring `4 at the vertical line through s. Mirroring `1 , . . . , `5 at the horizontal line
through s, we obtain the rectangles `01 , . . . , `05 , respectively. We then set Ks = {`i , `0i | 1 ≤ i ≤ 5}.
If l is diagonal, we create the candidates in the same manner as in the case that l is horizontal; see
Fig. 2(b). However, we only create the candidates `2 and `02 and the candidates that are horizontally
aligned.
If l is vertical, we place three copies `1 , . . . , `3 of R to the right of l as follows; see Fig. 2(c). We
rotate `1 by 45◦ counterclockwise and `2 by 45◦ clockwise. We place `1 , `2 , `3 at c such that the
midpoints of their left edges touch c. Mirroring `1 , `2 , `3 at the vertical line through s, defines the
rectangles `01 , `02 and `03 . We set Ks = {`i , `0i | 1 ≤ i ≤ 3}.
In case that s is a crossing of two metro lines, we create the candidates differently. If s is the crossing
of two diagonals, we create the candidates as shown in Fig. 2(d). If s is the crossing of a horizontal
and a vertical segment, we create the labels as shown in Fig. 2(e). If s is the crossing of a diagonal
and horizontal segment, we create the labels as shown in Fig. 2(f). We analogously create the labels,
if s is the crossing of a vertical and a diagonal segment.
Remark: If a stop lies on multiple metro lines, then we can apply similar constructions, where the
labels are placed on the angle bisectors of the crossing lines.
3
Computational Complexity
We first study the computational complexity of MetroMapLabeling assuming that the labels are
either based on OctilinStyle or CurvedStyle. In particular we show that the problem is NP-hard,
if the metro map consists of only one line. The proof uses a reduction from the NP-complete problem
monotone planar 3SAT [10]. Based on the given style, we create for the set C of 3SAT clauses a metro
map (M, S) such that (M, S) has a labeling if and only if C is satisfiable. The proof can be easily
adapted to other labeling styles. Note that the complexity of labeling points using a finite set of axisaligned rectangular label candidates is a well-studied NP-complete problem, e.g., see [7,6]. However,
c2
c1
x2
F
x3
c3
F
x4
Q
R
x5
x1
h
c4
P
Fig. 3: Illustration of NP-completeness proof. (a) 3-SAT formula ϕ with clauses c1 = x4 ∨ x1 ∨ x5 ,
c2 = x2 ∨ x4 ∨ x3 , c3 = x̄2 ∨ x̄1 ∨ x̄3 and c4 = x̄3 ∨ x̄5 ∨ x̄4 represented as a metro map. Truth assignment
is x1 = true, x2 = true, x3 = false, x4 = false and x5 = false. The gray graph G represents the
adjacencies of the used gadgets. The solid lines represent a spanning tree of G that can be used to
merge the polygons to one simple polygon. Exemplarily the polygons P and Q are merged into one
polygon R. The single components are illustrated in Fig. 4.
since we do not necessarily use axis-aligned rectangles as labels and since for the considered labeling
styles the labels are placed along metro lines, it is not obvious how to reduce a point-feature labeling
instance on an instance of MetroMapLabeling. In order to show the NP-hardness, we prove that
it is NP-complete to decide whether a metro map (M, S) has a labeling based on the given labeling
style.
Theorem 1. MetroMapLabeling is NP-complete, if the labels are based on OctilinStyle or CurvedStyle, even if the metro map has only one metro line.
Proof. For the illustrations we use OctilinStyle, but the same constructions can be done based
on CurvedStyle; see end of proof. We first show that the problem deciding whether (M, S) has a
labeling lies in NP. We first create for each stop s ∈ S its candidates Ks based on the given labeling
style. Recall that each candidate has constant size. We then guess for each stop s ∈ S the label `s that
belongs to the desired labeling L. Obviously, we can decide in polynomial time whether {`s | s ∈ S} is
a labeling of (M, S) performing basically intersection tests.
We now perform a reduction from the NP-complete Planar Monotone 3-Sat problem [10]. Let
ϕ be a Boolean formula in conjunctive normal form such that it consists of n variables and m clauses
and, furthermore, each clause contains at most three literals. The formula ϕ induces the graph Gϕ as
follows. Gϕ contains for each variable a vertex and it contains for each clause a vertex. Two vertices u
and v are connected by an edge {u, v} ∈ E if and only if u represents a variable x and v represents
a clause c, such that x is contained in c. We call a clause of ϕ positive (negative) if it contains only
positive (negative) literals.
The formula ϕ is an instance of Planar Monotone 3-Sat if it satisfies the following requirements.
1. ϕ is monotone, i.e., each clause is either positive or negative, and
2. the graph Gϕ is planar and has a rectilinear plane embedding such that
(a) the vertices representing variables are placed on a horizontal line h,
(b) the vertices representing negative clauses are placed below h,
(c) the vertices representing positive clauses are placed above h,
(a)
(b)
s1
`2
1
Negative
s2
s2
`1
1
`1
2
`2
1
`2
2
s3
`1
3
s3
`1
4
s4
`1
1
Positive
F
`2
4
(c)
F
s1
`2
3
c
(d)
x
x
F1
F2
H
x
F2
F1
x
x
x
Fig. 4: Illustration of the gadgets. Selectable labels are filled, while all other labels are not filled. Ports
are marked with a dashed square. (a) Chain gadget of length 4. (b) Fork gadget. (c) Clause gadget.
(d) Variable gadget with three negative and three positive ports.
(d) the edges are drawn on their respective side of h.
Planar Monotone 3-Sat then asks whether ϕ is satisfiable.
Using only stops lying on single horizontal and vertical segments, we construct a metro map (M, S)
that mimics the embedding of Gϕ . In particular M will only consist of one metro line C that connects
all stops such that the stops and their candidates simulate the variables and clauses of ϕ. We will
prove that (M, S) has a labeling if and only if ϕ is satisfiable. We refer to Fig. 3 for a sketch of
the construction. We first define gadgets simulating variables, clauses and connecting structures. Each
gadget consists of a set of stops that lie on the border of a simple polygon P . Later on we use this
polygon P to prescribe the shape of the metro line C.
Chain. The chain gadget represents and transmits truth values from variables to clauses mimicking
the embeddings of the edges in Gϕ . A chain consists of an even number of stops s1 , . . . , sk that lie
on vertical and horizontal segments; see Fig. 4(a). Hence, with respect to the given labeling style
each stop si has a predefined set of candidates Ksi . For each stop si there are two specially marked
candidates `1i and `2i that lie on opposite sides of si ’s segment; for an example see the filled blue labels
in Fig 4(a). We say that those labels are selectable, because we define the gadget such that those labels
are the only labels that can be selected for a labeling. To that end, we lay out the metro line such that
it does not intersect any selectable label, but all labels that are not selectable. The stops are placed
such that the following conditions are satisfied.
(1) The label `2i intersects the label `1i+1 for 1 ≤ i < k.
(2) Except the intersections mentioned in (1), there is no intersection between selectable labels of
different stops.
(3) The segments of the stops are connected by polylines s.t. the result is a simple polygon P intersecting all labels except the selectable labels.
The labels `11 and `2k do not intersect any selectable labels; we call them the ports of the chain. Later
on, we use the ports to connect other gadgets with the chain, i.e., we arrange the gadgets such that
two of their ports intersect, but no other selectable label. Further, we assign a polarization to each
selectable label. The labels `11 , . . . , `1k are negative and the labels `21 , . . . , `2k are positive.
Consider a labeling L of a chain assuming that P is interpreted as a metro line; we can cut P at
some point in order to obtain an open curve. By construction of P only selectable labels are contained
in L. In particular we observe that if the negative port `11 is not contained in L, then the positive
labels `21 , . . . , `2k belong to L. Analogously, if the positive port `2k is not contained in L, then the
negative labels `11 , . . . , `1k belong to L. We use this behavior to represent and transmit truth values
through the chain.
Fork. The fork gadget splits an incoming chain into two outgoing chains and transmits the truth
value represented by the incoming chain into the two outgoing chains. A fork consists of three stops s1 ,
s2 and s3 such that s1 and s2 are placed on vertical segments and s3 is placed on a horizontal segment;
see Fig. 4(b). Analogously to the chain, each stop si with (1 ≤ i ≤ 3) has two selectable labels `1i and
`2i . We arrange the stops such that the following conditions are satisfied.
1. The labels `12 and `13 intersect `21 . Apart from those two intersections no selectable label intersects
any other selectable label.
2. The segments of the stops are connected by polylines s.t. the result is a simple polygon P intersecting all labels except the selectable labels.
The label `11 is the incoming port and the labels `22 and `23 are the outgoing ports of the fork. We
distinguish two types of forks by assigning different polarizations to the selectable labels. In the negative
(positive) fork, the labels `11 , `12 and `13 are positive (negative) and the labels are `21 , `22 and `23 are
negative (positive). Hence, the incoming port is positive (negative) and the outgoings ports are negative
(positive).
Consider a labeling L of a fork assuming that P is interpreted as a metro line. By construction
of P only selectable labels belong to L. Further, if the incoming port `11 does not belong to L, then
the outgoing ports `21 and `21 belong to L. Finally, if one outgoing port does not belong to L, then the
incoming port belongs to L.
Clause. The clause gadget represents a clause c of the given instance. It forms a chain of length 2
with the addition that it has three ports instead of two ports; see Fig 4(c). To that end one of both
stops has three selectable labels; one intersecting a selectable label of the other stop, and two lying
on the opposite side of the stop’s segment without intersecting any selectable label of the other stop.
The gadget is placed at the position where the vertex of c is located in the drawing of Gϕ ; see Fig. 3.
We observe that a labeling L of a clause gadget always contains at least one port. Further, we do not
assign any polarization to its selectable labels.
Variable. The variable gadget represents a single variable x. It forms a composition of chains and
forks that are connected by their ports; see Fig 4(d). More precisely, let s be the number of clauses in
which the negative literal x̄ occurs and let t be the number of clauses in which the positive literal x
occurs. Along the horizontal line h on which the vertex of x is placed in the drawing of Gϕ , we place
a horizontal chain H. Further, we place a sequence of negative forks F 1 , . . . , F s−1 to the left of H
and a sequence of positive forks F1 , . . . , Ft−1 to the right of H. The negative incoming port of F1 is
connected to the positive port of H by a chain. Two consecutive forks Fi and Fi+1 are connected by a
chain H 0 such that H 0 connects a positive outgoing port of Fi with the negative incoming port of Fi+1 .
Analogously, the positive incoming port of F 1 is connected to the negative port of H by a chain. Two
consecutive forks F i and F i+1 are connected by a chain H 0 such that H 0 connects a negative outgoing
port of F i with the positive incoming port of F i+1 .
We observe that the gadget has s + t free ports. Further, we can arrange the forks such that the
free ports of F1 , . . . , Ft−1 lie above h and the free ports of F 1 , . . . , F s−1 lie below h.
Consider a labeling L of a variable. By construction of the forks and chains, if one positive free
port is not contained in L, then all negative free ports must be contained in L. Analogously, if one
negative free port is not contained in L, then all positive free ports must be contained in L.
Using additional chains we connect the positive free ports with the positive clauses and the negative
ports with the negative clauses correspondingly; see Fig. 3. More precisely, assume that the variable x
is contained in the positive clause c; negative clauses can be handled analogously. With respect to
the drawing of Gϕ , a positive free port of x’s gadget is connected with the negative port of a chain
whose positive port is connected with a free port of c’s gadget. Note that we can easily choose the
simple polygons enclosing the gadgets such that they do not intersect by defining them such that they
surround the gadgets tightly.
(a)
(b)
(c)
Fig. 5: Illustration of the gadgets based on OctilinStyle. Selectable labels are filled, while all other
labels are not filled. (a) Chain gadget of length 4. (b) Fork gadget. (c) Clause gadget.
One Metro Line. We construct the polygons enclosing the single gadgets such that they do not
intersect each other. We now sketch how the polygons can be merged to a single simple polygon P .
Cutting this polygon at some point we obtain a polyline prescribing the desired metro line.
We construct a graph H = (V, E) as follows. The polygons of the gadgets are the vertices of the
graph and an edge (P, Q) is contained in E if and only if the corresponding gadgets of the polygons P
and Q are connected by their ports; see Fig. 3. Since Gϕ is planar and the gadgets mimic the embedding
of Gϕ , it is not hard to see that H is also planar. We construct a spanning tree T of H. If an edge
(P, Q) of H is also contained in T , we merge P and Q obtaining a new simple polygon R; see for an
example Fig. 3. To that end we cut P and Q in polylines and connect the four end points by two new
polylines such that the result is a simple polygon. We in particular ensure that the new polygon does
not intersect any other polygon and that R intersects the same labels as P and Q together. In T we
correspondingly contract the edge. Note that by contracting edges, T remains a tree. We repeat that
procedure until T consists of a single vertex, i.e., only one simple polygon is left.
Soundness. It is not hard to see that our construction is polynomial in the size of the given 3SAT
formula ϕ.
Assume that ϕ is satisfiable. We show how to construct a labeling L of the constructed metro map.
For each variable x that is true (false) in the given truth assignment, we put all negative (positive)
labels of the corresponding variable gadget and its connected chains into L. By construction those
labels do not intersect. It remains to select labels for the clause gadgets. Consider a positive clause c;
negative clauses can be handled analogously. Since ϕ is satisfiable, c contains a variable x that is true
in the given truth assignment of ϕ. The set L contains only negative labels of the chain connecting the
gadget of x with the gadget of c, but no positive labels of that chain. Hence, we can add the port of c’s
gadget that is connected to that chain without creating intersections. For the second stop of the clause
we put that selectable label into L that is not a port. We can apply this procedure to all positive and
negative clauses without creating intersections, which yields the labeling L of the constructed metro
map.
Finally, assume that we are given a labeling L of the constructed metro map. Consider the clause
gadget of a positive clause c; negative clauses can be handled analogously. By construction L contains
at least one port ` of that gadget. This port is connected to a chain, which is then connected to a
gadget of a variable x. We set that variable x true. We apply this procedure to all clauses; for negative
clauses we set the corresponding variable to false. Since ` is contained in L, only negative labels of that
chain can be contained in L, but no positive labels. Hence, the positive ports of the variable gadget are
also not contained in L. By the previous reasoning this implies that all negative ports of the gadget are
contained in L. Consequently, by applying a similar procedure to negative clauses, it cannot happen
that x is set to false. Altogether, this implies a valid truth assignment of ϕ.
Remarks: Fig. 5 illustrates the construction of the gadgets for CurvedStyle. Note that only the
fork gadget, the clause gadget and the chain gadget rely on the concrete labeling style. Further, using
CurvedStyle, a stop s lying on a vertical segment lv has exactly two different distinguish labels; one
that lies to the left of lv and one that lies to the right of lv .
t
u
(a)
s2
consecutive
s s4 s5
3
(b)
s6
C
`0
`
s1
switchover
switchover
(c)
`00
s
C
s0
s00
`
`00 = `v
`0 =`u
s
s0
C
s00
Fig. 6: (a) Consecutive stops and switchovers. (b) The candidates satisfy the transitivity property.
(c) The candidates do not satisfy the transitivity property.
4
Labeling Algorithm for a Single Metro Line
We now study the case that the given instance I = (M, S, K, w) consists only of one metro line C.
Based on cartographic criteria we introduce three additional assumptions on I, which allows us to
efficiently solve MetroMapLabeling.
For each stop s ∈ S, we assume that each candidate ` ∈ Ks is assigned to one side of C; either ` is
a left candidate assigned to the left side of C, or ` is a right candidate assigned to the right side of C.
For appropriately defined candidate sets those assignments correspond with the geometric positions of
the candidates, i.e., left (right) candidates lie on the left (right) hand side of C.
Assumption 1 (Separated Labels) Candidates that are assigned to different sides of C do not
intersect.
This assumption is normally not a real restriction, because for appropriately defined candidate sets and
realistic metro lines, the line C separates both types of candidates geometrically. We further require
what we call the transitivity property.
Assumption 2 (Transitivity Property) For any three stops s, s0 , s00 ∈ S with s < s0 < s00 and any
three candidates ` ∈ Ks , `0 ∈ Ks0 and `00 ∈ Ks00 assigned to the same side of C, it holds that if neither
` and `0 intersect nor `0 and `00 intersect then also ` and `00 do not intersect; see also Fig. 6(b)–(c).
In our experiments we established Assumption 1 and Assumption 2 by removing candidates greedily.
In Section 8 we show that for real-world metro maps and the considered candidate sets we remove only
few labels, which indicates that those assumptions have only a little influence on the labelings.
Two stops s, s0 ∈ S with s < s0 are consecutive if there is no other stop s00 ∈ S with s < s00 < s0 ; see
Fig. 6(a). For two consecutive stops s1 , s2 ∈ S we say that each two candidates `1 ∈ Ks1 and `2 ∈ Ks2
are consecutive and denote the set that contains each pair of consecutive labels in L ⊆ K by PL ⊆ L×L.
Further, two consecutive labels `1 , `2 ∈ KC form a switchover (`1 , `2 ) if they are assigned to opposite
sides of C, where (`1 , `2 ) denotes an ordered set indicating the order of the stops of `1 and `2 . Two
switchovers of C are consecutive in L ⊆ K if there is no switchover in L in between of both. We define
the set of all switchovers in K by W and the set of consecutive switchovers in L ⊆ K by ΓL ⊆ W × W.
Based on cartographic criteria extracted from Imhof’s “general principles and requirements” for
map labeling [8], we require a cost function w : 2K → R+ of the following form; see also Section 5 for
a detailed motivation of w.
Assumption 3 (Linear Cost Function) For any L ⊆ K we require
X
X
X
w(L) =
w1 (`) +
w2 (`1 , `2 ) +
w3 (σ1 , σ2 ),
`∈L
(`1 ,`2 )∈PL
(σ1 ,σ2 )∈ΓL
where w1 : L → R rates a single label, w2 : PL → R rates two consecutive labels and w3 : ΓL → R rates
two consecutive switchovers.
In particular, we define w such that it penalizes the following structures to sustain readbility. (1) Steep
or highly curved labels. (2) Consecutive labels that lie on different sides of C, or that are shaped
differently. (3) Consecutive switchovers that are placed close to each other.
(a)
`1
`8 `9
`2
`4 `5
`3
`7
C
`6
(c)
⊥
(b)
`1
`4
`7
x
`2
`5
`8
`3
`6
`9
`2
`01
s02
d2
d1
y
hAσ,σ0
dummy
switchover
s1
s01
s2
σ
I(σ, σ 0 ]
`1
σ0
d3
>
d4
dummy
switchover
`02
= Aσ,σ0
Fig. 7: Illustrations for labeling a single metro line. (a) A one-sided instance and (b) the acyclic directed
graph G based on its labels. (c) A two-sided instance with a labeling. The switchovers σ 0 and σ separate
the labeling into a two-sided and a one-sided instance.
If I = ({C}, S, K, w) satisfies Assumption 1–3, we call MetroMapLabeling also SoftMetroLineLabeling. We now introduce an algorithm that solves this problem in O(n2 k 4 ) time, where n = |S|
and k = max{|Ks | | s ∈ S}. Note that k is typically constant. We assume w.l.o.g. that K contains only
candidates that do not intersect C.
Labels on One Side. We first assume that all candidate labels in K are assigned either to the left or to
the right side of C; without loss of generality to the left side of C. For two stops s, s0 ∈ S we denote
the instance restricted to the stops {s, s0 } ∪ {s00 ∈ S | s < s00 < s0 } by I[s, s0 ]. We denote the first stop
of C by s and the last stop by s. The transitivity property directly yields the next lemma.
Lemma 1. Let s, s0 and s00 be stops with s < s0 < s00 , L be a labeling of I[s, s0 ], ` ∈ L ∩ Ks and
`0 ∈ L ∩ Ks0 . Any `00 ∈ Ks00 intersecting ` also intersects `0 .
Proof. Recall for the proof that we assume that I satisfies Assumptions 1–3.
Assume for the sake of contradiction that there is a candidate `00 ∈ Ks00 such that `00 intersects `
but not `0 ; see Fig. 6(c). Since L is a labeling, the labels ` and `0 do not intersect. Hence, neither `
and `0 nor `0 and `00 intersect. Since all three labels are assigned to the same side of C, the transitivity
property holds, which directly contradicts that ` and `0 do not intersect.
t
u
Hence, the lemma states that `0 separates L from the candidates of the stops succeeding s0 . We use
this observation as follows. Based on K we define a directed acyclic graph G = (V, E); see Fig. 7(a)–(b).
This graph contains a vertex u for each candidate ` ∈ K and the two vertices x and y. We call x the
source and y the target of G. Let `u denote the candidate that belongs to the vertex u ∈ V \ {x, y}. For
each pair u, v ∈ V \ {x, y} the graph contains the edge (u, v) if and only if the stop of `u lies directly
before the stop of `v and, furthermore, `u and `v do not intersect. Further, for each vertex u of any
candidate of s the graph contains the edge (x, u), and for each vertex u of any candidate of s the graph
contains the edge (u, y). For an edge (u, v) ∈ E we define its cost we as follows. For u 6= x and v 6= y
we set we = w1 (`v ) + w2 (`u , `v ). For x = u we set we = w1 (`v ) and for v = y we set we = 0.P
An x-y path P ⊆ E in G is a path in G that starts at x and ends at y. Its costs are w(P ) = e∈P we .
The x-y path with minimum costs among all x-y paths is the shortest x-y path.
Lemma 2. For any x-y path P in G there is a labeling L of I with w(P ) = w(L) and for any labeling L
of I there is an x-y path P in G with w(P ) = w(L).
Proof. Recall for the proof that we assume that I satisfies Assumptions 1–3.
Let P = (VP , EP ) be an x-y path in G and let L = {`v ∈ K | v ∈ VP }, where VP denotes the
vertices of P and EP the edges of P . We show that L is a labeling of C with w(L) = w(P ). Obviously,
for each stop s ∈ S the set L contains exactly one candidate ` ∈ Ks . By construction for each edge
(u, v) ∈ EP the labels `u and `v do not intersect. Hence, by Lemma 1 the label `v cannot intersect any
label ` ∈ L of any stop that occurs before the stop of `u . Hence, the set L is a labeling. Let `1 , . . . , `n
be the labels in L in the order of their stops. It holds
w(P ) =
X
we = w1 (`1 ) +
e∈EP
n
X
(w1 (`i ) + w2 (`i−1 , `i )) = w(L)
(1)
i=2
Now, let L be an arbitrary labeling of C. We show that there is an x-y path P with w(P ) = w(L). Let
s and s0 be two consecutive stops with s < s0 and let ` and `0 be the corresponding labels in L. Since `
and `0 do not intersect, the corresponding vertices u and u0 of ` and `0 are adjacent in G. Hence, the
labels in L induce a path P in G. Let `1 , . . . , `n be the labels in L in the order of their stops. Using
Equation 1 we obtain w(P ) = w(L).
t
u
The lemma in particular proves that a shortest x-y path P in G corresponds with an optimal
labeling of I. Due to [4, Chapter 24], P can be constructed in O(|V | + |E|) time using a dynamic
programming approach, which we call MinPath. In particular MinPath considers each edge only
once. There are O(n · k) vertices in G and each vertex has at most k incoming edges, which implies
that there are O(n · k 2 ) edges. Since MinPath considers each edge only once, we compute the edges
of G on demand, which saves storage.
Theorem 2. If I is one-sided, SoftMetroLineLabeling can be optimally solved in O(nk 2 ) time
and O(nk) space.
Labels on Both Sides. If candidates lie on both sides of the metro line, we solve the problem utilizing
the algorithm for the one-sided case.
Consider a labeling L of I and let σ, σ 0 be two switchovers in L such that σ lies before σ 0 and
no other switchover lies in between both; see Fig. 7(c). Roughly spoken, σ and σ 0 induce a two-sided
instance that lies before σ and a one-sided instance that lies in between both switchovers σ and σ 0 .
Lemma 3. Let s, s01 , s02 and s00 be stops with s < s01 < s02 < s00 ; s01 and s02 are consecutive. Let L be a
labeling of I[s, s02 ], ` ∈ L ∩ Ks , `01 ∈ L ∩ Ks01 , `02 ∈ L ∩ Ks02 s.t. (`01 , `02 ) is a switchover. Any `00 ∈ Ks00
intersecting ` intersects `01 or `02 .
Proof. Recall for the proof that we assume that I satisfies Assumptions 1–3.
Assume for the sake of contradiction that there is a label `00 ∈ Ks00 such that `00 intersects ` without
intersecting `01 and `02 . Since ` and `00 intersect each other, due to Assumption 1 both are assigned to
the same side of C; w.l.o.g., let ` and `00 be assigned to the left hand side of C. Further, w.l.o.g., let `01
be a left candidate and `02 a right candidate; analogous arguments hold for the opposite case. Since L
is a labeling, the labels ` and `01 do not intersect. Hence, neither ` and `01 nor `01 and `00 intersect.
Since `, `01 and `00 are assigned to the same side of C, the transitivity property must hold. However,
this contradicts that ` and `00 intersect.
t
u
Hence, the lemma yields that for the one-sided instance we can choose any labeling; as long as this
labeling does not intersect any label of σ or σ 0 , it composes with σ, σ 0 and the labeling of the two-sided
instance to one labeling for the instance up to σ 0 . We use that observation as follows.
Let σ = (`1 , `2 ) and σ 0 = (`01 , `02 ) be two switchovers in W. Let s1 and s2 be the stops of `1 and `2 ,
and let s01 and s02 be the stops of `01 and `02 , respectively; see Fig. 7(c). We assume that σ < σ 0 , i.e.,
s1 < s01 . Let I(σ, σ 0 ] be the instance restricted to the stops {s ∈ S | s2 < s < s01 } ∪ {s01 , s02 }, where
(σ, σ 0 ] indicates that the stops of σ 0 belong to that instance, while the stops of σ do not.
The switchovers σ and σ 0 are compatible if `2 and `01 are assigned to the same side of C, and there
is a labeling for I[s1 , s02 ] such that it contains `1 , `2 , `01 and `02 and, furthermore, σ and σ 0 are the
only switchovers in that labeling. Let L be the optimal labeling among those labelings. We denote the
labeling L \ {`1 , `2 } of I(σ, σ 0 ] by Aσ,σ0 . Utilizing Theorem 2, we obtain Aσ,σ0 in O(n · k 2 ) time.
For any labeling L of an instance J let hL ∈ L be the label of the first stop in J and let tL ∈ L be
the label of the last stop in J; hL is the head and tL is the tail of L. For technical reasons we extend S
by the dummy stops d1 , d2 , d3 and d4 such that d1 < d2 < s < d3 < d4 for any stop s ∈ S. For d1
and d2 we introduce the dummy switchover ⊥ and for d3 and d4 the dummy switchover >. We define
that ⊥ and > are compatible to all switchovers in W and that ⊥ and > are compatible, if there is a
one-sided labeling for I. Conceptually, each dummy switchover consists of two labels that are assigned
to both sides of C. Further, neither ⊥ nor > has any influence on the cost of a labeling. Hence, w.l.o.g.
we assume that they are contained in any labeling.
Similar to the one-sided case we define a directed acyclic graph G0 = (V 0 , E 0 ). This graph contains
a vertex u for each switchover W ∪ {⊥, >}. Let σu denote the switchover that belongs to the vertex
u ∈ V . In particular let x denote the vertex of ⊥ and y denote the vertex of >. For each pair u, v ∈ V
the graph contains the edge (u, v) if and only if σu and σv are compatible and σu < σv . The cost we
of an edge e = (u, v) in G0 is we = w(Aσu ,σv ) + w3 (σu , σv ) + w2 (`u2 , hAσu ,σv ), where σu = (`u1 , `u2 ).
In the special case that σu and σv share a stop, we set we = w3 (σu , σv ) + w2 (`v1 , `v2 ) + w1 (`v2 ), where
σv = (`v1 , `v2 ).
Let P be an x-y path in G0 and let e1 = (x = v0 , v1 ), e2 = (v1 , v2 ), . . . , el = (vl−1 , vl = y) be
the edges of P . For a vertex vi of P with 0 ≤ i ≤ l we write σi instead of σvi . We denote the set
Sl
i=1 Aσi−1 ,σi by LP .
Lemma 4. a) The graph G0 has an x-y path if and only if I has a labeling.
b) Let P be a shortest x-y path in G0 , then LP is an optimal labeling of I.
Proof. Recall for the proof that we assume that I satisfies Assumption 1–3.
By construction of G0 , it directly follows that G0 has an x-y path P if and only if I has a labeling L.
We first show that LP is a labeling of I with w(LP ) = w(P ). Afterwards we prove that for any labeling L
of I it holds w(L) ≥ w(LP ).
Let e1 = (v0 , v1 ), e2 = (v1 , v2 ), . . . , el = (vl−1 , vl ) be the edges of P with v0 = x and vl = y. For a
vertex vi with 0 ≤ i ≤ l we write σi instead of σvi . We show by induction over m ∈ N with 1 ≤ m ≤ l
that
m
[
Lm :=
Aσi−1 ,σi
i=1
Pm
is a labeling of I(σ0 , σm ] with w(Lm ) = i=1 wei . Altogether this implies that LP = Ll is a labeling
with w(LP ) = w(P ).
For m = 1 we have L1 = Aσ0 ,σ1 . By the construction of G0 the set Aσ0 ,σ1 is a labeling. Since
σ0 = (`0 , `1 ) is a dummy switchover we have w3 (σ0 , σ1 ) = 0 and w2 (`2 , hAσ0 ,σ1 ) = 0. Hence, it holds
w(Aσ0 ,σ1 ) = we1 .
Now, consider the set Lm for m > 1. We first argue that Lm is a labeling of I(σ0 , σm ]. By induction
the set Lm−1 ⊆ Lm is a labeling of I(σ0 , σm−1 ]. Further, by construction the set Lm \Lm−1 = Aσm−1 ,σm
is a labeling of the instance I(σm−1 , σm ]. Since σm−1 and σm are compatible, no label of Lm \ Lm−1
intersects any label of σm−1 . Then
by Lemma 3 no two labels in Lm intersect P
each other.
Pm
m−1
We now show that w(Lm ) = i=1 wei . By induction we have w(Lm−1 ) = i=1 wei . Since Lm =
Lm−1 ∪ Aσm−1 ,σm it holds
w(Lm ) = w(Lm−1 ∪ Aσm−1 ,σm ).
(2)
We distinguish two cases. First assume that σm−1 and σm do not have any stop in common. Let σm−1 =
(`m−1
, `m−1
), then we derive from Equation (2)
1
2
w(Lm ) =w(Lm−1 ∪ Aσm−1 ,σm )
=w(Lm−1 ) + w(Aσm−1 ,σm ) + w3 (σm−1 , σm ) + w2 (`m−1
, hAσm−1 ,σm )
2
(I)
(II)
= w(Lm−1 ) + wem =
m
X
w ei .
i=1
Equality (I) holds due to the definition of wem and Equality (II) is by induction true. Now assume
m
that σm−1 and σm have a stop in common. Let σm = (`m
1 , `2 ), then we derive from Equation (2)
w(Lm ) =w(Lm−1 ∪ Aσm−1 ,σm )
m
m
=w(Lm−1 ) + w3 (σm−1 , σm ) + w2 (`m
1 , `2 ) + w1 (`2 )
m
(III)
(IV ) X
= w(Lm−1 ) + wem =
w ei .
i=1
Equality (III) holds due to the definition of wem and Equality (IV) is by induction true. Altogether
we obtain that Ll is a labeling with w(Ll ) = w(P ).
Finally we show, that there is no other labeling L with w(L) < w(LP ). Assume for the sake of
contradiction that there is such a labeling L. Let ⊥ = σ0 , σ1 , . . . , σl = > be the switchovers in L,
such that σi < σi+1 for each 0 ≤ i < l. We observe that two consecutive switchovers σi and σi+1
are compatible. Hence, for any i with 1 ≤ i ≤ l there is an edge ei = (u, v) in G0 with σu = σi−1
and σv = σi+1 . Consequently, the edges e1 , . . . , el form an x-y path Q. By the first two claims of this
lemma, there is a labeling LQ of I with w(Q) = w(LQ ). Since P is a shortest x-y path it holds
w(L) < w(LP ) = w(P ) ≤ w(Q) = w(LQ )
We now show that w(LQ ) ≤ w(L) deriving a contradiction. To that end recall that
LQ =
l
[
i=1
Aσi−1 ,σi .
For 1 ≤ i ≤ l the set Aσi−1 ,σi is an optimal labeling of the instance I(σi−1 , σi ] such that Aσi−1 ,σi ∪σi−1 is
a labeling and σi is the only switchover contained in Aσi−1 ,σi . Let Li ⊆ L and LQ
i ⊆ LQ be the labelings
restricted to I(σi−1 , σi ]. In particular we have LQ
=
A
and
σ
is
the
only
switchover in Li . If we
σ
,σ
i
i−1
i
i
had w(LQ ) > w(L), there must be two consecutive switchovers σi−1 and σi such that w(Li ) < w(LQ
i ).
However, this contradicts the optimality of Aσi−1 ,σi . Consequently, it holds w(LQ ) ≤ w(L) yielding
the claimed contradiction.
t
u
By Lemma 4 a shortest x-y path P in G0 corresponds with an optimal labeling L of C, if this
exists. Using MinPath we construct P in O(|V 0 | + |E 0 |) time. Since W contains O(nk 2 ) switchovers,
the graph G0 contains O(nk 2 ) vertices and O(n2 k 4 ) edges. As MinPath considers each edge only once,
we compute the edges of G0 on demand, which needs O(nk 2 ) storage. We compute the costs of the
incoming edges of a vertex v ∈ V 0 utilizing the one-sided case. Proceeding naively, we need O(nk 2 )
time per edge, which yields O(n3 k 6 ) time in total.
Reusing already computed information, we improve that result as follows. Let (u1 , v), . . . , (uk , v)
denote the incoming edges of v such that σu1 ≤ · · · ≤ σuk , i.e., the stop of σui ’s first label does not lie
after the stop of σuj ’s first label with i < j. Further, let σv = (`1v , `2v ) and let Gi be the graph for the
one-sided instance I[si , s] considering only candidates that lie on the same side as `1v , where si is the
stop of the second label of σui and s is the stop of `1v . Let Pi be the shortest xi -yi path in Gi , where xi
and yi denote the source and target of Gi , respectively. We observe that excluding the source and
target, the graph Gi is a sub-graph of G1 for all 1 ≤ i ≤ k. Further, since a sub-path of a shortest path
is also a shortest path among all paths having the same end vertices, we can assume without loss of
optimality that when excluding xi and yi from Pi , the path Pi is a sub path of P1 for all 1 ≤ i ≤ k. We
therefore only need to compute G1 and P1 and can use sub-paths of P1 in order to gain the costs of all
incoming edges of v. Hence, we basically apply for each vertex v ∈ V the algorithm for the one-sided
case once using O(nk 2 ) time per vertex. We then can compute the costs in O(1) time per edge, which
yields the next result.
Theorem 3. SoftMetroLineLabeling can be optimally solved in O(n2 k 4 ) time and O(nk 2 ) space.
1st
Step
2nd
Step
3rd
Step
4th
Step
Fig. 8: Schematic illustration of the presented workflow. 1st Step: generation of candidates. 2nd Step:
scaling and creation of initial labeling (red labels). 3rd Step: pre-selection of candidates. 4th Step:
Solving the metro lines independently.
5
Cost Function
In this section we motivate the cost function introduced in Section 4. For a given metro map ({C}, S)
that consists of a single metro line C, and generated candidates K, we rate each labeling L ⊆ K using
the following cost function:
X
X
X
w(L) =
w1 (`) +
w2 (`1 , `2 ) +
w3 (σ1 , σ2 ),
`∈L
(`1 ,`2 )∈PL
(σ1 ,σ2 )∈ΓL
where w1 : L → R rates a single label, w2 : PL → R rates two consecutive labels and w3 : ΓL → R rates
two consecutive switchovers; see Assumption 3. The definition of this function relies on the following
considerations, which are based on Imhof’s “general principles and requirements” for map labeling [8].
To respect that some (e.g., steep and highly curved) labels are more difficult to read than others,
we introduce a cost w1 (`) for each candidate label `.
Imhof further notes that “names should assist directly in revealing spatial situation” and exemplifies this principle with maps that show text only while still conveying the most relevant geographic
information. To transfer this idea to metro maps, we favor solutions where the labels for each two
consecutive stops on a metro line have similar properties. That is, the two labels should be placed on
the same side of the line and their slopes and curvatures should be similar. In a map satisfying this
criterion, a user need not find the point-text correspondence on a one-to-one basis. Instead, the user
can identify metro lines and sequences of stops based on label groups, which, for example, makes it
easier to count the stops till a destination. (Of course, this is also an improvement in terms of legibility.)
In our model, we consider the similarity of consecutive labels by introducing a cost w2 (`1 , `2 ) for each
pair (`1 , `2 ) of candidates that belong to consecutive stops on C. We penalize consecutive candidates
that lie on opposite sides of the metro line, because those disturb the overall label placement. We add
this cost to the objective value of a solution if both candidates are selected. Since we minimize the
total costs of the solution, the cost for a pair of candidates should be low if they are similar. Further,
if C has labels that do not lie on the same side of C, the implied switchovers should occur in regular
distances and not cluttered. Hence, for each pair σ1 and σ2 of two consecutive switchovers in a solution,
we add a cost w3 (σ1 , σ2 ) to the objective value of the solution that depends on the distance between
σ1 and σ2 ; the smaller the distance, the greater the cost of w3 (σ1 , σ2 ).
We now describe more precisely how we defined w1 , w2 and w3 for our evaluation. The definitions
depend on the applied labeling style.
Curved Labels. Using CurvedStyle for the labels we define the cost functions as follows. For
a label ` ∈ K let p1` = (x1` , y`1 ) be the start point of ` and let p2` = (x2` , y`2 ) be the end point of `;
recall that we derived the labels from Bézier curves. Let v ` be the vector connecting p1` with p2` and
let α ∈ [0, 2π] denote the angle of v ` . We define
0 ≤ α ≤ π2
α`
π − α
π
`
2 <α ≤π
δ` =
α` − π
π < α ≤ 23 π
2π − α` 32 π < α < 2π
Hence, the angle δ` is a measure for how horizontal the vector v ` is, whereby the smaller the value
of δ` , the more horizontal is v ` .
We defined the cost function w1 rating a single label ` ∈ K to be w1 = 10 · δ` . Hence, we penalize
steep labels. We defined w2 rating two consecutive labels `1 and `2 as follows. If `1 and `2 point into
different x-direction, w2 is 150. Else, if `1 and `2 are switchovers, the cost w2 is 0. In all other cases,
w2 is the difference between the angle α`1 and α`2 . Hence, in the latter case we penalize labels that
are differently aligned. Finally, w3 rating two switchovers σ1 and σ2 of the same line is defined as
, where dσ1 ,σ2 is the number of stops in between σ1 and σ2 . In particular w3 effects that
w3 = dσ200
1 ,σ2
a labeling with equally sized sequences of labels lying on the same side of their metro line are rated
better than a labeling where the sequences are sized irregularly.
Octilinear Labels. Using OctilinStyle for the labels we define the cost functions as follows.
Recall that we use OctilinStyle for octilinear metro maps. For a label ` let l` be the segment on
which its stop is placed. If l` is horizontal, but ` is not diagonal, we set w1 = 200. If l` is vertical or
diagonal, but ` is not horizontal, we set w1 = 100. In all other cases we set w1 = 0. The functions w2
and w3 are defined in the same way as for CurvedStyle.
6
Multiple Metro Lines
In this section we consider the problem that we are given a metro map (M, S) consisting of multiple
metro lines. We present an algorithm that creates a labeling for (M, S) in two phases. Each phase is
divided into two steps; see Fig. 8 for a schematic illustration. In the first phase the algorithm creates
the set K of label candidates and ensures that there exists at least one labeling for the metro map. In
the second phase it then computes a labeling L for (M, S, K). To that end it makes use of the labeling
algorithm for a single metro line; see Section 4. In order to rate L, we extend the cost function w for
a single metro line on multiple metro lines, i.e., for any L ⊆ K we require
X
X
X
X
w(L) =
w(C) with w(C) =
w1 (`) +
w2 (`1 , `2 ) +
w3 (σ1 , σ2 ),
C∈M
`∈LC
(`1 ,`2 )∈PLC
(σ1 ,σ2 )∈ΓLC
where LC = L ∩ KC is the labeling restricted to metro line C ∈ M, w1 : LC → R rates a single label,
w2 : PLC → R rates two consecutive labels and w3 : ΓLC → R rates two consecutive switchovers of C.
In particular w satisfies Assumption 3 for a single metro line.
Altogether, the workflow yields a heuristic that relies on the conjecture that using optimal algorithms in single steps is sufficient to obtain good labelings. In our evaluation we call that approach
DpAlg.
6.1
First Phase – Candidate Generation
First, we create the label candidates K. We then enforce that there is a labeling L for the given
instance I = (M, S, K).
1st Step – Candidate Creation. Depending on the labeling style, we generate a discrete set of candidate
labels for every stop. Hence, we are now given the instance (M, S, K, w). In particular we assume that
each candidate ` ∈ K is assigned to one side of its metro line C, namely to the left or right side of C,
and, furthermore, w is a cost function satisfying Assumption 3 for each metro line C.
2nd Step – Scaling. Since each stop of a metro map must be labeled, we first apply a transformation
on the given candidates to ensure that there is at least one labeling of the metro map. To that end we
first determine for each stop s ∈ S of each metro line C ∈ M two candidates of Ks that are assigned
to opposite sides of C. More specifically, among all candidates in Ks that are assigned to the right
hand side of C and that do not intersect any metro line of M, we take that candidate `R ∈ Ks with
minimum costs, i.e., w1 (`) is minimal. If such a candidate does not exist, because each label of Ks
intersects at least one metro line, we take a pre-defined label `R ∈ Ks that is assigned to the right hand
side of C. In the same manner we choose a candidate `L ∈ Ks that is assigned to the left hand side
of C. Let Ds = {`R , `L }, we now enforce that there is a labeling L of I such that for each stop s ∈ S
it contains a label of Ds .
S
We check whether the set s∈S Ds admits a labeling L for I. Later, we describe more specifically
how to do this. If L exists, we continue with the third step of the algorithm using I and L as input. Otherwise, we scale all candidates of K smaller by a constant factor and repeat the described
procedure. Sampling a pre-defined scaling range [xmin , xmax ], we find in that manner a scaling factor x ∈ [xmin , xmax ] for the candidates that admits a labeling L of I. We choose x as large as possible.
If we could not find x, e.g, because we have chosen [xmin , xmax ] or the sampling not appropriately, we
abort the algorithm, stating that the algorithm could not find a labeling.
S
Next, we describe how to check whether there is a labeling for s∈S Ds . Since for each s ∈ S the
set Ds contains two candidates, we can make use of a 2SAT formulation to model the labeling problem.
For each stop s ∈ S and each candidate
` ∈ Ds , we introduce the Boolean variable x` . Those Boolean
S
variables induce the set L = {` ∈ s∈S Ds | x` is true}. The following formulas are satisfiable if and
only if L is a labeling of M.
¬x`
¬x` ∨ ¬x`0
x` ∨ x`0 , ¬x` ∨ ¬x`0
∀s ∈ S ∀` ∈ Ds s.t. ∃C ∈ M that intersects `
[
∀`, `0 ∈
Ds s.t. ` and `0 intersect.
s∈S
∀s ∈ S ∀`, `0 ∈ Ds
The first formula ensures that there is no label of the solution that intersects any metro line. The
second one avoids overlaps between labels, while the two last formulas enforce that for each stop s ∈ S
there is exactly one label of Ds that is contained in the solution.
According to [2] in linear time with respect to the number of variables and formulas, the satisfiability
can be checked. We introduce O(n) variables and instantiate the second formula O(n2 ) times, because
each pair of candidates may overlap. The remaining formulas are instantiated in O(n2 ) time. Hence,
the total running time is in O(n2 · t), where t denotes the number of scaling steps.
6.2
Second Phase – Candidate Selection
We assume that we are given the scaled instance I = (M, S, K) and the labeling L for I of the
previous phase. We apply a pre-selection on the candidates by discarding candidates such that no
two stops’ candidates of different metro lines intersect and each metro line satisfies Assumption 1 and
Assumption 2. We never remove a label in L from the candidate set, however, to ensure that there is
always a feasible solution. Finally, considering the metro lines independently, we select for each stop a
candidate to be its label using the dynamic program described in Section 4.
3th Step – Candidate Pre-Selection. We first ensure that I satisfies Assumption 1 and Assumption 2.
If two candidates ` and `0 of the same metro line C intersect and if they are assigned to opposite sides
of C, we delete one of both labels as follows. If ` ∈ L, we delete `0 , and if `0 ∈ L we delete `. Otherwise,
if none of both is contained in L, we delete that label with higher costs; ties are broken arbitrarily.
Afterwards, Assumption 1 is satisfied. Further, for each metro line C of I we iterate through the stops
of C from its beginning to its end. Doing so, we delete candidates from K violating Assumption 2 as
described as follows. Let s be the currently considered stop. For each candidate ` ∈ Ks we check for
each stop s0 with s0 < s whether there is a label `0 ∈ Ks0 that intersects `. If `0 exists, we check whether
each candidate `00 ∈ Ks00 of any stop s00 with s0 < s00 < s intersects ` or `0 . If this is not the case, we
delete ` if ` is not contained in L, and otherwise we delete `0 . Note that not both can be contained
in L. By construction each metro line of instance I then satisfies Assumption 2, where L ⊆ K.
Finally, we ensure that the metro lines in M become independent in the sense that no candidates
of stops belonging to different metro lines intersect and no candidate intersects any metro line. Hence,
after this step, the metro lines can independently be labeled such that the resulting labelings compose
to a labeling of I.
We first rank the candidates of K as follows. For each metro line C ∈ M we construct a labeling LC
using the dynamic program for the two sided case as presented in Section 4. Due to the previous step,
those labelings exist. Note that for two metro lines C, C 0 ∈ M there may be labels ` ∈ LC and `0 ∈ LC 0
that intersect each other. For each candidate ` ∈ K we set val` = 1 if ` ∈ LC for a metro line C ∈ M,
and val` = 0 otherwise. A candidate ` ∈ K has a smaller rank than a candidate `0 ∈ K, if val` > val`0
or val` = val`0 and w1 (`) ≤ w1 (`0 ); ties are broken arbitrarily.
We now greedily remove candidates from K until all metro lines are independent. We create a conflict
graph G = (V, E) such that the vertices of G are the candidates and the edges model intersections
between candidates, i.e., two vertices are adjacent if and only if the corresponding labels intersect.
Then, we delete all vertices whose corresponding labels intersect any metro line. Afterwards, starting
with I = ∅, we construct an independent set I on G as follows. First, we add all vertices of G to I,
whose labels are contained in L, and delete them and their neighbors from G. Since the labels in
L do not intersect, I is an independent set of the original conflict graph. In the increasing order of
their ranks, we remove each vertex v and its neighbors from G. Each time we add v to I; obviously
sustaining that I is an independent set in G. We then update for each stop s ∈ S its candidate set Ks
to Ks ← {` ∈ Ks | vertex of ` is contained in I}. Since all labels of L are contained in I, there is a
labeling for (M, S, K, w) based on the new candidate set K.
4th Step – Final Candidate Selection. Let I = (M, S, K, w) be the instance after applying the third
step and L be the labeling that has been created in the first step. By the previous step the metro lines
are in the sense independent that candidates of stops belonging to different metro lines do not intersect.
Further, they all satisfy Assumption 1 and Assumption 2. Hence, we use the dynamic programming
approach of Section 4 in order to label them independently. The composition of those labelings is then
a labeling L of I.
7
Alternative Approaches
We now present the three approaches IlpAlg, ScaleAlg, GreedyAlg, which are adaptions of
our workflow. We use those to experimentally evaluate our approach against alternatives. While
GreedyAlg is a simple and fast greedy algorithm, IlpAlg and ScaleAlg are based on an ILP
formulation.
7.1
Integer Linear Programming Formulation
To assess the impact of the second phase of our approach, we present an integer linear programming
formulation that optimally solves MetroMapLabeling with respect to the required cost function.
Let (M, S, K, w) be an instance of that problem, which we obtain after the first phase of our approach.
We first note that we apply a specific cost function (see Section 5). The cost function w3 , which rates
two consecutive switchovers σ = (`1 , `2 ) and σ 0 = (`01 , `02 ) of a labeling L, does not rely on the actual
switchovers, but only on their positions on the corresponding metro line. Hence, we may assume that
w3 expects the stops of `1 and `01 . This assumption helps us reduce the number of variables.
For each candidate ` ∈ K we define a binary variable x` ∈ {0, 1}. If x` = 1, we interpret it such
that ` is selected for the labeling. We introduce the following constraints.
x` = 0
x ` + x `0 ≤ 1
X
x` = 1
`∈Ks
∀` ∈ K that intersect a metro line.
(3)
∀`, `0 ∈ K that intersect.
(4)
∀s ∈ S
(5)
Moreover, for each metro line C ∈ M we define the following variables. To that end, let P ∈ KC × KC
be the set of all consecutive labels and let s1 , . . . , sn be the stops of C in that particular order.
y`,`0 ∈ {0, 1},
∀(`, `0 ) ∈ P.
hi,j ∈ {0, 1},
1 ≤ i < n and i < j < n
zi ∈ {0, 1},
1≤i<n
If y`,`0 = 1, we interpret it such that both ` and `0 are selected for the labeling. If zi = 1, we interpret
it such that the selected labels of the stops si and si+1 form a switchover. If hi,j = 1, we interpret it
such that the selected labels at si , si+1 , sj and sj+1 form two consecutive switchovers. Further, for
each metro line C we introduce the following constraints. In these constraints L(K) denotes the set of
labels that lie to the left of C, and R(K) denotes the set of labels that lie to the right of C.
X
∀(`, `0 ) ∈ P
(6)
1≤i<n
(7)
x` − 1 ≤ zi
1≤i<n
(8)
X
1≤i<j<n
(9)
hi,j · w3 (si , sj )
(10)
x` + x`0 − 1 ≤ y`,`0
X
x` +
x` − 1 ≤ zi
`∈L(Ksi )
`∈R(Ksi+1 )
X
X
x` +
`∈R(Ksi )
`∈L(Ksi+1 )
zi + zj − 1 −
zk ≤ hi,j
i<k<j
We further define for each metro line C the following linear term.
w(C) :=
X
`∈KC
x` · w1 (`) +
X
y`,`0 · w2 (`, `0 ) +
(`,`0 )∈P
X
1≤i<n
i<j<n
Subject to the presented Constraints (3)–(9) we then minimize
X
w(C).
(11)
C∈M
Consider a variable assignment that minimizes (11) and satisfies Constraints (3)–(9). We show that
L = {` ∈ K | x` = 1}
is an optimal labeling of the given instance with respect to the given cost function. First of all, L is a
valid labeling. Constraint (3) ensures that no label in L intersects any metro line. By Constraint (4)
the labels in L are pairwise disjoint. Finally, by Constraint (5) for each stop there is exactly one label
contained in L. In particular, for a metro line C ∈ M, the set LC = L ∩ KC is a valid labeling of C.
We now show that w(LC ) = w(C) for any metro line C ∈ M. Since we minimize (11), this implies
the optimality of L. Obviously, for a label ` ∈ KC the cost w1 (`) is taken into account in w(C) if and
only if ` belongs to L.
By Constraint (6) for two consecutive labels ` and `0 of C we have y`,`0 = 1 if both are contained
in L. Further, by the minimality of (11), if at least one of both labels does not belong to L, it holds
y`,`0 = 0. Hence, w2 (`, `0 ) is taken into account in w(C) if and only if both ` and `0 belong to L.
By Constraint (7) and Constraint (8) it holds zi,j = 1 if the labels `i ∈ L and `i+1 ∈ L of the
consecutive stops si and si+1 form a switchover in L. Hence, by Constraint (9) it further holds hi,j = 1
if the labels of si and si+1 as well as sj and sj+1 form switchovers σi and σj in L, and, furthermore,
there is no other switchover in between σi and σj , i.e., both switchovers are consecutive. On the other
hand, by the minimality of w(C) in all other cases it holds hi,j = 0. Hence, w3 (si , sj ) is taken into
account in w(C) if and only if σi and σj are consecutive switchovers in L. Altogether we obtain the
following theorem.
Theorem 4. Given an optimal variable assignment for the presented ILP formulation, the set L =
{` ∈ K | x` = 1} is an optimal labeling of (M, S, K) with respect to w.
The approach IlpAlg simply replaces the second phase of DpAlg by that ILP formulation. Hence,
it solves the second phase optimally. The approach ScaleAlg samples a predefined scaling range
[xmin , xmax ], which is also used by DpAlg. For each scale x it scales the candidates correspondingly.
Using the ILP formulation it then checks whether the candidates admit a labeling. Hence, we approximately obtain the greatest scaling factor that admits a labeling.
7.2
Greedy-Algorithm
The algorithm GreedyAlg replaces the dynamic programming approach in our workflow as follows.
Starting with the solution L enforced by the 1st step, the greedy algorithm iterates once through the
stops of C. For each stop s of C it selects the candidate ` ∈ Ks that minimizes w1 (`)+w2 (`p , `)+w2 (`, `s )
among all valid candidates in Ks , where `p ∈ L is the candidate selected for the previous stop sp and
`s ∈ L is the candidate for the successive stop. It replaces the candidate of s in L with `.
8
Evaluation
To evaluate our approach presented in Section 6, we did a case study on the metro systems of Sydney
(173 stops) and Vienna (84 stops), which have been used as benchmarks before [5,11,14]. For Sydney we
took the curved layout from [5, Fig. 1a] and the octilinear layouts from [11, Fig. 9a,b][14, Fig. 10.], while
for Vienna we took the curved layout from [5, Fig. 8c] and the ocitlinear layouts from [11, Fig. 13a,b].
See also Table 1 for an overview of the instances. Since the metro lines of Sydney are not only paths,
we disassembled those metro lines into single paths. We did this by hand and tried to extract as long
paths as possible. Hence, the instances of Sydney decompose into 12 lines and the instances of Vienna
into 5 lines. We took the positions of the stops as presented in the corresponding papers. In the curved
layout of Sydney we removed the stops Tempe and Martin Place (in Fig. 9(a) marked as red dots),
because both stops are tightly enclosed by metro lines such that only the placement of very small labels
is possible. This is not so much a problem of our approach, but of the given layout. In a semi-automatic
approach the designer would then need to change the layout. For the curvilinear layouts we used labels
of CurvedStyle and for the octilinear layouts we used labels of OctilinStyle. For the layouts of
Sydney2, Sydney3 and Vienna2 the authors present labelings; see Fig. 10, Fig. 11 and Fig. 12. For any
other layout they do not present labelings.
s
: Ratio of
Table 1: Overview of considered instances. Style: Style of map and applied labels. smax
applied scale factors. The scale smax is a lower bound for the largest possible scaling factor (obtained
by ScaleAlg), and s is the scale computed by the first phase of our workflow.
Instance
Sydney1
Sydney2
Sydney3
Sydney4
Vienna1
Vienna2
Vienna3
Style
Octi.
Octi.
Octi.
Curved
Octi.
Octi.
Curved.
0.69
0.57
0.57
0.67
0.59
0.81
0.54
Reference [11, Fig. 9a] [11, Fig. 9b] [14, Fig. 10.] [5, Fig. 1a] [11, Fig. 13a] [11, Fig. 13b] [5, Fig. 8c]
s
smax
The experiments were performed on a single core of an Intel(R) Core(TM) i7-3520M CPU processor.
The machine is clocked at 2.9 GHz, and has 4096 MB RAM. Our implementation is written in Java.
For each instance and each algorithm we conducted 100 runs and took the average running times.
Each time before we started the 100 runs, we performed 50 runs without measuring the running time
in order to warm up the virtual machine (Java(TM) SE Runtime Environment, build 1.8.0 60-b27,
Table 2: Running times in seconds of the workflow broken down into its two phases and their single
steps (if applicable). Algo.: The applied algorithm. Times less than 0.01 seconds are marked with ?.
Algo.
Instance Layout
Phase 1: Creation Phase 2: Selection
P
P
1st 2nd
1st 2nd
Total
DpAlg
GreedyAlg
Sydney1 Octi.
ScaleAlg
IlpAlg
0.03
0.02
0.02
0.02
0.18
0.18
–
0.17
0.21
0.2
0.02
0.19
0.09
?
–
–
0.1
?
–
–
0.18
?
69.09
25.43
0.39
0.2
69.11
25.62
DpAlg
GreedyAlg
Sydney2 Octi.
ScaleAlg
IlpAlg
0.03
0.02
0.02
0.02
0.28
0.28
–
0.28
0.31
0.3
0.02
0.29
0.07
?
–
–
0.09
?
–
–
0.15
?
17.29
15.17
0.46
0.3
17.31
15.46
DpAlg
GreedyAlg
Sydney3 Octi.
ScaleAlg
IlpAlg
0.03
0.02
0.02
0.02
0.17
0.16
–
0.15
0.2
0.17
0.02
0.17
0.07
?
–
–
0.09
?
–
–
0.16
?
33.61
21.56
0.36
0.17
33.63
21.73
DpAlg
GreedyAlg
Sydney4 Curved
ScaleAlg
IlpAlg
0.03
0.03
0.03
0.02
0.33
0.33
–
0.34
0.36
0.35
0.05
0.36
0.1 0.1
?
?
–
–
–
–
0.2
?
243.58
98.81
0.56
0.35
243.63
99.17
DpAlg
GreedyAlg
Vienna1 Octi.
ScaleAlg
IlpAlg
0.01
0.01
0.01
0.01
0.06
0.06
–
0.06
0.07
0.07
0.01
0.08
0.01
?
–
–
?
?
–
–
0.02
?
3.31
0.55
0.09
0.07
3.32
0.63
DpAlg
GreedyAlg
Vienna2 Octi.
ScaleAlg
IlpAlg
0.01
0.01
0.01
0.01
0.09
0.1
–
0.08
0.1
0.11
0.01
0.09
?
?
–
–
?
?
–
–
?
?
4.65
0.36
0.11
0.11
4.66
0.45
DpAlg
GreedyAlg
Vienna3 Curved
ScaleAlg
IlpAlg
0.02
0.02
0.02
0.01
0.15
0.14
–
0.13
0.17
0.16
0.02
0.15
0.02
?
–
–
0.04
?
–
–
0.06
?
7.27
0.49
0.23
0.16
7.29
0.64
Table 3: Experimental results for Sydney and Vienna. Algo.: The applied algorithm. Candidates: Values
concerning candidates; No. of candidates after the first and third step, A12= No. of labels removed to
establish Assumption 1 and Assumption 2, SO= No. of switchovers. Cost: Ratio of costs; LA = labeling
obtained by procedure A ∈ {DpAlg, GreedyAlg, IlpAlg, ScaleAlg}, L = labeling obtained by
procedure IlpAlg. Sequence: Values concerning sequences of labels lying on the same side of their
metro line; min=length of shortest sequence, max =length of longest sequence, avg.=average length of
sequences.
Instance Layout Algo.
Candidates
Cost
Sequence
1st 3rd A12 SO
w1 (LA ) w2 (LA ) w3 (LA ) w(LA )
w1 (L)
w2 (L)
w3 (L)
w(L)
min max avg.
DpAlg
GreedyAlg
Sydney1 Octi.
ScaleAlg
IlpAlg
1164 1004 9 4
1164 1005 9 15
1164 –
– 13
1164 –
– 7
1.73
1.73
2.46
1
1.12
1.87
1.67
1
0.63
7.62
2.91
1
1.16
2.64
1.97
1
3
1
3
3
41
24
20
20
9.53
5.81
6.97
8.1
DpAlg
GreedyAlg
Sydney2 Octi.
ScaleAlg
IlpAlg
1104 963
1104 964
1104 –
1104 –
6
19
15
5
1.36
2.46
3.91
1
1.08
1.83
1.92
1
3.25
19.52
8.88
1
1.33
3.54
2.91
1
1
1
1
3
26
35
19
25
9.34
5.94
6.76
9.48
DpAlg
GreedyAlg
Sydney3 Octi.
ScaleAlg
IlpAlg
1132 1013 0 6
1132 1013 0 14
1132 –
– 12
1132 –
– 6
1
2.15
2.72
1
1
1.79
2.68
1
1
5.99
3.06
1
1
2.61
2.76
1
3
1
2
3
25
23
21
25
8.76
6.9
7.02
8.76
DpAlg
GreedyAlg
Sydney4 Curved
ScaleAlg
IlpAlg
1275 1020 2 8
1275 1021 2 28
1275 –
– 13
1275 –
– 7
1
1
1.05
1
1.13
1.08
1.69
1
1.15
12.81
2.28
1
1.01
1.16
1.1
1
3
1
2
3
27
24
20
27
7.38
4.82
6.21
7.92
DpAlg
GreedyAlg
Vienna1 Octi.
ScaleAlg
IlpAlg
466 403
466 403
466 –
466 –
0
0
–
–
2
7
6
2
0.74
1.26
2.04
1
1.29
2.43
3.14
1
0.99
8.59
6.22
1
1.07
2.48
2.99
1
8
1
1
8
21 12.8
21 9.41
19 10.05
21 12.8
DpAlg
GreedyAlg
Vienna2 Octi.
ScaleAlg
IlpAlg
468 324
468 324
468 –
468 –
1
1
–
–
8
18
6
4
2.63
4.79
2.63
1
1.03
1.72
1.07
1
4.01
14.23
2.55
1
1.52
3.36
1.4
1
1
1
1
5
16
12
12
16
8.12
4.31
7.87
10
DpAlg
GreedyAlg
Vienna3 Curved
ScaleAlg
IlpAlg
606 535
606 535
606 –
606 –
0
0
–
–
7
11
13
7
1
0.98
1.18
1
1
1.47
3.2
1
1
1.92
2.18
1
1
1.06
1.34
1
2
1
1
2
22
17
16
22
10.7
6.6
7.43
10.7
3
3
–
–
(a) Sydney4 : DpAlg, CurvedStyle
Layout from [5, Fig. 1a.].
(b) Sydney4 : GreedyAlg, CurvedStyle
Layout from [5, Fig. 1a].
Fig. 9: Labelings for Sydney using DpAlg and GreedyAlg.
Oracle). We did this in order to measure the actual running times of our algorithms and not to measure
the time that the virtual machine of Java spends for loading classes and optimizing byte code.
Table 2 and Table 3 present our quantitative results for the considered instances. For Sydney4
labelings are found in Fig. 9 and for Sydney2, Sydney3 and Vienna2 labelings created by DpAlg are
found in Fig. 10, Fig. 11, Fig. 12, respectively. The labelings of all instances are found in the appendix.
We first note that with respect to the total number of created candidates only few labels are
removed for enforcing Assumption 1 and Assumption 2; see Table 3, A12. This indicates that requiring
those assumptions is not a real restriction on a realistic set of candidates, even though they seem to
be artificial.
Running time. Even for large networks as Sydney, our algorithm DpAlg needs less than 0.6
seconds; see Table 2. This shows that our approach is applicable for scenarios in which the map
designer wants to adapt the layout and its labeling interactively. In particular in those scenarios not
every of the four steps must be repeated each time, which improves computing time. For example,
after once applying the scaling step (1st phase, 2nd step – the most time consuming step), the instance
does not need to be rescaled again, but the relation between label size and map size is determined.
Further, DpAlg is only moderately slower than GreedyAlg; 0.21 seconds in maximum, see Table 2,
Sydney4. On the other hand, the approaches IlpAlg and ScaleAlg are not alternatives, because
their running times are much worse; over 1 minute in maximum; see Table 2, Sydney4.
Quality. We observe that in all labelings created by DpAlg there are only few switchovers, namely
4–8; see Table 3, column SO. Hence, there are long sequences of consecutive labels that lie on the same
side of their metro line; see corresponding figures and Table 3, column Sequence. Together with the
ILP based approach IlpAlg, it yields the solution with the longest sequences in average. In particular
the switchovers are placed such that those sequences are regularly sized. The labels of a single sequence
are mostly directed into the same x-direction and in particular they are similarly shaped so that those
sequences of labels form regular patterns as desired. The alignment of the labels is chosen so that they
blend in with the alignment of their adjacent labels. In comparison with the solution of IlpAlg, the
DpAlg )
costs of DpAlg never exceed a factor of 1.52; see Table 3, column w(L
w(L) . For the instances Sydney4
and Vienna3 it even obtains a solution with the same costs. For the other instances, DpAlg basically
DpAlg )
spends its additional costs on the choice of the single labels ( w1w(L1 (L)
) and the distance of switchovers
DpAlg )
( w3w(L3 (L)
).
Berowra
Mt Kuring−gai
Mt Colah
Asquith
Richmond
East Richmond
Hornsby
Clarendon
Waitara
Windsor
Wahroonga
Mulgrave
Vineyard
Warrawee
Normanhurst
Turramurra
Riverstone
Thornleigh
Schofields
Quakers Hill
Pymble
Gordon
Pennant Hills
Killara
Marayong
Beecroft
Lindfield
Roseville
Cheltenham
R
d
n
s
rith
in
oo
gto
w
in
M
Pla
Pe
n
gs
err
St
W
Kin
Em
u
ary
s
M
tD
ru
itt
oo
ty
H
ill
D
oo
ns
id
e
Blacktown
Chatswood
Seven Hills
Epping
Toongabbie
Carlingford
Pendle Hill
Artarmon
St Leonards
Eastwood
Telopea
Wentworthville
Wollstonecraft
Denistone
Dundas
Westmead
Rydalmere
Parramatta
Waverton
West Ryde
North Sydney
Camellia
y
Meadowbank
Rosehill
rQ
ua
Harris Park
cti
ula
un
iJ
f
clif
nd
Bo
C
ro
gs
n
oo
d
ro
yd
o
As n
hfi
eld
Su
m
m
er
Le
H
ill
w
is
h
Pe am
te
rs
h
Sta am
nm
ore
N
ew
to
w
n
M
ac
do
na
ld
to
w
C
Bu
rw
om
eb
us
h
Museum
Central
H
Erskineville
lm
ore
rk
Green Square
St Peters
Mascot
C
Be
w
n
ow
l
ba
Pa
ke
m
La
on
a
hb
sto
y
nc
nk
go
ile
W
Pu
Ba
Ya
am
ps
ie
C
an
te
rb
ury
H
url
sto
ne
D
Pa
ulw
rk
ic
h
H
M
ill
arr
ic
kv
ille
Sy
de
nh
am
Se
fto
n
d
oo
hto
Warwick Farm
St James
Town Hall
Redfern
Strathfield
Birrong
C
m
ar
Regents Park
nfi
eld
he
ste
rH
ill
Cabramatta
arr
a
ge
Martin Place
Lidcombe
Olympic Park
Berala
Flemington
C
Ed
Concord West
Auburn
North Strathfield
Vill
aw
Kin
Wynyard
Guildford
Yennora
Fairfield
Canley Vale
ig
ss
irc
C
Rhodes
Clyde
Le
on
Milsons Point
Granville
Merrylands
Domestic
Liverpool
Tempe
International
Casula
Wolli Creek
Tu
rr
Arncliffe
Banksia
Ba
Be
rd
w
ell
ort
N
y
xle
ella
Pa
rk
h
H
ills
y
gs
gro
ve
Be
ve
rl
Kin
d
ee
arw
N
ds
to
w
erw
oo
R
iv
Pa
nia
hy
es
by
R
ev
H
ort
Pa
na
w
Ea
st
H
ols
ills
Glenfield
Macquarie Fields
Ingleburn
Rockdale
Minto
Kogarah
Leumeah
Carlton
Allawah
Campbelltown
Hurstville
Penshurst
Macarthur
Mortdale
Oatley
Como
Jannali
lla
are
nu
ro
C
h
a
ea
ba
nd
w
ym
ow
ng
ira
G
lo
M
ari
oo
ra
Kir
C
W
ee
Sutherland
Loftus
Engadine
Heathcote
Waterfall
(a) Original labeling by [11]
(b) Sydney2 : DpAlg, OctilinStyle
Layout from [11, Fig. 9b.].
Fig. 10: Labelings for Sydney.
In contrast, GreedyAlg yields significantly more switchovers; in maximum 20 switchovers more
than DpAlg, see Sydney4 . Consequently, there are many distracting switches of labels from one side
to the other of the metro line; e.g. see Fig. 9. Although the sequences of consecutive labels lying on
the same side may be longer in maximum compared to DpAlg, they are much shorter in average;
see Table 3, column Sequence. Further, several adjacent labels point in opposite x-directions, which
results in distracting effects; see corresponding figures. Altogether, the labelings that are obtained by
GreedyAlg do not look regular, but cluttered. DpAlg solves those problems since it considers the
metro line globally yielding an optimal labeling for a single line. This observation is also reflected in
DpAlg )
Table 3, column w(L
w(L) , which shows that the costs computed by GreedyAlg are significantly larger
than the costs computed by DpAlg. In particular costs for positioning the switchovers are much worse;
Sydney2 : w3 (LGreedyAlg )/w3 (L) = 19.52, w3 (LDpAlg )/w3 (L) = 3.25 and Vienna2 : w3 (LGreedyAlg )/w3 (L) = 14.23,
w3 (LDpAlg )/w3 (L) = 4.01. Hence, the better quality of DpAlg prevails the slightly better running time
of GreedyAlg.
Concerning the computed scale factor in the first phase of DpAlg, the labels are smaller than those
produced by ScaleAlg by a factor of 0.54–0.81; see Table 1. While this seems to be a drawback on the
first sight, the smaller size provides necessary space that is used to obtain a labeling of higher quality
with respect to the number and the placement of switchovers. Hence, the solutions of ScaleAlg
have more switchovers (except for Vienna2 ) and shorter sequences of labels lying on the same side in
average than DpAlg; see Table 3, column SO and Sequence.
We observe that both Nöllenburg and Wolff’s and our labelings of Sydney look quite similar,
whereas our labeling has less switchovers; see Fig. 10. The same applies for the labelings of the layout
of Vienna; see Fig. 12. Recall that their approach needed more than 10 hours to compute a labeled
metro map of Sydney. Since they need only up to 23 minutes to compute the layout without labeling,
it lends itself to first apply their approach to gain a layout and then to apply our approach to construct
a corresponding labeling.
Wang and Chi present in their paper [14] an approach that is divided into two phases. In the first
phase they compute the layout of the metro map and then in the second phase they create a labeling for
(a)
(b)
Fig. 11: Labelings for Sydney. (a) Original Labeling presented by Wang and Chi [14] (b) Sydney3 :
DpAlg.
Floridsdorf
Neue Donau
Handelskai
Dresdner Straße
Jägerstraße
Heiligenstadt
Währinger Straße Volksoper
Michelbeuern AKH
Ta
bo
r
Pr str
at aß
er e
st
er
n
Spittelau
Friedensbrücke
Roßauer Lände
Nußdorfer Straße
Messe
Trabrennstraße
Stadion
Schottenring
Alser Straße
Nestroyplatz
Josefstädter Straße
H K O
üt e t
te nd ta
ld le kr
or rs in
g
f
Sc J er traß
hw oh Str e
eg ns aß
le tra e
rs ß
tra e
ße
Museumsquartier
Gumpendorfer Straße
H
O üt
Br
b t
au U er eld
ns nte St orf
ch r S . V
w t. eit
eig V
M
eid
g eit
lin S Hieass
g ch t e
H ön zin
au b g
pt run
st n
ra
ße
Kettenbrückengasse
Pilgramgasse
Margaretengürtel
er
Schwedenplatz
St
Volkstheater
Westbahnhof
H
Zie
N gle
eu rg
ba as
ug se
as
se
Rathaus
re
St nga
ep s
ha se
ns
pla
tz
ub
en
to
r
Schottentor
Thaliastraße
Burggasse Stadthalle
Leopoldau
Großfeldsiedlung
Aderklaaer Straße
Rennbahnweg
Kagraner Platz
Kagran
Alte Donau
VIC Kaisermühlen
Donauinsel
Vorgartenstraße
Landstraße
Stadtpark
Karlsplatz
Taubstummengasse
Südtiroler Platz
Keplerplatz
Reumannplatz
Rochusgasse
Kardinal−Nagel−Platz
Schlachthausgasse
Erdberg
Gasometer
Zippererstraße
Enkplatz
Simmering
Längenfeldgasse
Niederhofstraße
Philadelphiabrücke
Tscherttegasse
Am Schöpfwerk
Alterlaa
Erlaaer Straße
Perfektastraße
Siebenhirten
(a)
(b)
Fig. 12: Labelings for Vienna. (a) Original Labeling presented by Nöllenburg and Wolff [11] (b) Vienna2 : DpAlg.
that layout. For both steps they formulate energy functions expressing their desired objectives, which
then are locally optimized. Figure 11(a) shows the metro map of Sydney created by their approach. In
comparison, Fig. 11(b) shows the same layout with a labeling created by our approach. Both labelings
look quite similar. While our approach needed 0.26s (see Table 2, Sydney3 ), their approach needed less
than 0.1s on their machine. However, their approach does not guarantee that the labels are occlusionfree, but labels may overlap with metro lines and other labels. This may result in illegible drawings
A
B
C
(a)
(b)
Fig. 13: Comparison of two labelings for the same line. (a) Labeling is created by the tool presented
by Wang and Chi[14]. (b) Labeling is created by DpAlg.
of metro maps. For example Fig. 13 shows two labelings of a metro line of Sydney that has been laid
out by the tool of Wang and Chi. Figure 13(a) shows a labeling that has been created by their tool,
while Fig. 13(b) shows a labeling that has been created by our approach. The labeling of Wang and
Chi has several serious defects that makes the map hardly readable. The marked regions A, B and C
show labels that overlap each other. Hence, some of the labels are obscured partly, while some of the
labels are completely covered by other labels. For example in region B the label St. Peters and the
label Erskinville overlap the label Macdonaldtown such that it is hardly viewable. Further, region C
contains two diagonal rows of stops aligned parallel. While the upper row is visible, the lower row is
almost completely covered by labels. Further, the labels of the upper row obscure the labels of the lower
row. In contrast our approach yields an occlusion-free labeling, such that each label and each stop is
easily legible. We therefore think that our approach is a reasonable alternative for the labeling step of
Wang and Chi’s approach. In particular, we think that the better quality of our approach prevails the
better running time of Wang and Chi’s approach.
In conclusion our workflow is a reasonable alternative and improvement for the approaches presented
both by Nöllenburg and Wolff, and by Wang and Chi. In the former case, our approach is significantly
faster, while in contrast to the latter case we can guarantee occlusion-free labelings.
Acknowledgment. We sincerely thank Herman Haverkort, Arlind Nocaj, Aidan Slingsby and Jo Wood
for helpful and interesting discussions.
References
1. P. K. Agarwal, M. van Kreveld, and S. Suri. Label placement by maximum independent set in rectangles.
Comp. Geom.-Theor. Appl., 11:209–218, 1998. 2
2. B. Aspvall, M. F. Plass, and R. E. Tarjan. A linear-time algorithm for testing the truth of certain quantified
boolean formulas. Inf. Process. Lett., 8(3):121–123, 1979. 16
3. J. Christensen, J. Marks, and S. Shieber. An empirical study of algorithms for point-feature label placement.
Acm. T. Graphic., 14(3):203–232, 1995. 2
4. T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein. Introduction to Algorithms (3. Edition). MIT
Press, 2009. 11
5. M. Fink, H. Haverkort, M. Nöllenburg, M. Roberts, J. Schuhmann, and A. Wolff. Drawing Metro Maps
Using Bézier Curves. In W. Didimo and M. Patrignani, editors, Graph Drawing, volume 7704 of LNCS,
pages 463–474. Springer Berlin Heidelberg, 2013. 1, 19, 22
6. M. Formann and F. Wagner. A packing problem with applications to lettering of maps. In ACM Sympos.
on Comput. Geom., pages 281–288, 1991. 2, 4
7. R. J. Fowler, M. S. Paterson, and S. L. Tanimoto. Optimal packing and covering in the plane are npcomplete. Inf. Process. Lett., 12(3):133 – 137, 1981. 4
8. E. Imhof. Positioning names on maps. Am. Cartographer, pages 128–144, 1975. 1, 2, 3, 9, 14
9. K. G. Kakoulis and I. G. Tollis. A unified approach to labeling graphical features. In Proceedings of
the Fourteenth Annual Symposium on Computational Geometry, SCG ’98, pages 347–356, New York, NY,
USA, 1998. ACM. 2
10. D. Lichtenstein. Planar formulae and their uses. SIAM J. Comput., 11(2):329–343, 1982. 4, 5
11. M. Nöllenburg and A. Wolff. Drawing and labeling high-quality metro maps by mixed-integer programming.
IEEE T. Vis. Comput. Gr., 17(5):626–641, 2011. 1, 19, 23, 24
12. J. Stott, P. Rodgers, J. Martinez-Ovando, and S. Walker. Automatic metro map layout using multicriteria
optimization. IEEE T. Vis. Comput. Gr., 17(1):101–114, 2011. 1
13. A. van Goethem, W. Meulemans, A. Reimer, H. Haverkort, and B. Speckmann. Topologically safe curved
schematisation. Cartogr. J., 50(3):276–285, 2013. 1
14. Y.-S. Wang and M.-T. Chi. Focus+context metro maps. IEEE T. Vis. Comput. Gr., 17(12):2528–2535,
2011. 1, 19, 23, 24, 25
15. A. Wolff. Graph drawing and cartography. In R. Tamassia, editor, Handbook of Graph Drawing and
Visualization, chapter 23, pages 697–736. CRC Press, 2013. 1
A
Labelings
(a) DpAlg
(b) GreedyAlg
(c) ScaleAlg
(d) IlpAlg
Fig. 14: Labelings for instance Sydney1.
(a) DpAlg
(b) GreedyAlg
(c) ScaleAlg
(d) IlpAlg
Fig. 15: Labelings for instance Sydney2.
(a) DpAlg
(b) GreedyAlg
(c) ScaleAlg
(d) IlpAlg
Fig. 16: Labelings for instance Sydney3.
(a) DpAlg
(b) GreedyAlg
(c) ScaleAlg
(d) IlpAlg
Fig. 17: Labelings for instance Sydney4.
(a) DpAlg
(b) GreedyAlg
(c) ScaleAlg
(d) IlpAlg
Fig. 18: Labelings for instance Vienna1.
(a) DpAlg
(b) GreedyAlg
(c) ScaleAlg
(d) IlpAlg
Fig. 19: Labelings for instance Vienna2.
(a) DpAlg
(b) GreedyAlg
(c) ScaleAlg
(d) IlpAlg
Fig. 20: Labelings for instance Vienna3.
| 8cs.DS
|
Sailfish: Alignment-free Isoform Quantification from
arXiv:1308.3700v1 [q-bio.GN] 16 Aug 2013
RNA-seq Reads using Lightweight Algorithms
Rob Patro1 , Stephen M. Mount2 , and Carl Kingsford1
1 Lane
Center for Computational Biology, School of Computer Science
Carnegie Mellon University
2 Department
of Cell Biology and Molecular Genetics and Center for
Bioinformatics and Computational Biology, University of Maryland
RNA-seq has rapidly become the de facto technique to measure gene expression. However,
the time required for analysis has not kept up with the pace of data generation. Here we
introduce Sailfish, a novel computational method for quantifying the abundance of previously
annotated RNA isoforms from RNA-seq data. Sailfish entirely avoids mapping reads, which is
a time-consuming step in all current methods. Sailfish provides quantification estimates much
faster than existing approaches (typically 20-times faster) without loss of accuracy.
The ability to generate genomic and transcriptomic data is accelerating beyond our ability to
process it. The increasingly widespread use and growing clinical relevance (e.g. [11]) of RNAseq measurements of transcript abundance will only serve to magnify the divide between our data
acquisition and data analysis capabilities.
The goal of isoform quantification is to determine the relative abundance of different RNA transcripts given a set of RNA-seq reads. In the analysis of RNA-seq data, isoform quantification is
1
one of the most computationally time-consuming steps, and it is commonly the first step in an analysis of differential expression among multiple samples [13]. There are numerous computational
challenges in estimating transcript-level abundance from RNA-seq data. Mapping the sequencing
reads to the genome or transcript sequences can require substantial computational resources. This
often leads to complicated models that account for read bias and error during inference, further
adding to the time spent on analysis. Finally, some reads, known as multireads [7, 15], can map
to multiple, sometimes many, different transcripts. The ambiguity resulting from these multireads
complicates the estimation of relative transcript abundances.
Existing approaches first use read-mapping tools, such as Bowtie [4], to determine potential
locations from which the RNA-seq reads originated. Given the read alignments, some of the
most accurate transcript quantification tools resolve the relative abundance of transcripts using
expectation-maximization (EM) procedures [15, 5, 10]. In such procedures, reads are first assigned
to transcripts, and these assignments are then used to estimate transcript abundances. The abundances are then used to re-estimate the read assignments, weighting potential matches in proportion
to the currently estimated relative abundances, and these steps are repeated until convergence. In
practice, both of these steps can be time consuming. For example, even when exploiting the parallel nature of the problem, mapping the reads from a reasonably sized (e.g. 100M reads) RNA-seq
experiment can take hours.
Recent tools, such as eXpress [10], aim to reduce the computational burden of isoform quantification from RNA-seq data by substantially altering the EM algorithm. However, even for such
advanced approaches, performing read alignment and processing the large number of alignments
that result from ambiguously mapping reads remains a significant bottleneck and fundamentally
limits the scalability of approaches that depend on mapping.
Sailfish, our software for isoform quantification from RNA-seq data, is based on the philosophy of lightweight algorithms, which make frugal use of data, respect constant factors, and effectively use concurrent hardware by working with small units of data where possible. Sailfish avoids
mapping reads entirely (Fig. 1), resulting in large savings in time and space. A key technical con-
2
tribution behind our approach is the observation that transcript coverage, which is essential for
isoform quantification, can be reliably and accurately estimated using counts of k-mers occurring
in reads. This results in the ability to obtain accurate quantification estimates more than an order of
magnitude faster than existing approaches, often in minutes instead of hours. For example, for the
data described in Figure 2, Sailfish is between 18 and 29 times faster than the next fastest method
while providing expression estimates of equal accuracy.
In Sailfish, the fundamental unit of transcript coverage is the k-mer. This is different from existing approaches, where the fragment or read is the fundamental unit of coverage. By working with
k-mers, we can replace the computationally intensive step of read mapping with the much faster
and simpler process of k-mer counting. We also avoid any dependence on read mapping parameters (e.g. mismatches and gaps) that can have a significant effect on both the runtime and accuracy
of conventional approaches. Yet, our approach is still able to handle sequencing errors in reads
because only the k-mers that overlap the erroneous bases will be discarded or mis-assigned, while
the rest of the read can be processed as if it were error-free. This also leads to Sailfish having only
a single explicit parameter, the k-mer length. Longer k-mers may result in less ambiguity, which
makes resolving their origin easier, but may be more affected by errors in the reads. Conversely,
shorter k-mers, though more ambiguous, may be more robust to errors in the reads (Supplementary
Fig. 1). Further, we can effectively exploit modern hardware where multiple cores and reasonably
large memories are common. Many of our data structures can be represented as arrays of atomic
integers (see Methods). This allows our software to be concurrent and lock-free where possible,
leading to an approach that scales well with the number of available CPUs (Supplementary Fig. 2).
Additional benefits of the Sailfish approach are discussed in Supplementary Note 1.
Sailfish works in two phases: indexing and quantification (Fig. 1). A Sailfish index is built
from a particular set of reference transcripts (a FASTA sequence file) and a specific choice of k-mer
length, k. The index consists of data structures that make counting k-mers in a set of reads and
resolving their potential origin in the set of transcripts efficient (see Methods). The most important
data structure in the index is the minimal perfect hash function [1] that maps each k-mer in the
3
(1) index (per reference & choice of k)
reference
transcripts
(2) quant
read data
perfect hash function
array of k-mer counts
transcript
k-mer
k-mers
transcripts
(per set of reads)
unhashable
k-mers
estimate
abundances
µ̂1
µ̂2
...
...
reallocate
counts
based on µ̂
µ̂M
µ̂1 µ̂2
µ̂M
Figure 1: The Sailfish pipeline consists of an indexing phase (1) that is invoked via the command
sailfish index and a quantification phase (2) invoked via the command sailfish quant. The
Sailfish index has four components: (a) a perfect hash function mapping each k-mer in the transcript set to a unique integer between 0 and N — the number of unique k-mers in the set of
transcripts; (b) an array recording the number of times each k-mer occurs in the reference set; (c)
an index mapping each transcript to the multiset of k-mers that it contains; (d) an index mapping
each k-mer to the set of transcripts in which it appears. The quantification phase consists of counting the indexed k-mers in the set of reads and then applying an EM procedure to determine the
maximum-likelihood estimates of relative transcript abundance.
4
reference transcripts to an index between 0 and the number of different k-mers in the transcripts
such that no two k-mers share an index. This allows us to quickly index and count any k-mer
from the reads that also appears in the transcripts. We find that pairing the minimum perfect
hash function with an atomically updateable array of k-mer counts allows us to count k-mers even
faster than with existing advanced lock-free hashes such as that used in Jellyfish [6]. The index
also contains a pair of look-up tables that allow fast access to the indexed k-mers appearing in a
specific transcript as well as the indexed transcripts in which a particular k-mer appears, both in
amortized constant time. Because the index depends only on the set of reference transcripts and
the choice of k-mer length, it only needs to be rebuilt when one of these factors changes.
The quantification phase of Sailfish takes as input the index described above and a set of RNAseq reads and produces an estimate of the relative abundance of each transcript in the reference,
measured in both Reads Per Kilobase per Million mapped reads (RPKM) and Transcripts Per Million (TPM); (see Methods for the definitions of these measures). First, Sailfish counts the number
of times each indexed k-mer occurs in the set of reads. Owing to efficient k-mer indexing by
means of the perfect hash function and the use of a lock-free counting data structure (Methods),
this process is efficient and scalable (Supplementary Fig. 2). Sailfish then applies an expectationmaximization (EM) procedure to determine maximum likelihood estimates for the relative abundance of each transcript. Conceptually, this procedure is similar to the EM algorithm used by
RSEM [5], except that k-mers rather than fragments are probabilistically assigned to transcripts,
and a two-step variant of EM is used to speed up convergence. The estimation procedure first
assigns k-mers proportionally to transcripts (i.e. if a transcript is the only potential origin for a
particular k-mer, then all observations of that k-mer are attributed to this transcript, whereas for a
k-mer that appears once in each of n different transcripts and occurs m times in the set of reads,
m/n observations are attributed to each potential transcript of origin). These initial allocations
are then used to estimate the expected coverage for each transcript (Methods, Eqn. 3). In turn,
these expected coverage values alter the assignment probabilities of k-mers to transcripts (Methods, Eqn. 1). Using these basic EM steps as a building block, we apply a globally-convergent
5
EM acceleration method, SQUAREM [16], that substantially increases the convergence rate of the
estimation procedure by modifying the parameter update step and step-length based on the current
solution path and the estimated distance from the fixed point (see Methods, Alg. 2).
Additionally, we reduce the number of variables that need to be fit by the EM procedure by
collapsing k-mers into equivalence classes. Two k-mers are equivalent from the perspective of the
EM algorithm if they occur in the same set of transcript sequences with the same rate (more details
available in Methods). This reduction in the number of active variables substantially reduces the
computational requirements of the EM procedure. For example, in the set of reference transcripts
for which we estimate abundance using the Microarray Quality Control (MAQC) [12] data (Fig. 2),
there are 60,504,111 k-mers (k = 20), of which 39,393,132 appear at least once in the set of reads.
However, there are only 151,385 distinct equivalence classes of k-mers with non-zero counts.
Thus, our EM procedure needs to optimize the allocations of 151,385 k-mer equivalence classes
instead of 39,393,132 individual k-mers, a reduction by a factor of ≈ 260.
Once the EM procedure converges, the estimated abundances are corrected for systematic errors due to sequence composition bias and transcript length using a regression approach similar to
Zheng et al. [18], though using random forest regression instead of a generalized additive model.
This correction is applied after initial estimates have been produced rather than at a read mapping
or fragment assignment stage, requiring fewer variables to be fit during bias correction.
To examine the efficiency and accuracy of Sailfish, we compared it to RSEM [5], eXpress [10]
and Cufflinks [15] using both real and synthetic data. Accuracy on real data was quantified by
the agreement between RNA-seq-based expression estimates computed by each piece of software
and qPCR measurements for the same sample (human brain tissue (HBR) in Fig. 2 and Supplementary Fig. 3, and universal human reference tissue (UHR) in Supplementary Fig. 4). These
paired RNA-seq and qPCR experiments were performed as part of the Microarray Quality Control
(MAQC) study [12]. qPCR abundance measurements are given at the resolution of genes rather
than isoforms. Thus, to compare these measurements with the transcript-level abundance estimates
produced by the software, we summed the estimates for all isoforms belonging to a gene to obtain
6
c
a
RPKM(log2 )
0
5
103
5
0
qPCR(log2 )
10
6.20h 6.75h
2.27h
102
15
15.72h
10.19h
Time in minutes
10
RPKM(log2 )
alignment
quantification
46.55h
10
b
104
0.26h
101
0.12h
5
0
0
5
10
Ground Truth(log2 )
d
15
fish RSEM Xpress fflinks
e
Cu
Sail
Human Brain Tissue
Pearson
Spearman
RMSE
medPE
Synthetic
SRX016366
100
fish RSEM Xpress fflinks
e
Cu
Sail
Synthetic
Sailfish
RSEM
eXpress
Cufflinks
Sailfish
RSEM
eXpress
Cufflinks
0.86
0.85
1.69
31.60
0.83
0.81
1.86
36.63
0.86
0.86
1.69
32.73
0.86
0.86
1.67
30.75
0.92
0.94
1.26
4.24
0.92
0.93
1.24
5.97
0.64
0.66
2.80
26.44
0.91
0.93
1.31
6.76
Figure 2: (a) The correlation between qPCR estimates of gene abundance (x-axis) and the estimates of Sailfish. The ground-truth results are taken from the microarray quality control study
(MAQC) [12]. The results shown here are for the human brain tissue and the RNA-seq based
estimates were computed using the reads from SRA accession SRX016366. The set of transcripts
used in this experiment were the curated RefSeq [9] genes (accession prefix NM) from hg18. (b)
The correlation between the actual number of transcript copies in a simulated dataset (x-axis) and
the abundance estimates of Sailfish. The transcripts used in this experiment were all Ensembl [2]
transcripts from hg19 that were annotated with a coding feature (CDS). (c) The total time taken
by each method, Sailfish, RSEM, eXpress and Cufflinks, to estimate isoform abundance on each
dataset. The total time taken by a method is the height of the corresponding bar, and the total is
further broken down into the time taken to perform read-alignment (for Sailfish, we instead measured the time taken to count the k-mers in the read set) and the time taken to quantify abundance
given the aligned reads (or k-mer counts). All tools were run in multi-threaded mode (where applicable) and were allowed to use up to 16 threads. The table (d) gives the accuracy of each of
the methods on both datasets, as measured by the Pearson and Spearman correlation coefficients,
root-mean-square error (RMSE) and median percentage error (medPE).
7
an estimate for that gene. We compare predicted abundances using correlation coefficients, rootmean-square error (RMSE), and median percentage error (medPE) (additional details available in
Supplementary Note 2). Figure 2 shows that the speed of Sailfish does not sacrifice any accuracy.
To show that Sailfish is accurate at the isoform level, we generated synthetic data using the Flux
Simulator [3], which allows versatile modeling of various RNA-seq protocols (see Supplementary
Note 3). Unlike synthetic test data used in previous work [5, 10], the procedure used by the Flux
Simulator is not based specifically on the generative model underlying our estimation procedure.
Sailfish remains accurate at the isoform level (Fig. 2).
The memory usage of Sailfish is comparable with that of other tools, using between 4 and 6 Gb
of RAM during isoform quantification for the experiments reported here.
Sailfish applies the idea of lightweight algorithms to the problem of isoform quantification
from RNA-seq reads and in doing so achieves a breakthrough in terms of speed. By eliminating
the necessity of read mapping from the expression estimation pipeline, we not only improve the
speed of the process but also simplify it considerably, eliminating the burden of choosing all but
a single external parameter (the k-mer length) from the user. As the size and number of RNA-seq
experiments grow, we expect Sailfish and its paradigm to remain efficient for isoform quantification
because the memory footprint is bounded by the size and complexity of the target transcripts and
the only phase that grows explicitly in the number of reads — k-mer counting — has been designed
to effectively exploit many CPU cores.
Sailfish is free and open-source software and is available at http://www.cs.cmu.edu/~ckingsf/
software/sailfish.
Methods
Indexing. The first step in the Sailfish pipeline is building an index from the set of reference transcripts T . Given a k-mer length k, we compute an index Ik (T ) containing four components. The
first component is a minimum perfect hash function h on the set of k-mers kmers(T ) contained
8
in T . A minimum perfect hash function is a bijection between kmers(T ) and the set of integers
{0, 1, . . . , |kmers(T )| − 1}. Sailfish uses the BDZ minimum perfect hash function [1]. The second
component of the index is an array C containing a count C (si ) for every si ∈ kmers(T ). Finally,
the index contains a lookup table F, mapping each transcript to the multiset of k-mers that it contains and a reverse lookup table R mapping each k-mer to the set of transcripts in which it appears.
The index is a product only of the reference transcripts and the choice of k, and thus needs only to
be recomputed when either of these changes.
Quantification. The second step in the Sailfish pipeline is the quantification of relative transcript
abundance; this requires the Sailfish index Ik (T ) for the reference transcripts T as well as a set of
RNA-seq reads R. First, we count the number of occurrences of each si ∈ kmers(T ) ∩ kmers(R).
Since we know exactly the set of k-mers that need to be counted and already have a perfect hash
function h for this set, we can perform this counting in a particularly efficient manner. We maintain
an array CR of the appropriate size |kmers(T )|, where CR (h(si )) contains the number of times
we have thus far observed si in R.
Sequencing reads, and hence the k-mers they contain, may originate from transcripts in either
the forward or reverse direction. To account for both possibilities, we check both the forward and
reverse-complement k-mers from each read and use a majority-rule heuristic to determine which of
the k-mers to increment in the final array of counts CR . If the number of k-mers appearing in h from
the forward direction of the read is greater than the number of reverse-complement k-mers, then we
only increment the counts for k-mers appearing in this read in the forward direction. Otherwise,
only counts for k-mers appearing in the reverse-complement of this read are incremented in the
array of counts. Ties are broken in favor of the forward directed reads. By taking advantage
of atomic integers and the compare-and-swap (CAS) operation provided by modern processors,
which allows many hardware threads to efficiently update the value of a memory location without
the need for explicit locking, we can stream through R and update the counts of CR in parallel
while sustaining very little resource contention.
9
We then apply an expectation-maximization algorithm to obtain estimates of the relative abundance of each transcript. We define a k-mer equivalence class as the set of all k-mers that appear in
the same set of transcripts with the same frequency. In other words, let χ(s) be a vector that so that
entry t of χ(s) gives how many times s appears in transcript t ∈ T . Then the equivalence class of a
k-mer si is given by [si ] = {s j ∈ kmers(T ) | χ(s j ) = χ(si )}. When performing the EM procedure,
we will allocate counts to transcripts according to the set of equivalence classes rather than the full
set of transcripts. We will let T ([si ]) = ∑s j ∈[si ] CR h(s j ) denote the total count of k-mers in R
that originate from equivalence class [si ]. We say that transcript t contains equivalence class [s] if
[s] is a subset of the multiset of k-mers of t and denote this by [s] ⊆ t.
Estimating abundances via an EM algorithm. The EM algorithm (Algo. 1) alternates between
estimating the fraction of counts of each observed k-mer that originates from each transcript (Estep) and estimating the relative abundances of all transcripts given this allocation (M-step).
The E-step of the EM algorithm computes the fraction of each k-mer equivalence class’ total
count that is allocated to each transcript. For equivalence class s j and transcript ti , this value is
computed by
µ̂i T ( s j )
,
α( j, i) =
∑t⊇[s j ] µ̂t
(1)
where µ̂i is the currently estimated relative abundance of transcript i. These allocations are then
used in the M-step of the algorithm to compute the relative abundance of each transcript. The
relative abundance of transcript i is estimated by
µ̂i =
µi
∑t j ∈T µ j
,
(2)
where µi is
µi =
∑[s j ]⊆ti α( j, i)
.
lˆi
(3)
The variable lˆi denotes the adjusted length of transcript i and is simply lˆi = li − k + 1 where li is the
length of transcript i in nucleotides.
10
Algorithm 1: An EM iteration. One iteration of the expectation-maximization procedure
that updates the estimated k-mer allocations α(·, ·) and computes new estimates of relative
µ.
transcript abundance µˆ 0 based on the current estimates of relative transcript abundance µ̂
µ)
1 function EM( µ̂
2 begin
3
for s j do
4
w = ∑t⊇[s j ] µ̂t
5
for t ⊇ s j do
6
α( j, i) = µ̂i T ( s j )/w
7
8
9
10
11
12
13
14
y=0
for t ∈ T do
Ct = ∑[s j ]⊆ti α( j, i)
µt0 = Ct /lˆi
y = y + µt
for t ∈ T do
µ̂ 0t = µt0 /y
return µˆ 0 = hµˆ 0 0 , µˆ 0 1 , . . . , µˆ 0 |T | i
Algorithm 2: A SQUAREM iteration. Updates the relative abundance estimates according to an accelerated EM procedure whose update direction and magnitude are dynamically
computed [16].
µ 1 = EM(µ̂
µ 0)
1 µ̂
µ 2 = EM(µ̂
µ 1)
2 µ̂
µ 1 − µ̂
µ0
3 r = µ̂
µ
µ 1) − r
4 v = ( µ̂ 2 − µ̂
5 γ = − ||r|| / ||v||
6 modify γ by backtracking, if necessary, to ensure global convergence
µ 3 = max(00, µ̂
µ 0 − 2γr + γ 2 v)
7 µ̂
µ 0 = EM(µ̂
µ 3)
8 µ̂
11
However, rather than perform the standard EM update steps, we perform updates according to
µ t = hµ̂0 , . . . , µ̂|T | i is a vector of relative
the SQUAREM procedure [16] described in Algo. 2. µ̂
abundance maximum-likelihood estimates, and EM(·) is a standard iteration of the expectationmaximization procedure as outlined in Algo. 1. For a detailed explanation of the SQUAREM
procedure and its proof of convergence, see [16]. Intuitively, the SQUAREM procedure builds an
µ from 3 successive steps along the EM solution path, and uses
approximation of the Jacobian of µ̂
the magnitude of the differences between these solutions to determine a step size γ by which to
update the estimates according to the update rule (line 7). The procedure is then capable of making
µ parameters, which substantially improves the speed of converrelatively large updates to the µ̂
gence. In Sailfish, the iterative SQUAREM procedure is repeated for a user-specified number of
steps (30 for all experiments reported in this paper; see Supplementary Fig. 5).
Bias Correction.
The bias correction procedure implemented in Sailfish is based on the model
introduced by Zheng et al. [18]. Briefly, it performs a regression analysis on a set of potential bias
factors where the response variables are the estimated transcript abundances (RPKMs). Sailfish
automatically considers transcript length, GC content and dinucleotide frequencies as potential
bias factors, as this specific set of features were suggested by Zheng et al. [18]. For each transcript, the prediction of the regression model represents the contribution of the bias factors to this
transcript’s estimated abundance. Hence, these regression estimates (which may be positive or
negative) are subtracted from the original estimates to obtain bias-corrected RPKMs. For further
details on this bias correction procedure, see [18]. The original method used a generalized additive
model for regression; Sailfish implements the approach using random forest regression to leverage
high-performance implementations of this technique. The key idea here is to do the bias correction
after abundance estimation rather than earlier in the pipeline. The bias correction of Sailfish can
be disabled with the --no-bias-correction command line option. Finally, we note that it is
possible to include other potential features, like normalized coverage plots that can encode positional bias, into the bias correction phase. However, in the current version of Sailfish, we have not
12
implemented or tested bias correction for these features.
Computing RPKM and TPM. Sailfish outputs both Reads Per Kilobase per Million mapped
reads (RPKM) and Transcripts Per Million (TPM) as quantities predicting the relative abundance
of different isoforms. The RPKM estimate is the most commonly used, and is ideally 109 times
the rate at which reads are observed at a given position, but the TPM estimate has also become
somewhat common [5, 17]. Given the relative transcript abundances µ̂i estimated by the EM
procedure described above, the TPM for transcript i is given by
TPMi = 106 µ̂i .
(4)
Let Ci = ∑[s j ]⊆ti α( j, i) be the number of k-mers mapped to transcript i. Then, the RPKM is given
by
RPKMi =
Ci
li/103
N
106
=
109 Clii
N
≈
109 µi
,
N
(5)
where N = ∑[si ] T ([si ]) and the final equality is approximate only because we replace li with lˆi .
Computing Accuracy Metrics. Since the RPKM (and TPM) measurements are only relative
estimates of isoform abundance, it is essential to put the ground-truth and estimated relative abundances into the same frame of reference before computing our validation statistics. While this centering procedure will not effect correlation estimates, it is important to perform before computing
M
RMSE and medPE. Let X = {xi }M
i=1 denote the ground-truth isoform abundances and Y = {yi }i=1
denote the estimated abundances. We transform the estimated abundances by aligning their centroid with that of the ground-truth abundances; specifically, we compute the centroid-adjusted
M
M
abundance estimates as Y 0 = {yi − ω}M
where
ω
=
x
−
y
/M. It is these centroid
∑
∑
i
i
i=i
i=1
i=1
adjusted abundance estimates on which we compute all statistics.
Simulated Data.
The simulated RNA-seq data was generated by the FluxSimulator [3] v1.2
with the parameters listed in Supplementary Note 3. This resulted in a dataset of 150M, 76 base13
pair pared-end reads. RSEM, eXpress and Cufflinks were given paired-end alignments since they
make special use of this data. Further, for this dataset, Bowtie [4] was given the additional flag
-X990 when aligning the reads to the transcripts, as 990 base-pairs was the maximum observed
insert size in the simulated data. TopHat [14] was provided with the option --mate-inner-dist
198, to adjust the expected mate-pair inner-distance to the simulated average. The read files were
provided directly to Sailfish without any extra information since the same quantification procedure
is used whether single or paired end reads are provided.
Software Comparisons. For all comparisons, both eXpress and RSEM were provided with the
same sets of aligned reads in BAM format. All reads were aligned with Bowtie [4] v0.12.9 using the parameters -aS and -v3, which allows up to three mismatches per read and reports all
alignments. To prepare an alignment for Cufflinks, TopHat was run using Bowtie 1 (--bowtie1)
and with options -N 3 and --read-edit-dist 3 to allow up to three mismatches per read. For
RSEM, eXpress and Cufflinks, the reported times were the sum of the times required for alignment (via Bowtie for RSEM and eXpress or via TopHat for Cufflinks) and the times required for
quantification. The time required for each method is further decomposed into the times for the
alignment and quantification steps in Fig. 2.
Choice of Software Options and Effect on Runtime.
Most expression estimation software,
including RSEM, eXpress and Cufflinks, provides a myriad of program options to the user which
allow for trade-offs between various desiderata. For example, the total time required by TopHat
and Cufflinks is lower when Cufflinks is run without bias correction (e.g. 1.92h as opposed to 2.27h
with bias correction on the SRX016366 data). However, without bias correction, Cufflinks yields
slightly lower accuracy (Pearson σ = 0.82, Spearman ρ = 0.81) than the other methods, while
still taking 16 times longer to run than Sailfish. Similarly, although aligned reads can be streamed
directly into eXpress via Bowtie, we empirically observed lower overall runtimes when aligning
reads and quantifying expressions separately (and in serial), so these times were reported. Also,
we found that on the synthetic data, the correlations produced by eXpress improved to Pearson
14
σ = 0.85, Spearman ρ = 0.87 and σ = 0.9, ρ = 0.92 when running the EM procedure for 10 or
30 extra rounds (-b 10 and -b 30). However, this also greatly increased the runtime (just for
the estimation step) from 1.9h to 16.71h and 42.12h respectively. In general, we attempted to run
each piece of software with the options that would be most common in a standard usage scenario.
However, despite the inherent difficulty of comparing a set of tools parameterized on an array
of potential options, the core thesis that Sailfish can provide accurate expression estimates much
faster than any existing tool remains true, as the fastest performing alternatives, when even when
sacrificing accuracy for speed, were over an order of magnitude slower than Sailfish.
Sailfish version 0.5 was used for all experiments, and all analyses were performed with a kmer size of k = 20. Bias correction was enabled in all experiments involving real but not simulated
data. The RPKM values reported by Sailfish were used as transcript abundance estimates.
RSEM [5] version 1.2.3 was used with default parameters, apart from being provided the alignment file (--bam), for all experiments, and the RPKM values reported by RSEM were used as
abundance estimates.
eXpress [10] version 1.3.1 was used for all experiments. It was run with default parameters
on the MACQ data, and without bias correction (--no-bias-correct) on the synthetic data. The
abundance estimates were taken as the FPKM values output by eXpress.
Cufflinks [15] version 2.1.1 was used for experiments and was run with bias correction (-b)
and multi-read recovery (-u) on the MACQ data, and with only multi-read recovery (-u) on the
synthetic data. The FPKM values output by Cufflinks were used as the transcript abundance estimates.
All experiments were run on a computer with 8 AMD OpteronTM 6220 processors (4 cores
each) and 256Gb of RAM. For all experiments, the wall time was measured using the built-in bash
time command.
Implementation of Sailfish. Sailfish has two basic subcommands, index and quant. The index
command initially builds a hash of all k-mers in the set of reference transcripts using the Jelly-
15
fish [6] software. This hash is then used to build the minimum perfect hash, count array, and
look-up tables described above. The index command takes as input a k-mer size via the -k option
and a set of reference transcripts in FASTA format via the -t parameter. It produces the Sailfish
index described above, and it can optionally take advantage of multiple threads with the target
number of threads being provided via a -p option.
The quant subcommand estimates the relative abundance of transcripts given a set of reads.
The quant command takes as input a Sailfish index (computed via the index command described
above and provided via the -i parameter). Additionally, it requires the set of reads, provided as
a list of FASTA or FASTQ files given by the -r parameter. Finally, just as in index command, the
quant command can take advantage of multiple processors, the target number of which is provided
via the -p option.
Sailfish is implemented in C++11 and takes advantage of several C++11 language and library
features. In particular, Sailfish makes heavy use of built-in atomic data types. Parallelization
across multiple threads in Sailfish is accomplished via a combination of the standard library’s
thread facilities and the Intel Threading Building Blocks (TBB) library [8]. Sailfish is available
as an open-source program under the GPLv3 license, and has been developed and tested on Linux
and Macintosh OS X.
Author Contributions
R.P., S.M.M. and C.K. designed the method and algorithms, devised the
experiments, and wrote the manuscript. R.P. implemented the Sailfish software.
Acknowledgments This work has been partially funded by National Science Foundation (CCF1256087, CCF-1053918, and EF-0849899) and National Institutes of Health (1R21AI085376 and
1R21HG006913). C.K. received support as an Alfred P. Sloan Research Fellow.
16
References
[1] F. C. Botelho, R. Pagh, and N. Ziviani. Simple and space-efficient minimal perfect hash
functions. In Algorithms and Data Structures, pages 139–150. Springer, 2007.
[2] P. Flicek, I. Ahmed, M. R. Amode, D. Barrell, K. Beal, S. Brent, D. Carvalho-Silva,
P. Clapham, G. Coates, S. Fairley, et al. Ensembl 2013. Nucleic Acids Research, 41(D1):D48–
D55, 2013.
[3] T. Griebel, B. Zacher, P. Ribeca, E. Raineri, V. Lacroix, R. Guigó, and M. Sammeth. Modelling and simulating generic RNA-Seq experiments with the flux simulator. Nucleic Acids
Research, 40(20):10073–10083, 2012.
[4] B. Langmead, C. Trapnell, M. Pop, S. L. Salzberg, et al. Ultrafast and memory-efficient
alignment of short DNA sequences to the human genome. Genome Biology, 10(3):R25,
2009.
[5] B. Li and C. Dewey. RSEM: accurate transcript quantification from RNA-Seq data with or
without a reference genome. BMC Bioinformatics, 12(1):323, 2011.
[6] G. Marçais and C. Kingsford. A fast, lock-free approach for efficient parallel counting of
occurrences of k-mers. Bioinformatics, 27(6):764–770, 2011.
[7] A. Mortazavi, B. A. Williams, K. McCue, L. Schaeffer, and B. Wold. Mapping and quantifying mammalian transcriptomes by RNA-Seq. Nature Methods, 5(7):621–628, 2008.
[8] C. Pheatt. Intel threading building blocks. Journal of Computing Sciences in Colleges,
23(4):298–298, Apr. 2008.
[9] K. D. Pruitt, T. Tatusova, G. R. Brown, and D. R. Maglott. NCBI Reference Sequences
(RefSeq): current status, new features and genome annotation policy. Nucleic Acids Research,
40(D1):D130–D135, 2012.
17
[10] A. Roberts, L. Pachter, et al. Streaming fragment assignment for real-time analysis of sequencing experiments. Nature methods, 10(1):71–73, 2013.
[11] S. Roychowdhury, M. K. Iyer, D. R. Robinson, R. J. Lonigro, Y.-M. Wu, X. Cao, S. KalyanaSundaram, L. Sam, O. A. Balbin, M. J. Quist, T. Barrette, J. Everett, J. Siddiqui, L. P. Kunju,
N. Navone, J. C. Araujo, P. Troncoso, C. J. Logothetis, J. W. Innis, D. C. Smith, C. D. Lao,
S. Y. Kim, J. S. Roberts, S. B. Gruber, K. J. Pienta, M. Talpaz, and A. M. Chinnaiyan. Personalized oncology through integrative high-throughput sequencing: A pilot study. Science
Translational Medicine, 3(111):111ra121, 2011.
[12] L. Shi, L. H. Reid, W. D. Jones, R. Shippy, J. A. Warrington, S. C. Baker, P. J. Collins,
F. De Longueville, E. S. Kawasaki, K. Y. Lee, et al. The MicroArray Quality Control (MAQC)
project shows inter-and intraplatform reproducibility of gene expression measurements. Nature Biotechnology, 24(9):1151–1161, 2006.
[13] C. Soneson and M. Delorenzi. A comparison of methods for differential expression analysis
of RNA-seq data. BMC Bioinformatics, 14(1):91, 2013.
[14] C. Trapnell, L. Pachter, and S. L. Salzberg. TopHat: discovering splice junctions with RNASeq. Bioinformatics, 25(9):1105–1111, 2009.
[15] C. Trapnell, B. A. Williams, G. Pertea, A. Mortazavi, G. Kwan, M. J. van Baren, S. L.
Salzberg, B. J. Wold, and L. Pachter. Transcript assembly and quantification by RNA-Seq
reveals unannotated transcripts and isoform switching during cell differentiation. Nature
Biotechnology, 28(5):511–515, 2010.
[16] R. Varadhan and C. Roland. Simple and globally convergent methods for accelerating the
convergence of any EM algorithm. Scandinavian Journal of Statistics, 35(2):335–353, 2008.
[17] G. P. Wagner, K. Kin, and V. J. Lynch. Measurement of mRNA abundance using RNA-seq
data: RPKM measure is inconsistent among samples. Theory in Biosciences, 131(4):281–
285, 2012.
18
[18] W. Zheng, L. M. Chung, and H. Zhao. Bias detection and correction in RNA-Sequencing
data. BMC Bioinformatics, 12(1):290, 2011.
19
Sailfish: Alignment-free Isoform Quantification from RNA-seq Reads using
Lightweight Algorithms
Rob Patro, Stephen M.Mount and Carl Kingsford
1
Supplementary Figure 1: Effect of k-mer length on retained data and k-mer
ambiguity
1.0
0.9
Ratio (count / total)
0.8
0.7
mapped k-mers / possible k-mers
unique k-mers / all k-mers
0.6
0.5
0.4
0.3
16
18
20
k-mer length
22
24
Supplementary Figure 1: As the k-mer length is varied in the range [15, 25] when processing the
synthetic dataset, we observe that a longer k-mer length results in a slight decrease in data retention
(denoted by the red line which shows the ratio of number of k-mers from the read set that were
hashable to the total number of k-mers appearing in the set of reads). Simultaneously, we observe
that the ratio of the number of unique k-mers (k-mers having a unique locus of origin) in the set of
transcripts to the total number of k-mers in the set of transcripts (blue line) increases as we make
k larger. It seems that, as expected, there is a trade-off in the choice of k, with a larger k resulting
in less robustness to sequencing error but a higher fraction unique k-mers and smaller k-mers
providing more robustness to errors in the data but at the cost of increased ambiguity. However,
since the differences are relatively small over a reasonably large range of k, we can expect the
inference procedure to be fairly robust to this parameter. We use k = 20 for all experiments,
and this is the default in Sailfish. However, we did not attempt to optimize this parameter when
performing our experiments.
2
Supplementary Figure 2: Speed of counting indexed k-mers
Supplementary Figure 2: The time to count all of the k-mers and quantify transcript abundance
in an 81M read dataset (SRX016366) as function of the number of concurrent hashing threads.
Even with only a single thread, the counts for all k-mers in the dataset can be processed in 34
minutes and 26 seconds, while with 32 processing threads, all k-mers can be counted in only 1
minute and 28 seconds.
3
Supplementary Note 1: Additional benefits of the Sailfish approach
An additional benefit of our lightweight approach is that the size of the indexing and counting
structures required by Sailfish are a small fraction of the size of the indexing and alignment files
required by most other methods. For example, for the MAQC dataset described in Figure 2, the
total size of the indexing and count files required by Sailfish for quantification was 2.4Gb, compared with much larger indexes and accompanying alignment files in BAM format used by other
approaches (e.g., the 15.5Gb index and alignment file produced by Bowtie [4]). Unlike alignment
files which grow with the number of reads, the Sailfish index files grow only with the number of
unique k-mers and the complexity of the transcriptome’s k-mer composition and are independent
of the number of reads.
4
Supplementary Figure 3: Correlation plots with qPCR on human brain
tissue and synthetic data
RSEM
Cufflinks
0
0
5
5
5
10
10
RPKM(log2 )
0
10
20
RPKM(log2 )
eXpress
5
qPCR(log2 )
0
10
10
5
qPCR(log2 )
0
10
20
20
15
15
15
10
10
10
5
5
5
0
0
5
15
Ground Truth(log2 )
10
20
0
0
5
15
Ground Truth(log2 )
10
20
0
0
5
5
0
10
15
qPCR(log2 )
Ground Truth(log2 )
20
Supplementary Figure 3: Correlation plots of RSEM, eXpress and Cufflinks for the data presented in Fig. 2. Each column is labeled with the method whose output was used to produce that
column’s plots. The top row of plots show the correlation between the computed RPKM and the
qPCR-based expression estimates for the human brain tissue. The bottom row of plots shows the
correlation between the computed RPKM and the true abundance of each transcript on the synthetic dataset. To generate the results shown here, eXpress was run using its default streaming
expression estimation algorithm. As reported in Methods, additional batch EM iterations improve
eXpress’s accuracy, but come at the cost of a substantial increase in runtime.
5
Supplementary Figure 4: Correlation with qPCR on universal human
reference tissue
RPKM(log2 )
Sailfish
RSEM
eXpress
Cufflinks
5
5
5
5
0
0
0
0
5
5
5
5
10
10
10
10
10
5
0
qPCR(log2 )
5
10
5
qPCR(log2 )
5
0
10
104
5
5
0
qPCR(log2 )
10
5
0
qPCR(log2 )
5
alignment
quantification
35.60h
103
Time in minutes
8.60h
2.60h
102
101
100
0.10h
fish
Sail
M
RSE
SRX016367
ress
eXp
s
flink
Cuf
Sailfish
RSEM
eXpress
Cufflinks
Pearson
0.87
0.85
0.87
0.87
Spearman
0.88
0.85
0.88
0.88
RMSE
1.64
1.81
1.65
1.68
medPE
29.95
34.77
31.03
27.33
Supplementary Figure 4: The accuracy of four methods on a second dataset from the MACQ [12]
study. The reads for this experiment were taken from SRA accession SRX016367 (≈ 93M reads)
and are from a mixture of different tissues (i.e. the Universal Human Reference or UHR). The
same set of reference transcripts were used as in Fig. 2 of the main text. The relative accuracy
and performance of the methods is similar to what we observed in the other MACQ dataset, with
Sailfish, eXpress and Cufflinks all achieving comparable accuracy (all slightly more accurate than
RSEM). Sailfish is ≈ 26 times faster then Cufflinks, the closest method in terms of speed.
6
Supplementary Note 2: Additional details of accuracy analysis
We compare predicted abundances using correlation coefficients (Pearson & Spearman), rootmean-square error (RMSE), and median percentage error (medPE). These metrics allow us to
gauge the accuracy of methods from different perspectives . For example, the Pearson correlation
coefficient measures how well trends in the true data are captured by the methods, but, because the
correlation is taken in the log scale, it discounts transcripts with zero (or very low) abundance in
either sample, while the RMSE includes transcripts with true or estimated abundance of zero. Both
eXpress and Cufflinks produced a few outlier transcripts, with very low but non-zero estimated
abundance, which significantly degraded the Pearson correlation measure. We discarded these
outliers by filtering the output of these methods, and setting to zero any estimated RPKM less than
or equal to 0.01, a cutoff chosen because it removed the outliers but did not seem to discard any
other truly expressed transcripts.
7
Supplementary Note 3: Parameters for simulated data
The simulated RNA-seq data was generated by the FluxSimulator [3] v1.2 with the following
parameters.
### Expression ###
NB_MOLECULES
5000000
REF_FILE_NAME
hg19_annotations.gtf
GEN_DIR
hg19/chrs
LOAD_NONCODING NO
TSS_MEAN
50
POLYA_SCALE
NaN
POLYA_SHAPE
NaN
### Fragmentation ###
FRAG_SUBSTRATE RNA
FRAG_METHOD
UR
FRAG_UR_ETA
350
FRAG_UR_D0
1
### Reverse Transcription ###
RTRANSCRIPTION YES
RT_PRIMER
RH
RT_LOSSLESS
YES
RT_MIN 500
RT_MAX 5500
### Amplification ###
GC_MEAN NaN
PCR_PROBABILITY 0.05
FILTERING
NO
### Sequencing ###
READ_NUMBER
150000000
READ_LENGTH
76
PAIRED_END
YES
ERR_FILE
76
FASTA
YES
UNIQUE_IDS
YES
8
Supplementary Figure 5: Convergence of relative abundance estimates
−5
×10
1.2
1.0
|µ̂0−µ̂1|
|T |
0.8
0.6
0.4
0.2
0.0
0
5
10
15
iteration
20
25
30
Supplementary Figure 5: The average difference between the relative abundance as estimated by
two successive applications of the EM step (Algo. 2 lines 1–2) versus iterations of the SQUAREM
algorithm (in the Universal Human Reference tissue experiment). We can see that the residual
drops off quickly, and appears to have converged before 30 iterations of the SQUAREM procedure
have been performed.
9
| 5cs.CE
|
Multiscale model reduction for shale gas transport in fractured
media
I. Y. Akkutlu
∗
Yalchin Efendiev†
Maria Vasilyeva‡
arXiv:1507.00113v1 [cs.CE] 1 Jul 2015
July 2, 2015
Abstract
In this paper, we develop a multiscale model reduction technique that describes shale gas transport
in fractured media. Due to the pore-scale heterogeneities and processes, we use upscaled models to
describe the matrix. We follow our previous work [1], where we derived an upscaled model in the form
of generalized nonlinear diffusion model to describe the effects of kerogen. To model the interaction
between the matrix and the fractures, we use Generalized Multiscale Finite Element Method [13, 17].
In this approach, the matrix and the fracture interaction is modeled via local multiscale basis functions.
In [17], we developed the GMsFEM and applied for linear flows with horizontal or vertical fracture
orientations on a Cartesian fine grid. In this paper, we consider arbitrary fracture orientations and use
triangular fine grid and developed GMsFEM for nonlinear flows. Moreover, we develop online basis
function strategies to adaptively improve the convergence. The number of multiscale basis functions in
each coarse region represents the degrees of freedom needed to achieve a certain error threshold. Our
approach is adaptive in a sense that the multiscale basis functions can be added in the regions of interest.
Numerical results for two-dimensional problem are presented to demonstrate the efficiency of proposed
approach.
1
Introduction
Shale gas transport is an active area of research due to a growing interest in producing natural gas from
source rocks. The shale systems have added complexities due to the presence of organic matter, known as
kerogen. The kerogen brings in new fluid storage and transport qualities to the shale. A number of authors,
e.g., Loucks et al. (2009), Sondergeld et al. (2010), and Ambrose et al. (2012), [26, 31, 3], have previously
discussed the physical properties of the kerogen using scanning electron microscopy (SEM) and showed the
co-existence of nanoporous kerogen and microporous conventional inorganic rock materials.
Gas transport in the kerogen typically develops at low Reynolds number and relatively high Knudsen
number values. Under these conditions, it is expected that the transport is not driven by laminar (Darcy)
flow dominantly but instead by the pore diffusion and other molecular transport mechanisms such as Knudsen
diffusion and the adsorbed phase (or surface) diffusion. The latter introduces nonlinear processes at the pore
scale that occur in heterogeneous pore geometry. Some types of upscaled models are needed to represent
these complex processes for reservoir simulations.
In large-scale simulations, the complex pore-scale transport needs to be coupled to the transport in fractures. This brings an additional difficulty in multiscale simulations. In particular, the multiscale simulations
of the processes describing the interaction between the fracture and the matrix require reduced-order model
approaches that work for problems without scale separation and high contrast. The objective of this paper
∗ Department
of Petroleum Engineering, Texas A & M University, College Station, TX
of Mathematics, Texas A&M University, College Station, TX.
‡ Department of Computational Technologies, Institute of Mathematics and Informatics, North-Eastern Federal University,
Yakutsk, Republic of Sakha (Yakutia), Russia, 677980 & Institute for Scientific Computation, Texas A&M University, College
Station, TX 77843-3368
† Department
1
is to discuss the development of such approaches for describing the fracture and the matrix interaction by
taking the upscaled matrix model following our previous work [1].
In our previous work [1], we proposed a set of macroscopic models that take into account the nanoporous
nature and nonlinear processes of the shale matrix. Our derivation uses multiple scale asymptotic analysis
applied to mass balance equations, equation of state (for free gas) and isotherm of adsorption. The fine-scale
microscopic description is largely based on the model formulated by Akkutlu and Fathi (2012), [2]. The
macroscopic parameters that appear in the equations require solutions of cell problem defined in representative volume elements (RVEs). These RVE problems take into account fine-scale variations and average their
effects on macro scale.
The multiscale approaches proposed in [1] are limited to representing the features that have scale separation. To represent the fracture network and the interaction between the fracture network and the matrix,
we present a multiscale approach following the framework of Generalized Multiscale Finite Element Method
(GMsFEM), [13]. The main idea of GMsFEM is to use multiscale basis functions to extract an essential
information in each coarse grid (computational grid) and develop a reduced-order model. In [17], we have
developed the GMsFEM and applied for linear flows with horizontal or vertical fracture orientations on a
Cartesian fine grid. In this paper, our contributions are: (1) the use of arbitrary fracture orientations and
use triangular fine grids; (2) the development of GMsFEM for nonlinear flows; and (3) the development of
online basis function strategies to adaptively improve the convergence.
To represent the fractures on the fine grid, we use Discrete Fracture Model (DFM) [35]. The fine grid
is constructed to resolve the fractures. For the coarse grid, we choose a rectangular grid. The GMsFEM
framework uses these fine-scale models in computing the snapshot space and the offline space. The nonlinear
models are handled with GMsFEM by locally updating multiscale basis functions.
The study of flows in fractured media has a long history. Some modeling techniques on the fine grid
include the Discrete Fracture Model (DFM), Embedded Fracture Model (EFM) [27, 25, 23], the singlepermeability model, the multiple-permeability models ([33, 4, 28, 23, 20, 34, 30, 10]), and hierarchical fracture
models [23]. Though these approaches are designed for fine-scale simulations, a number of these approaches
represent the fractures at a macroscopic level. For example, multiple-permeability models represent the
network of connected fractures macroscopically by introducing several permeabilities in each block. The
EFM ([27, 25, 23]) models the interaction of fractures with the fine-grid blocks separately for each block.
The main idea of hierarchical fracture modeling presented in [23] is to homogenize small-length fractures
(with the length smaller than the coarse block), while to represent the large-length fractures. Some of these
approaches can be generalized by incorporating the interaction of fractures and permeability heterogeneities
locally, which can lead to efficient upscaling techniques, [11, 19].
In recent papers [21], several multiscale approaches are proposed for representing the fracture effects.
These approaches share common concepts with the methods that we discuss here in a sense that they add
new degrees of freedom to represent the fractures on a coarse grid. The main difference is that our approaches
use local spectral problems accompanied by adaptivity to detect the regions, where to add new basis functions.
In this regard, the procedure of finding multiscale basis functions and the enrichment procedure is different
from existing techniques.
The proposed method constructs multiscale basis functions by appropriately selecting local snapshot space
and the local spectral problems for the underlying nonlinear problem. The local spectral problems allow us
to adaptively enrich in the regions with larger errors. In the paper, we discuss adaptivity issues and how to
add multiscale basis functions in some selected regions. To reduce the computational cost associated with
constructing the snapshot space, we follow [5] and use randomized boundary conditions. One of other novel
components of the paper is the use of online basis functions (see [8] for online basis functions for steady state
problems) for the time-dependent nonlinear problems. The online basis functions are constructed during the
simulation using the residual and they can reduce the error significantly. These basis functions are used if
the offline basis functions can not reduce the error below a desired threshold.
We present numerical results for some representative examples. In these examples, we use nonlinear
matrix and fracture models. Our numerical results show that the coarse-scale models with a fewer degrees
of freedom can be used to get an accurate approximation of the fine-scale solution. In particular, only 10 %
2
degrees of the freedom are needed to obtain an accurate representation of the fine-scale solution. We also
add a geomechanical contribution to the permeability term, where the permeability depends on the pressure.
Furthermore, we demonstrate the use of online basis functions and how they can reduce the error.
The paper is organized as follows. In the next section, we present a model problem. In Section 3, we
discuss the fine-scale model. Section 4 is devoted to the development of GMsFEM, in particular, the offline
spaces. In this section, we present numerical results for offline basis functions. In Section 5, we discuss
randomized snapshot spaces and show that their use can give similar accuracy while for less computational
cost. In Section 6, we develop online basis functions and present numerical results.
2
Model Problem
In this paper, we will study nonlinear gas transport in fractured media motivated by several applications
including shale gas. We are interested in the shale gas transport described in [2]. Similar equations arise in
other models, where one considers a free gas in the tight reservoirs. We will consider general equations
am (c)
∂c
= div(bm (c, x)∇c),
∂t
(1)
where c is the amount of free gas and a(c) and b(c) contain terms related to storage and adsorption coefficients.
In [2], the authors consider the nonlinear terms have the forms
am (c) = φ + (1 − φ)γ
∂F
,
∂c
bm (c, x) = φD + (1 − φ)γDs
κ
∂F
+ φ RT c,
∂c
µ
where γ is a parameter, which is unity in kerogen and is equal to Vgrain,k /Vgrain in the inorganic material
(Vgrain,k is grain volume and Vgrain is kerogen grain volume). Diffusivity D and porosity φ are defined for
the free fluid in the inorganic matrix and in the kerogen as follows
(
(
Dk in kerogen
φk in kerogen
D=
, φ=
.
Di
in inorganic matrix
φi
in inorganic matrix
For the free gas we have ideal gas assumption. The Darcy law of free gas flow in inorganic matrix is
used with permeability κ and gas viscosity µ. For the sorbed gas we can use Langmuir or Henry‘s isotherms
F = F (c). In [24, 35], the authors discuss a general framework, where the equations also include nonlinear
diffusivity due to adsorbed gas in a shale formation. In [22], the nonlinear terms appear due to barotropic
effects.
The nonlinear flows also contain components that are due to diffusion in the fractures. One needs
additional equations for modeling fractures. The fractures have high conductivity. We will use a general
equation of the form
∂c
af (c)
= div(bf (c, x)∇c)
(2)
∂t
to describe the flow within fractures. In [2], the authors use
af (c) = φf ,
bf (c, x) =
κf
RT c,
µ
where φf and κf are the fracture porosity and permeability. These problems are solved on a fine grid using
DFM as will be described in Section 3.
In many shale gas examples, the matrix heterogeneities can be upscaled and the resulting upscaled equation has the form (1). However, the interaction between matrix and fractures require some type of multiscale
modeling approach, where the effects of the fractures need to be captured more accurately. Approaches, such
as multicontinuum [24] are often used, but these approaches use idealized assumptions on fracture distributions. In this paper, we will use multiscale basis functions to represent fracture effects. In our previous work
3
[17], we have considered similar approaches for single-phase flow when fractures (which could be horizontal
or vertical) are aligned with Cartesian grid. In this paper, we consider arbitrary fracture distribution in the
context of nonlinear flow equations.
The overall model equations will be solved on a coarse grid. Next, we introduce the concepts of fine and
coarse grids. Let T H be a usual conforming partition of the computational domain Ω into finite elements
(triangles, quadrilaterals, tetrahedra, etc.). We refer to this partition as the coarse grid and assume that
each coarse element is partitioned into a connected union of fine grid blocks. The fine grid partition will be
denoted by T h , and is by definition a refinement of the coarse grid T H . We use {xi }N
i=1 (where N denotes
the number of coarse nodes) to denote the vertices of the coarse mesh T H and define the neighborhood of
the node xi by
[
(3)
ωi = {Kj ∈ T H ; xi ∈ K j }.
See Figure 1 for an illustration of neighborhoods and elements subordinated to the coarse discretization.
Figure 1: Illustration of a coarse neighborhood and coarse element
We emphasize the use of ωi to denote a coarse neighborhood, and K to denote a coarse element throughout
the paper.
3
Fine-scale discretization
To discretize the system on fine grid, we will use finite element method and use DFM for fractures. To
solve Problem (1) using finite element method (FEM), we need a fine grid discretization to capture the
fractures. These computations can be expensive. Here, we apply the discrete fracture network (DFM) model
for modeling flows in fractures [29].
In the discrete-fracture model, the aperture of the fracture appears as a factor in front of the one
dimensional integral for the consistency of the integral form. This is the main idea of the discrete-fracture
model, which can be applied in any complex configuration for fractured porous media.
To demonstrate it, we consider the two-dimensional problem of Equation (2). We simplify the fractures as
the lines with small aperture. Thus, one-dimensional element is needed to describe fractures in the discretefracture model. The system of equations (1) will be discretized in a two-dimensional form for the matrix
4
and in one-dimensional form for the fractures. The whole domain Ω can be represented by
Ω = Ωm ⊕i di Ωf,i ,
(4)
where m and f represent the matrix and the fracture of the permeability field κ, respectively. Here, di is the
aperture of the i th fracture and i is the index of the fractures. Note that Ωm is a two-dimensional domain
and Ωf,i is a one-dimensional domain (Figure 2). Then Equations (1) and (2) and can be written as follows
(for any test function v):
Z
XZ
∂c
∂c
∂c
am (c) v dx + di
m( , v) + a(c, v) =
af (c) v dx+
∂t
∂t
∂t
Ωm
Ωf,i
i
Z
Z
(5)
X
+
bm (c, x)∇c · ∇v dx + di
bm (c, x)∇c · ∇v dx = 0.
Ωm
i
Ωf,i
Figure 2: Fine grid with fractures
To solve (5), we will first linearize the system. We will use the following linearization
Z
XZ
cn+1 − cn
cn+1 − cn
cn+1 − cn
m(
, v) + a(cn+1 , v) =
am (cn )
v dx + di
af (cn )
v dx+
τ
τ
τ
Ωm
Ωf,i
i
(6)
Z
XZ
n
n+1
n
n+1
+
bm (c , x)∇c
· ∇v dx + di
bf (c , x)∇c
· ∇v dx = 0.
Ωm
i
Ωf,i
The standard fully-implicit finite difference scheme is used for the approximation with time step size τ and
superscripts n, n+1 denote previous and current time levels. This is a first-order in time and unconditionally
stable linearization.
PNf
For standard Galerkin finite element method, we write the solution as c = i=1
ci φi , where φi are the
standard linear element basis functions defined on T h and Nf denotes the number of the nodes on the fine
grid. The equation (6) can be presented in matrix form:
Mn
cn+1 − cn
+ An cn+1 = 0,
τ
5
(7)
where M is the mass matrix given by
Z
n
M = [mij ] =
n
am (c )φi φj dx + di
XZ
Ωm
af (cn )φi φj dx,
Ωf,i
i
and A is the stiffness matrix given by
Z
XZ
n
n
A = [aij ] =
bm (c , x)∇φi · ∇φj dx + di
Ωm
bf (cn , x)∇φi · ∇φj dx.
Ωf,i
i
Hence at each time step we have the following linear problem
Qn cn+1 = M n cn ,
(8)
where Qn = (M n + τ An ). This fine scale discretization yields large matrices of size Nf × Nf .
4
Coarse-grid discretization using GMsFEM. Offline spaces.
We use multiscale basis functions to represent the solution space. We will consider the continuous Galerkin
(CG) formulation and signify ωi as the support of basis functions. We denote the basis functions by ψkωi ,
which is supported in ωi , and the index k represents the numbering of these basis functions. In turn, the
CG solution will be sought as
X
cms (x, t) =
cik (t)ψkωi (x).
i,k
Once the basis functions are identified, the CG global coupling is given through the variational form
m(
∂c
, v) + a(cms , v) = 0,
∂t
for all v ∈ Voff ,
(9)
where Voff is used to denote the space spanned by those basis functions and
Z
XZ
m(c, v) =
am c v dx + di
af c v dx,
Ωm
Z
bm ∇c · ∇v dx + di
a(c, v) =
Ωf,i
i
Ωm
XZ
i
bf ∇c · ∇v dx.
Ωf,i
Let V be the conforming finite element space with respect to the fine-scale partition T h . We assume
c ∈ V is the fine-scale solution satisfying
m(
∂c
, v) + a(c, v) = 0,
∂t
v ∈ V.
(10)
Next, we describe GMsFEM. GMsFEM consists of offline and online stage. In the offline stage we
construct multiscale basis functions and after that in the online stage, we solve our problem for any input
parameters, such as right hand sides or boundary conditions.
Offline computations:
Step 1. Coarse grid generation.
Step 2. Construction of the snapshot space that will be used to compute an offline space.
Step 3. Construction of a “small” dimensional offline space by performing dimension reduction in the
space of local snapshots.
6
Given the computational domain, a coarse grid can be constructed and local problems are solved on
coarse neighborhoods to obtain the snapshot spaces. Then, smaller dimensional offline spaces are obtained
from the snapshot spaces by dimension reduction via some spectral problems [13, 12, 16, 14, 5]. After that
we can solve our problem in the constructed offline space. Moreover, we will construct online basis functions
that are problem dependent and are computed locally based on some local residuals [9, 8].
We now present the construction of the offline basis functions and the corresponding spectral problems
ω
for obtaining a space reduction. In the offline computation, we first construct a snapshot space Vsnap
. The
snapshot space can be the space of all fine-scale basis functions or the solutions of some local problems with
various choices of boundary conditions. For example, we can use the following κ-harmonic extensions to form
a snapshot space. For each fine-grid function, δjh (x), which is defined by δjh (x) = δj,k , ∀j, k ∈ Jh (ωi ), where
Jh (ωi ) denotes the fine-grid boundary node on ∂ωi . For simplicity, we omit the index i. Given a fine-scale
piecewise linear function defined on ∂ω (here ω is a generic coarse element), we define ψjω,snap by following
variational problem
Z
XZ
a(ψjω,snap , v) =
bm ∇ψjω,snap · ∇v dx + dj
bf ∇ψjω,snap · ∇v dx = 0 in ω,
(11)
ωm
ωf,j
j
and ψjω,snap = δjh (x) on ∂ω, ω = ωm ⊕j dj ωf,j .
For brevity of notation, we now omit the superscript ω, yet it is assumed throughout this section that the
offline space computations are localized to respective coarse subdomains. Let li be the number of functions
in the snapshot space in the region ω, and
Vsnap = span{ψjsnap :
1 ≤ j ≤ li },
for each coarse subdomain ω.
Denote
.
Rsnap = ψ1snap , . . . , ψlsnap
i
ω
In order to construct the offline space Voff
, we perform a dimension reduction of the snapshot space using
an auxiliary spectral decomposition. The analysis in [15] motivates the following eigenvalue problem in the
space of snapshots:
off off off
Aoff Ψoff
(12)
k = λk S Ψk ,
where
A
off
=
[aoff
mn ]
Z
=
snap
bm ∇ψm
·
∇ψnsnap
dx + dj
XZ
ωm
j
S off = [soff
mn ] =
Z
snap snap
bm ψ m
ψn dx + dj
snap
T
bf ∇ψm
· ∇ψnsnap dx = Rsnap
ARsnap ,
ωf,j
XZ
ωm
j
snap snap
T
bf ψ m
ψn dx = Rsnap
SRsnap ,
ωf,j
where A and S denote analogous fine scale matrices as defined by
Z
XZ
n
Aij =
bm (c , x)∇φi · ∇φj dx + di
bf (cn , x)∇φi · ∇φj dx,
Dm
Df,i
i
Z
Sij =
bm (cn )φi φj dx + di
Dm
XZ
i
bf (cn )φi φj dx,
Df,i
ω
where φi is the fine-scale basis function. To generate the offline space, we then choose the smallest Moff
eigenvalues from Eq. (12) and form the corresponding eigenvectors in the space of snapshots by setting
Pli
snap
ω
off
Ψoff
(for k = 1, . . . , Moff
), where Ψoff
ψkoff = j=1
kj ψj
kj are the coordinates of the vector Ψk .
Next, we create an appropriate solution space and variational formulation that for a continuous Galerkin
approximation. We begin with an initial coarse space V0init = span{χi }N
i=1 . Recall that N denotes the
7
number of coarse neighborhoods. Here, χi are the standard multiscale partition of unity functions defined
by
Z
XZ
a(χi , v) =
bm ∇χi · ∇v dx + dj
bf ∇χi · ∇v dx = 0 K ∈ ω
(13)
ωm
ωf,j
j
χi = gi
on ∂K,
for all K ∈ ω, where gi is a continuous function on ∂K and is linear on each edge of ∂K.
ωi
We then multiply the partition of unity functions by the eigenfunctions in the offline space Voff
to
construct the resulting basis functions
ψi,k = χi ψkωi ,off
ωi
for 1 ≤ i ≤ N and 1 ≤ k ≤ Moff
,
(14)
ωi
where Moff
denotes the number of offline eigenvectors that are chosen for each coarse node i. We note that
the construction in Eq. (14) yields continuous basis functions due to the multiplication of offline eigenvectors
with the initial (continuous) partition of unity. Next, we define the continuous Galerkin spectral multiscale
space as
ωi
Voff = span{ψi,k : 1 ≤ i ≤ N and 1 ≤ k ≤ Moff
}.
(15)
P
N
ωi
c
Using a single index notation, we may write Voff = span{ψi }N
i=1 , where Nc =
i=1 Moff denotes the total
number of basis functions in the space Voff . We also construct an operator matrix
R0T = [ψ1 , . . . , ψNc ] ,
where ψi are used to denote
the nodal values of each basis function defined on the fine grid.
P
We seek cms (x) = i ci ψi (x) ∈ Voff such that
m(
∂cms
, v) + a(cms , v) = 0
∂t
for all v ∈ Voff .
(16)
We note that variational form in (16) yields the following linear algebraic system
Qn0 cn+1
= M0n cn0 ,
0
(17)
where c0 denotes the nodal values of the discrete CG solution, and Qn0 = R0 Qn R0T and M0n = R0 M n . We
also note that the operator matrix may be analogously used in order to project coarse scale solutions onto
the fine grid cn+1 = R0T cn+1
. In our simulations presented next, we do not update basis functions. We
0
discuss basis function update in Section 6.
4.1
Numerical result
We present numerical results for the coarse-scale solution using offline basis functions. The basis functions
of the offline space are constructed following the procedure described above. Note that, the basis functions
are constructed only once at initial time and used for generating the stiffness matrix and the right hand side.
We consider the solution of problem with constant and nonlinear matrix-fracture coefficients in (6). As
constant coefficients (see previous section) representing matrix and fracture properties, we use following
am = 0.8,
bm = 1.3 · 10−7
and af = 0.001,
bf = 1.0.
For nonlinear matrix-fracture coefficients, we use
am (c) = φ + (1 − φ)
∂F
,
∂c
bm (c, x) = φD + (1 − φ)Ds
and
af (c) = φf ,
bf (c, x) =
8
κf
RT c,
µ
∂F
κ
+ φ RT c,
∂c
µ
(18)
(19)
where Dk = 10−7 [m2 /s], Di = 10−8 [m2 /s], φ = 0.04, T = 413[K], µ = 2 · 10−5 [kg/(m s)] and for fractures
kf = 10−12 [m2 ], φf = 0.001.
As for permeability κ in (18), we use constant κ = κ0 and stress-dependent model κ = κm (see [18, 32])
with
M !3
pc − αp
,
κm = κ0 1 −
p1
where κ0 = 10−18 [m2 ], p = RT c, pc = 109 [P a], p1 = 1.8 · 109 [P a], α = 0.5 and M = 0.5. For the sorbed
gas, we use Langmuir model
s
,
F (c) = cµs
(1 + sc)2
where s = 0.26 · 10−3 and cµs = 0.25 · 10−5 [mol/m3 ].
Figure 3: Coarse and fine grids. Coarse grid contains 50 cells, 85 facets and 36 vertices. Fine grid contains
7580 cells, 11470 facets and 3891 vertices.
Figure 4: Coarse and fine grids. Coarse grid contains 200 cells, 320 facets and 121 vertices. Fine grid contains
13036 cells, 19694 facets and 6659 vertices.
The equation is solved with Dirichlet boundary condition c(x, t) = 5000 on the left boundary and Neumann boundary conditions ∂c(x,t)
= 0 on other boundaries. The domain Ω has a length of 60 meters in both
∂n
9
directions. We calculate concentration for tmax = 5 years with the time step τ = 10 days. As for initial
condition, we use c(x, t = 0) = 10000[mol/m3 ]. For the numerical solution, we construct structured two
coarse grids with 36 nodes (Figure 3) and with 121 nodes (Figure 4). As for fine grids, we use unstructured
grids, which resolves the existing fractures.
Figure 5: Solution with constant matrix-fracture coefficients on coarse (top) and on fine (bottom) grids for
t=1, 3 and 5 year (from top to bottom)
In Figure 5, we show the pressure distribution for three concrete time level t = 1, 3 and 5 years. For
10
Figure 6: Coarse-scale (top) and fine-scale (bottom) solutions for t=5 year for the case of nonlinear permeability with κ = κ0
Figure 7: Coarse-scale (top) and fine-scale (bottom) solutions for t=5 year for the case of nonlinear coefficients
with κ = κm
the pressure and concentration, we have the following relationship: p = RT c. Pressure distribution for
nonlinear matrix-fracture coefficients in (6) is presented in Figures 6 - 7 for last time level. In these figures,
we show fine-scale (reference) and coarse-scale (multiscale) solutions. The coarse-scale solution is obtained
in an offline space of dimension 288 (using Mof f = 8 multiscale basis functions per coarse neighborhood) and
the fine-scale solution is obtained in a space of dimension 3891. Compared to the fine-scale solution on the
left with the coarse-scale solution on the right of the figures, we observe that the GMsFEM can approximate
the fine-scale solution accurately.
To compare the results, we use relative weighted errors
||ε||∗ = ||cms − ch ||∗ /||ch ||∗ ,
11
using L2a and Ha1 weighted norms that are defined as
Z
||ε||L2a =
1/2
,
a ε2 dx
Ω
1/2
.
(a ∇ε, ∇ε) dx
Z
||ε||Ha1 =
Ω
Figure 8: Multiscale basis functions corresponding to the first 3 smallest eigenvalues in the case with constant
fracture-matrix properties after multiplication to partition of unity functions, ψi,k = χi ψkωi ,off , i = 25 and
k = 0, 1, 2 (from left to right)
In Table 1, we present relative errors (in percentage) for last time level for constant fracture and matrix
properties in (6) using coarse grids with 36 and 121 nodes. For the coarse-scale approximation, we vary the
dimension of the spaces by selecting a certain number of offline basis functions (Mof f ) corresponding to the
smallest eigenvalues. In the Table 1, we recall that Vof f denotes the offline space, dim(Vof f ) is the offline
space dimension, Mof f is the number of the multiscale basis functions per coarse neighborhood (we use a
similar number of Mof f for each ωi ), cms and ch are the multiscale and reference solutions, respectively.
Figure 8 presents the multiscale basis functions corresponding to the first 3 smallest eigenvalues in the case
with constant fracture-matrix properties in (6). These offline basis functions are multiplied by partition of
unity functions. When we use Mof f = 8 and the case with 36 coarse nodes, the relative L2a and Ha1 weighted
errors are 0.3 % and 0.7%, respectively for final time level. The dimension of the corresponding offline space
is 288 and for reference solution is 3891. For coarse grid with 121 nodes, the relative errors are slightly
smaller 0.1% and 0.2% for L2a and Ha1 weighted errors, respectively. The dimension of the corresponding
offline space is 968 and for reference solution is 6659. The relative L2a and Ha1 errors at different time instants
for the cases with 36 and 121 coarse grids are presented in Figures 9 and 10. As we observe if we take 4 or
more basis functions per coarse node, the relative errors remain small.
Mof f
1
2
4
8
12
dim(Vof f )
36
72
144
288
432
λmin
9.0 10−9
4.5 10−8
1.1 10−7
2.2 10−6
0.19
L2a
24.484
12.229
1.068
0.303
0.083
Ha1
84.383
33.923
2.162
0.737
0.258
Mof f
1
2
4
8
12
dim(Vof f )
121
242
484
968
1452
λmin
2.5 10−8
9.6 10−8
1.6 10−7
0.37
1.23
L2a
17.136
3.975
0.651
0.110
0.060
Ha1
68.989
36.337
3.595
0.246
0.108
Table 1: Numerical results (relative errors (%) for the final time level). Left: for the case with 36 coarse
nodes. Right: for the case with 121 coarse nodes.
We present relative weighted errors in Tables 2 and 3 for different number of eigenvectors Mof f for the
case with nonlinear matrix-fracture coefficients in (6). We consider a case with 36 coarse nodes. When we
use Mof f = 8 and the case with κ = κ0 , the relative L2a and Ha1 errors are 0.2 % and 0.7%, respectively.
The dimension of the corresponding offline space is 288 and for reference solution is 3891. For the case with
κ = κm in (18), we have 0.4% and 1.0% of relative L2a and Ha1 errors, respectively. The dimension of coarse
spaces for the corresponding number of eigenvectors are 72, 144, 288, 432 and 576 for Mof f = 2, 4, 8 and 12.
12
Figure 9: Relative L2a and Ha1 weighted errors (%) for coarse grid in Figure 3 with 36 nodes. Constant
matrix-fracture properties.
Figure 10: Relative L2a and Ha1 weighted errors (%) for coarse grid in Figure 3 with 121 nodes. Constant
matrix-fracture properties.
We observe that as the dimension of the coarse space (the number of selected eigenvectors Mof f ) increases,
the respective relative errors decrease. Also we have similar error behaviour as for case with constant matrixfracture coefficients. Moreover, we see that the decrease in the relative error is fast initially and one can
obtain small errors using only a few basis functions.
Mof f
1
2
4
8
12
dim(Vof f )
36
72
144
288
432
λmin
4.8 10−9
2.4 10−8
6.0 10−8
1.1 10−6
0.19
L2a
21.717
10.772
0.933
0.270
0.123
Ha1
87.897
38.774
1.947
0.737
0.323
Mof f
1
2
4
8
12
dim(Vof f )
121
242
484
968
1452
λmin
2.5 10−8
9.6 10−8
1.6 10−7
0.37
1.23
L2a
14.333
3.673
0.646
0.110
0.063
Ha1
64.197
30.510
3.272
0.251
0.159
Table 2: Numerical results (relative weighted errors (%) for final time level) for case with κ = κ0 in (18).
Left: the case with 36 coarse nodes. Right: the case with 121 coarse nodes.
13
Mof f
1
2
4
8
12
dim(Vof f )
36
72
144
288
432
λmin
2.0 10−9
1.0 10−8
2.6 10−8
5.1 10−7
0.19
L2a
24.484
10.785
1.247
0.432
0.234
Ha1
92.039
35.874
2.423
1.069
0.712
Mof f
1
2
4
8
12
dim(Vof f )
121
242
484
968
1452
λmin
2.5 10−8
9.6 10−8
1.6 10−7
0.37
1.23
L2a
16.318
3.715
0.645
0.134
0.096
Ha1
60.267
22.872
3.000
0.386
0.282
Table 3: Numerical results (relative weighted errors (%) for the final time level) for case with κ = κm in
(18). Left: the case with 36 coarse nodes. Right: the case with 121 coarse nodes.
Remark 4.1. In our numerical simulations, we do not use empricial interpolation procedures for approximating the nonlinear functionals a· (c, ·) and b· (c, ·) (see [6] for more details). In the approaches of [6],
empirical interpolation concepts [7] are used to evaluate the nonlinear functions by dividing the computation
of the nonlinear function into coarse regions, evaluating the contributions of nonlinear functions in each
coarse region taking advantage of a reduced-order representation of the solution. By using these approaches,
we can reduce the computational cost associated with evaluating the nonlinear functions and consequently
making the computational cost to be independent of the fine grid.
5
Randomized oversampling GMsFEM
Next, we present numerical results for the oversampling and the randomized snapshots that can substantially
save the computational cost for snapshot calculations. In this algorithm, instead of solving local harmonic
problems (11) for each fine grid node on the boundary, we solve a small number of harmonic extension local
problems with random boundary conditions [5]. More precisely, we let
ψjωi ,rsnap = rj ,
x ∈ ∂ωi+ ,
where rj are independent identical distributed standard Gaussian random vectors on the fine grid nodes of
the boundary. When we use randomized snapshots, we only generate a fraction of the snapshot vectors by
using random boundary conditions.
For snapshot space calculations, we use the extended coarse grid neighborhood for m = 1, 2, . . . , by
ωi+ = ωi + m, where m is width of the fine-grid layer. Here, for example, ωi+ = ωi + 1 means the coarse
grid neighborhood plus all 1 layer of adjacent fine grid of ωi , and so on (see Figure 11 for illustration).
Calculations in the oversampled neighborhood domain ωi+ reduces the effects due to the artificial oscillation
in random boundary conditions.
5.1
Numerical results
The simulation results are presented in Tables 4 - 5 for 36 node coarse grid case. We use constant matrixfracture properties, see (6). We present the results for the randomized snapshot case for last time level. In
our simulations, we set the oversampling size m = 0, 2, 4, 6 for ωi+ = ωi + m and use different numbers of
multiscale basis functions Mof f = 2, 4, 8 and 12.
In Table 4, we investigate the effects of the oversampling ωi+ = ωi + m, as we increase the number of fine
grid extensions m = 0, 2, 4 and 6. We see that the oversampling helps to improve the results initially, but
the improvements slow and larger oversampling domains do not give significant improvement in the solution
accuracy. When we use a snapshot ratio of 25.6 % (between the standard number of snapshots and the
randomized algorithm for ωi+ = ωi + 4), the relative L2a and Ha1 weighted errors are 0.2 % and 0.8% for full
snapshots and 0.2 % and 0.9% for randomized snapshots. We observe that the randomized algorithm can
give similar errors as a full snapshots.
14
Figure 11: Neighborhood domain with oversampling (ωi+ = ωi + m, m = 1, 2, 4, 6) for the coarse grid with
36 nodes
Table 5 shows relative L2a and Ha1 errors for different number of randomized snapshots Mi . The oversampled region ωi+ = ωi + 4 is chosen, that is, the oversampled region contains an extra 4 fine-grid cell layers
around ωi . Our numerical results show that one can achieve a similar accuracy when using a fraction of
snapshots with randomized algorithms and thus, it can provide a substantial CPU savings.
6
Residual based adaptive online GMsFEM
In this section, we consider the construction of the online basis functions that are used in some regions
adaptively to reduce the error significantly. We follow earlier works [9, 8], which were done for linear timeindependent problems. The online basis functions are constructed based on a residual and take into account
distant effects. The construction of online basis functions is motivated by the analysis. Using the offline
computation, we construct multiscale basis functions that can be used for any input parameters to solve
the problem on the coarse grid. The fast convergence due to adding online basis functions depends on the
offline space. It is important that the offline space contains some essential features of the solution space. In
our numerical simulations, we demonstrate that with a sufficient number of offline basis functions, we can
achieve a rapid convergence for the proposed online procedure.
First, we derive the error indicator for the error (cn − cnms ) for time-dependent problem (21) in the energy
norm. Furthermore, we use the error indicator to develop an enrichment algorithm. The error indicator gives
an estimate of the local error on the coarse grid region ωi and we can then add basis functions to improve
the solution.
We assume, as before, V is the fine-scale finite element space. To find the fine-scale solution cn+1 ∈ V ,
we solve (as before)
cn+1 − cn
m(
, v) + a(cn+1 , v) = (f, v), ∀v ∈ V
(20)
τ
and for multiscale solution cn+1
ms ∈ Voff we have
m(
n
cn+1
ms − cms
, v) + a(cn+1
ms , v) = (f, v),
τ
∀v ∈ Voff .
We define a linear functional rn (v) for n-th time level by
n
n+1
rn (v) = τ (f, v) − m(cn+1
ms − cms , v) − τ a(cms , v).
15
(21)
2
4
8
12
full snapshots
randomized snapshots
L2a
Ha1
L2a
Ha1
without oversampling, ωi
100 %
39.7 %
12.229 33.923 8.303
33.237
1.068
2.162 1.704
4.730
0.303
0.737 1.005
2.962
0.083
0.258 0.557
1.643
2
4
8
12
with oversampling, ωi+ = ωi + 2
100 %
28.3 %
12.247 33.943 8.921
33.399
1.073
4.237 0.972
3.750
0.261
0.744 0.354
1.003
0.114
0.329 0.219
0.704
2
4
8
12
with oversampling, ωi+ = ωi + 4
100 %
25.6 %
12.216 33.657 9.334
28.213
1.015
4.576 0.626
2.561
0.262
0.841 0.264
0.949
0.114
0.349 0.153
0.441
2
4
8
12
with oversampling, ωi+ = ωi + 6
100 %
22.5 %
12.746 35.899 9.455
27.922
1.013
5.014 0.603
2.377
0.251
0.820 0.277
0.875
0.124
0.369 0.120
0.421
Mof f
Table 4: Randomized oversampling for GMsFEM with number of snapshots Mi = 24 (constant matrixfracture properties) in every ωi+ = ωi + n, n = 0, 2, 4, 6 for coarse mesh with 36 nodes (relative errors (%)
for final time level)
Mof f
2
4
8
12
12.8 % (Mi = 12)
L2
H1
8.228
24.878
1.908
4.208
0.861
1.777
-
17.0 % (Mi = 16)
L2
H1
9.449
28.895
1.381
3.581
0.589
1.563
0.313
0.781
21.3 % (Mi = 20)
L2
H1
7.346
22.774
0.779
2.692
0.292
1.189
0.217
0.581
25.6 % (Mi = 24)
L2
H1
9.334
28.213
0.626
2.561
0.264
0.949
0.153
0.441
29.8 % (Mi = 28)
L2
H1
9.335
26.973
0.843
4.439
0.245
0.894
0.110
0.393
Table 5: Randomized oversampling for GMsFEM with different number of snapshots Mi = 12, 16, 20, 24, 28
in every ωi+ = ωi + 4 (constant matrix-fracture properties) for coarse mesh with 36 nodes (relative errors
(%) for final time level)
Let ωi be a coarse region and Vi = H01 (ωi ) then
Z
Z
Z
n
n+1
n
ri (v) = τ
fv −
am (cms − cms ) v dx − τ
ωim
ωi
− dj
XZ
j
ωif,j
af (c) (cn+1
ms
ωim
−
cnms ) v dx
− τ dj
bm ∇cn+1
ms · ∇v dx
XZ
j
ωif,j
bf ∇cn+1
ms · ∇v dx,
where ωi = ωim ⊕j dj ωif,j and m and f represent the matrix and the fracture.
16
The solution at (n + 1) time level (cn+1
ms ) is the solution of the elliptic problem of the form
n+1
aτ (cms
, v) = τ (f, v) + m(cnms , v).
(22)
We use following notation
aτ (u, v) = m(u, v) + τ a(u, v).
Error estimators for the spatial discretization error take into account the dependence of the elliptic problem
(22) on the time step parameter τ and we will use the τ -weighted H1 norm
||v||2τ = τ ||v||2a + ||v||2m ,
where
||v||2τ = aτ (v, v),
||v||2a = a(v, v),
||v||2m = m(v, v).
We define the projection Π : V → Vof f by
Πv =
N
X
χi (Pi v),
i=0
where Pi : V → span{ψkwi ,of f } be the projection defined by
Pi v =
li
X
b v ψkwi ,of f ψkwi ,of f .
k=1
The projection Pi is the first li terms of spectral expansion in terms of eigenfunctions of following problem
Z
off
off
ωi
b |∇χi |2 Ψoff v dx.
(23)
aτ (Ψ , v) = λ τ
ωi
Then
Z
τ
b |∇χi |2 (v − Pi v)2 dx ≤
ωi
1 ωi
a (v − Pi v, v − Pi v).
λli+1 τ
and
ωi
i
aω
τ (v − Pi v, v − Pi v) ≤ aτ (v, v).
We note that this spectral problem is different from the original one formulated in (12); however, it involves
similar terms, such as energy norms and L2 norms.
Let en = cn − cnms is error for n-th time level and using (20) and (21), we have
m(en+1 − en , v) + τ a(en+1 , v) = (rn , v),
where the right hand side can be written as follows
n
n+1
rn (v) = τ (f, v) − m(cn+1
ms − cms , v) − τ a(cms , v)
n
n+1
≤ τ (f, v − Πv) + τ (f, Πv) − m(cn+1
ms − cms , Πv) − τ a(cms , Πv)
n
n+1
− m(cn+1
ms − cms , v − Πv) − τ a(cms , v − Πv)
n
n+1
= τ (f, v − Πv) − m(cn+1
ms − cms , v − Πv) − τ a(cms , v − Πv)
=
N
X
i=1
rin
(χi (v − Pi v)) ≤
N
X
||rin ||∗ ||χi (v − Pi v)||τ .
i=1
17
(24)
We have
||χi (v − Pi v)||2τ = ||χi (v − Pi v)||2m + τ ||χi (v − Pi v)||2a
Z
Z
= ||χi (v − Pi v)||2m + τ C
b |∇χi |2 (v − Pi v)2 dx + τ C
b χ2i |∇(v − Pi v)|2 dx
ωi
ωi
Z
i
≤ Caω
(v
−
P
v,
v
−
P
v)
+
τ
C
b |∇χi |2 (v − Pi v)2 dx
i
i
τ
ωi
C
C
i
i
aω
(v
−
P
v,
v
−
P
v)
≤
C
+
aω
≤ C+
i
i
τ
τ (v, v).
λli+1
λli+1
(25)
Therefore
m(en+1 − en , v) + τ a(en+1 , v) ≤
N
X
||rin ||∗ ||χi (v − Pi v)||τ ≤
i=1
≤
C+
N
X
1/2
C
Λmin
!1/2
||rin ||2∗
i=1
N
X
N
X
1/2
C
1/2
i
aω
||rin ||∗ C +
τ (v, v)
λ
li+1
i=1
1/2
i
aω
τ (v, v)
≤ C+
i=1
C
1/2
Λmin
N
X
!1/2
||rin ||2∗
aτ (v, v)1/2 ,
i=1
where Λmin = mini λli+1 .
Finally, we take v = en+1
||en+1 ||2τ
≤
C+
Λmin
≤
C+
N
X
1/2
C
||en+1 ||τ + m(en , en+1 )
i=1
N
X
1/2
C
!1/2
||rin ||2∗
Λmin
!1/2
||rin ||2∗
||en+1 ||τ + ||en ||m ||en+1 ||τ .
i=1
Then
||e
n+1
||τ ≤
C+
C
1/2
Λmin
N
X
!1/2
||rin ||2∗
+ ||en ||m .
i=1
This inequality residuals give a computable indicator of the error en+1 = cn+1 − cn+1
ms in the τ -weighted H1
norm.
Remark 6.1. We note that the analysis suggests the use of (23) as a local eigenvalue problem. This
eigenvalue problem is “slightly” different from (12) that we have used earlier. Our numerical simulations
show that the use of (23) improves the convergence of the offline or online procedures slightly in our numerical
examples. We will use the spectral problems based on (12) in our numerical simulations as it is independent
of time stepping.
Next, we consider online basis construction. We use the index m ≥ 1 to represent the enrichment level.
m
At the enrichment level m, we use Vms
to denote the corresponding space that can contains both offline and
m+1
m
online basis functions. We will consider a strategy for getting the space Vms
from Vms
. By the online
basis functions we mean basis functions that are computed during iterative process, contrary to offline basis
functions that are computed before iterative process. The online basis functions are computed based on
n+1,(m)
some local residuals for the current multiscale solution cms
.
m+1
m
Let Vms
= Vms
+span{ϕ} be the new approximate space that constructed by adding online basis ϕ ∈ Vi
n+1,(m+1)
m+1
on the i-th coarse neighborhood ωi and cms
∈ Vms
be the corresponding GMsFEM solution.
n,(m)
n+1
n+1
We define csemi = csemi (cms ), which satisfies
n
aτ (cn+1
semi , v) = τ (f, v) + m(cms , v).
18
n+1
For the error cn+1
ms − csemi , we have
n+1 n+1
n+1
n+1
n+1 n+1
n+1
n+1 n+1
aτ (cn+1
ms − csemi , cms − csemi ) = aτ (cms , cms − csemi ) − aτ (csemi , cms − csemi )
n+1
n+1
n+1
n+1
n+1
n
n+1
= aτ (cn+1
ms , cms − csemi ) − (τ f, cms − csemi ) − m(cms , cms − csemi )
n+1
= r(cn+1
ms − csemi ) =
N
X
n+1
n+1
n+1
ri (χi (Pi (cn+1
semi − cms ) + cms − csemi ))
i
≤
N
X
n+1
n+1
n+1
||ri ||∗ ||χi (Pi (cn+1
semi − cms ) + cms − csemi )||τ .
i
Using (25) we obtain
aτ (cn+1
ms
n+1,(m+1)
The solution cms
−
n+1
cn+1
semi , cms
−
cn+1
semi )
≤
C+
Λmin
n+1,(m)
||ri ||∗ .
(26)
i
satisfies
n+1 2
n+1,(m+1)
2
− cn+1
||cms
semi ||τ ≤ ||v − csemi ||τ ,
Taking v = cms
X
N
C
m+1
∀v ∈ Vms
.
+ αϕ, we have
2 ωi
2
n+1,(m)
2
n+1,(m)
2
ωi n+1,(m)
−cn+1
||cn+1,(m+1)
−cn+1
+αϕ−cn+1
−cn+1
ms
semi , ϕ)+α aτ (ϕ, ϕ).
semi ||τ ≤ ||cms
semi ||τ = ||cms
semi ||τ +2α aτ (cms
The last two terms in above inequality measure the amount of the reduction in error when the new basis
m
function ϕ is added to the space Vms
.
For α = −1 and ϕ ∈ Vi is the solution of
aτ (ϕ, v) = r(v),
n+1,(m+1)
Then for esemi
n+1,(m+1) 2
||τ
||esemi
n+1,(m+1)
= cms
n+1,(m) 2
||τ
n+1,(m) 2
||esemi
||τ
∀v ∈ Vi .
− cn+1
semi we have
≤ ||esemi
n+1,(m)
i
− cn+1
− 2 aω
τ (cms
semi , ϕ) + ri (ϕ)
≤
n+1,(m)
i
, ϕ) + 2τ (f, v) + 2m(cn,(m)
, ϕ) + r(ϕ) ≤ ||esemi
− 2 aω
τ (cms
ms
n+1,(m) 2
||τ
− ||r||2∗ .
To enhance the convergence and efficiency of the online adaptive GMsFEM, we consider enrichment on nonoverlapping coarse neighbothoods. Let I ⊂ {1, 2, . . . , N } be the index set of some non-overlapping coarse
m+1
m
neighborhoods. We define Vms
= Vms
+ span{ϕi , ı ∈ I} and obtain
X
n+1,(m+1) 2
n+1,(m) 2
||esemi
||τ ≤ ||esemi
||τ −
||ri ||2∗ .
i∈I
Finally, we combine this with (26) and obtain
n+1,(m)
n+1,(m+1) 2
||τ
||esemi
n+1,(m)
)
aωi (e
,e
||ri ||2∗ τ semi Psemi
N
C
2
C + Λmin
i∈I
i=1 ||ri ||∗
P
2
i∈I ||ri ||∗
||en+1,(m)
P
≤ 1 −
||2τ .
semi
N
C
2
C + Λmin
i=1 ||ri ||∗
n+1,(m) 2
||τ
≤ ||esemi
−
X
We will find online basis functions ϕ ∈ Vi to maximize the local resudial rin for current time level.
Moreover, the required ϕ is the solution of
aτ (ϕ, v) = ri (v),
19
∀v ∈ Vi ,
(27)
m
where Ri (v)is the local residual that defined using wms
n
n+1
ri (v) = τ (f, v) − m(cn+1
ms − cms , v) − τ a(cms , v)
and ||ri ||2∗ = ||ϕ||2τ according to the Riez representation theorem.
For solution in each time level, we iteratively enrich our offline space by residual based online basis
function. These basis functions are calculated using Equation (27) with zero Dirichlet boundary conditions
and the residual norm ||rin ||∗ provides a measure on the amount of reduction in energy error.
For the construction of the adaptive online basis functions, we first choose 0 < θ < 1, for each coarse
neighborhood ωi , find the online basis ϕi ∈ Vi using equation (27). After compute the norm of local residuals
and calculate ηi
ηi2 := kri k2? ,
2
where kri k? = kφi kτ , then arrange them in descending order, i.e. η12 ≥ η22 ≥ ... ≥ ηN
. Then, choose the
smallest k such that
N
k
X
X
θ
ηi2 ≤
ηi2 .
i=1
i=1
This implies that, for the coarse neighborhood ωj (j = 1, ...k), we add the corresponding online basis ϕj to
m
the original space Vms
.
6.1
Numerical results
Next, we present numerical results for residual based online basis functions. We consider a similar problem
as in the previous section with constant matrix-fracture properties in (6) and iteratively enrich the offline
space by online residual basis functions in some selected time steps. Our coarse and fine grid setups are the
same as in Section 4.1. Because re-generation of the matrix R is needed, when we add online basis function,
we add them for some selected time steps. We note that, when we add new online basis functions, which
are based on current residuals, we remove previously calculated online basis function and keep them till
we update the online basis functions. It will save computational time if we have small size of coarse scale
problem.
In Table 6, we present L2a and Ha1 errors. We consider three different cases. In the first case (we call
it Case 1), online basis functions are added at the first time step and after that in every 30-th time step.
In the second case (we call it Case 2), online basis functions are replaced at the first five consecutive time
steps, and after that, the online basis functions are updated in every 30-th time step. In the third case (we
call it Case 3), online basis functions are replaced at the first ten consecutive time steps, and after that,
the online basis functions are updated in every 30-th time step. More updates initially helps to reduce the
error due to the initial condition. As we mentioned that the offline space is important for the convergence,
and we present the results for different number of initial offline basis functions per coarse neighborhood. We
use multiscale basis functions from offline space as a initial basis functions. In Table 7, we show errors when
online basis functions are replaced at the first five consecutive time steps (as in Case 2), and afterwards,
online basis functions are updated at 10-th, 20-th and 30-th time step. For our calculations, we use tmax = 5
years with τ = 10 days. Calculations are performed in the coarse grid with 121 nodes for the case with
constant matrix-fracture properties. We observe from this table the following facts.
• Choosing 4 initial offline basis functions improves the convergence substantially. This indicates that
the choice of the initial offline space is important.
• Adding online basis functions less frequently (such as at every 30th time step) provides an accurate
approximation of the solution. This indicates that the online basis functions can be added only at
some selected time steps.
Next, we would like to show that one can use online basis functions adaptively and use the adaptivity
criteria discussed above. In Table 8, we present results for residual based online basis functions with adaptivity with θ = 0.7. In Figure 13, we show errors by time. We observe that applying adaptive algorithm can
much reduce errors.
20
Figure 12: Fine scale solution (right), coarse-scale using 2 offline basis functions (middle) and coarse-scale
after two online iteration for some time levels (left) for t=1, 3 and 5 year (from top to bottom) (constant
matrix-fracture coefficients). For fine-scale solution size of problem is 6659. For 2 offline basis functions is
242 and after two online iteration is 484
7
Conclusions
In this paper, we present a multiscale approach for shale transport in fractured media. Our approach uses an
upscaled model in the form of nonlinear parabolic equations to represent the matrix that consists of organic
and inorganic matter. The nonlinearities in the equation are due to the interaction of organic and inorganic
matter. The interaction of nonlinear matrix and the fracture is represented by multiscale basis functions. We
follow Generalized Multiscale Finite Element Method to extract the leading order terms that represent the
matrix and the fracture interaction. Multiscale basis functions are constructed locally in each coarse region
and they represent the interaction between the upscaled matrix and the fracture network. We show that
our proposed approach can effectively capture the small-scale effects and the overall system can be modeled
using a fewer degrees of freedom. Numerical results are presented. In some cases and some regions, the
offline procedure is insufficient to give accurate representations of the solution, due to the fact that offline
21
DOF
(# iter)
121
242 (1)
363 (2)
242
363 (1)
484 (2)
484
605 (1)
726 (2)
L2a
Mof f = 1
17.136
13.047
7.275
Mof f = 2
3.975
1.653
0.889
Mof f = 4
0.651
0.208
0.171
Ha1
DOF
(# iter)
68.989
43.662
12.262
121
242 (1)
363 (2)
36.337
6.705
0.972
242
363 (1)
484 (2)
3.595
0.307
0.056
484
605 (1)
726 (2)
L2a
Mof f = 1
17.136
13.209
6.603
Mof f = 2
3.975
1.716
0.546
Mof f = 4
0.651
0.165
0.105
Ha1
DOF
(# iter)
68.989
42.777
12.689
121
242 (1)
363 (2)
36.337
7.841
0.914
242
363 (1)
484 (2)
3.595
0.313
0.057
484
605 (1)
726 (2)
L2a
Mof f = 1
17.136
13.002
6.125
Mof f = 2
3.975
1.692
0.449
Mof f = 4
0.651
0.144
0.076
Ha1
68.989
42.091
13.186
36.337
8.709
0.862
3.595
0.318
0.056
Table 6: Convergence history using one, two and four offline basis functions (Mof f = 1, 2 and 4). We add
online basis functions for every 30 time step and for N −th first steps (Cases 1, 2, and 3). Left: N = 1.
Middle: N = 5. Right: N = 10. Here DOF for the last time step.
DOF
(# iter)
121
242 (1)
363 (2)
242
363 (1)
484 (2)
484
605 (1)
726 (2)
L2a
Mof f = 1
17.136
12.511
5.910
Mof f = 2
3.975
1.624
0.378
Mof f = 4
0.651
0.126
0.048
Ha1
DOF
(# iter)
68.989
41.477
13.153
121
242 (1)
363 (2)
36.337
8.225
0.934
242
363 (1)
484 (2)
3.595
0.303
0.033
484
605 (1)
726 (2)
L2a
Mof f = 1
17.136
12.925
6.275
Mof f = 2
3.975
1.669
0.474
Mof f = 4
0.651
0.147
0.080
Ha1
DOF
(# iter)
68.989
42.778
12.705
121
242 (1)
363 (2)
36.337
8.034
0.880
242
363 (1)
484 (2)
3.595
0.305
0.041
484
605 (1)
726 (2)
L2a
Mof f = 1
17.136
13.209
6.603
Mof f = 2
3.975
1.716
0.546
Mof f = 4
0.651
0.165
0.105
Ha1
68.989
42.777
12.689
36.337
7.841
0.914
3.595
0.313
0.057
Table 7: Convergence history using one, two and four offline basis functions (Mof f = 1, 2 and 4). We add
online basis functions for every N th time step and for first 5 steps Left: N = 10. Middle: N = 20. Right:
N = 30. Here DOF for last time step
computations are typically performed locally and global information is missing in these offline information.
These phenomena occur locally and in some of these regions that are identified using the proposed error
indicators, we need to develop online basis functions [8]. We discuss online basis functions and show that
this procedure converges fast.
8
Acknowledgements
We are grateful to Tat Leung for helpful discussions and suggestions regarding to online basis constructions.
MV’s work is partially supported by Russian Science Foundation Grant RS 15-11-10024.
22
DOF
(# iter)
121
242 (1)
363 (2)
242
363 (1)
484 (2)
484
605 (1)
726 (2)
P
iter
L2a
Mof f = 1
17.136
11
13.209
22
6.603
Mof f = 2
3.975
11
1.716
22
0.546
Mof f = 4
0.651
11
0.165
22
0.105
Ha1
DOF
68.989
42.777
12.689
121
243
381
36.337
7.841
0.914
242
376
504
3.595
0.313
0.057
484
635
737
P
iter
L2a
Mof f = 1
17.136
44
4.082
78
2.681
Mof f = 2
3.975
44
0.441
77
0.376
Mof f = 4
0.651
44
0.110
68
0.098
Ha1
68.989
6.418
2.871
36.337
0.720
0.325
3.595
0.044
0.039
Table 8: Convergence history using one, two and four offline basis functions (Mof f = 1, 2 and 4). We add
online basis functions for every 30 time step and for first 5 steps Left: without space adaptivity. Right: with
space adaptivity. Here DOF for last time step
References
[1] I Yucel Akkutlu, Yalchin Efendiev, and Viktoria Savatorova. Multi-scale asymptotic analysis of gas
transport in shale matrix. Transport in Porous Media, 107(1):235–260, 2015.
[2] I Yucel Akkutlu and Ebrahim Fathi. Multiscale gas transport in shales with local kerogen heterogeneities.
SPE Journal, 17(04):1–002, 2012.
[3] Raymond J Ambrose, Robert C Hartman, Mery Diaz-Campos, I Yucel Akkutlu, Carl H Sondergeld,
et al. Shale gas-in-place calculations part i: New pore-scale considerations. SPE Journal, 17(01):219–
229, 2012.
[4] R. Baca, R. Arnett, and D. Langford. Modeling fluid flow in fractured porous rock masses by finite
element techniques. Int. J. Num., 4:337348, 1984.
[5] V. Calo, Y. Efendiev, J. Galvis, and G. Li. Randomized oversampling for generalized multiscale finite
element methods. http://arxiv.org/pdf/1409.7114.pdf.
[6] Victor M Calo, Yalchin Efendiev, Juan Galvis, and Mehdi Ghommem. Multiscale empirical interpolation
for solving nonlinear pdes. Journal of Computational Physics, 278:204–220, 2014.
[7] S. Chaturantabut and D. C. Sorensen. Application of pod and deim to dimension reduction of nonlinear
miscible viscous fingering in porous media. Mathematical and Computer Modeling of Dynamical Systems,
17:337–353, 2011.
[8] Eric T Chung, Yalchin Efendiev, and Wing Tat Leung. Residual-driven online generalized multiscale
finite element methods. arXiv preprint arXiv:1501.04565, 2015.
[9] E.T. Chung, Y. Efendiev, and G. Li. An adaptive GMsFEM for high-contrast flow problems. Journal
of Computational Physics, 273:54–76, 2014.
[10] Peter Dietrich, Rainer Helmig, Martin Sauter, Heinz Hötzl, Jürgen Köngeter, and Georg Teutsch. Flow
and transport in fractured porous media. Springer Science & Business Media, 2005.
[11] L. J. Durlofsky. Numerical calculation of equivalent grid bock permeability tensors for heterogeneous
porous media. Water resources research, 27(5):699–708, 1991.
23
Figure 13: Dynamic of relative L2a (left) and Ha1 (right) weighted errors (%) for coarse grid in figure 4 with
121 nodes for the case with constant coefficients. Weighted errors using offline basis functions and online
basis functions with and without adaptivity. Top: 1 offline basis function. Middle: 2 offline basis functions.
Bottom: 4 offline basis functions.
[12] Y. Efendiev, J. Galvis, and E. Gildin. Local-global multiscale model reduction for flows in highly
heterogeneous media. Journal of Computational Physivs, 231 (24):8100–8113, 2012.
[13] Y. Efendiev, J. Galvis, and T. Hou. Generalized multiscale finite element methods. Journal of Computational Physics, 251:116–135, 2013.
[14] Y. Efendiev, J. Galvis, G. Li, and M. Presho. Generalized multiscale finite element methods. oversampling strategies. International Journal for Multiscale Computational Engineering, accepted, 2013.
[15] Y. Efendiev, J. Galvis, and X.H. Wu. Multiscale finite element methods for high-contrast problems
using local spectral basis functions. Journal of Computational Physics, 230:937–955, 2011.
[16] Y. Efendiev, T. Hou, and V. Ginting. Multiscale finite element methods for nonlinear problems and
their applications. Comm. Math. Sci., 2:553–589, 2004.
24
[17] Yalchin Efendiev, Seong Lee, Guanglian Li, Jun Yao, and Na Zhang. Hierarchical multiscale modeling for flows in fractured media using generalized multiscale finite element method. arXiv preprint
arXiv:1502.03828, 2015. to appear in International Journal on Geomathematics, (DOI) 10.1007/s13137015-0075-7.
[18] Anthony F Gangi. Variation of whole and fractured porous rock permeability with confining pressure.
In International Journal of Rock Mechanics and Mining Sciences & Geomechanics Abstracts, volume 15,
pages 249–257. Elsevier, 1978.
[19] B. Gong, M. Karimi-Fard, and L. J. Durlofsky. Upscaling discrete fracture characterizations to dualporosity, dual-permeability models for efficient simulation of flow with strong gravitational effects. SPE
J., 13(1):5867, 2008.
[20] H. Hajibeygi, D. Karvounis, and P. Jenny. A loosely coupled hierarchical fracture model for the iterative
multiscale finite volume method. Society of Petroleum Engineers. doi:10.2118/141991-MS.
[21] Hadi Hajibeygi, Dimitris Karvounis, Patrick Jenny, et al. A loosely coupled hierarchical fracture model
for the iterative multiscale finite volume method. In SPE Reservoir Simulation Symposium. Society of
Petroleum Engineers, 2011.
[22] SH Lee, C Jensen, and I Lunati. Modeling and simulation of shale gas production in multi-staged
hydraulic-fractured formations. In ECMOR XIII-13th European Conference on the Mathematics of Oil
Recovery, 2012.
[23] S.H. Lee, M.F. Lough, and C.L. Jensen. Hierarchical modeling of flow in naturally fractured formations
with multiple length scales. Water resources research, 37(3):443–455, 2001.
[24] Jianfang Li, Cong Wang, Didier Ding, Yu-Shu Wu, Yuan Di, et al. A generalized framework model for
simulation of gas production in unconventional gas reservoirs. In Paper SPE 163609 presented at SPE
Reservoir Simulation Symposium, The Woodlands, Texas, USA, pages 18–20, 2013.
[25] L. Li and S. H. Lee. Efficient field-scale simulation of black oil in naturally fractured reservoir through
discrete fracture networks and homogenized media. SPE Reservoir Evaluation & Engineering.
[26] Robert G Loucks, Robert M Reed, Stephen C Ruppel, and Daniel M Jarvie. Morphology, genesis, and
distribution of nanometer-scale pores in siliceous mudstones of the mississippian barnett shale. Journal
of Sedimentary Research, 79(12):848–861, 2009.
[27] M. F. Lough, S. H. Lee, and J. Kamath. A new method to calculate effective permeability of gridblocks
used in the simulation of naturally fractured reservoirs. SPE doi:10.2118/36730-PA.
[28] M. M. Karimi-Fard and A. Firoozabadi. Numerical simulation of water injection in 2d fractured media
using discrete-fracture model. SPE REE J., 4:117126, 2003.
[29] J. Noorishad and M. Mehran. An upstream finite element method for solution of transient transport
equation in fractured porous media. Water Resour. Res., 18(3):588–596, 1982.
[30] Volker Reichenberger, Hartmut Jakobs, Peter Bastian, and Rainer Helmig. A mixed-dimensional finite
volume method for two-phase flow in fractured porous media. Advances in Water Resources, 29(7):1020–
1036, 2006.
[31] Carl H Sondergeld, Raymond Joseph Ambrose, Chandra Shekhar Rai, Jason Moncrieff, et al. Microstructural studies of gas shales. In SPE Unconventional Gas Conference. Society of Petroleum Engineers,
2010.
[32] Asana Wasaki, I Yucel Akkutlu, et al. Permeability of organic-rich shale. In SPE Annual Technical
Conference and Exhibition. Society of Petroleum Engineers, 2014.
25
[33] Yu-Shu Wu, Yuan Di, Zhijiang Kang, and Perapon Fakcharoenphol. A multiple-continuum model
for simulating single-phase and multiphase flow in naturally fractured vuggy reservoirs. Journal of
Petroleum Science and Engineering, 78(1):13–22, 2011.
[34] Yu-Shu Wu, G. Qin, R. E. Ewing, Y. Efendiev, Z. Kang, and Y. Ren. A multiple-continuum approach
for modeling multiphase flow in naturally fractured vuggy petroleum reservoirs. SPE-104173, presented
at the 2006 SPE International Oil & Gas Conference and Exhibition in China held in Beijing, China.
[35] Jun Yao, Hai Sun, Dong-yan Fan, Chen-chen Wang, and Zhi-xue Sun. Numerical simulation of gas
transport mechanisms in tight shale gas reservoirs. Petroleum Science, 10(4):528–537, 2013.
26
| 5cs.CE
|
arXiv:1711.04177v1 [math.GT] 11 Nov 2017
COUNTING PROBLEMS IN GRAPH PRODUCTS AND RELATIVELY
HYPERBOLIC GROUPS
ILYA GEKHTMAN, SAMUEL J. TAYLOR, AND GIULIO TIOZZO
Abstract. We study properties of generic elements of groups of isometries of hyperbolic
spaces. Under general combinatorial conditions, we prove that loxodromic elements are
generic (i.e. they have full density with respect to counting in balls for the word metric)
and translation length grows linearly. We provide applications to a large class of relatively
hyperbolic groups and graph products, including right-angled Artin groups and rightangled Coxeter groups.
1. Introduction
Let G be a finitely generated group. One can learn a great deal about the geometric
and algebraic structure of G by studying its actions on various negatively curved spaces.
Indeed, Gromov’s theory of hyperbolic groups [Gro87] provides the clearest illustration
of this philosophy. However, weaker forms of negative curvature, ranging from relative
hyperbolicity [Far98, Bow12, Osi06] to acylindrical hyperbolicity [Osi15, Bow08], apply to
much larger classes of groups and still provide rather strong consequences. In all of these
theories, a special role is played by the loxodromic elements of the action, i.e. those elements
which act with sink-source dynamics. In this paper, we are interested in quantifying the
abundance of such isometries for the action of G on a hyperbolic space X. We emphasize
that in all but the simplest situations, the natural hyperbolic spaces that arise are not locally
compact. This includes actions associated to relatively hyperbolic groups [Far98], cubulated
groups [KK14, Hag14], mapping class groups [MM99], and Out(Fn ) [BF14, HM13], to name
only a few. Hence, in this paper we make no assumptions of local finiteness or discreteness
of the action.
Suppose that G y X is an action by isometries on a hyperbolic space X. We address
the question: How does a typical element of G act on X?
When G is not amenable, the word “typical” has no well defined meaning, and depends
heavily on the averaging procedure: a family of finitely supported measures exhausting
G. Although much is now known about measures generated from a random walk on G
[Mah11, CM15, MT14, MS14], very little is known about counting with respect to balls in
the word metric. This will be our main focus.
In more precise terms, fix a finite generating set S for the group G. Let Bn be the ball
of radius n about 1 with respect to the word metric d determined by S. Then we call a
property P generic if
#{g ∈ Bn : g has P }
→1
as n → ∞.
#Bn
In this language, a refinement of our questions asks when the loxodromic elements of a
particular action G y X are generic with respect to a generating set S. It is important to
Date: November 15, 2017.
1
2
I. GEKHTMAN, S.J. TAYLOR, AND G. TIOZZO
note that genericity in the counting model depends on the generating set: a priori, sets may
be generic with respect to one word metric, but not with respect to another.
The results of this paper are modeled on our previous work [GTT16], where we studied
the situation where G is itself hyperbolic. Recall that g ∈ G is loxodromic with respect
to the action G y X if and only if its translation length τX (g) = lim dX (x, g n x)/n is
strictly positive. In [GTT16], we prove that for any isometric action of a hyperbolic group
G on a hyperbolic metric space X, loxodromic elements are generic, and translation length
grows linearly. However, the genericity of loxodromic elements is in general false when
the hypothesis that G is hyperbolic is dropped (see Example 1). In the present paper, we
generalize this theorem to a much larger class of groups. Our general setup is discussed
below, but here is a sample:
Theorem 1.1. Suppose that either
(1) G is a finitely generated group which admits a geometrically finite action on a
CAT(−1) space with virtually abelian parabolic subgroups and S an admissible generating set, or
(2) G is a right-angled Artin or Coxeter group which does not split as a direct product,
and S is its standard vertex generating set.
Then for any nonelementary isometric action G y X on a separable hyperbolic metric space
X there is an L > 0 such that
#{g ∈ Bn : τX (g) ≥ Ln}
→ 1.
(1)
#Bn
In particular, loxodromic elements are generic.
In fact, our theorem applies to a more general class of relatively hyperbolic groups and
graph products (see Section 2 for precise statements and definitions) and in fact to any
group satisfying certain combinatorial conditions. Before moving to our general framework,
we state one more result which may be of independent interest. It is a direct generalization
of a theorem of Gouëzel, Mathéus, and Maucourant [GMM15] who consider the case where
G is hyperbolic.
Theorem 1.2. Let G be as in Theorem 1.1 with generating set S, and suppose that H is
an infinite index subgroup of G. Then
lim
n→∞
#(H ∩ Bn )
= 0.
#Bn
That is, the proportion of elements of G of length less than n which lie in H goes to 0 as
n → ∞.
1.1. General framework and results. Our general framework is as follows. We define
a graph structure to be a pair (G, Γ) where G is a countable group and Γ is a directed,
finite graph such that:
(1) there is a labeled vertex v0 , called the initial vertex ; for every other vertex v there
exists a directed path from v0 to v;
(2) every edge is labeled by a group element such that edges directed out of a fixed
vertex have distinct labels.
By (2), there exists an evaluation map ev : E(Γ) → G and this map extends to the set of
finite paths in Γ by concatenating edge labels. We denote by Ω0 the set of finite paths
starting at v0 , by Sn ⊂ Ω0 the set of paths of length n, and by #X the cardinality of X.
COUNTING LOXODROMICS
3
A graph structure is a geodesic combing if the evaluation map ev : Ω0 → G is bijective,
and each path in Ω0 evaluates to a geodesic in the associated Cayley graph. See Section 3.1
for details. For each set A ⊆ Ω0 we introduce the counting measure P n on Ω0 as
P n (A) :=
#(Sn ∩ A)
.
#Sn
The graph structure is almost semisimple if the number of paths of length n starting from
v0 has pure exponential growth, i.e. there exists c > 0, λ > 1 such that
c−1 λn ≤ #Sn ≤ cλn
for each n. See Section 3.2 for details.
Definition 1.3. For each vertex v of Γ, we denote by Lv the set of loops based at v, and
by Γv = ev(Lv ) its image in G. We call Γv the loop semigroup associated to v.
Consider an action G y X, where X is a hyperbolic metric space. A semigroup L < G
is nonelementary if it contains two independent loxodromics. A graph structure (G, Γ) is
nonelementary for the action G y X if for any vertex v of maximal growth (see Definition
3.1) the loop semigroup Γv is nonelementary.
We now introduce several criteria on a graph structure that guarantee it is nonelementary:
we call them thickness and quasitightness.
Definition 1.4 (Thickness). A graph structure is thick if for any vertex v of maximal
growth there exists a finite set B ⊆ G such that
G = BΓv B.
More generally, a graph structure is thick relatively to a subgroup H < G if for every vertex
v of maximal growth there exists a finite set B ⊆ G such that
H ⊆ BΓv B.
Given a path γ in Γ, we say it c–almost contains an element w ∈ G if γ contains a subpath
p such that w = a · ev(p) · b in G, with |a|, |b| ≤ c. We denote as Yw,c the set of paths in Γ
starting at the initial vertex which do not c–almost contain w. The following definition is
modeled on the one found in [AL02].
Definition 1.5 (Growth quasitightness). A graph structure (G, Γ) is called growth quasitight if there exists c > 0 such that for every w ∈ G the set Yw,c has density zero with
respect to P n ; that is,
P n (Yw,c ) → 0 as n → 0.
More generally, given a subgroup H < G we say that (G, Γ) is growth quasitight relative to
H if there exists a constant c > 0 such that for every w ∈ H the set Yw,c has density zero.
In the most general form, the main theorem we are going to prove is the following.
Theorem 1.6. Let G be a countable group of isometries of a separable, δ-hyperbolic metric
space X, and let (G, Γ) be an almost semisimple graph structure which is either:
(1) nonelementary;
(2) thick relative to a nonelementary subgroup H < G; or
(3) growth quasitight relative to a nonelementary subgroup H < G.
Then there exists L > 0 such that for every > 0 one has that:
4
I. GEKHTMAN, S.J. TAYLOR, AND G. TIOZZO
(i) Displacement grows linearly:
#{g ∈ Sn : dX (gx, x) ≥ (L − )n}
→1
#Sn
as n → ∞.
(ii) Translation length grows linearly:
#{g ∈ Sn : τX (g) ≥ (L − )n}
→1
#Sn
as n → ∞.
(iii) As a consequence, loxodromic elements are generic:
#{g ∈ Sn : g is X − loxodromic}
→1
#Sn
as n → ∞.
If we are interested in counting with respect to balls in the Cayley graph, we get the
following immediate consequence.
Corollary 1.7. Let G be a group with finite generating set S. Suppose that
(i) there is a geodesic combing for (G, S);
(ii) G has pure exponential growth with respect to S; and
(iii) the combing for (G, S) satisfies at least one of the conditions (1), (2), (3) above.
Then for any nonelementary action G y X on a hyperbolic space, the set of loxodromic
elements is generic with respect to S.
Note that the right-angled Artin group G = F2 × F3 of Example 1 with the standard
generators has a geodesic combing and has pure exponential growth but loxodromic elements
are not generic, so an additional dynamical condition (such as (1), (2), (3)) must be added.
In fact, we will show that for graph products such as RAAGs and RACGs this condition
amounts essentially to the group G not being a product. Moreover, we will prove that the
three conditions are related, namely (3) ⇒ (2) ⇒ (1).
2. Applications
2.1. Hyperbolic groups. By Cannon’s theorem [Can84], a hyperbolic group G admits a
geodesic combing for any generating set. In fact, the language recognized by the graph is
defined by choosing for each g ∈ G the smallest word (in lexicographic order) among all
words of minimal length which represent g. This is called the ShortLex representative. We
proved in [GTT16] that this graph structure is nonelementary, hence we can apply Theorem
1.6.
2.2. RAAGs, RACGs, and graph products. Let G be a right angled Artin or Coxeter
group, and let S its standard vertex generating set. A result of Hermiller and Meier [HM95]
implies that (G, S) is ShortLex automatic. In our language, (G, S) admits a geodesic combing. However, the graph Γ parameterizing this language of geodesics does not have the
correct dynamical properties needed to apply Theorem 1.6. In Section 10, we modify their
construction to show that when G is not a direct product, it has a graph structure with
respect to the standard generators with the strongest possible dynamical properties. We
then obtain the following:
Theorem 2.1. Let G be a right-angled Artin or Coxeter group which is not virtually cyclic
and does not split as a product, and consider an action of G on a hyperbolic, separable
metric space X. Then eq. (1) holds, and loxodromic elements are generic with respect to
the standard generators.
COUNTING LOXODROMICS
5
Actually, our theorem applies to all graph products of groups with geodesic combing
(Theorem 10.5). We refer the reader to Section 10 for details. Let us point out that RAAGs
which are products give in fact examples of actions where loxodromics are not generic:
Example 1 (Nongenericity in general). Denote the free group of rank n by Fn and fix a free
basis as a generating set. Let G = F2 × F3 and let X denote a Cayley graph for F2 . Give
G its standard generating set; that is, the generating set consisting of a basis for F2 and a
basis for F3 . Consider the action G y X in which the F2 factor acts by left multiplication
and the right factor acts trivially. If we denote the set of loxodromics for the action by
LOX, then
2
# (LOX ∩ Bn )
= 6= 1.
lim
n→∞
#Bn
3
Note that in the example above G has pure exponential growth and a geodesic combing,
so these two conditions are not sufficient to yield genericity of loxodromics. Moreover, the
complement of LOX is a subgroup H < G which has infinite index and positive density,
showing that conditions are needed also in Theorem 1.2.
Moreover, as a consequence of the geodesic combing that we produce in order to prove
the previous theorem, we also prove the following fine counting statement for the number
of elements in a ball with respect to the standard generating set. As far as we know, this
result is also new, and it may be of independent interest.
Theorem 2.2. Let G be a right-angled Artin group or Coxeter group which is not virtually
cyclic and does not split as a product. Then there exists λ > 1, C > 0 such that
lim
n→∞
#Sn
= C.
λn
We say that a group with a generating set with the previous property has exact exponential
growth. This is stronger than pure exponential growth (where one only requires C −1 λn ≤
#Sn ≤ Cλn ), and depends very subtly on the choice of generating set. In fact, in Theorem
11.1 we will establish this result also more generally for graph products.
2.3. Relatively hyperbolic groups. Our results also apply to a large class of relatively
hyperbolic groups. We need two hypotheses.
First, recall that a relatively hyperbolic group G is equipped with a compact metric space
∂G known as its Bowditch boundary, and such a space carries a natural Patterson-Sullivan
measure ν, defined with respect to the word metric on Cay(G, S) (see Section 9.2). We
call a relatively hyperbolic group G with a generating set S pleasant if the action of G on
∂G × ∂G is ergodic with respect to the measure ν × ν.
Second, we need a geodesic combing with respect to some generating set S. Let us call
a finite generating set S admissible if G admits a geodesic combing with respect to S. We
have the following general statement:
Theorem 2.3. Let G be a relatively hyperbolic group with an admissible generating set S
for which G is pleasant. Then, for each action of G on a hyperbolic, separable, metric space
X, there exists L > 0 such that
#{g ∈ Bn : τX (g) ≥ Ln}
→1
#Bn
As a consequence, X-loxodromic elements are generic.
as n → ∞.
6
I. GEKHTMAN, S.J. TAYLOR, AND G. TIOZZO
In fact, by [AC16] and [NS95], many relatively hyperbolic groups admit geodesic combings
as follows. Let us call a finitely generated group G geodesically completable if any finite
generating set S of G can be extended to a finite generating set S 0 ⊇ S for which there
exists a geodesic biautomatic structure. Antolı́n and Ciobanu ([AC16], Theorem 1.5) proved
that whenever G is hyperbolic relative to a collection of subgroups (Pi ) each of which is
geodesically completable, then G is geodesically completable. Moreover, from automata
theory ([HRR17], Theorem 5.2.7) one gets that if G admits a geodesic biautomatic structure
for S, then it also admits a geodesic combing for the same S. Hence one gets:
Proposition 2.4. Let (G, P) be a relatively hyperbolic group such that each parabolic subgroup P ∈ P is geodesically completable. Then every finite generating set S can be extended
to a finite generating set S 0 which admits a geodesic combing.
Let us note that in particular, virtually abelian groups are geodesically completable
([AC16], Proposition 10.1), hence any group hyperbolic relative to a collection of virtually
abelian subgroups is geodesically completable and admits a geodesic combing. Moreover,
we will prove (Proposition 9.17):
Proposition 2.5. If a group G acts geometrically finitely on a CAT(−1) proper metric
space, then G is pleasant with respect to any finite generating set.
In particular, geometrically finite Kleinian groups satisfy both hypotheses of Theorem
2.3, which establishes Theorem 1.1 (1) as a corollary of Theorem 2.3.
2.4. Actions with strongly contracting elements. Let us now remark that by combining our work with recent work of W. Yang one can apply our theorem in even more
general cases. Following [ACT15], [Yan16], we call an element g ∈ G strongly contracting
for the action on Cay(G, S) in the above sense if n 7→ g n is a quasigeodesic and there exists
C, D ≥ 0 such that for any geodesics γ in Cay(G, S) whose distance from hgi is at least C,
the diameter of the image of γ under the nearest point projection to hgi is bounded by D.
Wenyuan Yang [Yan16] has recently announced that whenever the action G y Cay(G, S)
has a strongly contracting element, G is growth quasitight and has pure exponential growth
with respect to S. Combining Theorem 1.7 with Yang’s result we obtain the following:
Corollary 2.6. Let G be a group with finite generating set S. Suppose that the Cayley
graph Cay(G, S) has a strongly contracting element and that (G, S) has a geodesic combing.
Then for any nonelementary action G y X on a hyperbolic space, the set of loxodromic
elements is generic with respect to S.
2.5. Genericity with respect to the Markov chain. Our approach is to deduce typical
properties of elements of G from typical long term behavior of long paths in the associated
graph structure. As a by-product, we also obtain a general theorem about generic elements
for sample paths in a Markov chain, which may be of independent interest. More precisely,
an almost semisimple graph Γ defines a Markov chain on the vertices of Γ (see 3.3), hence
it defines a Markov measure P on the set Ω0 of infinite paths from the initial vertex. For
such Markov chains, we prove the following:
Theorem 2.7. Let (G, Γ) be an almost semisimple, nonelementary graph structure for
G y X, and let x ∈ X. Then:
(1) For P-almost every sample path (wn ), the sequence (wn x) converges to a point in
∂X;
COUNTING LOXODROMICS
7
(2) There exist finitely many constants Li > 0 (i = 1, . . . , r) such that for P-almost
every sample path there exists an index i such that
d(wn x, x)
= Li ;
n
(3) If we denote L := min1≤i≤r Li , then for each > 0 one has
lim
n→∞
P(τX (wn ) ≥ n(L − )) → 1
as n → ∞. As a consequence,
P(wn is loxodromic) → 1
as n → ∞.
2.6. Non-backtracking random walks. An illustration of the previous result is given by
looking at non-backtracking random walks. Let G be a group and S = S −1 a generating
set. The non-backtracking random walk on G is the process defined by taking gn uniformly
at random among the elements of S \ {(gn−1 )−1 } and considering the sample path wn =
g1 g2 . . . gn . We prove the following, which answers a question of I. Kapovich.
Theorem 2.8. Let G be a nonelementary group of isometries of a separable hyperbolic
metric space X, and let S be a finite generating set. Consider the non-backtracking random
walk
wn := g1 . . . gn
defined as above, and let P be corresponding the measure on the set Ω0 of sample paths.
Then
P (wn is loxodromic on X) → 1
as n → ∞.
Proof. Let us consider F = F (S) the free group generated by S, with its standard word
metric. By composing the projection F → G with the action on X, we can think of F as a
group of isometries of X. Then F has a standard geodesic combing, whose graph Γ has only
one non-trivial component, hence (by Proposition 6.3) the graph structure (F, Γ) is thick,
hence nonelementary. The result then follows from Theorem 2.7.
2.7. Previous results. Beginning with Gromov’s influential works [Gro87, Gro93, Gro03],
there is a large literature devoted to studying typical behavior in finitely generated groups.
More recent developments can be found, for example, in [AO96, Arz98, BMR03, Cha95,
KMSS03, KS05, KRSS07, Ol’92].
If one takes the definition of genericity with respect to random walks, instead of using
counting in balls, then genericity of loxodromics has been established in many cases. In particular, the question of genericity of pseudo-Anosovs in the mapping class group goes back
to at least Dunfield-Thurston [DT06], and for random walks it has been proven independently by Rivin [Riv08] and Maher [Mah11]. This relates to our setup, as a mapping class
is pseudo-Anosov iff it acts loxodromically on the curve complex. Genericity of loxodromics
for random walks on groups of isometries of hyperbolic spaces has been established with
increasing level of generality in [CM15, Sis11, MT14]. Let us note that in general counting
in balls and counting with random walks need not yield the same result, and in fact it is
a very important problem to establish whether the harmonic measure for the random walk
can coincide with a Patterson-Sullivan-type measure, given by taking limits of counting
measures over balls. Many results in this area show that the two measures do not coincide
except in particular cases, while an existence result of a random walk for which harmonic
and PS measure coincide is due for hyperbolic groups to Connell-Muchnik [CM07].
8
I. GEKHTMAN, S.J. TAYLOR, AND G. TIOZZO
As for counting in balls, Wiest [Wie14] recently showed that if a group G satisfies a weak
automaticity condition and the action G y X on a hyperbolic space X satisfies a strong
geodesic word hypothesis, then the loxodromics make up a definite proportion of elements
of the n ball. This geodesic word hypothesis essentially requires geodesics in the group G,
given by the normal forms, to project to unparameterized quasigeodesics in the space X
under the orbit map. In our work, on the other hand, we do not assume any nice property
of the action except it being by isometries. Let us note that our theorems answer (when the
hypotheses of our two papers overlap) the open problems (3) (4) in ([Wie14], section 2.12)
Acknowledgments. We thank Yago Antolı́n for useful suggestions and clarifications. The
first author is partially supported by NSF grant DMS-1401875 and ERC advanced grant
“Moduli” of Prof. Ursula Hamenstädt, the second author is partially supported by NSF
grants DMS-1400498 and DMS-1744551, and the third author is partially supported by
NSERC and the Connaught fund.
3. Background material
Since graph structures play central role in our work, we begin by discussing some further
details. The reader will notices that much of this is inspired by the theory of regular
languages and automatics groups [EPC+ 92], but we place a special focus on the graph which
parameterizes the language. Thus our terminology may differ from that in the literature.
3.1. Graph structures. The general framework is as follows. We define a graph structure to be a pair (G, Γ) where G is a countable group and Γ is a directed, finite graph such
that:
(1) there is a labeled vertex v0 , called the initial vertex ; for every other vertex v there
exists a directed path from v0 to v;
(2) every edge is labeled by a group element such that edges directed out of a fixed
vertex have distinct label.
Thus, there exists an evaluation map ev : E(Γ) → G and this map extends to the set of
finite paths in Γ by concatenating edge labels. If S = ev(E(Γ)), we say that (G, Γ) is a graph
structure with respect to S. We denote by Ω0 the set of finite paths starting at v0 and by Ω
the set of all finite paths. When ev(Ω0 ) = G, we call the graph structure surjective; in this
case S = ev(E(Γ)) generates G as a semigroup. A surjective graph structure is geodesic
if for each path p ∈ Ω, the word length |ev(p)|S is equal to the length of the path. In this
case, all paths in Γ evaluate naturally to geodesic paths in the Cayley graph Cay(G, S).
Finally, the graph structure is called injective, if ev : Ω0 → G is injective. For example, if
each path in Ω0 labels the ShortLex geodesic representative of its evaluation (with respect
to some ordering on S), then (G, Γ) is injective. A bijective, geodesic graph structure (G, Γ)
with respect to S is called a geodesic combing of G with respect to S.
Note the evaluation map, restricted to Ω0 , factors through S ∗ , the set of all words in
the alphabet S. The image in S ∗ of Ω0 (i.e. all words which can be spelled starting at v0 )
is called the language parameterized (or recognized) by Γ. This language is prefix closed
by construction; an initial subword of a recognized word is also recognized. We warn the
reader that references differ on the exact meaning on some of these terms. For example,
Calegari–Fujiwara use the term “combing” to refer to the language of a bijective, geodesic
graph structure rather than the graph structure itself [CF10, Cal13]. Since we will be most
interested in dynamical properties of the graph parameterizing the language of geodesics,
we choose to emphasize the graph structure itself.
COUNTING LOXODROMICS
9
3.2. Almost semisimple graphs. Let us summarize some of the fundamental properties
about graphs and Markov chains. Much of this material appears in Calegari–Fujiwara
[CF10], and we refer to that article and [GTT16] for more details and proofs.
Let Γ be a finite, directed graph with vertex set V (Γ) = {v0 , v1 , . . . , vr−1 }. The adjacency
matrix of Γ is the r × r matrix M = (Mij ) defined so that Mij is the number of edges from
vi to vj .
Such a graph is almost semisimple of growth λ > 1 if the following hold:
(1) There is an initial vertex, which we denote as v0 ;
(2) For any other vertex v, there is a (directed) path from v0 to v;
(3) The largest modulus of the eigenvalues of M is λ, and for any eigenvalue of modulus
λ, its geometric multiplicity and algebraic multiplicity coincide.
We denote by Ω the set of all finite paths in Γ, Ωv for the set of finite paths starting at v,
and Ω0 = Ωv0 the set of finite paths starting at v0 . For a path g ∈ Ω, we use [g] to denote
its terminal vertex. Similarly, we denote as Ω∞ the set of all infinite paths in Γ, Ω∞
v the set
∞
.
of infinite paths starting at v and Ω∞
=
Ω
v0
0
Given two vertices v1 , v2 of a directed graph, we say that v2 is accessible from v1 and
write v1 → v2 if there is a path from v1 to v2 , and two vertices are mutually accessible
if v1 → v2 and v2 → v1 . Mutual accessibility is an equivalence relation, and equivalence
classes are called irreducible components of Γ.
For any subset A ⊆ Ω0 , we define the growth λ(A) of A as
p
λ(A) := lim sup n #(A ∩ Sn ),
n→∞
where Sn ⊂ Ω0 is the set of all paths starting at v0 that have length n.
For each vertex v of Γ which lies in a component C, let Pv (C) denote the set of finite
paths in Γ based at v which lie entirely in C. Moreover, for any path g from v0 to v, we let
Pg (C) = g · Pv (C) be the set of finite paths in Ω which can be written as a concatenation
of g with a path contained entirely in C.
Definition 3.1. An irreducible component C of Γ is called maximal if for some (equivalently,
any) g ∈ Ω0 with [g] ∈ C, the growth of Pg (C) equals λ. A vertex is maximal if it belong
to a vertex of maximal growth. Moreover, we say a vertex vi of Γ has large growth if there
exists a path from vi to a vertex in a maximal component, and it has small growth otherwise.
Definition 3.2. For every vertex v of Γ, the loop semigroup of v is the set Lv of loops in
the graph Γ which begin and end at v. It is a semigroup with respect to concatenation. A
loop in Lv is primitive if it is not the concatenation of two (non-trivial) loops in Lv .
Let Γ be an almost semisimple graph of growth λ > 1. Then there exist constants c > 0
and λ1 < λ such that ([GTT16], Lemma 2.3):
(1) For any vertex v of large growth and any n ≥ 0,
c−1 λn ≤ #{paths from v of length n} ≤ cλn
(2) For any vertex v of small growth and any n ≥ 0,
#{paths from v of length n} ≤ cλn1
(3) If v belongs to the maximal component C, then for any n ≥ 0
c−1 λn ≤ #{paths in Pv (C) of length n} ≤ cλn
10
I. GEKHTMAN, S.J. TAYLOR, AND G. TIOZZO
and also ([GTT16], Lemma 6.5)
c−1 λn ≤ #{paths in Lv of length n} ≤ cλn
3.3. Markov chains. Given an almost semisimple graph Γ of growth λ with edge set E(Γ),
one constructs a Markov chain on the vertices of Γ as follows. If vi has large growth, then
we set the probability µ(vi → vj ) of going from vi to vj as
(2)
µ(vi → vj ) =
Mij ρ(1)j
,
λρ(1)i
and if vi has small growth, we set µ(vi → vj ) = 0 for i 6= j and µ(vi → vi ) = 1.
Now, for each vertex v the measure µ induces measures Pnv on the space Ωv of finite paths
starting at v, simply by setting
Pnv (γ) = µ(e1 ) · · · µ(en )
for each path γ = e1 . . . en starting at v. Similarly, this measure can be extended to a
measure Pv on the space Ω∞
v of infinite paths starting at v. The most important cases for
us will be the measures on the set of (finite and infinite, respectively) paths starting at v0 ,
which we will denote as Pn = Pnv0 and P = Pv0 . Each measure Pv defines a Markov chain
on the space V (Γ), and we consider for each n the random variable
wn : Ω∞ → Ω
wn ((e1 , . . . , en , . . . )) = e1 . . . en
defined as the concatenation of the first n edges of the infinite path.
In order to compare the n-step distribution for the Markov chain to the counting measure,
let us denote as LG the set of paths from v0 ending at a vertex of large growth. Then we
note ([GTT16], Lemma 3.4) that there exists c > 1 such that, for each A ⊆ Ω0 ,
(3)
c−1 Pn (A) ≤ P n (A ∩ LG) ≤ c Pn (A).
It turns out (see [GTT16], Lemma 3.3) that, with respect to this choice of measure, a
vertex v belongs to a maximal irreducible component of Γ if and only if it is recurrent, i.e. :
(1) there is a path from v0 to v of positive probability; and
(2) whenever there is a path from v to another vertex w of positive probability, there
is also a path from w to v of positive probability.
For this reason, maximal components will also be called recurrent components.
It is well-known that for almost every path of the Markov chain there exists one recurrent
component C such that the path lies completely in C from some time on, and visits each
vertex of C infinitely many times. Thus, for each recurrent component C, we let ΩC be the
set of all infinite paths from the initial vertex which enter C and remain inside C forever,
and denote as PC the conditional probability of P on ΩC .
Moreover, for each recurrent vertex v the distribution of return times decays exponentially:
(4)
Pv τv+ = n ≤ e−cn
where τv+ = min{n ≥ 1 : [wn ] = v} denotes the first return time to vertex v.
We will associate to each recurrent vertex of the Markov chain a random walk, and use
previous results on random walks to prove statements about the asymptotic behavior of the
Markov chain.
COUNTING LOXODROMICS
11
For each sample path ω ∈ Ω∞ , let us define n(k, v, ω) as the k th time the path ω lies at
the vertex v. In formulas,
0
if k = 0
n(k, v, ω) :=
min (h > n(k − 1, v, ω) : [wh ] = v) if k ≥ 1
To simplify notation, we will write n(k, v) instead of n(k, v, ω) when the sample path ω is
fixed.
We now define the first return measure µv on the set of primitive loops by setting, for
each primitive loop γ = e1 . . . en with edges e1 , . . . , en ,
µv (e1 . . . en ) = µ(e1 ) . . . µ(en ).
Extend µv to the entire loop semigroup Lv by setting µv (γ) = 0 if γ ∈ Lv is not primitive.
Since almost every path starting at v visits v infinitely many times, the measure µv is a
probability measure.
By equation (4), for every recurrent vertex v, the first return measure µv has finite
exponential moment, i.e. there exists a constant α > 0 such that
Z
(5)
eα|γ| dµv (γ) < ∞.
Lv
3.4. Hyperbolic spaces. In this paper, X will always be a geodesic, separable metric
space. Such a space is called δ-hyperbolic for some δ ≥ 0 if for every geodesic triangle in X,
each side is contained within the δ–neighborhood of the other two sides. Each hyperbolic
space has a well-defined Gromov boundary ∂X, and we refer the reader to [BH09, Section
III.H.3], [GdlH90], or [KB02, Section 2] for definitions and properties.
If g is an isometry of X, its translation length is defined as
d(g n x, x)
n→∞
n
where the limit does not depend on the choice of x. In order to estimate the translation
length, we will use the following well-known lemma; see for example [MT14, Proposition
5.8].
τX (g) := lim
Lemma 3.3. There exists a constant c, which depends only on δ, such that for any isometry
g of a δ-hyperbolic space X and any x ∈ X with d(x, gx) ≥ 2(gx, g −1 x)x + c, the translation
length of g is given by
τX (g) = d(x, gx) − 2(gx, g −1 x)x + O(δ).
An isometry g of X is loxodromic if it has positive translation length; in that case, it has
two fixed points on ∂X. We say two loxodromic elements are independent if their fixed point
sets are disjoint. A semigroup (or a group) G < Isom X is nonelementary if it contains two
independent loxodromics. We will use the following criterion.
Proposition 3.4 ([DSU14, Proposition 7.3.1]). Let L be a semigroup of isometries of a
hyperbolic metric space X. If the limit set ΛL ⊂ ∂X of L on the boundary of X is nonempty
and L does not have a finite orbit in ∂X, then L is nonelementary.
Finally, we turn to the definition and basic properties of shadows in the δ-hyperbolic
space X. For x, y ∈ X, the shadow in X around y based at x is
Sx (y, R) = {z ∈ X : (y, z)x ≥ d(x, z) − R},
12
I. GEKHTMAN, S.J. TAYLOR, AND G. TIOZZO
where R > 0 and (y, z)x = 12 (d(x, y) + d(x, z) − d(y, z)) is the usual Gromov product. The
distance parameter of Sx (y, R) is by definition the number r = d(x, y) − R, which up to an
additive constant depending only on δ, measures the distance from x to Sx (y, R). Indeed,
z ∈ Sx (y, R) if and only if any geodesic [x, z] 2δ–fellow travels any geodesic [x, y] for distance
r + O(δ). The following observation is well-known.
Lemma 3.5. For each D ≥ 0, and each x, y in a metric space, we have
ND (Sx (y, R)) ⊆ Sx (y, R + D).
3.5. Random walks. A probability measure µ on G is said to be nonelementary with respect to the action G y X if the semigroup generated by the support of µ is nonelementary.
We will need the fact that a random walk on G whose increments are distributed according
to a nonelementary measure µ almost surely converge to the boundary of X and has positive
drift in X.
Theorem 3.6 ([MT14, Theorems 1.1, 1.2]). Let G be a countable group which acts by
isometries on a separable hyperbolic space X, and let µ be a nonelementary probability
distribution on G. Fix x ∈ X, and let (un ) be the sample path of a random walk with
independent increments with distribution µ. Then:
(1) almost every sample path (un x) converges to a point in the boundary of ∂X, and
the resulting hitting measure ν is nonatomic;
(2) moreover, if µ has finite first moment, then there is a constant L > 0 such that for
almost every sample path
lim
n→∞
d(x, un x)
= L > 0.
n
The constant L > 0 in Theorem 3.6 is called the drift of the random walk (un ).
4. Behavior of generic sample paths for the Markov chain
Let G be a group with a nonelementary action G y X on a hyperbolic space X. In
this section we assume that G has a graph structure (G, Γ) which is almost semisimple and
nonelementary.
4.1. Convergence to the boundary of X. In this section we show that almost every
sample path for the Markov chain converges to the boundary of X. Since we are assuming
that the graph structure is nonelementary, the exact same proof as in ([GTT16], Theorem
6.8) yields the following.
Theorem 4.1. For P-almost every path (wn ) in the Markov chain, the projection (wn x) to
the space converges to a point in the boundary ∂X.
As a consequence, we have for every vertex v of large growth a well-defined harmonic
measure νvX , namely the hitting measure for the Markov chain on ∂X: for each each (Borel)
A ⊆ ∂X we define
νvX (A) := Pv ( lim wn x ∈ A).
n→∞
The previous proof also provides a decomposition theorem for the harmonic measures
νvX . If R is the set of recurrent vertices of Γ, then we have:
X X
X
(6)
νvX =
µ(γ)γ∗ νw
w∈R γ:v→w
COUNTING LOXODROMICS
13
Here, the sum is over all finite paths from v to w which only meet a recurrent vertex at
their terminal endpoint. Note that if v is recurrent, then νvX is the harmonic measure for
the random walk on G generated by the measure µv , as discussed above.
Lemma 4.2. For any v of large growth, the measure νvX is non-atomic.
X
X
Proof. Since the random walk measures νw
are non-atomic, so are the measures γ∗ νw
for
X
each γ, hence by equation (6) the measure νv is also non-atomic as it is a linear combination
of non-atomic measures.
4.2. Positive drift along geodesics. In this section, we show that almost every sample
path has a well-defined and positive drift in X.
Theorem 4.3. For P-almost every sample path ω = (wn ) there exists a recurrent component
C = C(ω) for which we have
d(wn x, x)
= LC ,
lim
n→∞
n
where LC > 0 depends only on C.
Since Γ is finite, this gives at most finitely many potential drifts for the Markov chain.
Proof. Let v be a recurrent vertex. Since the graph structure is nonelementary, the loop
semigroup Γv is nonelementary, hence the random walk given by the return times to v has
positive drift. More precisely, from Theorem 3.6, there exists a constant `v > 0 such that
for almost every sample path which enters v,
d(wn(k,v) x, x)
lim
= `v .
k→∞
k
Morever, as the distribution of return times has finite exponential moment, for almost every
sample path one has
n(k, v, ω)
lim
= Tv .
k→∞
k
These two facts imply
d(wn(k,v) x, x)
`v
lim
=
.
k→∞
n(k, v)
Tv
Now, almost every infinite path visits every vertex of some recurrent component infinitely
often. Thus, for each recurrent vertex vi which belongs to a component C, there exists a
constant Li > 0 such that for PC -almost every path (wn ), there is a limit
d(wn(k,vi ) x, x)
.
n(k, vi )
Let C be a maximal component, and v1 , . . . , vk its vertices. Our goal now is to prove that
L1 = L2 = . . . Lk . Let us pick a path ω ∈ Ω0 such that the limit Li above exists for each
i = 1, . . . , k, and define Ai = {n(k, vi ), k ∈ N}, and the equivalence relation i ∼ j if Li = Lj .
Since wn(k,vi ) and wn(k,vi )+1 differ by one generator, d(wn(k,vi ) x, wn(k,vi )+1 x) is uniformly
bounded, hence
d(wn(k,vi )+1 x, x)
d(wn(k,vi ) x, x)
lim
= lim
= Li
k→∞
k→∞
n(k, vi ) + 1
n(k, vi )
so the equivalence relation satisfies the hypothesis of ([GTT16], Lemma 6.9), hence there is
a unique limit LC = Li so that
d(wn x, x)
lim
= LC .
n→∞
n
Li = lim
k→∞
14
I. GEKHTMAN, S.J. TAYLOR, AND G. TIOZZO
Corollary 4.4. For every vertex v of large growth, and for Pv -almost every sample path
(wn ) there exists a recurrent component C accessible from v such that we have
d(wn x, x)
n
Proof. By Theorem 4.3, for P-almost every path
LC for some recurrent component C. Let g0 be a
Then for any ω = (wn ) ∈ Ω∞
en )
v , the path (w
w
en+k = wn+k (g0 · ω) = g0 · wn (ω) where k = |g0 |.
lim
n→∞
= LC .
which passes through v, the drift equals
path from v0 to v of positive probability.
= g0 · ω belongs to Ω∞
0 , and moreover
Hence
d(wn x, x) − d(x, g0 x) ≤ d(w
en+k x, x) = d(wn x, g0 x) ≤ d(wn x, x) + d(x, g0 x)
and so by Theorem 4.3
lim
n→∞
d(w
en+k x, x)
d(wn x, x)
= lim
= LC
n→∞
n
n
as required.
For application in Section 5, we will need the following convergence in measure statement.
Let us denote
(7)
L :=
min
C recurrent
LC > 0
the smallest drift.
Corollary 4.5. For any > 0, and for any v of large growth,
d(x, wn x)
≤ L − → 0.
Pv
n
Proof. By the theorem, the sequence of random variables Xn = d(wnnx,x) converges almost
surely to a function X∞ with the finitely many values L1 , . . . , Lr . Moreover, for every n the
variable Xn is bounded above by the Lipschitz constant of the orbit map G → X. Thus,
Xn converges to X∞ in L1 , yielding the claim.
4.3. Decay of shadows for P. For any shadow S, we denote its closure in X ∪ ∂X by S.
Since the harmonic measures νvX for the Markov chain are nonatomic (by Lemma 4.2), we
get by the same proof as in [GTT16] the following decay of shadows results.
Proposition 4.6 (Decay of shadows, [GTT16], Proposition 6.18). There exists a function
f : R → R such that f (r) → 0 as r → ∞, and such that for each v of large growth,
νvX Sx (gx, R) ≤ f (r),
where r = d(x, gx) − R is the distance parameter of the shadow.
Proposition 4.7 ([GTT16], Proposition 6.19). There exists a function p : R+ → R+ with
p(r) → 0 as r → ∞, such that for each vertex v and any shadow Sx (gx, R) we have
Pv ∃n ≥ 0 : wn x ∈ Sx (gx, R) ≤ p(r),
where r = d(x, gx) − R is the distance parameter of the shadow.
5. Generic elements with respect to the counting measure
We now use the results about generic paths in the Markov chain to obtain results about
generic paths with respect to the counting measure.
COUNTING LOXODROMICS
15
5.1. Genericity of positive drift. The first result is that the drift is positive along generic
paths:
Theorem 5.1. Let (G, Γ) be an almost semisimple, nonelementary graph structure, and
L > 0 be the smallest drift as given by eq. (7). Then for every > 0 one has
#{g ∈ Sn : d(gx, x) ≥ (L − ) |g|}
→1
#Sn
as n → ∞.
The result follows from Corollary 4.5 similarly as in the proof of ([GTT16], Theorem 5.1).
Proof. Let AL denote the set of paths
AL := {g ∈ Ω0 : d(gx, x) ≤ L|g|}.
We know by Corollary 4.5 that for any L0 < L one has
Pn (AL0 ) → 0
as n → ∞.
Then, for each path g of length n we denote as gb the prefix of g of length n − blog nc, and
we observe that
#{g ∈ Sn : gb ∈
/ LG} #{g ∈ Sn ∩ AL− : gb ∈ LG}
P n (AL− ) ≤
+
#Sn
#Sn
and we know by [GTT16, Proposition 2.5] that the first term tends to 0. Now, by writing
g = gbh with |h| = log |g| we have that d(gx, x) ≤ (L − )|g| implies
d(b
g x, x) ≤ d(gx, x) + d(b
g x, gx) ≤ (L − )|g| + d(x, hx) ≤
hence, there exists C such that it is less than
≤ (L − )|g| + C log |g| ≤ L0 |b
g|
for any L − < L0 < L whenever |g| is sufficiently large. This proves the inclusion
{g ∈ Sn ∩ AL− : gb ∈ LG} ⊆ {g ∈ Sn : gb ∈ AL0 ∩ LG}
and by Lemma 3.2 (1)
#{g ∈ Sn : gb ∈ AL0 ∩ LG} ≤ cλlog n #(Sn−log n ∩ AL0 ∩ LG) ≤
hence by equation (3) and considering the size of Sn−log n
≤ c1 λlog n Pn−log n (AL0 )#Sn−log n ≤ c2 λn Pn−log n (AL0 ).
Finally, using that Pn−log n (AL0 ) → 0 we get
#{g ∈ Sn ∩ AL− : gb ∈ LG}
≤ lim sup c3 Pn−log n (AL0 ) = 0
#S
n→∞
n→∞
n
which proves the claim.
lim sup
5.2. Decay of shadows for the counting measure. For g ∈ G, we set
SxΓ (gx, R) = {h ∈ Ω0 : hx ∈ Sx (gx, R)} ,
where as usual, Sx (gx, R) is the shadow in X around gx centered at the basepoint x ∈ X
and hx = ev(h)x. We will need the following decay property for SxΓ (gx, R) ⊆ Ω0 .
Proposition 5.2. There is a function ρ : R+ → R+ with ρ(r) → 0 as r → ∞ such that for
every n ≥ 0
P n (SxΓ (gx, R)) ≤ ρ (d(x, gx) − R) .
We start with the following lemma in basic calculus.
16
I. GEKHTMAN, S.J. TAYLOR, AND G. TIOZZO
Lemma 5.3. Let p : R+ → [0, 1] be a decreasing function with p(x) → 0 as x → ∞. For
each α > 0 and each C > 0, there exists a function pe : R+ → R+ such that:
(1) pe(x) ≥ p(x) for each x ∈ R+ ;
(2) pe(x + mC) ≥ pe(x)e−αm for each x ∈ R+ , m ∈ N:
(3) pe(x) → 0 as x → ∞.
Proof. Let us define pe inductively as follows:
pe(x) := p(x)
if x ∈ [0, C)
pe(x + C) := max{p(x + C), pe(x)e−α }
Thus (1) and (2) are immediate from the definition. To prove (3), note that for each n there
exists k ≤ n such that pe(x + nC) = p(x + (n − k)C)e−αk . Hence as x + nC → ∞ either
x + (n − k)C → ∞ or k → ∞, thus the only limit point is 0.
Proof of Proposition 5.2. Pick a path h ∈ Ω0 of length n in SxΓ (gx, R), and let b
h denote
the longest subpath of h starting at the initial vertex and which ends in a vertex of large
growth. Let us write h = b
hl where l is the second part of the path. Note that we have
d(hx, b
hx) = d(b
hlx, b
hx) = d(lx, x) ≤ kC
where k := |l| and C is the Lipschitz constant of the orbit map, hence by Lemma 3.5
b
h ∈ SxΓ (gx, R0 + kC)
where R0 = R + D and D = O(δ). Note that for each element b
h there are at most cλk1
choices of the continuation l, hence
#(Sn ∩
SxΓ (gx, R))
≤c
n
X
λk1 #(Sn−k ∩ SxΓ (gx, R0 + kC) ∩ LG) ≤
k=0
and by using eq. (3) and Proposition 4.7
≤ c1
n
X
λk1 λn−k
n−k
P
(SxΓ (gx, R0
k=0
n
+ kC)) ≤ c1 λ
n
X
(λ1 /λ)k p(d(x, gx) − R0 − kC)
k=0
Now, by Lemma 5.3 we can replace p by pe, choosing α so that eα λ1 < λ, thus getting
pe(d(x, gx) − R0 − kC) ≤ eαk pe(d(x, gx) − R0 )
Thus, the previous estimate becomes
P n (SxΓ (gx, R)) ≤ c2
n
X
(eα λ1 /λ)k pe(d(x, gx) − R0 ) ≤ c3 pe(d(x, gx) − R0 )
k=0
which proves the lemma if one sets ρ(r) := c3 pe(r − D).
5.3. Genericity of loxodromics. We now use the previous counting results to prove that
loxodromic elements are generic with respect to the counting measure.
The strategy is to apply the formula of Lemma 3.3 to show that translation length grows
linearly as function of the length of the path: in order to do so, one needs to show that the
distance d(gx, x) is large (as we did in Theorem 5.1) and, on the other hand, the Gromov
product (gx, g −1 x)x is not too large. The trick to do this is to split the path g in two
subpaths of roughly the same length, and show that the first and second half of the paths
are almost independent.
COUNTING LOXODROMICS
17
To define this precisely, for each n let us denote n1 = b n2 c and n2 = n − n1 . For each
path g ∈ Ω, we define its initial part i(g) to be the subpath given by the first n1 edges
of g, and its terminal part t(g) to be the subpath given by the last n2 edges of g. With
this definition, g = i(g) · t(g) and |i(g)| = n1 , |t(g)| = n2 . Moreover, we define the random
variables in , tn : Ω∞ → Ω by in (w) = i(wn ) and tn (w) = t(wn ). Note that by definition
in = wn1 and by the Markov property we have for each paths g, h ∈ Ω:
P(in = g and tn = h) = P(wn1 = g)Pv (wn2 = h)
where v = [g]. In the next Lemma, we use the notation C(ω) to refer to the recurrent
component to which the sample path ω = (wn ) eventually belongs, as in Theorem 4.3.
Lemma 5.4. For any > 0 we have
d(x, tn (ω)x) LC(ω)
> →0
−
lim
P
n→∞
n
2
as n → ∞.
Proof. Note that by definition tn (ω) = wn2 (T n1 ω), where T : Ω∞ → Ω∞ is the shift in the
space of infinite paths. Note that for every A ⊆ Ω∞ by the Markov property we have
X
P(T −n A) =
P([wn ] = v)Pv (A)
v∈V
Let us define the function
Sn (ω, ω 0 ) :=
d(x, wn (ω)x)
− LC (ω 0 ) .
n
Note that from Corollary 4.4 for every vertex v of large growth and every > 0
Pv (Sn (ω, ω) ≥ ) → 0
Moreover, for every n, if the path (e1 , . . . , en , . . . ) lies entirely in the component C from some
point on, then the same is true for the shifted path (en+1 , en+2 , . . . ), i.e. C(T n ω) = C(ω)
almost surely, and so
Sn (ω, T k ω 0 ) = Sn (ω, ω 0 )
for all n, k
hence
=
X
P(Sn2 (T n1 ω, ω) > ) = P(Sn2 (T n1 ω, T n1 ω) > ) =
X
P([wn1 ] = v])P(Sn2 (ω, ω) > ) ≤
Pv (Sn2 (ω, ω) > )
v∈V
v∈V ∩LG
and the right-hand side tends to 0 by Corollary 4.4, proving the claim.
We now show that i(g) and t(g)−1 generically do not fellow travel. For the argument, let
Sn (v) denote the set of paths in Ω which start at v and have length n.
Lemma 5.5. Let f : R → R be any function such that f (n) → +∞ as n → +∞. Then
P n g ∈ Ω0 : (i(g)x, t(g)−1 x)x ≥ f (n) → 0
as n → ∞.
Proof. We compute
#{g ∈ Sn1 , [g] = v, h ∈ Sn2 (v) : (gx, h−1 x)x ≥ f (n)}
≤
P n g : (i(g)x, t(g)−1 x)x ≥ f (n) =
#Sn
18
I. GEKHTMAN, S.J. TAYLOR, AND G. TIOZZO
and by fixing v and forgetting the requirement that [g] = v we have
X #{g ∈ Sn , h ∈ Sn (v) : (gx, h−1 x)x ≥ f (n)}
1
2
≤
≤
#Sn
v∈V
then by fixing a value of h
X
X 1
≤
#Sn
v∈V
# g ∈ Sn1 : gx ∈ Sx (h−1 x, d(x, h−1 x) − f (n)) ≤
h∈Sn2 (v)
hence from decay of shadows (Proposition 5.2) follows that
X X ρ(f (n))#Sn
#V #Sn1 #Sn2 ρ(f (n))
1
≤
≤
≤ cρ(f (n)) → 0.
#Sn
#Sn
v∈V h∈Sn2 (v)
Once we have shown that i(g) and t(g)−1 are almost independent, we still need to show
that also g and g −1 are almost independent. In order to do so, we note that i(g) is the
beginning of g while t(g)−1 is the beginning of g −1 , and then we use the following trick from
hyperbolic geometry. See e.g. [TT15].
Lemma 5.6 (Fellow traveling is contagious). Let X be a δ–hyperbolic space with basepoint
x and let that A ≥ 0. If a, b, c, d are points of X with (a · b)x ≥ A, (c · d)x ≥ A, and
(a · c)x ≤ A − 3δ. Then (b · d)x − 2δ ≤ (a · c)x ≤ (b · d)x + 2δ.
In order to apply Lemma 5.6, we need to check that the first half of g (which is i(g)) and
the first half of g −1 (which is t(g)−1 ) generically do not fellow travel.
Lemma 5.7. For each η > 0, the probability
n(L − η)
n
−1
−1
P
g ∈ Ω0 : (t(g) x, g x)x ≤
→0
2
and
P
n
g ∈ Ω0
n(L − η)
: (i(g)x, gx)x ≤
2
→0
as n → ∞.
Proof. Consider the set BL := {g ∈ Ω0 : d(x, gx) − d(x, i(g)x) ≤
Theorem 4.3 for P-almost every sample path we have
|g|L
2 }.
We know by
LC
L
d(x, wn x) − d(x, wn1 x)
=
≥
n
2
2
hence for any L0 < L one has Pn (BL0 ) → 0 as n → ∞. Hence, as in the proof of Theorem
5.1 we get for any > 0,
n(L − )
n
(8)
P
d(x, gx) − d(x, i(g)x) ≥
→1
2
lim
n→∞
Finally, by writing out the Gromov product, the triangle inequality and the fact that the
action is isometric we get
(t(g)−1 x, g −1 x)x ≥ d(x, g −1 x) − d(t(g)−1 x, g −1 x) = d(x, gx) − d(x, i(g)x)
which combined with (8) proves the first half of the claim.
COUNTING LOXODROMICS
19
The second claim follows analogously. Namely, from Theorem 4.3 and Lemma 5.4, we
have for any > 0
n(LC(ω) − )
P d(x, wn (ω)x) − d(x, tn (ω)x) ≤
→0
2
which then implies as before
n(L − )
P n d(x, gx) − d(x, t(g)x) ≥
→1
2
and to conclude we use that
(i(g)x, gx)x ≥ d(x, gx) − d(i(g)x, gx) = d(x, gx) − d(x, t(g)x).
We now use Lemma 5.6 (fellow traveling is contagious) to show that the Gromov products
(gx, g −1 x)x do not grow too fast with respect to our counting measures.
Proposition 5.8. Let f : N → R be a function such that f (n) → +∞ as n → ∞. Then
P n (gx, g −1 x)x ≤ f (n) → 1
as n → ∞.
Proof. Define
n(L − η)
− 3δ
f1 (n) = min f (n) − 2δ,
2
It is easy to see that f1 (n) → ∞ as n → ∞. By Lemma 5.6, if we know that:
(1) (i(g)x, gx)x ≥ n(L − η)/2,
(2) (t(g)−1 x, g −1 x)x ≥ n(L − η)/2, and
(3) (i(g)x, t(g)−1 x)x ≤ f1 (n) ≤ n(L − η)/2 − 3δ,
then
(gx, g −1 x)x ≤ (i(g)x, t(g)−1 x)x + 2δ ≤ f1 (n) + 2δ.
Using Lemmas 5.5 and 5.7, the probability that conditions (1),(2), (3) hold tends to 1, hence
we have
P n ((gx, g −1 x)x ≤ f (n)) → 1
as n → ∞.
Finally, we put together the previous estimates and use Lemma 3.3 to prove that translation length grows linearly and loxodromic elements are generic.
Theorem 5.9 (Linear growth of translation length). Let (G, Γ) be an almost semisimple,
nonelementary graph structure, and L the smallest drift given by eq. (7). Then for any
> 0 we have
#{g ∈ Sn : τX (g) ≥ n(L − )}
→ 1,
#Sn
as n → ∞. As a consequence,
#{g ∈ Sn : g is X − loxodromic}
→ 1,
#Sn
as n → ∞.
20
I. GEKHTMAN, S.J. TAYLOR, AND G. TIOZZO
Proof. If we set f (n) = ηn with η > 0, then by Proposition 5.8 and Theorem 5.1 the events
(gx, g −1 x)x ≤ ηn and d(x, gx) ≥ n(L − η) occur with probability (P n ) which tends to 1,
hence by Lemma 3.3
P n τX (g) ≥ n(L − 3η) ≥ P n d(x, gx) − 2(gx, g −1 x)x + O(δ) ≥ n(L − 3η)
which approaches 1 as n → ∞. This implies the statement if we choose > 3η. The
second statements follows immediately since elements with positive translation length are
loxodromic.
5.4. Genericity of loxodromics for the Markov chain. We now remark that a very
similar proof yields that loxodromics are generic for P-almost every sample path of the
Markov chain. More precisely, we have the following (which is a reformulation of Theorem
2.7):
Theorem 5.10. Let (G, Γ) be an almost semisimple, nonelementary graph structure, and
let L be the smallest drift. Then for every > 0, one has
P τX (wn ) ≥ n(L − ) → 1,
as n → ∞. As a consequence,
P wn is loxodromic on X → 1
as n → ∞.
Proof. The proof is very similar to the proof of Theorem 5.9, so we will just sketch it. First,
by using the Markov property we establish that
lim P (in x, t−1
n x)x ≥ g(n) = 0
n→∞
for any choice of function g : N → R such that limn→+∞ g(n) = +∞. Then, by using
positivity of the drift as in the proof of Lemma 5.7 we prove that for each > 0, we have
lim P (wn−1 x, t−1
n x)x ≤ n(L − )/2 = 0
n→∞
and
lim P (in x, wn x)x ≤ n(L − )/2 = 0
n→∞
From the previous three facts, using Lemma 5.6 one proves:
lim P (wn x, wn−1 x)x ≥ f (n) = 0
n→∞
for any f : N → R such that limn→+∞ f (n) = +∞. The theorem then follows immediately
from this fact and Corollary 4.5, applying the formula of Lemma 3.3.
6. Thick graph structures
Definition 6.1. A graph structure (G, Γ) is thick if for every vertex v of maximal growth
there exists a finite set B ⊆ G such that
(9)
G = BΓv B
where Γv is the loop semigroup of v.
In greater generality, if H < G is a subgroup, we say that the graph structure (G, Γ) is
thick relatively to H if for any vertex v of maximal growth there exists a finite set B ⊆ G
such that
(10)
H ⊆ BΓv B.
COUNTING LOXODROMICS
21
6.1. The case of only one non-trivial component. We say that a component C is
non-trivial if there is at least one closed path of positive length entirely contained in C.
Proposition 6.2. If a graph structure (G, Γ) has only one non-trivial component, then it
is thick.
Proof. Let C be the unique maximal component of Γ. Every finite path γ in the graph
can be written as γ = h1 gh2 , where h1 is a path from the initial vertex to C, g is a path
entirely in C, and h1 is a path going out of C. By construction, the lengths of h1 and h2
are uniformly bounded. Fix some vertex v of C and let s be a shortest path from v to the
last vertex of h. Further, let t be a shortest path from the last vertex of g to v. Then one
can write
γ = h1 gh2 = h1 s−1 (sgt)t−1 h2
where h1 s−1 and t−1 h2 vary in a finite set, and sgt ∈ Γv . Hence G = BΓv B with B a finite
set.
6.2. Thick implies nonelementary.
Proposition 6.3. Fix an action G y X of G on a hyperbolic metric space X. Let (G, Γ)
be an almost semisimple graph structure, and H < G a nonelementary subgroup. If (G, Γ)
is thick relatively to H, then it is nonelementary, i.e. for any maximal vertex v the action
of the loop semigroup Γv on X is nonelementary.
Proof. Since the action of H is nonelementary, there exists a free subgroup F ⊆ H of rank
2 which quasi-isometrically embeds in X. Hence, the orbit map F → X extends to an
embedding ∂F → ∂X, and we identify ∂F with its image. Thickness implies F ⊆ BΓv B,
and taking limit sets in ∂X we see that
[
∂F ⊂
b · ΛΓv ,
b∈B
from which we conclude that ΛΓv is infinite. To complete the proof that Γv is nonelementary,
it suffices to show that Γv does not have a fixed point on ∂X. Suppose toward a contradiction
that p ∈ ∂X is such a fixed point.
Let us write F = hf, gi where f, g are free generators of F , and consider the sequence of
elements hi,j = f i g j in F . For each i, j there are ai,j , ci,j ∈ B such that hi,j = ai,j li,j ci,j for
some li,j in Γv . Since B is finite, we may pass to a subsequence and assume that ai,j = a
and ci,j = c for all i, j. Then li,j = a−1 hi,j c−1 fixes the point p for all i and so
hi,j (c−1 (p)) = a(p)
−i0 i−i0 j
for all i, j. Hence h−1
f
g is a sequence of elements of F which fix the point
i0 ,j0 hi,j = g
−1
q = c (p) ∈ ∂F ⊂ ∂X. Since F is a free group, this implies that g −i0 f i−i0 g j agree up to
powers for infinitely many i, j, a clear contradiction.
From Proposition 6.3 and Theorem 5.9 we get:
Theorem 6.4. Let G y X be a nonelementary action of a countable group on a separable,
hyperbolic metric space. Suppose that G has an almost semisimple graph structure Γ which
is thick with respect to a nonelementary subgroup H. Then loxodromic elements are generic:
lim
n→∞
#{g ∈ Sn : g is loxodromic on X}
=1
#Sn
22
I. GEKHTMAN, S.J. TAYLOR, AND G. TIOZZO
In fact, the translation length generically grows linearly: there exists L > 0 such that
lim
n→∞
#{g ∈ Sn : τX (g) ≥ Ln}
= 1.
#Sn
This completes the proof of Theorem 1.6 in the introduction.
7. Relative growth quasitightness
Fix a graph structure (G, Γ). In practice, we will often show that the graph structure is
thick by establishing the property of growth quasitightness. This property was introduced
in [AL02] and and further studied in [Yan16]. Our notion of quasitightness depends on the
particular graph structure.
Given a path γ in Γ, we say it c–almost contains an element w ∈ G if γ contains a subpath
p such that w = a · ev(p) · b in G, with |a|, |b| ≤ c. We denote as Yw,c the set of paths in Γ
starting at the initial vertex which do not c–almost contain w.
Definition 7.1. A graph structure (G, Γ) is called growth quasitight if there exists c > 0
such that for every w ∈ G the set Yw,c has density zero with respect to P n ; that is,
P n (Yw,c ) → 0
as n → 0.
More in general, given a subgroup H < G we say that (G, Γ) is growth quasitight relatively
to H if there exists a constant c > 0 such that for every w ∈ H the set Yw,c has density
zero.
7.1. Growth quasitight implies thick.
Proposition 7.2. Let (G, Γ) be an almost semisimple graph structure, and H < G a subgroup. If (G, Γ) is growth quasitight relatively to H, then it is thick relatively to H.
Proof. Let C be a component of maximal growth, let v a vertex in C, and let γ be some
path from the initial vertex to v. Denote the length of γ by d. Let w ∈ H. By growth
quasitightness plus maximal growth, there is a path of the form γγ1 , which c–almost contains
w and where γ1 is entirely contained in C. Since γ has length d, the path γ1 (c + d)–almost
contains w; that is,
γ1 = p1 p2 p3
where ev(p2 ) = awb for |a|, |b| ≤ c + d. Let q1 be a shortest path from v to the initial vertex
of p2 and q2 be a shortest path from the terminal vertex of p2 to v. Then
w = (a−1 ev(q1 )−1 ) · ev(q1 p2 q2 ) · (ev(q2 )−1 b−1 ),
where (a−1 ev(q1 )−1 ) and (ev(q2 )−1 b−1 ) vary in a finite set B. Since ev(q1 p2 q2 ) ∈ Γv , this
completes the proof.
Combining Proposition 7.2 with Theorem 6.4 we get:
Theorem 7.3. Let (G, Γ) be an almost semisimple graph structure which is growth quasitight
with respect to a nonelementary subgroup H. Then loxodromic elements are generic.
This completes the proof of Theorem 1.6.
COUNTING LOXODROMICS
23
8. Infinite index subgroups have zero density
In this section, we prove that in our general setup a subgroup H < G of infinite index
has zero density with respect to counting. Combined with what we are going to prove in
sections 9 and 10, this immediately implies Theorem 1.2 in the introduction. Recall that
ev : Ω0 → G is the evaluation map for paths starting at v0 .
Theorem 8.1. Let (G, Γ) be an injective, almost semisimple, thick graph structure. Let
H < G be an infinite index subgroup. Then
P n ({p ∈ Ω0 : ev(p) ∈ H}) → 0,
as n → ∞. That is, the proportion of paths starting at v0 which spell an element of H goes
to 0 as the length of the path goes to ∞.
The proof is an adaptation of ([GMM15], Theorem 4.3) to the non-hyperbolic case. We
will consider an extension ΓH of Γ = (V, E) defined as follows. The vertex set of ΓH is
V × H \ G. For any edge σ : x → y in Γ there is an edge in ΓH from (x, Hg) to (y, Hgg 0 )
where g 0 = ev(σ).
Lemma 8.2. Let C be a component of maximal growth. For any v1 ∈ C and g1 ∈ G there
are infinitely many Hg ∈ H \ G such that (v1 , Hg) can be reached from (v1 , Hg1 ) by a path
contained in C × H \ G.
Proof. Suppose not, so that the only points of H \ G that can be reached in this manner
are {Hz : z ∈ T } where T is a set of size D. Consider w ∈ G. By thickness, there exists a
finite set B ⊆ G and some path γ lying in C, starting and ending at v1 such that
ev(γ) = g2 wg3
where g2 , g3 lie in B. Then γ lifts to a path in ΓH from (v1 , Hg1 ) to (v1 , Hg1 ev(γ)). By
assumption, this implies Hg1 ev(γ) = Hz for some z ∈ T . Thus, there is an h ∈ H with
g2 wg3 = ev(γ) = g1−1 hz and hence w ⊂ B −1 g1−1 HT B −1 . Thus there is a finite subset
Υ = B −1 g1−1 ∪ T B −1 with G = ΥHΥ, so by Neumann’s theorem [Neu54] H must be of
finite index, giving a contradiction.
The following general result about Markov chains is Lemma 4.4 of [GMM15].
Lemma 8.3. Let Xn be a Markov chain on P
a countable set V , and m a stationary measure.
Let Ṽ be the set of points x ∈ V such that y:x→y m(y) = ∞ where x → y means there is
a positive probability path from x to y. Then for all x ∈ V and x0 ∈ Ṽ we have Px (Xn =
x0 ) → 0.
Combining Lemmas 8.3 and 8.2 we obtain:
Corollary 8.4. For any x1 , x2 ∈ Γ lying in a maximal component C and g1 , g2 ∈ G, the
number of paths of length n in ΓH from (x1 , Hg1 ) to (x2 , Hg2 ) is o(λn ).
Proof. The Markov chain µ on Γ restricts to a Markov chain µC on C, which in turn lifts to
a Markov chain µC,H on the induced graph CH on the vertex set C × H \ G ⊂ ΓH (obtained
by assigning to an edge the transition probability of its projection to C. A µC,H stationary
measure m̃ on CH is given by taking the product of the stationary measure m on C and
the counting measure on H \ Γ. Any vertex v ∈ C has positive m P
measure and all lifts of v
in CH have equal positive m̃ measure. Thus, Lemma 8.2 implies y:x→y m̃(y) → ∞. The
corollary now follows by applying Lemma 8.3 to the chain µC,H .
24
I. GEKHTMAN, S.J. TAYLOR, AND G. TIOZZO
Note, paths of length n in ΓH from (x1 , Hg1 ) to (x2 , Hg2 ) are in bijection with paths of
length n in Γ beginning at x1 , ending at x2 , and evaluating to elements of g1−1 Hg2 . Thus,
we obtain:
Corollary 8.5. For any x1 , x2 ∈ Γ lying in a maximal component C and g1 , g2 ∈ G, the
number of paths of length n in Γ beginning at x1 , ending at x2 , and evaluating to elements
of g1−1 Hg2 is o(λn ).
We now complete the proof of Theorem 8.1. Given k > 0, let Pn,k (resp. Qn,k ) be the set
of paths p ∈ Ω0 of length n which spend time at most k (resp. more than k) in non-maximal
components.
Note that there is a η < λ with |Qn,k | ≤ η k λn−k for all n and k. Now, consider a path γ
in Pn,k ∩ ev−1 H. We can decompose it as γ = γ1 γ2 γ3 where γ1 and γ3 have length adding
up to at most k and γ2 is contained in a maximal component C. Since a path in Pn,k spends
at most time k in nonmaximal components, there are only Dk possibilities for γ1 and γ3 ,
where D depends only on the graph. On the other hand, by Corollary 8.5, for a path in
ev−1 H, there are at most f (n) = o(λn ) possibilities for γ2 , thus |Pn,k ∩ ev−1 H| ≤ Dk f (n)
for all k < n where f (n) = o(λn ). Hence,
P n (ev−1 H) ≤ C 00 λ−n (|Pn,k ∩ ev−1 H| + |Qn,k |) ≤ C 00 Dk λ−n f (n) + C 00 (η/λ)k .
Fixing k we see that
lim sup P n (ev−1 H) ≤ C 00 (η/λ)k .
n→∞
As this is true for arbitrary k, we get limn→∞ P n (ev−1 H) = 0, as claimed.
9. Application to relatively hyperbolic groups
In this section, we show how our main theorem applies to a large class of relatively
hyperbolic groups.
Let G be a finitely generated group, and P be a collection of subgroups. Following
[Bow12], let us recall that G is hyperbolic relative to P if there is a compactum M on which
G acts geometrically finitely, and the maximal parabolic subgroups are the elements of P.
Such a compactum M is then unique up to G-equivariant homeomorphisms, and it is called
the Bowditch boundary of G. We will denote it as ∂G.
More precisely, let G act by homeomorphisms on a compact, perfect, metrizable space M .
Then a point ζ ∈ M is called conical if there is a sequence (gn ) and distinct points α, β ∈ M
such that gn ζ → α and gn η → β for all η ∈ M \ {ζ}. A point ζ ∈ M is called bounded
parabolic if the stabilizer of ζ in G is infinite, and acts cocompactly on M \ {ζ}. We say that
the action of G on M is a convergence action if G acts properly discontinuously on triples of
elements of M , and the action is geometrically finite if it is a convergence action and every
point of M is either a conical limit point or a bounded parabolic point. Note that there
are only countably many parabolic points. Finally, the maximal parabolic subgroups are the
stabilizers of bounded parabolic points. We refer the reader to [Far98, Bow12, Osi06, GM08]
for the relevant background material.
Fix a relatively hyperbolic group (G, P), a generating set S, and let dG denote distance
in G with respect to S. Let Ĝ be the vertices of Cay(G, S ∪ P) with the induced metric,
which we denote by dˆ or dĜ . Here Cay(G, S ∪ P) is the corresponding electrified
Cayley
S
graph, that is the Cayley graph of G with respect to the generating set S ∪ P. We remind
the reader that Cay(G, S ∪ P) is hyperbolic and that ∂Cay(G, S ∪ P) naturally includes as
a subspace into ∂G, the complement of which is the collection of parabolic fixed points.
COUNTING LOXODROMICS
25
Following [Yan13], the Bowditch boundary ∂G is equipped with a quasiconformal, nonatomic
measure ν, which is given by the Patterson-Sullivan construction by taking average on balls
for the word metric on G.
Definition 9.1. We define a relatively hyperbolic group G to be pleasant if its action on
∂G × ∂G with the measure ν × ν is ergodic.
We will also see (Proposition 9.17) that a relatively hyperbolic group is pleasant if it
admits a geometrically finitely action on a CAT (−1) proper metric space. For instance,
geometrically finite Kleinian groups satisfy this hypothesis. Note that, once G admits such
an action, the theorem works for isometric actions of G on any hyperbolic, metric space X.
In this section, we will prove the following result.
Theorem 9.2. Let G be a pleasant, relatively hyperbolic group, and let (G, Γ) be a geodesic
combing. Then for any nonelementary action of G on a hyperbolic metric space X, the
graph structure (G, Γ) is nonelementary.
Combining this result with Theorem 1.6, the discussion in section 2.3, and the fact that
relatively hyperbolic groups has pure exponential growth for any generating set ([Yan13],
Theorem 1.9) this establishes Theorem 2.3 in the introduction.
In fact, using very recent work of W. Yang [Yan16], the theorem may be extended to all
nontrivial relatively hyperbolic groups, as relatively hyperbolic groups contain strongly contracting elements by [ACT15] and [Yan16] (see Corollary 2.6 in the introduction). However,
we give a self-contained argument here.
9.1. Fellow traveling in the Cayley graph and coned-off space. We will need the
following proposition, which is certainly known to experts. We provide a proof for completeness.
Proposition 9.3. For K, C ≥ 0, there are D, L ≥ 0 such that the following holds. Suppose
that γ = [a, b] is a geodesic in Cay(G, S) with length at least L which projects to a K–
quasigeodesic in Cay(G, S ∪ P). Let γ 0 be any other geodesic in Cay(G, S) whose endpoints
have distance no more than C from a, b in Cay(G, S ∪ P). Then there are a0 , b0 ∈ γ 0 such
that
dG (a, a0 ) ≤ D and dG (b, b0 ) ≤ D.
We will use the following theorem of Osin:
Theorem 9.4 ([Osi06], Theorem 3.26). There is an ν ≥ 0 such that if p, q, r are sides of
a geodesic triangle in Cay(G, S ∪ P), then for any vertex v on p there exists a vertex u on
either q or r such that
dG (v, u) ≤ ν.
Proof of Proposition 9.3. Suppose that dG (a, b) > K(2C + 4ν + K) so that dĜ (a, b) >
2C + 4ν, for ν as in Theorem 9.4.
Let c, c0 be geodesics in Cay(G, S∪P) joining the endpoints of γ, γ 0 respectively. Note that
ˆ
by assumption the initial and terminal endpoints of these geodesics are at d–distance
less
than C from one another. Pick vertices ca , cb on c (ordered a, ca , cb , b) so that dĜ (a, ca ) =
dĜ (b, cb ) = C + 2ν. (This is possible since dĜ (a, b) > 2C + 4ν.) Consider a geodesic
quadrilateral with opposite sides c, c0 . Applying Theorem 9.4 twice, we may find vertices
c0a , c0b ∈ c0 such that dG (ca , c0a ) ≤ 2ν and dG (cb , c0b ) ≤ 2ν.
Now using, for example, ([Hru10], Lemma 8.8), we can find vertices γa , γb ∈ γ and
γa0 , γb0 ∈ γ 0 which have dG –distance at most L from ca , cb , c0a , c0b , respectively, where L ≥ 0
26
I. GEKHTMAN, S.J. TAYLOR, AND G. TIOZZO
depends only on (G, P) and S. Note that dG (γa , γa0 ) ≤ 2(L + ν) and dG (γb , γb0 ) ≤ 2(L + ν).
Moreover,
dĜ (a, γa ) ≤ dĜ (a, ca ) + dĜ (ca , γa ) ≤ C + 2ν + L,
and so dG (a, γa ) ≤ K(C + 2ν + L + 1) since both a and γa occur along γ. Similarly,
dG (b, γb ) ≤ K(C + 2ν + L + 1).
Putting everything together, after setting a0 = γa0 and b0 = γb0 , we see that each of
dG (a, a0 ) and dG (b, b0 ) are less than 2(L + ν) + K(C + 2ν + L + 1) and this completes the
proof.
9.2. Patterson-Sullivan measures and sphere averages. Continuing with the notation
from the previous section, let h be the exponent of convergence for Cay(G, S). That is,
h = lim log |Bn |/n,
n→∞
where Bn denotes the ball of radius n in G with respect to dG .
Definition 9.5. For g ∈ G, we define the large shadow Πr (g) at g to be the set of ζ ∈ ∂G
such that there exists some geodesic in Cay(G, S) from 1 converging to ζ intersecting Br (g).
Similarly, the small shadow πr (g) is the set of ζ ∈ ∂G such that every geodesic in Cay(G, S)
from 1 converging to ζ intersects Br (g).
In Theorem 1.7 and Proposition A4 of [Yan13] Yang constructs an h-quasiconformal
ergodic density ν without atoms for the word metric on the Bowditch boundary ∂G. In
Lemma 4.3 of [Yan13] he shows that this satisfies the shadow lemma: for large enough r:
ν(πr (g)) ' ν(Πr (g)) ' e−h·dG (1,g)
(11)
(up to a uniform multiplicative constant). In particular, ν has full support on ∂G. In what
follows, Sn denotes the set of elements g ∈ G with dS (1, g) = n.
Lemma 9.6. There is a C > 0 such that for any Borel set A ⊂ G ∪ ∂G one has
|A ∩ Sn |
≤ Cν(A),
lim sup
|Sn |
n→∞
where A denote the closure of A in G ∪ ∂G.
Proof. Let A ⊂ G ∪ ∂G be a Borel set. Since the number of elements in a ball of radius
r in Cay(G, S) is universally bounded, a point of ∂G lies in at most D small shadows
πr (g), g ∈ Sn where D depends only on r. Thus,
X
[
ν (πr (g)) ≤ Dν
πr (g)
g∈Sn ∩A
g∈Sn ∩A
Moreover, if we denote An := A \ Bn−1 then An+1 ⊆ An and
\ [
πr (g) ⊂ A.
n∈N g∈An
T
S
Indeed, if ζ ∈ n∈N g∈An πr (g), then there are gn ∈ A with |gn | ≥ n such that some (any)
geodesic from the identity to ζ meets Br (gn ). Hence, gn → ζ and so ζ ∈ A.
Thus, since Sn ∩ A ⊆ An we have for large enough n
[
[
ν
πr (g) ≤ ν
πr (g) ≤ 2ν(A)
g∈Sn ∩A
g∈An
COUNTING LOXODROMICS
27
so by exponential growth and the shadow lemma (11)
X
|Sn ∩ A|
' e−hn |Sn ∩ A| '
ν(πr (g)) . 2Dν(A).
|Sn |
g∈Sn ∩A
9.3. Growth quasitightness for relatively hyperbolic groups. We will now establish
a form of relative growth quasitightness for a relatively hyperbolic group G. S
Let w be an element of G. A w-path is an infinite path of the form lw = i∈Z wi γw in
the Cayley graph Cay(G, S), where γw = [1, w] is a geodesic segment joining the identity
and w. Of course there may be finitely many choices of lw for each w.
Definition 9.7. The element w is called K–bounded if some w-path lw (with the arc
length parameterization) in the Cayley graph Cay(G, S) projects to a K-quasigeodesic in
the electrified graph Cay(G, S ∪ P).
The following lemma is well-known. See for example [DMS10, ADT].
Lemma 9.8. For each K, there is a function f : N → N such that if w is K–bounded, then
every w-path lw is an f -stable quasigeodesic in the Cayley graph Cay(G, S).
Recall that lw being f –stable means that any K–quasigeodesic with endpoints on lw has
Hausdorff distance at most f (K) from the subpath of lw its endpoints span.
Given w ∈ G and c ≥ 0, we say that a (finite or infinite) geodesic γ c-almost contains w
if there exists g ∈ G such that dG (g, γ) ≤ c and dG (gw, γ) ≤ c. Let Xw,c be the set of h ∈ G
such that there exists a geodesic γ from identity to h which does not c-almost contain w.
That is, for every g ∈ Nc (γ), γ does NOT pass within distance c of gw.
Proposition 9.9. For each K ≥ 1, there is c ≥ 0 such that for every K–bounded w ∈ G
we have
|Bn ∩ Xw,c |
→0
ehn
as n → ∞.
We remark that, for fixed c, it suffices to prove the proposition for sufficiently long w, that
is, where |w|S is sufficiently large. We will prove this proposition by using the ergodicity of
the double boundary (Proposition 9.17). To do this, we will apply Proposition 9.3 several
times, for K the boundedness constant. Hence, we fix K once and for all, and consider the
constant D produced by that proposition as a function of C alone and write D = D(C).
Let Zw,c be the set of pairs (α, β) in ∂G × ∂G such that for every bi-infinite geodesic γ in
Cay(G, S) joining α and β there exist infinitely many x ∈ Nc (γ) such that γ passes within c
n
of xw. Let Zw,c
be the set of pairs (α, β) in ∂G × ∂G such that for every bi-infinite geodesic
γ joining α and β there are
n elements x ∈ Nc (γ) such that γ passes within c of
T at least
n
n
xw. By definition, Zw,c = n∈N Zw,c
. Moreover, for each n, w, c, the sets Zw,c and Zw,c
are
G-invariant subsets of ∂G × ∂G.
Furthermore we have
Lemma 9.10. For each K ≥ 0 there is a constant c0 = c0 (K) such that for all c ≥ c0 , Zw,c
contains a pair of conical points for every K–bounded w ∈ G.
Proof. Let f be the function given by Lemma 9.8. By definition, the w-path lw projects
to a K-quasigeodesic in Cay(G, S ∪ P), hence it has two distinct limit points (w−∞ , w∞ )
in the Bowditch boundary ∂G. Then, by connecting further and further points on lw by a
28
I. GEKHTMAN, S.J. TAYLOR, AND G. TIOZZO
geodesic in the Cay(G, S), using f -stability and the Ascoli-Arzelá theorem, one constructs
a geodesic in Cay(G, S) which connects w∞ to w−∞ and c-fellow travels lw , once c ≥ 2f (1).
Hence, (w−∞ , w∞ ) ∈ Zw,c .
Lemma 9.11. For each K, there is a c1 = c1 (K) and L1 = L1 (K) such that for c ≥ c1
n
and for any K–bounded w ∈ G with |w| ≥ L1 , the set Zw,c
has nonempty interior. More
n
n
precisely, the interior of Zw,c
contains every pair of conical points in Zw,c
where c0 is as
0
in Lemma 9.10.
n
Proof. Suppose that (α, β) ∈ Zw,c
is a pair of conical points, and pick a geodesic γ joining
0
α and β. Then by definition there are segments [xj , x0j ] ⊆ [α, β] for 1 ≤ j ≤ n and points zj
such that dG (xj , zj ), dG (x0j , zj w) ≤ c0 (for 1 ≤ j ≤ n).
Now, let αi , βi ∈ ∂G such that αi → α and βi → β. In particular, for some uniform R,
the projections of the geodesics [αi , βi ] to Cay(G, S ∪ P) R–fellow travel the projection of
[α, β] for longer and longer intervals. Hence, there is an N ≥ 0 so that for i ≥ N , each
ˆ
[αi , βi ] passes with d–distance
R from [xj , x0j ] ⊂ [α, β].
Since w is K–bounded, the geodesic segment zj · [1, w] = [zj , zj w] projects to a K–
quasigeodesic, so we may apply Proposition 9.3 (with C = R + c0 ) to find constants D, L
such that if |w| ≥ L there exist points yj , yj0 ∈ [αi , βi ] such that dG (yj , zj ), dG (yj0 , zj w) ≤ D.
n
and this
Setting c1 = D, L1 = L we see that for sufficiently large i ≥ 0, (αi , βi ) ∈ Zw,c
1
completes the proof.
Since G is pleasant, the action of G on ∂G × ∂G is ergodic, hence Lemma 9.11 implies
Lemma 9.12. For c ≥ c1 (K) and for any K–bounded w ∈ G with |w| ≥ L1 (K), the set
n
Zw,c
has full ν × ν measure. Hence, under the same hypotheses the set Zw,c has full ν × ν
measure.
n
is G-invariant, hence by ergodicity its measure is either 0
Proof. For each n, the set Zw,c
or 1. Since it has nonempty interior and the measure ν × ν has full support, then it must
n
.
have full measure. The second claim follows since Zw,c = ∩n Zw,c
Let Λw,c ⊂ ∂G be the set of conical points α ∈ ∂G such that for every geodesic ray γ
from the identity converging to α there are infinitely many points g ∈ Nc (γ) such that γ
passes within distance c of gw. Using Proposition 9.3 just as in Lemma 9.11 we have:
Lemma 9.13. For each K ≥ 0 there is a c2 = c2 (K) and L2 = L2 (K) such that for c ≥ c2 ,
and for any K–bounded w ∈ G with |w| ≥ L2 , if (α, β) ∈ Zw,c1 , then either α or β is in
Λw,c .
This implies
Corollary 9.14. For each c ≥ c2 and for any K-bounded w ∈ G with |w| ≥ L2 , the set
Λw,c has full ν measure.
Lemma 9.15. For each K ≥ 0 there is a c3 = c3 (K) and L3 = L3 (K) such that for each
c ≥ c3 and for any K–bounded w ∈ G with |w| ≥ L3 , the closure of Xw,c is contained in
∂G \ Λw,c2 .
Proof. If this were false, then for all large c ≥ 0 there would be a sequence (yi ) ⊆ Xw,c
converging to η ∈ Λw,c2 . Since η is not a parabolic fixed point, then one can view η as
belonging to the boundary of Cay(G, S ∪ P). Then the projections to Cay(G, S ∪ P) of any
geodesics γi = [1, yi ] must R–fellow travel the projection to Cay(G, S ∪ P) of [1, η] for longer
and longer intervals, where R is independent of w. If η were in Λw,c2 , then just as in the
COUNTING LOXODROMICS
29
proof of Lemma 9.11, we would obtain by applying Proposition 9.3 (with C = c2 + R) two
constants L3 , c3 such that for |w| ≥ L and for large i, the geodesic γi c3 –almost contains
w. Hence, for c ≥ c3 we obtain a contradiction to yi ∈ Xw,c for all i. This completes the
proof.
We are now in position to prove Proposition 9.9.
Proof of Proposition 9.9. By Lemma 9.15 and Corollary 9.14, for c ≥ c3
ν(Xw,c ) ≤ ν(∂G \ Λw,c2 ) = 0
Hence, applying Lemma 9.6, for large enough c > 0 we have
lim sup e−hn |Bn ∩ Xw,c | ≤ Cν(Xw,c ) = 0.
n
9.4. The loop semigroup is nonelementary. We will now assume that (G, P) is a pleasant relatively hyperbolic group which admits a geodesic combing for the generating set S.
Recall that by work of Antolı́n-Ciobanu [AC16], if the parabolic subgroups are geodesically completable, then every generating set for G can be extended to a generating set for
which G has a geodesic combing. From here on, we will use such a generating set.
Then for each w ∈ G and constant c, let us recall that Yw,c is the set of paths γ in the
directing graph from the initial vertex which do not c-almost contain w, i.e. such that one
cannot write ev(γ) = a1 wa2 in G, with |ai | ≤ c for i = 1, 2. By identifying paths from the
identity with group elements, it is immediate from the definition that Yw,c ⊆ Xw,c . Hence,
by Proposition 9.9, also Yw,c has zero density if w is K–bounded.
Proposition 9.16. Let (G, Γ) be a geodesic combing for a pleasant, relatively hyperbolic
group. Then (G, Γ) is nonelementary.
Proof. We are going to prove that the graph structure is thick relative to a nonelementary,
free subgroup F < G, which yields the claim by Proposition 6.3. Let v be a vertex of
maximal growth and w be any K–bounded word. Let d = diam Γ. Let c be the constant
from Proposition 9.9. Let h1 be a group element representing a path from the initial vertex
to v, and consider the set
Σ = {h1 h2 : h2 ∈ Γv }
Since v has maximal growth and Yw,c has zero density, the set Σ contains a path h which
does not belong to Yw,c . Then there is a path h = h1 h2 such that h1 has length ≤ d, h2
is entirely contained in the component Cv containing v, and h2 contains a subpath of the
form w0 = awb where a and b have length less than c + d. Let s be a path from v to the
start of w0 and t be a path from the end of w0 to v, each of length at most D. Then sw0 t is
in Γv and w = as−1 (swt)t−1 b ⊂ BΓv B where B is a finite set.
To complete the proof, it suffices to show that BΓv B contains a nonelementary subgroup
(Proposition 6.3). Using a standard ping-pong argument, construct a free subgroup H =
hf, gi ≤ G which K–quasi-isometrically embeds in Cay(G, S ∪ P) and which K–quasiisometrically embeds into X for some K ≥ 0. (Indeed, by [TT15], a random 2-generator
subgroup of G will have this property.) For this K, let B be the finite subset produced
above enlarged to contain f ± , g ± . Then for any w ∈ H, at least one of w, wf , or wg is
cyclically reduced in H and hence K–bounded in G. Hence w ∈ BΓv B and so H ≤ BΓv B,
as required.
30
I. GEKHTMAN, S.J. TAYLOR, AND G. TIOZZO
9.5. Double Ergodicity. We conclude this section by proving that a group G which admits
a geometrically finite action on a CAT(−1) space is pleasant.
Let us assume that G acts geometrically finitely on a CAT(−1) space Y . Recall that
an orbit map G → Y induces an embedding ∂G → ∂Y [Bow12, Theorem 9.4], so we can
identify ∂Y with the Bowditch boundary of G. We continue to denote the pushforward of
the measure ν to ∂Y by ν.
Proposition 9.17. Suppose G acts geometrically finitely on a CAT(−1) space Y . Then
the action of G on ∂Y × ∂Y is ergodic with respect to ν × ν.
We remind the reader that ν is quasiconformal with respect to the word metric rather
than the metric on Y .
Proof. Assume G acts geometrically finitely on a CAT(−1) space Y , with elements of P
being the parabolic subgroups. Recall that the Bowditch boundary ∂G is identified with
the Gromov boundary of Y. Let d = dG still denote the word metric on G. For ζ ∈ ∂G and
g, h ∈ G let
βζ (g, h) = lim sup (d(g, z) − d(h, z))
z→ζ
be the Busemann functions for the word metric on G. By W. Yang’s Lemma 2.20 in [Yan13]
there is a C > 0 such that for every conical ζ ∈ ∂G we have
| lim sup [d(g, z) − d(h, z)] − lim inf [d(g, z) − d(h, z)]| < C
(12)
z→ζ
z→ζ
Moreover, the Patterson-Sullivan measure ν (for the word metric) gives full measure to
conical points and is a quasiconformal density in the sense that there is a D > 0 such that:
dg? ν
(ζ) ≤ De−hβζ (g,e)
(13)
D−1 e−hβζ (g,e) ≤
dν
for all g ∈ G and ζ ∈ ∂G. We claim there is a G-invariant measure in the measure class of
ν × ν. Indeed, let
d(e, y) + d(e, z) − d(y, z)
ρe (ζ, ξ) = lim sup
.
2
z→ζ,y→ξ
Define a locally finite measure m0 on (∂G × ∂G) \ Diag by
dm0 (ζ, ξ) = e2hρe (ζ,ξ) dν(ζ) dν(ξ)
We now claim that the measure m0 is G quasi-invariant with a uniformly bounded derivative.
Indeed, we can compute
2ρe (g −1 ζ, g −1 ξ) − 2ρe (ζ, ξ) =
= lim sup d(e, g −1 y) + d(e, g −1 z) − d(g −1 y, g −1 z) − lim sup [d(e, y) + d(e, z) − d(y, z)] =
z→ζ,y→ξ
z→ζ,y→ξ
= lim sup [d(g, y) − d(e, y)] + lim sup [d(g, z) − d(e, z)] + O(1) = βξ (g, e) + βζ (g, e) + O(1)
y→ξ
z→ζ
(where we could distribute the limsup since the limsup and liminf are within bounded
difference (see (12)). Hence, combining this with (13) one gets that the Radon-Nykodym
cocycle is uniformly bounded, i.e.
−1
−1
dg? ν
dg? m0
dg? ν
(ζ)
(ξ) ∼
(ζ, ξ) = e2hρe (g ζ,g ξ)−2hρe (ζ,ξ)
=1
dm0
dν
dν
Hence, by a general fact in ergodic theory the Radon-Nykodym cocycle is also a coboundary
(see [Fur02], Proposition 1). Thus, there exists a G-invariant measure m on (∂G×∂G)\Diag
COUNTING LOXODROMICS
31
in the same measure class as m0 , hence also in the same measure class as ν × ν. By
[Yan13], the Patterson-Sullivan measure is supported on conical limit points. Thus, m is
also supported on pairs of conical limit points. By Theorem 2.6 of [Kai94], any quasi-product
G-invariant Radon measure on the double boundary of a CAT(−1) space which gives full
measure to pairs of conical limit points of G is ergodic. Thus, ν × ν is ergodic.
10. RAAGs, RACGs and graph products
Let Λ be a finite simplicial (undirected) graph. Recall that the corresponding right-angled
Artin group (RAAG) A(Λ) is the group given by the presentation
A(Λ) := hv ∈ V (Λ) : [v, w] = 1 ⇐⇒ (v, w) ∈ E(Λ)i.
The corresponding right-angled Coxeter group (RACG) C(Λ) is the group obtained from
A(Λ) by adding the relators v 2 = 1 for each v ∈ V (Λ). In each case, S = {v ±1 : v ∈ V (Λ)}
is called the set of standard (or vertex ) generators of the group.
In greater generality, let Λ be a finite simplicial graph, and for each vertex v of Λ let us
pick a finitely generated group Gv , which we call vertex group. Then we define the graph
product
G(Λ) := hg ∈ Gv : [g, h] = 1 ⇐⇒ g ∈ Gv , h ∈ Gw and (v, w) ∈ E(Λ)i
as the group generated by the vertex groups Gv with the relation that two vertex groups
commute if and only if the corresponding vertices are joined by an edge. Clearly, RAAGs
are special cases of graph products when Gx = Z for all x, and RACGs are graph products
with Gx = Z/2Z. Graph products were first introduced by Green [Gre90] and have received
much attention, see for example [BHP93, Chi94, Gre90, Hag08, HM95, HW+ 99, Mei96,
Mei95, Rad03].
In this section, we are going to apply our counting techniques to graph products.
10.1. Geodesic combing for graph products. Let us call a group admissible if it has
a geodesic combing with respect to some finite generating set (i.e., in the language of the
previous sections, if it has an admissible generating set). Recall that a recurrent component
is nontrivial if it contains at least one closed path. A component is terminal if there is no
path exiting it. A graph structure is recurrent if every vertex admits a directed path to
every vertex other than the initial one.
Recall that, given a graph Λ, the opposite graph is the graph Λop with the same vertex
set as Λ and such that (v, w) ∈ E(Λop ) if and only if (v, w) ∈
/ E(Λ). We will assume that
Λ is anticonnected, i.e. that the opposite graph Λop is connected. This implies that G(Λ) is
not a direct product of graph products associated to subgraphs of Λ.
Proposition 10.1. Let Λ be anticonnected, and choose for each vertex x a group Gx with
a geodesic combining (Gx , Γx ) for the generating set Sx . Then the graph product G(Λ) with
the generating set S = ∪x Sx admits a geodesic combing which is recurrent.
We call the generating set S in Proposition 10.1, the standard generating set for G(Λ).
Note that this agrees with the standard vertex generators for the special case of right-angled
Artin and Coxeter groups.
The proof of Proposition 10.1 will provide an explicit construction of a recurrent graph
structure for G(Λ) with the standard generators. Hermiller-Meier [HM95] provided a construction of a geodesic combing, which is however not recurrent. In the next few lemmas,
we will show that if Λ is anticonnected we can modify their construction in order to make
32
I. GEKHTMAN, S.J. TAYLOR, AND G. TIOZZO
it recurrent. Of course it is necessary to assume that Λ is anticonnected, as the counting
theorems fail for RAAGs which decompose as direct products (see Example 1).
Let us first review the construction of [HM95]. First introduce a total ordering on the
vertices of Λ such that the first two vertices in the ordering are not adjacent in Λ. Each
vertex of Λ will be labeled by a capital letter A, B, . . . .
Then for each pair of vertices (I, J) such that I and J are not adjacent in Λ and with
I > J one constructs the (I, J)-admissible tree in the following way. An (I, J)-admissible
word is a finite sequence IJK1 K2 . . . Kr (with r ≥ 0) such that:
(1)
J < K1 < K2 < · · · < Kr
and
(2) if Ki ≤ I for some i ≤ r, then Ki is not adjacent to at least one vertex among
I, J, K1 , . . . , Ki−1 .
Given (I, J), the (I, J)-admissible tree is the finite directed tree, whose vertices are labeled
by letters and whose paths spell exactly the (I, J)-admissible words. In particular, such a
tree will have I as a root and there is only one edge coming out of this vertex, with endpoint
J. Here, and in what follows, a directed edge always has the same label as its terminal
vertex.
Moreover, [HM95] define the header graph (the terminology is ours) as the graph with
one vertex for each letter, and an edge A → B if and only if A < B.
Finally, they construct the graph structure for C(Λ) (the corresponding RACG) as follows. Consider the union of an initial vertex v0 , the header graph and all (I, J)-admissible
trees. First, one identifies the vertex I of the header graph with the root of the (I, J)admissible tree for each possible J. Then, one adds one edge from v0 to each vertex of
the header graph, and if A > B and A, B are not adjacent, one joins by a directed edge
each vertex labeled A in the union of the (I, J)-admissible trees with the B vertex in the
(A, B)-admissible tree. As shown by Hermiller-Meier, this graph G gives a bijective, geodesic graph structure for C(Λ) [HM95, Section 5 and Proposition 3.3]. In fact, they show
that G recognizes the geodesic language of normal forms with respect to the ordering of the
vertices of Λ, but we will not need this stronger fact.
Let C be the subgraph of G obtained by removing all vertices in the header graph and
the initial vertex. That is, C is the subgraph induced on the vertices on all (I, J)–admissible
trees, excluding the initial vertex of the tree (which is labeled I).
Lemma 10.2. If Λ is anticonnected, then the graph C is irreducible, i.e. there is one directed
path from each vertex to any other vertex. Hence, G has a unique nontrivial recurrent
component C and this component is terminal.
Proof. We will show that C is indeed irreducible. This will suffice since the header graph
has no directed loops (it only has directed edges which increase in the ordering) and there
are no edges leaving C by construction.
Since the (unique) type J vertex in the (I, J)–admissible tree has a directed path to each
of its vertices and each vertex is in some (I, J)–admissible tree, it suffices to show that from
any vertex of C we can reach the type J vertex of any (I, J)–admissible tree. Hence, fix
some vertex v of G and I, J, which are vertices of Λ.
Here is a main point: Any type I vertex of any admissible tree is joined to the type J
vertex of the (I, J)–admissible tree. Hence, it suffices to get from v to any type I vertex of
any admissible tree. To do this let X be the type of v.
COUNTING LOXODROMICS
33
Fix a path X = X0 , X1 , . . . Xn = I in the complement graph of Λ. (Here we use that Λ
is anticonnected.) That is, Xi and Xi+1 are not adjacent in Λ. We have to get from v to
any type I vertex. We do this inductively as follows: We have either X < X1 or X > X1 .
In the first case, there is a type X1 vertex v1 in the admissible tree containing v along with
a directed path v → v1 . (Since no consecutive pair in our fixed path are adjacent in Λ,
condition (2) above holds automatically.) In the second case, there is an edge from v to the
unique X1 vertex of the (X, X1 )–admissible tree, call this vertex v1 . In either case we get a
directed path v → v1 , were v1 is a type X1 vertex.
We now repeat this argument to produce a path from v1 → v2 , where v2 is a type X2
vertex. Continuing in this manner, we produce v → v1 → . . . → vn , were vn has type I.
Since vn then has a directed edge to the type J vertex of the (I, J)–admissible tree (as
discussed above), this completes the proof,
We now know that the union of the admissible trees (excluding the initial vertices) is an
irreducible graph. However, the header graph by construction is not irreducible. However,
in the following lemma we observe that all words we can spell in the header graph can also
be spelled in one of the admissible trees. Hence, we can modify G (essentially, by removing
the header graph) in order to get a recurrent graph Gr which recognizes the same language
as G.
Lemma 10.3. If Λ is anticonnected, there exists a recurrent graph Gr which recognizes the
same language as G.
Proof. Assume that Λ is anticonnected and that its vertices are ordered so that the first
two vertices A, B do not commute (i.e. they are not adjacent). We modify G so that the
resulting graph Gr still recognizes the same language as G, and it is recurrent.
The modification is simple and requires only one observation: we note that any strictly
increasing sequence X1 . . . Xr can be spelled in the (B, A)–admissible tree, starting from
some vertex. In fact, if X1 = A then BX1 . . . Xr is (B, A)-admissible, since the only
required condition is that whenever Xi ≤ B the vertex Xi is not adjacent to some Xl with
l < i. However, the only two letters not greater than B are A, B, and A and B are not
adjacent by construction. Similarly, if X1 6= A then BAX1 . . . Xr is (B, A)-admissible.
Thus, the new graph Gr is given by removing the header graph and joining the initial
vertex v0 to each vertex of the (B, A)-admissible tree. Any word which is recognized by G is
made of an increasing word followed by a word spelled in the union of the admissible trees.
In Gr , such a word is spelled by spelling the increasing sequence in the (B, A)-admissible
tree, and the second part as before. This proves the claim.
The graph Gr is a recurrent graph which by Lemma 10.3 gives a bijective, geodesic graph
structure the right-angled Coxeter group C(Λ). We now modify the construction to produce
a geodesic combing for each graph product G(Λ).
Let ΓI be the graph structure of the vertex group GI , let v0,I be the initial vertex of
GI , let s1,I , . . . , sk,I be the labels of the edges going out of v0,I , and let v1,I , . . . , vI,k be
the targets of these edges, respectively. Moreover, let Γ0I be the subgraph of ΓI given by
removing the initial vertex.
To construct the graph structure for G(Λ), let us consider the disjoint union of a vertex
ve0 , which will serve as initial vertex, and a copy of Γ0I for each vertex v of type I in Gr .
Moreover, for any edge in Gr of type I → J let us connect each vertex of the corresponding
Γ0I with the vertices v1,J , . . . , vk,J of the corresponding Γ0J with edges labeled, respectively,
s1,J , . . . , sk,J . Finally, for each edge from v0 in Gr to some other vertex of type I, let
34
I. GEKHTMAN, S.J. TAYLOR, AND G. TIOZZO
us connect the new initial vertex ve0 with vertices v1,J , . . . , vk,J of Γ0J with edges labeled,
respectively, s1,J , . . . , sk,J .
This new graph ΓG gives a bijective, geodesic structure for G(Λ) with respect to the standard generators. This follows since, by construction, ΓG parameterizes the same language
of geodesic normal forms for G(Λ) given in [HM95]. Moreover, since ΓG is modeled on the
recurrent graph Gr one easily sees that ΓG is itself recurrent. This completes the proof of
Proposition 10.1.
Corollary 10.4. Let G(Λ) be a graph product of admissible groups, which does not decompose as a direct product. Then there exists a thick graph structure for its standard generating
set.
Proof. From Proposition 6.2, the graph structure given by the above proposition is thick
since ΓG is recurrent.
As a consequence of thickness, we are ready to establish the following counting result for
loxodromics.
Theorem 10.5. Let G be an infinite graph product of admissible groups which does not
decompose as a product of infinite groups, and let S be its set of standard (vertex) generators.
Then for any nonelementary action G y X on a separable hyperbolic space X, the set of
loxodromics for the action is generic with respect to S, i.e.
#{g ∈ G : |g|S ≤ n and g is X − loxodromic}
−→ 1,
#{g ∈ G : |g|S ≤ n}
as n → ∞.
11. Exact exponential growth for RAAGs and RACGs
We conclude by proving a fine estimate on the number of elements in a ball for RAAGs
and RACGs.
Theorem 11.1. Let G be a right-angled Artin or Coxeter group which is not virtually cyclic
and does not decompose as a product of infinite groups, and let us consider S its standard
generating set. Then there exists constants λ > 1, c > 0 such that the following limit exists:
(14)
lim
n→∞
#{g ∈ G : |g|S = n}
= c.
λn
We say that a group which satisfies (14) has exact exponential growth. Let us remark that
such a property is not invariant with respect to quasi-isometries of the metric, and hence it
depends very carefully on the generating set.
In fact, Theorem 11.1 will follow immediately from the following theorem for general
graph products.
Theorem 11.2. Let G(Λ) be a graph product of admissible groups, and assume that Λ is
anticonnected (so that the group does not split trivially as a product) and has at least 3
vertices. Then G(Λ) has exact exponential growth.
Note that it makes sense to assume that the number of vertices is at least 3. In fact, if
n = 1 then G(Λ) can be any group with a geodesic combing, while if n = 2 then G(Λ) can
be the free product of any two admissible groups. In particular, if it is a RAAG then it
must be the free group on 2 generators, which has exact exponential growth, and if it is a
RACG it must be Z/2Z ? Z/2Z, which is virtually cyclic.
COUNTING LOXODROMICS
35
Let us remark that the growth function for graph products has been worked out by
Chiswell [Chi94] (see also [AP14]); however, it does not seem obvious how to prove exact
exponential growth by this method.
Let us consider the recurrent graph Gr defined in the previous section, and denote as
r
G0 = Gr \ {v0 }. By the previous section, we know that Gr0 is irreducible. The final step in
the proof of Theorem 11.2 is the following lemma.
Lemma 11.3. If Λ is anticonnected and has at least 3 vertices, then the graph Gr0 is aperiodic.
Proof. Let us assume, consistently with the previous section, that the vertices of Λ are
ordered. Let us call A, B, C the three smallest vertices, with A < B < C, and assume that
A, B are not adjacent. Then let us observe that the sequences BAC and BABC are (B, A)
admissible, hence in the (B, A)-admissible tree there is a Y -shaped subtree with five vertices:
one labeled A, two labeled B (let us denote them B1 , B2 ) and two labeled C (let us denote
them C1 , C2 ) so that the paths in this subtree are B1 → A → C1 and B1 → A → B2 → C2 .
Now, since the graph is irreducible, there exists a path from C1 to A; let us denote its
vertices as C1 → v1 → v2 · · · → vk → A. Then by definition, the type of v1 is smaller than
C, and is not adjacent to C. Thus, by construction, there is also an edge from C2 to v1 ;
hence, in the graph there are two loops: one loop is given by A → C1 → v1 → · · · → vk → A
and the other is A → B → C2 → v1 → · · · → vk → A. Since the lengths of these two closed
paths differ by one, the greatest common divisor of the lengths of all paths is 1, hence Gr0
is aperiodic.
Note that the statement is false if the number of vertices is 2: indeed, then there is only
one loop of length 2, hence the period is 2.
Now, let us consider a general graph product G(Λ). By the previous section, by replacing
vertices of Gr with graphs which recognize the geodesic combings of vertex group, we get a
new graph ΓG which gives a geodesic combing for G(Λ). By the previous Lemma we get:
Corollary 11.4. If Λ is anticonnected and has at least three vertices, then the graph Γ0G =
ΓG \ {v0 } is irreducible and aperiodic.
Proof of Theorem 11.2. Since the graph Γ0G is irreducible and aperiodic, then by the PerronFrobenius theorem its adjacency matrix A has a unique eigenvalue λ > 1 of maximum
modulus, and that eigenvalue is real, positive, and simple. Moreover, the coordinates of
n
the corresponding eigenvector are all positive. Finally, the sequence A
λn converges to the
projection to the eigenspace. In particular, none of the basis vectors is orthogonal to the
eigenvector, hence for any i, j there exists cij > 0 such that
(An )ij
= cij .
n→∞
λn
lim
Now, each path of length n from the initial vertex starts with an edge to the irreducible
graph, hence
X #Sn−1 (vi )
X X (An−1 )ij
X X cij
#Sn
=
→
→
=c>0
λn
λn
λn
λ
v →v
v →v j
v →v j
0
i
0
i
which establishes exact exponential growth.
0
i
36
I. GEKHTMAN, S.J. TAYLOR, AND G. TIOZZO
References
[AC16]
Yago Antolin and Laura Ciobanu, Finite generating sets of relatively hyperbolic groups and
applications to geodesic languages, Trans. Amer. Math. Soc. 368 (2016), no. 11, 7965–8010.
[ACT15] G.N. Arzhantseva, C. Cashen, and J. Tao, Growth tight actions, Pacific J. Math. 278 (2015),
no. 1, 1–49.
[ADT]
Tarik Aougab, Matthew Gentry Durham, and Samuel J. Taylor, Pulling back stability with applications to Out(F) and relatively hyperbolic groups, J. Lond. Math. Soc.
[AL02]
Goulnara N Arzhantseva and IG Lysenok, Growth tightness for word hyperbolic groups, Mathematische Zeitschrift 241 (2002), no. 3, 597–611.
[AO96]
G. Arzhantseva and A. Olshanskii, Generality of the class of groups in which subgroups with a
lesser number of generators are free, Mat. Zametki 59 (1996), 489–496.
[AP14]
J. Athreya and A. Prasad, Growth in right-angled groups and monoids, available at
arXiv:1409.4142, 2014.
[Arz98]
GN Arzhantseva, Generic properties of finitely presented groups and Howson’s theorem, Communications in Algebra 26 (1998), no. 11, 3783–3792.
[BF14]
Mladen Bestvina and Mark Feighn, Hyperbolicity of the complex of free factors, Adv. Math. 256
(2014), 104–155. MR 3177291
[BH09]
Martin R. Bridson and Aandre Haefliger, Metric spaces of non-positive curvature, vol. 319,
Springer, 2009.
[BHP93] Young-Gheel Baik, Jim Howie, and SJ Pride, The identity problem for graph products of groups,
Journal of Algebra 162 (1993), no. 1, 168–177.
[BMR03] Alexandre V Borovik, Alexei G Myasnikov, and Vladimir N Remeslennikov, Multiplicative measures on free groups, Internat. J. Algebra Comput. 13 (2003), no. 06, 705–731.
[Bow08] Brian H. Bowditch, Tight geodesics in the curve complex, Invent. Math. 171 (2008), no. 2,
281–300.
[Bow12] Brian H Bowditch, Relatively hyperbolic groups, International Journal of Algebra and Computation 22 (2012), no. 03, 1250016.
[Cal13]
Danny Calegari, The ergodic theory of hyperbolic groups, Geometry and topology down under,
Contemp. Math 597 (2013), 15–52.
[Can84]
James W Cannon, The combinatorial structure of cocompact discrete hyperbolic groups, Geometriae Dedicata 16 (1984), no. 2, 123–148.
[CF10]
Danny Calegari and Koji Fujiwara, Combable functions, quasimorphisms, and the central limit
theorem, Ergodic Theory and Dynamical Systems 30 (2010), no. 05, 1343–1369.
[Cha95]
Christophe Champetier, Propriétés statistiques des groupes de présentation finie, Adv. Math.
116 (1995), no. 2, 197–262.
[Chi94]
Ian M Chiswell, The growth series of a graph product, Bulletin of the London Mathematical
Society 26 (1994), no. 3, 268–272.
[CM07]
Chris Connell and Roman Muchnik, Harmonicity of quasiconformal measures and Poisson
boundaries of hyperbolic spaces, Geom. Funct. Anal. 17 (2007), no. 3, 707–769.
[CM15]
Danny Calegari and Joseph Maher, Statistics and compression of scl, Ergodic Theory and Dynamical Systems 35 (2015), no. 01, 64–110.
[DMS10] Cornelia Druţu, Shahar Mozes, and Mark Sapir, Divergence in lattices in semisimple Lie groups
and graphs of groups, Transactions of the American Mathematical Society 362 (2010), no. 5,
2451–2505.
[DSU14] Tushar Das, David Simmons, and Mariusz Urbański, Geometry and dynamics in Gromov hyperbolic metric spaces: With an emphasis on non-proper settings, arXiv preprint arXiv:1409.2155
(2014).
[DT06]
Nathan Dunfield and William P Thurston, Finite covers of random 3-manifolds, Invent. Math.
166 (2006), no. 3, 457–521.
[EPC+ 92] David Epstein, Mike S Paterson, James W Cannon, Derek F Holt, Silvio V Levy, and William P
Thurston, Word processing in groups, AK Peters, Ltd., 1992.
[Far98]
Benson Farb, Relatively hyperbolic groups, Geom. Funct. Anal. 8 (1998), no. 5, 810–840.
[Fur02]
A. Furman, Coarse-geometric perspective on negatively curved manifolds and groups, Rigidity in
Dynamics and Geometry (M. Burger and A. Iozzi, eds.), Springer, 2002, pp. 149–166.
[GdlH90] É. Ghys and P. de la Harpe (eds.), Sur les groupes hyperboliques d’après Mikhael Gromov,
Progress in Mathematics, vol. 83, Birkhäuser Boston, Inc., Boston, MA, 1990, Papers from the
Swiss Seminar on Hyperbolic Groups held in Bern, 1988. MR 1086648 (92f:53050)
COUNTING LOXODROMICS
[GM08]
[GMM15]
[Gre90]
[Gro87]
[Gro93]
[Gro03]
[GTT16]
[Hag08]
[Hag14]
[HM95]
[HM13]
[HRR17]
[Hru10]
[HW+ 99]
[Kai94]
[KB02]
[KK14]
[KMSS03]
[KRSS07]
[KS05]
[Mah11]
[Mei95]
[Mei96]
[MM99]
[MS14]
[MT14]
[Neu54]
[NS95]
[Ol’92]
[Osi06]
37
Daniel Groves and Jason Fox Manning, Dehn filling in relatively hyperbolic groups, Israel Journal
of Mathematics 168 (2008), no. 1, 317–429.
Sébastien Gouëzel, Frédéric Mathéus, and François Maucourant, Entropy and drift in word hyperbolic groups, arXiv preprint arXiv:1501.05082 (2015).
Elisabeth Ruth Green, Graph products of groups, Ph.D. thesis, University of Leeds, 1990.
Mikhael Gromov, Hyperbolic groups, Springer, 1987.
M Gromov, Asymptotic invariants of infinite groups, Geometric group theory, vol. 2, vol. 182,
Cambridge University Press, 1993.
Mikhail Gromov, Random walk in random groups, Geometric and Functional Analysis 13 (2003),
no. 1, 73–146.
I. Gekhtman, S. Taylor, and G. Tiozzo, Counting loxodromics for hyperbolic actions, available
at arXiv:1605.02103, 2016.
F. Haglund, Finite index subgroups of graph products, Geometriae Dedicata 135 (2008), no. 1,
167–209.
Mark F Hagen, Weak hyperbolicity of cube complexes and quasi-arboreal groups, Journal of
Topology 7 (2014), no. 2, 385–418.
S. Hermiller and J. Meier, Algorithms and geometry for graph products of groups, Journal of
Algebra 171 (1995), no. 1, 230–257.
Michael Handel and Lee Mosher, The free splitting complex of a free group, i hyperbolicity,
Geometry & Topology 17 (2013), no. 3, 1581–1670.
Derek F. Holt, Sarah Rees, and Claas E. Röver, Groups, languages and automata, London
Mathematical Society Student Texts, vol. 88, Cambridge University Press, Cambridge, 2017.
G. Hruska, Relative hyperbolicity and relative quasiconvexity for countable groups, Algebr. Geom.
Topol 10 (2010), 1807–1856.
Tim Hsu, Daniel T Wise, et al., On linear and residual properties of graph products., The
Michigan Mathematical Journal 46 (1999), no. 2, 251–259.
Vadim Kaimanovich, Ergodicity of harmonic invariant measures for the geodesic flow on hyperbolic spaces, J. Reine Angew. Math. 455 (1994), 57–104.
Ilya Kapovich and Nadia Benakli, Boundaries of hyperbolic groups, Combinatorial and geometric
group theory (New York, 2000/Hoboken, NJ, 2001) 296 (2002), 39–93.
Sang-Hyun Kim and Thomas Koberda, The geometry of the curve graph of a right-angled Artin
group, International Journal of Algebra and Computation 24 (2014), no. 02, 121–169.
Ilya Kapovich, Alexei Myasnikov, Paul Schupp, and Vladimir Shpilrain, Generic-case complexity,
decision problems in group theory, and random walks, J. Algebra 264 (2003), no. 2, 665–694.
Ilya Kapovich, Igor Rivin, Paul Schupp, and Vladimir Shpilrain, Densities in free groups and
Zn , visible points and test elements, Math. Res. Lett 14 (2007), no. 2.
Ilya Kapovich and Paul Schupp, Genericity, the Arzhantseva-Ol’shanskii method and the isomorphism problem for one-relator groups, Math. Ann. 331 (2005), no. 1, 1–19.
Joseph Maher, Random walks on the mapping class group, Duke Math. J. 156 (2011), no. 3,
429–468.
Holger Meinert, The bieri-neumann-strebel invariant for graph products of groups, Journal of
Pure and Applied Algebra 103 (1995), no. 2, 205–210.
John Meier, When is the graph product of hyperbolic groups hyperbolic?, Geometriae Dedicata
61 (1996), no. 1, 29–41.
Howard A. Masur and Yair N. Minsky, Geometry of the complex of curves. I. Hyperbolicity,
Invent. Math. 138 (1999), no. 1, 103–149.
P Mathieu and A Sisto, Deviation inequalities for random walks, arXiv preprint arXiv:1411.7865
(2014).
Joseph Maher and Giulio Tiozzo, Random walks on weakly hyperbolic groups, To appear in J.
Reine Angew. Math. (2014).
Bernhard H. Neumann, Groups covered by finitely many cosets, Publ. Math. Debrecen 3 (1954).
Walter Neumann and Michael Shapiro, Automatic structures, rational growth and geometrically
finite hyperbolic groups, Invent. Math. 120 (1995), 259–287.
A Yu Ol’shanskii, Almost every group is hyperbolic, Internat. J. Algebra Comput. 2 (1992),
no. 01, 1–17.
Denis V Osin, Relatively hyperbolic groups: intrinsic geometry, algebraic properties, and algorithmic problems, vol. 843, American Mathematical Soc., 2006.
38
I. GEKHTMAN, S.J. TAYLOR, AND G. TIOZZO
[Osi15]
[Rad03]
[Riv08]
[Sis11]
[TT15]
[Wie14]
[Yan13]
[Yan16]
Denis Osin, Acylindrically hyperbolic groups, Trans. Amer. Math. Soc. (2015).
David G Radcliffe, Rigidity of graph products of groups, Algebraic & Geometric Topology 3
(2003), no. 2, 1079–1088.
Igor Rivin, Walks on groups, counting reducible matrices, polynomials, and surface and free
group automorphisms, Duke Math. J. 142 (2008), no. 2, 353–379.
Alessandro Sisto, Contracting elements and random walks, J. Reine Angew. Math. (2011).
Samuel J. Taylor and Giulio Tiozzo, Random extensions of free groups and surface groups are
hyperbolic, Int. Math. Res. Not. (2015).
Bert Wiest, On the genericity of loxodromic actions, arXiv preprint arXiv:1406.7041 (2014).
Wenyuan Yang, Patterson-Sullivan measures and growth of relatively hyperbolic groups, available
at arXiv:1308.6326, 2013.
, Statistically convex-cocompact actions of groups with contracting elements, available at
arXiv:1612.03648, 2016.
Department of Mathematics, Yale University, 10 Hillhouse Ave, New Haven, CT 06520, U.S.A,
E-mail address: [email protected]
Department of Mathematics, Temple University, 1805 North Broad Street Philadelphia, PA
19122, U.S.A,
E-mail address: [email protected]
Department of Mathematics, University of Toronto, 40 St George St, Toronto, ON, Canada,
E-mail address: [email protected]
| 4math.GR
|
E XPLORING THE S PACE OF B LACK - BOX ATTACKS
D EEP N EURAL N ETWORKS
arXiv:1712.09491v1 [cs.LG] 27 Dec 2017
Arjun Nitin Bhagoji ∗
Department of Electrical Engineering
Princeton University
ON
Warren He, Bo Li & Dawn Song
EECS Department
University of California, Berkeley
A BSTRACT
Existing black-box attacks on deep neural networks (DNNs) so far have largely
focused on transferability, where an adversarial instance generated for a locally
trained model can “transfer” to attack other learning models. In this paper, we
propose novel Gradient Estimation black-box attacks for adversaries with query
access to the target model’s class probabilities, which do not rely on transferability.
We also propose strategies to decouple the number of queries required to generate
each adversarial sample from the dimensionality of the input. An iterative variant
of our attack achieves close to 100% adversarial success rates for both targeted and
untargeted attacks on DNNs. We carry out extensive experiments for a thorough
comparative evaluation of black-box attacks and show that the proposed Gradient
Estimation attacks outperform all transferability based black-box attacks we tested
on both MNIST and CIFAR-10 datasets, achieving adversarial success rates similar
to well known, state-of-the-art white-box attacks. We also apply the Gradient
Estimation attacks successfully against a real-world Content Moderation classifier
hosted by Clarifai. Furthermore, we evaluate black-box attacks against state-of-theart defenses. We show that the Gradient Estimation attacks are very effective even
against these defenses.
1
I NTRODUCTION
The ubiquity of machine learning provides adversaries with both opportunities and incentives to
develop strategic approaches to fool learning systems and achieve their malicious goals. Many attack
strategies devised so far to generate adversarial examples to fool learning systems have been in the
white-box setting, where adversaries are assumed to have access to the learning model (Szegedy
et al. (2014); Goodfellow et al. (2015); Carlini & Wagner (2017); Moosavi-Dezfooli et al. (2015)).
However, in many realistic settings, adversaries may only have black-box access to the model, i.e.
they have no knowledge about the details of the learning system such as its parameters, but they may
have query access to the model’s predictions on input samples, including class probabilities. For
example, we find this to be the case in some popular commercial AI offerings, such as those from
IBM, Google and Clarifai. With access to query outputs such as class probabilities, the training loss
of the target model can be found, but without access to the entire model, the adversary cannot access
the gradients required to carry out white-box attacks.
Most existing black-box attacks on DNNs have focused on transferability based attacks (Papernot
et al. (2016); Moosavi-Dezfooli et al. (2016); Papernot et al. (2017)), where adversarial examples
crafted for a local surrogate model can be used to attack the target model to which the adversary has
no direct access. The exploration of other black-box attack strategies is thus somewhat lacking so
far in the literature. In this paper, we design powerful new black-box attacks using limited query
access to learning systems which achieve adversarial success rates close to that of white-box attacks.
These black-box attacks help us understand the extent of the threat posed to deployed systems by
adversarial samples. The code to reproduce our results can be found at https://github.com/
sunblaze-ucb/blackbox-attacks.
New black-box attacks. We propose novel Gradient Estimation attacks on DNNs, where the
adversary is only assumed to have query access to the target model. These attacks do not need any
∗
Work done while visiting UC Berkeley
1
Figure 1: Sample adversarial images of Gradient Estimation attacks on Clarifai’s Content Moderation
model. Left: original image, classified as ‘drug’ with a confidence of 0.99. Right: adversarial sample
with = 32, classified as ‘safe’ with a confidence of 0.96.
access to a representative dataset or any knowledge of the target model architecture. In the Gradient
Estimation attacks, the adversary adds perturbations proportional to the estimated gradient, instead
of the true gradient as in white-box attacks (Goodfellow et al. (2015); Kurakin et al. (2016)). Since
the direct Gradient Estimation attack requires a number of queries on the order of the dimension
of the input, we explore strategies for reducing the number of queries to the target model. We also
experimented with Simultaneous Perturbation Stochastic Approximation (SPSA) and Particle Swarm
Optimization (PSO) as alternative methods to carry out query-based black-box attacks but found
Gradient Estimation to work the best.
Query-reduction strategies We propose two strategies: random feature grouping and principal
component analysis (PCA) based query reduction. In our experiments with the Gradient Estimation
attacks on state-of-the-art models on MNIST (784 dimensions) and CIFAR-10 (3072 dimensions)
datasets, we find that they match white-box attack performance, achieving attack success rates up
to 90% for single-step attacks in the untargeted case and up to 100% for iterative attacks in both
targeted and untargeted cases. We achieve this performance with just 200 to 800 queries per sample
for single-step attacks and around 8,000 queries for iterative attacks. This is much fewer than the
closest related attack by Chen et al. (2017). While they achieve similar success rates as our attack,
the running time of their attack is up to 160× longer for each adversarial sample (see Section 3.5). A
further advantage of the Gradient Estimation attack is that it does not require the adversary to train a
local model, which could be an expensive and complex process for real-world datasets, in addition to
the fact that training such a local model may require even more queries based on the training data.
Attacking real-world systems. To demonstrate the effectiveness of our Gradient Estimation attacks
in the real world, we also carry out a practical black-box attack using these methods against the Not
Safe For Work (NSFW) classification and Content Moderation models developed by Clarifai, which
we choose due to their socially relevant application. These models have begun to be deployed for
real-world moderation (Liu, 2016), which makes such black-box attacks especially pernicious. We
carry out these attacks with no knowledge of the training set. We have demonstrated successful attacks
(Figure 1) with just around 200 queries per image, taking around a minute per image. In Figure 1, the
target model classifies the adversarial image as ‘safe’ with high confidence, in spite of the content that
had to be moderated still being clearly visible. We note here that due to the nature of the images we
experiment with, we only show one example here, as the others may be offensive to readers. The full set
of images can be found at https://sunblaze-ucb.github.io/blackbox-attacks/.
Comparative evaluation of black-box attacks. We carry out a thorough empirical comparison of
various black-box attacks (given in Table 7) on both MNIST and CIFAR-10 datasets. We study
attacks that require zero queries to the learning model, including the addition of perturbations that
are either random or proportional to the difference of means of the original and targeted classes,
as well as various transferability based black-box attacks. We show that the proposed Gradient
Estimation attacks outperform other black-box attacks in terms of attack success rate and achieve
results comparable with white-box attacks.
In addition, we also evaluate the effectiveness of these attacks on DNNs made more robust using
adversarial training (Goodfellow et al., 2015; Szegedy et al., 2014) and its recent variants including
ensemble adversarial training (Tramèr et al., 2017a) and iterative adversarial training (Madry
˛ et al.,
2017). We find that although standard and ensemble adversarial training confer some robustness
against single-step attacks, they are vulnerable to iterative Gradient Estimation attacks, with adversarial success rates in excess of 70% for both targeted and untargeted attacks. We find that our
2
methods outperform other black-box attacks and achieve performance comparable to white-box
attacks. Iterative adversarial training is quite robust against all black-box attacks we test.
In summary, our contributions include:
1. We propose new Gradient Estimation black-box attacks using queries to the target model.
We also investigate two methods to make the number of queries needed independent of the
input dimensionality.
2. We conduct a thorough evaluation of 10 different black-box attack strategies on state-ofthe-art classifiers on the MNIST and CIFAR-10 datasets, and we find that with a small
number of queries, our proposed Gradient Estimation attacks outperform transferability
based attacks and achieve attack success rates matching those of white-box attacks.
3. We carry out practical black-box attacks on Clarifai’s Not Safe For Work (NSFW) classification and Content Moderation models through public APIs, and show that the generated
adversarial examples can mislead these models with high confidence.
4. Finally, we evaluate these black-box attacks against state-of-the-art adversarial training
based defenses on DNNs, and we find that both standard and ensemble adversarial training
are not robust against Gradient Estimation attacks. Further, even against iterative adversarial
training our methods outperform transferability based attacks.
Related Work. Existing black-box attacks that do not use a local model were first proposed for
convex inducing two-class classifiers by Nelson et al. (2012). For malware data, Xu et al. (2016) use
genetic algorithms to craft adversarial samples, while Dang et al. (2017) use hill climbing algorithms.
These methods are prohibitively expensive for non-categorical and high-dimensional data such as
images. Papernot et al. (2017) proposed using queries to a target model to train a local surrogate
model, which was then used to to generate adversarial samples. This attack relies on transferability.
To the best of our knowledge, the only previous literature on query-based black-box attacks in the
deep learning setting is independent work by Narodytska & Kasiviswanathan (2016) and Chen et al.
(2017).
Narodytska & Kasiviswanathan (2016) propose a greedy local search to generate adversarial samples
by perturbing randomly chosen pixels and using those which have a large impact on the output
probabilities. Their method uses 500 queries per iteration, and the greedy local search is run for
around 150 iterations for each image, resulting in a total of 75,000 queries per image, which is
much higher than any of our attacks. Further, we find that our methods achieve higher targeted and
untargeted attack success rates on both MNIST and CIFAR-10 as compared to their method. Chen
et al. (2017) propose a black-box attack method named ZOO, which also uses the method of finite
differences to estimate the derivative of a function. However, while we propose attacks that compute
an adversarial perturbation, approximating FGSM and iterative FGS; ZOO approximates the Adam
optimizer, while trying to perform coordinate descent on the loss function proposed by Carlini &
Wagner (2017). Neither of these works demonstrates the effectiveness of their attacks on real-world
systems or on state-of-the-art defenses.
2
BACKGROUND AND E VALUATION SETUP
In this section, we will first introduce the notation we use throughout the paper and then describe
the evaluation setup and metrics used in the remainder of the paper. The full set of attacks that was
evaluated is given in Table 7 in Appendix C, which also provides a taxonomy for black-box attacks.
2.1
N OTATION
A classifier f (·; θ) : X → Y is a function mapping from the domain X to the set of classification
outputs Y. (Y = {0, 1} in the case of binary classification, i.e. Y is the set of class labels.) The
number of possible classification outputs is then |Y|. θ is the set of parameters associated with a
classifier. Throughout, the target classifier is denoted as f (·; θ), but the dependence on θ is dropped
if it is clear from the context. H denotes the constraint set which an adversarial sample must satisfy.
`f (x, y) is used to represent the loss function for the classifier f with respect to inputs x ∈ X and
their true labels y ∈ Y. The loss functions we use are the standard cross-entropy loss (denoted
3
xent) and the logit loss from Carlini & Wagner (2017) (denoted as logit). These are described in
Section 3.1.
An adversary can generate adversarial example xadv from a benign sample x by adding an appropriate
perturbation of small magnitude (Szegedy et al., 2014). Such an adversarial example xadv will either
cause the classifier to misclassify it into a targeted class (targeted attack), or any class other than the
ground truth class (untargeted attack).
Since the black-box attacks we analyze focus on neural networks in particular, we also define some
notation specifically for neural networks. The outputs of the penultimate layer of a neural network
f , representing the output of the network computed sequentially over all preceding layers, are
known as the logits. We represent the logits as a vector φf (x) ∈ R|Y| . The final layer of a neural
network f used for classification is usually a softmax layer represented as a vector of probabilities
f
P|Y|
eφi (x)
.
pf (x) = [pf1 (x), . . . , pf|Y| (x)], with i=1 pfi (x) = 1 and pfi (x) = P|Y|
f
φ (x)
j=1
2.2
e
j
E VALUATION SETUP
The empirical evaluation carried out in Sections 6 and 3 is on state-of-the-art neural networks on the
MNIST (LeCun & Cortes, 1998) and CIFAR-10 (Krizhevsky & Hinton, 2009) datasets. The details
of the datasets and the architecture and training procedure for all models are given below.
2.2.1
DATASETS
MNIST. This is a dataset of images of handwritten digits (LeCun & Cortes, 1998). There are 60,000
training examples and 10,000 test examples. Each image belongs to a single class from 0 to 9. The
images have a dimension d of 28 × 28 pixels (total of 784) and are grayscale. Each pixel value
lies in [0, 1]. The digits are size-normalized and centered. This dataset is used commonly as a
‘sanity-check’ or first-level benchmark for state-of-the-art classifiers. We use this dataset since it has
been extensively studied from the attack perspective by previous work.
CIFAR-10. This is a dataset of color images from 10 classes (Krizhevsky & Hinton, 2009). The
images belong to 10 mutually exclusive classes (airplane, automobile, bird, cat, deer, dog, frog, horse,
ship, and truck). There are 50,000 training examples and 10,000 test examples. There are exactly
6,000 examples in each class. The images have a dimension of 32 × 32 pixels (total of 1024) and
have 3 channels (Red, Green, and Blue). Each pixel value lies in [0, 255].
2.2.2
M ODEL TRAINING DETAILS
In this section, we present the architectures and training details for both the normally and adversarially
trained variants of the models on both the MNIST and CIFAR-10 datasets. The accuracy of each
model on benign data is given in Table 4.
MNIST. Each pixel of the MNIST image data is scaled to [0, 1]. We trained four different models on
the MNIST dataset, denoted Models A to D, which are used by Tramèr et al. (2017a) and represent
a good variety of architectures. For the attacks constrained with the L∞ distance, we vary the
adversary’s perturbation budget from 0 to 0.4, since at a perturbation budget of 0.5, any image can
be made solid gray. The model details for the 4 models trained on the MNIST dataset are as follows:
1. Model A (3,382,346 parameters): Conv(64, 5, 5) + Relu, Conv(64, 5, 5) + Relu,
Dropout(0.25), FC(128) + Relu, Dropout(0.5), FC + Softmax
2. Model B (710,218 parameters) - Dropout(0.2), Conv(64, 8, 8) + Relu, Conv(128, 6, 6) +
Relu, Conv(128, 5, 5) + Relu, Dropout(0.5), FC + Softmax
3. Model C (4,795,082 parameters) - Conv(128, 3, 3) + Relu, Conv(64, 3, 3) + Relu,
Dropout(0.25), FC(128) + Relu, Dropout(0.5), FC + Softmax
4. Model D (509,410 parameters) - [FC(300) + Relu, Dropout(0.5)] × 4, FC + Softmax
Models A and C have both convolutional layers as well as fully connected layers. They also have the
same order of magnitude of parameters. Model B, on the other hand, does not have fully connected
layers and has an order of magnitude fewer parameters. Similarly, Model D has no convolutional
4
layers and has fewer parameters than all the other models. Models A, B, and C all achieve greater
than 99% classification accuracy on the test data. Model D achieves 97.2% classification accuracy,
due to the lack of convolutional layers.
CIFAR-10. Each pixel of the CIFAR-10 image data is in [0, 255]. We choose three model architectures for this dataset, which we denote as Resnet-32, Resnet-28-10 (ResNet variants (He et al.,
2016; Zagoruyko & Komodakis, 2016)), and . For the attacks constrained with the L∞ distance, we
vary the adversary’s perturbation budget from 0 to 28.
As their name indicates, Resnet-32 and Resnet-28-10 are ResNet variants (He et al., 2016;
Zagoruyko & Komodakis, 2016), while Std.-CNN is a standard CNN (TensorFlow Authors, b). In
particular, Resnet-32 is a standard 32 layer ResNet with no width expansion, and Resnet-28-10 is
a wide ResNet with 28 layers with the width set to 10, based on the best performing ResNet from
Zagoruyko & Komodakis (TensorFlow Authors, a). The width indicates the multiplicative factor by
which the number of filters in each residual layer is increased. Std.-CNN is a standard CNN1 from
Tensorflow (Abadi et al., 2015) with two convolutional layers, each followed by a max-pooling and
normalization layer and two fully connected layers, each of which has weight decay.
Resnet-32 is trained for 125,000 steps, Resnet-28-10 is trained for 167,000 steps and Std.-CNN is
trained for 100,000 steps on the benign training data. Models Resnet-32 and Resnet-28-10 are
much more accurate than Std.-CNN. All models were trained with a batch size of 128. The two
ResNets achieve close to state-of-the-art accuracy (Benenson, 2016) on the CIFAR-10 test set, with
Resnet-32 at 92.4% and Resnet-28-10 at 94.4%. Std.-CNN, on the other hand, only achieves an
accuracy of 81.4%, reflecting its simple architecture and the complexity of the task.
2.3
M ETRICS
Throughout the paper, we use standard metrics to characterize the effectiveness of various attack
strategies. For MNIST, all metrics for single-step attacks are computed with respect to the test set
consisting of 10,000 samples, while metrics for iterative attacks are computed with respect to the first
1,000 samples from the test set. For the CIFAR-10 data, we choose 1,000 random samples from the
test set for single-step attacks and a 100 random samples for iterative attacks. In our evaluations of
targeted attacks, we choose target T for each sample uniformly at random from the set of classification
outputs, except the true class y of that sample.
Attack success rate. The main metric, the attack success rate, is the fraction of samples that meets
the adversary’s goal: f (xadv ) 6= y for untargeted attacks and f (xadv ) = T for targeted attacks with
target T (Szegedy et al., 2014; Tramèr et al., 2017a). Alternative evaluation metrics are discussed in
Appendix A.
Average distortion. We also evaluate the average distortion for adversarial examples using average
L2 distance between the benign samples and the adversarial ones as suggested by Gu & Rigazio
PN
(2014): ∆(Xadv , X) = N1 i=1 k(Xadv )i − (X)i k2 where N is the number of samples. This metric
allows us to compare the average distortion for attacks which achieve similar attack success rates,
and therefore infer which one is stealthier.
Number of queries. Query based black-box attacks make queries to the target model, and this
metric may affect the cost of mounting the attack. This is an important consideration when attacking
real-world systems which have costs associated with the number of queries made.
3
Q UERY BASED ATTACKS : G RADIENT E STIMATION ATTACK
Deployed learning systems often provide feedback for input samples provided by the user. Given query
feedback, different adaptive, query-based algorithms can be applied by adversaries to understand
the system and iteratively generate effective adversarial examples to attack it. Formal definitions of
query-based attacks are in Appendix B. We initially explored a number of methods of using query
feedback to carry out black-box attacks including Particle Swarm Optimization (Kennedy, 2011) and
Simultaneous Perturbation Stochastic Approximation (Spall, 1992). However, these methods were
1
https://github.com/tensorflow/models/tree/master/tutorials/image/
cifar10
5
not effective at finding adversarial examples for reasons detailed in Section 3.6, which also contains
the results obtained.
Given the fact that many white-box attacks for generating adversarial examples are based on gradient
information, we then tried directly estimating the gradient to carry out black-box attacks, and found
it to be very effective in a range of conditions. In other words, the adversary can approximate
white-box Single-step and Iterative FGSM attacks (Goodfellow et al., 2015; Kurakin et al., 2016)
using estimates of the losses that are needed to carry out those attacks. We first propose a Gradient
Estimation black-box attack based on the method of finite differences (Spall, 2005). The drawback
of a naive implementation of the finite difference method, however, is that it requires O(d) queries
per input, where d is the dimension of the input. This leads us to explore methods such as random
grouping of features and feature combination using components obtained from Principal Component
Analysis (PCA) to reduce the number of queries.
Threat model and justification: We assume that the adversary can obtain the vector of output
probabilities for any input x. The set of queries the adversary can make is then Qf = {pf (x), ∀x}.
Note that an adversary with access to the softmax probabilities will be able to recover the logits up to
an additive constant, by taking the logarithm of the softmax probabilities. For untargeted attacks, the
adversary only needs access to the output probabilities for the two most likely classes.
A compelling reason for assuming this threat model for the adversary is that many existing cloudbased ML services allow users to query trained models (Watson Visual Recognition, Clarifai, Google
Vision API). The results of these queries are confidence scores which can be used to carry out
Gradient Estimation attacks. These trained models are often deployed by the clients of these ML as a
service (MLaaS) providers (Liu (2016)). Thus, an adversary can pose as a user for a MLaaS provider
and create adversarial examples using our attack, which can then be used against any client of that
provider.
Comparing against existing black-box attacks: In the results presented in this section, we compare
our attacks against a number of existing black-box attacks in both ther targeted and untargeted case.
Detailed descriptions of these attacks are in Section 6. In particular, we compare against the following
attacks that make zero queries to the target model:
1. Baseline attacks: Random-Gaussian perturbations denoted as Rand. (Section 6.1.1) and
Difference-of-Means aligned perturbations denoted as D. of M. (Section 6.1.2)
2. Transferability attack (single local model) (Section 6.3.1) using Fast Gradient Sign (FGS)
and Iterative FGS (IFGS) samples generated on a single source model for both loss functions.
This is denoted as Transfer model FGS/IFGS-loss); e.g., Transfer Model A FGS-logit
3. Transferability attack (local model ensemble) (Section 6.3.2) using FGS and IFGS samples
generated on a source model for both loss functions (Transfer models FGS/IFGS-loss);
e.g., Transfer Model B, Model C IFGS-logit
We also compare against white-box attacks, descriptions of which are in Section 6.2 and results are in
Appendix D.
3.1
F INITE DIFFERENCE METHOD FOR GRADIENT ESTIMATION
In this section, we focus on the method of finite differences to carry out Gradient Estimation based
attacks. Let the function whose gradient is being estimated be g(x). The input to the function is a
d-dimensional vector x, whose elements are represented as xi , where i ∈ [1, . . . , d]. The canonical
basis vectors are represented as ei , where ei is 1 only in the ith component and 0 everywhere else.
Then, a two-sided estimation of the gradient of g with respect to x is given by
g(x+δe1 )−g(x−δe1 )
2δ
..
.
FDx (g(x), δ) =
.
(1)
g(x+δed )−g(x−δed )
2δ
δ is a free parameter that controls the accuracy of the estimation. A one-sided approximation can
also be used, but will be less accurate (Wright & Nocedal, 1999). If the gradient of the function g
exists, then limδ→0 FDx (g(x), δ) = ∇x g(x). The finite difference method is useful for a black-box
6
adversary aiming to approximate a gradient based attack, since the gradient can be directly estimated
with access to only the function values.
3.1.1
A PPROXIMATE FGS WITH FINITE DIFFERENCES
In the untargeted FGS method, the gradient is usually taken with respect to the cross-entropy loss
between the true label of the input and the softmax probability vector. The cross-entropy loss of a
P|Y|
network f at an input x is then `f (x, y) = − j=1 1[j = y] log pfj (x) = − log pfy (x), where y is
the index of the original class of the input. The gradient of `f (x, y) is
∇x `f (x, y) = −
∇x pfy (x)
pfy (x)
.
(2)
An adversary with query access to the softmax probabilities then just has to estimate the gradient of
pfy (x) and plug it into Eq. 2 to get the estimated gradient of the loss. The adversarial sample thus
generated is
!
FDx (pfy (x), δ)
xadv = x + · sign
.
(3)
pfy (x)
This method of generating adversarial samples is denoted as FD-xent. Targeted black-box adversarial
samples generated using the Gradient Estimation method are then
!
FDx (pfT (x), δ)
.
(4)
xadv = x − · sign
pfT (x)
The targeted version of the method is denoted as FD-xent-T.
3.1.2
E STIMATING THE LOGIT- BASED LOSS
We also use a loss function based on logits which was found to work well for white-box attacks by
Carlini & Wagner (2017). The loss function is given by
`(x, y) = max(φ(x + δ)y − max{φ(x + δ)i : i 6= y}, −κ),
(5)
where y represents the ground truth label for the benign sample x and φ(·) are the logits. κ is a
confidence parameter that can be adjusted to control the strength of the adversarial perturbation. If the
confidence parameter κ is set to 0, the logit loss is max(φ(x + δ)y − max{φ(x + δ)i : i 6= y}, 0).
For an input that is correctly classified, the first term is always greater than 0, and for an incorrectly
classified input, an untargeted attack is not meaningful to carry out. Thus, the loss term reduces to
φ(x + δ)y − max{φ(x + δ)i : i 6= y} for relevant inputs.
An adversary can compute the logit values up to an additive constant by taking the logarithm of the
softmax probabilities, which are assumed to be available in this threat model. Since the loss function
is equal to the difference of logits, the additive constant is canceled out. Then, the finite differences
method can be used to estimate the difference between the logit values for the original class y, and the
second most likely class y 0 , i.e., the one given by y 0 = argmaxi6=y φ(x)i . The untargeted adversarial
sample generated for this loss in the white-box case is xadv = x + · sign(∇x (φ(x)y0 − φ(x)y )).
Similarly, in the case of a black-box adversary with query-access to the softmax probabilities, the
adversarial sample is
xadv = x + · sign(FDx (φ(x)y0 − φ(x)y , δ)).
(6)
Similarly, a targeted adversarial sample is
xadv = x − · sign(FDx (max(φ(x)i : i 6= T ) − φ(x)T , δ)).
(7)
The untargeted attack method is denoted as FD-logit while the targeted version is denoted as FDlogit-T.
7
MNIST
Baseline
D. of M.
44.8 (5.6)
81.5 (5.6)
20.2 (5.6)
97.1 (5.6)
Model
A
B
C
D
CIFAR-10
Model
Resnet-32
Resnet-28-10
Std.-CNN
Baseline
D. of M.
9.3 (440.5)
6.7 (440.5)
20.3 (440.5)
Transfer from Model B
Gradient Estimation using Finite Differences
Rand.
8.5 (6.1)
7.8 (6.1)
4.1 (6.1)
38.5 (6.1)
Single-step
FD-xent
FD-logit
51.6 (3.3)
92.9 (6.1)
69.2 (4.5)
98.9 (6.3)
60.5 (3.8)
86.1 (6.2)
95.4 (5.8)
100.0 (6.1)
Iterative
IFD-xent
IFD-logit
75.0 (3.6)
100.0 (2.1)
86.7 (3.9)
100.0 (1.6)
80.2 (4.5)
100.0 (2.2)
98.4 (5.4)
100.0 (1.2)
Rand.
19.4 (439.4)
17.1 (439.4)
22.2 (439.4)
Iterative
IFGS-xent
IFGS-logit
89.8 (4.75)
88.5 (4.75)
79.5 (4.75)
78.7 (4.75)
73.3 (4.75)
71.4 (4.75)
Transfer from Resnet-28-10
Gradient Estimation using Finite Differences
Single-step
FD-xent
FD-logit
49.1 (217.1) 86.0 (410.3)
50.1 (214.8) 88.2 (421.6)
80.0 (341.3) 98.9 (360.9)
Single-step
FGS-xent
FGS-logit
66.3 (6.2)
80.8 (6.3)
49.5 (6.2)
57.0 (6.3)
76.3 (6.2)
87.6 (6.3)
Iterative
IFD-xent
IFD-logit
62.0 (149.9) 100.0 (65.7)
46.0 (120.4) 100.0 (74.9)
66.0 (202.5) 100.0 (79.9)
Single-step
FGS-xent
FGS-logit
74.5 (439.4) 76.6 (439.4)
37.4 (439.4) 37.7 (439.4)
Iterative
IFGS-xent
IFGS-logit
99.0 (275.4) 98.9 (275.6)
33.7 (275.4) 33.6 (275.6)
Table 1: Untargeted black-box attacks: Each entry has the attack success rate for the attack
method given in that column on the model in each row. The number in parentheses for each entry
is ∆(X, Xadv ), the average distortion over all samples used in the attack. In each row, the entry in
bold represents the black-box attack with the best performance on that model. Gradient Estimation
using Finite Differences is our method, which has performance matching white-box attacks. Above:
MNIST, L∞ constraint of = 0.3. Below: CIFAR-10, L∞ constraint of = 8.
MNIST
Baseline
Model
A
B
C
D
D. of M.
15.0 (5.6)
35.5 (5.6)
5.84 (5.6)
59.8 (5.6)
CIFAR-10
Baseline
Model
Resnet-32
Resnet-28-10
Std.-CNN
D. of M.
1.2 (440.3)
0.9 (440.3)
2.6 (440.3)
Gradient Estimation using Finite Differences
Single-step
FD-xent
FD-logit
30.0 (6.0)
29.9 (6.1)
29.5 (6.3)
29.3 (6.3)
34.1 (6.1)
33.8 (6.4)
61.4 (6.3)
60.8 (6.3)
Iterative
IFD-xent
IFD-logit
100.0 (4.2)
99.7 (2.7)
99.9 (4.1)
98.7 (2.4)
100.0 (4.3)
99.8 (3.0)
100.0 (3.7)
99.9 (1.9)
Transfer from Model B
Single-step
FGS-xent
FGS-logit
18.3 (6.3)
18.1 (6.3)
14.0 (6.3)
13.8 (6.3)
16.8 (6.3)
16.7 (6.3)
Transfer from Resnet-28-10
Gradient Estimation using Finite Differences
Single-step
FD-xent
FD-logit
23.8 (439.5) 23.0 (437.0)
29.2 (439.4) 28.0 (436.1)
44.5 (439.5) 40.3 (434.9)
Iterative
IFD-xent
IFD-logit
100.0 (110.9) 100.0 (89.5)
100.0 (123.2) 100.0 (98.3)
99.0 (178.8) 95.0 (126.8)
Iterative
IFGS-xent
IFGS-logit
54.5 (4.6)
46.5 (4.2)
34.0 (4.6)
26.1 (4.2)
36.4 (4.6)
32.8 (4.1)
Single-step
FGS-xent
FGS-logit
15.8 (439.4) 15.5 (439.4)
5.6 (439.4)
5.6 (439.4)
Iterative
IFGS-xent
IFGS-logit
71.8 (222.5) 80.3 (242.6)
5.1 (222.5)
5.9 (242.6)
Table 2: Targeted black-box attacks: adversarial success rates. The number in parentheses () for
each entry is ∆(X, Xadv ), the average distortion over all samples used in the attack. Above: MNIST,
= 0.3. Below: CIFAR-10, = 8.
3.1.3
I TERATIVE ATTACKS WITH ESTIMATED GRADIENTS
The iterative variant of the gradient based attack described in Section 6.2 is a powerful attack that
often achieves much higher attack success rates in the white-box setting than the simple single-step
gradient based attacks. Thus, it stands to reason that a version of the iterative attack with estimated
gradients will also perform better than the single-step attacks described until now. An iterative attack
with t + 1 iterations using the cross-entropy loss is:
!!
FDxtadv pfy (xtadv )
t+1
t
xadv = ΠH xadv + α · sign
,
(8)
pfy (xtadv )
where α is the step size and H is the constraint set for the adversarial sample. This attack is denoted
as IFD-xent. If the logit loss is used instead, it is denoted as IFD-logit.
3.1.4
E VALUATION OF G RADIENT E STIMATION USING F INITE D IFFERENCES
In this section, we summarize the results obtained using Gradient Estimation attacks with Finite
Differences and describe the parameter choices made.
FD-logit and IFD-logit match white-box attack adversarial success rates: The Gradient Estimation attack with Finite Differences (FD-logit) is the most successful untargeted single-step black-box
attack for MNIST and CIFAR-10 models. It significantly outperforms transferability-based attacks
(Table 1) and closely tracks white-box FGS with a logit loss (WB FGS-logit) on MNIST and CIFAR10 (Figure 2). For adversarial samples generated iteratively, the Iterative Gradient Estimation attack
with Finite Differences (IFD-logit) achieves 100% adversarial success rate across all models on both
datasets (Table 1). We used 0.3 for the value of for the MNIST dataset and 8 for the CIFAR-10
dataset. The average distortion for both FD-logit and IFD-logit closely matches their white-box
counterparts, FGS-logit and IFGS-logit as given in Table 8.
8
L∞ constrained strategies on Model A
L∞ constrained strategies on Resnet-32
100
Difference-of-means
Random-perturbation
Finite-difference xent
Finite-difference logit
Query-reduced PCA-100 logit
Transfer Model B FGS xent
Transfer Model B FGS logit
White-box FGS logit
White-box FGS xent
80
60
40
20
Adversarial success (%)
Adversarial success (%)
100
0
Difference-of-means
Random-perturbation
Finite-difference xent
Finite-difference logit
Query-reduced PCA-400 logit
Transfer Resnet-28-10 FGS xent
White-box FGS logit
White-box FGS xent
80
60
40
20
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
0
(a) Model A (MNIST)
4
8
12 16 20 24 28 32
(b) Resnet-32 (CIFAR-10)
Figure 2: Effectiveness of various single step black-box attacks on Model A (MNIST) and
Resnet-32 (CIFAR-10). The y-axis for both figures gives the variation in adversarial success as
is increased. The most successful black-box attack strategy in both cases is the Gradient Estimation
attack using Finite Differences with the logit loss (FD-logit), which coincides almost exactly with the
white-box FGS attack with the logit loss (WB FGS-logit). Also, the Gradient Estimation attack with
query reduction using PCA (GE-QR (PCA-k, logit)) performs well for both datasets as well.
FD-T and IFD-T achieve the highest adversarial success rates in the targeted setting: For targeted black-box attacks, IFD-xent-T achieves 100% adversarial success rates on almost all models as
shown by the results in Table 2. While FD-xent-T only achieves about 30% adversarial success rates,
this matches the performance of single-step white-box attacks such as FGS-xent-T and FGS-logit-T
(Table 9). The average distortion for samples generated using gradient estimation methods is similar
with that of white-box attacks.
Parameter choices: We use δ = 1.0 for FD-xent and IFD-xent for both datasets, while using
δ = 0.01 for FD-logit and IFD-logit. We find that a larger value of δ is needed for xent loss based
attacks to work. The reason for this is that the probability values used in the xent loss are not as
sensitive to changes as in the logit loss, and thus the gradient cannot be estimated since the function
value does not change at all when a single pixel is perturbed. For the Iterative Gradient Estimation
attacks using Finite Differences, we use α = 0.01 and t = 40 for the MNIST results and α = 1.0
and t = 10 for CIFAR-10 throughout. The same parameters are used for the white-box Iterative FGS
attack results given in Appendix D. This translates to 62720 queries for MNIST (40 steps of iteration)
and 61440 queries (10 steps of iteration) for CIFAR-10 per sample. We find these choices work well,
and keep the running time of the Gradient Estimation attacks at a manageable level. However, we find
that we can achieve similar adversarial success rates with much fewer queries using query reduction
methods which we describe in the next section.
3.2
Q UERY REDUCTION
The major drawback of the approximation based black-box attacks is that the number of queries
needed per adversarial sample is large. For an input with dimension d, the number of queries will be
exactly 2d for a two-sided approximation. This may be too large when the input is high-dimensional.
So we examine two techniques in order to reduce the number of queries the adversary has to make.
Both techniques involve estimating the gradient for groups of features, instead of estimating it one
feature at a time.
The justification for the use of feature grouping comes from the relation between gradients and
directional derivatives (Hildebrand, 1962) for differentiable functions. The directional derivative of a
function g is defined as ∇v g(x) = limh→0 g(x+hv)−g(x)
. It is a generalization of a partial derivative.
h
For differentiable functions, ∇v g(x) = ∇x g(x) · v, which implies that the directional derivative is
just the projection of the gradient along the direction v. Thus, estimating the gradient by grouping
features is equivalent to estimating an approximation of the gradient constructed by projecting it
ˆ x g(x) of any function g can be
along appropriately chosen directions. The estimated gradient ∇
computed using the techniques below, and then plugged in to Equations 3 and 6 instead of the finite
difference term to create an adversarial sample. Next, we introduce the techniques applied to group
the features for estimation.
9
3.2.1
Q UERY REDUCTION BASED ON RANDOM GROUPING
The simplest way to group features is to choose, without replacement, a random set of features. The
gradient can then be simultaneously estimated for all these features. If the size of the set chosen is k,
then the number of queries the adversary has to make is d kd e. When k = 1, this reduces to the case
where the partial derivative with respect to every feature is found, as in Section 3.1. In each iteration
of Algorithm 1, there is a set of indices S according to which v is determined, with vi = 1 if and
P
only if i ∈ S. Thus, the directional derivative being estimated is i∈S ∂g(x)
∂xi , which is an average of
partial derivatives. Thus, the quantity being estimated is not the gradient itself, but an index-wise
averaged version of it.
Algorithm 1 Gradient estimation with query reduction using random features
Input: x, k, δ, g(·)
ˆ x g(x) of g(·) at x
Output: Estimated gradient ∇
ˆ
1: Initialize empty vector ∇x g(x) of dimension d
2: for i ← 1 to d kd e − 1 do
i−1
3:
Choose a set of random k indices Si out of [1, . . . , d]/{∪j=1
Sj }
4:
Initialize v such that vj = 1 iff j ∈ Si
ˆ x g(x)j = g(x+δv)−g(x−δv) , which is the two-sided approximation of the directional
5:
For all j ∈ Si , set ∇
2δk
derivative along v
6: end for
d e−1
dk
7: Initialize v such that vj = 1 iff j ∈ [1, . . . , d]/{∪j=1
Sj }
d d e−1
k
8: For all j ∈ [1, . . . , d]/{∪j=1
3.2.2
ˆ x g(x)j =
Sj }, set ∇
g(x+δv)−g(x−δv)
2δk
Q UERY REDUCTION USING PCA COMPONENTS
A more principled way to reduce the number of queries the adversary has to make to estimate the
gradient is to compute directional derivatives along the principal components as determined by
principal component analysis (PCA) (Shlens, 2014), which requires the adversary to have access to a
set of data which is represetative of the training data. PCA minimizes reconstruction error in terms of
the L2 norm; i.e., it provides a basis in which the Euclidean distance to the original sample from a
sample reconstructed using a subset of the basis vectors is the smallest.
Concretely, let the samples the adversary wants to misclassify be column vectors xi ∈ Rd for
i ∈ {1, . . . , n}
X be the d × n matrix of centered data samples (i.e. X = [x̃1 x̃2 . . . x̃n ], where
Pnand let
1
i
j
x̃ = x − n j=1 x ). The principal components of X are the normalized eigenvectors of its sample
covariance matrix C = XXT. Since C is a positive semidefinite matrix, there is a decomposition
C = UΛUT where U is an orthogonal matrix, Λ = diag(λ1 , . . . , λd ), and λ1 ≥ . . . ≥ λd ≥ 0.
Thus, U in Algorithm 2 is the d × d matrix whose columns are unit eigenvectors of C. The eigenvalue
λi is the variance of X along the ith component.
Algorithm 2 Gradient estimation with query reduction using PCA components
Input: x, k, U, δ, g(·)
ˆ x g(x) of g(·) at x
Output: Estimated gradient ∇
1: for i ← 1 to k do
ui
2:
Initialize v such that v = ku
, where ui is the ith column of U
ik
3:
Compute
g(x + δv) − g(x − δv)
αi (v) =
,
2δ
which is the two-sided approximation of the directional derivative along v
ˆ x g(x)i = ∇
ˆ x g(x)i−1 + αi (v)v
4:
Update ∇
5: end for
ˆ x g(x) = ∇
ˆ x g(x)k
6: Set ∇
10
PCA-based query reduction for Model A
100
None
112
23
12
6
3
80
60
40
Adversarial success (%)
Adversarial success (%)
100
20
0
PCA-based query reduction for Resnet-32
100
None
400
200
100
50
10
80
60
40
Adversarial success (%)
Random feature groupings for Model A
20
0
0
0.1 0.2 0.3 0.4 0.5
60
40
20
0
0.1 0.15 0.2 0.25 0.3 0.35 0.4
(a) Gradient Estimation attack with
query reduction using random
grouping and the logit loss (GE-QR
(RG-k, logit)) on Model A (MNIST,
d = 784). The adversarial success rate decreases as the number
of groups d kd e is decreased, where
k is the size of the group and d is
the dimension of the input.
None
400
200
100
50
10
80
4
(b) Gradient Estimation attack with
query reduction using PCA components and the logit loss (GEQR (PCA-k, logit)) on Model A
(MNIST, d = 784). The adversarial
success rates decrease as the number
of principal components k used for
estimation is decreased. Relatively
high success rates are maintained
even for k = 50.
8 12 16 20 24 28 32
(c) Gradient Estimation attack with
query reduction using PCA components and the logit loss (GEQR (PCA-k, logit)) on Resnet-32
(CIFAR-10). Relatively high success rates are maintained even for
k = 400.
Figure 3: Adversarial success rates for Gradient Estimation attacks with query reduction (FDQR (Technique, logit)) on Model A (MNIST) and Resnet-32 (CIFAR-10), where Technique
is either PCA or RG. ‘None’ refers to FD-logit, the case where the number of queries is 2d, where d
is the dimension of the input.
In Algorithm 2, U is the d × d matrix whose columns are the principal components ui , where i ∈ [d].
The quantity being estimated in Algorithm 2 is an approximation of the gradient in the PCA basis:
(∇x g(x))k =
k
X
∇x g(x)T
i=1
ui
kui k
ui
,
kui k
where the term on the left represents an approximation of the true gradient by the sum of its projection
along the top k principal components. In Algorithm 2, the weights of the representation in the PCA
basis are approximated using the approximate directional derivatives along the principal components.
3.3
I TERATIVE ATTACKS WITH QUERY REDUCTION
Performing an iterative attack with the gradient estimated using the finite difference method (Equation 1) could be expensive for an adversary, needing 2td queries to the target model, for t iterations
with the two-sided finite difference estimation of the gradient. To lower the number of queries needed,
the adversary can use either of the query reduction techniques described above to reduce the number
of queries to 2tk ( k < d). These attacks using the cross-entropy loss are denoted as IGE-QR (RG-k,
xent) for the random grouping technique and IGE-QR (PCA-k, xent) for the PCA-based technique.
3.3.1
E VALUATION OF G RADIENT E STIMATION ATTACKS WITH QUERY REDUCTION
In this section, we first summarize the results obtained using Gradient Estimation attacks with query
reduction and then provide a more detailed analysis of the effect of dimension on attacks with query
reduction
Gradient estimation with query reduction maintains high attack success rates: For both datasets,
the Gradient Estimation attack with PCA based query reduction (GE-QR (PCA-k, logit)) is effective,
with performance close to that of FD-logit with k = 100 for MNIST (Figure 2a) and k = 400 for
CIFAR-10 (Figure 2b). The Iterative Gradient Estimation attacks with both Random Grouping and
PCA based query reduction (IGE-QR (RG-k, logit) and IGE-QR (PCA-k, logit)) achieve close to
100% success rates for untargeted attacks and above 80% for targeted attacks on Model A on MNIST
and Resnet-32 on CIFAR-10 (Figure 5). Figure 5 clearly shows the effectiveness of the gradient
11
estimation attack across models, datasets, and adversarial goals. While random grouping is not as
effective as the PCA based method for Single-step attacks, it is as effective for iterative attacks.
Effect of dimension on Gradient Estimation attacks: We consider the effectiveness of Gradient
Estimation with random grouping based query reduction and the logit loss (GE-QR (RG-k, logit))
on Model A on MNIST data in Figure 3a, where k is the number of indices chosen in each iteration
of Algorithm 1. Thus, as k increases and the number of groups decreases, we expect attack success
to decrease as gradients over larger groups of features are averaged. This is the effect we see in
Figure 3a, where the adversarial success rate drops from 93% to 63% at = 0.3 as k increases from
1 to 7. Grouping with k = 7 translates to 112 queries per MNIST image, down from 784. Thus, in
order to achieve high adversarial success rates with the random grouping method, larger perturbation
magnitudes are needed.
On the other hand, the PCA-based approach GE-QR (PCA-k, logit) is much more effective, as
can be seen in Figure 3b. Using 100 principal components to estimate the gradient for Model A on
MNIST as in Algorithm 2, the adversarial success rate at = 0.3 is 88.09%, as compared to 92.9%
without any query reduction. Similarly, using 400 principal components for Resnet-32 on CIFAR-10
(Figure 3c), an adversarial success rate of 66.9% can be achieved at = 8. At = 16, the adversarial
success rate rises to 80.1%.
Remarks: While decreasing the number of queries does reduce attack success rates, the proposed
query reduction methods maintain high attack success rates.
3.4
A DVERSARIAL SAMPLES
In Figure 4, we show some examples of successful untargeted adversarial samples against Model A
on MNIST and Resnet-32 on CIFAR-10. These images were generated with an L∞ constraint of
= 0.3 for MNIST and = 8 for CIFAR-10. Clearly, the amount of perturbation added by iterative
attacks is much smaller, barely being visible in the images.
WB FGS FD-logit GE-QR-PCA
MNIST
Single-step
CIFAR-10
WB IFGS IFD-logit IGE-QR-PCA
MNIST
Iterative
CIFAR-10
Figure 4: Untargeted adversarial samples on Model A on MNIST and Resnet-32 on CIFAR-10.
All attacks use the logit loss. Perturbations in the images generated using single-step attacks are far
smaller than those for iterative attacks. The ‘7’ from MNIST is classified as a ‘3’ by all single-step
attacks and as a ‘9’ by all iterative attacks. The dog from CIFAR-10 is classified as a bird by the
white-box FGS and Finite Difference attack, and as a frog by the Gradient Estimation attack with
query reduction.
3.5
E FFICIENCY OF GRADIENT ESTIMATION ATTACKS
In our evaluations, all models were run on a GPU with a batch size of 100. On Model A on MNIST
data, single-step attacks FD-xent and FD-logit take 6.2 × 10−2 and 8.8 × 10−2 seconds per sample
respectively. Thus, these attacks can be carried out on the entire MNIST test set of 10,000 images in
about 10 minutes. For iterative attacks with no query reduction, with 40 iterations per sample (α set
to 0.01), both IFD-xent and IFD-xent-T taking about 2.4 seconds per sample. Similarly, IFD-logit
12
Query-based attack
Attack success
No. of queries
Time per sample (s)
Finite Diff.
Gradient Estimation (RG-8)
Iter. Finite Diff.
Iter. Gradient Estimation (RG-8)
Particle Swarm Optimization
SPSA
92.9 (6.1)
61.5 (6.0)
100.0 (2.1)
98.4 (1.9)
84.1 (5.3)
96.7 (3.9)
1568
196
62720
8000
10000
8000
8.8 × 10−2
1.1 × 10−2
3.5
0.43
21.2
1.25
Table 3: Comparison of untargeted query-based black-box attack methods. All results are for
attacks using the first 1000 samples from the MNIST dataset on Model A and with an L∞ constraint
of 0.3. The logit loss is used for all methods expect PSO, which uses the class probabilities.
and IFD-logit-T take about 3.5 seconds per sample. With query reduction, using IGE-QR (PCA-k,
logit) with k = 100 the time taken is just 0.5 seconds per sample.
For Resnet-32 on the CIFAR-10 dataset, FD-xent, FD-xent-T, FD-logit and FD-logit-T all take
roughly 3s per sample. The iterative variants of these attacks with 10 iterations (α set to 1.0) take
roughly 30s per sample. Using query reduction, IGE-QR (PCA-k, logit) with k = 100 with 10
iterations takes just 5s per sample. The time required per sample increases with the complexity of the
network, which is observed even for white-box attacks.
All the above numbers are for the case when queries are not made in parallel. Our attack algorithm
allows for queries to be made in parallel as well. We find that a simple parallelization of the queries
gives us a 2 − 4× speedup. The limiting factor is the fact that the model is loaded on a single GPU,
which implies that the current setup is not fully optimized to take advantage of the inherently parallel
nature of our attack. With further optimization, greater speedups can be achieved.
Remarks: Overall, our attacks are very efficient and allow an adversary to generate a large number
of adversarial samples in a short period of time.
3.6
OTHER QUERY- BASED ATTACKS
We experimented with Particle Swarm Optimization (PSO),2 a commonly used evolutionary optimization strategy, to construct adversarial samples as was done by Sharif et al. (2016), but found it
to be prohibitively slow for a large dataset, and it was unable to achieve high adversarial success
rates even on the MNIST dataset. We also tried to use the Simultaneous Perturbation Stochastic
Approximation (SPSA) method, which is similar to the method of Finite Differences, but it estimates
the gradient of the loss along a random direction r at each step, instead of along the canonical basis
vectors. While each step of SPSA only requires 2 queries to the target model, a large number of steps
are nevertheless required to generate adversarial samples. A single step of SPSA does not reliably
produce adversarial samples. The two main disadvantages of this method are that i) the convergence
of SPSA is much more sensitive in practice to the choice of both δ (gradient estimation step size)
and α (loss minimization step size), and ii) even with the same number of queries as the Gradient
Estimation attacks, the attack success rate is lower even though the distortion is higher.
A comparative evaluation of all the query-based black-box attacks we experimented with for the
MNIST dataset is given in Table 3. The PSO based attack uses class probabilities to define the loss
function, as it was found to work better than the logit loss in our experiments. The attack that achieves
the best trade-off between speed and attack success is IGE-QR (RG-k, logit).
4
ATTACKING DEFENSES
In this section, we evaluate black-box attacks against different defenses based on adversarial training
and its variants. We focus on adversarial training based defenses as they aim to directly improve the
robustness of DNNs, and are among the most effective defenses demonstrated so far in the literature.
2
Using freely available code from http://pythonhosted.org/pyswarm/
13
IGE-QR (PCA-400, logit) across models
A
MNIST
A
Aadv-0.3
CIFAR-10
Resnet-32
Resnet-32 adv-8
Single-step
Iterative
PCA-100
RG-8
PCA-100
RG-8
88.1 (6.0)
4.1 (5.8)
61.5 (6.0)
2.0 (5.3)
99.9 (2.5)
50.7 (4.2)
98.4 (1.9)
27.5 (2.4)
Single-step
Untargeted
Resnet-32
Targeted
Aadv-0.3
Resnet-32adv-8
Iterative
PCA-400
RG-8
PCA-400
RG-8
66.9 (410.7)
8.0 (402.1)
66.8 (402.7)
7.7 (401.8)
98.0 (140.7)
97.0 (151.3)
99.0 (80.5)
98.0 (92.9)
0
40
20
60
80
100
Adversarial success (%)
(a) Adversarial success rates for untargeted attacks (b) Adversarial success rates of the Iterative Grausing query reduction. The parameters used for query dient Estimation attack using the logit loss, with
reduction are indicated in the table.
PCA (400) used for query reduction.
Figure 5: Adversarial success rates for query-reduced attacks. is set to 0.3 for Model A on
MNIST and 8 for Resnet-32 on CIFAR-10. Model Aadv-0.3 and Resnet-32 adv-8 are adversarially
trained variants (ref. Section 4.1) of the models. All attacks use roughly 8000 queries per sample for
both datasets.
We find that Iterative Gradient Estimation attacks perform much better than any single-step black-box
attack against defenses. Nevertheless, in Figure 6, we show that with the addition of an initial random
perturbation to overcome “gradient masking” (Tramèr et al., 2017a), the Gradient Estimation attack
with Finite Differences is the most effective single-step black-box attack on adversarially trained
models on MNIST.
Dataset (Model)
Benign
Adv
Adv-Ens
Adv-Iter
MNIST (A)
CIFAR-10 (Resnet-32)
99.2
92.4
99.4
92.1
99.2
91.7
99.3
79.1
Table 4: Accuracy of models on the benign test data
4.1
4.1.1
BACKGROUND AND EVALUATION SETUP
A DVERSARIAL TRAINING
Szegedy et al. (2014) and Goodfellow et al. (2015) introduced the concept of adversarial training,
where the standard loss function for a neural network f is modified as follows:
˜ y) = α`f (x, y) + (1 − α)`f (xadv , y),
`(x,
(9)
where y is the true label of the sample x. The underlying objective of this modification is to make the
neural networks more robust by penalizing it during training to count for adversarial samples. During
training, the adversarial samples are computed with respect to the current state of the network using
an appropriate method such as FGSM.
Ensemble adversarial training. Tramèr et al. (2017a) proposed an extension of the adversarial
training paradigm which is called ensemble adversarial training. As the name suggests, in ensemble
adversarial training, the network is trained with adversarial samples from multiple networks.
Iterative adversarial training. A further modification of the adversarial training paradigm proposes
training with adversarial samples generated using iterative methods such as the iterative FGSM attack
described earlier (Madry
˛ et al., 2017).
4.2
A DVERSARIALLY TRAINED MODELS
We train variants of Model A with the 3 adversarial training strategies described above using
adversarial samples based on an L∞ constraint of 0.3. Model Aadv-0.3 is trained with FGS samples,
while Model Aadv-iter-0.3 is trained with iterative FGS samples using t = 40 and α = 0.01. For the
model with ensemble training, Model Aadv-ens-0.3 is trained with pre-generated FGS samples for
Models A, C, and D, as well as FGS samples. The source of the samples is chosen randomly for
14
MNIST
Model
Model Aadv-0.3
Model Aadv-ens-0.3
Model Aadv-iter-0.3
CIFAR-10
Model
Resnet-32 adv-8
Resnet-32 adv-ens-8
Resnet-32 adv-iter-8
Baseline
D. of M.
6.5 (5.6)
2.0 (5.6)
3.0 (5.6)
Gradient Estimation using Finite Differences
Single-step
FD-xent
FD-logit
10.3 (2.6)
2.8 (5.9)
6.1 (3.5)
6.2 (6.3)
9.2 (7.4)
7.5 (7.2)
Rand.
1.3 (6.1)
1.2 (6.1)
1.0 (6.1)
Baseline
D. of M.
9.6 (440.5)
10.1 (440.5)
22.86 (440.5)
Transfer from Model B
Iterative
IFD-xent
IFD-logit
36.4 (3.1)
76.5 (3.1)
24.2 (4.1)
96.4 (2.7)
14.5 (0.96)
11.6 (3.5)
Single-step
FGS-xent
FGS-logit
14.6 (6.2)
14.63 (6.3)
3.1 (6.2)
3.1 (6.3)
11.5 (6.2)
11.0 (6.3)
Transfer from Resnet-28-10
Gradient Estimation using Finite Differences
Rand.
10.9 (439.4)
10.4 (439.4)
21.41 (439.4)
Single-step
FD-xent
FD-logit
2.4 (232.9)
8.5 (401.9)
7.7 (360.2) 12.2 (399.8)
45.5 (365.5) 47.5 (331.1)
Iterative
IFGS-xent
IFGS-logit
16.5 (4.7)
15.9 (4.7)
4.8 (4.7)
4.9 (4.7)
8.7 (4.7)
8.2 (4.7)
Iterative
IFD-xent
IFD-logit
69.0 (136.0) 100.0 (73.8)
95.0 (190.4) 100.0 (85.2)
55.0 (397.6) 54.6 (196.3)
Single-step
FGS-xent
FGS-logit
13.1 (439.4) 13.2 (439.4)
9.7 (439.4)
9.6 (439.4)
23.2 (439.4) 23.1 (439.4)
Iterative
IFGS-xent
IFGS-logit
30.2 (275.4) 30.2 (275.6)
15.9 (275.4) 15.5 (275.6)
22.3 (275.4) 22.3 (275.6)
Table 5: Untargeted black-box attacks for models with adversarial training: adversarial success
rates and average distortion ∆(X, Xadv ) over the samples. Above: MNIST, = 0.3. Below:
CIFAR-10, = 8.
L∞ constrained strategies on Model Aadv-0.3
L∞ constrained strategies on Resnet-32adv-0.3
100
Difference-of-means
Random-perturbation
Finite-difference xent
Finite-difference logit
Transfer Model B FGS xent
Transfer Model B FGS logit
White-box FGS logit
White-box FGS xent
Rand+finite-difference
Rand+FGS
80
60
40
20
Adversarial success (%)
Adversarial success (%)
100
0
Difference-of-means
Random-perturbation
Finite-difference xent
Finite-difference logit
Transfer Resnet-28-10 FGS xent
White-box FGS logit
White-box FGS xent
80
60
40
20
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
0
4
8
12
16
20
24
28
(a) Model Aadv-0.3 (MNIST)
(b) Resnet-32 adv-8 (CIFAR-10)
Figure 6: Effectiveness of various single step black-box attacks against adversarially trained
models. On the MNIST model, Model Aadv-0.3 the attack with the highest performance up till = 0.3
is the Gradient Estimation attack using Finite Differences with initially added randomness. Beyond
this, the Transferability attack (single local model) using samples from Model B performs better. On
the CIFAR-10 model Resnet-32 adv-8 , the best performing attack is the Transferability attack (single
local model) using samples from Resnet-28-10.
each minibatch during training. For all adversarially trained models, each training batch contains 128
samples of which 64 are benign and 64 are adversarial samples (either FGSM or iterative FGSM).
This implies that the loss for each is weighted equally during training; i.e., in Eq. 9, α is set to 0.5.
Networks using standard and ensemble adversarial training are trained for 12 epochs, while those
using iterative adversarial training are trained for 64 epochs.
We train variants of Resnet-32 using adversarial samples with an L∞ constraint of 8. Resnet-32
adv-8 is trained with FGS samples with the same constraint, and Resnet-32 ens-adv-8 is trained with
pre-generated FGS samples from Resnet-32 and Std.-CNN as well as FGS samples. Resnet-32
adv-iter-8 is trained with iterative FGS samples using t = 10 and α = 1.0. The adversarial variants of
Resnet-32 is trained for 80,000 steps
Table 4 shows the accuracy of these models with various defenses on benign test data.
4.3
S INGLE - STEP ATTACKS ON DEFENSES
In Figure 6a, we can see that both single-step black-box and white-box attacks have much lower
adversarial success rates on Model Aadv-0.3 as compared to Model A. The success rate of the Gradient
Estimation attacks matches that of white-box attacks on these adversarially trained networks as
well. To overcome this, we add an initial random perturbation to samples before using the Gradient
Estimation attack with Finite Differences and the logit loss (FD-logit). These are then the most
effective single step black-box attacks on Model Aadv-0.3 at = 0.3 with an adversarial success rate
of 32.2%, surpassing the Transferability attack (single local model) from B.
In Figure 6b, we again see that the Gradient Estimation attacks using Finite Differences (FD-xent
and FD-logit) and white-box FGS attacks (FGS-xent and FGS-logit) against Resnet-32. As
is increased, the attacks that perform the best are Random Perturbations (Rand.), Difference-ofmeans (D. of M.), and Transferability attack (single local model) from Resnet-28-10 with the latter
performing slightly better than the baseline attacks. This is due to the ‘gradient masking’ phenomenon
and can be overcome by adding random perturbations as for MNIST. An interesting effect is observed
15
Finite-difference vs RAND-FGSM for Model A variants
Adversarial success (%)
100
Finite-difference logit on Model Aadv-0.3
White-box FGS on Model Aadv-0.3
Finite-difference logit on Model Aadv-ens-0.3
White-box FGS on Model Aadv-ens-0.3
Finite-difference logit on Model Aadv-iter-0.3
White-box FGS on Model Aadv-iter-0.3
80
60
40
20
0
0.1 0.15 0.2 0.25 0.3 0.35 0.4
Figure 7: Increasing the effectiveness of FD-logit attacks on Model Aadv-0.3 , Model Aadv-ens-0.3
and Model Aadv-iter-0.3 (MNIST) by adding an initial L∞ constrained random perturbation of
magnitude 0.01.
at = 4, where the adversarial success rate is higher than at = 8. The likely explanation for this
effect is that the model has overfitted to adversarial samples at = 8. Our Gradient Estimation attack
closely tracks the adversarial success rate of white-box attacks in this setting as well.
Increasing effectiveness of single-step attacks using initial random perturbation: Since the
Gradient Estimation attack with Finite Differences (FD-xent and FD-logit) were not performing well
due the masking of gradients at the benign sample x, we added an initial random perturbation to
escape this low-gradient region as in the RAND-FGSM attack (Tramèr et al., 2017a). Figure 7 shows
the effect of adding an initial L∞ -constrained perturbation of magnitude 0.05. With the addition of
a random perturbation, FD-logit has a much improved adversarial success rate on Model Aadv-0.3 ,
going up to 32.2% from 2.8% without the perturbation at a total perturbation value of 0.3. It even
outperforms the white-box FGS (FGS-logit) with the same random perturbation added. This effect
is also observed for Model Aadv-ens-0.3 , but Model Aadv-iter-0.3 appears to be resistant to single-step
gradient based attacks. Thus, our attacks work well for single-step attacks on DNNs with standard
and ensemble adversarial training, and achieve performance levels close to that of white-box attacks.
4.4
I TERATIVE ATTACKS ON DIFFERENT ADVERSARIAL TRAINING DEFENSES
While single-step black-box attacks are less effective at lower than the one used for training, our
experiments show that iterative black-box attacks continue to work well even against adversarially
trained networks. For example, the Iterative Gradient Estimation attack using Finite Differences
with a logit loss (IFD-logit) achieves an adversarial success rate of 96.4% against Model Aadv-ens-0.3 ,
while the best transferability attack has a success rate of 4.9%. It is comparable to the white-box
attack success rate of 93% from Table 10. However, Model Aadv-iter-0.3 is quite robust even against
iterative attacks, with the highest black-box attack success rate achieved being 14.5%.
Further, in Figure 5, we can see that using just 4000 queries per sample, the Iterative Gradient
Estimation attack using PCA for query reduction (IGE-QR (PCA-400, logit)) achieves 100%
(untargeted) and 74.5% (targeted) adversarial success rates against Model Aadv-0.3 . Our methods far
outperform the other black-box attacks, as shown in Table 10.
Iterative black-box attacks perform well against adversarially trained models for CIFAR-10 as well.
IFD-logit achieves attack success rates of 100% against both Resnet-32 adv-8 and Resnet-32
adv-ens-8 (Table 5), which reduces slightly to 97% when IFD-QR (PCA-400, logit) is used. This
matches the performance of white-box attacks as given in Table 10. IFD-QR (PCA-400, logit) also
achieves a 72% success rate for targeted attacks at = 8 as shown in Figure 5.
The iteratively trained model has poor performance on both benign as well as adversarial samples.
Resnet-32 adv-iter-8 has an accuracy of only 79.1% on benign data, as shown in Table 4. The Iterative
16
Gradient Estimation attack using Finite Differences with cross-entropy loss (IFD-xent) achieves an
untargeted attack success rate of 55% on this model, which is lower than on the other adversarially
trained models, but still significant. This is in line with the observation by Madry
˛
et al. (2017) that
iterative adversarial training needs models with large capacity for it to be effective. This highlights a
limitation of this defense, since it is not clear what model capacity is needed and the models we use
already have a large number of parameters.
Remarks. Both single-step and iterative variants of the Gradient Estimation attacks outperform
other black-box attacks, achieving attack success rates close to those of white-box attacks even on
adversarially trained models.
5
ATTACKS ON C LARIFAI :
A REAL - WORLD SYSTEM
Since the only requirement for carrying out the Gradient Estimation based attacks is query-based
access to the target model, a number of deployed public systems that provide classification as a
service can be used to evaluate our methods. We choose Clarifai, as it has a number of models
trained to classify image datasets for a variety of practical applications, and it provides black-box
access to its models and returns confidence scores upon querying. In particular, Clarifai has models
used for the detection of Not Safe For Work (NSFW) content, as well as for Content Moderation.
These are important applications where the presence of adversarial samples presents a real danger: an
attacker, using query access to the model, could generate an adversarial sample which will no longer
be classified as inappropriate. For example, an adversary could upload violent images, adversarially
modified, such that they are marked incorrectly as ‘safe’ by the Content Moderation model.
We evaluate our attack using the Gradient Estimation method on the Clarifai NSFW and Content
Moderation models. An important point to note here is that given the lack of an easily accessible
open-source dataset for these tasks to train a local surrogate model, carrying out attacks based on
transferability is a challenging task. On the other hand, our attack can directly be used for any image
of the adversary’s choice. The Content Moderation model has five categories, ‘safe’, ‘suggestive’,
‘explicit’, ‘drug’ and ‘gore’, while the NSFW model has just two categories, ‘SFW’ and ‘NSFW.’
When we query the API with an image, it returns the confidence scores associated with each category,
with the confidence scores summing to 1. We use the random grouping technique in order to reduce
the number of queries and take the logarithm of the confidence scores in order to use the logit loss. A
large number of successful attack images can be found at https://sunblaze-ucb.github.
io/blackbox-attacks/. Due to their possibly offensive nature, they are not included in the
paper.
An example of an attack on the Content Moderation API is given in Figure 1, where the original
image on the left is clearly of some kind of drug on a table, with a spoon and a syringe. It is classified
as a drug by the Content Moderation model with a confidence score of 0.99. The image on the right
is an adversarial image generated with 192 queries to the Content Moderation API, with an L∞
constraint on the perturbation of = 32. While the image can still clearly be classified by a human as
being of drugs on a table, the Content Moderation model now classifies it as ‘safe’ with a confidence
score of 0.96.
Remarks. The proposed Gradient Estimation attacks can successfully generate adversarial examples
that are misclassified by a real-world system hosted by Clarifai without prior knowledge of the
training set or model.
6
E XISTING BLACK - BOX ATTACKS
In this section, we describe existing methods for generating adversarial examples. In all of these
attacks, the adversary’s perturbation is constrained using the L∞ distance.
6.1
BASELINE ATTACKS
Now, we describe two baseline black-box attacks which can be carried out without any knowledge of
or query access to the target model.
17
6.1.1
R ANDOM PERTURBATIONS
With no knowledge of f or the training set, the simplest manner in which an adversary may seek to
carry out an attack is by adding a random perturbation to the input (Szegedy et al., 2014; Goodfellow
et al., 2015; Fawzi et al., 2015). These perturbations can be generated by any distribution of the
adversary’s choice and constrained according to an appropriate norm. If we let P be a distribution
over X , and p is a random variable drawn according to P , then a noisy sample is just xnoise = x + p.
Since random noise is added, it is not possible to generate targeted adversarial samples in a principled
manner. This attack is denoted as Rand. throughout.
6.1.2
D IFFERENCE OF MEANS
A perturbation aligned with the difference of means of two classes is likely to be effective for an
adversary hoping to cause misclassification for a broad range of classifiers (Tramèr et al., 2017b).
While these perturbations are far from optimal for DNNs, they provide a useful baseline to compare
against. Adversaries with at least partial access to the training or test sets can carry out this attack.
An adversarial sample generated using this method, and with L∞ constraints, is xadv = x + ·
sign(µt − µo ), where µt is the mean of the target class and µo is the mean of the original ground
truth class. For an untargeted attack, t = argmini d(µi − µo ), where d(·, ·) is an appropriately
chosen distance function. In other words, the class whose mean is closest to the original class in terms
of the Euclidean distance is chosen to be the target. This attack is denoted as D. of M. throughout.
6.1.3
E FFECTIVENESS OF BASELINE ATTACKS
In the baseline attacks described above, the choice of distribution for the random perturbation attack
and the choice of distance function for the difference of means attack are not fixed. Here, we describe
the choices we make for both attacks. The random perturbation p for each sample (for both MNIST
and CIFAR-10) is chosen independently according to a multivariate normal distribution with mean 0,
i.e. p ∼ N (0, Id ). Then, depending on the norm constraint, either a signed and scaled version of the
random perturbation (L∞ ) or a scaled unit vector in the direction of the perturbation (L2 ) is added.
For an untargeted attack utilizing perturbations aligned with the difference of means, for each sample,
the mean of the class closest to the original class in the L2 distance is determined.
As expected, adversarial samples generated using Rand. do not achieve high adversarial success
rates (Table 1) in spite of having similar or larger average distortion than the other black-box attacks
for both the MNIST and CIFAR-10 models. However, the D. of M. method is quite effective at higher
perturbation values for the MNIST dataset as can be seen in Figure 2a. Also, for Models B and D,
the D. of M. attack is more effective than FD-xent. The D. of M. method is less effective in the
targeted attack case, but for Model D, it outperforms the transferability based attack considerably. Its
success rate is comparable to the targeted transferability based attack for Model A as well.
The relative effectiveness of the two baseline methods is reversed for the CIFAR-10 dataset, however,
where Rand. outperforms D. of M. considerably as is increased. This indicates that the models
trained on MNIST have normal vectors to decision boundaries which are more aligned with the
vectors along the difference of means as compared to the models on CIFAR-10.
6.2
S INGLE - STEP AND I TERATIVE FAST G RADIENT M ETHODS
Now, we describe two white-box attack methods, used in transferability-based attacks, for which
we constructed approximate, gradient-free versions in Section 3. These attacks are based on either
iterative or single-step gradient based minimization of appropriately defined loss functions of neural
networks. All results for white-box attacks are contained in Appendix D. Since these methods all
require the knowledge of the model’s gradient, we assume the adversary has access to a local model
f s . Adversarial samples generated for f s can then be transferred to the target model f t to carry out
a transferability-based attack (Papernot et al., 2016; Moosavi-Dezfooli et al., 2016). An ensemble
of local models (Liu et al., 2017) may also be used. Transferability-based attacks are described in
Section 6.3.
The single-step Fast Gradient method, first introduced by Goodfellow et al. (2015), utilizes a firstorder approximation of the loss function in order to construct adversarial samples for the adversary’s
surrogate local model f s . The samples are constructed by performing a single step of gradient ascent
18
for untargeted attacks. Formally, the adversary generates samples xadv with L∞ constraints (known
as the Fast Gradient Sign (FGS) method) in the untargeted attack setting as
xadv = x + · sign(∇x `f s (x, y)),
(10)
where `f s (x, y) is the loss function with respect to which the gradient is taken. The loss function
typically used is the cross-entropy loss (Goodfellow et al., 2016). Adversarial samples generated
using the targeted FGS attack are
xadv = x − · sign(∇x `f s (x, T )),
(11)
where T is the target class.
Iterative Fast Gradient methods are simply multi-step variants of the Fast Gradient method described
above (Kurakin et al., 2016), where the gradient of the loss is added to the sample for t + 1 iterations,
starting from the benign sample, and the updated sample is projected to satisfy the constraints H in
every step:
t
t
xt+1
adv = ΠH (xadv + α · sign(∇xtadv `f s (xadv , y))),
(12)
with x0adv = x. Iterative fast gradient methods thus essentially carry out projected gradient descent
(PGD) with the goal of maximizing the loss, as pointed out by Madry
˛
et al. (2017). Targeted
adversarial samples generated using iterative FGS are
t
t
xt+1
adv = ΠH (xadv − α · sign(∇xtadv `f s (xadv , T ))).
6.2.1
(13)
B EYOND THE CROSS - ENTROPY LOSS
Prior work by Carlini & Wagner (2017) investigates a variety of loss functions for white-box attacks
based on the minimization of an appropriately defined loss function. In our experiments with neural
networks, for untargeted attacks, we use a loss function based on logits which was found to work
well for white-box attacks in Carlini & Wagner (2017). The loss function is given by
`(x, y) = max(φ(x + δ)y − max{φ(x + δ)i : i 6= y}, −κ),
(14)
where y represents the ground truth label for the benign sample x, φ(·) are the logits. κ is a confidence
parameter that can be adjusted to control the strength of the adversarial sample. Targeted adversarial
samples are generated using the following loss term:
xadv = x − · sign(∇x (max(φ(x)i : i 6= T ) − φ(x)T )).
6.3
(15)
T RANSFERABILITY BASED ATTACKS
Here we describe black-box attacks that assume the adversary has access to a representative set of
training data in order to train a local model. One of the earliest observations with regards to adversarial
samples for neural networks was that they transfer; i.e, adversarial attack samples generated for one
network are also adversarial for another network. This observation directly led to the proposal of a
black-box attack where an adversary would generate samples for a local network and transfer these
to the target model, which is referred to as a Transferability based attack. Targeted transferability
attacks are carried out using locally generated targeted white-box adversarial samples.
6.3.1
S INGLE LOCAL MODEL
These attacks use a surrogate local model f s to craft adversarial samples, which are then submitted
to f in order to cause misclassification. Most existing black-box attacks are based on transferability
from a single local model (Papernot et al., 2016; Moosavi-Dezfooli et al., 2016). The different attack
strategies to generate adversarial instances introduced in Section 6.2 can be used here to generate
adversarial instances against f s , so as to attack f .
6.3.2
E NSEMBLE OF LOCAL MODELS
Since it is not clear which local model f s is best suited for generating adversarial samples that
transfer well to the target model f , Liu et al. (2017) propose the generation of adversarial examples
19
for an ensemble of local models. This method modifies each of the existing transferability attacks by
substituting a sum over the loss functions in place of the loss from a single local model.
Concretely, let the ensemble of m local models to be usedP
to generate the local loss be {f s1 , . . . , f sm }.
m
The ensemble loss is then computed as `ens (x, y) = i=1 αi `f si (x, y), where αi is the weight
given to each model in the ensemble. The FGS attack in the ensemble setting then becomes
xadv = x + · sign(∇x `ens (x, y)). The Iterative FGS attack is modified similarly. Liu et al. (2017)
show that the Transferability attack (local model ensemble) performs well even in the targeted attack
case, while Transferability attack (single local model) is usually only effective for untargeted attacks.
The intuition is that while one model’s gradient may not be adversarial for a target model, it is likely
that at least one of the gradient directions from the ensemble represents a direction that is somewhat
adversarial for the target model.
Untargeted Transferability to Model A
Source
B
B,C
B,C,D
Single-step
Iterative
FGS-xent
FGS-logit
IFGS-xent
IFGS-logit
66.3 (6.2)
80.8 (6.3)
89.8 (4.75)
88.5 (4.75)
68.1 (6.2)
89.8 (6.3)
95.0 (4.8)
97.1 (4.9)
56.0 (6.3)
88.7 (6.4)
73.5 (5.3)
94.4 (5.3)
Targeted Transferability to Model A
Single-step
Iterative
Source FGS-T (xent) FGS-T (logit) IFGS-T (xent) IFGS-T (logit)
B
18.3 (6.3)
18.1 (6.3)
54.5 (4.6)
46.5 (4.2)
23.0 (6.3)
23.0 (6.3)
76.7 (4.8)
72.3 (4.5)
B,C
B,C,D
25.2 (6.4)
25.1 (6.4)
74.6 (4.9)
66.1 (4.7)
Table 6: Adversarial success rates for transferability-based attacks on Model A (MNIST) at
= 0.3. Numbers in parentheses beside each entry give the average distortion ∆(X, Xadv ) over the
test set. This table compares the effectiveness of using a single local model to generate adversarial
examples versus the use of a local ensemble.
6.3.3
T RANSFERABILITY ATTACK RESULTS
For the transferability experiments, we choose to transfer from Model B for MNIST dataset and
from Resnet-28-10 for CIFAR-10 dataset, as these models are each similar to at least one of the
other models for their respective dataset and different from one of the others. They are also fairly
representative instances of DNNs used in practice.
Adversarial samples generated using single-step methods and transferred from Model B to the other
models have higher success rates for untargeted attacks when they are generated using the logit loss
as compared to the cross entropy loss as can be seen in Table 1. For iterative adversarial samples,
however, the untargeted attack success rates are roughly the same for both loss functions. As has been
observed before, the adversarial success rate for targeted attacks with transferability is much lower
than the untargeted case, even when iteratively generated samples are used. On the MNIST dataset,
the highest targeted transferability rate is 54.5% (Table 2) as compared to 89.8% in the untargeted
case (Table 1).
One attempt to improve the transferability rate is to use an ensemble of local models, instead of
a single one. The results for this on the MNIST data are presented in Table 6. In general, both
untargeted and targeted transferability increase when an ensemble is used. However, the increase is
not monotonic in the number of models used in the ensemble, and we can see that the transferability
rate for IFGS-xent samples falls sharply when Model D is added to the ensemble. This may be
due to it having a very different architecture as compared to the models, and thus also having very
different gradient directions. This highlights one of the pitfalls of transferability, where it is important
to use a local surrogate model similar to the target model for achieving high attack success rates.
20
7
C ONCLUSION
Overall, in this paper, we conduct a systematic analysis of new and existing black-box attacks on
state-of-the-art classifiers and defenses. We propose Gradient Estimation attacks which achieve high
attack success rates comparable with even white-box attacks and outperform other state-of-the-art
black-box attacks. We apply random grouping and PCA based methods to reduce the number of
queries required to a small constant and demonstrate the effectiveness of the Gradient Estimation
attack even in this setting. We also apply our black-box attack against a real-world classifier and
state-of-the-art defenses. All of our results show that Gradient Estimation attacks are extremely
effective in a variety of settings, making the development of better defenses against black-box attacks
an urgent task.
R EFERENCES
Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S.
Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew
Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath
Kudlur, Josh Levenberg, Dan Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike
Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent
Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg,
Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on
heterogeneous systems, 2015. URL http://tensorflow.org/. Software available from
tensorflow.org.
Rodrigo Benenson. Classification datasets results. http://rodrigob.github.io/are_we_
there_yet/build/#classification-dataset-type, 2016. Accessed: 2017-08-22.
Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In IEEE
Symposium on Security and Privacy, 2017, 2017.
Pin-Yu Chen, Huan Zhang, Yash Sharma, Jinfeng Yi, and Cho-Jui Hsieh. Zoo: Zeroth order
optimization based black-box attacks to deep neural networks without training substitute models.
arXiv preprint arXiv:1708.03999, 2017.
Clarifai. Clarifai | image & video recognition API. https://clarifai.com. Accessed:
2017-08-22.
Hung Dang, Huang Yue, and Ee-Chien Chang. Evading classifiers by morphing in the dark. 2017.
Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Analysis of classifiers’ robustness to adversarial
perturbations. arXiv preprint arXiv:1502.02590, 2015.
Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning. MIT Press, 2016.
Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial
examples. In International Conference on Learning Representations, 2015.
Google Vision API. Vision API - image content analysis | Google cloud platform. https://
cloud.google.com/vision/. Accessed: 2017-08-22.
Shixiang Gu and Luca Rigazio. Towards deep neural network architectures robust to adversarial
examples. arXiv preprint arXiv:1412.5068, 2014.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image
recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition,
pp. 770–778, 2016.
Francis Begnaud Hildebrand. Advanced calculus for applications, volume 63. Prentice-Hall Englewood Cliffs, NJ, 1962.
James Kennedy. Particle swarm optimization. In Encyclopedia of machine learning, pp. 760–766.
Springer, 2011.
21
Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009.
Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial examples in the physical world.
arXiv preprint arXiv:1607.02533, 2016.
Yann LeCun and Corrina Cortes. The MNIST database of handwritten digits. 1998.
Amy Liu. Clarifai featured hack: Block unwanted nudity in blog comments with disqus. https:
//goo.gl/TCCVrR, 2016. Accessed: 2017-08-22.
Yanpei Liu, Xinyun Chen, Chang Liu, and Dawn Song. Delving into transferable adversarial examples
and black-box attacks. In ICLR, 2017.
Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and
accurate method to fool deep neural networks. arXiv preprint arXiv:1511.04599, 2015.
Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal
adversarial perturbations. arXiv preprint arXiv:1610.08401, 2016.
Konda Reddy Mopuri, Utsav Garg, and R Venkatesh Babu. Fast feature fool: A data independent
approach to universal adversarial perturbations. arXiv preprint arXiv:1707.05572, 2017.
Aleksander Madry,
˛
Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu.
Towards deep learning models resistant to adversarial attacks. arXiv:1706.06083 [cs, stat], June
2017.
Nina Narodytska and Shiva Prasad Kasiviswanathan. Simple black-box adversarial perturbations for
deep networks. arXiv preprint arXiv:1612.06299, 2016.
Blaine Nelson, Benjamin IP Rubinstein, Ling Huang, Anthony D Joseph, Steven J Lee, Satish Rao,
and JD Tygar. Query strategies for evading convex-inducing classifiers. The Journal of Machine
Learning Research, 13(1):1293–1332, 2012.
Nicolas Papernot, Patrick McDaniel, and Ian Goodfellow. Transferability in machine learning: from
phenomena to black-box attacks using adversarial samples. arXiv preprint arXiv:1605.07277,
2016.
Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram
Swami. Practical black-box attacks against deep learning systems using adversarial examples. In
Proceedings of the 2017 ACM Asia Conference on Computer and Communications Security, 2017.
Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K Reiter. Accessorize to a crime: Real
and stealthy attacks on state-of-the-art face recognition. In Proceedings of the 2016 ACM SIGSAC
Conference on Computer and Communications Security, pp. 1528–1540. ACM, 2016.
Jonathon Shlens. A tutorial on principal component analysis. arXiv preprint arXiv:1404.1100, 2014.
James C Spall. Multivariate stochastic approximation using a simultaneous perturbation gradient
approximation. IEEE transactions on automatic control, 37(3):332–341, 1992.
James C Spall. Introduction to stochastic search and optimization: estimation, simulation, and
control, volume 65. John Wiley & Sons, 2005.
Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow,
and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning
Representations, 2014.
TensorFlow Authors. Tensorflow resnet models. https://github.com/tensorflow/
models/tree/master/resnet, a. Accessed: 2017-08-22.
TensorFlow Authors.
Tensorflow CIFAR-10 tutorial model.
https://github.com/
tensorflow/models/tree/master/tutorials/image/cifar10, b. Accessed:
2017-08-22.
22
Florian Tramèr, Alexey Kurakin, Nicolas Papernot, Dan Boneh, and Patrick McDaniel. Ensemble
adversarial training: Attacks and defenses. arXiv preprint arXiv:1705.07204, 2017a.
Florian Tramèr, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. The space of
transferable adversarial examples. arXiv preprint arXiv:1704.03453, 2017b.
Watson Visual Recognition. Watson visual recognition. https://www.ibm.com/watson/
services/visual-recognition/. Accessed: 2017-10-27.
Stephen J Wright and Jorge Nocedal. Numerical optimization. Springer Science, 35(67-68):7, 1999.
Weilin Xu, Yanjun Qi, and David Evans. Automatically evading classifiers. In Proceedings of the
2016 Network and Distributed Systems Symposium, 2016.
Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146,
2016.
A
A LTERNATIVE ADVERSARIAL SUCCESS METRIC
Note that the adversarial success rate can also be computed by considering only the fraction of inputs
that meet the adversary’s objective given that the original sample was correctly classified. That is,
one would count the fraction of correctly classified inputs (i.e. f (x) = y) for which f (xadv ) 6= y
in the untargeted case, and f t (xadv ) = T in the targeted case. In a sense, this fraction represents
those samples which are truly adversarial, since they are misclassified solely due to the adversarial
perturbation added and not due to the classifier’s failure to generalize well. In practice, both these
methods of measuring the adversarial success rate lead to similar results for classifiers with high
accuracy on the test data.
B
F ORMAL DEFINITIONS FOR QUERY- BASED ATTACKS
Here, we provide a unified framework assuming an adversary can make active queries to the model.
Existing attacks making zero queries are a special case in this framework. Given an input instance x,
the adversary makes a sequence of queries based on the adversarial constraint set H, and iteratively
adds perturbations until the desired query results are obtained, using which the corresponding
adversarial example xadv is generated.
We formally define the targeted and untargeted black-box attacks based on the framework as below.
Definition 1 (Untargeted black-box attack). Given an input instance x and an iterative active
query attack strategy A, a query sequence can be generated as x2 = A({(x1 , qf1 )}, H), ..., xi =
A({(x1 , qf1 ), . . . , (xi−1 , qfi−1 )}, H), where qfi denotes the ith corresponding query result on xi , and
we set x1 = x. A black-box attack on f (·; θ) is untargeted if the adversarial example xadv = xk
satisfies f (xadv ; θ) 6= f (x; θ), where k is the number of queries made.
Definition 2 (Targeted black-box attack). Given an input instance x and an iterative active
query attack strategy A, a query sequence can be generated as x2 = A({(x1 , qf1 )}, H), ...,
xi = A({(x1 , qf1 ), . . . , (xi−1 , qfi−1 )}, H), where qfi denotes the ith corresponding query result
on xi , and we set x1 = x. A black-box attack on f (·; θ) is targeted if the adversarial example
xadv = xk satisfies f (xadv ; θ) = T , where T and k are the target class and the number of queries
made, respectively.
The case where the adversary makes no queries to the target classifier is a special case we refer to as
a zero-query attack. In the literature, a number of these zero-query attacks have been carried out with
varying degrees of success (Papernot et al., 2016; Liu et al., 2017; Moosavi-Dezfooli et al., 2016;
Mopuri et al., 2017).
23
Zero-query
Abbreviation
Untargeted
Targeted
Rand.
D. of M.
+
+
+
Transfer model FGS-xent
Transfer model FGS-logit
Transfer model IFGS-xent
Transfer model IFGS-logit
+
+
+
+
+
+
+
+
FD-xent
FD-logit
IFD-xent
IFD-logit
+
+
+
+
+
+
+
+
GE-QR (RG-k, logit)
IGE-QR (RG-k, logit)
GE-QR (PCA-k, logit)
IGE-QR (PCA-k, logit)
+
+
+
+
+
+
+
+
WB FGS-xent
WB FGS-logit
WB IFGS-xent
WB IFGS-logit
+
+
+
+
+
+
+
+
Surrogate attack
FGS
Steps
Single-step
IFGS
Iterative
Finite-difference gradient estimation
Query based
Black-box
Attack
Random-Gaussian-perturbation
Difference-of-means
Transfer from model
Steps
Single-step
Iterative
Query-reduced gradient estimation
Technique
Random grouping
White-box
PCA
Fast gradient sign (FGS)
Steps
Single-step
Iterative
Single-step
Iterative
Steps
Single-step
Iterative
Loss function
Cross-entropy
Logit-based
Cross-entropy
Logit-based
Loss function
Cross-entropy
Logit-based
Cross-entropy
Logit-based
Loss function
Logit-based
Logit-based
Logit-based
Logit-based
Loss function
Cross-entropy
Logit-based
Cross-entropy
Logit-based
Table 7: Attacks evaluated in this paper
C
S UMMARY OF ATTACKS EVALUATED
Taxonomy of black-box attacks: To deepen our understanding of the effectiveness of black-box
attacks, in this work, we propose a taxonomy of black-box attacks, intuitively based on the number of
queries on the target model used in the attack. The details are provided in Table 7.
We evaluate the following attacks summarized in Table 7:
1. Zero-query attacks
(a) Baseline attacks: Random-Gaussian perturbations (Rand.) and Difference-of-Means
aligned perturbations (D. of M.)
(b) Transferability attack (single local model) using Fast Gradient Sign (FGS) and Iterative
FGS (IFGS) samples generated on a single source model for both loss functions
(Transfer model FGS/IFGS-loss); e.g., Transfer Model A FGS-logit
(c) Transferability attack (local model ensemble) using FGS and IFGS samples generated
on a source model for both loss functions (Transfer models FGS/IFGS-loss); e.g.,
Transfer Model B, Model C IFGS-logit
2. Query based attacks
(a) Finite-difference and Iterative Finite-difference attacks for the gradient estimation
attack for both loss functions (FD/IFD-loss); e.g., FD-logit
(b) Gradient Estimation and Iterative Gradient Estimation with Query reduction attacks
(IGE/GE-QR (Technique-k, loss)) using two query reduction techniques, random
grouping (RG) and principal component analysis components (PCA); e.g., GE-QR
(PCA-k, logit)
3. White-box FGS and IFGS attacks for both loss functions (WB FGS/IFGS (loss))
D
W HITE - BOX ATTACK RESULTS
In this section, we present the white-box attack results for various cases in Tables 8–10. Where
relevant, our results match previous work (Goodfellow et al., 2015; Kurakin et al., 2016).
24
MNIST
White-box
Model
A
B
C
D
Single-step
Iterative
FGS (xent) FGS (logit) IFGS (xent) IFGS (logit)
69.2 (5.9)
90.1 (5.9)
99.5 (4.4)
100.0 (2.1)
84.7 (6.2)
98.8 (6.3) 100.0 (4.75) 100.0 (1.6)
67.9 (6.1)
76.5 (6.6)
100.0 (4.7) 100.0 (2.2)
98.3 (6.3) 100.0 (6.5) 100.0 (5.6) 100.0 (1.2)
CIFAR-10
White-box
Model
Resnet-32
Resnet-28-10
Std.-CNN
Single-step
FGS (xent) FGS (logit)
82.6 (439.7) 86.8 (438.3)
86.4 (439.6) 87.4 (439.2)
93.9 (439.6) 98.5 (429.8)
Iterative
IFGS (xent) IFGS (logit)
100.0 (247.2) 100.0 (66.1)
100.0 (278.7) 100.0 (75.4)
98.0 (314.3) 100.0 (80.3)
Table 8: Untargeted white-box attacks: adversarial success rates and average distortion ∆(X, Xadv )
over the test set. Above: MNIST, = 0.3. Below: CIFAR-10, = 8.
MNIST
White-box
Model
A
B
C
D
Single-step
Iterative
FGS (xent) FGS (logit) IFGS (xent) IFGS (logit)
30.1 (6.1)
30.1 (6.1)
100.0 (4.7)
99.6 (2.7)
29.6 (6.2)
29.4 (6.3)
100.0 (4.6)
98.7 (2.4)
33.2 (6.4)
33.0 (6.4)
100.0 (4.6)
99.8 (3.0)
61.5 (6.4)
60.9 (6.3)
100.0 (4.7)
99.9 (2.0)
CIFAR-10
White-box
Model
Resnet-32
Resnet-28-10
Std.-CNN
Single-step
FGS (xent) FGS (logit)
23.7 (439.6) 23.5 (436.0)
28.0 (439.6) 27.6 (436.5)
43.8 (439.5) 40.2 (435.6)
Iterative
IFGS (xent) IFGS (logit)
100.0 (200.4) 100.0 (89.5)
100.0 (215.7) 100.0 (99.0)
99.0 (262.4) 95.0 (127.8)
Table 9: Targeted white-box attacks: adversarial success rates and average distortion ∆(X, Xadv )
over the test set. Above: MNIST, = 0.3. Below: CIFAR-10, = 8.
MNIST
White-box
Model
Model Aadv-0.3
Model Aadv-ens-0.3
Model Aadv-iter-0.3
Single-step
Iterative
FGS (xent) FGS (logit) IFGS (xent) IFGS (logit)
2.8 (6.1)
2.9 (6.0)
79.1 (4.2)
78.5 (3.1)
6.2 (6.2)
4.6 (6.3)
93.0 (4.1)
96.2 (2.7)
7.0 (6.4)
5.9 (7.5)
10.8 (3.6)
11.0 (3.6)
CIFAR-10
White-box
Model
Resnet-32 adv-8
Resnet-32 adv-ens-8
Resnet-32 adv-iter-8
Single-step
FGS (xent) FGS (logit)
1.1 (439.7) 1.5 (438.8)
7.2 (439.7) 6.6 (437.9)
48.3 (438.2) 50.4 (346.6)
Iterative
IFGS (xent) IFGS (logit)
100.0 (200.6) 100.0 (73.7)
100.0 (201.3) 100.0 (85.3)
54.9 (398.7) 57.3 (252.4)
Table 10: Untargeted white-box attacks for models with adversarial training: adversarial success
rates and average distortion ∆(X, Xadv ) over the test set. Above: MNIST, = 0.3. Below: CIFAR10, = 8.
25
| 1cs.CV
|
Collaborative Large-Scale Dense 3D Reconstruction
with Online Inter-Agent Pose Optimisation
Stuart Golodetz∗
Victor A. Prisacariu
Tommaso Cavallari∗
David W. Murray
Nicholas A. Lord∗
Philip H. S. Torr
arXiv:1801.08361v1 [cs.CV] 25 Jan 2018
Department of Engineering Science, University of Oxford
{smg,tommaso,nicklord,victor,dwm,phst}@robots.ox.ac.uk
Abstract
Reconstructing dense, volumetric models of real-world 3D
scenes is important for many tasks, but capturing large
scenes can take significant time, and the risk of transient
changes to the scene goes up as the capture time increases.
These are good reasons to want instead to capture several
smaller sub-scenes that can be joined to make the whole
scene. Achieving this has traditionally been difficult: joining sub-scenes that may never have been viewed from the
same angle requires a high-quality relocaliser that can cope
with novel poses, and tracking drift in each sub-scene can
prevent them from being joined to make a consistent overall
scene. Recent advances in mobile hardware, however, have
significantly improved our ability to capture medium-sized
sub-scenes with little to no tracking drift. Moreover, highquality regression forest-based relocalisers have recently
been made more practical by the introduction of a method to
allow them to be trained and used online. In this paper, we
leverage these advances to present what to our knowledge is
the first system to allow multiple users to collaborate interactively to reconstruct dense, voxel-based models of whole
buildings. Using our system, an entire house or lab can be
captured and reconstructed in under half an hour using only
consumer-grade hardware.
Figure 1: A globally consistent collaborative reconstruction
of a three-storey house, as produced by our approach. The
reconstruction involved relocalising 6 separate sequences
from the Priory subset of our dataset.
pensating for tracking drift [17, 46, 47, 24, 14]. However,
even with the most sophisticated voxel-based approaches
[24, 14], capturing the data needed to reconstruct a large
scene (e.g. a whole building) can take significant time and
planning, and require considerable concentration on the part
of the user. Moreover, the risk of transient changes to the
scene (e.g. people moving around) goes up as the capture
time increases, corrupting the model and forcing the user to
restart the capture. There are thus good reasons to want to
split the capture into several shorter sequences, which can
be captured either over multiple sessions or in parallel (by
multiple users) and then joined to make the whole scene.
Achieving this has traditionally been difficult: joining
the sub-scenes requires the ability to accurately determine
the relative transformations between them (the camera relo-
1. Introduction
Reconstructing dense, volumetric models of real-world 3D
scenes is an important task in computer vision and robotics,
with applications in content creation for films, games and
augmented reality [23], cultural heritage preservation [48]
and building information modelling [32]. Since the seminal
KinectFusion work of Newcombe et al. [33], which demonstrated real-time voxel-based reconstruction of a desk-sized
scene in real time using only a consumer-grade RGB-D sensor, huge progress has been made in increasing the size of
scene we are able to reconstruct [34, 36, 37] and in com∗ S.
Golodetz, T. Cavallari and N. Lord assert joint first authorship.
1
Figure 2: More examples of our approach, showing the Flat, House and Lab subsets of our dataset.
calisation problem), even though the areas in which they
overlap may never have been viewed from the same angles, and tracking drift in each sub-scene can prevent them
from being joined to make a consistent overall scene. Recent advances in mobile hardware, however, such as the
introduction of augmented reality smartphones that estimate their pose using visual-inertial odometry, have significantly improved our ability to capture medium-sized subscenes with little to no drift. Moreover, in RGB-D relocalisation, keyframe-based relocalisers such as random ferns
[20], which were previously widely used for relocalisation
in a single-user context but were unable to relocalise from
novel poses, have recently been giving way to high-quality
regression-based methods such as SCoRe forests [43, 45],
driven by recent work [2] that showed how they could be
trained and used online. Unlike keyframe-based methods,
such approaches have been shown to be much better suited
to relocalisation from novel poses, which is critical when
aligning sub-scenes captured from different angles.
In this paper, we leverage these advances to present
what to our knowledge is the first system to allow multiple users to collaborate interactively to reconstruct dense,
voxel-based models of whole buildings. Unlike previous
multi-agent mapping approaches, our approach is able to
reconstruct detailed, globally consistent dense models. Using our system, an entire house or lab can be captured and
reconstructed in under half an hour using only consumergrade hardware (see §4.5). We have integrated our approach
into the open-source SemanticPaint framework [22, 21],
making it easy for existing SemanticPaint and InfiniTAM
[37] users to benefit from our work, and constructed a new
dataset of sub-scenes to validate our approach, and we make
both our code and this dataset available online. Figures 1
and 2 show reconstructions produced by our approach.
approaches can be divided into two main categories:
2. Related Work
(ii) Centralised approaches, by contrast, can take advantage of the computing power provided by one or more central servers and their ability to communicate with all agents
at once to produce detailed, globally consistent maps. For
example, Chebrolu et al. [3] described a semi-dense approach based on LSD-SLAM [15] in which monocular
Although previous work on multi-agent mapping has
not focused on dense reconstruction, multi-agent mapping
itself has a rich research history in computer vision and
robotics, and several good surveys exist [40, 41]. Existing
(i) Decentralised approaches eschew the use of a central
server and instead produce a local map of the scene on each
agent, often transmitting these local maps between agents
when they meet to share knowledge of parts of the scene
that individual agents have not yet visited. For example,
Cunningham et al. [12] proposed an approach called DDFSAM, in which each robot produces a landmark-based map
and shares compressed, timestamped versions of it with
neighbouring robots. This was extended by [13], which
registered landmark-based maps together using an approach
based on Delaunay triangulation and RANSAC. The same
lab later proposed DDF-SAM 2.0 [11], which avoids repeated and expensive recreation of the combined neighbourhood map through the use of ‘anti-factors’. Cieslewski
et al. [8] presented a sophisticated decentralised collaborative mapping back-end based on distributed version control.
Choudhary et al. [6] described an approach that performs
decentralised mapping using object-based maps, decreasing the bandwidth required for map sharing, but depending
on the existence of pre-trained objects in the scene. Most
recently, Cieslewski et al. [7] aimed to minimise the bandwidth each robot uses for inter-agent relocalisation by first
establishing which other robots have relevant information,
and then only communicating with the best robot found.
Decentralised approaches have numerous applications, including search and rescue [30], agricultural robotics [4],
planetary exploration [1] and underwater mapping [35, 44],
but because of the limited computing power that tends to be
available on mobile agents, most existing approaches target robustness to unreliable network connections and mechanical failures, rather than reconstructing detailed scene
geometry, limiting their usefulness for tasks like building
information modelling or cultural heritage preservation.
reads frames
Mapping
Component 1
updates
Voxel Scene
Agent 1
Mapping Server
...
...
Local Relocaliser
accesses scenes
...
Mapping
Component N
Agent N
Scene Renderer
uses
Voxel Scene
reads poses from
Local Relocaliser
reads trajectories from
accesses scenes
and relocalisers
Pose Graph
Optimiser
~ f1
1 T2
~ f3
1T2
~
T
f4
1 2
Trajectory 1
...
Candidate
Relocalisation
Selector
feeds
Inter-Agent
Relocaliser
~
T
f2
1 2
~
T
f5
1 2
operates on
Pose Graph
...
adds relative transform samples
1T2
gT2
gT1
gT5
1T3
Trajectory N
~
T
f3
1 4
considers existing clusters
~
T
~ f2
1T4
~ f4
1T4
f1
1 4
1T4
gT4
3T4
gT3
3T5
3T6
gT6
Figure 3: The architecture of our system. Individual agents (whether local or remote) track their poses and feed posed RGB-D
frames to the mapping server. A separate mapping component is instantiated for each agent, which reconstructs a voxel scene
and trains a local relocaliser. Separately, the candidate relocalisation selector repeatedly selects a pose from one of the agents’
trajectories for relocalisation against another scene. The inter-agent relocaliser uses the scene renderer to render synthetic
RGB and depth images of the corresponding scene from the selected pose, and passes them to the local relocaliser of the
target scene. If a relocalisation succeeds (and is verified, see §3.3.1), a sample of the relative transform between the two
scenes is recorded. The relative transform samples for each scene pair are clustered for robustness (see §3.3.2). Whenever
the cluster to which a sample is added is sufficiently large, we construct a pose graph by blending the relative poses in the
largest clusters, and trigger a pose graph optimisation. The optimised poses are then used for rendering the overall scene.
clients produced pose-graph maps of keyframes that were
sent to a central server for optimisation. Riazuelo et al.
[39] described a cloud-based distributed visual SLAM system inspired by PTAM [26]. Their clients performed tracking only, with the more expensive mapping steps being performed in the cloud. Mohanarajah et al. [31] described another cloud-based approach based on the authors’ Rapyuta
robotics platform. The client robots estimated their local poses by running dense, keyframe-based visual odometry on RGB-D images from a PrimeSense camera. The
keyframes were then centrally optimised using g2 o [27].
Forster et al. [19] demonstrated centralised, keyframe-based
collaborative mapping from micro aerial vehicles (MAVs),
each equipped with a monocular camera and an IMU. More
recently, Schmuck and Chli [42] have shown how to incorporate server-to-client feedback into a multi-UAV collaborative approach, allowing agents to share information.
A few approaches do not fit cleanly into either category.
Reid et al. [38] described a distributed approach in which
multiple autonomous ground robots were controlled from a
centralised ground control station (GCS), but were able to
operate independently for a while if the connection to the
GCS failed. McDonald et al. [29] described a stereo approach in which a single agent reconstructed a scene over
multiple sessions, effectively collaborating with itself over
time. Chen et al. [5] described an approach that is initially
peer-to-peer, with each robot building a pose graph independently, storing its sensor data, and exchanging information
with other robots on rendezvous, but later client-server, with
the robots transferring their pose graphs and sensor data to
a central server for pose graph optimisation and the building of a coloured point cloud map. Fankhauser et al. [16]
described an approach that allowed an Asctec Firefly hexacopter and a quadrupedal ground robot to work together to
help the ground robot to navigate safely. Whilst no fixedposition central server was used, the ground robot had significant computing power and effectively played the role
of a server, performing bundle adjustment on data received
from the hexacopter to maintain a globally consistent map.
3. Our Approach
Since we target scenarios such as building information
modelling, where hardware failure is a minor concern, we
adopt a centralised approach with a number of lightweight
mobile clients and a powerful central server (e.g. a laptop
or desktop with one or more high-end GPUs): see Figure 3.
Each client estimates accurate local poses for a sequence of
RGB-D frames (see §3.1) and transmits both the frames and
the poses to the central server (see §3.2). The server constructs a sub-scene and trains a relocaliser for each client,
and relocalises the sub-scenes from different clients against
each other to establish a consistent global map (see §3.3).
Our system can run in either interactive or batch mode.
In batch mode, all of the sub-scenes are sent across to the
server (or simply read directly from disk) before any relocalisation is performed. In interactive mode, the server will
start relocalising clients against each other immediately, and
new clients can join on the fly to contribute to the map.
3.1. Local Tracking (Client-Side)
As mentioned above, each client in our system must estimate accurate local poses for a sequence of RGB-D frames
and then transmit both the frames and the poses to the central server. This is non-trivial, since traditional pose estimation approaches, particularly those based purely on visual tracking, have tended to be subject to significant tracking drift, particularly at larger scales. For collaborative
mapping, a common solution has been for clients to simply transmit inaccurate local poses and rely on the server
to perform global optimisation (e.g. pose graph optimisation between keyframes [38, 5, 31] or bundle adjustment
[3, 18, 42]) to achieve a globally consistent map (the optimised poses can then be sent back to the clients if desired
[3, 31, 42]). However, (i) such global optimisations do not
scale well, limiting the overall size of map that can be constructed, and (ii) the local poses on the clients are only corrected as global optimisations on the server finish, meaning
that for much of the time they cannot be fully trusted.
In our system, we instead leverage recent developments
in consumer-grade augmented reality hardware to place the
burden of accurate local pose estimation on the clients, freeing the server to focus on reconstructing a dense global map
in real time. In particular, we use a straightforward client
that captures the RGB-D images and poses provided by the
hardware (based on visual-inertial odometry) and transmits
them to the server (see §3.2). This important choice allows us to significantly simplify the design of the server (see
§3.3) and reduce server-side memory consumption, making
it possible to support more agents and larger global maps.
network bandwidth clients use, we only transmit frames that
were tracked successfully, and we compress the depth images in PNG format (lossless) and the RGB images in JPG
format (lossy). Moreover, to maintain a smooth, interactive
user experience on the client, we transmit messages containing the frames (and their accompanying poses) to the
server on a separate thread, which iteratively reads a frame
message from a pooled queue of reusable messages, compresses it and sends it to the server.1 The main thread writes
uncompressed frame messages into this queue based on the
current input. We discard messages that would overflow the
queue (e.g. when the network is too slow to keep up with the
client) to maintain interactivity and bound client memory
usage; the way in which this interacts with our compression
strategy is evaluated in the supplementary material.
At the server end, each client has a handler (running on
a separate thread) that maintains a pooled queue of uncompressed frame messages. When a compressed frame message arrives, it is immediately uncompressed and pushed
onto the queue (as on the client, we discard messages that
would overflow the queue). On the main thread, we run
a mapping component for each client (see §3.3), which
reads RGB-D frames and their accompanying poses from
its client’s queue as necessary and creates a local map.
3.3. Global Mapping (Server-Side)
3.2. RGB-D Frame Transmission (Client → Server)
The server has two jobs: (i) constructing a sub-scene and
training a relocaliser for each client, and (ii) determining
the relative transforms between these sub-scenes to establish a global coordinate system. To achieve the first, a separate mapping component for each client (i) runs the opensource InfiniTAM reconstruction engine [36, 37] on incoming (posed) RGB-D frames to construct a voxel-based map,
and (ii) trains a regression forest-based relocaliser for the
sub-scene online as per Cavallari et al. [2]. To achieve the
second, the server attempts to relocalise synthetic images
of one agent’s sub-scene using another agent’s relocaliser
to find estimates of the relative transform between the subscenes (see §3.3.1). These samples are clustered in transformation space to help suppress outliers, and a pose graph
is constructed and optimised in the background to further
refine the relative transforms between the sub-scenes (see
§3.3.2). This optimisation is inspired by the single-agent
sub-mapping approach of Kähler et al. [24], which showed
how to build globally consistent voxel-based models by dividing a scene into small sub-scenes and optimising the relative poses between them. However, we construct a pose
graph where each agent is represented by a single node,
and the edges denote the relative transforms between different agents’ sub-scenes. This differs from [24], where all
sub-scenes came from one agent and no relocalisation was
We use TCP between each client and the server to guarantee in-order delivery of RGB-D frames. To minimise the
1 Further details about the design of the pooled queue data structure can
be found in the supplementary material.
needed to establish the transforms between them.
3.3.1
Inter-Agent Relocalisation
To maintain a smooth, interactive experience on the server,
we attempt relocalisations between the sub-scenes of different clients on a separate thread and (if available) a separate
GPU. The way in which relocalisation attempts are scheduled depends on the mode in which the server is running.
In batch mode, relocalisations are attempted only once all
of the client sub-scenes have been fully created, at which
point they are attempted as quickly as possible (i.e. a new
relocalisation attempt is scheduled as soon as the previous
attempt finishes). In interactive mode, we relocalise whilst
the client sub-scenes are still being reconstructed, but at
most once every 20 frames: this is because relocalisers cannot be trained online and used simultaneously, so there is a
need to space out relocalisation attempts to allow sufficient
time in between attempts for the relocalisers to be trained.
To schedule an attempt, we first randomly generate a list
of 10 candidate relocalisations. Each candidate k = (a, f ),
where f = (b, i), denotes an attempt to relocalise frame i
of scene b in scene a’s coordinate system. To balance different pairs of scenes, which may have been reconstructed
from varying numbers of frames, we first uniformly sample an (a, b) scene pair from the set of all scene pairs, and
then uniformly sample a frame index i from scene b. Each
generated candidate k = (a, (b, i)) is then scored via
Φ(k) = φnew (k) − φconf (k) − φhomog (k).
(1)
In this, φnew aims to give a boost to candidates that might
connect new nodes to the pose graph, defined as 1 if one of
the sub-scenes a and b already has an optimised global pose
and the other does not, and 0 otherwise. φconf penalises
candidates that can only add a relative transform sample between a pair of sub-scenes that are already confidently relocalised with respect to each other:
φconf (a, (b, i)) = max
0,
max
(|c| − N ) .
c∈clusters(a,b)
(2)
Here, N = 2 is an empirically-chosen threshold on the
number of relocalisations needed before we become confident that the relative transform between a pair of sub-scenes
is correct. φhomog penalises candidates whose frame i has
a local pose in scene b that is too close to (within 5cm / 5◦
of) one that has already been tried against scene a; we use a
penalty of 5 for poses that are too close, and 0 otherwise.
Having scored the candidates, we schedule a relocalisation attempt for a candidate k = (a, f ) with maximum
score, where f = (b, i). This will proceed as shown in Figure 4. Let f T̃s denote the pose of frame f in the coordinate
system of a sub-scene s. First, we render synthetic RGB
and depth raycasts of sub-scene b from the known pose f T̃b ,
which we will call C̃bf and D̃bf . Then, we try to relocalise
them using sub-scene a’s relocaliser to obtain an estimated
pose f T̃a for f in a’s coordinate system. To verify this estimated pose, we first render a synthetic depth raycast D̃af
of sub-scene a from f T̃a , and compute a masked absolute
depth difference image ∆D between D̃af and D̃bf via
(
∆D(x) =
D̃af (x) − D̃bf (x)
0
if D̃af (x) > 0 and D̃bf (x) > 0
, (3)
otherwise
f
f
where x ranges over
PΩ, the domain of D̃a and D̃b . We
then compute µ = ( x∈Ω ∆D(x))/|Ω|, and add a relative
transform sample a T̃bf = (f T̃a )−1 f T̃b between a and b iff
µ < 5 (corresponding to 5cm). The effectiveness of this
verification step is evaluated in §4.2.
3.3.2
Inter-Agent Pose Optimisation
We incrementally cluster the relative transform samples
{a T̃bf } we add for each pair of sub-scenes (a, b) prior to
performing pose graph optimisation to suppress the effect
of outliers on the final result. After adding each sample,
we look to see if the cluster to which the sample has just
been added now has ≥ N samples (i.e. the sample has
contributed to a confident relative transform): if so, it is
worthwhile to run pose graph optimisation, since the pose
graph we construct may have changed since the last run. To
construct the pose graph, we first compute a binary relation R(a, b) between sub-scenes such that R(a, b) is true iff
the largest cluster between a and b has size ≥ N (i.e. there
is a confident relative transform between a and b). Next,
we compute the reflexive, transitive closure R∗ of this relation, such that R∗ (a, b) is true iff there is a chain (possibly empty) of confident relative transforms between a and
b. Finally, we denote the scene corresponding to the first
agent as the primary scene, and add to the pose graph a
node for each scene that is confidently connected to the primary scene by R∗ . To determine the edges to add to the
graph, we filter a list containing each pair of sub-scenes for
pairs (i) whose largest cluster has size ≥ N , and (ii) such
that both sub-scenes are confidently connected to the primary scene. For each surviving pair of sub-scenes (a, b),
we then blend the relative transform samples {a T̃bf } in the
largest cluster using dual quaternion blending [25] to form
an overall estimate a T̃b of the relative transform from b to a,
and add an edge to the graph. By adding only one edge for
each scene pair, we keep the pose graph we are optimising
small, allowing the optimisation to be run repeatedly in the
background to optimise the overall map.
The goal of optimising a pose graph G is to find an optimised global pose g T̂a for each scene a for which there
is a node in G. To perform the optimisation, we use the
approach of Kähler et al. [24], as implemented in the opensource InfiniTAM reconstruction engine [36, 37]. This uses
Failed
Renderer
Verifier
Renderer
Rejected
Figure 4: To relocalise the scene of an agent b against that of another agent a, we first choose an arbitrary frame from b’s
trajectory and render synthetic RGB and depth raycasts of b’s scene from the frame’s pose. Then, we try to relocalise them
using a’s relocaliser, which either fails, or produces an estimated pose for the frame in a’s coordinate system. If a pose is
proposed, we verify it by rendering a synthetic depth raycast of a’s scene from the proposed pose and comparing it to the
synthetic depth raycast of b’s scene. We accept the pose iff the two depth raycasts are sufficiently similar (see §3.3.1).
Levenberg-Marquardt to minimise the error function
X
(G) =
(a,b)∈edges(G)
v(g T̂b−1 g T̂a a T̃b )
2
,
(4)
where v(T ) = (q̂(T ), t(T ))> denotes the concatenation of
q̂(T ), the three imaginary components of a quaternion representing the rotational part of T , and t(T ), the translational
part of T . Implicitly, this optimisation is trying to achieve
−1
g T̂a
g T̂b = a T̃b for every a and b, i.e. to ensure that the
optimised global poses of all the scenes are consistent with
the estimated relative poses between them.
4. Experiments
We perform both quantitative and qualitative experiments to evaluate our approach. In §4.1, we evaluate the
quality of a reconstruction we obtain with our collaborative
approach by comparing a variety of measurements made on
a combined mesh of the scene with ground-truth measurements made with a laser range finder in the real world. In
§4.2, we evaluate the effectiveness of our depth differencebased verification approach (see §3.3.1), showing that our
verifier is able to prune large numbers of incorrect relative
transforms between the sub-scenes whilst rejecting practically no correct transforms. In §4.3, we demonstrate the
scalability of our approach, showing that by using synthetic
images for relocalisation and discarding data from the relocalisers we use when it is no longer needed, we can support up to 11 agents on a high-end PC. In §4.4, we evaluate
our strategy of using synthetic scene raycasts for relocalisation rather than real images from the original sequences,
by testing both approaches on the standard 7-Scenes relocalisation dataset of Shotton et al. [43]. Finally, in §4.5, we
time how long our approach takes to produce consistent reconstructions for four different subsets of our dataset. The
supplementary material contains further analysis.
4.1. Reconstruction Quality
To demonstrate our ability to achieve high-quality collaborative reconstructions, we combined 4 sequences from
the Flat subset of our dataset to reconstruct a combined map
of a two-bedroom flat (see Figure 5). Since a ground-truth
reconstruction of the flat was not available (and we did not
have access to a LIDAR scanner with which to obtain one),
we validated our reconstruction by comparing a variety of
measurements made on our combined map with groundtruth measurements made with a laser range finder (a Bosch
Professional GLM 40) in the real world. To achieve this, we
first converted the voxel-based maps of the sub-scenes produced by InfiniTAM [36, 37] into mesh-based maps using
Marching Cubes [28]. We then applied the relative transforms we had estimated between the sub-scenes during the
relocalisation process to these meshes to transform them
into a common coordinate system. Finally, we imported the
transformed meshes into MeshLab [9], and used its pointto-point measurement tool to make our measurements. As
shown in Figure 5, we found that the measurements on our
Figure 6: The GPU memory we use during a collaborative
reconstruction of the 6 Priory sequences from our dataset.
Figure 5: An example of the high-quality reconstructions
we can achieve using our collaborative approach. We joined
the 4 Flat sequences from our dataset to make a combined
map of a two-bedroom flat. The purple lines show the measurements we performed both using a laser range finder and
on a combined mesh of our model to validate our approach
(the ordering of the numbers is range finder then mesh).
reconstructed model were consistently within 5cm of the
ground truth, indicating that we are able to achieve reconstructions that correspond relatively well to real-world geometry. Further examples, this time showing collaborative
reconstructions of multi-storey houses and a large research
lab, can be found in the supplementary material.
4.2. Effectiveness of Depth Difference Verification
As mentioned in §3.3.1, we verify each proposed interagent relocalisation of a sub-scene b against a sub-scene a
by rendering a synthetic depth image of scene a from the
proposed pose and comparing it to the synthetic depth image from scene b that was passed to the relocaliser, masking
out pixels that do not have a valid depth in both images. To
evaluate the effectiveness of this approach, we first took 8
pairs of sequences {a, b} from our dataset that we were able
to successfully relocalise with respect to each other during
the normal operation of our approach, and recorded the relative transform between them as the ‘ground truth’ transform
for later use. We then attempted to relocalise every frame
from scene b using the relocaliser of scene a, and vice-versa.
Next, for each frame for which the relocaliser proposed a
relative transform, we ran our verification step on the proposed transform, thereby classifying it as either Verified or
Rejected. We then compared the proposed transform to the
‘ground truth’ transform, classifying it as Correct if it was
within 5◦ and 5cm of the ground truth, and as Incorrect
otherwise. Finally, we counted Verified/Correct transforms
as true positives (TP), Verified/Incorrect transforms as false
positives (FP), and similarly for the true negatives (TN) and
false negatives (FN), as shown in Table 1.
Our results show that our verifier has an extremely high
average recall rate (99.2%), meaning that it largely manages to avoid rejecting correct transforms. It also has a
reasonably good average specificity (75.9%), meaning that
it is fairly good at pruning the number of incorrect transforms we need to deal with. However, a fairly large number
of incorrect transforms still manage to pass the verification
stage: as mentioned in §3.3.2, these are dealt with later by
clustering the transforms and only making use of the transforms from the largest cluster. The additional effects of this
clustering step are shown in the supplementary material.
4.3. Scalability (Batch Mode)
To evaluate the scalability of our approach, we perform
separate experiments for the batch and interactive modes of
our system, since they target slightly different applications.
The interactive experiment is in the supplementary material.
Our batch mode experiment focuses on the server’s GPU
memory usage as the number of agents increases. To evaluate this, we performed a collaborative reconstruction of the
6 Priory sequences from our dataset (these represent a threestory house: see the supplementary material). To maximise
the number of agents we were able to handle, we added new
sub-scenes to the server one at a time and deleted the training data used by each sub-scene’s relocaliser once it had
been fully trained: this limited the maximum memory used
by the relocalisers, leaving us bounded primarily by the size
of the reconstructed voxel scenes. Figure 6 shows that the
final GPU memory we use for each forest-based relocaliser
[2] is < 500MB, meaning that we can potentially handle
around 24 relocalisers on a 12GB GPU such as an NVIDIA
Titan X (the actual number may be slightly less than this
due to driver overhead). Moreover, since the relocalisers we
use have the same structure for each scene [2], they could
potentially be merged to produce a global relocaliser and
reduce memory usage still further. The memory used by the
voxel scenes is currently the bottleneck preventing the scal-
Scene 1
Scene 2
Total Frames
Relocalised Frames
TP
FP
TN
Verifier Performance
FN
Precision
Recall
Specificity
f/bathroom2study
f/bathroom2study
f/kitchen2study
f/study2sittingroom
f/kitchen2study
f/study2sittingroom
f/study2sittingroom
f/turret2sittingroom
4901
4989
5544
5895
2190
519
1254
1350
517
24
325
555
931
59
317
211
741
436
609
579
1
0
3
5
35.7%
28.9%
50.6%
72.5%
99.8%
100.0%
99.1%
99.1%
44.3%
88.1%
65.8%
73.3%
h/frontroom2study
h/frontroom2study
h/hall2frontbedroom
h/hall2frontbedroom
h/hall2frontbedroom
h/hall2oldkitchen
h/hall2oldkitchen
h/mainbedroom2studio
7714
8171
7141
6942
1229
1307
653
839
245
373
95
90
90
175
106
131
894
757
451
615
0
2
1
3
73.1%
68.1%
47.3%
40.7%
100.0%
99.5%
99.0%
96.8%
90.9%
81.2%
81.0%
82.4%
–
–
–
–
–
–
52.1%
99.2%
75.9%
Average (all scene pairs)
Table 1: Evaluating the effectiveness of the depth difference verification we perform on proposed relocalisations between
pairs of sub-scenes (see §4.2). We attempt to relocalise every frame of each sub-scene against the other using [2], and record
the number of frames that we were able to relocalise, together with statistics on how many correct/incorrect relocalisations
were verified/rejected by the verifier.
ing of our approach to more agents: currently, each scene
takes up just over 1GB of memory, limiting us to around
11 agents on a Titan X (assuming the relocalisers are stored
on a secondary GPU). To scale further, we could reduce the
memory used by meshing each scene with Marching Cubes
[28] and discarding the voxel maps.
4.4. Relocalisation from Synthetic Images
As per §3.3.1, we relocalise the scenes of different agents
against each other using synthetic images, rather than the
real RGB-D frames originally captured by the agents. This
avoids the prohibitive memory cost of storing all RGB-D
frames acquired by each agent in RAM (easily hundreds of
MBs per agent). However, we might expect using synthetic
images to lower relocalisation performance, since we train
the relocalisers for each scene using the real input frames.
To verify that this is not a problem, we compared the results we are able to obtain using synthetic images to results
obtained using the approach of Cavallari et al. [2] with real
images on the standard 7-Scenes dataset [43]. Unlike [2],
who used at most 10 modes in each leaf of their regression
forests, we used at most 50 modes, since we found that this
gave better results (in both cases). To test our synthetic approach, we first reconstructed each 7-Scenes sequence from
the real training images as normal, then rendered synthetic
frames from the testing poses, rather than using the testing
images in the dataset. As Table 2 shows, our results using synthetic images were at least as good as, and in most
cases actually higher than, the results using the real images,
verifying that using synthetic images does not decrease relocalisation performance in practice. This is likely due to
the fact that by rendering synthetic images from the reconstructed scene, we implicitly remove noise from the frames
to be relocalised, improving pose estimation accuracy.
4.5. Timings
To evaluate how long it takes to produce consistent reconstructions using our approach, we computed the aver-
Sequence
Real Images
Synthetic Images
Chess
Fire
Heads
Office
Pumpkin
RedKitchen
Stairs
99.85%
99.20%
100.00%
99.80%
90.10%
91.34%
78.30%
100.00%
98.85%
100.00%
99.95%
93.00%
99.98%
88.10%
Table 2: Comparing relocalisation results obtained by rendering synthetic images of the scenes in the 7-Scenes
dataset [43] from the test poses to those obtained using the
real test images, by adapting a regression forest pre-trained
on Office [2]. Percentages denote proportions of test frames
with ≤ 5cm translation error and ≤ 5◦ angular error.
Flat
Priory
House
Lab
# Sequences
Longest Sequence: # Frames
Longest Sequence: Time (s)
Average Mapping Time (s)
4
3079
615.8
15.8
6
1640
328.0
38.4
4
4372
874.4
107.4
4
3188
637.6
272.0
Average Total Time (s)
631.6
366.4
981.8
909.6
Table 3: The times taken to collaboratively reconstruct four
different subsets of our dataset (see §4.5).
age times taken to collaboratively reconstruct four different
subsets of our dataset (see Table 3). We computed the time
taken to capture the sequences in each subset as 51 times the
length of the longest sequence, assuming parallel capturing
at 5Hz. The average mapping time is the time taken to relocalise the agents against each other and compute optimised
global poses for their maps. To account for the random selection of frames to relocalise, we globally mapped each
subset 5 times and reported the average time in the table.
The average total time is from the start of the capturing process to the output of a globally consistent map: this was
under half an hour for all subsets we tested.
5. Conclusion
In this paper, we have shown how to collaboratively reconstruct dense, volumetric models of 3D scenes using multiple agents. Existing collaborative mapping approaches
have traditionally suffered from an inability to trust the local poses produced by their mobile agents, forcing them to
perform costly global optimisations (e.g. on a server) to ensure a consistent map, and limiting their ability to perform
dense, volumetric mapping collaboratively.
By leveraging recent mobile hardware advances to construct rigid local sub-scenes that do not need further refinement, and joining them using a state-of-the-art regressionbased relocaliser, we avoid expensive global optimisations,
opting only to refine the relative poses between individual
agents’ overall maps. Our system allows multiple users to
collaboratively reconstruct consistent dense models of entire buildings in under half an hour using only consumergrade hardware, making it easier than ever before for users
to capture detailed 3D scene models at scale.
Acknowledgements
This work was supported by Innovate UK/CCAV project
#103700 (StreetWise), the EPSRC, ERC grant ERC2012-AdG 321162-HELIOS, EPSRC grant Seebibyte
EP/M013774/1 and EPSRC/MURI grant EP/N019474/1.
We would also like to thank Manar Marzouk and Maria
Anastasia Tsitsigkou for their help with the collaborative
dataset collection.
S UPPLEMENTARY M ATERIALS
A. Our Dataset
Our dataset comprises 4 different subsets – Flat, House,
Priory and Lab – each containing a number of different sequences that can be successfully relocalised against each
other. The name of each sequence is prefixed with a simple identifier indicating the subset to which it belongs: f/
for Flat, h/ for House, p/ for Priory and l/ for Lab. Basic
information about the sequences in each subset, their frame
counts and capture times can be found in Table 4. Illustrations of the sequences and how they fit together to make the
combined scenes are shown in Figures 7 to 10.
Each sequence was captured at 5Hz using an Asus ZenFone AR augmented reality smartphone, which produces
depth images at a resolution of 224 × 172, and colour images at a resolution of 1920 × 1080. To improve the speed
at which we were able to load sequences from disk, we resized the colour images down to 480 × 270 (i.e. 25% size)
to produce the collaborative reconstructions we show in the
paper, but we nevertheless provide both the original and resized images as part of the dataset. We also provide the
calibration parameters for the depth and colour sensors, the
6D poses for both sensors at each frame, and the optimised
global pose produced for each sequence when running our
approach on all of the sequences in each subset. Finally, we
provide a pre-built mesh of each sequence, pre-transformed
by its optimised global pose to allow the sequences from
each subset to be loaded into MeshLab [9] or CloudCompare [10] with a common coordinate system.
B. Additional Experiments
In this section, we describe some additional experiments
that we performed to evaluate our method. In §B.1, we evaluate the extent to which our approach of incrementally clustering the relative transform samples between different pairs
of sub-scenes is able to remove outliers and find consistent
relocalisations between the sub-scenes. In §B.2, we evaluate the scalability of our approach for interactive collaborative reconstruction. Finally, in §B.3, we evaluate the impact
of frame compression on the final reconstruction quality we
are able to achieve on the server. We show that with compression enabled, we can maintain real-time mapping whilst
discarding fewer frames, allowing us to reconstruct higherquality models.
Sequence
Frame Count
Capture Time (s)
f/bathroom2study
f/kitchen2study
f/study2sittingroom
f/turret2sittingroom
2173
2728
2816
3079
434.6
545.6
563.2
615.8
h/frontroom2study
h/hall2frontbedroom
h/hall2oldkitchen
h/mainbedroom2studio
4372
3342
3798
3600
874.4
668.4
759.6
720.0
p/bath2office
p/bed2office
p/dining2guest
p/guest2bath
p/kitchen2dining
p/living2dining
1319
1518
1232
1315
1142
1640
263.8
303.6
246.4
263.0
228.4
328.0
l/atrium
l/firstfloor
l/groundfloor
l/secondfloor
1709
3188
1985
2146
341.8
637.6
397.0
429.2
Table 4: The sequences in each subset of our dataset.
B.1. Effectiveness of Relative Transform Clustering
As described in §3.3.2 of the main paper, we incrementally cluster the relative transform samples {a T̃bf } we add
for each pair of sub-scenes (a, b) prior to performing pose
graph optimisation to suppress the effect of outliers on the
final result. This is achieved by checking each new relative transform sample to see if there is an existing cluster to
which it can be added (we specify that this is possible iff it
is within 20◦ and 10cm of any existing relative transform in
the cluster). If so, we add the sample to the first such cluster
we find; if not, we create a new cluster for it.
To evaluate the effectiveness of this approach, we took
the same 8 pairs of sequences {a, b} we used to evaluate our
depth difference verifier in §4.2 of the main paper, and again
relocalised every frame from scene b using the relocaliser
of scene a, and vice-versa. We then counted the number of
relative transform samples that were added during this process, and examined the clusters into which they had been
collected. In particular, we compared the size of the largest
cluster in each case (i.e. the largest ‘correct’ cluster) with
the size of the largest cluster whose blended transform (obtained by blending all of the relative transforms in the cluster using dual quaternion blending [25]) was not within 20◦
and 10cm of the blended transform of the correct cluster.
We refer to this latter cluster as the largest ‘incorrect’ cluster. The difference between these two sizes gave us a measure of the safety margin of our approach in each case, i.e.
Figure 7: A collaborative reconstruction of the 4 Flat sequences in our dataset (f/turret2sittingroom, f/study2sittingroom,
f/kitchen2study and f/bathroom2study). These collectively represent a single-storey, two-bedroom flat. The mono-colour
images show the individual sub-scenes we reconstruct from each sequence; the other images show the combined map.
Figure 8: A collaborative reconstruction of the 4 House sequences in our dataset (h/frontroom2study, h/hall2frontbedroom,
h/hall2oldkitchen and h/mainbedroom2studio). These collectively represent a two-storey, four-bedroom house. The monocolour images show the individual sub-scenes we reconstruct from each sequence; the other images show the combined map.
Figure 9: A collaborative reconstruction of the 6 Priory sequences in our dataset (p/bath2office, p/bed2office, p/guest2bath,
p/dining2guest, p/kitchen2dining and p/living2dining). These collectively represent a three-storey house. The mono-colour
images show the individual sub-scenes we reconstruct from each sequence; the other images show the combined map.
the number of consistent erroneous samples that would need
to have been added to the largest incorrect cluster to cause
it to have been chosen instead of the correct cluster.
As our results in Table 5 show, for most pairs of
scenes the size of the correct cluster was significantly
larger than the size of the largest incorrect cluster, indicating that in practice we are very likely to accumulate ≥ N samples from the correct cluster and become confident about it long before we accumulate
≥ N samples from an incorrect cluster. For two pairs
of scenes (f/bathroom2study & f/study2sittingroom, and
h/hall2frontbedroom & h/mainbedroom2studio), the safety
margins were much lower than in the other cases. However,
in both cases, the pairs of scenes in question have comparatively low overlap (see the green and yellow sequences in
both Figures 7 and 8). Moreover, whilst the blended transforms of the correct cluster and the largest incorrect cluster
in each case were not within 20◦ and 10cm of each other,
a manual inspection of the relevant transforms showed that
they were still comparatively close (within 50◦ and 50cm),
meaning that the safety margin before hitting a cluster with
a grossly incorrect blended transform is in practice somewhat higher in each case.
B.2. Scalability (Interactive Mode)
To evaluate the scalability of our approach in interactive mode, we time a collaborative reconstruction involving
a server and three different clients: (i) a desktop using a
Kinect camera and a relocaliser based on random ferns [20],
connected to the server over a wired network, (ii) a laptop
using an Orbbec Astra S camera and a regression forest relocaliser [2], connected to the server over WiFi, and (iii) a
local process on the same machine as the server that reads
posed RGB-D frames from disk.
The frame processing times on the server and each client
over the course of this experiment, together with a detailed
description of what is happening at each stage during the
process, are shown in Figure 11. As might be expected,
the time taken per frame on each client is unaffected by the
number of other agents connected to the server, allowing
local reconstruction to remain interactive even when many
clients are connected. Moreover, the time taken per frame
on the server increases by only a small amount for each additional agent added, allowing the server user to continue
to view the collaborative reconstruction interactively from
different angles even when multiple different clients are
connected. With a large number of clients connected, the
Figure 10: A collaborative reconstruction of the 4 Lab sequences in our dataset (l/groundfloor, l/firstfloor, l/atrium and
l/secondfloor). These collectively represent a three-storey research lab. The mono-colour images show the individual subscenes we reconstruct from each sequence; the other images show the combined map.
Scene 1
Scene 2
Total Frames
Samples Added
Correct Cluster
Largest Incorrect Cluster
Safety Margin
f/bathroom2study
f/bathroom2study
f/kitchen2study
f/study2sittingroom
f/kitchen2study
f/study2sittingroom
f/study2sittingroom
f/turret2sittingroom
4901
4989
5544
5895
1472
82
589
765
1317 (89.5%)
21 (25.6%)
506 (85.9%)
759 (99.2%)
18 (1.2%)
18 (22.0%)
5 (0.8%)
2 (0.3%)
1299
3
501
757
h/frontroom2study
h/frontroom2study
h/hall2frontbedroom
h/hall2frontbedroom
h/hall2frontbedroom
h/hall2oldkitchen
h/hall2oldkitchen
h/mainbedroom2studio
7714
8171
7141
6942
360
592
242
260
249 (69.2%)
442 (74.7%)
192 (79.3%)
102 (39.2%)
16 (4.4%)
10 (1.7%)
12 (5.0%)
69 (26.5%)
233
432
180
33
Table 5: Evaluating the extent to which our approach of incrementally clustering the relative transform samples between
different pairs of sub-scenes is able to remove outliers and find consistent relocalisations between the sub-scenes. We attempt
to relocalise every frame of each sub-scene against the other using [2], and record the total number of samples that were
added (equal to the number of relocalised frames that passed depth difference verification), and the sizes of the correct cluster
and the largest incorrect cluster produced by our method in each case (together with the percentages of the samples added
these sizes represent). The safety margin for each scene pair refers to the number of consistent erroneous samples that would
need to have been added to the largest incorrect cluster to cause it to have been chosen instead of the correct cluster.
server currently becomes less interactive because the mapping components for each client are being run sequentially:
this is something we plan to mitigate in the future by running all of the mapping components in parallel.
B.3. Frame Compression
To demonstrate the impact that compressing the RGB-D
frames we transmit over the network has on the final reconstruction quality on the server, we used CloudCompare
[10] to compare a locally-reconstructed reference model of
an office with two reconstructions performed from frame
Server allocates scene 3
Collaboration starts
Collaboration/rendering continues
Client 2 relocalises
Figure 11: An experiment showing how the frame processing times on the server and each of three individual clients change
over the course of a collaborative reconstruction using our system. Note that we smooth these times over a 100-frame window
to make the graph slightly easier to read. Prior to the start of the experiment, the server allocates a mapping component for
the first client and waits for it to connect. The experiment starts when the first client (a desktop using a Kinect camera and a
relocaliser based on random ferns [20]) connects to the server over a wired network. Reconstruction begins when the client
starts sending RGB-D frames to the server. Subsequently, a second client (a laptop using an Orbbec Astra S camera and a
regression forest relocaliser [2]) connects to the server over WiFi. This causes the server to allocate a mapping component for
the second client and start reconstructing a local scene based on the RGB-D frames that it sends across. In parallel, the server
starts trying to relocalise the local scenes of the two clients against each other on a separate thread. Around half way through
the experiment, the tracking on the second client fails, triggering a local relocalisation that causes a spike in processing time
on the client. Towards the end of the experiment, a third client (this time a local one on the same machine as the server that
reads posed RGB-D frames from disk) connects, and the server again allocates a mapping component for it. The processing
time on this client is lower because it does not need to perform camera tracking. At this point, the server is trying to relocalise
all three local scenes against each other. Finally, all three clients disconnect, and reconstruction stops. However, inter-agent
relocalisation continues on the server until a consistent reconstruction is achieved, and the server continues rendering all three
local scenes to allow the user to visualise what is happening.
data sent over a WiFi connection, one with frame compression disabled and the other with it enabled (see Figure 12). We measured the bandwidth of this WiFi connection to be roughly 8 Mbits/s, which is less than the roughly
25 Mbits/s that we needed to transmit the same sequence
in compressed form without loss over a wired connection,
and much less than the roughly 250 Mbits/s that would have
been needed to transmit the sequence without loss uncompressed. In other words, with compression enabled, we are
able to transmit around 1 in 3 of the frames in the sequence
over WiFi whilst maintaining real-time rates; without compression, this drops to more than 1 in 30. As Figure 12
shows, this has a significant effect on the resulting reconstruction quality: with compression disabled (b), we lose
some parts of the map completely and have a higher error
rate across the map as a whole; with compression enabled
(c), we manage to reconstruct more or less the entire map,
and the error rate is greatly reduced.
C. Additional Implementation Details
In this section, we describe some implementation details that might be relevant for anyone wanting to reimplement our approach, namely the inner workings of the pooled
queue data structure mentioned in §3.2 of the main paper,
and the way in which we can render a global map of all of
the sub-scenes we have reconstructed. These details can be
skipped by more casual readers.
(a) Reference Model
(b) Reference vs. Uncompressed
(c) Reference vs. Compressed
Figure 12: An example showing the impact of frame compression on the reconstruction quality we are able to achieve on
the server whilst maintaining real-time frame rates by discarding some frames: (a) a locally-reconstructed reference model
of an office; (b) the differences between the reference model and a model reconstructed from the frames that we managed
to transmit without using compression; (c) the differences between the reference model and a model reconstructed from the
frames that we managed to transmit with compression. The comparisons were made using CloudCompare [10]. The errors
range from blue = 0cm to red ≥ 5cm (or missing). With compression enabled, we are forced to discard far fewer frames,
allowing us to achieve a much lower error rate with respect to the reference model.
C.1. Pooled Queue Data Structure
A pooled queue is a data structure that pairs a normal
queue Q of objects of type T with a pool P of reusable
T objects, with the underlying goal of minimising memory reallocations for performance reasons. Normal queues
conventionally support the following range of operations:
empty checks whether the queue is empty, peek gets the
object (if any) at the front of the queue, pop removes the
object (if any) at the front of the queue, push adds an object to the back of the queue, and size gets the number of
objects on the queue. Pooled queues support the same range
of operations, but their implementations of push and pop
are necessarily complicated by interactions with the pool.
The pop operation is the more straightforward of the
two: this simply removes the object (if any) at the front of
the queue and returns it to the pool (with appropriate synchronisation in a multi-threaded context).
The push operation is more complicated, and we divide it into two parts: begin push and end push. The
begin push operation first checks whether a reusable object is currently available from the pool. If so, it removes
it from the pool and returns a push handler encapsulating
the object to the caller. The caller then modifies the object,
after which the push handler calls end push to actually
push the object onto the queue. If a reusable object is not
available from the pool when begin push is called, we
have a range of options, which are available as policies in
our implementation: (i) discard the object we are trying to
push onto the queue, (ii) grow the pool by allocating a new
object that can then be reused, (iii) remove a random element from the queue and return it to the pool so that it can
be reused, or (iv) wait for another thread to pop an object
from the queue and return it to the pool. In practice, we
found the discard strategy to work best for our approach:
growing the pool has the disadvantage that the memory usage can grow without bound over time, waiting makes the
client and/or server (as the case may be) less interactive, and
removing a random element from the queue functions much
like discarding but with different frames.
C.2. Global Map Rendering
To render a global map of all of the sub-scenes we have
reconstructed from a global pose Tg ∈ SE 3 , we first compute (for each agent a) the corresponding local pose Ta in
a’s local coordinate system:
Ta = Tg g T̂a
(5)
Using the standard raycasting approach for a voxel-based
scene [37], we then render synthetic colour and depth images (at 1280×960 resolution) for each agent from the local
pose we have computed for it, which we call Ca and Da respectively. Finally, we set the colour of each pixel in the
output image O based on per-pixel depth testing between
the agents, i.e.
O(x) = Ca∗ (x), where a∗ = argmin Da (x),
a
where x ranges over the domain of O.
(6)
References
[1] A. Bajpai, G. Burroughes, A. Shaukat, and Y. Gao. Planetary
Monocular Simultaneous Localization and Mapping. JFR,
33(2):229–242, 2016. 2
[2] T. Cavallari, S. Golodetz*, N. A. Lord*, J. Valentin, L. D.
Stefano, and P. H. S. Torr. On-the-Fly Adaptation of Regression Forests for Online Camera Relocalisation. In CVPR,
2017. 2, 4, 7, 8, 12, 13, 14
[3] N. Chebrolu, D. Marquez-Gamez, and P. Martinet. Collaborative Visual SLAM Framework for a Multi-Robot System.
In PPNIV, 2015. 2, 4
[4] F. A. A. Cheein and R. Carelli. Agricultural Robotics: Unmanned Robotic Service Units in Agricultural Tasks. IEEE
Industrial Electronics Magazine, 7(3), 2013. 2
[5] H. Chen, J. Zhong, Y. Fu, and Y. Lou. Pose-graph Based
3D Map Fusion with Distributed Robot System. In ROBIO,
2014. 3, 4
[6] S. Choudhary, L. Carlone, C. Nieto, J. Rogers, H. I. Christensen, and F. Dellaert. Distributed Mapping with Privacy
and Communication Constraints: Lightweight Algorithms
and Object-based Models. arXiv preprint arXiv:1702.03435,
2017. 2
[7] T. Cieslewski, S. Choudhary, and D. Scaramuzza. DataEfficient Decentralized Visual SLAM.
arXiv preprint
arXiv:1710.05772, 2017. 2
[8] T. Cieslewski, S. Lynen, M. Dymczyk, S. Magnenat, and
R. Siegwart. Map API – Scalable Decentralized Map Building for Robots. In ICRA, pages 6241–6247, 2015. 2
[9] P. Cignoni, M. Callieri, M. Corsini, M. Dellepiane, F. Ganovelli, and G. Ranzuglia. MeshLab: an Open-Source Mesh
Processing Tool. In Eurographics Italian Chapter Conference, pages 129–136, 2008. 6, 10
[10] CloudCompare v2.9 (GPL software), 2017. Retrieved from
http://www.danielgm.net/cc. 10, 13, 15
[11] A. Cunningham, V. Indelman, and F. Dellaert. DDF-SAM
2.0: Consistent Distributed Smoothing and Mapping. In
ICRA, pages 5220–5227, 2013. 2
[12] A. Cunningham, M. Paluri, and F. Dellaert. DDF-SAM:
Fully Distributed SLAM using Constrained Factor Graphs.
In IROS, pages 3025–3030, 2010. 2
[13] A. Cunningham, K. M. Wurm, W. Burgard, and F. Dellaert.
Fully Distributed Scalable Smoothing and Mapping with Robust Multi-robot Data Association. In ICRA, pages 1093–
1100, 2012. 2
[14] A. Dai, M. Nießner, M. Zollhöfer, S. Izadi, and C. Theobalt.
BundleFusion: Real-time Globally Consistent 3D Reconstruction using Online Surface Re-integration.
TOG,
36(3):24, 2017. 1
[15] J. Engel, T. Schöps, and D. Cremers. LSD-SLAM: LargeScale Direct Monocular SLAM. In ECCV, pages 834–849,
2014. 2
[16] P. Fankhauser, M. Bloesch, P. Krüsi, R. Diethelm, M. Wermelinger, T. Schneider, M. Dymczyk, M. Hutter, and
R. Siegwart. Collaborative Navigation for Flying and Walking Robots. In IROS, pages 2859–2866, 2016. 3
[17] N. Fioraio, J. Taylor, A. Fitzgibbon, L. D. Stefano, and
S. Izadi. Large-Scale and Drift-Free Surface Reconstruction Using Online Subvolume Registration. In CVPR, pages
4475–4483, 2015. 1
[18] P. Fleck, C. Arth, C. Pirchheim, and D. Schmalstieg. Tracking and Mapping with a Swarm of Heterogeneous Clients. In
ISMAR, 2015. 4
[19] C. Forster, S. Lynen, L. Kneip, and D. Scaramuzza. Collaborative Monocular SLAM with Multiple Micro Aerial Vehicles. In IROS, pages 3962–3970, 2013. 3
[20] B. Glocker, J. Shotton, A. Criminisi, and S. Izadi. RealTime RGB-D Camera Relocalization via Randomized Ferns
for Keyframe Encoding. TVCG, 21(5):571–583, May 2015.
2, 12, 14
[21] S. Golodetz*, M. Sapienza*, J. P. C. Valentin, V. Vineet, M.M. Cheng, A. Arnab, V. A. Prisacariu, O. Kähler, C. Y. Ren,
D. W. Murray, S. Izadi, and P. H. S. Torr. SemanticPaint: A
Framework for the Interactive Segmentation of 3D Scenes.
Technical Report TVG-2015-1, Department of Engineering
Science, University of Oxford, October 2015. Released as
arXiv e-print 1510.03727. 2
[22] S. Golodetz*, M. Sapienza*, J. P. C. Valentin, V. Vineet, M.M. Cheng, V. A. Prisacariu, O. Kähler, C. Y. Ren, A. Arnab,
S. L. Hicks, D. W. Murray, S. Izadi, and P. H. S. Torr.
SemanticPaint: Interactive Segmentation and Learning of
3D Worlds. In ACM SIGGRAPH Emerging Technologies,
page 22, 2015. 2
[23] J. Huang, A. Dai, L. Guibas, and M. Niessner. 3DLite: Towards Commodity 3D Scanning for Content Creation. TOG,
2017. 1
[24] O. Kähler, V. A. Prisacariu, and D. W. Murray. Real-Time
Large-Scale Dense 3D Reconstruction with Loop Closure. In
ECCV, pages 500–516, 2016. 1, 4, 5
[25] L. Kavan, S. Collins, C. O’Sullivan, and J. Zara. Dual
Quaternions for Rigid Transformation Blending. Technical
Report TCD-CS-2006-46, Trinity College Dublin, 2006. 5,
10
[26] G. Klein and D. Murray. Parallel Tracking and Mapping for
Small AR Workspaces. In ISMAR, pages 225–234, 2007. 3
[27] R. Kümmerle, G. Grisetti, H. Strasdat, K. Konolige, and
W. Burgard. g2 o: A General Framework for Graph Optimization. In ICRA, pages 3607–3613, 2011. 3
[28] W. E. Lorensen and H. E. Cline. Marching Cubes: A High
Resolution 3D Surface Construction Algorithm. ACM SIGGRAPH Computer Graphics, 21(4):163–169, 1987. 6, 8
[29] J. McDonald, M. Kaess, C. Cadena, J. Neira, and J. J.
Leonard. Real-time 6-DOF multi-session visual SLAM over
large-scale environments. RAS, 61(10):1144–1158, 2013. 3
[30] N. Michael, S. Shen, K. Mohta, Y. Mulgaonkar, V. Kumar,
K. Nagatani, Y. Okada, S. Kiribayashi, K. Otake, K. Yoshida,
K. Ohno, E. Takeuchi, and S. Tadokoro. Collaborative Mapping of an Earthquake-Damaged Building via Ground and
Aerial Robots. JFR, 29(5):832–841, 2012. 2
[31] G. Mohanarajah, V. Usenko, M. Singh, R. D’Andrea, and
M. Waibel. Cloud-based Collaborative 3D Mapping in RealTime with Low-Cost Robots. TASE, 12(2):423–431, 2015.
3, 4
[32] S. Murali, P. Speciale, M. R. Oswald, and M. Pollefeys. Indoor Scan2BIM: Building Information Models of House Interiors. In IROS, 2017. 1
[33] R. A. Newcombe, S. Izadi, O. Hilliges, D. Molyneaux,
D. Kim, A. J. Davison, P. Kohli, J. Shotton, S. Hodges,
and A. Fitzgibbon. KinectFusion: Real-Time Dense Surface
Mapping and Tracking. In ISMAR, pages 127–136, 2011. 1
[34] M. Nießner, M. Zollhöfer, S. Izadi, and M. Stamminger.
Real-time 3D Reconstruction at Scale using Voxel Hashing.
TOG, 32(6):169, 2013. 1
[35] L. Paull, G. Huang, M. Seto, and J. J. Leonard.
Communication-Constrained
Multi-UAV
Cooperative
SLAM. In ICRA, pages 509–516, 2015. 2
[36] V. A. Prisacariu, O. Kähler, M. M. Cheng, C. Y. Ren,
J. Valentin, P. H. S. Torr, I. D. Reid, and D. W. Murray. A
Framework for the Volumetric Integration of Depth Images.
arXiv preprint arXiv:1410.0925, 2014. 1, 4, 5, 6
[37] V. A. Prisacariu, O. Kähler, S. Golodetz, M. Sapienza,
T. Cavallari, P. H. S. Torr, and D. W. Murray. InfiniTAM
v3: A Framework for Large-Scale 3D Reconstruction with
Loop Closure. arXiv preprint arXiv:1708.00783v1, 2017. 1,
2, 4, 5, 6, 15
[38] R. Reid, A. Cann, C. Meiklejohn, L. Poli, A. Boeing, and
T. Braunl. Cooperative Multi-Robot Navigation, Exploration, Mapping and Object Detection with ROS. In IV,
2013. 3, 4
[39] L. Riazuelo, J. Civera, and J. M. M. Montiel. C2 TAM:
A Cloud framework for cooperative tracking and mapping.
RAS, 62(4):401–413, 2014. 3
[40] W. Rone and P. Ben-Tzvi. Mapping, localization and motion
planning in mobile multi-robotic systems. Robotica, 31:1–
23, 2013. 2
[41] S. Saeedi, M. Trentini, M. Seto, and H. Li. Multiple-robot
Simultaneous Localization and Mapping – A Review. JFR,
33(1):3–46, 2016. 2
[42] P. Schmuck and M. Chli. Multi-UAV Collaborative Monocular SLAM. In ICRA, pages 3863–3870, 2017. 3, 4
[43] J. Shotton, B. Glocker, C. Zach, S. Izadi, A. Criminisi, and
A. Fitzgibbon. Scene Coordinate Regression Forests for
Camera Relocalization in RGB-D Images. In CVPR, pages
2930–2937, 2013. 2, 6, 8
[44] L. Silveira, F. Guth, P. Drews-Jr, P. Ballester, M. Machado,
F. Codevilla, N. Duarte-Filho, and S. Botelho. An Opensource Bio-inspired Solution to Underwater SLAM. IFACPapersOnLine, 48(2):212–217, 2015. 2
[45] J. Valentin, M. Nießner, J. Shotton, A. Fitzgibbon, S. Izadi,
and P. Torr. Exploiting Uncertainty in Regression Forests
for Accurate Camera Relocalization. In CVPR, pages 4400–
4408, 2015. 2
[46] T. Whelan, M. Kaess, H. Johannsson, M. Fallon, J. J.
Leonard, and J. McDonald. Real-time large scale dense
RGB-D SLAM with volumetric fusion. IJRR, 34(4-5):598–
626, 2015. 1
[47] T. Whelan, S. Leutenegger, R. F. Salas-Moreno, B. Glocker,
and A. J. Davison. ElasticFusion: Dense SLAM Without A
Pose Graph. In RSS, 2015. 1
[48] M. Zollhöfer, C. Siegl, M. Vetter, B. Dreyer, M. Stamminger,
S. Aybek, and F. Bauer. Low-Cost Real-Time 3D Reconstruction of Large-Scale Excavation Sites. JOCCH, 9(1),
2016. 1
| 1cs.CV
|
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC
2-GROUPS
FRANK HIMSTEDT AND PETER SYMONDS
arXiv:1510.02370v1 [math.RT] 8 Oct 2015
Abstract. We prove a recursive formula for the exterior and symmetric powers of modules
for a cyclic 2-group. This makes computation straightforward. Previously, a complete
description was only known for cyclic groups of prime order.
1. Introduction
The aim of this paper is to provide a recursive procedure for calculating the exterior and
symmetric powers of a modular representation of a cyclic 2-group. Let G ∼
= C2n be a cyclic
n
n
group of order 2 and k a field of characteristic 2. Recall that there are 2 indecomposable
kG-modules V1 , V2 , . . . , V2n for which dim Vr = r.
Theorem 1.1. For all n ≥ 1, r ≥ 0 and 0 ≤ s ≤ 2n−1 we have
M i+j
Λr (V2n−1 +s ) ∼
Ω2n (Λi (Vs ) ⊗k Λj (V2n−1 −s )) ⊕ tV2n ,
=
i,j≥0
2i+j=r
where t is a non-negative integer chosen so that both sides have the same dimension.
Here Ω2n is the syzygy or Heller operator over C2n , so Ω2n Vs = V2n −s for 1 ≤ s ≤ 2n .
The group action on V1 , . . . , V2n−1 factors through C2n−1 so that exterior powers of these
modules can be computed by applying the formula for this smaller group. In particular, one
can determine the exterior powers on the right hand side of the formula in this way. We
also show that there is a simple recursive procedure for calculating tensor products. Since
Λ(A ⊕ B) ∼
= Λ(A) ⊗ Λ(B), we obtain a complete recursive procedure for calculating exterior
powers of all possible modules. It is sufficiently efficient that it is easy to calculate even by
hand far beyond the range that was previously attainable by machine computation.
For symmetric powers we use the following result from [21].
Theorem 1.2. For all n ≥ 1, r ≥ 0 and 0 ≤ s ≤ 2n−1 we have
′
′
S r (V2n−1 +s ) ∼
=ind Ωr n Λr (V2n−1 −s ),
2
where 0 ≤ r ′ < 2n and r ′ ≡ r (mod 2n ). Here the symbol ∼
=ind means up to direct summands
induced from subgroups H G.
Thus a knowledge of the exterior powers determines the symmetric powers up to induced
summands. In fact it is shown in [21] how such a formula determines the symmetric powers
completely, using a recursive procedure.
This project was supported by the Deutsche Forschungsgemeinschaft under the project KE 964/1-1 (“Invariantentheorie endlicher und algebraischer Gruppen”).
1
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
2
Formulas for the exterior and symmetric powers of a module for a cyclic group of prime
order p were given by Almkvist and Fossum [1] and Renaud [19]. These were extended to
cyclic p-groups by Hughes and Kemper [14] provided that the power is at most p − 1.
A formula for Λ2 in the case of cyclic 2-groups was given by Gow and Laffey [11]. Also
Kouwenhoven [15] obtained important results on exterior powers of modules for cyclic pgroups, including recursion formulas for Λ(Vq±1 ) where q is a power of p. For p = 2 these
formulas are special cases or direct consequences of Theorem 1.1, so we obtain independent
proofs for some of the results in [11, 15].
Our strategy is to consider Λ(V2n−1 +s ) as the quotient of S(V2n−1 +s ) by the ideal generated
by the squares of elements of V2n−1 +s . It turns out that we need to consider an intermediate
ring S̃(V2n−1 +s ), in which we only quotient out the squares of the elements of Vs ⊆ V2n−1 +s .
We show that S̃ r (V2n−1 +s ) ∼
=ind Λr (V2n−1 +s ) for r < 2n . But S̃(V2n−1 +s ) can be resolved by
the Koszul complex over S(V2n−1 +s ) on the squares of the elements of a basis for Vs . We
show that this Koszul complex is separated in the sense of [21], that is that the image of a
boundary map is contained in a projective submodule. This leads to the formula
M
Ωi2n (Λi (Vs ) ⊗k S j (V2n−1 +s )),
S̃ r (V2n−1 +s ) ∼
=proj
2i+j=r
where the symbol ∼
=proj means up to projective summands. Using Theorem 1.2, the right
hand side is easily seen to be equal to the right hand side of the formula in Theorem 1.1 modulo induced summands. This yields the formula of Theorem 1.1 modulo induced summands.
The strengthening to an equality modulo just projective summands is a formal inductive
argument.
We would like to thank Dikran Karagueuzian for the calculations that were very helpful
in discovering the formula of Theorem 1.1.
2. Koszul Complexes
Let G be a finite group, H a subgroup of G and k a field of characteristic p > 0. All tensor
products will be over k if not otherwise specified. We recall some general facts about chain
complexes of kG-modules from [21, Section 3].
Definition 2.1. ([21, Definition 3.2]) A chain complex C∗ of kG-modules is called:
(a) acyclic if it is 0 in negative degrees and it only has homology in degree 0;
(b) weakly induced from H if each module is induced from H, and weakly induced from H
except in degrees I if each Ci , i 6∈ I, is induced from H;
(c) separated at Ci if Im(di+1 ) → Ci factors through a projective kG-module;
(d) separated if it is separated at each Ci .
Write Bi for Im(di+1 ) ⊆ Ci . If the inclusion Bi ֒→ Ci factors through a projective then
it factors through the injective hull of Bi , call it Pi (injective is equivalent to projective for
modular representations), and Pi ֒→ Ci is injective since it is so on the socle. Thus we can
write Ci = Pi ⊕ Ci′ and Bi ⊆ Pi .
Lemma 2.2. ([21, Lemma 3.9]) If the chain complexes C∗ , C∗′ of kG-modules are separated
then so is the (total) tensor product C∗ ⊗ C∗′ . Similarly for a product of finitely many chain
complexes.
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
3
Proof. Let Pi be a projective module such that Im(di+1 ) ⊆ Pi ⊆ Ci and L
similarly for Pi′ .
′
′
′
′
Then,
L ′ summing over all degrees, Im(d ⊗ d ) ⊆ P ⊗′ C + C ⊗′ P where′ P := i Pi and P :=
i Pi . There is a short exact sequence 0 → P ⊗P → P ⊗C ⊕C ⊗P → P ⊗C +C ⊗P → 0.
The first two terms are projective, hence so is the third.
We need to consider tensor-induced complexes. For details of the construction see [4, II
4.1].
Lemma 2.3. Suppose that every elementary abelian p-subgroup of G is conjugate to a subgroup of H, and let C∗ be a complex of kH-modules that is separated. Then the tensor-induced
complex C∗ ↑⊗G
H is also separated.
Proof. By the proof of Lemma 2.2 above, the image Im(d↑⊗G
H ) is contained in
P ⊗ C ⊗ · · · ⊗ C + C ⊗ P ⊗ · · · ⊗ C + · · · + C ⊗ C ⊗ · · · ⊗ P,
which is a kG-submodule of C∗ ↑⊗G
H . But the same proof shows that this module is projective
on restriction to H, so it is projective, by Chouinard’s Theorem [7, Corollary 1.1].
The next two results comprise a variation on [21, Proposition 3.3] and have the same
proof.
Proposition 2.4. Let H be an arbitrary subgroup of G. Suppose that the complex K∗ :
Kw → · · · → K0 of kG-modules is:
(a) acyclic,
(b) weakly induced from H except in at most one degree and
(c) K∗ is separated on restriction to H.
Then K∗ is separated.
Recall that the Heller translate ΩV of a kG-module V is defined to be the kernel of the
projective cover P (V ) → V and Ωi V for i ≥ 1 denotes Ω iterated i times. Similarly Ω−1 V
is the cokernel of the injective hull V → I(V ) and Ω−i V for i ≥ 1 is its iteration. We
let Ω0 V denote V with any projective summands removed. These have the properties that
Ωi Ωj V ∼
= Ωi+j V and that if V is induced so is Ωi V .
Lemma 2.5. Suppose that the complex K∗ : Kw → · · · → K0 of kG-modules is:
(a) acyclic with H0 (K∗ ) = L, and
(b) separated.
Then L ∼
=proj K0 ⊕ Ω−1 K1 ⊕ Ω−2 K2 ⊕ · · · ⊕ Ω−w Kw .
Let V be a kG-module,
as a k-vector space, and W a submodule
L∞ finite-dimensional
L∞ rof V .
r
We write S = S(V ) = r=0 S (V ) for the symmetric algebra on V and Λ(W ) = r=0 Λ (W )
for the exterior algebra on W . For r < 0 let S r (V ) denote the 0 module.
Definition 2.6. Let W be a submodule of a kG-module V and let W p denote the kGsubmodule of S p (V ) spanned by the p-th powers of elements of W . Let K(V, W p ) denote
the Koszul complex of graded kG-modules:
d
d
d
d
· · · −→ S(V ) ⊗ Λ3 (W p ) −→ S(V ) ⊗ Λ2 (W p ) −→ S(V ) ⊗ W p −→ S(V ),
P
cp ∧ · · · ∧ w p for wj ∈ W and
where d(f ⊗ w1p ∧ w2p ∧ · · · ∧ wip ) = ij=1(−1)j+1 f wjp ⊗ w1p ∧ · · · ∧ w
j
i
f ∈ S(V ). We write K r (V, W p ) when we consider the complex K(V, W p ) in grading-degree r.
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
4
If k = F2 then the squaring map gives an isomorphism between W and W 2 , so we can
regard W 2 as a copy of W in degree 2 equipped with a squaring map into S 2 (V ). From this
P
point of view, the boundary map is given by d(f ⊗ w1 ∧ w2 ∧ · · · ∧ wi ) = ij=1 f wj2 ⊗ w1 ∧
···∧ w
cj ∧ · · · ∧ wi for wj ∈ W and f ∈ S(V ).
We will normally take the second point of view, so we will assume that k = F2 in a large
part of this paper. Since any kC2n -module can be written in F2 , this is not a significant
restriction.
Lemma 2.7. In the context of Definition 2.6, the complex K(V, W p ) is acyclic and its
homology in degree 0 is S(V )/(W p ), where (W p ) is the ideal generated by all elements w p ,
w ∈ W.
Proof. If {w1 , . . . , wr } is a basis for W then {w1p , . . . , wrp} is a regular sequence in S(V ) and
it spans W p . This is now a standard result about Koszul complexes.
Lemma 2.8. Let V , V ′ be kG-modules, finite-dimensional as k-vector spaces and let W , W ′
be submodules of V and V ′ , respectively. The complex K(V ⊕ V ′ , (W ⊕ W ′ )p ) is isomorphic
to the (total) tensor product K(V, W p ) ⊗ K(V ′ , W ′p ) as a complex of graded kG-modules.
Proof. This is analogous to [21, Lemma 3.8].
We also need to deal with tensor induction of graded modules and complexes.
∼
Lemma 2.9. Let H be a subgroup of G and let V, W be kH-modules. Then S(V ↑G
H) =
⊗G
G ∼
S(V ) ↑⊗G
H , Λ(V ↑H ) = Λ(V ) ↑H as graded kG-modules, and if the characteristic of k is 2
G
G 2 ∼
then K(V ↑H , (W ↑H ) ) = K(V, W 2 )↑⊗G
H as complexes of graded kG-modules.
Without the restriction on the characteristic of k we would have to deal with the sign
convention that appears in the definition of the action of G on the tensor-induced complex.
Proof. Let {gi } be a set of coset representatives for G/H and write V ↑G
H = ⊕gi ⊗ V . The
formulas now follow from the usual formulas for S and Λ of a sum and the definition of the
group action on a tensor induced module.
3. Modules for Cyclic 2-Groups
From now on, let G = hgi ∼
= C2n be a cyclic group of order 2n , n ≥ 1, and k a field
of characteristic 2. We write a(G) for the Green ring of kG-modules. Up to isomorphism,
there are 2n indecomposable kG-modules V1 , V2 , . . . , V2n and we choose the notation so that
dimk (Vi ) = i. For convenience we write V0 for the 0 module. The generator g ∈ G acts on Vi
with matrix a Jordan block with ones on the diagonal. Choose a k-basis {x1 , x2 , . . . , xn }
of Vn such that gxi = xi + xi−1 for all 2 ≤ i ≤ n and gx1 = x1 . Each element of S(Vi ) can
be written uniquely as a polynomial in x1 , . . . , xi , and for j ≤ i, we can identify Vj with
the kG-submodule of Vi spanned by x1 , x2 , . . . , xj . Each Vi is uniserial with composition
series 0 < V1 < V2 < · · · < Vi−1 < Vi . Note that for i ≤ 2n−1 the kernel of Vi is nontrivial
and so Vi can be identified with the i-dimensional indecomposable module for the quotient
group C2n−1 .
Decompositions of tensor products into indecomposables have been studied by several
authors, see for example [2, 3, 9, 13, 16, 17, 18, 20]. In our case, this decomposition can
easily be computed using the Heller translate Ω. We write Ω2n instead of Ω when we want
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
5
to emphasize that we are working with modules for the group C2n . It is easy to check that
Ω2n Vi ∼
=proj V2n −i for 0 ≤ i ≤ 2n , where proj means modulo projective modules for C2n .
Recall that Ω2n (V ⊗ V ′ ) ∼
=proj (Ω2n V ) ⊗ V ′ , where the projective part can be determined
by comparing dimensions. For cyclic groups, Ω22n ∼
=proj V ⊗ V ′ .
=proj Id and Ω2n V ⊗ Ω2n V ′ ∼
This provides an easy recursive method for calculating the decomposition of tensor products
in the case of cyclic 2-groups.
In order to calculate Va ⊗ Vb , we may assume a ≥ b and write a = 2r − a′ for the smallest
possible r such that a′ ≥ 0. Then Va ⊗ Vb ∼
= Ω2r (Va′ ⊗ Vb ) modulo copies
= (Ω2r Va′ ) ⊗ Vb ∼
r−1
r
then it is more efficient to write b = 2 − b′ too.
of V2r . If b ≥ 2
Example. For V9 , V13 ∈ a(C16 ) we have: V9 ⊗ V13 ∼
= (V7 ⊗ V3 )⊕ copies
= (Ω16 V7 ) ⊗ (Ω16 V3 ) ∼
∼
of V16 . By comparing dimensions we get V9 ⊗V13 = (V7 ⊗V3 )⊕6V16 . Now we consider the nonfaithful module V7 ⊗V3 as a module for the factor group ∼
= C8 and get V7 ⊗V3 ∼
= Ω8 (V1 ⊗V3 )⊕
∼
copies of V8 . Again by comparing dimensions we obtain V7 ⊗V3 = Ω8 (V1 ⊗V3 )⊕2V8 ∼
= V5 ⊕2V8
and hence V9 ⊗ V13 ∼
= V5 ⊕ 2V8 ⊕ 6V16 .
Let H = hg 2 i be the unique maximal subgroup of G. For 1 ≤ j ≤ 2n−1 we also denote the
indecomposable kH-module of dimension j by Vj . Of course, this is an abuse of notation,
but we will always make it clear whether we consider Vj as a kG-module or as a kH-module.
An elementary calculation with Jordan canonical forms shows that the restriction operator
G
↓G
H : a(G) → a(H) is given by Vi ↓H = Vi′ ⊕ Vi′′ where Vi′ is the kH-submodule generated
by {xi , xi−2 , xi−4 , . . . } and Vi′′ is the kH-submodule generated by {xi−1 , xi−3 , xi−5 , . . . }. In
particular, we have (i′ , i′′ ) = ( i+1
, i−1
) if i is odd, and (i′ , i′′ ) = ( 2i , 2i ) if i is even. The
2
2
G
n−1
induction operator ↑G
.
H : a(H) → a(G) is given by Vj ↑H = V2j for 1 ≤ j ≤ 2
We say that a kG-module is induced if it is induced from proper subgroups. Let aP (G)
be the submodule of a(G) generated by the projective modules and aI (G) the submodule
generated by the induced modules. Notice that aP (G) and aI (G) are ideals of a(G) and that
induction maps aP (H) into aP (G) and aI (H) into aI (G), but restriction only maps aP (G)
into aP (H).
The following lemmas deduce information on kG-modules and short exact sequences of
kG-modules from their restriction to H.
Lemma 3.1. Let A be a kG-module such that A↓G
H is induced from a proper subgroup of H.
Then A is induced from H.
Proof. We can assume that A is indecomposable. Since A↓G
H is induced, each indecomposable
direct summand of A↓G
has
even
dimension.
Thus
dim
(A)
is even and so A is induced.
k
H
As in the introduction we write ∼
=proj for isomorphisms modulo induced and
=ind and ∼
modulo projective summands, respectively.
Lemma 3.2. Let A, B be induced kG-modules such that A↓G ∼
=proj B.
=proj B↓G . Then A ∼
H
G∼
A↓G
H ↑H =
H
Proof. Since A is induced,
2A; the same is true for B. Inducing a projective yields
a projective, so we obtain 2A ∼
=proj 2B and the claim follows.
G
∼
Lemma 3.3. Let A and B be kG-modules such that A ∼
=ind B and A↓G
H =proj B ↓H . Then
A∼
=proj B.
Proof. We have A ⊕ X ∼
= B ⊕ Y for some induced modules X and Y . On restriction, we
G
∼
∼
obtain X ↓G
Y
↓
H =proj
H and so X =proj Y by Lemma 3.2. Now cancel the non-projective
summands of X and Y in the original formula.
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
6
Lemma 3.4. Let 0 → A → B → C → 0 be a short exact sequence of kG-modules that is
separated at B on restriction to H and such that C ∼
=ind B ⊕ Ω−1
2n A as kG-modules. Then
the sequence is separated at B (as a sequence of kG-modules).
Proof. The hypotheses imply that there are induced modules X and Y such that C ⊕ X ∼
=
−1
−1
G
G
G ∼
G ∼
A)↓
by
Lemma
2.5
(B
⊕
Ω
B↓
⊕
Ω
A↓
A
⊕
Y
and
also
C
↓
B ⊕ Ω−1
=
=
n
n
n−1
proj
proj
2
2
H
H
H
H
2
G
G
G ∼
applied to 0 → A↓G
H → B↓H → 0. It follows that X ↓H =proj Y ↓H , hence, by Lemma 3.2,
X∼
=proj B ⊕ Ω−1
=proj Y and then C ∼
2n A.
d
e
Thus our short exact sequence is 0 → A → B → B ⊕ Ω−1
2n A → 0, up to projective
summands. Consider the long exact sequence for Tate Ext:
d
e
∗
∗
1
(1) · · · → HomkG (A, A) →
HomkG (A, B) →
HomkG (A, B ⊕ Ω−1
2n A) → Ext (A, A) → · · · ,
where HomkG denotes homomorphisms modulo those that factorize through a projective.
Since Ext1 (A, A) ∼
= HomkG (A, Ω−1
2n A) we have
dim Im(e∗ ) + dim Ext1 (A, A) ≥ dim HomkG (A, B ⊕ Ω−1
2n A) =
dim HomkG (A, B) + dim HomkG (A, Ω−1
2n A) =
dim HomkG (A, B) + dim Ext1 (A, A)
and therefore dim Im(e∗ ) ≥ dim HomkG (A, B). Hence e∗ is injective and so d∗ = 0. But
d = d∗ (IdA ), so d factors through a projective, as required.
The next lemma describes tensor induction from H to G modulo induced modules and
gives information on the structure of the exterior algebra Λ(V2j ) as a kG-module in terms
of the kH-module Λ(Vj ).
Lemma 3.5. Let r, j be integers such that r ≥ 0 and 1 ≤ j ≤ 2n−1. We consider Vj as a
kH-module and V2j = Vj ↑G
H as a kG-module.
⊗G
∼
(a) Let A and B be kH-modules. Then (A ⊕ B) ↑⊗G
⊕B ↑⊗G
⊕X for some
H = A ↑H
H
induced kG-module X.
(b) There is an induced kG-module X such that Λ2r (V2j ) ∼
= Λr (Vj )↑⊗G
H ⊕ X.
r
(c) If r is odd, then the kG-module Λ (V2j ) is induced from H.
(d) If j is even, then the kG-module Vj ↑⊗G
H is induced from H.
∼
(e) If j is odd, then Vj ↑⊗G
V
.
=
ind 1
H
Proof. (a) follows from [4, I 3.15.2 (iii)].
(b) By the construction of induced modules, we have V2j = Vj ⊕ gVj as vector spaces and
the action of the generator g of G on V2j is given by g(v + gv ′) = g 2 v ′ + gv. So there is a
natural isomorphism
M
′
′′
(Λr (Vj ) ⊗ Λr (gVj ))
Λ2r (V2j ) = Λ2r (Vj ⊕ gVj ) ∼
=
r ′ ,r ′′ ≥0
r ′ +r ′′ =2r
of vector spaces, and thus
Λ2r (V2j ) ∼
= (Λr (Vj ) ⊗ gΛr (Vj )) ⊕
M
0≤r ′ <r ′′
r ′ +r ′′ =2r
′
′′
′
′′
((Λr (Vj ) ⊗ Λr (gVj )) ⊕ g(Λr (Vj ) ⊗ Λr (gVj ))).
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
7
Via this isomorphism, the right hand side becomes a kG-module and from the action of g,
′
′′
′
′′
we see that (Λr (Vj ) ⊗ Λr (gVj )) ⊕ g(Λr (Vj ) ⊗ Λr (gVj )) is a kG-submodule isomorphic to
′
′′
⊗G
r
r
r
(Λr (Vj ) ⊗ Λr (gVj ))↑G
H and Λ (Vj ) ⊗ gΛ (Vj ) is a submodule isomorphic to Λ (Vj )↑H .
(c) The proof is similar to that of (b). Note that, if r is odd, the summand corresponding
to r ′ = r ′′ which leads to the tensor induced submodule in (b) does not occur.
(d),(e) We say that a kG-module is induced except for possibly one trivial summand if it is
G
isomorphic to A↑G
H or A↑H ⊕V1 for some kH-module A. We prove (d) and (e) simultaneously
by showing that for all 1 ≤ j ≤ 2n−1 the kG-module Vj ↑⊗G
H is induced except for possibly
one trivial summand. The claim then follows from the fact that dimk (Vj ↑⊗G
H ) is even if and
only if j is even.
∼
The proof is by induction on j. Because V1 ↑⊗G
= V1 we can assume j > 1. If j is
H
even, then the kH-module Vj is induced from a proper subgroup of H. So [4, I 3.15.2 (iv)]
implies that Vj ↑⊗G
H is a direct sum of modules induced from H (even from proper subgroups
of H). Assume that j is odd. So we can write j = 2m + j ′ with 1 ≤ m < n − 1 and
1 ≤ j ′ < 2m . First, we treat the case j ′ = 1. By the Mackey formula for tensor induction
m
G∼
∼
[4, I 3.15.2 (v)] we have Vj ↑⊗G
H ↓H = V2m +1 ⊗ V2m +1 = V1 ⊕ (2 − 2)V2m ⊕ 2V2m+1 , and so
⊗G ∼
m−1
Vj ↑H = V1 ⊕ (2
− 1)V2m+1 ⊕ V2m+2 , which is induced up to one trivial summand. Now
′
assume j > 1. Then Vj ′ ⊗ V2m +1 ∼
= Vj ⊕ (j ′ − 1)V2m as kH-modules. By [4, I 3.15.2 (i)]
and (a) we get
(2)
⊗G ∼
⊗G
⊗G
′
(Vj ′ ↑⊗G
H ) ⊗ (V2m +1 ↑H ) =ind (Vj ↑H ) ⊕ (j − 1)(V2m ↑H ).
By induction and the case j ′ = 1, we know that the left hand side of (2) is induced except
for possibly one trivial summand. Hence, Vj ↑⊗G
H is induced except for possibly one trivial
summand.
We can now see that the symmetric and exterior powers of even dimensional indecomposable modules have a particularly restricted form.
Corollary 3.6. Suppose that we have non-negative integers j, s, t, u with u, j odd and s ≥ 1.
t
t
Furthermore, assume that 2t u < 2n and 2s j ≤ 2n . Then Λ2 u (V2s j ) and S 2 u (V2s j ) are both
t
t
induced unless t ≥ s. If t ≥ s then Λ2 u (V2s j ) ∼
= m′ V1 ⊕ Y , where
= mV1 ⊕ X and S 2 u (V2s j ) ∼
′
X, Y are induced modules and m and m are the numbers of non-induced indecomposable
t−s
t−s
summands in Λ2 u (Vj ) and Λ2 u (V2n−s −j ), respectively.
t
Proof. Using Lemma 3.5 (a),(b),(d) we see that, up to induced direct summands, Λ2 u (V2s j )
is tensor-induced from a subgroup of index 2min{s,t} . If t < s then, up to induced direct
summands, it is tensor-induced from Λu (V2s−t j ) and thus is induced, by part (c) of the same
lemma. If t ≥ s then, again up to induced direct summands, it is tensor-induced from
t−s
Λ2 u (Vj ); the description given is then seen to be valid using parts (a), (d) and (e). The
t
t
case of S 2 u (V2s j ) reduces to that of Λ2 u (V2n −2s j ), by Theorem 1.2.
Corollary 3.7. If X is a kG-module such that every direct summand has dimension divisible
by 4 then S 2 (X) is induced.
Proof. By the identity S 2 (A ⊕ B) ∼
= S 2 (A) ⊕ S 2 (B) ⊕ A ⊗ B, we may assume that X is
indecomposable, say X = V4u . The claim now follows from Corollary 3.6.
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
8
In the proof of our main result we will often have information only modulo induced direct
summands. The following definition and lemmas deal with the splitting of maps in such
situations.
Recall that a map f : A → B of kG-modules is split injective, if there is a map g : B → A
of kG-modules such that g ◦ f = IdA . For maps f : A → B, f ′ : A → B ′ of kG-modules we
write (f, g) : A → B ⊕ B ′ , a 7→ (f (a), f ′ (a)).
Definition 3.8. Let f : A → B be a map of kG-modules. We say that f is split injective
modulo induced summands if there exists an induced kG-module X and a map f ′ : A → X
of kG-modules such that (f, f ′ ) : A → B ⊕ X is split injective.
Split injective modulo induced summands behaves in much the same way as split injective.
Lemma 3.9. Given maps f : A → B, g : B → C and h : D → E of kG-modules:
(a) if f and g are split injective modulo induced summands then so is g ◦ f ,
(b) if g ◦ f is split injective modulo induced summands then so is f ,
(c) if f , h are split injective modulo induced summands then so is f ⊗h : A⊗D → B ⊗E.
Proof. (a) By assumption, we have induced modules X, Y and maps f ′ : A → X, u : B → A,
u′ : X → A, g ′ : B → Y , v : C → B, v ′ : Y → B such that u ◦ f + u′ ◦ f ′ = IdA and
v ◦ g + v ′ ◦ g ′ = IdB . We define (g ◦ f )′ : A → X ⊕ Y, a 7→ (f ′ (a), g ′ ◦ f (a)), w : C → A, c 7→
u ◦ v(c) and w ′ : X ⊕ Y → A, (x, y) 7→ u′ (x) + u ◦ v ′ (y). Then w ◦ (g ◦ f ) + w ′ ◦ (g ◦ f )′ = IdA .
Parts (b) and (c) are proved in a similar way; the proofs are left to the reader.
Lemma 3.10. Let f : A → B be a map of kG-modules and write A = A′ ⊕ A′′ , where A′ has
only non-induced summands and A′′ has only induced summands. Let i denote the inclusion
of A′ in A. Then f is split injective modulo induced summands if and only if f ◦ i is split
injective.
Proof. Suppose that f is split injective modulo induced summands; we want to show that
f ◦ i is split injective. By Lemma 3.9 (a), the map f ◦ i is split injective modulo induced
summands, so we can assume that A = A′ and we have to show that f is split injective.
Since f is split injective modulo induced summands we have an induced module X and
maps f ′ : A → X, u : B → A, u′ : X → A such that u◦f +u′ ◦f ′ = IdA . Since X and A have
no summands
in common, we know that u′ ◦ f ′ lies in the radical of EndkG (A) (note that
L
if A =
Ai with Ai indecomposable and we write elements of EndkG (A) as matrices with
entries in HomkG (Ai , Aj ) then the radical consists of the morphisms for which no component
is an isomorphism). Thus u ◦ f is surjective, hence an automorphism of A, and f is split
injective.
Conversely, suppose that f ◦ i : A′ → B is split injective, so there is a map g : B → A′
such that g ◦ (f ◦ i) = IdA′ . Let j denote the inclusion of X := A′′ in A and f ′ the projection
of A onto A′′ . We define v := i ◦ g : B → A, and v ′ : X → A, x 7→ −(i ◦ g ◦ f ◦ j)(x) + j(x).
Then v ◦ f + v ′ ◦ f ′ = IdA , so f is split injective modulo induced summands.
Remark. The proof above shows that the induced module X in Definition 3.8 can always be
chosen in such a way that X only contains indecomposable direct summands that also occur
in A.
Remark. Definition 3.8 makes sense for any finite group and any class of indecomposable
modules and Lemmas 3.9(a,b) and 3.10 remain true.
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
9
It will turn out that certain symmetric and exterior powers of modules for cyclic 2-groups
are contained in the Z-submodule c(G) of the Green ring a(G) spanned by the indecomposable modules Vr for r satisfying r 6≡ 2 (mod 4). We describe some properties of c(G).
Lemma 3.11. The submodule c(G) is
(a) a subring of a(G) and
(b) closed under Ω2n .
Proof. Part (b) is clear from the definitions.
For part (a) we need to show that Vi ⊗ Vj ∈ c(G) for all 0 ≤ i, j ≤ 2n , i, j 6≡ 2 (mod 4).
For n = 1 we only have V1 ⊗ V1 = V1 ∈ c(G). Suppose that n > 1. By the remarks on
the computation of tensor products at the beginning of this section, we have Vi ⊗ Vj =
′
′ ′
n−1
′
and i′ ≡ ±i (mod 4)
Ωm
2n (Vi′ ⊗ Vj ′ ) ⊕ m V2n for some integers m, m , where 0 ≤ i , j ≤ 2
′
and j ≡ ±j (mod 4). We can consider Vi′ and Vj ′ as modules for H ∼
= C2n−1 and the claim
follows from induction and part (b).
4. Main Theorem
From now on we assume that k = F2 is a field with 2 elements and G is a cyclic group of
order 2n . For 0 ≤ s ≤ 2n−1 , we know from Lemma 2.7 that K(V2n−1 +s , Vs2 ) is acyclic and
that its homology in degree 0 is S(V2n−1 +s )/(Vs2 ). It will turn out that S(V2n−1 +s )/(Vs2 ) is
closely related to the exterior algebra Λ(V2n−1 +s ), so it is natural to study the structure of
L
2
e 2n−1 +s ) =
er
the graded ring S(V
r≥0 S (V2n−1 +s ) := S(V2n−1 +s )/(Vs ) as a kG-module. For a
L
er
non-negative integer m write Se<m (V2n−1 +s ) = m−1
r=0 S (V2n−1 +s ) and use a similar notation
e (V2n−1 +s ) denote the kernel of the natural epimorphism
for other graded modules. Let N
e
S(V2n−1 +s ) → Λ(V2n−1 +s ).
Choose a k-basis {x1 , x2 , . . . , x2n−1 +s } of V2n−1 +s as in Section 3. For simplicity, write
xtop := x2n−1 +s , xtop−1 := x2n−1 +s−1 and so on. Each element of S(V2n−1 +s ) can be written
Qn
uniquely as a polynomial in x1 , x2 , . . . , xtop . Set a := 2i=1 (g i xtop ) ∈ S(V2n−1 +s ). If s < 2n−1
e 2n−1 +s ), and if s = 2n−1 let e
let e
a be the image of a in S(V
a be the image of the element
Q
n−1
2
e
n
a is still invariant, because g 2 x2n = x2n +x2n−1
t∈G/C2 (txtop ) in S(V2 ). In the latter case e
e 2n ) is invariant under C2 . In all cases, a is homogeneous of degree 2n , has
and so x2top ∈ S(V
degree 2n when considered as a polynomial in xtop , the elements a and e
a are invariant under
n−1
the action of G, and the image of e
a in Λ(V2n−1 +s ) is 0. If s = 2
then we also write eb for
Q2n i
e 2n ).
the image of the element i=1 (g xtop ) in S(V
The next theorem is our main result. Since any representation of G ∼
= C2n over a field of
characteristic 2 can be written in F2 , part (d) implies Theorem 1.1, but we record the other
parts since they are also of interest and they form an integral part of the proof.
Theorem 4.1. Let n and s be integers such that n ≥ 1 and 0 ≤ s ≤ 2n−1 .
(a) (Separation) The complex K(V2n−1 +s , Vs2 ) of kG-modules is separated.
n
e 2n−1 +s ) ∼
a] ⊗ Se<2 (V2n−1 +s ) as graded
(b) (Periodicity) For s < 2n−1 we have S(V
=ind k[e
n
e 2n ) ∼
a] ⊗ (Se<2 (V2n ) ⊕ keb). In both cases
kG-modules. For s = 2n−1 we have S(V
=ind k[e
e 2n−1 +s ).
the isomorphism from right to left is induced by the product in S(V
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
10
(c) (Splitting) The short exact sequence of graded kG-modules
e (V2n−1 +s ) −→ S(V
e 2n−1 +s ) −→ Λ(V2n−1 +s ) −→ 0
0 −→ N
e (V2n−1 +s ) = e
e 2n−1 +s ) ⊕ I,
e where Ie is a kG-module induced from H.
is split and N
aS(V
(d) (Exterior powers) For each r ≥ 0 we have the following isomorphism of kG-modules
M i+j
Ω2n (Λi (Vs ) ⊗ Λj (V2n−1 −s )).
Λr (V2n−1 +s ) ∼
=proj
i,j≥0
2i+j=r
The case s = 0 is a little unnatural, but we need it for the induction, because the restriction
of V2n−1 +1 is V2n−2 +1 ⊕ V2n−2 .
It is sometimes more succinct to consider Hilbert series with coefficients in the Green ring
(possibly modulo projectives or induced modules). For more details see [12]. In particular,
we consider the following series associated to a kG-module V :
∞
∞
P
P
λt (V ) =
Λr (V )tr ,
σt (V ) =
S r (V )tr ,
σ
et (V ) =
r=0
∞
P
r=0
er
r
λΩ
t (V
S (V )t ,
)=
r=0
∞
P
Ωr Λr (V )tr .
r=0
The last of these requires G to be specified in order for the Ω to be determined; it is
naturally considered modulo projectives. They all commute with restriction and turn direct
sums of modules into products of series. This is all an easy consequence of the corresponding
properties of the corresponding functors on modules, except perhaps for λΩ
t (V ⊕ W ), where
r+s
r
s
∼
we need the formula Ω V ⊗ Ω W =proj Ω (V ⊗ W ).
Many of our statements about modules imply Hilbert series versions.
n
(3)
2 −1
σt (V2n−1 +s ) =ind λΩ
t (V2n−1 −s )(1 − t )
σ
et (V2n−1 +s ) =ind λΩ
t2 (Vs )σt (V2n−1 +s )
n
σ
et (V2n−1 +s ) =ind λt (V2n−1 +s )(1 − t2 )−1
Ω
λt (V2n−1 +s ) =proj λΩ
t2 (Vs )λt (V2n−1 −s )
Theorem 1.2
Separation 4.1(a)
Splitting and periodicity 4.1(b),(c)
Exterior powers 4.1(d),
where the symbols =ind and =proj mean that we consider the coefficients only modulo induced
or projective direct summands, respectively. The first and last of the above identities are, in
fact, equivalent to the original versions. The second identity follows from Theorem 4.1 (a),
Lemma 2.5 and Lemma 2.7 (once the theorem is proved).
Remark. An easy calculation shows that, for fixed n and s, the last of the formulas in (3)
follows formally from the first three if we are satisfied with only =ind .
Remark. The proof of Theorem 1.2 in [21] actually gives a more precise formula than the
first one in (3). It works by showing that the complex K(V2n , V2n−1 −s ) defined in [21] is
separated and then applying Lemma 2.5; note that the definition of K(V2n , V2n−1 −s ) in [21]
is different from our Definition 2.6. The result is that σt (V2n−1 +s ) =proj σt (V2n )λΩ
t (V2n−1 −s ).
Since V2n can be given a basis that is permuted by G, each S r (V2n ) has a monomial basis
that is permuted. For small n, the decomposition of σt (V2n ) can be calculated by hand;
in general the calculation can be organized using [21, Proposition 2.2]. Alternatively, [21,
Proposition 2.2] can be applied directly to σt (V2n−1 +s ).
The next six sections are devoted to the proof of Theorem 4.1 by induction on n.
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
11
5. The Case n = 1
In this section we start the inductive proof of Theorem 4.1. Suppose that n = 1, so we
have to prove the statements in Theorem 4.1 for s ∈ {0, 1}. With these assumptions on n
and s, parts (b)-(d) of Theorem 4.1 can easily be verified by a direct calculation. In fact,
in (b) one obtains isomorphisms of kG-modules (not only modulo induced summands), and
in (c) one gets Ie = 0. Separation for s = 0 is trivial.
Let us consider part (a) for n = s = 1. We have to show that for each r > 0 the short
exact sequence 0 → S r−2 (V2 ) → S r (V2 ) → Ser (V2 ) → 0 of kG-modules is separated at S r (V2 ).
If r is odd, then S r (V2 ) is induced by Theorem 1.2, hence projective, and so separation is
obviously true. Separation is trivial for r = 0. For even r > 0, a direct calculation and
r−2
(V2 ), and so separation
Theorem 1.2 show that Ser (V2 ) ∼
=ind S r (V2 ) ⊕ Ω−1
= V1 ⊕ V1 ∼
2 S
follows from Lemma 3.4.
Sections 6-10 comprise the inductive step in the proof of Theorem 4.1. In these sections
we always assume that n > 1 is an integer and that Theorem 4.1 holds for all smaller values
of n. Throughout these sections the notation remains the same as in Sections 3 and 4; thus
G = hgi ∼
= C2n is a cyclic group of order 2n , k = F2 is a field with two elements and s is an
integer such that 0 ≤ s ≤ 2n−1 .
6. Periodicity
In this section we prove part (b) of Theorem 4.1, assuming that parts (a)-(d) of the
theorem hold for all smaller values of n.
Let H be the unique maximal subgroup of G and let {x1 , x2 , . . . , xtop } be a k-basis
n
of V2n−1 +s as in Section 4. We choose G-invariant elements a ∈ S 2 (V2n−1 +s ) and e
a ∈
n
2
e
S (V2n−1 +s ) as in Section 4. Let T (V2n−1 +s ) be the kG-submodule of S(V2n−1 +s ) spanned
n
by the monomials in x1 , . . . , xtop that are not divisible by x2top . We have S(V2n−1 +s ) ∼
=
<2n
<2n
n−1
n−1
n−1
(V2 +s ) = S (V2 +s ).
k[a] ⊗ T (V2 +s ) as kG-modules; see [21, Lemma 1.1]. So T
n
Notice that the periodicity of S(V2n−1 +s ) in [21, Theorem 1.2] is equivalent to T ≥2 (V2n−1 +s )
being induced. In fact, we know something stronger from [21, Corollary 3.11], namely that
n−1
T >2 −s (V2n−1 +s ) is induced.
e 2n−1 +s ), obtaining S(V
e 2n−1 +s ) ∼
a]⊗Te(V2n−1 +s )
We can make the same construction for S(V
= k[e
as kG-modules.
Define L(V2n−1 +s , Vs2 ) to be the subcomplex of K(V2n−1 +s , Vs2 ) defined using T (V2n−1 +s )
instead of S(V2n−1 +s ), that is
d
d
d
· · · −→ T (V2n−1 +s ) ⊗ Λ2 (Vs ) −→ T (V2n−1 +s ) ⊗ Vs −→ T (V2n−1 +s ),
where the boundary morphisms are as in Definition 2.6 (this can be done since the xtop used
in the definition of T (V2n−1 +s ) is not contained in Vs ). Thus L(V2n−1 +s , Vs2 ) is a complex of
graded kG-modules; it is exact except in degree 0, where the homology is H0 (L(V2n−1 +s , Vs2 )),
which is isomorphic to Te(V2n−1 +s ) as a kG-module. Notice that, by construction, the complexes K(V2n−1 +s , Vs2 ) and k[a]⊗L(V2n−1 +s , Vs2 ) of kG-modules are isomorphic. In particular,
note for later use that one of them is separated (over G or over H) if and only if the other
is so too.
From now on we fix s and abbreviate the notation to just S, T , K, L, etc.
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
12
Suppose that s < 2n−1 . We claim that Lri = T r−2i ⊗ Λi (Vs ) is induced for all i ≥ 0, r ≥ 2n .
We may assume that i ≤ s. Then r − 2i ≥ 2n − 2s > 2n−1 − s and so T r−2i is induced. Thus
Lr is a complex of induced kG-modules for each r ≥ 2n .
Consider the restriction of the complex K to the subgroup H. It decomposes as a tensor
product of two complexes, by Lemma 2.8. Each of these is separated, by induction and
Theorem 4.1 (a), hence so is their product, by Lemma 2.2. It follows that for each r ≥ 0,
the complex Lr is separated on restriction to H. We have just seen that Lr is a complex
of induced modules for all r ≥ 2n . Thus, for each r ≥ 2n , the complex Lr is separated, by
n
n
Proposition 2.4. Now Lemma 2.5 shows that H0 (L≥2 ) is induced. But this is exactly Te≥2 ,
n
n
e 2n−1 +s ) ∼
so S(V
a] ⊗ Te ∼
a] ⊗ Te<2 = k[e
a] ⊗ Se<2 is periodic if s < 2n−1 .
= k[e
=ind k[e
n
Now suppose that s = 2n−1 . By the same argument as for s < 2n−1 , we see that Te>2 is
n
a ⊕ keb.
induced. To complete the proof of Theorem 4.1 (b) we have to show that Se2 ∼
=ind ke
n
n
2 −i n
Set yi := g
x2 for i = 1, 2, . . . , 2 , so {y1 , . . . , y2n } is a k-basis of V2n which is permuted
n−1
by G. A basis for V2n−1 < V2n is given by g 2 yi − yi = yi+2n−1 − yi for i = 1, 2, . . . , 2n−1 .
e 2n ), so yei2 = ye2 n−1 for i = 1, . . . , 2n−1 . The set consisting
Write yei for the image of yi in S(V
i+2
n
of all monomials of degree 2 in all the yei such that ye1 , . . . , ye2n−1 only occur to the power
e 2n ). The group G permutes these monomials and it is
at most 1 forms a k-basis for S(V
straightforward to check that there are two invariant monomials, namely ye1 ye2 · · · ye2n = eb and
a; the rest span induced submodules. This completes the proof of
ye22n−1 +1 ye22n−1 +2 · · · ye22n = e
periodicity.
7. Splitting
In this section we prove part (c) of Theorem 4.1, assuming the whole of the theorem for
smaller n.
Let H be the unique maximal subgroup of G and {x1 , x2 , . . . , xtop } a k-basis of V2n−1 +s as
e 2n−1 +s ) for the kernel of the natural surjection
in Section 4. As in Theorem 4.1 we write N(V
f
e 2n−1 +s ) → Λ(V2n−1 +s ). The following proposition deals with the structure of S(V
e 2n−1 +s )
S(V
n
in degrees less than 2 .
Proposition 7.1. For any integer s such that 0 ≤ s ≤ 2n−1 , the short exact sequence
(4)
n
f
e <2n (V2n−1 +s ) −→ Se<2n (V2n−1 +s ) −→
0 −→ N
Λ<2 (V2n−1 +s ) −→ 0
e <2n (V2n−1 +s ) is induced from H.
of graded kG-modules is split, and N
Before starting with the proof of Proposition 7.1 we introduce some further notation. As
described at the beginning of Section 3, we have V2n−1 +s ↓G
H = V2n−2 +s′ ⊕ V2n−2 +s′′ where
′ ′′
n−2
′
′′
′
′′
0 ≤ s ,s ≤ 2
and s = s or s = s + 1. The kH-submodule V2n−2 +s′ of V2n−1 +s
has the k-basis {xtop , xtop−2 , xtop−4 , . . . } and the kH-submodule V2n−2 +s′′ has the k-basis
{xtop−1 , xtop−3 , xtop−5 , . . . }. We write Se′ (V2n−2 +s′ ) and Se′′ (V2n−2 +s′′ ) for S(V2n−2 +s′ )/(Vs2′ ) and
S(V2n−2 +s′′ )/(Vs2′′ ), respectively. So the xi with odd i and the xi with even i provide natural
e 2n−1 +s ) of kH-modules, and
e 2n−1 +s ) and Se′′ (V2n−2 +s′′ ) → S(V
embeddings Se′ (V2n−2 +s′ ) → S(V
we have
e 2n−1 +s ) ∼
S(V
= Se′ (V2n−2 +s′ ) ⊗ Se′′ (V2n−2 +s′′ )
as kH-modules, where the isomorphism is given by f1 ⊗ f2 7→ f1 · f2 .
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
13
e 2n−2 +s′ ) according to the description preceding Theoa′ ∈ S(V
Choose a′ ∈ S(V2n−2 +s′ ), e
rem 4.1, but working over H. Thus a′ is homogeneous of degree 2n−1 , has degree 2n−1 when
considered as a polynomial in xtop . Furthermore, a′ and e
a′ are invariant under the action of H
a′′ ∈ Se′′ (V2n−2 +s′′ ).
and e
a′ has image 0 in Λ(V2n−2 +s′ ). Similarly, choose a′′ ∈ S(V2n−2 +s′′ ) and e
So a′′ is homogeneous of degree 2n−1 , has degree 2n−1 when considered as a polynomial in
xtop−1 and a′′ and e
a′′ are invariant under the action of H and e
a′′ has image 0 in Λ(V2n−2 +s′′ ).
By induction, e
a′ and e
a′′ are periodicity generators of Se′ (V2n−2 +s′ ) and Se′′ (V2n−2 +s′′ ), respectively. That is, we have
n−1
n−1
a′′ ] ⊗ Se′′<2 (V2n−2 +s′′ )
a′ ] ⊗ Se′<2 (V2n−2 +s′′ ) and Se′′ (V2n−2 +s′′ ) ∼
Se′ (V2n−2 +s′ ) ∼
=ind k[e
=ind k[e
or the variant with eb′ or eb′′ if s′ = 2n−2 or s′′ = 2n−2 .
Lemma 7.2. Let s be an integer such that 0 ≤ s ≤ 2n−1 and let e
a′ be a periodicity generator
for Se′ (V2n−2 +s′ ) as above. Then
n
n−1
n−1
n−1
e
Se<2 (V2n−1 +s )↓G = Se<2 (V2n−1 +s ) ⊕ e
a′ Se<2 (V2n−1 +s ) ⊕ (ge
a′ )Se<2 (V2n−1 +s ) ⊕ X
H
e is generated as a k-vector space by the images
as kH-modules, whereLthe
kH-submodule X
2n −1
r
of all monomials x ∈ r=2n−1 S (V2n−1 +s ) such that x has degree strictly less than 2n−1 when
considered as a polynomial in xtop and x has degree strictly less than 2n−1 when considered
as a polynomial in xtop−1 .
Proof. We give all monomials in S(V2n−1 +s ) the lexicographic order with xtop−1 > xtop >
L2n −1
n−1
r
xtop−2 > · · · > x1 . Let h ∈
a′ has leading term x2top−1 we
r=2n−1 S (V2n−1 +s ). Since ge
n−1
n
can write h as h = h1 · ge
a′ + h2 where h1 ∈ Se<2 (V2n−1 +s ) and h2 ∈ Se<2 (V2n−1 +s ) such
that h2 has degree < 2n−1 when considered as a polynomial in xtop−1 . Then, because e
a′
n−1
has leading term x2top and only involves monomials in xtop , xtop−2 , xtop−4 , . . . , we can find
L2n −1
n−1
r
e such that h2 = h3 · e
h3 ∈ Se<2 (V2n−1 +s ) and h4 ∈ X
a′ + h4 . Thus, r=2
n−1 S (V2n−1 +s ) is
n−1
n−1
e Comparing dimensions, we see
the sum of e
a′ Se<2 (V2n−1 +s ), (ge
a′ )Se<2 (V2n−1 +s ) and X.
that this sum has to be direct and Lemma 7.2 follows.
We are now ready to prove Proposition 7.1.
Proof. (of Proposition 7.1) We study the restriction of the sequence (4) to the maximal
subgroup H. By Lemma 7.2, the middle term is
n
n−1
e
Se<2 (V2n−1 +s )↓G = Se<2 (V2n−1 +s ) ⊕ Je ⊕ X,
H
n−1
where Je := e
a′ Se
(V2n−1 +s ) ⊕ (ge
a′ )Se<2 (V2n−1 +s ). Owing to the choice of e
a′ , we have
n
n
e <2 (V2n−1 +s ) and therefore Je ⊆ N
e <2 (V2n−1 +s ). In fact, by construction, Je is a kGe
a′ ∈ N
L2n −1
r
e (V2n−1 +s ) and is induced from H. We consider the exact sequence
submodule of r=2n−1 N
of kG-modules
n −1
2M
Ser (V2n−1 +s ) −→ X −→ 0,
(5)
0 −→ Je −→
<2n−1
L2n −1
r=2n−1
er
e
where X :=
r=2n−1 S (V2n−1 +s )/J. We know from Lemma 7.2 that the sequence (5) is
∼ e
e
split when restricted to H and X ↓G
H = X as kH-modules. Since J is induced from H it
is relatively H-injective, and so the sequence (5) splits over kG (see [8, Theorem (19.2)]).
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
14
n
Thus Je is a direct summand of Se<2 (V2n−1 +s ) over kG, so there is a kG-submodule Je′′ of
n
n
e <2n (V2n−1 +s ), it follows that
Se<2 (V2n−1 +s ) such that Se<2 (V2n−1 +s ) = Je ⊕ Je′′ . Since Je ⊆ N
e <2n (V2n−1 +s ) = Je ⊕ Je′ , where Je′ := Je′′ ∩ N
e <2n (V2n−1 +s ). We have
N
n−1
n−1
a′ Se<2 (V2n−1 +s )↓G
a′ )Se<2 (V2n−1 +s )↓G
Je↓G
H =e
H ⊕ (ge
H.
n−1
<2n−1
∼ e′
e′′
, we have, by induction and
Because Se<2 (V2n−1 +s )↓G
H = (S (V2n−2 +s′ ) ⊗ S (V2n−2 +s′′ ))
ignoring the grading,
n−1
n−1
(6)
Je↓G ∼
=ind (Λ′ ⊗ Λ′′ )<2 ⊕ (Λ′ ⊗ Λ′′ )<2 .
H
′
′′
Here we write Λ ⊗ Λ for the graded kH-module Λ(V2n−2 +s′ ) ⊗ Λ(V2n−2 +s′′ ). Restricting the
sequence (4) to H, we obtain the sequence
e <2n (V2n−1 +s )↓G → (Se′ (V2n−2 +s′ ) ⊗ Se′′ (V2n−2 +s′′ ))<2n → (Λ′ ⊗ Λ′′ )<2n → 0
0→N
H
of kH-modules, which is split by induction. Thus, by induction again, we obtain
n−1
∼ a′ (Λ′ ⊗ Λ′′ )<2n−1 ⊕ e
e <2n (V2n−1 +s )↓G
(7)
N
a′′ (Λ′ ⊗ Λ′′ )<2 .
H =ind e
e′ G ∼ e e′ G ∼ e <2n (V2n−1 +s )↓G ∼
e G
Equations (6) and (7) imply that Je↓G
H ⊕ J ↓H = (J ⊕ J )↓H = N
H =ind J ↓H . It
′
G
follows that Je ↓H is induced from proper subgroups of H. By Lemma 3.1 the kG-module Je′
e <2n (V2n−1 +s ) = Je⊕ Je′ is induced from H. We have just seen
is induced from H, and hence N
e <2n (V2n−1 +s ) is relatively H-injective
that sequence (4) is split on restriction to H; since N
the sequence must split over kG. This completes the proof of Proposition 7.1.
e 2n−1 +s ) and the exterior powers
The following corollary provides a connection between S(V
of V2n−1 +s in degrees less than 2n .
Corollary 7.3. For r and s integers such that 0 ≤ s ≤ 2n−1 and 0 ≤ r < 2n , the map f
induces an isomorphism of kG-modules modulo induced summands
Ser (V2n−1 +s ) ∼
=ind Λr (V2n−1 +s ).
Proof. This is clear from Proposition 7.1 (for n > 1) and Section 5 (for n = 1).
We can now prove Theorem 4.1 (c). For s < 2n−1 we have Se (V2n−1 +s ) ∼
= Λ(V2n−1 +s )⊕X,
where X is induced, so part (c) of Theorem 4.1 follows from part (b). For s = 2n−1 we have
n
Se<2 (V2n−1 +s ) ⊕ keb ∼
= Λ(V2n−1 +s ) ⊕ X ′ , where X ′ is induced. Note that eb maps to a generator
2n
of Λ (V2n−1 +s ). Again, part (c) of Theorem 4.1 is a consequence of (b).
<2n
8. Preparation for Separation
In this section we prepare for the proof of part (a) of Theorem 4.1, assuming the whole of
the theorem for smaller n.
Let H be the unique maximal subgroup of G and let {x1 , x2 , . . . , xtop } be a k-basis of
V2n−1 +s as in Section 4. The main goal of this section is to develop useful criteria for the
complex K(V2n−1 +s , Vs2 ) to be separated.
Lemma 8.1. Let r, s be non-negative integers such that 0 ≤ s ≤ 2n−1 . Suppose that for each
′
′
0 ≤ r ′ < r with r ′ ≡ r mod 2, the complex K r (V2n−1 +s , Vs2 ) is separated at K0r (V2n−1 +s , Vs2 ) =
′
S r (V2n−1 +s ). Then K r (V2n−1 +s , Vs2 ) is separated at Kir (V2n−1 +s , Vs2 ) = S r−2i (V2n−1 +s )⊗Λi (Vs )
for all i ≥ 1. The same is true when K and S are replaced by L and T from Section 6.
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
15
Proof. We only demonstrate the proof for K and S; the proof for L and T is analogous.
We write V := V2n−1 +s and W := Vs for short. Fix i ≥ 1 and consider the boundary morphism di+1 : S r−2i−2 (V )⊗Λi+1 (W ) → S r−2i (V )⊗Λi (W ) in K r (V, W 2 ). We have to show that
Im(di+1 ) → S r−2i (V ) ⊗ Λi (W ) factors through a projective kG-module. Since K r−2i (V, W 2 )
is separated at K0r−2i (V, W 2 ) the inclusion (W 2 )r−2i → S r−2i (V ) factors through a projective
kG-module P r−2i . We can write the inclusion Im(di+1 ) → S r−2i (V )⊗Λi (W ) as a composition
of inclusions
Im(di+1 ) → (W 2 )r−2i ⊗ Λi (W ) → S r−2i (V ) ⊗ Λi (W ),
where the last map factors through the projective kG-module P r−2i ⊗ Λi (W ).
Lemma 8.2. Let s and r be integers such that 0 ≤ s ≤ 2n−1 and 0 < r < 2n , and suppose
that the complex K i (V2n−1 +s , Vs2 ) is separated for all 0 ≤ i < r. Then the following statements
are equivalent:
(a) K r (V2n−1 +s , Vs2 ) is separated,
g
(b) The natural map S r (V2n−1 +s ) −→ Λr (V2n−1 +s ) is split injective modulo induced summands,
L
i+j
i
j
(c) Λr (V2n−1 +s ) ∼
i,j≥0 Ω2n (Λ (Vs ) ⊗ Λ (V2n−1 −s )).
=ind
2i+j=r
Proof. We write S r := S r (V2n−1 +s ), Ser := Ser (V2n−1 +s ) and K i := K i (V2n−1 +s , Vs2 ). The
conditions on K i and Lemma 8.1 show that K r is separated except, perhaps, at K0r = S r .
The restriction of the complex K r to H decomposes as a tensor product of two complexes,
by Lemma 2.8. Each of these is separated by our continuing induction hypothesis, hence so
is their product, by Lemma 2.2, and so K r is separated on restriction to H. Thus the short
exact sequence
(8)
j
i
0 −→ Im(d1 ) −→ S r −→ Ser −→ 0
from K r is separated at S r on restriction to H (the maps i, j should not be confused with the
indices in part (c) of the lemma). The separation
of K r in positive (complex-) degrees and
L
i−1
i
j
Lemma 2.5 yield the formula Im(d1 ) ∼
i≥1, j≥0 Ω2n (Λ (Vs )) ⊗ S (V2n−1 +s ). Theorem 1.2
=proj
2i+j=r
now shows that
M i+j−1
Ω n (Λi (Vs ) ⊗ Λj (V2n−1 −s )).
(9)
Im(d1 ) ∼
=ind
2
i≥1, j≥0
2i+j=r
(a) ⇒ (b) Let f : Ser → Λr be the natural surjection, so g = f ◦ j. By Proposition 7.1,
the map f is split injective modulo induced summands, and, by Lemma 3.9, it is enough
to show that j is split injective modulo induced summands. By assumption, S r = X ⊕ M
for some submodules X and M of S r such that X is projective and ker(j) = Im(d1 ) ⊆ X.
Let j ′ : S r → X be the projection onto X and u′ : X → S r the natural embedding. Define
u : Ser = j(X) ⊕ j(M) → S r , j(x) + j(m) 7→ m (note that the restriction of j to M is
injective). Then u ◦ j + u′ ◦ j ′ = IdS r and so j is split injective modulo induced summands.
(b) ⇒ (c) Assume (b). The factorization g = f ◦j and Lemma 3.9 (b) imply that j is also split
injective modulo induced summands. Write S r = A′ ⊕ A′′ , where A′ has only non-induced
summands and A′′ is induced. By Lemma 3.10, the restriction of j to A′ is split injective,
so j maps A′ injectively into Ser and j(A′ ) is a direct summand of Ser . Factoring out A′ and
j
i
j(A′ ) in (8) we obtain the short exact sequence 0 −→ Im(d1 ) −→ S r /A′ −→ Ser /j(A′ ) −→ 0.
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
16
As we have seen at the beginning of the proof, i factors through a projective on restriction
to H, and so the same is true for i. Thus the complex Im(d1 ) → S r /A′ is separated on restriction to H. Because S r /A′ ∼
= A′′ is induced, the complex is separated, by Proposition 2.4.
Lemma 2.5 yields Ser /j(A′ ) ∼
=ind Ω2n Im(d1 ). Using (9) we obtain
=proj S r /A′ ⊕ Ω2n Im(d1 ) ∼
M i+j
Ω2n (Λi (Vs ) ⊗ Λj (V2n−1 −s )).
(10)
Ser /j(A′ ) ∼
=ind
i≥1, j≥0
2i+j=r
Theorem 1.2 implies j(A′ ) ∼
=ind Ωr2n Λr (V2n−1 −s ). Adding the summand j(A′ )
=ind S r ∼
= A′ ∼
to both sides of (10) and using Corollary 7.3 gives us the formula in (c).
(c) ⇒ (a) Assume that (c) holds. From Corollary 7.3, Theorem 1.2 and (9) we get Ser ∼
=ind
r
Im(d
).
Separation
of
K
now
follows
from
applying
Lemma
3.4
to
the
short
exact
S r ⊕ Ω−1
1
2n
sequence (8).
Separation of K r (V2n−1 +s , Vs2 ) for r = 0, 1 is trivial. We will now prove it for r = 2. Notice
that if a non-zero map Va → Vb of kG-modules is to factor through a projective over C2n ,
then we must have a + b > 2n . This is because the map must factor through the projective
cover V2n ։ Vb , which has kernel V2n −b , into which Va will certainly be mapped if a ≤ 2n − b.
Lemma 8.3. For any integer s such that 0 ≤ s ≤ 2n−1 the complex K 2 (V2n−1 +s , Vs2 ) is
separated.
Proof. The complex in question is Vs ֒→ S 2 (V2n−1 +s ). By induction, the map factors through
a projective on restriction to H. Write S 2 (V2n−1 +s ) = A′ ⊕A′′ , where A′ has only non-induced
summands and A′′ has only induced summands. The component Vs → A′′ factors through a
projective, by Proposition 2.4.
We claim that the component Vs → A′ must be 0. From Theorem 1.2, we know that
2
S (V2n−1 +s ) =ind Λ2 (V2n−1 −s ); but V2n−1 −s is a module for C2n−1 , and it follows that A′
contains only summands of dimension ≤ 2n−1 . Let Vt be such a summand, so t ≤ 2n−1 and
suppose that there is a non-zero component Vs → Vt . It must factor through a projective on
restriction, where it is a map Vs′ ⊕ Vs′′ → Vt′ ⊕ Vt′′ , with s′ , s′′ , t′ , t′′ ≤ 2n−2 . By the discussion
above, none of the components can factor through a projective module over C2n−1 unless
they are 0.
We can readily prove separation when s is even.
Lemma 8.4. For any even integer s such that 0 ≤ s ≤ 2n−1 , the complex K(V2n−1 +s , Vs2 ) of
kG-modules is separated.
Proof. Write s = 2s′ . From Lemma 2.9 we know that K(V2n−1 +s , Vs2 ) ∼
= K(V2n−2 +s′ , Vs2′ )↑⊗G
H .
The right hand side is separated by Lemma 2.3 and our induction hypothesis.
In view of this lemma, we assume now that s is odd.
Lemma 8.5. Let s be an odd integer such that 0 < s < 2n−1 . Then (given our induction
hypothesis):
(a) Λr (V2n−1 −s ) ∈ c(G) for all r ≥ 0 and
(b) S r (V2n−1 +s ) ∈ c(G) for all 0 ≤ r < 2n−1 .
Here c(G) is the subgroup of the Green ring in Lemma 3.11.
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
17
Proof. For part (a), the dimension d = 2n−1 − s of the module is in the range where we
know that our formula for exterior powers (see Theorem 4.1 (d)) is valid by our continuing
induction hypothesis. The statement is clearly true for d = 1 and we can employ induction
on d, using the formula and the properties of c(G) in Lemma 3.11.
For part (b) we use the formula σt (V2n−1 +s ) =proj σt (V2n )λΩ
t (V2n−1 −s ) from the remark at
the end of Section 4 and part (a). The summands of S(V2n ) are permutation modules on a
monomial basis, so are in c(G) unless the stabilizer of a monomial is of index 2. But this first
happens in degree 2n−1 , because if a monomial fixed by a subgroup of order 2n−1 contains yi ,
it must also contain all 2n−1 elements of the orbit of yi .
9. Separation
First we make some general constructions related to symmetric and exterior powers of
vector spaces. It is convenient to do this integrally first and then reduce modulo 2. Let U
be a free module over the integers localized at 2, Z(2) . For r ≥ 0 set
T r (U) = U ⊗Z(2) · · · ⊗Z(2) U
(r times).
Let the symmetric group Σr act on T r (U) by permuting the factors. Factoring out the action
of Σr we get S r (U) = T r (U)/Σr = T r (U) ⊗Z(2) Σr Z(2) . We can also let Σr act on T r (U) by
permuting the factors and multiplying by the signature of the permutation, in which case we
write T r (U)σ . Similarly, on factoring out the action of Σr we obtain Λr (U) = T r (U)σ /Σr .
For any subset I ⊆ {1, . . . , r} we set ΣI := {π ∈ Σr | π(i) = i for all i 6∈ I} (so ΣI is a
subgroup of Σr isomorphic to Σ|I| ). For r ≥ 2, write r = 2p + t with 1 ≤ t ≤ 2p . Consider
the subgroup
(
Qr :=
Σ{1,...,2p } × Σ{2p +1,...,r}
if t < 2p
(Σ{1,...,2p } × Σ{2p +1,...,2p+1 } ) ⋊ hτ i if t = 2p
of Σr , where τ ∈ Σr is the involution mapping i to 2p + i for i = 1, 2, . . . , 2p . The importance of Qr lies the fact that the index |Σr : Qr | is odd. This can been seen as folp
p
lows: |Σ2p +t |/(|Σ2p | · |Σt |) = 2 t+t , which is equal to the coefficient of xt in (1 + x)2 +t ≡
p+1
p
p
p+1
p
(mod 4).
(1 + x2 )(1 + x)t (mod 2); also (1 + x)2 = (1 + x2 + 2X))2 ≡ 1 + 2x2 + x2
r
r
r
r
Define LS (U) := T (U)/Qr and LΛ (U) := T (U)σ /Qr . There are natural quotient maps
have sections trS : S r (U) → LrS (U)
qS : LrS (U) → S r (U) and qΛ : LrΛ (U) → Λr (U), which
P
and trΛ : Λr (U) → LrΛ (U) given by tr x := |Σr1:Qr | π∈Σr /Qr πx. These have the property
that qS ◦ trS = IdS r (U ) and qΛ ◦ trΛ = IdΛr (U ) . These maps are all natural transformations of
functors on free Z(2) -modules.
Writing r = 2p + t as before, we see from the description of Qr that
( p
S 2 (U) ⊗ S t (U)
if t < 2p
LrS (U) ∼
=
p
p
p
(S 2 (U) ⊗ S 2 (U))/C2 ∼
= S 2 (S 2 (U)) if t = 2p .
Similarly, if r ≥ 3 we have
( p
Λ2 (U) ⊗ Λt (U)
if t < 2p
LrΛ (U) ∼
=
p
p
p
(Λ2 (U) ⊗ Λ2 (U))/C2 ∼
= S 2 (Λ2 (U)) if t = 2p ,
because the involution τ has signature 1, provided that p ≥ 1.
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
18
Now let V be an F2 -vector space and let U be a free Z(2) -module such that V ∼
= F2 ⊗Z(2) U.
r
r
r
r
r
Let L denote one of the functors S , Λ , LS , LΛ above and use it to define a functor with the
same name on F2 -vector spaces by Lr (V ) = F2 ⊗Z(2) Lr (U). This gives the expected result
for S r (V ) and Λr (V ).
In order to verify that Lr is really a functor on vector spaces, notice that if U and U ′ are
two free Z(2) -modules then the natural map HomZ(2) (U, U ′ ) → HomF2 (F2 ⊗Z(2) U, F2 ⊗Z(2) U ′ )
is surjective, so all maps of vector spaces lift. Furthermore, a map in the kernel has image
in 2U ′ , so factors through multiplication by 2 on U ′ . But multiplication by 2 on U ′ induces
multiplication by 2r on T r (U ′ )σ , thus it induces 0 on F2 ⊗Z(2) Lr (U ′ ).
It follows that the formulas above are also valid for F2 -vector spaces. A difference is that
we now have natural transformations er : S r → Λr and Lre : LrS → LrΛ induced by reducing
modulo squares.
The above functors induce functors on modules for a group in the obvious way.
Remark. Any representation of G over a field of characteristic 2 can be written in F2 , so
this is sufficient for our purposes. If we really needed functors on vector spaces over a bigger
field, this could be achieved by starting with a larger ring than Z(2) .
In the rest of this section we prove part (a) of Theorem 4.1, assuming the whole of the
theorem for smaller n. We use the same notation as before.
Lemma 9.1. Suppose that V is a kG-module, r ≥ 2 and Lre : LrS (V ) → LrΛ (V ) is split
injective modulo induced summands. Then er : S r (V ) → Λr (V ) is split injective modulo
induced summands.
Proof. Consider the commutative diagram
er
S r (V ) −−−→ Λr (V )
tr
trS y
y Λ
Lr
LrS (V ) −−−e→ LrΛ (V ).
The map trS is split injective, so if Lre is split injective modulo induced summands then so
is Lre ◦ trS , by Lemma 3.9 (a). But this is equal to trΛ ◦ er and Lemma 3.9 (b) shows that er
is split injective modulo induced summands.
Lemma 9.2. If s is an odd integer such that 0 < s < 2n−1 and r is an integer such that
0 ≤ r < 2n , then er : S r (V2n−1 +s ) → Λr (V2n−1 +s ) is split injective modulo induced summands.
Proof. We use induction on r. The cases r = 0, 1 are trivial and r = 2 is covered by
Lemma 8.3 combined with Lemma 8.2. Let r ≥ 3 and write r = 2p + t with 1 ≤ t ≤ 2p .
Abbreviate V2n−1 +s to V . By Lemma 9.1, it is sufficient to check that Lre : LrS (V ) → LrΛ (V )
is split injective modulo induced summands.
p
p
p
If t < 2p then Lre = e2 ⊗ et : S 2 (V ) ⊗ S t (V ) → Λ2 (V ) ⊗ Λt (V ). This is split injective
modulo induced summands by induction and Lemma 3.9 (c).
p
p
p
p
If t = 2p then Lre = S 2 (e2 ) : S 2 (S 2 (V )) → S 2 (Λ2 (V )). By induction, e2 is split injective
p
p
modulo induced summands, so it extends to a split injective map M : S 2 (V ) → Λ2 (V ) ⊕ X
with left inverse N, where X is induced. By the remark after Lemma 3.10, we may assume
p
that X only contains summands that are also summands of S 2 (V ); by Lemma 8.5 and the
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
19
assumption that r < 2n (and hence 2p < 2n−1, since t = 2p ), these are of dimension divisible
p
by 4. Applying S 2 , we see that S 2 (e2 ) extends to
p
p
p
S 2 (M) : S 2 (S 2 (V )) → S 2 (Λ2 (V )) ⊕ S 2 (X) ⊕ (Λ2 (V ) ⊗ X),
p
with left inverse S 2 (N). Certainly Λ2 (V ) ⊗ X is induced, and S 2 (X) is induced, by Corollary 3.7. Thus Lre and thus er are split injective modulo induced summands.
Again, let s be an odd integer such that 0 < s < 2n−1 . It follows from Lemmas 8.1, 8.2
and 9.2 that the complex K r (V2n−1 +s , Vs2 ) is separated for all 0 ≤ r < 2n .
Recall that K(V2n−1 +s , Vs2 ) is separated if and only if the complex L(V2n−1 +s , Vs2 ) from
Section 6 is separated. For the rest of this section we will write just K, L etc. . Now Lr is
separated for all 0 ≤ r < 2n , because it coincides with K r in this range. We will show that
Lr is separated for r ≥ 2n by induction on r, so let r ≥ 2n and assume that the complex is
separated in all lower degrees.
By Lemma 8.1, we can also assume that Lr is separated in positive (complex-)degrees, so
it is enough to prove that the short exact sequence
(11)
0 → Im(dr1 ) → T r → Ter → 0
is separated at T r . By Lemma 2.8, the restriction of K to the maximal subgroup H of G
decomposes as a tensor product of two complexes, and each of these is separated, by our
continuing induction hypothesis and Theorem 4.1(a). Their product is also separated, by
Lemma 2.2, hence so is Lr . It follows that the sequence (11) is separated at T r on restriction
to H.
But T r is induced for this range of r. Separation of (11) follows immediately from Proposition 2.4 applied to Im(dr1 ) → T r .
This proves that the complex K r (V2n−1 +s , Vs2 ) is separated for all r ≥ 0, and part (a) of
Theorem 4.1 follows.
10. Exterior Powers
In this section we prove part (d) of Theorem 4.1, assuming the whole of the theorem for
smaller n.
Because we have already proved separation, periodicity and splitting we know that
Ω
λt (V2n−1 +s ) =ind λΩ
t2 (Vs )λt (V2n−1 −s );
see the first remark at the end of Section 4. In order to obtain the formula with =proj , we
first consider the restriction to the subgroup H of index 2. Writing V2n−1 +s ↓G
H = V2n−2 +s′ ⊕
V2n−2 +s′′ , the two sides of the formula become
Ω
Ω
Ω
λt (V2n−2 +s′ )λt (V2n−2 +s′′ ) and λΩ
t2 (Vs′ )λt2 (Vs′′ )λt (V2n−2 −s′ )λt (V2n−2 −s′′ ).
′′
Ω
′
But we know, by induction, that λt (V2n−2 +s′ ) =proj λΩ
t2 (Vs )λt (V2n−2 −s′ ) and similarly for s .
Thus, on restriction, the two sides are equal modulo projectives. Now use Lemma 3.3 in
order to see that the two sides are equal modulo projectives even before restriction. This
finally completes the proof of Theorem 4.1.
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
20
11. A Bound on the Number of Non-Induced Summands
The description of the tensor product given in Section 3 shows that the decomposition of
Vr ⊗ Vs into indecomposable summands involves a summand of odd dimension if and only
both r and s are odd, in which case it contains precisely one odd-dimensional summand.
Let us write summ(V ) for the number of indecomposable summands in the module V .
Proposition 11.1. The number of non-induced summands in Λ(V ) is at most
q
1
(dim(V
)+summ(V
))
22
= 2summ(V ) dim(Λ(V )).
Proof. Let f (V ) denote the number of non-induced summands in Λ(V ). The comment about
the tensor product above shows that f (V ⊕ W ) = f (V )f (W ). The proposed bound also
turns sums into products, so it suffices to consider the case when V = Vr is indecomposable
1
and show that f (Vr ) ≤ 2 2 (r+1) .
We use induction on r. Since the cases r = 0, 1 are trivial we can assume that r ≥ 2,
and we can write r = 2n−1 + s, where 1 ≤ s ≤ 2n−1 . Setting t = 1 in the formula
Ω
λt (V2n−1 +s ) =proj λΩ
t2 (Vs )λt (V2n−1 −s ) and using induction we obtain
Ω
f (Λ(V2n−1 +s )) = f (λ1 (V2n−1 +s )) = f (λΩ
1 (Vs ))f (λ1 (V2n−1 −s )) = f (λ1 (Vs ))f (λ1 (V2n−1 −s ))
1
n−1 −s+1)
≤ 2 2 (2
1
1
n−1 +2)
2 2 (s+1) = 2 2 (2
1
n−1 +s+1)
≤ 2 2 (2
.
For an indecomposable kG-module Vr , if we assume that the group acts faithfully then
r > 12 |G|, and the dimension of any direct summand of Λ(Vr ) is at most |G|. It follows that
the dimension of the non-induced part of Λ(Vr ) divided by the dimension of the whole of
1
Λ(Vr ) is at most 2 2 (3−r) r.
12. Remarks
(a) As already mentioned in the introduction, the formula in Theorem 1.1 reduces the computation of Λr (V2n−1 +s ) to the computation of tensor products of exterior powers of modules
of smaller dimension. Since tensor products can easily be determined recursively (see Section 3), this gives an efficient recursive method for calculating the decomposition of exterior
powers of modules for cyclic 2-groups into indecomposables. A program based on this recurrence relation was implemented in GAP [10] by the first author.
A restriction on the use of Theorem 1.1 is the growth of the multiplicities of direct summands of the form V2m . For example, the multiplicity of V128 as a direct summand of Λ57 (V147 )
is 8197519886357582844587268803532720. If one is only interested in the non-induced part
of Λr (V2n−1 +s ) the recurrence relation can be applied modulo induced summands to keep the
multiplicities relatively small.
Together with the results in [21], the recurrence relation in Theorem 1.1 also provides an
algorithm for computing the decomposition of the symmetric powers S r (Vm ) into indecomposables for arbitrary m and r.
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
21
Example. We determine the decomposition of Λ6 (V13 ) into indecomposables:
Λ6 (V13 ) ∼
=proj Ω0+6 (Λ0 (V5 ) ⊗ Λ6 (V3 )) ⊕ Ω1+4 (Λ1 (V5 ) ⊗ Λ4 (V3 )) ⊕
16
16
3
0
2
2
3+0
Ω2+2
16 (Λ (V5 ) ⊗ Λ (V3 )) ⊕ Ω16 (Λ (V5 ) ⊗ Λ (V3 ))
(Λ2 (V5 ) ⊗ V3 ) ⊕ Ω16 (Λ3 (V5 )).
∼
=proj
3
Furthermore Λ (V5 ) ∼
= Λ2 (V5 ), by duality, and
Λ2 (V5 ) ∼
= V3 ⊕ V7 .
= V3 ⊕ Ω8 (V1 ) ∼
= Ω0+2 (Λ0 (V1 ) ⊗ Λ2 (V3 )) ⊕ Ω1+0 (Λ1 (V1 ) ⊗ Λ0 (V3 )) ∼
8
8
Thus
Λ6 (V13 ) ∼
=proj (V3 ⊗ V3 ) ⊕ (V7 ⊗ V3 ) ⊕ Ω16 (V3 ⊕ V7 )
=proj (V3 ⊕ V7 ) ⊗ V3 ⊕ Ω16 (V3 ⊕ V7 ) ∼
∼
=proj (V1 ⊕ 2V4 ) ⊕ (V5 ⊕ 2V8 ) ⊕ (V13 ⊕ V9 ) ∼
=proj V1 ⊕ 2V4 ⊕ V5 ⊕ 2V8 ⊕ V9 ⊕ V13 .
Comparing dimensions, we obtain Λ6 (V13 ) ∼
= V1 ⊕ 2V4 ⊕ V5 ⊕ 2V8 ⊕ V9 ⊕ V13 ⊕ 104V16.
(b) Obviously, Gow and Laffey’s formula for exterior squares [11, Theorem 2] is the special
case r = 2 of Theorem 1.1. Furthermore, setting s = 2n−1 − 1 in Theorem 1.1 gives
Kouwenhoven’s formula [15, Theorem 3.4] for Λr (Vq−1 ) when q is a power of 2 (for all r).
In [15, Theorem 3.5] Kouwenhoven proved the formula
λt (Vq+1 − Vq−1 ) = 1 + (Vq+1 − Vq−1 )t + t2
(12)
in a(Cq·p )[[t]], where q is a power of a prime p; see Section 4 for a definition of λt . We will
show how this can be derived from Theorem 1.1 in the case that p = 2. Note that, since
the dimension series of the two sides match, it is sufficient to prove this modulo projectives.
The theorem gives us:
2n
λt (V2n−1 +1 ) = (1 + V2n −1 t2 )λΩ
(V2n−1 −1 )
t
modulo V2n and
Ω
λt (V2n−1 −1 ) = (1 + V2n−1 −1 t)λt22n−1 (V2n−2 −1 )
modulo V2n−1 . The latter can be written as
Ω
λt (V2n−1 −1 ) = (1 + V2n−1 −1 t)(λt22n−1 (V2n−2 −1 ) + V2n−1 f (t))
exactly (the last term can be written inside the parentheses, since (1+V2n−1 −1 t) is invertible).
Applying Ω2n in odd degrees we obtain
Ω
2n
λΩ
(V2n−1 −1 ) = (1 + V2n−1 +1 t)(λt22n−1 (V2n−2 −1 ) + V2n−1 f (t)).
t
Substituting into the left hand side of (12) yields
(1 + V2n −1 t2 )(1 + V2n−1 +1 t)(1 + V2n−1 −1 t)−1
modulo V2n , and it is easy to verify that
(1 + V2n −1 t2 )(1 + V2n−1 +1 t) = (1 + V2n−1 −1 t)(1 + (V2n−1 +1 − V2n−1 −1 )t + t2 )
modulo V2n .
(c) Theorem 1.1 can also be used to calculate the Adams operations on the Green ring
a(C2n ), as was shown to us by Roger Bryant and Marianne Johnson. For each r > 0 and
j ∈ {1, . . . , 2n }, define an element ψΛr (Vj ) ∈ a(C2n ) by
ψΛ1 (Vj ) − ψΛ2 (Vj )t + ψΛ3 (Vj )t2 − · · · = ψΛ,t (Vj ) =
d
dt
log λt (Vj ).
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
22
By Z-linear extension there is a map ψΛr : a(C2n ) → a(C2n ), called the rth Adams operation
defined by the exterior powers. It can be shown that if r is odd, then ψΛr is the identity map
i
i
i
on a(C2n ) and that ψΛ2 r = ψΛ2 for all i ≥ 1, so all that remains is to describe ψΛ2 for i ≥ 1
(see [5] and [6] for details). For j ≥ 2, write j = 2m + s with m ≥ 0 and 1 ≤ s ≤ 2m ; then
i
i−1
i
ψΛ2 (V2m +s ) = 2ψΛ2 (Vs ) + ψΛ2 (V2m −s )
for all i ≥ 2 and
ψΛ2 (V2m +s ) = 2V2m+1 − 2V2m+1 −s + ψΛ2 (V2m −s ).
This can be seen by applying the definition of the Adams operations to the Hilbert series
form of Theorem 1.1, obtaining (in the obvious notation)
Ω
Ω
ψΛ,t (V2m +s ) =proj 2tψΛ,t
2 (Vs ) + ψΛ,t (V2m −s ).
References
[1] G. Almkvist, R.M. Fossum, Decompositions of exterior and symmetric powers of indecomposable Z/pZmodules in characteristic p and relations to invariants, in Sém. d’Algèbre P. Dubreil, Lecture Notes in
Math. 641, Springer-Verlag, Berlin, Heidelberg, New York, 1976-1977, 1–111.
[2] M. J. J. Barry, Decomposing tensor products and exterior and symmetric squares, J. Group Theory 14
(2011), 59–82.
[3] M. J. J. Barry, Generators for decompositions of tensor products of modules, Arch. Math. 97 (2011),
503–512.
[4] D. J. Benson, Representations and cohomology I and II, Cambridge studies in advanced mathematics
30, 31 (1991), Cambridge Univ. Press, Cambridge.
[5] R. M. Bryant, M. Johnson, Adams operations on the Green ring of a cyclic group of prime-power order,
J. Algebra, 323 (2010), 2818–2833
[6] R. M. Bryant, M. Johnson, Periodicity of Adams operations on the Green ring of a finite group, J. Pure
Appl. Algebra, 215 (2011), 989–1002
[7] L. Chouinard, Projectivity and relative projectivity over group rings, J. Pure and Applied Algebra 7
(1976), 278-302.
[8] C.W. Curtis, I. Reiner, Methods of representation theory I, Wiley, New York, 1987.
[9] J. A. Green, The modular representation algebra of a finite group, Illinois J. Math. 6 (1962), 607–619.
[10] The GAP Group, ‘GAP – Groups, Algorithms, and Programming‘, Version 4.5.4 (2012),
http://www.gap-system.org.
[11] R. Gow, T. Laffey, On the decomposition of the exterior square of an indecomposable module of a cyclic
p-group, J. Group Theory 9 (2006), 659–672.
[12] F. Himstedt, P. Symonds, Equivariant Hilbert series, Algebra & Number Theory 3 (2009), 423–443.
[13] X.-D. Hou, Elementary divisors of tensor products and p-ranks of binomial matrices, Linear Algebra
Appl. 374 (2003), 255–274.
[14] I. Hughes, G. Kemper, Symmetric powers of modular representations, Hilbert series and degree bounds,
Comm. in Algebra, 28 (2000), 2059–2088.
[15] F. M. Kouwenhoven, The λ-structure of the Green rings of cyclic p-groups, Proc. of Symposia in Pure
Math., 47 (1987), 451–466.
[16] C. W. Norman, On the Jordan form of the tensor product over fields of prime characteristic, Linear and
Multilinear Algebra 38 (1995), 351–371.
[17] C. W. Norman, On Jordan bases for the tensor product and Kronecker sum and their elementary divisors
over fields of prime characteristic, Linear Multilinear Algebra 56 (2008), 415-451.
[18] J.-C. Renaud, The decomposition of products in the modular representation ring of a cyclic group of
prime power order, J. Algebra 58 (1979), 1–11.
[19] J.-C. Renaud, Recurrence relations in a modular representation algebra, Bull. Austral. Math. Soc. 26
(1982), 215–219.
EXTERIOR AND SYMMETRIC POWERS OF MODULES FOR CYCLIC 2-GROUPS
23
[20] B. Srinivasan, The modular representation ring of a cyclic p-group, Proc. London Math. Soc. (3) 14
(1964), 677–688.
[21] P. Symonds, Cyclic group actions on polynomial rings, Bulletin London Math. Soc. 39 (2007), 181–188.
Technische Universität München, Zentrum Mathematik - M11, Boltzmannstr. 3, 85748
Garching, Germany
E-mail address: [email protected]
School of Mathematics, University of Manchester, Manchester M13 9PL, United Kingdom
E-mail address: [email protected]
| 0math.AC
|
Identification of functionally related enzymes by learning-to-rank
methods
arXiv:1405.4394v1 [cs.LG] 17 May 2014
Michiel Stock, Thomas Fober, Eyke Hüllermeier, Serghei Glinca, Gerhard Klebe,
Tapio Pahikkala, Antti Airola, Bernard De Baets, Willem Waegeman ∗
Abstract
Enzyme sequences and structures are routinely used in the biological sciences as queries to search for functionally related enzymes in online databases. To this end, one usually departs from some notion of similarity,
comparing two enzymes by looking for correspondences in their sequences, structures or surfaces. For a given
query, the search operation results in a ranking of the enzymes in the database, from very similar to dissimilar
enzymes, while information about the biological function of annotated database enzymes is ignored.
In this work we show that rankings of that kind can be substantially improved by applying kernel-based
learning algorithms. This approach enables the detection of statistical dependencies between similarities of the
active cleft and the biological function of annotated enzymes. This is in contrast to search-based approaches,
which do not take annotated training data into account. Similarity measures based on the active cleft are known
to outperform sequence-based or structure-based measures under certain conditions. We consider the Enzyme
Commission (EC) classification hierarchy for obtaining annotated enzymes during the training phase. The
results of a set of sizeable experiments indicate a consistent and significant improvement for a set of similarity
measures that exploit information about small cavities in the surface of enzymes.
1
Introduction
Modern high-throughput technologies in molecular biology are generating more and more protein sequences and tertiary structures, of which only a small fraction can ever be experimentally annotated w.r.t. functionality. Predicting
the biological function of enzymes1 remains extremely challenging, and especially novel functions are hard to detect,
despite the large number of automated annotation methods that have been introduced in the last decade [1, 2].
Existing online services, such as BLAST or ReliBase, often provide tools to search in databases that contain
collections of annotated enzymes. These systems rely on some notion of similarity when searching for related
enzymes, but the definition of similarity differs from system to system. Indeed, a vast number of measures for
expressing similarity between two enzymes exists in the literature, performing calculations on different levels of
abstraction. One can make a major subdivision of these measures into approaches that solely use the sequence of
amino acids, approaches that also take into account the tertiary structure, and approaches that consider local fold
information by analyzing small cavities (hypothetical binding sites) at the surface of an enzyme.
Sequence-based measures such as BLAST [3] and PSI-BLAST [4] can be computed in an efficient manner and
are able to find enzymes with related functions under certain conditions. In addition to these, several kernelbased methods have been developed to make predictions for proteins at the sequence level - see e.g. [5, 6]. A high
sequence similarity usually results in a high structural similarity, and proteins with a sequence identity (the number
of matches in an alignment) above 40 % are generally considered to share the same structure [7]. However, this
assumption becomes less reliable in the twilight zone, when the sequence identity is situated between 25 and 40 %.
Furthermore, enzymes with comparable functions can exhibit sequences with very low sequence identity [8, 9].
For these reasons, and because three-dimensional crystal structures are becoming more and more available in
online databases, the comparison of proteins at the structural level has gained increasing attention. The secondary
∗ M. Stock, B. De Baets and W. Waegeman are with the Department of Mathematical Modelling, Statistics and Bioinformatics,
Ghent University, Coupure links 653, 9000 Ghent, Belgium, email: [email protected]. T. Fober and E. Hüllermeier are with
the Philipps-Universität of Marburg, Department of Mathematics and Computer Science, Hans-Meerwein-Straße 6, D-35032 Marburg,
Germany. S. Glinca and G. Klebe are also with the Philipps-Universität of Marburg, Department of Pharmacy, Marbacher Weg 6-10,
D-35032 Marburg, Germany. T. Pahikkala and A. Airola are with the Department of Information Technology and the Turku Centre
for Computer Science, University of Turku, Joukahaisenkatu 3-5 B 20520 Turku, Finland.
1 Enzymes are biomolecules that catalyze chemical reactions. All the enzymes we consider in this work are proteins and vice versa.
Both notions will be used interchangeably.
1
structure of an enzyme is known to highly influence its biological function [10] and contains valuable information
that is missing at the sequence level [11, 12, 13]. Many approaches that perform calculations on the overall fold
of the protein have been developed - see e.g. [14, 15, 16]. Unfortunately, such approaches are also not optimal
for determining the function of enzymes. They require knowledge of active site residues and usually lead to a
quite coarse representation, especially for enzymes, where often only a few specific residues are responsible for the
catalytic mechanism [17]. For example, the vicinal-oxygen-chelate superfamily shows a large functional diversity
while having only a limited sequence diversity [18, 19]. It has also been shown that some parts of the protein
structure space have a high functional diversity [20], further limiting the use of global fold similarity. For these
reasons, many methods consider local structural features, such as evolutionary conserved residues [21, 22].
The most appropriate similarity measures for the prediction of enzyme functions focus on surface regions in
which ligands, co-factors and substrates can bind [23]. Cavities in the surface are known to contain valuable
information, and exploiting similarities between those cavities helps finding functionally related enzymes. By
considering structural and physico-chemical information of binding sites, one can detect relationships that cannot
be found using traditional sequence- and fold-based methods, making such similarities of particular interest for
applications in drug discovery [24, 25]. In addition to providing a complementary notion of protein families [26],
these methods also allow for extracting relationships between cavities of unrelated proteins [27]. Similarity measures
that highlight cavities and binding sites can be further subdivided into graph-based approaches such as [28, 29, 30,
31, 32], geometric approaches such as [33, 34] and feature-based approaches such as [35, 36]. These measures will
be discussed more thoroughly in Section 2.
This paper aims to show that the search for functionally related enzymes can be substantially improved by
applying learning-to-rank methods. These algorithms use training data to build a mathematical model for ranking
objects, such as enzymes, that are not necessarily seen among the training data. While our methods can be applied
to all types of data, as long as a meaningful similarity measure can be constructed, we only demonstrate its power
using cavity-based measures, for the reasons explained above. Ranking-based machine learning algorithms are often
used for applications in information retrieval [37]. Due to their proven added value for search engines, ranking-based
machine learning methods have gained some popularity in bioinformatics, for example in drug discovery [38, 39]
or to find similarities between proteins [40, 41]. Despite this, many online services such as BLAST, PDB, Dali
and CavBase solely rely on similarity measures to construct rankings, without utilizing annotated enzymes and
learning algorithms to steer the search process during a training phase. However, due to the presence of annotated
enzymes in online databases, improvements can be made by applying ranking-based machine learning algorithms.
This amounts to a transition from an unsupervised to a supervised learning scenario.
Using four different cavity-based similarity measures and one based on sequence alignment as input for RankRLS [42],
a kernel-based ranking algorithm, we demonstrate a significant improvement for each of these measures. RankRLS
works in a similar way as competitors such as RankSVM [43], because it uses annotated training data to learn
rankings during a training phase. The training data is annotated via the Enzyme Commission (EC) functional
classification hierarchy, a commonly used way to subdivide enzymes into functional classes. EC numbers adopt a
four-number hierarchical structure, representing different levels of catalytic detail. Importantly, this representation focuses on the chemical reactions that are performed, and not on structure or homology. As explained more
elaborately in Section 2, the EC numbers are used to construct a ground-truth catalytic similarity measure, and
subsequently to generate ground-truth rankings. In addition to obtaining annotated training data, this procedure
also allows for a fair comparison with the more traditional approach, using conventional performance measures for
rankings. This way of evaluating also characterizes the difference between our search engine approach and previous
work in which supervised learning algorithms for EC number assignment have been considered – for a far from
complete list see e.g. [44, 29, 45, 46, 47]. In this work we are unable to compare to such methods, because they
do not return rankings as output. Nonetheless, similar to some of these approaches, we do take the hierarchical
structure of the EC numbers into account. Instead of predicting one EC number, a ranking of functionally related
enzymes is returned for a given query. In this scheme the top of the obtained ranking is expected to contain
enzymes with functions similar to the query enzyme with an unknown EC number. A ranking provides end users
with a generally well-known and easily understandable output, while still useful results can be retrieved when an
enzyme with a new EC number is encountered.
2
2.1
Material and methods
Database
Our work builds upon CavBase, a database that is made commercially available as part of ReliBase [48]. CavBase
can be used for the automated detection, extraction, and storage of protein cavities from experimentally determined
2
protein structures, which are available through the Protein Data Bank (PDB).
The geometrical arrangement of the pocket and its physico-chemical properties are first represented by predefined pseudocenters – spatial points that characterize the geometric center of a functional group specified by a
particular property. The type and the spatial position of the pseudocenters depend on the amino acids that border
the binding pocket and expose their functional groups. They are derived from the protein structure using a set
of predefined rules [49]. Hydrogen-bond donor, acceptor, mixed donor/acceptor, hydrophobic aliphatic, metal ion,
pi (accounts for the ability to form π-π interactions) and aromatic properties are considered as possible types of
pseudocenters. These pseudocenters can be regarded as a compressed representation of surface areas where certain
protein-ligand interactions are encountered. Consequently, a set of pseudocenters is an approximate representation
of a spatial distribution of physico-chemical properties.
To build and test our models we require an appropriate data set that contains sufficiently many proteins and
EC classes. Based on the experience of local pharmaceutical experts, we chose the data set of EC classes depicted
in Table 1. To generate the first data set (data set I), we retrieved all proteins from the PDB which got assigned
one of these EC classes. Thus, we ended up with a set of 5,257 proteins. To ensure that only unique proteins
were contained in our data set, we used the protein culling server2 with its default parameterization. As such, all
proteins that have high pairwise homology were filtered out. This procedure resulted in a data set of cardinality
1,714. To extract the active site of the protein we used the assumption that the largest binding site of a protein
does contain its catalytic center [23]. Hence, for each protein we took the binding site from the database CavBase
which maximized the volume. From our data set, 158 proteins were not contained in the CavBase (e.g., because
the structure was determined by NMR instead of X-ray). Therefore these proteins were removed from the data
set, resulting in a final data set of size 1,556.
The first data set comes with two drawbacks. First of all, the binding site containing the catalytic centre was
determined by a pure heuristic, namely by taking the largest binding site among all binding sites a protein exhibits.
Moreover, sufficient resolution was not a criterium for selecting the cavities. This may lead to a data set of low
quality. Therefore, relying on the expertise of pharmaceutical experts we compiled another data set referred to as
data set II, containing the same EC classes. For this data set, all proteins from the PDB that have a resolution of
at least 2.5 Å were considered. Moreover the binding site volume was required to range between 350 Å3 and 3500
Å3 . Structures not meeting these conditions were eliminated since resolutions below 2.5 Å usually lead to a too
coarse representation, while binding sites with volumes outside the above-mentioned range are usually artefacts
produced by the algorithm used for their detection. From the resulting set of 24,102 proteins the active site was
selected. This resulted in a data set of 1730 enzymes on which we applied the protein culling server to finally end
up with a second data set of 561 enzymes.
A pairwise sequence similarity matrix and phylogenetic tree of our data sets can be found in the supplementary
materials.
2.2
Similarity measures for cavities
In the introduction we have motivated why our analysis is restricted to similarity measures for cavities, which are
three-dimensional objects that can be represented in multiple ways. Some measures are graph-based, transforming
cavities into node-labeled and edge-weighted graphs. This allows to apply traditional techniques to compare graphs
(e.g. [28]). Unfortunately, techniques that construct a boolean similarity measure, such as those based on graph
isomorphisms, are not appropriate for comparing noisy and flexible protein structures. Computing the maximum
common subgraph [50] can be considered as a more appropriate alternative, and this method will be used in this
paper as a baseline (see below). The graph edit distance [51] is another measure to compare graphs, specifying the
number of edit operations needed to transform a given graph into another graph. This distance can be calculated
in different ways, e.g., by using a greedy heuristic [52] or quadratic programming [53]. Unfortunately, the graph
edit distance is very hard to parameterize and often quite inefficient. More efficient approaches belong to the
class of graph kernels. They have gained a lot of attention in bioinformatics, as they allow for a sufficiently high
degree of error tolerance. Different realizations are available, such as the shortest path kernel [54], the random
walk kernel [30] and the graphlet kernel [31, 32]. Graph kernels work particularly well for small molecules such
as ligands, but they are less useful for larger molecules such as proteins. They gave rather poor results in [36],
which explains why we concentrated here on the maximum common subgraph as a representative for graph-based
approaches.
As a second category of measures for cavities, geometric methods directly process the labeled spatial coordinates
of the functional parts, denoted as point clouds, instead of transforming a protein cavity into a graph. Remarkably,
only few approaches have been proposed that build on this representation. In [33] geometric hashing is employed to
2 http://www.bioinf.manchester.ac.uk/leaf/
3
Table 1: List of the 21 EC numbers with their accepted name and the number of examples of each class for the
two data sets.
EC
EC
EC
EC
EC
EC
EC
EC
EC
EC
EC
EC
EC
EC
EC
EC
EC
EC
EC
EC
EC
EC
number
1.1.1.1
1.1.1.21
1.5.1.3
1.11.1.5
1.14.15.1
2.1.1.45
2.1.1.98
2.4.1.1
2.4.2.29
2.7.11.1
3.1.1.7
3.1.3.48
3.4.21.4
3.4.21.5
3.5.2.6
4.1.2.13
4.2.1.1
4.2.1.20
5.3.1.5
5.3.3.1
6.3.2.1
accepted name
alcohol dehydrogenase
aldehyde reductase
dihydrofolate reductase
cytochrome-c peroxidase
camphor 5-monooxygenase
thymidylate synthase
diphthine synthase
phosphorylase
tRNA-guanine transglycosylase
non-specific serine/threonine enzyme kinase
acetylcholinesterase
enzyme-tyrosine-phosphatase
trypsin
thrombin
β-lactamase
fructose-bisphosphate aldolase
carbonate dehydratase
tryptophan synthase
xylose isomerase
steroid ∆-isomerase
pantoate-β-alanine ligase
# set I
23
35
110
92
30
63
5
43
32
304
23
151
118
87
153
48
186
13
18
14
8
# set II
15
30
6
31
36
22
43
40
16
24
13
28
72
51
8
4
76
7
21
10
8
calculate a superposition of protein cavities that can be used to derive an alignment and a similarity score. A similar
approach was used in [34], in which an optimization problem was solved instead of applying geometric hashing.
Beside these two approaches, several other methods exist for comparing two point clouds [55]. Unfortunately, the
majority of these methods cannot cope with biological data, due to a very high complexity or error intolerance.
As a third family of approaches, one can also represent the protein cavity as a feature vector, taking both the
geometry of the cavity and physico-chemical properties into account – see e.g. [36, 35]. Subsequently, traditional
or specialized measures can be applied on these vectors to obtain similarity scores between protein cavities [56, 57].
In the experiments we selected representative method for each of the three groups: one graph-based measure,
one geometric measure and one feature-based measure. We also considered the original CavBase measure and a
measure obtained from the Smith-Waterman protein sequence alignment. This lead to a comparison of five different
measures, four based on cavities and one based on sequence alignment. Below, these measures are explained more
in detail:
Labeled Point Cloud Superposition (LPCS) [34]. This value is obtained by processing labeled point clouds.
Hence, the CavBase data can be used directly without a need for transforming it into another representation.
Intuitively, two labeled point clouds are considered similar if they can be spatially superimposed. More
specifically, an approximate superposition of the two structures is obtained by fixing the first point cloud and
moving the second point cloud as a whole. Two point clouds are well superimposed when each point in the
first cloud can be matched with a point in the second point cloud, while the distances of these points are
small and their labels consistent. This concept is used to define a fitness function that is maximized using a
direct search approach [58]. The obtained maximal fitness is taken as the similarity between the two labeled
point clouds. A similar measure was also proposed in [59], but a convolution kernel is suggested to obtain
similarities between the point clouds.
Maximum Common Subgraph (MCS) [50]. Using the MCS, the original representation in the form of a
labeled point cloud must be transformed into a node-labeled and edge-weighted graph. Each pseudocenter
is becoming a node labeled with the corresponding physico-chemical property. To capture the geometry,
a complete graph is considered, where each edge is weighted with the Euclidean distance between the two
pseudocenters it is adjacent to. The problem of measuring similarity between protein cavities now boils down
to the problem of measuring similarity between graphs. A well-known approach here is to search for the
maximum common subgraph of the two input graphs and to define similarity as the size of the maximum
common subgraph relative to the size of the larger graph. In case of noisy data, a threshold is required,
4
defining two edges as equal if their weight differs at most by . In this paper, this parameter is set to 0.2 Å,
as recommended by several authors [52, 60].
CavBase (CB) similarity [49]. CavBase also makes use of an algorithm for the detection of common subgraphs.
Instead of considering the largest common subgraph, as done in the case of MCS, the 100 largest common
subgraphs are considered. Each common subgraph is used to determine a transformation rule by means of
the Kabsch algorithm [61], which superimposes both proteins. In a post-processing step the surface points are
also superimposed according to the transformation rule, and a similarity score is derived using these surface
points. Eventually, a set of 100 similarity values is obtained, from which the highest value is returned as
similarity between the two protein cavities.
Fingerprints (FP) Fingerprints are a well-known concept and have been used successfully in many domains.
For the comparison of protein binding sites, the authors in [62] transformed the protein binding site into a
node-labeled and edge-weighted graph as described above. Moreover they defined generically a set of features,
namely complete node-labeled and edge-weighted graphs of size 3. For each such feature, a test is performed
to decide whether or not the feature is contained in the graph representing the protein. This is done by
subgraph isomorphism, to checks whether the labels are identical. The nodes of the features are labeled by
the set of physiochemical properties. Edges of patterns are labeled by intervals or bins and instead of testing
for equivalence, a test is performed whether edge weight of the graph representing the protein falls into the
bin of the pattern. The thus generated fingerprints are compared by means of the Jaccard similarity measure,
as proposed by [56].
Smith-Waterman (SW) Beside using structure-based approaches to compare protein binding sites, we used
also sequence alignment in our experimental study. To calculate sequence alignments we used the SmithWaterman algorithm [63] which was parameterized with the Blosum-62 matrix. From the sequence alignment
we derived the sequence identity which was subsequently used to perform experiments.
2.3
Unsupervised ranking
In the introduction we have explained why existing online services such as BLAST, PDB, Dali and CavBase
construct rankings in an unsupervised way. These systems create a ranking by means of a similarity measure only,
without training a model that uses annotated enzymes. The annotated enzymes in a database are simply ranked
according to their similarity with an enzyme query with unknown function. In the case of CavBase, the enzymes
having a high cavity-based similarity appear on top of the ranking, those exhibiting a low cavity-based similarity
end up at the bottom. More formally, let us represent the similarity between a pair of enzymes by K : V 2 → R+ ,
where V represents the set of all potential enzymes. Given the similarities K(v, v 0 ) and K(v, v 00 ), we compose the
ranking of v 0 and v 00 conditioned on the query v as:
00
0
00
v 0 K
v v ⇔ K(v, v ) ≥ K(v, v ) ,
K
v
(1)
indicates the relation is ranked higher than, for query v, based on the similarity K. Note that this is
where
a relation between two enzymes conditioned on a third enzyme. In our context, there is no meaningful ranking
possible between enzymes v 0 and v 00 without referring to another enzyme v. This approach adopts the same
methodology as a nearest neighbor classifier, but a ranking rather than a class label should be seen as the output
of the algorithm.
The quality of such rankings can be evaluated when the database contains annotated enzymes and annotated
queries. In an evaluation phase, we compare the obtained ranking with the ground truth ranking, which can be
constructed from the EC numbers for annotated enzymes. This ground truth ranking can be deduced from the
catalytic similarity (i.e., ground truth similarity) between the query and all database enzymes, by counting the
number of successive matches in the EC label of the query and the database enzymes. Thus the catalytic similarity
is a property of only a pair of enzymes. In contrast, in order to create the ground truth ranking of two enzymes,
the catalytic similarity has to be calculated w.r.t. a third enzyme. For example, an enzyme with EC number
EC 2.4.2.23 has a catalytic similarity of two compared to an enzyme labeled as EC 2.4.99.12, since both enzymes
belong to the family of glycosyltransferases. Conversely, the same enzyme manifests a similarity value of only
one with an enzyme labeled as EC 2.8.2.23. Both are transferases in this case, but they show no further relevant
similarity in the chemistry of the reactions to be catalyzed.
More formally, let us represent the catalytic similarity between two enzymes by a relation Q : V 2 → {0, 1, 2, 3, 4}.
Q is defined by:
4 Y
i
X
Q(v, v 0 ) =
qi qj ,
i=1 j=1
5
0
EC 2.7.7.34
3
EC ?.?.?.?
2
0
1
EC 4.2.3.90
0
0
0
EC 4.6.1.11
2
EC 2.7.7.12
EC 2.7.1.12
Figure 1: Six enzyme structures are shown, five of which correspond to a known EC number. The catalytic similarity
Q is depicted on the edges of the graph. The algorithm that we present allows us to infer for the unannotated
query (denoted as EC ?.?.?.?) a ranking of the annotated enzymes. To this end, the unsupervised approach solely
uses cavity-based similarity measures, whereas the supervised approach also takes the EC numbers of annotated
enzymes into account.
zondag, 13 mei 2012
where qi equals 1 if the ith digit of the EC numbers of v and v 0 are the same and 0 otherwise. Figure 1 gives an
example for six enzymes, five of which correspond to a known EC number. The catalytic similarity Q is depicted
on the edges of the graph. The proposed algorithm allows us to infer for an unannotated query a ranking of the
annotated enzymes, some of which the algorithm may not have encountered among the training data.
Given the similarities Q(v, v 0 ) and Q(v, v 00 ), we compose similar to Eq. (1) the ground truth ranking of v 0 and
00
v conditioned on the query v as:
00
0
00
v 0 Q
v v ⇔ Q(v, v ) ≥ Q(v, v ) .
As a result, an entire ground truth ranking of database enzymes with known EC numbers can be constructed, given
an annotated query enzyme.
2.4
Supervised ranking
In contrast to unsupervised ranking approaches, supervised algorithms do take ground truth information into
account during a training phase. We perform experiments with so-called conditional ranking algorithms [64, 65]
using the RankRLS implementation [42]. Let us introduce the short-hand notation e = (v, v 0 ) to denote a couple
consisting of an enzyme query v and a database enzyme v 0 . RankRLS produces a linear basis function model of
the type:
h(e) = h(v, v 0 ) = hw, Φ(e)i ,
(2)
in which w denotes a vector of parameters and Φ(e) an implicit feature representation for the couple e = (v, v 0 ).
RankRLS differs from more conventional kernel-based methods because it optimizes a convex and differentiable
approximation of the rank loss in bipartite ranking (i.e., area under the ROC curve) instead of the zero-one loss.
Together with the standard L2 regularization term on the parameter vector w and a regularization parameter λ,
the following loss is minimized:
X X
L(h, T ) =
(Qe − Qē − h(e) + h(ē))2 ,
(3)
v∈V e,ē∈Ev
for a given training set T = {(e, Qe ) | e ∈ E}. Here Qe = Q(v, v 0 ) denotes the ground truth similarity as defined
above, E the set of training query-object couples for which ground truth information is available and Ev the subset
of E containing results for the query v. The outer sum in Eq. (3) takes all queries into account, and the inner
sum analyzes all pairwise differences between the ranked results for a given query. This loss can be minimized in
a computationally efficient manner, using analytic shortcuts and gradient-based methods, as shown in [64, 65].
According to the representer theorem [66], one can rewrite Eq. (2) in the following dual form:
X
h(e) = hw, Φ(e)i =
aē K Φ (e, ē) .
ē∈E
with K Φ (e, ē) a kernel function with four enzymes as input and aē the weights in the dual space. In this paper we
adopt the Kronecker product feature mapping, containing information on couples of enzymes:
Φ(e) = Φ(v, v 0 ) = φ(v) ⊗ φ(v 0 ) ,
6
with φ(v) a feature mapping of an individual enzyme and ⊗ the Kronecker product. One can easily show that this
pairwise feature mapping yields the Kronecker product pairwise kernel in the dual representation:
K Φ (e, ē) = K Φ (v, v 0 , v̄, v̄ 0 ) = K φ (v, v̄)K φ (v 0 , v̄ 0 ) ,
with K φ a traditional kernel for enzymes. Specifying a universal kernel for K φ leads to a universal kernel for
K Φ [67], indicating that one can use the kernel to represent any arbitrary relation, provided that the learning
algorithm has access to training data of sufficient quality. This kernel has been introduced in [68] for modelling
protein-protein interactions. We consider this kernel because of its universal approximation property, but also other
pairwise kernels exist, such as the cartesian pairwise kernel [69], the metric learning pairwise kernel [70] and the
transitive pairwise kernel [71, 64]. Nonetheless, it is probably not very surprising that such kernels only yield an
improvement if the concepts to be learned satisfy the restrictions that are imposed by the kernels [67].
With the exception of the FP measure, none of the similarity measures discussed in Section 2.2 are strictly
speaking valid kernels. Using the above construction, all similarity measures can be converted into kernels of type
K φ , when they are made symmetric and positive definite. These attributes guarantee a numerically stable and
unique solution of the learning algorithm. We simply enforced symmetry by averaging the similarity matrix with
its transpose. Subsequently, we made the different similarity matrices positive definite by performing an eigenvalue
decomposition and setting all eigenvalues smaller than 10−10 equal to zero. This method leads to a negligible loss
of information compared to the numerical accuracy of our algorithms and data storage. Finally, each kernel matrix
was normalized so that all diagonal elements have a value equal to one. Since these procedures were performed on
the whole data set, one arrives at a so-called transductive learning setting [72]. Minor adjustments would obtain a
more traditional inductive learning setting. Note that overfitting is prevented when applying this procedure, since
the EC numbers of the enzymes in the data set are not taken into account.
Since the catalytic similarity is a symmetric measure we also perform a post-processing to the output of our
algorithm. The matrix with the predicted values used for ranking the enzymes is made symmetric by averaging it
with its transpose.
2.5
Performance measures for ranking
The ranking obtained with unsupervised or supervised learning algorithms can be compared to the ground truth
ranking by applying performance measures that are commonly used in information retrieval.
First of all, the ranking accuracy (RA) is considered, and it is defined as follows:
RA =
1
1 X
|V | v∈V |{Ev | ye > yē }|
X
I(h(e) − h(ē)) ,
e,ē∈Ev :
ye <yē
with I(x) the Heaviside function, returning one when its argument is positive, zero when its argument is negative
and 1/2 when its argument is zero. The ranking accuracy can be considered as a generalization of the area under
the ROC curve for more than two ordered classes [73].
Our interest in the ranking accuracy is motivated by two reasons. Firstly, unlike most other performance
measures we consider, all levels of the EC hierarchy are taken into account to determine the performance of
different algorithms. Predicted rankings can be interpreted as layered or multipartite rankings – see e.g. citeWaegeman2008e,Furnkranz2009. The ranking accuracy preserves this hierarchical structure by counting all pairwise comparisons. The second reason of interest is based on the fact that the ranking accuracy is optimized by
the RankRLS software, using the convex and differentiable approximation given in Eq. (3). This loss function
characterizes the most important difference with more traditional kernel-based algorithms, such as support vector machines, resulting in an information retrieval setting instead of a more traditional classification or network
inference setting.
Since the ranking accuracy is not generally known in bioinformatics, we also evaluated our algorithms using
three more conventional performance measures that are commonly considered for bipartite rankings (i.e., rankings
containing relevant versus irrelevant objects). These three measures are the area under the ROC curve (AUC),
mean average precision (MAP) and normalized discounted cumulative gain (nDCG) [74]. For AUC and MAP
all ground truth rankings had to be converted into bipartite rankings, leading to a decrease in granularity for
performance estimation. We chose a cut-off threshold of three in ground truth similarity: a retrieved enzyme is
relevant to the enzyme query if at least the first three parts of its EC number are identical to the query.
2.6
Experimental setup
We selected two data sets of enzymes from CavBase, as described in Section 2.1. The ground-truth catalytic
similarity of all enzyme pairs was computed for each data set. Each data set was further randomized and split in
7
Table 2: Summary of the results obtained for unsupervised and supervised ranking for both data sets. For each
combination of similarity and type of performance measure, the performance is averaged over the different folds
and queries, with the standard deviation between parentheses.
Unsupervised
Supervised
Unsupervised
Supervised
RA
MAP
AUC
nDCG
RA
MAP
AUC
nDCG
CB
0.6096
0.4883
0.7110
0.7161
0.7717
0.5659
0.8242
0.6550
(0.1286)
(0.2912)
(0.1845)
(0.3107)
(0.1960)
(0.3384)
(0.1857)
(0.4195)
FP
0.6411
0.4478
0.7221
0.6615
0.7175
0.5966
0.8003
0.6928
RA
MAP
AUC
nDCG
RA
MAP
AUC
nDCG
CB
0.7216
0.9156
0.9415
0.9599
0.9997
0.9995
0.9999
0.9983
(0.1911)
(0.1662)
(0.1209)
(0.0708)
(0.0011)
(0.0037)
(0.0005)
(0.0185)
FP
0.7212 (0.1546)
0.7156 (0.2478)
0.8714 (0.1447)
0.8189 (0.2003)
0.931 (0.1096)
0.905 (0.2397)
0.9844 (0.0502)
0.8887 (0.2243)
(0.1670)
(0.2692)
(0.1961)
(0.2709)
(0.2098)
(0.3219)
(0.2225)
(0.3984)
Set I
MCS
0.5960
0.4967
0.6967
0.7686
0.7988
0.6854
0.8585
0.7717
Set II
MCS
0.8070
0.9094
0.9622
0.9330
0.9944
0.9980
0.9997
0.9744
(0.1010)
(0.2659)
(0.1728)
(0.3001)
(0.1906)
(0.3001)
(0.1802)
(0.3635)
LPCS
0.6173
0.4826
0.7183
0.7364
0.7741
0.6324
0.8306
0.7324
(0.1244)
(0.2572)
(0.1565)
(0.276)
(0.1945)
(0.3064)
(0.1954)
(0.3724)
SW
0.6576
0.5201
0.7397
0.7401
0.7899
0.6963
0.8813
0.7518
(0.1446)
(0.2964)
(0.1980)
(0.2597)
(0.1789)
(0.3103)
(0.1591)
(0.3669)
(0.1736)
(0.1514)
(0.0847)
(0.0948)
(0.0253)
(0.0265)
(0.0036)
(0.0773)
LPCS
0.7515
0.8303
0.8937
0.8986
0.9883
0.9663
0.9926
0.9467
(0.1647)
(0.1768)
(0.1165)
(0.1146)
(0.0583)
(0.1172)
(0.0354)
(0.1183)
SW
0.7856
0.9167
0.9523
0.9589
0.9985
0.9980
0.9998
0.9968
(0.1591)
(0.1634)
(0.1146)
(0.0768)
(0.0076)
(0.0393)
(0.0042)
(0.0344)
four cross-validation folds of equal size. In the unsupervised case each subset was used individually to allow for a
comparison with the supervised model. Of such a subset, each enzyme was used as a query to rank the remaining
enzymes, as described in Section 2.3. The performance for each of these rankings was averaged to obtain the global
performance over the folds.
In the supervised setting, each fold was withheld as a test set, while the other three parts of the data set were
used for training and model selection. This process was repeated for each part, so that every instance was used for
training and testing (thus, four-fold outer cross-validation). Neither the query nor the database enzymes are thus
used for building a model, which allows us to demonstrate that our methods can generalize to new enzymes. In
addition, a 10-fold inner cross-validation loop was implemented for estimating the optimal regularization parameter
λ, as recommended in [75]. The value of this hyperparameter, which controls model complexity, was selected from
a grid containing all powers of 10 from 10−4 to 105 . The final model was trained using the whole training set and
the median of the best hyperparameter values over the ten folds. We used the implementation RLScore3 in Python
to train the models.
3
Results and discussion
3.1
Differences between cavity-based similarities and data sets
Table 2 gives a global summary of the results obtained for the unsupervised and the supervised ranking approach
for both data sets. One can note a sizeable difference between the performances for the different cavity-based
similarities, data sets and the performance measures used. Despite this variation, it is clear that data set I is
considerably harder than data set II. This can easily be explained by the fact that data set II only contains
enzymes with a certain resolution of the active site. Furthermore, for set II the active site is determined by an
expert, while for set I the active site is resolved by heuristically choosing the largest cavity. It is likely that some
mistakes are made in this annotation process. Consequently, inferring functional similarity of data set I will be
harder.
The cavity-based similarity measure based on fingerprints usually results in the worst performance, except
for the ranking error in the unsupervised setting. It seems that the performance of the FP does not improve as
much in the supervised approach, compared to other cavity-based similarity measures. This is likely because the
fingerprints cause a high loss of information, since even functionally dissimilar enzyme cavities can be considered
similar according to this metric.
Comparing the two graph-based similarities (MCS and CB), we see some differences between the data sets.
Though both perform relatively well, MCS performs better for data set I, while CB is the clear champion of data
3 See
http://staff.cs.utu.fi/~aatapa/software/RLScore/ for this software.
8
set II. The good performance of CavBase for data set II can be explained easily. CavBase computes the 100
largest common subgraphs, which could be used to construct a cavity-based similarity measure. However, a graph
representation leads to a loss of information, since the coordinates of pseudocenters cannot be restored. Moreover,
since the size of the (maximum) common subgraph is an integer that usually lies in the range of 4 to 12 nodes,
there is a loss of resolution by mapping many different pairs of cavities to the same similarity score. In theory,
MCS suffers from these drawbacks. Even though the resolution problem is to a certain extent solved if the size of
the maximum common subgraph is divided by the size of the larger binding site, the graph representation could
still lead to a slight loss of information.
On the other hand, the LPCS measure uses geometric information, hence, no loss of information is introduced
by transforming the pseudocenter representation into a graph representation. Moreover, this transformation does
not cause the resolution problem. Yet, the measure is computed via solving a multimodal optimization problem, so
it is possible to get stuck in a local optimum, resulting in a similarity score that is too low. Similar to MCS, LPCS
seems to perform relatively better for data set I compared to data set II, probably because the local optimum
becomes less of an issue in the former case. This can be explained by the fact that data set I contains larger
cavities, on average, hence making it harder to find the global optimum.
Finally, we consider the measure based on sequence alignment. For data set I, the SW similarity measure
competes with the MCS as one of the best measures, depending on the performance measure. For data set II in the
supervised case, it is only outperformed by the CB measure. It is clear that the SW measure is a powerful method
for comparing cavities, as it is also limited by a bad resolution of the cleft.. Like MCS, SW seeks to quantify the
largest similar region, here as a local alignment. As this contains, information about the common residues of the
cavity, this is a simple, though powerful measure.
3.2
The benefits of supervised ranking
The ranking of both data set I and data set II showed a considerable improvement in the supervised approach.
Three important reasons can be put forward to explain the improvement in performance. In this section we will
illustrate this using data set II, as this one showed the most clear effects of learning. First of all, the traditional
benefit of supervised learning plays an important role. One can expect that supervised ranking methods outperform unsupervised methods, because they take annotations into account during the training phase to guide the
model towards retrieval of enzymes with a similar EC number. Conversely, unsupervised methods solely rely on a
meaningful similarity measure between enzymes, while ignoring EC numbers.
Second, we also advocate that supervised ranking methods have the ability to preserve the hierarchical structure
of EC numbers in their predicted rankings. Figure 2 supports this claim. It summarizes the values used for ranking
one fold of the test set obtained by the different models. A higher value (indicated by a lighter color) in a row means
that this enzyme is considered to have a higher catalytic similarity w.r.t. the query enzyme. So, for unsupervised
ranking it visualizes K φ (v, v 0 ), for supervised ranking the values h(v, v 0 ) are shown. Each row of the heatmap
corresponds to one query. For the supervised models one notices a much better correspondence with the ground
truth. Furthermore, the different levels of catalytic similarity can be better distinguished. In addition, an example
of the distributions of the predicted values within one query is visualized in Figure 3 by means of box plots. The
different populations within a plot correspond to the different levels of the catalytic similarity w.r.t. the query
enzyme. This illustrates again that supervised models can make a better discrimination between enzymes that are
functionally more similar and those that are dissimilar. For this example query, no quartiles are overlapping in any
supervised model, unlike the unsupervised approach, which only detects a good ranking for exact matches (i.e.,
enzymes having an EC number identical to the query).
A third reason for improvement by the supervised ranking method can be found in the exploitation of dependencies between different catalytic similarity values. Roughly speaking, if one is interested in the catalytic
similarity between enzymes v and v 0 , one can try to compute this catalytic similarity in a direct way based on
mutual relationships in cavities, or derive it in an indirect way from the cavity-based similarity with a third enzyme
v 00 . This division into a direct and an indirect approach shows a certain correspondence to similar discussions in
the context of inferring protein-protein interaction and signal transduction networks – see e.g. [76, 70, 77]. Unsupervised ranking boils in a certain sense down to a direct approach, while supervised ranking should be interpreted
as indirect. Especially when the similarity matrix contains noisy values, one can expect that the indirect approach
allows for detecting back bone entries and correcting the noisy ones.
3.3
Differences between performance measures
Table 2 indicates that the different performances are to some degree influenced by the similarity measure and data
set used. This is especially clear for the supervised ranking approach. One can observe a very clear distinction
9
(b) Ground truth
(a) Unsupervised and supervised measures used for ranking
Figure 2: (a) Heatmaps of the values used for ranking the data set II for one fold in the testing phase. Each row
of the heatmap corresponds to one query. The four figures on top visualize the cavity-based similarities K φ (v, v 0 )
that are used to construct an unsupervised ranking. The four figures at the bottom visualize the model output
h(v, v 0 ), which is used to derive a supervised ranking. (b) The ground truth catalytic similarity that had to be
learned.
between the ranking accuracy and area under the ROC curve, which treat every position as equally important,
and the other two measures, which emphasize the top of the ranking. This should not come as a surprise, as
an approximation of the ranking error is optimized in our algorithms. The AUC is very related to the RA, as
they coincide for bipartite rankings. In the latter we only make a distinction between ’relevant’ enzymes, which
have three or more EC numbers in common, and enzymes which do not. Since the RA uses a finer fragmentation
of functional similarity, this is a more severe performance measure compared to the AUC. For data set II, both
AUC and RA are both very close to the theoretical optimum for nearly all cavity-based similarity measures in the
supervised case.
Figure 4 shows the ROC curves that are obtained by applying the cut-off threshold for data set I, which defines
a database enzyme a hit if at least the first three digits of the EC number are correct. In contrast to the scalar
performance measures of Table 2, the ROC curve gives information about the quality of the ranking at all positions.
It is immediately clear that supervised ranking outcompetes unsupervised ranking, as the former’s curves are closer
to the upper left corner. Typically for these curves, there is a part in the beginning where the line has a very high
slope, showing that a certain fraction of relevant objects can be detected with very high sensitivity and specificity.
This fraction that can be detected nearly without mistakes increases after the supervised learning step, as indicated
by the higher ‘offset’ of these curves (most clear for SW, LPCS and FP in the ROC curve). The next section of
the curve is usually a part with a smaller average slope, indicating that at some point it becomes harder to detect
signal. For the unsupervised curves, this is nearly a straight line, which means that from that point the detection
of catalytically similar enzymes is essentially random. The supervised curves still have a concave shape at their
second part, which shows that relevant enzymes can still be detected in that piece.
From Table 2 it is also clear that supervised ranking usually scores worse for MAP and nDCG compared to
RA and AUC. For the nDCG the performance sometimes even decreases after learning a model! This can easily
be explained by the fact that our model optimizes the quality of the complete ranking, in contrast to only the top
which is assessed by MAP and nDGC. Note that the top functionally similar enzymes (i.e. the same EC number)
can likely be detected based on the cavity-based similarity alone. Hence, training a model might not be required
to perform good in this section. One can see this learning effect nicely in Figure 4 for the CB similarity measure.
The quality of the supervised ranking of the top for this measure is worse than for all other measures (indicated
by the low nDCG). The overall ranking (indicated by the AUC) is quite good in comparison, as the lower part
compensates for the bad ranking at the top. Depending on the application, the top or the general ranking might
be more of interest.
4
Related work
Since the comparison of enzymes has become an important task in functional bioinformatics, a vast number of
similarity measures for proteins has been proposed so far. As mentioned in the introduction, a reliable method
10
Unsupervised
FP
Unsupervised
LPCS
Unsupervised
MCS
Unsupervised
SW
1.0
1.0
1.0
1.0
1.0
●
3
0
3
0.6
prediction
0.2
●
0
0.0
0.0
0.0
●
0
●
0.2
0.2
0.2
0.2
0.4
0.4
0.8
0.8
●
0.6
prediction
0.6
prediction
0.4
0.6
prediction
0.8
0.8
●
0.4
0.6
0.4
prediction
0.8
Unsupervised
CB
3
0
3
0
3
cat.
similarity
cat.
similarity
cat.
similarity
cat.
similarity
cat.
similarity
Supervised
CB
Supervised
FP
Supervised
LPCS
Supervised
MCS
Supervised
SW
3
●
0
2
0
1
prediction
2
0
0
0
●
1
prediction
2
prediction
●
1
1
prediction
2
1
prediction
2
3
3
3
3
4
●
●
●
3
cat.
similarity
0
3
cat.
similarity
0
−1
●
0
−1
●
3
cat.
similarity
0
3
cat.
similarity
0
3
cat.
similarity
Figure 3: Box-and-whisker plots of the values used for ranking data set II for one randomly chosen query as an
example. The different populations on the x-axis denote the groups that are formed by subdividing the database
enzymes according to the number of EC number digits they share with the query. Given a query v and a database
enzyme v 0 , the y-axis shows the distribution of the values h(v, v 0 ) and K φ (v, v 0 ) for the supervised and the unsupervised approach, respectively. For nearly all cavity-based similarity measures, one can observe a much better
separation of the groups for the supervised approach.
will focus on the geometry and the physico-chemical properties of certain regions of an enzyme. However, methods
that are based on the sequence or the fold usually exhibit a much lower complexity and they can also lead to good
results, especially when the sequence identity is above a certain threshold. ProFunc [78] is in this regard a very
interesting tool, in which a bulk of different methods is applied, such as sequence alignment, motif and template
search, and also a comparison of active sites. The biological function of enzymes is derived from the closest match
in different databases such as PDB, UniProt and PROCAT, and finally returned by the program. Despite being
very powerful, this approach becomes nevertheless very inefficient, with runtimes up to several hours fro a single
protein. Since we considered a sizeable data set for which nearly 3,000,000 pairwise similarity scores had to be
computed, it became impossible to compare our results to ProFunc.
In addition to focusing on individual enzymes, one can also take protein-protein interactions into account for
inferring the function. Proteins that are close to each other in a protein-protein interaction network are expected to
have similar functions, so one can try to infer the function of an unanotated protein by looking at its neighbors [79].
Similarly, one can also solve optimization problems over the global network, such as maximizing the number of edges
that connect proteins sharing the same function [80, 81]. Other approaches make use of probabilistic graphical
models such as Markov random fields [82, 83]. Conceptually, these methods might also enrich the predictions
obtained by an unsupervised search-based approach, but they usually do not consider cavity and binding site
information to predict the function of proteins.
5
Conclusion
In this paper we have recast the EC annotation problem as a conditional ranking problem. We have shown that
retrieval of enzymes w.r.t. functionality can be substantially improved by applying a supervised ranking method
that takes advantage of ground truth EC numbers during the training phase. In contrast, more traditional methods
rely heavily on a notion of similarity to search for functionally related enzymes in online databases. Such methods
lead to an unsupervised approach in which annotations are not taken into account.
We focused specifically on cavity-based similarity measures, because their benefits compared to sequence-based
and structure-based approaches have been demonstrated in previous work, although our method can work with
any meaningful similarity measure defined on enzymes. In the experiments we could demonstrate a considerable
improvement of the quality of the overall ranking. The results were influenced by the type of data used and the way
11
0.6
0.4
CB sup.
FP sup.
LPCS sup.
MCS sup.
SW sup.
CB unsup.
FP unsup.
LPCS unsup.
MCS unsup.
SW unsup.
0.0
0.2
Average true positive rate
0.8
1.0
ROC curve for the different enzyme similarity
measurements of data set I
0.0
0.2
0.4
0.6
0.8
1.0
False positive rate
Figure 4: Receiver operating characteristic curves for the unsupervised and supervised ranking methods for data
set I. An enzyme is considered functionally similar to the query if the first three digits of the EC number are
identical to those of the query.
the ranking was evaluated, indicating that the most optimal method is highly dependent of the specific problem
setting. Nevertheless, our supervised ranking algorithm outperformed the unsupervised ranking algorithm for all
cavity-based similarities for most performance measures considered. While the unsupervised approach succeeded
quite well in returning exact matches to a query, the hierarchical structure of EC numbers was better preserved in
the rankings predicted by the supervised approach. As such, supervised ranking can be interpreted as a powerful
alternative for retrieval methods that are more traditionally used in bioinformatics.
Acknowledgements
MS, BDB and WW acknowledge the support of Ghent University (MRP Bioinformatics: from nucleotides to
networks). T.F., E.H., S.G. and G.K. gratefully acknowledge financial support by the German Research Foundation
(DFG) and the LOEWE Research Center for Synthetic Microbiology, Marburg. T.P. and A.A. are both supported
for this work by the Academy of Finland (grant 134020 and 128061, respectively).
References
[1] I. Friedberg, “Automated protein function prediction–the genomic challenge,” Briefings in Bioinformatics,
vol. 7, pp. 225–42, Sept. 2006.
[2] D. Dunaway-Mariano, “Enzyme function discovery,” Structure, vol. 16, no. 11, pp. 1599–1600, 2008.
[3] S. F. Altschul, W. Gish, W. Miller, E. W. Myers, and D. J. Lipman, “Basic Local Alignment Search Tool,”
Journal of Molecular Biology, vol. 215, no. 3, pp. 403–410, 1990.
[4] S. F. Altschul, T. L. Madden, A. A. Schäffer, J. Zhang, Z. Zhang, W. Miller, and D. J. Lipman, “Gapped
Blast and PsiBlast: a new generation of protein database search programs,” Nucleic Acid Research, vol. 25,
no. 17, pp. 3389–3402, 1997.
[5] C. Leslie, E. Eskin, and W. S. Noble, “The spectrum kernel: a string kernel for SVM protein classification.,”
in Pacific Symposium on Biocomputing, pp. 566–575, Jan. 2002.
[6] C. S. Leslie, E. Eskin, A. Cohen, J. Weston, and W. S. Noble, “Mismatch string kernels for discriminative
protein classification,” Bioinformatics, vol. 20, pp. 467–76, Mar. 2004.
12
[7] R. Powers, J. C. Copeland, K. Germer, K. A. Mercier, V. Ramanathan, and P. Revesz, “Comparison of protein
active site structures for functional annotation of proteins and drug design,” Proteins: Structure, Function,
and Bioinformatics, vol. 65, pp. 124–135, 2006.
[8] B. Rost, “Enzyme function less conserved than anticipated,” Journal of Molecular Biology, vol. 318, no. 2,
pp. 595–608, 2002.
[9] A. J. Chalk, C. L. Worth, J. P. Overington, and A. W. E. Chan, “PDBLIG: Classification of small molecular
protein binding in the protein data bank,” Journal of Medical Chemistry, vol. 47, no. 15, pp. 3807–3816, 2004.
[10] K. Kinoshita, Y. Murakami, and H. Nakamura, “eF-seek: prediction of the functional sites of proteins by
searching for similar electrostatic potential and molecular surface shape,” Nucleic Acid Research, vol. 35,
no. suppl 2, pp. W398–W402, 2007.
[11] J. Thornton, A. Todd, D. Milburn, N. Borkakoti, and C. Orengo, “From structure to function: Approaches
and limitations,” Nature Structural Biology, vol. 7 Suppl, pp. 991–994, 2000.
[12] M. Kotera, Y. Okuno, M. Hattori, S. Goto, and M. Kanehisa, “Computational assignment of the EC numbers for genomic-scale analysis of enzymatic reactions,” Journal of the American Chemical Society, vol. 126,
pp. 16487–98, Dec. 2004.
[13] V. Egelhofer, I. Schomburg, and D. Schomburg, “Automatic assignment of EC numbers,” PLoS Computational
Biology, vol. 6, pp. 1000661+, 2010.
[14] M. Shatsky, R. Niussinov, and H. J. Wolfson, “A method for simultaneous alignment of multiple protein
structures,” Proteins: Structure, Function, and Bioinformatics, vol. 56, pp. 143–156, 2004.
[15] M. Shatsky, The Common Point Set Problem with Applications to Protein Structure Analysis. PhD thesis, Tel
Aviv University, Tel Aviv, Israel, 2006.
[16] A. Harrison, F. Pearl, I. Sillitoe, T. Slidel, R. Mott, J. Thornton, and C. Orengo, “Recognizing the fold of a
protein structure,” Bioinformatics, vol. 19, no. 14, pp. 1748–1759, 2003.
[17] A. C. Martin, C. A. Orengo, E. G. Hutchinson, S. Jones, M. Karmirantzou, R. A. Laskowski, J. B. Mitchell,
C. Taroni, and J. M. Thornton, “Protein folds and functions,” Structure, vol. 6, pp. 875–84, July 1998.
[18] P. C. Babbitt and J. Gerlt, “Understanding enzyme superfamilies,” Journal of Biological Chemistry, vol. 272,
pp. 30591–30594, Dec. 1997.
[19] J. A. Gerlt and P. C. Babbitt, “Divergent evolution of enzymatic function: mechanistically diverse superfamilies
and functionally,” Annual review of biochemistry, vol. 70, pp. 209–246, 2001.
[20] M. Osadchy and R. Kolodny, “Maps of protein structure space reveal a fundamental relationship between
protein structure and function,” Proceedings of the National Academy of Sciences of the United States of
America, vol. 108, no. 30, pp. 12301–12306, 2011.
[21] D. M. Kristensen, R. M. Ward, A. M. Lisewski, S. Erdin, B. Y. Chen, V. Y. Fofanov, M. Kimmel, L. E.
Kavraki, and O. Lichtarge, “Prediction of enzyme function based on 3D templates of evolutionarily important
amino acids,” BMC Bioinformatics, vol. 9, Jan. 2008.
[22] S. Erdin, A. M. Lisewski, and O. Lichtarge, “Protein function prediction: towards integration of similarity
metrics,” Current Opinion in Structural Biology, vol. 21, pp. 180–8, Apr. 2011.
[23] R. A. Laskowski, N. M. Luscombe, M. B. Swindells, and J. M. Thornton, “Protein clefts in molecular recognition and function,” Protein Science, vol. 5, no. 12, pp. 2438–2452, 1996.
[24] S. Pérot, O. Sperandio, M. A. Miteva, A.-C. Camproux, and B. O. Villoutreix, “Druggable pockets and binding
site centric chemical space: a paradigm shift in drug discovery,” Drug Discovery Today, vol. 15, no. 15-16,
pp. 656–667, 2010.
[25] C. Andersson, B. Y. Chen, and A. Linusson, “Mapping of ligand-binding cavities in proteins,” Proteins, vol. 78,
no. 6, pp. 1408–1422, 2010.
13
[26] M. Weisel, J. M. Kriegl, and G. Schneider, “Architectural repertoire of ligand-binding pockets on protein
surfaces,” Chembiochem, vol. 11, no. 4, pp. 556–563, 2010.
[27] A. Weber, A. Casini, A. Heine, D. Kuhn, C. T. Supuran, A. Scozzafava, and G. Klebe, “Unexpected nanomolar
inhibition of carbonic anhydrase by COX-2-selective celecoxib: new pharmacological opportunities due to
related binding site recognition,” Journal of Medical Chemistry, vol. 47, no. 3, pp. 550–557, 2004.
[28] J. Huan, D. Bandyopadhyay, W. Wang, J. Snoeyink, J. Prins, and A. Tropsha, “Comparing graph representations of protein structure for mining family-specific residue-based packing motifs,” Journal of computational
biology : a journal of computational molecular cell biology, vol. 12, no. 6, pp. 657–71, 1999.
[29] K. M. Borgwardt, C. S. Ong, S. Schönauer, S. V. N. Vishwanathan, A. J. Smola, and H.-P. Kriegel, “Protein
function prediction via graph kernels,” Bioinformatics, vol. 21 Suppl 1, pp. i47–56, June 2005.
[30] T. Gärtner, Kernels for Structured Data. Singapore: World Scientific, 2008.
[31] N. Shervashidze, “Efficient graphlet kernels for large graph comparison,” in Proceedings of the 12th International Conference on Artificial Intelligence and Statistics (AISTATS), vol. 5, pp. 488–495, 2009.
[32] V. Vacic, L. I. ans S. Lonardi, and P. Radivojac, “Graphlet kernels for prediction of functional residues in
protein structures,” Journal of Computational Biology, vol. 17, no. 1, pp. 55–72, 2010.
[33] M. Shatsky, A. Shulman-Peleg, R. Nussinov, and H. J. Wolfson, “The multiple common point set problem
and its application to molecule binding pattern detection,” Journal of Computational Biology, vol. 13, no. 2,
pp. 407–428, 2006.
[34] T. Fober, S. Glinca, G. Klebe, and E. Hüllermeier, “Superposition and alignment of labeled point clouds,”
IEEE/ACM Transactions on Computational Biology and Bioinformatics, vol. 8, pp. 1653–1666, November/December 2011.
[35] N. Weill and D. Rognan, “Alignment-free ultra-high-throughput comparison of druggable protein-ligand binding sites,” Journal of Chemical Information and Modeling, vol. 50, no. 1, pp. 123–135, 2010.
[36] T. Fober, M. Mernberger, G. Klebe, and E. Hüllermeier, “Fingerprint kernels for protein structure comparison,”
Molecular Informatics, vol. 31, no. 6-7, pp. 443–452, 2012.
[37] E. Hüllermeier and J. Fürnkranz, Preference Learning. Springer, 2010.
[38] F. Rathke, K. Hansen, U. Brefeld, and K.-R. Müller, “StructRank: a new approach for ligand-based virtual
screening,” Journal of Chemical Information and Modeling, vol. 51, pp. 83–92, Dec. 2010.
[39] S. Agarwal, D. Dugar, and S. Sengupta, “Ranking chemical structures for drug discovery: a new machine
learning approach.,” Journal of Chemical Information and Modeling, vol. 50, pp. 716–31, May 2010.
[40] J. Weston, A. Eliseeff, D. Zhou, C. Leslie, and W. S. Noble, “Protein ranking: from local to global structure
in the protein similarity network,” Proceedings of the National Academy of Science, vol. 101, pp. 6559–6563,
2004.
[41] R. Kuang, J. Weston, W. S. Noble, and C. Leslie, “Motif-based protein ranking by network propagation,”
Bioinformatics, vol. 21, pp. 3711–8, Oct. 2005.
[42] T. Pahikkala, E. Tsivtsivadze, A. Airola, J. Järvinen, and J. Boberg, “An efficient algorithm for learning to
rank from preference graphs,” Machine Learning, vol. 75, no. 1, pp. 129–165, 2009.
[43] T. Joachims, “A support vector method for multivariate performance measures,” in Proceedings of the International Conference on Machine Learning, Bonn, Germany, pp. 377–384, 2005.
[44] P. D. Dobson and A. J. Doig, “Predicting enzyme class from protein structure without alignments.,” Journal
of Molecular Biology, vol. 345, pp. 187–99, Jan. 2005.
[45] J. Rousu, C. Saunders, S. Szedmak, and J. Shawe-Taylor, “Kernel-based learning of hierarchical multilabel
classification models,” Journal of Machine Learning Research, vol. 7, pp. 1601–1626, 2006.
[46] A. Sokolov and A. Ben-Hur, “A structured-outputs method for prediction of protein function,” in Proceedings
of the 3rd International Workshop on Machine Learning in Systems Biology, 2008.
14
[47] A. K. Arakaki, Y. Huang, and J. Skolnick, “EFICAz2: enzyme function inference by a combined approach
enhanced by machine learning,” BMC Bioinformatics, vol. 10, p. 107, Jan. 2009.
[48] M. Hendlich, A. Bergner, J. Günther, and G. Klebe, “Relibase: design and development of a database for
comprehensive analysis of proteinligand interactions,” Journal of Molecular Biology, vol. 326, pp. 607–620,
Feb. 2003.
[49] S. Schmitt, D. Kuhn, and G. Klebe, “A new method to detect related function among proteins independent
of sequence and fold homology,” Journal of Molecular Biology, vol. 323, no. 2, pp. 387–406, 2002.
[50] H. Bunke and K. Shearer, “A graph distance metric based on the maximal common subgraph,” Pattern
Recognition Letters, vol. 19, no. 3-4, pp. 255–259, 1998.
[51] A. Sanfeliu and K. Fu, “A distance measure between attributed relational graphs for pattern recognition,”
IEEE Transactions on Systems, Man and Cybernetics, vol. 13, no. 3, pp. 353–362, 1983.
[52] N. Weskamp, E. Hüllermeier, D. Kuhn, and G. Klebe, “Multiple graph alignment for the structural analysis of
protein active sites.,” IEEE/ACM Transactions on Computational Biology and Bioinformatics, vol. 4, no. 2,
pp. 310–20, 2007.
[53] M. Neuhaus and H. Bunke, Bridging the Gap between Graph Edit Distance and Kernel Machines. New Jersey:
World Scientific, 2007.
[54] K. Borgwardt, C. Ong, S. Schonauer, S. Vishwanathan, A. Smola, and H. Kriegel, “Protein function prediction
via graph kernels,” Bioinformatics, vol. 21, no. 1, pp. i47–i56, 2005.
[55] H. Alt and L. J. Guibas, “Discrete geometric shapes: Matching, interpolation, and approximation: A survey,”
tech. rep., Handbook of Computational Geometry, 1996.
[56] P. Mahé, N. Ueda, T. Akutsu, J.-L. Perret, and J.-P. Vert, “Graph kernels for molecular structure-activity
relationship analysis with support vector machines,” Journal of chemical information and modeling, vol. 45,
no. 4, pp. 939–51, 2005.
[57] M. M. Deza and E. Deza, Encyclopedia of Distances. Heidelberg, Germany: Springer, 2009.
[58] H.-G. Beyer and H.-P. Schwefel, “Evolution strategies: A comprehensive introduction,” Natural Computing,
vol. 1, no. 1, pp. 3–52, 2002.
[59] B. Hoffmann, M. Zaslavskiy, J.-P. Vert, and V. Stoven, “A new protein binding pocket similarity measure
based on comparison of clouds of atoms in 3D: application to ligand prediction,” BMC Bioinformatics, vol. 11,
Jan. 2010.
[60] T. Fober, M. Mernberger, G. Klebe, and E. Hüllermeier, “Evolutionary construction of multiple graph alignments for the structural analysis of biomolecules.,” Bioinformatics, vol. 25, pp. 2110–7, Aug. 2009.
[61] W. Kabsch, “A solution of the best rotation to relate two sets of vectors,” Acta Crystallographica, vol. 32,
pp. 922–923, 1976.
[62] T. Fober, M. Mernberger, R. Moritz, and E. Hüllermeier, “Graph-kernels for the comparative analysis of
protein active sites,” in German Conference on Bioinformatics, (Halle (Saale), Germany), pp. 21 – 31, 2009.
[63] T. F. Smith and M. S. Waterman, “Identification of common molecular subsequences,” Journal of Molecular
Biology, vol. 147, pp. 195–197, 1981.
[64] T. Pahikkala, W. Waegeman, E. Tsivtsivadze, T. Salakoski, and B. De Baets, “Learning intransitive reciprocal
relations with kernel methods,” European Journal of Operational Research, vol. 206, pp. 676–685, Nov. 2010.
[65] T. Pahikkala, A. Airola, M. Stock, B. De Baets, and W. Waegeman, “Efficient regularized least-squares
algorithms for conditional ranking on relational data,” Machine Learning, vol. 93, no. 2-3, pp. 321–356, 2013.
[66] B. Schölkopf and A. Smola, Learning with Kernels, Support Vector Machines, Regularisation, Optimization
and Beyond. The MIT Press, 2002.
[67] W. Waegeman, T. Pahikkala, A. Airola, T. Salakoski, M. Stock, and B. De Baets, “A kernel-based framework
for learning graded relations from data,” IEEE Transactions on Fuzzy Systems, vol. 20, pp. 1090–1101, 2012.
15
[68] A. Ben-Hur and W. S. Noble, “Kernel methods for predicting protein-protein interactions,” Bioinformatics,
vol. 21, pp. i38–46, June 2005.
[69] H. Kashima, S. Oyama, Y. Yamanishi, and K. Tsuda, “On pairwise kernels: An efficient alternative and
generalization analysis.,” in PAKDD (T. Theeramunkong, B. Kijsirikul, N. Cercone, and T. B. Ho, eds.),
vol. 5476 of Lecture Notes in Computer Science, pp. 1030–1037, Springer, 2009.
[70] J.-P. Vert, J. Qiu, and W. S. Noble, “A new pairwise kernel for biological network inference with support
vector machines.,” BMC Bioinformatics, vol. 8, Jan. 2007.
[71] R. Herbrich, T. Graepel, and K. Obermayer, “Large margin rank boundaries for ordinal regression,” in Advances in Large Margin Classifiers (A. Smola, P. Bartlett, B. Schölkopf, and D. Schuurmans, eds.), pp. 115–132,
MIT Press, 2000.
[72] O. Chapelle, B. Schölkopf, and A. Zien, Semi-Supervised Learning. MIT Press, 2006.
[73] W. Waegeman, B. De Baets, and L. Boullart, “ROC analysis in ordinal regression learning,” Pattern Recognition Letters, vol. 29, pp. 1–9, 2008.
[74] K. Järvelin and J. Kekäläinen, “Cumulated gain-based evaluation of IR techniques,” ACM Transactions on
Information Systems (TOIS), vol. 20, no. 4, pp. 422–446, 2002.
[75] S. Varma and R. Simon, “Bias in error estimation when using cross-validation for model selection,” BMC
Bioinformatics, vol. 7, Jan. 2006.
[76] R. Albert, B. DasGupta, R. Dondi, S. Kachalo, E. Sontag, A. Zelikovsky, and K. Westbrooks vol. 14, pp. 927–
949, 2007.
[77] P. Geurts, N. Touleimat, M. Dutreix, and F. D’Alché-Buc, “Inferring biological networks with output kernel
trees,” BMC Bioinformatics, vol. 8, no. 2, p. S4, 2007.
[78] R. Laskowski, J. Watson, and J. Thornton, “Profunc: a server for predicting protein function from 3d stucture,”
NAR, vol. 33, pp. W89–W93, 2005.
[79] B. Schwikowski, P. Uetz, and S. Fields, “A network of protein-protein interactions in yeast,” Nature Biotechnology, vol. 18, pp. 3548–3557, 2000.
[80] A. Vazquez, A. Flammini, A. Maritan, and A. Vespignani, “Global protein function prediction from proteinprotein interaction networks,” Bioinformatics, vol. 21, pp. ii59–ii65, 2003.
[81] U. Karaoz, T. Murali, S. Letovsky, Y. Zheng, C. Ding, C. Cantor, and S. Kasif, “Whole-genome annotation by
using evidence integration in functional-linkage networks,” Proceedings of the National Academy of Sciences,
vol. 101, pp. 2888–2893, 2004.
[82] M. Deng, K. Zhang, S. Mehta, T. Chen, and F. Sun, “Prediction of protein function using protein-protein
interaction data,” Journal of Computational Biology, vol. 10, pp. 947–960, 2003.
[83] S. Letovsky and S. Kasif, “Predicting protein function from protein/protein interaction data: a probabilistic
approach,” Bioinformatics, vol. 19, pp. i197–i204, 2003.
16
| 5cs.CE
|
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
1
A Generalized-Impedance Based Stability
Criterion for Three-Phase Grid-Connected
Voltage Source Converters
Huanhai Xin, Ziheng Li, Wei Dong, Zhen Wang, and Leiqi Zhang
Abstract—The output impedance matrices of three-phase
grid-connected voltage source converters (VSCs) are widely used
in power system stability analysis. Regardless of how the
impedance is modeled, there always exist coupling terms in the
impedance matrix, which makes the system a multi-inputmulti-output (MIMO) system. As a practical approximation, the
coupling terms are generally omitted and the impedance-based
stability criterion for the resultant single-input-single-output
(SISO) system can be applicable. However, such handling may
result in analytical errors. This letter proposes a new stability
criterion based on an equivalent SISO system by introducing the
concept of generalized- impedances, which can completely keep
the coupling terms. Further, the effects of the phase-locked-loop
(PLL) parameters on system stability are studied based on the
proposed criterion. The effectiveness of the proposed criterion is
verified by a hardware-in-the-loop (HIL) simulation based on a
RT-LAB platform.
Index Terms—Grid-connected converters, impedance modeling,
small-signal stability, sub-synchronous oscillation.
I. INTRODUCTION
RID-CONNECTED voltage source converters (VSCs) are
widely used for renewable energy integration and
high-voltage DC transmission systems [1]. However, the
interaction between the power electronic controllers and the
transmission lines can lead to power oscillations, which poses
new challenges for power system stability [2-5].
When analyzing the oscillation problems introduced by
power electronic devices, the impedance-based methods are
widely used, in which a grid-connected VSC is modeled as an
output impedance connected to a voltage/current source and the
grid is represented similarly [2-7]. In a three-phase system, the
impedance of the VSC or the grid can be mathematically
written as a matrix, which relates the voltage vector to the
current vector. Currently, most impedance-based methods for
grid-tied VSCs can be classified into two categories according
to the reference frames where the impedance matrices are
formulated: the synchronous frame (dq frame) based methods
G
This work is jointly supported by the National Key Research and
Development Program (No. 2016YFB0900104), the National Science
Foundation of China (No. 51577168) and Science & Technology Project of
Yunnan Electric Power Company (yndw (2016) 000303DD00124).
Huanhai Xin, Ziheng Li, Wei Dong, Zhen Wang, and Leiqi Zhang are with
the College of Electrical Engineering, Zhejiang University, Hangzhou 310027,
China. Email: [email protected].
[2-3] and the stationary frame (abc frame) based methods [4-5].
In the dq frame, there usually exist coupling terms
(non-diagonal elements) in the impedance matrices of the
VSCs and the grid, which makes the system a MIMO system.
Usually, the generalized Nyquist stability criterion (GNSC) is
used for stability performance analysis in these methods [2-3].
Similarly, in the abc frame, the coupling terms also exist
between the positive and the negative sequence impedances in
the VSC system and the grid. Although the two sequence
impedances of a symmetrical grid can be strictly decoupled (the
impedance matrix is diagonal), the impedance matrix of the
VSC usually has coupling terms [5]. As a practical
approximation, the coupling terms are generally omitted and
the impedance-based stability criterion for the resultant SISO
system can be applicable [4, 6]. Unfortunately, ignoring the
coupling terms may lead to inaccurate analysis results in a
mirror frequency coupled system [5, 7].
This letter presents a generalized-impedance based stability
criterion (GISC) for three-phase grid-connected VSCs. By
mathematically manipulating the characteristic equation, the
system can be transformed into an equivalent SISO system,
which can be regarded as a generalized-impedance based series
circuit. In essence, the small-signal instability of the system can
be explained as a resonance of the equivalent circuit and
evaluated by the classical Nyquist stability criterion.
II. GRID-CONNECTED VSC MODEL
The grid-connected VSC considered in this letter is shown in
Fig. 1. The converter uses a conventional LCL filter. The
grid-side network includes C f and LLine , where C f is the filter
capacitor, LLine denotes the total inductance of the filter and the
transmission line. The converter is controlled by a dual-loop
vector controller based on a PLL. The current controller of the
VSC has voltage feed-forward (VFF) and decoupling terms.
The positions of the phasors and the reference frames of the
system are shown in Fig. 2, where the dq frame represents the
converter-side rotating frame introduced by the PLL and the xy
frame represents the grid-side synchronous rotating frame [2].
In steady state, the dq frame is aligned with the xy frame.
This paper focuses on the oscillations induced by the PLL
and the inner current loop (similar to [5]). The following
assumptions are considered to simplify the problem: 1) The
time delay of the sampling circuits and the PWM are neglected
due to their tiny time scales [8]. 2) The dynamics of the filter in
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
the VFF are neglected, since the oscillation frequency under
study is much lower than the cut-off frequency of the filter.
VSC Side
U sdq
Vdc
U
node 1
Lf
PWM
I abc
Cf
I abc
U abc
abc/dq
abc/dq
I dq
U sdqref
Grid Side
IG
I
LLine
U abc
abc/dq
Uq
U dq
I dqref
Current
Control
PLL
PLL
Fig. 1. Block diagram of the grid-connected converter
q y
I
I
U
U
d
pll
x
0
xy : grid-side rotating frame
(global frame)
dq : inverter-side rotating frame
(local frame)
0:synchronous speed
: speed obtained by PLL
Fig. 2. Reference frames and phasor positions.
It should be emphasized that there is no constraint on the
power factor (PF) of the VSC, i.e., the VSC can be the interface
of a STATCOM (PF=0) or a renewable energy (usually PF 1).
A. Dynamic Model of the VSC in Polar Coordinate
For simplicity of analysis, resistances of the filter inductor
and the transmission line are neglected. In the dq frame, the
dynamic model of the VSC, including the PLL, filter inductor
and the current controller, can be represented by:
pll G pll ( s )U q
U sdref ( I dref I d )Gi ( s) U d 0 L f I q
(1)
U sqref ( I qref I q )Gi ( s) U q 0 L f I d ,
U =U L sI L I
d
f
d
f q
sd
U sq =U q L f sI q L f I d
where Gpll (s) and Gi (s) are the transfer functions of the PLL
and the current controller, respectively. The meanings of other
symbols are shown in Fig. 1 and Fig. 2.
To obtain the small-signal model, the following steps can be
followed: 1) Linearizing (1) and solving the linearized
equations, the dynamic model in the dq frame can be obtained;
2) The model is transformed to the xy frame using angle
relationships shown in Fig. 2; 3) The model in the xy frame is
written in the form of the polar coordinate as follows:
0 U
I 0
ΔI
0 Y ( s) U ,
(2)
I
g
where Yg (s) Gi (s)Gpll (s) I 0 / [(Gi (s) L f s)(1 Gpll (s)U 0 )] and the
subscript “0” denotes steady-state values. The detailed
derivation is shown in Appendix A.
B. Dynamic Model of the Grid in Polar Coordinate
According to the topology shown in Fig. 1, the model of the
network can be formulated as follows:
I G
U
ΔIG
ΔI Y
(3)
I G G
U
where Y = YLine + YC . The expressions of YLine and YC are
shown in Appendix B. In particular, YC is 0 if the output filter
only has an inductor.
III. GENERALIZED-IMPEDANCES OF THE SYSTEM
A. Definition of the Generalized-impedance
It follows from (2) and (3) that the VSC model and the grid
model have the common structure shown as follows:
I A( s) C ( s) U
I C (s) D( s) U .
Grid
Controller
2
(4)
Based on (4), the generalized-admittances ( Ygi (s), i 1,2,3 )
are defined as follows:
Yg1 (s) : D(s) A(s), Yg 2 (s) : A(s) jC(s), Yg 3 (s) : A(s) jC(s). (5)
The generalized-impedance is defined as the inverse of the
generalized-admittance (i.e., Z gi (s) Ygi1 (s) ).
B. Generalized-impedances of the VSC and the Grid
It follows from (2) and (5) that the generalized-admittance of
the VSC can be calculated as follows:
Yg1_ VSC (s) Yg (s), Yg 2 _ VSC (s) 0, Yg 3 _ VSC (s) 0.
(6)
Note that matrix Y in (3) has the same structure as matrix
0
a b
a b -1 a jb
, it can be
b a . Since T b a T 0
a
jb
obtained that
TYT-1 : diag(Y+ ,Y ) diag(Z1,Z1 ),
where T
(7)
1 1 j
, j 1 . As mentioned in Section I,
2 1 j
matrix Y exists coupling terms and thus the system becomes a
MIMO system.
From the definition in (5), the generalized-impedance of the
grid can be obtained as follows:
Yg1_ grid (s) 0, Yg 2 _ grid (s) Y , Yg 3 _ grid (s) Y .
(8)
-1
Note that TYT implies the admittance is changed from the
synchronous frame to the stationary frame [7], so Y (or Y )
corresponds to the positive (or negative) sequence admittance.
IV. GENERALIZED-IMPEDANCE BASED STABILITY ANALYSIS
In this section, we derive the characteristic equations of the
system, and analyze its small-signal stability.
It follows from (2) and (3) that the characteristic equation
can be expressed as
0 0
det
Y 0,
0
Y
g
(9)
where det() is the determinant function.
Since T is invertible, (9) is equivalent to
0 0
det T
Y T-1 0,
0 Yg
(10)
which is further equivalent to
Y 1 1 Y+ 0
det g
(11)
0.
2 1 1 0 Y
When Yg (s) 0 , (11) implies det(Y (s)Y (s)) 0 . This is a
special case indicating the VSC and the network have the same
resonance point, which is not the focus of this paper. Excluding
this case, (11) is equivalent to
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
1 0 Z Z
1 0 Yg I
= det
2Z g I
0
1 1 Z Z
1 1 2
(12)
It follows from (12) that the characteristic equation of the
closed-loop system can be simplified as
1 Yg (Z Z ) 2 0.
(13)
From the definitions of the generalized-impedance, we can
rearrange (13) as 2Z g1_ VSC Z g 2 _ grid Z g 3 _ grid 0 , which can be
explained as the series circuit of the generalized-impedances of
the VSC and the grid. Therefore, the oscillation of the
grid-connected VSC can be regarded as the series resonance of
the equivalent circuit. The system described by (13) is a SISO
system, thus the complexity of analysis is reduced distinctly.
It is worth noting that the VSC and the grid are usually
assumed to be stable respectively when analyzing the
interaction between the VSC and the grid [6]. We also assume
the VSC is stable when connected to an ideal grid, so Yg ( s) and
(Z Z ) 2 have no pole in the right half plane. Consequently,
it follows from the Nyquist criterion that
GISC: If the Nyquist curve of Yg (s)(Z Z ) 2 does not
encircle the point (1, j0) , then the system is stable.
Note that, although the form of GISC looks similar to the
criterion given in [6], their meanings are different. On one hand,
the open-loop transfer function used to plot the Nyquist curve is
not the loop-gain Zsource s / Zload s of the source and load
impedances seen at their interface, but the loop-gain of the
equivalent generalized-impedance circuit. On the other hand,
the criterion in [6] performs very well for single-phase system
(which is a SISO system), but there may be errors resulting
from the coupling terms [5] in three-phase systems. In contrast,
the GISC uses the special structure of the impedance matrix in
(2) and established an equivalent SISO system from a MIMO
system without any approximation.
Moreover, the distance between (1, j0) and the Nyquist
curve represents the stability margin, so the GISC can also be
used to guide the oscillation suppression. For example, the
influences of the PLL’s parameters on system stability can be
analyzed and tuned via the GISC. The studied system is shown
in Fig. 1, in which LLine is set as 0.26p.u., other parameters of
the VSC are listed in Table I. The Nyquist plots of the
open-loop transfer function with three groups of PLL
parameters, i.e., the proportional gain K ppll and the integral
gain Kipll , are given in Fig. 3. It shows that, in the neighborhood
of the PLL parameters given in Table I, larger proportional gain
( K ppll ) has more positive effect on system stability (see the
curves of P2 and P3 in Fig. 3), while larger integral gain ( Kipll )
has more negative effect (see the curves of P1 and P2).
0.26p.u. at t=2s. The reference signals of the current control
loop are assumed to be constant, and I dref 1p.u. , I qref 0p.u.
The stability of this system is analyzed using the GISC. The
Nyquist curve of Yg (s)(Z Z ) 2 is shown in Fig. 4. The point
(1, j0) is not encircled initially, thus the system is stable. With
the additional inductor, the plot shows that the system has two
poles in the right half plane, so the system is unstable. The HIL
simulation results in Fig. 5 also verify the above analysis,
where the sub-synchronous oscillation (42Hz) and supersynchronous oscillation (58Hz) occur after the impedance is
increased from 0.2p.u. to 0.26p.u.
0.8
6 dB
4 dB
2 dB
P 1 (Kppll=2.5,
ipll=3020)
-2 dB
-4 K
dB
0 dB
P 2 (Kppll=2.5, Kipll=2000)
P 3 (Kppll=5, Kipll=2000)
0.6
10 dB
0.4
Imaginary Axis
Y Y
Z
2 Z g I det g
Z
Y
2
0.2
-10 dB
20 dB
-20 dB
0
P1
P3
P2
-0.2
-0.4
-0.6
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
Real Axis
Fig. 3. Nyquist plots of the open-loop transfer function with different PLL
parameters.
4 dB
0.6 6 dB
0.4
Imaginary Axis
Z
det
Z
3
2 dB
0 dB
-2 dB
Lline = 0.20
Lline = 0.26
10 dB
-10 dB
0.2
20 dB
-20 dB
0
-0.2
-0.4
-0.6
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
Real Axis
Fig. 4. Nyquist plots of the open-loop transfer function with different LLine .
Voltage (Phase A)
Current (Phase A)
Divergent Oscillation
V. EXPERIMENTAL RESULTS
The hardware-in-the-loop (HIL) simulation of a three-phase
grid-connected VSC is conducted in RT-LAB, where the VSC
and the grid are simulated by OP5600 and the converter is
controlled by a controller based on TMS320F28335. The
impedance of the transmission line is changed from 0.20p.u. to
t =2s
Voltage spectrum
Fig. 5. Instability due to the change of inductance from 0.20p.u. to 0.26p.u.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
TABLE I. PARAMETERS OF THE VSC
Symbol
Sb
Ub
Lf
Gi(s)
Gpll(s)
linearized state equation of LLine can be expressed as:
Description
Base value of power
Base value of voltage
Inductance of the inverter side filter
Transfer function of the current controller
Transfer function of the PLL
4
1
I Linex
s U x
I Liney L s 2 L 2 s U y .
Line
Line
Value
500kVA
690V
0.2pu
0.6+15/s
2.5/s+3020/s2
Denote Line to be the power factor angle at the node 1 of LLine .
Then (B.2) can be converted to the polar coordinates as follows:
I Line
U
I Line Line = YLine U ,
where YLine
VI. CONCLUSION
The generalized-impedance based stability criterion is
proposed for the stability analysis of a three-phase gridconnected VSC system. By rigorous mathematical derivation, a
MIMO system composed of the VSC and the grid is
transformed to an equivalent SISO system composed of the
generalized-impedances of the grid and the VSC. Consequently,
the oscillation can be explained as the series resonance of the
equivalent circuit. Such GISC can be used for the stability
analysis of a three-phase system. GISC based study shows the
significant effects of the PLL parameters on system stability.
HIL simulation validates the effectiveness of the proposed
criterion. Future work will extend this criterion to systems with
multiple VSCs and use this criterion to guide VSC control
design.
By linearizing (1), the small-signal model of the inverter can be
obtained in the dq frame, and can be rewritten in the form of polar
coordinates as follows:
(A.1)
U sdref G s I dref L T G s T I U ,
0 f Θ2 i Θ1 I U
i
I
U
U sqref
I qref
(A.2)
U sd U
0 I q U
U sq U U L f sG pll s 0 I d U U
,
I
I
0 L f TΘ2
L
s
T
f
Θ1
I I
I I
(A.3)
cos I sin I
sin I cos I
and TΘ2
.
sin I cos I
cos I sin I
It follows from the assumptions that U sdref U sd , U sqref U sq ,
and I dref =I qref =0 are satisfied. Thus, from (A.2) and (A.3) the
following equation is obtained:
I
I
0 sG pll ( s) L f U U q ( sL f Gi (s ))TΘ1
.
I I
Id
(A.4)
Considering that the angles in (A.4) satisfy U pll
and I pll (shown in Fig. 2), pll in (A.1) can be further
expressed as pll Gpll (s)U / (1 Gpll (s)U ) in terms of . Thus, by
eliminating U , , and pll in (A.4), (A.5) holds:
I 0 0 U
I 0 Yg ( s) U ,
where Yg (s) Gi (s)Gpll (s)I 0 / [(Gi (s) L f s)(1 Gpll (s)U 0 )] and
subscript “0” denotes steady-state values.
(A.5)
the
APPENDIX B
The linearized state equation of an inductor in xy reference frame
can be expressed as:
1
I kx
s U ix U jx
I ky L s 2 L 2 s U iy U jy ,
k
k
(B.1)
where i and j are the nodes next to the inductor.
Since the voltage of the ideal grid is considered to be constant, the
1
s cos Line sin Line
.
LLine s 2 LLine 2 s sin Line cos Line
I kx
I ky Ck
s U ix U jx
s U iy U jy ,
(B.4)
where i and j are the nodes next to the inductor.
Denote C to be the power factor angle at the node 1 of C f . Since
the voltage of the ground is also constant, the linearized state equation
of C f can be expressed as:
I C
U
I C C = YC U ,
(B.5)
s cos C sin C
s sin C cos C .
where YC CC
REFERENCES
[2]
[3]
[4]
[5]
[6]
where TΘ1
(B.3)
Similarly, the linearized state equation of a capacitor in xy
coordinates can be expressed as:
[1]
APPENDIX A
pll G pll ( s)U U
,
s
pll
(B.2)
[7]
[8]
F. Blaabjerg, Z. Chen, and S. B. Kjaer, "Power electronics as efficient
interface in dispersed power generation systems," IEEE Trans. Power
Electron., vol. 19, no. 5, pp. 1184-1194, Sep. 2004.
B. Wen, D. Boroyevich, R. Burgos, P. Mattavelli, and Z. Shen, "Analysis
of D-Q small-signal impedance of grid-tied inverters," IEEE Trans.
Power Electron., vol. 31, no. 1, pp. 675-687, Jan. 2016.
B. Wen, D. Dong, D. Boroyevich, R. Burgos, P. Mattavelli, and Z. Shen,
"Impedance-based analysis of grid-synchronization stability for
three-phase paralleled converters," IEEE Trans. Power Electron., vol.
31, no. 1, pp. 26-37, Jan. 2016.
M. Cespedes, J. Sun, "Impedance modeling and analysis of grid- connected
voltage-source converters," IEEE Trans. Power Electron., vol. 29, no.
3, pp. 1254-1261, Mar. 2014.
M. K. Bakhshizadeh, X. Wang, F. Blaabjerg, et. al., "Couplings in phase
domain impedance modeling of grid-connected converters," IEEE Trans.
Power Electron., vol. 31, no. 10, pp. 6792-6796, Oct. 2016.
J. Sun, "Impedance-based stability criterion for grid-connected inverters,"
IEEE Trans. Power Electron., vol.26, no.11, pp.3075-3078, Nov. 2011.
A. Rygg, M. Molinas, Z. Chen, and X. Cai, "A modified sequence domain
impedance definition and its equivalence to the dq-domain impedance
definition for the stability analysis of ac power electronic systems," IEEE J.
Emerging Sel. Topics Power Electron., vol. 4, no. 4, pp. 1383-1396, Dec.
2016.
M. Zhao, X. Yuan, J. Hu, and Y. Yan, "Voltage dynamics of current
control time-scale in a VSC-connected weak grid," IEEE Trans. Power
Syst., vol. 31, no.4, pp. 2925-2937, Jul. 2016.
| 3cs.SY
|
arXiv:1610.02523v2 [math.GR] 7 Jun 2017
THE kR -PROPERTY ON FREE TOPOLOGICAL GROUPS
FUCAI LIN, SHOU LIN, AND CHUAN LIU
Abstract. A space X is called a kR -space, if X is Tychonoff and the necessary and sufficient
condition for a real-valued function f on X to be continuous is that the restriction of f on
each compact subset is continuous. In this paper, we mainly discuss the kR -property on the
free topological groups, and generalize some well-known results of K. Yamada’s in the free
topological groups.
1. Introduction
Recall that X is called a k-space, if the necessary and sufficient condition for a subset A of
X to be closed is that A ∩ C is closed for every compact subset C. It is well-known that the
k-property generalizing metrizability has been studied intensively by topologists and analysts.
A space X is called a kR -space, if X is Tychonoff and the necessary and sufficient condition for a
real-valued function f on X to be continuous is that the restriction of f on each compact subset
is continuous. Clearly every Tychonoff k-space is a kR -space. The converse is false. Indeed, for
any uncountable cardinal κ the power Rκ is a kR -space but not a k-space, see [11, Theorem 5.6]
and [10, Problem 7.J(b)]. Now, the kR -property has been widely used in the study of topology,
analysis and category, see [3, 4, 5, 6, 13, 16, 17].
Results of our research will be presented in two separate papers. In the paper [14], we mainly
extend some well-known results in k-spaces to kR -spaces, and then seek some applications in the
study of free Abelian topological groups. In the current paper, we shall deter the kR -property
on free topological groups and extend some results of K. Yamada’s on free topological groups.
The paper is organized as follows. In Section 2, we introduce the necessary notation and
terminologies which are used for the rest of the paper. In Section 3, we investigate the kR property on free topological groups, and generalize some results of K. Yamada’s. In section 4,
we pose some interesting questions about kR -spaces in the class of free topological groups which
are still unknown for us.
2. Preliminaries
In this section, we introduce the necessary notation and terminologies. Throughout this
paper, all topological spaces are assumed to be Tychonoff, unless otherwise is explicitly stated.
First of all, let N be the set of all positive integers and ω the first countable order. For a space
X, we always denote the set of all the non-isolated points by NI(X). For undefined notation
and terminologies, the reader may refer to [2], [8], [9] and [15].
Let X be a topological space and A ⊆ X be a subset of X. The closure of A in X is denoted
by A. Moreover, A is called bounded if every continuous real-valued function f defined on A
is bounded. The space X is called a k-space provided that a subset C ⊆ X is closed in X if
C ∩ K is closed in K for each compact subset K of X. A space X is called a kR -space, if X
is Tychonoff and the necessary and sufficient condition for a real-valued function f on X to be
continuous is that the restriction of f on each compact subset is continuous. Note that every
Tychonoff k-space is a kR -space. A subset P of X is called a sequential neighborhood of x ∈ X,
2000 Mathematics Subject Classification. Primary 54H11, 22A05; Secondary 54E20; 54E35; 54D50; 54D55.
Key words and phrases. kR -space; k-space; stratifiable space; Lašnev space; k-network; free topological group.
The first author is supported by the NSFC (Nos. 11571158, 11201414, 11471153), the Natural Science
Foundation of Fujian Province (Nos. 2017J01405, 2016J05014, 2016J01671, 2016J01672) of China.
1
2
FUCAI LIN, SHOU LIN, AND CHUAN LIU
if each sequence converging to x is eventually in P . A subset U of X is called sequentially open
if U is a sequential neighborhood of each of its points. A subset F of X is called sequentially
closed if X \ F is sequentially open. The space X is called a sequential space if each sequentially
open subset of X is open. The space X is said to be Fréchet-Urysohn if, for each x ∈ A ⊂ X,
there exists a sequence {xn } such that {xn } converges to x and {xn : n ∈ N} ⊂ A.
Definition 2.1. [4] Let X be a topological space.
• A subset U of X is called R-open if for each point x ∈ U there is a continuous function
f : X → [0, 1] such that f (x) = 1 and f (X \ U ) ⊂ {0}. It is obvious that each R-open is open.
The converse is true for the open subsets of Tychonoff spaces.
• A subset U of X is called a functional neighborhood of a set A ⊂ X if there is a continuous
function f : X → [0, 1] such that f (A) ⊂ {1} and f (X \ U ) ⊂ {0}. If X is normal, then each
neighborhood of a closed subset A ⊂ X is functional.
Definition 2.2. Let λ be a cardinal. An indexed family {Xα }α∈λ of subsets of a topological
space X is called
• point-countable if for any point x ∈ X the set {α ∈ λ : x ∈ Xα } is countable in X;
• compact-countable if for any compact subset K in X the set {α ∈ λ : K ∩ Xα 6= ∅} is
countable in X;
• locally finite if any point x ∈ X has a neighborhood Ox ⊂ X such that the set {α ∈ λ :
Ox ∩ Xα 6= ∅} is finite in X;
• compact-finite in X if for each compact subset K ⊂ X the set {α ∈ λ : K ∩ Xα 6= ∅} is
finite in X;
• strongly compact-finite [4] in X if each set Xα has an R-open neighborhood Uα ⊂ X such
that the family {Uα }α∈λ is compact-finite in X;
• strictly compact-finite [4] in X if each set Xα has a functional neighborhood Uα ⊂ X such
that the family {Uα }α∈λ is compact-finite in X.
Definition 2.3. [4] Let X be a topological space and λ be a cardinal. An indexed family
{Fα }α∈λ of subsets of a topological space X is called a fan (more precisely, a λ-fan) in X if this
family is compact-finite but not locally finite in X. A fan {Fα }α∈λ is called strong (resp. strict)
if each set Fα has a R-open neighborhood (resp. functional neighborhood) Uα ⊂ X such that
the family {Uα }α∈λ is compact-finite in X.
If all the sets Fα of a λ-fan {Fα }α∈λ belong to some fixed family F of subsets of X, then
the fan will be called an F λ -fan. In particular, if each Fα is closed in X, then the fan will be
called a Cldλ -fan.
Clearly, we have the following implications:
strict fan ⇒ strong fan ⇒ fan.
Let P be a family of subsets of a space X. Then, P is called a k-network if for every compact
′
subset K of X and
S an′ arbitrary open set U containing K in X there is a finite subfamily P ⊆ P
such that K ⊆ P ⊆ U . Recall that a space X is an ℵ-space (resp. ℵ0 -space) if X has a
σ-locally finite (resp. countable) k-network. Recall that a space X is said to be Lašnev if it is
the continuous closed image of some metric space.
Definition 2.4. [7] A topological space X is a stratifiable space if for each open subset U in X,
one can assign a sequence {Un }∞
n=1 of open subsets of X such that
⊂
U
;
(a) U
Sn∞
(b) n=1 Un = U ;
(c) Un ⊂ Vn whenever U ⊂ V .
Note: Each Lašnev space is stratifiable [9].
Let X be a non-empty space. Throughout this paper, X −1 = {x−1 : x ∈ X}, which is just
a copy of X. For every n ∈ N, the Fn (X) denotes the subspace of F (X) that consists of all
THE kR -PROPERTY ON FREE TOPOLOGICAL GROUPS
3
the words of reduced length at most n with respect to the free basis X. Let e be the neutral
element of F (X), that is, the empty
For every n ∈ N, an element x1 x2 · · · xn is also called
L word.
L
a form for (x1 , x2 , · · · , xn ) ∈ (X
X −1 {e})n . The word g is called reduced if it does not
contain e or any pair of consecutive symbol of the form xx−1 . It follows that if the word g is
reduced and non-empty, then it is different from the neutral element e of F (X). In particular,
each element g ∈ F (X) distinct from the neutral element can be uniquely written in the form
ǫ1 ǫ2
ǫn
g = xǫ11 xǫ22 · · · xǫnn , where n ≥ 1, ǫ1 ∈ {−1, 1}, xi ∈ X, and the support of g = xS
1 x2 · · · xn is
defined as supp(g) = {x1 , · · · , xn }. Given a subset K of F (X), we put supp(K) = g∈K supp(g).
For every n ∈ N, let
M
M
in : (X
X −1
{e})n → Fn (X)
be the natural mapping defined by
for each (x1 , x2 , · · · , xn ) ∈ (X
L
in (x1 , x2 , ...xn ) = x1 x2 · · · xn
L
−X {0})n .
3. The kR -property on free topological groups
In this section, we investigate the kR -property on free topological groups, and generalize some
results of K. Yamada’s. Recently, T. Banakh in [4] proved that F (X) is a k-space if F (X) is
a kR -space for a Lašnev space X. Indeed, he obtained this result in the class of more weaker
spaces. However, he did not discuss the following quesiton:
Question 3.1. Let X be a space. For some n ∈ ω, if Fn (X) is a kR -space, is Fn (X) a k-space?
First, we give the following Theorem 3.3, which gives a complementary for the result of T.
Banakh’s.
Lemma 3.2. Let F (X) be a kR -space. If each Fn (X) is a normal k-space, then F (X) is a
k-space.
Proof. It is well-known that each compact subset of F (X) is contained in some Fn (X) [2,
Corollary 7.4.4]. Hence it follows from [13, Lemma 2] that F (X) is a k-space.
Theorem 3.3. Let X be a paracompact σ-space. Then F (X) is a kR -space and each Fn (X) is
a k-space if and only if F (X) is a k-space.
Proof. Since X is a paracompact σ-space, it follows from [2, Theorem 7.6.7] that F (X) is also
a paracompact σ-space, hence each Fn (X) is normal. Now it is legal to apply Lemma 3.2 and
finish the proof.
Next, we shall show that for arbitrary a metrizable space X, the kR -property of F8 (X) implies
that F (X) is a k-space, see Theorem 3.6. We first prove two technic propositions. To prove
them, we need the description of a neighborhood base of e in F (X) obtained in [20]. Let
2n
H0 (X) = {h = xε11 xε22 · · · xε2n
∈ F (X) :
2n
X
εi = 0, xi ∈ X for i ∈ {1, 2, · · · , n}, n ∈ N}.
i=1
Obviously, the subset H0 (X) is a clopen normal subgroup of F (X). It is easy to see that each
h ∈ H0 (X) can be represented as
h = g1 xε11 y1−ε1 g1−1 g2 xε22 y2−ε1 g2−1 · · · gn xεnn yn−ε1 gn−1 ,
for some n ∈ N, where xi , yi ∈ X, εi = ±1 and gi ∈ F (X) for i ∈ {1, 2, · · · , n}. Let P (X) be
the set of all continuous pseudometrics on a space X. Then take an arbitrary r = {ρg : g ∈
F (X)} ∈ P (X)F (X) . Let
n
X
ρgi (xi , yi ) : h = g1 xε11 y1−ε1 g1−1 g2 xε22 y2−ε1 g2−1 · · · gn xεnn yn−ε1 gn−1 , n ∈ N}
pr (h) = inf{
i=1
for each h ∈ H0 (X). In [20], Uspenskii proved that
4
FUCAI LIN, SHOU LIN, AND CHUAN LIU
(a) ρr is a continuous on H0 (X) and
(b) {{h ∈ H0 (X) : pr (h) < δ} : r ∈ P (X)F (X) , δ > 0} is a neighborhood base of e in F (X).
Moreover, pr (e) = 0 for each r ∈ P (X)F (X) .
Proposition 3.4. For a stratifiable k-space X if F8 (X) is a kR -space, then X is separable or
discrete.
Proof. Assume
L on the contrary that X is neither separable nor discrete. Then X contains a
space Y = C
D as a closed subset, where C = {xn : n ∈ ω}∪{x} is a convergent sequence with
its limit point {x} and D = {dα : α ∈ ω1 } is a discrete closed subset of X. Since D is a discrete
closed subset of X, we choose a discrete family {Oα }α∈ω1 of open subsets suchSthat dα ∈ Oα
for each α ∈ ω1 . We may assume that xn 6= xm for arbitrary n 6= m and C ∩ α∈ω1 Oα = ∅.
Since X is stratifiable and Y is closed in X, it follows from [20] that F (Y ) is homeomorphic to
a closed subgroup of F (X). Hence F8 (Y ) is closed subspace of F8 (X). Next we shall show that
F8 (X) contains a strict Cldω -fan, a contradiction. Indeed, since F8 (X) is normal, it suffices to
construct a strong Cldω -fan in F8 (X).
For each α ∈ ω1 choose a function fα : ω1 → ω such that fα |α : α → ω is a bijection. For
each n ∈ ω, let
−1
−1
Fn = {d−1
xm dβ dα xn x−1 dα
: fα (β) = n, α, β ∈ ω1 , m ≤ n}.
β x
We claim that the family {Fn : n ∈ ω} is a strong Cldω -fan in F8 (X). We divide into the proof
by the following three statements.
(1) For each n ∈ ω, the set Fn is closed in F8 (X).
Fix an arbitrary n ∈ ω. It suffices to show that the set Fn is closed in F8 (Y ). Let Z = suppFn .
Then it is obvious that Z is a closed discrete subspace of Y . It follows from [20] that F (Z) is
homeomorphic to a closed subgroup of F (Y ), and thus F8 (Z) is a closed subspace of F8 (Y ).
Since F (Z) is discrete and Fn ⊂ F8 (Z), the set Fn is closed in F8 (Y ) (and thus closed in F (X)).
(2) The family {Fn : n ∈ ω} is strong compact-finite in F8 (X).
By induction, choose two families of open neighborhoods {Wn }n∈ω and {Vn }n∈ω in X satisfy
the following conditions:
(a) for each n ∈ ω, xn ∈ Wn ;
(b) for each n ∈ ω, x ∈ Vn and Vn+1 ⊂ Vn ;
(c) for each n ∈ ω, we have Wn ∩ (C ∪ D) = {xn }, Vn ∩ (D ∪ Wn ) = ∅ and Vn ∩ C ⊂ Cn ,
where Cn = {xm : m > n} ∪ {x};
S
S
(d) V1 ∩ α∈ω1 Oα = ∅ and Wn ∩ α∈ω1 Oα = ∅ for each n ∈ ω.
For each n ∈ ω, let
[
Un = {Oβ−1 Vn−1 Wm Oβ Oα Wn Vn−1 Oα−1 : fα (β) = n, α, β ∈ ω1 , m ≤ n}.
Obviously, each Un contains in F8 (X) \ F7 (X), and since F8 (X) \ F7 (X) is open in F8 (X), it
follows from [2, Corollary 7.1.19] that each Un is open in F8 (X). We claim that the family
{Un : n ∈ ω} is compact-finite in F8 (X). Suppose not, there exist a compact subset K in F8 (X)
and an increasing sequence {nk } such that K ∩ Unk 6= ∅ for each k ∈ ω. Since X is stratifiable,
F (X) is paracompact, then the closure of the set supp(K) is compact in X. However, for each
k ∈ ω, since K ∩ Unk 6= ∅, there exist mk ∈ ω, αk ∈ ω1 and βk ∈ ω1 such that fαk (βk ) = nk and
6= ∅, hence supp(K) ∩ Oαk 6= ∅ and supp(K) ∩ Oβk 6= ∅.
Oα−1
Wmk Oβk Oαk Wnk Vn−1
K ∩ Oβ−1
Vn−1
k
k
k
k
Therefore, the set supp(K) intersects each element of the family {Oαk , Oβk : k ∈ ω}. Since the
set {nk : k ∈ ω} is infinite and fαk (βk ) = nk for each k ∈ ω, the set {αk , βk : k ∈ ω} is infinite.
Then supp(K) intersects infinitely many Oα ’s, which is a contradiction with the compactness of
supp(K). Therefore, the family {Un : n ∈ ω} is compact-finite in F8 (X). Therefore, the family
{Fn : n ∈ ω} is strong compact-finite in F8 (X).
(3) The family {Fn : n ∈ ω} is not locally finite at the point e in F8 (Y ) (and thus not locally
finite in F8 (X)).
THE kR -PROPERTY ON FREE TOPOLOGICAL GROUPS
S
S
n∈ω
{h ∈ F8 (Y ) : pr (h) < δ} ∩
[
Indeed, it suffices to show that e ∈
r = {ρg : g ∈ F (Y )}, we shall show that
n∈ω
Fn \
5
Fn in F8 (Y ). For any δ > 0 and
Fn 6= ∅.
n∈ω
are continuous pseudometrics on Y
Since the sequence {xn } converges to x and ρdα and ρd−1
α
(xn , x) < 2δ for each
for each α ∈ ω1 , there is n(α) ∈ ω such that ρdα (xn , x) < δ2 and ρd−1
α
n ≥ n(α). Therefore, there are n0 ∈ ω and uncountable set A ⊂ ω1 such that ρdα (xn , x) < 2δ
and ρd−1
(xn , x) < 2δ for each n ≥ n0 and α ∈ A. Choose α ∈ A that has infinitely many
α
predecessors in A. Since fα (α ∩ A) is an infinite set, there exist m > n0 and β ∈ α ∩ A such
that fα (β) = m. Then the word
−1
g = d−1
xn0 dβ dα xfα (β) x−1 d−1
α ∈ Ffα (β) = Fm .
β x
Furthermore, we have
δ
δ
−1
(xn0 , x) + ρdα (xm , x) < + = δ.
pr (g) = pr (d−1
xn0 dβ dα xfα (β) x−1 d−1
α ) ≤ ρd−1
β x
β
2 2
S
Hence e ∈ n∈ω Fn . Then The family {Fn : n ∈ ω} is not locally finite at the point e in
F8 (Y ).
Proposition 3.5. For a metrizable space X if F8 (X) is a kR -space, then X is locally compact.
Proof. Assume on the contrary that X is not locally compact. Then there exists a closed
hedgehog subspace
[
J = {x} ∪ (
Xn ) ∪ {zn : n ∈ ω}
n∈ω
such that
(1) Xn = {yn } ∪ {xn,j : j ∈ ω} is a closed discrete subset of J for each n ∈ ω;
(2) {zn : n S
∈ ω} is a closed discrete subset of J; and
(3) {{x} ∪ n≥k Xn : k ∈ ω} is a neighborhood base of x in J.
By Proposition 3.4, the space X is separable. Next we shall show that F8 (X) contains a strict
Cldω -fan, which is a contradiction with F8 (X) is a kR -space. Since F8 (X) is an ℵ0 -space by
[1, Theorem 4.1], the subspace F8 (X) is normal, hence it suffices to show that F8 (X) contains
a strong Cldω -fan. Furthermore, it follows from [4, Proposition 2.9.2] that each compact-finite
family of subsets of X is strongly compact-finite, hence it suffices to show that F8 (X) contains
a Cldω -fan. For any n, j ∈ ω, put
En,j = {zn−1 yj−1 xzn xn,j }.
Then it is obvious that each En,j is closed. Furthermore, it follows from the proof of [19,
S
S
Proposition 2.1] that x ∈ n,j∈ω En,j \ n,j∈ω En,j , and thus the family {En,j : n, j ∈ ω} is not
locally finite at the point x. Next we claim that the family {En,j : n, j ∈ ω} is compact-finite.
Suppose not, there exist a compact subset K and two sequences {ni } and {ji } such that
K ∩ Eni ,ji 6= ∅. Then the closure of the set supp(K) is compact since F8 (X) is paracompact.
Since the family {En,j : n, j ∈ ω} are disjoint each other, one of the sequences {ni } and {ji } is
an infinite set. If {ni } is an infinite set, then the closed discrete set {zni : i ∈ ω} is contained
is an infinite set and
in supp(K), which is a contradiction since supp(K) is compact. If {ji } S
{ni } is a finite set, then there exists N ∈ ω such that {xni ,ji : i ∈ N} ⊂ j<N Xj . Obviously,
the closed discrete set {xni ,ji : i ∈ N} is an infinite set and contains in supp(K), which is a
contradiction.
Now we can show one of our main theorems in this paper.
Theorem 3.6. For a metrizable space X, the following are equivalent:
(1) F (X) is a k-space;
(2) F8 (X) is a k-space;
6
FUCAI LIN, SHOU LIN, AND CHUAN LIU
(3) F8 (X) is a kR -space;
(4) the space X is locally compact separable or discrete.
Proof. The equivalence of (1) and (4) was showed in [19]. It is obvious that (2) ⇒ (3). By
Propositions 3.4 and 3.5, we have (3) ⇒ (4).
By Theorem 3.6, it is natural to ask the following question:
Question 3.7. Let X be a metrizable space. If Fn (X) is a kR -space for some n ∈ {4, 5, 6, 7},
then is F8 (X) a kR -space?
Note For each n ∈ {2, 3}, the answer to the above question is negative. Indeed, for an
arbitrary metrizable space X, since i2 is a closed mapping, F2 (X) is a Fréchet-Urysohn space
(and thus a k-space). For n = 3, we have the following Theorem 3.9. However, for each
n ∈ {4, 5, 6, 7}, the above question is still unknown for us.
Proposition 3.8. For a metrizable space X if F3 (X) is a kR -space, then X is locally compact
or NI(X) is compact.
Proof. Assume on the contrary that neither X is locally compact nor the set of all non-isolated
points of X is compact. Then X contains a closed subspace
[
M
Y = {x} ∪
Xn ⊕
Cn ,
n∈ω
n∈ω
where for every n ∈ ω
Xn = {x
Sn,i : i ∈ ω} is a closed discrete subset of X,
{{x} ∪ m≥n Xm : n ∈ ω} is a neighborhood base of x in Y ,
Cn = {cn,i : i ∈ ω} ∪ {cn } is a convergent sequence with its limit cn , and
Cn is contained in the open subset Un of X such that the family {Un : n ∈ ω} is discrete in
S
S
X and ({x} ∪ m∈ω Xm ) ∩ m∈ω Um = ∅.
In order to obtain a contradiction, we shall construct a strict Cldω -fan in F3 (X). For any
n, i ∈ ω, choose an open neighborhood Oni of the point xn,i in X such that the family {Oni : i ∈ ω}
S
S
is discrete, Oni ∩ m∈ω Um = ∅ and Oni ∩ ({x} ∪ n∈ω Xn ) = {xn,i }.
For each n, i ∈ ω, let
E(n, i) = {gn,i = cn c−1
n,i xn,i }
and
U (n, i) = Vni (Wni )−1 Oni ,
where Vni and Wni are two arbitrary open neighborhoods of cn and cn,i in X respectively such that
Vni ∪Wni ⊂ Un and Vni ∩Wni = ∅. Obviously, each E(n, i) is closed and it follows from [2, Corollary
7.1.19] that U (n, i) is an open neighborhood of E(n, i) for each n, i ∈ ω. In [19], the author has
S
S
showed that x ∈ n,i E(n, i) \ n,i E(n, i), hence the family {E(n, i) : n, i ∈ ω} is not locally
finite in F3 (X). To complete the proof, it suffices to show that the family {U (n, i) : n, i ∈ ω} is
compact-finite in F3 (X). Suppose not, there exists a compact subset K and two sequences {nj }
and {ij } such that K ∩ U (nj , ij ) 6= ∅. Similar to the proof of Proposition 3.5, we can obtain a
contradiction.
Theorem 3.9. For a metrizable space X the following are equivalent:
(1) F3 (X) is a k-space;
(2) F3 (X) is a kR -space;
(3) the space X is locally compact or NI(X) is compact.
Proof. The equivalence of (1) and (3) was showed in [19]. The implication of (1) ⇒ (2) is
obvious. By Proposition 3.8, we have (2) ⇒ (3).
The following theorem was proved in [14].
THE kR -PROPERTY ON FREE TOPOLOGICAL GROUPS
7
Theorem 3.10. [14] Let X be a stratifiable space such that X 2 is a kR -space. If X satisfies
one of the following conditions, then either X is metrizable or X is the topological sum of
kω -subspaces.
(1) X is a k-space with a compact-countable k-network;
(2) X is a Fréchet-Urysohn space with a point-countable k-network.
Since F2 (X) contains a closed copy of X × X, it follows from Theorem 3.10 that we have the
following theorem.
Theorem 3.11. Let X be a stratifiable k-space with a compact-countable k-network. Then the
following are equivalent:
(1) F2 (X) is a k-space;
(2) F2 (X) is a kR -space;
(3) either X is metrizable or X is the topological sum of kω -subspaces.
The following proposition shows that we can not replace “F3 (X)” with “F4 (X)” in Theorem 3.9. First, we recall a special space. Let
M
M3 =
{Cα : α < ω1 },
where Cα = {c(α, n) : n ∈ N} ∪ {cα } with c(α, n) → cα as n → ∞ for each α ∈ ω1 .
Proposition 3.12. The subspace F4 (M3 ) is not a kR -space.
Proof. It suffices to show that F4 (M3 ) contains a strict Cldω -fan. It follows from [12, Theorem
20.2] that we can find two families A = {Aα : α ∈ ω1 } and B = {Bα : α ∈ ω1 } of infinite
subsets of ω such that
(a) Aα ∩ Bβ is finite for all α, β ∈ ω1 ;
(b) for no A ⊂ ω, all the sets Aα \ A and Bα ∩ A, α ∈ ω1 are finite.
For each n ∈ ω, put
−1
Xn = {c(α, n)c−1
α c(β, n)cβ : n ∈ Aα ∩ Bβ , α, β ∈ ω1 }.
It suffices to show the following three statements.
(1) The family {Xn } is strictly compact-finite in F4 (M3 ).
Since M3 is a Lašnev space, it follows from [2, Theorem 7.6.7] that F (M3 ) is also a paracompact σ-space, hence F4 (M3 ) is paracompact (and thus normal). Hence it suffices to show
that the family {Xn } is strongly compact-finite in F4 (M3 ). For each α ∈ ω1 and n ∈ ω, let
Cαn = Cα \ {c(α, m) : m ≤ n}, and put
Un = {c(α, n)x−1 c(β, n)y −1 : n ∈ Aα ∩ Bβ , α, β ∈ ω1 , x ∈ Cαn , y ∈ Cβn }.
Obviously, each Xn ⊂ F4 (M3 ) \ F3 (M3 ). Since F4 (M3 ) \ F3 (M3 ) is open in F4 (M3 ), it follows
from [2, Corollary 7.1.19] that each Un is open in F4 (M3 ). We claim that the family {Un } is
compact-finite in F4 (M3 ). Suppose not, then there exist a compact subset K in F4 (M3 ) and
a subsequence {nk } of ω such that K ∩ Unk 6= ∅ for each k ∈ ω. For each k ∈ ω, choose an
arbitrary point
−1
zk = c(αk , nk )x−1
k c(βk , nk )yk ∈ K ∩ Unk ,
n
where xk ∈ Cαnkk and yk ∈ Cβkk . Since F4 (M3 ) is paracompact, it follows from [1] that the closure
of the set supp(K) is compact in M3 . Therefore, there exists N ∈ ω such that
[
supp(K) ∩ {Cα : α ∈ ω1 \ {αi ∈ ω1 : i ≤ N }} = ∅,
S
that is, supp(K) ⊂ α∈{γi ∈ω1 :i≤N } Cα . Since each zk ∈ K, there exists
αk , βk ∈ {αi ∈ ω1 : i ≤ N }
such that Aαk ∩ Bβk is an infinite set, which is a contradiction with Aα ∩ Bβ is finite for all
α, β < ω1 .
(2) Each Xn is closed in F4 (M3 ).
8
FUCAI LIN, SHOU LIN, AND CHUAN LIU
Fix an arbitrary n ∈ ω. Next we shall show that Xn is closed in F4 (M3 ). Let Z = supp(Xn ).
Then Z is a closed discrete subset of M3 . Since M3 is metriable, it follows from [20] that F (Z)
is homeomorphic to a closed subgroup of F (M3 ), hence F4 (Z) is a closed subspace of F4 (M3 ).
Since F (Z) is discrete and Xn ⊂ F4 (Z), the set Xn is closed in F4 (Z) (and thus closed in
F4 (M3 )).
(3) The family {Xn } is not locally finite at the point e in F4 (M3 ).
S
S
Indeed, it suffices to show that e ∈ n∈ω Xn \ n∈ω Xn . For any δ > 0 and r = {ρg : g ∈
F (M3 )}, we shall show that
[
{h ∈ F4 (M3 ) : pr (h) < δ} ∩
Xn 6= ∅.
n∈ω
since ρe is continuous, we can choose a function f : ω1 → ω such that ρe (c(α, n), cα ) < 2δ
for any α ∈ ω1 and n ≥ f (α). For each α < ω1 , put A′α = {n ∈ Aα : n ≥ f (α)} and
Bα′ = {n ∈ Bα : n ≥ f (α)}. By the condition (b) of the families A and B, it is easy to see that
there exist α, β ∈ ω1 such that A′α ∩ Bβ′ 6= ∅. So, choose n ∈ A′α ∩ Bβ′ . Then ρe (c(α, n), cα ) < 2δ
−1
and ρe (c(β, n), cβ ) < 2δ . Let z = c(α, n)c−1
α c(β, n)cβ . Then z ∈ Xn and
δ
δ
pr (z) ≤ ρe (c(α, n), cα ) + ρe (c(β, n), cβ ) < + = δ,
2 2
S
hence z ∈ {h ∈ F4 (M3 ) : pr (h) < δ} ∩ n∈ω Xn .
Theorem 3.13. Let X be a metrizable space. If F4 (X) is a kR -space, then NI(X) is separable.
Proof. Suppose not, then X contains a closed copy of M3 . Use the same notation in Theorem 3.12. Since X is metrizable, there exists a discrete family {Uα }α∈ω1 of open subsets in X
such that Cα ⊂ Uα for each α ∈ ω1 . For arbitrary (α, n) ∈ ω1 × ω, choose open neighborhoods
Vα,n and Wα,n of the point c(α, n) and cα in X respectively such that Vα,n ∩ Wα,n = ∅ and
Vα,n ∪ Wα,n ⊂ Uα . For each n ∈ N, put and
[
−1
−1
On = {Vα,n Wα,n
Vβ,n Wβ,n
: n ∈ Aα ∩ Bβ , α, β ∈ ω1 }.
Similar to the proof of Theorem 3.12, we can show that the family {On }n∈ω of open subsets
is compact-finite in F4 (X), hence the family {Xn }n∈ω is a strict Cldω -fan in F4 (X), which is a
contradiction.
4. Open questions
In this section, we pose some interesting questions about kR -spaces in the class of free topological groups, which are still unknown for us.
By Theorems 3.6, 3.9 and 3.11, it is natural to pose the following question.
Question 4.1. Let X be a metrizable space. For each n ∈ {4, 5, 6, 7}, if Fn (X) is a kR -space,
is Fn (X) a k-space?
In [18], Yamada also made the following conjecture:
Yamada’s Conjecture: The subspace F4 (X) is Fréchet-Urysohn if the set of all non-isolated
points of a metrizable space X is compact.
Indeed, we know no answer for the following question.
Question 4.2. Is F4 (X) a kR -space if the set of all non-isolated points of a metrizable space
X is compact?
In particular, we have the following question.
L
Question 4.3. Let X = C
D, where C is a non-trivial convergent sequence with its limit
point and D is an uncountable discrete space. Is F4 (X) a kR -space?
In [3], the authors showed that each closed subspace of a stratifiable kR -space is a kR -subspace.
However, the following two questions are still open.
THE kR -PROPERTY ON FREE TOPOLOGICAL GROUPS
9
Question 4.4. Is each closed subgroup of a kR -free topological group kR ?
Question 4.5. Is each subspace Fn (X) of a kR -free topological group kR ?
Acknowledgements. The authors wish to thank the reviewers for careful reading preliminary version of this paper and providing many valuable suggestions.
References
[1] A.V. Arhangel’skiı̌, O. G. Okunev and V. G. Pestov, Free topological groups over metrizable spaces, Topology
Appl., 33(1989), 63–76.
[2] A. Arhangel’skiı̌ and M. Tkachenko, Topological groups and related structures, Atlantis Press, Paris; World
Scientific Publishing Co. Pte. Ltd., Hackensack, NJ, 2008.
[3] T. Banakh and S. Gabriyelyan, On the Ck -stable closure of the class of (separable) metrizable spaces,
Monatshefte für Mathematik, 180(2016), 39–64.
[4] T. Banakh, Fans and their applications in General Topology, Functional Analysis and Topological Algebra,
arXiv:1602.04857v2.
[5] J.L. Blasco, On µ-Spaces and kR -Spaces, Proc. Amer. Math. Soc., 67(1)(1977), 179–186.
[6] C.R. Borges, A stratifiable kR -space which is not a k-space, Proc. Amer. Math. Soc., 81(1981), 308–310.
[7] J.G. Ceder, Some generalizations of metric spaces, Pacific J. Math., 11(1961), 105-C125.
[8] R. Engelking, General Topology (revised and completed edition), Heldermann Verlag, Berlin, 1989.
[9] G. Gruenhage, Generalized metric spaces, In: K. Kunen, J. E. Vaughan(Eds.), Handbook of Set-Theoretic
Topology, Elsevier Science Publishers B.V., Amsterdam, 1984, 423–501.
[10] J. Kelley, General Topology, Springer-Verlag, New York-Berlin, 1975.
[11] N. Noble, The continuity of functions on Cartesian products, Trans. Amer. Math. Soc., 149(1970), 187C198.
[12] W. Just, M. Weese, Discovering Modern Set Theorey II, Graduate Studies in Mathematics, 18, 1997.
[13] J. Li, On k-spaces and kR -spaces, Czech. Math. J., 55(2006), 941–945.
[14] F. Lin, S. Lin, C. Liu, The kR -property in sequential fans and the applications in the free Abelian topological
groups, arXiv:1610.02522v1.
[15] S. Lin, Point-countable Covers and Sequence-covering mappings (second edition), Beijing, China Science
Press, 2015.
[16] S. Lin, Note on kR -spaces, Q & A in General Topoloy, 9(1991), 227–236.
[17] E. Michael,, On k-spaces, kR -spaces and k(X), Pacific J. Math.., 47(1973), 487–498.
[18] K. Yamada, Fréchet-Urysohn spaces in free topological groups, Proc. Amer. Math. Soc., 130(2002), 2461–
2469.
[19] K. Yamada, The natural mappings in and k-subspaces of free topological groups on metrizable spaces,
Topology Appl., 146-147(2005), 239–251.
[20] V. Uspenskii, Free topological groups of metrizable spaces, Math. USSR Izv. 37(1991), 657–680.
(Fucai Lin): School of mathematics and statistics, Minnan Normal University, Zhangzhou 363000,
P. R. China
E-mail address: [email protected]; [email protected]
(Shou Lin): Institute of Mathematics, Ningde Teachers’ College, Ningde, Fujian 352100, P. R.
China
E-mail address: [email protected]
(Chuan Liu): Department of Mathematics, Ohio University Zanesville Campus, Zanesville, OH
43701, USA
E-mail address: [email protected]
| 4math.GR
|
1
Binary Hypothesis Testing with Byzantine
Sensors: Fundamental Trade-off Between
Security and Efficiency
arXiv:1707.09592v3 [cs.IT] 19 Jan 2018
Xiaoqiang Ren∗ , Jiaqi Yan∗ , and Yilin Mo∗†
Abstract
This paper studies binary hypothesis testing based on measurements from a set of sensors, a subset
of which can be compromised by an attacker. The measurements from a compromised sensor can be
manipulated arbitrarily by the adversary. The asymptotic exponential rate, with which the probability
of error goes to zero, is adopted to indicate the detection performance of a detector. In practice, we
expect the attack on sensors to be sporadic, and therefore the system may operate with all the sensors
being benign for an extended period of time. This motivates us to consider the trade-off between the
detection performance of a detector, i.e., the probability of error, when the attacker is absent (defined as
efficiency) and the worst-case detection performance when the attacker is present (defined as security).
We first provide the fundamental limits of this trade-off, and then propose a detection strategy that
achieves these limits. We then consider a special case, where there is no trade-off between security and
efficiency. In other words, our detection strategy can achieve the maximal efficiency and the maximal
security simultaneously. Two extensions of the secure hypothesis testing problem are also studied and
fundamental limits and achievability results are provided: 1) a subset of sensors, namely “secure” sensors,
are assumed to be equipped with better security countermeasures and hence are guaranteed to be benign;
2) detection performance with unknown number of compromised sensors. Numerical examples are given
to illustrate the main results.
Index Terms
Hypothesis testing, Security, Secure detection, Efficiency, Trade-off, Byzantine attacks, Fundamental
limits
∗: Xiaoqiang Ren, Jiaqi Yan and Yilin Mo are with the School of Electrical and Electronics Engineering, Nanyang
Technological University, Singapore. Emails: [email protected], [email protected], [email protected]
†: Corresponding Author.
January 22, 2018
DRAFT
2
I. I NTRODUCTION
Background and Motivations: Network embedded sensors, which are pervasively used to
monitor the system, are vulnerable to malicious attacks due to their limited capacity and sparsely
spatial deployment. An attacker may get access to the sensors and send arbitrary messages, or
break the communication channels between the sensors and the system operator to tamper with
the transmitted data. Such integrity attacks have motivated many researches on how to infer
useful information from corrupted sensory data in a secure manner [1]–[3]. In this paper, we
follow this direction but with the focus on the trade-off between the performance of the inference
algorithm when the attacker is absent and the “worst-case” performance when the attacker, which
has the knowledge of the inference algorithm, is present. We define two metrics, efficiency and
security, to characterize the performance of the hypothesis testing algorithm (or detector) under
the two scenarios respectively and analyze the trade-off between security and efficiency.
Our Work and its Contributions: We consider the sequential binary hypothesis testing based
on the measurements from m sensors. It is assumed that n out of these m sensors may be
compromised by an attacker, the set of which is chosen by the attacker and fixed over time.
The adversary can manipulate the measurements sent by the compromised sensors arbitrarily.
According to Kerckhoffs’s principle [4], i.e., the security of a system should not rely on its
obscurity, we assume that the adversary knows exactly the hypothesis testing algorithm used by
the fusion center. On the other hand, the fusion center (i.e., the system operator) only knows the
number of malicious sensors n, but does not know the exact set of the compromised sensors.
At each time k, the fusion center needs to make a decision about the underlying hypothesis
based on the possibly corrupted measurements collected from all sensors until time k. Given a
hypothesis testing algorithm at the fusion center (i.e., a measurements fusion rule), the worst-case
probability of error is investigated, and the asymptotic exponential decay rate of the error, which
we denote as the “security” of the system, is adopted to indicate the detection performance. On
the other hand, when the attacker is absent, the detection performance of a hypothesis testing
algorithm, i.e., the asymptotic exponential decay rate of the error probability, is denoted by the
“efficiency”.
We focus on the trade-off between efficiency and security. In particular, we are interested in
characterizing the fundamental limits of the trade-offs between efficiency and security and the
detectors that achieve these limits.
January 22, 2018
DRAFT
3
The main contributions of this work are summarized as follows:
1) To the best of our knowledge, this is the first work that studies the trade-off between the
efficiency and security of any inference algorithm.
2) With mild assumptions on the probability distributions of the measurements, we provide the
fundamental limits of the trade-off between the efficiency and security (Corollaries 1 and 2).
Furthermore, we present detectors, with low computational complexity, that achieve these
limits (Theorem 4). Therefore, the system operator can easily adopt the detectors we
proposed to obtain the best trade-off between efficiency and security. Interestingly, in
some cases, e.g., Gaussian random variables with same variance and different mean, the
maximal efficiency and the maximal security can be achieved simultaneously (Theorem 5).
3) Similar results, i.e., the fundamental limits of the trade-off and the detectors that possess
these limits, are established with several different problem settings (Section V). This shows
that our analysis techniques are insightful and may be helpful for the future related studies.
Related Literature: A sensor is referred to as a Byzantine sensor if its messages to the fusion
center are fully controlled by an adversary1 . Recently, detection with Byzantine sensors has
been studied extensively in [5]–[14], among which [5]–[7] took the perspective of an attacker
and aimed to find the most effective attack strategy, [8]–[11] focused on designs of resilient
detectors, and [12]–[14] formulated the problem in a game-theoretic way. The main results
of [5] are the critical fraction of Byzantine sensors that blinds the fusion center, which is
just the counterpart of the breakdown point in robust statistics [15], and the most effective
attack strategy that minimizes the asymptotic error exponent in the Neyman-Pearson setting, i.e.,
the Kullback–Leibler (K–L) divergence. Since the Byzantine sensors are assumed to generate
independent and identical distributed (i.i.d.) data, the resulting measurements with minimum
K–L divergence and the corresponding robust detector coincide with those in [16]. Similar
results were obtained in [6], [7] by considering non-asymptotic probability of error in Bayesian
setting and asymptotic Bayesian performance metric, i.e., Chernoff information, respectively.
The authors in [8] focused on computation efficient algorithms to determine optimal parameters
of the q-out-of-m procedure [17] in large scale networks for different fractions of Byzantine
sensors. More than two types of sensors were assumed in [9], [10]. The authors thereof proposed
1
In practice, to manipulate the data of a sensor, an adversary may attack the sensor node itself or break the communication
channel between the sensor and the fusion center. In this paper, we do not distinguish these two approaches.
January 22, 2018
DRAFT
4
a maximum likelihood procedure, which is based on the iterative expectation maximization
(EM) algorithm [18], simultaneously classifying the sensor nodes and performing the hypothesis
testing. The authors in [11] showed that the optimal detector is of a threshold structure when
the fraction of Byzantine sensors is less than 0.5. A zero-sum game was formulated in each
of [12]–[14], among which a closed-form equilibrium point of attack strategy and detector was
obtained in [14], computation efficient and nearly optimal equilibrium point (exact equilibrium
point only in certain cases) was obtained in [12], and numerical simulations were used to study
the equilibrium point in [13].
While in [5]–[10], [13] the Byzantine sensors are assumed to generate malicious data independently, this work, as in [11], [12], [14], assumes that the Byzantine sensors may collude
with each other. The collusion model is more reasonable since the attacker is malicious and
will arbitrarily change the messages of the sensors it controlls. Notice also that compared to the
independence model, the collusion model complicates the analysis significantly. Unlike [6]–[10],
[12], [13], where the sensors only send binary messages, this work, as in [5], [11], [14], assumes
that the measurements of a benign sensor can take any value. Since the binary message model
simplifies the structure of corrupted measurements, and, hence, implicitly limits the capability of
an attacker, it is easier to be dealt with. This work differs from [11], [14] as follows. The authors
in [11] focused on one time step scenario. The analysis is thus fundamentally different and more
challenging. On the contrary, in this work the hypothesis testing is performed sequentially and
an asymptotic regime performance metric, i.e., the Chernoff information, is concerned. A similar
setting as in this work was considered in our recent work [14]. However, [14] focused on the
equilibrium point. The performance (i.e., the security and efficiency) of the obtained equilibrium
detection rule is merely one point of the admissible set that will be characterized in this paper.
Finally, we should remark that the aforementioned literature mainly focuses on designing
algorithms in adversarial environment. However, those algorithms may perform poorly in the
absence of the adversary comparing to the classic Neyman-Pearson detector or Naive Bayes
detector. A fundamental question, which we seek to answer in this paper, is that how to design
a detection strategy which performs “optimally” regardless of whether the attacker is present.
Organization: In Section II, we formulate the problem of binary hypothesis testing in adversarial environments, in which the attack model, the performance indices and the notion of the
efficiency and security are defined. For the sake of completeness, we give a brief introduction
the large deviation theory in Section III, which is a key supporting technique for the later
January 22, 2018
DRAFT
5
analysis. The main results are presented in Section IV. We first provide the fundamental limits
of the trade-off between the efficiency and security. We then propose detectors that achieve these
limits. At last, we show that the maximal efficiency and the maximal security can be achieved
simultaneously in some cases. Two extensions are investigated in Section V. After providing
numerical examples in Section VI, we conclude the paper in Section VII.
Notations: R (R+ ) is the set of (nonnegative) real numbers. Z+ is the set of positive integers.
The cardinality of a finite set I is denoted as |I|. For a set A ∈ Rn , int(A) denotes its interior.
Pk
For any sequence {x(k)}∞
k=1 , we denote its average at time k as x̄(k) ,
t=1 x(t)/k. For a
vector x ∈ Rn , the support of x, denoted by supp(x), is the set of indices of nonzero elements:
supp(x) , {i ∈ {1, 2, . . . , n} : xi 6= 0}.
II. P ROBLEM F ORMULATION
Consider the problem of detecting a binary state θ ∈ {0, 1} using m sensors’ measurements.
Define the measurement y(k) at time k to be a row vector:
h
i
y(k) , y1 (k) y2 (k) · · · ym (k) ∈ Rm ,
(1)
as a vector of all measurements from time 1 to time k:
h
i
Y(k) , y(1) y(2) · · · y(k) ∈ Rmk .
(2)
where yi (k) is the scalar measurement from sensor i at time k. For simplicity, we define Y(k)
Given θ, we assume that all measurements {yi (k)}i=1,...,m, k=1,2,... are independent and identically
distributed (i.i.d.). The probability measure generated by yi (k) is denoted as ν when θ = 0 and it
is denoted as µ when θ = 1. In other words, for any Borel-measurable set A ⊆ R, the probability
that yi (k) ∈ A equals ν(A) when θ = 0 and equals µ(A) when θ = 1. We denote the probability
space generated by all measurements y(1), y(2), . . . as (Ωy , Fy , Poθ )2 , where for any l ≥ 1
2
Poθ (yi1 (k1 ) ∈ A1 , . . . , yil (kl ) ∈ Al )
ν(A1 )ν(A2 ) . . . ν(Al )
=
µ(A1 )µ(A2) . . . µ(Al )
if θ = 0
,
if θ = 1
The superscript “o” stands for original, which is contrasted with corrupted measurements.
January 22, 2018
DRAFT
6
when (ij , kj ) 6= (ij ′ , kj ′ ) for all j 6= j ′ . The expectation taken with respect to Poθ is denoted by
Eoθ . We further assume that ν and µ are absolutely continuous with respect to each other. Hence,
the log-likelihood ratio λ : R → R of yi (k) is well defined as
dµ
(yi ) ,
λ(yi ) , log
dν
(3)
where dµ/dν is the Radon-Nikodym derivative.
We define fk : Rmk → [0, 1], the detector at time k, as a mapping from the measurement
space Y(k) to the interval [0, 1]. When fk (Y(k)) = 0, the system makes a decision θ̂ = 0, and
when fk (Y(k)) = 1, θ̂ = 1. When fk (Y(k)) = γ ∈ (0, 1), the system then “flips a biased coin”
to choose θ̂ = 1 with probability γ and θ̂ = 0 with probability 1 − γ. The system’s strategy
f , (f1 , f2 , · · · ) is defined as an infinite sequence of detectors from time 1 to ∞.
A. Attack Model
Let the (manipulated) measurements received by the fusion center at time k be
y′ (k) = y(k) + ya (k),
(4)
where ya (k) ∈ Rm is the bias vector injected by the attacker at time k. In the following,
Assumptions 1–3 are made on the attacker, among which Assumption 1 is in essence the only
limitation we pose.
Assumption 1 (Spare Attack). There exists an index set I ⊂ M , {1, 2, . . . , m} with |I| = n
S
a
such that ∞
k=1 supp(y (k)) = I. Furthermore, the system knows the number n, but it does not
know the set I.
We should remark that the above assumption does not pose any restrictions on the value of
yia (k) if sensor i is compromised at time k, i.e., the bias injected into the data of a compromised
sensor can be arbitrary.
Assumption 1 says that the attacker can compromise up to n out of m sensors at each time. It is
practical to assume that the attacker possesses limited resources, i.e., the number of compromised
sensors is (non-trivially) upper bounded, since otherwise it would be too pessimistic and the
problem becomes trivial. The quantity n might be determined by the a priori knowledge about
the quality of each sensor. Alternatively, the quantity n may be viewed as a design parameter,
which indicates the resilience level that the system is willing to introduce; the details of which are
January 22, 2018
DRAFT
7
in Remark 1. Notice also that since the worst-case attacks (over the set of compromised sensors
and the attack strategy) are concerned (the performance metric will be introduced shortly), it is
equivalent to replace the cardinality requirement |I| = n by |I| ≤ n. We should note that in [6],
[8], [12], it was also assumed that the number/fraction of malicious sensor nodes is known to
the system operator.
Moreover, the set of compromised sensors is assumed to be fixed over time. Notice that if
we assume that the set of compromised sensors has a fixed cardinality but is time-varying, i.e.,
there exists no a set like I to bound the compromised sensors, the attacker would be required
to abandon the sensor nodes it has compromised, which is not sensible. Notice that in [8]–
[10], it was assumed the set of malicious/misbehaving sensors is fixed as well. We should also
note that though this work is concerned with asymptotic performances (i.e., the security and
efficiency introduced later), the numerical simulations in Section VI show that our algorithm
indeed perform quite well in an non-asymptotic setup. Actually, if a finite-time horizon problem
is considered and the time required for an attacker to control a benign sensor is large enough,
then it is reasonable to assume that the set of compromised sensors is fixed.
In fact, the exactly same sparse attack model as in Assumption 1 has been widely adopted by
literature dealing with Byzantine sensors, e.g., state estimation [19], [20], and quickest change
detection [21].
Finally, we should note that we do not assume any pattern of the bias yia (k) for i ∈ I, i.e., the
malicious bias injected may be correlated across the compromised sensors and correlated over
time. Compared to the independence assumption in [5]–[10], [13], our assumption improves the
effectiveness of the attacker and is more realistic in the sense that the attacker is malicious and
will do whatever it wants.
Remark 1. The parameter n can also be interpreted as how many bad sensors the system
can and is willing to tolerate, which is a design parameter for the system operator. In general,
increasing n will increase the resilience of the detector under attack. However, as is shown in
the rest of the paper, a larger n may result in more conservative design and is likely to cause
a performance degradation during normal operation when no sensor is compromised.
Assumption 2 (Model Knowledge). The attacker knows the probability measure µ and ν and
the true state θ.
January 22, 2018
DRAFT
8
By the knowledge about the sensor, the attacker can develop the probability measure µ and
ν. To obtain the true state, the attacker may deploy its own sensor network. Though it might
be difficult to satisfy in practice, this assumption is in fact conventional in literature concerning
the worst-case attacks, e.g., [5], [21]. Nevertheless, this assumption is in accordance with the
Kerckhoffs’s principle.
Assumption 3 (Measurement Knowledge). At time k, the attacker knows the current and all the
historical measurements available at the compromised sensors.
Since the attacker knows the true measurement of a compromised sensor i, yi (k), it may set
the fake measurement arrived at the fusion center yi′ (k) to any value it wants by injecting yia (k).
One may also verify that all the results in this paper remain even if the attacker is “strong”
enough where at time k, it knows measurements from all the sensors Y(k).
An admissible attack strategy is any causal mapping from the attacker’s available information
to a bias vector that satisfies Assumption 1. This is formalized as follows. Let I = {i1 , i2 , . . . , in }.
Define the true measurements of the compromised sensors from time 1 to k as
h
i
YI (k) , yI (1) yI (2) · · · yI (k) ∈ R|I|k
with
i
h
yI (k) , yi1 (k) yi2 (k) · · · yin (k) ∈ R|I| .
Similar to Y(k), Y ′ (k) (Y a (k)) is defined as all the manipulated (bias vector) from time 1 to
k. The bias ya (k) is chosen as a function of the attacker’s available information at time k:
ya (k) , g(YI (k), Y a (k − 1), I, θ, k),
(5)
where g is a function3 of YI (k), Y a(k − 1), I, θ, k such that ya (k) satisfies Assumption 1. We
denote g as an admissible attacker’s strategy. Notice that since time k is an input variable and the
available measurements YI (k), Y a (k − 1) are “increasing” with respect to time k, the definition
in (5) does not exclude the time-varying attack strategy. Denote the probability space generated
by all manipulated measurements y′ (1), y′ (2), . . . as (Ω, F , Pθ ). The expectation taken with
respect to the probability measure Pθ is denoted by Eθ .
3
The function g is possibly random. For example, given the available information, the adversary can flip a coin to decide
whether to change the measurement or not.
January 22, 2018
DRAFT
9
B. Asymptotic Detection Performance
Given the strategy of the system and the attacker, the probability of error at time k can be
defined as
e(θ, I, k) ,
E0 fk (Y ′ (k))
1 − E1 fk (Y ′ (k))
when θ = 0,
(6)
when θ = 1.
Notice that fk could take any value from [0, 1]. Hence, the expected value of fk is used to
compute the probability of error. In this paper, we are concerned with the worst-case scenario.
As a result, let us define
ǫ(k) ,
max
θ=0,1,|I|=n
e(θ, I, k).
(7)
In other words, ǫ(k) indicates the worst-case probability of error considering all possible sets
of compromised sensors and the state θ given the detection rule f and attack strategy g. Notice
also that in accordance with Assumption 1, the set I in the above equation is fixed over time.
Ideally, for each time k, the system wants to design a detector fk to minimize ǫ(k). However,
such a task can hardly be accomplished analytically since the computation of the probability of
error usually involves numerical integration. Thus, in this article, we consider the asymptotic
detection performance in hope to provide more insight on the detector design. Define the rate
function as
ρ , lim inf −
k→∞
log ǫ(k)
.
k
(8)
Clearly, ρ is a function of both the system strategy f and the attacker’s strategy g. As such,
we will write ρ as ρ(f, g) to indicate such relations. Since ρ indicates the rate with which the
probability of error goes to zero, the system would like to maximize ρ in order to minimize the
detection error. On the contrary, the attacker wants to decrease ρ to increase the detection error.
C. Interested Problems
In practice, the attacker may not be present consistently. As a result, the system will operate
for an extended period of time with all sensors being benign. Thus, a natural question arises: is
there any detection rule that has “decent” performance regardless of the presence of the attacker?
Or is there a fundamental trade-off between security and efficiency? In other words, a detector
that is “good” in the presence of an adversary will be “bad” in a benign environment. This paper
is devoted to answering this question.
January 22, 2018
DRAFT
10
Informally, the performance of a detection rule when there is no attacker at all is referred
to by “efficiency”, while the performance when the worst-case attacker (provided that the
attacker knows the detection rule used by the system) is present is referred to by “security”.
Mathematically speaking, given a system strategy f , denote by E(f ) and S (f ) its efficiency and
security respectively, which are formalized as follows:
E(f ) ,ρ(f, g = 0),
(9)
S (f ) , inf ρ(f, g)
(10)
g
where 0 ∈ Rm is the zero vector.
III. P RELIMINARY: L ARGE D EVIATION T HEORY
In this section, we introduce the large deviation theory, which is a key supporting technique
of this paper.
To proceed, we first introduce some definitions. Let Mω (w) ,
R
Rd
ew·X dω(X), w ∈ Rd be the
moment generating function for the random vector X ∈ Rd that has the probability measure ω,
where w · X is the dot product. Let domω , {w ∈ Rd |Mω (w) < ∞} be the support such that
Mω (w) is finite. Define the Fenchel–Legendre transform of the function log Mω (w) as
Iω (x) = sup {x · w − log Mω (w)}, x ∈ Rd .
(11)
w∈Rd
Theorem 1 (Multidimensional Cramér’s Theorem [22]). Suppose X(1), . . . , X(k), . . . be a
sequence of i.i.d. random vectors and X(k) ∈ Rd has the probability measure ω. Let X(k) ,
Pk
t=1 X(t)/k, k ∈ Z+ be the empirical mean. Then if 0 ∈ int(domω ), the probability P(X(k) ∈
A) with A ⊆ Rd satisfies the large deviation principle, i.e.,
1) if A is closed,
lim sup
k→∞
1
log P(X(k) ∈ A) ≤ − inf Iω (x).
x∈A
k
2) if A is open,
lim inf
k→∞
January 22, 2018
1
log P(X(k) ∈ A) ≥ − inf Iω (x).
x∈A
k
DRAFT
11
IV. M AIN R ESULTS
A. Technical Preliminaries
Denote the moment generating function of the log-likelihood ratio λ under each hypothesis
as:
M0 (w) ,
M1 (w) ,
Z
∞
y=−∞
Z ∞
exp(wλ(y))dν(y),
(12)
exp(wλ(y))dµ(y).
(13)
y=−∞
Furthermore, define dom0 as the region where M0 (w) is finite and I0 (x) as the Fenchel–Legendre
transform of log M0 (w). The quantities M1 (w), dom1 and I1 (x) are defined similarly.
Denote the the Kullback-Leibler (K–L) divergences by
Z ∞
Z
D(1k0) ,
λ(y)dµ, D(0k1) , −
y=−∞
∞
λ(y)dν.
y=−∞
To apply the multidimensional Cramér’s Theorem and avoid degenerate problems, we adopt the
following assumptions:
Assumption 4. 0 ∈ int(dom0 ) and 0 ∈ int(dom1 ).
Assumption 5. The K–L divergences are well-defined, i.e., 0 < D(1k0), D(0k1) < ∞.
With the above assumptions, we have the following properties of I0 (x) and I1 (x). the proof
of which is provided in Appendix A.
Theorem 2. Under Assumptions 4 and 5, the followings hold:
1) I0 (x) (I1 (x)) is twice differentiable, strictly convex and strictly increasing (strictly decreasing) on [−D(0k1), D(1k0)].
2) The following equalities hold:
I1 (D(1k0)) = 0,
(14)
I0 (D(1k0)) = D(1k0),
(15)
I0 (−D(0k1)) = 0,
(16)
I1 (−D(0k1)) = D(0k1).
(17)
I0 (0) = I1 (0).
January 22, 2018
(18)
DRAFT
12
1.8
1.6
I0 (D(1||0)) = D(1||0)
I0 (x)
I1 (x)
1.4
1.2
1
0.8
I1 (−D(0||1)) = D(0||1)
0.6
0.4
I0 (0) = I1 (0) = C
0.2
0
−1
−0.5
0
0.5
1
1.5
2
Fig. 1: Illustration of I0 (x) and I1 (x). The figure is plotted by assuming y1 (1) to be Bernoulli
distributed under both hypotheses with Po0 (y1 (1) = 1) = 0.02 and Po1 (y1 (1) = 1) = 0.6.
Since I0 (0) = I1 (0), let us define
C , I0 (0).
(19)
To make the presentation clear, we illustrate I0 (x) and I1 (x) in Fig. 1.
The “inverse functions” of I0 (x) and I1 (x) are defined as follows: for z ≥ 0,
I0−1 (z) = max{x ∈ R : I0 (x) = z},
I1−1 (z) = min{x ∈ R : I1 (x) = z}.
Let Dmin , min{D(0k1), D(1k0)}. We further define h(z) : (0, (m − n)Dmin ) 7→ (0, (m − n)Dmin )
as
h(z)
,(m − n) min{I0 (I1−1 (z/(m − n)), I1 (I0−1 (z/(m − n))}.
January 22, 2018
(20)
DRAFT
13
B. Fundamental Limits
We are ready to provide the fundamental limitations between efficiency and security. The
proof is provided in Appendix B.
Theorem 3. For any detection rule f , the following statements on E(f ) and S (f ) are true:
1) E(f ) ≤ mC,
2) S (f ) ≤ (m − 2n)+ C, where (m − 2n)+ = max{0, m − 2n}.
3) S (f ) ≤ E(f ),
4) Let E(f ) = z, we have
S (f ) ≤
(
h(z)
if 0 < z < (m − n)Dmin
(21a)
0
if z ≥ (m − n)Dmin .
(21b)
Remark 2. Theorem 3 indicates that mC is the maximum efficiency that can be achieved by
any detector, while (m − 2n)+ C is the maximum security that can be achieved. Therefore, if
m ≤ 2n, i.e., no less that half of the sensors are compromised, then S (f ) = 0 for any f , which
implies that all detectors will have zero security. In that case, the naive Bayes detector will be
the optimal choice since it has the optimal efficiency and the analysis becomes trivial. Therefore,
without any further notice, we assume m > 2n for the rest of the paper.
Notice that fourth constraint in Theorem 3 indicates a trade-off between security and efficiency.
For general cases, the maximum security and efficiency may not be achieved simultaneously.
However, in Section IV-D, we will prove that for a special case, there exist detectors that can
achieve the maximum security and efficiency at the same time.
Notice that I0 (x) (I1 (x)) is strictly increasing (decreasing) on [−D(0k1), D(1k0)]. Therefore,
combining (15) and (17), one obtains that h(z) is strictly decreasing. Then the dual version
of (21a) is obtained as follows. Let S (f ) = z we have that if 0 < z ≤ (m − 2n)C,
E(f ) ≤ h−1 (z) = h(z),
(22)
where h−1 (z) is the inverse function of h(z), and the equality holds because h(z) is an involutory
function, i.e., h(h(z)) = z for every z ∈ (0, (m − n)Dmin ).
We then have the following two corollaries. The results follow straightforwardly from Theorem 3 and (22), we thus omit the proofs.
January 22, 2018
DRAFT
14
Corollary 1. Suppose the security of a detector f satisfies
S (f ) = z ∈ [0, (m − 2n)C],
then the maximum efficiency of f satisfies the following inequality:
mC
if z = 0
max E(f ) ≤
,
f ∈{f :S (f )=z}
he (z) if z > 0
where he (z) , min{mC, h(z)}.
Corollary 2. Suppose the efficiency of a detector f satisfies
E(f ) = z ∈ [0, mC],
then the maximum security of f satisfies the following inequality:
hs (z) if 0 < z < z ′ ,
max
S (f ) ≤
f ∈{f :E(f )=z}
0
if z ≥ z ′ or z = 0,
where hs (z) = min{z, (m − 2n)C, h(z)}, and z ′ = (m − n)Dmin .
C. Achievability
In this section, we propose a detector that achieves the upper bounds in Corollaries 1 and 2.
Let zs ≤ (m−2n)C. At time k ≥ 1, the algorithm, denoted by fz∗s , is implemented as follows.
Remark 3. We here discuss about the computational complexity of the detection rule fz∗s . The
computational complexity for the step 1 is O(m). Notice that the quantity λ̄i (k) is computed in a
recursive fashion. The complexity for the step 2 is O(m log m). To compute δ(0, k) and δ(1, k),
we can first sort I0 (λ̄i (k)) and I1 (λ̄i (k)) in ascending order, respectively, and then sum the first
m − 2n elements of each. The computational complexity for the step 3 and step 4 is fixed, and
the step 5 has computational complexity O(m). Therefore, the total computational complexity
for each time step is O(m log m).
We now show the performance of fz∗s and the proof is provided in Appendix C.
January 22, 2018
DRAFT
15
Algorithm 1 Hypothesis testing algorithm fz∗s
1: Compute the empirical mean of the likelihood ratio from time 1 to time k for each sensor i:
λ̄i (k) ,
k
X
λ(yi′ (t))/k
t=1
1
k−1
λ̄i (k − 1) + λ(yi′ (k))
=
k
k
(23)
with λ̄i (0) = 0.
2: Compute I0 (λ̄i (k)) and I1 (λ̄i (k)) for each i. Compute the following sum:
X
δ(0, k) ,
min
I0 (λ̄i (k)),
|O|=m−n,O⊂M
δ(1, k) ,
min
|O|=m−n,O⊂M
i∈O
X
I1 (λ̄i (k)).
i∈O
3: If δ(0, k) < zs , make a decision θ̂ = 0; go to the next step otherwise.
4: If δ(1, k) < zs , make a decision θ̂ = 1; go to the next step otherwise.
P
5: If m
i=1 λ̄i (k) < 0, make a decision θ̂ = 0; make a decision θ̂ = 1 otherwise.
Definition 1. (ze , zs ) are called an admissible pair if the following inequalities holds:
0 ≤zs ≤ (m − 2n)C,
mC
if zs = 0
ze ≤
,
he (zs ) if zs > 0
where he (zs ) is defined in Corollary 1.
Theorem 4. Let (ze , zs ) be any admissible pair of efficiency and security. Then there holds
E(fz∗s ) ≥ ze ,
S (fz∗s ) ≥ zs .
The above theorem means that the upper bounds in Corollaries 1 and 2 are achieved by
fz∗s . Hence, we provide a tight characterization on admissible efficiency and security pair. We
illustrate the shape of admissible region in Fig 2.
Remark 4. The optimal detector may not be necessarily unique, in the sense that there may exist
other detectors, other than the one defined by Algorithm 1, that can achieve the same efficiency
January 22, 2018
DRAFT
16
3
2.5
((m − n)C, (m − n)C)
Security
2
(m − 2n)C
1.5
1
0.5
0
mC
0
0.5
1
1.5
2
Efficiency
2.5
3
Fig. 2: Achievable efficiency and security region for any detector. The figure is plotted by
assuming y1 (1) to be Bernoulli distributed under both hypotheses with Po0 (y1 (1) = 1) = 0.02
and Po1 (y1 (1) = 1) = 0.6. The shaded area is the admissible pair (E(f ), S (f )) for any detector
f . The red dashed line is the function (m − n)I0 (I1−1 (z/(m − n))), while the blue dotted line
the function (m − n)I1 (I0−1 (z/(m − n))).
and security limits. By definition, the detectors achieving the limits have the same asymptotic
performance. However, the finite-time performance (in terms of detection error) may be different
and we are planning to investigate this in the future work.
D. Special Case: Symmetric Distribution
In this subsection, we discuss a case where the maximum security and efficiency can be
achieved simultaneously by a detector.
Notice that by the definition of admissible pair, if we have
h((m − 2n)C) ≥ mC,
(24)
then we know that (ze = mC, zs = (m − 2n)C) is an admissible pair and hence the detector
∗
f(m−2n)C
defined in Section IV-C can achieve maximum security (m − 2n)C and efficiency
January 22, 2018
DRAFT
17
mC simultaneously. In other words, adding security will not deteriorate the performance of the
system in the absence of the adversary.
The following theorem provides a sufficient condition for (24), which is based on the first
order derivative of I0 (·) and I1 (·). The proof is presented in Appendix D for the sake of legibility.
(1)
(1)
∗
Theorem 5. If I0 (x)|x=0 = −I1 (x)|x=0 , then h((m−2n)C) ≥ mC holds. Therefore, f(m−2n)C
possesses not only the maximal security but also the maximal efficiency.
Notice that whether or not the above sufficient condition is satisfied merely depends on the
probability distribution of the original observations, which is independent of the number of the
compromised sensors.
If there exists “symmetry” between distribution µ and ν, then the sufficient condition can be
satisfied. To be specific, if there exists a constant a such that for any Borel measurable set A,
we have
µ(a + A) = ν(a − A),
then one can prove that
M0 (w) = M0 (−w),
which further implies that
(1)
(1)
I0 (x) = I1 (−x) ⇒ I0 (x)|x=0 = −I1 (x)|x=0 .
We provide two examples of pairs of “symmetric” distributions as follows:
1) Each yi (k) is i.i.d. Bernoulli distributed and
θ
with probability p0
yi (k) =
1 − θ with probability 1 − p0
2) Each yi (k) satisfies the following equation:
yi (k) = aθ + vi (k),
where a 6= 0 and vi (k) ∼ N (v̄, σ 2 ) is i.i.d. Gaussian distributed.
V. E XTENSION
In this section, we consider two extensions to the problem settings discussed in Section IV.
January 22, 2018
DRAFT
18
A. Secure Sensors
Consider that there is a subset of “secure” sensors that are well protected and cannot be
compromised by the attacker. We would like to study the trade-off between security and efficiency
when those “secure” sensors are deployed.
Let ms out of the total m sensors be “secure” and the remaining m − ms sensors are “normal”
ones that can be compromised by an adversary. In this subsection, n can take any value in
{0, 1, . . . , m − ms } and does not necessarily satisfy 2n < m. The other settings are the same as
in Section II. Denote by Es (f ) and Ss (f ) the efficiency and security of a detection rule f under
such case.
Then one obtains the following results as in Theorem 3.
Theorem 6. For any detection rule f , the following statements on Es (f ) and Ss (f ) are true:
1) Es (f ) ≤ mC,
2) Ss (f ) ≤ max((m − 2n)C, ms C),
3) Ss (f ) ≤ Es (f ),
4) Let Es (f ) = z, we have
Ss (f ) ≤
h(z)
0
if 0 < z < (m − n)Dmin
.
if z ≥ (m − n)Dmin
The above theorem is proved in the same manner as in Appendix B. Notice that the essential
difference is the range of Ss (f ), i.e., the statement in the second bullet. This is due to the fact
that the ms secure sensors cannot be compromised.
Remark 5. From the above theorem, one sees that replacing ms normal sensors with secure
sensors does not change the fundamental trade-off between the security and efficiency. However,
the benefit of these ms secure sensors are that the security itself is improved when 2n > m−ms .
Also, one notice that when m − ms ≥ 2n, there are no gains of deploying secure sensors.
Intuitively, in such case the redundancy of the m − ms normal sensors is enough.
Furthermore, the detector fzss in Algorithm 2, which is a slight variation of fz∗s and treats the
ms secure sensors separately, achieves the limits. This is stated in the following theorem, which
is proved in the same manner as in Appendix C.
January 22, 2018
DRAFT
19
Theorem 7. If the pair (ze , zs ) satisfies
0 ≤zs ≤ max((m − 2n)C, ms C),
mC
if zs = 0
ze ≤
,
he (zs ) if zs > 0
then there holds
Es (fzss ) ≥ ze ,
Ss (fzss ) ≥ zs .
Algorithm 2 Hypothesis testing algorithm fzss when there are secure sensors
1: Compute λ̄i (k) for each sensor.
2: Compute I0 (λ̄i (k)) and I1 (λ̄i (k)) for each sensor. Compute the minimum sum from the
“normal” sensors:
δ(0, k) =
δ(1, k) =
3: If δ(0, k) +
Pi=m
min
|O|=m−ms −n,O⊂{1,...,m−ms }
min
|O|=m−ms −n,O⊂{1,...,m−ms }
m−ms +1 I0 (λ̄i (k))
X
I0 (λ̄i (k)),
i∈O
X
I1 (λ̄i (k)).
i∈O
< zs , make a decision θ̂ = 0; go to the next step otherwise.
Pi=m
4: If δ(1, k) + m−ms +1 I1 (λ̄i (k)) < zs , make a decision θ̂ = 1; go to the next step otherwise.
P
5: If m
i=1 λ̄i (k) < 0, make a decision θ̂ = 0; make a decision θ̂ = 1 otherwise.
B. Unknown Number of Compromised Sensors
In the previous section, we assume that if the system is being attacked, then n sensors are
compromised. However, in practice, the exact number of compromised sensors is likely to
be unknown. In this subsection, we assume that we know an estimated upper bound on the
compromised sensors, denoted by n. Let na denote the number of the sensors that are actually
compromised. Therefore, na may take value in Na , {0, 1, 2, . . . n}4 .
4
In Section II-A, we remark that the requirement |I| = n can be equivalently replaced by |I| ≤ n. The implicit assumption
is that the estimated upper bound n is tight and the worst-case number of compromised sensors is in indeed n. Therefore, na
in this section may also be interpreted as the tight upper bound of the number of actually compromised senors.
January 22, 2018
DRAFT
20
Given a detector f , denote by Dna (f ) the detection performance when the number of compromised sensor is na . Then, one has D0 (f ) = E(f ) and Dn (f ) = S (f ). In the following, we
present the pairwise trade-off between Dna (f ) and Dn′a (f ) for any 0 ≤ na , n′a ≤ n, and propose
an algorithm to achieve these performance limits. A similar argument as in Section IV can be
adopted to obtain these results, the details of which are omitted.
We define h̃ : Na × Na × (0, ∞) 7→ (0, ∞) as
h̃(na , n′a , z)
n
o
,(m − ña ) min h̃0 (z/(m − ña )), h̃1 (z/(m − ña )) ,
where ña = na + n′a , and
h̃0 (z) =
h̃1 (z) =
I0 (I −1 (z))
1
0
I1 (I −1 (z))
0
0
if 0 < z < D(0||1)
,
if z ≥ D(0||1)
if 0 < z < D(1||0)
.
if z ≥ D(1||0)
Then one obtains that for any detector f and na , n′a ∈ N, there hold
Dna (f ) ≤ (m − 2na )C,
(25)
Dna (f ) ≤ h̃ na , n′a , Dn′a (f ) .
(26)
Let z , (z0 , z1 , . . . , zn ) be a n-tuplet of admissible detection performance, i.e.,
zna ≤ (m − 2na )C,
zna ≤ h̃(na , n′a , zn′a ).
Then the detector in Algorithm 3, which is a variation of fz∗s in Section IV-C and is denoted by
fz∗ , can achieve these performance, i.e., Dna (fz∗ ) ≥ zna for any na ∈ N.
VI. N UMERICAL E XAMPLES
A. Asymptotic Performance
We simulate the performance of the detector fz∗s proposed in Section IV-C (i.e., its efficiency
and security) and compare the empirical results to the theoretical ones shown in Fig. 2.
The same parameters as in Fig. 2 are used, i.e., Po0 (y1 (1) = 1) = 0.02, Po1 (y1 (1) = 1) = 0.6,
m = 9 and n = 2. To simulate the security, it is assumed that the following attack strategy is
January 22, 2018
DRAFT
21
Algorithm 3 Hypothesis testing algorithm fz∗
initialization: na = n.
1: Compute λ̄i (k), I0 (λ̄i (k)), I1 (λ̄i (k)) for each sensor i.
2: While na ≥ 1
1) Compute these two minima:
δ̃(0, k, na ) ,
δ̃(1, k, na ) ,
min
|O|=m−na ,O⊂M
min
|O|=m−na ,O⊂M
X
I0 (λ̄i (k)),
i∈O
X
I1 (λ̄i (k)).
i∈O
2) If δ̃(0, k, na ) < zna , make a decision θ̂ = 0 and stop.
3) If δ̃(1, k, na ) < zna , make a decision θ̂ = 1 and stop.
4) Replace na with na − 1.
P
3: If m
i=1 λ̄i (k) < 0, make a decision θ̂ = 0; make a decision θ̂ = 1 otherwise.
adopted. If θ = 0, the attacker modifies the observations of the compromised sensors such that
for every i ∈ I and k ≥ 1
I0 (λ̄i (k)) ≥ zs .
On the other hand, if θ = 1, the attack strategy is such that I1 (λ̄i (k)) ≥ zs holds for every i ∈ I
and k ≥ 1.
To simulate the performance with high accuracy, we adopt the importance sampling approach [23]. To plot Fig. 3, we let zs be in [0, (m − 2n)C = 1.5987]. Notice that the theoretical
performance of fz∗s coincides exactly with the fundamental limits in Fig. 2. Therefore, Fig. 3
verifies that our algorithm fz∗s indeed achieves the fundamental limits.
B. Non-asymptotic Performance
We have proved that our algorithm is optimal in the sense that it achieves the fundamental
trade-off between the security and efficiency. However, notice that both the security and efficiency
are asymptotic performance metrics. In this example, we show that our algorithm possesses quite
“nice” finite-time performance as well by comparing it to the naive Bayes detector. We should
remark that while the Bayes detector is strictly optimal (i.e., optimal for any time horizon) in
the absence of attackers, its security is zero. The results are in Fig. 4, where zs is chosen to
January 22, 2018
DRAFT
22
1.6
1.4
Security
1.2
1
theoretical
empirical
0.8
0.6
0.4
0.2
0
2.72 2.74 2.76 2.78 2.8 2.82 2.84 2.86 2.88 2.9
Efficiency
Fig. 3: Comparison between the empirical and theoretical performance of the detector fz∗s when
zs ∈ [0, (m − 2n)C].
be 1.4282. Fig. 4 illustrates that the algorithm fz∗s with zs = 1.4282 has a finite-time detection
performance comparable to that of naive Bayes detector when the attacker is absent. The finitetime performance metric ǫ(k) is defined in (7), where the attacker is absent, i.e., g = 0, and the
detector is fz∗s =1.4282 or the naive Bayes. One should note that the security of fz∗s is 1.4282. As
a result, adopting the secure detector fz∗s increases the security of the system while introducing
minimum performance loss in the absence of the adversary.
C. Comparison with Other Detectors
To simulate the detectors introduced later, we use the same sensor network model as in
Fig. 3. The asymptotic performances, i.e., the efficiency and security, are summarized in Table I,
while the non-asymptotic performances when the attacker is absent are in Fig. 5. Table I is
consistent with the statement that our algorithm achieves the best trade-off between the security
and efficiency, while Fig. 5 shows that it is preferable to adopt our algorithm as well with respect
to the finite-time performance when the attacker is absent. In the following, we present the two
detectors to be compared to in detail.
The first detector is the equilibrium detection rule that is proposed in [14] for cases where
January 22, 2018
DRAFT
23
Naive Bayes
∗
f1.4282
10−2
10−4
ǫ(k)
10−6
10−8
10−10
10−12
10−14
1
2
3
4
5
6
time k
7
8
9
10
Fig. 4: Finite-time performance of fz∗s in the absence of the adversary.
m > 2n. This detection rule, which shares the same spirit with the α-trimmed mean in robust
statistics [15], first removes the largest n and smallest n log-likelihood ratios, and then compares
the mean of the remaining m − 2n log-likelihood ratios to 0, just as in the classic probability
ratio test. The details of the detection rule, denoted by ftrim , are formalized as follows.
0 if Pm−n λ̄[i] (k) < 0,
i=n+1
ftrim (Y ′ (k)) =
P
m−n
1 if
i=n+1 λ̄[i] (k) ≥ 0,
where λ̄[i] (k) is the i-th smallest element of {λ̄1 (k), λ̄2 (k), . . . , λ̄m (k)} with λ̄i (k) being the
empirical mean of log-likelihood ratio from time 1 to k for senor i, which is defined in (23).
It was shown in [14] that the security and efficiency of ftrim are
S (ftrim ) = (m − 2n)C,
E(ftrim ) = (m − n)C.
Since for any z < C, there hold
I0 (I1−1 (z)) > C,
I1 (I0−1 (z)) > C.
Then by the definition of h(z) and Theorem 4, one obtains that if the security of our algorithm
is (m − 2n)C, its efficiency is larger than (m − n)C, i.e.,
E(fz∗s =(m−2n)C ) > (m − n)C.
January 22, 2018
DRAFT
24
Therefore, our algorithm is preferable since, with the same security, it achieves the larger
efficiency than the algorithm ftrim . In particular, by Theorem 5, the efficiency gain of our
algorithm in certain cases is nC.
The next detector is the q-out-of-m procedure [17], which has been studied in [8] by assuming
that the malicious sensor nodes generate fictitious data randomly and independently, and the
probability that the compromised sensor flips the binary message is known. The q-out-of-m
procedure is simple and works as follows. At time k, after receiving the mk (binary) messages,
the fusion center makes a decision
1 if Pk Pm y ′ (k) ≥ qk ,
i=1 i
t=1
θ̂ =
0 otherwise.
(27)
Let q = [q1 , . . . , qk , . . .] be a sequence of thresholds used in the above detector from time 1 to
infinity. In the sequel, we denote the above detector as fqom (q). Notice that fqom (q) is just the
naive Bayesian detector, which minimizes the weighted sum of miss detection and false alarm
at each time k (the weight is determined by qk ). It is clear that if fqom (q) is used at the fusion
center, the worst-case attack is always sending 0 (1) if the true state θ is 1 (0). Therefore, at
time k, the performance (i.e., the probability of detection error) of the detector fqom (q) under
the worst-case attacks is as follows.
mk − nk
pj1 (1 − p1 )mk−nk−j ,
P1 (fqom (q) = 0) =
j
j=0
qk
X
P0 (fqom (q) = 1)
mk − nk
pj0 (1 − p0 )mk−nk−j ,
=
j
j=max(0,qk −nk)
mk−nk
X
where p0 , Po0 (y1 (1) = 1) = 0.02, p1 , Po1 (y1 (1) = 1) = 0.6. Then it is reasonable to set
nk < qk < mk − nk, since otherwise the worst-case (over θ) detection error will be 1. However
it is challenging to obtain the optimal qk analytically to minimize the worst-case detection error;
we do this by brute-force numerical simulations. By varying the time k from 1 to 40, we obtain
∗
the (approximate) security and the optimal parameters q1∗ , . . . , q40
. Then we further simulate the
performance of the q-out-m algorithm when the optimal parameters obtained above are used and
the attacker is absent.
January 22, 2018
DRAFT
25
TABLE I: The asymptotic performances of our algorithm fz∗s with zs = 1.4282, the trimmed
mean detector in [14] ftrim , and the optimal q-out-m procedure fqom (q∗ ).
fz∗s =1.4282
ftrim
fqom (q∗ )
security
1.43
1.43
0.69
efficiency
2.88
2.00
1.68
∗
f1.4282
ftrim
fqom (q∗ )
10−2
ǫ(k)
10−5
10−8
10−11
10−14
1
2
3
4
5
6
time k
7
8
9
10
Fig. 5: Finite-time performance of fz∗s , ftrim , and fqom (q∗ ) in the absence of the adversary.
VII. C ONCLUSION
AND
F UTURE WORK
In this paper, we studied the trade-off between the detection performance of a detector when
the attacker is absent (termed efficiency) and the “worst-case” detection performance when
the attacker, knowing the detector, is present (termed security). The setting is that a binary
hypothesis testing is conducted based on measurements from a set of sensors, some of which
can be compromised by an attacker and their measurements can be manipulated arbitrarily. We
first provided the fundamental limits of the trade-off between the efficiency and security of
any detector. We then presented detectors that possesses the limits of the efficiency and security.
Therefore, a clear guideline on how to balance the efficiency and security has been established for
the system operator. An interesting point of the fundamental trade-off is that in some cases, the
maximal efficiency and the maximal security can be achieved simultaneously, i.e., the maximal
January 22, 2018
DRAFT
26
efficiency (security) can be achieved without compromising any security (efficiency). In addition,
two extensions were investigated: secure sensors are assumed for the first one, and the detection
performance beyond the efficiency and security is concerned for the second one. The main results
were verified by numerical examples. Investigating the problem when the measurements from
the benign sensors are not i.i.d. is a future direction.
A PPENDIX A
T HE P ROOF
OF
T HEOREM 2
The following lemma is needed to prove Theorem 2:
Lemma 1. If Assumption 4 and 5 hold, then the following statement is true:
1) For any w,
M0 (w + 1) = M1 (w).
(28)
2) There exists a small enough ǫ > 0, such that log M0 (w) is well-defined on [−ǫ, 1 + ǫ], and
log M1 (w) is well-defined on [−1 − ǫ, ǫ].
3) log M0 (w), log M1 (w) are strictly convex.
4) The derivative of log M0 (w) and log M1 (w) satisfy
(log M0 (w))(1) |w=1 = D(1k0).
(29)
(log M0 (w))(1) |w=0 = −D(0k1).
(30)
(log M1 (w))(1) |w=0 = D(1k0).
(31)
(log M1 (w))(1) |w=−1 = −D(0k1).
(32)
Proof. By definition,
M0 (w + 1) =
=
Z
Z
∞
−∞
∞
−∞
dµ
(y)
dν
dµ
(y)
dν
w
w
dµ
(y)dν(y)
dν
dµ(y) = M1 (w),
which proves (28).
Assuming M0 (w1 ), M0 (w2 ) < ∞ and w1 < w2 , by the convexity of the exponential function,
we know that for any λ and 0 < α, β < 1 and α + β = 1,
0 < exp [(αw1 + βw2 )λ] ≤ αew1 λ + βew2 λ .
January 22, 2018
DRAFT
27
Therefore 0 < M0 (αw1 + βw2 ) ≤ αM0 (w1 ) + βM0 (w2 ) is well-defined, which proves that the
domain of log M0 (w) is convex.
Furthermore, by Assumption 4, 0 ∈ int(dom1 ) gives
1 ∈ int(dom0 ),
(33)
Hence, [0, 1] ⊂ int(dom0 ), which proves that log M0 (w) is well-defined on [−ǫ, 1 + ǫ] if ǫ is
small enough.
It is well known that log M0 (w) is infinitely differentiable on int(dom0 ) (see [22, Exercise
2.2.24]). Basic calculations give that
(log M0 (w))(2)
w
2
Z
dµ
dµ
(y)
log
(y)
dν(y) > 0
=
dν
dν
R
(34)
(35)
always holds, where (log M0 (w))(2) is the second derivative. The above quantity is strictly
positive since the KL divergence between probability measure µ and ν are strictly positive
by Assumption 5. Therefore, log M0 (w) is strictly convex.
The domain and the strict convexity of log M1 (w) can be proved similarly.
Take the derivative of log M0 (w) at w = 1 yields
Z
dµ
(1)
(log M0 (w)) |w=1 =
λ(y) (y)dν(y)
dν
ZR
=
λ(y)dµ(y) = D(1k0).
R
Equation (30), (31) and (32) can be proved similarly.
We are now ready to prove Theorem 2:
Proof of Theorem 2. Define the derivative of log M0 (w) to be ψ(w). Since log M0 (w) is strictly
convex, we know that ψ(w) is strictly increasing and therefore, its inverse function is well defined
on [−D(0k1), D(1k0)]. Denote the inverse function as ϕ(x). By the convexity of log M0 (w), we
have that
log M0 (w) ≥ log M0 (w∗ ) + ψ(w∗ )(w − w∗ ).
(36)
Hence, for any x ∈ [−D(0k1), D(1k0)], suppose that ψ(w∗ ) = x, we have
wx − log M0 (w) = [w∗ ψ(w∗ ) − log M0 (w∗ )]
+ [(w − w∗ )ψ(w∗ ) + log M0 (w∗ ) − log M0 (w)] .
January 22, 2018
DRAFT
28
Notice the last term on the RHS of the equation is non-positive. Hence, we can prove that
I0 (x) = w∗ ψ(w∗ ) − log M0 (w∗ ) = ϕ(x)x − log M0 (ϕ(x)).
(37)
Take the derivative and second order derivative of I0 (x) we have
d2 I0 (x)
1
dI0 (x)
= ϕ(x),
= (1)
> 0,
2
dx
dx
ψ (ϕ(x))
where the last inequality is due to the fact that log M0 (w) is strictly convex, and thus its second
derivative ψ (1) is strictly positive. Hence we prove that I0 (x) is twice differentiable and strictly
convex on [−D(0k1), D(1k0)]. Notice that
dI0 (x)
dx
= ϕ(−D(0k1)) = 0,
x=−D(0k1)
we can prove that I0 (x) is also strictly increasing. Similarly we can prove the properties for
I1 (x).
Combining (37), (29) and (30), we can prove (14) and (15). Equation (16) and (17) can be
proved similarly.
Since
I0 (0) = sup 0 · w − log M0 (w) = sup − log M0 (w),
w
w
I1 (0) = sup 0 · w − log M1 (w) = sup − log M1 (w),
w
w
and
M0 (w + 1) = M1 (w),
We can conclude I0 (0) = I1 (0).
A PPENDIX B
T HE P ROOF
OF
T HEOREM 3
The proof is divided into four parts, each of which is devoted to one of the statements in
Theorem 3.
Part I. For any index set O ⊂ M and χ ∈ R, define the following Bayesian like detector:
0 if P
i∈O λ̄i (k) < χ,
fk,χ,O (Y ′(k)) =
(38)
1 if P
λ̄
(k)
≥
χ,
i∈O i
January 22, 2018
DRAFT
29
where λ̄i (k) is the empirical mean of the log-likelihood ratio from time 1 to k for sensor i,
which is defined in (23). Denote
fχ,O = (f1,χ,O (Y ′ (1)), · · · , fk,χ,O (Y ′ (k)), · · · )
and
fO∗ , f0,O .
(39)
∗
It is well known that fM
minimize the average error probability [24]: e(θ = 0, O = ∅, k)+e(θ =
1, O = ∅, k), where, recall, e(θ, O, k) is defined in (6). Notice that
log(e(θ = 0, O = ∅, k) + e(θ = 1, O = ∅, k))
lim inf −
k→∞
k
log maxθ e(θ, O = ∅, k)
.
= lim inf −
k→∞
k
∗
Hence, when the attacker is absent, fM
is optimal in the sense that the rate ρ defined in (8)
∗
is maximized. Furthermore, Cramér’s Theorem gives that E(fM
) = mI0 (0) = mC. Therefore,
E(f ) ≤ mC holds for any detector f .
Part II. In this part, we show S (f ) ≤ (m−2n)+ C. The proof is by construction: we construct
a attack strategy g ∗ such that, for any detection rule f , the following inequality holds:
ρ(f, g ∗ ) ≤ (m − 2n)+ C.
(40)
Let O′ = {1, . . . , n} and O′′ = {m − n + 1, . . . , m}. The attack strategy g ∗ is as follows.
(i). When θ = 0, sensors in O′ are compromised and the distributions are flipped, i.e., the
measurements of sensors in O′ are i.i.d. as µ.
(ii). When θ = 1, sensors in O′′ \ O′ are compromised and the distributions are flipped.
Thus, under attack g ∗ , for either θ = 1 or θ = 0, sensors in O′ will follow distribution µ and
sensors in O′′ \ O′ will follow distribution ν. In other words, only sensors in M \ (O′ ∪ O′′ )
have different distributions under different θ. Notice that when m ≤ 2n, M \ (O′ ∪ O′′ ) = ∅,
∗
which means that ρ(f, g ∗ ) = 0. If m > 2n, by the optimality of the detection rule fM\(O
′ ∪O ′′ )
defined in (39), one obtains ρ(f, g ∗ ) ≤ (m − 2n)C. Equation (40) is thus obtained.
Part III. It is clear from the definitions of E(f ) and S (f ) that S (f ) ≤ E(f ) holds.
Part IV. Consider the following product measures:
µa = µ × µ . . . × µ × ν × ν . . . × ν ,
{z
}
{z
} |
|
m−n
n
µka = µa × µa . . . × µa
{z
}
|
k
January 22, 2018
DRAFT
30
and
ν∗ = ν × ν . . . × ν ,
{z
}
|
m
ν∗k = ν∗ × ν∗ . . . × ν∗ .
{z
}
|
k
The measure µa is generated by an attack that flips the distribution on the last n sensors, when
the true hypothesis is θ = 1. The measure ν∗ is generated by benign sensors when the true
hypothesis is θ = 0.
Now let us consider the following problem: given φ > 0, find the detection rule f such that
Eν∗k fk + φk Eµka (1 − fk )
(41)
is minimized for every k ≥ 1. Let fφ = (fφ,1 , . . . , fφ,k , . . .) with fk,φ given by
fk,φ (Y ′ (k)) = fk,− log φ,{1,2,...,m−n} (Y ′ (k)),
(42)
where, recall, the function fk,χ,O (Y ′ (k)) is defined in (38). Then by the Bayesian decisiontheoretic detection theory, fφ is a solution to the above problem. Let
Eφ , lim inf −
k→∞
and
Sφ , lim inf −
k→∞
log Eν∗k fφ,k
k
log Eµka (1 − fφ,k )
.
k
Then from the optimality of fφ , for any φ > 0 and any detector f = (f1 , . . . , fk , . . .), the
following hold for any k:
If Eν∗k fk ≤ Eν∗k fφ,k , then Eµka (1 − fk ) ≥ Eµka (1 − fφ,k ).
This implies that
log Eν∗k fk
≥ Eφ ,
k→∞
k
log Eµka (1 − fk )
then lim inf −
≤ Sφ .
k→∞
k
If lim inf −
Furthermore, the definitions of E(f ) and S (f ) yield
log Eν∗k fk
,
k→∞
k
log Eµka (1 − fk )
S (f ) ≤ lim inf −
.
k→∞
k
E(f ) ≤ lim inf −
January 22, 2018
DRAFT
31
Therefore, for any φ > 0 and any detector f , the following hold:
If E(f ) ≥ Eφ , then S (f ) ≤ Sφ .
Now let us evaluate Eφ and Sφ . Let φ̃ = − log φ/(m − n), then Cramér’s theorem yields
0
if φ̃ ≤ −D(0k1),
Eφ =
(m − n)I0 (φ̃) if φ̃ > −D(0k1),
and
Sφ =
0
if φ̃ ≥ D(1k0),
(m − n)I1 (φ̃) if φ̃ < D(1k0).
Notice that the monotonicity of I0 (·) on [−D(0k1), ∞) implies that if 0 < Eφ < (m −
n)I0 (D(1k0)) = (m − n)D(1k0), φ̃ ∈ (−D(0k1), D(1k0)) holds. Therefore, if 0 < Eφ <
(m − n)D(1k0), there holds
Sφ = (m − n)I1 (I0−1 (Eφ /(m − n))).
One thus obtains that for any detector f , if 0 < E(f ) < (m − n)D(1k0)
S (f ) ≤ (m − n)I1 (I0−1 (E(f )/(m − n))).
(43)
Also, it is easy to see that if Eφ ≥ (m − n)D(1k0), Sφ = 0 holds. Thus,
S (f ) = 0 if E(f ) ≥ (m − n)D(1k0).
(44)
Similarly, one considers the detection problem for the measures µk∗ and νak and obtains that
for any detector f , if 0 < E(f ) < (m − n)D(0k1)
S (f ) ≤ (m − n)I0 (I1−1 (E(f )/(m − n)))
(45)
S (f ) = 0 if E(f ) ≥ (m − n)D(0k1).
(46)
and
Then equation (21a) follows from (43) and (45), and equation (21b) from (44) and (46).
January 22, 2018
DRAFT
32
A PPENDIX C
T HE P ROOF
OF
T HEOREM 4
This theorem is proved by showing that fz∗s = 0 (or 1) if certain conditions are satisfied (i.e.,
Lemma 4). Furthermore, the special structure of these conditions can ensure that, under any
attacks, fz∗s = 0 (or 1) if the measurements of sensors in an attack free environment belong to
a certain set, to which the Cramér’s Theorem is applied.
Before proceeding, we need to define the following subsets of Rm :
Definition 2. Define B− , B+ ⊂ Rm as
)
(
)
(
m
m
X
X
−
m
+
m
λi < 0 , B , λ ∈ R :
λi ≥ 0 .
B , λ∈R :
i=1
i=1
Definition 3. Let O ⊂ M, j ∈ {0, 1} and z ∈ R+ , define a ball as
n
o
X
Bal(O, j, z) = λ ∈ Rm :
Ij (λi ) < z .
i∈O
Definition 4. Let j ∈ {0, 1} and z ∈ R+ , define an extended ball as
[
EBal(j, z, n) ,
Bal(O, j, z).
|O|=m−n
From the definition of extended balls, it is clear that
h
i
λ1 . . . λm ∈ EBal(j, z, n)
if and only if the following inequality holds:
X
min
Ij (λi ) < z.
|O|=m−n
Combining with the definition of
fz∗s ,
i∈O
we know that at time k, the output of fz∗s is 0 if and
only if
where Λ− (zs ) is defined as
h
i
λ̄(k) , λ̄1 (k) · · · λ̄m (k) ∈ Λ− (zs ),
Λ− (zs ) , EBal(0, zs , n)
The output is 1 if λ̄(k) ∈ Λ+ (zs ), where
[
B− \ EBal(1, zs , n) .
Λ+ (zs ) , Rm \Λ− (zs )
[
= B+ EBal(1, zs , n) \ EBal(0, zs , n)
January 22, 2018
DRAFT
33
We first need the following supporting lemma.
Lemma 2. Given O1 , O2 ⊂ M , {1, 2, . . . , m} with |O1
T
O2 | = p > 0, z ≤ pD(1k0), the
optimal value of the following optimization problem is given by pI1 (I0−1 (z/p)):
X
I1 (xi )
infm
x∈R
i∈O1
X
s.t.
(47)
I0 (xi ) < z.
i∈O2
Proof. Since I1 (·) is nonnegative, I1 (D(1k0)) = 0 and xi can take any value when i ∈
/ O2 , one
can equivalently rewrite (47) as
X
infm
x∈R
I1 (xi )
i∈O1 ∩O2
X
s.t.
I0 (xi ) < z,
i∈O2
xi = D(1k0), i ∈ O1 \ O2 .
By the nonnegativity of I0 (·), the above equation is equivalent to
X
I1 (xi )
inf
m
′
x∈R ,0<z ≤z
i∈O1 ∩O2
X
s.t.
I0 (xi ) < z ′ ,
i∈O1 ∩O2
X
(48)
I0 (xi ) ≤ z − z ′ ,
i∈O2 \O1
xi = D(1k0), i ∈ O1 \ O2 .
To obtain the solution to the above equation, let us fist focus on the following optimization
problem:
minm
x∈R
X
I1 (xi )
i∈O
s.t.
X
(49)
′
I0 (xi ) = z ,
i∈O
where O = O1 ∩ O2 . Denotes its optimal value by ψ(z ′ ). In the following, we show that
ψ(z ′ ) = pI1 (I0−1 (z ′ /p)).
(50)
We claim that a solution to (49) is
xi =
January 22, 2018
(
I0−1 (z ′ /p)
if i ∈ O,
(51a)
whatever
if i ∈
/ O.
(51b)
DRAFT
34
With this claim,
(50) clearly holds. In the following we show that this claim is correct.
Equation (51b) is trivial. We then focus on (51a). Due to the convexity of the functions I0 (·)
and I1 (·), one obtains that for any x ∈ Rm ,
pI0
X
xi /p
i∈O
pI1
X
xi /p
i∈O
!
≤
!
≤
X
I0 (xi ),
i∈O
X
I1 (xi )
i∈O
Therefore, without any performance loss, one may restrict the solution to the set X∗ as follows:
X∗ , {x ∈ Rm : x1 = x2 = · · · = xp }.
Then it is clear from the monotonicity of I0 and I1 that (51a) holds. This thus proves (50).
Notice that ψ(z ′ ) in (50) is decreasing with respect to z ′ . Then the fact that I0 (−D(0k1)) = 0
yields that (48) is equivalent to
minm
x∈R
s.t.
X
I1 (xi )
i∈O1 ∩O2
X
I0 (xi ) = z,
i∈O1 ∩O2
xi = −D(0k1), i ∈ O2 \ O1 ,
xi = D(1k0), i ∈ O1 \ O2 ,
which concludes Lemma 2 by (50).
Lemma 3. Assume that (ze , zs ) are an admissible pair, then the following statements are true:
1) Bal(M, 0, ze ) ⊆ B− .
2) Bal(M, 1, ze ) ⊆ B+ .
T
3) EBal(0, zs , n) EBal(1, zs , n) = ∅.
T
4) EBal(1, zs , n) Bal(M, 0, ze ) = ∅.
T
5) EBal(0, zs , n) Bal(M, 1, ze ) = ∅.
Proof. 1): It suffices to prove that given any x ∈ Rm , if x ∈ B+ , then x 6∈ Bal(M, 0, ze). By
the convexity of I0 (x), one obtains that
X
X
I0 (xi ) ≥ mI0 (1/m
xi ) ≥ mI0 (0) = mC,
i∈M
January 22, 2018
i∈M
DRAFT
35
where the second inequality follows from x ∈ B+ and the fact that I0 (x) is increasing when
x ≥ 0. Notice that by its definition, ze ≤ mC holds. The proof is done.
2): This can be proved similarly to 1).
3): By the definition of EBal, we need to prove that for any O1 , O2 with |O1 | = |O2 | = m−n,
Bal(O1 , 0, zs ) ∩ Bal(O2 , 1, zs ) = ∅ holds. Notice that when z ≤ pD(1k0), pI1 (I0−1 (z/p)) is
increasing with respect to p. Thus by Lemma 2, it suffices to prove that (m−2n)I1 (I0−1 (zs /(m−
2n))) ≥ zs , which is true because 0 ≤ zs ≤ (m−2n)C, pI1 (I0−1 (z/p)) is decreasing with respect
to z when z ≤ pD(1k0), and (m − 2n)I1 (I0−1 (0)) = (m − 2n)C.
4): Similar to 3), it suffices to prove that for any O1 with |O1 | = m − n, Bal(O1 , 0, zs ) ∩
Bal(M, 0, ze ) = ∅ holds. By Lemma 2, it suffices to prove that (m−n)I1 (I0−1 (ze /(m−n))) ≥ zs .
Then it is equivalent to prove that (m − n)I1 (I0−1 (he (zs )/(m − n))) ≥ zs , which follows from
the definition of he (z) and the fact that pI1 (I0−1 (z/p)) is decreasing with respect to z when
z ≤ pD(1k0).
5): This can be proved similarly to 4).
From Lemma 3, one obtains straightforwardly the following lemma.
Lemma 4. Assume that (ze , zs ) are an admissible pair, then the following set inclusions are
true:
1) EBal(0, zs , n) ⊆ Λ− (zs ).
2) EBal(1, zs , n) ⊆ Λ+ (zs ).
3) Bal(M, 0, ze ) ⊆ Λ− (zs ).
4) Bal(M, 1, ze ) ⊆ Λ+ (zs ).
We are now ready to prove Theorem 4.
Proof of Theorem 4. We focus on the proof of S (fz∗s ) ≥ zs , and a similar (and simpler) approach
can be used to prove E(fz∗s ) ≥ ze . Notice that EBal(0, zs , n) ⊆ Λ− (zs ) in Lemma 4 gives that,
January 22, 2018
DRAFT
36
under any attacks, there holds Bal(M, 0, zs ) ⊆ Λ− (zs ). Therefore,
1
log P0 (fz∗s ,k = 1)
k→∞ k
1
≤ lim sup log Po0 (λ̄(k) ∈ Rm \ Bal(M, 0, zs ))
k→∞ k
m
X
≤−
inf
I0 (xi )
m
lim sup
x∈R \Bal(M,0,zs )
i=1
= −zs ,
(52)
where the second inequality holds because of the Cramér’s Theorem and the fact that Rm \
Bal(M, 0, zs ) is closed.
Similarly, by EBal(1, zs , n) ⊆ Λ+ (zs ) in Lemma 4, one obtains
lim sup
k→∞
1
log P1 (fz∗s ,k = 0) ≤ −zs .
k
(53)
It follows from (52) and (53) that S (fz∗s ) ≥ zs . The proof is thus complete.
A PPENDIX D
T HE P ROOF
OF
T HEOREM 5
Define the following two functions h0 (z), h1 (z) : (0, Dmin) 7→ (0, Dmin):
h0 (z) = I0 (I1−1 (z)),
h1 (z) = I1 (I0−1 (z)).
Then we have the following two lemmas on h0 (z) and h1 (z).
Lemma 5. Both h0 (z) and h1 (z) are convex. Furthermore, the following equality holds:
h0 (C) = h1 (C) = C.
(54)
Proof. The equation (54) follows directly from (18) and (19). To prove the convexity of h0 (z)
and h1 (z), we first need to prove that I0−1 (x) is convex and I1−1 (x) is concave on [0, Dmin ].
Notice that if ψ is the inverse function of φ and φ are twice differentiable, then by chain rule
ψ (2) (x) = −
φ(2) (ψ(x))
[φ(1) (ψ(x))]
3.
Therefore, since I0 (x) (I1 (x)) is strictly convex and strictly decreasing (increasing) on [−D(0k1), D(1k0)],
I0−1 (x) (I1−1 (x)) is convex (concave) on [0, Dmin ].
January 22, 2018
DRAFT
37
The convexity of h0 (z) and h1 (z) then follows the fact that the composition of a convex and
increasing (decreasing) function with a convex (concave) function is convex [25].
We are now ready to prove Theorem 5
Proof. By chain rule, we know that
1
dh0 (z)
(1)
|z=C = I0 (x)|x=0 × (1)
= −1.
dz
I1 (x)|x=0
Therefore, by the convexity of h0 (z), we know that
h0 (z) ≥ h0 (C) − (z − C) ×
dh0 (z)
|z=C = 2C − z.
dz
Similarly, one can prove that
h1 (z) ≥ 2C − z.
Hence, by the definition of h(z),
h(z) ≥ 2(m − n)C − z,
∗
which implies that h((m − 2n)C) ≥ mC holds and f(m−2n)C
achieves maximum security and
efficiency simultaneously.
R EFERENCES
[1] A. S. Rawat, P. Anand, H. Chen, and P. K. Varshney, “Collaborative spectrum sensing in the presence of byzantine attacks
in cognitive radio networks,” IEEE Transactions on Signal Processing, vol. 59, no. 2, pp. 774–786, 2011.
[2] Y. Mo and B. Sinopoli, “Secure estimation in the presence of integrity attacks,” IEEE Transactions on Automatic Control,
vol. 60, no. 4, pp. 1145–1151, 2015.
[3] A. Teixeira, K. C. Sou, H. Sandberg, and K. H. Johansson, “Secure control systems: A quantitative risk management
approach,” IEEE Control Systems, vol. 35, no. 1, pp. 24–45, 2015.
[4] C. E. Shannon, “Communication theory of secrecy systems,” Bell Labs Technical Journal, vol. 28, no. 4, pp. 656–715,
1949.
[5] S. Marano, V. Matta, and L. Tong, “Distributed detection in the presence of byzantine attacks,” IEEE Transactions on
Signal Processing, vol. 57, no. 1, pp. 16–29, 2009.
[6] B. Kailkhura, Y. S. Han, S. Brahma, and P. K. Varshney, “Distributed bayesian detection in the presence of byzantine
data,” IEEE transactions on signal processing, vol. 63, no. 19, pp. 5250–5263, 2015.
[7] ——, “Asymptotic analysis of distributed bayesian detection with byzantine data,” IEEE Signal Processing Letters, vol. 22,
no. 5, pp. 608–612, 2015.
[8] M. Abdelhakim, L. E. Lightfoot, J. Ren, and T. Li, “Distributed detection in mobile access wireless sensor networks under
byzantine attacks,” IEEE Transactions on Parallel and Distributed Systems, vol. 25, no. 4, pp. 950–959, 2014.
January 22, 2018
DRAFT
38
[9] E. Soltanmohammadi, M. Orooji, and M. Naraghi-Pour, “Decentralized hypothesis testing in wireless sensor networks in
the presence of misbehaving nodes,” IEEE Transactions on Information Forensics and Security, vol. 8, no. 1, pp. 205–215,
2013.
[10] E. Soltanmohammadi and M. Naraghi-Pour, “Fast detection of malicious behavior in cooperative spectrum sensing,” IEEE
Journal on Selected Areas in Communications, vol. 32, no. 3, pp. 377–386, 2014.
[11] Y. Mo, J. P. Hespanha, and B. Sinopoli, “Resilient detection in the presence of integrity attacks,” IEEE Transactions on
Signal Processing, vol. 62, no. 1, pp. 31–43, 2014.
[12] K. G. Vamvoudakis, J. P. Hespanha, B. Sinopoli, and Y. Mo, “Detection in adversarial environments,” IEEE Transactions
on Automatic Control, vol. 59, no. 12, pp. 3209–3223, 2014.
[13] A. Abrardo, M. Barni, K. Kallas, and B. Tondi, “A game-theoretic framework for optimum decision fusion in the presence
of byzantines,” IEEE Transactions on Information Forensics and Security, vol. 11, no. 6, pp. 1333–1345, 2016.
[14] J.
Yan,
X.
Ren,
and
Y.
Mo,
“Sequential
detection
in
adversarial
environments,”
Online;
http://yilinmo.github.io/papers/cdc17-1.html.
[15] P. J. Huber, Robust statistics. Springer, 2011.
[16] ——, “A robust version of the probability ratio test,” The Annals of Mathematical Statistics, vol. 36, no. 6, pp. 1753–1758,
1965.
[17] R. Viswanathan and V. Aalo, “On counting rules in distributed detection,” IEEE Transactions on Acoustics, Speech, and
Signal Processing, vol. 37, no. 5, pp. 772–775, 1989.
[18] A. P. Dempster, N. M. Laird, and D. B. Rubin, “Maximum likelihood from incomplete data via the em algorithm,” Journal
of the royal statistical society. Series B (methodological), pp. 1–38, 1977.
[19] H. Fawzi, P. Tabuada, and S. Diggavi, “Secure estimation and control for cyber-physical systems under adversarial attacks,”
IEEE Transactions on Automatic Control, vol. 59, no. 6, pp. 1454–1467, 2014.
[20] S. Mishra, Y. Shoukry, N. Karamchandani, S. N. Diggavi, and P. Tabuada, “Secure state estimation against sensor attacks
in the presence of noise,” IEEE Transactions on Control of Network Systems, vol. 4, no. 1, pp. 49–59, 2017.
[21] G. Fellouris, E. Bayraktar, and L. Lai, “Efficient byzantine sequential change detection,” IEEE Transactions on Information
Theory, 2017.
[22] A. Dembo and O. Zeitouni, Large deviations techniques and applications.
Springer Science & Business Media, 2009,
vol. 38.
[23] R. Y. Rubinstein and D. P. Kroese, Simulation and the Monte Carlo method.
John Wiley & Sons, 2016.
[24] S. Key, “Fundamentals of statistical signal processing, volume ii: Detection theory,” 1993.
[25] S. Boyd and L. Vandenberghe, Convex optimization.
January 22, 2018
Cambridge university press, 2004.
DRAFT
| 7cs.IT
|
A1 -TYPE SUBGROUPS CONTAINING REGULAR
UNIPOTENT ELEMENTS
arXiv:1706.05212v1 [math.GR] 16 Jun 2017
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
Abstract. Let G be a simple exceptional algebraic group of adjoint type over an algebraically closed field of characteristic p > 0 and let X = PSL2 (p) be a subgroup of
G containing a regular unipotent element x of G. By a theorem of Testerman, x is
contained in a connected subgroup of G of type A1 . In this paper we prove that with
two exceptions, X itself is contained in such a subgroup (the exceptions arise when
(G, p) = (E6 , 13) or (E7 , 19)). This extends earlier work of Seitz and Testerman, who
established the containment under some additional conditions on p and the embedding
of X in G. We discuss applications of our main result to the study of the subgroup
structure of finite groups of Lie type.
1. Introduction
Let G be a simple algebraic group of adjoint type over an algebraically closed field K of
characteristic p > 0. Let X = PSL2 (q) be a subgroup of G, where q > 4 is a p-power, and
let x ∈ X be an element of order p. By the main theorem of [25], x is contained in a closed
connected subgroup of G of type A1 , unless G = G2 , p = 3 and x belongs to the conjugacy
(3)
class of G labelled A1 as in [14]. With a view towards applications to the study of the
subgroup structure of finite groups of Lie type, it is desirable to seek natural extensions
of this result. In particular, under what conditions can one embed the full subgroup X in
an A1 subgroup of G?
As a special case of the main theorem of [28], this question has a positive answer when G
is classical and X is not contained in a proper parabolic subgroup of G (for G = SLn (K),
this is a well-known theorem of Steinberg [31]). One can see that the condition on the
embedding of X is necessary by considering indecomposable representations of X which
do not arise as restrictions of indecomposable representations of an algebraic A1 . In [28],
Seitz and Testerman also provide a positive answer if G is a simple exceptional algebraic
group (of type G2 , F4 , E6 , E7 or E8 ) and p is large enough, still under the same assumption
that X is not contained in a proper parabolic subgroup of G. More precisely, the approach
in [28] requires p > N (G) where
N (G2 ) = 19, N (F4 ) = 43, N (E6 ) = 43, N (E7 ) = 67, N (E8 ) = 113.
(1)
More general results on the embedding of finite quasisimple subgroups in exceptional
algebraic groups are established by Liebeck and Seitz in [18]. For instance, if X = PSL2 (q)
and q is sufficiently large, then [18, Theorem 1] implies that X is contained in a proper
closed positive dimensional subgroup of G. Here “sufficiently large” means that q >
t(G).(2, p − 1) with
t(G2 ) = 12, t(F4 ) = 68, t(E6 ) = 124, t(E7 ) = 388, t(E8 ) = 1312.
(2)
It is natural to seek an extension of [28, Theorem 2] by removing the conditions on p
and the embedding of X in G when G is of exceptional type and X = PSL2 (q). In [29],
Seitz and Testerman study the case where x ∈ X is semiregular in G (that is, CG (x)
is a unipotent group). Notice that if x is not semiregular then x ∈ CG (s)0 for some
Date: March 27, 2018.
1
2
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
non-trivial semisimple element s ∈ G and one can hope to answer the question in the
proper reductive subgroup CG (s)0 ; so the semiregular case, where such a reduction is
not possible, is particularly interesting. In this situation, the main result of [29] states
that X is contained in a connected subgroup of type A1 if either q > p, or if q = p and
PGL2 (q) 6 NG (X).
In this paper, we extend the results in [29] by studying the remaining case where
X = PSL2 (p) and PGL2 (p) 66 NG (X). In order to do this, we will assume x ∈ X is regular
in G, which means that CG (x) is an abelian unipotent group of dimension r, where r is
the rank of G (equivalently, x is contained in a unique Borel subgroup of G). It is well
known that regular unipotent elements exist in all characteristics and they form a single
conjugacy class. Since the order of x is the smallest power of p greater than the height
of the highest root of G (see [34, Order Formula 0.4]), our hypothesis implies that p > h,
where h is the Coxeter number of G. (Recall that h = dimr G − 1 = ht(α0 ) + 1, where
ht(α0 ) is the height of the highest root of G.) In particular, given the aforementioned
earlier work in [28], we may assume that
h 6 p 6 N (G),
(3)
where N (G) is given in (1).
Our main result is the following (in this paper, an A1 -type subgroup is a closed connected
subgroup isomorphic to SL2 (K) or PSL2 (K)).
Theorem 1. Let G be a simple exceptional algebraic group of adjoint type over an algebraically closed field of characteristic p > 0. Let X = PSL2 (p) be a subgroup of G
containing a regular unipotent element of G. Then exactly one of the following holds:
(i) X is contained in an A1 -type subgroup of G;
(ii) G = E6 , p = 13 and X is contained in a D5 -parabolic subgroup of G;
(iii) G = E7 , p = 19 and X is contained in an E6 -parabolic subgroup of G.
In all three cases, X is uniquely determined up to G-conjugacy.
Remark 1. Let us make some comments on the statement of Theorem 1.
(a) To see the uniqueness of X in part (i), it suffices to show that every subgroup
Y = PSL2 (p) of G containing x is conjugate to X. Write X < A and Y < B,
where A and B are A1 -type subgroups of G. By Proposition 2.11(ii), A and B are
G-conjugate, say A = B g , so X, Y g < A. Finally, by applying [17, Theorem 5.1]
and Lang’s theorem, we deduce that X and Y g are A-conjugate.
(b) The interesting examples arising in (ii) and (iii) were found by Craven [9] in his
recent study of the maximal subgroups with socle PSL2 (q) in finite exceptional
groups of Lie type. The action of such a subgroup X on the adjoint module Lie(G)
is described in Theorem 8.1 (see Section 8) and its construction is explained in [9,
Section 9]. Let us say a few words on the construction in (ii), where G = E6 and
p = 13. Let P = QL be a D5 -parabolic subgroup of G and identify the unipotent
radical Q with a 16-dimensional spin module for L0 = D5 . Take a subgroup
Y = PSL2 (p) < L0 containing a regular unipotent element of L0 and consider
the semidirect product QY < P (note that Y is uniquely determined up to L0 conjugacy). Now one checks that Q|Y has an 11-dimensional composition factor
W with dim H 1 (Y, W ) = 1. It follows that there is a complement X = PSL2 (p) to
Q in QY that is not QY -conjugate to Y . Moreover, one can show that X contains
a regular unipotent element of G and there is a unique P -class of such subgroups
X (hence X is uniquely determined up to G-conjugacy). We will show that the
subgroup X constructed in this way is not contained in an A1 -type subgroup of G
(this follows from Theorem 2 below). A similar construction can be given in (iii)
A1 -TYPE SUBGROUPS
3
and again one can show that such a subgroup is both unique up to conjugacy and
is not contained in an A1 -type subgroup.
(c) The conclusion of Theorem 1 for G = G2 can be deduced from the proof of [29,
Lemma 3.1]. It also follows from Kleidman’s classification of the maximal subgroups of G2 (p) in [13]. However, for completeness we will provide an alternative
proof, following the same approach we use for the other exceptional groups.
(d) Finally, let us comment on the adjoint hypothesis in the statement of the theorem.
Let G be a simple exceptional algebraic group and let Gad be the corresponding
adjoint group. Suppose Y = PSL2 (p) or SL2 (p) is a subgroup of G containing
a regular unipotent element y of G. The regularity of y implies that Z(Y ) 6
Z(G) and thus Y Z(G)/Z(G) = PSL2 (p) is a subgroup of Gad containing a regular
unipotent element, so it is determined by Theorem 1.
The next result shows that the subgroups X in part (i) of Theorem 1 are G-irreducible
in the sense of Serre (that is, X is not contained in a proper parabolic subgroup of G).
The proof is given at the end of Section 2. By [35, Theorem 1.2], any connected reductive
subgroup of a reductive algebraic group G containing a regular unipotent element is Girreducible, so we can view Theorem 2 as a partial analogue for subgroups isomorphic to
PSL2 (p) in simple exceptional groups.
Theorem 2. Let G be a simple exceptional algebraic group of adjoint type over an algebraically closed field of characteristic p > 0 and let x ∈ G be a regular unipotent element
such that
x ∈ X = PSL2 (p) < A < G,
where A is an A1 -type subgroup of G. Then X is G-irreducible.
Remark 2. As in Theorem 1, let X = PSL2 (p) be a subgroup of G containing a regular
unipotent element. By combining Theorems 1 and 2, we deduce that X is contained in an
A1 -type subgroup of G if and only if X is not contained in a proper parabolic subgroup
of G. In particular, the examples arising in parts (ii) and (iii) of Theorem 1 are genuine
exceptions to the containment in (i).
The next result follows by combining Theorem 1 with the main results of [28, 29].
Corollary 1. Let G be a simple algebraic group of adjoint type over an algebraically closed
field of characteristic p > 0 and let X = PSL2 (q) be a subgroup of G containing a regular
unipotent element of G, where q > 4 is a p-power. In addition, if G is classical assume
that X is G-irreducible. Then either
(a) X is contained in an A1 -type subgroup of G, or
(b) q = p and (G, p, X) is one of the cases in parts (ii) and (iii) in Theorem 1.
Next we present some further applications of Theorem 1. Let G be a simple algebraic
group as in Theorem 1 and recall that a finite subgroup of G is Lie primitive if it is not
contained in a proper closed subgroup of positive dimension. In [11, Section 3], Guralnick
and Malle determine the maximal Lie primitive subgroups H of G containing a regular
unipotent element (the maximal closed positive dimensional subgroups of G containing a
regular unipotent element were determined in earlier work of Saxl and Seitz [26]). More
precisely, they give a list of possibilities for H, but they do not claim that all cases actually
occur. In particular, their proof relies on [28] and thus H = PSL2 (p) arises as a possibility
when G ∈ {F4 , E6 , E7 , E8 } and h 6 p 6 N (G). Therefore, by combining [11, Theorems
3.3, 3.4] with Theorem 1, we obtain the following refinement.
Corollary 2. Let G be a simple exceptional algebraic group of adjoint type over an algebraically closed field of characteristic p > 0. Suppose H is a maximal Lie primitive
subgroup of G containing a regular unipotent element. Let H0 denote the socle of H.
4
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
(i) If G = G2 , then one of the following holds:
(a) p = 2 and H = J2 ;
(b) p = 7 and H = 23 .SL3 (2), G2 (2) or PSL2 (13);
(c) p = 11 and H = J1 .
(ii) If G = F4 , then one of the following holds:
(a) p = 2 and H0 = PSL3 (16), PSU3 (16) or PSL2 (17);
(b) p = 13 and H = 32 .SL3 (3), or H0 = PSL2 (25), PSL2 (27) or 3 D4 (2).
(iii) If G = E6 , then one of the following holds:
(a) p = 2 and H0 = PSL3 (16), PSU3 (16) or Fi22 ;
(b) p = 13 and either H = 33 .SL3 (3) or H0 = 2 F4 (2)0 .
(iv) If G = E7 , then p = 19 and H0 = PSU3 (8) or PSL2 (37).
(v) If G = E8 , then one of the following holds:
(a) p = 2 and H0 = PSL2 (31);
(b) p = 7 and H0 = PSp8 (7) or Ω9 (7);
(c) p = 31 and H = 25+10 .SL5 (2) or 53 .SL3 (5), or H0 = PSL2 (32), PSL2 (61) or
PSL3 (5).
Remark 3. By Corollary 2, there are no Lie primitive subgroups containing a regular
unipotent element if p > 31. This lower bound is best possible: the case (G, p) = (E8 , 31)
with H0 = PSL2 (32) is a genuine example (this can be deduced from recent work of
Litterick [22]). However, we are not claiming that all of the possibilities listed in Corollary
2 are Lie primitive and contain regular unipotent elements (indeed, we expect that this
list can be reduced further).
We can also use Theorem 1 to shed new light on the subgroup structure of finite exceptional groups of Lie type. Let G be a simple exceptional algebraic group of adjoint type
over F̄p with p prime and let F : G → G be a Steinberg endomorphism of G with fixed
point subgroup GF , an almost simple group over Fq . The maximal subgroups of the Ree
groups 2 G2 (q) and 2 F4 (q) (and their automorphism groups) have been determined up to
conjugacy by Kleidman [13] and Malle [23], respectively, and similarly G2 (q) is handled
in [8] for q even and in [13] for q odd. Therefore, we may assume GF is one of F4 (q),
E6 (q), 2 E6 (q), E7 (q) and E8 (q). In these cases, through the work of many authors, the
problem of determining the maximal subgroups H of GF has essentially been reduced to
the case where H is an almost simple group of Lie type with socle H0 over a field Fq0 of
characteristic p (see [24, Section 29.1] and the references therein). Here one of the main
problems is to determine if such a subgroup is of the form M F , where M is maximal
among positive dimensional F -stable closed subgroups of G. Significant restrictions on
the rank of H0 and the size of q0 are established in [16, 18], but the problem of obtaining
a complete classification is still open.
The special case H0 = PSL2 (q0 ) is of particular interest. If q0 > t(G).(2.p − 1), where
t(G) is the integer in (2), then the aforementioned work of Liebeck and Seitz [18] shows that
q = q0 and H = M F for some maximal connected subgroup M of G of type A1 . Further
results in this direction have recently been obtained by Craven [9] when GF is one of F4 (q),
E6 (q), 2 E6 (q) or E7 (q). Using the maximality of H, he proves that H = M F in almost
every case, but his approach is unable to eliminate certain values of q0 . In particular, the
case where H = PSL2 (h + 1) contains a regular unipotent element of G is problematic
(the existence of such subgroups, in a much more general setting, was established by Serre
[30], which explains why they are called Serre embeddings in [9]). Using Theorem 1, one
can show that all maximal Serre embeddings are of the form (M F )0 (we can also handle
A1 -TYPE SUBGROUPS
5
G = E8 , which is excluded in [9]). In particular, it follows that part (ii) in [9, Theorem
1.1] is a subcase of part (iii), and the same conclusion holds for [9, Theorem 1.3].
To conclude the introduction, let us briefly describe the main steps in the proof of
Theorem 1 (we refer the reader to Section 2.5 for more details). Suppose x ∈ X =
PSL2 (p) < G is a regular unipotent element of G and let A < G be an A1 -type subgroup
containing x with maximal torus T = {t(c) | c ∈ K × }. Set V = Lie(G) and F×
p = hξi.
Without loss of generality, replacing X by a suitable G-conjugate, we show that we may
assume that X contains the toral element t(ξ) ∈ T , which corresponds to a diagonalizable
element s ∈ SL2 (p) with eigenvalues ξ and ξ −1 (see Lemma 2.19). We can use the known
action of A on V to determine the eigenvectors and eigenspaces of s on V and this severely
restricts the possibilities for V |X . It is possible to obtain further restrictions on the
indecomposable summands of V |X by considering the trace on V of semisimple elements
in X of small order (typically, we only need to work with elements of order 2 and 3).
In this way, in almost all cases, we are able to reduce to the situation where V |X is
compatible with the action of a PSL2 (p) subgroup of A. In this situation, V |X is given
in Table 2 (our notation for indecomposable summands in Table 2 is explained in Section
2.1) and we observe that the socle of V |X has a 3-dimensional simple summand
W = hw2 , w0 , w−2 i,
where wi is an eigenvector for s with eigenvalue ξ i . Let Ei be the ξ i -eigenspace of s on
V . Without loss of generality, we may assume that the action of x on W (in terms of this
basis) is given by the matrix
1 1 1
0 1 2
0 0 1
and thus
w2 ∈ ker(x − 1) ∩ E2 ,
w0 ∈ ker((x − 1)2 ) \ ker(x − 1) ∩ E0 ,
w−2 ∈ ker((x − 1)3 ) \ ker((x − 1)2 ) ∩ E−2 .
Our main goal is to show that W is an sl2 -subalgebra of V .
To do P
this, we may assume that x is obtained by exponentiating the regular nilpotent
element γ∈Π(G) eγ ∈ V with respect to a fixed Chevalley basis
B = {eα , fα , hγ | α ∈ Φ+ (G), γ ∈ Π(G)}
for V (see Section 2.5 for more details). This allows us to explicitly identify a maximal
torus T of an A1 -type subgroup of G containing x, which means that we can compute
eigenvectors and eigenspaces for s in terms of the Chevalley basis. With the aid of Magma
[3] to simplify the computations, we can describe the action of x on V in terms of a dim G×
dim G matrix with respect to B and then compute bases for the subspaces ker((x − 1)i )
for i > 1. In this way, we obtain expressions for w2 , w0 and w−2 in terms of B, but
with undetermined coefficients. We then derive relations between these coefficients by
considering the action of x on W , and further relations can be found by using the fact
that CV (x) = ker(x − 1) is an abelian subalgebra. Apart from a handful of special cases,
this allows us to reduce to the case where W is an sl2 -subalgebra and we complete the
argument by showing that the stabilizer of W in G is an A1 -type subgroup.
This process of elimination and extension comprises the bulk of the proof of Theorem
1 (see Sections 3–7). However, there are a handful of possibilities for (G, p) which require
further attention; these are the cases arising in part (ii) of Theorem 2.22 and they are
handled in Section 8. In each of these cases, the action of X on V is known (up to one of
6
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
three possibilities if (G, p) = (E6 , 13) or (E7 , 19)) and X stabilizes a non-zero subalgebra
of heα | α ∈ Φ+ (G)i. This allows us to reduce to the case where X is contained in
a proper parabolic subgroup P = QL of G. Let π : P → P/Q be the quotient map.
Using π, we identify L with P/Q and so we may view π(X) as a subgroup of L0 . We
may as well assume that P is a minimal parabolic (with respect to containing X), so
π(X) is not contained in a proper parabolic subgroup of L0 . Now π(x) ∈ L0 is a regular
unipotent element which is contained in an A1 -type subgroup H of L0 (this follows by
combining Theorem 2.22 with the aforementioned earlier work of Seitz and Testerman
[28] for classical groups). By inspecting [15], we can determine the action of H on V ,
which must be compatible with the action of X on V given in Theorem 2.22. In this way
we deduce that (G, p, L0 ) = (E6 , 13, D5 ) and (E7 , 19, E6 ) are the only possibilities, and
this completes the proof of Theorem 1.
Remark 4. We anticipate that some of the new methods introduced in this paper will
be applicable more generally. Indeed, in future work we will investigate subgroups of
exceptional algebraic groups of the form X = PSL2 (p) containing semiregular unipotent
elements, with a view to removing the conditions in [28, Theorem 2].
Notation. Our notation is fairly standard. For a simple algebraic group G we write Φ(G),
Φ+ (G) and Π(G) = {α1 , . . . , αr } for the set of roots, positive roots and simple roots of G,
with respect to a fixed Borel subgroup, and we follow Bourbaki [4] in labelling the simple
roots. We will often denote a root α = a1 α1 + · · · + ar αr by writing α = a1 · · · ar . If V is a
module for a group then soc(V ) and rad(V ) denote the socle and radical of V , respectively,
and we write V m to denote V ⊕ · · · ⊕ V (with m summands). It will be convenient to write
[An1 1 , . . . , Ank k ] for a block-diagonal matrix with a block Ai occurring with multiplicity ni .
In addition, we will write Ji for a standard (upper triangular) unipotent Jordan block of
size i.
Acknowledgments. Testerman was supported by the Fonds National Suisse de la
Recherche Scientifique, grant numbers 200021 146223 and 200021 153543. Burness thanks
the Section de Mathématiques and the Centre Interfacultaire Bernoulli at EPFL for their
generous hospitality. It is a pleasure to thank David Craven and Jacques Thévenaz for
many useful discussions. We also thank Bob Guralnick, Gunter Malle and Iulian Simion
for helpful comments on an earlier version of this paper.
2. Preliminaries
In this section we record some preliminary results that will be needed in the proof of
Theorem 1. We start by recalling some well known results from the modular representation
theory of the simple groups PSL2 (p). Our main reference is Alperin [1].
2.1. Representation theory. Let K be an algebraically closed field of characteristic
p > 5, let X = PSL2 (p) and let P = hxi ∼
= Zp be a Sylow p-subgroup of X.
The subgroup P of X has exactly p indecomposable KP -modules, say W1 , . . . , Wp ,
where dim Wi = i and Wp is the unique projective indecomposable KP -module. The
element x has Jordan form [Ji ] on Wi . In particular, if W is a projective KP -module,
then dim W = ap for some a > 1, and x has Jordan form [Jpa ] on W .
There are precisely (p + 1)/2 simple KX-modules, labelled V1 , V3 , . . . , Vp in [1], where
dim Vi = i. In particular, every simple KX-module is odd-dimensional. Here V1 is the
trivial module and Vp is the Steinberg module. It is easy to see that x has Jordan form
[Ji ] on Vi . By a theorem of Steinberg, each Vi is the restriction of a simple module for
the corresponding algebraic group of type A1 (see [32, Section 13]), so we can refer to the
highest weight of Vi with respect to a maximal torus of the algebraic A1 . We identify
A1 -TYPE SUBGROUPS
7
the weights of this 1-dimensional torus with the set of integers, and we will often write
Vi = LX (i − 1) to highlight the highest weight of Vi .
Similarly, there are precisely (p+1)/2 projective indecomposable KX-modules, labelled
P1 , P3 , . . . , Pp in [1], where Pp = Vp is simple and the remainder are reducible. Here
dim P1 = dim Pp = p and dim Pi = 2p for 1 < i < p. The element x has Jordan form [Jp ]
on P1 and Pp , and Jordan form [Jp2 ] on the remaining Pi . The structure of these modules
is described by Alperin [1, pp.48–49]. In terms of composition factors, we have
P1 = V1 |Vp−2 |V1
and
Pi = Vi |(Vp−i+1 ⊕ Vp−i−1 )|Vi
where 1 < i < p is odd. (Here this notation indicates that soc(Pi ) ∼
= Pi /rad(Pi ) ∼
= Vi and
∼
rad(Pi )/soc(Pi ) = Vp−i+1 ⊕ Vp−i−1 .) It will be convenient to define
U = P1 = LX (0)|LX (p − 3)|LX (0)
(4)
W (i) = Pi+1 = LX (i)|(LX (p − i − 1) ⊕ LX (p − i − 3))|LX (i)
(5)
and
for i ∈ {2, 4, . . . , p − 3}.
The Green correspondence (see [1, Section 11]) implies that if V is an indecomposable
KX-module then V |P = W ⊕W 0 where W is projective (or zero) and W 0 is indecomposable
(or zero). In particular, the following lemma holds.
Lemma 2.1. Let V be an n-dimensional indecomposable KX-module and write n = ap+b,
where a > 0 and 0 6 b < p. Then x has Jordan form [Jpa , Jb ] on V .
The main result on the structure of indecomposable KX-modules is the following theorem. Here we define a subtuple of an n-tuple [m1 , . . . , mn ] to be a tuple of the form
[mi , mi+1 , . . . , mj ] for some 1 6 i 6 j 6 n. We denote this by writing
[mi , mi+1 , . . . , mj ] ⊆ [m1 , . . . , mn ].
Theorem 2.2. Let V be a reducible indecomposable non-projective KX-module. Then
there exists an integer ` > 2 and a subtuple
[a1 , . . . , a` ] ⊆ [1, p − 2, 3, p − 4, . . . , p − 2, 1]
such that
soc(V ) = Va1 ⊕ Va3 ⊕ · · · ⊕ Va`− , V /soc(V ) = Va2 ⊕ Va4 ⊕ · · · ⊕ Va`−1+
where = 1 if ` is even, otherwise = 0.
Proof. This follows from the discussion in [12, Section 3]. Also see [9, Proposition 5.14].
Corollary 2.3. Let V be an indecomposable KX-module with precisely two composition
factors. If soc(V ) = LX (i) then V /soc(V ) ∈ {LX (p − i − 3), LX (p − i − 1)} for some
i ∈ {0, 2, . . . , p − 3}, hence dim V = p ± 1.
Corollary 2.4. Let V be a reducible indecomposable KX-module. Then dim V > p − 1.
Moreover, if V has at least four composition factors, then dim V > 2(p − 1).
There is a very similar description of the indecomposable modules for SL2 (p). Indeed,
if we write X = SL2 (p)/Z then each indecomposable KX-module corresponds to an
indecomposable module for SL2 (p) with kernel Z = {±I2 }.
8
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
2.2. Traces. Let x2 and x3 be representatives of the unique conjugacy classes of elements
of order 2 and 3 in X, respectively (note that xi is semisimple since p > 5). Let V be a
KX-module and let tr(V, xi ) denote the trace of xi on V .
Lemma 2.5. If V = LX (i) then
tr(V, x2 ) = (−1)i/2 , tr(V, x3 ) =
1 i ≡ 0 (mod 3)
−1 i ≡ 1 (mod 3)
0 i ≡ 2 (mod 3)
Proof. This is a straightforward calculation, using the fact that we can identify LX (i) with
the i-th symmetric power Symi (M ), where M is the natural module for SL2 (K).
If V is a KX-module with composition factors V1 , . . . , Vk , then
tr(V, xi ) =
k
X
tr(Vj , xi )
j=1
since the action of xi is diagonalizable. Therefore, the next two results are immediate
corollaries of Lemma 2.5 (here we use the notation U and W (i) defined in (4) and (5)).
Lemma 2.6. We have
1 p ≡ 1 (mod 4)
1 p ≡ 1 (mod 3)
tr(U, x2 ) =
, tr(U, x3 ) =
3 p ≡ 3 (mod 4)
2 p ≡ 2 (mod 3)
Lemma 2.7. We have
tr(W (i), x2 ) =
2 i ≡ 0 (mod 4)
−2 i ≡ 2 (mod 4)
and
2 p ≡ 1 (mod 3)
1 p ≡ 2 (mod 3)
−1 p ≡ 1 (mod 3)
tr(W (i), x3 ) =
−2 p ≡ 2 (mod 3)
−1 p ≡ 1 (mod 3)
1 p ≡ 2 (mod 3)
i ≡ 0 (mod 3)
i ≡ 1 (mod 3)
i ≡ 2 (mod 3)
Let G be a simple algebraic group over K of adjoint type, let V be a KG-module and
let m be a positive integer. Define
Tm (G, V ) = {tr(V, x) | x ∈ G has order m}.
Recall that the adjoint module for G is the Lie algebra Lie(G), on which G acts via the
adjoint representation.
Proposition 2.8. Let G be a simple exceptional algebraic group of adjoint type over an
algebraically closed field of characteristic p > 5. Let V = Lie(G) be the adjoint module.
Then Tm (G, V ) is recorded in Table 1 for m ∈ {2, 3}.
Proof. This follows by inspecting the dimensions of the centralizers of elements of order
m in G (see [10, Tables 4.3.1 and 4.7.1]), using the fact that
dim CV (g) = dim CG (g)
for every semisimple element g ∈ G (see [6, Section 1.14], for example).
A1 -TYPE SUBGROUPS
G
9
T2 (G, V ) T3 (G, V )
G2 −2
−1, 5
F4
−4, 20
−2, 7
E6
−2, 14
−3, 3, 6, 15, 30
E7
−7, 5, 25
−2, 7, 34, 52
E8
−8, 24
−4, 5, 14, 77
Table 1. Traces of elements of order 2 and 3 on the adjoint module
Remark 2.9. Suppose X = PSL2 (p) is contained in G = E6 , where p > 5 and G is
adjoint. Write G = Ĝ/S and X = X̂/S, where Ĝ is the simply connected group of type
E6 and S = Z3 is the centre of Ĝ. Now X has Schur multiplier Z2 , which implies that
X̂ = Z3 × X. Therefore, every element y ∈ X of order 3 lifts to an element in Ĝ of order
3. In particular, if y ∈ X has order 3 then CG (y)0 = A5 T1 , D4 T2 or A32 (see [10, Table
4.7.1]), whence tr(V, y) ∈ {−3, 6, 15} with respect to V = Lie(G).
Remark 2.10. In a few cases it is helpful to know the eigenvalue multiplicities on V of
elements in G of order m > 3 for certain values of m; the relevant cases are the following:
(G, m) ∈ {(F4 , 7), (E6 , 7), (E7 , 5), (E8 , 19)}.
It is straightforward to obtain this information with the aid of Magma [3], using an
algorithm of Litterick (see [21, Section 3.3.1] – we thank Dr. Litterick for his assistance
with these computations.
2.3. A1 -type subgroups. Let G be a simple algebraic group and recall that p is a good
prime for G if p > 2 in types B, C and D, p > 3 for G2 , F4 , E6 and E7 , and p > 5 when G
is of type E8 (all primes are good in type A).
Proposition 2.11. Let G be a simple algebraic group of adjoint type over an algebraically
closed field of good characteristic p > 0. Let x ∈ G be an element of order p.
(i) There is an A1 -type subgroup of G containing x.
(ii) If x is regular then the subgroup in (i) is unique up to G-conjugacy.
Proof. Part (i) follows from the main theorem of [34]. Part (ii), for G exceptional, follows
from [15, Theorem 4]. Now assume G is classical and let H be an A1 -type subgroup of
G containing x. Let V be the natural module for G. By [35, Theorem 1.2], H is not
contained in a proper parabolic subgroup of G. In particular, if G is of type A, B or C
then H acts irreducibly and tensor indecomposably (see [35, Proposition 2.3]) on V and
the conjugacy statement follows from representation theory.
Finally, let us assume G = Dr (with r > 4). We claim that H < L < G, where L = Br−1
is the stabilizer of a non-singular 1-space. The result then follows since H is unique in L
up to L-conjugacy, and L itself is unique up to G-conjugacy. To justify the claim, first
observe that x has Jordan form [J2r−1 , J1 ] on V , using [26, Lemma 1.2(ii)] and the fact
that x has order p, so p 6= 2. In particular, H acts reducibly on V and we complete the
argument by applying [20, Lemma 2.2].
Proposition 2.12. Let G be a simple exceptional algebraic group of adjoint type and let
x ∈ G be a regular unipotent element such that
x ∈ X = PSL2 (p) < A < G,
10
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
where A is an A1 -type subgroup. Then the action of X on the adjoint module V = Lie(G)
is given in Table 2.
Proof. A precise description of V |A as a tilting module is given in [19, Table 10.1] (for
G = E6 we may assume that A < F4 < G so the action of A on V can be deduced from the
actions of A on Lie(F4 ) and the minimal module for F4 (see [19, Table 10.2])). Following
[19], we write T (l; µ; . . .) for a tilting module having the same composition factors as the
direct sum of Weyl modules for A with highest weights l, µ, . . .. In terms of this notation,
we get
T (10; 2)
G = G2
G = F4
T (22; 14; 10; 2)
T (22; 16; 14; 10; 8; 2)
G = E6
V |A =
T
(34;
26;
22;
18;
14;
10;
2)
G = E7
T (58; 46; 38; 34; 26; 22; 14; 2) G = E8
As explained at the start of [19, Section 10], we can express T (l; µ; . . .) as a direct sum
of indecomposable tilting modules of the form T (c), where the highest weight c is at most
2p − 2. For example, suppose G = F4 and p = 19, so V |A = T (22; 14; 10; 2) as above.
The highest weight is 22, so one summand is T (22), which is a uniserial module of shape
14|22|14 (see [27, Lemma 2.3]). The highest weight not already accounted for is 10, so
T (10) = LA (10) is a summand and we deduce that V |A = T (22) ⊕ LA (10) ⊕ LA (2) and
thus
V |X = T (22)|X ⊕ LX (10) ⊕ LX (2).
By [27, Lemma 2.3], T (22)|X is a projective indecomposable KX-module of dimension
2p = 38, so T (22)|X = W (i) for some i. By comparing socles, it follows that i = 14 and
thus
V |X = W (14) ⊕ LX (10) ⊕ LX (2)
as recorded in Table 2. The other cases are entirely similar and we omit the details.
For the remainder of this section, let G be a simple exceptional algebraic group of
adjoint type over an algebraically closed field K of characteristic p > 0, and let r and
h = h(G) be the rank and Coxeter number of G, respectively. We will assume G contains
a regular unipotent element of order p, which means that
p > h.
(6)
We need to recall the construction of A1 -type subgroups of G containing regular unipotent
elements, following the treatment in [30, 33, 34].
First we need some new notation. Let LC be a simple Lie algebra over C of type Φ(G).
Fix a Chevalley basis
B = {eα , fα , hγ | α ∈ Φ+ (G), γ ∈ Π(G)}
of LC and write zi = zαi for z ∈ {e, f, h} and Π(G) = {α1 , . . . , αr }. It will be convenient
to define fα = e−α for each α ∈ Φ+ (G). Let LZ be the Z-span of B and set LK = LZ ⊗Z K.
(By abuse of notation, we also write eα , fα , ei , fi , hi for the elements eα ⊗ 1, fα ⊗ 1, etc.,
in LK .) Fix a root α ∈ Φ(G). As in the familiar Chevalley construction, we have
(ad(eα )j /j!)(LZ ) ⊆ LZ
for all j > 0, and this allows us to construct the element exp(ad(xeα )) ∈ GLdim G (Z[x]),
where x is an indeterminate. By passing to K, we obtain a 1-dimensional unipotent
subgroup
Uα = {exp(ad(γeα )) | γ ∈ K} 6 Aut(LK )0 = G
(see [6, Proposition 4.4.2]). Note that G = hUα | α ∈ Φ(G)i.
A1 -TYPE SUBGROUPS
V |X
G
p
G2
p > 11 LX (10) ⊕ LX (2)
p=7
F4
11
W (2)
p > 23 LX (22) ⊕ LX (14) ⊕ LX (10) ⊕ LX (2)
p = 19 W (14) ⊕ LX (10) ⊕ LX (2)
p = 17 W (10) ⊕ LX (14) ⊕ LX (2)
p = 13 W (10) ⊕ W (2)
E6
p > 23 LX (22) ⊕ LX (16) ⊕ LX (14) ⊕ LX (10) ⊕ LX (8) ⊕ LX (2)
p = 19 W (14) ⊕ LX (16) ⊕ LX (10) ⊕ LX (8) ⊕ LX (2)
p = 17 W (10) ⊕ LX (16) ⊕ LX (14) ⊕ LX (8) ⊕ LX (2)
p = 13 W (10) ⊕ W (8) ⊕ W (2)
E7
p > 37 LX (34) ⊕ LX (26) ⊕ LX (22) ⊕ LX (18) ⊕ LX (14) ⊕ LX (10) ⊕ LX (2)
p = 31 W (26) ⊕ LX (22) ⊕ LX (18) ⊕ LX (14) ⊕ LX (10) ⊕ LX (2)
p = 29 W (22) ⊕ LX (26) ⊕ LX (18) ⊕ LX (14) ⊕ LX (10) ⊕ LX (2)
p = 23 W (18) ⊕ W (10) ⊕ LX (22) ⊕ LX (14) ⊕ LX (2)
p = 19 W (14) ⊕ W (10) ⊕ W (2) ⊕ LX (18)
E8
p > 59 LX (58) ⊕ LX (46) ⊕ LX (38) ⊕ LX (34) ⊕ LX (26) ⊕ LX (22) ⊕ LX (14) ⊕ LX (2)
p = 53 W (46) ⊕ LX (38) ⊕ LX (34) ⊕ LX (26) ⊕ LX (22) ⊕ LX (14) ⊕ LX (2)
p = 47 W (34) ⊕ LX (46) ⊕ LX (38) ⊕ LX (34) ⊕ LX (26) ⊕ LX (22) ⊕ LX (14) ⊕ LX (2)
p = 43 W (38) ⊕ W (26) ⊕ LX (34) ⊕ LX (22) ⊕ LX (14) ⊕ LX (2)
p = 41 W (34) ⊕ W (22) ⊕ LX (38) ⊕ LX (26) ⊕ LX (14) ⊕ LX (2)
p = 37 W (34) ⊕ W (26) ⊕ W (14) ⊕ LX (22) ⊕ LX (2)
p = 31 W (26) ⊕ W (22) ⊕ W (14) ⊕ W (2)
Table 2. The action of X on V = Lie(G) in Proposition 2.12
Given the lower bound on p in (6), we can make a similar construction for more general
elements of LZ . To do this, let LZ(p) be the Z(p) -span of B, where Z(p) is the localization
of Z at the prime ideal (p) = pZ, so that LK = LZ(p) ⊗Z(p) K. By [34, Proposition 1.5] we
have
ad(e)j /j! (LZ(p) ) ⊆ LZ(p)
P
for all e ∈ α∈Φ+ (G) Zeα and all j > 0. Then as in the Chevalley construction, for any
P
P
non-zero element y in α∈Φ+ (G) Zeα or α∈Φ+ (G) Zfα , we can produce
xy (x) = exp(ad(xy)) ∈ GLdim G (Z(p) [x]).
In particular, by passing to K, we define
Uy = {xy (γ) = exp(ad(γy)) | γ ∈ K} ⊆ Aut(LK )0 = G.
(7)
We will use this general set-up to construct certain A1 -type subgroups of our group G,
following [33, 34]. In order to state the main result (Proposition 2.13 below), recall that
an ordered triple of elements (e, h, f ) chosen from LK (or from LZ ) is an sl2 -triple if they
satisfy the commutation relations between the standard generators of the Lie algebra sl2 ,
namely
[h, e] = 2e, [h, f ] = −2f, [e, f ] = h.
12
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
We have the following result (in part (iii), we use the notation xy (γ) from (7)).
Pr
Proposition
Pr 2.13. Suppose p > h(G) and (e, h, f ) is an sl2 -triple of LZ , with e = i=1 ei
and f ∈ i=1 Zfi . Then the following hold:
(i) Ue and Uf are 1-dimensional subgroups of G.
(ii) A = hUe , Uf i is an A1 -type subgroup of G.
(iii) T = {t(c) | c ∈ K × } is a maximal torus of hUe , Uf i, where
t(c) = xe (c)xf (−c−1 )xe (c)xe (−1)xf (1)xe (−1),
and the map t : Gm → T is a morphism of algebraic groups.
(iv) The action of T on the basis {v̄ = v ⊗ 1 | v ∈ B} of LK is given by
t(c) · ēα = cα(h) ēα , t(c) · h̄i = h̄i
for all α ∈ Φ(G), 1 6 i 6 r. Moreover, αi (h) = 2 for all 1 6 i 6 r.
(v) T normalizes Ue and Uf .
(vi) Ue contains a regular unipotent element of G.
Proof. This follows by combining [33, Lemmas 1 and 2] and [34, Lemma 1.2].
The following proposition will play an important role in the proof of Theorem 1.
Pr
Proposition
2.14.
Suppose
p
>
h(G)
and
(e,
h,
f
)
is
an
sl
-triple
of
L
,
with
e
=
2
Z
i=1 ei
Pr
and f ∈ i=1 Zfi . Let W be the 3-dimensional subalgebra of LK generated by {e, f } and
let H be the stabilizer of W in G. Then H is an A1 -type subgroup of G.
Proof. Let A be the A1 -type subgroup of G constructed in Proposition 2.13(ii). Note that
A contains a regular unipotent element and it clearly stabilizes W by construction, so
A 6 H. Let M1 be a maximal closed positive dimensional subgroup of G with H 6 M1 .
By the main theorem of [35], A is not contained in a proper parabolic subgroup of G, so
Borel-Tits [2, Corollary 3.9] (also see Weisfeiler [36]) implies that M1 is reductive. By [26,
Theorem A], either M1 is an A1 -type subgroup (and thus A = H = M1 ), or G = E6 and
M1 = F4 . In the latter case, H 6= M1 since M1 does not stabilize a 3-dimensional subspace
of LK , so let M2 be a maximal closed positive dimensional subgroup of M1 with H 6 M2 .
As above, M2 is reductive and by applying [26, Theorem A] once again, we conclude that
A = H = M2 .
We would like to be able to use Proposition 2.14 to identify the stabilizers of other
sl2 -subalgebras of LK . With this aim in mind, we present Proposition 2.15 below. In
order to state this result, we need some additional notation.
Suppose we have an sl2 -triple (e, h, f ) as in Proposition 2.13. Let T be the 1-dimensional
torus constructed in part (iii) of the proposition. Let α0 ∈ Φ(G) be the highest root and
recall thatPh(G) = ht(α0 ) + 1, where
ht : Φ(G) → N is the familiar height function (that
P
is, if α = i ai αi then ht(α) = i ai ). Then
{2i | −ht(α0 ) 6 i 6 ht(α0 )}
is the set of weights of T on both LZ and LK . For each T -weight m, write (LZ )m for the
corresponding T -weight space and similarly for LK . In both the statement and proof of
the following result, we use the notation ā = a ⊗ 1 ∈ LK for a ∈ LZ .
Pr
Proposition
2.15.
Suppose
p
>
h(G)
and
(e,
h,
f
)
is
an
sl
-triple
of
L
,
with
e
=
2
Z
i=1 ei
P
and f ∈ ri=1 Zfi . Suppose y ∈ (LZ )p−1 ∩ CLZ (e) and z ∈ (LZ )p−3 are chosen so that
(i) [y, z] = 0 in LZ ; and
(ii) (ē, h̄ + γ ȳ, f¯ + γ z̄) is an sl2 -triple in LK for some γ ∈ K.
A1 -TYPE SUBGROUPS
13
Then there exists g ∈ CG (ē) such that g · h̄ = h̄ + γ ȳ and g · f¯ = f¯ + γ z̄ in LK . Moreover,
the stabilizer in G of the subalgebra W of LK generated by {ē, f¯ + γ z̄} is an A1 -type
subgroup.
P
Proof. First observe that y ∈ α∈Φ+ (G) Zeα since y ∈ (LZ )p−1 , so we can take g = xy (γ) ∈
G as in (7). Note that g ∈ CG (ē) since y ∈ CLZ (e). Now y is an eigenvector for ad(h)
(since y is a T -weight vector), so [y, [y, h]] = 0 and thus
g · h̄ = h̄ + γ[ȳ, h̄] = h̄ − γ[h̄, ȳ] = h̄ − γ(p − 1)ȳ = h̄ + γ ȳ.
The maximum T -weight in LZ is 2ht(α0 ), which is at most 2(p−1) since p > h(G) > ht(α0 ),
so ad(y)i (f ) ∈ (LZ )i(p−1)−2 = 0 for all i > 3 and thus
1
g · f¯ = f¯ + γ[ȳ, f¯] + γ 2 [ȳ, [ȳ, f¯]].
2
In addition, since [y, z] = 0 and z ∈ (LZ )p−3 , we have
[h + y, f + z] = −2f + (p − 3)z + [y, f ].
(8)
¯
¯
The sl2 commutation relations imply that [h̄ + ȳ, f + z̄] = −2(f + z̄), which is equal to
−2f¯ − 3z̄ + [ȳ, f¯] by (8). Therefore, [ȳ, f¯] = z̄ and thus [ȳ, [ȳ, f¯]] = 0. We conclude that
g · h̄ = h̄ + γ ȳ and g · f¯ = f¯ + γ z̄ in LK , as required. The final statement concerning the
stabilizer of W follows immediately from Proposition 2.14.
2.4. Exponentiation. In this section we turn to a different notion of “exponentiation”,
following Seitz [27]. As before, let G be a simple exceptional algebraic group of adjoint
type over an algebraically closed field K of characteristic p > 0 and let r and h denote the
rank and Coxeter number of G. Let U = hUα | α ∈ Φ+ (G)i be the unipotent radical of
a fixed Borel subgroup B of G corresponding to our choice of base Π(G) = {α1 , . . . , αr },
where the root subgroup Uα is defined as in (7). As explained in [27, Section 5], we may
view Lie(U ) as an algebraic group via the Hausdorff formula. Set V = Lie(G).
We start by recalling [27, Proposition 5.3].
Proposition 2.16. Suppose p > h. Then there exists a unique isomorphism of algebraic
groups
θ : Lie(U ) → U
(9)
whose tangent map is the identity and which is B-equivariant; that is, θ(b · n) = bθ(n)b−1
for all n ∈ Lie(U ), b ∈ B.
Suppose G contains a regular unipotent element x of order p, so p > h and we are in a
position to use Proposition 2.16 to study the structure of CG (x). Replacing x by a suitable
conjugate, we may assume that
x = xe (1) = exp(ad(e)),
Pr
where e = i=1 ei . As in Proposition 2.13, let A be an A1 -type subgroup of G containing
x, and let T = {t(c) | c ∈ K × } be the given maximal torus of A. Without loss of generality,
we may assume that T is contained in the Borel subgroup B defined above. From the
description of the action of A on V = Lie(G) in the proof of Proposition 2.12, it follows
that t(c) acts on the 1-eigenspace CV (x) = Lie(CG (x)) as
diag(cd1 , . . . , cdr ),
(10)
where the di are recorded in Table 3 (we label the di so that they form a decreasing
sequence).
Proposition
2.17. Let x = xe (1) ∈ G be a regular unipotent element of order p, where
P
e = ri=1 ei , and let T = {t(c) | c ∈ K × } be the torus constructed in Proposition 2.13.
14
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
G
di
G2 10, 2
F4
22, 14, 10, 2
E6
22, 16, 14, 10, 8, 2
E7
34, 26, 22, 18, 14, 10, 2
E8
58, 46, 38, 34, 26, 22, 14, 2
Table 3. The integers d1 , . . . , dr in (10)
(i) There exist connected 1-dimensional unipotent subgroups Xi = {xi (γ) | γ ∈ K}
such that CG (x) = hXi | 1 6 i 6 ri. In particular,
each z ∈ CG (x) can be written
Qr
as a commuting product of the form z = i=1 xi (γi ) for some γi ∈ K.
(ii) We have
t(c)xi (γ)t(c)−1 = xi (cdi γ)
(11)
for all c ∈ K × , γ ∈ K, 1 6 i 6 r.
Proof. First note that p > h since x has order p. As above, let U = hUα | α ∈ Φ+ (G)i
be the unipotent radical of a Borel subgroup B of G and note that x ∈ U and T 6 B.
Moreover, we have CG (x) 6 U and thus CV (x) = Lie(CG (x)) ⊆ Lie(U ). Choose vr ∈
Lie(U ) such that θ(vr ) = x, where θ is the map in Proposition 2.16. Extend to a basis
{v1 , . . . , vr } of the 1-eigenspace CV (x), where t(c) · vi = cdi vi for each i, and construct the
corresponding connected 1-dimensional unipotent subgroups
Xi = {xi (γ) = θ(γvi ) | γ ∈ K} 6 G.
Recall that CG (x) is abelian, so CV (x) = Lie(CG (x)) is an abelian subalgebra and the
proof of [27, Proposition 5.4] implies that each Xi is contained in CG (x). Therefore,
H = hXi | 1 6 i 6 ri is a closed connected unipotent subgroup of CG (x). Moreover,
vi ∈ Lie(Xi ) for each i, so dim H > r and thus H = CG (x) (note that CG (x) is connected
since G is adjoint). Part (i) now follows since CG (x) is abelian. Finally, part (ii) follows
from the B-equivariance of θ (see Proposition 2.16).
Proposition 2.18. Let U be the unipotent radical of a Borel subgroup B of G, let W be
a proper non-zero subalgebra of Lie(U ) and let H be the stabilizer of W in G. Assume H
contains a regular unipotent element of G of order p. Then either
(i) H is contained in a proper parabolic subgroup of G; or
(ii) H is contained in an A1 -type subgroup of G.
Proof. Since p > h, we can consider the isomorphism θ : Lie(U ) → U in (9). Let Z = Z(W )
be the centre of W , which is a non-zero abelian subalgebra of W stabilized by H. We
claim that θ(Z) 6 H. To see this, let z ∈ Z, w ∈ W and note that θ(z) and θ(w) commute
since [z, w] = 0 in Lie(U ) (see the proof of [27, Proposition 5.4]). The B-equivariance of θ
implies that
θ(θ(z) · w) = θ(z)θ(w)θ(z)−1 = θ(w),
so θ(z) · w = w and the claim follows. Therefore, H is a positive dimensional subgroup of
G containing a regular unipotent element.
To complete the argument, we proceed as in the proof of Proposition 2.14, using BorelTits [2, Corollary 3.9]. Let us assume H is not contained in a proper parabolic subgroup of
G. Then H 6 M1 , where M1 is a maximal closed reductive positive dimensional subgroup
A1 -TYPE SUBGROUPS
15
of G. By the main theorem of [26], either M1 is an A1 -type subgroup, or G = E6 and
M1 = F4 , so we may assume that we are in the latter situation. Suppose H = M1 . Since
V |M1 = Lie(M1 ) ⊕ V26 ,
where V26 is the minimal module for M1 , it follows that W = V26 is the only possibility.
But V26 must contain a non-zero semisimple element (just by comparing the ranks of G
and M1 ) and this is a contradiction. Therefore H is a proper subgroup of M1 and thus
H 6 M2 for some maximal closed reductive subgroup M2 of M1 . By a further application
of [26] we conclude that H is contained in an A1 -type subgroup of G.
2.5. Methods. In this section we discuss the proof of Theorem 1, highlighting the main
steps and ideas.
Let G be a simple exceptional algebraic group of adjoint type defined over an algebraically closed field K of characteristic p > 0. Let r be the rank of G and let V = Lie(G)
be the adjoint module. Suppose x ∈ X = PSL2 (p) < G is a regular unipotent element of
G, so p > h where h is the Coxeter number of G. The embedding of X in G corresponds
to an abstract homomorphism ϕ : SL2 (p) → G with kernel Z = Z(SL2 (p)) and image X.
As before, let LC be a simple Lie algebra over C of type Φ(G) and fix a Chevalley basis
B = {eα , fα , hγ | α ∈ Φ+ (G), γ ∈ Π(G)}.
(12)
Since p > h, we can view B as a basis for V , where eα , fα are in the appropriate root
spaces with respect to the Cartan subalgebra spanned by the hγ . It will be convenient to
write zi = zαi for z ∈ {e, f, h} and Π(G) = {α1 , . . . , αr }.
P
Set e = ri=1 ei and let (e, h, f ) be an sl2 -triple as in Proposition 2.13. Let A be the
corresponding A1 -type subgroup of G constructed in Proposition 2.13, with maximal torus
T = {t(c) | c ∈ K × } and associated morphism t : Gm → T . By replacing X by a suitable
G-conjugate, we may assume that x = exp(ad(e)) ∈ A. Let v : Ga → A be a morphism of
algebraic groups such that
t(c)v(γ)t(c)−1 = v(c2 γ)
for all c ∈ K × , γ ∈ K. We may assume v : Ga → im(v) is an isomorphism of algebraic
groups.
Consider the elements
ξ 0
1 1
u=
, s=
(13)
0 1
0 ξ −1
in SL2 (p), where F×
p = hξi = {1, . . . , p − 1}. Without loss of generality, we may assume
that x = ϕ(u) = v(1) so ϕ(sus−1 ) = ϕ(um ) = v(m) with m = ξ 2 . Then t(ξ)xt(ξ)−1 =
ϕ(s)xϕ(s)−1 and thus ϕ(s) = t(ξ)z for some z ∈ CG (x). Set s̄ = ϕ(s) ∈ X.
Lemma 2.19. There exists a CG (x)-conjugate of X containing x and t(ξ).
Proof. As noted above, we have s̄ = t(ξ)z
Qr for some z ∈ CG (x). By Proposition 2.17
thereQare scalars γi ∈ K such that z = i=1 xi (γi ). Let us consider a general element
y = ri=1 xi (δi ) ∈ CG (x). In view of (11), we get
Y
Y
Y
Y
ys̄y −1 =
xi (δi )t(ξ)
xi (γi − δi )t(ξ)−1 t(ξ) =
xi (δi )
xi (ξ di (γi − δi ))t(ξ)
i
i
i
=
Y
i
xi (δi + ξ di (γi − δi ))t(ξ)
i
where the di are the integers appearing in Table 3.
di
Since p > h and F×
p = hξi, it is easy to see that there is at most one i such that ξ = 1.
If there is no such i then we can set δi = ξ di γi /(ξ di − 1) for all i, so ys̄y −1 = t(ξ) and X y is
the desired conjugate of X. Finally, suppose ξ dj = 1 and γj 6= 0 for some j. By defining δi
16
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
as above for all i 6= j, we get ys̄y −1 = t(ξ)xj (γj ) with [t(ξ), xj (γj )] = 1. But this implies
that ys̄y −1 is a non-semisimple element, which contradicts the semisimplicity of s̄.
In view of the lemma, we may assume that X contains t(ξ), which corresponds to a
diagonalizable element s ∈ SL2 (p) with eigenvalues ξ and ξ −1 . Since t(ξ) ∈ T , we can
use the known action of A on V (see the proof of Proposition 2.12) to determine the
eigenvectors and eigenspaces of s on V . For example,
{ξ d1 , ξ d2 , . . . , ξ dr }
(14)
is the collection of eigenvalues of s on CV (x), where the di are given in Table 3. We set
s̄ = t(ξ) = sZ ∈ X, where Z is the centre of SL2 (p). Note that A contains the Borel
subgroup hs̄, xi of X.
The proof of Theorem 1 has three main steps, which we now describe.
Step 1: Elimination. Our initial aim is to reduce to the situation where the action of X on
V is compatible with the decomposition of V as an A-module given in Table 2. In almost all
cases, we are able to achieve this goal. To do this, we consider the possible decompositions
of V |X as a direct sum of indecomposable KX-modules, using the description of these
modules given in Section 2.1, with the aim of eliminating all but one possibility.
First we use the fact that the decomposition of V |X has to be compatible with the
Jordan form of x on V (this can be read off from the relevant tables in [14]). In addition,
it must be compatible with the known eigenvalues of s on V (as noted above, these are
just the eigenvalues of t(ξ) on V , which we can compute from the known action of A on
V ). Note that if M is an indecomposable summand of V |X then the restriction of M
to hsi is completely reducible, so we just need to identify the KX-composition factors of
M in order to compute the eigenvalues of s on this summand. Often it is sufficient to
compare the eigenvalues of s on CV (x) with the expected eigenvalues in (14), and we can
also use our earlier calculations on the traces of elements of order 2 and 3 to obtain further
restrictions on V |X (see Section 2.2). With this approach in mind, the following lemma
will be useful.
Lemma 2.20. Let M be an indecomposable KX-module of the form LX (i), U or W (j),
where i ∈ {0, 2, . . . , p − 1} and j ∈ {2, 4, . . . , p − 3}. Then the eigenvalues of s on CM (x)
are ξ i , 1 and {ξ j , ξ −j }, respectively.
Proof. First recall that x has Jordan form [Ji+1 ], [Jp ] and [Jp2 ] on LX (i), U and W (j),
respectively. The fixed point of x on the simple module LX (i) has highest weight i, so the
result is clear in this case. Similarly, soc(U ) = LX (0) so s has eigenvalue ξ 0 on CU (x).
Finally, suppose M = W (j). The highest weight of soc(M ) = LX (j) is j, so ξ j is one of
the eigenvalues of s on CM (x). To determine the second eigenvalue, it is helpful to view
W (j) as the restriction to X of the tilting module T (2p − 2 − j) for the ambient algebraic
group of type A1 (see [27, Lemma 2.3]). On the latter module, x has a fixed point of
weight 2p − 2 − j (the high weight), so the eigenvalue of s is ξ 2p−2−j = ξ −j as required.
Let us illustrate how Step 1 is carried out in the specific case (G, p) = (E8 , 31).
8 ] on V (see [14,
Example 2.21. Suppose G = E8 and p = 31, so x has Jordan form [J31
Table 9]). In particular, V |X is projective and thus every indecomposable summand of
V |X is also projective. In terms of the notation introduced in Section 2.1, the possibilities
A1 -TYPE SUBGROUPS
17
for V |X are as follows
M 1 ⊕ M 2 ⊕ M 3 ⊕ M 4 ⊕ M5 ⊕ M6 ⊕ M7 ⊕ M8
W (a1 ) ⊕ M1 ⊕ M2 ⊕ M3 ⊕ M4 ⊕ M5 ⊕ M6
W (a1 ) ⊕ W (a2 ) ⊕ M1 ⊕ M2 ⊕ M3 ⊕ M4
W (a1 ) ⊕ W (a2 ) ⊕ W (a3 ) ⊕ M1 ⊕ M2
W (a1 ) ⊕ W (a2 ) ⊕ W (a3 ) ⊕ W (a4 )
where Mi ∈ {LX (30), U } and ai ∈ {2, 4, . . . , 28}. If V |X has an Mi summand, then s has
an eigenvalue ξ 30 = ξ 0 on CV (x), which contradicts (14), so we must have
V |X = W (a1 ) ⊕ W (a2 ) ⊕ W (a3 ) ⊕ W (a4 ).
Since ξ 30 = 1 and s has eigenvalues ξ i , ξ −i on CW (i) (x) (see Lemma 2.20), it follows that
{ξ a1 , ξ −a1 , ξ a2 , ξ −a2 , ξ a3 , ξ −a3 , ξ a4 , ξ −a4 } = {ξ 28 , ξ 16 , ξ 8 , ξ 4 , ξ 26 , ξ 22 , ξ 14 , ξ 2 }
(see Table 3). Up to a re-ordering of summands, this immediately implies that
a1 ∈ {2, 28}, a2 ∈ {4, 26}, a3 ∈ {8, 22}, a4 ∈ {14, 16}.
Let y ∈ X be an involution. Since tr(W (i), y) = ±2 and tr(V, y) ∈ {−8, 24} (see Lemma
2.7 and Proposition 2.8), it follows that tr(W (ai ), y) = −2 for all i, whence ai ≡ 2 (mod 4)
and thus (a1 , a2 , a3 , a4 ) = (2, 26, 22, 14) is the only possibility. We have now reduced to
the case where the decomposition of V |X is compatible with V |A (see Table 2).
Step 2: Extension. Next observe that if V |X has the decomposition given in Table 2 then
the socle of V |X has a simple summand W = LX (2). To complete the argument, we aim
to show that W is an sl2 -subalgebra of V and its stabilizer in G is an A1 -type subgroup.
We can do this in almost every case; the exceptions are the two special cases appearing in
the statement of Theorem 1.
Let {w2 , w0 , w−2 } be a basis for W , where wi is an eigenvector for s with eigenvalue ξ i .
We may assume that the action of x on W is given by the matrix
1 1 1
0 1 2
(15)
0 0 1
with respect to this basis (that is, x(w0 ) = w0 + w2 , etc.). If we define s̄ = sZ ∈ X
as above then hs̄, xi is a Borel subgroup of X and we can consider the opposite Borel
subgroup hs̄, x0 i of X, where x0 ∈ X is also a regular unipotent element of order p. With
respect to the above basis, we may assume that x0 acts on W via the matrix
1 0 0
2 1 0
(16)
1 1 1
If all these conditions are satisfied, then we will say that {w2 , w0 , w−2 } is a standard basis
for W .
With the aid of Magma [3] we can construct a dim G × dim G matrix to represent the
action of x on V with respect to our Chevalley basis B. Moreover, we can use Proposition
2.13(iv) to compute the eigenvalues and eigenvectors of t(ξ) (and thus s) on V in terms of
B. For i ∈ Z, it will be convenient to write Ei for the ξ i -eigenspace of s on V (so wi ∈ Ei
for the elements in a standard basis of W ).
Next we identify a basis {v1 , . . . , vr } of the 1-eigenspace CV (x) = ker(x − 1) in terms
of B, where vi ∈ Edi (see Table 3). Since w2 ∈ CV (x) ∩ E2 we can write
r
X
w2 =
ai vi
i=1
18
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
for some ai ∈ K, where ai 6= 0 only if ξ di = ξ 2 . Similarly,
w0 ∈ ker((x − 1)2 ) \ ker(x − 1) ∩ E0 ,
w−2 ∈ ker((x − 1)3 ) \ ker((x − 1)2 ) ∩ E−2 .
Using Magma it is straightforward to compute bases for the relevant kernels; these computations can be done by hand, but it is much quicker and more efficient to use a machine.
Given these bases, say B2 , B0 and B−2 , we can write
X
X
X
w2 =
av v, w0 =
bv v, w−2 =
cv v
v∈B2
v∈B0
v∈B−2
for av , bv , cv ∈ K and our goal is to determine these scalars. To do this, we can use the
specified actions of x and x0 on W to derive relations between the coefficients. Further
relations can be determined by exploiting the fact that x and x0 are regular unipotent
elements. For example, we observe that x0 · w−2 = w−2 and
x0 · [w−2 , w0 ] = [w−2 , w0 + w−2 ] = [w−2 , w0 ],
where [ , ] is the Lie bracket on V , so w−2 , [w−2 , w0 ] ∈ CV (x0 ). Since the regularity of
x0 implies that CG (x0 ) is abelian, it follows that CV (x0 ) = Lie(CG (x0 )) is an abelian
subalgebra of V (for the latter equality, recall that p > h) and thus
[w−2 , [w−2 , w0 ]] = 0.
(17)
Proceeding in this way, our P
goal is to reduce toP
the case where W = hw2 , w0 , w−2 i is
an sl2 -subalgebra, with w2 = ri=1 ei and w−2 ∈ ri=1 Zfi . Moreover, we want to find
integers l, µ such that (w2 , lw0 , µw−2 ) is an sl2 -triple over Z (that is, an sl2 -triple of LZ in
the notation of Section 2.3). Indeed, if we can do this, then Proposition 2.14 implies that
the stabilizer of W in G is an A1 -type subgroup and so we are in the genericP
situation
described in part (i) of Theorem 1. In a few cases, we are unable to force w−2 ∈ ri=1 Zfi ,
but by appealing to Proposition 2.15 we can still show that the same conclusion holds.
In the remaining cases where W is not an sl2 -subalgebra, or the action of X on V is
incompatible with V |A , we will show that X stabilizes a non-zero subalgebra of heα | α ∈
Φ+ (G)i. More precisely, we will establish the following result, which reduces the proof of
Theorem 1 to the handful of cases appearing in Table 4 (see Remark 1(a) for the conjugacy
statement in part (i)).
Theorem 2.22 (Reduction Theorem). Let G be a simple exceptional algebraic group of
adjoint type over an algebraically closed field of characteristic p > 0. Let X = PSL2 (p)
be a subgroup of G containing a regular unipotent element of G and set V = Lie(G) with
Chevalley basis as in (12). Then one of the following holds:
(i) X is contained in an A1 -type subgroup of G and X is uniquely determined up to
G-conjugacy;
(ii) X stabilizes a non-zero subalgebra of heα | α ∈ Φ+ (G)i and (G, p, V |X ) is one of
the cases in Table 4.
We will prove the Reduction Theorem in Sections 3–7, considering each possibility for
G in turn.
Step 3: Parabolic analysis. The final step in our proof of Theorem 1 concerns the cases
arising in Theorem 2.22(ii), given in Table 4. In view of Proposition 2.18, we may assume
that X is contained in a proper parabolic subgroup P = QL of G and we proceed by
studying the possible embeddings of X in such a subgroup. Take P to be a minimal such
parabolic and let π : P → P/Q be the quotient map. By identifying L with P/Q, we
may view π(X) as a subgroup of L0 . Now we can show that π(X) < H, where H is an
A1 -TYPE SUBGROUPS
19
V |X
G
p
F4
13 W (10) ⊕ W (2)
E6 13 W (10) ⊕ W (8) ⊕ W (2)
W (10) ⊕ W (4) ⊕ W (2)
W (10)2 ⊕ W (4)
E7 19 W (8) ⊕ W (4) ⊕ W (2) ⊕ U
W (16) ⊕ W (10) ⊕ W (4) ⊕ U
W (16) ⊕ W (14) ⊕ W (8) ⊕ U
E8 37 W (34) ⊕ W (26) ⊕ W (14) ⊕ LX (22) ⊕ LX (2)
Table 4. The exceptional cases (G, p, V |X ) in Theorem 2.22
A1 -type subgroup of L0 containing a regular unipotent element of L0 (namely, π(x)), so
we can use [15] to study the composition factors of V |H for each (G, L0 ). In turn, this
imposes restrictions on the decomposition of V |X . But the possibilities for V |X are listed
in Table 4 and in this way we arrive at the two special cases in the statement of Theorem
1. See Section 8 for the details. (Notice that we adopt a similar approach in the proof of
Theorem 2 below.)
Example 2.23. To illustrate some of the above ideas, let us explain how we handle the
case (G, p) = (E8 , 31). Recall that in Example 2.21 we reduced to the situation where
V |X = W (2) ⊕ W (26) ⊕ W (22) ⊕ W (14),
which is compatible with the decomposition of V |A .
Let W = soc(W (2)) = LX (2) and let {w2 , w0 , w−2 } be a standard basis of W as above.
First consider w2 ∈ CV (x). Now CV (x) ∩ E2 is 1-dimensional (indeed, by inspecting Table
3 we see that there is a unique di which is congruent to 2 modulo 30), spanned by the sum
of the simple root vectors, so we must have
w2 = a1 (e1 + e2 + e3 + e4 + e5 + e6 + e7 + e8 )
for some non-zero scalar a1 ∈ K. Similarly, w0 is contained in the 1-dimensional space
ker((x − 1)2 ) ∩ E0 and we take
w0 = a2 (h1 + 19h2 + 4h3 + 9h4 + 28h5 + 18h6 + 10h7 + 4h8 ).
Finally, w−2 is in the 2-dimensional space ker((x − 1)3 ) ∩ E−2 (note that ξ −2 = ξ 58 since
p = 31) and it follows that
w−2 = a3 (8f1 + 28f2 + f3 + 10f4 + 7f5 + 20f6 + 18f7 + f8 ) + a4 eα0
where α0 = 2α1 + 3α2 + 4α3 + 6α4 + 5α5 + 4α6 + 3α7 + 2α8 is the highest root. Note that
a3 6= 0 since w−2 ∈ ker((x − 1)3 ) \ ker((x − 1)2 ) and eα0 ∈ CV (x).
By considering the action of x on W (see (15)) we quickly deduce that a2 = 16a1 and
a3 = 4a1 . Finally, one checks that the condition in (17) yields a4 = 0, so setting a1 = 1
we have
w2 = e1 + e2 + e3 + e4 + e5 + e6 + e7 + e8
w0 = 16(h1 + 19h2 + 4h3 + 9h4 + 28h5 + 18h6 + 10h7 + 4h8 )
w−2 = 4(8f1 + 28f2 + f3 + 10f4 + 7f5 + 20f6 + 18f7 + f8 )
and it is easy to see that w2 , w0 and w−2 satisfy the relations
[w2 , w−2 ] = 2w0 , [w2 , w0 ] = w2 , [w0 , w−2 ] = w−2 ,
(18)
20
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
and thus W = hw2 , w0 , w−2 i is an sl2 -subalgebra. Moreover, if we set
0
w20 = w2 , w00 = −2w0 , w−2
= −w−2 ,
(19)
0 ) is an sl -triple. Moreover, working mod p, we have
then (w20 , w00 , w−2
2
0
w−2
= 92f1 + 136f2 + 182f3 + 270f4 + 220f5 + 168f6 + 114f7 + 58f8
0 ) is an sl -triple over Z (see the proof of [34, Proposition 2.4]). Since
and thus (w20 , w00 , w−2
2
X stabilizes W , it is contained in an A1 -type subgroup of G by Proposition 2.14. This
completes the proof of Theorem 1 for G = E8 with p = 31.
We close this section by giving a proof of Theorem 2.
Proof of Theorem 2. Let V = Lie(G) be the adjoint module for G. Seeking a contradiction,
suppose X < P , where P = QL is a proper parabolic subgroup of G with unipotent
radical Q and Levi factor L. We may as well assume that P is minimal with respect to
the containment of X. In particular, if π : P → P/Q is the quotient map and we identify
L with P/Q, then π(X) is not contained in a proper parabolic subgroup of L0 . Now π(x)
is a regular unipotent element of L0 (see [35, Lemma 2.6]), so by combining Theorem 2.22
with the main theorem of [28], we deduce that π(X) is contained in an A1 -type subgroup
H of L0 . The KH-composition factors of V |H can be read off from the information in [15,
Tables 1–5] and we can use this to determine the KX-composition factors of V |X . Indeed,
each composition factor of V |P is an irreducible KL0 -module (the unipotent radical Q acts
trivially on the KP -composition factors of V |P ), so the decompositions of V |X and V |H
have to be compatible. But the decomposition of V |X is given in Table 2 and in this way
we will reach a contradiction.
To see this, first observe that X has at least one trivial composition factor on V , coming
from Z(L). By inspecting Table 2, this immediately implies that
(G, p) ∈ {(F4 , 13), (E6 , 13), (E8 , 37)}.
Suppose (G, p) = (E8 , 37). From Table 2, the KX-composition factors of V |X are as
follows:
LX (34)2 , LX (26)2 , LX (22)2 , LX (20), LX (14)2 , LX (10), LX (8), LX (2)2 , LX (0).
(20)
By inspecting [15, Table 5], using the fact that V |X has a unique trivial composition
factor, we deduce that L0 = A4 A2 A1 , A4 A3 or A5 A2 . However, in each of these cases
we see that V |X has an LX (6) composition factor, which is incompatible with (20). The
other two possibilities for (G, p) can be eliminated in a similar fashion. For example, if
(G, p) = (E6 , 13) then the composition factors of V |X are
LX (10)3 , LX (8)3 , LX (4), LX (2)4 , LX (0).
By inspecting [15, Table 3], just considering trivial composition factors, we deduce that
L0 = A22 A1 , A4 A1 or D5 , but in each case we find that V |X has two or more LX (4) factors.
This is a contradiction.
As mentioned above, the proof of Theorem 2.22 will be given in Sections 3–7, where
we carry out Steps 1 and 2 for each group in turn. We handle Step 3 in Section 8, thus
completing the proof of Theorem 1.
3. The case G = G2
We begin the proof of Theorem 1 by handling the case G = G2 . As noted in Remark
1(c), the result in this case can be deduced from the proof of [29, Lemma 3.1] (it also
follows from Kleidman’s classification of the maximal subgroups of G2 (p) in [13]).
A1 -TYPE SUBGROUPS
21
Theorem 3.1. Let G be a simple algebraic group of type G2 over an algebraically closed
field of characteristic p > 0. Let X = PSL2 (p) be a subgroup of G containing a regular
unipotent element x of G. Then X is contained in an A1 -type subgroup of G.
Proof. As noted in Section 1, we may assume that 7 6 p 6 19 (see (3)). Let V = Lie(G)
be the adjoint module for G and fix a Chevalley basis for V as in (12). We will use the
notation introduced in Section 2.5. In particular, hs̄, xi is a Borel subgroup of X, where
s̄ = t(ξ) = sZ and
{ξ 10 , ξ 2 }
(21)
i
are the eigenvalues of s ∈ SL2 (p) on CV (x), where F×
p = hξi. Let Ei be the ξ -eigenspace
of s on V and recall from Section 2.3 that we may assume x is obtained by exponentiating
the regular nilpotent element e = e1 + e2 ∈ V (that is, we will assume x = exp(ad(e))).
According to [14, Table 2], the Jordan form of x on V is as follows:
[J11 , J3 ] p > 11
(22)
[J72 ]
p=7
We will use the notation U and W (i) for the projective indecomposable KX-modules
defined in (4) and (5), respectively.
Case 1. V |X is semisimple
First assume V |X is semisimple and recall that x has Jordan form [Jm+1 ] on LX (m)
(for 0 6 m < p). In view of (22), it follows that
(
LX (10) ⊕ LX (2) p > 11
V |X =
LX (6)2
p = 7.
If p = 7 then the above decomposition implies that ξ 6 is an eigenvalue of s on CV (x), but
this is not compatible with (21).
Now assume p > 11, so V |X = LX (10) ⊕ LX (2). As in Section 2.5, let {w2 , w0 , w−2 }
be a standard basis for the summand W = LX (2), so wi ∈ Ei and the action of x on
W is given by the matrix in (15). Our goal is to show that W is an sl2 -subalgebra with
P
w2 = e and w−2 ∈ 2i=1 Zfi . Furthermore, we seek integers l, µ so that (w2 , lw0 , µw−2 ) is
an sl2 -triple over Z, which will allow us to apply Proposition 2.14.
For p > 17 we find that each space
ker(x − 1) ∩ E2 , ker((x − 1)2 ) ∩ E0 , ker((x − 1)3 ) ∩ E−2
(23)
is 1-dimensional, which gives us
w2 = a1 (e1 + e2 ), w0 = a2 (h1 + 13h2 ), w−2 = a3 (4f1 + f2 )
for some non-zero scalars ai ∈ K (in the expressions for w0 and w−2 , the specific coefficients
of the hi and fi will depend on the characteristic p; the coefficients presented here are for
p = 17). If we set a1 = 1 then by considering the action of x on V we deduce that a2 = 14
and a3 = 7. Now w2 , w0 and w−2 satisfy the relations in (18), so we get an sl2 -triple
0 ) as in (19). Here w 0
0
0
0
(w20 , w00 , w−2
−2 = −w−2 = 6f1 + 10f2 and thus (w2 , w0 , w−2 ) is an
sl2 -triple over Z (see the proof of [34, Proposition 2.4]). Finally, by applying Proposition
2.14, we conclude that X is contained in an A1 -type subgroup of G.
Next assume p = 13. Once again ker(x−1)∩E2 and ker((x−1)2 )∩E0 are 1-dimensional,
but now ker((x − 1)3 ) ∩ E−2 is 2-dimensional, spanned by the vectors 11f1 + f2 and e32
(here we use the notation e32 for eγ with γ = 3α1 + 2α2 ). Therefore
w2 = a1 (e1 + e2 ), w0 = a2 (h1 + 6h2 ), w−2 = a3 (11f1 + f2 ) + a4 e32
22
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
for some ai ∈ K. By considering the action of x on W we deduce that a2 = 10a1 and
a3 = 3a1 . Moreover, (17) implies that a4 = 0 and by arguing as above, setting a1 = 1 and
using Proposition 2.14, we deduce that X is contained in an A1 -type subgroup of G.
Now suppose p = 11. Here we have
w2 = a1 (e1 + e2 ), w0 = a2 (h1 + 9h2 ) + a3 e32 , w−2 = a4 (5f1 + f2 ) + a5 e31
and by considering the action of x on W we deduce that a2 = 8a1 , a4 = a1 and a5 = 2a3 .
We may as well set a1 = 1, so
w2 = e1 + e2 , w0 = 8(h1 + 9h2 ) + γe32 , w−2 = 5f1 + f2 + 2γe31
for some γ ∈ K. One now checks that the relations in (18) are satisfied (for all γ), so
W = hw2 , w0 , w−2 i is an sl2 -subalgebra. Moreover, if we take
e = w2 , h = −2(8(h1 + 9h2 )) = 6h1 + 4h2 , f = −(5f1 + f2 ) = 6f1 + 10f2 ,
then (e, h, f ) is an sl2 -triple over Z and we can apply Proposition 2.15 (with y = e32 and
z = e31 ). It follows that the stabilizer of W in G is an A1 -type subgroup.
Case 2. rad(V |X ) 6= 0, p > 11
To complete the proof of the theorem we may assume that rad(V |X ) 6= 0. Suppose
p > 11 and W is a reducible indecomposable summand of V |X . If p > 13 then dim W > 12
(see Corollary 2.4) and thus Lemma 2.1 implies that x has a Jordan block of size n > 12
on W , but this is incompatible with (22). Now assume p = 11. Here (22) implies that W
has at least three composition factors (if there were only two, then x would have Jordan
form [J10 ] or [J11 , J1 ] on W , which contradicts (22)). By Lemma 2.1, it follows that x has
Jordan form [J11 , Ji ] on W with i ∈ {0, 3}, so dim W ∈ {11, 14}. By considering Theorem
2.2, it is easy to see that i = 0 is the only possibility, so W = U is projective and thus
V |X = U ⊕ LX (2).
However, this implies that an involution x2 ∈ X has trace 2 on V (see Section 2.2), which
is incompatible with Proposition 2.8. This is a contradiction.
Case 3. rad(V |X ) 6= 0, p = 7
Finally, let us assume p = 7. Here V |X is projective and thus each indecomposable
summand is also projective. Since the eigenvalues of s on CV (x) are {ξ 4 , ξ 2 }, we deduce
that V |X = W (2) or W (4). In fact, by considering the trace of x2 , we see that V |X = W (2)
is the only option. This is compatible with the decomposition of V with respect to an
A1 -type subgroup of G containing a regular unipotent element (see Table 2).
Let W be the LX (2) summand in the socle of V |X and let {w2 , w0 , w−2 } be a standard
basis. The spaces ker(x − 1) ∩ E2 and ker((x − 1)2 ) ∩ E0 are 1-dimensional, whereas
ker((x − 1)3 ) ∩ E−2 is 2-dimensional and we get
w2 = a1 (e1 + e2 ), w0 = a2 (h1 + 4h2 ), w−2 = a3 (2f1 + f2 ) + a4 e32
for some ai ∈ K. Set a1 = 1, so w2 = e. By considering the action of x on W we
deduce that a2 = a3 = 4. Moreover, (17) implies that a4 = 0 and we deduce that
W = hw2 , w0 , w−2 i is an sl2 -subalgebra and the relations in (18) are satisfied. As before,
the desired result now follows by applying Proposition 2.14.
This completes the proof of Theorem 3.1.
A1 -TYPE SUBGROUPS
23
4. A reduction for G = F4
In this section our goal is to establish Theorem 2.22 when G = F4 . The proof of
Theorem 1 in this case will be completed in Section 8. Our main result is the following.
Theorem 4.1. Let G be a simple algebraic group of type F4 over an algebraically closed
field of characteristic p > 0. Let X = PSL2 (p) be a subgroup of G containing a regular
unipotent element x of G and set V = Lie(G). Then one of the following holds:
(i) X is contained in an A1 -type subgroup of G;
(ii) p = 13, V |X = W (10) ⊕ W (2) and X stabilizes a non-zero subalgebra of heα | α ∈
Φ+ (G)i.
Proof. In view of (3), we may assume that 13 6 p 6 43. We set up the standard notation
as before. In particular,
{ξ 22 , ξ 14 , ξ 10 , ξ 2 }
(24)
×
are the eigenvalues of s on CV (x), where Fp = hξi, and
[J , J15 , J11 , J3 ] p > 23
23
2 ,J ,J ]
[J19
p = 19
11 3
(25)
2
[J
,
J
,
J
]
p
= 17
15 3
17
4
[J13 ]
p = 13
is the Jordan form of x on V (see [14, Table 4]). We may assume that x is obtained by
exponentiating the regular nilpotent element e = e1 + e2 + e3 + e4 in V , with respect to a
Chevalley basis for V as in (12). It will also be useful to note that V |X is self-dual.
Case 1. V |X is semisimple
If p ∈ {13, 17, 19} then (25) implies that
2
LX (18) ⊕ LX (10) ⊕ LX (2) p = 19
V |X =
LX (16)2 ⊕ LX (14) ⊕ LX (2) p = 17
LX (12)4
p = 13
but none of these decompositions are compatible with the eigenvalues of s on CV (x) given
in (24). For example, if p = 19 then the given decomposition implies that the relevant
eigenvalues are {ξ 0 , ξ 0 , ξ 10 , ξ 2 }, but this contradicts (24).
Now assume p > 23, so
V |X = LX (22) ⊕ LX (14) ⊕ LX (10) ⊕ LX (2).
Let W be the LX (2) summand and let {w2 , w0 , w−2 } be a standard basis for W as in
Section 2.5, so wi ∈ Ei (the ξ i -eigenspace of s on V ) and the action of x and x0 on W
is given by the matrices in (15) and (16), respectively, where hs̄, x0 i is the opposite Borel
subgroup of X. If p > 29 then the spaces in (23) are 1-dimensional and we get
w2 = a1 (e1 + e2 + e3 + e4 )
w0 = a2 (h1 + 23h2 + 4h3 + 6h4 )
w−2 = a3 (5f1 + 28f2 + 20f3 + f4 )
for some ai ∈ K (in the expressions for w0 and w−2 , the specific coefficients depend on the
characteristic p; the ones given here are for p = 29). If we set a1 = 1 then we can use the
action of x on V to deduce that a2 = 18 and a3 = 13. Moreover, the relations in (18) are
0 ) is an sl -triple, where these elements are defined
satisfied and it follows that (w20 , w00 , w−2
2
in (19). Now
0
w−2
= −w−2 = −13(5f1 + 28f2 + 20f3 + f4 ) = 22f1 + 42f2 + 30f3 + 16f4
24
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
0 ) is an sl -triple over Z (see the proof of [34, Proposition
working mod p, so (w20 , w00 , w−2
2
2.4]). By applying Proposition 2.14, we conclude that X is contained in an A1 -type
subgroup of G.
Now suppose p = 23. Here
w2 = a1 (e1 + e2 + e3 + e4 )
w0 = a2 (h1 + 4h2 + 16h3 + 7h4 ) + a3 e2342
w−2 = a4 (10f1 + 17f2 + 22f3 + f4 ) + a5 e1342
for some ai ∈ K (we use the notation e2342 for eγ with γ = 2α1 + 3α2 + 4α3 + 2α4 , and
similarly for e1342 ). By considering the action of x on W we deduce that a2 = 12a1 ,
a4 = 7a1 and a5 = 2a3 . Setting a1 = 1 we get
w2 = e1 + e2 + e3 + e4
w0 = 12(h1 + 4h2 + 16h3 + 7h4 ) + γe2342
w−2 = 7(10f1 + 17f2 + 22f3 + f4 ) + 2γe1342
for some γ ∈ K, and one can check that the relations in (18) are satisfied. In particular,
W is an sl2 -subalgebra of V . Set
e = w2 , h = −2(12(h1 + 4h2 + 16h3 + 7h4 )) = 22h1 + 15h2 + 14h3 + 9h4
and
f = −(7(10f1 + 17f2 + 22f3 + f4 )) = 22f1 + 19f2 + 7f3 + 16f4 .
Then (e, h, f ) is an sl2 -triple over Z and by applying Proposition 2.15 (with y = e2342 and
z = e1342 ) we deduce that the stabilizer of W in G is an A1 -type subgroup.
Case 2. rad(V |X ) 6= 0, p > 19
For the remainder we may assume that rad(V |X ) 6= 0. First assume p > 19. By arguing
as in Case 2 in the proof of Theorem 3.1, it is straightforward to reduce to the case p = 19.
For example, suppose p = 23 and W is a reducible indecomposable summand of V |X . The
Jordan form of x on V (see (25)) implies that W has at least three composition factors and
we can use Lemma 2.1 to see that x has Jordan form [J23 , Ji ] on W with i ∈ {0, 3, 11, 15},
so dim W ∈ {23, 26, 34, 38}. Using Theorem 2.2, we deduce that i = 0 is the only option,
so
V |X = U ⊕ LX (14) ⊕ LX (10) ⊕ LX (2).
But this implies that an involution x2 ∈ X has trace 0 on V , which is incompatible with
Proposition 2.8.
Now assume p = 19. Suppose W is a reducible non-projective indecomposable summand
of V |X . By combining Lemma 2.1 and Theorem 2.2 we deduce that x has Jordan form
2 , J ] or [J 2 , J ] on W , so there is a unique such summand (and the other summand
[J19
11
19 3
is simple). However, this is incompatible with the self-duality of V |X . For example, if x
2 , J ] on W , then V | = W ⊕ L (10) and Theorem 2.2 implies that
has Jordan form [J19
3
X
X
soc(W ) = LX (0) ⊕ LX (2) ⊕ LX (4), W/soc(W ) = LX (16) ⊕ LX (14)
(up to duality) so V |X is not self-dual.
Therefore, we may assume that each indecomposable summand is either simple or projective, so the possibilities for V |X are as follows:
U ⊕ LX (18) ⊕ LX (10) ⊕ LX (2)
U 2 ⊕ LX (10) ⊕ LX (2)
W (i) ⊕ LX (10) ⊕ LX (2)
A1 -TYPE SUBGROUPS
25
with i ∈ {2, 4, . . . , 16}. As in (24), the eigenvalues of s on CV (x) are {ξ 4 , ξ 14 , ξ 10 , ξ 2 }.
Since s has eigenvalues 1 and {ξ i , ξ −i } on CU (x) and CW (i) (x), respectively (see Lemma
2.20), it follows that
V |X = W (i) ⊕ LX (10) ⊕ LX (2)
with i ∈ {4, 14}. The case i = 4 can be ruled out by considering the trace of x2 , hence
i = 14 and V |X is compatible with the containment of X in an A1 -type subgroup of G
(see Table 2). We need to show that X is contained in such a subgroup. To do this we
can repeat the argument in Case 1 for p > 29 (the details are entirely similar).
Case 3. rad(V |X ) 6= 0, p = 17
Now assume p = 17. Suppose W is a reducible non-projective indecomposable summand
2 , J ] or [J , J ],
of V |X . It is easy to check that the Jordan form of x on W is either [J17
3
17 15
2
so there is a unique such summand. If x has Jordan form [J17 , J3 ] on W then Theorem
2.2 implies that V |X = W ⊕ LX (14) (up to duality) where
soc(W ) = LX (0) ⊕ LX (2) ⊕ LX (4), W/soc(W ) = LX (16) ⊕ LX (14),
but this is incompatible with the self-duality of V |X . Similarly, in the other case we have
V |X = W ⊕ V1 ⊕ LX (2) and
soc(W ) = LX (i) ⊕ LX (i + 2), W/soc(W ) = LX (14 − i) ⊕ LX (12 − i)
with i ∈ {0, 2, 4, . . . , 12} and V1 ∈ {LX (16), U }. By self-duality, i = 6 is the only option.
But this implies that x2 has trace 0 on V , which contradicts Proposition 2.8.
It follows that each indecomposable summand of V |X is either simple or projective. By
arguing as above (the case p = 19), using the fact that s has eigenvalues {ξ 6 , ξ 14 , ξ 10 , ξ 2 }
on CV (x), we deduce that
V |X = W (i) ⊕ LX (14) ⊕ LX (2)
with i ∈ {6, 10}. If i = 6 then one can check that an element x3 ∈ X of order 3 has trace
1 on V , so Proposition 2.8 implies that i = 10. Therefore, the action of X is compatible
with an A1 -type subgroup of G (see Table 2) and it remains to establish the desired
containment.
As before, let {w2 , w0 , w−2 } be a standard basis of the LX (2) summand W in the
decomposition of V |X . In the usual manner we deduce that
w2 = a1 (e1 + e2 + e3 + e4 )
w0 = a2 (h1 + 5h2 + 6h3 + 10h4 )
for some non-zero scalars a1 , a2 ∈ K. We may assume a1 = 1. Now w−2 is contained in
ker((x − 1)3 ) ∩ E−2 , which is 2-dimensional, and we get
w−2 = a3 (12f1 + 9f2 + 4f3 + f4 ) + a4 (e1231 − e1222 ).
Since the action of x on W is given by the matrix in (15) we deduce that a2 = 6 and
a3 = 1. Finally, the condition in (17), which is obtained by considering the action of x0
on W , implies that a4 = 0. It is easy to see that the relations in (18) are satisfied and we
complete the argument in the usual manner, via Proposition 2.14.
Case 4. rad(V |X ) 6= 0, p = 13
Finally, let us assume that p = 13. Here V |X is projective and thus each indecomposable
summand is also projective. Since the eigenvalues of s on CV (x) are {ξ 10 , ξ 2 , ξ 10 , ξ 2 }, we
quickly deduce that V |X is one of the following:
W (2) ⊕ W (10), W (2)2 , W (10)2 .
26
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
Let y = ŷZ ∈ X be an element of order 7, where ŷ ∈ SL2 (13) is SL2 (K)-conjugate to
a diagonal matrix diag(ω, ω −1 ) and ω ∈ K is a non-trivial 7-th root of unity. For each
decomposition we can compute the eigenvalues of y on V and then compare the results with
the list of eigenvalue multiplicities of all elements in G of order 7 (as noted in Remark
2.10, the latter can be computed using Litterick’s algorithm in [21]). For example, if
V |X = W (2)2 then y ∈ GL52 (K) is conjugate to the diagonal matrix
[I8 , ωI8 , ω 2 I8 , ω 3 I6 , ω 4 I6 , ω 5 I8 , ω 6 I8 ],
but one checks that no element in G of order 7 acts on V with these eigenvalues. In this
way, we deduce that V |X = W (2) ⊕ W (10) is the only possibility.
Let W be the LX (2) summand in the socle of V |X and let {w2 , w0 , w−2 } be a standard
basis. The spaces ker(x − 1) ∩ E2 and ker((x − 1)2 ) ∩ E0 are 2-dimensional and we get
w2 = a1 (e1 + e2 + e3 + e4 ) + a2 (e1231 − e1222 )
w0 = a3 (h1 + 9h2 + 12h3 + 9h4 ) + a4 (e1221 + 10e1122 )
for some ai ∈ K. Finally, one checks that ker((x − 1)3 ) ∩ E−2 is 4-dimensional and we take
w−2 = a5 (3f1 + f2 + 10f3 + f4 ) + a6 (e1220 + 3e0122 ) + a7 (e1121 + 8e0122 ) + a8 e2342 .
In the usual manner, by considering the action of x on W , we get a3 = 2a1 , a4 = a2 ,
a5 = 10a1 and a7 = 2a2 + a6 . In addition, the condition in (17) yields the following system
of equations:
a21 a2 + 8a21 a6 = 0
7a21 a2 + 11a21 a6 = 0
11a21 a2 + 5a21 a6 = 0
4a21 a8 + 8a1 a22 + 7a1 a2 a6 = 0
If a1 6= 0 then these equations imply that a2 = a6 = a8 = 0, so
w2 = a1 (e1 + e2 + e3 + e4 )
w0 = −a1 (h1 + 9h2 + 12h3 + 9h4 )
w−2 = 10a1 (3f1 + f2 + 10f3 + f4 )
and by setting a1 = 1 we can use Proposition 2.14 to show that X is contained in an
A1 -type subgroup. On the other hand, if a1 = 0 then we can set a2 = 1, so
w2 = e1231 − e1222
w0 = e1221 + 10e1122
w−2 = a6 (e1220 + 3e0122 ) + (2 + a6 )(e1121 + 8e0122 ) + a8 e2342
One checks that [w2 , w0 ] = 0, so
0 = x0 · [w2 , w0 ] = [w2 + 2w0 + w−2 , w0 + w−2 ]
and we deduce that a6 = 7, hence
w−2 = 7e1220 + 9e1121 + 2e0122 + a8 e2342 .
It is now easy to check that W ⊆ heα | α ∈ Φ+ (G)i is a subalgebra, which gives case (ii)
in the statement of the theorem.
This completes the proof of Theorem 4.1.
A1 -TYPE SUBGROUPS
27
5. A reduction for G = E6
The following result, which we prove in this section, establishes Theorem 2.22 for groups
of type E6 .
Theorem 5.1. Let G be a simple adjoint algebraic group of type E6 over an algebraically
closed field of characteristic p > 0. Let X = PSL2 (p) be a subgroup of G containing a
regular unipotent element x of G and set V = Lie(G). Then one of the following holds:
(i) X is contained in an A1 -type subgroup of G;
(ii) p = 13, V |X is one of
W (10) ⊕ W (8) ⊕ W (2), W (10) ⊕ W (4) ⊕ W (2), W (10)2 ⊕ W (4)
and X stabilizes a non-zero subalgebra of heα | α ∈ Φ+ (G)i.
Proof. Here 13 6 p 6 43, V |X is self-dual and
{ξ 22 , ξ 16 , ξ 14 , ξ 10 , ξ 8 , ξ 2 }
(26)
F×
p
are the eigenvalues of s on CV (x), where
= hξi (see Section 2.5). By inspecting [14,
Table 6], we see that
[J , J17 , J15 , J11 , J9 , J3 ] p > 23
23
2 ,J ,J ,J ,J ]
[J19
p = 19
17 11 9 3
(27)
3
,
J
,
J
,
J
]
p = 17
[J
15 9 3
17
6 ]
[J13
p = 13
is the Jordan form of x on V . We freely adopt the notation introduced in Section 2.
Case 1. V |X is semisimple
If p ∈ {13, 17, 19} then
LX (18)2 ⊕ LX (16) ⊕ LX (10) ⊕ LX (8) ⊕ LX (2) p = 19
LX (16)3 ⊕ LX (14) ⊕ LX (8) ⊕ LX (2)
p = 17
V |X =
LX (12)6
p = 13
but not one of these decompositions is compatible with eigenvalues of s on CV (x) (see
(26)) so we may assume p > 23 and
VX = LX (22) ⊕ LX (16) ⊕ LX (14) ⊕ LX (10) ⊕ LX (8) ⊕ LX (2).
Let W be the LX (2) summand and let {w2 , w0 , w−2 } be a standard basis for W . If p > 29
then one checks that each of the spaces in (23) are 1-dimensional and the result quickly
follows via Proposition 2.14. For example, if p = 29 then
w2 = a1 (e1 + e2 + e3 + e4 + e5 + e6 )
w0 = a2 (h1 + 5h2 + 20h3 + 28h4 + 20h5 + h6 )
w−2 = a3 (f1 + 5f2 + 20f3 + 28f4 + 20f5 + f6 )
and by setting a1 = 1 and considering the action of x on W (see (15)), we deduce that
a2 = 21 and a3 = 13. One now checks that (w2 , −2w0 , −w−2 ) is an sl2 -triple over Z (see
the proof of [34, Proposition 2.4]) and by applying Proposition 2.14 we deduce that X is
contained in an A1 -type subgroup of G.
Now assume p = 23. Here ker((x − 1)2 ) ∩ E0 and ker((x − 1)3 ) ∩ E−2 are 2-dimensional
and we get
w2 = a1 (e1 + e2 + e3 + e4 + e5 + e6 )
w0 = a2 (h1 + 10h2 + 22h3 + 17h4 + 22h5 + h6 ) + a3 e122321
w−2 = a4 (f1 + 10f2 + 22f3 + 17f4 + 22f5 + f6 ) + a5 e112321
28
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
where a1 a2 a4 6= 0. Set a1 = 1. From the action of x on W we deduce that a2 = 15, a4 = 7
and a5 = 2a3 , so
w2 = e1 + e2 + e3 + e4 + e5 + e6
w0 = 15(h1 + 10h2 + 22h3 + 17h4 + 22h5 + h6 ) + γe122321
w−2 = 7(f1 + 10f2 + 22f3 + 17f4 + 22f5 + f6 ) + 2γe112321
for some γ ∈ K. If we take
e = w2 , h = −2(15(h1 + 10h2 + 22h3 + 17h4 + 22h5 + h6 ))
and
f = −7(f1 + 10f2 + 22f3 + 17f4 + 22f5 + f6 ) = 16f1 + 22f2 + 7f3 + 19f4 + 7f5 + 16f6
then (e, h, f ) is an sl2 -triple over Z and using Proposition 2.15 we conclude that X is
contained in an A1 -type subgroup of G.
Case 2. rad(V |X ) 6= 0, p > 19
If p > 23 then we can essentially repeat the argument in the proof of Theorem 4.1 (see
the first paragraph in Case 2). Indeed, it is easy to reduce to the case where p = 23 and
V |X = U ⊕ LX (16) ⊕ LX (14) ⊕ LX (10) ⊕ LX (8) ⊕ LX (2),
but this is not compatible with (26).
Now assume p = 19. Suppose V |X has a reducible non-projective indecomposable
summand W . By applying Lemma 2.1 and Theorem 2.2, we deduce that the Jordan form
of x on W is one of the following:
2
2
2
[J19
, J11 ], [J19
, J9 ], [J19
, J3 ], [J19 , J17 ].
In particular, V |X has a unique such summand. The structure of W is described in
Theorem 2.2 and it is easy to see that the existence of such a summand contradicts the
2 , J ] on W . Then up to
self-duality of V |X . For instance, suppose x has Jordan form [J19
9
duality we have
soc(W ) = LX (6) ⊕ LX (8) ⊕ LX (10), W/soc(W ) = LX (10) ⊕ LX (8)
and thus V |X = W ⊕ LX (16) ⊕ LX (10) ⊕ LX (2) is not self-dual. The other cases are very
similar.
Therefore, we may assume that each indecomposable summand of V |X is either simple
or projective. By considering the eigenvalues of s in (26), we deduce that
V |X = W (i) ⊕ LX (16) ⊕ LX (10) ⊕ LX (8) ⊕ LX (2)
with i ∈ {4, 14}. If i = 4 then we find that x2 has trace 2 on V , which contradicts
Proposition 2.8, hence i = 14 is the only possibility. In the usual manner, we now construct
a basis
w2 = a1 (e1 + e2 + e3 + e4 + e5 + e6 )
w0 = a2 (h1 + 18h2 + 9h3 + 5h4 + 9h5 + h6 )
w−2 = a3 (f1 + 18f2 + 9f3 + 5f4 + 9f5 + f6 ) + a4 (e112211 − e111221 )
(with a1 a2 a3 6= 0) of the summand W = LX (2) of V |X . If we set a1 = 1 and consider the
action of x on W (see (15)) we deduce that a2 = 11 and a3 = 3, and one checks that the
condition in (17) gives a4 = 0. The result now follows in the usual manner via Proposition
2.14.
Case 4. rad(V |X ) 6= 0, p = 17
A1 -TYPE SUBGROUPS
29
First assume that V |X has a reducible indecomposable non-projective summand W . In
the usual way, by combining Lemma 2.1 and Theorem 2.2, we deduce that the Jordan
form of x on W is one of the following:
3
2
2
[J17
, J3 ], [J17
, J9 ], [J17
, J3 ], [J17 , J15 ].
3 , J ] on W , so V | = W ⊕ L (14) ⊕ L (8). By
Suppose that x has Jordan form [J17
3
X
X
X
applying Theorem 2.2, using the self-duality of V |X , we deduce that
soc(W ) = W/soc(W ) = LX (10) ⊕ LX (8) ⊕ LX (6)
is the only possibility, but this is incompatible with the eigenvalues of s on CV (x). We can
2 , J ] and [J 2 , J ] by the self-duality of V | , so let us assume x has Jordan
rule out [J17
9
X
17 3
form [J17 , J15 ] on W . By self-duality it follows that
soc(W ) = W/soc(W ) = LX (8) ⊕ LX (6)
and thus V |X is one of the following:
(
W ⊕ M1 ⊕ M2 ⊕ LX (8) ⊕ LX (2)
W ⊕ W (i) ⊕ LX (8) ⊕ LX (2)
where Mj ∈ {LX (16), U } and i ∈ {2, 4, . . . , 14}. However, it is clear that none of these
decompositions are compatible with (26).
For the remainder, we may assume that each indecomposable summand of V |X is either
simple or projective. By considering the eigenvalues of s, we deduce that
V |X = W (i) ⊕ M1 ⊕ LX (14) ⊕ LX (8) ⊕ LX (2)
with i ∈ {6, 10} and M1 ∈ {LX (16), U }. By computing the trace of x3 and appealing to
Proposition 2.8 (and also Remark 2.9), it follows that i = 10 and M1 = LX (16) is the
only possibility. In particular, we have now reduced to the case where the decomposition
of V |X is compatible with containment in an A1 -type subgroup of G (see Table 2).
As before, let W be the LX (2) summand of V |X and let {w2 , w0 , w−2 } be a standard
basis of W . The reader can check that
w2 = a1 (e1 + e2 + e3 + e4 + e5 + e6 )
w0 = a2 (h1 + 12h2 + 4h3 + 9h4 + 4h5 + h6 ) + a3 (e112211 − e111221 )
w−2 = a4 (f1 + 12f2 + 4f3 + 9f4 + 4f5 + f6 ) + a5 (e112210 + e011221 )
+ a6 (e111211 + 15e011221 )
with a1 a2 a4 6= 0. Set a1 = 1. By considering the action of x on W we deduce that a2 = 9,
a4 = 1 and a6 = 2a3 + a5 . The condition in (17) yields a5 = 15a3 , so a6 = 0 and thus
w2 = e1 + e2 + e3 + e4 + e5 + e6
w0 = 9(h1 + 12h2 + 4h3 + 9h4 + 4h5 + h6 ) + γ(e112211 − e111221 )
w−2 = f1 + 12f2 + 4f3 + 9f4 + 4f5 + f6 + 15γ(e112210 + e011221 )
for some γ ∈ K. In addition, the relations in (18) are satisfied and W is an sl2 -subalgebra
of V . Set
e = w2 , h = −2(9(h1 + 12h2 + 4h3 + 9h4 + 4h5 + h6 ))
and
f = −(f1 + 12f2 + 4f3 + 9f4 + 4f5 + f6 ) = 16f1 + 5f2 + 13f3 + 8f4 + 13f5 + 16f6 .
Then (e, h, f ) is an sl2 -triple over Z and by applying Proposition 2.15 (with y = e112211 −
e111221 and z = −(e112210 + e011221 )) we conclude that X is contained in an A1 -type
subgroup of G.
30
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
Case 5. rad(V |X ) 6= 0, p = 13
Here (27) implies that V |X is projective, so each indecomposable summand is also
projective. In view of (26), we must have V |X = W (i) ⊕ W (j) ⊕ W (k) with i, j ∈ {2, 10}
and k ∈ {4, 8}. In each case, the traces of x2 and x3 are −2 and −3, respectively, so
we need to work harder to eliminate some of these decompositions. Let y = ŷZ ∈ X
be an element of order 7, where ŷ ∈ SL2 (13) is SL2 (K)-conjugate to a diagonal matrix
diag(ω, ω −1 ) and ω ∈ K is a non-trivial 7-th root of unity. We can compute the eigenvalues
of y on V and then compare with the eigenvalue multiplicities of all elements in G of order
7, which we obtain using the algorithm in [21]. In this way, we deduce that V |X is one of
the following:
W (10) ⊕ W (8) ⊕ W (2), W (10) ⊕ W (4) ⊕ W (2), W (10)2 ⊕ W (4).
Case 5(a). p = 13, V |X = W (10) ⊕ W (8) ⊕ W (2)
Here V |X is compatible with the containment of X in an A1 -type subgroup of G (see
Table 2). Let W be the LX (2) summand in the socle of V |X and let {w2 , w0 , w−2 } be a
standard basis. In the usual manner, we deduce that
w2 = a1 (e1 + e2 + e3 + e4 + e5 + e6 ) + a2 (e112210 + e111211 − e011221 )
w0 = a3 (h1 + 3h2 + 10h3 + h4 + 10h5 + h6 ) + a4 (e111210 + 3e111111 − e011211 )
w−2 = a5 (f1 + 3f2 + 10f3 + f4 + 10f5 + f6 ) + a6 (e111110 + 7e011210 − e011111 )
+ a7 (e101111 + 9e011210 ) + a8 (e122321 )
for some scalars ai ∈ K. By considering the action of x on W , together with the condition
in (17), we see that
a3 = 5a1 , a4 = a2 , a5 = 10a1 , a7 = 6a2 + 2a6
and either a1 = 0 or a2 = a6 = a8 = 0. In the latter situation, we set a1 = 1 and then
check that the relations in (18) are satisfied – this allows us to apply Proposition 2.14 to
conclude that X is contained in an A1 -type subgroup of G. Now assume a1 = 0 and set
a2 = 1. Here one checks that [w−2 , w2 ] = 0, so [w−2 , w2 +2w0 +w−2 ] = 0 since x0 preserves
the Lie bracket on V . This yields a6 = 9, so
w2 = e112210 + e111211 − e011221
w0 = e111210 + 3e111111 − e011211
(28)
w−2 = 9e111110 + 11e101111 + 6e011210 + 4e011111 + a8 (e122321 )
We conclude that W is an X-invariant subalgebra of heα | α ∈ Φ+ (G)i, as in part (ii) in
the statement of the theorem.
Case 5(b). p = 13, V |X = W (10) ⊕ W (4) ⊕ W (2) or W (10)2 ⊕ W (4)
Let W be the LX (4) summand in the socle of V |X and let {w4 , w2 , w0 , w−2 , w−4 } be a
basis of W with wi ∈ Ei . We may assume that the actions of x and x0 on W are given by
the matrices
1 1 1 1 1
1 0 0 0 0
0 1 2 3 4
4 1 0 0 0
0 0 1 3 6 and 6 3 1 0 0
(29)
0 0 0 1 4
4 3 2 1 0
0 0 0 0 1
1 1 1 1 1
A1 -TYPE SUBGROUPS
31
respectively (in terms of this basis). One checks that ker(x − 1) ∩ E4 is 1-dimensional,
whereas the following spaces
ker((x − 1)2 ) ∩ E2 , ker((x − 1)3 ) ∩ E0 , ker((x − 1)4 ) ∩ E−2 , ker((x − 1)5 ) ∩ E−4
have dimension 3, 3, 5 and 6, respectively, and we get
w4 = a1 (e112211 − e111221 )
w2 = a2 (e112210 + e011221 ) + a3 (e111211 + 11e011221 ) + a4 (e1 + e2 + e3 + e4 + e5 + e6 )
w0 = a5 (e111210 + e011211 ) + a6 (e111111 + 8e011211 ) + a7 (h1 + 3h2 + 10h3 + h4 + 10h5 + h6 )
w−2 = a8 (e111110 + e011111 ) + a9 (e101111 + 10e011111 ) + a10 (e011210 + 9e011111 )
+ a11 (f1 + 3f2 + 10f3 + f4 + 10f5 + f6 ) + a12 (e122321 )
w−4 = a13 (e111100 + 3e001111 ) + a14 (e101110 + e001111 ) + a15 (e011110 + 6e001111 )
+ a16 (e010111 + 3e001111 ) + a17 (12f101000 + 10f010100 + 2f001100 + 11f000110 + f000011 )
+ a18 (e112321 )
for some ai ∈ K.
Set a1 = 1 and consider the relations among the ai obtained from the action of x on
this basis. It is also helpful to note that x0 is a regular unipotent element, so CV (x0 ) is
abelian and thus [w−4 , [w−4 , w−2 ]] = 0. In this way, we deduce that
w4 = e112211 − e111221
w2 = a2 (e112210 + e011221 ) + (1 + a2 )(e111211 + 11e011221 )
w0 = 2a2 (e111210 + e011211 ) + (6 + 6a2 )(e111111 + 8e011211 )
w−2 = a8 (e111110 + e011111 ) + (4 + 5a2 + 2a8 )(e101111 + 10e011111 )
+ (6a2 + 12a8 )(e011210 + 9e011111 ) + a12 (e122321 )
w−4 = a13 (e111100 + 3e001111 ) + (2a2 + 8a8 + a13 )(e101110 + e001111 )
+ (11a2 + 9a8 )(e011110 + 6e001111 ) + (1 + 2a2 + 2a8 + 12a13 )(e010111 + 3e001111 )
+ 4a12 (e112321 )
Next one checks that [w2 , w−2 ] = 0, so x0 · [w2 , w−2 ] = 0 and thus
[w2 + 3w0 + 3w−2 + w−4 , w−2 + w−4 ] = 0
since x0 preserves the Lie bracket. This yields a13 = 12a2 + 2a22 + 12a2 a8 . Similarly,
[w4 , w2 ] = 0 and thus
[w4 + 4w2 + 6w0 + 4w−2 + w−4 , w2 + 3w0 + 3w−2 + w−4 ] = 0.
This relation implies that a22 + 12a2 a8 + 9a2 + 12a8 = 0 and it is now straightforward to
check that W ⊆ heα | α ∈ Φ+ (G)i is a subalgebra.
This completes the proof of Theorem 5.1.
6. A reduction for G = E7
In this section we establish the following result, which proves Theorem 2.22 for groups
of type E7 .
Theorem 6.1. Let G be a simple adjoint algebraic group of type E7 over an algebraically
closed field of characteristic p > 0. Let X = PSL2 (p) be a subgroup of G containing a
regular unipotent element x of G and set V = Lie(G). Then one of the following holds:
(i) X is contained in an A1 -type subgroup of G;
32
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
(ii) p = 19, V |X is one of
W (8) ⊕ W (4) ⊕ W (2) ⊕ U, W (16) ⊕ W (10) ⊕ W (4) ⊕ U,
W (16) ⊕ W (14) ⊕ W (8) ⊕ U
and X stabilizes a non-zero subalgebra of heα | α ∈ Φ+ (G)i.
Proof. Recall that we may assume 19 6 p 6 67 (see (3)) and
{ξ 34 , ξ 26 , ξ 22 , ξ 18 , ξ 14 , ξ 10 , ξ 2 }
(30)
F×
p
is the collection of eigenvalues of s on CV (x), where
= hξi. By [14, Table 8], the Jordan
form of x on V is as follows:
[J , J27 , J23 , J19 , J15 , J11 , J3 ] p > 37
35
2 ,J ,J ,J ,J ,J ]
p = 31
[J31
23 19 15 11 3
2 ,J ,J ,J ,J ,J ]
[J29
p = 29
(31)
27 19 15 11 3
5
[J , J15 , J3 ]
p = 23
23
7 ]
[J19
p = 19
Note that V |X is self-dual.
Case 1. V |X is semisimple
If p < 37 then the eigenvalues of s on CV (x) are incompatible with (30), so we may
assume p > 37 and thus
V |X = LX (34) ⊕ LX (26) ⊕ LX (22) ⊕ LX (18) ⊕ LX (14) ⊕ LX (10) ⊕ LX (2)
in view of (31). Let W be the LX (2) summand and let {w2 , w0 , w−2 } be a standard basis
for W . In the usual manner, it is straightforward to show that W is an appropriate sl2 subalgebra and we can use Proposition 2.14 to show that (i) holds in the statement of the
theorem. For example, if p = 37 we get
w2 = a1 (e1 + e2 + e3 + e4 + e5 + e6 + e7 )
w0 = a2 (h1 + 33h2 + 15h3 + 5h4 + 12h5 + 32h6 + 28h7 )
w−2 = a3 (4f1 + 21f2 + 23f3 + 20f4 + 11f5 + 17f6 + f7 ) + a4 e2234321
If we set a1 = 1, then by considering the action of x on W , we deduce that a2 = 20
and a3 = 10. Furthermore, the relation in (17) implies that a4 = 0 and we deduce that
(w2 , −2w0 , −w−2 ) is an sl2 -triple over Z (see the proof of [34, Proposition 2.4]). Now apply
Proposition 2.14.
Case 2. rad(V |X ) 6= 0, p > 29
If p > 37 then a combination of Lemma 2.1 and Corollary 2.4 implies that x has a
Jordan block of size n > 36 on V , but this contradicts (31).
Next assume p = 31. In the usual way, by applying Lemma 2.1 and Theorem 2.2, and by
appealing to the self-duality of V |X , we can reduce to the case where each indecomposable
summand of V |X is either simple or projective. By considering the eigenvalues in (30), it
follows that
V |X = W (i) ⊕ LX (22) ⊕ LX (18) ⊕ LX (14) ⊕ LX (10) ⊕ LX (2)
with i ∈ {4, 26}. If i = 4 then an involution x2 ∈ X has trace −3 on V , which contradicts
Proposition 2.8. Therefore i = 26 and it is entirely straightforward to show that the LX (2)
summand of V |X is an appropriate sl2 -subalgebra. The result follows via Proposition 2.14
in the usual fashion.
A1 -TYPE SUBGROUPS
33
A similar argument applies when p = 29. If V |X has a reducible non-projective summand then the self-duality of V |X implies that
V |X = M1 ⊕ M2 ⊕ LX (18) ⊕ LX (14) ⊕ LX (10) ⊕ LX (2)
is the only possibility, where
soc(M1 ) ∼
= M1 /soc(M1 ) = LX (12) ⊕ LX (14)
and M2 ∈ {LX (28), U }. However, this implies that x2 has trace −3 on V , which is a
contradiction. Therefore, the indecomposable summands of V |X are simple or projective,
and by considering the eigenvalues in (30) we deduce that
V |X = W (i) ⊕ LX (26) ⊕ LX (18) ⊕ LX (14) ⊕ LX (10) ⊕ LX (2)
with i ∈ {6, 22}. We can rule out i = 6 by computing the trace of x3 , so i = 22 and we
complete the argument as in the previous case.
Case 3. rad(V |X ) 6= 0, p = 23
As before, it is not difficult to reduce to the case where each indecomposable summand
of V |X is either simple or projective. By considering the eigenvalues in (30) we deduce
that
V |X = W (i) ⊕ W (j) ⊕ M1 ⊕ LX (14) ⊕ LX (2)
where i ∈ {4, 18}, j ∈ {10, 12} and M1 ∈ {LX (22), U }. By computing the trace of x2 we
see that (i, j) = (4, 12) or (18, 10), and we can rule out the first possibility by considering
the trace of x3 . This calculation with x3 also implies that M1 = LX (22), so
V |X = W (18) ⊕ W (10) ⊕ LX (22) ⊕ LX (14) ⊕ LX (2).
Let W be the LX (2) summand and fix a standard basis {w2 , w0 , w−2 }. By considering
the spaces
ker(x − 1) ∩ E2 , ker((x − 1)2 ) ∩ E0 , ker((x − 1)3 ) ∩ E−2
we deduce that
w2 = a1 (e1 + e2 + e3 + e4 + e5 + e6 + e7 )
w0 = a2 (h1 + 17h2 + 6h3 + 15h4 + 11h5 + 11h6 + 15h7 )
+ a3 (e1223210 + e1123211 − e1122221 )
w−2 = a4 (20f1 + 18f2 + 5f3 + f4 + 13f5 + 13f6 + f7 )
+ a5 (e1123210 + 2e1122211 + 20e1112221 )
Setting a1 = 1 and using the action of x on W , we deduce that a2 = 6, a4 = 19 and
a5 = 2a3 , so we have
w2 = e1 + e2 + e3 + e4 + e5 + e6 + e7
w0 = 6(h1 + 17h2 + 6h3 + 15h4 + 11h5 + 11h6 + 15h7 )
+ γ(e1223210 + e1123211 − e1122221 )
w−2 = 19(20f1 + 18f2 + 5f3 + f4 + 13f5 + 13f6 + f7 )
+ 2γ(e1123210 + 2e1122211 + 20e1112221 )
for some γ ∈ K. One can check that the relations in (18) are satisfied, so W is an
sl2 -subalgebra of V . Set
e = w2 , h = −2(6(h1 + 17h2 + 6h3 + 15h4 + 11h5 + 11h6 + 15h7 )),
f = −19(20f1 + 18f2 + 5f3 + f4 + 13f5 + 13f6 + f7 )
= 11f1 + 3f2 + 20f3 + 4f4 + 6f5 + 6f6 + 4f7
34
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
and
y = e1223210 + e1123211 − e1122221 , z = e1123210 + 2e1122211 + 20e1112221 .
Then (e, h, f ) is an sl2 -triple over Z and we can use Proposition 2.15 to deduce that X is
contained in an A1 -type subgroup of G.
Case 4. rad(V |X ) 6= 0, p = 19
Finally, let us assume p = 19 so V |X is projective and each indecomposable summand
is also projective. By considering the eigenvalues in (30), it follows that
V |X = W (i) ⊕ W (j) ⊕ W (k) ⊕ M1
where i ∈ {2, 16}, j ∈ {4, 14}, k ∈ {8, 10} and M1 ∈ {LX (18), U }. By computing the
trace of x2 we see that (i, j, k, M1 ) is one of the following:
(2, 14, 10, LX (18)), (16, 4, 8, LX (18)), (16, 4, 10, U ), (16, 14, 8, U ), (2, 4, 8, U ).
In all of these cases, x3 has trace 2 on V , which is compatible with Proposition 2.8. If
V |X = W (16) ⊕ W (4) ⊕ W (8) ⊕ LX (18) then there is an element y ∈ X of order 5 with
eigenvalues [I25 , ωI27 , ω 2 I27 , ω 3 I27 , ω 4 I27 ] on V , but one checks that there are no elements
in G that act on V in this way (for example, see [7, Table 6]), so this possibility is ruled
out.
If V |X is one of
W (8) ⊕ W (4) ⊕ W (2) ⊕ U, W (16) ⊕ W (10) ⊕ W (4) ⊕ U,
W (16) ⊕ W (14) ⊕ W (8) ⊕ U
then X stabilizes the 1-dimensional subalgebra of V spanned by the vector
w = e1122111 − e1112211 + e0112221 .
Indeed, X stabilizes soc(U ) = LX (0), which is spanned by a vector in CV (x) ∩ E0 . But
one checks that CV (x) ∩ E0 = hwi so we are in case (ii) in the statement of the theorem.
Finally, suppose V |X = W (2) ⊕ W (14) ⊕ W (10) ⊕ LX (18), which is compatible with
the containment of X in an A1 -type subgroup of G (see Table 2). Let W be the LX (2)
summand in the socle of V |X and let {w2 , w0 , w−2 } be a standard basis. In the usual way
we obtain
w2 = a1 (e1 + e2 + e3 + e4 + e5 + e6 + e7 )
w0 = a2 (h1 + 2h2 + 12h3 + 14h4 + 5h5 + 6h6 + 17h7 ) + a3 (e1122111 − e1112211 + e0112221 )
w−2 = a4 (9f1 + 18f2 + 13f3 + 12f4 + 7f5 + 16f6 + f7 )
+ a5 (e1122110 − e1112210 + 13e1112111 + 12e0112211 ) + a6 (e2234321 )
and we may assume a1 = 1. By considering the action of x on W , together with the
condition in (17), we deduce that a2 = 2, a4 = 11, a5 = 16a3 and a6 = 13a23 , so we have
w2 = e1 + e2 + e3 + e4 + e5 + e6 + e7
w0 = 2(h1 + 2h2 + 12h3 + 14h4 + 5h5 + 6h6 + 17h7 ) + γ(e1122111 − e1112211 + e0112221 )
w−2 = 11(9f1 + 18f2 + 13f3 + 12f4 + 7f5 + 16f6 + f7 )
+ 16γ(e1122110 − e1112210 + 13e1112111 + 12e0112211 ) + 13γ 2 (e2234321 )
for some γ ∈ K. Set
e = w2 , h = −2(2(h1 + 2h2 + 12h3 + 14h4 + 5h5 + 6h6 + 17h7 )),
f = −11(9f1 + 18f2 + 13f3 + 12f4 + 7f5 + 16f6 + f7 )
= 15f1 + 11f2 + 9f3 + f4 + 18f5 + 14f6 + 8f7
A1 -TYPE SUBGROUPS
35
and
y = e1122111 − e1112211 + e0112221
z1 = 8(e1122110 − e1112210 + 13e1112111 + 12e0112211 )
z2 = 11e2234321
Then (e, h, f ) is an sl2 -triple over Z, but we cannot directly apply Proposition 2.15. However, a minor modification of the argument in the proof of that proposition will work.
First observe that y ∈ (LZ )p−1 ∩ CLZ (e), z1 ∈ (LZ )p−3 and z2 ∈ (LZ )2p−4 (in terms of
the notation used in the proof of Proposition 2.15). Moreover,
(w2 , −2w0 , −w−2 ) = (ē, h̄ + δ ȳ, f¯ + δz̄1 + δ 2 z̄2 )
is an sl2 -triple in LK , where δ = −2γ. Set g = exp(ad(δy)) ∈ G and note that
1
g · ē = ē, g · h̄ = h̄ + δ[ȳ, h̄] = h̄ + δ ȳ, g · f¯ = f¯ + δ[ȳ, f¯] + δ 2 [ȳ, [ȳ, f¯]]
2
(for the final equality, note that all higher degree terms are zero since the maximum
T -weight on LZ is 2ht(α0 ) 6 2(p − 1)). Now calculating (in LZ ), we have
[h + y, f + z1 + z2 ] = −2f + (p − 3)z1 + (2p − 4)z2 + [y, f ] + [y, z1 ]
and passing to LK we get
−2f¯ − 3z̄1 − 4z̄2 + [ȳ, f¯] + [ȳ, z̄1 ] = −2(f¯ + z̄1 + z̄2 ).
Therefore [ȳ, f¯] + [ȳ, z̄1 ] = z̄1 + 2z̄2 and by comparing T -weights we deduce that [ȳ, f¯] = z̄1
and [ȳ, z̄1 ] = 2z̄2 . Finally, this implies that
1
g · f¯ = f¯ + δz̄1 + δ 2 [ȳ, z̄1 ] = f¯ + δz̄1 + δ 2 z̄2
2
and we can now conclude as in the proof of Proposition 2.15. In particular, X is contained
in an A1 -type subgroup of G.
This completes the proof of Theorem 6.1.
7. A reduction for G = E8
In this section we complete the proof of the Reduction Theorem (see Theorem 2.22).
Our main result is the following:
Theorem 7.1. Let G be a simple algebraic group of type E8 over an algebraically closed
field of characteristic p > 0. Let X = PSL2 (p) be a subgroup of G containing a regular
unipotent element x of G and set V = Lie(G). Then one of the following holds:
(i) X is contained in an A1 -type subgroup of G;
(ii) p = 37, V |X = W (34) ⊕ W (26) ⊕ W (14) ⊕ LX (22) ⊕ LX (2) and X stabilizes a
non-zero subalgebra of heα | α ∈ Φ+ (G)i.
Proof. First note that we may assume that 31 6 p 6 113 (see (3)). In fact, we may assume
p > 37 since the case p = 31 was handled in Section 2 (see Examples 2.21 and 2.23). Recall
that
{ξ 58 , ξ 46 , ξ 38 , ξ 34 , ξ 26 , ξ 22 , ξ 14 , ξ 2 }
(32)
36
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
is the collection of eigenvalues of s on CV (x) and note that V |X is self-dual. The Jordan
form of x on V is as follows:
[J59 , J47 , J39 , J35 , J27 , J23 , J15 , J3 ] p > 59
2 ,J ,J ,J ,J ,J ,J ]
[J53
p = 53
39 35 27 23 15 3
3
p = 47
[J47 , J39 , J27 , J23 , J15 , J3 ]
4 ,J ,J ,J ,J ]
[J43
p = 43
(33)
35 23 15 3
4 ,J ,J ,J ,J ]
[J
p
=
41
41 39 27 15 3
6 ,J ,J ]
p = 37
[J37
23 3
8
[J31 ]
p = 31
(see [14, Table 9]).
Case 1. V |X is semisimple
By considering the eigenvalues in (32) we deduce that p > 59 and
V |X = LX (58) ⊕ LX (46) ⊕ LX (38) ⊕ LX (34) ⊕ LX (26) ⊕ LX (22) ⊕ LX (14) ⊕ LX (2).
Let W be the LX (2) summand and let {w2 , w0 , w−2 } be a standard basis. If p > 61
then it is straightforward to show that W is an appropriate sl2 -subalgebra and the result
follows by applying Proposition 2.14 (note that if p = 61 then ker((x − 1)3 ) ∩ E−2 is
2-dimensional, but this does not cause any special difficulties). Now assume p = 59. Here
ker((x − 1)2 ) ∩ E0 and ker((x − 1)3 ) ∩ E−2 are both 2-dimensional and we get
w2 = a1 (e1 + e2 + e3 + e4 + e5 + e6 + e7 + e8 )
w0 = a2 (h1 + 22h2 + 52h3 + 35h4 + 46h5 + 48h6 + 41h7 + 25h8 ) + a3 (e23465432 )
w−2 = a4 (26f1 + 41f2 + 54f3 + 25f4 + 16f5 + 9f6 + 4f7 + f8 ) + a5 (e23465431 )
Set a1 = 1 and consider the action of x on W (see (15)). We deduce that a2 = 13, a4 = 1
and a5 = 2a3 , so
w2 = e1 + e2 + e3 + e4 + e5 + e6 + e7 + e8
w0 = 13(h1 + 22h2 + 52h3 + 35h4 + 46h5 + 48h6 + 41h7 + 25h8 ) + γ(e23465432 )
w−2 = 26f1 + 41f2 + 54f3 + 25f4 + 16f5 + 9f6 + 4f7 + f8 + 2γ(e23465431 )
for some γ ∈ K. Set
e = w2 , h = −2(13(h1 + 22h2 + 52h3 + 35h4 + 46h5 + 48h6 + 41h7 + 25h8 ))
and
f = −(26f1 + 41f2 + 54f3 + 25f4 + 16f5 + 9f6 + 4f7 + f8 )
= 33f1 + 18f2 + 5f3 + 34f4 + 43f5 + 50f6 + 55f7 + 58f8
Then (e, h, f ) is an sl2 -triple over Z (see the proof of [34, Proposition 2.4]) and by applying
Proposition 2.15 (with y = e23465432 and z = e23465431 ) we deduce that X is contained in
an A1 -type subgroup of G.
Case 2. rad(V |X ) 6= 0, p > 53
If p > 61 then the dimension of each indecomposable summand of V |X is at least 60,
which implies that the Jordan form of x has a block of size n > 60. This is a contradiction.
Next assume p = 59. Suppose W is a reducible indecomposable non-projective summand of V |X , so dim W > 58 (see Corollary 2.4). In view of (33) and Lemma 2.1, we
deduce that x has Jordan form [J59 , Ji ] on W for some i ∈ {3, 15, . . . , 47}. But this implies that dim W is even, so W has at least four composition factors and thus i > 57
A1 -TYPE SUBGROUPS
37
(again, by Corollary 2.4). This is a contradiction. Therefore, we may assume that each
indecomposable summand of V |X is either simple or projective. Clearly,
V |X = U ⊕ LX (46) ⊕ LX (38) ⊕ LX (34) ⊕ LX (26) ⊕ LX (22) ⊕ LX (14) ⊕ LX (2)
is the only possibility. However, this implies that x2 has trace −4 on V , which contradicts
Proposition 2.8.
Now assume p = 53. As in the previous case, by applying Lemma 2.1 and Theorem
2.2, and by appealing to the self-duality of V |X , it is straightforward to reduce to the case
where the indecomposable summands of V |X are either simple or projective. Moreover,
by considering the eigenvalues in (32), we deduce that
V |X = W (i) ⊕ LX (38) ⊕ LX (34) ⊕ LX (26) ⊕ LX (22) ⊕ LX (14) ⊕ LX (2)
with i ∈ {6, 46}. By computing the trace of x3 ∈ X, it follows that i = 46. It is now
entirely straightforward to show that the LX (2) summand of V |X is an appropriate sl2 subalgebra and the result follows via Proposition 2.14.
Case 3. rad(V |X ) 6= 0, p = 47
As in the previous case, we can quickly reduce to the situation where each indecomposable summand of V |X is simple or projective, in which case
V |X = W (i) ⊕ M2 ⊕ LX (38) ⊕ LX (26) ⊕ LX (22) ⊕ LX (14) ⊕ LX (2)
with i ∈ {12, 34} and M1 ∈ {LX (46), U }. By computing the trace of x2 we deduce that
i = 34 and M1 = LX (46), in which case V |X is compatible with the containment of X in
an A1 -type subgroup of G (see Table 2).
As usual, let W be the LX (2) summand of V |X and let {w2 , w0 , w−2 } be a standard
basis for W . We get
w2 = a1 (e1 + e2 + e3 + e4 + e5 + e6 + e7 + e8 )
w0 = a2 (h1 + 26h2 + 3h3 + 6h4 + 31h5 + 10h6 + 37h7 + 18h8 ) + a3 (e23354321 − e22454321 )
w−2 = a4 (34f1 + 38f2 + 8f3 + 16f4 + 20f5 + 11f6 + 36f7 + f8 )
+ a5 (e22354321 − 2e13354321 )
We may set a1 = 1. By considering the action of x on this basis we deduce that a2 = 1,
a4 = 36 and a5 = 45a3 , so
w2 = e1 + e2 + e3 + e4 + e5 + e6 + e7 + e8
w0 = h1 + 26h2 + 3h3 + 6h4 + 31h5 + 10h6 + 37h7 + 18h8 + γ(e23354321 − e22454321 )
w−2 = 36(34f1 + 38f2 + 8f3 + 16f4 + 20f5 + 11f6 + 36f7 + f8 )
+ 45γ(e22354321 − 2e13354321 )
for some γ ∈ K. One now checks that the relations in (18) are satisfied and thus W is an
sl2 -subalgebra of V . Set
e = w2 , h = −2(h1 + 26h2 + 3h3 + 6h4 + 31h5 + 10h6 + 37h7 + 18h8 )
and
f = −36(34f1 + 38f2 + 8f3 + 16f4 + 20f5 + 11f6 + 36f7 + f8 )
= 45f1 + 42f2 + 41f3 + 35f4 + 32f5 + 27f6 + 20f7 + 11f8
Then (e, h, f ) is an sl2 -triple over Z and we can use Proposition 2.15 to conclude that X
is contained in an A1 -type subgroup of G.
Case 4. rad(V |X ) 6= 0, p = 43
38
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
By arguing in the usual manner, it is straightforward to reduce to the case where
each indecomposable summand of V |X is either simple or projective. By considering the
eigenvalues in (32), we deduce that
V |X = W (i) ⊕ W (j) ⊕ LX (34) ⊕ LX (22) ⊕ LX (14) ⊕ LX (2)
with i ∈ {4, 38} and j ∈ {16, 26}. By computing the trace of x2 , we see that (i, j) = (38, 26)
is the only option, in which case V |X is compatible with the desired containment of X in
an A1 -type subgroup of G. As usual, we now construct the summand W = LX (2) of V |X
in terms of a standard basis {w2 , w0 , w−2 }; it is easy to show that W is an appropriate
sl2 -subalgebra and we can conclude by applying Proposition 2.14.
Case 5. rad(V |X ) 6= 0, p = 41
First assume that V |X has a reducible indecomposable non-projective summand W . In
the usual way, we deduce that the Jordan form of x on W is one of the following:
4
4 ,J ]
, J27 ], [J41
15
[J41
3
[J41 , J3 ]
2 , J ], [J 2 , J ]
[J 2 , J27 ], [J41
15
41 3
41
[J41 , J39 ],
4 , J ] or [J 4 , J ] then there is a unique such summand.
If the Jordan form is either [J41
27
41 15
Moreover, W has an odd number of composition factors and it is easy to see that this is
incompatible with the self-duality of V |X . Similar reasoning rules out the cases where x
2 , J ]. Finally, suppose x has Jordan form [J 3 , J ] or [J , J ]. Here
has Jordan form [J41
i
41 39
41 3
the self-duality of V |X implies that
soc(W ) ∼
= W/soc(W ) = LX (22) ⊕ LX (20) ⊕ LX (18)
or
soc(W ) ∼
= W/soc(W ) = LX (20) ⊕ LX (18),
respectively. However, the existence of such a summand would mean that ξ 20 is an eigenvalue of s on CV (x), which is not the case (see (32)). Therefore, we conclude that every
indecomposable summand of V |X is either simple or projective. More precisely, in view
of (32), it follows that
V |X = W (i) ⊕ W (j) ⊕ LX (38) ⊕ LX (26) ⊕ LX (14) ⊕ LX (2)
with i ∈ {6, 34} and j ∈ {18, 22}. By computing the trace of x3 we deduce that
(i, j) = (34, 22), in which case V |X is compatible with the containment of X in an A1 -type
subgroup.
Let W be the LX (2) summand of V |X and let {w2 , w0 , w−2 } be a standard basis. In
the usual manner we deduce that
w2 = a1 (e1 + e2 + e3 + e4 + e5 + e6 + e7 + e8 )
w0 = a2 (h1 + 30h2 + 10h3 + 27h4 + 22h5 + 25h6 + 36h7 + 14h8 )
w−2 = a3 (3f1 + 8f2 + 30f3 + 40f4 + 25f5 + 34f6 + 26f7 + f8 )
+ a4 (e22343221 − e12343321 + e12244321 )
Set a1 = 1. By considering the action of x on this basis we get a2 = 36 and a3 = 24.
Finally, one can check that the condition in (17) implies that a4 = 0 and now the desired
result follows from Proposition 2.14.
Case 6. rad(V |X ) 6= 0, p = 37
A1 -TYPE SUBGROUPS
39
To complete the proof of the theorem, we may assume that p = 37 (recall that the case
p = 31 was handled earlier in Examples 2.21 and 2.23). As usual, let us first assume that
V |X has a reducible indecomposable non-projective summand W . By applying Lemma
2.1 and Theorem 2.2, we deduce that the Jordan form of x on W is one of the following:
6
4
3
2
2
[J37
, J23 ], [J37
, J23 ], [J37
, J3 ], [J37
, J23 ], [J37
, J3 ].
3 , J ] is the only possibility, with
In fact, the self-duality of V |X implies that [J37
3
soc(W ) ∼
= W/soc(W ) = LX (16) ⊕ LX (18) ⊕ LX (20).
But if this is a summand of V |X then ξ 20 is an eigenvalue of s on CV (x), contradicting
(32). Therefore, we have reduced to the case where each indecomposable summand of V |X
is simple or projective. Again, by considering (32) we deduce that
V |X = W (i) ⊕ W (j) ⊕ W (k) ⊕ LX (22) ⊕ LX (2)
(34)
with i ∈ {2, 34}, j ∈ {10, 26} and k ∈ {14, 22}.
We claim that (i, j, k) = (34, 26, 14), in which case the decomposition of V |X is compatible with the containment of X in an A1 subgroup of G. One can check that all of the
eight above decompositions are compatible with the trace of x2 and x3 , so we consider the
traces of elements of larger order. Let y ∈ X be an element of order 19. In each case it is
straightforward to compute the eigenvalues of y on V . Using Litterick’s algorithm in [21],
we can compute the eigenvalues on V of every element in G of order 19 and in this way
we deduce that (i, j, k) = (34, 26, 14) as claimed.
Let W be the LX (2) summand of V |X with standard basis {w2 , w0 , w−2 }. The spaces
ker(x − 1) ∩ E2 , ker((x − 1)2 ) ∩ E0 , ker((x − 1)3 ) ∩ E−2
have respective dimensions 2, 2 and 3, which gives
w2 = a1 (e1 + e2 + e3 + e4 + e5 + e6 + e7 + e8 ) + a2 (e22343221 − e12343321 + e12244321 )
w0 = a3 (h1 + 24h2 + 6h3 + 15h4 + 4h5 + 34h6 + 31h7 + 32h8 )
+ a4 (e22343211 + 24e12343221 + 25e12243321 )
w−2 = a5 (22f1 + 10f2 + 21f3 + 34f4 + 14f5 + 8f6 + 16f7 + f8 )
+ a6 (e22343210 + 13e12243221 − e12233321 ) + a7 (e12343211 + 23e12243221 + 2e12233321 )
By considering the action of x on this basis, we deduce that a3 = 28a1 , a4 = 3a2 , a5 = 16a1
and a7 = 6a2 + a6 . The condition in (17) yields the equations
16a21 a2 + 3a21 a6 = 0
19a21 a2 + 6a21 a6 = 0
If a1 6= 0 then these equations imply that a2 = a6 = 0, so we can set a1 = 1 and then
apply Proposition 2.14 to show that X is contained in an A1 -type subgroup. On the other
hand, if a1 = 0 then we may assume a2 = 1, so
w2 = e22343221 − e12343321 + e12244321
w0 = 3(e22343211 + 24e12343221 + 25e12243321 )
w−2 = a6 (e22343210 + 13e12243221 − e12233321 ) + (6 + a6 )(e12343211 + 23e12243221 + 2e12233321 )
It is straightforward to check that W ⊆ heα | α ∈ Φ+ (G)i is a subalgebra and this puts us
in case (ii) of the theorem.
This completes the proof of Theorem 7.1.
40
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
8. Proof of Theorem 1
In this final section we complete the proof of Theorem 1. In view of Theorem 3.1, we
may assume that G is of type F4 , E6 , E7 or E8 . Moreover, by our work in Sections 4–7,
it remains to handle the cases appearing in Table 4. In each of these cases, X stabilizes a
non-zero subalgebra W ⊆ heα | α ∈ Φ+ (G)i of V = Lie(G) and by applying Proposition
2.18 we can assume that X is contained in a proper parabolic subgroup P = QL of G
with unipotent radical Q and Levi factor L. The following result, when combined with
Theorem 2, completes the proof of Theorem 1. (Recall that Craven [9] has constructed
a subgroup X satisfying the conditions in parts (ii) and (iii) of Theorem 1, and he has
established its uniqueness up to conjugacy; see Remark 1(b).)
Theorem 8.1. Let G be a simple exceptional algebraic group of adjoint type over an
algebraically closed field of characteristic p > 0. Let X = PSL2 (p) be a subgroup of G
containing a regular unipotent element of G and let V = Lie(G) be the adjoint module. If
X is contained in a proper parabolic subgroup P = QL of G, then either
(i) G = E6 , p = 13, L0 = D5 and V |X = W (10)2 ⊕ W (4); or
(ii) G = E7 , p = 19, L0 = E6 and V |X = W (16) ⊕ W (14) ⊕ W (8) ⊕ U .
Proof. We may assume that P is minimal with respect to containing X. Let π : P → P/Q
be the quotient map and identify L with P/Q. By arguing as in the first paragraph in the
proof of Theorem 2 (see the end of Section 2), we deduce that π(X) is contained in an
A1 -type subgroup H of L0 . In addition, Theorem 2 implies that X is not contained in an
A1 -type subgroup of G, so (G, p, V |X ) must be one of the cases in Table 4. As noted in
the proof of Theorem 2, the composition factors of V |H can be read off from [15, Tables
1–5] and this imposes restrictions on the composition factors of V |X . By considering each
possibility for (G, L0 ) in turn, comparing composition factors with Table 4, we will show
that the cases labelled (i) and (ii) in the statement of the theorem are the only compatible
options.
First assume (G, p) = (F4 , 13), so the composition factors of V |X are LX (10)3 , LX (8),
LX (2)3 and LX (0). By inspecting [15, Table 2] it is easy to see that there is no compatible
Levi subgroup L. Similarly, if (G, p) = (E8 , 37) then the composition factors of V |X are
given in (20) and thus we can eliminate this case by repeating the argument in the proof
of Theorem 2.
Next suppose (G, p) = (E6 , 13). The three possibilities for V |X are as follows:
W (10) ⊕ W (8) ⊕ W (2) : LX (10)3 , LX (8)3 , LX (4), LX (2)4 , LX (0)
W (10) ⊕ W (4) ⊕ W (2) : LX (10)3 , LX (8)2 , LX (6), LX (4)2 , LX (2)3 , LX (0)
W (10) ⊕ W (10) ⊕ W (4) : LX (10)4 , LX (8), LX (6), LX (4)2 , LX (2)2 , LX (0)2
In all three cases, we see that V |X has at most two trivial composition factors, so [15,
Table 3] implies that
L0 = A2 A21 , A22 A1 , A4 A1 or D5 .
If L0 = A2 A21 then V |X has five or more LX (2) factors, which is incompatible with all
three possibilities for V |X . Similarly, if L0 = A4 A1 then there are too many LX (4) factors,
and we can rule out L0 = A22 A1 because we would get LX (1) composition factors, which
is absurd. Finally, suppose L0 = D5 . Since the Weyl module WX (14) has an LX (10)
composition factor, we see that V |X has four such factors and thus
V |X = W (10) ⊕ W (10) ⊕ W (4)
is the only option.
A1 -TYPE SUBGROUPS
41
Finally, let us assume (G, p) = (E7 , 19). The three possibilities for V |X are as follows:
W (8) ⊕ W (4) ⊕ W (2) ⊕ U : LX (16)2 , LX (14)2 , LX (12), LX (10), LX (8)3 , LX (4)2 ,
LX (2)2 , LX (0)2
W (16) ⊕ W (10) ⊕ W (4) ⊕ U : LX (16)3 , LX (14), LX (12), LX (10)2 , LX (8), LX (6),
LX (4)2 , LX (2), LX (0)3
W (16) ⊕ W (14) ⊕ W (8) ⊕ U : LX (16)3 , LX (14)2 , LX (10), LX (8)3 , LX (4), LX (2)2 ,
LX (0)3
By inspecting [15, Table 4], counting the number of trivial composition factors, we quickly
reduce to a small number of possibilities for L0 . By considering non-trivial composition
factors, it is straightforward to reduce further to the case L0 = E6 . For example, we can
rule out L0 = A6 because there would be too many LX (4) factors. Similarly, L0 = D5 A1
is out because we would have an LX (6) and at least three LX (8) factors, which is not
compatible with any of the three possibilities above. We can rule out L0 = D6 because it
would imply that V |X has an LX (5) factor. Finally, suppose L0 = E6 . Here V |X has at
least three LX (16) and LX (8) composition factors, so
V |X = W (16) ⊕ W (14) ⊕ W (8) ⊕ U
is the only possibility.
References
[1] J.L. Alperin, Local Representation Theory, Cambridge Stud. Adv. Math., vol. 11, Cambridge Univ.
Press, 1986.
[2] A. Borel and J. Tits, Éléments unipotents et sous-groupes paraboliques de groupes réductifs. I, Invent.
Math. 12 (1971), 95–104.
[3] W. Bosma, J. Cannon and C. Playoust, The Magma algebra system I: The user language, J. Symbolic
Comput. 24 (1997), 235–265.
[4] N. Bourbaki, Groupes et algebrès de Lie (Chapitres 4, 5 et 6), Hermann, Paris, 1968.
[5] R.W. Carter, Simple groups of Lie type, John Wiley and Sons, London, 1972.
[6] R.W. Carter, Finite Groups of Lie Type: Conjugacy Classes and Complex Characters, John Wiley
and Sons, London, 1985.
[7] A.M. Cohen and R.L. Griess, On finite subgroups of the complex Lie group of type E8 , Proc. Sympos.
Pure Math. 47 (1987), 367–405.
[8] B.N. Cooperstein, Maximal subgroups of G2 (2n ), J. Algebra 70 (1981), 23–36.
[9] D.A. Craven, Maximal PSL2 subgroups of exceptional groups of Lie type, preprint (arXiv:1610.07469)
[10] D. Gorenstein, R. Lyons and R. Solomon, The Classification of the Finite Simple Groups, Number
3, Mathematical Surveys and Monographs, vol. 40, Amer. Math. Soc., 1998.
[11] R. Guralnick and G. Malle, Rational rigidity for E8 (p), Compos. Math. 150 (2014), 1679–1702.
[12] G.J. Janusz, Indecomposable representations of groups with a cyclic Sylow subgroup, Trans. Amer.
Math. Soc. 125 (1966), 288–295.
[13] P.B. Kleidman, The maximal subgroups of the Chevalley groups G2 (q) with q odd, the Ree groups
2
G2 (q), and their automorphism groups, J. Algebra 117 (1988), 30–71.
[14] R. Lawther, Jordan block sizes of unipotent elements in exceptional algebraic groups, Comm. Algebra
23 (1995), 4125–4156.
[15] R. Lawther and D.M. Testerman, A1 subgroups of exceptional algebraic groups, Mem. Amer. Math.
Soc. 141 (1999).
[16] M.W. Liebeck, J. Saxl and D.M. Testerman, Simple subgroups of large rank in groups of Lie type,
Proc. London Math. Soc. 72 (1996), 425–457.
[17] M.W. Liebeck and G.M. Seitz, Subgroups generated by root elements in groups of Lie type, Annals
of Math. 139 (1994), 293–361.
[18] M.W. Liebeck and G.M. Seitz, On the subgroup structure of exceptional groups of Lie type, Trans.
Amer. Math. Soc. 350 (1998), 3409–3482.
[19] M.W. Liebeck and G.M. Seitz, The maximal subgroups of positive dimension in exceptional algebraic
groups, Mem. Amer. Math. Soc. 802 (2004).
[20] M.W. Liebeck and D.M. Testerman, Irreducible subgroups of algebraic groups, Q. J. Math. 55 (2004),
47–55.
42
TIMOTHY C. BURNESS AND DONNA M. TESTERMAN
[21] A.J. Litterick, Finite simple subgroups of exceptional algebraic groups, Ph.D. thesis, Imperial College
London, 2013.
[22] A.J. Litterick, On non-generic finite subgroups of exceptional algebraic groups, Mem. Amer. Math.
Soc., to appear.
[23] G. Malle, The maximal subgroups of 2 F4 (q 2 ), J. Algebra 139 (1991), 52–69.
[24] G. Malle and D. Testerman, Linear Algebraic Groups and Finite Groups of Lie Type, Cambridge
Advanced Studies in Mathematics vol. 133, Cambridge University Press, 2011.
[25] R. Proud, J. Saxl and D.M. Testerman, Subgroups of type A1 containing a fixed unipotent element
in an algebraic group, J. Algebra 231 (2000), 53–66.
[26] J. Saxl and G.M. Seitz, Subgroups of algebraic groups containing regular unipotent elements, J.
London Math. Soc. 55 (1997), 370–386.
[27] G.M. Seitz, Unipotent elements, tilting modules, and saturation, Invent. Math. 141 (2000), 467–502.
[28] G.M. Seitz and D.M. Testerman, Extending morphisms from finite to algebraic groups, J. Algebra
131 (1990), 559–574.
[29] G.M. Seitz and D.M. Testerman, Subgroups of type A1 containing semiregular unipotent elements, J.
Algebra 196 (1997), 595–619.
[30] J.-P. Serre, Exemples de plongements des groupes PSL2 (Fp ) dans des groupes de Lie simples, Invent.
Math. 124 (1996), 525–562.
[31] R. Steinberg, Representations of algebraic groups, Nagoya Math. J. 22 (1963), 33–56.
[32] R. Steinberg, Endomorphisms of linear algebraic groups, Mem. Amer. Math. Soc. 80 (1968).
[33] D.M. Testerman, The construction of the maximal A1 ’s in the exceptional algebraic groups, Proc.
Amer. Math. Soc. 116 (1992), 635–644.
[34] D.M. Testerman, A1 -type overgroups of elements of order p in semisimple algebraic groups and the
associated finite groups, J. Algebra 177 (1995), 34–76.
[35] D.M. Testerman and A. Zalesski, Irreducibility in algebraic groups and regular unipotent elements,
Proc. Amer. Math. Soc. 141 (2013), 13–28.
[36] B. Weisfeiler, On one class of unipotent subgroups of semisimple algebraic groups, preprint
(arxiv:0005149)
T.C. Burness, School of Mathematics, University of Bristol, Bristol BS8 1TW, UK
E-mail address: [email protected]
D.M. Testerman, Institute of Mathematics, Station 8, École Polytechnique Fédérale de
Lausanne, CH-1015 Lausanne, Switzerland
E-mail address: [email protected]
| 4math.GR
|
BOIJ-SÖDERBERG DECOMPOSITIONS OF LEX-SEGMENT IDEALS
SEMA GÜNTÜRKÜN
arXiv:1404.0118v2 [math.AC] 20 Aug 2015
A BSTRACT. Boij-Söderberg theory describes the scalar multiples of Betti diagrams of graded modules
over a polynomial ring as a linear combination of pure diagrams with positive coefficients. There are a few
results that describe Boij-Söderberg decompositions explicitly. In this paper, we focus on the Betti diagrams
of lex-segment ideals. Mainly, we characterize the Boij-Söderberg decomposition of a lex-segment ideal
and describe it by using Boij-Söderberg decompositions of some other related lex-segment ideals.
1. I NTRODUCTION
Boij-Söderberg is very recent theory which addresses the characterization of Betti diagrams of graded
modules in polynomial rings. It is originated in a pair of conjectures by Boij and Söderberg in [3] and
in [4] whose proof is given by Eisenbud and Schreyer in [6]. Their result gives a characterization of
Betti diagrams of graded modules up to scalar multiples. For more information about Boij-Söderberg
theory, we refer to a very informative survey written by Fløystad in [8]. This theory brings up an idea of
decomposition of the Betti diagrams of graded modules whose resolutions are not pure resolution. If the
resolution is pure, then the decomposition consists of only one pure diagram with a positive coefficient
as expected. There is not much known about the behavior of the Boij-Söderberg decomposition of an
ideal in a polynomial ring. Any characterization of Boij-Söderberg decompositions, either about BoijSöderberg coefficients or about the chain of the degree sequences associated with the pure diagrams,
would also assist us in understanding and interpreting the structural consequences of Boij-Söderberg
decomposition of the Betti diagrams. Although the theory is quite recent and has a lot of open problems,
improvements and contributions to this theory are quite impressive. Cook in [5] and Berkesch, Erman,
Kumini, and Sam in [1] discuss Boij-Söderberg theory in the perspective of poset structures. In [14],
Nagel and Sturgeon examine the Boij-Söderberg decomposition of some ideals that are raised from
some combinatorial objects. They show the combinatorial importance of the coefficients of the pure
diagrams in the Boij-Söderberg decompositions of their interest of ideals. In [10], results of Gibbons,
Jeffries, Mayes, Rauciu, Stone and White provide a relation between decomposition of the Betti diagrams
of complete intersections and degrees of their minimal generators. Another recent work is done by
Francisco, Mermin, and Schweig in [9]. In their paper, they study the behavior of the Boij-Söderberg
coefficients of Borel ideals.
For the sake of simplicity, the abbreviation BS is used for the Boij-Söderberg . In this paper, we study
the behavior of BS decompositions of lex-segment ideals and obtain a neat relation between the BS
decompositions of a given lex ideal and some other related lex ideals. Throughout the paper, our main
focus will be the BS chain of the degree sequences in the decomposition and we also provide a strong
correlation of the coefficients of the pure diagrams as well. The reason of why we are interested in the
BS decomposition of lex-segment ideals is based on the fact that lex ideals have very particular Betti
diagrams. The Bigatti-Hulett-Pardue, in [2, 13, 15], prove that the lex-segment ideals have the largest
Betti numbers among the ideals with the same Hilbert function. This pivotal property of lex-segment
ideals makes their BS decompositions worthy of study. Moreover, Eliahou-Kervaire formula gives a
nice formulation for the Betti diagram of lex ideals. The main goal is to obtain a pattern for the BS
decomposition of a lex ideal by using decompositions of some other related lex-segment ideals.
In what follows, let R = k[x, y, z] be a polynomial ring of 3 variables, with the lexicographic order,
x >lex y >lex z and L be a lex-segment ideal in R. The ideal L can be decomposed as L = xa + J
where a is also a lex-segment ideal in R and J is a lex-segment ideal in k[y, z]. The first main result
2010 Mathematics Subject Classification. 13D02, 13C99, 13A99.
Key words and phrases. Boij-Söderberg decompositions, Betti diagrams, Lex-segment ideals.
1
2
SEMA GÜNTÜRKÜN
of this paper describes the “beginning” of the BS decomposition of L in terms of the decomposition
of a. The algorithm of BS decomposition itself provides a chain of degree sequences. The first degree
sequence in the chain is the top degree sequence of the Betti diagram of L. By the algorithm, the second
degree sequences is the top degree sequence of the remaining diagram after the subtraction of the first
pure diagram with a suitable coefficient from the Betti diagram. It continues until the Betti diagram is
decomposed completely. Thus, by saying that the “beginning” of the BS decomposition, we mean the
several degree sequences, that is pure diagrams, are obtained in the beginning of the BS decomposition
of L. We now state our first result.
Theorem 1.1. Let R = k[x, y, z] and L be a lex-segment ideal of codimension 3 in R. Suppose 1 6= a =
L : (x). Write the Boij-Söderberg decomposition of a as
β(a) =
t
X
αi πdi + Ra ,
t
X
α̃i πd̄i + RL
i=0
where
<
< ... < < ... <
are all top degree sequences of length 3, that is, di = (di0 , di1 , di2 )
for i = 0, 1, ..., t, and Ra is the linear combination of the pure diagrams greater that πdt . Then the
Boij-Söderberg decomposition of L has the form
d0
d1
dl
dt
β(L) =
i=0
d̄i
di
(di0
+ 1, di1
+ 1, di2
where = + 1 =
+ 1), and α̃i = αi for i = 0, 1, ..., t and α̃t ≥ αt , and RL is a
linear combination of pure diagrams greater than πd̄t .
The second main result in this article is devoted to the pure diagrams - that is to say, degree sequences
- of the BS decomposition of the Betti diagrams of L and (J, x) = (L, x) in the polynomial ring R =
k[x, y, z]. Like in Theorem 1.1, we notice some similarities of the BS decompositions of lex ideal L and
(L, x). We reveal that the entire part of the BS decomposition of (L, x) containing all pure diagrams of
length less than 3 shows up precisely at the end of the BS decomposition of L. That is, all pure diagrams
of length less than 3 are exactly the same with the coefficients. In particular, we prove that
Theorem 1.2. Let L ⊂ R = k[x, y, z] be an Artinian lex-segment ideal of codimension 3. Suppose that L
cannot be decomposed as L = x(x, y, z t ) + J where J is different from (y, z)Gmin (J) and 1 < t < k − 1.
Let a = L : (x) be a lex-segment ideal of R. Then L = xa + J where J ∈ k[y, z] is a stable ideal of
codim 2. The ideal (J, x) = (L, x) is also a codim 3 Artinian, lex-segment ideal in R.
β(L, x) = R(L,x) +
n
X
αi πdi
i=t+1
where dt+1 < dt+2 < ... < dn are all top degree sequences of length less than 3, with the coefficients
αi , i = t + 1, ..., n. R(L,x) is the linear combination of the pure diagrams associated with the degree
sequences of length 3.
Then the BS decomposition of L is
n
X
αi πdi
β(L) = RL +
i=t+1
where the chain dt+1
dt+2
dn
<
< ... <
of degree sequences of length 2 and 1 exactly with the same coefficients αi and RL is the linear combination of the pure diagrams associated with the degree sequences
of length 3.
As a plan of this paper, we first discuss some useful relations of the Betti numbers of the ideals L,
a and J in section 2. We also describe the entire Betti diagram of the lex ideal L in terms of the Betti
numbers of the colon ideal a = L : (x) and the stable ideal J in the same section. In section 3, we give
the proof of Theorem 1.1 which gives the relation between the beginning of the BS decompositions of L
and a. Proof of Theorem 1.2 is given in section 4.
BOIJ-SÖDERBERG DECOMPOSITIONS OF LEX-SEGMENT IDEALS
3
Combining the results of Theorems 1.1 and 1.2, in the case of R = k[x, y, z], we give the following
diagram to summarize nicely the relation between the degree sequences in the BS decompositions of L
and the ideals a and (L, x).
all length < 3
all length 3
The chain of
more length 3
degree sequences
< degree sequences < degree sequences .
degree sequences =
coming from
coming from
of L
(section 5)
(L, x)
a(−1)
We will see that, most of the time, the decompositions of a and (L, x) may not enough to cover all pure
diagrams in the decomposition of L since there might be some pure diagrams of length 3 which may not
obtained by the ideal a.
One naturally hopes to obtain a description of the entire BS decomposition of lex-segment ideal L
in terms of some other related ideals. Section 5 includes further observations for a possible way to
describe the entire BS chain of degree sequences in the BS decomposition of L. The lexicographic order
x >lex y >lex z makes us think about the colon ideals b = L : (y) and c = L : (z). As in the case
a = L : (x) in section 3, one may expect similar results for the lex ideals b and c. Indeed, the examples
show that there is a relation between the BS decompositions of the lex ideal L and the colon ideals b and
c. This allows us to give an almost full description of the pure diagrams appearing in the decomposition
of L.
2. BACKGROUND
AND
P RELIMINARIES
Throughout this section we assume that R is a graded polynomial ring with n < ∞ variables over a
field k with each variable has degree one. In the case of n = 3, we will see the description of the Betti
diagram L = xa + J in terms of the Betti numbers of a and J.
Let M be a graded R-module. The minimal graded free resolution of M is written as
F : 0 → Fn → . . . → Fi → . . . → F1 → F0 → M → 0
where
Fi =
M
R(−α)βi,α .
α≥0
The numbers βi,α are the Betti numbers of M and are considered in the Betti diagram β(M ) of M whose
entry in row i and column j is βi,i+j . Let d = (d0 , d1 , ..., dn−1 ) ∈ Zn≥0 be a sequence of non-negative
integers of length n+1 with d0 < ... < dn−1 . The graded free resolution of M is called a pure resolution
of type d = (d0 , ..., di , ..., dn−1 ) if, for all i = 0, 1, ..., n − 1, the i-th syzygy module of M is generated
only by elements of degree di , in other words, all Betti numbers are zero except βi,di (M ). Then the Betti
diagram of this module is called a pure diagram of type d. The formula for the pure diagram associated
by d is based on the Herzog-Kühl equations introduced in [12],
n−1
1
λ Q
if α = di
where λ ∈ Q>0 .
βi,α =
i=0,i6=k |di − dk |
0
otherwise
We define a partial order on the degree sequences so that ds < dt if dsi ≤ dti for all i = 0, 1, ..., n − 1.
The order on the degree sequences induces an order of the pure diagrams πds < πdt if ds < dt . The BS
decomposition of a Betti diagram of an R- module is a linear combination of pure diagrams with positive
coefficients.
Algorithm 2.1 (Boij-Söderberg Decomposition Algorithm). The algorithm to decompose for a given
(non-pure) Betti diagram has the following steps
(1) Determine the top degree sequence d = (d0 , ..., di , ..., dn−1 ) of the Betti diagram of the Rmodule M , say β(M ).
4
SEMA GÜNTÜRKÜN
(2) Determine the coefficient α of the pure diagram πd by
βi,di (M )
min
, for i = 0, 1, .., n − 1 .
βi,di (πd )
(3) Subtract απd from the Betti diagram β(M ) so that the new entries will be all positive.
(4) Repeat the first and second steps for the remaining diagram β(M )−απd until the Betti diagram
is completely decomposed into pure diagrams.
Thus the BS decomposition of a graded R-module M gives an ordered decomposition of the Betti
diagram,
X
β(M ) =
as πds where πds < πdt if s < t.
s
Example 2.2. For instance, let I = (x2 , xy, xz, y 2 ) be an ideal in k[x, y, z], the BS decomposition of
R/I is given as
β(R/I) = (8)πd0 + (4)πd1
where
0 1 2
0 1 2 3
1
πd0 = 0 24 - - - < πd1 = 0 61 - - as d0 = (0, 2, 3, 4) < d1 = (0, 2, 3).
1 - 41 13 81
1 - 12 31
Consider a monomial ideal I in R. We will denote the set of minimal monomial generators of I by
G(I). Then G(I)i will denote the subset of G(I) containing the minimal generators of degree i. The
notation Gmin (I) will be used for the initial degree of the monomials in I and Gmax (I) will stand for
the maximum degree of the monomials in G(I) throughout the paper. We next state the definitions of
graded lexicographic monomial order and lex-segment ideal.
Definition 2.3. Let m = xs11 ...xsnn and n = xt11 ...xtnn be two monomials in R = k[x1 , ..., xn ]. If either
deg m > deg n or deg m = deg n and si − ti > 0 for the first index i such that si − tt 6= 0, then it is said
that m >glex n in graded lexicographic order.
Definition 2.4. Let R be a polynomial ring and L be a monomial ideal in R generated by the monomials
m1 , ..., ml . The ideal L is called a lex-segment ideal (lexicographic ideal, or lex ideal) in R if for each
monomial m ∈ R the existence of some mi ∈ G(L) with m >glex mi and deg(m) = deg(mi ) implies
m ∈ L.
For simplicity, we will use > for the lex order >glex unless the order is different than lexicographic
order. In this section, we make some observations about the Betti diagrams of lex-segment ideals. We
aim to get some correlations between Betti numbers of the lex ideals L = xa + J, a = L : x and J in
k[x, y, z]. Next lemma shows that the colon ideal a is also a lex-segment ideal in k[x, y, z].
Lemma 2.5. Let L be a lex-segment ideal in R = k[x1 , .., xn ]. Consider the colon ideals ai = L : (xi ),
for i = 1, ..., n. Then each ai is also a lex-segment ideal in R.
Proof. Let m′ ∈ ai be a monomial, for any i = 1, ..., n. Let m be a monomial in R and deg m = deg m′
and m >glex m′ . Then xi m′ ∈ L as ai = L : (xi ), and xi m >glex xi m′ . This implies xi m ∈ L and
hence m ∈ L : (xi ) = ai .
Let u be a monomial in R = k[x1 , .., xn ], we define m(u) to be the largest index i such that xi divides
u. Recall that a monomial ideal I is said to be stable if, for every monomial u ∈ G(I) and all i < m(u),
xi u/xm(u) is also in G(I).
Next we quote a proposition from [7].
Proposition 2.6. (Eliahou-Kervaire formula) Let I ⊂ R be a stable ideal. Then
P
m(u)−1
(a) βi,i+j (I) =
,
i
u∈G(I)j
(b) proj dim R/I = max{m(u) : u ∈ G(I)},
(c) reg (I) = max{deg (u) : u ∈ G(I)}.
BOIJ-SÖDERBERG DECOMPOSITIONS OF LEX-SEGMENT IDEALS
5
From now on, we always assume n = 3, that is, R = k[x, y, z] unless otherwise is stated. We follow
with a lemma that indicates the relation between the minimal generators of ideals L = xa + J, a = L : x
and J. Then the next lemma provide a crucial short exact sequence of these ideals.
Lemma 2.7. If L is lex-segment ideal in R, then there are unique monomial ideals a ⊂ R and J ⊂ k[y, z]
such that L = xa + J. Moreover, the ideal a is also a lex-segment ideal since a = L : (x) and J is stable
in R, and G(L) = xG(a) ∪ G(J).
Proof. The proof follows immediately from the fact that L is a lex-segment ideal with graded lex order
x > y > z.
Lemma 2.8. Let 0 → F1 → F0 → J and 0 → G2 → G1 → G0 → a be graded free resolutions for the
ideals J and a. If L = xa + J, then there is a short exact sequence 0 → J(−1) → a(−1) ⊕ J → L → 0.
Moreover,
0 → G2 (−1) ⊕ F1 (−1) → G1 (−1) ⊕ F1 ⊕ F0 (−1) → G0 (−1) ⊕ F0 → L
is the graded minimal free resolution of L.
Proof. The form of the lex-segment ideal L implies the short exact sequence 0 → J(−1) → a(−1) ⊕
J → L → 0. The mapping cone for the short exact sequence provides a free resolution for L. Let
m ∈ G(a) ∩ G(J). Then m ∈ G(J) implies either m = ym′ or m = zm′ for some monomial
m′ ∈ k[y, z]. As L is a lex-segment ideal and m ∈ L, xm′ ∈ G(L). So m′ ∈ G(a). Therefore m is
divisible by m′ and m cannot be a minimal generator of a. Therefore the ideals J and a do not have
common minimal generators. This tells us that there is no cancellation in the mapping cone structure. So
the resulting graded free resolution for L is minimal.
We now analyze the Betti numbers of the ideals L, a = L : (x) and J. We know that the lex-segment
ideals L and a are stable and in addition to this, J is a lex ideal in k[y, z]. Thus, Eliahou-Kervaire formula
gives rise to the following decomposition,
βi,i+j (L) =
P
u∈G(L)j
m(u)−1
i
X m(u) − 1
+
i
=
u∈G(L)j
x|u
|
{z
= βi,i+j−1 (a)
}
X m(u) − 1
i
u∈G(L)j
x∤u
|
{z
say Di,i+j
We name the initial degree of J, Gmin (J) := k and the Betti numbers of β(a) and β(J) as
}
ai,i+j := βi,i+j (a), and ci,i+j := βi,i+j (J).
The following lemmas provide some relations and identities about the Betti numbers of L, a. They help
us to describe the entire Betti diagram of L with respect to the Betti numbers of a and J.
Lemma 2.9. As L = xa + J in R = k[x, y, z], if Gmin (L) ≥ 2, then Gmin (L) = Gmin (a) + 1 by
stability of the ideals L and a = L : (x) 6= 1.
Lemma 2.10. The Eliahou-Kervaire formula gives the following identities for the Betti numbers of the
J
• c0,k = c1,k+1 + 1,
• c0,j = c1,j+1 for all j ≥ k + 1,
• if c0,k = k + 1 then c1,k+1 = k and ci,i+j = 0 for all i = 0, 1 and j ≥ k + 1.
We know that βi,i+j (L) = ai,i+j−1 + Di,i+j . Thus, it follows that
(
0,
when j ≤ k − 1,
Di,i+j =
βi,i+j (J, x), when j ≥ k.
That is, D0,j = c0,j , D1,j+1 = c0,j + c1,j+1 , and D2,j+2 = c1,j+1 .
6
SEMA GÜNTÜRKÜN
Lemma 2.11. Gmin (J) ≥ Gmax (a) + 1 where J 6= 0.
Proof. Say Gmax (a) = t. Suppose k = Gmin (J) < t, then y k ∈ G(L)k . So, by lex-order, all monomials
u of degree k divisible by x are in L. Thus, u is in the form xi y j z s where i ≥ 1, i + j + s = k. As
Gmax (a) = t > k, there is a minimal generator v ∈ L of degree t + 1 such that x|v. Therefore, v can be
written as v = w1 · w2 where w1 and w2 are two monomials such that deg w1 = k and w2 is not divisible
by x and deg w2 = t − k + 1. Since all degree k monomials divisible by x are in L, w1 is in L and so v
cannot be a minimal generator. Thus k ≥ t.
Now, we need to show that the equality is not possible, i.e. k = t is not possible. We prove this by
contradiction. To this end, suppose k = t. So y k is a minimal generator in L and since t = k we can
find at least one minimal generator u of a with degree k. Then xu becomes a minimal generator in L of
degree k + 1. As all monomials v of degree k divisible by x are in L then there is a monomial w such that
v = xw and w|u. This contradicts that u is a minimal generator of a. Hence k 6= t. i.e. k ≥ t + 1
Suppose Betti diagrams for a and J are
β(a)
1
2
..
.
0
a0,1
a0,2
..
.
1
a1,2
a1,2
..
.
2
a2,3
a2,4
..
.
and
k − 1 a0,k−1 a1,k a2,k+1
β(J)
k
k+1
..
.
0
c0,k
c0,k+1
..
.
1
c0,k − 1
c0,k+1 ,
..
.
Gmax (J) c0,Gmax (J) c0,Gmax (J)
TABLE 1. The Betti diagrams of a and J.
Therefore the short exact sequence in Lemma 2.8 together with all other Lemmas 2.7, 2.9 and 2.11 we
discuss in this section yield that the Betti diagram for L has the following form:
β(L)
2
3
..
.
0
a0,1
a0,2
..
.
k−1
k
k+1
..
.
1
a1,2
a1,2
..
.
2
a2,3
a2,4
..
.
a0,k−2
a1,k−1
a2,k
a0,k−1 + c0,k a1,k + 2c0,k − 1 a2,k+1 + c0,k − 1
c0,k+1
2c0,k+1
c0,k+1
..
..
..
.
.
.
Gmax (L) = Gmax (J)
c0,Gmax (L)
2c0,Gmax (L)
c0,Gmax (L)
TABLE 2. The Betti diagram of L
If the Betti diagrams of the ideals a and J “overlap” then they do only at the k th row of the β(L) as in
the above diagram. In other words, the Betti numbers of β(L) in the kth row may be expressed in terms
of both the Betti numbers of β(a) and β(J) in their k − 1th and first row, respectively.
3. T HE B OIJ -S ÖDERBERG D ECOMPOSITIONS
OF
L AND L : (x)
In this section we prove Theorem 1.1. Let d0 < d1 < ... < di < ... < dt be the BS chain of
all length 3 top degree sequences for a = L : (x). Suppose the chain of the first t + 1 top degree
sequences of the BS decomposition of the Betti diagram of L is d̄0 < d̄1 < ... < d̄i < ... < d̄t . Then
d̄i = di + 1 = (di0 + 1, di1 + 1, di2 + 1) for all i = 0, 1, ..., t with exactly the same coefficients, except
possibly the coefficient of πd̄t .
BOIJ-SÖDERBERG DECOMPOSITIONS OF LEX-SEGMENT IDEALS
7
Recall that, for a given top degree sequence d = (d0 , d1 , d2 ), the “normalized” pure diagram πd can
be obtained as
if i + j 6= di
0
!
2
2
Q
Q
βi,i+j (πd ) =
1
|di − dr |, i = 0, 1, 2 .
if i + j = di , where λ = lcm
λ
|di −dr |
r=0,r6=i
r=0,r6=i
Thus, this formula provides pure diagrams with integer entries. From now on, pure diagrams have integer
entries. Let d0 = (d00 , d01 , d02 ) be the top degree sequence for the Betti diagram of a. if d02 < k +1, that is,
d00 < d01 < d02 < k + 1, so d00 < k − 1 then we see that βi,i+j (a) = βi,i+j+1 (L) for all j = 0, 1, ..., k − 2
as in the table 2. This essentially follows from the fact that the Betti diagrams of a and J may only
overlap on the k-th row in the Betti diagram of L. As L = xa + J and degree shift due to multiplication
by x the top degree sequence of β(L) will be d0 + 1. Thus β(L) − α0 πd0 +1 becomes the first step of
the BS-decomposition of β(L). In fact, we could repeat this process for all degree sequence ds such that
ds2 < k + 1.
Suppose ds2 < k + 1 for s = 0, 1, ..., l − 1 and ds2 ≤ k + 1 for s ≥ l for some l > 0. So we assume
that the next degree sequence after dl−1 is dl = (dl0 , dl1 , dl2 = k + 1). Therefore, after l steps in the
decomposing both β(a) and β(L), we would get the remaining diagrams
β(a) −
l−1
X
αs πds =: β̃(a)
and
β(L) −
l−1
X
αs πds +1 =: β̃(L).
s=0
s=0
Let dl = (dl0 , dl1 , dl2 ) be the next top degree sequence of the Betti diagram for a and dl2 = k + 1 so
above paragraph shows that dl + 1 becomes the next top degree sequence of Betti diagram for L. Then
the remaining diagram after the first l steps of the BS decompositions for both a and L look like as
following,
β(a) −
l−1
P
s=0
αs πds
β̃(a)
0
1
2
dl0
β̃0,dl0 (a)
..
..
..
..
.
.
.
.
=
dl1 − 1
a0,dl1 −1 β̃1,dl1 (a)
..
..
..
..
.
.
.
.
dl2 − 2 = k − 1 a0,dl2 −2 a1,dl2 −1 a2,dl2
TABLE 3. Remaining diagram after l steps for β(a)
and similarly,
β(L) −
l−1
P
s=0
αs πds +1
β̃(L)
dl0 + 1
..
.
dl1
=
..
.
dl2 − 1 = k
dl2
..
.
0
β̃0,dl0 +1 (L)
..
.
a0,dl1 −1
..
.
a0,dl2 −2 + c0,dl2 −1
c0,dl2
..
.
1
..
.
β̃1,dl1 +1 (L)
..
.
a1,dl2 −1 + c0,dl2 −1 + c1,dl2
c0,dl2 + c1,dl2 +1
..
.
TABLE 4. Remaining diagram after l step for β(L).
2
..
.
..
.
β̃2,dl2 +1 (L)
c1,dl2 +1
..
.
8
SEMA GÜNTÜRKÜN
By construction of β(L), we deduce that
β̃0,dl +1 (L) = β̃0,dl (a) as dl0 + 1 < k,
0
0
β̃1,dl +1 (L) = β̃1,dl (a) as dl1 < k, and
1
1
β̃2,dl +1 (L) = a2,dl + c1,dl as dl2 − 1 = k.
2
2
2
The algorithm 2.1 exposes the coefficient of the pure diagram πdl to be
(3.1)
αl = min
(
β̃0,dl (a)
β̃1,dl (a)
a2,dl
2
1
0
,
,
β0,dl (πdl ) β1,dl (πdl ) β2,dl (πdl )
0
1
2
)
.
and similarly for the BS-decomposition of β(L) there is a rational number α̃l as the coefficient of the
pure diagram πdl +1 such that
(3.2)
α̃l = min
(
β̃0,dl (a)
0
β̃1,dl (a)
1
a2,dl + c1,dl
2
,
,
β0,dl (πdl ) β1,dl (πdl ) β2,dl (πdl )
0
1
2
2
)
Hence we just need to look at the k-th row of the Betti diagram of L if β(a) and β(L) overlap. Thus, we
only need to think about the top degree sequences ds of length 3 of β(a) such that ds2 = k + 1.
Case 1: Let a2,k+1 be eliminated in the (l + 1)-th step of the decomposition algorithm of β(a). In
l+1
other words, dl = (dl0 , dl1 , dl2 ) is of length 3, whereas dl+1 = (dl+1
0 , d1 ) has length 2. It shows that
0
1
i
l
d < d < ... < d < ... < d are all length 3 degree sequences in the decomposition of β(a). Hence,
BS decomposition of β(a) is
β(a) =
l
X
αs πds + [all pure diagrams of length less than 3].
s=0
Recall that we only focus on the degree sequences of length three. Since the length of dl+1 is two, we
do not need to pay attention to the (l + 2)-th step in the decomposition. Besides that the table 4 already
shows that dl + 1 is top degree sequence of the remaining diagram of L, β̃(L). Therefore, the first
(l + 1)-th top degree sequences of Boij-Söderberg decomposition of β(L) is
d0 + 1 < d1 + 1 < ... < dl + 1
where the coefficients α̃i = αi for i = 0, 1, ..., l − 1.
Case 2: Suppose that a2,k+1 is not eliminated in the (l + 1)-th step of the decomposition of β(a).
Moreover we assume that it will vanish in the (t + 1)-th step for some t > l. That is, the chain of the
degree sequences in the BS decomposition of β(a) is
d0 < d1 < ... < dl < ... < dt < ... < dn
where
• for s = 0, 1, ..., l − 1, ds = (ds0 , ds1 , ds2 ) has length 3 such that ds2 < k + 1,
• for s = l, ..., t, ds = (ds0 , ds1 , ds2 ) has length 3 such that ds2 = k + 1,
• for s = t + 1, ..., n, ds = (ds0 , ds1 ) has length 2.
As the entries only above the (k − 1)-th row are eliminated until the (t + 1)-th step of the decomposition,
it is easy to observe the remaining diagram of L. In table 2, we have seen that the entries of both β(a)
and β(L) above the k-th row are the same. Therefore, the remaining diagram of β(a) after subtracting
the first t pure diagrams is
BOIJ-SÖDERBERG DECOMPOSITIONS OF LEX-SEGMENT IDEALS
β(a) −
t−1
P
αs πds
s=0
where β̃i,dti (a) = βi,dti (a) −
t−1
P
s=0
β̃(a)
0
t
d0
β̃0,dt0 (a)
..
..
.
.
=
dt1 − 1
a0,dt1 −1
..
..
.
.
dt2 − 2 = k − 1 a0,dt2 −2
1
..
.
β̃1,dt1 (a)
..
.
a1,dt2 −1
9
2
..
.
..
.
β̃2,dt2 (a)
αs βi,dti (πds ), for i = 0, 1, 2. Furthermore, as in (3.1) and (3.2), we have
similar relations between the coefficients in both BS decomposition of β(a) and β(L) during their first t
steps. The coefficients of the pure diagrams πds in the decomposition of β(a) for s = l, ..., t − 1 is
)
(
β̃i,dsi (a)
, for i = 0, 1, 2 .
αs = min
βi,dsi (πds )
Similarly, the corresponding coefficient α˜s of the pure diagram πds +1 in the decomposition of β(L)
becomes
(
)
β̃i,dsi +1 (L)
α˜s = min
, for i = 0, 1, 2
βi,dsi +1 (πds +1 )
)
(
β̃1,ds1 (a) β̃2,ds2 (a) + c1,ds2
β̃0,ds0 (a)
.
,
,
= min
β0,ds0 (πds ) β1,ds1 (πds )
β2,ds2 (πds )
We assume that any of the entries of the corresponding to dsi for i = 0, 1 is eliminated where s =
l, .., t − 1. Thus
β̃2,ds2 (a)
αs <
, where ds2 = k + 1.
β2,k+1 (πds )
So it follows that
β̃2,ds2 (a) + c1,k+1
β̃2,ds2 (a)
<
.
β2,k+1 (πds )
β2,k+1 (πds )
Hence α̃s = αs for s = l, ..., t − 1. However, this equality may not be true for the coefficients αt and α̃t
since β̃2,dt2 (a) will be eliminated in the next step. So αt ≤ α̃t . Hence the remaining diagram of β(L) is
β̃(L) := β(L) −
t−1
P
s=0
αs πds +1 =
dt0 + 1
..
.
dt1
..
.
dt2 − 1 = k
k+1
..
.
0
β̃0,dt0 +1 (L)
..
.
a0,dt1 −1
..
.
a0,dt2 −2 + c0,dt2 −1
c0,k+1
..
.
1
..
.
β̃1,dt1 +1 (L)
..
.
a1,dt2 −1 + c0,dt2 −1 + c1,dt2
c0,k+1 + c1,k+2
..
.
2
..
.
..
.
β̃2,dt2 (L)
c1,k+2
..
.
where
β̃0,dt0 +1 (L) = β̃0,dt0 (a) as dt0 + 1 < k,
β̃1,dt1 +1 (L) = β̃1,dt1 (a) as dt1 < k, and
β2,dt2 +1 (L) = β̃2,dt2 (a) + c1,dt2 as dt2 − 1 = k.
This will bring us back to Case 1; dt = (dt0 , dt1 , dt2 ) is the last top degree sequence of length 3 in the
BS decomposition of β(a). The remaining diagram above clearly shows us that dt + 1 = (dt0 + 1, dt1 +
1, dt2 + 1) shows up as a degree sequence in the BS decomposition of β(L) in the next step.
As a summary, if d0 < d1 < ... < dt is the chain of the all top degree sequences of length 3 in
the BS-decomposition of β(a) with coefficients αs for s = 0, 1, .., t. Then d0 + 1 < d1 + 1 < ... <
10
SEMA GÜNTÜRKÜN
dt + 1 becomes the first t top degree sequences of length 3 in the BS-decomposition of β(L) with
α̃s = αs if s < t and α̃t ≥ αt .
Hence we have shown that the ”beginnings“ of the chain of the degree sequences in the BS decompositions of β(L) = xa + J and β(a) are identical. We believe that there is a analogous result for a
lex-ideal in k[x1 , ..., xn ].
Remark 3.3. Let L = (x1 )a + J in R = k[x1 , x2 , ..., xn ] be a lex-segment ideal, then a is also lexsegment ideal in R and J turns out to be a stable ideal of codim n − 1 in k[x2 , ..., xn ].
Suppose
Fn−1 −
→··· −
→ Fi −
→··· −
→ F1 −
→J −
→0
Gn −
→··· −
→ Gi −
→··· −
→ G1 −
→a −
→0
are the minimal free resolutions of J and a, respectively. We get the same short exact sequence in
Lemma 2.8, then by mapping cone we have the following minimal free resolution for L
0 → Gn (−1) ⊕ Fn−1 (−1) → ... → G2 (−1) ⊕ F2 ⊕ F1 (−1) → G1 (−1) ⊕ F1 → L.
So it yields
βi,i+j (L) =
βi,i+j−1(a)
βi,i+j−1(a) +
where i = 0, 1, ..., n − 1 and i + j < Gmin (J),
i
P
βt,j+t (J)
where i = 0, 1, ..., n − 1 and i + j ≥ Gmin (J).
t=i−1
By using lex-order properties of L and a, as we did in case n = 3, we conclude that the Betti diagrams
of a and J either overlap only on the Gmin (J)-th row of the Betti diagram of L or do not overlap at all.
Identify k := Gmin (J). Therefore, the Betti diagram of L in k[x1 , ..., xn ] is
β(L)
2
3
..
.
0
a0,1
a0,2
..
.
1
a1,2
a1,3
..
.
2
a2,3
a2,4
..
.
k−1
k
k+1
..
.
a0,k−2
a0,k−1 + c0,k
c0,k+1
..
.
a1,k−1
a1,k + c0,k + c1,k+1
c0,k+1 + c1,k+2
..
.
a2,k
a2,k+1 + c1,k+1 + c2,k+2
c1,k+2 + c2,k+3
..
.
Gmax (L) = Gmax (J)
c0,Gmax (L)
...
...
...
n-1
an−1,n
an−1,n+1
..
.
...
...
an−1,k+n−3
... an−1,k+n−2 + cn−1,k+n−1
...
cn−1,k+n−1
..
...
.
c0,Gmax (L) + c1,Gmax (L)+1 c1,Gmax (L)+1 + c2,Gmax (L)+2 ...
cn−1,Gmax (L)+n−1
TABLE 5. Betti diagram of L in k[x1 , ..., xn ]
Henceforth, proof of Theorem 1.1 can be easily modified for the polynomial ring of n variables.
Corollary 3.4. Let L = (x1 )a + J in R = k[x1 , x2 , ..., xn ] be a lex-segment ideal. If πd0 < πd1 <
... < πdt are all pure diagrams of length n in the BS decomposition of a, where di = (di0 , di1 , ..., din−1 )
for i = 0, 1, ..., t. Then the chain of pure diagrams
πd̄0 < πd̄1 < ... < πd̄t
appears in the beginning of the BS decomposition of L such that
d̄i = di + 1 = (di0 + 1, di1 + 1, .., din−1 + 1).
BOIJ-SÖDERBERG DECOMPOSITIONS OF LEX-SEGMENT IDEALS
4. T HE B OIJ -S ÖDERBERG D ECOMPOSITION
FOR
11
L AND (L, x)
In Theorem 1.1, we have showed that if πdi , a type-di = (di0 , di1 , di2 ) pure diagram, appears as a
summand in the BS decomposition of β(a) then, π(d+1)i for (d + 1)i = (di0 + 1, di1 + 1, di2 + 1) show up
as the i-th summand of the BS decomposition of β(L), with the same BS coefficient possible except for
the last one. In this section, we now consider the end of the BS decomposition of L in R = k[x, y, z] and
show that all degree sequences of length less than 3 in the decomposition of β(L, x) = β(J, x) occurs
precisely as all degree sequences of length less that 3 in the decomposition for L. We prove our claim
for all Artinian lex-segment ideals L = a(x) + J except the ones of the form L = x(x, y, z t ) + J where
J is different that (y, z)Gmin (J) and 1 < t < k − 1. The main idea of proof is induction whose base step
also requires some tedious case analyzing of the decompositions of both β(L) and β(L, x). Then finally
schemes of case analyzing help us to demonstrate how all degree sequences of length less than 3 for both
L and (L, x) coincide entirely with their coefficients.
Furthermore, we conjecture that the statement of the Theorem 1.2 is also true for L = x(x, y, z t ) + J,
whereas proof of that situation requires a case analyzing which becomes infeasible.
4.1. Decomposing the Betti diagram of (L, x). First we observe that the same pure diagrams πd , for
all d of length less than 3, as in the decomposition of the Betti diagram of L. To show this, it suffices to
check on the remaining diagrams after several steps of the decomposition algorithm for (L, x). We also
notice that, for all i > k, i-th row of the Betti diagram of (L, x) has the form |c0,i , 2c0,i+1 c0,i | where
c0,i < k.
Say Gmax (L, x) = Gmax (L) = Gmax (J) =: n. Assume k = Gmin (J) > 2 and n ≥ k + 1. Then
the Betti diagram of the lex ideal (L, x) is
0
1
−
..
.
1
2
..
.
β(L, x) = k − 1
−
k
c0,k
k + 1 c0,k+1
..
..
.
.
n
c0,n
1
−
−
..
.
2
−
−
..
.
−
2c0,k − 1
2c0,k+1
..
.
−
c0,k − 1
c0,k+1
..
.
2c0,n
c0,n
First degree sequence is d̄0 = (1, k + 1, k + 2), then we have β(L, x) − γ0 πd̄0 where
πd̄0 =
1
.
.
.
k
0
1
.
.
.
−
1
−
.
.
.
k+1
2
−
.
.
.
k
and
2c0,k − 1 c0,k − 1
c0,k − 1
γ0 = min 1,
,
.
=
k+1
k
k
c
−1
Notice that Artinian lex ideal property of L yields c0,k ≤ k + 1. Thus 0,kk ≤ 1.
If c0,k = k + 1, i.e. Gmax (L) = k, then the BS decomposition of (L, x) becomes
β(L, x) = (1)π(1,k+1,k+2) + (k)π(k,k+1) + (1)π(k) .
If c0,k < k + 1, then
c0,k −1
k
< 1. Therefore, the remaining diagram after the first step becomes
12
SEMA GÜNTÜRKÜN
0
..
.
c0,k − 1
k
..
.
k
c0,k
k+1
..
.
n
c0,k+1
..
.
c0,n
1
β(L, x) − γ0 πd̄0 =
1−
1
2
-
-
..
.
(k − 1)c0,k + 1
k
2c0,k+1
..
.
2c0,n
..
.
c0,k+1
..
.
c0,n
Then the next pure diagram is
πd̄1 =
1
.
.
.
k
k+1
0
2
.
.
.
−
−
1
−
.
.
.
k+2
−
2
−
.
.
.
−
k
d̄1 = (1, k + 1, k + 3).
for
The coefficient for πd̄1 is
γ1 = min
1 c0,k − 1 c0,k (k − 1) + 1 c0,k+1
−
,
,
2
2k
k(k + 2)
k
.
Case I: Let
γ1 =
1 c0,k − 1
−
.
2
2k
This is implied by the following inequalities;
k
+ 1 ≤ c0,k and 2c0,k+1 + c0,k > k + 1.
3
Thus the algorithm eliminates the entry β0,1 (L, x)and the remaining diagram is
c0,k − 1
1 c0,k − 1
−
π(1,k+1,k+2) −
π(1,k+1,k+3)
β(L, x) −
k
2
2k
0
k
c0,k
= k+1 c
0,k+1
..
.
n
Next if
d̄2
..
.
c0,n
1
3c0,k − k − 3
2
2
−
1 c0,k − 1
)
c0,k+1 − k( −
2
2k
..
.
2c0,k+1
..
.
2c0,n
c0,n
= (k, k + 1, k + 3) and then the coefficient of the pure diagram π(k,k+1,k+3) becomes
k + 1 − c0,k
c0,k 3c0,k − k − 3
,
, c0,k+1 −
.
γ2 = min
2
6
2
This creates two possible sub-cases and we observe the remaining diagrams for each case;
3(c0,k−1 ) − k
which is a result of k3 < c0,k+1 , then we obtain
Case I.1: If γ2 =
6
0
k
β(L, x) − γ0 πd̄0 − γ1 πd̄1 − γ2 πd̄2 = k + 1
..
.
n
k
3 +1
c0,k+1
1
-
..
.
2c0,k+1
..
.
c0,n
2c0,n
2
c0,k+1 −
..
.
c0,n
k
3
BOIJ-SÖDERBERG DECOMPOSITIONS OF LEX-SEGMENT IDEALS
k+1−c0,k
2
Case I.2: If γ2 = c0,k+1 −
as c0,k+1 <
β(L, x) − γ0 πd̄0 − γ1 πd̄1 − γ2 πd̄2
k
3
13
, then
0
k
k + 1 − 2c0,k
k+1
c0,k+1
= k+2
c0,k+2
..
..
.
.
n
c0,n
1
k − 3c0,k+1
2c0,k+1
2c0,k+2
..
.
2
c0,k+2
..
.
2c0,n
c0,n
Case II: Let
γ1 =
c0,k (k − 1) + 1
,
k(k + 2)
because
k
+ 1 > c0,k and (k − 1)c0,k + 1 < (k + 2)c0,k+1 .
3
Then the algorithm gives
β(L, x) −
c0,k − 1
k
π(1,k+1,k+2) −
c0,k (k − 1) + 1
k(k + 2)
π(1,k+1,k+3)
1
2
-
-
..
.
k
0
k − 3(c0,k − 1)
1−
k+2
..
.
c0,k
..
.
-
k+1
c0,k+1
2c0,k+1
..
.
n
..
.
..
.
2c0,n
..
.
(k − 1)c0,k + 1
c0,k+1 −
k+2
..
.
c0,n
1
=
c0,n
The next top degree sequence is d̄2 = (1, k + 2, k + 3) and its coefficient is
n k − 3(c − 1) 2c
(k − 1)c0,k + 1 o
0,k+1 c0,k+1
0,k
.
,
,
−
γ2 = min
k+2
k+2 k+1
(k + 1)(k + 2)
Therefore, it splits into two sub-cases :
Case II.1: γ2 =
β(L, x) −
k − 3(c0,k − 1
as k + 2 < c0,k+1 + 2c0,k . Then
k+2
2
P
i=0
γi πd̄i
k
k
+
1
=
..
.
n
0
c0,k
c0,k+1
..
.
c0,n
1
2
2c0,k+1 − k + 3(c0,k − 1) c0,k+1 + 2c0,k − (k + 2)
..
..
.
.
2c0,n
c0,n
d̄3
The next degree sequence is
= (k, k + 2, k + 3) and its coefficient is
n
2
k
1
k + 2o
γ3 = min c0,k , c0,k+1 − + c0,k − 1, co,k+1 + c0,k −
.
3
3
2
2
If c0,k < 12 co,k+1 + c0,k − k+2
2 , then it implies k + 2 < c0,k+1 which is not possible. Thus the coefficient
must be
c0,k+1 (k − 1)c0,k
γ3 =
−
1(k + 1)(k + 2).
k+1
+
Then the remaining diagram looks like
14
SEMA GÜNTÜRKÜN
0
β(L, x) −
3
P
k+2−c0,k+1
2
k
γi πd̄i = k + 1
c0,k+1
..
.
..
.
n
i=0
c0,n
1
-
2
..
.
c0,k+1 +k
2
..
.
2c0,n
c0,n
c0,k+1 (k − 1)c0,k + 1
−
, then the algorithm
k+1
(k + 1)(k + 2)
≥ 1. So this case does not exist if the maximum degree is k + 1. Suppose
Case II.2: Note that if Gmax (L, x) = k + 1 and γ2 =
contradicts the assumption c0,k
Gmax (L, x) > k + 1, so
γ2 =
c0,k+1 (k − 1)c0,k + 1
−
if 2c0,k + c0,k+1 < k + 2.
k+1
(k + 1)(k + 2)
Then the remaining diagram is
0
β(L, x) −
2
P
γi πd̄i
i=0
1
1−
..
.
k
=
k+1
k+2
..
.
n
c0,k+1
c0,k+2
..
.
kc0,k+1 +(k−1)c0,k +1
k+1
2
..
.
-
2c0,k+2
..
.
c0,k+2
..
.
c0,n
2c0,n
c0,n
(2c0,k +c0,k+1 )−1
k+1
..
.
c0,k
1
..
.
-
The top degree sequence of the remaining diagrams is (1, k + 2, k + 4) and its coefficient is
γ3 = min
n1
2
−
2c0,k + c0,k+1 − 1 kc0,k+1 + (k − 1)c0,k + 1 c0,k+2 o
.
,
,
2(k + 1)
(k + 1)(k + 3)
k+1
Next we observe each possible sub-cases for γ3 .
1 2c0,k + c0,k+1 − 1
Case II.2.a : If γ3 = −
as a result of
2
2(k + 1)
k + 4 < 3c0,k+1 + 4c0,k and k + 2 < 2c0,k + c0,k+1 + 2c0,k+2 .
Then we get the remaining diagram
0
β(L, x) −
3
P
i=0
γi πd̄i
k
k+1
= k+2
..
.
n
c0,k
c0,k+1
c0,k+2
..
.
c0,n
1
-
2
-
3c0,k+1 +4c0,k −(k+4)
2
2c0,k+2
..
.
c0,k+2 −
2c0,n
c0,k+1+2c0,k −1
2
−
..
.
k+1
2
c0,n
In this case, we would like to continue to decompose one more step. The coefficient of the pure
diagram π(k,k+2,k+4) comes from
n
3c0,k+1 + 4c0,k − (k + 4)
c0,k+1 + 2c0,k − 2 − k o
γ4 = min c0,k ,
.
, c0,k+2 +
4
2
(a1): If γ4 = co,k , that is,
k+4
< c0,k+1 and k + 2 < c0,k+1 + 2c0,k+2
3
So the remaining diagram becomes
BOIJ-SÖDERBERG DECOMPOSITIONS OF LEX-SEGMENT IDEALS
4
P
β(L, x) −
γi πd̄i
i=0
(a2): If γ4 =
k+1
= k+2
..
.
n
0
1
c0,k+1
c0,k+2
..
.
3c0,k+1 −(k+4)
2
c0,n
2c0,n
2c0,k+2
..
.
15
2
c0,k+2 −
k+2−c0,k+1
2
..
.
c0,n
3c0,k+1 + 4c0,k − (k + 4)
, which is forced by
4
k+4
> c0,k+1 and c0,k+1 + k < 4c0,k+2.
3
Then
0
β(L, x) −
4
P
γi πd̄i
i=0
k
k+1
= k+2
..
.
k+4−3c0,k+1
4
c0,k+1
1
-
c0,k+2
..
.
2c0,k+2
..
.
2
k+c0,k+1
c0,k+2 −
4
..
.
c0,n
2c0,n
c0,n
n
(a3): If γ4 = c0,k+2 +
c0,k+1 + 2c0,k − 2 − k
, which is caused by
2
2c0,k+2 + c0,k+1 < k + 2 and 4c0,k+2 < k + c0,k+1 .
Then the remaining diagram becomes
0
β(L, x) −
4
P
γi πd̄i
i=0
k
k+1
= k+2
k+3
..
.
c0,k+1
c0,k+2
c0,k+3
..
.
c0,k+1 −4c0,k+2 +k
2
2c0,k+2
2
-
2c0,k+3
..
.
c0,k+3
..
.
c0,n
2c0,n
c0,n
k+2−2c0,k+2 −c0,k+1
2
n
Case II.2.b : If γ3 =
1
-
kc0,k+1 + (k − 1)c0,k + 1
because of the following inequalities
(k + 1)(k + 3)
3c0,k+1 + 4c0,k < k + 4 and kc0,k+1 + (k − 1)c0,k + 1 < (k + 3)c0,k+2 .
Thus, the remaining diagram is
β(L, x) −
3
X
γi πd̄i
i=0
0
1
..
.
k
=
k+1
k+2
..
.
n
(k+1)(k+2)−2c0,k (2k+1)+3c0,k+1 (k−1)
(k+1)(k+2)
Case II.2.c : If γ3 =
..
.
c0,k
c0,k+1
c0,k+2
..
.
c0,n
1
..
.
-
2
..
.
-
2c0,k+2
..
.
2c0,n
(k+3)c0,k+2 −kc0,k+1 −(k−1)c0,k −1
k+3
..
.
c0,n
1 2c0,k + c0,k+1 − 1
−
as
2
2(k + 1)
2c0,k+2 + c0,k+1 + 2c0,k < k + 2 and (k + 3)c0,k+2 < kc0,k+1 + (k − 1)c0,k + 1.
16
SEMA GÜNTÜRKÜN
Then
β(L, x) −
2
X
γi πd̄i
i=0
0
1
=
2c0,k + c0,k+1 − 1
1−
−
k+1
.
.
.
k
c0,k
k+1
c0,k+1
2c0,k+2
k+1
.
.
.
1
2
-
-
.
.
.
-
.
.
.
-
kc0,k+1 + (k − 1)c0,k + 1
−
k+1
k+2
c0,k+2
2c0,k+2
k+3
.
.
.
n
c0,k+3
.
.
.
c0,n
2c0,k+3
.
.
.
2c0,n
(k+3)c0,k+2
k+1
c0,k+2 −
(k+1)c0,k+2
k+1
=0
c0,k+3
.
.
.
c0,n
We notice that the pattern of the above remaining diagram is similar to the one in the beginning of the
Case II.2. All possible top degree sequences (∗, ∗, k + 4) in Case II.2.a and Case II.2.b will be replaced
by (∗, ∗, k + 5) in the next steps of Case II.2.c.
Case III: If Gmax (L, x) = k + 1 then by Eliahou-Kervaire formula we get c0,k + c0,k+1 = k + 1 as
c
L is an Artinian lex segment ideal. Let γ1 = 0,k+1
k , then it requires that 2c0,k+1 + c0,k < k + 1 and
so c0,k+1 < 1 which is a contradiction. Thus this case does not exits if the maximum degree is k + 1.
Suppose that Gmax (L, x) > k + 1. Thus
c0,k+1
γ1 =
k
since
2c0,k+1 + c0,k < k + 1 and (k + 2)c0,k+1 < (k − 1)c0,k + 1.
Then the remaining diagram after subtracting two pure diagrams with corresponding coefficients is
c
c0,k − 1
0,k+1
β(L, x) −
π(1,k+1,k+3) =
π(1,k+1,k+2) −
k
k
..
.
0
k + 1 − (c0,k + 2co,k+1 )
k
..
.
k
c0,k
k+1
k+2
..
.
n
c0,k+1
c0,k+2
..
.
c0,n
1
1
2
-
-
..
.
(k − 1)c0,k + 1 − (k + 2)c0,k+1
k
2c0,k+1
2c0,k+2
..
.
2c0,n
..
.
c0,k+2
..
.
c0,n
Then the next top degree sequence is d̄2 = (1, k + 1, k + 4) and the coefficient
k + 1 − (c0,k + 2co,k+1 ) (k − 1)c0,k + 1 − (k + 2)c0,k+1 c0,k+2
γ2 = min
,
,
.
3k
k(k + 3)
k
Case III.1: Let γ2 =
β(L, x) −
2
P
i=0
k + 1 − (c0,k + 2co,k+1 )
. Then
3k
γi πd̄i
k
k+1
= k+2
..
.
n
0
c0,k
c0,k+1
c0,k+2
..
.
c0,n
1
2
-
(4(c0,k −1)−(c0,k+1 +k)
3
2c0,k+1
2c0,k+2
..
.
2c0,n
c0,k+2 −
(k+1)−(2c0,k+1 +c0,k )
3
..
.
c0,n
BOIJ-SÖDERBERG DECOMPOSITIONS OF LEX-SEGMENT IDEALS
17
Thus d̄3 = (k, k + 1, k + 4) and
c0,k (4(c0,k − 1) − (c0,k+1 + k)
(k + 1) − (2c0,k+1 + c0,k )
γ3 = min
,
, c0,k+2 −
.
3
12
3
We next observe one step more in the decomposition. So we get the following sub-cases sinceγ3 has two
possible cases:
c0,k
, the remaining diagram turns into
Case III.1.a: When γ3 =
3
0
β(L, x) −
3
P
γi πd̄i
i=0
k
1
k+1
= k+2
c0,k+2
..
..
.
.
n
c0,n
Case III.1.b: When γ3 = c0,k+2 −
β(L, x) −
3
P
γi πd̄i
i=0
c +k
+ 0,k4
c0,k+1
2
P
3c0,k+1 −k
4
..
.
c0,n
0
1
k
k + 1 − (3c0,k+2 + 2c0,k+1 ) k − (4c0,k+2 + 3c0,k+1 )
k+1
c0,k+1
2c0,k+1
k
+
2
c
2c
0,k+2
0,k+2
=
k+3
c0,k+3
2c0,k+3
..
..
..
.
.
.
c0,n
2c0,n
2
c0,k+3
..
.
c0,n
(k − 1)c0,k + 1 − (k + 2)c0,k+1
and then
k(k + 3)
γi πd̄i
i=0
c0,k+2 +
2c0,n
0
β(L, x) −
2c0,k+1
2c0,k+2
..
.
2
-
(k + 1) − (2c0,k+1 + c0,k )
, the remaining diagram has the form
3
n
Case III.2: Let γ2 =
1
-
1
..
.
k
=
k+1
k+2
..
.
n
1
..
.
-
k+4−(c0,k +co,k+1 )
k+3
..
.
c0,k
c0,k+1
c0,k+2
..
.
c0,n
2c0,k+1
2c0,k+2
..
.
2c0,n
2
..
.
c0,k+2 −
(k−1)c0,k +1−(k+2)c0,k+1
k+3
..
.
c0,n
c0,k+2
Case III.3: Let γ2 =
and then
k
0
1
..
.
β(L, x) −
2
P
i=0
γi πd̄i
k
= k+1
k+2
k+3
..
.
n
k+1−(c0,k +2co,k+1 +3c0,k+2 )
k
..
.
c0,k
c0,k+1
c0,k+2
c0,k+3
..
.
c0,n
1
..
.
2
..
.
(k−1)c0,k +1−(k+2)c0,k+1 −(k+3)c0,k+2
k
2c0,k+1
-
2c0,k+2
2c0,k+3
..
.
c0,k+3
..
.
2c0,n
c0,n
Hence we could pause the decomposing process since we have observed enough part of BS decomposition of the Betti diagram (L, x) so that we can compare each possible remaining diagram with the ones
will be obtained from BS decompositon of the Betti diagram of L.
We examine the BS decomposition of the lex ideal L. First of all, as a trivial case, we notice that if
Gmin (L) = 1, then the statement is vacuously true since L = (L, x).
18
SEMA GÜNTÜRKÜN
We next induct on the difference of the initial degrees Gmin (J) − Gmin (a) ≥ 1.
Base Step: In this step, we show that the statement is true for the lex ideals L = xa+J when Gmin (J)−
Gmin (a) = 1. That is, if Gmin (J) = k ≥ 2 then Gmin (a) = k − 1. So a = (x, y, z)k−1 since L is a lex
ideal. To this end, we modify the Betti diagram of L in the table 2 to this particular lex ideal L;
0
1
2
k(k+1)
k(k−1)
k
+ c0,k (k − 1)(k + 1) + 2c0,k − 1
+ c0,k − 1
2
2
k
+
1
c
2c
c
0,k+1
0,k+1
0,k+1
β(L) =
..
..
..
..
.
.
.
.
n
c0,n
2c0,n
c0,n
k(k − 1)
+
2
c0,k − 1, respectively. Then d1 = (k, k + 1, k + 3) becomes the next degree sequence with the coefficient
nk
o
α1 = min
, c0,k+1 .
3
Now, analyze each possible cases for the next step in the decomposition.
k
(i) If α1 = < c0,k+1 then the remaining diagram of of β(L) after three steps becomes
3
0
1
2
k
+
1
k
3
1
P
k
β(L) −
αi πdi = k + 1 c0,k+1 2c0,k+1 c0,k+1 − 3
.
.
.
.
i=0
..
..
..
..
Obviously the first top degree sequence and its coefficient are d0 = (k, k+1, k+2) and α0 =
n
c0,n
2c0,n
c0,n
k
k
If c0,k > + 1 and 2c0,k+1 + c0,k > k + 1 as in the Case I of (L, x) and since α1 = < c0,k+1 this
3
3
remaining diagrams matches with the one in the Case I.1,
β(L) −
1
X
αi πdi = β(L, x) −
2
X
γi πd̄i .
i=0
i=0
Hence we are done as the BS decompositions for L and (L, x) have the same remaining diagram.
Otherwise the BS decomposition of (L, x) results in either as in Case II or Case III, we keep decomposing the Betti diagram of L.
The next degree sequence is d2 = (k, k + 2, k + 3) and the coefficient of π(k,k+2,k+3) is
2c0,k+1 c0,k+1 k
k
α2 = min
+ 1,
,
−
.
3
3
2
6
As
k
3
+1<
c0,k+1
2
−
k
6
implies k + 2 < c0,k+1 which is not possible. So
α2 =
c0,k+1 k
− .
2
6
Therefore the remaining diagram is
0
β(L) −
2
P
i=0
αi πdi
k
k+1
=
k+2
..
.
n
If we have Case II for (L, x), which means
k+2−c0,k+1
2
c0,k+1
c0,k+2
..
.
c0,n
1
c0,k+1 +k
2
2c0,k+2
..
.
2c0,n
2
c0,k+2
..
.
c0,n
k
+ 1 > c0,k and (k − 1)c0,k + 1 < (k + 2)c0,k+1
3
BOIJ-SÖDERBERG DECOMPOSITIONS OF LEX-SEGMENT IDEALS
19
and also k + 2 < c0,k+1 + 2c0,k , so
β(L) −
2
X
αi πdi = β(L, x) −
3
X
γi πei , as in Case II.1.
i=0
i=0
If k + 2 > c0,k+1 + 2c0,k , which corresponds to Case II.2 for (L, x), again we move on to the next
step. So d3 = (k, k + 2, k + 4) and then
nk + 2 − c
o
0,k+1 c0,k+1 + k
α3 = min
,
, c0,k+2 .
2
4
This splits into three cases;
4
2
P
P
k+2−c0,k+1
πd3 = β(L, x) −
γi πei , as in Case II.2.a1..
• β(L) −
αi πdi −
2
• β(L) −
• β(L) −
i=0
2
P
i=0
2
P
i=0
αi πdi −
c0,k+1 +k
πd3
4
= β(L, x) −
4
P
γi πei , as in Case II.2.a2..
i=0
αi πdi − c0,k+2 πd3 = β(L, x) −
i=0
4
P
γi πei , as in Case II.2.a3..
i=0
We might need to recall that Case II.2.a for (L, x) requires k + 4 < 3c0,k+1 + 4c0,k and k + 2 <
2c0,k + c0,k+1 + 2c0,k+2 . If k + 4 > 3c0,k+1 + 4c0,k as in Case II.2.b, it contradicts the first assumption
k
3 < c0,k+1 . As Case II.2.c is a case similar to Case II.2, continuing the algorithm will leads us again the
same remaining diagram as we have Case II.2.a.
Thus, if either Case I or Case II holds for the decomposition of (L, x), we always end up with the
same remaining diagrams, even with the ones having size 3.
If the decomposition for (L, x) turns into the Case III, we get
2c0,k+1 + c0,k < k + 1 and (k + 2)c0,k+1 < (k − 1)c0,k + 1.
On the other hand, by the assumption k3 < c0,k+1 , the former and the latter inequalities imply c0,k < k3 +1
and k3 + 1 < c0,k , respectively and we have a contradiction. Thus, this situation cannot come true at all.
(ii) If α1 = c0,k+1 < k3 , and if the BS decomposition of (L, x) follows Case I then we have
0
1
2
k
k + 1 − 2c0,k+1 k − 3c0,k+1
1
2
P
k
+
1
c
2c
- = β(L, x) − P γi π i .
0,k+1
0,k+1
β(L) −
αi πdi =
d̄
..
..
..
..
i=0
i=0
.
.
.
.
n
c0,n
2c0,n
c0,n
If the decomposition of (L, x) follows different path, the next degree sequence in the decomposition
for L becomes d = (k, k + 1, k + 4) and the coefficient is
k + 1 − 2c0,k+1 k − 3c0,k+1
,
, c0,k+2 .
α2 = min
3
4
We move on the next sub-cases for each possibility for the coefficient α2 .
k − 3c0,k+1
(ii.a) If α2 =
thanks to the inequality
4
k < 3c0,k+1 + 4c0,k+2 ,
we have
0
β(L) −
2
P
i=0
αi πdi
k
k+1
=
k+2
..
.
n
k+4+c0,k+1
4
c0,k+1
1
-
2
-
c0,k+2
..
.
2c0,k+1
2c0,k+2
..
.
4c0,k+2 +3c0,k+1 −k
4
c0,n
2c0,n
c0,n
..
.
20
SEMA GÜNTÜRKÜN
If the algorithm follows through as in the Case III for the decomposition of (L, x), the relation
k < 3c0,k+1 + 4c0,k+2 yields the same the remaining diagram. Hence
β(L) −
2
X
αi πdi = β(L, x) −
3
X
γi πd̄i .
i=0
i=0
If the Betti diagram of (L, x) is decomposed as in the Case II, we keep decomposing the
Betti diagram of L and get three sub-cases for the next coefficient
α3 = min
nk + 4 + c
0,k+1
, c0,k+1 ,
4c0,k+2 + 3c0,k+1 − k
}.
4
4
k + 4 + c0,k+1
If α3 =
, that is,
4
k + 4 < 3c0,k+1 and k + 2 < 2c0,k+2 + c0,k+1
then
0
1
2
3c0,k+1 −k−4
k + 1 c0,k+1
2
3
P
k+2c0,k+1
2c0,k+2
c0,k+2−
β(L) −
αi πdi = k + 2 c0,k+2
2
i=0
..
..
..
..
.
.
.
.
n
c0,n
2c0,n
c0,n
4
P
which is equal to β(L, x) −
γi πd̄i since the same relations are also required for the Case
i=0
II.2.a.1 for the decomposition of (L, x).
If α3 = c0,k+1 , that is, the relations are
3c0,k+1 < k + 4 and k < 4c0,k+2 − c0,k+1 ,
then
0
β(L) −
3
P
αi πdi
i=0
k
k+1
=
k+2
..
.
k+4−3c0,k+1
4
c0,k+1
1
-
2
-
c0,k+2
..
.
2c0,k+2
..
.
4c0,k+2 −c0,k+1 −k
4
..
.
n
c0,n
2c0,n
c0,n
and it is the same remaining diagram as in Case II.2.a.2 for (L, x) because of the same required
relations.
4c0,k+2 + 3c0,k+1 − k
If α3 =
, which is a consequence of the following inequalities
4
k + 2 < 2c0,k+2 + c0,k+1 and 4c0,k+2 − c0,k+1 < k.
Then
0
β(L) −
3
P
i=0
αi πdi
k
k+1
= k+2
k+3
..
.
n
= β(L, x) −
4
X
i=0
c0,k+1
c0,k+2
c0,k+3
..
.
c0,k+1 −4c0,k+2 +k
2
2c0,k+2
2
-
2c0,k+3
..
.
c0,k+3
..
.
c0,n
2c0,n
c0,n
k+2−(2c0,k+2 +c0,k+1 )
2
1
-
γi πd̄i as in Case II.2.a.3 for (L, x).
BOIJ-SÖDERBERG DECOMPOSITIONS OF LEX-SEGMENT IDEALS
21
(ii.b) If α2 = c0,k+2 then the remaining diagram becomes
0
1
2
k
k + 1 − 2c0,k+1 − 3c0,k+2 k − 3c0,k+1 − 4c0,k+2
k+1
c0,k+1
2c0,k+1
2
P
k
+
2
c
2c
0,k+2
0,k+2
β(L) −
αi πdi =
i=0
k+3
c0,k+3
2c0,k+3
c0,k+3
..
..
..
..
.
.
.
.
n
c0,n
2c0,n
c0,n
The above diagram matches with remaining diagrams as in the Case III.1.b for the decomposition of (L, x) if (L, x) decomposes as in Case III. Now suppose that BS decomposition of
β(L, x) after several steps ends up as in the Case II. We want to show that same remaining
diagram occurs for L as well. If Gmax is not k + 3, we may have another case which gives us
similar pattern like above diagram such as
2
P
β(L) −
αi πdi − c0,k+3 π(k,k+1,k+5) =
i=0
0
1
2
k
k + 1 − 2c0,k+1 − 3c0,k+2 − 4c0,k+3 k − 3c0,k+1 − 4c0,k+2 − 5c0,k+3
k+1
c0,k+1
2c0,k+1
k+2
c0,k+2
2c0,k+2
k+3
c0,k+3
2c0,k+3
k+3
c0,k+4
2c0,k+4
c0,k+4
..
..
..
..
.
.
.
.
n
c0,n
2c0,n
c0,n
k − (3c0,k+1 + 4c0,k+2 )
Thus, assume the maximum degree is k+3. Therefore the next coefficient is α3 =
5
with the pure diagram πk,k+1,k+5. Then the remaining diagram turns into
0
β(L) −
3
P
k
αi πdi = k + 1
i=0
k+2
k+3
k+5+2c0,k+2 +c0,k+1
2
c0,k+1
c0,k+2
c0,k+3
1
2c0,k+1
2c0,k+2
2c0,k+3
2
5c0,k+3 +4c0,k+2 +3c0,k+1 −k
5
Also, we have assumed that the Betti diagram of (L, x) decomposes as in Case II. We notice
that the entries of the remaining diagram is closely related to the one in Case II.2.a3.
0
β(L) −
3
P
i=0
αi πdi
k
= k+1
k+2
k+3
k+2−2c0,k+2 −c0,k+1
2
c0,k+1
c0,k+2
c0,k+3
+ 3D
1
c0,k+1 −4c0,k+2 +k
2
2c0,k+2
2c0,k+3
+ 5D
2
c0,k+3 + 2D
4c0,k+2 +3c0,k+1 −k
.
10
where D =
Thus the next coefficient is
c0,k+1 − 4c0,k+2 + k
c0,k+3
k + 2 − 2c0,k+2 − c0,k+1
+ D,
+ D,
+D
α4 = min
6
10
2
which follows the same paths for the decomposition of (L, x) in Case II.2.a3.
BS decompositions of β(L) and β(L, x) always come up with the same remaining diagrams after
several steps of the decomposition. Moreover, we observe that they share not only the length two pure
diagrams and their coefficients but also some length three pure diagrams.
Hence, in every possible case we end up with the same remaining diagrams for both β(L) and β(L, x).
In other words, the BS decompositions of L and (L, x) coincide precisely after several steps of the
algorithm. Thus the statement holds for the case of Gmin (J) − Gmin (a) = 1.
Induction Hypothesis: Let the statement be true for all lex ideals L = xa+J with Gmin (J)−Gmin (a) =
22
SEMA GÜNTÜRKÜN
N ≥ 1. We need to show that it is also true for the lex ideals satisfying Gmin (J) − Gmin (a) = N + 1.
To this end, we identify the initial degrees of J and a by Gmin (J) = k and Gmin (a) = m.
Suppose that L = xa + J is a lex ideal such that k − m = N + 1. So k − m = N + 1 ≥ 2. We prove
this into two cases.
Case A: If y m ∈
/ a. Since a is a lex ideal, we write a = xb + I. Then we notice that Gmin (I) 6= k.
Otherwise it contradicts to y k ∈ G(J). Thus k Gmin (I) m as y m ∈
/ a. Define ã ⊂ a as the ideal
containing all monomials of a of degree greater or equal to m + 1 and note that ã is also a lex ideal with
Gmin (ã) = m + 1. Define L̃ = xã + J and it is a lex ideal with Gmin (J) − Gmin (ã) = k − (m + 1) =
k − m − 1 = N + 1 − 1 = N . Therefore, by the induction hypothesis, β(L̃) and β(L, x) have the same
ends in their BS decompositions, i.e. same pure diagrams of length less than 3 with same coefficients,
X
X
β(L̃) −
α̃i πd̃i = β(L, x) −
γi πdi .
di , all length 3 degree sequences
d̃i , all length 3 degree sequences
˜ It is easy to see that I˜ = I as y m ∈
On the other hand, ã can be decomposed as ã = xb̃ + I.
/ a and
Gmin (b̃) = m. Clearly, Gmin (I) − Gmin (b̃) ≤ (k − 1) − m = N . Again by the induction hypothesis
BS decompositions of β(ã) and β(I, x) have the same ends.
Recall that a = xb + I, so we get Gmin (I) − Gmin (b) ≤ (k − 1) − (m − 1) = k − m = N + 1.
Suppose that Gmin (I) − Gmin (b) < N + 1, then thanks to the induction hypothesis, BS decompositions of a and (I, x) have the same ends, so do a and ã. That is,
X
X
D := β(a) −
All length 3 pure diagrams = β(ã) −
All length 3 pure diagrams.
Also using the Theorem 1.1 BS decompositions for the ideals L and L̃ can be observed as;
0
β(L) =
di
X
αi πdi +
with l(di )=3
1
Remaining
2
..
.
diagram, D
k
..
.
βi,i+j (L, x), i ≥ k
2
and
0
β(L̃) =
d̃i
X
with l(d̃i )=3
α̃i πd̃i +
1
Remaining
2
..
.
diagram, D
k
..
.
βi,i+j (L, x), i ≥ k
2
This shows that β(L) and β(L̃) have same ends and we also know that β(L̃) and β(L, x) have the same
ends. Hence the statement is true.
It remains to study when Gmin (I) − Gmin (b) = (k − 1) − (m − 1) = k − m = N + 1, which
means Gmin (I) = k − 1. It follows from Gmin (J) = k that I = (y, z)k−1 . Then a = xb + I and
b = xb + I where Gmin (I) − Gmin (b) ≥ (k − 2) − (m − 2) ≥ N + 1. If it is a strict inequality then
applying the same process as we have done for L can be applied to a to prove the statement. If there is
an equality, we end up with the same situation. L = xa + J where Gmin (J) = k, Gmin (a) = m and
k − m = N + 1, and a = xb + I where I = (y, z)k−1 , Gmin (b) = m − 1, and b = xb + I where
I = (y, z)k−2 , Gmin (b) = m − 2. We repeat this until we get
c = x(x, y, z t−1 ) + K where K = (y, z)s , s = k − m + 1, 1 ≤ t ≤ k − m.
For this form of the lex ideal, one can check the BS decomposition of the ideal c.
BOIJ-SÖDERBERG DECOMPOSITIONS OF LEX-SEGMENT IDEALS
β(c) =
2
..
.
0
2
..
.
1
1
..
.
2
..
.
t
..
.
1
..
.
2
..
.
1
..
.
23
s s + 1 2s + 1 s
1
t
=
+
+ 1
2: t−1 t −
t:
−
− 1
2: s−t+1 −
−
1
−
s
−
s t:
s:
−
− t−1
+
1
t
+
t−1
s
t: 1
−
−
+
s : − s−t+2 s−t+1
2: 1
−
−
t : − t−1 t
2: 1 −
−
s: − s s−1
s s: 1 1 +1 s: 1
|
{z
}
same end as in the decomposition of (L,x)
Therefore the statement is true for the ideal c. So we may assume that, without loss of generality, a is in
the form of c, i.e., Gmin (I) − Gmin (a) < N + 1. This observation completes the proof for Case I.
Case B: Let y m ∈ a.
(i) If m ∈
/ 1; we write a = xb + I and Gmin (I) = m > 1. This implies that b = (x, y, z)m−1 .
Consider ã = (a, x) = x(1)+I. Clearly ã = (ã, x), so the statement is trivially true for the ideal
ã. Moreover, Gmin (I) − Gmin (a) = m − (m − 1) = 1. By the base case, the decompositions
of β(a) and β(I, x) have the same ends. Hence,
X
X
β(a) −
all length 3 pure diagrams =
β(I, x) −
all length 3 pure diagrams
X
all length 3 pure diagrams.
=
β(ã) −
Similar to the Case 1, consider the lex ideal L̃ = xã + J and y Gmin (ã) = y ∈
/ ã. Thus by the
result of the Case 1, the statement is true for L̃. We do exactly the same trick as in Case 1 to
show that β(L) and β(L̃) have the same ends and it follows that the statement holds for L.
(ii) If m = 1; that is, a = (x, y, z t ) where 1 ≤ t ≤ k − 1. In the Case 1 we have already
shown that the BS decomposition of the β(L) satisfy the statement if L = x(x, y, z t ) + J where
J = (y, z)k . Nevertheless, for more general stable ideal J ⊂ k[y, z] we have already assumed
that L cannot be in that form in the statement.
Conjecture 4.1. The statement of Theorem 1.2 holds for all Artinian lex-ideals in k[x, y, z].
Theorem 1.2 shows that the ends of the Boij-Söderbeg decompositions of L and (L, x) = (J, x) are
exactly the same for all Artinian lex ideals L in R except the ones in the form of L = x(x, y, z t ) + J
where J is different from (y, z)Gmin (J) and 1 < t < k −1. On the other hand, based on the computations
we have done using the BoijSoederberg packages of the computer algebra software Macaulay2, see
[11], we strongly believe that this result is also true for the lex ideals in that particular form.
5. F URTHER O BSERVATIONS
AND
E XAMPLES
For an Artinian lex ideal L ⊂ k[x, y, z] of codimension 3, we have shown that the summands of length
3 pure diagrams of the Boij-Söderberg decomposition of a where a = L : (x), and the summands of pure
diagrams of length less than 3 in the Boij-Söderberg decomposition of (L, x) appear in the decomposition
24
SEMA GÜNTÜRKÜN
of the ideal L = a(x) + J in the beginning and in the end, respectively.
length 3 degree
all length < 3 degree
extra length 3
β(L) = sequences coming +
+ sequences coming .
degree sequences
from a(−1)
from (L, x)
There might be also some other pure diagrams of length 3 other than the ones coming from the BS
decomposition of a. However, how this middle part containing pure diagrams of length 3 comes out is
not quite clear. One might ask whether or not the ideals b = L : (y) and c = L : (z) help to describe
the middle part. In fact, examples show that there is a quite strong relation between them as well.
Nevertheless, both the results obtained in sections 3 and 4 about β(a) and β(L, x) and the observations
we discuss in this section about β(b) and β(c) provide a very close approximation for the BS chain
of degree sequences for β(L). Either our observation in this section is not enough to cover the entire
middle part of the decomposition of β(L) or the BS decompositions of b and c may give redundant
degree sequences.
Now in this section we illustrate the elation between the BS decompositions of the ideals b, c and L
via examples.
Example 5.1. Let
L = (x2 , xy 2 , xyz, xz 2 , y 8 , y 7 z, y 6 z 2 , y 5 z 3 , y 4 z 4 , y 3 z 5 , y 2 z 6 , yz 7 , z 8 )
be a lex segment ideal in R. Then a = L : x = (x, y 2 , yz, z 2 ) is lex segment ideal such that L = xa + J
where J = (y, z)8 is stable in R and lex segment in k[y, z]. Similarly the ideals
b = L : y = (x2 , xy, xz, y 7 , y 6 z, y 5 z 2 , y 4 z 3 , y 3 z 4 , y 2 z 5 , yz 6 , z 7 ) = L : z = c
are lex segment ideals such that L = yb + I = zc + K where I = (x2 , xz 2 , z 8 ) and K = (x2 , xy 2 , y 8 ).
We construct similar short exact sequences like in Lemma 2.8 for the ideals b and c. Unlike the case for
a, we might have cancellations in the mapping cone of the short exact sequences for ideals. It means we
can have cancellations in the Betti diagram since the mapping cone structure may not yield the minimal
free resolution. This situation causes different degree sequences that do not appear in BS decomposition
of L.
We first notice that b = c and find the BS decomposition of β(a)
β(a) = (1)π(1,3,4) + [pure diags. of length < 3],
Then we consider the short exact sequence for the ideals b and L = yb + I
0 −
→
I(−1)
x
→
−
b(−1) ⊕ I
x
R(−3) ⊕ R(−4) ⊕ R(−9)
x
R(−2) ⊕ R4 (−3) ⊕ R9 (−8)
x
0
R(−5) ⊕ R7 (−10)
x
R(−5) ⊕ R(−10)
x
R4 (−4) ⊕ R16 (−9)
x
→
−
L
x
R3 (−3)
z }| {
✘✘ ⊕R9 (−8)
R4✘
(−3)
R(−2) ⊕ ✘
x
✘ ⊕ R5 (−4) ⊕ R17 (−9)
✘✘
R(−3)
✘
x
R2 (−5) ⊕ R8 (−10)
x
0
0
The mapping cone of the short exact sequence for ideals b and L (so the same for c and L = zc + K)
ends up with “one” cancellation in the first degree. So we interpret this as ignoring one pure diagram
→ 0
−
BOIJ-SÖDERBERG DECOMPOSITIONS OF LEX-SEGMENT IDEALS
25
at the beginning, which is the one corresponding to the degree sequence (2, 3, 4) at the beginning of the
decomposition of β(b). Therefore,
8
1
✘✘
β(b) = β(c) = ✘
(1)π
✘✘
(2,3,4) + ( )π(2,3,9) + ( )π(2,8,9) + [pure diags. length < 3].
7
7
The pure diagrams of length less than 3 are coming from the ideal
β(L, x) = [length 3 pure diags.] + (8)π(8,9) + (1)π(8) .
Hence we claim that the summands (with coefficients) in the BS decomposition of β(L) are
β(L) ≈ (1)π(2,4,5) +(α2 ) π(3,4,10) +(α3 ) π(3,9,10) + (8)π(8,9) + (1)π(8) ,
| {z }
| {z }
|
{z
}
| {z }
from a(−1)
from b(−1)
from b(−1)
from (L,x)
for some coefficients α2 , α3 in Q. Indeed, the BS decomposition of L is,
2
9
β(L) = (1)π(2,4,5) + ( )π(3,4,10) + ( )π(3,9,10) + (8)π(8,9) + (1)π(8) .
7
7
The impressive point of this example is that we are able to describe the entire BS chain of degree
sequences of L from its the colon ideals a, b and the ideal (L, x).
Example 5.2. This example shows that some different situations might occur other than the previous
example.
Let L = (x2 , xy 2 , xyz, xz 2 , y 4 , y 3 z, y 2 z 2 , yz 6 , z 9 ) be lex-segment ideal in R. Then a = L : x =
(x, y 2 , yz, z 2 ), b = L : y = (x2 , xy, xz, y 3 , y 2 z, yz 2 , z 6 ) and c = L : (z) = (x2 , xy, xz, y 3 , y 2 z, yz 5 , z 8 ).
We observe that one cancellation occurs in the mapping cone process of each ideal b and c. BS decompositions of a, b, c and (L, x) are
β(a) = 1π(1,3,4) + [pure diags. of length < 3],
5
1
7
1
✘✘ + π(2,3,5) + π(2,4,5) + π(2,4,8) +
π
β(b) = ✘
1π✘
(2,3,4)
3
6
4
20 (3,4,8)
1
+ π(3,7,8) + [pure diags. of length < 3],
10
1
1
1
1
1
✘✘ + π(2,3,5) + π(2,4,5) + π(2,4,8) +
β(c) = ✘
1π✘
π(3,4,8) + π(3,7,8)
(2,3,4)
3
3
2
10
10
3
1
+ π(3,7,10) + π(3,9,10) + [pure diags. of length < 3]
14
42
and
β(L, x) = [pure diags. of length 3] + 1π(4,10) + 1π(7,10) + 1π(9) .
So, the BS decomposition for the ideal L is likely to be
β(L) ≈ 1π(2,4,5) + α2 π(3,4,6) + α3 π(3,5,6) + α4 π(3,5,9) + α5 π(4,5,9)
+ α6 π(4,8,9) + α7 π(4,8,11) + α8 π(4,10,11) + 1π(4,10) + 1π(7,10) + 1π(9)
where αi ∈ Q, i = 2, ...8. Thus it seems that we almost obtain the actual BS decomposition for L
which is
2
2
+
π(3,4,6)
π(3,5,6)
β(L) = 1π(2,4,5) +
3
3
| {z }
| {z }
| {z }
from a(−1)
from b(−1) and c(−1)
from b(−1) and c(−1)
3
1
1
+
+
+
π(3,5,9)
π(4,5,9)
π(4,8,9)
2
10
20
| {z }
| {z }
| {z }
from b(−1) and c(−1)
from b(−1) and c(−1)
from b(−1) and c(−1)
1
+ 1π(4,10) + 1π(7,10) + 1π(9) .
+
π
4 | (4,8,11)
|
{z }
{z
}
from c(−1)
from (L,x)
26
SEMA GÜNTÜRKÜN
Apparently, the BS decomposition of c provides an additional pure diagram, π(4,10,11) , which does not
appear in the BS decomposition of L. Nevertheless it still supports the idea of the covering up the middle
part of the decomposition of β(L) by using decompositions of β(b) and β(c).
Example 5.3. In the previous example we saw that our approximation for L gives a longer BS chain of
the degree sequences than the actual BS chain of the degree sequences via the BS decomposition of the
ideals a, b, c and (L, x).
Consider the lex-segment ideal L = (x2 , xy, xz 2 , y 6 , y 5 z, y 4 z 3 , y 3 z 4 , y 2 z 5 , yz 6 , z 9 ) in R. Then the
colon ideals are a = L : x = (x, y, z 2 ), b = L : y = (x, y 5 , y 4 z, y 3 z 3 , y 2 z 4 , yz 5 , z 6 ), and c = L : z =
(x2 , xy, xz, y 5 , y 4 z 2 , y 3 z 3 , y 2 z 4 , yz 5 , z 8 ). The mapping cone for the ideal c requires two cancellations,
so we ignore the first two degree sequences. Then,
1
1
β(a) = π(1,2,4) + π(1,3,4) + [pure diags. of length < 3],
3
3
1
9
2
1
β(b) = π(1,6,7) + π(1,6,8) + π(1,7,8) + π(5,7,8) + [pure diags. of length < 3],
5
35
7
2
19
1
1
1 ✟✟ 1
✘
✘ + π
π(2,7,8) + π(2,7,10) + π(5,7,10)
β(c) = ✘
1π✘
✟(2,3,8) + π(2,6,8) +
(2,3,4)
✟
6
3
30
15
3
+ [pure diags. of length < 3], and
1
β(L, x) = [pure diags. length 3] + π(6,8) + 2π(7,8) + 2π(7,10) + 1π(9) .
2
Then, we get the following chain of degree sequences in order to set up the approximate BS decomposition for L
β(L) ≈ (2, 3, 5) < (2, 4, 5) < (2, 7, 8) < (2, 7, 9) < (2, 8, 9) < (6, 8, 9)
{z
} |
{z
}
|
from a(−1)
from b(−1)
< (3, 7, 9) < (3, 8, 9) < (3, 8, 11) < (6, 8, 11) < (7, 9) < (8, 9) < (8, 11) < (10) .
|
{z
} |
{z
}
from c(−1)
from (L,x)
However, we know that the degree sequences in the decomposition must be a partial ordered chain, so
the ones that violate the partial order are needed to be eliminated. From the decomposition of β(c), we
get (3, 7, 9) as the first degree sequence, but (2, 8, 9) and (6, 8, 9) cannot be before (3, 7, 9). So we have
to ignore the sequences (2, 8, 9) and (6, 8, 9). Then we get an approximate decomposition such as
1
1
β(L) ≈ π(2,3,5) + π(2,4,5) + α3 π(2,7,8) + α4 π(2,7,9) + α7 π(3,7,9)
3
3
1
1
+ α8 π(3,8,9) + α9 π(3,8,11) + α10 π(6,8,11) + π(6,8) + π(7,9)
2
2
+ 2π(8,9) + 2π(8,11) + 1π(10) .
The BS decomposition of β(L) is
1
1
1
2
1
β(L) = π(2,3,5) + π(2,4,5) + π(2,4,8) + π(2,7,8) + π(2,7,9)
3
3
3
15
10
1
1
1
1
1
+ π(3,7,9) + π(3,8,9) + π(6,8,11) + π(6,8) + π(7,9)
2
2
2
2
2
+ 2π(8,9) + 2π(8,11) + 1π(10) .
The degree sequence (3, 8, 11) associated with (2, 7, 10), which is coming from the decomposition of
β(c), does not show up in the decomposition of β(L), similar to the situation in Example 5.2. Moreover,
for this lex ideal L, we realize another different situation. The degree sequence (2, 4, 8) shows up in the
BS chain of degree sequences of β(L), but (2 − 1, 4 − 1, 8 − 1) = (1, 3, 7) does not appear in any of the
decompositions of β(a), β(b) and β(c).
An explanation for that extra degree sequence (2, 4, 8) might be possible for this example. We see that
(2, 4, 5) is the last degree sequence coming from a(−1) and the next degree sequence (2, 7, 8) is from
b(−1). If we assume that there is no other degree sequence between (2, 4, 5) and (2, 7, 8), it implies
BOIJ-SÖDERBERG DECOMPOSITIONS OF LEX-SEGMENT IDEALS
27
that simultaneous elimination of the entries in the positions of β1,4 and β2,5 in the Betti diagram of L by
the algorithm of BS decomposition. However, this is not possible because otherwise there would not be
a pure diagram of length 2 in the BS decomposition of a. Hence again by the partial order, it must be
(2, 4, 5) < (2, 4, 8) < (2, 7, 8).
Examples 5.2 and 5.3 show that the BS decompositions of a, b, c and (L, x) may not be enough to
provide the entire chain of degree sequences in the BS decomposition of L. Therefore, it is possible
that there are some gaps and redundant degree sequences in the approximation of BS chain of degree
sequences of L. In view of the explanations, such as the cancellations in mapping cone, the necessity of
the order of the chain of the degree sequences, we are able to provide the entire chain of degree sequences
in the BS decomposition of L.
Problem 5.4. Is it true that the the Boij-Söderberg decomposition of a lex ideal L can be described by the
decompositions of its colon ideals a, b, c and (L, x) precisely. That is, in terms of all the pure diagrams
and their coefficients?
The relation between BS decompositions of a (Artinian) lex-segment ideal L and the lex ideals a =
L : x and (L, x) is pointed out in Theorems 1.1 and 1.2. Furthermore, the examples we have observed
in this section show that if we know the BS decompositions of the colon ideals b = L : y and c = L : z,
then almost the entire BS chain of the degree sequences for the lex-segment ideal L may be revealed. In
other words, we try formalize the full chain of degree sequences of the BS decomposition of the ideal L
by using the BS chains of degree sequences of the colon ideals a,b, c and the lex ideal (L, x). Studying
what the observations indicate is a further direction for our research on BS decomposition of lex-segment
ideals.
A natural follow-up work which aims to describe all BS coefficients of a lex ideal L in terms of the
coefficients of its colon ideals a, b, c and the larger lex ideal (L, x) may arise at this point, i.e., we narrow
our attention on the degree sequences, that is, pure diagrams. Although the results about a and (L, x)
involve the coefficients as well, we do not have a foresight regarding a relation between the coefficients
in the decompositions of b, c and L based on the observations mentioned in this section.
ACKNOWLEDGMENT
The author would like to thank to her Ph.D. adviser Uwe Nagel for proposing the problem about
Boij-Söderberg decompositions of lex-segment ideals and his contributions and discussions during the
course of preparing this manuscript. She would also like to thank Daniel Erman for introducing her to
the Boij-Söderberg Theory during a 2011 Summer Graduate Workshop at MSRI.
R EFERENCES
[1] Christine Berkesch, Daniel Erman, Manoj Kummini, and Steven V. Sam. Poset structures in Boij-Söderberg theory. Int.
Math. Res. Not. IMRN, (22):5132–5160, 2012.
[2] Anna Maria Bigatti. Upper bounds for the Betti numbers of a given Hilbert function. Comm. Algebra, 21(7):2317–2334,
1993.
[3] Mats Boij and Jonas Söderberg. Graded Betti numbers of Cohen-Macaulay modules and the multiplicity conjecture. J.
Lond. Math. Soc. (2), 78(1):85–106, 2008.
[4] Mats Boij and Jonas Söderberg. Betti numbers of graded modules and the multiplicity conjecture in the non-CohenMacaulay case. Algebra Number Theory, 6(3):437–454, 2012.
[5] David Cook, II. The structure of the Boij-Söderberg posets. Proc. Amer. Math. Soc., 139(6):2009–2015, 2011.
[6] David Eisenbud and Frank-Olaf Schreyer. Betti numbers of graded modules and cohomology of vector bundles. J. Amer.
Math. Soc., 22(3):859–888, 2009.
[7] Shalom Eliahou and Michel Kervaire. Minimal resolutions of some monomial ideals. J. Algebra, 129(1):1–25, 1990.
[8] Gunnar Fløystad. Boij-Söderberg theory: introduction and survey. In Progress in commutative algebra 1, pages 1–54. de
Gruyter, Berlin, 2012.
[9] Chirstopher Francisco, Jeffrey Mermin, and Jay Scheweig. Boij-söderberg and veronese decompositions.
https://math.okstate.edu/people/mermin/papers/Boij-Soderberg_and_Veronese_decompositions.pdf.
[10] Courtney Gibbons, Jack Jeffries, Sarah Mayes, Claudiu Raicu, Branden Stone, and Bryan White. Non-simplicial decompositions of betti diagrams of complete intersections. arXiv:1301.3441.
[11] Daniel R. Grayson and Michael E. Stillman. Macaulay2, a software system for research in algebraic geometry.
http://www.math.uiuc.edu/Macaulay2/.
28
SEMA GÜNTÜRKÜN
[12] J. Herzog and M. Kühl. On the Betti numbers of finite pure and linear resolutions. Comm. Algebra, 12(13-14):1627–1646,
1984.
[13] Heather A. Hulett. Maximum Betti numbers of homogeneous ideals with a given Hilbert function. Comm. Algebra,
21(7):2335–2350, 1993.
[14] Uwe Nagel and Stephen Sturgeon. Combinatorial interpretations of some Boij-Söderberg decompositions. J. Algebra,
381:54–72, 2013.
[15] Keith Pardue. Deformation classes of graded modules and maximal Betti numbers. Illinois J. Math., 40(4):564–585, 1996.
E-mail address: [email protected]
D EPARTMENT OF M ATHEMATICS , U NIVERSITY
MI 48109-1043, USA
BOR ,
OF
M ICHIGAN , 2074 E AST H ALL , 530 C HURCH S TREET A NN A R -
| 0math.AC
|
1
Offline signature authenticity verification through
unambiguously connected skeleton segments
arXiv:1711.03082v1 [cs.CV] 8 Nov 2017
Jugurta Montalvão, Luiz Miranda, and Jânio Canuto
Abstract—A method for offline signature verification is presented in this paper. It is based on the segmentation of the
signature skeleton (through standard image skeletonization) into
unambiguous sequences of points, or unambiguously connected
skeleton segments corresponding to vectorial representations of
signature portions. These segments are assumed to be the fundamental carriers of useful information for authenticity verification,
and are compactly encoded as sets of 9 scalars (4 sampled
coordinates and 1 length measure). Thus signature authenticity is
inferred through Euclidean distance based comparisons between
pairs of such compact representations. The average performance
of this method is evaluated through experiments with offline
versions of signatures from the MCYT-100 database. For comparison purposes, three other approaches are applied to the
same set of signatures, namely: (1) a straightforward approach
based on Dynamic Time Warping distances between segments,
(2) a published method by [18], also based on DTW, and (3)
the average human performance under equivalent experimental
protocol. Results suggest that if human performance is taken as a
goal for automatic verification, then we should discard signature
shape details to approach this goal. Moreover, our best result
– close to human performance – was obtained by the simplest
strategy, where equal weights were given to segment shape and
length.
Index Terms—Pseudo-online signatures, Skeletonization, Mean
Opinion Score (MOS).
I. I NTRODUCTION
ANDWRITTEN signature is a form of personal identification widely accepted, both socially and legally, and
it has been used for centuries to authenticate documents such
as bank checks, letters, contracts and many other that require
proof of authorship. By signing, a person may provide unique
information regarding the way she or he converts gesture
intentions into spontaneous hand movement. Writing speed,
traversed path, pen tilt, pressure applied, all these data are
articulated to result in a static figure on signed documents
[12].
Signature analysis can be divided in two categories: offline
and online. In the offline mode, either signatures are available
through the traditional wet ink method (such as in paper
documents), or they are available in scanned form, through
optical devices, such as scanners and digital cameras. In both
cases, all available data corresponds to static signature images.
This type of approach is not the most efficient for verifying
signatures due to the fact that relevant dynamic information is
discarded.
In the online mode, a person uses a digitizing device (e.g.
digitizing tablets or touchsceen devices) to directly record
signals from the hand movement. This provides much more
information than a static image, for the digitizing device
H
typically can record several complementary signals, such as
the path travelled by the pen tip, as well as its instantaneous
speed, applied pressure and pen tilt. This approach is the one
that dominates research on signature verification now, due to
the worldwide spreading of affordable acquisition devices [15].
However, the offline approach still has some attractive
aspects. For instance, even today, many contracts and credit
card authorization are performed through traditional signatures
on paper. Indeed, although online signature verification has
higher reliability, in many practical situations, for economical
or practical reasons, wet ink signatures are yet useful biometric
signals. And even in the unlikely scenario of a complete
substitution of wet ink signatures by electronically acquired
ones, at least the task of signature verification from ancient
ink on paper documents should remain a relevant topic, due to
the large amount of old signed documents, whose authenticity
are potentially waiting to be verified [16].
To give some fundamental definitions and jargon, we assume that a signature verification is a process that determines
whether a tested signature was produced by a target individual,
from which at least one genuine signature is available. If
under some chosen criteria the tested signature is similar
to the genuine references, below a pre-established similarity
threshold, it is labelled true, or a genuine signature. Otherwise,
the signature is labelled false, or a forgery. Moreover, Coetzer
et al. [4] classify forgeries as
•
•
•
Random forgery: The forger does not know the author’s
name neither the original signature. Thus the false signature is completely random.
Simple forgery: The forger knows the author’s name, but
she/he does not have access to the original signature.
Skilled forgery: The forger has access to samples of
genuine signatures, and also knows the name of the
author. It can also be divided into two classes: Amateur
and Professional. The Professional Skilled Forgery is
produced by a person with professional expertise in
handwriting analysis, being able to produce a higher
quality forgery than the Amateur.
In general, the offline signature verification process can be
divided into four steps [1]: Acquisition, Preprocessing, Feature
Extraction and Comparison. In the Preprocessing step, image
quality is improved and pixels are transformed to reduce the
computational burden of the subsequent steps. Examples of
techniques applied in this step are: thinning, color conversion,
noise reduction, smoothing, morphological operations and resizing. For instance, Shah et al. [17] cropped images to exclude
redundant white regions. The Feature Extraction step is where
2
most works propose innovations. According to Batista et al.
[2], an ideal feature extraction technique extracts a minimal
feature set that maximizes interpersonal variability amongst
signature samples from various subjects, whereas it minimizes
intrapersonal variability amongst samples belonging to the
same subject. Lee and Pan [10] divide the features into three
classes: Global Features, Local Features and Geometrical
Features.
Typical features extracted from offline signatures are
marginal projections. Shanker and Rajagopalan [18] extracts
vertical projection of bitmaps corresponding to signatures, thus
yielding profiles which are compared through Dynamic Time
Warping (DTW). Likewise, Coetzer et al. [4] pushes a bit further the same idea, by using many marginal projections of the
same signature, over different angles, what they call Discrete
Radon Transform, whose behaviour is modelled with a Hidden
Markov Model. Nguyen et al. [13] also use similar projections.
Indeed, they use two techniques for global features extraction:
the first is derived from the total energy a writer uses to create a
signature, whereas the second technique employs information
from the vertical and horizontal projections of a signature,
focusing on the proportion of the distance between key strokes
in the image, and the height/width of the signature.
Although marginal projections are more commonly used
in literature, straightforward approaches to feature extraction
may also rely upon image skeletonization [7]. Typically, skeletonization is used to filter foreground pixels in bitmaps. But it
can also be used to map offline signatures into sets of points,
similar to online representations, which is appealing because
online verification techniques may be deployed, such as the
use of DTW to compare segments of points from different
signatures. Indeed, this straightforward approach corresponds
to the baseline method implemented in this paper, as explained
in Section III.
Once features are available, signature authenticity verification can be performed. To simulate actual verification, most
academic works randomly select a small number of genuine
signature samples from each user (typically from 5 to 15) to
play the role of a set of enrolled signatures. Then, samples
from the remaining dataset of false and genuine signatures
are randomly taken to simulate verification attempts. These
test samples are compared to the enrolled samples, and a
decision is made. If a genuine signature is rejected, it is called
a false rejection error. By contrast, if a forgery is accepted, it
is called a false acceptance error. The experimental protocol
for False Acceptance Rate (FAR) and False Rejection Rate
(FRR) computation used in this work is explained in Section
V.
Furthermore, in this paper, we propose a method inspired
by online approaches, through the compact codification of
segments of skeletonized offline signatures, as explained in
Section IV. These skeletonized segments are the basic aspect
of this work, and they are detailed in Section II. Moreover,
the segmentation used in this work induces a straightforward method similar to classical online verification strategies,
through DTW.
As for our experiments, we take into account that, in
biometrics, comparing performances measures using different
databases can be misleading, which we consider an important
issue. Therefore, to allow a direct comparison under the same
database and the same experimental protocol for all compared
methods, we use the well-known MCYT-100 online signature
database, where each online sample was converted to an offline
(bitmap) representation, as explained in Section V, where
comparative results are provided along with our experimental
setup. It is worth noting that human baseline performance is
presented for the same dataset, in the spirit of works such as
[5] and [11]. Finally, in Section VI, we conclude by discussing
our results and the usefulness of Mean Opinion Scores (MOS)
as a potential goal to automatic verification performances.
II. U NAMBIGUOUSLY
CONNECTED SKELETON SEGMENTS
Raw online signature signals are frequently represented by
two vectors of samples: a sequence of regularly sampled
horizontal positions, xON LI N E (n), and another sequence of
corresponding vertical positions, yON LI N E (n), where n stands
for sample counter through time. As compared to offline representation, signature verification through signals xON LI N E (n)
and yON LI N E (n) is significantly better.
Although we know that velocity information may not be
completely recovered from offline representations, we address
the offline signature verification problem by first recovering
horizontal and vertical signals, which may be regarded as
pseudo-versions of xON LI N E (n) and yON LI N E (n). This is
done through standard skeletonization, as described in [7].
Unlike true online representations, skeletons from each signature are sets of unordered points. For instance, as in Figure
1, most offline signature skeletons can be regarded as sets
of unordered pixels in a bitmap, even though some subsets
of these pixels form segments that are clearly created in an
unambiguous sequential hand gesture.
Comparisons between online signatures are straightforward,
because points (xON LI N E , yON LI N E ) are ordered through
time. Analogously, the comparison between two offline signatures may also be done through the comparison of sequences
of points (xOF F LI N E , yOF F LI N E ), representing black pixel
coordinates in skeletons. However, ambiguities concerning the
ordering of points turns this task into a combinatorial optimization problem whose computational cost may be prohibitive.
To significantly reduce this cost, both methods proposed
in this paper decompose offline signatures skeletons into
Unambiguously Connected Skeleton Segments (UCSS), as illustrated in Figure 1. To define UCSS, an offline signature
skeleton is regarded as connected graph G(V, E) where the
vertices V are the points of the skeleton and the edges E
are bidirectional connection between neighbouring points (8connected neighbourhood). We also consider that the degree
of a vertex V is the number of neighbouring vertices that V is
connected to. Therefore, each UCSS is a sequences of directly
connected vertices found between:
a) two vertices with degrees greater than 2 (internal segments),
b) an one-degree vertex and a vertex with degree greater
than 2 (extremities), or
c) two one-degree vertices (isolated lines).
3
Fig. 1. (a) Inside the circles one finds the points of the skeleton that delimits
UCSS. (b) Full skeleton as a set of extracted features(UCSS).
We assume that each segment, sm,n,k , is a portion of
signature where points are unambiguously ordered, apart from
a single ambiguity in the overall direction of the pen movement (i.e. one does not know in what end of the UCSS
the movement of the pen begins). Thus, the n-th signature
sample, n = 1, 2, . . . , N, from the m-th signer, m = 1, 2, . . . , M
is represented by a set {sm,n } = {sm,n,1, sm,n,2, . . . , sm,n,K },
where sm,n,k is the k-th UCSS, or a sequence of Lm,n,k pairs
of coordinate points, (xi, yi ), 1 < i ≤ Lm,n,k .
Moreover, to take into account the single ambiguity in
the overall direction of the pen movement, each UCSS is
represented twice: first with the sequence of pairs in a given
order, sm,n,k
=
[(x1, y1 ), (x2, y2 ), . . . , (x Lm, n, k , y Lm, n, k )],
∗
and
then
in
reversed
order,
sm,n,k
=
[(x Lm, n, k , y Lm, n, k ), (x Lm, n, k −1, y Lm, n, k −1 ), . . . , (x1, y1 )].
III. BASELINE M ETHOD
Two methods for automatic offline signature verification are
proposed in this paper. The first method is considered as a
baseline, for it is a straightforward application of Dynamic
Time Warping to compute distances between UCSS. In this
method, the standard DTW method under Itakura’s restrictions
[9] is applied to systematically compare every segment sm,n,k
∗
) to every segments of a given bag of
(and its reverse, sm,n,k
segments, extracted from reference signatures. Consider, for
instance, a test signature, {sm,t est } with Kt est UCSS, and a
bag of segments, {B}, with KB UCSS (i.e. all segments, from
all references, are merged into the single set {B}). Then, each
UCSS in the test signature is compared to all KB UCSS from
{B}, and the minimum distance is taken. In other words, each
UCSS in a test signature is associated to the single UCSS in
{B} which yields the minimum DTW distance. More precisely,
Cm,t est (k, {B}) =
1
Lm,t est,k
min DTW(sm,t est,k, sB, j ),
j
(1)
and the average distance between sets {sm,t est } and {B} is
given by
C̄m,t est ({B}) =
1
Kt est
KÕ
t est
Cm,t est (k; {B})
k=1
where DTW(sm,t est,k, sB, j ) stands for Dynamic Time Warping
distance under Itakura’s restriction between sm,t est,k and sB, j ,
or its reversed version, s∗B, j , depending on which one yields the
lowest distance. Moreover, k ∈ 1, 2, . . . , Kt est , j ∈ 1, 2, . . . , KB ,
and test is a pointer to a signature from the test set. Moreover,
Lm,t est,k is the length (number of points) of the k-th UCSS,
of the tested signature.
In this work, we randomly take 5 genuine
signatures from each individual as reference set,
denoted as {sm,r e f1 },{sm,r e f2 },..., {sm,r e f5 }. Therefore,
{B} is the union of all 5 references, namely:
Ð
Ð
Ð
{B} = {{sm,r e f1 } {sm,r e f2 } . . . {sm,r e f5 }}, and KB
is the resulting cardinality of {B}.
To provide a better score for tested signatures, we also
define 5 partial bags of segments, {B}r e fi , where segments
of the i-th reference signature are excluded from {B}. As a
result, we are able to compute an average distance between
each reference signature, {sm,r e fi }, and the corresponding
remaining bag of segments, {B}r e fi , as follows
C̄0 ({B}) = (1/5)
5
Õ
C̄m,r e fi ({B}r e fi ).
i=1
Finally, the total distance between a tested signature
{sm,t est } and a genuine set of references, summarised by the
bag of segments {B}, is defined as:
Jm,r e f =
C̄m,t est ({B})
,
C̄0 ({B})
(2)
where C̄0 ({B}) plays the role of a normalization score.
An important drawback of this baseline method is the
computation of more than KB × Kt est DTW distances in order
to obtain a single score/cost for each tested signature, where
KB is roughly 5 times the average number of segments (UCSS)
in a genuine signature. As a consequence, this method has a
high computational burden.
IV. P ROPOSED
METHOD WITH
UCSS
SUBSAMPLING
To significantly alleviate the computational load of the baseline method, we encode each UCSS as an eight-dimensional
(8D) vector that roughly represents the shape and the position
of the encode UCSS, plus a scalar corresponding to the UCSS
length (the length is given in terms of number of points in the
UCSS), as illustrated in Figure 2.
This UCSS encoding strategy is the main aspect of the
proposed method. We assume that almost all UCSS are short
enough to prevent strong warping, therefore, one may get rid
of the high DTW computation cost by replacing each UCSS
with 4 subsampled points. In other words, it is assumed that
UCSS comparisons through DTW are almost equivalent to
the much faster Euclidean distance computation between the
corresponding 8D vectors (UCSS length is not taken into
account when plain DTW is used). Indeed, a UCSS sm,n,k can
be regarded as a composition of two sampled signals, say xn
and yn , with n = 1, 2, . . . , LUCSS , and the proposed coding
scheme just takes 4 equally spaced subsamples of xn and
yn , thus yielding an 8-dimensional vector um,n,k . For practical
purposes, we then assume that:
DTW(sm,n,i , sm,n, j ) D2 (um,n,i , um,n, j )
≈
,
Lm,n,i
4
4
(ii) The oversampled set of points are numerically rounded
to integer values, and
(iii) they are also dilated until segments are approximately
four pixels wide.
These three steps are enough to convert the entire online data
into offline signatures. Indeed, the resulting images are used
in the Mean Opinion Score experiment detailed in Subsection
V-A. As for the other experiments, another sequence of steps
are still taken to convert offline signatures into pseudo-online
ones, namely:
(i) The standard skeletonization method described in [7] is
applied to each dilated signature image.
(ii) Resulting skeletons, or sets of 2D points from each
signature sample are centered at the origin, whereas their
variance in both vertical and horizontal directions are
scaled to one.
Fig. 2. Proposed method: illustration of 8D vectors and lengths of two
segments (segments 10 and 17). Each segment (UCSS) is coded with 8+1 = 9
scalars. The last scalar represents the segment length.
where D2 (um,n,i , um,n, j ) is the accumulated squared Euclidean
distance between the 4 corresponding points in the two
compared UCSS. We highlight that, as for the definition of
function DTW, in Equation 1, D2 is also defined as the
minimum of two distances, where both um,n, j and its reversed
version, u∗m,n, j are considered. For the non-reversed um,n, j
8
Í
only, we obtain D2 (um,n,i , um,n, j ) =
(um,n,i (k) − um,n, j (k))2 .
k=1
Therefore, the comparison between two signatures is significantly simplified through the use of the following distance, as
compared to that in Equation 1:
Cm,t est (k, {U}) =
1
min D2 (um,t est,k, uU, j ),
4 j
(3)
where {U} is the set of all 4-points segments from the
reference signatures. Analogously, every subset {B}r e fi is
replaced with {U}r e fi , and the final score/cost for a given
signature can be computed as in Equation 2.
V. E XPERIMENTAL R ESULTS
In this work, we use the well-known MCYT-100 online signatures database, for which error rates (for online verification
task) can be abundantly found in literature [6], [17], [8]. This
database is a subcorpus of the MCYT database [14], acquired
from M = 100 different writers. Each writer provided 25
genuine signatures, whereas 5 different volunteers provided
25 skilled forgeries per signature (N = 50). All signatures
were acquired through a WACOM Intuos A6 USB Tablet at
constant sampling rate of 100 Hz.
To obtain the offline signatures we need for our experiments,
we first convert each MCYT online sample into an image
where only horizontal (xON LI N E ) and vertical (yON LI N E ) pen
tip positions through time are considered, as follows:
(i) Points in each online signature are interpolated using
splines to allow for oversampling of the otherwise sparse
representation (due to the relatively low sampling rate of
100 samples per second).
Fig. 3. (a) Online signature. (b) Oversampled online signature. (c) Dilated
image – approximately four pixels wide lines. (d) Skeleton of the signature.
From the resulting pseudo-online versions of the signatures,
represented by a set of coordinate pairs, UCSS are extracted. It
is noteworthy that the order in which these pairs are presented
no longer stands for a discrete time counter, as in true online
representation, for it now represents a mere skeleton point
counter, whose correspondence to time ordering is unknown.
In all experiments, to simulate an actual biometric system,
we randomly choose 5 genuine signatures from the database
to form the enrolment card. Afterwards, the other signatures of
the same user are randomly sampled and compared with those
into the enrolment card. Given a decision threshold, the proportion of true signatures whose costs are above this threshold
(thus wrongly rejected as not genuine) is the estimated FRR,
whereas the proportion of false signatures whose distances are
below the threshold is the estimated FAR.
We also compared the computational burden of the two
proposed methods to process and compare one enrolment
card, with 5 genuine samples per signature, and to compute
scores for all 20 + 25 = 45 remaining signature samples per
reference. The processing time for the baseline method was
thousands times greater than that for the proposed method with
UCSS subsampling. Surprisingly, the lighter method yielded
a significantly better performance, as presented in Figure 4.
An interesting by-product of both methods is that after two
UCSS are associated (according either to Equation 1 or to
Equation 3), instead of comparing their shapes, one may just
5
compare their lengths, thus yielding a new score corresponding
to the absolute difference of associated lengths. This can be
regarded as a third method, here referred to as the length based
one, in Figure 4 and Table 1.
TABLE I
R ESULTS FOR EACH METHOD AFTER 50 INDEPENDENT RUNS . E ACH RUN
CORRESPONDS TO A RANDOM PARTITION OF 5 REFERENCE AND 10 TEST
SIGNATURES (5 GENUINE ONES ).
Method
I-Baseline
II- Method by [18]
III-Length based
IV- Shape based (proposed)
V-Fusion (mean) of scores
from III and IV
EER
32.2 %
36.7 %
25.1 %
19.4 %
std. dev.
1.4 %
0.6 %
1.2 %
1.3 %
18.7 %
1.0 %
A. Mean Opinion Score Extraction Protocol
Fig. 4. ROC curves for all tested methods. In this set of experiments, we use 5
randomly chosen reference signatures to simulate enrolment, and a pool of 20
remaining genuine signatures, along with 25 false ones to test the simulated
system. A black dot also indicates the MOS performance, as explained in
Subsection V-A, for a subjective decision threshold that cannot be handled to
yield a ROC.
To provide a wider comparison scenario for the proposed
methods based on UCSS, we also included in the comparison
experiments the method published by [18], which extracts projections of bitmaps corresponding to signatures, and compare
them through a modified DTW, where so called stability measures are included to improve performances. We reproduced
the best implementation of their method (as explained in their
paper), and we applied it to the the same offline signatures
we used to test our methods, under the same experimental
protocol. We highlight that although the method by [18] was
implemented and revised to strictly follow their instructions,
we only apply it to the database of pseudo-offline signatures,
as illustrated in Figure 3. Moreover, instead of 10 reference
genuine signatures, as in [18], we use 5 ones to assure the
same protocol to all compared methods.
In a second round of experiments, we repeated 50 times,
for each method, the following protocol: 5 genuine signatures
were randomly chosen as references (enrolment), whereas a
set of 10 (5 from the remaining genuine signatures, plus 5
randomly chosen false ones) were used to test the simulated
system. For each independent trial, we adjusted the threshold
decision to obtain the operational point where FAR equals
FRR, the Equal Error Rate (EER). Table 1 presents all average
results per method, in terms of EER and its standard deviation
over the 50 independent trials.
In both sets of experiments, scores of the methods III and
IV are also fused through simple arithmetic mean, yielding an
improved performance, as shown in Figure 4 and Table 1.
To quantify the human performance for the same task, we
also prepared 239 two-page cards, each one corresponding to a
genuine source of signatures – a genuine signer – in the MCYT
database. These cards contain five genuine signatures on its
left side, and ten signatures randomly chosen in its right side.
Only 5 out of the 10 signatures on the right side are genuines.
An example of these cards can be seen in Figure 5. These
cards were presented to different students and lecturers in
our university (willing volunteers), and these volunteers were
carefully instructed to study the genuine signatures presented
on the left part of the page, and then to label all signatures on
the right part, by writing in the boxes next to each signature
a T, for true (genuine), or a F, for false. No further a priori
was provided. We highlight that the volunteers did not know
the proportions of true and false signatures in each card.
Fig. 5. (a) Reference panel with five randomly chosen genuine signatures (b)
Panel with five randomly chosen forgeries and five randomly chosen genuine
signatures.
A total of 239 cards were filled by 103 volunteers, and
after comparing all provided labels to the true hidden labels,
the estimated Mean Opinion Score (MOS) obtained was consolidated as:
•
•
25.6% of False Rejection Rate and
11.0% of False Acceptance Rate.
The dot in Figure 4 allows the visual comparison between
the rates FAR and FRR from MOS (to which the decision
threshold cannot be known or handled) and the ROC curves
6
from the automatic methods, for a range of possible decision
thresholds.
VI. C ONCLUSION
This brief work was inspired by the evident superiority of
online signature verification methods, as compared to offline
ones. Therefore all methods proposed here are based on skeletonization, possibly the most straightforward method to obtain pseudo-online signature representations from images. The
baseline method cross-compares all unambiguously connected
sequences of points from skeletons, where the conception of a
Unambiguously Connected Skeleton Segment (UCSS - whose
formal definition was given in Section II) plays a pivotal
role. By assuming that UCSS shapes and position are relevant
information for biometric verification, we should expect that
the systematic cross-comparison of UCSS, through DTW,
would yield the best performance, at a high computational
cost.
However, it was noticed instead that the alternative method
initially proposed to alleviate the high computational burden of
the baseline method – by considering only 4 points per UCSS
– yielded a significantly better performance, as compared to
the baseline method. Moreover, even the very simple byproduct method based on the comparison between UCSS
lengths performed better than the baseline method. Indeed, the
baseline method only outperformed the method by [18], which
uses an improved DTW, but relies upon marginal projections
of signatures instead of segments like the UCSS proposed in
this work.
From these results, we conjecture that, in average, UCSS is
a good segmentation option, but UCSS shape details are not
relevant information carriers for the biometric verification purpose. Indeed, given the comparative performances, we should
even conclude that UCSS shape details are disturbing noises,
for biometric verification task. This point is an interesting
matter for future works.
Regarding the by-product method based on UCSS length, it
is noteworthy that the use of either DTW or Euclidean distance
to match UCSS is a necessary step. In other words, behind the
apparent simplicity of this method, one should be aware that
the matching of UCSS is a not so simple step of it, which also
rises interesting questions. For instance, the superiority of the
joint approach – where UCSS shape and length are combined
– may have a connection to the lost signal of pen tip velocity,
which in turn is the main signal for biometric verification in
[3]. Indeed, UCSS shape (straightness) and length are expected
to be somehow dependent on pen tip velocity, either through
two-thirds power law, or through isochrony [19], and this
dependency is also an attractive subject for further works.
For now, in this letter, we just conjecture that the fusion of
length and shape based scores is somehow related to inferred
velocity signal, given a signature image, which may explain
its relatively good performance.
Clearly, even the best performance presented here is far
from performances typically found in literature for online
verification, with the same database in its original (online)
form. By contrast, in this work we assume that, for the offline
verification task, the best possible performances are not far
from what a crowd of willing and attentive humans can do.
Therefore, we choose to use a MOS-based rate as basis for
comparison, and it indicates that the best performances we
obtained are indeed close to that of a crowd of humans.
Nonetheless, further experiments using actual offline signature
databases and other competing methods are intended to be
done in the sequel of this work.
ACKNOWLEDGMENT
This work was supported by grant from the Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq)
to J.M.
R EFERENCES
[1] Y. M. Al-Omari, S. N. H. S. Abdullah, and K. Omar, “State-of-the-art in
offline signature verification system,” in Pattern Analysis and Intelligent
Robotics (ICPAIR), 2011 International Conference on, vol. 1. IEEE,
2011, pp. 59–64.
[2] L. Batista, D. Rivard, R. Sabourin, E. Granger, and P. Maupin, “State of
the art in off-line signature verification,” Pattern Recognition Technologies
and Applications: Recent Advances, pp. 39–62, 2008.
[3] J. Canuto, B. Dorizzi, J. Montalvão, and L. Matos, “On the infinite
clipping of handwritten signatures,” Pattern Recogn. Lett., vol. 79, no. C,
pp. 38–43, Aug. 2016.
[4] J. Coetzer, B. M. Herbst, and J. A. du Preez, “Offline signature verification using the discrete radon transform and a hidden markov model,”
EURASIP Journal on Advances in Signal Processing, vol. 2004, no. 4,
pp. 1–13, 2004.
[5] J. Coetzer, B. Herbst, and J. Du Preez, “Off-line signature verification:
a comparison between human and machine performance,” in Tenth International Workshop on Frontiers in Handwriting Recognition. Suvisoft,
2006.
[6] M. A. Ferrer, J. F. Vargas, A. Morales, and A. Ordóñez, “Robustness
of offline signature verification based on gray level features,” IEEE
Transactions on Information Forensics and Security, vol. 7, no. 3, pp.
966–977, 2012.
[7] R. C. Gonzalez and R. E. Woods, “Image processing,” Digital image
processing, vol. 2, 2007.
[8] L. G. Hafemann, R. Sabourin, and L. S. Oliveira, “Offline handwritten
signature verification-literature review,” arXiv preprint arXiv:1507.07909,
2015.
[9] F. Itakura, “Minimum prediction residual principle applied to speech
recognition,” IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 23, no. 1, pp. 67–72, 1975.
[10] S. Lee and J. C. Pan, “Offline tracing and representation of signatures,”
Systems, Man and Cybernetics, IEEE Transactions on, vol. 22, no. 4, pp.
755–771, 1992.
[11] D. Morocho, A. Morales, J. Fierrez, and R. Tolosana, “Signature recognition: establishing human baseline performance via crowdsourcing,” in
2016 4th International Conference on Biometrics and Forensics (IWBF).
IEEE, 2016, pp. 1–6.
[12] E. Nel, J. A. du Preez, and B. M. Herbst, “Estimating the pen trajectories
of multi-path static scripts using hidden markov models,” in Document
Analysis and Recognition, 2005. Proceedings. Eighth International Conference on. IEEE, 2005, pp. 41–45.
[13] V. Nguyen, M. Blumenstein, and G. Leedham, “Global features for
the off-line signature verification problem,” in 2009 10th International
Conference on Document Analysis and Recognition. IEEE, 2009, pp.
1300–1304.
[14] J. Ortega-Garcia, J. Fierrez-Aguilar, D. Simon, J. Gonzalez, M. FaundezZanuy, V. Espinosa, A. Satue, I. Hernaez, J.-J. Igarza, C. Vivaracho et al.,
“Mcyt baseline corpus: a bimodal biometric database,” IEE ProceedingsVision, Image and Signal Processing, vol. 150, no. 6, pp. 395–401, 2003.
[15] R. Plamondon and S. N. Srihari, “Online and off-line handwriting
recognition: a comprehensive survey,” Pattern Analysis and Machine
Intelligence, IEEE Transactions on, vol. 22, no. 1, pp. 63–84, 2000.
[16] Y. Qiao, M. Nishiara, and M. Yasuhara, “A framework toward restoration
of writing order from single-stroked handwriting image,” IEEE transactions on pattern analysis and machine intelligence, vol. 28, no. 11, pp.
1724–1737, 2006.
7
[17] A. S. Shah, M. Khan, F. Subhan, M. Fayaz, and A. Shah, “An offline signature verification technique using pixels intensity levels,” International
Journal of Signal Processing, Image Processing and Pattern Recognition,
vol. 9, no. 8, pp. 205–222, 2016.
[18] A. P. Shanker and A. Rajagopalan, “Off-line signature verification using
dtw,” Pattern recognition letters, vol. 28, no. 12, pp. 1407–1414, 2007.
[19] P. Viviani and T. Flash, “Minimum-jerk, two-thirds power law, and
isochrony: converging approaches to movement planning.,” Journal of
Experimental Psychology: Human Perception and Performance, vol. 21,
no. 1, pp. 32–53, 1995.
| 1cs.CV
|
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
1
Difficulty Adjustable and Scalable Constrained
Multi-objective Test Problem Toolkit
arXiv:1612.07603v2 [cs.NE] 16 Sep 2017
Zhun Fan, Senior Member, IEEE, Wenji Li, Xinye Cai, Hui Li, Caimin Wei, Qingfu Zhang, Fellow, IEEE,
Kalyanmoy Deb, Fellow, IEEE, and Erik D. Goodman
Abstract—Multi-objective evolutionary algorithms (MOEAs)
have achieved great progress in recent decades, but most of them
are designed to solve unconstrained multi-objective optimization
problems. In fact, many real-world multi-objective problems
usually contain a number of constraints. To promote the research
of constrained multi-objective optimization, we first propose three
primary types of difficulty, which reflect the challenges in the
real-world optimization problems, to characterize the constraint
functions in CMOPs, including feasibility-hardness, convergencehardness and diversity-hardness. We then develop a general
toolkit to construct difficulty adjustable and scalable constrained
multi-objective optimization problems (CMOPs) with three types
of parameterized constraint functions according to the proposed
three primary types of difficulty. In fact, combination of the
three primary constraint functions with different parameters can
lead to construct a large variety of CMOPs, whose difficulty
can be uniquely defined by a triplet with each of its parameter
specifying the level of each primary difficulty type respectively.
Furthermore, the number of objectives in this toolkit are able to
scale to more than two. Based on this toolkit, we suggest nine
difficulty adjustable and scalable CMOPs named DAS-CMOP1-9.
To evaluate the proposed test problems, two popular CMOEAs
- MOEA/D-CDP and NSGA-II-CDP are adopted to test their
performances on DAS-CMOP1-9 with different difficulty triplets.
The experiment results demonstrate that none of them can solve
these problems efficiently, which stimulate us to develop new
constrained MOEAs to solve the suggested DAS-CMOPs.
Index Terms—Constrained problems, multi-objective optimization, test problems, controlled difficulties.
I. I NTRODUCTION
RACTICAL optimization problems usually involve simultaneous optimization of multiple and conflicting objectives with many constraints. Without loss of generality,
constrained multi-objective optimization problems (CMOPs)
can be defined as follows:
P
minimize F(x) = ( f1 (x), . . . , fm (x))T
subject to
(1)
gi (x) ≥ 0, i = 1, . . . , q
h j (x) = 0, j = 1, . . . , p
x ∈ Rn
where F(x) = ( f1 (x), f2 (x), . . . , fm (x))T ∈ Rm is a mdimensional objective vector, gi (x) ≥ 0 defines i-th of q
inequality constraints, h j (x) = 0 defines j-th of q equality
constraints. If m is greater than three, we usually call it a
constrained many-objective optimization problem (CMaOP).
A solution x is said to be feasible if it meets gi (x) ≥ 0, i =
1, . . . , q and h j (x) = 0, j = 1, . . . , p at the same time. For two
feasible solutions x1 and x2 , solution x1 is said to dominate x2
if fi (x1 ) ≤ fi (x2 ) for each i ∈ {1, . . . , m} and f j (x1 ) < f j (x2 )
for at least one j ∈ {1, . . . , m}, denoted as x1 x2 . For a
feasible solution x∗ ∈ Rn , if there is no other feasible solution
x ∈ Rn dominating x∗ , x∗ is said to be a feasible Pareto-optimal
solution. The set of all the feasible Pareto-optimal solutions
is called Pareto Set (PS). Mapping the PS into the objective
space results in a set of objective vectors, denoted as the Pareto
Front (PF), where PF = {F(x) ∈ Rm |x ∈ PS}.
For CMOPs, more than one objective need to be optimized simultaneously subject to constraints. Generally speaking, CMOPs are much more difficult to solve than their
unconstrained counterparts - unconstrained multi-objective
optimization problems (MOPs). Constrained multi-objective
evolutionary algorithms (CMOEAs) are particularly designed
to solve CMOPs, with the capability of balancing the search
between the feasible and infeasible regions in the search space
[1]. In fact, two basic issues need to be considered carefully
when designing a CMOEA. One is to balance the feasible
solutions and the infeasible solutions, the other is to balance
the convergence and diversity of a CMOEA.
To address the former issue, constraint handling mechanisms need to be carefully designed by researchers. The
existing constraint handling methods can be broadly classified
into five different types, including feasibility maintenance,
use of penalty functions, separation of constraint violation
and objective values, multi-objective constraint handling and
hybrid methods [2].
The feasibility maintenance methods usually adopt special
encoding and decoding techniques to guarantee that a newly
generated solution is feasible. The penalty function-based
method is one of the most popular approaches. The overall
constraints violation is added to each objective with a predefined penalty factor, which indicates a preference between
the constraints and the objectives. The penalty function-based
method includes static penalties [3], dynamic penalties [4],
death penalty functions [3], co-evolutionary penalty functions [5], adaptive penalty functions [6]–[8] and self-adaptive
penalty functions [9], [10] etc. In the methods using separation
of constraint violation and objective values, the constraint
functions and the objective functions are treated separately.
Variants of this type include stochastic ranking (SR) [11], constraint dominance principle (CDP) [12], epsilon-constrained
methods [13], [14]. In the multi-objective constraint handling
method, the constraint functions are transformed to one extra
objective function. Representative methods of this type include infeasibility driven evolutionary algorithm (IDEA) [15],
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
COMOGA [16] and Cai and Wang’s Method (CW) [17],
etc. The hybrid methods of constraint handling usually adopt
several constraint-handling methods. Representative methods
include adaptive trade-off model (ATM) [18] and ensemble of
constraint handling methods (ECHM) [19].
To address the second issue, the selection methods need to
be designed to balance the performance of convergence and
diversity in MOEAs. At present, MOEAs can be generally
classified into three categories based on the selection strategies. They are Pareto-dominance (e.g., NSGA-II [20], PAES-II
[21] and SPEA-II [22]), decomposition-based (e.g., MOEA/D
[23], MOEA/D-DE [24], MOEA/D-M2M [25] and EAGMOEA/D [26]) and indicator based methods (e.g., IBEA [27],
R2-IBEA [28], SMS-EMOA [29] and HypE [30]). In the group
of Pareto-dominance based methods, such as NSGA-II [20],
the set of the first non-dominated level solutions is selected
to improve the performance of convergence, and the crowding
distance is adopted to maintain the performance of diversity. In
the decomposition-based methods, the performance of convergence is maintained by minimizing the aggregation functions
and the performance of diversity is obtained by setting the
weight vectors uniformly. In the indicator based methods, such
as HypE [30], the performance of convergence and diversity
is achieved by the using the hypervolume metric.
A CMOP includes objectives and constraints. A number of
features have already been identified to define the difficulty of
objectives, which include:
1) Geometry of PF (linear, convex, concave, degenerate,
disconnected and mixed of them)
2) Search space (biased, or unbiased)
3) Unimodal or multi-modal objectives
4) Dimensionality of variable space and objective space
The first one is the geometry of PF. The geometry of
PF of a MOP can be linear, convex, concave, degenerate,
disconnected and mixed of the them. Representative MOPs
reflecting this type of difficulty include ZDT [31], F1-9 [32]
and DTLZ [33]. The second one is the biased or unbiased
search space. Representative MOPs in this category include
MOP1-7 [34] and IMB1-14 [35]. The third one is the modality
of objectives. The objectives of a MOP can be either uni-modal
(DTLZ1 [33]) or multi-modal (F8 [32]). Objectives with multimodal have multiple local optimal solutions which increase
the likelihood of an algorithm being trapped in local optima.
The high dimensionality of variable space and objective space
are also critical features to define the difficulty of objectives.
LSMOP1-9 [36] have high dimensionality in the variable
space. DTLZ [33] and WFG [37] have high dimensionality
in the objective space.
On the other hand, constraint functions in general greatly
increase the difficulty of solving CMOPs. However, as far as
we know, only several test suites (CTP [38], CF [39]) are
designed for CMOPs.
CTP test problems [38] are have the capability of adjusting
the difficulty of the constraint functions. They offer two types
of difficulties: the difficulty near the Pareto front and the difficulty in the entire search space. The test problem CTP1 gives
the difficulty near the PF, because the constraint functions
of CTP1 make the search region close to the Pareto front
2
infeasible. Test problems CTP2-CTP8 provide an optimizer
the difficulty in the entire search space.
CF test problems [39] are also commonly used benchmarks,
which provide two types of difficulties. For CF1-CF3 and CF8CF10, their PFs are a part of their unconstrained PFs. The
rest of CF test problems CF4-CF7 have difficulties near their
PFs, and many constrained Pareto optimal points lie on some
boundaries of the constraints.
Even though CDP [38] and CF [39] offer the abovementioned advantages. They have some limitations:
•
•
•
•
The number of decision variables in the constraint functions can not be extended.
The difficulty level of each type is not adjustable.
No constraint functions with low ratios of feasible regions
in the entire search space are suggested.
The number of objectives is not scalable.
Some other used two-objective test problems, include BNH
[40], TNK [41], SRN [42] and OSY [43] problems, which
are not scalable to the number of objectives, and difficult to
identify types of difficulties.
In this paper, we propose a general framework to construct
difficulty adjustable and objective scalable CMOPs which can
overcome the limitations of existing CMOPs. CMOPs constructed by this toolkit can be classified into three major types,
which are feasibility-hard, convergence-hard and diversityhard CMOPs. Feasibility-hard CMOP is a type of problem
that presents difficulty for CMOEAs to find feasible solutions
in the search space. CMOPs with feasibility-hardness usually
have small portions of feasible regions in the entire search
space. In addition, CMOPs with convergence-hardness mainly
suggest difficulty for CMOEAs to approach the PFs efficiently
by setting many obstacles before the PFs, while CMOPs with
diversity-hardness mainly provide difficulty for CMOEAs to
distribute their solutions along the complete PFs. In our work,
the three types of difficulty are embedded into the CMOPs
through proper construction of constraint functions.
In summary, the contribution of this paper is as follows:
1) This paper defines three primary types of difficulty
for constraints in CMOPs. When designing new constraint handling mechanisms for a CMOEA, one has
to investigate the nature of constraints in a CMOP that
the CMOEA is aiming to address, including the types
and levels of difficulties embedded in the constraints.
Therefore, a proper definition on the types of difficulty
for constraints in CMOPs is necessary and desirable.
2) This paper also defines the level of difficulty, regarding
each type of difficulty for constraints in the constructed
CMOPs, which can be adjusted by users. A difficulty
level is uniquely defined by a triplet with each of its
parameter specifying the level of each primary difficulty
type respectively. Combination of the three primary
constraint types with different difficulty triplets can lead
to construction of a large variety of constraints for
CMOPs.
3) Based on the proposed three primary types of difficulty
for constraints, nine difficulty adjustable CMOPs named
DAS-CMOP1-9 are constructed.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
The remainder of this paper is organized as follows. Section
II discusses the effects of constraints on PFs. Section III
introduces the types and levels of difficulties provided by
constraints in CMOPs. Section IV explains the proposed
toolkit of construction methods for generating constraints in
CMOPs with different types and levels of difficulty. Section
V realizes the scalability to the number of objectives in
CMOPs using the proposed toolkit. Section VI generates a
set of difficulty adjustable CMOPs using the proposed toolkit.
In Section VII, the performance of two CMOEAs on DASCMOP1-9 with different difficulty levels are compared by
experimental studies, and Section VIII concludes the paper.
II. E FFECTS OF CONSTRAINTS ON PF S
Constraints define the infeasible regions in the search space,
leading to different types and levels of difficulty for the
resulting CMOPs. Some major effects of the constraints on
PFs in CMOPs include the following [44]:
1) Infeasible regions make the original unconstrained PF
partially feasible. This can be further divided into two
situations. In the first situation, the PF of the constrained
problem consists of a part of its unconstrained PF and
a set of solutions on some boundaries of constraints,
as illustrated by Fig. 1(a). In the second situation, the
PF of the constrained problem is only a part of its
unconstrained PF, as illustrated by Fig. 1(b).
2) Infeasible regions block the way towards the PF, as
illustrated by Fig. 1(c).
3) The complete original PF is covered by infeasible regions and becomes no more feasible. Every constrained
Pareto optimal point lies on some constraint boundaries,
as illustrated by Fig. 1(d).
4) Constraints may reduce the dimensionality of the PF
with one example illustrated by Fig. 1(e). In general, although the problem is M−dimensional, constraints make
the constrained PF K−dimensional (where K < M). In
the particular case of Fig. 1(e), M = 2, K = 1.
III. D IFFICULTY TYPES AND LEVELS OF CMOP S
Three primary difficulty types have been identified, including convergence-hardness, diversity-hardness, and feasibilityhardness. A difficulty level for each primary difficulty type
can be defined as a parameter ranging from 0 to 1. Three
difficulty levels, corresponding to three primary difficulty
types respectively, form a triplet that depicts the nature of
the difficulty of a CMOP.
A. Difficulty 1: Diversity-hardness
Generally, the PFs of CMOPs with diversity-hardness have
many discrete segments, or some parts more difficult to be
achieved than the other parts by imposing large infeasible
regions near them. As a result, achieving the complete PF
is difficult for CMOPs.
3
B. Difficulty 2: Feasibility-hardness
For the feasibility-hard CMOPs, the ratios of feasible regions in the search space are usually very low. It is difficult to
generate a feasible solution for a CMOEA on the feasibilityhard CMOPs. Often in the initial stage of a CMOEA, most
solutions in the population are infeasible.
C. Difficulty 3: Convergence-hardness
CMOPs with convergence-hardness hinder the convergence
of CMOEAs towards the PFs. Usually, CMOEAs encounter
more difficulty to approach the PFs. Because infeasible regions
block the way of CMOEAs converging to the PFs. In other
words, the generational distance (GD) metric [45], which
indicates the performance of convergence, is difficult to be
minimized in the evolutionary process.
D. Difficulty level of each primary difficulty type
A difficulty level of each primary difficulty type can be
defined by a parameter in the parameterized constraint function
corresponding to the primary difficulty type. Each parameter
is normalized from 0 to 1. Three parameters, corresponding
to the difficulty level of the three primary difficulty types
respectively, form a triplet (η, ζ, γ) that exactly defines the
nature of difficulty of a CMOP constructed by the three
parameterized constraint functions.
If each element of the triplet can only take value of either 0
or 1, then a simple combination of the three primary difficulty
types will give rise to seven basic different difficulty types.
This is analogous to a simple combination of three primary
colors gives rise to seven basic colors. But if we allow the three
parameters to take any value between 0 and 1, then we can
literally get countless difficulty nature (analogous to countless
colors in the color space). A difficulty nature here is then
precisely depicted by a triplet (η, ζ, γ).
IV. CONSTRUCTION TOOLKIT
As we know, constructing a CMOP is composed of constructing two major parts - objective functions and constraint
functions. Li, et al. [46] suggested a general framework for
constructing objective functions. It is stated as follows:
fi (x) = αi (x1:m−1 ) + βi (x1:m−1, xm:n )
)T ,
(2)
where x1:m−1 = (x1, . . . , xm−1 xm:n = (xm, . . . , xn are two
sub-vectors of x = (x1, . . . , xn )T . The function αi (x1:m−1 ) is
called the shape function, and βi (x1:m−1, xm:n ) is called the
nonnegative distance function. The objective function fi (x), i =
1, . . . , m is the sum of the shape function αi (x1:m−1 ) and the
nonnegative distance function βi (x1:m−1, xm:n ). We adopt Li,
et al.’s method [46] in this work.
In terms of constructing the constraint functions, three
different types of constraint functions are suggested in this
paper, corresponding to the proposed three primary types
of difficulty of CMOPs. More specifically, Type-I constraint
functions provide the difficulty of diversity-hardness, TypeII constraint functions introduce the difficulty of feasibilityhardness, and Type-III constraint functions generate the difficulty of convergence-hardness. The detailed definition of
)T
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
1.5
1.5
infeasible area
PF without constraints
PF with constraints
infeasible area
PF with constraints
PF without constraints
0.8
1
f2
1
f2
f2
1
infeasible area
PF with constraints
PF without constraints
feasbile area
PF with constraints
PF without constraints
0.6
f2
1
1
1.5
infeasbile area
PF with constraints
PF without constraints
f2
1.5
4
0.4
0.5
0.5
0.5
0.5
0.2
0
0
0.5
1
1.5
0
0
0.5
f1
(a)
1
1.5
0
0
0.5
f1
1
1.5
0
0
0.5
f1
(b)
(c)
1
1.5
0
0
0.5
f1
f1
(d)
(e)
Fig. 1. Illustration of the effects of constraints on PFs. (a) Infeasible regions makes the original unconstrained PF partially feasible. Many constrained Pareto
optimal solutions lie on the constraint boundaries. (b) Infeasible regions makes the original unconstrained PF partially feasible. The constrained PF is a part
of its unconstrained PF. (c) Infeasible regions blocks the way of converging to the PF. The constrained PF is same to its unconstrained PF. (d) The complete
original PF is no more feasible. Every constrained Pareto optimal solution lies on the constraint boundaries. (e) Constraints reduce the dimensionality of the
PF. A two-objective optimization problem is transformed into a constrained single optimization problem.
TABLE I
BASIC DIFFICULTY TYPES OF THE CMOP S
Basic Difficulty Types
T1: Diversity-hardness
T2: Feasibility-hardness
T3: Convergence-hardness
T4: Diversity-hardness and feasibility-hardness
T5: Diversity-hardness and convergence-hardness
T6: Feasibility-hardness and convergence-hardness
T7: Diversity-hardness, feasibility-hardness and convergence-hardness
Comment
Distributing the feasible solutions in the complete PF is difficult.
Obtaining a feasible solution is difficult.
Approaching a Pareto optimal solution is difficult.
Obtaining a feasible solution and the complete PF is difficult.
Approaching a Pareto optimal solution and the complete PF is difficult.
Obtaining a feasible solution and approaching a Pareto optimal solution is difficult.
Obtaining a Pareto optimal solution and the complete PF is difficult.
A. Type-I Constraint Functions: Diversity-hardness
Fig. 2. The illustration of three primary difficulty types and their combination
resulting in seven basic difficulty types (as shown in Table I), using an analogy
of three primary colors and their combination towards seven basic colors.
Fig. 3. The illustration of combining three parameterized constraint functions
using a triplet composing of three parameters. The three primary constraint
functions correspond to the three primary difficulty types respectively.
the three types of constraint functions are given in detail as
follows:
Type-I constraint functions are defined to limit the boundary
of sub-vector x1:m−1 . More specifically, this type of constraint
functions divides the PF of a CMOP into a number of
disconnected segments, generating the difficulty of diversityhardness. Here, we use a parameter η ∈ [0, 1] to represent the
level of difficulty. η = 0 means the constraint functions impose
no effects on the CMOP, while η = 1 means the constraint
functions provide their maximum effects.
An example of CMOP with diversity-hardness is suggested
as follows:
minimize f1 (x) = x1 + g(x)
minimize f2 (x) = 1 − x12 + g(x)
Ín
(3)
g(x) = i=2
(xi − sin(0.5πx1 ))2
subject to c(x) = sin(aπx1 ) − b ≥ 0
xi ∈ [0, 1]
where a > 0, b ∈ [0, 1]. As an example a = 10, n = 2 are
set here. The parameter η indicating the level of difficulty is
set to η = (b + 1)/2. The number of disconnected segments in
the PF is controlled by a. Moreover, the value of b controls
the width of each segment. The width of segments reaches
its maximum when b = −1. When b increases, the width of
segments decreases, and the difficulty level increases, so does
the parameter of the difficulty level η. As a result, if η is
set to 0.55, the PF is shown in Fig. 4(a). If η = 0.75, the
PF is shown in Fig.4 (b). It can be observed that the width
of segments of the PF is reduced as η keeps increasing. If
η = 1.0, the width of segments shrinks to zero, which provides
the maximum level of difficulty to the CMOP. The PF of a
three-objective CMOP with Type-I constraint functions is also
shown in Fig. 4(d), with the difficult level η = 0.75. It can be
1
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
seen that Type-I constraint functions can be applied in more
than two-objective CMOPs, which means that a CMOP with
the scalability to the number of objectives can be constructed
using this type of constraints.
B. Type-II Constraint Functions: Feasibility-hardness
Type-II constraint functions are set to limit the reachable
boundary of the distance function of βi (x1:m−1, xm:n ), and
thereby control the ratio of feasible regions. As a result, TypeII constraint functions generate the difficulty of feasibilityhardness. Here, we use a parameter ζ to represent the level
of difficulty, which ranges from 0 to 1. ζ = 0 means the
constraints are the weakest, and ζ = 1.0 means the constraint
functions are the strongest.
For example, a CMOP with Type-II constraint functions can
be defined as follows:
minimize f1 (x) = x1 + g(x)
minimize f2 (x) = 1 − x12 + g(x)
Ín
g(x) = i=2
(xi − sin(0.5πx1 ))2
(4)
subject to c1 (x) = g(x) − a ≥ 0
c2 (x) = b − g(x) ≥ 0
n = 30, xi ∈ [0, 1]
1
, and a ≥ 0, b ≥ 0 and b ≥ a.
where ζ equals to exp(b−a)
The distance between the constrained PF and unconstrained
PF is controlled by a, and a = 0.5 in this example. The ratio
of feasible regions is controlled by b − a. If b − a = +∞,
ζ = 0, the feasible area reaches maximum as shown in Fig.
5(a). If b − a = 0.1, ζ = 0.905, the feasible area is decreased
as shown in Fig. 5(b). If b − a = 0, ζ = 1.0, the feasible area
in the objective space is very small. The PF of this problem
is shown in Fig. 5(c). Type-II constraints can be also applied
to CMOPs with three objectives as shown in Fig. 5(d).
C. Type-III Constraint Functions: Convergence-hardness
Type-III constraint functions limit the reachable boundary
of objectives. As a result, infeasible regions act like ’blocking’
hindrance for searching populations of CMOEAs to approach
the PF. As a result, Type-III constraint functions generate the
difficulty of convergence-hardness. Here, we use a parameter
γ to represent the level of difficulty, which ranges from 0 to 1.
γ = 0 means the constraints are the weakest, γ = 1 means the
constraints are the strongest, and the difficulty level increases
as γ increases.
For example, a CMOP with Type-III constraint functions
can be defined as follows:
min
f1 (x) = x1 + g(x)
min
f2 (x) = 1 − x12 + g(x)
Ín
where g(x) = i=2
(xi − sin(0.5πx1 ))2
s.t.
ck (x) = (( f1 − pk ) cos θ k − ( f2 − qk ) sin θ k )2 /ak2
(5)
+(( f1 − pk ) sin θ k + ( f2 − qk ) cos θ k )2 /b2k ≥ r
pk = [0, 1, 0, 1, 2, 0, 1, 2, 3]
q
k = [1.5, 0.5, 2.5, 1.5, 0.5, 3.5, 2.5, 1.5, 0.5]
ak2 = 0.4, b2k = 1.6, θ k = −0.25π
c = 20, n = 30, xi ∈ [0, 1], k = 1, . . . , 9
5
where the level of difficulty parameter γ is defined as γ = 2r.
If γ = 0.1, the PF is shown in Fig. 6(a). If γ = 0.5, the
infeasible regions are increased and shown in Fig. 6(b). If
γ = 0.75, the infeasible regions become bigger than those of
γ = 0.5 as shown in Fig. 6(c). The constraints of Type-III can
be also applied to CMOPs with three objectives as shown in
Fig. 6(d).
Type-III constraint functions can be expressed in a matrix
form, which can be defined as follows:
(F(x) − Hk )T Sk (F(x) − Hk ) ≥ r
(6)
where F(x) = ( f1 (x), . . . , fm (x))T . Hk is a translation vector.
Sk is a transformational matrix, which control the degree of
rotation and stretching of the vector (F(x) − Hk ). According
to the Type-III constraint functions in Eq. (5), Hk = (pk , qk )T ,
and Sk can be expressed as follows:
cos2 θk + sin2 θk
a2
b2
Sk = k sin 2θk k
b k2
k
− sina2θ
2
cos2 θk
b k2
2
+ sina2θk
k
k
It is worthwhile to point out that by using this approach we
can further extend the number of objectives to be more than
three, even though more sophisticated visualization approach
is needed to show the resulting CMOPs in the objective space.
To summarize, the three types of constraint functions discussed above correspond to the three primary difficulty types
of CMOPs respectively. In particular, Type-I constraint function corresponds to diversity-hardness, Type-II corresponds to
feasibility-hardness, and Type-III corresponds to convergencehardness. The level of each primary difficulty type can be
decided by a parameter. In this work, three parameters are
defined in a triplet (η, ζ, γ), which specifies the difficulty level
of a particular difficulty type. It is noteworthy to point out
that this approach of constructing toolkit for CMOPs can
also be scaled to generate CMOPs with more than three
objective functions. The scalability to the number of objectives
is discussed in more detail in Section V.
V. S CALABILITY TO THE NUMBER OF OBJECTIVES
Recently many-objective optimization attracts a lot of research interests, which makes the feature of scalability to
the number of objectives of CMOPs desirable. A general
framework to construct CMOPs with the scalability to the
number of objectives is given in Eq. (7).
In Eq. (7), we borrow the idea of WFG toolkit [37] to
construct objectives, which can be scaled to any number
of objectives. More specifically, the number of objectives is
controlled by a user-defined parameter m.
Three different types of constraint functions proposed in
Section V can be combined together with the scalable objectives to construct difficulty adjustable and scalable CMOPs
(DAS-CMOPs). More specifically, the first K constraint functions with Type-I are defined to limit the reachable boundary
of each decision variable in the shape functions (α1 (x1:m−1 ) to
αm (x1:m−1 )), which have the ability to control the difficulty
level of diversity-hardness by η. The (K + 1) to (K + P)
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
2
f2
f2
2
3
PF without constraint
Feasible area
PF with Type−I constraint
PF without constraint
Feasible area
PF with Type−I constraint
2
Feasible area
PF with Type−I constraint
2
1.5
1
f3
3
PF without constraint
Feasible area
PF with Type−I constraint
f2
3
6
0.5
1
1
1
0
2
2
0
0
1
2
0
0
3
1
f1
2
1
0
0
3
1
f1
(a) η = 0.55
2
3
0 0
f2
f1
(b) η = 0.75
1
(c) η = 1.0
f1
(d) m = 3, η = 0.75
Fig. 4. Illustrations on the influence of Type-I constraint functions. When the parameter of difficulty level η increases, the width of segments in the PF
decreases, and the difficulty level of a CMOP increases. Because the PF of a CMOP with Type-I constraint is disconnect and usually has many discrete
segments, obtaining the complete PF is difficult. Thus a CMOP with Type-I constraints is diversity-hard. (a) Two-objective CMOP with η = 0.55. (b)
Two-objective CMOP with η = 0.75. (c) Two-objective CMOP with η = 1.0. (d) Three-objective CMOP with η = 0.75.
2
2
PF without constraint
PF with Type−II constraint
Feasible area
1.5
PF without constraint
PF with Type−II constraint
Feasible area
1.5
2
PF without constraint
PF with constraint
Feasible area
1
1
0.5
f2
f2
f2
f3
1.5
2
PF without constraint
PF with Type−II constraint
Feasible area
1
0.5
1
0
0.5
−1
2
1
0
0
0.5
1
f1
1.5
0
0
2
(a) ζ = 0
0.5
1
f1
1.5
0
0
2
0.5
(b) ζ = 0.905
1
f1
1.5
2
(c) ζ = 1.0
1.5
1
0.5
0 0
f2
2
f1
(d) m = 3, ζ = 0.905
1
. Here, the default value of a is set
Fig. 5. Illustrations on the influence of Type-II constraint functions. The parameter of difficulty degree ζ = exp(b−a)
to 0.5. The ratio of feasible regions is controlled by ζ . When the parameter ζ increases, the portion of feasible regions decreases, and the difficulty level
of feasibility increases. (a) ζ = 0.0. (b) ζ = 0.905. (c) ζ = 1.0. (d) The Type-II constraint can be applied into three-objective optimization problems, and
ζ = 0.905.
4
4
PF with Type−III constraint
Infeasible area
3
4
PF with Type−III constraint
Infeasible area
3
PF with Type−III constraint
Infeasible area
PF with Type−III constraint
Infeasible area
1.5
3
2
f3
f2
f2
f2
1
2
2
0.5
1
1
1
0
0
0
0
0
0
0
0
1
2
f1
(a) γ = 0.1
3
4
1
2
f1
3
(b) γ = 0.5
4
0
0.5
1
2
f1
(c) γ = 0.75
3
4
0.5
1
1
1.5
f2
1.5
f1
(d) γ = 0.5, m = 3
Fig. 6. Illustrations on the influence of Type-III constraint functions. Infeasible regions block the way of converging to the PF. The gray parts of each figure
are infeasible regions. A parameter γ is adopted to represent the level of difficulty, which ranges from 0 to 1. γ = 0 means the constraints are the weakest, and
γ = 1 means the constraints are the strongest. When γ increases, the difficulty level of convergence-hardness of a CMOP increases. (a) γ = 0.1. (b) γ = 0.5.
(c) γ = 0.75. (d) The Type-III constraint can also be applied into three-objective optimization problems, here with γ = 0.5.
constraint functions belong to Type-II, which limit the reachable boundary of the distance functions (β1 (x1:m−1, xm:n ) to
β1 (x1:m−1, xm:n )). They have the ability to control the the difficulty level of feasibility-hardness by ζ. The last Q constraint
functions are set directly on each objective, and belong to
Type-III. They generate a number of infeasible regions, which
hinder the working population of a CMOEA approaching to
the PF. The difficulty level of convergence-hardness generated
by Type-III constraint functions is controlled by γ. The rest
of parameters in Eq. (7) are illustrated as follows.
Three parameters K, P and Q are used to control the
number of each type of constraint functions, respectively.
K ≤ m − 1, P ≤ m and Q ≥ 1. The total number of constraint
functions is controlled by (K +P+Q). n decides the dimensions
of decision variables, and n ≥ m. a decides the number of
disconnected segments in the PF. d indicates the distance
between the constrained PF and the unconstrained PF. The
difficulty level of a DAS-CMOP is controlled by a difficulty
triplet (η, ζ, γ), with each of its component ranging from 0 to
1. When each of parameter in the difficult triplet increases,
the difficulty level of a DAS-CMOP increases.
It is worth noting that the number of objectives of DASCMOPs can be easily scaled by tuning the parameter of m.
The difficulty level of DAS-CMOPs can be also easily adjusted
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
7
by assigning a difficulty triplet (η, ζ, γ) with three parameters
ranging from 0 to 1.
min
min
min
s.t
f1 (x) = α1 (x1:m−1 ) + β1 (x1:m−1, xm:n )
f2 (x) = α2 (x1:m−1 ) + β2 (x1:m−1, xm:n )
..
.
fm (x) = αm (x1:m−1 ) + βm (x1:m−1, xm:n )
ck (x) = sin(aπxk ) − b ≥ 0, if k is odd
ck (x) = cos(aπxk ) − b ≥ 0, if k is even
cK+p (x) = (e − β p )(β p − d) ≥ 0
cK+P+q (x) = (F(x) − Hq )T Sq (F(x) − Hq ) ≥ r
k = 1, . . . , K, and K ≤ m − 1
p = 1, . . . , P, and P ≤ m
q = 1, . . . , Q
η = (b + 1)/2
ζ = exp(d − e), d ≤ e
γ = 2r
(7)
VI. A SET OF DIFFICULTY ADJUSTABLE AND SCALABLE
CMOP S
In this section, as an example, a set of nine difficulty
adjustable and scalable CMOPs (DAS-CMOP1-9) is suggested
through the proposed toolkit.
As mentioned in Section IV, constructing a CMOP composes of constructing objective functions and constraint functions. According to Eq. (7), we suggest nine multi-objective
functions, including convex, concave and discrete PF shapes,
to construct CMOPs. A set of difficulty adjustable constraint
functions is generated by Eq. (7). Nine difficulty adjustable
and scalable CMOPs named DAS-CMOP1-9 are generated by
combining the suggested objective functions and the generated
constraint functions. The detailed definitions of DAS-CMOP19 are shown in Table II.
In Table II, DAS-CMOP1-3 have the same constraint functions. For DAS-CMOP4-6, they also have the same constraint
functions. The difference between DAS-CMOP1-3 and DASCMOP4-6 is that they have different distance functions. For
DAS-CMOP1-6, they have two objectives. The number of
objectives in Eq. (7) are able to scale to more than two. For
example, DAS-CMOP7-9 have three objectives. The constraint
functions of DAS-CMOP8 and DAS-CMOP9 are the same as
those of DAS-CMOP7.
It is worth noting that the value of difficulty triplet elements can be set by users. If we want to reduce/increase the
difficulty levels of DAS-CMOP1-9, we only need to adjust the
parameters of the triplet elements to smaller/larger values, and
generate a new set of test instances.
VII. EXPERIMENTAL STUDY
A. Experimental Settings
To test the performance of CMOEAs on the DAS-CMOPs,
two commonly used CMOEAs (i.e., MOEA/D-CDP and
NSGA-II-CDP) are tested on DAS-CMOP1-9 with sixteen
different difficulty triplets in the experiment. As descripted in
Section IV, three parameters are defined in a triplet (η, ζ, γ),
which specifies the difficulty level of a particular difficulty
type. More specifically, η represents the difficulty level of
diversity-hardness, ζ denotes the difficulty level of feasibilityhardness, and γ indicates the difficulty level of convergencehardness. The difficulty triplets for each DAS-CMOP are listed
in Table III.
The detailed parameters of the algorithms are summarized
as follows.
1) Setting for reproduction operators: The mutation probability Pm = 1/n (n is the number of decision variables).
For the polynomial mutation operator, the distribution
index is set to 20. For the simulated binary crossover
(SBX) operator, the distribution index is set to 20. The
rate of crossover CR = 0.9.
2) Population size: For DAS-CMOP1-6, N = 200, and for
DAS-CMOP7-9, N = 105.
3) Number of runs and stopping condition: Each algorithm
runs 30 times independently on each test problem with
sixteen different difficulty triplets. The maximum function evaluations is 100000 for DAS-CMOP1-6, 200000
for DAS-CMOP7-9.
4) Neighborhood size: T = 20 for DAS-CMOP1-6, T = 10
for DAS-CMOP7-9.
5) Probability use to select in the neighborhood: δ = 0.9.
6) The maximal number of solutions replaced by a child:
nr = 2.
B. Performance Metric
To measure the performance of MOEA/D-CDP and NSGAII-CDP on DAS-CMOP1-9 with different difficulty triplets,
the inverted generation distance (IGD) [47] is adopted. The
detailed definition of IGD is given as follows:
• Inverted Generational Distance (IGD):
The IGD metric simultaneously reflects the performance of
convergence and diversity, and it is defined as follows:
Í
IGD(P∗, A) =
y∗ ∈ P∗
d(y ∗, A)
| P∗ |
(8)
qÍ
m
2
∗
d(y ∗, A) = min{
i=1 (yi − yi ) }
y∈A
where P∗ is the ideal PF set, A is an approximate PF
set achieved by an algorithm. m represents the number of
objectives. It is worth noting that the smaller value of IGD
represents the better performance of both diversity and convergence.
C. Performance Comparisons on two-objective DAS-CMOPs
Table IV presents the statistic results of IGD values
for MOEA/D-CDP and NSGA-II-CDP on DAS-CMOP13. We can observe that for DAS-CMOP1 with difficulty triplets (0.0, 0.25, 0.0), (0.0, 0.5, 0.0) and (0.0, 0.75, 0.0),
NSGA-II-CDP is significantly better than MOEA/D-CDP,
which indicates that NSGA-II-CDP is more suitable for
solving DAS-CMOP1 with feasibility-hardness. For DASCMOP1 with difficulty triplets (0.0, 0.0, 0.25), (0.0, 0.0, 0.5)
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
8
TABLE II
DAS-CMOP S T EST SUITE : THE OBJECTIVE FUNCTIONS AND CONSTRAINT FUNCTIONS OF DAS-CMOP1-9.
Problem
Objectives
DAS-CMOP1
min
min
where
DAS-CMOP2
min
min
where
DAS-CMOP3
min
min
where
DAS-CMOP4
DAS-CMOP5
DAS-CMOP6
DAS-CMOP7
DAS-CMOP8
DAS-CMOP9
min
min
where
min
min
where
min
min
where
f1 (x) = x1 + g1 (x)
f2 (x) = 1 − x12 + g2 (x)
Í
g1 (x) = j ∈J1 (x j − sin(0.5π x1 ))2
Í
g2 (x) = j ∈J2 (x j − cos(0.5π x1 ))2
J1 = { j | j is odd and 2 ≤ j ≤ n}
J2 = { j | j is even and 2 ≤ j ≤ n}
n = 30, x ∈ [0, 1] n
f1 (x) = x1 + g1 (x)
√
f2 (x) = 1 − x1 + g2 (x)
Í
g1 (x) = j ∈J1 (x j − sin(0.5π x1 ))2
Í
g2 (x) = j ∈J2 (x j − cos(0.5π x1 ))2
J1 = { j | j is odd and 2 ≤ j ≤ n}
J2 = { j | j is even and 2 ≤ j ≤ n}
n = 30, x ∈ [0, 1] n
f1 (x) = x1 + g1 (x)
√
f2 (x) = 1 − x1 + 0.5 ∗ | sin(5π x1 ) | + g2 (x)
Í
g1 (x) = j ∈J1 (x j − sin(0.5π x1 ))2
Í
g2 (x) = j ∈J2 (x j − cos(0.5π x1 ))2
J1 = { j | j is odd and 2 ≤ j ≤ n}
J2 = { j | j is even and 2 ≤ j ≤ n}
n = 30, x ∈ [0, 1] n
f1 (x) = x1 + g(x)
f2 (x) = 1 − x12 + g(x)
Í
g(x) = (n − 1) + nj=2 (x j − 0.5)2 − cos(20π(x j − 0.5))
n = 30, x ∈ [0, 1] n
f1 (x) = x1 + g(x)
√
f2 (x) = 1 − x1 + g(x)
Í
g(x) = (n − 1) + nj=2 (x j − 0.5)2 − cos(20π(x j − 0.5))
n = 30, x ∈ [0, 1] n
f1 (x) = x1 + g(x)
√
f2 (x) = 1 − x1 + 0.5 ∗ | sin(5π x1 ) | + g(x)
Í
g(x) = (n − 1) + nj=2 (x j − 0.5)2 − cos(20π(x j − 0.5))
n = 30, x ∈ [0, 1] n
min
min
min
where
f1 (x) = x1 ∗ x2 + g(x)
f2 (x) = x2 ∗ (1 − x1 ) + g(x)
f3 (x) = 1 − x2 + g(x)
Í
g(x) = (n − 2) + nj=3 (x j − 0.5)2 − cos(20π(x j − 0.5))
n = 30, x ∈ [0, 1] n
min
min
min
where
min
min
min
where
f1 (x) = cos(0.5π x1 ) ∗ cos(0.5π x2 ) + g(x)
f2 (x) = cos(0.5π x1 ) ∗ sin(0.5π x2 ) + g(x)
f3 (x) = sin(0.5πx1 ) + g(x)
Í
g(x) = (n − 2) + nj=3 (x j − 0.5)2 − cos(20π(x j − 0.5))
n = 30, x ∈ [0, 1] n
f1 (x) = cos(0.5π x1 ) ∗ cos(0.5π x2 ) + g(x)
f2 (x) = cos(0.5π x1 ) ∗ sin(0.5π x2 ) + g(x)
f3 (x) = sin(0.5πx1 ) + g(x)
2
Í
0.25 j
g(x) = nj=3 (x j − cos( n π(x1 + x2 )))
n = 30, x ∈ [0, 1] n
Constraints
c1 (x) = sin(aπ x1 ) − b ≥ 0
c2 (x) = (e − g1 )(g1 − d) ≥ 0
c3 (x) = (e − g2 )(g2 − d) ≥ 0
ck+3 (x) = (( f1 − pk ) cos θk − ( f2 − qk ) sin θk )2 /a2
k
+(( f1 − pk ) sin θk + ( f2 − qk ) cos θk )2 /b k2 ≥ r
pk = [0, 1, 0, 1, 2, 0, 1, 2, 3], a2k = 0.3, b k2 = 1.2, θk = −0.25π
qk = [1.5, 0.5, 2.5, 1.5, 0.5, 3.5, 2.5, 1.5, 0.5]
a = 20, d = 0.5, η = (b + 1)/2, ζ = exp(d − e), γ = 2r
It is the same as that of DAS-CMOP1
It is the same as that of DAS-CMOP1
c1 (x) = sin(aπ x1 ) − b ≥ 0
c2 (x) = (e − g(x))(g(x) − d) ≥ 0
ck+2 (x) = (( f1 − pk ) cos θk − ( f2 − qk ) sin θk )2 /a2k
+(( f1 − pk ) sin θk + ( f2 − qk ) cos θk )2 /b k2 ≥ r
pk = [0, 1, 0, 1, 2, 0, 1, 2, 3], a2k = 0.3, b k2 = 1.2, θk = −0.25π
qk = [1.5, 0.5, 2.5, 1.5, 0.5, 3.5, 2.5, 1.5, 0.5]
a = 20, d = 0.5, η = (b + 1)/2, ζ = exp(d − e), γ = 2r
It is the same as that of DAS-CMOP4
It is the same as that of DAS-CMOP4
c1 (x) = sin(aπ x1 ) − b ≥ 0
c2 (x) = cos(aπx2 ) − b ≥ 0
c3 (x) = (e − g(x))(g(x) − d) ≥ 0
Í
ck+3 (x) = 3j=1, j,k fj2 + ( fk − 1)2 − r 2 ≥ 0
Í3
c7 (x) = j=1 ( fj − √1 )2 − r 2 ≥ 0
3
a = 20, d = 0.5, k = 1, 2, 3
η = (b + 1)/2, ζ = exp(d − e), γ = 2r
It is the same as that of DAS-CMOP7
It is the same as that of DAS-CMOP7
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
TABLE III
T HE DIFFICULTY TRIPLETS FOR EACH DAS-CMOP
Difficulty Triplets
(0.0,0.0,0.0)
(0.0,0.25,0.0)
(0.0,0.0,0.25)
(0.25,0.0,0.0)
(0.0,0.5,0.0)
(0.0,0.0,0.5)
(0.5,0.0,0.0)
(0.0,0.75,0.0)
(0.0,0.0,0.75)
(0.75,0.0,0.0)
(0.0,1.0,0.0)
(0.0,0.0,1.0)
(1.0.0,0.0,0.0)
(0.25,0.25,0.25)
(0.5,0.5,0.5)
(0.75,0.75,0.75)
and (0.0, 0.00, 0.75), MOEA/D-CDP is significantly better than
NSGA-II-CDP, which indicates that MOEA/D-CDP is more
suitable for solving DAS-CMOP1 with convergence-hardness.
For DAS-CMOP1 with simultaneous convergence-, feasibilityand convergence-hardness, i.e., the difficulty triplets are
(0.25, 0.25, 0.25), (0.5, 0.5, 0.5) and (0.75, 0.75, 0.75), NSGAII-CDP is significantly better than MOEA/D-CDP.
The final populations with the best IGD values in
30 independent runs by using MOEA/D-CDP and
NSGA-II-CDP on DAS-CMOP1 with difficulty triplets
(0.5, 0.0, 0.0), (0.0, 0.5, 0.0), (0.0, 0.0, 0.5), (0.5, 0.5, 0.5) and
(0.75, 0.75, 0.75) are plotted in Fig. 7. We can observe that
each type of constraint functions in DAS-CMOP1 indeed
generates corresponding difficulties for MOEA/D-CDP and
NSGA-II-CDP. With the increasing of each elements in the
difficulty triplet, the problem is more difficult to solve, as
illustrated by Fig. 7(d)-(e) and Fig. 7(i)-(j).
For DAS-CMOP2 with feasibility-hardness, for example,
the difficulty triplets are (0.0, 0.25, 0.0), (0.0, 0.5, 0.0) and
(0.0, 0.75, 0.0), NSGA-II-CDP is significantly better than
MOEA/D-CDP. For DAS-CMOP2 with difficulty triplets
(0.0, 0.0, 0.5) and (0.0, 0.0, 0.75), MOEA/D-CDP is significantly better than NSGA-II-CDP, which indicates that
MOEA/D-CDP is more suitable for solving DAS-CMOP2 with
convergence-hardness. For DAS-CMOP2 with the difficulty
triplet (1.0, 0.0, 0.0), MOEA/D-CDP is also significantly better than NSGA-II-CDP. For DAS-CMOP2 with simultaneous
convergence-, feasibility- and convergence-hardness, NSGAII-CDP is significantly better than MOEA/D-CDP.
Fig. 8 shows the final populations with the best IGD
values in 30 independent runs by using MOEA/D-CDP
and NSGA-II-CDP on DAS-CMOP2 with difficulty triplets
(0.5, 0.0, 0.0), (0.0, 0.5, 0.0), (0.0, 0.0, 0.5), (0.5, 0.5, 0.5) and
(0.75, 0.75, 0.75). Both MOEA/D-CDP and NSGA-II-CDP can
only achieve a few parts of the PFs. With the increasing of
each element in the difficulty triplet, it is more difficult for
MOEA/D-CDP and NSGA-II-CDP to find the whole PFs of
DAS-CMOP2.
For DAS-CMOP3 with the difficulty triplet (0.0, 0.0, 0.0),
that is, there are no constraints in DAS-CMOP3, NSGAII-CDP is significantly better than MOEA/D-CDP. For
DAS-CMOP3 with difficulty triplets (0.0, 0.0, 0.75) and
(0.0, 0.0, 1.0), MOEA/D-CDP is significantly better than
NSGA-II-CDP, which indicates that MOEA/D-CDP is more
suitable for solving DAS-CMOP3 with larger difficulty levels in terms of convergence-hardness. For DAS-CMOP3
with the difficulty triplet (1.0, 0.0, 0.0), MOEA/D-CDP performs better than NSGA-II-CDP. For DAS-CMOP3 with
feasibility-hardness, for example, the difficulty triplets are
(0.0, 0.25, 0.0), (0.0, 0.5, 0.0) and (0.0, 0.75, 0.0), NSGA-II-
9
CDP is significantly better than MOEA/D-CDP. For DASCMOP3 with simultaneous convergence-, feasibility- and
convergence-hardness, NSGA-II-CDP is significantly better
than MOEA/D-CDP.
Fig. 9 shows the final populations with the best IGD
values in 30 independent runs by using MOEA/D-CDP
and NSGA-II-CDP on DAS-CMOP3 with difficulty triplets
(0.5, 0.0, 0.0), (0.0, 0.5, 0.0), (0.0, 0.0, 0.5), (0.5, 0.5, 0.5) and
(0.75, 0.75, 0.75). For DAS-CMOP3 with diversity- or
feasibility- or convergence-hardness, MOEA/D-CDP and
NSGA-II-CDP can not find the whole PFs. With the increasing
of difficulty triplets, DAS-CMOP3 is becoming more difficult
for MOEA/D-CDP and NSGA-II-CDP to solve, as illustrated
by Fig. 9(d)-(e) and Fig. 9(i)-(j).
The statistic results of IGD values for MOEA/D-CDP and
NSGA-II-CDP on DAS-CMOP4-7 are shown in Table V.
From this Table, we can observe that MOEA/D-CDP is significantly better than MOEA/D-CDP on DAS-CMOP4 with the
difficulty triplet (0.0, 0.0, 0.0). In other words, MOEA/D-CDP
works better than NSGA-II-CDP on DAS-CMOP4 without any
constraints. For DAS-CMOP4 with feasibility-hard difficulty
triplets (0.0, 0.25, 0.0), (0.0, 0.5, 0.0) and (0.0, 0.0, 1.0), NSGAII-CDP is significantly better than MOEA/D-CDP.
For DAS-CMOP4 with diversity-hard difficulty triplets
(0.25, 0.0, 0.0), (0.5, 0.0, 0.0), (0.75, 0.0, 0.0) and (1.0, 0.0, 0.0),
MOEA/D-CDP performs significantly better than NSGAII-CDP. For DAS-CMOP4 with convergence-hard difficulty triplets (0.0, 0.0, 0.25), (0.0, 0.0, 0.5) and (0.0, 0.0, 0.75),
MOEA/D-CDP also performs significantly better than NSGAII-CDP.
For DAS-CMOP5 with feasibility-hard difficulty
triplets (0.0, 0.25, 0.0), (0.0, 0.5, 0.0), (0.0, 0.0, 1.0) and
(0.75, 0.75, 0.75), NSGA-II-CDP performs significantly better
than MOEA/D-CDP. For DAS-CMOP5 with diversity- or
convergence-hardness, MOEA/D-CDP is significantly better
than NSGA-II-CDP.
For DAS-CMOP6 with the difficulty triplet (0.0, 0.0, 1.0),
NSGA-II-CDP is significantly better than MOEA/D-CDP.
For DAS-CMOP6 with difficulty triplets (0.0, 0.0, 0.25),
(0.0, 0.0, 0.5), (0.5, 0.0, 0.0) and (1.0, 1.0, 1.0), MOEA/D-CDP
is significantly better than NSGA-II-CDP. For DAS-CMOP5
with the rest of difficulty triplets, MOEA/D-CDP and NSGAII-CDP have not any significantly difference.
D. Performance Comparisons on three-objective DAS-CMOPs
The statistic results of IGD values for MOEA/D-CDP
and NSGA-II-CDP on DAS-CMOP7-9 are presented in Table IV. For DAS-CMOP7 without any constraints, that
is, the difficulty triplet is (0.0, 0.0, 0.0), NSGA-II-CDP is
significantly better than MOEA/D-CDP. For DAS-CMOP7
with diversity- or feasibility-hardness, for example, the difficulty triplets are (0.0, 0.25, 0.0), (0.25, 0.0, 0.0) (0.0, 0.5, 0.0),
(0.5, 0.0, 0.0), (0.0, 0.75, 0.0) and (0.75, 0.0, 0.0), NSGA-IICDP is also significantly better than MOEA/D-CDP. For DASCMOP7 with difficulty triplets (0.0, 0.0, 1.0) and (1.0, 0.0, 0.0),
MOEA/D-CDP is significantly better than NSGA-II-CDP. For
DAS-CMOP7 with simultaneous diversity-, feasibility- and
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
convergence-hardness, NSGA-II-CDP performs significantly
better than NSGA-II-CDP.
For DAS-CMOP8 with convergence-hardness, for example, the difficulty triplets are (0.0, 0.0, 0.25), (0.0, 0.0, 0.5),
(0.0, 0.0, 0.75) and (0.0, 0.0, 1.0), MOEA/D-CDP performs significantly better than NSGA-II-CDP. For DAS-CMOP8 with
the difficulty triplet (1.0, 0.0, 0.0), MOEA/D-CDP is also significantly better than NSGA-II-CDP. For DAS-CMOP8 with
the rest of difficulty triplets, NSGA-II-CDP is better or significantly better than MOEA/D-CDP.
For DAS-CMOP9 with feasibility-hardness, i.e, the
difficulty triplets are (0.0, 0.25, 0.0), (0.0, 0.5, 0.0) and
(0.0, 0.75, 0.0), NSGA-II-CDP performs significantly better
than MOEA/D-CDP. For DAS-CMOP9 with convergenceor diversity-hardness, i.e., the difficulty triplets are
(0.0, 0.0, 0.25), (0.0, 0.0, 0.5) (0.5, 0.0, 0.0), (0.0, 0.0, 0.75),
(0.75, 0.0, 0.0), (0.0, 0.0, 1.0) (1.0, 0.0, 0.0), MOEA/D-CDP
is significantly better than NSGA-II-CDP. For DAS-CMOP9
with simultaneous diversity-, feasibility- and convergencehardness, NSGA-II-CDP performs significantly better than
MOEA/D-CDP. The final populations with the best IGD
values in 30 independent runs by using MOEA/D-CDP
and NSGA-II-CDP on DAS-CMOP9 with difficulty triplets
(0.5, 0.0, 0.0), (0.0, 0.5, 0.0), (0.0, 0.0, 0.5), (0.5, 0.5, 0.5) and
(0.75, 0.75, 0.75) are plotted in Fig. 10. We can observe that
both MOEA/D-CDP and NSGA-II-CDP only achieve a few
parts of PFs of DAS-CMOP9.
E. Analysis of Experimental Results
From the above performance comparisons on the nine test
instances DAS-CMOPs, it is clear that each type of constraint
functions generates corresponding difficulties for MOEA/DCDP and NSGA-II-CDP. With the increasing of each elements
in the difficulty triplet, the problem is becoming more difficult
for MOEA/D-CDP and NSGA-II-CDP to solve. Furthermore,
it can be concluded that NSGA-II-CDP performs better than
MOEA/D-CDP on DAS-CMOPs with feasibility-hardness.
MOEA/D-CDP performs better than NSGA-II-CDP on DASCMOPs with diversity- or convergence-hardness. In the case
of DAS-CMOPs with simultaneous diversity-, feasibility- and
convergence-hardness, NSGA-II-CDP performs better than
MOEA/D-CDP on most of test instances.
VIII. C ONCLUSION
In this work, we proposed a construction toolkit to build
difficulty adjustable and scalable CMOPs. The method used
to design the construction toolkit is based on three primary
constraint functions identified to correspond to three primary
difficulty types of DAS-CMOPs. The method is also scalable because both number of objectives and constraints can
be conveniently extended. As an example, a set of DASCMOPs (DAS-CMOP1-9) was generated using this construction toolkit. To verify the effectiveness of the suggested test
instances, comprehensive experiments were conducted to test
the performance of two popular CMOEAs (MOEA/D-CDP
and NSGA-II-CDP) on DAS-CMOPs with different difficulty
triplets. Through analyzing the performance of the two test
10
algorithms, it is found that the three primary types of difficulties did exist in the corresponding test problems, and the algorithms under test showed different behaviors in reaching the
PFs. The observation demonstrates that the proposed method
of constructing the CMOPs is very efficient and effective to
help evaluate the performance of the tested algorithms.
ACKNOWLEDGMENT
This research work was supported by Guangdong Key
Laboratory of Digital Signal and Image Processing, the National Natural Science Foundation of China under Grant
(61175073, 61300159, 61332002, 51375287), Jiangsu Natural Science Foundation (BK20130808) and Science and
Technology Planning Project of Guangdong Province, China
(2013B011304002).
R EFERENCES
[1] T. P. Runarsson and X. Yao, “Search biases in constrained evolutionary
optimization,” Systems, Man, and Cybernetics, Part C: Applications and
Reviews, IEEE Transactions on, vol. 35, no. 2, pp. 233–243, 2005.
[2] C. A. Coello Coello, “Theoretical and numerical constraint-handling
techniques used with evolutionary algorithms: a survey of the state of
the art,” Computer Methods in Applied Mechanics and Engineering, vol.
191, no. 11-12, pp. 1245–1287, Jan. 2002.
[3] F. Hoffmeister and J. Sprave, “Problem-independent handling of constraints by use of metric penalty functions,” 1996.
[4] J. A. Joines and C. R. Houck, “On the use of non-stationary penalty
functions to solve nonlinear constrained optimization problems with
ga’s,” in Evolutionary Computation, 1994. IEEE World Congress on
Computational Intelligence., Proceedings of the First IEEE Conference
on. IEEE, 1994, pp. 579–584.
[5] F.-z. Huang, L. Wang, and Q. He, “An effective co-evolutionary differential evolution for constrained optimization,” Applied Mathematics and
computation, vol. 186, no. 1, pp. 340–356, 2007.
[6] J. C. Bean and A. ben Hadj-Alouane, A dual genetic algorithm for
bounded integer programs, 1993.
[7] D. W. Coit, A. E. Smith, and D. M. Tate, “Adaptive penalty methods for
genetic optimization of constrained combinatorial problems,” INFORMS
Journal on Computing, vol. 8, no. 2, pp. 173–182, 1996.
[8] A. Ben Hadj-Alouane and J. C. Bean, “A genetic algorithm for the
multiple-choice integer program,” Operations research, vol. 45, no. 1,
pp. 92–101, 1997.
[9] B. Tessema and G. G. Yen, “A self adaptive penalty function based
algorithm for constrained optimization,” in Evolutionary Computation,
2006. CEC 2006. IEEE Congress on. IEEE, 2006, pp. 246–253.
[10] Y. G. Woldesenbet, G. G. Yen, and B. G. Tessema, “Constraint handling
in multiobjective evolutionary optimization,” Evolutionary Computation,
IEEE Transactions on, vol. 13, no. 3, pp. 514–525, 2009.
[11] T. P. Runarsson and X. Yao, “Stochastic ranking for constrained evolutionary optimization,” Evolutionary Computation, IEEE Transactions
on, vol. 4, no. 3, pp. 284–294, 2000.
[12] K. Deb, “An efficient constraint handling method for genetic algorithms,”
Computer methods in applied mechanics and engineering, vol. 186,
no. 2, pp. 311–338, 2000.
[13] M. Laumanns, L. Thiele, and E. Zitzler, “An efficient, adaptive parameter
variation scheme for metaheuristics based on the epsilon-constraint
method,” European Journal of Operational Research, vol. 169, no. 3,
pp. 932–942, 2006.
[14] T. Takahama, S. Sakai, and N. Iwane, “Constrained optimization by the ε
constrained hybrid algorithm of particle swarm optimization and genetic
algorithm,” in AI 2005: Advances in Artificial Intelligence. Springer,
2005, pp. 389–400.
[15] T. Ray, H. K. Singh, A. Isaacs, and W. Smith, “Infeasibility driven
evolutionary algorithm for constrained optimization,” in Constrainthandling in evolutionary optimization. Springer, 2009, pp. 145–165.
[16] P. D. Surry and N. J. Radcliffe, “The comoga method: constrained
optimisation by multi-objective genetic algorithms,” Control and Cybernetics, vol. 26, pp. 391–412, 1997.
[17] Z. Cai and Y. Wang, “A multiobjective optimization-based evolutionary
algorithm for constrained optimization,” Evolutionary Computation,
IEEE Transactions on, vol. 10, no. 6, pp. 658–675, 2006.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
11
TABLE IV
M EAN AND STANDARD DEVIATION OF I GD VALUES OBTAINED BY MOEA/D-CDP AND NSGA-II-CDP ON DAS-CMOP1-3. W ILCOXON ’ S RANK SUM
TEST AT 0.05 SIGNIFICANCE LEVEL IS PERFORMED BETWEEN MOEA/D-CDP AND NSGA-II-CDP. † AND ‡ DENOTE THAT THE PERFORMANCE OF
NSGA-II-CDP IS SIGNIFICANTLY WORSE THAN OR BETTER THAN THAT OF MOEA/D-CDP, RESPECTIVELY.
DAS-CMOP1
MOEA/D-CDP
NSGA-II-CDP
1.367E-01
1.494E-01
2.934E-02
3.723E-02
1.375E-01
9.878E-02‡
2.591E-02
1.185E-02
1.351E-01
2.006E-01†
3.515E-02
2.100E-02
1.497E-01
1.524E-01
2.609E-02
4.081E-02
1.578E-01
1.174E-01‡
1.516E-02
1.304E-02
1.809E-01
2.362E-01†
3.005E-02
2.125E-02
1.602E-01
1.605E-01
4.631E-02
3.716E-02
1.858E-01
1.420E-01‡
3.527E-02
1.546E-02
1.769E-01
3.147E-01†
3.770E-02
5.324E-02
1.658E-01
1.510E-01
4.960E-02
4.018E-02
3.682E-01
3.636E-01
1.230E-02
6.217E-03
6.998E-01
4.625E-01‡
3.767E-01
2.755E-02
4.515E-01
1.531E+00†
1.169E-01
9.908E-01
2.674E-01
2.119E-01‡
3.893E-02
4.025E-02
4.160E-01
3.624E-01‡
7.441E-02
4.702E-02
7.750E-01
7.079E-01‡
5.452E-02
9.615E-02
(0.0,0.25,0.0)
(0.0,0.0,0.25)
(0.25,0.0,0.0)
(0.0,0.5,0.0)
(0.0,0.0,0.5)
(0.5,0.0,0.0)
(0.0,0.75,0.0)
(0.0,0.0,0.75)
(0.75,0.0,0.0)
(0.0,1.0,0.0)
(0.0,0.0,1.0)
(1.0.0,0.0,0.0)
(0.25,0.25,0.25)
(0.5,0.5,0.5)
(0.75,0.75,0.75)
1
1.5
True PF
MOEA/D-CDP
0.4
1
1.5
0.6
0.4
1
0.2
0.4
0.6
0.8
0.5
0.5
1
1
0
1.5
0
0.5
f1(x)
0.2
0.5
1
f1(x)
(f) (0.5,0.0,0.0)
1.5
0.5
0.5
1
f1(x)
(g) (0.0,0.5,0.0)
1.5
1.6
True PF
NSGA-II-CDP
1.5
1
0
0.5
1
f1(x)
(h) (0.0,0.0,0.5)
1.5
0.5
0.5
1.5
(e) (0.75,0.75,0.75)
True PF
NSGA-II-CDP
1.4
2
0.4
1
f1(x)
2.5
0.6
0
0.6
0.5
1.5
(d) (0.5,0.5,0.5)
0.2
0
1
True PF
NSGA-II-CDP
0.8
f2(x)
f2(x)
1
1
f1(x)
1
True PF
NSGA-II-CDP
1.5
0.4
0.5
0.5
1.5
(c) (0.0,0.0,0.5)
2
True PF
NSGA-II-CDP
0.6
1
f1(x)
(b) (0.0,0.5,0.0)
1
0.8
1.2
0.8
f2(x)
0
(a) (0.5,0.0,0.0)
f2(x)
True PF
1.4
0.2
f1(x)
0
1.6
True PF
MOEA/D-CDP
f2(x)
0.2
0
2
True PF
MOEA/D-CDP
0.8
f2(x)
0.6
f2(x)
f2(x)
1
True PF
MOEA/D-CDP
0.8
DAS-CMOP3
MOEA/D-CDP
NSGA-II-CDP
1.909E-01
1.536E-01‡
4.315E-02
2.340E-02
1.744E-01
1.224E-01‡
4.493E-02
2.374E-02
2.076E-01
1.888E-01‡
4.827E-02
3.296E-02
2.300E-01
2.146E-01
5.525E-02
5.225E-02
1.847E-01
1.292E-01‡
3.752E-02
2.305E-02
2.127E-01
2.095E-01
4.019E-02
1.883E-02
4.185E-01
3.263E-01‡
1.068E-01
1.168E-01
2.320E-01
1.527E-01‡
4.604E-02
2.729E-02
2.201E-01
2.656E-01†
2.596E-02
5.844E-02
2.258E-01
1.956E-01‡
5.725E-02
5.785E-02
4.353E-01
4.389E-01
4.106E-02
2.952E-02
6.590E-01
6.615E-01†
2.879E-04
1.982E-03
4.331E-01
1.842E+00†
9.567E-02
1.431E+00
2.589E-01
1.686E-01‡
3.644E-02
4.206E-02
4.680E-01
4.027E-01‡
5.516E-02
3.817E-02
7.473E-01
2.808E-01‡
1.864E-01
1.242E-01
f2(x)
(0.0,0.0,0.0)
DAS-CMOP2
MOEA/D-CDP
NSGA-II-CDP
1.678E-01
1.720E-01
2.675E-02
3.710E-02
1.203E-01
7.994E-02‡
2.097E-02
1.316E-02
1.709E-01
1.606E-01
3.485E-02
4.088E-02
1.719E-01
1.796E-01
3.073E-02
4.273E-02
1.359E-01
9.157E-02‡
2.820E-02
1.670E-02
1.580E-01
2.164E-01†
4.670E-02
4.156E-02
1.796E-01
1.799E-01
3.874E-02
3.604E-02
1.548E-01
1.073E-01‡
2.665E-02
1.557E-02
1.554E-01
3.354E-01†
2.601E-02
1.110E-01
2.206E-01
1.712E-01‡
4.483E-02
4.269E-02
3.250E-01
3.235E-01
7.175E-03
4.915E-03
7.388E-01
7.177E-01
1.901E-01
1.254E-01
4.102E-01
1.413E+00†
1.029E-01
8.854E-01
1.404E-01
9.860E-02‡
2.999E-02
1.632E-02
1.656E-01
1.123E-01‡
4.357E-02
1.638E-02
1.867E-01
1.029E-01 ‡
3.758E-02
1.201E-02
f2(x)
Instance
Difficulty Triplet
1.2
1
0.8
1
f1(x)
(i) (0.5,0.5,0.5)
1.5
0.6
0.5
1
1.5
f1(x)
(j) (0.75,0.75,0.75)
Fig. 7. The final populations with the best I GD values in 30 independent runs by using MOEA/D-CDP and NSGA-II-CDP on DAS-CMOP1 with different
difficulty triplets. (a)-(e) and (f)-(j) show the best populations achieved by MOEA/D-CDP and NSGA-II-CDP on DAS-CMOP1 with difficulty triplets
(0.5, 0.0, 0.0), (0.0, 0.5, 0.0), (0.0, 0.0, 0.5), (0.5, 0.5, 0.5) and (0.75, 0.75, 0.75), respectively.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
12
TABLE V
M EAN AND STANDARD DEVIATION OF I GD VALUES OBTAINED BY MOEA/D-CDP AND NSGA-II-CDP ON DAS-CMOP4-6. W ILCOXON ’ S RANK SUM
TEST AT 0.05 SIGNIFICANCE LEVEL IS PERFORMED BETWEEN MOEA/D-CDP AND NSGA-II-CDP. † AND ‡ DENOTE THAT THE PERFORMANCE OF
NSGA-II-CDP IS SIGNIFICANTLY WORSE THAN OR BETTER THAN THAT OF MOEA/D-CDP, RESPECTIVELY.
DAS-CMOP4
MOEA/D-CDP
NSGA-II-CDP
1.121E-02
5.282E-02†
8.906E-03
2.977E-02
2.913E-03
2.412E-03‡
1.290E-03
1.460E-04
7.880E-02
3.380E-01†
4.741E-02
5.514E-02
1.849E-02
6.475E-02†
1.324E-02
3.587E-02
2.915E-03
2.345E-03‡
1.108E-03
7.926E-05
2.380E-01
9.456E-01†
8.551E-02
3.330E-01
2.321E-02
5.856E-02†
1.611E-02
3.015E-02
2.668E-03
3.105E-03
7.843E-04
3.840E-03
4.260E-01
1.266E+00†
1.443E-01
1.709E-01
1.690E-02
5.761E-02†
1.301E-02
2.894E-02
5.229E-02
1.019E-01
1.307E-01
1.272E-01
1.824E+00
1.633E+00‡
9.650E-02
2.638E-01
4.795E-01
7.564E+00†
1.055E-01
4.254E+00
3.392E-03
5.863E-02†
1.941E-03
1.292E-01
1.219E-02
3.658E-01†
3.938E-02
1.246E-01
2.341E-01
2.449E-01
1.670E-03
5.065E-02
(0.0,0.25,0.0)
(0.0,0.0,0.25)
(0.25,0.0,0.0)
(0.0,0.5,0.0)
(0.0,0.0,0.5)
(0.5,0.0,0.0)
(0.0,0.75,0.0)
(0.0,0.0,0.75)
(0.75,0.0,0.0)
(0.0,1.0,0.0)
(0.0,0.0,1.0)
(1.0.0,0.0,0.0)
(0.25,0.25,0.25)
(0.5,0.5,0.5)
(0.75,0.75,0.75)
1
2
True PF
MOEA/D-CDP
0.4
f2(x)
0.6
1
0.4
0.2
0.4
0.6
0.8
1
1
0
1.5
(a) (0.5,0.0,0.0)
1
0.2
(b) (0.0,0.5,0.0)
0.5
0.5
1
1
1
0.2
1
0.5
0.5
1.5
f1(x)
1
(d) (0.5,0.5,0.5)
(e) (0.75,0.75,0.75)
1.5
True PF
NSGA-II-CDP
True PF
NSGA-II-CDP
0.6
0.4
1.5
f1(x)
1.5
0.8
f2(x)
f2(x)
0.8
(c) (0.0,0.0,0.5)
True PF
NSGA-II-CDP
0.4
0.6
1
True PF
NSGA-II-CDP
0.6
0.4
f1(x)
1.5
0.8
0
f1(x)
f2(x)
0
1
f2(x)
0.6
True PF
MOEA/D-CDP
0.2
0.5
0.5
f1(x)
0
1.5
True PF
MOEA/D-CDP
True PF
NSGA-II-CDP
f2(x)
0.2
0
1.5
True PF
MOEA/D-CDP
0.8
1.5
f2(x)
f2(x)
1
True PF
MOEA/D-CDP
0.8
DAS-CMOP6
MOEA/D-CDP
NSGA-II-CDP
6.304E-02
6.778E-02
7.337E-02
3.402E-02
6.410E-02
5.489E-02
6.666E-02
5.249E-02
1.202E-01
3.907E-01†
5.496E-02
2.475E-01
9.321E-02
1.291E-01
7.435E-02
6.290E-02
7.189E-02
5.139E-02
4.962E-02
5.530E-02
4.169E-01
9.448E-01†
3.895E-01
3.083E-01
9.403E-02
1.369E-01†
5.364E-02
5.124E-02
6.957E-02
6.104E-02
7.664E-02
5.444E-02
1.260E+00
1.258E+00
3.391E-01
2.116E-01
1.446E-01
1.268E-01
1.056E-01
6.071E-02
1.816E-01
1.616E-01
1.024E-01
1.222E-01
1.983E+00
1.656E+00‡
1.490E-01
2.554E-01
5.126E-01
8.314E+00†
9.984E-02
6.611E+00
1.719E-01
1.418E-01
1.534E-01
1.734E-01
7.863E-01
5.457E-01
5.005E-01
3.147E-01
1.014E+00
9.106E-01
3.383E-01
4.101E-01
f2(x)
(0.0,0.0,0.0)
DAS-CMOP5
MOEA/D-CDP
NSGA-II-CDP
1.464E-02
5.356E-02†
1.365E-02
2.640E-02
3.034E-03
2.269E-03‡
1.265E-03
7.276E-05
4.158E-02
4.192E-01†
2.298E-02
1.665E-01
2.010E-02
6.510E-02†
1.235E-02
4.145E-02
2.940E-03
2.390E-03‡
1.187E-03
6.758E-04
1.010E-01
8.761E-01†
7.600E-02
2.850E-01
1.822E-02
5.840E-02†
9.361E-03
2.992E-02
2.587E-03
3.078E-03
1.237E-03
4.450E-03
5.002E-01
1.223E+00†
1.142E-01
2.086E-01
1.992E-02
7.616E-02†
1.608E-02
3.299E-02
1.178E-02
1.114E-01†
3.037E-02
1.017E-01
1.946E+00
1.612E+00‡
3.058E-01
2.741E-01
4.465E-01
7.917E+00†
9.627E-02
7.613E+00
2.754E-03
4.428E-02
6.112E-04
1.206E-01
5.861E-01
2.919E-01
7.668E-01
4.291E-01
1.319E+00
4.312E-01‡
4.447E-01
5.137E-01
f2(x)
Instance
Difficulty Triplet
1
1
0.2
0
0.2
0.4
0.6
f1(x)
(f) (0.5,0.0,0.0)
0.8
1
0.5
0.5
1
f1(x)
(g) (0.0,0.5,0.0)
1.5
0
0
0.2
0.4
0.6
f1(x)
(h) (0.0,0.0,0.5)
0.8
1
0.5
0.5
1
f1(x)
(i) (0.5,0.5,0.5)
1.5
0.5
0.5
1
1.5
f1(x)
(j) (0.75,0.75,0.75)
Fig. 8. The final populations with the best I GD values in 30 independent runs by using MOEA/D-CDP and NSGA-II-CDP on DAS-CMOP2 with different
difficulty triplets. (a)-(e) and (f)-(j) show the best populations achieved by MOEA/D-CDP and NSGA-II-CDP on DAS-CMOP2 with difficulty triplets
(0.5, 0.0, 0.0), (0.0, 0.5, 0.0), (0.0, 0.0, 0.5), (0.5, 0.5, 0.5) and (0.75, 0.75, 0.75), respectively.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
13
TABLE VI
M EAN AND STANDARD DEVIATION OF I GD VALUES OBTAINED BY MOEA/D-CDP AND NSGA-II-CDP ON DAS-CMOP7-9. W ILCOXON ’ S RANK SUM
TEST AT 0.05 SIGNIFICANCE LEVEL IS PERFORMED BETWEEN MOEA/D-CDP AND NSGA-II-CDP. † AND ‡ DENOTE THAT THE PERFORMANCE OF
NSGA-II-CDP IS SIGNIFICANTLY WORSE THAN OR BETTER THAN THAT OF MOEA/D-CDP, RESPECTIVELY.
DAS-CMOP7
MOEA/D-CDP
NSGA-II-CDP
6.052E-02
5.150E-02‡
1.018E-03
4.413E-03
5.154E-02
4.946E-02‡
1.787E-03
2.179E-03
5.048E-02
5.288E-02†
9.856E-04
5.451E-03
5.852E-02
4.869E-02‡
7.656E-04
4.057E-03
5.178E-02
4.891E-02‡
3.099E-03
1.839E-03
4.583E-02
6.307E-02†
2.608E-04
1.707E-02
5.546E-02
4.548E-02‡
1.450E-03
4.871E-03
5.197E-02
5.008E-02‡
2.550E-03
2.209E-03
4.523E-02
5.550E-02†
8.689E-04
1.165E-02
5.030E-02
4.257E-02‡
1.483E-03
4.277E-03
6.275E-02
5.037E-02
5.399E-02
2.946E-03
4.052E-02
3.359E-01†
3.495E-03
3.097E-02
5.309E-01
1.045E+01†
1.013E-01
9.771E+00
5.149E-02
4.687E-02‡
2.355E-03
2.668E-03
5.015E-02
4.335E-02‡
2.605E-03
2.033E-03
4.671E-02
3.979E-02‡
2.854E-03
3.892E-03
(0.0,0.0,0.25)
(0.25,0.0,0.0)
(0.0,0.5,0.0)
(0.0,0.0,0.5)
(0.5,0.0,0.0)
(0.0,0.75,0.0)
(0.0,0.0,0.75)
(0.75,0.0,0.0)
(0.0,1.0,0.0)
(0.0,0.0,1.0)
(1.0.0,0.0,0.0)
(0.25,0.25,0.25)
(0.5,0.5,0.5)
(0.75,0.75,0.75)
1
1.6
True PF
MOEA/D-CDP
1
0.8
0.2
0.2
0.4
0.6
0.8
0.4
0.5
1
(a) (0.5,0.0,0.0)
1
0
1.5
1
0
(b) (0.0,0.5,0.0)
0.2
0.6
0.8
1
f1(x)
(f) (0.5,0.0,0.0)
1.5
0.5
0.5
1
f1(x)
(g) (0.0,0.5,0.0)
1.5
0.6
0.4
1.5
1.5
1
0.2
0.4
0.6
f1(x)
(h) (0.0,0.0,0.5)
0.8
1
0.5
0.5
1
1.2
1.4
(e)(0.75,0.75,0.75)
True PF
NSGA-II-CDP
0.4
0.8
f1(x)
True PF
NSGA-II-CDP
1.4
2
0
0.6
1.6
True PF
NSGA-II-CDP
0.6
0
1
(d)(0.5,0.5,0.5)
0.2
0.5
1
f1(x)
2.5
0.8
f2(x)
1
0.2
0
0.5
0.5
1
(c)(0.0,0.0,0.5)
True PF
NSGA-II-CDP
f2(x)
0.4
0.4
1
True PF
NSGA-II-CDP
1.2
0.8
f1(x)
1.5
0.6
0
0.4
f1(x)
1
0.8
0.6
True PF
MOEA/D-CDP
1.4
1.5
f2(x)
0
1.6
True PF
MOEA/D-CDP
0.2
0.6
f1(x)
f2(x)
f2(x)
0.4
2
True PF
MOEA/D-CDP
0.8
1.2
0.6
0
1
True PF
MOEA/D-CDP
1.4
f2(x)
f2(x)
0.8
f2(x)
(0.0,0.25,0.0)
DAS-CMOP9
MOEA/D-CDP
NSGA-II-CDP
2.737E-01
3.158E-01
2.447E-01
1.784E-01
3.499E-01
2.084E-01‡
1.447E-01
1.231E-01
9.470E-02
3.315E-01†
6.021E-02
1.805E-01
3.127E-01
3.458E-01
2.813E-01
1.977E-01
3.552E-01
2.347E-01‡
1.265E-01
1.042E-01
8.891E-02
3.847E-01†
2.556E-02
1.532E-01
1.875E-01
4.130E-01†
1.502E-01
1.874E-01
3.309E-01
1.880E-01‡
1.275E-01
1.016E-01
1.430E-01
4.165E-01†
1.100E-01
1.053E-01
1.963E-01
3.675E-01†
1.720E-01
2.286E-01
6.635E-01
6.360E-01
5.863E-02
3.965E-02
3.265E-01
4.129E-01†
1.061E-01
4.490E-02
6.505E-01
3.184E+00†
1.102E-01
2.713E+00
3.636E-01
2.390E-01‡
1.293E-01
1.075E-01
3.704E-01
2.902E-01‡
1.294E-01
1.054E-01
3.874E-01
3.016E-01‡
1.501E-01
1.283E-01
f2(x)
(0.0,0.0,0.0)
DAS-CMOP8
MOEA/D-CDP
NSGA-II-CDP
6.679E-02
6.679E-02
5.616E-04
5.958E-03
6.759E-02
6.453E-02‡
2.889E-03
1.798E-03
6.399E-02
7.098E-02†
1.117E-03
3.921E-03
6.756E-02
6.326E-02‡
1.122E-03
3.801E-03
6.718E-02
6.455E-02‡
2.691E-03
2.882E-03
6.502E-02
7.726E-02†
1.267E-03
7.831E-03
6.663E-02
5.746E-02‡
1.498E-03
4.938E-03
6.746E-02
6.401E-02‡
2.562E-03
2.269E-03
6.681E-02
7.974E-02†
5.307E-03
8.422E-03
6.151E-02
5.014E-02‡
1.504E-03
5.565E-03
9.187E-02
6.621E-02
9.316E-02
2.866E-03
5.852E-02
6.803E-02†
8.136E-03
1.571E-02
6.650E-01
8.974E+00†
9.562E-02
6.236E+00
6.997E-02
6.184E-02‡
2.735E-03
2.544E-03
6.961E-02
5.518E-02‡
3.527E-03
2.564E-03
6.909E-02
4.677E-02‡
3.572E-03
3.447E-03
f2(x)
Instance
Difficulty Triplet
1.2
1
0.8
1
f1(x)
(i) (0.5,0.5,0.5)
1.5
0.6
0.4
0.6
0.8
1
1.2
1.4
f1(x)
(j) (0.75,0.75,0.75)
Fig. 9. The final populations with the best I GD values in 30 independent runs by using MOEA/D-CDP and NSGA-II-CDP on DAS-CMOP3 with different
difficulty triplets. (a)-(e) and (f)-(j) show the best populations achieved by MOEA/D-CDP and NSGA-II-CDP on DAS-CMOP3 with difficulty triplets
(0.5, 0.0, 0.0), (0.0, 0.5, 0.0), (0.0, 0.0, 0.5), (0.5, 0.5, 0.5) and (0.75, 0.75, 0.75), respectively.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015
f2(x)
f1(x)
(a) (0.5,0.0,0.0)
f3(x)
f3(x)
0
2
2
f2(x)
1
0 0
f1(x)
(f) (0.5,0.0,0.0)
1.5
1
0.5
2
f2(x)
1
0 0.5
(g) (0.0,0.5,0.0)
f1(x)
0
2
f2(x)
2
1
f1(x)
(h) (0.0,0.0,0.5)
0.5
1.5
f2(x)
f1(x)
1
0.5
2
f2(x)
2
1
f1(x)
(i) (0.5,0.5,0.5)
f1(x)
True PF
NSGA-II-CDP
1.5
1.5
0 0
1
0.5 0.5
(e) (0.75,0.75,0.75)
True PF
NSGA-II-CDP
1
1.5
1
1
0 0.5
2
1
0 0
1
(d) (0.5,0.5,0.5)
0.5
1
f3(x)
f2(x)
f1(x)
True PF
NSGA-II-CDP
1.5
1.5
1
1
0 0
1.5
1
(c) (0.0,0.0,0.5)
True PF
NSGA-II-CDP
2
0.5
1
f2(x)
f1(x)
(b) (0.0,0.5,0.0)
True PF
NSGA-II-CDP
1
1
0.5 0
1
0.5
2
2
1
True PF
MOEA/D-CDP
1.5
f3(x)
1
0 0
1
0.5
0
2
2
1
True PF
MOEA/D-CDP
1.5
f3(x)
1
f3(x)
f2(x)
1.5
0.5
1.5
2
1
True PF
MOEA/D-CDP
1.5
f3(x)
f3(x)
f3(x)
1
0.5
0
2
True PF
MOEA/D-CDP
2
f3(x)
True PF
MOEA/D-CDP
1.5
14
1
0.5
2
2
1
f2(x)
1
0 0
f1(x)
(j) (0.75,0.75,0.75)
Fig. 10. The final populations with the best I GD values in 30 independent runs by using MOEA/D-CDP and NSGA-II-CDP on DAS-CMOP9 with
different difficulty triplets. (a)-(e) and (f)-(j) show the best populations achieved by MOEA/D-CDP and NSGA-II-CDP on DAS-CMOP9 with difficulty triplets
(0.5, 0.0, 0.0), (0.0, 0.5, 0.0), (0.0, 0.0, 0.5), (0.5, 0.5, 0.5) and (0.75, 0.75, 0.75), respectively.
[18] Y. Wang, Z. Cai, Y. Zhou, and W. Zeng, “An adaptive tradeoff model
for constrained evolutionary optimization,” Evolutionary Computation,
IEEE Transactions on, vol. 12, no. 1, pp. 80–92, 2008.
[19] B. Y. Qu and P. N. Suganthan, “Constrained multi-objective optimization
algorithm with an ensemble of constraint handling methods,” Engineering Optimization, vol. 43, no. 4, pp. 403–416, 2011.
[20] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist
multiobjective genetic algorithm: NSGA-II,” Evolutionary Computation,
IEEE Transactions on, vol. 6, no. 2, pp. 182–197, 2002.
[21] D. W. Corne, N. R. Jerram, J. D. Knowles, M. J. Oates et al., “PESA-II:
Region-based selection in evolutionary multiobjective optimization,” in
Proceedings of the Genetic and Evolutionary Computation Conference
(GECCO’2001). Citeseer, 2001.
[22] E. Zitzler, M. Laumanns, L. Thiele, E. Zitzler, E. Zitzler, L. Thiele,
and L. Thiele, “SPEA2: Improving the strength pareto evolutionary
algorithm,” 2001.
[23] Q. Zhang and H. Li, “MOEA/D: A multiobjective evolutionary algorithm
based on decomposition,” Evolutionary Computation, IEEE Transactions
on, vol. 11, no. 6, pp. 712–731, 2007.
[24] H. Li and Q. Zhang, “Multiobjective optimization problems with complicated pareto sets, MOEA/D and NSGA-II,” Evolutionary Computation,
IEEE Transactions on, vol. 13, no. 2, pp. 284–302, 2009.
[25] H.-L. Liu, F. Gu, and Q. Zhang, “Decomposition of a multiobjective
optimization problem into a number of simple multiobjective subproblems,” Evolutionary Computation, IEEE Transactions on, vol. 18, no. 3,
pp. 450–455, 2014.
[26] X. Cai, Y. Li, Z. Fan, and Q. Zhang, “An external archive guided
multiobjective evolutionary algorithm based on decomposition for combinatorial optimization,” Evolutionary Computation, IEEE Transactions
on, vol. 19, no. 4, pp. 508–523, Aug 2015.
[27] E. Zitzler and S. Künzli, “Indicator-based selection in multiobjective
search,” in Parallel Problem Solving from Nature-PPSN VIII. Springer,
2004, pp. 832–842.
[28] D. H. Phan and J. Suzuki, “R2-IBEA: R2 indicator based evolutionary
algorithm for multiobjective optimization,” in Evolutionary Computation
(CEC), 2013 IEEE Congress on. IEEE, 2013, pp. 1836–1845.
[29] N. Beume, B. Naujoks, and M. Emmerich, “SMS-EMOA: Multiobjective selection based on dominated hypervolume,” European Journal of
Operational Research, vol. 181, no. 3, pp. 1653–1669, 2007.
[30] J. Bader and E. Zitzler, “HypE: An algorithm for fast hypervolumebased many-objective optimization,” Evolutionary computation, vol. 19,
no. 1, pp. 45–76, 2011.
[31] K. Deb, “Multi-objective Genetic Algorithms: Problem Difficulties and
Construction of Test Problems,” Evolutionary Computation, vol. 7, no. 3,
pp. 205–230, 1999.
[32] H. Li and Q. Zhang, “Multiobjective Optimization Problems With Complicated Pareto Sets, MOEA/D and NSGA-II.” IEEE Trans. Evolutionary
Computation, 2009.
[33] K. Deb, L. Thiele, M. Laumanns, and E. Zitzler, Scalable test problems
for evolutionary multiobjective optimization. Springer, 2005.
[34] H.-L. Liu, F. Gu, and Q. Zhang, “Decomposition of a Multiobjective
Optimization Problem Into a Number of Simple Multiobjective Sub-
problems,” IEEE Transactions on Evolutionary Computation, vol. 18,
no. 3, pp. 450–455, 2014.
[35] H.-L. Liu, L. Chen, K. Deb, and E. Goodman, “Investigating the Effect
of Imbalance Between Convergence and Diversity in Evolutionary Multiobjective Algorithms,” IEEE Transactions on Evolutionary Computation,
vol. PP, no. 99, pp. 1–1, 2016.
[36] R. Cheng, Y. Jin, M. Olhofer, and B. Sendhoff, “Test Problems for
Large-Scale Multiobjective and Many-Objective Optimization,” IEEE
Transactions on Cybernetics, pp. 1–14, 2016.
[37] S. Huband, P. Hingston, L. Barone, and L. While, “A review of multiobjective test problems and a scalable test problem toolkit,” Evolutionary
Computation, IEEE Transactions on, vol. 10, no. 5, pp. 477–506, 2006.
[38] K. Deb, Multi-objective optimization using evolutionary algorithms.
John Wiley & Sons, 2001, vol. 16.
[39] Q. Zhang, A. Zhou, S. Zhao, P. N. Suganthan, W. Liu, and S. Tiwari,
“Multiobjective optimization test instances for the cec-2009 special session and competition,” University of Essex, Colchester, UK and Nanyang
technological University, Singapore, special session on performance
assessment of multi-objective optimization algorithms, technical report,
pp. 1–30, 2008.
[40] T. T. Binh and U. Korn, “Mobes: A multiobjective evolution strategy
for constrained optimization problems,” in The Third International
Conference on Genetic Algorithms (Mendel 97), vol. 25. Citeseer,
1997, p. 27.
[41] M. Tanaka, H. Watanabe, Y. Furukawa, and T. Tanino, “Ga-based
decision support system for multicriteria optimization,” in Systems, Man
and Cybernetics, 1995. Intelligent Systems for the 21st Century., IEEE
International Conference on, vol. 2. IEEE, 1995, pp. 1556–1561.
[42] N. Srinvas and K. Deb, “Multi-objective function optimization using
non-dominated sorting genetic algorithms,” Evolutionary Computation,
vol. 2, no. 3, pp. 221–248, 1994.
[43] A. Osyczka and S. Kundu, “A new method to solve generalized
multicriteria optimization problems using the simple genetic algorithm,”
Structural optimization, vol. 10, no. 2, pp. 94–99, 1995.
[44] H. Jain and K. Deb, “An Evolutionary Many-Objective Optimization Algorithm Using Reference-Point Based Nondominated Sorting Approach,
Part II - Handling Constraints and Extending to an Adaptive Approach.”
IEEE Trans. Evolutionary Computation, vol. 18, no. 4, pp. 602–622,
2014.
[45] D. A. Van Veldhuizen and G. B. Lamont, “Evolutionary computation and
convergence to a pareto front,” in Late breaking papers at the genetic
programming 1998 conference. Citeseer, 1998, pp. 221–228.
[46] H. Li, Q. Zhang, and J. Deng, “Multiobjective test problems with
complicated pareto fronts: Difficulties in degeneracy,” in Evolutionary
Computation (CEC), 2014 IEEE Congress on. IEEE, 2014, pp. 2156–
2163.
[47] P. A. Bosman and D. Thierens, “The balance between proximity
and diversity in multiobjective evolutionary algorithms,” Evolutionary
Computation, IEEE Transactions on, vol. 7, no. 2, pp. 174–188, 2003.
| 9cs.NE
|
arXiv:1707.08644v2 [math.ST] 24 Oct 2017
Sharpening Jensen’s Inequality
J. G. Liao and Arthur Berg
Division of Biostatistics and Bioinformatics
Penn State University College of Medicine
October 26, 2017
Abstract
This paper proposes a new sharpened version of the Jensen’s inequality. The
proposed new bound is simple and insightful, is broadly applicable by imposing minimum assumptions, and provides fairly accurate result in spite of its simple form.
Applications to the moment generating function, power mean inequalities, and RaoBlackwell estimation are presented. This presentation can be incorporated in any
calculus-based statistical course.
Keywords: Jensen gap, Power mean inequality, Rao-Blackwell Estimator, Taylor series
1
1
Introduction
Jensen’s inequality is a fundamental inequality in mathematics and it underlies many important statistical proofs and concepts. Some standard applications include derivation
of the arithmetic-geometric mean inequality, non-negativity of Kullback and Leibler divergence, and the convergence property of the expectation-maximization algorithm (Dempster et al.,
1977). Jensen’s inequality is covered in all major statistical textbooks such as Casella and Berger
(2002, Section 4.7) and Wasserman (2013, Section 4.2) as a basic mathematical tool for
statistics.
Let X be a random variable with finite expectation and let ϕ(x) be a convex function,
then Jensen’s inequality (Jensen, 1906) establishes
E [ϕ (X)] − ϕ (E [X]) ≥ 0.
(1)
This inequality, however, is not sharp unless var(X) = 0 or ϕ(x) is a linear function of x.
Therefore, there is substantial room for advancement. This paper proposes a new sharper
bound for the Jensen gap E[ϕ(X)] − ϕ (E[X]). Some other improvements of Jensen’s inequality have been developed recently; see for example Walker (2014), Abramovich and Persson
(2016); Horvath et al. (2014) and references cited therein. Our proposed bound, however,
has the following advantages. First, it has a simple, easy to use, and insightful form in
terms of the second derivative ϕ′′ (x) and var(X). At the same time, it gives fairly accurate
results in the several examples below. Many previously published improvements, however,
are much more complicated in form, much more involved to use, and can even be more
difficult to compute than E[ϕ(X)] itself as discussed in Walker (2014). Second, our method
requires only the existence of ϕ′′ (x) and is therefore broadly applicable. In contrast, some
other methods require ϕ(x) to admit a power series representation with positive coefficients (Abramovich and Persson, 2016; Dragomir, 2014; Walker, 2014) or require ϕ(x) to
be super-quadratic (Abramovich et al., 2014). Third, we provide both a lower bound and
an upper bound in a single formula.
We have incorporated the materials in this paper in our classroom teaching. With only
slightly increased technical level and lecture time, we are able to present a much sharper
version of the Jensen’s inequality that significantly enhances students’ understanding of
2
the underlying concepts.
2
Main result
Theorem 1. Let X be a one-dimensional random variable with mean µ, and P (X ∈ (a, b)) =
1, where −∞ ≤ a < b ≤ ∞. Let ϕ(x) is a twice differentiable function on (a, b), and define
function
h(x; ν) ,
Then
ϕ (x) − ϕ (ν) ϕ′ (ν)
−
.
x−ν
(x − ν)2
inf {h(x; µ)}var(X) ≤ E [ϕ (X)] − ϕ (E[X]) ≤ sup {h(x; µ)}var(X).
x∈(a,b)
(2)
x∈(a,b)
Proof. Let F (x) be the cumulative distribution function of X. Applying Taylor’s theorem
to ϕ(x) about µ with a mean-value form of the remainder gives
ϕ(x) = ϕ(µ) + ϕ′ (µ)(x − µ) +
ϕ′′ (g(x))
(x − µ)2 ,
2
where g(x) is between x and µ. Explicitly solving for ϕ′′ (g(x))/2 gives ϕ′′ (g(x))/2 = h(x; µ)
as defined above. Therefore
b
E [ϕ (X)] − ϕ (E [X]) =
Z
=
Z
=
Z
b
{ϕ(x) − ϕ(µ)} dF (x)
a
a
a
ϕ′ (µ)(x − µ) + h(x; µ)(x − µ)2 dF (x)
b
h(x; µ)(x − µ)2 dF (x),
and the result follows because inf x∈(a,b) h(x; µ) ≤ h(x; µ) ≤ supx∈(a,b) h(x; µ).
Theorem 1 also holds when inf h(x; µ) is replaced by inf ϕ′′ (x)/2 and sup h(x; µ) replaced
by sup ϕ′′ (x)/2 since
inf
ϕ′′ (x)
≤ inf h(x; µ) and
2
sup
ϕ′′ (x)
≥ sup h(x; µ).
2
These less tight bounds are implied in the economics working paper Becker (2012). Our
lower and upper bounds have the general form J · var(X), where J depends on ϕ. Similar
3
forms of bounds are presented in Abramovich and Persson (2016); Dragomir (2014); Walker
(2014), but our J in Theorem 1 is much simpler and applies to a wider class of ϕ.
Inequality (2) implies Jensen’s inequality when ϕ′′ (x) ≥ 0. Note also that Jensen’s
inequality is sharp when ϕ(x) is linear, whereas inequality (2) is sharp when ϕ(x) is a
quadratic function of x.
In some applications the moments of X present in (2) are unknown, although a random
sample x1 , . . . , xn from the underlying distribution F is available. A version of Theorem 1
suitable for this situation is given in the following corollary.
Corollary 1.1. Let x1 , . . . , xn be any n datapoints in (−∞, ∞), and let
n
1X
x̄ =
xi ,
n i=1
n
1X
ϕ(xi ),
ϕx =
n i=1
n
1X
S =
(xi − x̄)2 .
n i=1
2
Then
inf h(x; x̄)S 2 ≤ ϕx − ϕ(x̄) ≤ sup h(x; x̄)S 2 ,
x∈[a,b]
x∈[a,b]
where a = min{x1 , . . . , xn } and b = max{x1 , . . . , xn }.
Proof. Consider the discrete random variable X with probability distribution P (X = xi ) =
1/n, i = 1, . . . , n. We have E[X] = x̄, E[ϕ(X)] = ϕx , and var(X) = S 2 . Then the corollary
follows from application of Theorem 1.
Lemma 1. If ϕ′ (x) is convex, then h(x; µ) is monotonically increasing in x, and if ϕ′ (x) is
concave, then h(x; µ) is monotonically decreasing in x.
Proof. We prove that h′ (x; µ) ≥ 0 when ϕ′ (x) is convex. The analogous result for concave
ϕ′ (x) follows similarly. Note that
dh(x; µ)
=
dx
ϕ′ (x)+ϕ′ (µ)
− ϕ(x)−ϕ(µ)
2
x−µ
2
1
(x − µ) ,
2
so it suffices to prove
ϕ (x) − ϕ (µ)
ϕ′ (x) + ϕ′ (µ)
≥
.
2
x−µ
Without loss of generality we assume x > µ. Convexity of ϕ′ (x) gives
ϕ′ (y) ≤ϕ′ (µ) +
ϕ′ (x) − ϕ′ (µ)
(y − µ)
x−µ
4
for all y ∈ (µ, x). Therefore we have
Z x
ϕ (x) − ϕ (µ) =
ϕ′ (y) dy
µ
Z x
ϕ′ (x) − ϕ′ (µ)
′
≤
(y − µ) dy
ϕ (µ) +
x−µ
µ
ϕ′ (x) + ϕ′ (µ)
=
(x − µ) .
2
and the result follows.
Lemma 1 makes Theorem 1 easy to use as the follow results hold:
inf h(x; µ) = lim h(x; µ)
x→a
, when ϕ′ (x) is convex
sup h(x; µ) = lim h(x; µ)
x→b
inf h(x; µ) = lim h(x; µ)
x→b
, when ϕ′ (x) is concave.
sup h(x; µ) = lim h(x; µ)
x→a
Note the limits of h(x; µ) can be either finite or infinite. The proof of Lemma 1 borrows
ideas from Bennish (2003). Examples of functions ϕ(x) for which ϕ′ is convex include
ϕ(x) = exp(x) and ϕ(x) = xp for p ≥ 2 or p ∈ (0, 1]. Examples of functions ϕ(x) for which
ϕ′ is concave include ϕ(x) = − log x and ϕ(x) = xp for p < 0 or p ∈ [1, 2].
3
Examples
Example 1 (Moment Generating Function). For any random variable X supported on (a, b)
with a finite variance, we can bound the moment generating function E[etX ] using Theorem
1 to get
inf {h(x; µ)}var(X) ≤ E[etX ] − etE[X] ≤ sup {h(x; µ)}var(X),
x∈(a,b)
x∈(a,b)
where
h(x; µ) =
For t > 0 and (a, b) = (−∞, ∞), we have
tetµ
etx − etµ
−
.
(x − µ)2 x − µ
inf h(x; µ) = lim h(x; µ) = 0 and
x→−∞
5
sup h(x; µ) = lim h(x; µ) = ∞.
x→∞
So Theorem 1 provides no improvement over Jensen’s inequality. However, on a finite
domain such as a non-negative random variable with (a, b) = (0, ∞), a significant improvement in the lower bound is possible because
inf h(x; µ) = h(0; µ) =
1 − etµ + tµetµ
> 0.
µ2
Similar results hold for t < 0. We apply this to an example from Walker (2014), where
X is an exponential random variable with mean 1 and ϕ(x) = etx with t = 1/2. Here the
√
actual Jensen’s gap is E[etX ] − et E[X] = 2 − e ≈ .351. Since var(X) = 1, we have
.176 ≈ h(0; µ) ≤ E[etX ] − et E[X] ≤ lim h(x; µ) = ∞.
x→∞
The less sharp lower bound using inf ϕ′′ (x)/2 is 0.125. Utilizing elaborate approximations
and numerical optimizations Walker (2014) yielded a more accurate lower bound of 0.271.
Example 2 (Arithmetic vs Geometric Mean). Let X be a positive random variable on
interval (a, b) with mean µ. Note that − log(x) is convex whose derivative is concave.
Applying Theorem 1 and Lemma 1 leads to
lim h(x; µ)var(X) ≤ −E{log(X)} + log µ ≤ lim h(x; µ)var(X),
x→a
x→b
where
h(x; µ) =
1
− log x + log µ
+
.
2
(x − µ)
µ(x − µ)
Now consider a sample of n positive data points x1 , . . . , xn . Let x̄ be the arithmetic
1
mean and x̄g = (x1 x2 · · · xn ) n be the geometric mean. Applying Corollary 1.1 gives
exp{S 2 h(b; x̄)} ≤
x̄
≤ exp{S 2 h(a; x̄)},
x̄g
where a, b, S 2 are as defined in Corollary 1.1. To give some numerical results, we generated
100 random numbers from uniform distribution on [10,100]. For these 100 numbers, the
arithmetic mean x̄ is 54.830 and the geometric mean x̄g is 47.509. The above inequality
becomes
1.075 ≤
x̄
1
(x1 x2 · · · xn ) n
= 1.154 ≤ 1.331,
which are fairly tight bounds. Replacing h(xn ; x̄) by ϕ′′ (xn )/2 and h(x1 ; x̄) by ϕ′′ (x1 )/2
leads to a less accurate lower bound 1.0339 and upper bound 21.698.
6
Example 3 (Power Mean). Let X be a positive random variable on a positive interval (a, b)
with mean µ. For any real number s 6= 0, define the power mean as
Ms (X) = (EX s )1/s
Jensen’s inequality establishes that Ms (X) is an increasing function of s. We now give a
sharper inequality by applying Theorem 1. Let r 6= 0, Y = xr , µy = EY , p = s/r and
ϕ(y) = y p . Note that EX s = E{ϕ(Y )}. Applying Theorem 1 leads to
inf h(y; µy )var(Y ) ≤ E[X s ] − (EX r )p ≤ sup h(y; µy )var(Y ),
where
h(y; µy ) =
y p − µpy
pµp−1
y
.
−
2
(y − µy )
y − µy
To apply Lemma 1, note that ϕ′ (y) is convex for p ≥ 2 or p ∈ (0, 1] and is concave for
p < 0 or p ∈ [1, 2] as noted in Section 2.
Applying the above result to the case of r = 1 and s = −1, we have Y = X, p = −1.
Therefore
−1
−1
−1
−1 −1
−1
(EX) + lim h(y; µy )var(X)
≤ EX
≤ (EX) + lim h(y; µy )var(X)
.
y→a
y→b
For the same sequence x1 , . . . , xn generated in Example 2, we have x̄harmonic = 39.113.
Applying Corollary 1.1 leads to
25.337 ≤ x̄harmonic = 39.113 ≤ 48.905.
Note that the upper bound 48.905 is much smaller than the arithmetic mean x̄ = 54.830
by the Jensen’s inequality. Replacing h(b; x̄) by ϕ′′ (b)/2 and h(a; x̄) by ϕ′′ (a)/2 leads to a
less accurate lower bound 0.8298 and 51.0839.
In a recent article published in the American Statistician, de Carvalho (2016) revisited
Kolmogorov’s formulation of generalized mean as
Eϕ (X) = ϕ−1 (E [ϕ(X)]),
(3)
where ϕ is a continuous monotone function with inverse ϕ−1 . The Example 2 corresponds
to ϕ(x) = − log(x) and Example 3 corresponds to ϕ(x) = xs . We can also apply Theorem
1 to bound ϕ−1 (Eϕ(X)) for a more general function ϕ(x).
7
Example 4 (Rao-Blackwell Estimator). Rao-Blackwell theorem (Theorem 7.3.17 in Casella
and Berger, 2002; Theorem 10.42 in Wasserman, 2013) is a basic result in statistical estimation. Let θ̂ be an estimator of θ, L(θ, θ̂) be a loss function convex in θ̂, and T a sufficient
statistic. Then the Rao-Blackwell estimator, θ̂∗ = E[θ̂ | T ], satisifies the following inequality in risk function
E[L(θ, θ̂)] ≥ E[L(θ, θ̂∗ )].
(4)
We can improve this inequality by applying Theorem 1 to ϕ(θ̂) = L(θ, θ̂) with respect to
the conditional distribution of θ̂ given T :
E[L(θ, θ̂) | T ] − L(θ, θ̂∗ ) ≥ inf h(x; θ̂∗ )var(θ̂ | T ),
x∈(a,b)
where function h is defined as in Theorem 1 for ϕ(θ̂) and P (θ̂ ∈ (a, b) | T ) = 1. Further
taking expectations over T gives
∗
E[L(θ, θ̂)] − E[L(θ, θ̂ )] ≥ E
∗
inf h(x; θ̂ )var(θ̂ | T ) .
x∈(a,b)
In particular for square-error loss, L(θ, θ̂) = (θ̂ − θ)2 , we have
h
i
E[(θ − θ̂) ] − E[(θ − θ̂ ) ] = E var(θ̂ | T ) .
2
∗ 2
Using the original Jensen’s inequality only establishes the cruder inequality in Equation
(4).
4
Improved bounds by partitioning
As discussed in Example 1 above, Theorem 1 does not improve on Jensen’s inequality if
inf h(x; µ) = 0. In such cases, we can often sharpen the bounds by partitioning the domain
(a, b) following an approach used in Walker (2014). Let
a = x0 < x1 < · · · < xm = b,
Ij = [xj−1 , xj ), ηj = P (X ∈ Ij ), and µj = E(X | X ∈ Ij ). It follows from the law of total
expectation that
8
E[ϕ(X)] =
=
m
X
j=1
m
X
ηj E[ϕ(X) | X ∈ Ij ]
ηj ϕ(µj ) +
j=1
m
X
j=1
ηj (E[ϕ(X) | X ∈ Ij ] − ϕ(µj )) .
Let Y be a discrete random variable with distribution P (Y = µj ) = ηj , j = 1, 2, . . . , m. It
is easy to see that EY = EX. It follows by Theorem 1 that
m
X
j=1
ηj ϕ(µj ) = E [ϕ (Y )] ≥ ϕ(EY ) +
inf
y∈[µ1 ,µm ]
h(y; µy )var(Y ).
We can also apply Theorem 1 to each E[ϕ(X | X ∈ Ij )] − ϕ(µj ) term:
E[ϕ(X | X ∈ Ij )] − ϕ(µj ) ≥ inf h(x; µj )var(X | X ∈ Ij ).
x∈Ij
Combining the above two equations, we have
E[ϕ(X)] − ϕ(EX) ≥
inf
y∈[µ1 ,µm ]
h(y; µy )var(Y ) +
m
X
j=1
ηj inf h(x; µj )var(X | X ∈ Ij ).
x∈Ij
(5)
Replacing inf by sup in the righthand side gives the upper bound.
The Jensen gap on the left side of (5) is positive if any of the m + 1 terms on the right
is positive. In particular, the Jensen gap is positive if there exists an interval I ⊂ (a, b)
that satisfies inf x∈I ϕ′′ (x) > 0, P (X ∈ I) > 0 and var(X | X ∈ I) > 0. Note that a
finer partition does not necessarily lead to a sharper lower bound in (5). The focus of the
partition should therefore be on isolating the part of interval (a, b) in which ϕ′′ (x) is close
to 0.
Consider example X∼N (µ, σ 2 ) with µ = 0 and σ = 1 and ϕ(x) = ex . We divide
(−∞, ∞) into three intervals with equal probabilities. This gives
Ij
ηj
(−∞, −.431)
1/3
-1.091
0.280
0.000
0.212
(−0.431, 0.431) 1/3
0.000
0.060
0.435
0.580
1.091
0.280
1.209
∞
(0.431, ∞)
1/3
E[X | X ∈ Ij ] var(X | X ∈ Ij )
9
inf x∈Ij h(x; µj ) supx∈Ij h(x; µj )
σ
The actual Jensen gap is eµ+ 2 − eµ = 0.649. The lower bound from (5) is 0.409, which is
a huge improvement over Jensen’s bound of 0. The upper bound ∞, however, provides no
improvement over Theorem 1.
To summarize, this paper proposes a new sharpened version of the Jensen’s inequality.
The proposed bound is simple and insightful, is broadly applicable by imposing minimum
assumptions on ϕ(x), and provides fairly accurate result in spite of its simple form. It can
be incorporated in any calculus-based statistical course.
References
Abramovich, S. and L.-E. Persson (2016). Some new estimates of the Jensen gap. Journal
of Inequalities and Applications 2016 (1), 39.
Abramovich, S., L.-E. Persson, and N. Samko (2014). Some new scales of refined Jensen
and Hardy type inequalities. Mathematical Inequalities & Applications 17 (3), 1105–1114.
Becker, R. A. (2012). The variance drain and Jensen’s inequality. Technical report, CAEPR
Working Paper No. 2012-004. Available at SSRN: https://ssrn.com/abstract=2027471 or
http://dx.doi.org/10.2139/ssrn.2027471.
Bennish, J. (2003). A proof of Jensens inequality. Missouri Journal of Mathematical
Sciences, 15 (1).
Casella, G. and R. L. Berger (2002). Statistical inference, Volume 2. Duxbury Pacific
Grove, CA.
de Carvalho, M. (2016). Mean, what do you mean? The American Statistician 70 (3),
270–274.
Dempster, A. P., N. M. Laird, and D. B. Rubin (1977). Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society. Series B
(methodological), 1–38.
Dragomir, S. (2014). Jensen integral inequality for power series with nonnegative coefficients and applications. RGMIA Res. Rep. Collect 17, 42.
10
Horvath, L., K. A. Khan, and J. Pecaric (2014). Refinement of Jensen’s inequality for
operator convex functions. Advances in inequalities and applications 2014.
Jensen, J. L. W. V. (1906). Sur les fonctions convexes et les inégalités entre les valeurs
moyennes. Acta mathematica 30 (1), 175–193.
Walker, S. G. (2014). On a lower bound for the Jensen inequality. SIAM Journal on
Mathematical Analysis 46 (5), 3151–3157.
Wasserman, L. (2013). All of statistics: a concise course in statistical inference. Springer
Science & Business Media.
11
| 10math.ST
|
Stochastic k-Server: How Should Uber Work?*
Sina Dehghani
∗†
Soheil Ehsani
arXiv:1705.05755v2 [cs.DS] 30 May 2017
Vahid Liaghat
‡
∗†
MohammadTaghi HajiAghayi
Saeed Seddighin
∗†
∗†
Abstract
In this paper we study a stochastic variant of the celebrated k-server problem. In the kserver problem, we are required to minimize the total movement of k servers that are serving an
online sequence of t requests in a metric. In the stochastic setting we are given t independent
distributions hP1 , P2 , . . . , Pt i in advance, and at every time step i a request is drawn from Pi .
Designing the optimal online algorithm in such setting is NP-hard, therefore the emphasis of
our work is on designing an approximately optimal online algorithm. We first show a structural
characterization for a certain class of non-adaptive online algorithms. We prove that in general
metrics, the best of such algorithms has a cost of no worse than three times that of the optimal
online algorithm. Next, we present an integer program that finds the optimal algorithm of
this class for any arbitrary metric. Finally by rounding the solution of the linear relaxation
of this program, we present an online algorithm for the stochastic k-server problem with an
approximation factor of 3 in the line and circle metrics and factor of O(log n) in a general
metric of size n. In this way, we achieve an approximation factor that is independent of k, the
number of servers.
Moreover, we define the Uber problem, motivated by extraordinary growth of online network
transportation services. In the Uber problem, each demand consists of two points -a source and
a destination- in the metric. Serving a demand is to move a server to its source and then to
its destination. The objective is again minimizing the total movement of the k given servers.
We show that given an α-approximation algorithm for the k-server problem, we can obtain an
(α + 2)-approximation algorithm for the Uber problem. Motivated by the fact that demands
are usually highly correlated with the time (e.g. what day of the week or what time of the day
the demand has arrived), we study the stochastic Uber problem. Using our results for stochastic
k-server we can obtain a 5-approximation algorithm for the stochastic Uber problem in line and
circle metrics, and a O(log n)-approximation algorithm for general metrics.
Furthermore, we extend our results to the correlated setting where the probability of a
request arriving at a certain point depends not only on the time step but also on the previously
arrived requests.
1
Introduction
The k-server problem is one of the most fundamental problems in online computation that has
been extensively studied in the past decades. In the k-server problem we have k mobile servers on
a metric space M. We receive an online sequence of t requests where the ith request is a point
ri ∈ M. Upon the arrival of ri , we need to move a server to ri , at a cost equal to the distance
∗
University of Maryland
Supported in part by NSF CAREER award CCF-1053605, NSF BIGDATA grant IIS-1546108, NSF AF:Medium
grant CCF-1161365, DARPA GRAPHS/AFOSR grant FA9550-12-1-0423, and another DARPA SIMPLEX grant.
‡
Facebook
†
from the current position of the server to ri . The goal is to minimize the total cost of serving all
requests.
Manasse, McGeoch, and Sleator [31] introduced the k-server problem as a natural generalization
of several online problems, and a building block for other problems such as the metrical task systems.
They considered the adversarial model, in which the online algorithm has no knowledge of the future
requests. Following the proposition of Sleator and Tarjan [34], they evaluate the performance of an
online algorithm using competitive analysis. In this model, an online algorithm ALG is compared
to an offline optimum algorithm OPT which is aware of the entire input in advance. For a sequence
of requests ρ, let |ALG(ρ)| and |OPT(ρ)| denote the total cost of ALG and OPT for serving ρ. An
algorithm is c-competitive if for every ρ, |ALG(ρ)| ≤ c |OPT(ρ)| + c0 where c0 is independent of ρ.
Manasse et al. [31] showed a lower bound of k for the competitive ratio of any deterministic
algorithm in any metric space with at least k + 1 points. The celebrated k-server conjecture states
that this bound is tight for general metrics. For several years the known upper bounds were all
exponential in k, until a major breakthrough was achieved by Koutsoupias and Papadimitriou [29],
who showed that the so-called work function algorithm is (2k − 1)-competitive. Proving the tight
competitive ratio has been the “holy grail” of the field in the past two decades. This challenge has
led to the study of the problem in special spaces such as the uniform metric (also known as the
paging problem), line, circle, and trees metrics (see [15, 16] and references therein). We also refer
the reader to Section 1.3 for a short survey of randomized algorithms, particularly the recent result
of Bansal, Buchbinder, Madry, and Naor [7] which achieves the competitive ratio of O(log3 n log2 k)
for discrete metrics that comprise n points.
The line metric (or Euclidean 1-dimensional metric space) is of particular interest for developing
new ideas. Chrobak, Karloof, Payne, and Vishwnathan [15] were the first to settle the conjecture
in the line by designing an elegant k-competitive algorithm. Chrobak and Larmore [16] generalized
this approach to tree metrics. Later, Bartal and Koutsoupias [10] proved that the work function
algorithm is also k-competitive in line. Focusing on the special case of k = 2 in line, Bartal et al. [9]
show that, using randomized algorithms, one can break the barrier of lower bound k by giving a
1.98-competitive algorithm for the case where we only have two servers.
Despite the strong lower bounds for the k-server problem, there are heuristics algorithms that
are constant competitive in practice. For example, for the paging problem- the special case of
uniform metric- the least recently used (LRU) strategy is shown to be experimentally constant
competitive (see Section 1.3). In this paper we present an algorithm an run it on real world
data to measure its empirical performance. In particular we use the distribution of car accidents
obtained from road safety data. Our experiments illustrate our algorithm is performing even better
in practice.
The idea of comparing the performance of an online algorithm (with zero-knowledge of the
future) to the request-aware offline optimum has led to crisp and clean solutions. However, that
is not without its downsides. The results in the online model are often very pessimistic leading
to theoretical guarantees that are hardly comparable to experimental results. Indeed, one way to
tighten this gap is to use stochastic information about the input data as we describe in this paper.
We should also point out that the competitive analysis is not the only possible or necessarily the
most suitable approach for this problem. Since the distributions from which the input is generated
are known, one can use dynamic programming (or enumeration of future events) to derive the
optimal movement of servers. Unfortunately, finding such an optimal online solution using the
distributions is an NP-hard problem 1 , thus the dynamic programming or any other approach takes
1
Reduction from k-median to Stochastic k-server: to find the k median of set S of vertices, one can construct an
instance of stochastic k-server with t = 1 and P1 (v) = 1/|S| for every v ∈ S. The best initialization of the servers
2
exponential time. This raises the question that how well one can perform in comparison to the best
online solution. In the rest of the paper we formally define the model and address this question.
A natural and well-motivated generalization of k-server is to assume the demands are two points
instead of just one, consisting of a source and a destination. To serve a demand we need to move a
server to the source and then move it to the destination. We call this problem the Uber problem.
One can see, the Uber problem is the same as k-server when the sources and the destinations are
the same. We also show that, given an α-approximation algorithm for the k-server problem, we
can obtain a (α + 2)-approximation algorithm for the Uber problem. Thus our results for k-server
also apply to the Uber problem.
1.1
The Stochastic Model
In this paper, we study the stochastic k-server problem where the input is not chosen adversarially,
but consists of draws from given probability distributions. This problem has lots of applications
such as network transportations and equipment replacement in data centers. The current mega
data centers contain hundreds of thousands of servers and switches with limited life-span. For
example servers usually retire after at most three years. The only efficient way to scale up the
maintenance in data centers is by automation, and robots are designed to handle maintenance
tasks such as repairs or manual operations on servers. The replacement process can be modeled as
requests that should be satisfied by robots, and robots can be modeled as servers. This problem
also has applications in physical networks. As an example, suppose we model a shopping service
(e.g. Google Express) as a k-server problem in which we receive an online sequence of shopping
requests for different stores. We have k shopping cars (i.e., servers) that can serve the requests by
traveling to the stores. It is quiet natural to assume that on a certain time of the week/day, the
requests arrive from a distribution that can be discovered by analyzing the history. For example,
an Uber request is more likely to be from suburb to midtown in the morning, and from midtown
to suburb at night. We formalize this stochastic information as follows.
For every i ∈ [1 · · · t], a discrete probability distribution Pi is given in advance from which
request ri will be drawn at time step i. The distributions are chosen by the adversary and are
assumed to be independent but not necessarily identical. This model is inspired by the well-studied
model of prophet inequalities 2 [30, 25]. As mentioned before, the case of line metric has proven
to be a very interesting restricted case for studying the k-server problem. In this paper, we focus
mainly on the class of line metric though our results carry over to circle metric and general metrics
as well.
In the adversarial model, the competitive ratio seems to be the only well-defined notion for
analyzing the performance of online algorithms. However, in the presence of stochastic information,
one can derive a much better benchmark that allows us to make fine-grained distinctions between
the online algorithms. We recall that in the offline setting, for a class of algorithms C, the natural
notion to measure the performance of an algorithm ALG ∈ C is the approximation ratio defined
as the worse case ratio of |ALG| to |OPT(C)| where OPT(C) is the optimal algorithm in the class.
In this paper, we also measure the performance of an online algorithm by its approximation ratio–
compared to the optimal online solution. We note that given distributions P1 , . . . , Pt , one can
iteratively compute the optimal online solution by solving the following exponential-size dynamic
gives the optimum solution to k-median of S.
2
In the prophet inequality setting, given (not necessarily identical) distributions P1 , . . . , Pt , an online sequence of
values x1 , . . . , xn where xi is drawn from Pi , an onlooker has to choose one item from the succession of the values,
where xi is revealed at step i. The onlooker can choose a value only at the time of arrival. The goal is to maximize
the chosen value.
3
program: for every i ∈ [0 · · · t] and every possible placement A of k servers (called a configuration)
on the metric, let τ (i, A) denote the minimum expected cost of an online algorithm for serving the
first i requests and then moving the servers to configuration A. Note that τ (i, A) can inductively
be computed via the following recursive formula
τ (i, A) = min τ (i − 1, B) + Eri ∼Pi [min. distance from B to A subject to serving ri ] ,
B
where τ (0, A) is initially zero for every A.
1.2
Our Results
Our first main result is designing a constant approximation algorithm in the line metric when the
distributions for different time steps are not necessarily identical.
Theorem 1.1 There exists a 3-approximation online algorithm for the stochastic k-server problem
in the line metric. The running time is polynomial in k and the sum of the sizes of the supports of
input distributions. The same guarantee holds for the circle metric.
For the general metric, we present an algorithm with a logarithmic approximation guarantee.
Theorem 1.2 There exists a O(log n)-approximation online algorithm for the stochastic k-server
problem in a general metric of size n.
We prove the theorems using two important structural results. The first key ingredient is a
general reduction from class of online algorithms to a restricted class of non-adaptive algorithms
while losing only a constant factor in the approximation ratio. Recall that a configuration is a
placement of k-servers on the metric. We say an algorithm ALG is non-adaptive if it follows the
following procedure: ALG pre-computes a sequence of configurations A0 , A1 , . . . , At . We start by
placing the k-servers on A0 . Upon the arrival of ri , (i) we move the servers to configuration Ai ;
next (ii) we move the closest server s to ri ; and finally (iii) we return s to its original position in
Ai . We first prove the following structural result.
Theorem 1.3 For the stochastic k-server problem in the general metric, the optimal non-adaptive
online algorithm is within 3-approximation of the optimal online algorithm.
Using the aforementioned reduction, we focus on designing the optimal non-adaptive algorithm.
We begin by formulating the problem as an integer program. The second ingredient is to use the
relaxation of this program to formalize a natural fractional variant of the problem. In this variant,
a configuration is a fractional assignment of server mass to the points of the metric such that the
total mass is k. To serve a request at point ri , we need to move some of the mass to have at
least one amount of server mass on ri . The cost of moving the server mass is naturally defined
as the integral of the movement of infinitesimal pieces of the server mass. By solving the linear
relaxation of the integer program, we achieve the optimal fractional non-adaptive algorithm. We
finally prove Theorems 1.1 and 1.2 by leveraging the following rounding techniques. The rounding
method in line has been also observed by Türkoglu [35]. We provide the proof for the case of line
in Section 5 for the sake of completeness. The rounding method for general metrics is via the wellknown embedding of a metric into a distribution of well-separated trees while losing a logarithmic
factor in the distortion. Bansal et al. [7] use a natural rounding method similar to that of Blum,
Burch, and Kalai [12] to show that any fractional k-server movement on well-separated trees can
be rounded to an integral counterpart by losing only a constant factor.
4
Theorem 1.4 (first proven in [35]) Let ALGf denote a fractional k-server algorithm in the
line, or circle. One can use ALGf to derive a randomized integral algorithm ALG such that for every request sequence σ, E [|ALG(σ)|] = |ALGf (σ)|. The expectation is over the internal randomness
of ALG. Furthermore, in the stochastic model ALG can be derandomized.
Theorem 1.5 (proven in [7]) Let ALGf denote a fractional k-server algorithm in any metric.
One can use ALGf to derive a randomized integral algorithm ALG such that for every request
sequence σ, E [|ALG(σ)|] ≤ O(log n) |ALGf (σ)|.
We further show that in the stochastic setting, if the number of possible input scenarios is m,
even if the distributions are correlated, one can compute the best fractional online competitive
algorithm in time polynomial in m and n. Note that since the number of placements of k servers
on n points is exponential, it is not possible to enumerate all the possible choices of an online
algorithm. We solve this problem by presenting a non-trivial LP relaxation of the problem with
size polynomial in n and m; therefore obtaining the following result. We present the formal model
and analysis in Appendix A.
Theorem 1.6 The optimal online algorithm of the stochastic k-server problem with correlated
setting in line and circle can be computed in polynomial time w.r.t. the number of possible scenarios.
In general metrics, an O(log n)-approximation algorithm can be obtained.
We also show that having an α-approximation algorithm for k-server, we can obtain a (α + 2)approximation for the Uber problem, using a simple reduction.
Theorem 1.7 Let ALG denote an α-approximation algorithm for k-server. One can use ALG to
derive a (α + 2)-approximation algorithm for the Uber problem.
Proof. Consider an instance of the Uber problem IU . Let si and ti denote the i-th source and
destination, respectively. We generate an instance of the k-server problem Ik by removing every ti
from IU . In other words the demands are si ’s. We use ALG to provide a solution for IU as follows.
For satisfying the i-th demand, we use ALG to move a server to si . Then using the shortest path
from si to ti , we move that server to ti and then return it back to si . Let OPTU and OPTk denote
the cost of the optimal solutions for IU and Ik , respectively. Let d(si , ti ) denote the distance of ti
from si in the metric. Let C denote the total movement of the servers. We have,
OPTU ≥ OPTk .
X
OPTU ≥
d(si , ti ).
i
C ≤ α OPTk +2
X
d(si , ti ) ≤ (α + 2) OPTU .
i
1.3
Further Related Work
The randomized algorithms often perform much better in the online paradigm. For the k-server
problem, a lower bound of Ω(log k) is shown by [28] for the competitive ratio of randomized algorithms in most common metrics. Despite the exponential gap, compared to the lower bound of
deterministic algorithms, very little is known about the competitiveness of randomized algorithms.
5
In fact, the only known algorithms with competitive ratios below k, work either in the uniform
metric (also known as the paging problem [21, 32, 2, 8]), a metric comprising k + 1 points [23], and
two servers on the line [9]. Two decades after the introduction of the k-server problem, a major
breakthrough was achieved by Bansal et al. [7] in discrete metrics with sub-exponential size. If M
comprise n points, their randomized algorithm achieves a competitive ratio of O(log3 n log2 k).
The case of uniform metric has been extensively studied under various stochastic models motivated by the applications in computer caching. Koutsoupias and Papadimitriou [29] consider two
refinements of the competitive analysis for server problems. First, they consider the diffuse adversary model. In this model, at every step i the adversary chooses a distribution Di over the uniform
metric of the paging problem. Then the ith request is drawn from Di which needs to be served. The
distribution Di is not known to the online algorithm and it may depend on the previous requests.
However, in their paper, they consider the case wherein it is guaranteed that for every point p,
Di (p) ≤ for a small enough ; i.e., the next request is not predictable with absolute certainty for
the adversary. The results of Koutsoupias and Papadimitriou and later Young [36] shows that the
optimum competitive ratio in this setting is close to 1 + Θ(k).
The second refinement introduced in [29] restricts the optimal solution to having lookahead at
most `. Hence, one can define a comparative ratio which indicates the worst-case ratio of the cost
of the best online solution to the best solution with lookahead `. They show that for the k-server
problem, and more generally the metrical task system problem, there are online algorithms that
admit a comparative ratio of 2` + 1; for some instances this ratio is tight.
Various other models of restricting the adversary (access graph model [14, 26, 22], fault rate
model [27, 6, 19], etc) have also been considered for the paging problem (see [33, 11] and references therein for a further survey of these results). Unfortunately, many of the stochastic settings
considered for the paging problem do not seem to have a natural generalization beyond the uniform metric setting. For example, in the diffuse adversary model, most of the studied distributions
do not weaken the adversary in the general metric. In this paper, we look for polynomial-time
approximation algorithms in the class of online algorithms that have access to the distributions.
We would like to mention that various online problems have been previously considered under
prophet inequality model or i.i.d. model (where all distributions are identical). The maximum
matching problem, scheduling, and online network design has been extensively studied in these
models(see e.g. [3, 5, 4, 17, 1, 18]). In the graph connectivity problems, Garg, Gupta, Leonardi,and
Sankowski [24] consider the online variants of Steiner tree and several related problems under the
i.i.d. stochastic model. In the adversarial model, there exists an Ω(log n) lower bound on the
competitive ratio of any online algorithm, where n is the number of demands. However, Garg et
al. show that under the i.i.d. assumption, these problems admit online algorithms with constant
or O(log log n) competitive ratios. We refer the reader to the excellent book by Borodin and ElYaniv [13] for further study of online problems.
2
Preliminaries
In this section we formally define the stochastic k-server problem. The classical k-server problem is
defined on a metric M which consists of points that could be infinitely many. For every two points
x and y in metric M, let d(x, y) denote the distance of x from y which is a symmetric function and
6
satisfies the triangle inequality. More precisely for every three points x, y, and z we have
d(x, x) = 0
(1)
d(x, y) = d(y, x)
(2)
d(x, y) + d(y, z) ≥ d(x, z).
(3)
In the k-server problem the goal is to place k servers on k points of the metric, and move
these servers to satisfy the requests. We refer to every placement of the servers on the metric
points by a configuration. Let ρ = hr1 , r2 , . . . , rt i be a sequence of requests, the goal of the kserver problem is to find configurations hA0 , A1 , A2 , . . . , At i such that for every i there exists a
server on point ri in configuration Ai . We say such a list of configurations
is valid for the given
P
list of requests. A valid sequence of configurations is optimal if
d(Ai−1 , Ai ) is minimized where
d(X, Y ) stands for the minimum cost of moving servers from configuration X to configuration
Y . An optimal sequence hA0 , A1 , . . . , At i of configurations is called an optimal offline solution of
OFKS(M, ρ) when
Pρ is known in advance. We refer to the optimal cost of such movements with
| OFKS(M, ρ)| = d(Ai−1 , Ai ).
We also define the notion of fractional configuration as an assignment of the metric points to
non-negative real numbers. More precisely, each number specifies a mass of fractional server on a
point. Every fractional solution adheres to the following condition: The total sum of the values
assigned to all points is exactly equal to k. Analogously, a fractional configuration serves a request
ri if there is a mass of size at least 1 of server assigned to point ri . An offline fractional solution
of the k-server problem for a given sequence of requests ρ is defined as a sequence of fractional
configurations hA0 , A1 , . . . , At i such that Ai serves ri .
In the online k-server problem, however, we are not given the whole sequence of requests in the
beginning, but we will be informed of every request once its realization is drawn. An algorithm
A is an online algorithm for the k-server problem if it reports a configuration A0 as an initial
configuration and upon realization of every request ri it returns a configuration Ai such that
hA0 , A1 , . . . , Ai i is valid for hr1 , r2 , . . . , ri i. If A is deterministic, it generates a unique sequence of
configurations for every sequence of requests. Let A(M, ρ) be the sequence that A generates for
requests in ρ and |A(M, ρ)| denote its cost.
In the online stochastic k-server problem, in addition to metric M, we are also given t independent probability distributions hP1 , P2 , . . . , Pt i which show the probability that every request ri
is realized on a point of the metric at each time. An algorithm A is an online algorithm for such
a setting, if it generates a configuration for every request ri not solely based on hr1 , r2 , . . . , ri i and
hA0 , A1 , . . . , Ai−1 i but also with respect to the probability distributions. Similarly, we define the
cost of an online algorithm A for a given sequence of requests ρ with |A(M, ρ, hP1 , P2 , . . . , Pt i)|.
We define the expected cost of an algorithm A on metric M and with probability distributions
hP1 , P2 , . . . , Pt i by
|A(M, hP1 , P2 , . . . , Pt i)| = E∀i,ri ∼Pi |A(M, ρ, hP1 , P2 , . . . , Pt i)|.
For every metric M and probability distributions hP1 , P2 , . . . , Pt i we refer to the online algorithm
with the minimum expected cost by OPTM,hP1 ,P2 ,...,Pt i .
An alternative way to represent a solution of the k-server problem is as a vector of configurations
hB0P
, B1 , . . . , Bt i such
P that Bi does not necessarily serve request ri . The cost of such solution is equal
to d(Bi−1 , Bi ) + 2d(Bi , ri ) where d(Bi , ri ) is the minimum distance of a server in configuration
Bi to request ri . The additional cost of 2d(Bi , ri ) can be thought of as moving a server from Bi to
serve ri and returning it back to its original position. Thus, every such representation of a solution
7
can be transformed to the other representation. Similarly, d(Bi , ri ) for a fractional configuration
Bi is the minimum cost which is incurred by placing a mass 1 of server at point ri . We use letter
B for the configurations of such solutions throughout the paper.
In this paper the emphasis is on the stochastic k-server problem on the line metric. We define
the line metric L as a metric of points from −∞ to +∞ such that the distance of two points x
and y is always equal to |x − y|. Moreover, we show that deterministic algorithms are as powerful
as randomized algorithms in this setting, therefore we only focus on deterministic algorithms in
this paper. Thus, from here on, we omit the term deterministic and every time we use the word
algorithm we mean a deterministic algorithm unless otherwise is explicitly mentioned.
3
Structural Characterization
Recall that an online algorithm A has to fulfill the task of reporting a configuration Ai upon
arrival of request ri based on hA0 , A1 , . . . , Ai−1 i, hr1 , r2 , . . . , ri i, and hP1 , P2 , . . . , Pt i. We say an
algorithm B is request oblivious, if it reports configuration Bi regardless of request ri . As such,
B generates configurations
t i for a sequence of requests hr1 , r2 , . . . , rt i and the cost
P hB0 , B1 , . . . , BP
of such configuration is
d(Bi−1 , Bi ) + 2d(Bi , ri ). More precisely, no matter what request ri
is, B will generate the same configuration for a given list of past configurations hB0 , B1 , . . . , Bi−1 i,
a given sequence of past requests hr1 , r2 , . . . , ri−1 i, and the sequence of probability distributions
hP1 , P2 , . . . , Pt i. In the following we show that every online algorithm A can turn into a request
oblivious algorithm BA that has a cost of at most |3A(M, ρ, hP1 , P2 , . . . , Pt i)| for a given sequence
of requests ρ.
Lemma 3.1 Let A be an online algorithm for the stochastic k-server problem. For any metric M,
there exists a request oblivious algorithm BA such that
|BA (M, hP1 , P2 , . . . , Pt i)| ≤ 3|A(M, hP1 , P2 , . . . , Pt i)|.
Proof. Let ρ be a sequence of requests. We define online algorithm BA as follows: The configuration that BA reports for a given list of input arguments hB0 , B1 , . . . , Bi i, hr1 , r2 , . . . , ri i,
and hP1 , P2 , . . . , Pt i is the output of algorithm A on inputs hB0 , B1 , . . . , Bi i, hr1 , r2 , . . . , ri−1 i, and
hP1 , P2 , . . . , Pt i (The same input except that ri is dropped from the sequence of requests). We show
the cost of such algorithm for input ρ is at most 3 times the cost of A for the same input.
Let hA0 , A1 , . . . , At i be the sequence of configurations that A generates for requests ρ and
hB0 , B1 , . . . , Bt i be the output of algorithm BA . According to the construction of BA , B0 = A0 and
Bi = Ai−1 for all 1 ≤ i ≤ t. Note that for algorithm A, we assume
P every Ai serves
P request ri . By
definition, the cost of solution hB0 , B1 , B2 , . . . , Bt i is equal to
d(Bi−1 , Bi ) + 2 d(Bi , ri ). Since
B0 = B1 = A0 and Bi = Ai−1 ,
t
X
d(Bi−1 , Bi ) =
i=1
t−1
X
d(Ai−1 , Ai ) ≤
i=1
t
X
d(Ai−1 , Ai ) = |A(M, ρ, hP1 , P2 , . . . , Pt i)|.
(4)
i=1
Moreover, since every Ai servers request ri , d(Bi , ri ) ≤ d(Bi , Ai ) = d(Ai−1 , Ai ). Hence,
2
t
X
i=1
d(Bi , ri ) ≤ 2
t
X
i=1
d(Bi , Ai ) = 2
t
X
d(Ai−1 , Ai ) = 2|A(M, ρ, hP1 , P2 , . . . , Pt i)|.
i=1
Inequality (4) along with Equation (5) implies
|BA (M, ρ, hP1 , P2 , . . . , Pt i)| ≤ 3|A(M, ρ, hP1 , P2 , . . . , Pt i)|.
8
(5)
Since this holds for all requests ρ ∼ hP1 , P2 , . . . , Pt i, we have
|BA (M, hP1 , P2 , . . . , Pt i)| ≤ 3|A(M, hP1 , P2 , . . . , Pt i)|
and the proof is complete.
An immediate corollary of Lemma 3.1 is that the optimal request oblivious algorithm has a cost
of at most |3 OPTM,hP1 ,P2 ,...,Pt i (M, hP1 , P2 , . . . , Pt i)|. Therefore, if we only focus on the request
oblivious algorithms, we only lose a factor of 3 in comparison to the optimal online algorithm. The
following lemma states a key structural lemma for an optimal request oblivious algorithm.
Lemma 3.2 For every request oblivious algorithm B, there exists a randomized request oblivious
algorithm B 0 with the same expected cost which is not only oblivious to the last request, but also
oblivious to all requests that have come prior to this.
Proof. For any given request oblivious online algorithm B, we construct an online algorithm B 0 which is oblivious to all of the requests as follows: For an input hB1 , B2 , . . . , Bi−1 i
of configurations and probability distributions hP1 , P2 , . . . , Pt i, draw a sequence of requests
hr1 , r2 , . . . , ri i from hP1 , P2 , . . . , Pt i conditioned on the constraint that B would generate configurations hB1 , B2 , . . . , Bi−1 i for requests hr1 , r2 , . . . , ri−1 i. Now, report the output of B for inputs
hB1 , B2 , . . . , Bi−1 i, hr1 , r2 , . . . , ri i, and hP1 , P2 , . . . , Pt i.
We define the cost of step i of algorithm B 0 as d(Bi−1 , Bi ) + 2d(Bi , ri ). Due to the construction
of algorithm B 0 , the expected cost of this algorithm at every step i for a random sequence of
requests is equal to the expected cost of algorithm B for a random sequence of requests drawn from
hP1 , P2 , . . . , Pt i. Therefore, the expected cost of both algorithms for a random sequence of requests
are equal and thus |B(M, hP1 , P2 , . . . , Pt i)| = |B 0 (M, hP1 , P2 , . . . , Pt i)|.
Lemma 3.2 states that there always exists an optimal randomized request oblivious online
algorithm that returns the configurations regardless of the requests. We call such an algorithm
non-adaptive. Since a non-adaptive algorithm is indifferent to the sequence of the requests, we can
assume it always generates a sequence of configurations just based on the distributions. For an
optimal of such algorithms, all such sequence of configurations should be optimal as well. Therefore,
there always exists an optimal non-adaptive online algorithm which is deterministic. By Lemma
3.1 not only do we know the optimal request oblivious algorithm is at most 3-approximation, but
also the same holds for the optimal non-adaptive algorithm.
Theorem 3.3 There exists a sequence of configurations hB0 , B1 , . . . , Bt i such that an online algorithm which starts with B0 and always returns configuration Bi upon arrival of request ri has an
opproximation factor of at most 3.
4
Fractional Solutions
In this section we provide a fractional online algorithm for the k-server problem that can be implemented in polynomial time. Note that by Theorem 3.3 we know that there exist configurations
hB1 , B2 , . . . , Bt i such that the expected cost of a non-adaptive algorithm that always returns these
configurations is at most 3 times the cost of an optimal online algorithm. Therefore, we write an
integer program to find such configurations with the least expected cost. Next, we provide a relaxed
LP of the integer program and show that every feasible solution of such LP corresponds to a fractional online algorithm for the stochastic k-server problem. Hence, solving such a linear program,
that can be done in polynomial time, gives us a fractional online algorithm for the problem.
9
4.1
Linear Program
Recall that given t independent distributions hP1 , . . . , Pt i for online stochastic k-server, an adaptive
algorithm can be represented by t + 1 configurations hB0 , . . . , Bt i. Upon the arrival of each request
ri , we move the servers from configuration Bi−1 to Bi and then one server serves ri and goes back
to its position in Bi . The objective is to find the configurations such that the cost of moving to
new configurations in addition to the expected cost of serving the requests is minimized. Therefore
the problem can formulated in an offline manner. First we provide an integer program in order to
find a vector of configurations with the least cost.
The decision variables of the program represent the configurations, the movement of servers
from one configuration to another, and the way that each possible request is served. In particular,
at each time step τ :
• For each node v there is a variable bτ,v ∈ N denoting the number of servers on node v.
• For each pair of nodes u and v, there is a movement variable fτ,u,v ∈ N denoting the number
of servers going from u to v for the next round.
• For each node v and possible request node r, there is a variable xτ,v,r ∈ {0, 1} denoting
whether r is served by v or not.
In the following integer program, the first set of constraints ensures the number of servers on
nodes at each time is updated correctly according to the movement variables. The second set of
constraints ensures that each possible request is served by at least one server. The third set of
constraints ensures that no possible request is served
PtBy the definition, the
Pt by an empty node.
cost of a sequence of configurations hB0 , . . . , Bt i is i=1 d(Bi−1 , Bi ) + 2 i=1 d(Bi , ri ). Thus the
objective is to minimize the expression
XX
XXX
fτ,u,v d(u, v) + 2
xτ,v,r Pr(z ∼ Pτ = r)d(v, r)
τ
u,v
τ
v
r
, where Pr(z ∼ Pτ = r) denotes the probability that r is requested at time τ .
XX
XXX
min.
fτ,u,v d(u, v) + 2
xτ,v,r Pr(z ∼ Pτ = r)d(v, r)
τ
∀τ, v
u,v
bτ +1,v = bτ,v +
X
τ
v
fτ,u,v −
X
u
∀τ, u, v
X
r
fτ,v,u .
u
xτ,v,r ≥ 1.
v
∀τ, v, r
∀τ
xτ,v,r ≤ bτ,v .
X
bτ,v ≤ k.
∀τ, v, r
xτ,v,r ∈ {0, 1}.
∀τ, u, v
fτ,u,v ∈ N.
∀τ, v
bτ,v ∈ N.
v
Now we consider the following relaxation of the above integer program.
10
min.
XX
τ
∀τ, v
fτ,u,v d(u, v) + 2
u,v
bτ +1,v = bτ,v +
X
XXX
τ
v
fτ,u,v −
X
u
∀τ, u, v
X
xτ,v,r Pr(z ∼ Pτ = r)d(v, r)
r
fτ,v,u .
u
xτ,v,r ≥ 1.
v
∀τ, v, r
∀τ
xτ,v,r ≤ bτ,v .
X
bτ,v ≤ k.
v
5
Reduction from Integral k-server to Fractional k-server
In this section we show how we can obtain an integral algorithm for the stochastic k-server problem
from a fractional algorithm. We first show that every fractional algorithm for the line metric can
be modified to an integral algorithm with the same cost. Next, we study the problem on HST
metrics; we give a rounding method that produces an integral algorithm from a fractional algorithm
while losing a constant factor. Finally, we leverage the previously known embedding techniques to
show every metric can be embedded into HST’s with a distortion of at most O(log n). This will
lead to a rounding method for obtaining an integral algorithm from every fractional algorithm on
general metrics while losing a factor of at most O(log n). Combining this with the 3 approximation
fractional algorithm that we provide in Section 4, we achieve an O(log n) approximation algorithm
for the stochastic k-server problem on general graphs.
5.1
Integrals Are as Strong as Fractionals On the Line
In this section we show every fractional algorithm on the line metric can be derandomized to an
integral solution with the same expected cost. The rounding method is as follows: For every
fractional configuration A, we provide an integral configuration I(A) such that (i) the distance of
two configurations A1 and A2 is equal to the expected distance of two configurations I(A1 ) and
I(A2 ). (ii) for every point x in the metric that A has a server mass of size at least 1 on x, there
exists a server on point x in I(A).
Let for every point x in the metric, A(v) denote the amount of server mass on node v of the
line. For every fractional configuration B, we define a mass function
fA : (0, k] →
as follows.
P
PV
j
fA (x) = vj if and only if j is the minimum integer such that j−1
A(i)
<
x
and
i=1
i=1 A(i) ≥ x.
Intuitively, if one gathers the server mass by sweeping the line from left to right, fA (x) is the first
position on which we have gathered x amount of server mass. The rounding algorithm is as follows:
• Pick a random real number r in the interval [0, 1).
• I(A) contains k servers on positions fA (r), fA (r + 1), fA (r + 2), . . . , fA (r + k − 1).
Note that the rounding method uses the same r for all of the configurations. More precisely,
we draw r from [0, 1) at first and use this number to construct the integral configurations from
fractional configurations. The following two lemmas show that both of the properties hold for the
rounding algorithm we proposed.
Lemma 5.1 Let A be a fractional configuration and x be a point such that A(x) ≥ 1. Then I(A)
has a server on x.
11
Proof. Due to the construction of our rounding method, for every two consecutive servers a and
b in I(A), the total mass of servers after a and before b in the fractional solution is less than 1.
Therefore, I(A) should put a server on point x, otherwise the total mass of servers in the fractional
solution between the first server before x and the first server after x would be at least 1.
The next lemma shows that the rounding preserves the distances between the configurations in
expectation.
Lemma 5.2 Let A1 and A2 be two fractional configurations and |A1 − A2 | be their distance. The
following holds for the distances of the configurations
E| I(A1 ) − I(A2 )| = |A1 − A2 |.
Proof. The key point behind the proof of this lemma is that the distance of two fractional
configurations A1 and A2 can be formulated as follows
Z 1
|A1 − A2 | =
| Iω (A1 ) − Iω (A2 )|dω
0
where Iω (A) stands for an integral configurations which places the servers on points fA (ω), fA (ω+1),
fA (ω + 2), . . ., fA (ω + k − 1). Since at the beginning of the rounding method we draw r uniformly
at random, the expected distance of the two rounded configurations is exactly equal to
Z 1
| Iω (A1 ) − Iω (A2 )|dω
0
which is equal to the distance of A1 from A2 .
Theorem 5.3 For any given fractional online algorithm A for the k-server problem on the line
metric, there exists an online integral solution for the same problem with the same expected cost.
5.2
Reduction for General Graphs
An HST is a undirected rooted tree in which every leaf represents a point in the metric and the
distance of a pair of points in the metric is equal to the distance of the corresponding leaves in the
tree. In an HST, weights of the edges are uniquely determined by the depth of the vertices they
connect. More precisely, in a σ-HST the weight of an edges between a vertex v and its children is
equal to σ h−dv where h stands for the height of the tree and dv denotes the depth of vertex v.
Since HSTs are very well structured, designing algorithms on HSTs is relatively easier in comparison to a more complex metric. Therefore, a classic method for alleviating the complexity of the
problems is to first embed the metrics into HSTs with a low distortion and then solve the problems
on these trees.
Perhaps the most important property of the HSTs is the following:
Observation 5.1 For every pair of leaves u, v ∈ T of an HST, the distance of u and v is uniquely
determined by the depth of their deepest common ancestor.
Note that, the higher the depth of the common ancestor is, the lower the distance of the leaves will
be. Therefore, the closest leaves to a leaf v are the ones that share the most common ancestors with
v. Bansal et al. propose a method for rounding every fractional solution of the k-server problem
to an integral solution losing at most a constant factor [7].
12
Theorem 5.4 [7] Let T be a σ-HST with n leaves, σ > 5, and let A = hA0 , A1 , A2 , . . . , At i be a
sequence of fractional configurations. There is an online procedure that maintains a sequence of
randomized k-server configurations S = hS0 , S1 , S2 , . . . , St i satisfying the following two properties:
• At any time i, the state Si is consistent with the fractional state Ai .
• If the fractional state changes from xi−1 to xi at time i, incurring a movement cost of ci , then
the state Si−1 can be modified to a state Si while incurring a cost of O(ci ) in expectation.
Embedding general metrics into trees and in particular HSTs has been the subject of many
studies. The seminal work of Fakcharoenphol et al. [20] has shown that any metric can be randomly
log n
embedded to σ-HSTs with distortion O( σlog
σ ).
Theorem 5.5 [20] There exists a probabilistic method to embed an arbitrary metric M into σlog n
HSTs with distortion σlog
σ .
Therefore, to round a fractional solution on a general metric, we first embed it into 6-HSTs with a
distortion of at most O(log n) and then round the solution while losing only a constant factor. This
will give us an integral algorithm that has an expected cost of at most O(log n) times the optimal.
Theorem 5.6 For any given fractional online algorithm A for the k-server problem on an arbitrary
metric, there exists an online integral solution for the same problem having a cost of no worse that
O(log n) times the cost of A in expectation.
6
Acknowledgment
We would like to thank Shi Li for having helpful discussions.
References
[1] M. Abolhasani, S. Ehsani, H. Esfandiari, M. Hajiaghayi, R. Kleinberg, and B. Lucier. Beating
1 − 1/e for ordered prophets. arXiv preprint arXiv:1704.05836, 2017.
[2] D. Achlioptas, M. Chrobak, and J. Noga. Competitive analysis of randomized paging algorithms. Theoretical Computer Science, 234(1):203–218, 2000.
[3] S. Alaei, M. Hajiaghayi, and V. Liaghat. Online prophet-inequality matching with applications
to ad allocation. In Proceedings of the 13th ACM Conference on Electronic Commerce, pages
18–35. ACM, 2012.
[4] S. Alaei, M. Hajiaghayi, and V. Liaghat. The online stochastic generalized assignment problem. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and
Techniques, pages 11–25. Springer, 2013.
[5] S. Alaei, M. T. Hajiaghayi, V. Liaghat, D. Pei, and B. Saha. Adcell: Ad allocation in cellular
networks. In Algorithms–ESA 2011, pages 311–322. Springer, 2011.
[6] S. Albers, L. M. Favrholdt, and O. Giel. On paging with locality of reference. In Proceedings
of the thiry-fourth annual ACM symposium on Theory of computing, pages 258–267. ACM,
2002.
13
[7] N. Bansal, N. Buchbinder, A. Madry, and J. Naor. A polylogarithmic-competitive algorithm
for the k-server problem. 2011.
[8] N. Bansal, N. Buchbinder, and J. S. Naor. A primal-dual randomized algorithm for weighted
paging. Journal of the ACM (JACM), 59(4):19, 2012.
[9] Y. Bartal, M. Chrobak, and L. L. Larmore. A randomized algorithm for two servers on the
line. Information and Computation, 158(1):53–69, 2000.
[10] Y. Bartal and E. Koutsoupias. On the competitive ratio of the work function algorithm for
the k-server problem. Theoretical computer science, 324(2):337–345, 2004.
[11] L. Becchetti. Modeling locality: A probabilistic analysis of lru and fwf. In Algorithms–ESA
2004, pages 98–109. Springer, 2004.
[12] A. Blum, C. Burch, and A. Kalai. Finely-competitive paging. In Foundations of Computer
Science, 1999. 40th Annual Symposium on, 1999.
[13] A. Borodin and R. El-Yaniv. Online computation and competitive analysis. cambridge university press, 2005.
[14] A. Borodin, S. Irani, P. Raghavan, and B. Schieber. Competitive paging with locality of
reference. Journal of Computer and System Sciences, 50(2):244–258, 1995.
[15] M. Chrobak, H. Karloff, T. Payne, and S. Vishwnathan. New ressults on server problems.
SIAM Journal on Discrete Mathematics, 4(2):172–181, 1991.
[16] M. Chrobak and L. L. Larmore. An optimal on-line algorithm for k servers on trees. SIAM
Journal on Computing, 20(1):144–148, 1991.
[17] S. Dehghani, S. Ehsani, M. Hajiaghayi, V. Liaghat, and S. Seddighin. Online survivable
network design and prophets. 2015.
[18] S. Dehghani, I. A. Kash, and P. Key. Online stochastic scheduling and pricing the clouds.
2017.
[19] P. J. Denning. The working set model for program behavior. Communications of the ACM,
26(1):43–48, 1983.
[20] J. Fakcharoenphol, S. Rao, and K. Talwar. A tight bound on approximating arbitrary metrics by tree metrics. In Proceedings of the thirty-fifth annual ACM symposium on Theory of
computing, pages 448–455. ACM, 2003.
[21] A. Fiat, R. M. Karp, M. Luby, L. A. McGeoch, D. D. Sleator, and N. E. Young. Competitive
paging algorithms. Journal of Algorithms, 12(4):685–699, 1991.
[22] A. Fiat and M. Mendel. Truly online paging with locality of reference. In Foundations of
Computer Science, 1997. Proceedings., 38th Annual Symposium on, pages 326–335. IEEE,
1997.
[23] A. Fiat and M. Mendel. Better algorithms for unfair metrical task systems and applications.
SIAM Journal on Computing, 32(6):1403–1422, 2003.
14
[24] N. Garg, A. Gupta, S. Leonardi, and P. Sankowski. Stochastic analyses for online combinatorial
optimization problems. In Proceedings of the nineteenth annual ACM-SIAM symposium on
Discrete algorithms, pages 942–951. Society for Industrial and Applied Mathematics, 2008.
[25] M. T. Hajiaghayi, R. Kleinberg, and T. Sandholm. Automated online mechanism design and
prophet inequalities. In AAAI, volume 7, pages 58–65, 2007.
[26] S. Irani, A. R. Karlin, and S. Phillips. Strongly competitive algorithms for paging with locality
of reference. SIAM Journal on Computing, 25(3):477–497, 1996.
[27] A. R. Karlin, S. J. Phillips, and P. Raghavan. Markov paging. SIAM Journal on Computing,
30(3):906–922, 2000.
[28] H. Karloff, Y. Rabani, and Y. Ravid. Lower bounds for randomized k-server and motionplanning algorithms. SIAM Journal on Computing, 23(2):293–312, 1994.
[29] E. Koutsoupias and C. H. Papadimitriou. On the k-server conjecture. Journal of the ACM
(JACM), 42(5):971–983, 1995.
[30] U. Krengel, L. Sucheston, et al. Semiamarts and finite values. Bull. Amer. Math. Soc, 83(4),
1977.
[31] M. S. Manasse, L. A. McGeoch, and D. D. Sleator. Competitive algorithms for server problems.
Journal of Algorithms, 11(2):208–230, 1990.
[32] L. A. McGeoch and D. D. Sleator. A strongly competitive randomized paging algorithm.
Algorithmica, 6(1-6):816–825, 1991.
[33] K. Panagiotou and A. Souza. On adequate performance measures for paging. In Proceedings
of the thirty-eighth annual ACM symposium on Theory of computing, pages 487–496. ACM,
2006.
[34] D. D. Sleator and R. E. Tarjan. Amortized efficiency of list update and paging rules. Communications of the ACM, 28(2):202–208, 1985.
[35] D. Türkoglu. The k-Server Problem and Fractional Analysis. PhD thesis, Masters Thesis,
The University of Chicago, 2005. http://people. cs. uchicago. edu/ duru/papers/masters. pdf,
2005.
[36] N. E. Young. Bounding the diffuse adversary. In SODA, volume 98, pages 420–425, 1998.
15
A
Correlated Setting
In this section, we study the k-server problem when the probability distributions are not independent. Recall that in the independent setting the sequence of requests is referred to by
ρ = hr1 , . . . , rt i. In the correlated model we assume all different possibilities for ρ have been
given in the form of a set R = {ρ1 , . . . , ρm } of m sequences ρi = hri,1 , . . . , ri,t i. Moreover, we
assume the probability of each scenario ρi is denoted by pi and given in advance. Given the list of
different scenarios and probabilities, the goal is to design an online algorithm to serve each request
ri,j prior to arrival of the next request such that the overall movement of the servers is minimized.
We model this problem by an integer program. We first write an integer program and show
that every solution of this program is uniquely mapped to a deterministic online algorithms for the
problem. Moreover, every online algorithm can be mapped to a feasible solution of the program.
More precisely, each solution of the program is equivalent to an online algorithm of the problem.
Furthermore, we show how to derive an online algorithm from the solution of the integer program.
These two imply that the optimal deterministic online algorithm can be obtained from the optimal
solution of the program.
A.1
Program
To better convey the idea behind the integer program, we first introduce the tree T which is a
trie containing all sequences ρ1 to ρm . Let us use w(v) to denote the path from the root to a
node v. With these notations, a node v ∈ T represents a request which may occur conditioning
all requests in w(v) occur beforehand. Besides, every leaf of T uniquely represents one of the ρi ’s.
Let us use l(v) to denote the set of those indices i for which ρi is a leaf of the subtree of v. At
each step t, only those ρi ’s can be a final option for R that hρi,1 , . . . , ρi,t i = hr1 , . . . , rt i. Hence, a
new request rt can be informative since we know that none of the ρi ’s in l(rt−1 )\l(rt au) will occur
anymore.PFor a node v we define P r(v) as the probability of all requests in w(v) happening i.e.
P r(v) = i∈l(v) P r(R = ρi ).
We extend the tree T by adding k − 1 additional nodes. As shown in Figure 1, these nodes
form a path leading to the root of T . These nodes plus the root represent the initial configuration
of the k servers. Let us call these nodes the initial set I. Now we can show the movement of the
servers in our metric space by means of k tokens in T . To do so, we begin with putting one token
on each of the k nodes of I. Each token corresponds to one of the servers. After a server moves to
serve a request rt , we move its corresponding token to a node of T which represents the request rt .
Note that at this step, there is no discrimination between any of the sequences in l(rt ) in terms of
occurrence. This causes a deterministic online algorithm A to serve the first |w(rt )| requests of R
in the same way if R is going to be one of ρi ’s (i ∈ l(rt )). A result of this uniquely serving is that
we can use some downward links on T in order to show how each request v gets served. In the next
paragraphs we explain about these links and how we construct the integer program.
Let us use xu,v to denote a link from a node u ∈ T to its descendant v. xu,v is one if and only if
A uses the same server to serve u and then v without using that server to serve any other request
between u and v. This consecutive serving may occur with probability P r(v) = P r(u)P r(v|u). In
this case, the algorithm moves a server from u to v and pays |u − v| as the distance cost between
the two points of the metric space corresponding to u and v.
There are two conditions for these links that we must care about. First, since each request v
should be served with a server, at least one of the xu,v ’s should be one for all u in w(v). Without
loss of generality, we assume this is exactly one of them, i.e. there is no need to serve a request
with more than one server. Second, after serving a request u, a server can go for serving at most
16
one other request. That is, for each i ∈ l(u), there should be at most one v ∈ ρi such that xu,v = 1.
This condition guarantees that in serving the sequence of requests R, a server which serves rt1 ∈ R
has always at most one other request rt2 ∈ R as the next serving request.
The following integer program maintains both conditions for xu,v ’s and has the expected overall
movement of all servers as the objective function:
X
P r(v)|u − v|xu,v
min.
u,v∈T ;u∈w(v)
∀v ∈ T \I
X
xu,v = 1.
u∈w(v)
∀u ∈ T, i ∈ l(u)
X
xu,v ≤ 1.
v∈ρi
∀u, v ∈ T, u ∈ w(v) xu,v ∈ {0, 1}
Next, we can relax the constraints of the program to make it linear. Therefore, instead of
assigning either {0} or {1}, to each xu,v we let it be a real number between 0 and 1. Thus, the
integer program turns to the following linear program with the same objective function but more
relaxed constraints.
X
min.
P r(v)|u − v|xu,v
u,v∈T ;u∈w(v)
∀v ∈ T \I
X
xu,v = 1.
u∈w(v)
∀u ∈ T, i ∈ l(u)
X
xu,v ≤ 1.
v∈ρi
∀u, v ∈ T, u ∈ w(v) xu,v ≤ 1
∀u, v ∈ T, u ∈ w(v) xu,v ≥ 0
Note that every feasible solution of the linear program is corresponding to a fractional solution of
the problem. Since the optimal solution of the linear program can be found in polynomial time,
using the rounding methods presented in Section 5 we obtain an optimal online algorithm for the
line metric and a O(log n) approximation algorithm for general metrics as stated in Theorem 1.6.
B
Experimental Results
The goal of this section is to make an evaluation of our method for the line on a real world data set.
The line can be an appropriate model for a plenty of applications. For example, it could be sending
road maintenance trucks to different points of a road or sending emergency vehicles to accident
scenes along a highway. For this experiment, we take the case of car accidents.
Data sets. We use Road Safety Data3 to find the distribution of the accidents along the A14
road in Great Britain. In 2015, over 1600 accidents occurred on this highway, with an average of
140 accidents per month. We assume a point every 10 miles along the highway. That is 40 points
in total. Then we build the distributions with respect to how the accidents are spread over the
days of month. In this way, we achieve 30 distributions for 40 points along the line.
3
4
https://data.gov.uk/dataset/road-accidents-safety-data/
https://en.wikipedia.org/wiki/A1 road (Great Britain)
17
Number of Servers
Algorithm
Optimum
2
6.5
0.2
3
7.6
0.8
4
6.7
3.1
5
7.1
8.4
6
7.5
29.4
7
8.3
57.9
8
8.5
126.3
9
8.4
406.7
10
9.3
1477.1
11
8.2
6173.6
Table 1: The running time of our algorithm and the optimum algorithm in seconds. For higher
number of servers, the optimum solution was not calculable within 5 hours.
Algorithms. We compare the performance of our method to that of the optimum algorithm.
To find the optimum solution we use backtracking. The running time of the algorithm is exponential
to k. However, we use techniques such as branch and bound and exponential dynamic programming
to get a fast implementation.
Results. We run different experiments with k from 2 to 11 on the line and distributions
explained above. In previous sections we showed an upper bound of 3 for the approximation factor
of our algorithm. Interestingly, in these experiments we can observe a better performance as shown
by Figure 1. We compare the running time of the algorithms in Table 1. Note that the size of
our LP our method solvers does not vary by k. This is in fact the reason behind why its running
time remains almost the same. In contrast, the running time of the optimum algorithm grows
exponentially.
Figure 1: Performance of our algorithm compared to the optimum. The dashed curve indicates
two times the optimum.
18
| 8cs.DS
|
Learning-Theoretic Foundations of Algorithm Configuration for
Combinatorial Partitioning Problems∗
Maria-Florina Balcan
Vaishnavh Nagarajan
Ellen Vitercik
Colin White
arXiv:1611.04535v3 [cs.DS] 17 May 2017
May 18, 2017
Abstract
Max-cut, clustering, and many other partitioning problems that are of significant importance to machine learning and other scientific fields are NP-hard, a reality that has motivated
researchers to develop a wealth of approximation algorithms and heuristics. Although the best
algorithm to use typically depends on the specific application domain, a worst-case analysis is
often used to compare algorithms. This may be misleading if worst-case instances occur infrequently, and thus there is a demand for optimization methods which return the algorithm
configuration best suited for the given application’s typical inputs. We address this problem
for clustering, max-cut, and other partitioning problems, such as integer quadratic programming, by designing computationally efficient and sample efficient learning algorithms which
receive samples from an application-specific distribution over problem instances and learn a
partitioning algorithm with high expected performance. Our algorithms learn over common
integer quadratic programming and clustering algorithm families: SDP rounding algorithms
and agglomerative clustering algorithms with dynamic programming. For our sample complexity analysis, we provide tight bounds on the pseudodimension of these algorithm classes, and
show that surprisingly, even for classes of algorithms parameterized by a single parameter, the
pseudo-dimension is superconstant. In this way, our work both contributes to the foundations
of algorithm configuration and pushes the boundaries of learning theory, since the algorithm
classes we analyze consist of multi-stage optimization procedures and are significantly more
complex than classes typically studied in learning theory.
∗
Authors’ addresses: {ninamf,vaishnavh,vitercik,crwhite}@cs.cmu.edu.
1
Introduction
NP-hard problems arise in a variety of diverse and oftentimes unrelated application domains. For
example, clustering is a widely-studied NP-hard problem in unsupervised machine learning, used to
group protein sequences by function, organize documents in databases by subject, and choose the
best locations for fire stations in a city. Although the underlying objective is the same, a “typical
problem instance” in one setting may be significantly different from that in another, causing approximation algorithms to have inconsistent performance across the different application domains.
We study how to characterize which algorithms are best for which contexts, a task often referred to in the AI literature as algorithm configuration. This line of work allows researchers
to compare algorithms according to an application-specific metric, such as expected performance
over their problem domain, rather than a worst-case analysis. If worst-case instances occur infrequently in the application domain, then a worst-case algorithm comparison could be uninformative
and misleading. We approach application-specific algorithm configuration via a learning-theoretic
framework wherein an application domain is modeled as a distribution over problem instances. We
then fix an infinite class of approximation algorithms for that problem and design computationally
efficient and sample efficient algorithms which learn the approximation algorithm with the best
performance over the distribution, and therefore an algorithm with high performance in the specific application domain. Gupta and Roughgarden [24] introduced this learning framework to the
theory community, but it has been the primary model for algorithm configuration and portfolio
selection in the artificial intelligence community for decades [33] and has led to breakthroughs in
diverse fields including combinatorial auctions [28], scientific computing [16], vehicle routing [12],
and SAT [42].
In this framework, we study two important, infinite algorithm classes. First, we analyze approximation algorithms based on semidefinite programming (SDP) relaxations and randomized
rounding procedures, which are used to approximate integer quadratic programs (IQPs). These
algorithms can be used to find a nearly optimal solution to a variety of combinatorial partitioning
problems, including the seminal max-cut and max 2-SAT problems. Second, we study agglomerative clustering algorithms followed by a dynamic programming step to extract a good clustering.
These techniques are widely used in machine learning and across many scientific disciplines for data
analysis. We begin with a concrete problem description.
Problem description. In this learning framework, we fix a computational problem, such as
max-cut or k-means clustering, and assume that there exists an unknown, application-specific
distribution D over a set of problem instances Π. We denote an upper bound on the size of the
problem instances in the support of D by n. For example, the support of D might be a set of
social networks over n individuals, and the researcher’s goal is to choose an algorithm with which
to perform a series of clustering analyses. Next, we fix a class of algorithms A. Given a cost
function cost : A × Π → [0, H], the learner’s goal is to find an algorithm h ∈ A that approximately
optimizes the expected cost with respect to the distribution D, as formalized below.
Definition 1 ([24]). A learning algorithm L (, δ)-learns the algorithm class A with respect to
the cost function cost if, for every distribution D over Π, with probability at least
1− δ i
over
h
m
the choice of a sample S ∼ D , L outputs an algorithm ĥ ∈ A such that Ex∼D cost ĥ, x −
minh∈A {Ex∼D [cost(h, x)]} < . We require that the number of samples be polynomial in n, 1 , and
1
δ , where n is an upper bound on the size of the problem instances in the support of D. Further, we
say that L is computationally efficient if its running time is also polynomial in n, 1 , and 1δ .
We derive our guarantees by analyzing the pseudo-dimension of the algorithm classes we study
1
(see Section 1.1, [31, 32, 2]). We then use the structure of the problem to provide efficient algorithms
for most of the classes we study.
SDP-based methods for integer quadratic programming. Many NP-hard problems, such as
max-cut, max-2SAT, and correlation clustering, can be represented as an integer quadratic program
(IQP) of the following form. The input is an n × n matrix A with nonnegative diagonal entries and
the
P output is a binary assignment to each variable in the set X = {x1 , . . . , xn } which maximizes
i,j∈[n] aij xi xj . In this formulation, xi ∈ {−1, 1} for all i ∈ [n]. (When the diagonal entries are
allowed to be negative, the ratio between the semidefinite relaxation and the integral optimum can
become arbitrarily large, so we restrict the domain to matrices with nonnegative diagonal entries.)
IQPs appear frequently in machine learning applications, such as MAP inference [25, 44, 20]
and image segmentation and correspondence problems in computer vision [15, 10]. Max-cut is
an important IQP problem, and its applications in machine learning include community detection
[7], variational methods for graphical models [34], and graph-based semi-supervised learning [39].
The seminal Goemans-Williamson max-cut algorithm is now a textbook example of semidefinite
programming [21, 41, 38]. Max-cut also arises in many other scientific domains, such as circuit
design [43] and computational biology [36].
The best approximation algorithms for IQPs relax the problem to anPSDP, where the input is
the same matrix A, but the output is a set of unit vectors maximizing i,j aij hui , uj i. The final
step is to transform, or “round,” the set of vectors into an assignment of the binary variables in X.
This assignment corresponds to a feasible solution to the original IQP. There are infinitely many
rounding techniques to choose from, many of which are randomized. These algorithms make up
the class of Random Projection, Randomized Rounding algorithms (RPR2 ), a general framework
introduced by [18]. RPR2 algorithms are known to perform well in theory and practice. When the
integer quadratic program is a formulation of the max-cut problem, the class of RPR2 algorithms
contain the groundbreaking Goemans-Williamson algorithm, which achieves a 0.878 approximation
ratio [21]. Assuming the unique games conjecture and P 6= N P , this approximation is optimal
to within any additive constant [27]. More generally, if A is any real-valued n × n matrix with
nonnegative diagonal entries, then there exists an RPR2 algorithm that achieves an approximation
ratio of Ω(1/ log n) [14], and in the worst case, this ratio is tight [1]. Finally, if A is positive
semi-definite, then there exists an RPR2 algorithm that achieves a 2/π approximation ratio [9].
We analyze several classes of RPR2 rounding function classes, including s-linear [18], outward
rotation [45], and ˜-discretized rounding functions [30]. For each class, we derive bounds on the
number of samples needed to learn an approximately optimal rounding function with respect to
an underlying distribution over problem instances using pseudo-dimension. We also provide a
computationally efficient and sample efficient learning algorithm for learning an approximately
optimal s-linear or outward rotation rounding function in expectation. We note that our results
also apply to any class of RPR2 algorithms where the first step is to find some set of vectors on the
unit sphere, not necessarily the SDP embedding, and then round those vectors to a binary solution.
This generalization has led to faster approximation algorithms with strong empirical performance
[26].
Clustering by agglomerative algorithms with dynamic programming. Given a set of n
datapoints and the pairwise distances between them, at a high level, the goal of clustering is to
partition the points into groups such that distances within each group are minimized and distances
between each group are maximized. A classic way to accomplish this task is to use an objective
function. Common clustering objective functions include k-means, k-median, and k-center, which
we define later on. We focus on a very general problem where the learner’s main goal is to minimize
an abstract cost function such as the cluster purity or the clustering objective function, which is
2
the case in many clustering applications such as clustering biological data [19, 29]. We study
infinite classes of two-step clustering algorithms consisting of a linkage-based step and a dynamic
programming step. First, the algorithm runs one of an infinite number of linkage-based routines
to construct a hierarchical tree of clusters. Next, the algorithm runs a dynamic programming
procedure to find the pruning of this tree that minimizes one of an infinite number of clustering
objectives. For example, if the clustering objective is the k-means objective, then the dynamic
programming step will return the optimal k-means pruning of the cluster tree.
For the linkage-based procedure, we consider several parameterized agglomerative procedures
which induce a spectrum of algorithms interpolating between the popular single-, average-, and
complete-linkage procedures, which are prevalent in practice [3, 35, 40] and known to perform
nearly optimally in many settings [4, 5, 6, 22]. For the dynamic programming step, we study an
infinite class of objectives which include the standard k-means, k-median, and k-center objectives,
common in applications such as information retrieval [11, 13]. We show how to learn the best
agglomerative algorithm and pruning objective function pair, thus extending our work to multiparameter algorithms. We provide tight pseudo-dimension bounds, ranging from Θ(log n) for simpler
algorithm classes to Θ(n) for more complex algorithm classes, so our learning algorithms are sample
efficient.
Key challenges. One of the key challenges in analyzing the pseudo-dimension of the algorithm
classes we study is that we must develop deep insights into how changes to an algorithm’s parameters
affect the solution the algorithm returns on an arbitrary input. For example, in our clustering
analysis, the cost function could be the k-means or k-median objective function, or even the distance
to some ground-truth clustering. As we range over algorithm parameters, we alter the merge step by
tuning an intricate measurement of the overall similarity of two point sets and we alter the pruning
step by adjusting the way in which the combinatorially complex cluster tree is pruned. The cost
of the returned clustering may vary unpredictably. Similarly, in integer quadratic programming, if
a variable flips from positive to negative, a large number of the summands in the IQP objective
will also flip signs. Nevertheless, we show that in both scenarios, we can take advantage of the
structure of the problems to develop our learning algorithms and bound the pseudo-dimension.
In this way, our algorithm analyses require more care than standard complexity derivations
commonly found in machine learning contexts. Typically, for well-understood function classes used
in machine learning, such as linear separators or other smooth curves in Euclidean spaces, there is
a simple mapping from the parameters of a specific hypothesis to its prediction on a given example
and a close connection between the distance in the parameter space between two parameter vectors
and the distance in function space between their associated hypotheses. Roughly speaking, it is
necessary to understand this connection in order to determine how many significantly different
hypotheses there are over the full range of parameters. Due to the inherent complexity of the
classes we consider, connecting the parameter space to the space of approximation algorithms and
their associated costs requires a much more delicate analysis. Indeed, the key technical part of
our work involves understanding this connection from a learning-theoretic perspective. In fact, the
structure we discover in our pseudo-dimension analyses allows us to develop many computationally
efficient meta-algorithms for algorithm configuration due to the related concept of shattering. A
constrained pseudo-dimension of O(log n) often implies a small search space of 2O(log n) = O(n) in
which the meta-algorithm will uncover a nearly optimal configuration.
We bolster the theory of algorithm configuration by studying algorithms for problems that are
ubiquitous in machine learning and optimization: integer quadratic programming and clustering.
In this paper, we develop techniques for analyzing randomized algorithms, whereas the algorithms
analyzed in the previous work were deterministic. We also provide the first pseudo-dimension
3
lower bounds in this line of work, which require an involved analysis of each algorithm family’s
performance on carefully constructed instances. Our lower bounds are somewhat counterintuitive,
since for several of the classes we study, they are of the order Ω(log n), even if the corresponding
classes of algorithms are defined by a single real-valued parameter.
1.1
Preliminaries and definitions
In this section, we provide the definition of pseudo-dimension in the context of algorithm classes.
Consider a class of algorithms A and a class of problem instances X . Let the cost function cost(h, x)
denote the abstract cost of running an algorithm h ∈ A on a problem instance x ∈ X . Similarly,
define the function class HA = {cost(h, ·) : X → R|h ∈ A}. Then, recall that a finite subset of
problem instances S = {x1 , x2 , . . . xm } is shattered by the function class H, if there exist real-valued
witnesses r1 , . . . , rm such that for all subsets T ⊆ S, there exists a function cost (hT , ·) ∈ H, or in
other words, an algorithm hT ∈ A such that cost (hT , xi ) ≤ ri if and only if i ∈ T . Then, we can
define the pseudo-dimension of the algorithm class A to be the pseudo-dimension P dim(H) of H
i.e., the cardinality of the largest subset of X shattered by H.
By bounding P dim(H), clearly we can derive sample complexity guarantees in the context
of algorithm classes [17]: for every distribution D over X , every > 0, and every δ ∈ (0, 1],
2
m ≥ c H
P dim(H) + log 1δ for a suitable constant c (independent of all other parameters),
then with probability at least 1 − δ over m samples x1 , . . . , xm ∼ D,
m
1 X
cost(h, xi ) − Ex∼D [cost(h, x)] <
m
i=1
for every algorithm h ∈ A. Therefore, if a learning algorithm receives as input a sufficiently large
set of samples and returns the algorithm which performs best on that sample, we can be guaranteed
that this algorithm is close to optimal with respect to the underlying distribution.
2
SDP-based methods for integer quadratic programming
In this section, we study several IQP approximation algorithms. These classes consist of SDP
rounding algorithms and are a generalization of the seminal Goemans-Williamson (GW) max-cut
algorithm [21]. We prove that it is possible to learn the optimal algorithm from a fixed class over
a specific application domain, and for many of the classes we study, this learning procedure is
computationally efficient and sample efficient.
P
We focus on integer quadratic programs of the form
i,j∈[n] aij xi xj , where the input is a
matrix A with nonnegative diagonal entries and the output is an assignment of the binary variables
X = {x1 , . . . , xn } maximizing this sum. Specifically, each variable in X is set to either −1 or 1. This
problem is also known as MaxQP [14]. Most algorithms with the best approximation guarantees
use an SDP relaxation. The SDP relaxation has the form
X
maximize
aij hui , uj i
subject to ui ∈ S n−1 .
(1)
i,j∈[n]
Given the set of vectors {u1 , . . . , un }, we must decide how they represent an assignment of the
binary variables in X. In the GW algorithm, the vectors are projected onto a random vector Z
drawn from an n-dimensional Gaussian. Next, if the directed distance of the resulting projection
is greater than 0, then the corresponding binary variable is set to 1, and otherwise it is set to −1.
4
In some cases, the GW algorithm can be improved upon by probabilistically assigning each
binary variable to 1 or −1. In the final rounding step, any rounding function r : R → [−1, 1] can
be used to specify that a variable xi is set to 1 with probability 12 + 21 · r (hZ, ui i) and −1 with
probability 21 − 12 · r (hZ, ui i). See Algorithm 1 for the pseudocode. This is known as a Random
Algorithm 1 SDP rounding algorithm with rounding function r
Input: Matrix A ∈ Rn×n .
1: Solve the SDP (1) for the optimal embedding U = (u1 , . . . , un ).
2: Choose a random vector Z ∈ Rn according to the n-dimensional Gaussian distribution.
3: Define the fractional assignment h : X → [−1, 1] such that h(xi ) = r (hZ, ui i).
Output: h.
Projection, Randomized Rounding (RPR2 ) algorithm, so named by the seminal work of [18]. The
randomized assignment h produced by an RPR2 algorithm is called a fractional assignment. Based
on the output h, we can derive a proper assignment of the variables x1 , . . . , xn where xi is set to 1
with probability 21 + 21 h(xi ) and −1 with probability 21 − 12 h(xi ).
In this section, we analyze the class of s-linear round functions. For the max-cut problem, [18]
proved that when the maximium cut in the graph is not very large, a worst-case approximation
ratio above the GW ratio is possible using an s-linear rounding function. For example, they proved
that if the optimal cut contains at most a 0.6 fraction of the edges, then the ratio is at least 0.9128.
The optimal choice of s depends on the graph, but we give an efficient algorithm to learn a nearly
optimal value for s in expectation over a distribution of problem instances. In Appendix B, we
consider other rounding functions, including ˜-discretized rounding functions [30], outward rotation
algorithms [45], and “sigmoid-like” rounding functions, which include the classes of s-linear and
outward rotation functions.
An s-linear rounding function φs : R → [−1, 1] is parameterized by s > 0, as follows:
−1 if y < −s
φs (y) = y/s if − s ≤ y ≤ s
1
if y > s.
Figure 1: A graph of the 2-linear function φ2 .
Our goal is to devise an algorithm Lslin that (, δ)-learns the best s-linear rounding function
with respect to a distribution D over MaxQP problem instances. Specifically, let slin∗s (A) be the
expected value of the solution returned by RPR2 using the rounding function φs when evaluated
on the MaxQP problem instance defined by the matrix A. We instantiate the cost function cost
to be cost(A, s) = −slin∗s (A). Here, we take the negative of slin∗s (A) since our goal is to find the
s parameter that will maximize this value in expectation, and minimizing cost(A, s) now amounts
to maximizing slin∗s (A). We require that Lslin returns a value ŝ such that with probability at least
1 − δ, if s∗ maximizes E [slin∗s (A)], then E [slin∗s∗ (A)] − E [slin∗ŝ (A)] < .
A∼D
A∼D
A∼D
One might expect that the first step would be to bound the pseudo-dimension of the class
Hslin∗ = {slin∗s : A → [−H, H] | s > 0} (Pdim(Hslin∗ )), where A is the set of all real-valued n × n
matrices with nonnegative diagonal entries and H is an upper bound on the range of slin∗s when
restricted to the support of the distribution over problem instances. We pursue an alternative route
5
that provides a simpler sample complexity and algorithmic analysis. We instead bound the pseudodimension of the class Hslin = {slins : A × Rn → [−H, H] | s > 0}, where slins (A, Z) is the value
of the fractional assignment produced by projecting the SDP embedding of A onto Z and rounding
the directed distances
function φs . ExplicPof the projections (multiplied by ||Z||) using the rounding
∗
itly, slins (A, Z) = i,j aij φs (hui , Zi) · φs (hui , Zi) . Notice that slins (A) = EZ∼Z [slins (A, Z)],
where Z denotes the standard n-dimensional Gaussian distribution. We prove tight bounds on the
pseudo-dimension of Hslin and derive generalization guarantees for the algorithm class we ultimately
care about: Hslin∗ , as follows. All omitted proofs can be found in Appendix A.
Theorem 2. Suppose that Lslin is an algorithm that takes
as input m samples (A(i) , Z(i) ) ∼
1 Pm
D × Z and returns the parameter ŝ which maximizes m i=1 slins (A(i) , Z(i) ). Further, supposePthat m is sufficiently large to ensure that with probability at least 1 − δ, for all s > 0,
m
1
(i)
(i)
|m
i=1 slins (A , Z ) − E(A,Z) [slins (A, Z)] | ≤ 2 . Then Lslin (, δ)-learns the class of s-linear
rounding functions with respect to the cost function −slin∗s .
Now, we will show that Pdim(Hslin ) = Θ(log n) and then we will present Algorithm 2, a
computationally efficient and sample efficient algorithm. We will often fix a tuple (A, Z) and
consider slins (A, Z) as a function of s, which we denote as slinA,Z (s). We begin with a helpful
lemma.
Lemma 1. The function slinA,Z : R>0 → R is made up of n + 1 piecewise components of the
form sa2 + sb + c for some a, b, c ∈ R. Moreover, if the border between two components falls at some
s ∈ R>0 , then it must be that s = |hui , Zi| for some ui in the optimal SDP embedding of A.
n−1 be the optimal embedding of A. Then we may write
Proof. First, let
PX = {u1 , . . . , un } ⊂ S
slinA,Z (s) = i,j aij φs (hui , Zi) · φs (hui , Zi) . For any ui ∈ X, the specific form of φs (hZ, ui i)
depends solely on whether hZ, ui i < −s, −s ≤ hZ, ui i ≤ s, or s < hZ, ui i. Of course, if hZ, ui i ≥ 0
or hZ, ui i < 0, we can disregard the possibility that hZ, ui i < −s or hZ, ui i > s, respectively. Then
so long as |hZ, ui i| > s, we have that φs (hZ, ui i) = ±1, where the sign depends on the sign of hZ, ui i.
Further, when s grows to the point where s > |hZ, uj i|, we have that φs (hZ, ui i) = hZ, uj i/s.
Therefore, if we order the set of real values {|hZ, u1 i| , . . . , |hZ, un i|}, then so long as s falls between
two consecutive elements of this ordering, the form of slinA,Z (s) is fixed. In particular, each
summand is either a constant, a constant multiplied by 1s , or a constant multiplied by s12 , perhaps
accompanied by an additive constant term. This means that we may partition the positive real line
into n + 1 intervals where the form of slin(A,Z) (s) is a fixed quadratic function, as claimed.
Lemma 1 allows us to prove the following bound on Pdim(Hslin ).
Lemma 2. Pdim(Hslin ) = Θ(log n).
Lemma 2 follows from Lemmas 3 and 4, where we prove Pdim(Hslin ) = O(log n) and Pdim(Hslin ) =
Ω(log n).
Lemma 3. Pdim(Hslin ) = O(log n).
Proof. We prove this upper bound by showing that if a set S of size m is shatterable, then
m = O(log n). This means that the largest shatterable set must be of size O(log n),so the pseudodimension of Hslin is O(log n). We arrive at this bound by fixing a tuple A(i) , Z(i) ∈ S and analyzing slinA,Z (s). In particular, we make use of Lemma 1, from which we know that slin(A,Z) (s)
is composed of n + 1 piecewise
quadratic components. Therefore, if ri is the witness corresponding
to the element A(i) , Z(i) , we can partition the positive real line into at most 3(n + 1) intervals
6
At most 3(n + 1) intervals
for sample (A(1) , Z (1) ) and witness r1
s>0
At most 3(n + 1) intervals
for sample (A(2) , Z (2) ) and witness r2
s>0
..
At most 3(n + 1) intervals
for sample (A(|S|) , Z (|S|) ) and witness r|S|
s>0
Merging the intervals of all samples and witnesses
At most (3n + 2)m + 1 intervals
for samples S and their respective witnesses
s>0
Figure 2: Partitioning s > 0 into intervals given a set S of m tuples A(i) , Z(i) and witnesses ri
such that within each interval for each i, slin(A(i) ,Z (i) ) (s) is always greater than ri or lesser than
ri .
where slinA,Z (s) is always either less than its witness ri or greater than ri as s varies over one
fixed interval. The constant 3 term comes from the fact that for a single, continuous quadratic
component of slin(A,Z) (s), the function may equal ri at most twice, so there are at most three
subintervals where the function is less than or greater than ri .
Now, S consists of m tuples A(i) , Z(i) , each of which corresponds to its own partition of the
positive real line. If we merge these partitions (as shown in Figure 2), simple algebra shows that
we are left with at most (3n + 2)m + 1 intervals such that for all i ∈ [m], slinA(i) ,Z(i) (s) is always
either less than its witness ri or greater than ri as s varies over one fixed interval. In other words,
in one interval, the binary labeling of S, defined by whether each sample is less than or greater
than its witness, is fixed. This means that if S is shatterable, the 2m values of s which induce all
2m binary labelings of S must come from distinct intervals. Therefore 2m ≤ (3n + 2)m + 1, so
m = O(log n).
Lemma 4. Pdim(Hslin ) = Ω(log n).
Proof sketch. In order
to prove that the pseudo dimension of Hslin is at least c log n for some c, we
present a set S = (G(1) , Z(1) ), . . . , (G(m) , Z(m) ) of m = c log n graphs and projection vectors that
can be shattered by Hslin . In other words, there exist m witnesses r1 , . . . , rm and 2m = nc values
of s, H = {s1 , . . . , snc }, such that for all T ⊆ [m], there exists sT ∈ H such that if j ∈ T , then
slinST (G(j) , Z(j) ) > rj and if j 6∈ T , then slinST (G(j) , Z(j) ) ≤ rj .
To build S, we use the same graph G for all G(j) and we vary Z(j) . We set G to be the graph
composed of bn/4c disjoint copies of K4 . Via a careful choice of the vectors Z(j) and witnesses rj ,
we pick out 2m critical values of s, which we call C, such that slinG,Z(1) (s) switches from above
to below its witness for every other element of the critical values in C. Meanwhile, slinG,Z(2) (s)
switches from above to below its witness half as often as slinG,Z(1) (s). Similarly, slinG,Z(3) (s)
switches from above to below its witness half as often as slinG,Z(2) (s), and so on. Therefore, we
achieve every binary labeling of S using the functions {slins | s ∈ C}, so S is shattered.
Our lower bound is particularly strong because it holds for a family of positive semidefinite
matrices, rather than a more general family of real-valued matrices. We now prove that our
learning algorithm, Algorithm 2 is correct, computationally efficient, and sample efficient.
To prove that Algorithm 2 (, δ)-learns the class of s-linear rounding functions, we first prove
that Algorithm 2 ideed outputs the empirically optimal value of ŝ.
7
Algorithm 2 An algorithm for finding an empirical value maximizing s-linear rounding function
Input: Sample S = {(A(1) , Z(1) ), . . . , (A(m) , Z(m) )}
(i)
(i)
1: For all i, solve for the SDP embedding U (i) of A(i) , where U (i) = (u1 , . . . , un ).
2: Let T = s1 , . . . , s|T | be the set of all values s > 0 such that there exists a pair of indices
(i)
j ∈ [n], i ∈ [m] with |hZ(i) , uj i| = s.
1 Pm
3: For i ∈ [|T | − 1], let ŝi be the value in [si , si+1 ] which maximizes m
i=1 slinA(i) ,Z(i) (s).
1 Pm
4: Let ŝ be the value in {ŝ1 , . . . , ŝ|T |−1 } that maximizes m
i=1 slinA(i) ,Z(i) (s).
Output: ŝ
1 Pm
(i)
(i) over the
Lemma 5. Algorithm 2 produces the value ŝ which maximizes m
i=1 slins A , Z
input sample S. Algorithm 2 has running time polynomial in m and n.
1 Pm
Proof. First, define hS,Z (s) = m
i=1 slinA(i) ,Z(i) (s), which we claim that Algorithm 2 maximizes.
In Lemma 1, we proved that each function slinA(i) ,Z(i) (s) is made up of at most n + 1 piecewise
components of the form sa2 + sb + c for some a, b, c ∈ R. Therefore, hS,Z (s) is made up of at most
mn + 1 piecewise components of the form sa2 + sb + c as well. Moreover, by Lemma 1, if the border
(i)
(i)
between two components falls at some s ∈ R>0 , then it must be that hZ(i) , xj i = s for some xj
in the optimal max-cut SDP embedding of A(i) . These are the thresholds which are computed in
Step 2 of Algorithm 2. Therefore, as we increase s starting at 0, s will be a fixed quadratic function
between the thresholds, so it is simple to find the optimal value of sP
between any pair of consecutive
1
(i)
(i) (Step 4),
thresholds (Step 3), and then the value maximizing hS,Z (s) = m m
i=1 slins A , Z
which is the global optimum.
We are now ready to prove our main result.
2
drawn from (D × Z)m , AlgoTheorem 3. Given a sample of size m = O H
log (n) + log 1δ
rithm 2 (, δ)-learns the class of s-linear rounding functions with respect to the cost function −slin∗s
and it is computationally efficient.
Proof. Let S = A(1) , Z(1) , . . . , A(m) , Z(m)
be a sample of size m. In Lemma 5, we prove
1 Pm
(i)
(i) in
that Algorithm 2 on input S returns the value ŝ which maximizes m
i=1 slins A , Z
polynomial
time. Then from our pseudodimension bound in Lemma 3 we have that with m =
H 2
O
log n + log 1δ
samples, with probability at least 1 − δ, for all s > 0,
1 X
slins A(i) , Z(i) −
E
[slins (A, Z)] < .
m
2
(A,Z)∼D×Z
m
i=1
Then, from Lemma 12, E [slins∗ (A)] − E [slinŝ (A)] < . Thus, Algorithm 2 (, δ)-learns the
A∼D
best s-linear function with respect to D.
3
A∼D
Agglomerative algorithms with dynamic programming
We begin with an overview of agglomerative algorithms with dynamic programming, which include
many widely-studied clustering algorithms, and then we define several parameterized classes of such
algorithms. As in the previous section, we prove it is possible to learn the optimal algorithm from
8
a fixed class for a specific application, and for many of the classes we analyze, this procedure is
computationally efficient and sample efficient. We focus on agglomerative algorithms with dynamic
programming for clustering problems. A clustering instance V = (V, d) consists of a set V of n
points and a distance metric d : V ×V → R≥0 specifying all pairwise distances between these points.
The overall goal of clustering is to partition the points into groups such that distances within each
group are minimized and distances between each group are maximized. Clustering is typically
performed using an objective function Φ, such as k-means, k-median, k-center, or the distance to
the ground truth clustering (a scenario we discuss in more detail in Section 3.2). Formally, an
objective function Φ takes as input a set of points c = {c1 , . . . , ck } ⊆ V which we call centers,
as well as a partition C = {C1 , . . . , CP
V which we call a clustering. We define the rich class
k } of P
of clustering objectives Φ(p) (C, c) = ki=1 ( q∈Ci d(q, ci )p )1/p for p ∈ [1, ∞) ∪ {∞}. The k-means,
k-median, and k-center objective functions are Φ(2) , Φ(1) , and Φ(∞) , respectively.1
Next, we define agglomerative clustering algorithms with dynamic programming, which are
prevalent in practice [3, 35, 40] and enjoy strong theoretical guarantees in a variety of settings
[4, 5, 6, 22]. Examples of these algorithms include the popular single-, complete-, and averagelinkage algorithms with dynamic programming. An agglomerative clustering algorithm with dynamic programming is defined by two functions: a merge function and a pruning function. A merge
function ξ(A, B) → R≥0 defines a distance between two sets of points A, B ⊆ V . The algorithm
builds a cluster tree T by starting with n singleton leaf nodes, and iteratively merging the two sets
with minimum distance until there is a single node remaining, consisting of the set V . The children
of any node T in this tree correspond to the two sets of points that were merged to form T during
the sequence of P
merges. Common choices for the merge function ξ include mina∈A,b∈B d(a, b) (single
1
linkage), |A|·|B|
a∈A,b∈B d(a, b) (average linkage) and maxa∈A,b∈B d(a, b) (complete linkage).
A pruning function Ψ takes as input a k 0 -pruning of any subtree of T and returns a score R≥0
for that pruning. A k 0 -pruning for a subtree T is a partition of the points contained in T ’s root
into k 0 clusters such that each cluster is an internal node of T . Pruning functions may be similar to
objective functions, though the input is a subtree. The k-means, -median, and -center objectives
are standard pruning functions. The algorithm returns the k-pruning of the tree T that is optimal
according to Ψ, which can be found in polynomial time using dynamic programming. Algorithm 3
details how the merge function and pruning function work together to form an agglomerative
clustering algorithm with dynamic programming. In the dynamic programming step, to find the
1-pruning of any node T , we only need to find the best center c ∈ T . When k 0 > 1, we recursively
find the best k 0 -pruning of T by considering different combinations of the best i0 -pruning of the left
child TL and the best k 0 − i0 -pruning of the right child TR for i0 ∈ {1, . . . , k − 1} and choosing the
best combination.
Pictorially, Figure 3 depicts an array of available choices when designing an agglomerative
clustering algorithm with dynamic programming. Each path in the chart corresponds to an alternative choice of a merging function ξ and pruning function Ψ. The algorithm designer’s goal is to
determine the path that is optimal for her specific application domain.
In Section 3.1, we analyze several classes of algorithms where the merge function comes from an
infinite family of functions while the pruning function is an arbitrary, fixed function. In Section 3.2,
we expand our analysis to include algorithms defined over an infinite family of pruning functions
in conjunction with any family of merge functions. Our results hold even when there is a fixed
preprocessing step that precedes the agglomerative merge step (as long as it is independent of ξ
1
There have been several papers that provide theoretical guarantees for clustering under this family of objective functions for other values of p. For instance, see Gupta and Tangwongsan’s work [23] which provides a O(p)
approximation algorithm when p < log n and Bateni et al.’s work [8] which studies distributed clustering algorithms.
9
Algorithm 3 Agglomerative algorithm with dynamic programming
Input: Clustering instance V = (V, d), merge function ξ, pruning function Ψ.
1: Agglomerative merge step to build a cluster tree T according to ξ:
• Start with n singleton sets {v} for each v ∈ V .
• Iteratively merge the two sets A and B which minimize ξ(A, B) until a single set remains.
• Let T denote the cluster tree corresponding to the sequence of merges.
2: Dynamic programming to find the k-pruning of T minimizing Ψ :
0
• For each node
T , find the best k -pruning of the subtree rooted at T in T , denoted by
CT,k0 , cT,k0 according to following dynamic programming recursion:
(
minc∈T Ψ ({T }, c)
if k 0 = 1,
Ψ CT,k0 , cT,k0 =
mini0 ∈[k0 −1] Ψ CTL ,i0 ∪ CTR ,k0 −i0 , cTL ,i0 ∪ cTR ,k0 −i0
otherwise.
where TL and TR denote the left and right children of T , respectively.
Output: The best k-pruning of the root node Troot of T .
Figure 3: A schematic for a class of agglomerative clustering algorithms with dynamic programming.
10
and Ψ), therefore our analysis carries over to algorithms such as in [6].
3.1
Linkage-based merge functions
We now define three infinite families of merge functions and provide sample complexity bounds
for these families with any fixed but arbitrary pruning function. The families A1 and A3 consist
of merge functions ξ(A, B) that depend on the minimum and maximum of all pairwise distances
between A and B. The second family, denoted by A2 , is a richer class which depends on all pairwise
distances. All classes are parameterized by a single value α.
(
)
1/α
A1 =
min (d(u, v))α + max (d(u, v))α
α ∈ R ∪ {∞, −∞} ,
u∈A,v∈B
u∈A,v∈B
1/α
X
1
A2 =
(d(u, v))α
α ∈ R ∪ {∞, −∞} ,
|A||B| u∈A,v∈B
A3 = α min d(u, v) + (1 − α) max d(u, v) α ∈ [0, 1] .
u∈A,v∈B
u∈A,v∈B
For i ∈ {1, 2, 3}, we define Ai (α) as the merge function in Ai defined by α. A1 and A3 define
spectra of merge functions ranging from single-linkage (A1 (−∞) and A3 (1)) to complete-linkage
(A1 (∞) and A3 (0)). A2 defines a richer spectrum which includes average-linkage in addition to
single- and complete-linkage. Given a pruning function Ψ, we denote (Ai (α), Ψ) as the algorithm
which builds a cluster tree using Ai (α), and then prunes the tree according to Ψ. To reduce
notation, when Ψ is clear from context, we often refer to the algorithm (Ai (α), Ψ) as Ai (α) and
the set of algorithms {(Ai (α), Ψ) | α ∈ R ∪ {−∞, ∞}} as Ai . For example, when the cost function
is Φ(p) , then we always set Ψ to minimize the Φ(p) objective, so the pruning function is clear from
context.
Recall that for a given class of merge functions and a cost function (a generic clustering objective
Φ), our goal is to learn a near-optimal value of α in expectation over an unknown distribution of
clustering instances. One might wonder if there is some α that is optimal across all instances, which
would preclude the need for a learning algorithm. In Theorem 4, we prove that this is not the case;
for each p ∈ [1, ∞) ∪ {∞} and i ∈ {1, 2, 3}, given any α, there exists a distribution over clustering
instances for which Ai (α) is the best algorithm in Ai with respect to Φ(p) . Crucially, this means
that even if the algorithm designer sets p to be 1, 2, or ∞ as is typical in practice, the optimal
choice of the tunable parameter α could be any real value. The optimal value of α depends on the
underlying, unknown distribution, and must be learned, no matter the value of p.
To formally describe this result, we set up notation similar to Section 2. Let V denote the
set of all clustering instances over at most n points. With a slight abuse of notation, we will use
Φ(Ai (α),Ψ) (V) to denote the abstract cost of the clustering produced by (Ai (α), Ψ) on the instance
V.
Theorem 4. For i ∈ {1, 2, 3} and a permissible
value iof α for Ahi , there exists
h
i a distribution D
(p)
(p)
over clustering instances V such that EV∼D ΦAi (α) (V) < EV∼D ΦAi (α0 ) (V) for all permissible
values of α0 6= α for Ai .
For all omitted proofs in this section, see Appendix C. Now for an arbitrary objective function
Φ and arbitrary pruning function Ψ, we analyze the complexity of the classes
HA1 ×{Ψ},Φ = Φ(A1 (α),Ψ) : V → R≥0 α ∈ R ∪ {∞, −∞} ,
11
HA2 ×{Ψ},Φ = Φ(A2 (α),Ψ) : V → R≥0 α ∈ R ∪ {∞, −∞} , and
HA3 ×{Ψ},Φ = Φ(A3 (α),Ψ) : V → R≥0 α ∈ [0, 1] .
We drop the subscript Φ from HAi ×{Ψ},Φ when the objective function is clear from context. Furthermore, in our analysis we will often fix a tuple V = (V, d) and use the notation ΦAi ,V (α) to
analyze how ΦAi (α) (V) changes as a function of α. We start with A1 and A3 .
Theorem 5. For all objective functions Φ(p) , Pdim HA1 ,Φ(p) = Θ(log n) and Pdim HA3 ,Φ(p) =
2 and all pruning functions Ψ, Pdim H
Θ(log n). For all other objective
functions
Φ
=
A
×{Ψ},Φ
1
O(log n) and Pdim HA3 ×{Ψ},Φ = O(log n).
This theorem follows from Lemma 7 and Lemma 8. We begin with the following structural
lemma, which will help us prove Lemma 7.
Lemma 6. ΦA1 ,V : R ∪ {−∞, ∞} → R>0 is made up of O(n8 ) piecewise constant components.
Proof sketch. Note that for α 6= α0 , the clustering returned by A1 (α) and the associated cost are
both identical to that of A1 (α0 ) if both the algorithms construct the same merge tree. As we range
α across R and observe the run of the algorithm for each α, at what values of α do we expect A1 (α)
to produce different merge trees? To answer this, suppose that at some point in the run of algorithm
A1 (α), there are two pairs of subsets of V , (A, B) and (X, Y ), that could potentially merge. There
exist eight points p, p0 ∈ A, q, q 0 ∈ B, x, x0 ∈ X, and y, y 0 ∈ Y such that the decision of which pair
to merge depends on the sign of ((d(p, q))α + d(p0 , q 0 )α )1/α − ((d(x, y))α + d(x0 , y 0 )α )1/α . Using a
consequence of Rolle’s Theorem, which we provide in Appendix C, we show that the sign of the
above expression as a function of α flips at most four times across R. Since each merge decision
is defined by eight points, iterating over all (A, B) and (X, Y ) it follows that we can identify all
O(n8 ) unique 8-tuples of points which correspond to a value of α at which some decision flips. This
means we can divide R ∪ {−∞, ∞} into O(n8 ) intervals over each of which the merge tree, and
therefore the output of ΦA1 ,V (α), is fixed.
In Appendix C, we show a corresponding statement for A3 (Lemma 18). These lemmas allow
us to upper bound the pseudo-dimension of HA1 , Φ and HA1 , Φ by O(log n) in a manner similar to
Lemma 3, where we prove a pseudo-dimension upper bound on the class of s-linear SDP rounding
algorithms. Thus we obtain the following lemma.
Lemma 7. For any Φ, Pdim(HA1 , Φ) = O(log n) and Pdim(HA3 , Φ) = O(log n).
Next, we give lower bounds for the pseudo-dimension of the two classes.
Lemma 8. For any objective function Φ(p) , Pdim HA1 ,Φ(p) = Ω(log n) and Pdim HA3 ,Φ(p) =
Ω(log n).
Proof sketch. We give
a general proof outline that applies to both classes. Let b ∈ {1, 3}. We
(1)
construct a set S = V , . . . , V (m) of m = log n − 3 clustering instances that can be shattered by
Ab . There are 2m = n/8 possible labelings for this set, so we need to show there are n/8 choices
of α such that each of these labelings is achievable by some Ab (α) for some α. The crux of the
proof lies in showing that given a sequence α0 < α1 < · · · < αn0 < αn0 +1 (where n0 = Ω(n)), it is
2
Recall that although k-means, k-median, and k-center are the most popular choices, the algorithm designer
can use other objective functions such as the distance to the ground truth clustering (which we discuss further in
Section 3.2).
12
Algorithm 4 An algorithm for finding an empirical cost minimizing algorithm in A1 or A3
Input: Sample S = V (1) , . . . , V (m) , b ∈ {1, 3}
1: Let T = ∅. For each sample V (i) = V (i) , d(i) ∈ S, and for each ordered set of 8 points
{v1 , . . . , v8 } ⊆ V (i) , solve for α (if a solution exists) in the following equation and add the
solutions to T : d(v1 , v2 )α + d(v3 , v4 )α = d(v5 , v6 )α + d(v7 , v8 )α .
If b = 1 :
d(v1 , v2 )α + d(v3 , v4 )α = d(v5 , v6 )α + d(v7 , v8 )α .
If b = 3 : αd(v1 , v2 ) + (1 − α)d(v3 , v4 ) = αd(v5 , v6 ) + (1 − α)d(v7 , v8 ).
Order the elements of set T ∪ {−∞, +∞} as α1 < . . . < α|T | . For each 0 ≤ i ≤ |T |, pick an
arbitrary α in the interval (αi , αi+1 ) and run Ab (α) on P
all clustering instances in S to compute
P
V∈S ΦAb (α) (V). Let α̂ be the value which minimizes
V∈S ΦAb (α) (V).
Output: α̂
2:
possible to design an instance V = (V, d) over n points and choose a witness r such that ΦAb (α) (V)
alternates n0 /2 times above and below r as α traverses the sequence of intervals (αi , αi+1 ).
Here is a high level description of our construction. There will be two “main” points, a and a0
in V . The rest of the points are defined in groups of 6: (xi , yi , zi , x0i , yi0 , zi0 ), for 1 ≤ i ≤ (n − 2)/6.
We will define the distances between all points such that initially for all Ab (α), xi merges to yi
to form the set Ai , and x0i merges to yi0 to form the set A0i . As for (zi , zi0 ), depending on whether
α < αi or not, Ab (α) merges the points zi and zi0 with the sets Ai and A0i respectively or vice versa.
This means that there are (n − 2)/6 values of α such that Ab (α) has a unique behavior in S
the merge
step. Finally,Sfor all α, sets Ai merge to {a}, and sets A0i merge to {a0 }. Let A = {a} ∪ i Ai and
A0 = {a0 } ∪ i A0i . There will be (n − 2)/6 intervals (αi , αi+1 ) for which Ab (α) returns a unique
partition {A, A0 }. By carefully setting the distances, we cause the cost Φ({A, A0 }) to oscillate above
and below a specified value r along these intervals.
The upper bound on the pseudo-dimension implies a computationally efficient and sample
effi
cient learning algorithm for Ai for i ∈ {1, 3}. See Algorithm 4. First, we know that m = Õ (H/)2
samples are sufficient to (, δ)-learn the optimal algorithm in Ai . Next, as a consequence of Lemmas 6 and 18, the range of feasible values of α can be partitioned into O(mn8 ) intervals, such that
the output of Ai (α) is fixed over the entire set of samples on a given interval. Moreover, these
intervals are easy to compute. Therefore, a learning algorithm can iterate over the set of intervals,
and for each interval I, choose an arbitrary α ∈ I and compute the average cost of Ai (α) evaluated
on the samples. The algorithm then outputs the α that minimizes the average cost.
Theorem 6. Let Φ be a clustering objective and let Ψ be a pruning function
computable in poly
H 2
1
log n + log δ , and a value i ∈ {1, 3},
nomial time. Given an input sample of size m = O
Algorithm 4 (, δ)-learns the class Ai × {Ψ} with respect to the cost function Φ and it is computationally efficient.
Proof.
Algorithm 4 finds the empirically best α by solving for the O(mn8 ) discontinuities of
P
V∈S ΦAb (α) (V) and evaluating the function over the corresponding intervals, which are guaranteed to be constant by Lemmas 6 and 18. Therefore, we can pick any arbitrary α within each
interval to evaluate the empirical cost over all samples, and find the empirically best α. This can be
done in polynomial time because there are polynomially many intervals, and the runtime of Ai (α)
on a given instance is polynomial time.
13
Then it follows from Theorem 5 that m samples are sufficient for Algorithm 4 to (, δ)-learn the
optimal algorithm in Ai for i ∈ {1, 3}.
Now we turn to A2 . We obtain the following bounds on the pseudo-dimension.
Theorem 7. For objective functions Φ(p) , Pdim(HA2 ,Φ(p) ) = Θ(n). For all other objective functions
Φ and all pruning functions Ψ, Pdim HA2 ×{Ψ},Φ = O(n).
This theorem follows from Lemmas 9 and 10.
Lemma 9. For all objective functions Φ and all pruning functions Ψ, Pdim HA2 ×{Ψ},Φ = O(n).
Proof. Recall the proof of Lemma 7. We are interested in studying how the merge trees constructed
by A2 (α) changes over m instances as we increase α over R. To do this, as in the proof of Lemma 7,
we fix an instance and consider two pairs of sets A, B and X, Y that could be potentially merged.
Now, P
the decision to merge one pair
is determined by the sign of the expression
P before the other
1
1
α −
α . First note that this expression has O(n2 )
(d(p,
q))
(d(x,
y))
p∈A,q∈B
x∈X,y∈Y
|A||B|
|X||Y |
terms, and by a consequence of Rolle’s
which we prove in Appendix C, it has O(n2 ) roots.
Theorem
Therefore, as we iterate over the O (3n )2 possible pairs (A, B) and (X, Y ), we can determine
O 32n unique expressions each with
O(n2 ) values of α at which the corresponding decision flips.
Thus we can divide R into O n2 32n intervals over each of which the output of ΦA3 ,V (α) is fixed.
In fact, suppose S = V (1) , . .. , V (m) is a shatterable set of size m with witnesses r1 , . . . , rm . We
can divide R into O mn2 32n intervals over each of which ΦA2 ,V (i) (α) is fixed for all i ∈ [m] and
therefore the corresponding labeling of S according to whether or not ΦA2 (α) V (i) ≤ ri is fixed as
well for all i ∈ [m]. This means that HA2 can achieve only O mn2 32n labelings, which is at least
2m for a shatterable set S, so m = O(n).
Lemma 10. For all objective functions Φ(p) , Pdim HA2 ,Φ(p) = Ω(n).
Proof sketch. The crux of the proof is to show that there exists a clustering instance V over n
points, a witness r, and a set of α’s 1 = α0 < α1 < · · · < α2N < α2N +1 = 3, where N = b(n − 8)/4c,
such that ΦA2 ,V (α) oscillates above and below r along the sequence of intervals (αi , αi+1 ). We
finish the proof in a manner similar to Lemma 8 by constructing instances with fewer oscillations.
To construct V, first we define two pairs of points which merge together regardless of the value
of α. Call these merged pairs A and B. Next, we define a sequence of points pi and qi for 1 ≤ i ≤ N
with distances set such that merges involving points in this sequence occur one after the other. In
particular, each pi merges with one of A or B while qi merges with the other. Therefore, there
are potentially 2N distinct merge trees which can be created. Using induction to precisely set the
distances, we show there are 2N distinct values of α, each corresponding to a unique merge tree,
thus enabling A2 to achieve all possible merge tree behaviors. Finally, we carefully add more points
to the instance to control the oscillation of the cost function over these intervals as desired.
In Appendix C.1, we prove more results about A2 assuming a natural restriction on the instance
space V. In particular, we show the pseudo-dimension can be drastically reduced if the number of
unique distances in each problem instance is not too large. In Appendix C.2, we analyze classes
of algorithms that interpolate between A1 and A2 . A summary of results for all of our algorithm
classes can be found in Table 2.
14
3.2
Dynamic programming pruning functions
In the previous section, we analyzed several classes of linkage-based merge functions assuming a
fixed pruning function in the dynamic programming step of the standard linkage-based clustering
algorithm, i.e. Step 2 of Algorithm 3. In this section, we analyze an infinite class of dynamic programming pruning functions and derive comprehensive sample complexity guarantees for learning
the best merge function and pruning function in conjunction.
By allowing an application-specific choice of a pruning function, we significantly generalize
the standard linkage-based clustering algorithm framework. Recall that in the algorithm selection
model, we instantiated the cost function to be a generic clustering objective Φ. In the standard
clustering algorithm framework, where Φ is defined to be any general Φ(p) (which include objectives
like k-means), the best choice of the pruning function for the algorithm selector is Φ(p) itself as it
would return the optimal pruning of the cluster tree for that instantiation of cost. However, when
the goal of the algorithm selector is, for example, to provide solutions that are close to a ground
truth clustering for each problem instance, the best choice for the pruning function is not obvious.
In this case, we assume that the learning algorithm’s training data consists of clustering instances
that have been labeled by an expert according to the ground truth clustering. For example, this
ground truth clustering might be a partition of a set of images based on their subject, or a partition
of a set of proteins by function. On a fresh input data, we no longer have access to the expert or
the ground truth, so we cannot hope to prune a cluster tree based on distance to the ground truth.3
Instead, the algorithm selector must empirically evaluate how well pruning according to alternative objective functions, such as k-means or k-median, approximate the ground truth clustering
on the labeled training data. In this way, we instantiate cost to be the distance of a clustering
from the ground truth clustering. We guarantee that the empirically best pruning function from
a class of computable objectives is near-optimal in expectation over new problem instances drawn
from the same distribution as the training data. Crucially, we are able to make this guarantee even
though it is not possible to compute the cost of the algorithm’s output on these fresh instances
because the ground truth clustering is unknown.
Along these lines, we can also handle the case where the training data consists of clustering
instances, each of which has been clustered according to an objective function that is NP-hard
to compute. In this scenario, our learning algorithm returns a pruning objective function that is
efficiently computable and which best approximates the NP-hard objective on the training data,
and therefore will best approximate the NP-hard objective on future data. Hence, in this section,
we analyze a richer class of algorithms defined by a class of merge functions and a class of pruning
functions. The learner now has to learn the best combination of merge and pruning functions from
this class.
To define this more general class of agglomerative clustering algorithms, let A denote a generic
class of linkage-based merge functions (such as any of the classes Ai defined in Sections 3.1 and C)
parameterized by α. We also
define a rich class of center-based clustering objectives for the dynamic
(p)
programming step: F = Ψ | p > 0 where Ψ(p) takes as input a partition C = {C1 , C2 , . . . , Ck0 }
of n0 points and a set of centers c = {c1 , c2 , . . . , ck0 } such that ci ∈ Ci . The function Ψ(p) is defined
3
If Φ is the distance to ground truth clustering, then Φ cannot be directly measured when the clustering algorithm
is used on new data. However, we assume that the learning algorithm has access to training data which consists of
clustering instances labeled by the ground truth clustering. The learning algorithm uses this data to optimize the
parameters defining the clustering algorithm family. With high probability, on a new input drawn from the same
distribution as the training data, the clustering algorithm will return a clustering that is close to the unknown ground
truth clustering.
15
Figure 4: Cluster tree corresponding to Table 1.
1
2
3
Clusters
Centers
Clusters
Centers
Clusters
Centers
A
B
C
D
E
A, B
D, E
A, B, C
A, B, C, D, E
{A}
{A}
{B}
{B}
{C}
{C}
{D}
{D}
{E}
{E}
{A, B}
{A}
{A}, {B}
{A}, {B}
{D, E}
{E}
{D}, {E}
{D}, {E}
{A, B, C}
{C}
{A, B}, {C}
{A, C}
{A}, {B}, {C}
{A, B, C}
{A, B, C, D, E}
{C}
{A, B, C}, {D, E}
{C, E}
{A, B}, {C}, {D, E}
{A, C, E}
Table 1: Example dynamic programming table corresponding to the cluster tree in Figure 4 for
k = 3.
such that
(p)
Ψ
(C, c) =
sX X
p
(d(q, ci ))p .
(2)
Ci ∈C q∈Ci
Note that the definition of Ψ(p) is identical to Φ(p) , but we use this different notation so as not to
confuse the dynamic programming function with the clustering objective function. Let A(α) denote
(p)
the α-linkage merge function from A and F(p) denote the pruning function
Ψ . Earlier, for an
abstract objective Φ, we bounded the pseudodimension of HA×{Ψ},Φ = Φ(A(α),Ψ) : V → R≥0 ,
where Φ(A(α),Ψ) (V) denoted the cost of the clustering produced by building the cluster tree on V
using the merge function A(α) and then pruning the tree using a fixed pruning function Ψ. Now, we
are interested in doubly-parameterized algorithms of the form (A(α), F(p)) which uses the merge
function A(α) to build a cluster tree and then use the pruning function F(p) to prune it. To
analyze the resulting class of algorithms,
which we will denote by A × F, we have to bound the
pseudodimension of HA×F ,Φ = Φ(A(α),F (p)) : V → R≥0 p > 0 .
Recall that in order to show that pseudodimension of HA×{Ψ},Φ is upper bounded by dHA , we
proved that,
given
a sample of m clustering instances over n nodes, we can split the real line into at
dHA
most O m2
intervals such that as α ranges over a single interval, the m cluster trees returned
by the α-linkage merge function are fixed. To extend this analysis to HA×F ,Φ , we first prove a
similar fact in Lemma 11. Namely, given a single cluster tree, we can split the real line into a fixed
number of intervals such that as p ranges over a single interval, the pruning returned by using the
function Ψ(p) is fixed. We then show in Theorem 8 how to combine this analysis of the rich class
of dynamic programming algorithms with our previous analysis of the possible merge functions to
obtain a comprehensive analysis of agglomerative algorithms with dynamic programming.
We visualize the dynamic programming step of Algorithm 3 with pruning function Ψ(p) using
a table such as Table 1, which corresponds to the cluster tree in Figure 4. Each row of the
table corresponds to a sub-clustering value k 0 ≤ k, and each column corresponds to a node of the
corresponding cluster tree. In the column corresponding to node T and the row corresponding to
the value k 0 , we fill in the cell with the partition of
T into k 0 clusters that corresponds to the best
k 0 -pruning of the subtree rooted at T , CT,k0 , cT,k0 as defined in Step 2 of Algorithm 3.
16
Figure
5: Partition of the positive real line based on whether or not
P
p
q∈T (d(q, v2 )) as p ranges R>0 .
P
p
q∈T (d(q, v1 ))
≤
Lemma 11. Given a cluster tree T for a clustering instance V = (V, d) of n points, the positive
real line can be partitioned into a set I of O(n2(k+1) k 2k ) intervals such that for any I ∈ I, the
cluster tree pruning according to Ψ(p) is identical for all p ∈ I.
Proof. To prove this claim, we will examine the dynamic programming (DP) table corresponding
to the given cluster tree and the pruning function Ψ(p) as p ranges over the positive real line. As
the theorem implies, we will show that we can split the positive real line into a set of intervals so
that on a fixed interval I, as p ranges over I, the DP table under Ψ(p) corresponding to the cluster
tree is invariant. No matter which p ∈ I we choose, the DP table under Ψ(p) will be identical, and
therefore the resulting clustering will be identical. After all, the output clustering is the bottomright-most cell of the DP table since that corresponds to the best k-pruning of the node containing
all points (see Table 1 for an example). We will prove that the total number of intervals is bounded
by O(n2(k+1) k 2k ).
We will prove this lemma using induction on the row number k 0 of the DP table. Our induc0
tive
will be the following. The positive real line can be partitioned into a set I (k ) of
hypothesis
Q 0
0
0
0
O n2 kj=1 n2 j intervals such that for any I (k ) ∈ I (k ) , as p ranges over I (k ) , the first k 0 rows of
the DP table corresponding to Ψ(p) are
Notice
that this means that the positive real line
invariant.
Qk
2
2
2
= O n2(k+1) k 2k intervals such that for any
can be partitioned into a set I of O n
j=1 n j
I ∈ I, as p ranges over I, the DP table corresponding to Ψ(p) is invariant. Therefore, the resulting
output clustering is invariant as well.
Base case (k 0 = 1). Let p be a positive real number. Consider the first row of the DP table
corresponding to Ψ(p) . Recall that each column in the DP table corresponds to a node T in the
clustering tree where T ⊆ V . In the first row of the DP table and the column corresponding to node
T , we fill in the cell with the single node T and the point c ∈ T which minimizes Ψ(p) ({T }, {c}) =
P
p
q∈T (d(q, c)) . The only thing that might change as we vary p is the center minimizing this
objective.
Let v1 and v2 be two points in T . The point v1 is a better candidate for the center of T
than v2 if and only if Ψ(p) ({T }, {v1 }) ≤PΨ(p) ({T }, {v2 }) which
means that Ψ(p) ({T }, {v1 }) −
P
(p)
p
p
Ψ
0, or in other words,
q∈T (d(q, v1 )) −
q∈T (d(q, v2 )) ≤ 0. The equation
P ({T }, {v2 })p ≤ P
p
q∈T (d(q, v1 )) −
q∈T (d(q, v2 )) has at most 2|T | zeros, so there are at most 2|T | + 1 intervals
I1 , . . . , It which partition the positive real line such that for any Ii , as p ranges over Ii , whether or
not Ψ(p) ({T }, {v1 }) ≤ Ψ(p) ({T }, {v2 }) is fixed. For example, see Figure 5. Every pair of points in T
similarly partitions the positive real line into 2|T | + 1 intervals. If we merge all |T |2 /2 partitions —
2
one partition for each pair of points in T — then we are left with at most |T2| · 2|T | + 1 = |T |3 + 1
intervals I1 , . . . , Iw partitioning the positive real line such that for any Ii , as p ranges over Ii , the
point v ∈ T which minimizes Ψ(p) ({T }, {v}) is fixed.
17
Since T is arbitrary, we can thus partition the real line for each node T 0 in the cluster tree.
Again, this partition defines the center of the cluster T 0 as p ranges
Pover the3 positive real4 line. If we
merge the partition for every node T ∈ T , then we are left with
T ∈T |T | + 1 = O(n ) intervals
I1 , . . . , I` such that as p ranges over any one interval Ii , the centers of all nodes in the cluster tree
are fixed. In other words, for each T , the point vi ∈ T which minimizes Ψ(p) ({T }, {vi }) is fixed.
Of course, this means that the first row of the DP table is fixed as well. Therefore, the inductive
hypothesis holds for the base case.
Inductive step. Consider the k 0 th row of the DP table. We knowfrom the inductive
hypothesis
Qk0 −1 2 2
0 −1)
(k
2
that the positive real line can be partitioned into a set I
of O n
intervals such
j=1 n j
0
0
0
that for any I (k −1) ∈ I (k −1) , as p ranges over I (k −1) , the first k 0 − 1 rows of the DP table
corresponding to Ψ(p) are invariant.
0
0
Fix some interval I (k −1) ∈ I (k −1) . Let T be a node in the cluster tree T and let TL and
TR be the left and right children of T in T respectively. Notice that the pruning which belongs in the cell in the ith row and the column corresponding to T does not depend on the other
cells in the ith row, but only on the cells in rows 1 through i − 1. In particular, the pruning
which belongs in this cell depends on the
inequalities defining which i0 ∈ {1, . . . , k 0 − 1} minimizes
Ψ(p) CTL ,i0 ∪ CTR ,k0 −i0 , cTL ,i0 ∪ cTR ,k0 −i0 . We will now examine this objective function and show
that the minimizing i0 , and therefore the optimal pruning, only changes a small number of times
0
as p ranges over I (k −1) .
For an arbitrary i0 ∈ {1, . . . , k 0 − 1}, since i0 and k 0 − i0 are both strictly less than k 0 , the best
0
i -pruning of TL (CTR ,i0 , cTR ,i0 ) is exactly the entry in the i0 th row of the DP table and the column
corresponding to TL . Similarly, the best k 0 − i0 -pruning of TR , (CTR ,k0 −i0 , cTR ,k0 −i0 ) is exactly the
entry in the k 0 − i0 th row of the DP table and the column corresponding to TR . Crucially, these
0
entries do not change as we vary p ∈ I (k −1) , thanks to the inductive hypothesis.
0
Therefore, for any i0 , i00 ∈ {1, . . . , k 0 − 1}, we know that for all p ∈ I (k −1) , the k 0 -pruning of
T corresponding to the combination of the best i0 -pruning of TL and the best k 0 − i0 pruning of
TR is fixed and can be denoted as (C 0 , c0 ). Similarly, the k 0 -pruning of T corresponding to the
combination of the best i00 -pruning of TL and the best k 0 − i00 pruning of TR is fixed and can
0
be denoted as (C 00 , c00 ). Then, for any p ∈ I (k −1) , (C 0 , c0 ) is a better pruning than (C 00 , c00 ) if
and only if Ψ(p) (C 0 , c0 ) ≤ Ψ(p) (C00 , c00 ). In order to analyze this inequality,
let us consider
p the
p
p
p
equivalent inequality Ψ(p) (C 0 , c0 ) ≤ Ψ(p) (C 00 , c00 ) i.e., Ψ(p) (C 0 , c0 ) − Ψ(p) (C 00 , c00 ) ≤ 0.
Now, to expand this expression let C 0 = {C10 , C20 , . . . , Ck0 0 } and c0 = {c01 , c02 . . . , c0k0 } and similarly
C 00 = {C100 , C200 , . . . , Ck0 0 } and c00 = {c001 , c002 . . . , c00k0 }. Then, this inequality can then be written as,
0
k X
X
i=1 q∈Ci0
The equation
0
Pk0 P
i=1
q∈Ci0
p
d(q, c0i )
(d(q, c0i ))p −
0
−
k X
X
i=1 q∈Ci00
Pk 0 P
i=1
q∈Ci00
d(q, c00i )
p
≤ 0.
(d(q, c00i ))p has has at most 2n zeros as p
0
ranges over I (k −1) . Therefore, there are at most 2n + 1 subintervals partitioning I (k −1) such that
as p ranges over one subinterval, the smaller of Ψ(p) (C 0 , c0 ) and Ψ(p) (C 00 , c00 ) is fixed. In other
words, as p ranges over one subinterval, either the combination of the best i0 -pruning of T ’s left
child and the best (k 0 − i0 )-pruning of T ’s right child is better than the combination of the best
i00 -pruning of T ’s left child with the best (k − i00 )-pruning of T ’s right child, or vice versa. For
all pairs i0 , i00 ∈ {1, . . . , k 0 − 1}, we can similarly partition I into at most 2n + 1 subintervals
0
defining the better of the two prunings. If we merge all (k 0 − 1)2 /2 partitions of I (k −1) , we have
0
2
0
(k −1)
· 2n + 1 = (k 0 − 1)2 n + 1 total subintervals of I (k −1) such that as p ranges over a single
2
18
subinterval,
argmini0 ∈{1,...,k0 −1} Ψ(p) CTL ,i0 ∪ CTR ,k0 −i0 , cTL ,i0 ∪ cTR ,k0 −i0
is fixed. Since these equations determine the entry in the ith row of the DP table and the column
corresponding to the node T , we have that this entry is also fixed as p ranges over a single subinterval
0
in I (k −1) .
0
The above partition of I (k −1) corresponds to only a single cell in the k 0 th row of the DP table.
Considering the k 0 th row of the DP table as a whole, we must fill in at most 2n entries, since there
are at most 2n columns of the DP table. For each column, there is a corresponding partition of
0
I (k −1) such that as p ranges over a single subinterval in the partition, the entry in the k 0 th row
0
and that column is fixed. If we merge all such partitions, we are left with a partition of I (k −1)
consisting of at most 2n2 (i − 1)2 + 1 intervals such that as p ranges over a single interval, the entry
0
in every column of the k 0 th row is fixed. As these intervals are subsets of I (k −1) , by assumption,
the first k 0 − 1 rows of the DP table are also fixed. Therefore, the first k 0 rows are fixed.
0
0
0
To recap, we fixed an interval I (k −1) such that as p ranges overI (k −1) , the first
k − 1 rows of
Q
0 −1
k
the DP table are fixed. By the inductive hypothesis, there are O n2 j=1 n2 j 2 such intervals.
0
Then, we showed that I (k −1) can be partitioned into 2n2 (k 0 − 1)2 + 1 intervals such that for any
(k0 )
(k0 )
0
one subinterval
IQ 0 , as p ranges over I , the first k rows of the DP table are fixed. Therefore,
k
there are O n2 j=1 n2 j 2 total intervals such that as p ranges over a single interval, the first k 0
rows of the DP table are fixed.
Aggregating this analysis over all k rows of the DP table, we have that there are
!
k
Y
2
= O n2(k+1) k 2k
O n2
n2 k 0
k=1
intervals such that the entire DP table is fixed so long as p ranges over a single interval.
We are now ready to prove our main theorem in this section.
Theorem 8. Given m clustering instances, suppose there are O m · 2O(dHA ) intervals that partition the domain of α such that as α ranges over a single interval, the m cluster trees returned
by the α-linkage merge function from A are fixed. Then, for all sets of m samples S ∼ Dm ,
P dim (HA×F ,Φ ) = O (dHA + k log n).
Proof. Let S be a set of clustering instances. Fix a single interval of α (as shown along the horizontal
axis in Figure 6) where the set of cluster trees returned by the α-linkage merge function from A
is fixed across all samples. We know from Lemma 11 that we can split the real line into a fixed
number of intervals such that as p ranges over a single interval (as shown along the vertical axis
in Figure 6), the dynamic programming (DP) table is fixed for all the samples, and therefore the
resulting set of clusterings
is fixed. In particular, for a fixed α interval, each of the m samples has
its own O n2(k+1) k 2k intervals of p, and when we merge them, we are left with O mn2(k+1) k 2k
intervals such that as p ranges over a single interval, each DP table for each sample
is fixed,
and
therefore the resulting clustering for each sample is fixed. Since there are O m2dHA such α
intervals, each inducing O mn2(k+1) k 2k such p intervals in total, we have O 2dHA m2 n2(k+1) k 2k
cells in R2 such that if (α, p) is in one fixed cell, the resulting
clustering across
all samples is
dHA 2 2(k+1) 2k
m
fixed. If HA×F shatters S, then it must be that 2 = O 2
m n
k , which means that
m = O log 2dHA n2(k+1) k 2k
= O (dHA + k log n).
19
O mn2(k+1) k 2k intervals of p
for a fixed α interval
p>0
Fixed set of pruned
cluster trees on samples S
in each cell
α
Fixed set of cluster trees
on samples S
for all α in this interval.
O(m · 2O(dHA ) ) intervals of α
Figure 6: Illustration of the partition of the parameter space as described in the proof of Theorem 9.
Theorem 9. Given a sample of size m = O (H/)2 (dHA + log n + log(1/δ)) and a clustering
objective Φ, it is possible to (, δ)-learn the class of algorithms A×F with respect to the cost function
Φ. Moreover, this procedure is efficient if the following conditions hold:
1. k is constant, which ensures that the partition of p values is polynomial in n.
2. 2dHA is polynomial in n, which ensures that the partition of α values is polynomial in n.
3. It is possible to efficiently compute the partition of α into intervals so that on a single interval
I, for all α ∈ I, the m cluster trees returned by α-linkage performed on S are fixed.
Proof. A technique for finding the empirically best algorithm from A × F follows naturally from
Lemma 11; we partition the range of feasible values of α as described in Section 3.1, and for each
resulting interval of α, we find the fixed set of cluster trees on the samples. We then partition
the values of p as discussed in the proof for Lemma 11. For each interval of p, we use Ψ(p) to
prune the trees and determine the fixed empirical cost corresponding to that interval of p and
α. This is illustrated in Figure 6. Iterating over all partitions of the parameter space, we can
find parameters with the best empirical cost. In Theorem 8, we use Lemma 11 to show that
Pdim(HA×F ,Φ ) = O (dHA + k log n) and thus arrive at our sample complexity bound when k is
constant.
4
Discussion and open questions
In this work, we show how to learn near-optimal algorithms over several infinite, rich classes of
SDP rounding algorithms and agglomerative clustering algorithms with dynamic programming.
We provide computationally efficient and sample efficient learning algorithms for many of these
problems and we push the boundaries of learning theory by developing techniques to compute the
pseudo-dimension of intricate, multi-stage classes of IQP approximation algorithms and clustering
algorithms. We derive tight pseudo-dimension bounds for the classes we study, which lead to strong
20
sample complexity guarantees. We hope that our techniques will lead to theoretical guarantees in
other areas where empirical methods for algorithm configuration have been developed.
There are many open avenues for future research in this area. In this work, we focused on
algorithm families containing only computationally efficient algorithms. However, oftentimes in
empirical AI research, the algorithm families in question contain procedures that are too slow to
run to completion on many training instances. In this situation, we would not be able to determine
the exact empirical cost of an algorithm on the training set. Could we still make strong, provable
guarantees for application-specific algorithm configuration in this scenario? This work also leaves
open the potential for data-dependent bounds over well-behaved distributions, such as those over
clustering instances satisfying some form of stability, be it approximation stability, perturbation
resilience, or so on.
Acknowledgments. This work was supported in part by grants NSF-CCF 1535967, NSF CCF1422910, NSF IIS-1618714, a Sloan Fellowship, a Microsoft Research Fellowship, a NSF Graduate
Research Fellowship, a Microsoft Research Women’s Fellowship, and a National Defense Science
and Engineering Graduate (NDSEG) fellowship.
We thank Sanjoy Dasgupta, Anupam Gupta, and Ryan O’Donnell for useful discussions.
References
[1] Noga Alon, Konstantin Makarychev, Yury Makarychev, and Assaf Naor. Quadratic forms on
graphs. Inventiones mathematicae, 163(3):499–522, 2006.
[2] Martin Anthony and Peter L Bartlett. Neural network learning: Theoretical foundations.
Cambridge University Press, 2009.
[3] Pranjal Awasthi, Maria-Florina Balcan, and Konstantin Voevodski. Local algorithms for interactive clustering. In Proceedings of the International Conference on Machine Learning (ICML),
pages 550–558, 2014.
[4] Pranjal Awasthi, Avrim Blum, and Or Sheffet. Center-based clustering under perturbation
stability. Information Processing Letters, 112(1):49–54, 2012.
[5] Maria-Florina Balcan, Nika Haghtalab, and Colin White. k-center clustering under perturbation resilience. In Proceedings of the Annual International Colloquium on Automata, Languages, and Programming (ICALP), 2016.
[6] Maria-Florina Balcan and Yingyu Liang. Clustering under perturbation resilience. SIAM
Journal on Computing, 45(1):102–155, 2016.
[7] Afonso S. Bandeira, Nicolas Boumal, and Vladislav Voroninski. On the low-rank approach for
semidefinite programs arising in synchronization and community detection. In Proceedings of
the Conference on Learning Theory (COLT), pages 361–382, 2016.
[8] MohammadHossein Bateni, Aditya Bhaskara, Silvio Lattanzi, and Vahab Mirrokni. Distributed balanced clustering via mapping coresets. In Proceedings of the Annual Conference
on Neural Information Processing Systems (NIPS), pages 2591–2599, 2014.
[9] Ahron Ben-Tal and Arkadi Nemirovski. Lectures on modern convex optimization: analysis,
algorithms, and engineering applications, volume 2. Siam, 2001.
21
[10] William Brendel and Sinisa Todorovic. Segmentation as maximum-weight independent set.
In Proceedings of the Annual Conference on Neural Information Processing Systems (NIPS),
pages 307–315, 2010.
[11] Fazli Can. Incremental clustering for dynamic information processing. ACM Transactions on
Information Systems (TOIS), 11(2):143–164, 1993.
[12] Yves Caseau, François Laburthe, and Glenn Silverstein. A meta-heuristic factory for vehicle routing problems. In International Conference on Principles and Practice of Constraint
Programming (CP), pages 144–158. Springer, 1999.
[13] Moses Charikar, Chandra Chekuri, Tomás Feder, and Rajeev Motwani. Incremental clustering
and dynamic information retrieval. In Proceedings of the Annual Symposium on Theory of
Computing (STOC), pages 626–635, 1997.
[14] Moses Charikar and Anthony Wirth.
Maximizing quadratic programs: extending
Grothendieck’s inequality. In Proceedings of the Annual Symposium on Foundations of Computer Science (FOCS), pages 54–60, 2004.
[15] Timothee Cour, Praveen Srinivasan, and Jianbo Shi. Balanced graph matching. In Proceedings
of the Annual Conference on Neural Information Processing Systems (NIPS), pages 313–320,
2006.
[16] Jim Demmel, Jack Dongarra, Victor Eijkhout, Erika Fuentes, Antoine Petitet, Rich Vuduc,
R Clint Whaley, and Katherine Yelick. Self-adapting linear algebra algorithms and software.
Proceedings of the IEEE, 93(2):293–312, 2005.
[17] R.M Dudley. The sizes of compact subsets of Hilbert space and continuity of Gaussian processes. Journal of Functional Analysis, 1(3):290 – 330, 1967.
[18] Uriel Feige and Michael Langberg. The RPR2 rounding technique for semidefinite programs.
Journal of Algorithms, 60(1):1–23, 2006.
[19] Darya Filippova, Aashish Gadani, and Carl Kingsford. Coral: an integrated suite of visualizations for comparing clusterings. BMC bioinformatics, 13(1):276, 2012.
[20] Roy Frostig, Sida Wang, Percy S Liang, and Christopher D Manning. Simple MAP inference
via low-rank relaxations. In Proceedings of the Annual Conference on Neural Information
Processing Systems (NIPS), pages 3077–3085, 2014.
[21] Michel X Goemans and David P Williamson. Improved approximation algorithms for maximum cut and satisfiability problems using semidefinite programming. Journal of the ACM
(JACM), 42(6):1115–1145, 1995.
[22] Anna Grosswendt and Heiko Roeglin. Improved analysis of complete linkage clustering. In
European Symposium of Algorithms, volume 23, pages 656–667. Springer, 2015.
[23] Anupam Gupta and Kanat Tangwongsan. Simpler analyses of local search algorithms for
facility location. arXiv preprint arXiv:0809.2554, 2008.
[24] Rishi Gupta and Tim Roughgarden. A PAC approach to application-specific algorithm selection. In Proceedings of the 2016 ACM Conference on Innovations in Theoretical Computer
Science (ITCS), pages 123–134, 2016.
22
[25] Qixing Huang, Yuxin Chen, and Leonidas Guibas. Scalable semidefinite relaxation for maximum a posterior estimation. In Proceedings of the International Conference on Machine
Learning (ICML), pages 64–72, 2014.
[26] Fredrik D Johansson, Ankani Chattoraj, Chiranjib Bhattacharyya, and Devdatt Dubhashi.
Weighted theta functions and embeddings with applications to max-cut, clustering and summarization. In Proceedings of the Annual Conference on Neural Information Processing Systems
(NIPS), pages 1018–1026, 2015.
[27] Subhash Khot, Guy Kindler, Elchanan Mossel, and Ryan O’Donnell. Optimal inapproximability results for MAX-CUT and other 2-variable CSPs? SIAM Journal on Computing, 37(1):319–
357, 2007.
[28] Kevin Leyton-Brown, Eugene Nudelman, and Yoav Shoham. Empirical hardness models:
Methodology and a case study on combinatorial auctions. Journal of the ACM (JACM),
56(4):22, 2009.
[29] Marina Meilă. Comparing clusterings: an information based distance. Journal of multivariate
analysis, 98(5):873–895, 2007.
[30] Ryan O’Donnell and Yi Wu. An optimal SDP algorithm for max-cut, and equally optimal
long code tests. In Proceedings of the Annual Symposium on Theory of Computing (STOC),
pages 335–344, 2008.
[31] David Pollard. Convergence of stochastic processes. Springer-Verlag, 1984.
[32] David Pollard. Empirical processes. Institute of Mathematical Statistics, 1990.
[33] John R Rice. The algorithm selection problem. Advances in computers, 15:65–118, 1976.
[34] Andrej Risteski and Yuanzhi Li. Approximate maximum entropy principles via GoemansWilliamson with applications to provable variational methods. In Proceedings of the Annual
Conference on Neural Information Processing Systems (NIPS), pages 4628–4636, 2016.
[35] Mehreen Saeed, Onaiza Maqbool, Haroon Atique Babri, Syed Zahoor Hassan, and S Mansoor
Sarwar. Software clustering techniques and the use of combined algorithm. In Proceedings of
the European Conference on Software Maintenance and Reengineering, pages 301–306. IEEE,
2003.
[36] Sagi Snir and Satish Rao. Using max cut to enhance rooted trees consistency. IEEE/ACM
Transactions on Computational Biology and Bioinformatics (TCBB), 3(4):323–333, 2006.
[37] Timo Tossavainen. On the zeros of finite sums of exponential functions. Australian Mathematical Society Gazette, 33(1):47–50, 2006.
[38] Vijay V Vazirani. Approximation algorithms. Springer Science & Business Media, 2013.
[39] Jun Wang, Tony Jebara, and Shih-Fu Chang. Semi-supervised learning using greedy max-cut.
Journal of Machine Learning Research, 14(Mar):771–800, 2013.
[40] James R White, Saket Navlakha, Niranjan Nagarajan, Mohammad-Reza Ghodsi, Carl Kingsford, and Mihai Pop. Alignment and clustering of phylogenetic markers-implications for microbial diversity studies. BMC bioinformatics, 11(1):152, 2010.
23
[41] David P Williamson and David B Shmoys. The design of approximation algorithms. Cambridge
University press, 2011.
[42] Lin Xu, Frank Hutter, Holger H. Hoos, and Kevin Leyton-Brown. SATzilla: portfolio-based
algorithm selection for SAT. Journal of Artificial Intelligence Research, 32:565–606, June 2008.
[43] Chihiro Yoshimura, Masanao Yamaoka, Masato Hayashi, Takuya Okuyama, Hidetaka Aoki,
Ken-ichi Kawarabayashi, and Hiroyuki Mizuno. Uncertain behaviours of integrated circuits
improve computational performance. Scientific reports, 5, 2015.
[44] Mingjun Zhong, Nigel Goddard, and Charles Sutton. Signal aggregate constraints in additive
factorial HMMs, with application to energy disaggregation. In Proceedings of the Annual
Conference on Neural Information Processing Systems (NIPS), pages 3590–3598, 2014.
[45] Uri Zwick. Outward rotations: a tool for rounding solutions of semidefinite programming
relaxations, with applications to max cut and other problems. In Proceedings of the Annual
Symposium on Theory of Computing (STOC), pages 679–687, 1999.
A
Proofs from Section 2
Theorem 2. Suppose that Lslin is an algorithm that takes
as input m samples (A(i) , Z(i) ) ∼
1 Pm
D × Z and returns the parameter ŝ which maximizes m i=1 slins (A(i) , Z(i) ). Further, supposePthat m is sufficiently large to ensure that with probability at least 1 − δ, for all s > 0,
m
1
(i)
(i)
|m
i=1 slins (A , Z ) − E(A,Z) [slins (A, Z)] | ≤ 2 . Then Lslin (, δ)-learns the class of s-linear
rounding functions with respect to the cost function −slin∗s .
Proof. Theorem 2 follows directly from Lemma 12.
Lemma 12. Suppose that m is sufficiently large to ensure that with probability at least 1 − δ over
a draw of m samples A(i) , Z(i) ∼ D × Z, for all s-linear functions φs ,
1 X
E
[slins (A, Z)] < .
slins A(i) , Z(i) −
m
2
(A,Z)∼D×Z
m
i=1
1 Pm
(i)
(i) and s∗ maximizes
Then with probability at least 1 − δ, if ŝ maximizes m
i=1 slins A , Z
E [slins (A)], then E [slins∗ (A)] − E [slinŝ (A)] < .
A∼D
A∼D
A∼D
Proof. Notice that since D × Z is a product distribution, we have that
E
[slins (A, Z)] = E
E [slins (A, Z)] = E [slins (A)] ,
A∼D Z∼Z
(A,Z)∼D×Z
A∼D
so we know that with probability at least 1 − δ, for all s-linear functions φs ,
1 X
slins A(i) , Z(i) − E [slins (A)] < .
A∼D
m
2
m
i=1
By assumption, we know that with probability at least 1 − δ,
1 X
slinŝ A(i) , Z(i) − E [slinŝ (A)] <
A∼D
m
2
m
i=1
24
and
1 X
slins∗ A(i) , Z(i) − E [slins∗ (A)] < ,
A∼D
m
2
m
i=1
which means that E [slins∗ (A)] − E [slinŝ (A)] < .
A∼D
A∼D
Lemma 4. Pdim(Hslin ) = Ω(log n).
Proof. In order to prove
that the
pseudo dimension
of Hslin is at least c log n for some c, we must
present a set S = G(1) , Z(1) , . . . , G(m) , Z(m)
of m = c log n graphs and projection vectors
that can be shattered by Hslin . In other words, there exist m witnesses r1 , . . . , rm and 2m = nc s
values H = {s1 , . .. , snc } such that for all T ⊆ [m], there exists
sT ∈ H such that if j ∈ T , then
(j)
(j)
(j)
(j)
slinST G , Z
> rj and if j 6∈ T , then slinST G , Z
≤ rj .
To build S, we will use the same graph G for all G(j) and we will vary Z(j) . We set G to be the
graph composed of bn/4c disjoint copies of K4 . If n = 4, then a simple calculation confirms that
an optimal max-cut SDP embedding of G is
−1/3
−1/3
−1/3
1
√
√
√
− 2/3 − 2/3
0
2/3
2
, p
, p
.
,
0 0 2/3 − 2/3
0
0
0
0
Therefore, for n > 4, an optimal embedding is the set of n vectors SDP (G) such that for all
i ∈ {0, . . . , bn/4c − 1},
r
r
√
√
√
2
2
2
2
1
2 2
1
1
e4i+1 , − e4i+1 +
e4i+2 , − e4i+1 −
e4i+2 +
e4i+3 , − e4i+1 −
e4i+2 −
e4i+3
3
3
3
3
3
3
3
3
are elements SDP (G).
We now define the set of m vectors Z(j) . First, we set Z(1) to be the vector
Z(1) = 70 , 5 · 70 , 5 · 70 , 70 , 71 , 5 · 71 , 5 · 71 , 71 , 72 , 5 · 72 , 5 · 72 , 72 , 73 , 5 · 73 , 5 · 73 , 73 , . . . .
In other words, it is the concatenation the vector 7i (1, 5, 5, 1) for all i > 0. Next, Z(2) is defined as
Z(2) = 70 , 5 · 70 , 5 · 70 , 70 , 0, 0, 0, 0, 72 , 5 · 72 , 5 · 72 , 72 , 0, 0, 0, 0, . . . ,
so Z(2) is the same as Z(1) for all even powers of 7, and otherwise its entries are 0. In a similar
vein,
Z(3) = 70 , 5 · 70 , 5 · 70 , 70 , 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 74 , 5 · 74 , 5 · 74 , 74 , . . . .
j−1
To pin down this pattern, we set Z(j) to be the same as Z(1) for all entries of the form 7i2 (1, 5, 5, 1)
for i ≥ 0, and otherwise its entries are 0.
We set the following positive, increasing constants which will appear throughout the remaining
analysis:
a=
(1, 0, 0, 0) · (1, 5, 5, 1) = 1
√
p
p
√
b=
(−1/3, − 2/3, 2/3, 0) · (1, 5, 5, 1) = 5 √2/3 − 5 32+1
√
c=
(−1/3, 2 2/3, 0, 0) · (1, 5, 5, 1) = 10 32−1
√
p
p
√
d = (−1/3, − 2/3, 2/3, 0) · (1, 5, 5, 1) = 5 2/3 + 5 32+1 .
25
Figure 7: Depiction of slinG,Z(j) (s) as s increases from 0. A black dot means that slinG,Z(j) (s) ≥ rj
and a white dot means that slinG,Z(j) (s) < rj . Here, ai = 7i−1 c and bi = 7i−1 d.
We also set c̃ = b + c + bc − d − bd − cd and we claim that the witnesses
b
1
1
−1
r1 = −
2 3n c2
1
c̃
rj = −
j>1
j−1
2 3n72 −2 d2
are sufficient to prove that this set is shatterable, and we will spend the remainder of the proof
showing that this is true.
Now, the domain of slinG,Z(j) (s) can be split into intervals on which it has a simple, fixed form.
h j−1
j−1
These intervals begin at 1 and have the form 7i2 , 7(i+1)2
, for i ≥ 0. It is straightforward
h j−1
j−1
matter of calculations to check that for s ∈ 7i2 , 7(i+1)2
,
1
slinG,Z(j) (s) =
h P2
i
j−1
i−1 2k2j−1
72i2
1
1
7
+
c̃
−
1
2
k=0
3n s h
s
i
1 12 c̃ Pi−1 72k2j−1 + 72i2j−1 b − 1
k=0
3n s h
i
−
Pi−1 2k2j−1
1
1
2i2j−1 (b + c + bc) − y
7
+
7
c̃
2
k=0
3n s h
i
1 12 c̃ Pi 72k2j−1
k=0
3n s
2i2j−1
h j−1
j−1
if s ∈ 7i2 , 7i2 b
h j−1
j−1
if s ∈ 7i2 b, 7i2 c
h j−1
j−1
if s ∈ 7i2 c, 7i2 d
h j−1
j−1
if s ∈ 7i2 d, 7(i+1)2
,
(1+b+c)
where y = 7
. (We note here that the power of 7 pattern was chosen so that these
s
intervals are well defined, since 7i d < 7i+1 .)
We call the following increasing
sequence of numbers points of interest, which we use to prove
that this set is shattered: 70 c, 70 d, 71 c, 71 d, 72 c, 72 d, . . . , 7i c, 7i d, . . .
We make two claims about these points of interest:
1. slinG,Z(1) (s) is above its witness whenever s = 7i c and it is below its witness whenever
s = 7i d for i ≥ 0.
2. Let j > 1 and consider slinG,Z(j) (s). There are 2j points of interest per interval
h j−1
j−1
7i2 , 7(i+1)2
.
On the first half of these points of interest, slinG,Z(j) (s) is greater than its witness and on
the second half, slinG,Z(j) (s) is less than its witness.
These claims are illustrated by the dots in Figure 7. Together, these claims imply that S can be
shattered because for any vector b ∈ {0, 1}m , there exists a point of interest s such that slins (S)
induces the binary labeling b on S.
26
The first claim is true because
slinG,Z(1)
" i−1
#
!
X
1
c̃
72k + 72i b − 1
72i c2
k=0
1
72i − 1
1
1
2i
c̃ · 2
= −
+7 b −1 ,
2 3n 72i c2
7 −1
1
1
7i c = −
2 3n
which is an increasing function of i, so it is minimized when i = 0, where
1
b
1
0
slinG,Z(1) 7 c = −
− 1 = r1
2 3n c2
so slinG,Z(1) 7i c is always at least its witness. Further,
slinG,Z(1)
1
1
7i d = −
2 3n
1
1
−
2 3n
=
" i
#!
X
1
c̃
72k
72i d2
k=0
"
#!
1
72(i+1) − 1
c̃ ·
),
72i d2
48
which is again an increasing function in i, with a limit of
1
49c̃
−
< r1 .
2 144nd2
Therefore, slinG,Z(1) 7i d is always less than its witness, and we may conclude that the first claim
is always true.
For the second claim, notice that
j−1
7i2
j−1
c < 7i2
j−1 +1
j−1 +1
c < 7i2
j−1 +2
d < 7i2
j−1
j−1
j−1
c · · · < 7i2 +2 c = 7(i+1)2 c,
h j−1
j−1
so there are 2j points of interest per interval 7i2 c, 7(i+1)2 c , as claimed. The first two points
j−1
d < 7i2
j−1
of interest, 7i2 c and 7i2 d, fall in an interval where slinG,Z(j) is decreasing in s. Therefore, it
j−1
is minimized when s = 7i2 d, where
" i
#!
j−1 1
X
1
1
j−1
slinG,Z(j) 7i2 d = −
c̃
72k2
2 3n 72i2j−1 d2
k=0
"
#!
j
1
1
7(i+1)2 − 1
1
.
= −
c̃ ·
2 3n 72i2j−1 d2
72j − 1
j−1
Simple calculations show that slinG,Z(j) 7i2 d is an increasing function in i, so it is minimized
when i = 0, where slinG,Z(j) (d) =
1
2
−
c̃
3nd2
> rj , as desired.
h j−1
j−1
The remaining points of interest fall in the interval 7i2 d, 7(i+1)2
, so slinG,Z(j) (s) has
h P
i
j−1
1
1
c̃ ik=0 72k2
. This segment of the function has a negative derivative, so
the form 12 − 3n
s2
it is decreasing.
j−1
j−1
If j = 2, then the points of interest weh already considered,
7i2 c and 7i2 d, make up half
j−1
j−1
of the 2j points of interest in the interval 7i2 c, 7(i+1)2 c . Therefore, we only need to show
27
j−1 +1
that when s equals 7i2
j−1 +1
c and 7i2
d, then slinG,Z(j) (s) is less than its witness. As we saw,
j−1
slinG,Z(j) is decreasing on this segment, so it is enough to show that slinG,Z(j) 7i2 +1 c is less
than its witness. To this end,
" i
#!
j−1
1
X
1
1
i2
+1
2k2j−1
slinG,Z(j) 7
c = −
c̃
7
2 3n 72i2j−1 +2 c2
k=0
"
#!
j
1
1
7(i+1)2 − 1
1
c̃ ·
.
= −
2 3n 72i2j−1 +2 c2
72j − 1
This is an increasing function of i with a limit of
Therefore, when s equals 7
i2j−1 +1
b and 7
i2j−1 +1
1
2
−
1
3n
1
72 c2
h
j
72
j
72 −1
c̃ ·
i
< rj when j = 2.
c, then slinG,Z(j) (s) is less than its witness.
h j−1
j−1
Finally, if j > 2, since slinG,Z(j) (s) is decreasing on the interval 7i2 c, 7(i+1)2
, we must
j−1 j−2
th
only check that at the 2j−1 − 1
point of interest 7i2 +2 −1 d , slinG,Z(j) (s) is greater than
j−1 j−2
th
its witness and at the 2j−1 + 1
point of interest 7i2 +2 c , slinG,Z(j) (s) is less than its
witness. To this end,
" i
#!
j−1 j−2
1
X
1
1
i2
+2
−1
2k2j−1
d = −
slinG,Z(j) 7
c̃
7
2 3n 7i2j +2j−1 −2 d2
k=0
"
#!
j
1
1
7(i+1)2 − 1
1
c̃ ·
.
= −
2 3n 7i2j +2j−1 −2 d2
72j − 1
This function is increasing in i, so it is minimized when i = 0, where
"
#!
j
j−2
1
1
1
72 − 1
2
−1
d = −
slinG,Z(j) 7
c̃ · 2j
2 3n 72j−1 −2 d2
7 −1
1
c̃
1
= −
= rj .
j−1 −2 2
2
2 3n 7
d
j−1 j−2
Therefore, slinG,Z(j) 7i2 +2 −1 d ≥ rj for all i.
Next,
i2j−1 +2j−2
slinG,Z(j) 7
1
1
c = −
2 3n
1
1
= −
2 3n
1
7i2j +2j−1 c2
1
7i2j +2j−1 c2
which is an increasing function in i, with a limit of
j−1
1
c̃72
−
2 3nc2 72j − 1
28
"
"
c̃
i
X
2k2j−1
7
k=0
j
#!
7(i+1)2 − 1
c̃ ·
72j − 1
#!
.
as i tends toward infinity. Therefore,
j−1
j−1 j−2 1
c̃72
< rj
slinG,Z(j) 7i2 +2
≤ −
2 3nc2 72j − 1
for all i, so the second claim holds.
B
More algorithm classes for MaxQP
B.1
˜-discretized functions for max-cut
The class of ˜-discretized rounding functions are a finite yet rich class of functions for the RPR2
paradigm. They were introduced by O’Donnell and Wu as a tool for characterizing the SDP gap
curve for the max-cut problem, which we define shortly in order to describe O’Donnell and Wu’s
guarantees for RPR2 using ˜-discretized rounding functions. However, we first define the max-cut
SDP value of a graph G to be
X
1 1
wij
Sdp(G) = max
,
− g(vi ) · g(vj )
g:V →Bn
2 2
(vi ,vj )∈E
where n = |V | and Bn denotes {x ∈ Rn | ||x|| ≤ 1}.
Now, suppose that a graph G has an SDP value Sdp(G) ≥ c. The SDP gap curve GapSDP (c) is a
function that measures the smallest optimal max-cut value among all graphs such that Sdp(G) ≥ c.
In other words, given that Sdp(G) ≥ c, we are guaranteed that the optimal max-cut value of G is
at least GapSDP (c). Formally,
Definition 10. For 12 ≤ s ≤ c ≤ 1, we call the pair (c, s) an SDP gap if there exists a graph G
with Sdp(G) ≥ c and Opt(G) ≤ s. We define the SDP gap curve by
GapSDP (c) = inf{s | (c, s) is an SDP gap}.
O’Donnell and Wu prove that if G is a graph such that Sdp(G) ≥ c, if one runs RPR2 iteratively
with all ˜-discretized rounding functions, then with high probability, at least one will result in a
cut with value GapSDP (c) − ˜. We now formally state the definition of an ˜-discretized rounding
function as well as O’Donnell and Wu’s algorithm guarantees.
Definition 11. Given ˜ > 0, let I˜ denote the partition of R \ {0} into intervals,
I˜ = ±(−∞, −B], ±(−B, −B + ˜2 ], ±(−B + ˜2 , −B + 2˜
2 ], . . . , ±(−2˜
2 , −˜
2 ], ±(−˜
2 , ˜2 ) ,
p
where B = B(˜
) is the smallest integer multiple of ˜2 exceeding 2 ln(1/˜
). We say that a function
r : R → [−1, 1] is ˜-discretized if the following hold:
1. r is identically −1 on (−∞, −B], 0 at 0, and identically 1 on [B, ∞).
2. r’s values on the finite intervals in I˜ are from the set ˜Z ∩ (−1, 1).
2
Note that there are 2O(1/˜ ) ˜-discretized functions.
Theorem 12 (Corollary 5.4 in [30]). There is an algorithm which, given any graph G with Sdp(G) ≥
2
c and any ˜ > 0, runs in time poly(|V |)2O(1/˜ ) and with high probability outputs a proper cut in G
with value at least GapSDP (c) − ˜.
29
Namely, the algorithm alluded to in Theorem 12 takes as input a graph, runs RPR2 using all
˜-discretized rounding functions, and returns the cut with the maximum value. We define cut˜(G)
to be the value of the resulting cut.
It is well-known that the pseudo-dimension of a finite function class F has pseudo-dimension
log |F|. This immediately implies the following theorem.
Theorem 13. Given an input sample of size m = O 12 ˜12 + log 1δ there exists an algorithm that
(, δ)-learns the class of ˜-discretized rounding functions with respect to the cost function −cut˜.
B.2
Outward rotations
Next we study a class of “outward rotation” based algorithms proposed by Zwick [45]. For the maxcut problem, outward rotations are proven to work better than the random hyperplane technique
of Goemans and Williamson [21] on graphs with “light” max-cuts where the max-cut does not
constitute a large proportion of the edges. As stated earlier, though Feige and Langberg later
showed that there exists a class of rounding functions for which RPR2 becomes equivalent to
outward rotations [18], we will analyze this class as it was originally presented by Zwick [45].
The class of outward rotation algorithms is characterized by an angle γ ∈ [0, π/2] varying
which results in a range of algorithms between the random hyperplane technique of Goemans and
Williamson and the naive approach of outputting a random binary assignment [21]. Unlike RPR2 ,
the output here is a binary assignment. An outward rotation algorithm in essence extends the
optimal SDP embedding in Rn to R2n . The way this is done can be understood as follows. The
original embedding is first carried over to the first n co-ordinates of a 2n-dimensional space while the
remaining co-oordinates are set to zero. Suppose un+1 , un+2 , . . . , u2n are the orthonormal vectors
along each of the last n co-ordinates. Each embedding ui is rotated “out” of the original space,
towards un+i by an angle of γ. After performing these outward rotations, the new embedding is
projected onto a random hyperplane in R2n . The binary assignment is then defined deterministically
based on the sign of the projections like in the GW algorithm [21]. Intuitively, the parameter γ
determines how far the SDP embedding is used to determine the final projection of vi as against
an arbitrary value drawn from the normal distribution, which is contributed by ui . We formally
define the class below.
Algorithm 5 SDP rounding algorithm using γ-outward rotation
Input: Matrix A ∈ Rn×n
1: Solve the SDP (1) for the optimal embedding U = (u1 , . . . , un ) of A.
2: Define a new embedding u0i in R2n such that the first n co-ordinates correspond to ui cos γ and
the following n co-ordinates are set to 0 except the (n + i)th co-ordinate which is set to sin γ.
3: Choose a random vector Z ∈ R2n according to the 2n-dimensional Gaussian distribution.
4: For each decision variable xi , assign xi = sgn (hu0i , Zi) .
Output: x1 , . . . , xn .
We will now set up notations similar to Section 2. Let owrγ (A, Z) be the value of the binary
assignment produced by projecting the SDP embedding of A onto Z after rotating it outwardly by
γ. That is,
owrγ (A, Z) =
X
i,j
aij sgn hu0i , Zi sgn hu0j , Zi .
We will use owrγ (A) to denote the expected value of owrγ (A, Z) when Z is sampled from Z,
the 2n-dimensional normal distribution. It can be easily seen that a fact very similar to Lemma
30
12 in Appendix A will apply here. Therefore, we will again use samples of the form A(i) , Z(i) for
the pseudo-dimension
analysis.
Let Howr = owrγ : A × R2n → [0, 1] | γ ∈ [0, π/2] . We first prove in Section B.2.1 that the
pseudo-dimension of Howr is O(log n). Next, in Section B.2.2 we present an efficient learning
algorithm.
B.2.1
The pseudo-dimension of the class of outward rotation based algorithms
We show an upper bound on the pseudo-dimension of the class of outward rotation based algorithms.
In the following discussion, we will use the notation owrA,Z (γ) in order to examine how the value
changes as a function of γ for a fixed (A, Z).
Theorem 14. Pdim(Howr ) = O(log n).
Proof. Suppose S = A(1) , Z(1) , . . . , A(m) , Z(m) is shatterable. This means that there exist
m thresholds {r1 , . . . , rm } ⊂ R such that for each T ⊆ [m], there exists a parameter γT such that
owrA(i) ,Z(i) (γT ) > ri if and only if i ∈ T .
We claim that for each sample A(i) , Z(i) , owrA(i) ,Z(i) (γ) is a piecewise constant function in
γ with at most n values of γ at which it is discontinuous. If this were to be true, it means that
there exists n values a1 , a2 , . . . an such that for γ within a given interval in [0, a1 ], (a1 , a2 ], . . . [an , 1]
owrA(i) ,Z(i) (γ) is identical and so is the label given by the witness ri . Therefore, there are at most
mn values of γ which define mn + 1 intervals such that the labels given by the witnesses for the set
of m samples is identical within each interval i.e., only at most mn + 1 distinct labelings of S are
achievable for any choice of the witnesses. However, since S is shatterable, we need 2m < mn + 1.
Thus, s = O(log n)
Now, we only need prove our claim about owrA,Z (γ) given A and γ = Z. Observe that as
γ increases, owrA,Z (γ) will change only when sgn (hu0i , Zi) changes for some vi . Now, note that
hu0i , Zi = hui , Z[1,...,n] i cos γ + zn+i sin γ where Z[1,...,n] is the projection of Z over the first n coordindates. Clearly, hu0i , Zi is a monotone function in γ ∈ [0, π/2] and attains zero at
hui , Z[1,...,n] i
−1
γ = tan
−
.
zn+i
This implies that for each i ∈ [n], sgn (hu0i , Zi) changes at most once within [0, π/2]. Therefore,
owrA,Z (γ) is a piecewise constant function with at most n discontinuities.
B.2.2
A learning algorithm
We now present Algorithm 6 that efficiently learns the best value of γ for outward rotation with
respect to samples drawn from D × Z.
1 Pm
Lemma 13. Algorithm 6 produces the value γ̂ which maximizes m
i=1 owrA(i) ,Z(i) (γ) given the
(1) (1)
(m)
(m)
sample S = A , Z
,..., A ,Z
. Algorithm 6 has running time polynomial in m and
n.
Proof. Recall from the proof of Theorem 14 that T defines intervals over [0, π/2] within each of
which the behavior of any γ is constant across all samples in S. Therefore, we only need to examine
the performance of a single value of γ within each interval to exhaustively evaluate all possibilities,
and single out the best one.
Also observe that since there are only O(mn) values in T (in Step 2) and since computing the
binary assignment on a set of m instances for a particular value of γ takes polynomial time in m
and n, Step 3 should also take only polynomial time in m and n.
31
Algorithm 6 An algorithm for finding the empirical value maximizing γ
Input: Sample S = A(1) , Z(1) , . . . , A(m) , Z(m)
(1)
1: Solve for U , . . . , U (m) the optimal SDP embeddings for A(1) , . . . , A(m) , where U (i) =
(i)
(i)
u1 , . . . , un .
2: Let T = γ1 , . . . , γ|T | be the set of all values γ ∈ [0, π/2] such that there exists a pair of indices
(j)
hu(j) ,Z[1,...,n] i
−1
= γ.
i ∈ [n], j ∈ [m] with tan
−
(j)
zn+i
1 Pm
(i)
(i) .
3: Let γ̂ = argmax
i=1 owrγ A , Z
m
γ∈T ∪{π/2}
Output: γ̂
Together with Theorem 2 and Theorem 14, Lemma 13 implies the following theorem.
Theorem 15. Given an input sample of size m = O 12 log n log 1 + log 1δ drawn from (D×Z)m ,
Algorithm 6 (, δ)-learns the class of outward rotation algorithm with respect to the cost function
−owrγ .
B.3
A general analysis of RPR2 algorithms
In Sections 2 and B.2, we investigated two specific classes of RPR2 algorithms. We will now
present an analysis that can be applied to a wide range of classes of RPR2 algorithms including
that of s-linear functions and outward rotations. In particular, we show that for most classes of
“sigmoid-like” rounding functions, the pseudo-dimension of RPR2 is Θ(log n).
As a first step towards this goal, we define what it means for a class of rounding functions to be
“sigmoid-like” (Section B.3.1). Next, in order to both generalize and simplify the sample complexity
analysis, we provide an alternative randomized procedure to RPR2 , which we call Randomized
Projection Randomized Thresholding (RPRT), which produces a binary assignment rather than a
fractional assignment (Section B.3.2). We prove that RPRT is equivalent to RPR2 by showing
that in expectation, the assignment produced by RPRT has the same value as the assignment
produced by RPR2 on an arbitrary problem instance. We work in the RPRT framework to bound
the sample complexity required to (, δ)-learn the best rounding function in a fixed class for either
RPR2 or RPRT. We prove this by showing that the pseudo-dimension of RPRT with a sigmoid-like
rounding function is Θ(log n) (Section B.3.3). Finally, in Section B.3.4, we present an algorithm
which (, δ)-learns the best rounding function in a fixed class.
B.3.1
A generic class of sigmoid-like rounding functions
We say that a class of functions F is “sigmoid-like” if the functions in the class are parameterized
by a single constant s ∈ R+ and there exists a baseline function f1 ∈ F such that for every function
fs ∈ F, fs (x) := f1 (sx). Clearly, such a representation is rich enough to encompass classes of most
common sigmoid-like functions including s-linear functions or any piece-wise linear function. In
order for a class of sigmoid-like functions to qualify as SDP rounding functions, we additionally
require that each function is non-decreasing and has a limit of −1 as x approaches −∞, and has a
limit of 1 as x approaches ∞. In particular,
Definition 16. A class of real-valued functions F = {fs | s ∈ R+ } consists of sigmoid-like rounding
functions if there exists a baseline function f1 (x) : R → [−1, 1] such that
1. f1 (x) is non-decreasing,
32
2. lim f1 (x) = 1 and lim f1 (x) = −1,
x→∞
x→−∞
and for any function fs ∈ F, we can write fs (x) = f1 (sx).
Before we define RPRT, we make the following observation about sigmoid-like functions, as
it will be useful for our algorithm design and sample complexity analysis. In particular, observe
that for each s, Fs (x) := fs (x)+1
is a cumulative density function associated with some probability
2
distribution. This is because Fs is non-decreasing and its limits are 0 and 1 as x tends to −∞ and
+∞, respectively. We denote the probability density function associated with Fs by ps .
B.3.2
Randomized Projection Randomized Thresholding
RPRT differs from RPR2 primarily in that it produces a binary assignment rather than a fractional
assignment. In essence, RPRT simply samples a binary assignment from the distribution over
binary assignments that RPR2 effectively outputs. Like RPR2 , RPRT first projects the embedding
of an IQP instance A onto a random vector Z. However, RPRT then draws a random threshold for
each variable and assigns the variable to either 1 or −1 depending on whether the directed distance
of the projection (multiplied by ||Z||) is less than or greater than its threshold. The distribution
from which these thresholds are picked are designed to mimic RPR2 in expectation.
Algorithm 7 RPRT-based SDP rounding algorithm with rounding function fs ∈ F
Input: Matrix A ∈ Rn×n .
1: Solve the SDP (1) for the optimal embedding U = (u1 , . . . , un ) of A.
2: Choose a random vector Z ∈ Rn according to the n-dimensional Gaussian distribution.
n
3: Draw Q = (q1 , q2 , . . . , qn ) ∼ (p1 ) where p1 is the probability density function corresponding
to f1 (x).
4: For each decision variable xi , output the assignment xi = sgn (qi − shui , Zi).
Output: x1 , . . . , xn .
Now we show that the expected value of the binary assignment produced by RPRT is equal to
that of RPR2 for a given rounding function fs . We define rprts (A, Z, Q) to be the deterministic
value of the binary assignment produced by RPRT using the rounding function fs , given the values
of both Z and Q. Similarly, we define rprts (A, Z) to be the expected value of the binary assignment
produced by RPRT given the value of Z. In other words, rprts (A, Z) = EQ∼(p1 )n [rprts (A, Z, Q)] .
Finally, we define rprts (A) to be the expected value of the binary assignment produced by RPRT
and by rpr2s (A) the value of the binary assignment produced by RPR2 . In other words,
rprts (A) = EQ∼(p1 )n ,Z∼Z [rprts (A, Z, Q)] and rpr2s (A) = EZ∼Z [rpr2s (A, Z)]
Theorem 17. Given a MaxQP problem with input matrix A and a rounding function fs , the
expected value of the binary assignment produced by RPRT equals the value of the fractional assignment produced by RPR2 , i.e. rprts (A) = rpr2s (A) .
Proof. Let xi and x0i denote the binary and fractional assignments produced by RPR2 and RPRT
for a given Z and Q respectively. We claim that for all i, EQ∼(p1 )n [x0i ] = xi . If we can prove
this, then it implies that rprts (G, Z) = rpr2s (G, Z). To prove this, we will make use of the fact
that given Z, and hence given the projection hui , Zi, x0i and x0j are independent random variables.
Therefore, the expectation of x0i x0j over all random draws of Q can be written as the product of
33
their individual expected values.
rprts (A, Z) =
=
E
Q∼(p1 )n
X
X
aij x0i x0j
i,j
aij EQ∼(p1 )n [x0i ]EQ∼(p1 )n [x0j ]
i,j
=
X
aij xi xj = rpr2s (A, Z).
i,j
Once we have rprts (A, Z) = rpr2s (A, Z), it is clear that rprts (A) = rpr2s (A) as Z is sampled
from the same distribution in both the algorithms.
Now we only need to show that for a given Z, EQ∼(p1 )n [x0i ] = xi i.e., for a given Z, for any
variable, the expected binary assignment of the randomized thresholding step is equal to the fractional assignment of RPR2 for that variable. Let us examine the expected binary assignment of
RPRT for a particular variable xi . The probability that xi is assigned +1 is P rqi ∼p1 [qi < shui , Zi].
Since f1 is the cumulative density function of p1 , this is equal to f1 (shui , Zi). However, due to the
way we defined fs , this value is in fact equal to fs (hui , Zi). We complete the proof by noting that
fs (hui , Zi) is precisely the fractional assignment of RPR2 for the choice of the rounding function
fs .
Now, taking advantage of this equivalence between RPR2 and RPRT for a given class of sigmoidlike functions, we will show that any learning algorithm that (, δ)-learns the best rounding functions
with respect to RPRT also (, δ)-learns the best rounding functions with respect to RPR2 . Recall
from Section 2 that in the analysis for pseudo-dimension, we incorporated Z in the sample set
S. In the same vein, we will now also incorporate Q
sample set. In other words, we will
in(1)our (1)
provide to our learning algorithm the set of samples, A , Z , Q(1) , . . . , A(m) , Z(m) , Q(m) ∼
(D × Z × (p1 )n )m . In order to prove that this indeed works, we now state Theorem 18 and Lemma
14 which parallel Theorem 2 and Lemma 12 respectively.
Theorem 18. Suppose F = {fs |s > 0} is a class of sigmoid-like rounding functions. Let Hrprt =
{rprts : A × Rn × Rn → [0, 1]|s > 0} and let dHrprt be the pseudo-dimension of Hrprt . Suppose
that Lrprt is an algorithm that takes as input m samples A(i) , Z(i) , Q(i) ∼ D × Z × (p1 )n , where
m = O 12 (dHrprt log 1 + log 1δ ) , and returns the parameter ŝ that maximizes
1 X
rprts A(i) , Z(i) , Q(i) .
m
m
i=1
Then Lrprt (, δ)-learns the class of rounding functions F with respect to the cost function −rpr2s
and is computationally efficient.
The proof for Theorem 18 follows from the following Lemma that is similar to Lemma 12.
Lemma 14. Suppose that m is sufficiently
large to ensure that with probability at least 1 − δ over
(i)
(i)
(i)
a draw of m samples A , Z , Q
∼ D × Z × (p1 )n , for all functions fs ∈ F,
1 X
rprts A(i) , Z(i) , Q(i) −
E
[rprts (A, Z, Q)] < .
m
2
(A,Z,Q(i) )∼D×Z×(p1 )n
i=1
1 Pm
(i)
(i)
(i) , and s∗ maxThen with
least 1 − δ, if ŝ maximizes m
i=1 rprts A , Z , Q
probability at
imizes
E [rpr2s (A)] , then E [rpr2s∗ (A)] − E [rpr2ŝ (A)] < .
m
A∼D
A∼D
A∼D
34
Proof of Lemma 14. Since D × Z × (p1 )n is a product distribution, we have that
E
(A,Z,Q)∼D×Z×(p1 )n
[rprts (A, Z, Q)] = E [rprts (A)] .
A∼D
But from Theorem 17 we know that rprts (A) = rpr2s (A), which implies that
E [rprts (A)] = E [rpr2s (A)] .
A∼D
A∼D
Hence, we can restate the assumption in the theorem statement as follows. With probability at
least 1 − δ, for all functions fs ,
1 X
rprts A(i) , Z(i) , Q(i) − E [rpr2s (A)] < .
A∼D
m
2
m
i=1
Since this true for ŝ and s∗ , we can say that with probability 1 − δ,
1 X
rprtŝ A(i) , Z(i) , Q(i) − E [rpr2ŝ (A)] <
A∼D
m
2
m
i=1
and
1 X
rprts∗ A(i) , Z(i) , Q(i) − E [rpr2s∗ (A)] < .
A∼D
m
2
m
i=1
Now, by definition of ŝ we have that,
1 X
1 X
rprtŝ A(i) , Z(i) , Q(i) ≤
rprts∗ A(i) , Z(i) , Q(i) .
m
m
m
m
i=1
i=1
Combining the previous three inequalities, we get that E [rpr2s∗ (A)] − E [rpr2ŝ (A)] < .
A∼D
B.3.3
A∼D
The pseudo-dimension of a general class of RPRT algorithms
The main result of this section is a tight bound on the pseudo-dimension of any general class of
RPRT algorithms.
Theorem 19. Let Hrprt = {rprts : A × Rn × Rn → [0, 1] | s > 0}. Pdim(Hrprt ) = Θ(log n).
This follows from Lemma 15 and Lemma 16 where we provide matching upper and lower bounds
on Pdim(Hrprt ).
Lemma 15. Pdim(Hrprt ) = O(log n).
Proof. Suppose S = A(1) , Z(1) , Q(1) , . . . , A(m) , Z(m) , Q(m) is shatterable. Then there exist
m thresholds {r1 , . . . , rm } ⊂ R such that for all T ⊆ [m], there exists a parameters sT such that
vals0 T A(i) , Z(i) > ri if and only if i ∈ T .
We first claim that for a given sample A(i) , Z(i) , Q(i) the domain of s i.e., (0, ∞) can be parti(i)
(i)
(i)
(i)
tioned into at most n + 1 intervals (0, a1 ], (a1 , a2 ], . . . (an , ∞) such that the binary assignment
produced by RPRT on the sample is identical across all parameter settings of s within a given
interval. Once this is proved, we can then consider a partition of the domain of s into at most
m
S
(i)
(i)
mn + 1 intervals based on the points
{a1 , . . . , an }. Now, we know that for any value of s within
i=1
35
a given interval of these mn + 1 intervals, the labeling induced by the witnesses on S is the same.
Therefore, there are at most mn + 1 possible labelings that can be produced by any choice of the
witnesses over any set of m samples. However, since we picked a shatterable instance, it must be
that 2m ≤ mn + 1 i.e., m = O(log n).
Now to complete
the proof, we need to examine the behavior of RPRT on a single sample
A(i) , Z(i) , Q(i) for different configurations of s. Observe that as we increase s from 0 to ∞
keeping all else constant, we expect RPRT to produce a different binary assignment only when the
assigned value changes for some vertex vi . However, this happens only when qi = shg(vi ), Zi i.e.,
when s = qi /hg(vi ), Zi. Since there are only n vertices, and at most one value of s at which the
algorithm changes behavior, we expect the algorithm to change its behavior at most n times as s
increases. This proves our claim.
We now prove a lower bound on the pseudo-dimension of Hrprt .
Lemma 16. Pdim(Hrprt ) = Ω(log n).
Proof. In order to prove that the pseudo-dimension
ofHrprt is at least c log n for some constant c,
(1) (1) (i)
we must devise a set of samples S = G , Z , Q
, . . . , G(m) , Z(m) , Q(i) of size m = c log n
that can be shattered by Hrprt . This means that we should be able to find m witnesses r1 , . . . , rm
and 2m = nc s values H = {s1 , . . ., snc } such that for all T ⊆ [m], there exists sT ∈ H such that if
i ∈ T , then rprtsT G(i) , Z(i) , Q(i) > ri and if i 6∈ T , then rprtsT G(i) , Z(i) , Q(i) ≤ ri .
In our solution, we will use the same S as designed in the proof for Lemma 4. That is, all G(i) are
identical and each consists of n/4 disjoint K4 graphs over vertices Vk = {v4k+1 , v4k+2 , v4k+3 , v4k+4 }
for k = 0, 1, . . . bn/4c − 4. We will later pick Z(i) and Q(i) such that as the value of s increases,
we change the cuts on V1 , V2 , . . . in that order and the change in the cuts is alternatingly better
and worse. By appropriately choosing different values of Z(i) , we can carefully place the intervals
in which these oscillations occur across all samples so as to be able to shatter a sample of size
Ω(log n).
In order to define Z(i) and Q(i) , we make use of the following increasing sequence defined
i−1
recursively: c1 = 8 and ci = c2i−1 . Note that we can also write ci = 82
but we will use the
1
(i)
notation ci for the sake of convenience. Let aik = ck . We will now define Z in terms of aik as
i
follows. For even k, we define
1
3
(i)
(i)
(i)
(i)
· aik , aik , · aik , aik
Z4k+1 , Z4k+2 , Z4k+3 , Z4k+4 =
2
2
and for odd k we define,
(i)
(i)
(i)
(i)
Z4k+1 , Z4k+2 , Z4k+3 , Z4k+4 = (aik , −aik , −aik , aik )
The rationale behind choosing the above value for Z(i) will become evident as we proceed with
the proof. First, by a simple calculation we can confirm that the directed distance of the projections
of the vertices on Z(i) lie in certain intervals as stated below.
For even k,
1
j = 4k + 1,
2 aik , 2aik
1 a , 2a
j = 4k + 2,
ik
ik
hg (i) (vj ), Z(i) i ∈ 12
j = 4k + 3,
2 aik , 2aik
−2a , − 1 a j = 4k + 4.
ik
2 ik
36
For odd k,
1
aik , 2aik
2
−2a , − 1 a
ik
(i)
(i)
2 ik
hg (vj ), Z i ∈
−2aik , − 12 aik
1 a , 2a
ik
2 ik
j
j
j
j
= 4k + 1,
= 4k + 2,
= 4k + 3,
= 4k + 4.
As for the choice of Q(i) , we pick the same value for all i. In particular, we choose Q(i) as
follows. For even k,
(i)
(i)
(i)
(i)
Q4k+1 , Q4k+2 , Q4k+3 , Q4k+4 = (−1, +1, +1, −1) ,
and for odd k,
(i)
(i)
(i)
(i)
Q4k+1 , Q4k+2 , Q4k+3 , Q4k+4 = (−1, −1, −1, +1) ,
Again, our choiceof Q(i) will make sense as we proceed with the proof. We now analyze
how
(i)
(i)
(i)
(i)
(i)
(i)
rprts G , Z , Q
varies as a function of s. We will show that rprts G , Z , Q
oscillates
above and below a threshold ri , bn/4c/2 times and these oscillations are spaced in such a manner
that we can pick 2m values of s that can shatter S.
First, let us examine the values of s at which we expect the behavior of rprts G(i) , Z(i) , Q(i)
to change. We know that this can only consist of values of s equal to 1/hg (i) (vj ), Z(i) i for some j.
Based on this fact,
we observe that for the vertices in Vk the values of s at which the cut changes
lie in 12 cki , 2 · cki . Thus, note that the intervals of s in which the cut changes for V1 , V2 , . . . are
spaced far apart in increasing order because 2cki < 12 ck+1
given ci > 4.
i
Our analysis will henceforth focus on the values of s outside these intervals. We now claim
that there exists values bmax and bmin such that the value of the cut is only one of these values
in
k , 2 · ck and
between
the
above
intervals.
In
particular,
the
value
is
equal
to
b
between
0.5c
max
i
i
h
i
0.5ck+1
, 2 · ck+1
for odd k and is equal to bmin for even k. By simple calculation, it can be verified
i
i
that the cut (S, S c ) assigned to Vk varies as follows. For even k,
(
{v4k+1 , v4k+4 }
s < 0.5cki ,
S=
{v4k+1 , v4k+2 , v4k+3 } s > 2 · cki ,
and for odd k,
(
{v4k+1 , v4k+2 , v4k+3 }
S=
{v4k+1 , v4k+4 }
s < 0.5cki ,
s > 2 · cki .
Now, when the cut is defined by S = {v4k+1 , v4k+4 } is 2/3. On the other hand when S =
{v4k+1 , v4k+2 , v4k+3 },
value of the cut is 1/2. Thus we can make a crucial observation here: as s
1the
k
crosses the interval 2 ci , 2 · cki for even k, the net value of the cut over the whole graph decreases,
and for odd k it increases.
More precisely, we claim that rprts G(i) , Z(i) , Q(i) as a function of s outside these intervals
takes one of two values bmin and bmax . To state these values explicitly, let us define a set of variables
bk such that for even k, bk = 2/3 and for odd k, bk = 1/2. Observe that when s < 1/2, the value
1 Pbn/4c−1
bk as every odd graph contributes a value of 1/2 and every even graph
of the cut is bn/4c
k=0
contributes a value of 2/3 to the total cut. We will denote this value by bmax . Now when s ∈ (2, 12 ci ),
37
the value then decreases to
1
bn/4c
P
bn/4c−1
bk
k=0
− 2/3 + 1/2 which we will call bmin . We can extend
this observation as follows to any interval Ii,k = (2 · cki , 21 ck+1
)
i
(
bmin s ∈ Ii,k for even k ≤ bn/4c − 1,
(i)
(i)
(i)
rprts G , Z , Q
=
bmax s ∈ Ii,k for odd k ≤ bn/4c − 1,
Now we only need to choose ri to be (bmin + bmax )/2. To complete the proof, we need to show
that the oscillations with respect to these witnesses for all the samples are spaced in a manner that
we can pick 2m different s covering all possible oscillations for m = Ω(log n).
This becomes clear from two observations. First, Ii,2k ∪ Ii,2k+1 ⊂ Ii+1,k because ci = 2c2i−1 .
Secondly, values of s injIi,2k and kIi,2k+1 induce different labelings on the ith sample.
bn/4c
samples. It turns out that c2m /2 ≤ c1
. That is, Im,0 ∪ Im,1
Now consider m = log(bn/4c)
3
contains I1,0 ∪ . . . I1,bn/4c−1 . We then claim that the intervals I1,k each induce a different labeling
on S with respect to the witnesses. To see why this is true, observe that since Ii,k ⊂ Ii+1,bk/2c , the
labeling induced by s ∈ I1,k is defined by which terms of the sequence k1 = k, ki+1 = bki /2c are
odd. This however is in fact the binary equivalent of k. Since the binary equivalent for a given
k is unique, the labeling induced by I1,k is unique. Thus, we have Ω(log n) samples that can be
shattered.
B.3.4
A learning algorithm
We now present a learning algorithm(Algorithm 8) which (, δ)-learns the best rounding function
with respect to D from a class of sigmoid-like rounding functions.
Algorithm 8 An algorithm for finding an empirical value maximizing rounding function fs
n
o
(m)
Input: Sample S = A(1) , Z(1) , Q(1) , . . . , A(m) , Z(m),Q
1: Solve for X (1) , . . . , X (m) the optimal SDP embeddings of A(1) , . . . , A(m) , where X (i) =
(i)
(i)
x1 , . . . , xn .
2: Let T = s1 , . . . , s|T | be the set of all values s > 0 such that there exists a pair of indices
(j)
(j)
i ∈ [n], j ∈ [m] with hZ(j) , xi i = sqi .
1 Pm
(i)
(i)
(i) .
3: Let ŝ = argmax
i=1 rprts A , Z , Q
m
Output: ŝ
s∈T ∪{s|T | +1}
In particular, we prove the following guarantee regarding Algorithm 8’s performance.
1 Pm
Lemma 17. Algorithm 2 produces the value ŝ which maximizes m
rpr2s A(i) , Z(i) , Q(i)
i=1
given the sample S = A(1) , Z(1) , Q(1) , . . . , A(m) , Z(m) , Q(m) . Algorithm 8 has running time
polynomial in m and n.
Proof. The correctness of the algorithm is evident from the proof for Lemma 15. The algorithm
identifies the n values of s at which the behavior of RPRT changes and proceeds to exhaustively
evaluate all the possible binary assignments which are only polynomially many. Therefore the
algorithm successfully finds the best value of s in polynomial time.
38
Corollary 1. Given an input sample of size m = O 12 log n log 1 + log 1δ drawn from (D × Z ×
(p1 )n )m , Algorithm 8 (, δ)-learns the class of rounding functions F with respect to the cost function
−rpr2s and is computationally efficient.
C
Proofs from Section 3
Theorem 4. For i ∈ {1, 2, 3} and a permissible
value iof α for Ahi , there exists
h
i a distribution D
(p)
(p)
over clustering instances V such that EV∼D ΦAi (α) (V) < EV∼D ΦAi (α0 ) (V) for all permissible
values of α0 6= α for Ai .
Proof. We give a general proof for all three classes A1 , A2 , and A3 . We will point out a few places
in the proof where the details for b = 1, 2, 3 are different, but the general structure of the argument
is the same. For each value of b, we construct a single clustering instance V = (V, d) that has the
desired property; the distribution D is merely the single clustering instance with probability 1.
Consider some permissible value of α, denoted α∗ . Set k = 4 and n = 210. The clustering
instance consists of two well-separated ‘gadgets’ of two clusters each. The class Ab results in
different 2-clusterings of the first gadget depending on whether α ≤ α∗ or not. Similarly, Ab results
in different 2-clusterings of the second gadget depending on whether α ≥ α∗ or not. By ensuring
that for the first gadget α ≤ α∗ results in the lowest cost 2-clustering, and for the second gadget
α ≥ α∗ results in the lowest cost 2-clustering, we ensure that α = α∗ is the optimal parameter
overall.
The first gadget is as follows. We define five points a1 , b1 , c1 , x1 and y1 . For the sake of
convenience, we will group the remaining points into four sets A1 , B1 , X1 , and Y1 each containing 25
points. We set the distances as follows: d(a1 , b1 ) = d(x1 , y1 ) = 1, d(a1 , c1 ) = 1.1, and d(b1 , c1 ) = 1.2.
For a ∈ A1 ∪ B1 , d(c1 , a) = 1.51 and d(a1 , a) = d(b1 , a) = 1.6. For x ∈ X1 ∪ Y1 , d(x1 , x) = d(y1 , x) =
1.6. For a ∈ A1 , b ∈ B1 , x ∈ X1 , and y ∈ Y1 , d(a, b) = d(x, y) = 1.6. We also define special points
x∗1 ∈ X1 and y1∗ ∈ Y1 , which have the same distances as the rest of the points in X1 and Y1
respectively, except that d(x1 , x∗1 ) = 1.51 and d(y1 , y1∗ ) = 1.51. If two points p and q belong to the
same set (A1 , B1 , X1 , or Y1 ), then d(p, q) = 1.5.
The distances d(x1 , c1 ) and d(y1 , c1 ) are defined in terms of b and α∗ , but they will always be
between 1.1 and 1.2. For b = 1, we set d(x1 , c1 ) = d(y1 , c1 ) = 1.2 − .1 · α∗ . For b = 2 and b = 3,
1
∗
∗
d(x1 , c1 ) = d(y1 , c1 ) = ((1.1α + 1.2α )/2) α∗ .
So far, all of the distances we have defined are in [1, 2], therefore they trivially satisfy the triangle
inequality. We set all of the rest of the distances to be the maximum distances allowed under the
triangle inequality. Therefore, the triangle inequality holds over the entire metric.
Now, let us analyze the merges caused by Ab (α) for various values of α. Regardless of the values
of α and b, since the distances between the first five points are the smallest, merges will occur over
these initially. In particular, regardless of α and b, a1 is merged with b1 , and x1 with y1 . Next, by
a simple calculation, if α ≤ α∗ , then c1 merges with a1 ∪ b1 . If α > α∗ , then c1 merges with x1 ∪ y1 .
Denote the set containing a1 and b1 by A01 , and denote the set containing x1 and y1 by X10 (one of
these sets will also contain c1 ). Between A01 and X10 , the minimum distance is ≥ 1.1 + 1.1 ≥ 2.2.
All other subsequent merges (except for the very last merge) will involve all distances smaller than
2.2, so we never need to consider A01 merging to X10 .
The next smallest distances are all 1.5, so all points in A1 will merge together, and similarly
for B1 , X1 , and Y1 . At this point, the algorithm has created six sets: A01 , X10 , A1 , B1 , X1 , and Y1 .
We claim that if α ≤ α∗ , A01 will merge to A1 and B1 , and X10 will merge to X1 and Y1 . This is
because the maximum distance between sets in each of these merges is 1.6, whereas the minimum
39
distance between {A01 , A1 , B1 } and {X10 , X1 , Y1 } is ≥ 2.2. Therefore, for all three values of b, the
claim holds true.
Next we claim that the 2-clustering cost of gadget 1 will be lowest for clusters A01 ∪ A1 ∪ B1 }
and X10 ∪ X1 ∪ Y1 and when c1 ∈ A01 , i.e., when α ≤ α∗ . Clearly, since the distances within
A01 ∪ A1 ∪ B1 and X10 ∪ X1 ∪ Y1 are much less than the distances across these sets, the best 2clustering is A01 ∪ A1 ∪ B1 and X10 ∪ X1 ∪ Y1 (with all points at distance ≤ 1.6 to their center).
We proved this will be a pruning of the tree when α ≤ α∗ . Therefore, we must argue the cost
of this 2-clustering is lowest when c1 ∈ A01 . The idea is that c1 can act as a very good center for
A01 ∪ A1 ∪ B1 . But if c1 ∈ X10 , then the best center for A01 ∪ A1 ∪ B1 will be an arbitrary point
in A1 ∪ B1 . The cost in the first case is 1.51p · 50 + 1.1p + 1.2p . The cost in the second case is
1.5p · 24 + 1.6p · 27.
For X10 ∪ X1 ∪ Y1 , the center does not change depending on α (x∗1 and y1∗ tie for the best center),
so the only difference in the cost is whether or not to include c1 . If α ≤ α∗ , then the cost is
1.5p · 24 + 1.51p + 1.6p · 26, otherwise the cost is 1.5p · 24 + 1.51p + 1.6p · 26 + (1.6 + 1.2 − 0.1α∗ )p .
Putting it all together, if α ≤ α∗ , the cost is 1.51p · 50 + 1.1p + 1.2p + 1.5p · 24 + 1.51p + 1.6p · 26.
Otherwise the cost is 1.5p · 48 + 1.51p + 1.6p · 53 + (1.6 + 1.2 − 0.1α∗ )p . Subtracting off like terms,
we conclude that the first case is always smaller because 1.51p · 49 + 1.1p + 1.2p < 1.5p · 24 + 1.6p ·
26 + (1.6 + 1.2 − 0.1α∗ )p for all p ≥ 1.
Next, we will construct the second gadget arbitrarily far away from the first gadget. The second
gadget is very similar to the first. There are points a2 , b2 , c2 , x2 , y2 , x∗2 , y2∗ and sets to A2 , B2 , X2 ,
Y2 . d(a2 , b2 ) = d(x2 , y2 ) = 1, d(x2 , c2 ) = 1.1, d(y2 , c2 ) = 1.2, and for b = 1, d(a2 , c2 ) = d(b2 , c2 ) =
1
∗
∗
1.2 − .1 · α∗ . For b = 2 or b = 3, d(a2 , c2 ) = d(b2 , c2 ) = ((1.1α + 1.2α )/2) α∗ . The rest of the
distances are the same as in gadget 1. Then c2 joins {a2 , b2 } if α ≥ α∗ , not α ≤ α∗ . The rest of the
argument is identical. So the conclusion we reach, is that the cost for the second gadget is much
lower if α ≥ α∗ .
Therefore, the final cost of the 4-clustering is minimized when α = α∗ , and the proof is complete.
Now we will show a structural lemma for A3 , which is similar to Lemma 6. Then we will provide
the full details for the proof of Lemma 6.
Lemma 18. ΦA3 ,V : [0, 1] → R>0 is made up of O(n8 ) piecewise constant components.
Proof. First note that for α 6= α0 , the clustering returned by A1 (α) and the associated cost are both
identical to that of A1 (α0 ) if both the algorithms construct the same merge tree. Now, as we increase
α from 0 to 1 and observe the run of the algorithm for each α, at what values of α do we expect
A1 (α) to produce different merge trees? To answer this, suppose that at some point in the run
of algorithm A1 (α), there are two pairs of subsets of V , (A, B) and (X, Y ), that could potentially
merge. There exist eight points p, p0 ∈ A, q, q 0 ∈ B, x, x0 ∈ X, and y, y 0 ∈ Y such that the decision
of which pair to merge depends on whether αd(p, q) + (1 − α)d(p0 , q 0 ) or αd(x, y) + (1 − α)d(x0 , y 0 )
is larger. Clearly, there is at most one value of α for which these expressions are equal, unless
the difference of the expressions is zero for all α. Assuming that ties are broken arbitrarily but
consistently, this implies that there is at most one α ∈ [0, 1] such that the choice of whether to
merge (A, B) before (X, Y ) is identical for all α < α0 , and similarly identical for α ≥ α0 . Since each
merge decision is defined by eight points, iterating over all pairs (A, B) and (X, Y ) it follows that
we can identify all O(n8 ) unique 8-tuples of points which correspond to a value of α at which some
decision flips. This means we can divide [0, 1] into O(n8 ) intervals over each of which the merge
tree, and therefore the output of ΦA1 ,V (α), is fixed.
40
Now we will provide the details of Lemma 6. In the argument for the structure of HA3 ,Φ ,
we relied on the linearity of A3 ’s merge equation to prove that for any eight points, there is
exactly one value of α such that αd(p, q) + (1 − α)d(p0 , q 0 ) = αd(x, y) + (1 − α)d(x0 , y 0 ). Now
we will use a consequence of Rolle’s Theorem (Theorem 20) to bound the values of α such that
((d(p, q))α + d(p0 , q 0 )α )1/α = ((d(x, y))α + d(x0 , y 0 )α )1/α .
P
x
Theorem 20 (ex. [37]). Let f be a polynomial-exponential sum of the form f (x) = N
i=1 ai bi ,
where bi > 0, ai ∈ R, and at least one ai is non-zero. The number of roots of f is upper bounded
by N .
Lemma 6. ΦA1 ,V : R ∪ {−∞, ∞} → R>0 is made up of O(n8 ) piecewise constant components.
Proof. As was the case for HA3 , the clustering returned by A1 (α) and the associated cost are
identical to that of A1 (α0 ) as long as both algorithms construct the same merge trees. Our objective
is to understand the behavior of A1 (α) over m instances. In particular, as α varies over R we
want to count the number of times the algorithm outputs a different merge tree on one of these
instances. For some instance V we will consider two pairs of sets A, B and X, Y that can be
potentially merged. The decision to merge one pair before the other is determined by the sign of
dα (p, q) + dα (p0 , q 0 ) − dα (x, y) + dα (x0 , y 0 ). This expression, as before, is determined by a set of 8
points p, p0 ∈ A, q, q 0 ∈ B, x, x0 ∈ X and y, y 0 ∈ Y chosen independent of α.
Now, from Theorem 20, we have that the sign of the above expression as a function of α flips 4
times across R. Since the expression is defined by exactly 8 points, iterating over all pairs (A, B)
and (X, Y ) we can list only O(n8 ) such unique expressions, each of which correspond to O(1) values
of α at which the corresponding decision flips. Thus, we can divide R into O(n8 ) intervals over
each of which the output of ΦA1 ,V (α) is fixed.
Now we give the full details for Lemmas 7 and 8.
Lemma 7. For any Φ, Pdim(HA1 , Φ) = O(log n) and Pdim(HA3 , Φ) = O(log n).
Proof. Suppose S = V (1) , . . . , V (m) is a set of clustering instances that can be shattered by
HA1 using the witnesses r1 , . . . , rm . We must show that m = O(log n). For each value of α ∈
R ∪ {−∞, ∞},
V (i) , based on whether or not
the algorithm A1 (α) induces a binary labeling on each
(i)
(i)
ΦA1 (α) V
≤ ri . From Lemma 6, we know that every sample V partitions R ∪ {∞, −∞} into
8
O(n ) intervals in this way. Merging all m partitions, we can divide R ∪ {∞, −∞} into O(mn8 )
intervals over each of which ΦA3 ,V (i) (α), and therefore the labeling induced by the witnesses, is
fixed for all i ∈ [m] (similar to Figure 7). This means that HA1 can achieve only O(mn8 ) binary
labelings, which is at least 2m since S is shatterable, so m = O(log n).
The details for HA3 , Φ are identical, by using Lemma 18.
Lemma 8. For any objective function Φ(p) , Pdim HA1 ,Φ(p) = Ω(log n) and Pdim HA3 ,Φ(p) =
Ω(log n).
We first prove this lemma for the center-based objective cost denoted by Φ(p) for p ∈ [1, ∞) ∪
{∞}. We later note how this can be extended cluster purity based cost. We first prove the following
useful statement which helps us construct general examples with desirable properties. In particular,
the following lemma guarantees that given a sequence of values of α of size O(n), it is possible to
construct an instance V such that the cost of the output of A1 (α) on V as a function of α, that
(p)
is ΦA1 ,V (α), oscillates above and below some threshold as α moves along the sequence of intervals
(αi , αi+1 ). Given this powerful guarantee, we can then pick appropriate sequences of α and generate
41
y1 0
1
x1 0
1.41
1
xi 0
1.41
z1
yi 0
1.4 − 0.1αi
2
z1 0
zi 0
zi
1.4
1.3
yi
y1
1
1
xi
x1
Figure 8: The clustering instance used in Lemma 19
a sample set of Ω(log n) instances that correspond to cost functions that oscillate in a manner that
helps us pick Ω(n) values of s that shatters the samples.
Lemma 19. Given n ∈ N, b ∈ {1, 3}, and given a sequence of n0 ≤ bn/7c α’s such that 0 = α0 <
α1 < · · · < αn0 < αn0 +1 = .7, there exists a real valued witness r > 0 and a clustering instance
(p)
V = (V, d), |V | = n, such that for 0 ≤ i ≤ n0 /2 − 1, ΦAb (α) (V) < r for α ∈ (α2i , α2i+1 ), and
(p)
ΦAb (α) (V) > r for α ∈ (α2i+1 , α2i+2 ), for k = 2.
Proof. First, we focus on A3 , and we discuss A1 at the end of the proof. First of all, in order for
d to be a metric, we set all distances in [1, 2] so that the triangle inequality is trivially satisfied.
In particular, the following are the distances of the pairs of points within each group for 1 ≤ i ≤
(n − 2)/6.
d(xi , yi ) = d(x0i , yi0 ) = 1,
d(xi , zi ) = 1.3, d(yi , zi ) = 1.4,
d(x0i , zi ) = d(yi0 , zi ) = 1.4 − .1 · αi ,
d(xi , x0i ) = d(yi , yi0 ) = 2.
We set the distances to zi0 as follows (see Figure 8).
d(xi , zi0 ) = d(yi , zi0 ) = d(x0i , zi0 ) = d(yi0 , zi0 ) = 1.41,
d(zi , zi0 ) = 2.
Then the first merges will be xi to yi and x0i to yi0 , no matter what α is set to be (when each
point is a singleton set, each pair of points with the minimum distance in the metric will merge).
Next, zi will either merge to Ai or A0i based on the following equation:
α · 1.3 + (1 − α) · 1.4 ≶ α · (1.4 − .1 · αi ) + (1 − α)(1.4 − .1 · αi )
42
1.4 − .1 · α ≶ 1.4 − .1 · αi
=⇒
=⇒
αi ≶ α
If α < αi , then zi merges to A0i , otherwise it will merge to Ai . Notice that the merge expression
for Ai to A0i could be as small as α · 1.3 + (1 − α) · 2 = 2 − .7 · α, but we do not want this
merge to occur. If we ensure all subsequent merges have maximum distance less than 1.5, then
Ai will not merge to A0i (until A and A0 merge in the very final step) as long as α < .7, because
α · 1.5 + (1 − α) · 1.5 = 1.5 < 2 − .7 · .7.
These distances ensure zi0 merges after zi regardless of the value of α, since zi is closer than zi0
to xi , x0i , yi , and yi0 . Furthermore, zi0 will merge to the opposite set of zi , since we set d(zi , zi0 ) = 2.
The merge expression for zi0 to merge to the opposite set is α · 1.41 + (1 − α) · 1.41, while the merge
expression to the same set is ≥ α · 1.41 + (1 − α) · 2.
Now we set the distances to a and a0 as follows.
d(a, xi ) = d(a, yi ) = d(a0 , x0i ) = d(a0 , yi0 ) = 1.42,
d(a, x0i ) = d(a, yi0 ) = d(a0 , xi ) = d(a0 , yi0 ) = 2.
We also set all distances between Ai and A0j to be 2, for all i and j, and all distances between Ai
and Aj to be 1.5, for all i 6= j. We will set the distances from a and a0 to zi and zi0 later, but they
will all fall between 1.45 and 1.5. By construction, every set Ai will merge to the current superset
containing {a}, because the merge expression is α · 1.42 + (1 − α)1.5, and any other possible merge
will have value ≥ α · 1.3 + (1 − α) · 2, which is larger for α < .7. Similarly, all A0i sets will merge to
{a0 }.
Therefore, the final two sets in the linkage tree are A and A0 . Given 1 ≤ i ≤ (n − 2)/6, by con0 , . . . z0
0
0
struction, for α ∈ (αi , αi+1 ), {z1 , . . . , zi , zi+1
(n−2)/6 } ⊆ A and {z1 , . . . , zi , zi+1 , . . . z(n−2)/6 } ⊆
A0 .
Finally, we set the distances between a, a0 , zi , and zi0 to ensure the cost function oscillates.
∀1 ≤ j ≤ (n − 2)/12,
∀i,
d(a, zi0 ) = d(a0 , zi ) = 1.46
0
d(a, z2j−1 ) = d(a0 , z2j
) = 1.47,
0
and d(a, z2j ) = d(a0 , z2j+1
) = (2 · 1.46p − 1.47p )1/p .
Now we calculate
cost of (A, A0 ) for α’s in different ranges. Regardless of α, all
Pthe 2-clustering
partitions will pay i (d(a, xi )p + d(a, yi )p + d(a0 , x0i )p + d(a0 , yi0 )p ) = (n − 2)/6 · (4 · 1.42p ),but the
distances for zi and zi0 differ. For α ∈ (α0 , α1 ), all of the z’s pay 1.46p , so the cost is (n − 2)/6 · (4 ·
1.42p + 2 · 1.46p ). Denote this value by rlow .
When α ∈ (α1 , α2 ), the only values that change are z1 and z10 , which adds d(a, z1 ) + d(a0 , z10 ) −
d(a, z10 ) − d(a0 , z1 ) = 2 · (1.47p − 1.46p ) > 0 to the cost (the inequality is always true for p ∈ [1, ∞]).
Denote rlow + 2 · (1.47p − 1.46p ) by rhigh . When α ∈ (α2 , α3 ), the values of z2 and z20 change, and
the cost changes by d(a, z2 ) + d(a0 , z20 ) − d(a, z20 ) − d(a0 , z2 ) = 2 · ((2 · 1.46p − 1.47p ) − 1.46p ) =
−2 · (1.47p − 1.46p ), decreasing it back to rlow . P
In general, the cost for α ∈ (αi , αi+1 ) is rlow + 1≤j≤i (−1)i+1 · 2(1.47p − 1.46p ) = rlow + (1.47p −
1.46p ) + (−1)i+1 · (1.47p − 1.46p ). If α ∈ (α2j , α2j+1 ), then the cost is rlow , and if α ∈ (α2j+1 , α2j+2 ),
the cost is rhigh . We set r = (rlow + rhigh )/2, and conclude that the cost function oscillates above
and below r as specified in the lemma statement.
The pruning step will clearly pick (A, A0 ) as the optimal clustering, since the only centers with
more than 3 points at distance < 1.5 are a and a0 , and (A, A0 ) are the clusters in which the most
points can have a and a0 as centers. This argument proved the case where n0 = (n − 2)/6. If
43
n0 < (n − 2)/6, then we set d(a, zi ) = d(a0 , zi0 ) = 1.46 for all i > n0 , which ensures the cost function
oscillates exactly n0 times. This completes the proof.
It is straightforward to modify this proof to work for A1 . The only major change is to set
1
d(x0i , zi ) = d(yi0 , zi ) = ((1.3αi + 1.4αi )/2) αi .
Now we can prove Lemma 8.
Proof of Lemma 8. Given b ∈ {1, 2}, we prove the claim for HAb ,Φ(p) by constructing a set of
samples S = {V (1) , . . . , V (m) } where m = log n − 3 that can be shattered by HAb ,Φ(p) . That is,
we should be able to choose 2m = n/8 different values of α such that there exists some witnesses
(p)
r1 , . . . , rm with respect to which ΦAb (α) (·) induces all possible labelings on S.
Choose a sequence of 2m distinct α’s arbitrarily in the range (0, .7). We will index the terms of
this sequence using the notation αx for all x ∈ {0, 1}m , such that αx < αy iff x1 x2 . . . xm < y1 y2 ym .
Then the α’s satisfy
0 < α[0
... 0 0]
< α[0
... 0 1]
< α[0
... 1 0]
< · · · < α[1
... 1 1]
< .7.
Given x, denote by n(x) the vector corresponding to x1 x2 . . . xs + 1, therefore, αn(x) is the smallest
α greater than αx .
Now, the crucial step is that we will use Lemma 19 to define our examples V (1) , . . . V (m) and
witnesses r1 , . . . rm so that when α ∈ (αx , αn(x) ) the labeling induced by the witnesses on S corre(p)
sponds to the vector x. This means that for α ∈ (αx , αn(x) ) the cost function ΦAb (α) (V (i) ) must be
greater than ri if the ith term in x is 1, and less than ri otherwise. Since there are only 2m = n8
x’s, it implies that for any sample V (i) there at most n/8 values of α at which we want its cost to
flip above/below ri . We can we can accomplish this using Lemma 19 by choosing αx ’s for which
V (i) is supposed to switch labels. In this manner, we pick each V (i) and ri thus creating a sample
of size Ω(log n) that is shattered by HAb ,Φ(p) .
Note C.1. Lemma 8 assumes that the pruning step fixes a partition, and then the optimal centers
can be chosen for each cluster in the partition, but points may not switch clusters even if they are
closer to the center in another cluster. This is desirable, for instance, in applications which much
have a balanced partition.
If it is desired that the pruning step only outputs the optimal centers, and then the clusters are
determined by the Voronoi partition of the centers, we modify the proof as follows. We introduce
2n0 more points into the clustering instance: c1 , . . . , cn0 , and c01 , . . . , c0n0 . Each ci will merge to
cluster A, and each c0i will merge to cluster A0 . We set the distances so that ci and c0i will be the
best centers for A and A0 when α ∈ (αi , αi+1 ). The distances are also set up so that the cost of
the Voronoi tiling induced by c2i and c02i is rlow , and the cost for c2i+1 and c02i+1 is rhigh . This is
sufficient for the argument to go through.
Furthermore, the lower bound holds even ifSthe cost function is theSsymmetric distance to the
0
0 } be the ground
ground truth clustering. For this proof, let A ∪ i {z2i , z2i+1
} and A0 ∪ i {z2i+1 , z2i
truth clustering. Then in each interval as α increases, the cost function switches between having
(n − 2)/3 errors and having (n − 2)/3 − 2 errors.
Now we restate and prove Lemma 10.
Lemma 10. For all objective functions Φ(p) , Pdim HA2 ,Φ(p) = Ω(n).
44
To prove this, we start with a helper lemma.
Lemma 20. Given n, and setting N = b(n−8)/2c, then there exists a clustering instance V = (V, d)
of size |V | = n and a set of 2N + 2 α’s for which α-linkage creates a unique merge tree.
Proof. Here is the outline of our construction. At the start, two specific pairs of points will always
merge first: pa merges with qa , and pb merges with qb . The sets {pa , qa } and {pb , qb } will stay
separated until the last few merge operations. Throughout the analysis, at any point in the merging
procedure, we denote the current superset containing {pa , qa } by A, and we similarly denote the
superset of {pb , qb } by B. The next points to merge will come in pairs, (pi , qi ) for 1 ≤ i ≤ N . We
construct the distances so that pi and qi will always merge before pj and qj , for i < j. Furthermore,
for all i, {pi } will first merge to A or B, and then {qi } will merge to the opposite set as pi . Let
these two merges be called ‘round i’, for 1 ≤ i ≤ N . Finally, there will be a set CA of size N + 2
which merges together and then merges to A, and similarly a set CB which merges to B.
Thus, in our construction, the only freedom is whether pi merges to A or to B, for all i. This
is 2N combinations total. The crux of the proof is to show there exists an α for each of these
behaviors.
We attack the problem as follows. In round 1, the following equation specifies whether p1 merges
to A or B:
1
1
(d(pa , p1 )α + d(qa , p1 )α ) ≶ (d(pb , p1 )α + d(qb , p1 )α )
2
2
If the LHS is smaller, then p1 merges to A, otherwise B. By carefully setting the distances, we
will ensure there exists a value α0 which is the only solution to the equation in the range (1, 3).
Then p1 merges to A for all α ∈ (1, α0 ), and B for all α ∈ (α0 , 3). For now, assume it is easy to
force q1 to merge to the opposite set.
In round 2, there are two equations:
1
(d(pa , p2 )α + d(qa , p2 )α + d(p1 , p2 )α ) ≶
3
1
(d(pa , p2 )α + d(qa , p2 )α + d(q1 , p2 )α ) ≶
3
1
(d(pb , p2 )α + d(qb , p2 )α + d(q1 , p2 )α ),
3
1
(d(pb , p2 )α + d(qb , p2 )α + d(p1 , p2 )α ).
3
The first equation specifies where p2 merges in the case when p1 ∈ A, and the second equation is
the case when p1 ∈ B. So we must ensure there exists a specific α[−1] ∈ (1, α0 ) which solves equation
1, and α[1] ∈ (α0 , 3) which solves equation 2, and these are the only solutions in the corresponding
intervals.
In general, round i has 2i−1 equations corresponding to the 2i−1 possible states for the partially
constructed tree. For each state, there is a specific α interval which will cause the algorithm to
reach that state. We must ensure that the equation has exactly one solution in that interval. If we
achieve this simultaneously for every equation, then the next round will have 2 · 2i−1 states. See
Figure 9 for a schematic of the clustering instance.
For 1 ≤ i ≤ N , given x ∈ {−1, 1}i−1 , let Ex denote the equation in round i which determines
where pi merges, in the case where for all 1 ≤ j < i, pj merged to A if xj = −1, or B if xj = 1
(and let E 0 denote the single equation for round 1). Let αx ∈ (1, 3) denote the solution to Ex = 0.
Then we need to show the α’s are well-defined and follow a specific ordering, shown in Figure
10. This ordering is completely specified
by two conditions: (1) α[x −1] < α[x] < α[x 1] and (2)
S
α[x −1 y] < α[x 1 z] for all x, y, z ∈ i<N {−1, 1}i and |y| = |z|.
Now we show how to set up the distances to achieve all of these properties. To enhance
readability, we start with an example for n = 10, and then move to the general construction.
45
qb
pb
1
≈ 1.151
≈ 1.151
p1 2
pi
q1
qi
pN
qN
1.2
All distances here ≈ 1.5
1.1
1
qa
pa
Figure 9: The clustering instance used in Lemma 10
α0
p1
q1
α[−1]
p2
q2
α[−1 −1]
α[−1 −1]
α[1]
p2
α[−1 1]
α[−1]
α[−1 1]
q2
α[1 −1]
α0
α[1 −1]
α[1 1]
α[1]
α[1 1]
Figure 10: A schematic for the α intervals. Each edge denotes whether to merge pi to A or qi to A.
46
We give the construction round by round. All distances are in [1, 2] which ensures the triangle
inequality is always satisfied. Set d(pa , qa ) = d(pb , qb ) = 1, and set all pairwise distances between
{pa , qa } and {pb , qb } to 2. We also set d(pi , qi ) = 2 for all i 6= j.
Here are the distances for the first round.
d(pa , p1 ) = d(pa , q1 ) = 1.1, d(qa , p1 ) = d(qa , q1 ) = 1.2,
p
d(pb , p1 ) = d(pb , q1 ) = d(qb , p1 ) = d(qb , q1 ) = (1.12 + 1.22 )/2 ≈ 1.151.
Say we break ties by lexicographic order. Then p1 will merge first, and E 0 is the following.
1
1
(1.1α + 1.2α ) ≶ (1.151α + 1.151α ).
2
2
The unique solution in (1, 3) is α0 = 2. p1 merges to A for α ∈ (1, 2), and B for α ∈ (2, 3).
The merge equations for q1 are
1
(1.1α + 1.2α + 2α ) ≶
3
1
(1.1α + 1.2α ) ≶
2
1
(1.151α + 1.151α ) if p1 ∈ A, or
2
1
(1.151α + 1.151α + 2α ) otherwise.
3
As long as α ∈ (1, 3), q1 will merge to the opposite cluster as p1 . This is because we set d(p1 , q1 )
to be significantly larger than the other relevant distances.
We set the round two distances as follows.
d(pa , p2 ) = d(pa , q2 ) = 1.1,
d(qa , p2 ) = d(qa , q2 ) = 1.2,
p
d(pb , p2 ) = d(pb , q2 ) = d(qb , p2 ) = d(qb , q2 ) = (1.12 + 1.22 )/2 ≈ 1.151,
d(p1 , p2 ) = d(p1 , q2 ) = 1.5 + 10−4 ,
d(q1 , p2 ) = d(q1 , q2 ) = 1.5 − 10−4 .
Note the distances from {pa , qa , pb , qb } are the same to p1 as to p2 . Since we break ties lexicographically, this ensures p2 and q2 merge in round 2. Alternatively, we can add tiny perturbations
to the distances such that they do not affect our analysis, but ensure the correct merge orders regardless of the tiebreaking rule. The values were picked so that the pa , qa , pb , qb distances will have
the most influence in the merge equations. If it were just these four distances, then the value of α
at equality would be 2 again. But in this round, the p1 and q1 distances show up in the equation.
We set d(p1 , p2 ), d(q1 , p2 ) ≈ 1.5 but with small ‘offsets’ so that the values of α are on either side of
α0 = 2. Equations E[−1] and E[1] are the following.
1
1
(1.1α + 1.2α + 1.5001α = (2 · 1.151α + 1.4999α ),
3
3
1
1
(1.1α + 1.2α + 1.4999α ) = (2 · 1.151α + 1.5001α ).
3
3
Then α[−1] ≈ 1.884 and α[1] ≈ 2.124.
As in the previous round, it is straightforward to check all four merge equations for q2 send q2
to the opposite cluster as p2 , as long as α < 3. So far, our four intervals of α which lead to distinct
behavior are (1, 1.884), (1.884, 2), (2, 2.124), and (2.124, 3).
47
Now we specify the distances for the third round, the final round in our example. Again, we
set the distances from {pa , qa , pb , qb } the same way as in previous rounds. The new distances are
as follows.
d(p1 , p3 ) = d(p1 , q3 ) = 1.5 + 10−4 ,
d(q1 , p3 ) = d(q1 , q3 ) = 1.5 + 10−4 ,
d(p2 , p3 ) = d(p2 , q3 ) = 1.5 + 10−6 ,
d(q2 , p3 ) = d(q2 , q3 ) = 1.5 − 10−6 .
Again, if it were just the first four points, the value of α would be 2. The distances for {p1 , q1 }
differ by a small offset, and the distances for {p2 , q2 } differ by an even smaller offset, which causes
the latter distances to have less influence in the merge equations. This forces the α’s into the
correct intervals. In general, the offset value will decrease in higher and higher rounds.
The equations E[−1 −1] , E[−1 1] , E[1 −1] , and E[1 1] are as follows:
1
(1.1α + 1.2α + (1.5 + 10−4 )α + (1.5 + 10−6 )α ) =
4
1
(1.1α + 1.2α + (1.5 + 10−4 )α + (1.5 − 10−6 )α ) =
4
1
(1.1α + 1.2α + (1.5 − 10−4 )α + (1.5 + 10−6 )α ) =
4
1
(1.1α + 1.2α + (1.5 − 10−4 )α + (1.5 − 10−6 )α ) =
4
1
(2 · 1.151α + (1.5 − 10−4 )α + (1.5 − 10−6 )α ),
4
1
(2 · 1.151α + (1.5 − 10−4 )α + (1.5 + 10−6 )α ),
4
1
(2 · 1.151α + (1.5 + 10−4 )α + (1.5 − 10−6 )α ),
4
1
(2 · 1.151α + (1.5 + 10−4 )α + (1.5 + 10−6 )α ).
4
Solving for the α’s, we obtain α[−1 −1] ≈ 1.882, α[−1 1] ≈ 1.885, α[1 −1] ≈ 2.123, α[1 1] ≈ 2.125.
Furthermore, solving the equations for q3 , we find that it will merge to the opposite cluster for
α < 3. Therefore, we have 8 different ranges corresponding to the 8 equations.
This example suggests a general argument by induction which follows the same intuition. In
each round, the new distances have less and less influence in the merge equations, ensuring the α’s
stay in the correct ranges to double the number of behaviors.
In our argument, we will utilize the following fact (true by elementary calculus).
Fact 1. For all 0 ≤ z ≤ .01 and α ∈ (1, 3), the following are true about g(z, α) = (1.5 − z)α −
α
(1.5 + z)α and h(z, α) = (1.1 − z)α + (1.1 + z)α − 2 · (((1.1 − z)α + (1.1 + z)α )/2) 2 .
1. For z > 0, g(z, α) < 0,
2. for a fixed z, g is nonincreasing in α,
3. for a fixed α, g is nonincreasing in z,
4. h(0, α) = 0 and h is nondecreasing in z.
Here are the details for the general construction. All distances will be between 1 and 2 so that
the triangle inequality is satisfied. Given N , for all i,
d(pa , qa ) = d(pb , qb ) = 1,
d(pa , qa ) = d(pa , qb ) = d(pb , qa ) = d(pb , qb ) = 2,
∀i ≤ N, d(pa , pi ) = d(pa , qi ) = 1.1 − q, d(qa , pi ) = d(qa , qi ) = 1.1 + q,
r
1
d(pb , pi ) = d(pb , qi ) = d(qb , pi ) = d(qb , qi ) =
((1.1 − q)2 + (1.1 + q)2 ),
2
d(pi , qi ) = 2,
∀1 ≤ j < i ≤ N, d(pi , pj ) = d(pi , qj ) = 1.5 + oj
d(qi , pj ) = d(qi , qj ) = 1.5 − oj .
48
where q and oj are offset values in (0, .01) which we will specify later. Then for α ∈ (1, 3), the
following are true.
• The first two merges are pa to qa and pb to qb ,
• {pi } and {qi } will always prefer merging to A or B instead of merging to another singleton
{pj } or {qj }.
After the first two merges occur, all pi and qi are tied to first merge to A or B. For convenience,
we specify the tiebreaking order as {p1 , q1 , . . . , pN , qN }. Alternatively, at the end we can make tiny
perturbations to the distances so that tiebreaking does not occur.
Next, we choose the value for q, which must be small enough to ensure that qi always merges
to the opposite cluster as pi . Consider
!
X
N +2
h(α, q, o1 , . . . , oN , x) =
(1.1 + q)α + (1.1 − q)α +
xi (1.5 + oi )α + 1.5α
N +3
i<N
X
α
2
2
− 2 · (((1.1 + q) + (1.1 − q) )/2) 2 −
xi (1.5 + oi )α .
i<N
If this equation is positive for all x ∈ {−1, 1}N −1 , then qN will always merge to the opposite
cluster as pN (and qi will always merge to the opposite cluster as pi , which we can similarly show
by setting oj = 0 in h for all j > i).
Note
h(α, 0, 0, . . . , 0, x) =
N +2
(2 · 1.1α + (N + 1) · 1.5α ) − 2 · 1.1α − N · 1.5α > 0
N +3
for all x and all α ∈ (1, 3). Fact 1 implies there exists a 0 < q ∗ < .01 such that h(α, q, 0, . . . , 0, x)
stays positive. Similarly, there exists a cutoff value δ > 0 such that for all 0 < o1 , . . . , oN < δ,
α ∈ (1, 3), and x ∈ {−1, 1}N −1 , h(α, q ∗ , o1 , . . . , ok , x) > 0. Therefore, as long as we set all the
offsets oi less than δ, the merges will be as follows:
1. pa merges to qa and pb merges to qb .
2. For 1 . . . , N , pi merges to A or B, and qi merges to the opposite cluster. Then qN will always
merge to the opposite cluster as pN .
N
Now
S we showi that there are 2 intervals for α ∈ (1, 3) which give unique behavior. Recall for
x ∈ i<N {−1, 1} , Ex is defined as
X
α
1
(1.1 − q ∗ )α + (1.1 + q ∗ )α − 2 · ( ((1.1 − q ∗ )2 + (1.1 + q ∗ )2 )) 2 +
xi ((1.5 − oi )α − (1.5 + oi )α ).
2
i<N
For brevity, we denote
1
1
d = ( ((1.1 − q ∗ )2 + (1.1 + q ∗ )2 )) 2 .
2
We show the αs are correctly ordered by proving the following three statements with induction.
The first statement is sufficient to order the αs, and the second two will help to prove
S the first. i
1. There exist 0 < o1 , . . . , oN < δ such that if we solve Ex = 0 for αx for all x ∈ i<N {−1, 1} ,
then the α’s satisfy α[x −1] < α[x] < α[x 1] and for all i < N , α[x 1] < α[y −1] for x, y ∈ {−1, 1}i
and x1 . . . xi < y1 . . . yi .
2. For all k 0 ≤ N and α ∈ (1, 3),
X
(1.5 + ok0 )α − (1.5 − ok0 )α +
((1.5 − oi )α − (1.5 + oi )α ) > 0.
k0 <i<N
49
3.
X
(1.1 − q ∗ )3 + (1.1 + q ∗ )3 − 2 · d3 +
∗
∗
(1.1 − q ) + (1.1 + q ) − 2 · d +
i<N
X
i<N
((1.5 − oi )3 − (1.5 + oi )3 ) > 0, and
((1.5 + oi ) − (1.5 − oi )) < 0.
We proved the base case in our earlier example for n = 10. Assume for k ≤ N , there exist
0 < o1 , . . . , ok < δ which satisfy the three properties. We first prove the inductive step for the
second and third statements.
By inductive hypothesis, we know for all k 0 ≤ k and α ∈ (1, 3),
X
(1.5 + ok0 )α − (1.5 − ok0 )α +
((1.5 − oi )α − (1.5 + oi )α ) > 0,
k0 <i≤k
Since there are finite integral values of k 0 ≤ k, and the expression is > 0 for all values of k 0 , then
there exists an > 0 such that the expression is ≥ for all values of k 0 . Then we define za such
that (1.5 + za )α − (1.5 − za )α < 2 for α ∈ (1, 3). Then for all 0 < z < za , k 0 ≤ k + 1, and α ∈ (1, 3),
(1.5 + ok0 )α − (1.5 − ok0 )α +
X
((1.5 − oi )α − (1.5 + oi )α ) > 0.
k0 <i≤k+1
So as long as we set 0 < ok+1 < za , the inductive step of the second property will be fulfilled. Now
we move to the third property. We have the following from the inductive hypothesis:
X
(1.1 − q ∗ )3 + (1.1 + q ∗ )3 − 2 · d3 +
((1.5 − oi )3 − (1.5 + oi )3 ) > 0,
i≤k0
(1.1 − q ∗ ) + (1.1 + q ∗ ) − 2 · d +
X
i≤k0
((1.5 + oi ) − (1.5 − oi )) < 0.
We may similarly find zb such that for all 0 < ok+1 < zb ,
(1.1 − q ∗ )3 + (1.1 + q ∗ )3 − 2 · d3 +
∗
∗
X
((1.5 − oi )3 − (1.5 + oi )3 ) > 0,
i≤k+1
(1.1 − q ) + (1.1 + q ) − 2 · d +
X
((1.5 + oi ) − (1.5 − oi )) < 0.
i≤k+1
Now we move to proving the inductive step of the first property. Given x ∈ {−1, 1}k , let
p(x), n(x) ∈ {−1, 1}k denote the vectors which sit on either side of αx in the ordering, i.e., αx is
the only αy in the range (αp(x) , αn(x) ) such that |y| = k. If x = [1 . . . 1], then set αn(x) = 3, and
if x = [0 . . . 0], set αp(x) = 1. Define
f (α, x, z) = Ex + (1.5 − z)α − (1.5 + z)α .
By inductive hypothesis, we have that f (αx , x, 0) = 0. We must show there exists zx such that
for all 0 ≤ z ≤ zx , f (αx , x, z) < 0 and f (αn(x) , x, z) > 0. This will imply that if we choose
0 < ok+1 < zx , then α[x 1] ∈ (αx , αn(x) ).
50
Case 1: x 6= [1 . . . 1]. Since f (αx , x, 0) = 0, and by Fact 1, then for all 0 < z < .01, f (αx , x, z) <
0. Now denote i∗ as the greatest index such that xi∗ = −1. Then n(x) = [x1 . . . xi∗ −1 1 − 1 · · · − 1].
By statement 1 of the inductive hypothesis (αn(x) is a root of En(x) = 0),
(1.1 − q ∗ )αn(x) + (1.1 + q ∗ )αn(x) − 2 · dαn(x) +
X
(n(x)i (1.5 − oi )αn(x) − n(x)i (1.5 + oi )αn(x) ) = 0
i≤k
From statement 2 of the inductive hypothesis, we know that
X
(1.5 − oi∗ )αn(x) − (1.5 + oi∗ )αn(x) +
((1.5 + oi )αn(x) − (1.5 − oi )αn(x) ) < 0.
i∗ <i≤k
It follows that
(1.1 − q ∗ )αn(x) + (1.1 + q ∗ )αn(x) − 2 · dαn(x) +
X
i<i∗
(n(x)i (1.5 − oi )αn(x) − n(x)i (1.5 + oi )αn(x) ) > 0,
and furthermore,
(1.1 − q ∗ )αn(x) + (1.1 + q ∗ )αn(x) − 2 · dαn(x) +
X
i<i∗
(xi (1.5 − oi )αn(x) − xi (1.5 + oi )αn(x) ) > 0.
Therefore, f (αn(x) , 0) > 0, so denote f (αn(x) , 0) = > 0. Then because of Fact 1, there exists zx
such that ∀0 < z < zx , f (αn(x) , z) > 0.
Case 2: x = [1 . . . 1]. Since f (αx , 0) = 0, and by Fact 1, then for all 0 < z < .01, f (αx , z) < 0.
By property 3 of the inductive hypothesis, we have
X
(1.1 − q ∗ )3 + (1.1 + q ∗ )3 − 2 · d3 +
((1.5 − oi )3 − (1.5 + oi )3 ) > 0,
i≤k
so say this expression is equal to some > 0. Then from Fact 1, there exists zx such that for all
0 < z < zx , 0 < (1.5 + z)3 − (1.5 − z)3 < 2 . Combining these, we have f (3, z) > 0 for all 0 < z < zx .
To recap, in both cases we showed there exists zx such that for all 0 < z < min(.01, zx ),
f (αx , z) < 0 and f (αn(x) , z) > 0. We may perform a similar analysis on a related function f 0 ,
defined as f 0 (α, x, z) = Ex + (1.5 + z)α − (1.5 − z)α to show there exists zx0 such that for all
0 < z < zx0 , f 0 (αp(x) , z) < 0 and f 0 (αx , z) > 0. We perform this analysis over all x ∈ {−1, 1}k .
Finally, we set ok+1 = minx (zx , zx0 , za , zb , .01). Given x ∈ {−1, 1}k , since f (αx , ok+1 ) < 0 and
f (αn(x) , ok+1 ) > 0, there must exist a root α[x 1] ∈ (αx , αn(x) ) (and by Fact 1, the function is
monotone in α in the short interval (αx , αn(x) ), so there is exactly one root). Similarly, there must
exist a root α[x −1] ∈ (αp(x) , αx ). Then we have shown α[x −1] and α[x 1] are roots of E[x −1] and
E[x 1] , respectively. By construction, α[x −1] < αx < α[x 1] , so condition 1 is satisfied. Now we
need to show condition 2 is satisfied. Given x, y ∈ {−1, 1}k , let k 0 be the largest number for which
xi = yi , ∀i ≤ k 0 . Let z = x[1...k0 ] = y[1...k0 ] . Then by the inductive hypothesis,
αx < αn(x) ≤ αz ≤ αp(y) < αy .
It follows that
α[x
−1]
< α[x
1]
< αz < α[y
proving condition 2. This completes the induction.
51
−1]
< α[y 1] ,
Now we are ready to prove Lemma 10.
Proof of Lemma 10. Given n, and setting N = b(n − 8)/4c, we will show there exists a clustering
instance (V, d) of size |V | = n, a witness r, and a set of 2N + 2 α’s 1 = α0 < α1 < · · · < α2N <
(p)
α2N +1 = 3, such that ΦA3 (α) (V) oscillates above and below r between each interval (αi , αi+1 ).
We start by using the construction from Lemma 10, which gives a clustering instance with
2N + 8 points and 2N + 2 values of α for which α-linkage creates a unique merge tree. The next
part is to add 2N more points and define a witness r so that the cost function alternates above
and below r along each neighboring α interval, for a total of 2N oscillations. Finally, we will finish
off the proof in a manner similar to Lemma 8.
Starting with the clustering instance (V, d) from Lemma 10, we add two sets of points, CA and
CB , which do not interfere with the previous merges, and ensure the cost functions alternates. Let
CA = {ca , c0a , a1 , a2 , . . . , aN } and CB = {cb , c0b , b1 , b2 , . . . , bN }. All distances between two points in
CA are 1, and similarly for CB . All distances between a point in CA and a point in CB are 2. The
distances between CA ∪ CB and A ∪ B are as follows (we defined the sets A and B in Lemma 10).
d(pa , ca ) = d(pa , c0a ) = d(qa , ca ) = d(qa , c0a ) = 1.51,
d(pb , cb ) = d(pb , c0b ) = d(qb , cb ) = d(qb , c0b ) = 1.51,
d(pa , cb ) = d(pa , c0b ) = d(qa , cb ) = d(qa , c0b ) = 2,
d(pb , ca ) = d(pb , c0a ) = d(qb , ca ) = d(qb , c0a ) = 2,
d(pa , c) = d(qa , c) = d(pb , c) = d(qb , c) = 2 ∀c ∈ CA ∪ CB \ {ca , c0a , cb , c0b },
d(c, pi ) = d(c, qi ) = 1.51 ∀1 ≤ i ≤ N − 1 and c ∈ CA ∪ CB .
We will specify the distances between {ca , c0a , cb , c0b } and {pN , qN } soon, but they will be in
[1.6, 2]. So at the start of the merge procedure, all points in CA merge together, and all points in
CB merge together. Then all merges from Lemma 10 take place, because all relevant distances are
smaller than 1.51. We end up with four sets: A, B, CA , and CB . The pairs (A, B) and (CA , CB ) are
dominated by distances of length 2, so the merges (CA , A) and (CB , B) will occur, which dominate
(CA , B) and (CB , A) because of the distances between {pa , qa , pb , qb } and {ca , c0a , cb , c0b }. The final
merge to occur will be (CA ∪ A, CB ∪ B), however, the 2-median pruning step will clearly pick the
2-clustering CA ∪ A, CB ∪ B, since no other clustering in the tree has almost all distances ≤ 1.51.
Then by construction, ca or c0a will be the best center for CA ∪ A, which beat pa and qa because
1.51 · (2N ) < 1.1 · N + 2 · N = 1.55 · (2N ). Similarly, cb or c0b will be the best center for CB ∪ B.
Note that centers {ca , c0a } and {cb , c0b } currently give equivalent 2-median costs. Denote this cost
by r0 (i.e., the cost before we set the distances to pN and qN ).
Now we set the final distances as follows.
d(ca , pN ) = d(cb , qN ) = 1.6,
d(c0a , pN ) = d(c0b , qN ) = 1.7,
d(c0a , qN ) = d(c0b , pN ) = 1.8,
d(ca , qN ) = d(cb , pN ) = 1.9.
If pN ∈ A and qN ∈ B, then ca and cb will be the best centers, achieving cost r0 + 3.2 for
(CA ∪ A, CB ∪ B). If pN ∈ B and qN ∈ A, then c0a and c0b will be the best centers, achieving cost
r0 + 3.6 for (CA ∪ A, CB ∪ B).
52
The distances are also constructed so that in the variant where the pruning outputs the optimal
centers, and then all points are allowed to move to their closest center, the cost still oscillates. First
note that no points other than pN and qN are affected, since d(ca , pi ) = d(ca , qi ) for i < N , and
similarly for cb . Then pN will move to the cluster with ca or c0a , and qN will move to the cluster
with cb or c0b . If pN was originally in A, then the cost is r0 + 3.2, otherwise the cost is r0 + 3.4.
In either scenario, we set r = r0 + 3.3. Then we have ensured for all x ∈ {−1, 1}N −1 , the cost
for α ∈ (αp(x) , αx ) is < r, and the cost for α ∈ (αx , αn(x) ) is > r.
We have finished our construction of a clustering instance whose cost function alternates 2N
times as α increases. To finish the proof, we will show there exists a set S = {V1 , . . . , Vs } of size
s = N = b(n − 8)/4c ∈ Ω(n) that is shattered by A. Such a set has 2N orderings total. For V1 ,
we use the construction which alternates 2N times. For V2 , we use the same construction, but we
eliminate (pN , qN ) so that there are only N − 1 rounds (the extra two points can be added to CA
and CB to preserve |V2 | = n). Then V2 ’s cost will alternate 12 · 2N times, between the intervals
(αp(x) , αx ) and (αx , αn(x) ), for x ∈ {−1, 1}N −2 . So V2 oscillates every other time V1 oscillates, as α
N
increases. In general, Vi will be the construction with only N − i + 1 rounds, oscillating 2 2i−1 times,
and each oscillation occurs every other time Vi−1 oscillates. This ensures for every x ∈ {−1, 1}N −1 ,
(αp(x) , αx ) and (αx , αn(x) ) will have unique labelings, for a total of 2N labelings. This completes
the proof.
Note C.2. As in Lemma 8, this lower bound holds even if the cost function is the symmetric
distance to the ground truth clustering. Merely let pN and qN belong to different ground truth
clusters, but for all i < N , pi and qi belong to the same ground truth cluster. Since in each
adjacent α interval, pN and qN switch clusters, this shows the symmetric distance to the ground
truth clustering oscillates between every interval.
Now we give an ERM algorithm for A2 , similar to Algorithm 4.
Algorithm 9 An algorithm for finding an empirical cost minimizing algorithm in A2
Input: Sample S = V (1) , . . . , V (m) .
1: Let T = ∅. For each sample V (i) = V (i) , d(i) ∈ S, and for all A, B, X, Y ⊆ V (i) , solve for α (if
a solution exists) in the following equation and add the solutions to T :
1
|A||B|
X
(d(p, q))α =
p∈A,q∈B
1
|X||Y |
X
(d(x, y))α .
x∈X,y∈Y
Order the elements of set T ∪ {−∞, +∞} as α1 < . . . < α|T | . For each 0 ≤ i ≤ |T |, pick an
arbitrary
α in the interval (αi , αi+1 ) and run A2 (α) on P
all clustering instances in S to compute
P
Φ
(V).
Let
α̂
be
the
value
which
minimizes
V∈S ΦA2 (α) (V).
V∈S A2 (α)
Output: α̂
2:
Theorem 21. Let Φ be a clustering objective
and let Ψ be a pruning function. Given an input
H 2
1
sample of size m = O
n + log δ , Algorithm 9 (, δ)-learns the class A2 × {Ψ} with respect
to the cost function Φ.
Proof. The sample complexity analysis follows the same logic as the proof of Theorem 6. To
prove that Algorithm 9 indeed finds the empirically best α, recall from the pseudo-dimension
analysis that the cost as a function of α for any instance is a piecewise constant function with
53
O(n2 32n ) discontinuities. In Step 1 of Algorithm 9, we solve for the values of α at which the
2 2n
discontinuities occur and add them to the set
PT . T therefore partitions α’s range into O(mn 3 )
subintervals. Within each of these intervals, V∈S ΦA2 (α) (V) is a constant function. Therefore, we
pick any arbitrary α within each interval to evaluate the empirical cost over all samples, and find
the empirically best α.
C.1
Restricted classes of clustering instances
Below we consider restricted classes of clustering instances and improve on the pseudo-dimension
bounds of HA2 as compared to Lemma 9. In particular, we consider the class Vβ that consists of
clustering instances in which the distances take one of at most β (β ∈ N) real values. A natural
example would be one in which all distances are integers and are less than some value H. Here,
β = H. For this case, we show a tight bound.
Theorem 22. For any objective function Φ, let
HA2 ,Φ = ΦA2 (α) : Vβ → R≥0 | α ∈ R ∪ {∞, −∞} .
Then Pdim(HA2 , Φ) = O (min(β log n, n)).
Proof. The proof for the upper bound follows a similar line of reasoning as that of Lemma 9. For a
particular instance V, let {d1 , d2 , . . . , dβ } denote the set of β values that its distances can take. The
Pβ
1
α
linkage criterion for merging A, B can be expressed as |A||B|
i=1 ai di where, each ai can take one
of at most O(n2 ) values corresponding to the number of pairs of points at a distance di . Therefore,
iterating over allpairs of subsets (A, B) like we did in the proof of Lemma 9, we can only list at
most O (n2 )β+1 potential linkage criteria. Therefore, the set of all pairs of subsets (A, B) and
(X, Y ) will induce at most O (n2 )2(β+1) unique comparisons between two linkage criteria. By the
same argument as in the proof of Lemma 9, since each such comparison has O(n2 ) roots, if S ∈ Vm
is a shatterable set, then 2m ≤ n2 ((n2 )2(β+1) ), which means that m = O(β log n).
Theorem 23. For any objective function Φ(p) , Pdim(HA2 ,Φ(p) ) = Ω (min(β, n)).
Proof. For the lower bound, we use a similar line of reasoning as in Lemma 10. In each round of
the construction in Lemma 10, only a constant number of distinct edge lengths are added. I.e. the
offsets oi define new distances 1.5 + oi and 1.5 − oi per round, but the set of the rest of the distances
has constant size. Therefore, we can easily modify the proof to construct a clustering instance with
Ω(m) rounds using m distinct distances. This instance will have 2Ω(β) distinct behaviors depending
on α, however this reasoning is only consistent for β = o(n). For β = Ω(n), we may inherit the
lower bound from Lemma 10, so the final pseudo-dimension lower bound is Ω (min(β, n)).
C.2
An Interpolation
So far, the linkage criteria have been based on the distances between either two pairs of points (A1
and A3 ), or every single pair between two sets considered for merging (A2 ). Now we provide an
interpolation between these two extremes. In particular, we define a linkage criterion which uses
σ different distances between the sets for comparison. In particular, for any two sets A and B, we
define an abstract rule to pick σ pairs of points (pi , qi ) from A × B. For example, a natural choice
would be to pick the (σ − 1)-th quantiles of the set of distances between points in A and B along
with the maximum and minimum distances. On picking these points, we define the criterion as a
function of these σ distances as follows.
54
A1,σ =
A2,σ =
( σ
X
i=1
αi d(pi , qi ) | α = (α1 , . . . , ασ ) ∈ Rσ
σ
X
d(p, q))α
i=1
!1/α
)
| α ∈ R ∪ {∞, −∞}
,
(3)
.
(4)
Observe that A1,σ has multiple parameters unlike any of the classes of algorithms we have
discussed. Therefore, the analysis for A1,σ is considerably different from the rest as shown below.
We use notations similar to the previous sections.
Theorem 24. For any Φ, let
HA1,σ ,Φ = ΦA1,σ (α) : V → R≥0 | α ∈ Rσ .
The pseudo-dimension of Pdim(HA1,σ ,Φ ) = O(σ 2 log n).
Proof. The proof parallels that of Lemma 7. Consider two pairs of sets A, B and X, Y that can be
potentially merged. Regardless of the parameters chosen, we know that the linkage criterion first
chooses σ pairs of points (pi , qi ) ∈ A × B and (xi , yi ) ∈P
X × Y . Now, theP
decision to merge A, B
before X, Y or vice versa is determined by the sign of σi=1 αi d(pi , qi ) − σi=1 αi d(xi , yi ). If this
expression evaluates to zero, we will break ties arbitrarily but consistently.
Observe that for a given set of values for d(pi , qi ) and d(xi , yi ), the above expression is either 0 for all α, or is equal to zero for the hyperplane passing through the origin and normal to
(d(p1 , q1 ) − d(x1 , y1 ), . . . , d(pσ , qσ ) − d(xσ , yσ )) in the parameter space Rσ . This hyperplane divides
the parameter space into two half-spaces each of which correspond to merging one pair of sets
before the other.
Next we note that, for a given
problem instance, as we iterate over all pairs of sets (A, B) and
(X, Y ), we can list only O n4σ possible choices for the hyperplane as there are only so many
4σ-tuples of points pi , qi , xi , yi . Thus, for m problem instances, we can list at most O mn4σ
different hyperplanes
in Rσ . These hyperplanes, can partition the parameter space into at most
σ
4σ
O mn
regions such that all the parameter settings in a given region correspond to identical
merge trees and hence identical costs/labeling induced by the witnesses. By an argument similar
to the proof of Lemma 7, we can conclude that m = O(σ 2 log n).
Theorem 25. For any Φ, let
HA2,σ ,Φ = ΦA2,σ (α) : V → R | α ∈ R≥0 ∪ {∞, −∞ .
Then Pdim(HA2,σ ) = O (min(σ log n, n)).
Proof. This proof follows the same reasoning as in Lemma 9. The decision of whether to merge
any two pairs of sets
Pσ(A, αB) and (X,
PYσ ), isαdetermined by the sign of the difference in their linkage
criterion which is i=1 d (pi , qi ) − i=1 d (xi , yi ) where the points pi , qi , xi , yi are fixed for a given
pairs of set. We know from Theorem 20 that this expression has only O(n2 ) roots. Furthermore, as
we iterate over all pairs of sets (A, B) and (X, Y ), we can only generate as many such expressions
as there are 4σ-tuples of points pi , qi , xi , yi . In particular, we will list only O(n4σ ) such expression
each with O(n2 ) roots. In summary, similar to the proof of Lemma 9, we can argue that for a set
of samples in V m to be shattered, we need that 2m ≤ n4σ · n2 i.e., m = O(σ log n).
55
Theorem 26. For any Φ(p) Pdim(HA2,σ ,Φ(p) ) = Ω(σ).
Proof. We use a similar line of reasoning as in Lemma 10. In round i of the construction in Lemma
10, the merges were between a set of size 1 and a set of size i + 2. Therefore, we can easily modify
the proof to construct a clustering instance with Ω(σ) rounds, and the merge equations will be the
same as in Lemma 10. This instance will have 2Ω(σ) distinct behaviors depending on α, and so the
pseudo-dimension is Ω(σ).
Algorithm
Class
Linkage rule
Pseudodimension
Runtime
of
Learning Algorithm
A1
Θ(log n)
Õ(n8 talg )
Θ(n)
Õ(n2 32n talg )
A3
minp∈A,q∈B (d(p, q))α
+
α
maxp∈A,q∈B (d(p, q)) for α ∈ R
P
1
α
p∈A,q∈B d(A, B) for α ∈ R
|A||B|
α minp∈A, q∈B d(p, q) + (1 −
α) maxp∈A, q∈B d(p, q) for α ∈ (0, 1)
Θ(log n)
Õ(n8 talg )
A2
with β unique distances
Pσ
i=1 αi d(pi , qi ) for αi ∈ R
Pσ
α
i=1 d(pi , qi ) for αi ∈ R
Θ̃(min(β, n))
Õ(n4(β+1) talg )
A2
A1,σ
A2,σ
O(σ 2 log n)
Θ̃(σ)
Õ(n2σ+2 talg )
Table 2: The different classes of algorithms and their corresponding linkage rule, pseudo-dimension,
and runtime. talg denotes the runtime of the α-linkage algorithm for an arbitrary linkage rule.
56
| 8cs.DS
|
arXiv:1711.06735v1 [math.GR] 17 Nov 2017
ENDOMORPHISMS OF REGULAR ROOTED TREES INDUCED BY
THE ACTION OF POLYNOMIALS ON THE RING Zd OF d-ADIC
INTEGERS
ELSAYED AHMED AND DMYTRO SAVCHUK
Abstract. We show that every polynomial in Z[x] defines an endomorphism of the dary rooted tree induced by its action on the ring Zd of d-adic integers. The sections of this
endomorphism also turn out to be induced by polynomials in Z[x] of the same degree. In the
case of permutational polynomials acting on Zd by bijections the induced endomorphisms are
automorphisms of the tree. In the case of Z2 such polynomials were completely characterized
by Rivest in [Riv01]. As our main application we utilize the result of Rivest to derive the
condition on the coefficients of a permutational polynomial f (x) ∈ Z[x] that is necessary
and sufficient for f to induce a level transitive automorphism of the binary tree, which is
equivalent to the ergodicity of the action of f (x) on Z2 with respect to the normalized Haar
measure.
1. Introduction
For a fixed integer d ≥ 2 every polynomial f (x) ∈ Z[x] naturally induces mappings
fn : Z/(dn Z) → Z/(dn Z) for all positive integers n. Equivalently, these mappings are induced
by the action of f on the ring of d-adic integers Zd . These two equivalent approaches
to study polynomials in Z[x] have been used in different contexts during the last several
decades. One of the first motivations came from the constructions of the generators for
pseudo-random sequences and goes back to Knuth [Knu81, Ana98, Lar02]. In this and most
of the other applications it is crucial to consider polynomials acting by permutations on
Z/(dn Z). Such polynomials are generally called permutational polynomials, however it is
important to emphasize the distinction of these polynomials from the class of permutation
polynomials that permute elements of finite fields Fpn (see [LN83, Chapter 7] for a survey).
In many cases the stronger condition of transitivity of this action is required. Another
type of applications come from cryptography. Rivest in [Riv01] completely characterized
polynomials that act by permutations on Z/(2n Z) for all n ≥ 1 and points out the use of
one of them (namely, f (x) = 2x2 + x) in the symmetric block cipher RC6 [RRSY] that was
one of the five finalists of the AES competition. The questions of ergodicity of the action of
permutational polynomials have been studied in the context of dynamical systems on Zp by
Anashin [Ana06]. We refer the reader to a nice survey paper [Fan15] for a background and
history in polynomial dynamics on Z/(dn Z).
In this paper we offer another view to the polynomials acting on Zd and on Z/(dn Z).
Namely, we use the tools from the theory of groups acting on rooted trees by automorphisms and groups generated by Mealy automata. This theory exploded in 1980’s after
many counterexamples to well known conjectures in group theory were found among such
groups. For example, Grigorchuk group was the first example of a group of intermediate
growth [Gri83], as well as one of the first examples of infinite finitely generated torsion
1
groups [Gri80] (see also [Ale72, Sus79, GS83]). Now this is a rich theory with connections
to combinatorics [GŠ06], analysis [GLSŻ00], holomorphic dynamics [Nek05], dynamical systems [GS16], computer science [MS15], and many other areas. We refer the reader to the
survey article [GNS00] for history and references. The key idea in many of the arguments
in this theory is understanding automorphisms of rooted trees by describing their sections
(terms states and restrictions are also widely used) at the subtrees hanging down from vertices of the original rooted tree. If the original tree was regular (i.e., every vertex has the
same number of children), these subtrees are canonically isomorphic to the original tree, and
these sections can be treated as the automorphisms of the original tree as well. We utilize
this approach to analyze the action of polynomials on Zd .
Note that the connection between the functions on the boundary of X ∗ induced by automata and 1-Lipschitz functions on Zd was also established by Anashin in [Ana12], where
a criterion for finiteness of the corresponding automaton in terms of the van der Put series
of the function was developed. This criterion provided an application of the p-adic analysis
to the theory of automata. In this paper we suggest a converse application.
The set of vertices of a rooted d-ary tree can be identified with the set X ∗ of all finite
words over an alphabet X = {0, 1, . . . , d−1}. In this case the n-th level of X ∗ corresponds to
Z/(dn Z) by identifying x0 x1 . . . xn−1 ∈ X n with x0 + d · x1 + · · · + dn−1 · xn−1 ∈ Z/(dn Z). The
boundary of X ∗ (consisting of all infinite paths without backtracking initiating at the root)
is then identified with the ring Zd of d-adic integers. We show in Proposition 3.1 that in this
interpretation each polynomial in Z[x] induces an endomorphism of the tree X ∗ , while each
permutational polynomial induces an automorphism. Our first result describes the structure
of these endomorphisms.
Theorem 3.2. Given a polynomial f (x) = a0 + a1 x + · · · + at xt ∈ Z[x] inducing an endomorphism of X ∗ , the image of a vertex x0 ∈ X under f is f (x0 ) (mod d) and the section of
f at x0 is again a polynomial given by the equation:
t
X
f (i) (x0 ) i−1 i
(1)
f |x0 (x) = f (x0 ) div d +
d x,
i!
i=1
where f (i) denotes the ith derivative of a polynomial f , and f (x0 ) div d is the quotient of
f (x0 ) and d.
Note that the case of linear polynomials was partially considered by Bartholdi and Šunı́c
in [BŠ06].
Our main application deals with permutational polynomials acting transitively on Z/(2n Z)
for all n. In terms of the action on the tree this condition is equivalent to being level
transitive. Equivalently, f ∈ Z[x] induces a level transitive automorphism if and only if the
corresponding dynamical system (Z2 , f ) is minimal (i.e., the orbit of each element of Z2 under
f is dense in Z2 ), or ergodic with respect to the Haar measure on Z2 (coinciding with the
uniform Bernoulli measure on Z2 viewed as a Cantor set) [GNS00, Proposition 6.5],[Ana06].
In order to state our main result, we first review the history of the problem. The following
theorem prover by Larin in [Lar02] gives the conditions that f has to satisfy in order to be
transitive mod 2n for each positive integer n.
Theorem 4.7 ([Lar02]). A polynomial f (x) = a0 + a1 x + · · · + at xt ∈ Z[x] is transitive
mod 2n for every positive integer n if and only if it satisfies the following conditions:
2
(i)
(ii)
(iii)
(iv)
a0 ≡ 1 (mod 2)
a1 ≡ 1 (mod 2)
a3 + a5 + a7 + · · · ≡ 2a2 (mod 4)
a4 + a6 + a8 + · · · ≡ a1 + a2 − 1 (mod 4)
Rivest in [Riv01] (see an alternative proof in [MŠG10]) derived the following conditions
that are necessary and sufficient for a polynomial f ∈ Z[x] to induce a permutation of each
level of {0, 1}∗ and hence an automorphism of the tree.
Theorem 3.8 ([Riv01]). A polynomial f (x) = a0 + a1 x + · · · + at xt in Z[x] induces a
permutation of Z/(2n Z) for each positive integer n if and only if it satisfies the following
conditions:
(i) a1 ≡ 1 (mod 2)
(ii) a2 + a4 + a6 + · · · ≡ 0 (mod 2)
(iii) a3 + a5 + a7 + · · · ≡ 0 (mod 2)
Using Theorem 3.2 we study level transitivity of a permutational polynomial f by counting
the number of nontrivial actions of the sections of f in each level. In [BOERT96, Proposition
(4.6)], it was shown that f acts level transitively on the rooted binary tree if and only if the
number of nontrivial actions of the sections of f in each level of {0, 1}∗ is odd. Using this
fact, we determine the conditions that f has to meet in order to be level transitive. These
conditions are summarized in the following theorem which is the main result of the paper.
Theorem 4.6. Let f (x) = a0 + a1 x + · · · + at xt be a permutational polynomial acting on the
rooted binary tree. Then this action is level transitive if and only if the following conditions
hold:
(i) a0 ≡ 1 (mod 2)
(ii) 2a2 ≡ a3 + a5 + · · · (mod 4)
(iii) a2 + a1 − 1 ≡ a4 + a6 + · · · (mod 4)
Combining the conditions of Theorems 3.8 and 4.6, we obtain the conditions of Theorem 4.7 using a completely different approach. We hope that this new tool can be utilized
to attack other problems from p-adic analysis. For example, the suggested approach may
work to characterize ergodicity of polynomials acting on Zp for p > 2.
The structure of the paper is as follows. In Section 2 we set up necessary notation regarding rooted trees and their automorphisms. Section 3 describes endomorphisms and automorphisms of rooted trees arising from polynomial actions on the ring of d-adic integers.
Finally, Section 4 contains the main result on the conditions for a permutational polynomial
to act level transitively on the binary tree, which is equivalent to an ergodic action of the
polynomial on Z2 .
Acknowledgement. The authors would like to thank Zoran Šunı́c and Said Sidki for
fruitful and motivating discussions during the preparation of the manuscript.
2. Preliminaries
We start this section with the notation and terminology that will be used throughout the
paper. A tree is a connected graph with no cycles. A rooted tree is a tree with one vertex
selected to be the root. In any connected graph, a metric (called combinatorial metric) can
be defined such that the distance between any pair of vertices is the number of edges in a
3
shortest path (geodesic) connecting them. The nth level of a rooted tree is defined as the set
of vertices whose distance from the root is n. Since the tree has no cycles, for each vertex
v of the nth level there is only one path from the root to v. The vertex u in this path that
lies in level n − 1 is called the parent of v. The vertex v is called a child of u. Hence every
vertex except for the root has exactly one parent and may have some children.
A rooted tree is said to be d-regular or d-ary if there exists some positive integer d such
that each vertex of the tree has exactly d children. So all these trees have infinitely many
levels. In case d = 2, such a tree is called a rooted binary tree which represents the main
interest in our paper. We will always visualize those trees such that they grow from top to
bottom. So the root is the highest vertex and the children of each vertex v are located right
under v.
We will label the vertices of a rooted d-ary tree by finite words over a finite alphabet
X = {0, 1, . . . , d − 1}. Equivalently the set X ∗ of all finite words over X can be given the
structure of a rooted d-ary tree by declaring that v is adjacent to vx for each v ∈ X ∗ and
x ∈ X. Thus the empty word corresponds to the root and for each positive integer n the set
X n corresponds to the nth level of the tree. The example of the rooted binary tree is shown
in Figure 1.
0
0
1
0
1
00
01
0
000
0
10
1
001
1
010
2
11
2
011
3
100
4
3
101
5
110
6
111
7
Figure 1. Standard Numbering of vertices in the binary tree
We will identify the nth level of X ∗ with the ring Z/(dn Z) by identifying a vertex
x0 x1 . . . xn−1 ∈ X n with x0 + dx1 + · · · + dn−1xn−1 ∈ Z/(dn Z). For example, the vertices 00,
01, 10 and 11 of the second level of the rooted binary tree are identified with 0, 2, 1 and
3, respectively as shown in Figure 2. Moreover, the boundary of the tree can be naturally
identified with the ring Zd of d-adic integers.
The way we identified the nth level of X ∗ with Z/(dn Z) may not be the most natural way.
A more natural way to do that is to identify a vertex x0 x1 . . . xn−1 with its d-ary expansion
xn−1 + dxn−2 + · · · + dn−1x0 . So the vertices 00, 01, 10 and 11 of the second level of the
binary tree are identified with 0, 1, 2 and 3, respectively, as shown in Figure 1. However, we
adopt the first identification so that mappings induced by polynomials on X ∗ preserve the
adjacency relation as we will see later.
Definition 1. An endomorphism of X ∗ is a map from the set of vertices of X ∗ to itself which
preserves the adjacency relation. If such a map is bijective, it is called an automorphism of
X ∗.
An automorphism of X ∗ preserves the degree of each vertex as well as the distance from
each vertex to the root. Since the root is the only vertex with degree d, it is invariant under
4
0
+1
0
1
+2
+1
0
2
1
+4
0
+2
+1
3
+4
4
2
+4
+4
6
1
5
3
7
+8
0
8
4
12
2
10
6
14
1
9
5
13
3
11
7
15
+1 6
0 16
8 24
4 20 12 28
2 18 10 26
6 22 14 30
1 17
Even
9 25
5 21 13 29
3 19 11 27
7 23 15 31
Odd
Figure 2. Dyadic Numbering of vertices in the binary tree
all the tree automorphisms. Also the levels of the tree are invariant under automorphisms
since the distance is preserved. The group of all automorphisms of X ∗ is denoted by Aut X ∗ .
Another important concept that we want to introduce here is the definition of the section
of an endomorphism of X ∗ at a vertex.
Definition 2. Let h be an endomorphism of X ∗ and x ∈ X. For any word v ∈ X ∗ , we have
h(xv) = h(x)v ′ for some v ′ ∈ X ∗ . The map h|x : X ∗ → X ∗ given by
(2)
h|x (v) = v ′
clearly defines an endomorphism of X ∗ and is called the section of h at vertex x. Inductively
we can define the section of h at a vertex v = x1 x2 . . . xn as h|v = h|x1 |x2 . . . |xn .
In order to fully define the action of an endomorphism h on X ∗ , we need to specify its
action on the first level as well as its sections at the vertices of the first level. To see the
action of h on the second level, we do the same with the sections of h at the vertices of the
first level and so on. In case of an endomorphism h of a rooted binary tree, the action of
h on the first level is either trivial or is a switch of 0 and 1. Another language to define an
endomorphism of X ∗ is to give its wreath recursion which also specifies its action on the first
level and its sections at the vertices of the first level. This language makes computations
easier. Since we will not do any computations with endomorphisms, this approach will not
be used here.
We can visualize the action of h on X ∗ using what is called the full portrait of h. The full
portrait of h is a labeled infinite rooted d-ary tree with the root labeled by the name of the
endomorphism h and each vertex v labeled by h|v . Under each vertex v, we usually write
the name of the mapping that h|v defines on the first level of the subtree hanging from v.
In the case of a rooted binary tree, we draw a little arc (called switch) connecting the two
edges hanging down from v if h|v acts nontrivially on the first level of the subtree hanging
from v. If there is no switch, it means the action is trivial.
5
τ
τ
τ
τ
τ
τ
PSfrag replacements
Figure 3. The full portrait of the adding machine
One of the basic examples of automorphisms of the rooted binary tree is the adding
machine which we will denote throughout the paper by τ . It gets its name from the fact
that its action on the boundary of the tree, which is identified by the ring Z2 , is equivalent
to adding one to the input. The sections of τ at the vertices 0 and 1 are respectively e and
τ , where e is the identity automorphism. And τ acts nontrivially on the first level. The full
portrait of τ is shown in Figure 3.
The next definition introduces the notion of level transitivity which is the core concept
of our paper. Also a necessary and sufficient condition for an automorphism to act level
transitively on a rooted binary tree is provided in the next theorem, that is a partial case of
a more general result [BOERT96, Proposition (4.6)].
Definition 3. An automorphism h of X ∗ is said to act level transitively on X ∗ if it acts
transitively on each level of X ∗ .
Proposition 2.1. Let h be an automorphism of the rooted binary tree {0, 1}∗. Then h acts
level transitively if and only if the full portrait of h has an odd number of switches (nontrivial
actions) in each level including the zeroth level.
Remark 2.2. The proof of this theorem involves induction on the level of the tree and is
quite short. We leave it as an exercise and only point out that even though this is (the only)
external result that we use in this paper, its proof can be included here essentially at no
extra cost, so it does not affect our claim that our proof is self contained (modulo the result
of Rivest [Riv01] on the characterization of permutational polynomials).
Example 1. As we can see in Figure 3, the adding machine τ has exactly one switch in each
level. So the last theorem asserts that τ acts level transitively on the rooted binary tree.
3. Endomorphisms of Rooted Trees Arising from Polynomials over Z
For a fixed integer d ≥ 2, each polynomial f (x) ∈ Z[x] induces mappings fn : Z/(dn Z) →
Z/(dn Z) for all positive integers n by taking the evaluation map modulo dn . Identifying the
nth level of the rooted d-ary tree X ∗ with the ring Z/(dn Z), the polynomial f gives rise to a
mapping on the whole tree. In the next proposition, we show that this mapping is always an
6
endomorphism of the rooted tree as it preserves the root and the adjacency relation. If for
some d the mapping fn is a bijection for each n, then f induces a permutation of each level
of the rooted d-ary tree X ∗ and hence induces an automorphism of X ∗ . From now on, we
will use the term permutational polynomial to denote a polynomial in Z[x] that induces an
automorphism of X ∗ and to denote the automorphism it induces on X ∗ as well. The same
letter will be used to refer to both functions and no confusion will arise.
Proposition 3.1. Let f (x) ∈ Z[x]. Then f induces an endomorphism on the rooted d-ary
tree X ∗ . Moreover, different polynomials over Z induce different endomorphisms on X ∗ .
Proof. We pick two adjacent vertices u and v of the tree such that u is the parent of v. So
u = x0 + dx1 + · · · + dn−1 xn−1 ∈ X n
and
v = x0 + dx1 + · · · + dn−1 xn−1 + dn xn ∈ X n+1
for some x0 , . . . , xn ∈ {0, 1, . . . , d − 1} and n ≥ 0. Let
f (u) = y0 + dy1 + · · · + dn−1 yn−1
and
′
f (v) = y0′ + dy1′ + · · · + dn−1 yn−1
+ dn yn′ ,
where y0 , . . . , yn−1 , y0′ , . . . , yn′ ∈ {0, 1, . . . , d − 1}. Since u ≡ v (mod dn ), it follows that
f (u) ≡ f (v) (mod dn ). Thus
′
y0 + dy1 + · · · + dn−1 yn−1 = y0′ + dy1′ + · · · + dn−1 yn−1
By uniqueness of the d-ary expansion, we obtain yi = yi′ for i = 0, . . . , n−1. Hence f (u) is the
parent of f (v). This means that f preserves the adjacency relation so it is an endomorphism
of X ∗ .
Now we consider two different polynomials g, h ∈ Z[x]. We can find an integer k such that
g(k) 6= h(k). Let N be the smallest positive integer such that g(k) 6≡ h(k) (mod dN ). Then
the actions of g and h on the N th level of X ∗ are different.
Our next goal is to completely describe endomorphisms induced by polynomials by explicitly describing their sections at all the vertices of X ∗ . Before we proceed to the next theorem,
we need to introduce some basic notation to make our expressions less cumbersome.
Notation. Given two integers d > 0 and a, we can use the division algorithm to find two
unique integers q and r such that a = dq + r where r ∈ {0, 1, . . . , d − 1}. We will adopt the
notation Qd (a) = q and Rd (a) = r i.e., Qd (a) = a div d and Rd (a) is the remainder of a when
divided by d. So for each integer a we always have
(3)
a = Rd (a) + dQd (a).
Theorem 3.2. Given a polynomial f (x) = a0 +a1 x+· · ·+at xt ∈ Z[x] inducing an endomorphism of X ∗ , the image of a vertex x0 ∈ X under the induced endomorphism is Rd (f (x0 ))
and the section of f at x0 is again a polynomial given by the equation:
(4)
f |x0 (x) = Qd (f (x0 )) +
t
X
f (i) (x0 )
i=1
7
i!
di−1 xi
Proof. We pick a vertex u = x0 + dx1 + d2 x2 + · · · + dn xn which has x0 as a prefix. We can
write u as u = x0 + d · v, where v = x1 + dx2 + · · · + dn−1 xn corresponds to the suffix of
the word u as in equation (2). Using the Taylor expansion of f about x0 and the fact that
according to equality (3), f (x0 ) = Rd (f (x0 )) + dQd (f (x0 )) we obtain
!
t
t
X
X
f (i) (x0 ) i i
f (i) (x0 ) i−1 i
f (u) = f (x0 + d · v) =
d v = Rd (f (x0 )) + d · Qd (f (x0 )) +
d v
i!
i!
i=0
i=1
Therefore,
f (x0 ) = Rd (f (x0 + d · v)) = Rd (f (x0 )),
where with a slight abuse of notation in the left hand side x0 and f (x0 ) denote vertices of
the first level of X ∗ , and in the right hand side x0 is an element of Z/(dZ). And finally,
f |x0 (v) = Qd (f (x0 + d · v)) = Qd (f (x0 )) +
t
X
f (i) (x0 )
i=1
i!
di−1 v i
Remark 3.3. Equation (4) immediately implies that all the sections of f are polynomials
of the same degree as the degree of f . Also since f (t) (x) = t!at , all the sections of f at the
same level n have the same leading coefficient (dt−1 )n at .
Example 2. The sections of the polynomial f (x) = 1 + 3x + 2x2 at the vertices of the first
three levels of the tree are depicted in Figure 4.
Proposition 3.4. A polynomial f acting on X ∗ has finitely many distinct sections if and
only if it is linear.
Proof. First we show that a linear polynomial f (x) = ax + b acting on X ∗ has finitely many
sections. From equation (4), we see that the section of f at a vertex x0 ∈ X is given by
f |x0 (x) = ax + Qd (b + x0 a). Hence all the sections of f are linear polynomials with the same
leading coefficient a. So the number of sections of f is bounded up by the number of distinct
constant terms of these sections. Since the constant term can be written as Qd (b+a+· · ·+a)
where we have exactly x0 ≤ d−1 summands equal to a, it is enough to notice that for any collection of integers n1 , n2 , . . . , nd we have |Qd (n1 + n2 + · · · + nd )| ≤ max{|n1 |, |n2 |, . . . , |nd |}.
Therefore, any section of f has the form ax + c where |c| ≤ max{|a|, |b|}.
The fact that any nonlinear polynomial f (x) = a0 + a1 x + · · · + at xt acting on X ∗ has
infinitely many sections follows immediately from Remark 3.3 and Proposition 3.1.
Now we turn our attention to the permutational polynomials inducing automorphisms of
X ∗ . First of all, we recall the definition.
Definition 4. A polynomial f (x) ∈ Z[x] is said to be d-permutational (or, simply, permutational if d is clear from the context) if for each n ≥ 1 the mapping fn : Z/(dn Z) → Z/(dn Z)
induced by the evaluation homomorphism is a permutation.
According to Proposition 3.1 each permutational polynomial induces an automorphism of
the d-ary tree X ∗ . The following simple remark follows immediately from the definition.
Remark 3.5. The sections of a permutational polynomial acting on X ∗ are again permutational.
8
1 + 3x + 2x2
ag replacements
3 + 7x + 4x2
3x + 4x2
3 + 11x + 8x2
3x + 8x2
3x + 16x2
1 + 11x + 16x2
5 + 19x + 16x2
1 + 7x + 8x2
7x + 16x2
11 + 27x + 16x2
7 + 15x + 8x2
3 + 15x + 16x2
8 + 23x + 16x2
15 + 31x + 16x2
Figure 4. Sections of an automorphism of {0, 1}∗ induced by a polynomial
f (x) = 1 + 3x + 2x2
For each d ≥ 2, the set of all linear d-permutational polynomials obviously forms a group
under the operation of composition. However, the set of all d-permutational polynomials
does not form a group as it is not closed under taking inverses. In fact, it has the structure
of a cancellative monoid as shown in the next proposition.
Proposition 3.6. For each d ≥ 2, the set of all d-permutational polynomials forms a cancellative monoid under the operation of composition.
Proof. It is clear that the composition of two d-permutational polynomials is again a polynomial and that it induces an automorphism of the rooted d-ary tree X ∗ . The polynomial
id(x) = x plays the role of the identity as it clearly induces the identity automorphisms of
the tree. so we have the structure of a monoid.
The inverse of an automorphism of X ∗ induced by a permutational polynomial always
exists but it cannot always be induced by a polynomial. Although this is the case for linear
permutational polynomials, the inverse of a nonlinear permutational polynomial cannot be
induced by a polynomial. If this were true, we would have two polynomials not both linear
acting on the ring Zd (the boundary of X ∗ ) with a trivial composition which is impossible.
Still the existence of the inverse makes cancellation legitimate.
Remark 3.7. If we consider the action of a d-permutational polynomial f on some specific
level n, identified with Z/(dn Z), the inverse of f on Z/(dn Z) can always be induced by a
permutational polynomial as shown in [MŠG10].
9
For the rest of the paper, we will consider only permutational polynomials acting on the
rooted binary tree {0, 1}∗. The next theorem, introduced by Rivest in [Riv01], determines
the conditions under which a polynomial f (x) ∈ Z[x] induces a permutation of Z/(2n Z) for
each n and hence an automorphism of {0, 1}∗ (i.e., is a 2-permutational polynomial).
Theorem 3.8 ([Riv01, MŠG10]). A polynomial f (x) = a0 + a1 x + · · · + at xt ∈ Z[x] induces
a permutation of Z/(2n Z) for each n if and only if it satisfies the following conditions:
(i) a1 ≡ 1 (mod 2)
(ii) a2 + a4 + a6 + · · · ≡ 0 (mod 2)
(iii) a3 + a5 + a7 + · · · ≡ 0 (mod 2)
The theorem does not put any restriction on the constant term a0 of the polynomial f .
Assuming that f satisfies the conditions of the theorem i.e., it induces a permutation of
Z/(2n Z) and that a0 is even (odd), all the evens of Z/(2n Z) are mapped to evens (odds) and
all the odds of Z/(2n Z) are mapped to odds (evens). A special case of the theorem is when
f (x) = ax + b i.e., when f is linear. It defines a permutation of Z/(2n Z) if and only if a is
odd.
From now on, a permutational polynomial will always mean a 2-permutational polynomial.
Also we will drop the 2 subscript in the notation of the function Q2 and write Q(a) = a div 2.
According to equation (4), the sections of a permutational polynomial f (x) = a0 + a1 x +
· · · + at xt acting on {0, 1}∗ at the vertices 0 and 1 are given by the two equations:
(5)
f |0 (x) = Q(a0 ) + a1 x + 2a2 x2 + 22 a3 x3 · · · + 2t−1 at xt
and
(6) f |1(x) = Q(a0 + a1 + · · · + at ) + (a1 + 2a2 + · · · + tat )x
t
X
f (i) (1) i−3 i
+ 2 · 1a2 + 3 · 2a3 + · · · + t(t − 1)at x2 + 4
2 x
i!
i=3
In the case of a linear polynomial f (x) = ax + b, we have f |0 (x) = ax + Q(b) and f |1 (x) =
ax + Q(a + b). For example, the adding machine τ , which was introduced in the preliminary
section, can be represented by a linear permutational polynomial τ (x) = x + 1. The sections
of τ at 0 and 1 are respectively x and x + 1.
According to Theorem 3.8, the group of all linear 2-permutational polynomials is isomorphic to the group
a b
| a, b ∈ Z, a odd .
0 1
A generating set of this group is {p, q−1 , q3 , q5 , q7 , . . .}, where p(x) = x + 1 and qm (x) = mx
for m = −1, 3, 5, 7, . . .. All the generators have infinite orders except for q−1 which is an
involution. It was shown in [BŠ06] that for each m = −1, 3, 5, 7, . . ., the subgroup generated
by p and qm is the Baumslag-Solitar group BS(1, m) = hp, qm | qm pm = pqm i, where we
adopt the convention that when the expression f g is used to denote the composition of
two functions it means the function f acts first. In the same paper, Bartholdi and Šunı́c
considered sections of some linear polynomials acting on {0, 1}∗.
We end this section by introducing one more notation and stating a simple lemma that
will be used in Section 4 to simplify the proof of the main result.
10
Notation. Suppose f (x) = a0 + a1 x + · · ·+ at xt ∈ Z[x] is a permutational polynomial. Then
it uniquely defines the following integers kf , mf and nf such that
(7)
a1 = 2kf + 1,
a2 + a4 + · · · = 2mf ,
a3 + a5 + · · · = 2nf .
The following lemma, in particular, shows that not every permutational polynomial can
be a section of another permutational polynomial.
Lemma 3.9. Let f (x) = a0 + a1 x + · · · + at xt be a permutational polynomial acting on
the rooted binary tree {0, 1}∗. Then for each non-root vertex v ∈ {0, 1}∗, the permutational
polynomial f |v (x) = b0 + b1 x + · · · + bt xt corresponding to the section of f at v, satisfies
nf |v ≡ 0 (mod 2) and b2 ≡ 0 (mod 2).
Proof. Clearly, it is enough to check the conditions only for vertices of the first level. But this
trivially follows from equations (5) and (6). We only have to notice that all the coefficients
at xi for i ≥ 3 in both f |0 and f |1 are divisible by 4 and that the coefficients at x2 in both
f |0 and f |1 are divisible by 2.
4. Level Transitivity of Permutational Polynomials Acting on a Rooted
Binary Tree
We start this section by presenting a couple of basic number theoretic facts that will be
used many times in the proof of our main theorem. We introduce a bunch of lemmas without
proofs. The proofs are straight forward so we leave them as simple exercises. The first lemma
gives some properties of the function Q(x) which was defined in the last section. All the
other lemmas follow from the fact that for any positive integer n we have ax ≡ x (mod n)
if a and x are integers and a ≡ 1 (mod n).
Lemma 4.1. For any three integers x, m and n, we have:
(i) Q(x + 2n) = Q(x) + n
(ii) Q(x) + Q(x + 1) = x
(iii) Q(x) + Q(x + 1 + 2n) = x + n
(iv) Q(x + 4n) ≡ Q(x) (mod 2)
(v) Q(x) + Q(x + 1 + 4n) ≡ x (mod 2)
(vi) If n ≡ m (mod 4), then Q(x + n) ≡ Q(x + m) (mod 2)
Lemma 4.2. Let a1 , a2 , . . . , an be a collection of odd numbers. Then for any sequence of
integers x1 , x2 , . . . , xn , we have a1 x1 + a2 x2 + · · · + an xn ≡ x1 + x2 + · · · + xn (mod 2) i.e.,
a1 x1 + a2 x2 + · · · + an xn has the same parity as x1 + x2 + · · · + xn .
Lemma 4.3. Let a1 , a2 , . . . , an and x1 , x2 , . . . , xn be integers such that ai ≡ 1 (mod 4) for
i = 1, 2, . . . , n. Then a1 x1 + a2 x2 + · · · + an xn ≡ x1 + x2 + · · · + xn (mod 4).
Lemma 4.4. Let a1 , a2 , . . . , an be a collection of odd numbers and x1 , x2 , . . . , xn be any
sequence of integers. Then a21 x1 + a22 x2 + · · · + a2n xn ≡ x1 + x2 + · · · + xn (mod 4).
Now it is time to introduce our main theorem that determines the conditions under which a
permutational polynomial acts level transitively on the rooted binary tree {0, 1}∗ . Obviously
11
a permutational polynomial acts nontrivially on the first level of the tree if and only if its
constant term is odd. According to Proposition 2.1, a permutational polynomial f acts level
transitively on {0, 1}∗ if and only if in each level of the tree, the number of sections of f with
odd constant terms is odd, or equivalently, the sum of these constant terms is odd. The next
proposition determines the conditions that a linear permutational polynomial has to meet in
order to act level transitively on {0, 1}∗. The general result is given by Theorem 4.6 which
we prove by induction on level. The proof of the linear case is provided first so that the idea
of the induction becomes clear and then the general case is considered. Actually the proof
given here is essentially different from the proofs of similar results introduced in [Ana06] and
in [Lar02].
Proposition 4.5. Let f (x) = ax + b be a permutational polynomial acting on the rooted
binary tree {0, 1}∗. Then this action is level transitive if and only if the following conditions
hold:
(i) b ≡ 1 (mod 2)
(ii) a ≡ 1 (mod 4)
Proof. We first show that the conditions are necessary. If condition (i) is not satisfied, then
f does not act transitively on the first level. So we assume that condition (i) is satisfied but
condition (ii) is not. We can also use the fact that a is odd since f is permutational. Thus
we can write b = 2s + 1 and a = 2r + 1, where s and r are integers and r is odd. The sections
of f at 0 and 1 are respectively ax + Q(b) = ax + s and ax + Q(a + b) = ax + r + s + 1.
The sum of the two constant terms is 2s + r + 1 which is even. Therefore, f does not act
transitively on the second level.
To prove sufficiency of the conditions, we use induction on level. First we can write
a = 4k + 1 for some integer k which follows from condition (ii). Since b is odd, f acts
transitively on the first level. Now assume that the sections of f at some level n ≥ 0 are
respectively ax + α1 , . . . , ax + αm , where m = 2n and suppose that α1 + · · · + αm ≡ 1
(mod 2) which serves as the induction hypothesis. The sections of f at level n + 1 are
respectively ax+σ1 , ax+σ2 , . . . , ax+σ2m−1 , ax+σ2m , where σ1 = Q(α1 ), σ2 = Q(α1 +4k +1),
. . . , σ2m−1 = Q(αm ), σ2m = Q(αm + 4k + 1). Hence the sum of constant terms of the
sections of level n + 1 is (σ1 + σ2 ) + · · · + (σ2m−1 + σ2m ) = (Q(α1 ) + Q(α1 + 4k + 1)) + · · · +
(Q(αm ) + Q(αm + 4k + 1)) ≡ α1 + · · · + αm ≡ 1 (mod 2), where we have used Lemma 4.1
part (v). The proof is now complete.
The adding machine τ (x) = x + 1 as well as all its odd powers τ r (x) = x + r, where r
is odd, satisfy the conditions of the last proposition. So they act level transitively on the
rooted binary tree {0, 1}∗ .
Theorem 4.6. Let f (x) = a0 + a1 x + · · · + at xt be a permutational polynomial acting on
the rooted binary tree {0, 1}∗. Then this action is level transitive if and only if the following
conditions hold:
(i) a0 ≡ 1 (mod 2)
(ii) 2a2 ≡ a3 + a5 + · · · (mod 4)
(iii) a2 + a1 − 1 ≡ a4 + a6 + · · · (mod 4)
Proof. To show that the given conditions are necessary, we first notice that if condition
(i) is not satisfied then f does not act transitively on the first level. So we assume that
12
condition (i) is satisfied and hence we can write a0 = 2s + 1 for some integer s. Since f is
permutational, it satisfies the three conditions of Theorem 3.8. These three conditions could
be written respectively as a1 = 2k + 1, a2 + a4 + · · · = 2m and a3 + a5 + · · · = 2n for some
integers k, m and n. By adding a2 to both sides in condition (iii) and doing some simple
algebraic manipulation, conditions (ii) and (iii) can be rewritten as:
(ii′ ) a2 ≡ n (mod 2)
(iii′ ) k + a2 ≡ m (mod 2)
Equations (5) and (6) tell us that the constant terms of the sections of f at 0 and 1 are
respectively Q(a0 ) and Q(a0 + a1 + · · · + at ). Their sum is Q(a0 ) + Q(a0 + a1 + · · · + at ) =
Q(a0 ) + Q(a0 + 2k + 1 + 2m + 2n) = a0 + k + m + n = 2s + (k + n) + m + 1, where we have
used Lemma 4.1 part (iii). For f to act transitively on the second level, we must have
(8)
(k + n) ≡ m
(mod 2).
If exactly one of the two conditions (ii′ ) and (iii′ ) is satisfied, then (k + n) 6≡ m (mod 2).
However, if the two conditions are both not satisfied, we have (k + n) ≡ m (mod 2) so f
does act transitively on the second level. We will show that in the latter case, f does not
act transitively on the third level by proving that the sum of constant terms of the sections
of f at the second level is even.
From equations (5) and (6), we can infer that the constant terms of the sections of f at
the vertices 00, 01, 10 and 11 are respectively:
• Q(Q(a0 )) = Q(s)
• Q(s + a1 + 2a2 + · · · + 2t−1 at ) = Q(s + 2k + 1 + 2a2 + · · · + 2t−1 at )
• Q(Q(a0 + a1 + · · · + at )) = Q(Q(2s + 1 + 2k + 1 + 2m + 2n)) = Q(1 + s + k + m + n)
• Q((1 + s + k + m + n) + (a1 + 2a2 + · · · + tat ) + (2 · 1a2 + 3 · 2a3 + · · · + t(t − 1)at ) +
P
(i)
4 ti=3 f i!(1) 2i−3 )
The sum of these constant terms modulo 2 is
Q(s)+Q(s+2k+1+2a2 )+Q(1+s+k+m+n)+Q((1+s+k+m+n)+a1 +22 a2 +32 a3 +· · ·+t2 at ),
where we have used Lemma 4.1 part (iv). Using part (iii) of the same lemma and reusing
part (iv) as well, the sum modulo 2 simplifies to
s+k+a2 +Q(1+s+k+m+n)+Q((1+s+k+m+n)+1+2k+(22a2 +42 a4 +· · · )+(32a3 +52 a5 +· · · ))
≡ s + k + a2 + Q(1 + s + k + m + n) + Q((1 + s + k + m + n) + 1 + 2k + (32 a3 + 52 a5 + · · · ))
≡ s + k + a2 + Q(1 + s + k + m + n) + Q((1 + s + k + m + n) + 1 + 2k + (a3 + a5 + · · · ))
≡ s + k + a2 + Q(1 + s + k + m + n) + Q((1 + s + k + m + n) + 1 + 2k + 2n) (mod 2)
Where we have used Lemma 4.4 to obtain 32 a3 + 52 a5 + · · · ≡ a3 + a5 + · · · . Again we apply
part (iii) of Lemma 4.1 to have the sum modulo 2 as
s + k + a2 + (1 + s + k + m + n) + k + n ≡ 2s + 2k + (k + n) + m + (n + a2 + 1)
≡ ((k + n) + m) + (n + a2 + 1) ≡ (k + n) + m ≡ 0 (mod 2)
where we have used the fact that the two conditions (ii′ ) and (iii′ ) are both not satisfied.
Now we prove sufficiency of the conditions. Again we use the fact that f is permutational
to write a1 = 2k + 1, a2 + a4 + · · · = 2m and a3 + a5 + · · · = 2n for some integers k, m and n.
We can easily infer that 2a2 ≡ 2n (mod 4) and 2n + 2k ≡ 2m (mod 4). The last equality is
equivalent to equality 8. We will show that f acts transitively on the first and second levels
13
and then use induction for lower levels. Condition (i) guarantees transitivity on the first
level.
We deduced in the proof of necessity that the transitivity on the second level is equivalent
to equality (8), which is automatically satisfied in our case as shown above.
Before we formulate the induction hypothesis, we recall that at each vertex v of {0, 1}∗ the
(f | )
(f | )
(f | )
section f |v is induced by a permutational polynomial f |v (x) = a0 v +a1 v x+· · ·+at v xt , so
(f | )
(f | )
(f | )
it defines the integers kf |v , mf |v , and nf |v such that a1 v = 2kf |v +1, a2 v +a4 v +· · · = 2mf |v
(f | )
(f | )
and a3 v + a5 v + · · · = 2nf |v . Also if v is a non-root vertex, we have nf |v ≡ 0 (mod 2) and
(f | )
a2 v ≡ 0 (mod 2) by Lemma 3.9.
Let us assume that the sections of f at some level l ≥ 1 of the tree are ϕ1 , . . . , ϕr , where
(ϕ )
(ϕ )
r = 2l . Suppose that a0 1 +· · ·+a0 r ≡ 1 (mod 2) which serves as our induction hypothesis.
Let ψ1 , ψ2 , . . . , ψ2r−1 , ψ2r be the sections of f at level l + 1. Then
2r
X
(ψ )
a0 i
=
r
X
(ψ
)
a0 2j−1
+
(ψ )
a0 2j
j=1
i=1
r
X
(ϕ )
(ϕ )
(ϕ )
(ϕ )
=
Q a0 j + Q a0 j + a1 j + · · · + at j
j=1
=
r
X
(ϕ )
(ϕ )
Q a0 j + Q a0 j + 2kϕj + 1 + 2mϕj + 2nϕj
j=1
r
X
(ϕ )
=
a0 j + kϕj + mϕj + nϕj
j=1
=
r
X
(ϕj )
a0
+
r
X
k ϕj +
j=1
j=1
≡1+
r
X
j=1
k ϕj +
r
X
j=1
r
X
mϕj
mϕj +
r
X
nϕj
j=1
(mod 2)
j=1
by induction hypothesis, where we have used Lemma 4.1 part (iii), then rearranged the
terms, and finally exploited Lemma 3.9 at the last transition. We claim that (kϕi + kϕi+1 ) +
(mϕi + mϕi+1 ) ≡ 0 (mod 2) for i = 1, 3, . . . , r − 1.
We will show that for every permutational polynomial g(x) = b0 + b1 x + b2 x2 + · · · + bt xt
satisfying the two conditions ng ≡ 0 (mod 2) and b2 ≡ 0 (mod 2), we have (kg|0 + kg|1 ) +
(mg|0 + mg|1 ) ≡ 0 (mod 2). This will be enough to prove the claim.
(g| )
(g| )
First we have a1 0 + a1 1 = 2kg|0 + 1 + 2kg|1 + 1, thus from equations (5) and (6) we have
(g|0 )
a1
(g|1 )
+ a1
2
b1 + (b1 + 2b2 + · · · + tbt )
−1=
2
3b3 + 5b5 + · · ·
b1 + b2 + (2b4 + 3b6 + 4b8 + · · · ) +
−1=
2
3b3 + 5b5 + · · ·
−1
2kg + 1 + b2 + (2b4 + 4b8 + · · · ) + (3b6 + 5b10 + · · · ) +
2
kg|0 + kg|1 =
−1=
14
Using Lemma 4.2, we can write the sum modulo 2 as
3b3 + 5b5 + · · ·
kg|0 + kg|1 ≡ (b6 + b10 + · · · ) +
(mod 2)
2
Again from equations (5) and (6), we can write mg|0 + mg|1 modulo 2 as
mg|0 + mg|1 ≡ b2 + (b2 + 3 · 1b3 + 2 · 3b4 + 5 · 2b5 + 3 · 5b6 + · · · )
≡ 2b2 + (2 · 3b4 + 5 · 2b5 + · · · ) + (3 · 1b3 + 3 · 5b6 + · · · )
≡ b3 + b6 + b7 + b10 + b11 + b14 + b15 + · · · (mod 2)
The last equivalence comes from applying Lemma 4.2. Therefore the sum
(kg|0 + kg|1 ) + (mg|0 + mg|1 )
3b3 + 5b5 + · · ·
2
5(b3 + b5 ) + 9(b7 + b9 ) + · · ·
(mod 2)
≡
2
From Lemma 4.3, we have 5(b3 +b5 )+9(b7 +b9 )+· · · ≡ (b3 +b5 )+(b7 +b9 )+· · · ≡ 2n (mod 4)
5(b3 + b5 ) + 9(b7 + b9 ) + · · ·
≡ n ≡ 0 (mod 2). The proof is now complete.
and thus
2
If a mapping f acting on the rooted binary tree {0, 1}∗ is level transitive, then the orbit
of every element in the boundary of the tree Z2 under f is dense in Z2 . Thus the dynamical
system (Z2 , f ) is minimal. Also Anashin proved in [Ana06] that a polynomial f ∈ Z[x] is
ergodic with respect to the normalized Haar measure on Z2 if and only if f is transitive
mod 2n for every positive integer n. Combining Theorem 3.8 and Theorem 4.6, we thus
obtain a new and more elementary proof of the result of Larin [Lar02].
≡ (b3 + b7 + b11 + b15 + · · · ) +
Theorem 4.7 ([Lar02]). Let f (x) = a0 + a1 x + · · · + at xt ∈ Z2 [x]. Then the dynamical
system (Z2 , f ) is minimal, or equivalently, f is ergodic with respect to the normalized Haar
measure of Z2 if and only if the following conditions are satisfied:
(i) a0 ≡ 1 (mod 2)
(ii) a1 ≡ 1 (mod 2)
(iii) 2a2 ≡ a3 + a5 + · · · (mod 4)
(iv) a2 + a1 − 1 ≡ a4 + a6 + · · · (mod 4)
References
[Ale72]
[Ana98]
[Ana06]
[Ana12]
[BGK+ 08]
S. V. Alešin. Finite automata and the Burnside problem for periodic groups. Mat. Zametki,
11:319–328, 1972.
V. S. Anashin. Uniformly distributed sequences in computer algebra or how to construct program generators of random numbers. J. Math. Sci. (New York), 89(4):1355–1390, 1998. Computing mathematics and cybernetics, 2.
Vladimir Anashin. Ergodic transformations in the space of p-adic integers. In p-adic mathematical physics, volume 826 of AIP Conf. Proc., pages 3–24. Amer. Inst. Phys., Melville, NY,
2006.
V. Anashin. Automata finiteness criterion in terms of van der Put series of automata functions.
p-Adic Numbers Ultrametric Anal. Appl., 4(2):151–160, 2012.
I. Bondarenko, R. Grigorchuk, R. Kravchenko, Y. Muntyan, V. Nekrashevych, D. Savchuk, and
Z. Šunić. Classification of groups generated by 3-state automata over 2-letter alphabet. Algebra
Discrete Math., (1):1–163, 2008. (available at http://arxiv.org/abs/0803.3555 ).
15
[BOERT96] Hyman Bass, Maria Victoria Otero-Espinar, Daniel Rockmore, and Charles Tresser. Cyclic
renormalization and automorphism groups of rooted trees, volume 1621 of Lecture Notes in
Mathematics. Springer-Verlag, Berlin, 1996.
[BŠ06]
Laurent I. Bartholdi and Zoran Šuniḱ. Some solvable automaton groups. In Topological and
Asymptotic Aspects of Group Theory, volume 394 of Contemp. Math., pages 11–29. Amer.
Math. Soc., Providence, RI, 2006.
[Fan15]
Ai-Hua Fan. p-adic polynomial dynamics, 2015.
https://cantorsalta2015.sciencesconf.org/conference/cantorsalta2015/Fan.pdf.
[GLSŻ00] Rostislav I. Grigorchuk, Peter Linnell, Thomas Schick, and Andrzej Żuk. On a question of
Atiyah. C. R. Acad. Sci. Paris Sér. I Math., 331(9):663–668, 2000.
[GNS00]
R. I. Grigorchuk, V. V. Nekrashevich, and V. I. Sushchanskiı̆. Automata, dynamical systems,
and groups. Tr. Mat. Inst. Steklova, 231(Din. Sist., Avtom. i Beskon. Gruppy):134–214, 2000.
[Gri80]
R. I. Grigorchuk. On Burnside’s problem on periodic groups. Funktsional. Anal. i Prilozhen.,
14(1):53–54, 1980.
[Gri83]
R. I. Grigorchuk. On the Milnor problem of group growth. Dokl. Akad. Nauk SSSR, 271(1):30–
33, 1983.
[GS83]
Narain Gupta and Said Sidki. On the Burnside problem for periodic groups. Math. Z.,
182(3):385–388, 1983.
[GŠ06]
Rostislav Grigorchuk and Zoran Šuniḱ. Asymptotic aspects of Schreier graphs and Hanoi Towers
groups. C. R. Math. Acad. Sci. Paris, 342(8):545–550, 2006.
[GS16]
Rostislav Grigorchuk and Dmytro Savchuk. Ergodic decomposition of group actions on rooted
trees. Tr. Mat. Inst. Steklova, 292(Algebra, Geometriya i Teoriya Chisel):100–117, 2016.
[Knu81]
Donald E. Knuth. The art of computer programming. Vol. 2. Addison-Wesley Publishing Co.,
Reading, Mass., second edition, 1981. Seminumerical algorithms, Addison-Wesley Series in Computer Science and Information Processing.
[Lar02]
M. V. Larin. Transitive polynomial transformations of residue rings. Diskret. Mat., 14(2):20–32,
2002.
[LN83]
Rudolf Lidl and Harald Niederreiter. Finite fields, volume 20 of Encyclopedia of Mathematics
and its Applications. Addison-Wesley Publishing Company, Advanced Book Program, Reading,
MA, 1983. With a foreword by P. M. Cohn.
[MS15]
Alexei Miasnikov and Dmytro Savchuk. An example of an automatic graph of intermediate
growth. Ann. Pure Appl. Logic, 166(10):1037–1048, 2015.
[MŠG10]
Smile Markovski, Zoran Šunić, and Danilo Gligoroski. Polynomial functions on the units of Z2n .
Quasigroups Related Systems, 18(1):59–82, 2010.
[Nek05]
Volodymyr Nekrashevych. Self-similar groups, volume 117 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2005.
[Riv01]
Ronald L. Rivest. Permutation polynomials modulo 2w . Finite Fields Appl., 7(2):287–292, 2001.
[RRSY]
Ronald L. Rivest, M. J. B. Robshaw, R. Sidney, and Y. L. Yin. The RC6 block cipher. Posted on
the RC6 site of RSA Laboratories. Slides from NIST AES1 (1998-08-21) and AES3 (2000-04-14)
conferences.
[Sus79]
V. I. Sushchansky. Periodic permutation p-groups and the unrestricted Burnside problem. DAN
SSSR., 247(3):557–562, 1979. (in Russian).
Department of Mathematics and Statistics, University of South Florida, 4202 E Fowler
Ave, Tampa, FL 33620-5700, [email protected]
Department of Mathematics and Statistics, University of South Florida, 4202 E Fowler
Ave, Tampa, FL 33620-5700, [email protected]
16
| 4math.GR
|
Fully Decentralized Policies for Multi-Agent Systems:
An Information Theoretic Approach
Roel Dobbe∗, David Fridovich-Keil∗and Claire Tomlin
†
August 1, 2017
arXiv:1707.06334v2 [cs.SY] 29 Jul 2017
Abstract
1
Learning cooperative policies for multi-agent systems is often challenged by partial observability and a lack of coordination.
In some settings, the structure of a problem allows a distributed solution with limited communication. Here, we consider
a scenario where no communication is available, and instead we learn local policies for all agents that collectively mimic
the solution to a centralized multi-agent static optimization problem. Our main contribution is an information theoretic
framework based on rate distortion theory which facilitates analysis of how well the resulting fully decentralized policies are
able to reconstruct the optimal solution. Moreover, this framework provides a natural extension that addresses which nodes
an agent should communicate with to improve the performance of its individual policy.
Introduction
Finding optimal decentralized policies for multiple agents is often a hard problem hampered by partial observability and a
lack of coordination between agents. The distributed multi-agent problem has been approached from a variety of angles,
including distributed optimization [Boyd et al., 2011], game theory [Aumann and Dreze, 1974] and decentralized or networked
partially observable Markov decision processes (POMDPs) [Oliehoek and Amato, 2016, Goldman and Zilberstein, 2004, Nair
et al., 2005]. In this paper, we analyze a different approach consisting of a simple learning scheme to design fully decentralized
policies for all agents that collectively mimic the solution to a common optimization problem, while having no access to a
global reward signal and either no or restricted access to other agents’ local state. This algorithm is a generalization of that
proposed in our prior work [Sondermeijer et al., 2016] related to decentralized optimal power flow (OPF). Indeed, the success
of regression-based decentralization in the OPF domain motivated us to understand when and how well the method works in
a more general decentralized optimal control setting.
The key contribution of this work is to view decentralization as a compression problem, and then apply classical results from
information theory to analyze performance limits. More specifically, we treat the ith agent’s optimal action in the centralized
problem as a random variable u∗i , and model its conditional dependence on the global state variables x = (x1 , . . . , xn ), i.e.
p(u∗i |x), which we assume to be stationary in time. We now restrict each agent i to observe only the ith state variable xi .
Rather than solving this decentralized problem directly, we train each agent to replicate what it would have done with full
information in the centralized case. That is, the vector of state variables x is compressed, and the ith agent must decompress
xi to compute some estimate ûi ≈ u∗i . In our approach, each agent learns a parameterized Markov control policy ûi = π̂i (xi )
via regression. The π̂i are learned from a data set containing local states xi taken from historical measurements of system
state x and corresponding optimal actions u∗i computed by solving an offline centralized optimization problem for each x.
In this context, we analyze the fundamental limits of compression. In particular, we are interested in unraveling the
relationship between the dependence structure of u∗i and x and the corresponding ability of an agent with partial information
to approximate the optimal solution, i.e. the difference – or distortion – between decentralized action ûi = π̂i (xi ) and u∗i .
This type of relationship is well studied within the information theory literature as an instance of rate distortion theory [Cover
and Thomas, 2012, Chapter 13]. Classical results in this field provide a means of finding a lower bound on the expected
distortion as a function of the mutual information – or rate of communication – between u∗i and xi . This lower bound is
valid for each specified distortion metric, and for any arbitrary strategy of computing ûi from available data xi . Moreover,
we are able to leverage a similar result to provide a conceptually simple algorithm for choosing a communication structure –
letting the regressor π̂i depend on some other local states xj6=i – in such a way that the lower bound on expected distortion is
minimized. As such, our method generalizes [Sondermeijer et al., 2016] and provides a novel approach for the design and
analysis of regression-based decentralized optimal policies for general multi-agent systems. We demonstrate these results on
synthetic examples, and on a real example drawn from solving OPF in electrical distribution grids.
∗ Indicates
equal contribution.
Dobbe, David Fridovich-Keil and Claire Tomlin are with the Department of Electrical Engineering and Computer Sciences, University of
California, Berkeley, CA 94720, USA. [dobbe, dfk, tomlin]@eecs.berkeley.edu.
† Roel
1
2
Related Work
Decentralized control has long been studied within the system theory literature, e.g. [Lunze, 1992, Siljak, 2011]. Recently,
various decomposition based techniques have been proposed for distributed optimization based on primal or dual decomposition
methods, which all require iterative computation and some form of communication with either a central node [Boyd et al.,
2011] or neighbor-to-neighbor on a connected graph [Pu et al., 2014, Raffard et al., 2004, Sun et al., 2013]. Distributed
model predictive control (MPC) optimizes a networked system composed of subsystems over a time horizon, which can be
decentralized (no communication) if the dynamic interconnections between subsystems are weak in order to achieve closed-loop
stability as well as performance [Christofides et al., 2013]. The work of Zeilinger et al. [2013] extended this to systems
with strong coupling by employing time-varying distributed terminal set constraints, which requires neighbor-to-neighbor
communication. Another class of methods model problems in which agents try to cooperate on a common objective without
full state information as a decentralized partially observable Markov decision process (Dec-POMDP) [Oliehoek and Amato,
2016]. Nair et al. [2005] introduce networked distributed POMDPs, a variant of the Dec-POMDP inspired in part by the
pairwise interaction paradigm of distributed constraint optimization problems (DCOPs).
Although the specific algorithms in these works differ significantly from the regression-based decentralization scheme we
consider in this paper, a larger difference is in problem formulation. As described in Sec. 3, we study a static optimization
problem repeatedly solved at each time step. Much prior work, especially in optimal control (e.g. MPC) and reinforcement
learning (e.g. Dec-POMDPs), poses the problem in a dynamic setting where the goal is to minimize cost over some time
horizon. In the context of reinforcement learning (RL), the time horizon can be very long, leading to the well known tradeoff
between exploration and exploitation; this does not appear in the static case. Additionally, many existing methods for the
dynamic setting require an ongoing communication strategy between agents – though not all, e.g. [Peshkin et al., 2000]. Even
one-shot static problems such as DCOPs tend to require complex communication strategies, e.g. [Modi et al., 2005].
Although the mathematical formulation of our approach is rather different from prior work, the policies we compute
are similar in spirit to other learning and robotic techniques that have been proposed, such as behavioral cloning [Sammut,
1996] and apprenticeship learning [Abbeel and Ng, 2004], which aim to let an agent learn from examples. In addition, we
see a parallel with recent work on information-theoretic bounded rationality [Ortega et al., 2015] which seeks to formalize
decision-making with limited resources such as the time, energy, memory, and computational effort allocated for arriving at a
decision. Our work is also related to swarm robotics [Brambilla et al., 2013], as it learns simple rules aimed to design robust,
scalable and flexible collective behaviors for coordinating a large number of agents or robots.
3
General Problem Formulation
Consider a distributed multi-agent problem defined by a graph G = (N , E), with N denoting the nodes in the network with
cardinality |N | = N , and E representing the set of edges between nodes. Fig. 1a shows a prototypical graph of this sort. Each
node has a real-valued state vector xi ∈ Rαi , i ∈ N . A subset of nodes C ⊂ N , with cardinality |C| = C, are controllable
and
P
>
βi
i∈N αi = X
∈
R
hence are termed “agents.” Each of these agents has
an
action
variable
u
∈
R
,
i
∈
C.
Let
x
=
(x
,
.
.
.
,
x
)
i
i
N
P
denote the full network state vector and u ∈ R i∈C βi = U the stacked network optimization variable. Physical constraints
such as spatial coupling are captured through equality constraints g(x, u) = 0. In addition, the system is subject to inequality
constraints h(x, u) ≤ 0 that incorporate limits due to capacity, safety, robustness, etc. We are interested in minimizing a
convex scalar function fo (x, u) that encodes objectives that are to be pursued cooperatively by all agents in the network, i.e.
we want to find
u∗ = arg min
fo (x, u) ,
u
(1)
s.t.
g(x, u) = 0, h(x, u) ≤ 0.
Note that (1) is static in the sense that it does not consider the future evolution of the state x or the corresponding future
values of cost fo . We apply this static problem to sequential control tasks by repeatedly solving (1) at each time step. Note
that this simplification from an explicitly dynamic problem formulation (i.e. one in which the objective function incorporates
future costs) is purely for ease of exposition and for consistency with the OPF literature as in [Sondermeijer et al., 2016]. We
could also consider the optimal policy which solves a dynamic optimal control or RL problem and the decentralized learning
step in Sec. 3.1 would remain the same.
Since (1) is static, applying the learned decentralized policies repeatedly over time may lead to dynamical instability.
Identifying when this will and will not occur is a key challenge in verifying the regression-based decentralization method,
however it is beyond the scope of this work.
3.1
Decentralized Learning
We interpret the process of solving (1) as applying a well-defined function or stationary Markov policy π ∗ : X −→ U that
maps an input collective state x to the optimal collective control or action u∗ . We presume that this solution exists and can
be computed offline. Our objective is to learn C decentralized policies ûi = π̂i (xi ), one for each agent i ∈ C, based on T
historical measurements of the states {x[t]}Tt=1 and the offline computation of the corresponding optimal actions {u∗ [t]}Tt=1 .
2
x1
u2
x2
x3
x4
x5
x6
u5
u6
(a) Distributed multi-agent problem.
u1*
ui*
u C*
x1
xj
xN
û1
ûi
ûC
(b) Graphical model of dependency structure.
Figure 1: (a) shows a connected graph corresponding to a distributed multi-agent system. The circles denote the local state
xi of an agent, the dashed arrow denotes its action ui , and the double arrows denote the physical coupling between local state
variables. (b) shows the Markov Random Field (MRF) graphical model of the dependency structure of all variables in the
decentralized learning problem. Note that the state variables xi and the optimal actions u∗i form a fully connected undirected
network, and the local policy ûi only depends on the local state xi .
Local training sets
{x2 [t], u⇤2 [t]}Tt=1
Multi-Agent
System
Centralized
Optimization
2
Data gathering
Decentralize
Decentralize
Decentralized
dLearning
dLearning
Learning
Optimal data
3 2
3 2 x⇤ 3 2 ⇤ 3
x1
1
u2
u2
6 .. 7 4
6 . 7
4 . 5 , u5 5 4 .. 5 , 4 u⇤5 5
u6
u⇤6
x6
x⇤6
approximate
Local policies
û2 = ⇡
ˆ2 (x2 )
û5 = ⇡ˆ5 (x5 )
û6 = ⇡
ˆ6 (x6 )
Figure 2: A flow diagram explaining the key steps of the decentralized regression method, depicted for the example system in
Fig. 1a. We first collect data from a multi-agent system, and then solve the centralized optimization problem using all the
data. The data is then split into smaller training and test sets for all agents to develop individual decentralized policies π̂i (xi )
that approximate the optimal solution of the centralized problem. These policies are then implemented in the multi-agent
system to collectively achieve a common global behavior.
Although each policy π̂i individually aims to approximate u∗i based on local state xi , we are able to reason about how well
their collective action can approximate π ∗ . Figure 2 summarizes the decentralized learning setup.
More formally, we describe the dependency structure of the individual policies π̂i : Rαi −→ Rβi with a Markov Random
Field (MRF) graphical model, as shown in Fig. 1b. The ûi are only allowed to depend on local state xi while the u∗i may
depend on the full state x. With this model, we can determine how information is distributed among different variables
and what information-theoretic constraints the policies {π̂i }i∈C are subject to when collectively trying to reconstruct the
centralized policy π ∗ . Note that although we may refer to π ∗ as globally optimal, this is not actually required for us to reason
about how closely the π̂i approximate π ∗ . That is, our analysis holds even if (1) is solved using approximate methods. In a
dynamical reformulation of (1), for example, π ∗ could be generated using techniques from deep RL.
3.2
A Rate-Distortion Framework
We approach the problem of how well the decentralized policies π̂i can perform in theory from the perspective of rate distortion.
Rate distortion theory is a sub-field of information theory which provides a framework for understanding and computing the
minimal distortion incurred by any given compression scheme. In a rate distortion context, we can interpret the fact that the
output of each individual policy π̂i depends only on the local state xi as a compression of the full state x. For a detailed
overview, see [Cover and Thomas, 2012, Chapter 10]. We formulate the following variant of the the classical rate distortion
problem
D∗ = min∗
p(û|u )
s.t.
E [d(û, u∗ )] ,
I(ûi ; u∗j ) ≤ I(xi ; u∗j ) , γij ,
I(ûi ; ûj ) ≤ I(xi ; xj ) , δij , ∀i, j ∈ C ,
3
(2)
where I(·, ·) denotes mutual information and d(·, ·) an arbitrary non-negative distortion measure. As usual, the minimum
distortion between random variable u∗ and its reconstruction û may be found by minimizing over conditional distributions
p(û|u∗ ).
The novelty in (2) lies in the structure of the constraints. Typically, D∗ is written as a function D(R), where R is the
maximum rate or mutual information I(û; u∗ ). From Fig. 1b however, we know that pairs of reconstructed and optimal
actions cannot share more information than is contained in the intermediate nodes in the graphical model, e.g. û1 and u∗1
cannot share more information than x1 and u∗1 . This is a simple consequence of the data processing inequality [Cover and
Thomas, 2012, Thm. 2.8.1]. Similarly, the reconstructed optimal actions at two different nodes cannot be more closely related
than the measurements xi ’s from which they are computed. The resulting constraints are fixed by the joint distribution of the
state x and the optimal actions u∗ . That is, they are fully determined by the structure of the optimization problem (1) that
we wish to solve.
We emphasize that we have made virtually no assumptions about the distortion function. For the remainder of this paper,
we will measure distortion as the deviation between ûi and u∗i . However, we could also define it to be the suboptimality gap
fo (x, û) − fo (x, u∗ ), which may be much more complicated to compute. This definition could allow us to reason explicitly
about the cost of decentralization, and it could address the valid concern that the optimal decentralized policy may bear no
resemblance to π ∗ . We leave further investigation for future work.
3.3
Example: Squared Error, Jointly Gaussian
To provide more intuition into the rate distortion framework, we consider an idealized example in which the xi , ui ∈ R1 .
Let d(û, u∗ ) = kû − u∗ k22 be the squared error distortion measure, and assume the state x and optimal actions u∗ to be
jointly Gaussian. These assumptions allow us to derive an explicit formula for the optimal distortion D∗ and corresponding
regression policies π̂i . We begin by stating an identity for two jointly Gaussian X, Y ∈ R with correlation ρ: I(X; Y ) ≤
γ ⇐⇒ ρ2 ≤ 1 − e−2γ , which follows immediately from the definition of mutual information and the formula for the entropy
of a Gaussian random variable. Taking ρûi ,u∗i to be the correlation between ûi and u∗i , σû2 i and σu2 ∗i to be the variances of ûi
and u∗i respectively, and assuming that u∗i and ûi are of equal mean (unbiased policies π̂i ), we can show that the minimum
distortion attainable is
D∗ = min∗ E ku∗ − ûk22 : ρ2ûi ,u∗i ≤ 1 − e−2γii = ρ2u∗i ,xi , ∀i ∈ C ,
(3)
p(û|u )
X
=
min
σu2 ∗i + σû2 i − 2ρûi ,u∗i σu∗i σûi : ρ2ûi ,u∗i ≤ ρ2u∗i ,xi ,
(4)
{ρûi ,u∗ },{σûi }
i
= min
{σûi }
=
X
i
X
i
σu2 ∗i
i
+ σû2 i − 2ρu∗i ,xi σu∗i σûi
σu2 ∗i (1 − ρ2u∗i ,xi ) .
,
(5)
(6)
In (4), we have solved for the optimal correlations ρûi ,u∗i . Unsurprisingly, the optimal value turns out to be the maximum
allowed by the mutual information constraint, i.e. ûi should be as correlated to u∗i as possible, and in particular as much as
u∗i is correlated to xi . Similarly, in (5) we solve for the optimal σûi , with the result that at optimum, σûi = ρu∗i ,xi σu∗i . This
means that as the correlation between the local state xi and the optimal action u∗i decreases, the variance of the estimated
action ûi decreases as well. As a result, the learned policy will increasingly “bet on the mean” or “listen less” to its local
measurement to approximate the optimal action.
Moreover, we may also provide a closed form expression for the regressor which achieves the minimum distortion D∗ . Since
we have assumed that each u∗i and the state x are jointly Gaussian, we may write any u∗i as an affine function of xi plus
independent Gaussian noise. Thus, the minimum mean squared estimator is given by the conditional expectation
ûi = π̂i (xi ) = E [u∗i |xi ] = E [u∗i ] +
ρu∗i xi σu∗i
(xi − E [xi ]) .
σxi
(7)
Thus, we have found a closed form expression for the best regressor π̂i to predict u∗i from only xi in the joint Gaussian case
with squared error distortion. This result comes as a direct consequence of knowing the true parameterization of the joint
distribution p(u∗ , x) (in this case, as a Gaussian).
3.4
Determining Minimum Distortion in Practice
Often in practice, we do not know the parameterization p(u∗ |x) and hence it may be intractable to determine D∗ and the
corresponding decentralized policies π̂i . However, if one can assume that p(u∗ |x) belongs to a family of parameterized functions
(for instance universal function approximators such as deep neural networks), then it is theoretically possible to attain or at
least approach minimum distortion for arbitrary non-negative distortion measures.
4
Practically, one can compute the mutual information constraint I(u∗i , xi ) from (2) to understand how much information a
regressor π̂i (xi ) has available to reconstruct u∗i . In the Gaussian case, we were able to compute this mutual information in
closed form. For data from general distributions however, there is often no way to compute mutual information analytically.
Instead, we rely on access to sufficient data {x[t], u∗ [t]}Tt=1 , in order to estimate mutual informations numerically. In such
situations (e.g. Sec. 5), we discretize the data and then compute mutual information with a minimax risk estimator, as
proposed by Jiao et al. [2014].
4
Allowing Restricted Communication
Suppose that a decentralized policy π̂i suffers from insufficient mutual information between its local measurement xi and the
optimal action u∗i . In this case, we would like to quantify the potential benefits of communicating with other nodes j 6= i
in order to reduce the distortion limit D∗ from (2) and improve its ability to reconstruct u∗i . In this section, we present an
information-theoretic solution to the problem of how to choose optimally which other data to observe, and we provide a lower
bound-achieving solution for the idealized Gaussian case introduced in Sec. 3.3. We assume that in addition to observing its
own local state xi , each π̂i is allowed to depend on at most k other xj6=i .
Theorem 1. (Restricted Communication)
If Si is the set of k nodes j 6= i ∈ N which ûi is allowed to observe in addition to xi , then setting
Si = arg max I(u∗i ; xi , {xj : j ∈ S}) : |S| = k ,
S
(8)
minimizes the best-case expectation of any distortion measure. That is, this choice of Si yields the smallest lower bound D∗
from (2) of any possible choice of S.
Proof. By assumption, Si maximizes the mutual information between the observed local states {xi , xj : j ∈ Si } and the
optimal action u∗i . This mutual information is equivalent to the notion of rate R in the classical rate distortion theorem
[Cover and Thomas, 2012]. It is well-known that the distortion rate function D(R) is convex and monotone decreasing in R.
Thus, by maximizing mutual information R we are guaranteed to minimize distortion D(R), and hence D∗ .
Theorem 1 provides a means of choosing a subset of the state {xj : j 6= i} to communicate to each decentralized policy
π̂i that minimizes the corresponding best expected distortion D∗ . Practically speaking, this result may be interpreted as
formalizing the following intuition: “the best thing to do is to transmit the most information.” In this case, “transmitting the
most information” corresponds to allowing π̂i to observe the set S of nodes {xj : j 6= i} which contains the most information
about u∗i . Likewise, by “best” we mean that Si minimizes the best-case expected distortion D∗ , for any distortion metric d.
As in Sec. 3.3, without making some assumption about the structure of the distribution of x and u∗ , we cannot guarantee
that any particular regressor π̂i will attain D∗ . Nevertheless, in a practical situation where sufficient data {x[t], u∗ [t]}Tt=1 is
available, we can solve (8) by estimating mutual information [Jiao et al., 2014].
4.1
Example: Joint Gaussian, Squared Error with Communication
Here, we reexamine the joint Gaussian-distributed, mean squared error distortion case from Sec. 3.3, and apply Thm. 1.
We will take u∗ ∈ R1 , x ∈ R10 and u∗ , x jointly Gaussian with zero mean and arbitrary covariance. The specific covariance
matrix Σ of the joint distribution p(u∗ , x) is visualized in Fig. 3a. For simplicity, we show the squared correlation coefficients
of Σ which lie in [0, 1]. The boxed cells in Σ in Fig. 3a indicate that x9 solves (8), i.e. j = 9 maximizes I(u∗ ; x1 , xj ) the
mutual information between the observed data and regression target u∗ . Intuitively, this choice of j is best because x9 is
highly correlated to u∗ and weakly correlated to x1 , which is already observed by û; that is, it conveys a significant amount of
information about u∗ that is not already conveyed by x1 .
Figure 3b shows empirical results. Along the horizontal axis we increase the value of k, the number of additional variables
xj which regressor π̂i observes. The vertical axis shows the resulting average distortion. We show results for a linear regressor
of the form of (7) where we have chosen Si optimally according to (8), as well as uniformly at random from all possible sets of
unique indices. Note that the optimal choice of Si yields the lowest average distortion D∗ for all choices of k. Moreover, the
linear regressor of (7) achieves D∗ for all k, since we have assumed a Gaussian joint distribution.
5
Application to Optimal Power Flow
In this case study, we aim to minimize the voltage variability in an electric grid caused by intermittent renewable energy
sources and the increasing load caused by electric vehicle charging. We do so by controlling the reactive power output of
distributed energy resources (DERs), while adhering to the physics of power flow and constraints due to energy capacity
and safety. Recently, various approaches have been proposed, such as [Farivar et al., 2013] or [Zhang et al., 2014]. In these
methods, DERs tend to rely on an extensive communication infrastructure, either with a central master node [Xu et al.,
5
1
u∗
x1
x2
x3
x4
x5
x6
x7
x8
x9
x10
25
optimal strategy
20
0.6
15
average random strategy
MSE
0.8
0.4
10
0.2
5
0
u∗ x1 x2 x3 x4 x5 x6 x7 x8 x9 x10
0
2
4
6
8
10
Additional Observations k
(a) Squared correlation coefficients.
(b) Comparison of communication strategies.
Figure 3: Results for optimal communication strategies on a synthetic Gaussian example. (a) shows squared correlation
coefficients between of u∗ and all xi ’s. The boxed entries correspond to x9 , which was found to be optimal for k = 1. (b)
shows that the optimal communication strategy of Thm. 1 achieves the lowest average distortion and outperforms the average
over random strategies.
2017] or between agents leveraging local computation [Dall’Anese et al., 2014]. We study regression-based decentralization as
outlined in Sec. 3 and Fig. 2 to the optimal power flow (OPF) problem [Low, 2014], as initially proposed by Sondermeijer
et al. [2016]. We apply Thm. 1 to determine the communication strategy that minimizes optimal distortion to further improve
the reconstruction of the optimal actions u∗i .
Solving OPF requires a model of the electricity grid describing both topology and impedances; this is represented as a
graph G = (N , E). For clarity of exposition and without loss of generality, we introduce the linearized power flow equations
over radial networks, also known as the LinDistFlow equations [Baran and Wu, 1989]:
X
Pij =
Pjk + pcj − pgj ,
(9a)
(j,k)∈E,k6=i
Qij =
X
(j,k)∈E,k6=i
Qjk + qjc − qjg ,
vj = vi − 2 (rij Pij + ξij Qij )
(9b)
(9c)
In this model, capitals Pij and Qij represent real and reactive power flow on a branch from node i to node j for all branches
g
g
(i, j) ∈ E, lower case pci and qic are the real and reactive
√ power consumption at node i, and pi and qi are its real and reactive
power generation. Complex line impedances rij + −1ξij have the same indexing as the power flows. The LinDistFlow
equations use the squared voltage magnitude vi , defined and indexed at all nodes i ∈ N . These equations are included as
constraints in the optimization problem to enforce that the solution adheres to laws of physics.
To formulate our decentralized learning problem, we will treat xi , (pci , qic , pgi ) to be the local state variable, and, for all
controllable nodes, i.e. agents i ∈ C, we have ui , qig , i.e. the reactive power generation can be controlled (vi , Pij , Qij are
treated as dummy variables). We assume that for all nodes i ∈ N , consumption pci , qic and real power generation pgi are
predetermined respectively by the demand and the power generated by a potential photovoltaic (PV) system. The action
space is constrained by the reactive power capacity |ui | = |qig | ≤ q̄i . In addition, voltages are maintained within ±5% of 120V ,
which is expressed as the constraint v ≤ vi ≤ v . The OPF problem now reads
X
|vi − vref | ,
(10)
u∗ = arg g min
qi , ∀i∈C
s.t.
i∈N
(9) , |qig | ≤ q̄i , v ≤ vi ≤ v .
Following Fig. 2, we employ models of real electrical distribution grids (including the IEEE Test Feeders [IEEE PES, 2017]),
which we equip with with T historical readings {x[t]}Tt=1 of load and PV data, which is composed with real smart meter
measurements sourced from Pecan Street Inc. [2017]. We solve (10) for all data, yielding a set of minimizers {u∗ [t]}Tt=1 . We
then separate the overall data set into C smaller data sets {xi [t], u∗i [t]}Tt=1 , ∀i ∈ C and train linear policies with feature
kernels φi (·) and parameters θi of the form π̂i (xi ) = θi> φi (xi ). Practically, the challenge is to select the best feature kernel
φi (·). We extend earlier work which showed that decentralized learning for OPF can be done satisfactorily via a hybrid
forward- and backward-stepwise selection algorithm [Friedman et al., 2001, Chapter 3] that uses a quadratic feature kernels.
Figure 4a shows the result for an electric distribution grid model based on a real network from Arizona. This network
has 129 nodes and, in simulation, 53 nodes were equipped with a controllable DER (i.e. N = 129, C = 53). In Fig. 4a we
show the voltage deviation from a normalized setpoint on a simulated network with data not used during training. The
improvement over the no-control baseline is striking, and performance is nearly identical to the optimum achieved by the
6
2
×10−3
linear, random
linear, optimal
quadratic, random
quadratic, optimal
1.8
MSE
1.6
1.4
1.2
1
0
1
2
3
4
5
Additional Observations k
(a) Voltage output with and without control.
(b) Comparison of OPF communication strategies.
Figure 4: Results for decentralized learning on an OPF problem. (a) shows an example result of decentralized learning the shaded region represents the range of all voltages in a network over a full day. As compared to no control, the fully
decentralized regression-based control reduces voltage variation and prevents constraint violation (dashed line). (b) shows that
the optimal communication strategy Si outperforms the average for random strategies on the mean squared error distortion
metric. The regressors used are stepwise linear policies π̂i with linear or quadratic features.
centralized solution. Concretely, we observed: (i) no constraint violations, and (ii) a suboptimality deviation of 0.15% on
average, with a maximum deviation of 1.6%, as compared to the optimal policy π ∗ .
In addition, we applied Thm. 1 to the OPF problem for a smaller network [IEEE PES, 2017], in order to determine the
optimal communication strategy to minimize a squared error distortion measure. Fig. 4b shows the mean squared error
distortion measure for an increasing number of observed nodes k and shows how the optimal strategy outperforms an average
over random strategies.
6
Conclusions and Future Work
This paper generalizes the approach of Sondermeijer et al. [2016] to solve multi-agent static optimal control problems with
decentralized policies that are learned offline from historical data. Our rate distortion framework facilitates a principled
analysis of the performance of such decentralized policies and the design of optimal communication strategies to improve
individual policies. These techniques work well on a model of a sophisticated real-world OPF example.
There are still many open questions about regression-based decentralization. It is well known that strong interactions
between different subsystems may lead to instability and suboptimality in decentralized control problems [Davison and
Chang, 1990]. There are natural extensions of our work to address dynamic control problems more explicitly, and stability
analysis is a topic of ongoing work. Also, analysis of the suboptimality of regression-based decentralization should be possible
within our rate distortion framework. Finally, it is worth investigating the use of deep neural networks to parameterize both
the distribution p(u∗ |x) and local policies π̂i in more complicated decentralized control problems with arbitrary distortion
measures.
References
P. Abbeel and A. Y. Ng. Apprenticeship Learning via Inverse Reinforcement Learning. In International Conference on
Machine Learning, New York, NY, USA, 2004. ACM.
R. J. Aumann and J. H. Dreze. Cooperative games with coalition structures. International Journal of Game Theory, 3(4):
217–237, Dec. 1974.
M. Baran and F. Wu. Optimal capacitor placement on radial distribution systems. IEEE Transactions on Power Delivery, 4
(1):725–734, Jan. 1989.
S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein. Distributed Optimization and Statistical Learning via the Alternating
Direction Method of Multipliers. Foundations and Trends R in Machine Learning, 3(1):1–122, July 2011.
M. Brambilla, E. Ferrante, M. Birattari, and M. Dorigo. Swarm robotics: a review from the swarm engineering perspective.
Swarm Intelligence, 7(1):1–41, Mar. 2013.
P. D. Christofides, R. Scattolini, D. M. de la Pena, and J. Liu. Distributed model predictive control: A tutorial review and
future research directions. Computers & Chemical Engineering, 51:21–41, 2013.
7
T. M. Cover and J. A. Thomas. Elements of information theory. John Wiley & Sons, 2012.
E. Dall’Anese, S. V. Dhople, and G. Giannakis. Optimal dispatch of photovoltaic inverters in residential distribution systems.
Sustainable Energy, IEEE Transactions on, 5(2):487–497, 2014. URL http://ieeexplore.ieee.org/xpls/abs_all.jsp?
arnumber=6719562.
E. J. Davison and T. N. Chang. Decentralized stabilization and pole assignment for general proper systems. IEEE Transactions
on Automatic Control, 35(6):652–664, 1990.
M. Farivar, L. Chen, and S. Low. Equilibrium and dynamics of local voltage control in distribution systems. In 2013 IEEE
52nd Annual Conference on Decision and Control (CDC), pages 4329–4334, Dec. 2013. doi: 10.1109/CDC.2013.6760555.
J. Friedman, T. Hastie, and R. Tibshirani. The elements of statistical learning, volume 1. Springer series in statistics Springer,
Berlin, 2001.
C. V. Goldman and S. Zilberstein. Decentralized control of cooperative systems: Categorization and complexity analysis. J. Artif.
Int. Res., 22(1):143–174, Nov. 2004. ISSN 1076-9757. URL http://dl.acm.org/citation.cfm?id=1622487.1622493.
IEEE PES. IEEE Distribution Test Feeders, 2017. URL http://ewh.ieee.org/soc/pes/dsacom/testfeeders/.
J. Jiao, K. Venkat, Y. Han, and T. Weissman. Minimax Estimation of Functionals of Discrete Distributions. arXiv preprint,
June 2014. arXiv: 1406.6956.
S. Low. Convex Relaxation of Optimal Power Flow; Part I: Formulations and Equivalence. IEEE Transactions on Control of
Network Systems, 1(1):15–27, Mar. 2014.
J. Lunze. Feedback Control of Large Scale Systems. Prentice Hall PTR, Upper Saddle River, NJ, USA, 1992. ISBN 013318353X.
P. J. Modi, W.-M. Shen, M. Tambe, and M. Yokoo. Adopt: Asynchronous distributed constraint optimization with
quality guarantees. Artif. Intell., 161(1-2):149–180, Jan. 2005. ISSN 0004-3702. doi: 10.1016/j.artint.2004.09.003. URL
http://dx.doi.org/10.1016/j.artint.2004.09.003.
R. Nair, P. Varakantham, M. Tambe, and M. Yokoo. Networked Distributed POMDPs: A synthesis of distributed constraint
optimization and POMDPs. In AAAI, volume 5, pages 133–139, 2005.
F. A. Oliehoek and C. Amato. A Concise Introduction to Decentralized POMDPs. Springer International Publishing, 1 edition,
2016.
P. A. Ortega, D. A. Braun, J. Dyer, K.-E. Kim, and N. Tishby. Information-Theoretic Bounded Rationality. arXiv preprint,
2015. arXiv:1512.06789.
Pecan Street Inc. Dataport, 2017. URL http://www.pecanstreet.org/.
L. Peshkin, K.-E. Kim, N. Meuleau, and L. P. Kaelbling. Learning to cooperate via policy search. In Proceedings of the
Sixteenth Conference on Uncertainty in Artificial Intelligence, UAI’00, pages 489–496, San Francisco, CA, USA, 2000.
Morgan Kaufmann Publishers Inc. ISBN 1-55860-709-9. URL http://dl.acm.org/citation.cfm?id=2073946.2074003.
Y. Pu, M. N. Zeilinger, and C. N. Jones. Inexact fast alternating minimization algorithm for distributed model predictive
control. In Conference on Decision and Control, Los Angeles, CA, USA, 2014. IEEE.
R. L. Raffard, C. J. Tomlin, and S. P. Boyd. Distributed optimization for cooperative agents: Application to formation flight.
In Conference on Decision and Control, Nassau, The Bahamas, 2004. IEEE.
C. Sammut. Automatic construction of reactive control systems using symbolic machine learning. The Knowledge Engineering
Review, 11(01):27–42, 1996.
D. D. Siljak. Decentralized control of complex systems. Dover Books on Electrical Engineering. Dover, New York, NY, 2011.
URL http://cds.cern.ch/record/1985961.
O. Sondermeijer, R. Dobbe, D. B. Arnold, C. Tomlin, and T. Keviczky. Regression-based Inverter Control for Decentralized
Optimal Power Flow and Voltage Regulation. In Power and Energy Society General Meeting, Boston, MA, USA, July 2016.
IEEE.
A. X. Sun, D. T. Phan, and S. Ghosh. Fully decentralized AC optimal power flow algorithms. In Power and Energy Society
General Meeting, Vancouver, Canada, July 2013. IEEE.
Y. Xu, Z. Y. Dong, R. Zhang, and D. J. Hill. Multi-Timescale Coordinated Voltage/Var Control of High RenewablePenetrated Distribution Systems. IEEE Transactions on Power Systems, PP(99):1–1, 2017. ISSN 0885-8950. doi:
10.1109/TPWRS.2017.2669343.
8
M. N. Zeilinger, Y. Pu, S. Riverso, G. Ferrari-Trecate, and C. N. Jones. Plug and play distributed model predictive control
based on distributed invariance and optimization. In Conference on Decision and Control, Florence, Italy, 2013. IEEE.
B. Zhang, A. Lam, A. Dominguez-Garcia, and D. Tse. An Optimal and Distributed Method for Voltage Regulation
in Power Distribution Systems. IEEE Transactions on Power Systems, PP(99):1–13, 2014. ISSN 0885-8950. doi:
10.1109/TPWRS.2014.2347281.
9
| 7cs.IT
|
Permutation monoids and MB-homogeneous structures
arXiv:1802.04166v1 [math.GR] 12 Feb 2018
Thomas D. H. Coleman1, David M. Evans2, Robert D. Gray3,4
February 13, 2018
Abstract
In this paper, we investigate the connection between infinite permutation monoids
and bimorphism monoids of first-order structures. Taking our lead from the study
of automorphism groups of structures as infinite permutation groups and the more
recent developments in the field of homomorphism-homogeneous structures, we
establish a series of results that underline this connection. Of particular interest
is the idea of MB-homogeneity; a relational structure M is MB-homogeneous if
every monomorphism between finite substructures of M extends to a bimorphism
of M.
The results in question include a characterisation of closed permutation monoids,
a Fraı̈ssé-like theorem for MB-homogeneous structures, and the construction of 2ℵ0
pairwise non-isomorphic countable MB-homogeneous graphs. We prove that any
finite group arises as the automorphism group of some MB-homogeneous graph
and use this to construct oligomorphic permutation monoids with any given finite
group of units. We also consider MB-homogeneity for various well-known examples of homogeneous structures and in particular give a complete classification of
countable homogeneous undirected graphs that are also MB-homogeneous.
Keywords: bimorphisms, MB-homogeneous, cancellative monoids, permutation monoids, oligomorphic transformation monoids, homomorphism-homogeneous
structures, infinite graph theory.
2010 Mathematics Subject Classification: 20M99, 03C15, 05C63
Let M be a first-order structure with automorphism group Aut(M). As “structure
is whatever is preserved by automorphisms” [17], the automorphism group is a key
concept in understanding the model theory of M. Every automorphism of M is a
permutation of the domain M of M; hence we can view Aut(M) as an infinite permutation group. Much of the existing literature in this field explores connections between
infinite permutation group theory and model theory; see [2] and [7] for instance. More
recent work has studied the endomorphism monoid End(M) of a first-order structure
M; analogously, these are examples of infinite transformation monoids. By imposing additional conditions on the type of endomorphism of M, we can obtain various
1
School of Mathematics and Statistics, University of St Andrews, St Andrews, KY16 9SS, United
Kingdom. Email: [email protected].
2
Department of Mathematics, Imperial College London, South Kensington Campus, London, SW7
2AZ, United Kingdom. Email: [email protected].
3
School of Mathematics, University of East Anglia, Norwich, NR4 7TJ, United Kingdom. Email:
[email protected].
4
This work was supported by the EPSRC grant EP/N033353/1 ‘Special inverse monoids: subgroups,
structure, geometry, rewriting systems and the word problem’.
1
submonoids of End(M). These various natural monoids of transformations associated
with a first-order structure M have been studied by Lockett and Truss [20, 21].
The principal aim of papers by Cameron and Nešetřil [9], Mašulović [23], [24],
Mašulović and Pech [25], Lockett and Truss [20, 21], and Hartman, Hubička and
Mašulović [16] focus on generalising the current theory on infinite permutation groups
to the case of infinite transformation monoids; particularly in the case of End(M).
Research on End(M) is not restricted to finding analogues of results about Aut(M);
understanding End(M) is key to the study of the polymorphism clone Pol(M) of M
and hence complexity of constraint satisfaction problems [5]. This connection provides
motivation for studying endomorphism monoids of first-order structures, and has been
extensively studied by Bodirsky [3, 4], Bodirsky and Nešetřil [5] and Bodirsky and
Pinsker [6].
The collection of all bijective endomorphisms of a first-order structure M with domain M forms a monoid under the composition operation; we call this the bimorphism
monoid of the structure M and denote it by Bi(M). Of course, every automorphism
of M is a bijective homomorphism; but in general, the converse does not hold. Hence
Bi(M) contains the automorphism group Aut(M) (as its group of units), but is also
contained in the symmetric group Sym(M ) since every element of Bi(M) is a bijection
from M to itself. Therefore, the bimorphism monoid Bi(M) gives a natural example of a permutation monoid; a monoid where each element is a permutation. In
fact, as we shall see for any countable set M , the closed submonoids of the symmetric
group Sym(M ) under the pointwise convergence topology are precisely the bimorphism
monoids of first-order structures M on the domain M (Theorem 2.1).
Although it is a natural concept, permutation monoids have not received much attention in the literature. By definition, every permutation monoid is a group-embeddable
monoid. The study of group-embeddable monoids was a principal interest of early semigroup theorists; a well known result (Ore’s Theorem, see [27]) says that a monoid T is
embeddable in a group if it is cancellative and satisfies Ore’s condition (for a, b ∈ T ,
then aT ∩ bT 6= ∅). Furthermore, a monoid T is group-embeddable if and only if M
has a faithful representation as a monoid of permutations. Equivalently, M is groupembeddable if and only if M is isomorphic to a submonoid of some symmetric group;
that is, a permutation monoid.
A result of Reyes [30] (see [7]) states that a subgroup of the infinite symmetric group
is closed if and only if it is the automorphism group of some first-order structure. Examples of highly symmetric infinite permutation groups are abundant in the literature,
often arising as automorphism groups of structures with “nice” structural conditions;
specifically, ℵ0 -categoricity and homogeneity. A structure M is ℵ0 -categorical if and
only if M is the unique (up to isomorphism) countable model of the theory Th(M) of
M, and it is homogeneous if every isomorphism between finite substructures of M extends to an automorphism of M. The two are related; every homogeneous structure M
over a finite relational language is ℵ0 -categorical, and an ℵ0 -categorical structure M is
homogeneous if and only if Th(M) has quantifier elimination [22]. Finding examples of
ℵ0 -categorical structures provides corresponding examples of interesting permutation
groups. It is a famous theorem of Engeler, Ryll-Nardzewski and Svenonius (see [17])
that a structure M is ℵ0 -categorical if and only if Aut(M) has finitely many orbits on
M n for every n ∈ N; such groups are called oligomorphic permutation groups. It follows
that a potential source of oligomorphic permutation groups are automorphism groups
of homogeneous structures over a finite relational language. The celebrated theorem of
2
Fraı̈ssé [14] gives a characterisation of homogeneous structures, which can then be used
to construct many examples of structures with an oligomorphic automorphism group.
Such structures include the countable dense linear order without endpoints (Q, <), the
random graph R, and the generic poset P . This was followed in time by complete
classification results for countable homogeneous structures; for posets by Schmerl [32],
undirected graphs by Lachlan and Woodrow [19] and directed graphs by Cherlin [10].
In the study of infinite permutation monoids and bimorphism monoids of first-order
structures, the natural analogue of homogeneity is MB-homogeneity. A structure M is
MB-homogeneous if every finite partial monomorphism of M extends to a bimorphism
of M. This notion of homogeneity was first introduced by Lockett and Truss in [20], in
which they determined conditions using MB-homogeneity for the existence of generic
bimorphisms of a structure. Further results were shown by the same authors [21]; where
they classified homomorphism-homogeneous (and hence MB-homogeneous) posets in
a wide-ranging result. MB-homogeneity is a natural extension of MM-homogeneity –
where every finite partial monomorphism of M extends to a monomorphism of M
– which has been more widely considered; for instance, Cameron and Nešetřil [9]
demonstrated a Fraı̈ssé-theoretic result about construction and uniqueness of MMhomogeneous structures.
The aim of this paper is to develop the theory of infinite permutation monoids,
with a particular focus on bimorphism monoids and MB-homogeneous structures. We
begin in Section 1 by recalling the definition of an oligomorphic transformation monoid
from [25], and extend results of the same paper by considering other self-map monoids
as introduced by Lockett and Truss [21]. Section 2 investigates permutation monoids
and MB-homogeneity in more detail, including a characterisation of closed permutation
monoids (Theorem 2.1) and a Fraı̈ssé-like theorem for MB-homogeneity (Propositions
2.6, 2.7, and 2.8). Section 3 is devoted to MB-homogeneous graphs; including establishing some useful properties of MB-homogeneous graphs, introducing the notion of
bimorphism equivalence (Definition 3.7) and demonstrating that there exist 2ℵ0 nonisomorphic MB-homogeneous graphs, each of which is bimorphism equivalent to the
random graph R (Theorem 3.20). Furthermore, we show that for any finite group H,
there exists a MB-homogeneous graph Γ such that Aut(Γ) ∼
= H (Theorem 3.21); consequently there exists an oligomorphic permutation monoid with H as its group of units.
Finally, Section 4 assesses a selection of previously known homogeneous structures to
determine whether or not they are MB-homogeneous, culminating in a complete classification of countably infinite graphs that are both homogeneous and MB-homogeneous
(Theorem 4.11).
Throughout the article, maps act on the right of their arguments, and we compose
maps from left to right. A relational first-order signature σ consists of a collection of
relations {R̄i : i ∈ I} where each Ri has an arity ni ∈ N for all i ∈ I. A σ-structure
with equality M consists of a domain M and subsets Ri ⊆ M ni interpreting R̄i in
σ for each i ∈ I. We follow usual convention for notation regarding relations; for a
σ-structure M, we write RiM (x̄) if x̄ ∈ Ri ⊆ M ni . All structures M will be countably
infinite unless stated otherwise.
3
1
Oligomorphic transformation monoids
Let X be a countably infinite set and let T be an infinite submonoid of End(X). Then
T is a transformation monoid acting on the set X via (x, α) 7→ xα for all x ∈ X
and α ∈ T . We can extend this to an action of T on tuples x̄ ∈ X n , with T acting
componentwise on x̄. We begin by outlining some important definitions regarding the
notion of orbits on tuples for a transformation monoid; versions of these appear in [33].
Definition 1.1. Let T ⊆ End(X) be a transformation monoid acting on tuples X n as
above, and let U be the group of units of T .
• Define the forward orbit of a tuple x̄ to be the set F (x̄) = {ȳ ∈ X n : (∃s ∈
T )(x̄s = ȳ)}.
• Define the strong orbit of x̄ to be the set S(x̄) = {ȳ ∈ X n : (∃s, t ∈ T )(x̄s = ȳ
and ȳt = x̄)}.
• Define the group orbit of x̄ to be the set U (x̄) = {ȳ ∈ X n : (∃s ∈ U )(x̄s = ȳ)}.
We note immediately that U (x̄) ⊆ S(x̄) ⊆ F (x̄) for any tuple x̄ and that if ȳ ∈ F (x̄)
then F (ȳ) ⊆ F (x̄). Furthermore, the relation x̄ ∼ ȳ if and only if x̄ and ȳ are in the
same strong orbit is an equivalence relation [33]. In comparison, the forward orbit
is reflexive and transitive (and thus a preorder), but it may not be symmetric. This
notion therefore differs slightly from Steinberg’s weak orbits, obtained by taking the
symmetric-transitive closure of the forward orbit preorder [33]. We outline a basic
lemma regarding these orbits which will be useful throughout the section; the proof of
this is omitted.
Lemma 1.2. Let T be a transformation monoid acting on a set of tuples X n . For any
tuple x̄ ∈ X n , we have the following:
S
(1) F (x̄) = ȳ∈F (x̄) S(ȳ);
S
(2) S(x̄) = ȳ∈S(x̄) U (ȳ).
Recall that a permutation group G ⊆ Sym(X) acts oligomorphically on X if and
only if there are finitely many orbits on X n for every n ∈ N [7]. If the action of
G componentwise on tuples of X is oligomorphic, we say that G is an oligomorphic
permutation group. The next definition, originally of [25], places these concepts in the
context of transformation monoids.
Definition 1.3. We say that a transformation monoid T ⊆ End(X) acts oligomorphically on X if and only if there are finitely many strong orbits on X n for every
n ∈ N. If the componentwise action of T on tuples of X is oligomorphic, then call T
an oligomorphic transformation monoid.
We note that if T is itself a group, then the strong orbits are the group orbits and
the definitions coincide; so any oligomorphic permutation group is an oligomorphic
transformation monoid. Our next result provides more connections between oligomorphic permutation groups and oligomorphic transformation monoids, generalising [25,
Lemma 2.10].
4
Proposition 1.4. Let T ⊆ End(X) be a transformation monoid with group of units U .
If U is an oligomorphic permutation group then T is an oligomorphic transformation
monoid.
Proof. As U is an oligomorphic permutation group, there are finitely many group orbits
U (ȳ) with y ∈ X n for every n ∈ N. As every strong orbit S(x̄) arises as the union of
group orbits U (ȳ), we conclude that there are at most finitely many strong orbits of T
acting on X n for every natural number n by Lemma 1.2.
Remark. By the Ryll-Nardzewski theorem (see [17]), U is an oligomorphic permutation
group if and only if it is the automorphism group of some ℵ0 -categorical structure
M. By this proposition and the fact that Aut(M) acts as the group of units for
any endomorphism monoid of M, we conclude that if M is ℵ0 -categorical then T ∈
{End(M), Epi(M), Mon(M), Bi(M), Emb(M)} is an oligomorphic transformation
monoid. (See [21] for the definitions of these various transformation monoids.)
This result provides us with numerous examples of oligomorphic transformation
monoids, with the caveat that they are closely related to ℵ0 -categorical structures via
their group of units. The main result of this section distances the notion of oligomorphicity in monoids from ℵ0 -categoricity by providing a different source of suitable
examples; but first we detail some preliminary conditions. In the same way that homogeneous structures over a finite language provide examples of ℵ0 -categorical structures
(and hence oligomorphic permutation groups), we turn to homomorphism-homogeneity
to provide examples of oligomorphic transformation monoids. In the table below, we
recall the eighteen different notions of homomorphism-homogeneity as presented in the
two papers of Lockett and Truss [20, 21].
isomorphism (I)
monomorphism (M)
homomorphism (H)
End(M) (H)
IH
MH
HH
Epi(M) (E)
IE
ME
HE
Mon(M) (M)
IM
MM
HM
Bi(M) (B)
IB
MB
HB
Emb(M) (I)
II
MI
HI
Aut(M) (A)
IA
MA
HA
Table 1: Table of XY-homogeneity: a first-order structure M is XY-homogeneous if a
finite partial map of type X (column) extends to a map of type Y (row) in the associated
monoid.
We note that if a structure M is XY-homogeneous then it is also IY-homogeneous.
For shorthand in our next result, denote the monoid of maps of type Y by Y(M) for
some structure M. For example, H(M) is the endomorphism monoid of M, and A(M)
is the automorphism group. A previous observation of Lockett and Truss in [21] says
that an endomorphism of a finite relational structure is an automorphism if and only
if it is a bijection.
Lemma 1.5. If A and B are finite σ-structures and f : A −
→ B and g : B −
→ A are
∼
bijective homomorphisms, then A = B and f, g are isomorphisms.
5
Proof. The composition map f g : A −
→ A is a bijective endomorphism of A; by the
observation above, f g must be an automorphism of A. For some ā ∈ Ani , if ¬RiA (ā)
and RiB (āf ), then RiA (āf g) as g is a homomorphism. Since f g is an automorphism this
is a contradiction; so f must preserve non-relations and is therefore an isomorphism.
A similar argument applies to show that g is an isomorphism.
Proposition 1.6. Let M be an XY-homogeneous structure with domain M . Then two
tuples ā = (a1 , . . . , an ) and b̄ = (b1 , . . . , bn ) are in the same strong orbit of Y(M) if
and only if there exists a partial isomorphism f of M such that āf = b̄.
Proof. Suppose that α, β ∈ Y(M) are maps such that āα = b̄ and b̄β = ā respectively;
so ai α = bi and bi β = ai for 1 ≤ i ≤ n. If α sends elements ai 6= aj of ā to elements
ai α = aj α of b̄, then we have that ai αβ = aj αβ and so ai = aj ; a contradiction. Hence
the restrictions α|ā and β|b̄ are injective maps and so they are also bijections. Now, we
→ A, where A, B are the structures induced
consider the maps α|ā : A −
→ B and β|b̄ : B −
by M on ā, b̄ respectively. By Lemma 1.5, we see that A ∼
= B and so we can define
f = α|ā . Conversely, we see that if f is an isomorphism between the structures A and
B; by XY-homogeneity (and hence IY-homogeneity) of M, we extend f to a map α ∈
Y(M) such that āα = b̄. Similarly, we extend the map f −1 : B −
→ A to a map β ∈
Y(M) such that b̄β = ā and so ā and b̄ are in the same strong orbit.
We now move on to proving the final result of this section. A result of [7, ch2] states
that, for Aut(M) acting on M n , the number of group orbits on n-tuples is finite if the
number of group orbits on k-tuples of distinct elements is finite for every k ≤ n. Using
the fact from Proposition 1.6 that maps between two tuples in the same strong orbit
are bijections, it is not hard to show that a similar result holds for the number of strong
orbits of n-tuples when Y(M) is acting on M n . This, together with Proposition 1.6,
proves the following theorem.
Theorem 1.7. If M is an XY-homogeneous structure over a finite relational language,
then Y(M) is an oligomorphic transformation monoid.
Proof. As M is over a finite relational language, it has finitely many isomorphism types
on k-tuples of distinct elements for any k ∈ N; by Proposition 1.6, there are finitely
many strong orbits on k-tuples of distinct elements for every k ∈ N. The result then
follows from the observation above.
Using this theorem, we can find examples of structures with oligomorphic transformation monoids that are not ℵ0 -categorical; for instance, any homomorphism-homogeneous
poset not in Schmerl’s classification (see [23] or [21]) has an oligomorphic endomorphism
monoid. A notable instance is Corollary 2.2(a) in [9]; there exists a countably infinite
graph Γ with oligomorphic endomorphism (and monomorphism) monoid but a trivial
automorphism group.
Of particular interest to this paper is the idea of an oligomorphic permutation
monoid. This follows as a special case of Definition 1.3, where the transformation
monoid T is also a permutation monoid. It is a corollary of Theorem 1.7 that if M
is an MB-homogeneous structure over a finite relational language, then Bi(M) is an
oligomorphic permutation monoid. It follows that finding MB-homogeneous structures
give interesting examples of permutation monoids; this motivates the study of MBhomogeneous structures in Section 2.
6
2
Permutation monoids and MB-homogeneity
Among the endomorphism monoids of a first-order structure M mentioned in the introduction was the collection of bimorphisms Bi(M). As a set of endomorphisms it
is a submonoid of End(M); however as a collection of bijective maps it is a submonoid of Sym(M ), the symmetric group on the domain of M . It follows that Bi(M)
is a group-embeddable monoid and so it acts on M by permutations as opposed to
transformations. Thus Bi(M) (for a countably infinite M) expressed in this fashion
is an infinite permutation monoid. This section is devoted to the study of infinite
permutation monoids, outlining some general properties and presenting a method for
constructing interesting infinite permutation monoids via MB-homogeneous structures.
To start, recall that the symmetric group Sym(M ) on a countably infinite set M
has a natural topology given by pointwise convergence, with a basis for open sets given
by the cosets of stabilizers of tuples. It is well-known ([30], see 4.1.4 in [17]) that
G ≤ Sym(M ) is closed under the pointwise convergence topology if and only if it is
the automorphism group of some first-order structure M on domain M . This was
generalised by Cameron and Nešetřil [9] to closed submonoids T ≤ End(M ) under the
product topology; this occurs if and only if T is the endomorphism monoid of some
first-order structure M on domain M . Our first proposition provides an analogous
result for closed permutation monoids; the proof of which is similar to other results
along these lines. To do this, we recall the following standard result from point-set
topology [29, Theorem 17.2]; if Y is a subspace of some topological space X, then a set
A is closed in Y under the subspace topology inherited from X if and only if A = Y ∩B,
where B is some closed set in X. Furthermore, the pointwise convergence topology on
Sym(M ) is the same as the topology induced on the set by End(M ) via the inclusion
map: so Sym(M ) is a subspace of End(M ).
Theorem 2.1. Let M be a countable set. A submonoid T of Sym(M ) is closed under
the pointwise convergence topology if and only if it is the bimorphism monoid of some
structure M on domain M .
Proof. We begin with the converse direction. Suppose that T = Bi(M) is the bimorphism monoid of a structure M on domain M . As Sym(M ) is a subspace of End(M ),
and Bi(M) is the intersection of Sym(M ) and the closed set End(M) of End(M ), it
follows from the result above that Bi(M) is closed in Sym(M ).
For the forward direction, assume that T is a closed submonoid of Sym(M ). Define
an n-ary relation Rx̄ by
Rx̄ (ȳ) ⇔ (∃s ∈ T )(x̄s = ȳ)
for each n ∈ N and x̄ ∈ M n . Let M be the relational structure on M with relations
Rx̄ for all tuples x̄ ∈ M n and all n ∈ N. The proof that T = Bi(M) is by containment
both ways.
As every element of T is already a permutation of the domain M of M, proving
that T acts as endomorphisms on M is enough to show that T ⊆ Bi(M). Assume
then that s ∈ T and ȳ ∈ M n such that Rx̄ (ȳ) holds; as this happens, there exists an
s′ ∈ T such that x̄s′ = ȳ. Therefore x̄s′ s = ȳs and so Rx̄ (ȳs) holds; so T ⊆ End(M)
and hence T ⊆ Bi(M).
It remains to show that Bi(M) ⊆ T ; so suppose that α ∈ Bi(M). Our aim is to
show that α is a limit point of T ; as T is closed, it must contain all its limit points.
7
Note that each n-tuple x̄ defines a neighbourhood of α, consisting of all functions β
such that x̄α = x̄β. As T is a monoid, it follows that Rx̄ (x̄) holds and so Rx̄ (x̄α) holds.
By definition of Rx̄ , there exists s ∈ T such that x̄α = x̄s; hence α is a limit point of
T . Therefore α ∈ T and so Bi(M) ⊆ T , completing the proof.
Remark. It is a well-known result from descriptive set theory that any closed subset
A of a Polish space X is itself a Polish space with the induced topology from X (see
[18]). As Sym(M ) is a Polish space, it follows that Bi(M) is also a Polish space; so
bimorphisms of first-order structures provide natural examples of Polish monoids. We
leave this area of investigation open.
Our aim now is to determine a cardinality result for closed submonoids of Sym(X).
For any x̄ ∈ X n , define the pointwise stabilizer of x̄ to be the set
St(x̄) = {α ∈ Bi(M) : x̄α = x̄}
Note also that as Bi(M) is a group-embeddable monoid, it is therefore cancellative.
Proposition 2.2. For any countably infinite first-order structure M, either |Bi(M)| ≤
ℵ0 or |Bi(M)| = 2ℵ0 , the first alternative holding if and only if the pointwise stabilizer
of some tuple is the identity e ∈ Bi(M).
Proof. Assume that St(x̄) = {e} for some x̄ ∈ M n . Let γ1 , γ2 ∈ Bi(M) where x̄γ1 =
x̄γ2 = ȳ. As Bi(M) is a collection of permutations, it embeds in Sym(M ) = G,
and the action of Bi(M) on M extends to an action of G on M. As a consequence,
there exists a unique γ1−1 ∈ G such that γ1 γ1−1 = e. As ȳγ1−1 = x̄, we have that
x̄γ2 γ1−1 = x̄γ1 γ1−1 = x̄. But St(x̄) = {e} and so γ2 γ1−1 = e; hence γ1 = γ2 . By the fact
there are only countably many tuples in F (x̄), we are forced to conclude that Bi(M)
is countable in this case.
On the other hand, suppose that St(x̄) 6= {e} for all tuples x̄ ∈ M n . As M
is countably infinite, we can enumerate elements of M = {x1 , x2 , . . .}. Using this
enumeration, we define a sequence of tuples (x̄k )k∈N where x̄k = (x1 , . . . , xk ) for all
k ∈ N. Since St(x̄) 6= {e} for all tuples x̄ of M, for each element x̄k of (x̄k )k∈N there
exists tk ∈ Bi(M) such that tk 6= e and x̄k tk = x̄k . This means that the sequence
(tk )k∈N of elements of Bi(M) is an infinite sequence where each bimorphism tk is
not the identity element of Bi(M). As the sequence of tuples (x̄k )k∈N will eventually
encapsulate every element of M , the sequence of bimorphisms (tk )k∈N approaches the
pointwise stabilizer of M. This is the identity element and so e is a limit point of
Bi(M).
Now, for some α ∈ Bi(M), consider the sequence (tk α)k∈N . Here, tk α 6= α for any
k ∈ N; for if tk α = α for some k, then cancellativity of Bi(M) implies that tk = e,
contradicting our earlier assumption. Then α is a limit point for the sequence (tk α)k∈N ,
and so every element of Bi(M) is a limit point. This means that Bi(M) is a perfect
set and thus has cardinality of the continuum (ch 6, [18]).
2.1
MB-homogeneity
As mentioned in the introduction, automorphism groups of homogeneous structures
are examples of infinite permutation groups whose orbits on n-tuples of distinct elements are determined by isomorphism types of substructures (see Proposition 1.6). It
8
is a consequence of Theorem 1.7 that bimorphism monoids of MB-homogeneous structures provide examples of infinite permutation monoids whose orbits on n-tuples are
defined in the same way. Motivated by this, we aim to find a way of constructing MBhomogeneous structures; our inspiration for this task is Fraı̈ssé’s theorem for finding
homogeneous structures (see [2]). Building on a similar result of Cameron and Nešetřil
[9] for MM-homogeneous structures, the rest of this section is devoted to providing
a Fraı̈ssé-like theorem for constructing MB-homogeneous structures. Throughout this
section σ is a relational signature, C is a class of finite σ-structures and, following
convention, we write A, B to mean the σ-structures on domains A, B. We refer the
reader to [17] for further background on model theory.
There are two main properties that a class of finite structures C has that guarantee
the existence of a countable homogeneous structure with age C . The first is the joint
embedding property (JEP); this property ensures that we can construct a countable
structure M with age C . The second is the amalgamation property (AP); this ensures
our constructed structure M is homogeneous by showing that it has the equivalent
extension property (EP). When building a countable MB-homogeneous structure M
with age C , we still need the JEP to ensure that M has the age we want but we need
a different amalgamation property to ensure MB-homogeneity rather than standard
homogeneity. As we are aiming to extend to bijective endomorphisms, we require a
“back and forth” style argument to ensure that the extended map is indeed bijective.
Due to the fact that bimorphisms are not automorphisms in general, we require two
amalgamation conditions, and hence two extension conditions, to ensure that M is MBhomogeneous. We begin this section by examining the required extension properties.
If M is MB-homogeneous then M is MM-homogeneous; it follows (by Proposition
4.1(b) of [9]) that M must have the mono-extension property (MEP), which takes care
of the forward extension:
(MEP) For all A, B ∈ Age(M) with A ⊆ B and monomorphism f : A −
→ M,
there exists a monomorphism g : B −
→ M extending f .
Any suitable “back” condition should express in its statement the key difference
between a monomorphism of M and a bimorphism. This difference is the existence of
a preimage; for any α ∈ Bi(M) and substructure A ⊆ M there exists a substructure
B ⊆ M such that Bα = A. Note that this is not true for an arbitrary monomorphism
β of M as Mβ is not required to equal M. The existence of a preimage for every
extended map is the condition we wish to ensure in our “back” extension property, and
this motivates our next definition.
Definition 2.3. Let A, B be two σ-structures. We say that an injective map f : A −
→B
is an antimonomorphism if and only if ¬RA (a1 , . . . , an ) implies ¬RB (a1 f, . . . , an f ) for
all n-ary relations R of σ.
Remarks. Note that any map that is both a monomorphism and an antimonomorphism
is an embedding.
It is an easy exercise to show that the function composition of two antimonomorphisms is again an antimonomorphism; in particular, it is important to note that the
composition of an antimonomorphism and an embedding is an antimonomorphism.
We now state and prove a lemma stating that a preimage of a monomorphism is
an antimonomorphism. For any bijective function f : A → B, we write f¯ : B → A to
be the unique bijection such that f f¯ = 1A and f¯f = 1B . Notice here that f¯ = f .
9
Lemma 2.4. Let A, B be two σ-structures, and suppose that f : A −
→ B is a bijection.
¯
Then f is a monomorphism if and only if f is an antimonomorphism.
Proof. Suppose f is a monomorphism and that ¬RB (b1 , . . . , bn ) = ¬RB (a1 f, . . . , an f )
holds in B. Since f preserves relations, it follows that ¬RA (a1 , . . . , an ) holds in A; as
¯ . . . , an f f¯) = ¬RA (b1 f¯, . . . , bn f¯).
f f¯ = 1A we have that ¬RA (a1 , . . . , an ) = ¬RA (a1 f f,
In the converse direction, suppose that f¯ is an antimonomorphism and that RA (a1 , . . . , an )
is a relation with ai = bi f¯ for all 1 ≤ i ≤ n. As f¯ preserves non-relations, it follows that RB (b1 , . . . , bn ) holds. Finally, as f¯f = 1B , we have that RB (b1 , . . . , bn ) =
RB (b1 f¯f, . . . , bn f¯f ) = RB (a1 f, . . . , an f ) and therefore f is an monomorphism.
Remarks. By restricting the codomain of a monomorphism f : A −
→ B to the image,
we see that f ′ : A −
→ Af is a bijective homomorphism and therefore f¯′ : Af −
→ A is
an antimonomorphism by the above lemma. Similarly, by restricting the codomain of
an antimonomorphism f¯ : B −
→ A to the image, we see that f¯′ : B −
→ B f¯ and so we
obtain a bijective homomorphism f ′ : B f¯ −
→ B.
It is an immediate corollary of this result that if A, B, C are σ-structures, and
f : A → B, g : B → C are bijective homomorphisms, then (f g) = ḡ f¯ : C → A is a
bijective antimonomorphism.
We use antimonomorphisms to express the backward extension in the bi-extension
property (BEP):
(BEP) For all A, B ∈ Age(M) with A ⊆ B and antimonomorphism
f¯ : A −
→ M, there exists a antimonomorphism ḡ : B −
→ M extending f¯.
These properties prove to be necessary and sufficient for MB-homogeneity.
Proposition 2.5. Let M be a countable structure with age C . Then M is MBhomogeneous if and only if M has the BEP and the MEP.
Proof. Suppose that M is MB-homogeneous; then M is also MM-homogeneous and
has the MEP from Proposition 4.1(a) of [9]. Let A ⊆ B ∈ C and f¯ : A → M be
an antimonomorphism. As Age(M) = C , there exist copies A′ ⊆ B ′ ⊆ M of A and
B and isomorphisms θ : B −
→ B ′ and θ −1 : B ′ → B. Restrict the codomain of f¯ to
its image to find an antimonomorphism f¯′ : A → Af¯; so θ −1 f¯′ = h̄ : A′ → Af¯ is a
¯ = h : Af¯ → A′ is a monomorphism; as
bijective antimonomorphism. By Lemma 2.4, h̄
M is MB-homogeneous, extend h to a bimorphism α of M. So α|B ′ θ −1 : B ′ → B is a
bijective homomorphism; by Lemma 2.4, define θ ᾱ = ḡ : B → B ′ . It remains to show
that θ ᾱ extends f¯; indeed, for a ∈ A we have that
af¯ = af¯αᾱ = af¯f θ ᾱ = aθ ᾱ
as α extends f θ. Therefore M has the BEP.
Conversely, suppose that M has the BEP and the MEP, and let f : A −
→ B be a
monomorphism between finite substructures of M. We shall extend f : A → B to a
bimorphism α of M in stages using a back and forth argument. Set A0 = A, B0 = B
and f0 = f . At a typical stage, we have a bijective monomorphism fk : Ak −
→ Bk
extending f , where Ai ⊆ Ai+1 and Bi ⊆ Bi+1 for all i ≤ k. As M is countable, we
can enumerate the points M = {m0 , m1 , . . .}. When k is even, pick a point mi , where
i is the smallest number such that mi ∈
/ dom f . So, Ak ∪ {mi } is a substructure of M
10
′
: Ak ∪ {mi } −
→ M.
containing Ak . Using the MEP, extend fk to a monomorphism fk+1
′
′
→ Bk ∪{mi fk+1 } is a
By restricting the codomain of fk+1 we have that fk+1 : Ak ∪{mi } −
bijective homomorphism extending fk . If k is odd, note that f¯k : Bk −
→ Ak is a bijective
antimonomorphism by Lemma 2.4. Select a point mi , where i is the least number such
that mi ∈
/ dom f¯k ; so Bk ∪ {mi } is a substructure of M. Using the BEP, extend f¯k
′
: Bk ∪ {mi } −
→ M. Again, by restricting the codomain
to an antimonomorphism f¯k+1
′
¯
we obtain a bijective antimonomorphism fk+1 : Bk ∪ {mi } −
→ Ak ∪ {mi f¯k+1
} extending
′
¯
¯
fk ; so fk+1 : Ak ∪ {mi fk+1 } −
→ Bk ∪ {mi } is a bijective homomorphism extending fk .
By ensuring that every mi appears at both an odd and even stage, countably many
applications of this procedure yield a bimorphism α of M extending f .
Remark. We note here that this process of extending one point at a time eventually
creates a bimorphism. In a similar fashion to Dolinka [12, Section 3] we can present
equivalent conditions to Proposition 2.5 using one point extensions.
(1PBEP/1PMEP) Suppose that A ⊆ B ∈ C and there is a (anti)monomorphism
f : A −
→ M. If |B r A| = 1 then there exists a (anti)monomorphism
g:B−
→ M extending f .
A straightforward proof by induction shows that a structure M has both these conditions, if and only if it also has the BEP and MEP. These easier-to-handle properties
are useful in determining whether or not a structure is MB-homogeneous.
We now turn our attention to the actual process of construction. As we also need
to ensure MM-homogeneity, it would make sense to take our “forth” amalgamation
condition to be the mono-amalgamation property introduced in [9].
(MAP) Let C be a class of finite structures. For any A, B1 , B2 ∈ C and
any maps fi : A −
→ Bi (for i = 1, 2) such that f1 is a monomorphism and
f2 is an embedding, there exists C ∈ C and monomorphisms gi : Bi −
→C
(for i = 1, 2) such that f1 g1 = f2 g2 and g1 is an embedding.
Similar to the BEP, we use antimonomorphisms to set out the “back” amalgamation
condition, known as the bi-amalgamation property (BAP).
(BAP) Let C be a class of finite structures. For any elements A, B1 , B2 ∈
C , antimonomorphism f¯1 : A −
→ B1 and embedding f2 : A −
→ B2 , there
exists D ∈ C , embedding g1 : B1 −
→ D and antimonomorphism ḡ2 : B2 −
→D
such that f¯1 g1 = f2 ḡ2 . (see Figure 1.)
We have enough now to prove the first of our propositions that detail our Fraı̈ssé-like
construction.
Proposition 2.6. Let M be an MB-homogeneous structure. Then Age(M) has the
MAP and the BAP.
Proof. As M is MB-homogeneous it is necessarily MM-homogeneous; by Proposition
4.1(b) of [9], Age(M) has the MAP.
Suppose then that A, B1 , B2 are structures in Age(M), and assume we have an
antimonomorphism f¯1 : A −
→ B1 and an embedding f2 : A −
→ B2 . Without loss
of generality, further assume that A, B1 , B2 ⊆ M and that f2 is the inclusion map.
11
g1
∃D
ḡ2
B1
B2
f¯1
f2
A
Figure 1: The bi-amalgamation property (BAP)
From Lemma 2.4, f¯1 induces a bijective monomorphism f1 : Af¯1 −
→ A. As M is MBhomogeneous we extend f1 to some α ∈ Bi(M). Restricting α to B1 yields a bijective
monomorphism α|B1 : B1 −
→ B1 α, where B1 α ⊇ A. Let D be the structure induced by
M on B1 α∪B2 ; it follows that A ⊆ D. As α is surjective, we define C to be the structure
such that Cα = D; set g1 : B1 −
→ C to be the inclusion map. As α : M −
→ M is a
bijective homomorphism, then ᾱ : M −
→ M is an antimonomorphism by Lemma 2.4.
Hence, ᾱ|B2 : B2 −
→ B2 ᾱ ⊆ C and define ḡ2 : B2 −
→ C to be this antimonomorphism.
It is easy to check that f¯1 g1 = f2 ḡ2 and so Age(M) has the BAP.
For our next result, recall that a class of finite structure C has the JEP if for all
A, B ∈ C there exists a D ∈ C such that A, B ⊆ D.
Proposition 2.7. If C is a class of finite relational structures that is closed under
isomorphisms and substructures, has countably many isomorphism types and has the
JEP, MAP and BAP, then there exists an MB-homogeneous structure M such that
Age(M) = C .
Proof. We build M over countably many stages, assuming that Mk has been constructed at some stage k ∈ N. Note that as C has countably many isomorphism types,
we can enumerate them by C = {A0 , A1 , . . .}.
If k ≡ 0 mod 3, select a structure Ai ∈ C , where k = 3i. Use the JEP to find
a structure D that embeds both Mk and Ai ; define Mk+1 to be this structure D. If
k ≡ 1 mod 3, select a triple (A, B, f ) such that A ⊆ B ∈ C and f : A −
→ Mk is a
monomorphism. Using the MAP, we find a structure Mk+1 ∈ C such that there is an
embedding e : Mk → Mk+1 and f extends to some monomorphism g : B −
→ Mk+1 . If
¯
¯
k ≡ 2 mod 3 select a triple (X, Y, f ) such that X ⊆ Y ∈ C and f : X −
→ Mk is an
antimonomorphism. Using the BAP we find a structure Mk+1 such that there is an
embedding e′ : Mk → Mk+1 and f¯ is extended to an antimonomorphism ḡ : Y −
→ Mk+1 .
We can arrange the steps such that:
• every structure A ∈ C appears at some 0 mod 3 stage;
• every triple (A, B, f ) appears at some 1 mod 3 stage, where for every such k, for
every A ⊆ B ∈ C and every monomorphism f : A → Mk , there exists ℓ ≥ k and
embedding ek,ℓ : Mk → Mℓ such that f extends to a monomorphism g : B → Mℓ .
• every triple (X, Y, f¯) appears at some 2 mod 3 stage, where for every such k, for
every X ⊆ Y ∈ C and every antimonomorphism f¯ : X → Mk , there exists ℓ ≥ k
and embedding ek,ℓ : Mk → Mℓ such that f¯ extends to a antimonomorphism
ḡ : Y → Mℓ .
12
S
Following this, define M = k∈N Mk . All that remains to show is that M has age C
and that M is MB-homogeneous.
Our construction ensures that every Ai ∈ C appears at a 0 mod 3 stage, so it
follows that C ⊆ Age(M). Conversely, we have that Mk ∈ C for all k ∈ N; as C is
closed under substructures, Age(M) ⊆ C and so they are equal. Now suppose that
A ⊆ B ∈ C and f : A −
→ M is a monomorphism. From the arrangement of steps
above, there exists a k such that Af ⊆ Mk . From the construction, there exists an
ℓ ∈ N such that Mℓ ⊃ Mk and the monomorphism g : B → Mℓ extends f ; so M has
the MEP. We can use a similar argument to show that M has the BEP in a similar
fashion and so M is MB-homogeneous by Proposition 2.5.
A major consequence of Fraı̈ssé’s original theorem is the fact that any two Fraı̈ssé
limits with the same age are isomorphic. While we cannot guarantee that any two
structures with the same age constructed in the manner of Proposition 2.7 are isomorphic (see Section 3 for some examples) we can provide a uniqueness condition for this
method of construction using a weaker notion of equivalence. Once again, we extend
an idea of [9] to achieve this goal.
Let M, N be two countable structures with the same signature σ. We say that
M, N are bi-equivalent if:
• Age(M) = Age(N ), and;
• every embedding from a finite structure of M into N extends to a bijective
monomorphism α : M −
→ N , and vice versa.
Note that bi-equivalence is an equivalence relation on structures with the same
signature. Two σ-structures M and N can be bi-equivalent without being isomorphic;
see Example 3.11 for more details. We now show this is the relevant equivalence relation
for MB-homogeneity.
Proposition 2.8. 1) Let M, N be two bi-equivalent structures. If M is MB-homogeneous,
so is N .
2) If M, N are MB-homogeneous and Age(M) = Age(N ), then M and N are
bi-equivalent.
Proof. 1) By Proposition 2.5 it suffices to show that N has the MEP and BEP. If M
and N are bi-equivalent, they are certainly mono-equivalent in the sense of Proposition
4.2 of [9]; as M is MM-homogeneous, so is N (by the same result of [9]) and thus N
has the MEP.
Suppose then that A ⊆ B ∈ Age(N ) and there exists an antimonomorphism f¯ :
A −
→ A′ ⊆ N . Note that A need not be isomorphic to A′ . As Age(M) = Age(N )
there exists a copy A′′ of A′ in M; fix an isomorphism e : A′ → A′′ between the
two. Therefore, e is a isomorphism from a finite structure of N into M; as the two
are bi-equivalent, we extend this to a bijective homomorphism α : N → M. This
in turn induces a bijective antimonomorphism ᾱ : M −
→ N by Lemma 2.4. Now,
define an antimonomorphism h̄ = f¯e : A → A′′ ; this is an antimonomorphism from A
into M. Since M is MB-homogeneous, it has the BEP by Proposition 2.5 and so we
extend h̄ to an antimonomorphism h̄′ : B → M. Now, the map h̄′ ᾱ : B −
→ N is a
antimonomorphism; we need to show it extends f¯. So, for all a ∈ A, and using the
facts that α extends e and h̄′ extends h̄ = f¯e, we have that
af¯ = af¯αᾱ = af¯eᾱ = ah̄′ ᾱ.
13
Therefore N has the BEP.
2) Let f : A −
→ B be a bijective embedding from a finite structure of M into N . We
utilise a back and forth argument constructing the bijection over countably many steps;
so set A0 = A, B0 = B and f0 = f . At some stage, we have a bijective monomorphism
fk : Ak −
→ Bk extending f , where Ai ⊆ Ai+1 and Bi ⊆ Bi+1 for all i ≤ k. As
both M and N are countable, we can enumerate the points M = {m0 , m1 , . . .} and
N = {n0 , n1 , . . .}.
If k is even, select a point mi ∈ Mr dom fk , where i is the smallest natural number
such that mi ∈
/ dom fk . We have that Ak ∪ {mi } ⊆ M and so is an element of Age(N )
by assumption. As N is MB-homogeneous it has the MEP and so fk can be extended
′
: Ak ∪ {mi } −
→ N . Restricting the codomain to the image of
to a monomorphism fk+1
′
′
} extending fk .
gives a bijective monomorphism fk+1 : Ak ∪ {mi } −
→ Bk ∪ {mi fk+1
fk+1
¯
If k is odd, note that fk : Bk −
→ Ak is a bijective antimonomorphism. Select a point
nj ∈ N r dom f¯k , where j is the least natural number such that nj ∈
/ im fk . Therefore
Bk ∪ {nj } ⊆ N and is an element of Age(M). As M is MB-homogeneous, use the
′
: Bk ∪ {nj } −
→ M. Restricting the
BEP to extend f¯k to an antimonomorphism f¯k+1
codomain to the image once more delivers the required bijective antimonomorphism
′
f¯k+1 : Bk ∪ {nj } −
→ Ak ∪ {nj f¯k+1
} extending f¯k ; taking the inverse shows that fk+1 is
a bijective homomorphism extending fk as required. Repeating this process countably
many times, ensuring that each mi is in the domain and each nj in the image, provides
a bijective homomorphism α : M −
→ N extending f . The converse direction is entirely
analogous.
We conclude the section by observing that due to Proposition 2.7 and Proposition 2.8,
a MB-homogeneous structure is determined by its age up to bi-equivalence.
3
MB-homogeneous graphs
In this section, we focus on MB-homogeneous graphs in more detail. In this article,
a graph Γ is a set of vertices V Γ together with a set of edges EΓ, where this edge
set interprets a irreflexive and symmetric binary relation E. If {u, v} ∈ EΓ, we say
that u and v are adjacent and write u ∼ v ∈ Γ. If {u, v} ∈
/ EΓ, we say that {u, v}
is a non-edge, that they are non-adjacent, and write u ≁ v. For a graph Γ, define
the complement Γ̄ of Γ to be the graph with vertex set V Γ̄ = V Γ and edges given by
u ∼ v ∈ Γ̄ if and only if u ≁ v ∈ Γ. For n ∈ N ∪ {ℵ0 }, recall that a complete graph
K n on n vertices is a graph on vertex set V K n (with |V K n | = n) with edges given by
u ∼ v ∈ K n if and only if u 6= v ∈ V K n . The complement of the graph K n is called the
null graph on n vertices, and is denoted by K̄ n . Say that ∆ is an induced subgraph of a
graph Γ if V ∆ ⊆ V Γ and u ∼ v ∈ ∆ if and only if u ∼ v ∈ ∆. In this paper, whenever
we say ∆ is a subgraph of Γ, we mean that ∆ is an induced subgraph of Γ. The only
exception is the case of a spanning subgraph; say that ∆ is a spanning subgraph of Γ if
V ∆ = V Γ and u ∼ v ∈ ∆ implies that u ∼ v ∈ Γ.
A logical place to start our investigation is by demonstrating some properties of
MB-homogeneous graphs.
Proposition 3.1. Let Γ be an MB-homogeneous graph. Then its complement Γ̄ is also
MB-homogeneous.
14
Proof. We note that A ∈ Age(Γ) if and only if Ā ∈ Age(Γ̄). Since Γ is MB-homogeneous,
it has the MEP and BEP by Proposition 2.5. Now suppose that A ⊆ B ∈ Age(Γ) and
that f¯ : A −
→ Γ is an antimonomorphism. Any such f¯ preserves non-edges and may
change edges to non-edges; so f¯ : Ā −
→ Γ̄ is a monomorphism. As Γ has the BEP, f¯ can
be extended to an antimonomorphism ḡ : B −
→ Γ; this in turn induces a monomorphism
ḡ : B̄ −
→ Γ̄ and hence Γ̄ has the MEP. The proof that Γ̄ has the BEP is similar.
Following this, we can guarantee that certain subgraphs appear in an MB-homogeneous
graph. Cameron and Nešetřil [9, Proposition 2.5] prove that every MM-homogeneous
graph must contain an infinite complete subgraph; we expand this proposition.
Corollary 3.2. Any infinite non-complete, non-null MB-homogeneous graph Γ contains both an infinite complete and an infinite null subgraph.
Proof. Any MB-homogeneous graph is necessarily MM-homogeneous and hence it contains an infinite complete subgraph from the aforementioned result of [9]. By Proposition 3.1
we have that Γ̄ is also MB-homogeneous and so contains an infinite complete subgraph;
the result follows from this.
A consequence of this argument is that an MB-homogeneous graph Γ is neither a
locally finite graph (where for all v ∈ V Γ, there are only finitely many edges involving
v) nor the complement of a locally finite graph. In fact, we can say more than this.
Recall that the diameter of a graph Γ is the greatest length of the shortest path between
any two points u, v ∈ V Γ. The next result is a restatement of [9, Proposition 1.1(c)];
the proof follows as every MB-homogeneous graph is also an MH-homogeneous graph.
Corollary 3.3 (Proposition 1.1(c), [9]). Suppose that Γ is a connected MB-homogeneous
graph. Then Γ has diameter at most 2 and every edge is contained in a triangle.
We now examine cases where the graph is disconnected (and non-null to avoid
triviality). It is shown in [9] that any disconnected MH-homogeneous graph is a disjoint
union of complete graphs of all the same size. We use this result in conjunction with
Corollary 3.2 to see that the only candidates for a disconnected MB-homogeneous graph
must be disjoint unions of infinite complete graphs. Building on an observation of [31]
that any disconnected MM-homogeneous graph is a disjoint union of infinite complete
graphs, we classify disconnected MB-homogeneous graphs in our next result.
F
Proposition 3.4. Let Γ = i∈I Ki , where Ki ∼
= K ℵ0 for all i in some index set I.
1) If I is finite with size n > 1, then Γ is MM-homogeneous but not MB-homogeneous.
2) If I is countably infinite, then Γ is MB-homogeneous.
Proof. In both cases, we note that every A ∈ Age(Γ) can be decomposed
as finite
F
disjoint union of finite complete graphs; so we can write that A = kj=1 Cj , where Cj
is a complete graph of some finite size.
F
1) As |I| = n we note that k ≤ n for all A = kj=1 Cj in the age of Γ. Suppose that
A ⊆ B ∈ Age(Γ) with B r A = {b}. We have two choices for b; either b is completely
independent of A or b is related to exactly one Cj for some 1 ≤ j ≤ k. If it is the
former, we can extend any monomorphism f : A −
→ Γ to a monomorphism g : B −
→Γ
by sending b to any vertex v ∈ V Γ r Af . If it is the latter, then we can extend any
monomorphism f : A −
→ Γ to a monomorphism g : B −
→ Γ by sending b to a vertex
15
v ∈ Ki r Cj f , where j is defined as above. Hence Γ has the MEP. However, note
that Γ does not embed an independent n + 1 set and so Γ is not MB-homogeneous by
Corollary 3.2.
2) The proof that Γ is MM-homogeneous when I is infinite is as above. Assume then
that A ⊆ B ∈ Age(Γ) with B r A = {b}, and let f¯ : A −
→ Γ be an antimonomorphism.
We note that as A is finite then Af¯ is finite; since I is infinite, there will always exist
i ∈ I such that Ki ∩ Af¯ = ∅. Therefore, regardless of how b is related to A, we can
extend f¯ to an antimonomorphism ḡ : B −
→ Γ by mapping b to some v ∈ Ki , where i
is as stated above. Hence Γ is MB-homogeneous by Proposition 2.5.
Remark.
We note that from Proposition 3.4 and Proposition 3.1 that the complement
F
of i∈N Kiℵ0 , which is the complete multipartite graph with infinitely many partitions
each of infinite size, is also MB-homogeneous.
F
The proof that Γ = i∈N Kiℵ0 is MB-homogeneous relied on the existence of an
independent element to every image of a finite antimonomorphism. Our aim now is
to obtain some sufficient conditions for MB-homogeneity along these lines in order to
construct some new examples.
Definition 3.5. Let Γ be an infinite graph.
• Say that Γ has property (△) if for every finite set U ⊆ V Γ there exists u ∈ V Γ
such that u is adjacent to every member of U .
• Say that Γ has property (∴) if for every finite set V ⊆ V Γ there exists v ∈ V Γ
such that v is non-adjacent to every member of V .
(See Figure 2 for a diagram.)
property (△)
property (∴)
U
V
u
Γ
v
Figure 2: A diagram of Definition 3.5
We note here that if a graph Γ has property (△) then it is algebraically closed (see
[13]); due to the self-complementary nature of these properties, if Γ has property (∴)
then its complement Γ is algebraically closed. These properties prove to be sufficient
for MB-homogeneity.
Proposition 3.6. Let Γ be an infinite graph. If Γ has both properties (△) and (∴)
then Γ is MB-homogeneous.
Proof. Suppose once more that A ⊆ B ∈ Age(Γ) with B r A = {b}, and that f : A −
→Γ
is a monomorphism. As A is finite, Af is a finite set of vertices in Γ and so by property
16
(△) there exists a vertex v of Γ such that v is adjacent to every element of Af . We have
then that v is a potential image point of b, and the map g : B −
→ Γ extending f and
sending b to v is a monomorphism; so Γ has the MEP. Using property (∴) in a similar
fashion shows that Γ has the BEP and so is MB-homogeneous by Proposition 2.5.
Remark. The converse of this result is not true. Proposition 3.4 2) is an example of
an MB-homogeneous graph with property (∴) but not property (△). Its complement
is an example of an MB-homogeneous graph with property (△) but not property (∴).
This shows that any algebraically closed graph Γ whose complement is also algebraically closed is MB-homogeneous. We now present a notion of equivalence that
extends an idea of [13].
Definition 3.7. Let Γ, ∆ be two graphs. We say that Γ is bimorphism equivalent to
∆ if there exist bijective homomorphisms α : Γ −
→ ∆ and β : ∆ −
→ Γ.
Remark. Informally, this definition says that you can start with Γ, and draw some
number (possibly infinite) of extra edges onto Γ to get ∆; from there, you can draw
some number (possibly infinite) of extra edges onto ∆, to get a graph isomorphic to Γ.
This is a weaker version of bi-equivalence introduced in Proposition 2.8. Every pair
of bi-equivalent graphs are bimorphism equivalent by definition, but the converse is
not true; specifically, bimorphism equivalent graphs need not have the same age (see
Corollary 3.10 and Example 3.11). Note that this definition is equivalent to saying
that Γ, ∆ contain each other as spanning subgraphs. Justifying the name, this is an
equivalence relation of graphs (up to isomorphism); we denote this relation by ∼b .
The product αβ : Γ −
→ Γ of the two bijective homomorphisms induces a bimorphism
on Γ, and so if Bi(Γ) = Aut(Γ) we have that α and β are necessarily isomorphisms.
If this occurs, then [Γ]∼b is a singleton equivalence class. We show that bimorphism
equivalence preserves properties (△) and (∴).
Proposition 3.8. Let Γ, ∆ be bimorphism equivalent graphs via bijective homomorphisms α : Γ −
→ ∆ and β : ∆ −
→ Γ. Then Γ has properties (△) and (∴) if and only if ∆
does.
Proof. Suppose that Γ has property (△) and X ⊆ V Γ. From this, there exists a vertex
v ∈ V Γ adjacent to every element of X. As α is a homomorphism, Xα is a finite
subset of V ∆ and vα is adjacent to every element of Xα; since α is bijective, every
finite subset Y of V ∆ can be written as Xα for some X ⊆ V Γ. These observations
show that ∆ has property (△). Using the fact that α is a bijective monomorphism,
by Lemma 2.4 there exists an antimonomorphism ᾱ : ∆ −
→ Γ. Since Γ has property
(∴), for any finite X ⊆ V Γ there exists a vertex w ∈ Γ independent of every element
in X. Due to the fact that ᾱ preserves non-edges, any Y ⊆ V ∆ has a vertex x ∈ V ∆
independent of every element of Y ; and as α is bijective this happens for every finite
set X ⊆ V Γ and so ∆ has property (∴). The converse direction is symmetric.
In fact, we can say more about the connection between bimorphism equivalence
and properties (△) and (∴).
Proposition 3.9. If Γ, ∆ are two graphs with properties (△) and (∴), then Γ and ∆
are bimorphism equivalent.
17
Proof. Assume that Γ, ∆ are two graphs with properties (△) and (∴). We use a back
and forth argument to construct a bijective homomorphism α : Γ → ∆ and a bijective
antimonomorphism β̄ : Γ → ∆, which by Lemma 2.4 will be the converse of a bijective
homomorphism β : ∆ → Γ. Suppose that f : {c} → {d} is a function sending a
vertex c of Γ to a vertex d of ∆; this is a bijective homomorphism. Now set {c} = C0 ,
{d} = D0 , f = f0 , and assume that we have extended f to a bijective homomorphism
fk : Ck → Dk , where Ci and Di are finite and Ci ⊆ Ci+1 and D ⊆ Di+1 for all
0 ≤ i ≤ k − 1. Furthermore, as both Γ and ∆ are countable, we can enumerate their
vertices as V Γ = {c0 , c1 , . . .} and V ∆ = {d0 , d1 , . . .}.
If k is even, select a vertex cj ∈ Γ where j is the smallest number such that cj ∈
/ Ck .
As ∆ has property (△), there exists a vertex u ∈ ∆ such that u is adjacent to every
element of Dk . Define a map fk+1 : Ck ∪{cj } → Dk ∪{u} sending cj to u and extending
f ; this map is a bijective homomorphism as any edge from cj to some element of Ck is
preserved (see Figure 3 for a diagram of an example).
Ck
Dk
cj
fk+1
u
Γ
∆
Figure 3: k even in proof of Proposition 3.9
Now, if k is odd, choose a vertex dj ∈ ∆ where j is the smallest number such that
dj ∈
/ Dk . As Γ has property (∴), there exists a vertex v ∈ Γ such that v is independent
of every element of Ck . Define a map fk+1 : Ck ∪ {v} → Dk ∪ {dj } sending c to dj
and extending fk . Then fk+1 is a bijective homomorphism as fk is and every edge
between c and Ck is preserved; because there are none. See Figure 4 for a diagram of
an example of this stage.
Ck
Dk
v
fk+1
dj
Γ
∆
Figure 4: k odd in proof of Proposition 3.9
18
Repeating this process infinitely many times, ensuring that each vertex of Γ appears
at an even stage and each vertex of ∆ appears at an odd stage, defines a bijective
homomorphism α : Γ → ∆. We can construct a bijective antimonomorphism β̄ :
Γ → ∆ in a similar fashion; replacing homomorphism with antimonomorphism and
using property (∴) of ∆ at even steps and property (△) of Γ at odd steps. So the
converse map β : ∆ → Γ is a bijective homomorphism and so Γ and ∆ are bimorphism
equivalent.
Recall that the random graph R is the countable universal homogeneous graph.
It is well known (see [8]) that R is characterised up to isomorphism by the following
extension property (EP):
(EP) For any two finite disjoint sets of vertices U, V of R there exists x ∈ V R
such that x is adjacent to every vertex in U and non-adjacent to any vertex
in V . (see Figure 5.)
U
R
V
x
Figure 5: (EP) in the random graph R
The next result extends [9, Proposition 2.1(i)] and establishes a complementary
condition to the graph case of [13, Corollary 2.2].
Corollary 3.10. Suppose that Γ is a countable graph. Then Γ has properties (△) and
(∴) if and only if it is bimorphism equivalent to R.
Proof. As R has both properties (△) and (∴), the converse direction follows from
Proposition 3.8, and the forward direction follows from Proposition 3.9.
Remarks. These three results together show that the equivalence class [R]∼b is precisely
the set of all countable graphs Γ with properties (△) and (∴).
In particular, if Γ ∈ [R]∼b and Γ ≇ R (such a Γ exists: see Example 3.11), then Γ
is bimorphism equivalent to R. This means that you can draw infinitely many extra
edges on R (a necessity, see Proposition 2 of [8]) to get an MB-homogeneous graph
Γ ≇ R, and then draw infinitely many more edges on Γ to get a graph isomorphic to
R.
19
3.1
Constructing uncountably many examples
The aim of this subsection is to use these properties of MB-homogeneous graphs in
order to construct uncountably many examples. To begin with, we construct an initial
example of an MB-homogeneous graph that is not homogeneous. This is important; if
we wish to describe 2ℵ0 many examples of MB-homogeneous graphs, then they cannot
also be homogeneous; as there are only countably many non-isomorphic homogeneous
graphs [19].
Example 3.11. Let P = (pn )n∈N0 be an infinite binary sequence with infinitely many
0’s and 1’s. Define the graph Γ(P ) on the infinite vertex set V Γ(P ) = {v0 , v1 , . . .} with
edge relation vi ∼ vj if and only if pmax(i,j) = 0. From this, we can observe that:
• if pi = 0 then vi ∼ vj for all natural numbers j < i;
• if pi = 1 then vi ≁ vj for all j < i;
where < is the natural ordering on N. Say that Γ(P ) is the graph determined by the
binary sequence P . An example (where P = (0, 1, 0, 1, . . .)) is given in Figure 6.
v0
v1
v2
v3
v4
v5
v6
v7
v8
Figure 6: Γ(P ), with P = (0, 1, 0, 1, . . .)
As P has infinitely many of each term, we have that for every finite subsequence
A = {ai1 , . . . , aik } of P there exist natural numbers c, d > ik such that pc = 0 and
pd = 1. This together with the manner of the construction ensures that Γ(P ) has both
properties (△) and (∴) and is therefore MB-homogeneous.
We check that for any sequence P satisfying the above conditions the graph Γ(P )
is not isomorphic to a homogeneous graph by using the classification of countable
homogeneous graphs by Lachlan and Woodrow [19]. As Γ(P ) has both an edge and a
non-edge it is neither K ℵ0 nor its complement K̄ ℵ0 . Since Γ(P ) is connected, it is not a
disjoint union of complete graphs; as it has property (∴), it is not the complement of a
disjoint union of complete graphs. From Corollary 3.2, as Γ(P ) is neither complete nor
null it contains both an infinite complete subgraph and an infinite null subgraph; so it
is not isomorphic to the Kn -free graph Hn or its complement H̄n for any n. Finally,
we note that as no term p of P can be both 0 and 1 simultaneously, there is no vertex
vp that is adjacent to {v0 } and non-adjacent to {v1 }. Hence Γ(P ) does not satisfy the
extension property characteristic of the random graph. This accounts for all countable
graphs in the classification; so Γ is not a homogeneous graph.
20
Remarks. Let P, Q be two infinite binary sequences with infinitely many 0’s and 1’s.
Any such infinite binary sequence contains every finite binary sequence X as a subsequence. The finite induced subgraphs Γ(X) of Γ(P ) are those induced on V Γ(X) by the
edge relation of P . As this is true for all such binary sequences P and Q, we conclude
that Γ(P ) and Γ(Q) have the same age. It can be shown from here that for any two
such sequences P and Q, then Γ(P ) and Γ(Q) are bi-equivalent (see Proposition 2.8).
Throughout the rest of this section, any binary sequences P, Q have infinitely many
0’s and 1’s. This guarantees that any graph Γ(P ) determined by P has properties (△)
and (∴).
Many natural questions arise from this construction. Perhaps the most pertinent of
these is: to what extent does Γ(P ) depend on the binary sequence P ? Answering this
question will tell us exactly how many new MB-homogeneous graphs there are of this
kind. We achieve a solution by investigating the automorphism group as an invariant of
Γ(P ). Our first lemma establishes a convention for the zeroth place of such a sequence.
We denote the neighbourhood set of a vertex v ∈ V Γ by N (v).
Lemma 3.12. Suppose that Γ(P ) and Γ(Q) are the graphs on the vertex sets V =
{v0 , v1 , . . .} and W = {w0 , w1 , . . .}, with edges determined by the binary sequences
P = (pn )n∈N0 and Q = (qn )n∈N0 respectively, and assume that pi = qi for all i > 0.
Then Γ(P ) ∼
= Γ(Q).
Following this, we can take p0 = p1 for any binary sequence P without loss of
generality; we adopt this as convention for the rest of the section. Now, if P is a binary
sequence, denote the kth consecutive string of 0’s and 1’s by Ok and Ik respectively,
and denote the vertex sets corresponding to these subsequences by V Ok and V Ik (see
Figure 7). Furthermore, denote the graph induced by Γ(P ) on any subset V X of V Γ
by Γ(X).
1
1
0
I1
v0
1
O1
v1
V I1
0
v2
1
0
I2
v3
V O1
1
v4
0
O2
v5
v6
V I2
v7
v8
V O2
1
0
I3
O3
v9
v10
V I3
V O3
Figure 7: Γ(P ), with P = (1, 1, 0, 0, 1, 1, 1, 0, 0, 1, 0, . . .)
21
...
Lemma 3.13. Let P = (pn )n∈N0 be a binary sequence, and let Γ(P ) be the graph
determined by this binary sequence. Suppose that vj is a vertex in V In for some n ∈ N.
Then Γ(N (vj )) ∼
= K ℵ0 .
Proof. From the assumption that pj = 1, the observation of Example 3.11, and the
definition of an edge in Γ(P ), we have that vj ∼ vk if and only if j < k and pk = 0;
so N (vj ) = {vk : k > j, pk = 0}. As there are infinitely many 0’s in P , this set is
infinite as j is finite. Now, take va , vb ∈ N (vj ). Here, va ∼ vb ∈ Γ(N (vj )) if and only if
pmax(a,b) = 0; but pa = pb = 0 and so any two vertices of N (vj ) are adjacent.
Lemma 3.14. Let P = (pn )n∈N0 be a binary sequence. Suppose that v, w are vertices in
some V Ok = {vk1 , . . . , vkn } corresponding to some Ok = {pk1 , . . . , pkn }. Then N (v) ∪
{v} = N (w) ∪ {w} and Γ(N (v)) ∼
= Γ(N (w)).
Proof. Define the sets X = {vj : j < i1 } and Y = {vk : k ≥ i1 , ak = 0}. Due to the
construction of Γ in Example 3.11, we have that N (v) ∪ {v} = X ∪ Y = N (w) ∪ {w}.
For all u ∈ V Γ(V Ok ), it is easy to see that N (u) = X ∪ (Y r {u}). Using a similar
argument to the proof of Lemma 3.13, we have that Γ(Y r {u}) is an infinite complete
graph for any u ∈ Γ(V Ok ), and every element in Y is adjacent to every element of
X. For any v, w ∈ V Ok define a map f : X ∪ (Y r {v}) −
→ X ∪ (Y r {w}) fixing X
pointwise and sending (Y r {v}) to (Y r {w}) in any fashion; this is an isomorphism
between Γ(N (v)) and Γ(N (w)).
Before our next proposition, recall that if there exists γ ∈ Aut(Γ) such that vγ = w,
then the graphs induced on N (v) and N (w) are isomorphic. Furthermore, recall that
an independent set U of a graph Γ is a maximum independent set if there does not
exist a subset W of Γ such that W is a independent set with |W | > |U |.
Proposition 3.15. Let P = (pn )n∈N0 be a binary sequence. Then
Y
Aut(Γ(P )) =
(Aut(Γ(Ok )) × Aut(Γ(Ik )))
i∈N
the infinite direct product of automorphism groups on each Γ(Ok ) and Γ(Ik ).
Proof. Using Lemma 3.12, we set the convention that p0 = p1 throughout; hence either
O1 or I1 (depending on whether P starts with a 0 or 1) has size at least 2. For some
i, write Oi = {pi1 , . . . , pin } and Ii = {qi1 , . . . , qim }, and we write V Oi = {vi1 , . . . , vin }
and V Ii = {wi1 , . . . , wim }.
We first show that any automorphism of Γ(P ) fixes both V Ok and V Ik setwise for
all k ∈ N, using a series of claims. Then, we prove that any bijective map from Γ(P )
to itself fixing every point except those in a single V Ok (or V Ik ) and acting as an
automorphism on that V Ok (or V Ik ) is an automorphism of Γ(P ). We begin with our
first claim.
Claim 1. If v ∈ V Oi and w ∈ V Oj with i 6= j, then Γ(N (v)) ≇ Γ(N (w)).
Proof of Claim 1. Without loss of generality, suppose that i < j and so i1 < i2 < . . . <
in < j1 < j2 < . . . < jm . We define the following sets:
Xi = {vk : k < i1 }
Xj
= {vk : k < j1 }
Yi = {vk : k ≥ i1 , pk = 0}
Yj = {vk : k ≥ j1 , pk = 0}.
22
Lemma 3.14 asserts that N (v) = Xi ∪ (Yi r {v}) and N (w) = Xj ∪ (Yj r {w}). By
reasoning outlined in Lemma 3.14, if va ≁ vb ∈ Γ(N (v)) then va , vb ∈ Xi . Note
also that both Xi and Xj are both finite sets, and so any maximum independent set
of Xi , Xj (and hence of N (v), N (w)) must be finite. Consider the sets A = {va :
a < i1 , pa = 1} ∪ {v0 } and B = {vb : b < j1 , pb = 1} ∪ {v0 }, contained in Xi and Xj
respectively. So if va , vd ∈ A, then pmax(a,d) = 1 and so va ≁ vd . Now, if vc ∈ Xi r A
then pc = 0, and as vc 6= v0 it follows that pmax(c,0) = 0 and so vc ∼ v0 . Hence A
is the maximum independent set in Xi ; using a similar argument, we can show that
B is the maximum independent set in Xj . Since i < j there exists an Ik between
Oi and Oj ; so there is a e ∈ N where in < e < j1 and pe = 1. Hence ve ∈ B r A
and so |B| > |A|. Therefore, as Γ(N (v)) and Γ(N (w)) have different sizes of maximum independent sets, they are not isomorphic and this concludes the proof of Claim 1.
This shows that there is no automorphism α of Γ(P ) sending a vertex v ∈ V Oi to a
vertex w ∈ V Oj where i 6= j, as they have non-isomorphic neighbourhoods. We move
on to our next claim.
Claim 2. Suppose that v ∈ V Oi (with i ≥ 2) and w ∈ V Ij . Then there is no
automorphism of Γ(P ) sending v to w.
Proof of Claim 2. Define the set A = {va : a < i1 , pa = 1} ∪ {v0 } as in the previous
claim. As i ≥ 2, there exists a c such that 0 < c < i1 with pc = 1. Hence we have
that |A| ≥ 2 and so Γ(N (v)) has at least one non-edge. However, Γ(N (w)) ∼
= K ℵ0 by
Lemma 3.13 and so there cannot possibly be an automorphism sending v to w. This
proves the second claim.
From these claims, we see that any automorphism α of Γ(P ) fixes V Oi setwise for
i ≥ 2. Now, for k ≥ 2, any V Ik (or V Ik+1 ) sandwiched between V Ok and V Ok+1 , as
V Ik (or V Ik+1 ) are the only vertices not adjacent to every vertex in V Ok and adjacent
to every vertex in V Ok+1 . As V Ok and V Ok+1 are fixed setwise, we conclude that V Ik
(or V Ik+1 ) is fixed setwise for k ≥ 2.
Claim 3. V O1 and V I1 are fixed setwise by any automorphism of Γ(P ).
Proof of Claim 3: There are two cases to consider; where P either begins with a 0 or
a 1.
Case 1 (p0 = 0). Suppose for a contradiction that there exists an α ∈ Aut(Γ) sending
a v ∈ V O1 to a w ∈ V I1 . Due to our case we have |V O1 | ≥ 2 and so Γ(V O1 ) has an
edge from v to some u ∈ V O1 . But w ∈ V I1 and thus is independent of everything
in V O1 ∪ V I1 ; so the edge between v and u is not preserved by α and so V O1 is fixed
setwise. From this, we can use a similar argument to above to show that V I1 is fixed
setwise.
Case 2 (p0 = 1). From this assumption, we have that |V I1 | ≥ 2 and so Γ(V I1 )
contains a non-edge. Hence for some v ∈ V O1 , the graph Γ(N (v)) contains a nonedge. Therefore, from Lemma 3.13 and Claim 1, Γ(N (v)) ≇ Γ(N (w)) for some w ∈
V Γ(P ) r V O1 and so V O1 is fixed setwise. As this happens, both V I1 and V I2 are
fixed setwise. This concludes the proof of Claim 3.
23
So for any k ∈ N, we have that V Ok and V Ik are fixed setwise by any automorphism of
Γ(P ). It remains to prove that any automorphism of Γ(V Ok ) or Γ(V Ik ) whilst fixing
every other point in Γ(P ) is an automorphism of Γ(P ). Here, as Γ(V Ok ) is a complete
graph and Γ(V Ik ) is a null graph, it follows that Aut(Γ(V Ok )) ∼
= Sym(|V Ok |) and
∼
Aut(Γ(V Ik )) = Sym(|V Ik |) for all k ∈ N.
So suppose that f : Γ(P ) −
→ Γ(P ) is a bijective map that acts as an automorphism on some Γ(V Ok ) and fixes V Γ r V Ok . By Lemma 3.14 any two elements
v, w ∈ V Ok have the same extended neighbourhood N (v) ∪ {v} = N (w) ∪ {w} (and
Γ(N (v)) ∼
= Γ(N (w))). It follows that f preserves all edges and non-edges between V Ok
and V Γ(P ) r V Ok and so it is an automorphism of Γ(P ). The proof that the bijective
map g : Γ(P ) −
→ Γ(P ) that acts as an automorphism of some Γ(V Ik ) and fixes every
other point is similar.
This proposition guarantees the existence of countably many MB-homogeneous
graphs. For instance, define a binary sequence P n = (pi )i∈N by the following:
(
1 if i = 0, 1, 2, . . . , n − 1, n + 1, n + 3, . . .
pi =
0 if i = n, n + 2, . . .
Here, P n is a sequence of n many 1’s followed by alternating 0’s and 1’s. It follows that
Aut(Γ(P n )) is the infinite direct product of one copy of Sym(n) together with infinitely
many trivial groups; so Aut(Γ(P n )) ∼
= Sym(n). This is true for all n ≥ 2. So if m 6= n,
then Aut(Γ(P m )) ≇ Aut(Γ(P n )) and so Γ(P m ) ≇ Γ(P n ).
Furthermore, it is straightforward to see that if two monoids T, T ′ have nonisomorphic groups of units U, U ′ respectively, then T ≇ T ′ . Hence we have constructed
several examples of non-isomorphic oligomorphic permutation monoids on bi-equivalent
structures. By Proposition 1.6, this means that each of these oligomorphic permutation
monoids have the same strong orbits.
We now aim to use the basic framework established here to construct 2ℵ0 nonisomorphic examples of MB-homogeneous graphs. To achieve this, the idea is to add in
pairwise non-embeddable finite structures into the age of some Γ(P ) to ensure uniqueness up to isomorphism. To this end, let A = (an )n∈N0 be a strictly increasing sequence
of natural numbers. We use A to recursively define a binary sequence P A = (pi )i∈N as
follows;
Base: 0 followed by a1 many 1’s.
Inductive: Assuming that the nth stage of the sequence has been constructed, add a
0 followed by an+1 many 1’s to the right hand side of the sequence.
For instance, if A = (2, 3, 5, . . .) then P A = (0, 1, 1, 0, 1, 1, 1, 0, 1, 1, 1, 1, 1, . . .). Using
such a binary sequence P A, we construct Γ(P A) in the fashion of Example 3.11. As
P A has infinitely many 0’s and 1’s then Γ(P A), the graph determined by P A, is MBhomogeneous. The eventual plan is to induce finite graphs onto the independent sets
induced on Γ(P A) by strings of consecutive 1’s in P A. By selecting a suitable countable
family of pairwise non-embeddable graphs that do not appear in the age of Γ(P A), we
ensure graphs of different ages.
We prove two lemmas which form the basis for our construction of 2ℵ0 non-isomorphic
examples of MB-homogeneous graphs. The second lemma is folklore; and the proof is
omitted.
24
Lemma 3.16. Let P be any binary sequence. Then Γ(P ) does not embed any cycle
graph of size m ≥ 4.
Proof. Let V X = {vi1 , . . . , vim } ⊆ V Γ(P ), with m ≥ 4. Let X be the graph on
V X with edges induced by Γ(P ), and suppose X is an m-cycle. As each vertex in
a cycle graph has degree 2, we see that pij 6= 0 for ij > i3 . But this means that
pi4 = . . . = pim = 1 and so vi4 has degree 0 in X. This is a contradiction and so X is
not an m-cycle.
Lemma 3.17. Let Cm and Cn be two cycle graphs with m, n ≥ 3. Then Cm embeds in
Cn (and vice versa) if and only if m = n; in which case they are isomorphic.
Hence C = {Cn : n ≥ 4} is a countable family of pairwise non-embeddable graphs
such that Age(Γ(P A)) ∩ C = ∅. Now suppose that A = (an )n∈N is a strictly increasing
sequence of natural numbers with ai ≥ 4; construct Γ(P A) in the usual fashion. We
have that the size of each Ik in P A is ak . For each independent set Γ(IK ) ⊆ Γ(P A),
draw an ak -cycle on its vertices, thus creating a new graph Γ(P A)′ (see Figure 8).
Figure 8: Γ(P A)′ corresponding to the sequence A = (4, 5, 6, . . .), with added cycles
highlighted in red
In particular, note that even with these additional structures, Γ(P A) is still MBhomogeneous as it has properties (△) and (∴). Since we have added so many structures
to the age, we must be careful that we have not added extra cycles of sizes not expressed
in the sequence A. The next proposition alleviates this concern.
Proposition 3.18. Suppose that A = (an )n∈N is a strictly increasing sequence of
natural numbers with a1 ≥ 4 and suppose that m ≥ 4 is a natural number such that
m 6= an for all n ∈ N. Then the graph Γ(P A)′ outlined above does not contain an
m-cycle.
Proof. Suppose that M ⊆ Γ(P A)′ is an m-cycle. Then the edge set of M is a combination of the edges of the graph induced by the finite subsequence Q = (qi1 , . . . , qim )
of P A (where i1 < i2 < . . . < im ) on V M = {vi1 , . . . , vim } and the edges from cycles
added onto Γ(P A). We aim to show that M = Γ(In ) for some n ∈ N; that is, the
only cycles of size ≥ 4 in Γ(P A)′ are precisely those we added in the construction. So
assume here that Q contains a 0. As M is an m-cycle, dM (v) = 2 for all v ∈ M and
so the only elements of Q that can be 0 are qi1 , qi2 and qi3 . We split our consideration
into cases.
25
Case 1 (qi3 = 0). As we assume this, vi3 is adjacent to both vi1 and vi2 . If qi2 = 0,
then vi2 ∼ vi1 creating a 3-cycle; this is a contradiction as M does not embed a 3-cycle.
Therefore, qi2 = 1 and so vi1 is adjacent to some vij where 3 < j ≤ m. Since qij = 1 for
all ij > i3 , it follows that the edge between vi1 and vij was induced by an added cycle.
This implies that vi1 , vij ∈ V Γ(Ik ) for some k. Therefore, (bi1 , . . . , bij ) is a sequence
of 1’s where i1 < i2 < . . . < ij are consecutive natural numbers; a contradiction as
i1 < i3 < ij and so 1 = qi3 = 0. Hence, qi3 6= 0.
Case 2 (qi2 = 0). Since this happens, vi2 is adjacent to vi1 and some vertex vij 6= vi1 ;
so ij > i2 . If qij is 1, then vij is non-adjacent to any vertex vik with ik < ij and qik = 0,
as the construction of Γ(P A)′ ensures that no edges are drawn in this case. But this is
a contradiction as qi2 = 0 and vij ∼ vi2 . So qij must be 0 in this case; but ij > i2 and
so qij = 1 by Case 1 and the argument preceding Case 1. This is a contradiction and
so qi2 = 1.
A similar argument to that of Case 2 holds for when qi1 = 0 and so no element of
Q is 0. Hence, Q is made up of 1’s; however, it is still a possibility that these originate
from from different Ik ’s. We now show however that every vertex of M comes from a
single V Ik for some k. As no element of Q is 0, we conclude that two vertices in M
have an edge between them only if they are contained in the same V Ik and have an
edge between them in Γ(Ik ). As M is connected, we have that M ⊆ Γ(Ik ) for some k.
Finally, as Γ(Ik ) is an ak -cycle embedding an m-cycle, we are forced to conclude that
m = ak by Lemma 3.17 and so we are done.
Corollary 3.19. Suppose that A = (an )n∈N and B = (bn )n∈N are two different strictly
increasing sequences of natural numbers with a1 , b1 ≥ 4. Then Γ(P A)′ ≇ Γ(P B)′ .
Proof. As A and B are different sequences there exists a j ∈ N such that aj 6= bj ; without loss of generality assume that aj < bj . Hence Γ(P A)′ embeds an aj -cycle; but as
aj ∈
/ B, by Proposition 3.18 Γ(P B)′ does not embed an aj -cycle. Hence Age(Γ(P A)′ ) 6=
Age(Γ(P B)′ ) and so they are not isomorphic.
These results prove the following:
Theorem 3.20. There are 2ℵ0 many non-isomorphic, non-bi-equivalent MB-homogeneous
graphs, each of which is bimorphism equivalent to the random graph R.
Proof. As there are 2ℵ0 strictly increasing sequences of natural numbers we have continuum many examples of Γ(P A)′ by Corollary 3.19. As these graphs have different
ages, this means we have constructed 2ℵ0 many non-isomorphic, non-bi-equivalent MBhomogeneous graphs. Furthermore, as each these examples has property (△) and (∴),
they are bimorphism equivalent to R by Corollary 3.10.
Remark. As a consequence of this, |[R]∼b | = 2ℵ0 . This means that there are 2ℵ0 pairwise
non-isomorphic graphs Γ with the property stated in the remark after Corollary 3.10.
Finally in this section, we utilise this technique of overlaying finite graphs in order
to prove the second main theorem of the section. Recall that a graph Γ is n-regular if
every v ∈ V Γ has degree n for some n ∈ N.
Theorem 3.21. Any finite group H arises as the automorphism group of an MBhomogeneous graph Γ.
26
Proof. By a version of Frucht’s theorem [15] there exists countably many 3-regular
graphs Γ such that Aut(Γ) ∼
= H; and so, as there are only finitely many graphs of
size less than or equal to 5, there exists a 3-regular graph ∆ of size n ≥ 6 such that
Aut(∆) ∼
= H. By the handshake lemma (see p5 [11]), such a graph must have a total
of 3n/2 edges out of a total of (n2 − n)/2 possible edges; as n ≥ 6, this means that ∆
must induce at least 6 non-edges.
Define a binary sequence P = (pi )i∈N by the following:
(
1 if i = 0, 1, 2, . . . , n − 1, n + 1, n + 3, . . .
pi =
0 if i = n, n + 2, . . .
So P is a sequence of n many 1’s followed by alternating 0’s and 1’s. Using the notation
established above for Lemma 3.12, it follows that |I1 | = n and |Ok | = |Im | = 1 for k ≥ 1
and m ≥ 2. Construct Γ(P ) on V Γ(P ) = {v0 , v1 , . . .} as illustrated in Example 3.11,
and draw in edges on V I1 such that Γ(V I1 ) ∼
= ∆ to obtain a graph Γ(P )′ (see Figure 9).
v0
v1
v2
vn−2
vn−1
vn
vn+1
vn+2
vn+3
vn+4
Figure 9: Γ(P )′ , with ∆ highlighted in red
In a similar fashion to Proposition 3.15, we aim to show that V Oi and V Ii are fixed
setwise for any automorphism α of Γ(P )′ through a series of claims.
Claim 1. If va ∈ V Oi and vb ∈ V Oj with i 6= j, then Γ(N (v)) ≇ Γ(N (w)).
Proof of Claim 1. As |V Ok | = 1 for all k ∈ N, we have that V Oi = {va } and V Oj =
{vb }. Assume without loss of generality that a < b. We define the following sets:
Xa = {vk : k < a}
Xb = {vk : k < b}
Ya = {vk : k ≥ a, pk = 0}
Yb = {vk : k ≥ b, pk = 0}.
Lemma 3.14 applies in this situation; so N (va ) = Xa ∪ Ya and N (vb ) = Xb ∪ Yb .
As in the proof of Proposition 3.15, any maximum independent set of N (va ), N (vb ) is
contained in Xa , Xb respectively. Now, as Γ(V I1 ) ∼
= ∆ is a 3-regular graph on more
than six vertices, there exists some maximal independent set M ⊆ V I1 of Γ(V I1 ) with
size greater than or equal to 2.
27
Now we consider the sets
A = {vc ∈ V Γ(P )′ : n − 1 < c < a, pc = 1} ∪ M
and
B = {vc ∈ V Γ(P )′ : n − 1 < c < b, pc = 1} ∪ M,
the maximum independent sets of Xa and Xb respectively. As i < j, there exists d
such that a < d < b with pd = 1. Hence vd ∈ B r A and so |B| > |A|. Since A, B are
maximum independent sets of Γ(N (va )) and Γ(N (vb )) respectively with different sizes,
we conclude that Γ(N (va )) ≇ Γ(N (vb )). This ends the proof of Claim 1.
This shows that there exists no automorphism γ of Γ(P )′ sending any v ∈ V Oi to
w ∈ V Oj with i 6= j.
Claim 2. There exists no automorphism sending v ∈ V Ok to w ∈ V Im for all k, m ∈ N.
Proof of Claim 2. We split the proof into two cases; where m = 1 and where m ≥ 2.
For the latter, Γ(N (w)) ∼
= K ℵ0 for any w ∈ V Im with m ≥ 2 by Lemma 3.13. But
as Γ(V I1 ) is not a complete graph, we have that Γ(V Ok ) contains a non-edge for all
k ∈ N and so Γ(N (v)) ≇ Γ(N (w)) in this case. It remains to show that there is no
automorphism sending v ∈ V Ok to w ∈ V I1 . In this case Γ(N (w)) is the union of an
infinite complete graph K and G = Γ(NΓ(V I1 ) (w)), with every vertex of K connected
to every vertex of G. This means any non-edge of Γ(N (w)) must be induced by G;
as |NΓ(V I1 ) (w)| = 3, there are at most 3 of them for any w ∈ V I1 . However, as Xa
contains V I1 , we have that Γ(N (v)) contains ∆ as an induced subgraph. By the reasoning above, ∆ has at least 6 non-edges and therefore so does Γ(N (v)). This means
that Γ(N (v)) ≇ Γ(N (w)) for any v ∈ V Ok and any w ∈ V I1 and so we are done.
Here, V Ok is fixed setwise for all k ∈ N; as |V Ok | = 1 for all such k we have that they
are also fixed pointwise. As in the proof of Proposition 3.15, we have that any V Ik
sandwiched between V Ok−1 and V Ok are the only vertices not adjacent to every vertex
in V Ok and adjacent to every vertex in V Ok+1 . As V Ok and V Ok+1 are fixed setwise,
we deduce that V Ik is fixed setwise (and hence pointwise) for k ≥ 2. We conclude that
V I1 is fixed setwise under automorphisms of Γ(P )′ .
Finally, we show that any bijective map γ : Γ(P )′ → Γ(P )′ acting as an automorphism on V I1 and fixing everything else is an automorphism of Γ(P )′ . As every
v ∈ V Ok for all k is connected to each u ∈ V I1 and every w ∈ V Im for all m is
independent of each u ∈ V I1 , we have that this map preserves all edges and non-edges
of Γ(P )′ and so is an automorphism of Γ(P )′ .
Using this together with Theorem 1.7, it follows that for any finite group U there
exists an oligomorphic permutation monoid that has U as a group of units.
4
Worked examples
This section is devoted to determining MB-homogeneity of examples of homogeneous
structures. This investigation concludes in a complete classification of countable homogeneous graphs that are also MB-homogeneous.
28
There are a variety of ways to demonstrate that a structure M is MB-homogeneous.
Firstly, if M is a homogeneous structure where each finite partial monomorphism h is
also a finite partial isomorphism, then M is MB-homogeneous as we can extend h to
an automorphism. If M is a graph, then it suffices to show that M has properties (△)
and (∴) by Proposition 3.6. Finally, recall the remarks of Proposition 2.5; to prove
that M is MB-homogeneous it suffices to show that M has both the 1PBEP and
the 1PMEP. All of these techniques are used at points in demonstrating the following
positive examples.
Example 4.1. Let Γ = K ℵ0 , the complete graph on countably many vertices; this
graph is homogeneous [19]. Suppose that h : A −
→ B is a monomorphism between
two finite substructures of Γ. As there are no non-edges to preserve, it must preserve
non-edges and so h is a finite partial isomorphism. Using homogeneity of Γ, we extend
h to an automorphism (and hence a bimorphism) of Γ and so Γ is MB-homogeneous.
Its complement Γ̄, the infinite null graph, is also MB-homogeneous by Proposition 3.1.
Example 4.2. A tournament is defined to be an oriented, loopless complete graph. By
a similar argument to the complete graph in Example 4.1, every finite partial monomorphism of a tournament is a finite partial isomorphism. So the three homogeneous tournaments (Q, <), the random tournament T and the local order S(2) (see Cherlin [10])
are all MB-homogeneous.
Example 4.3. Let R be the countable universal homogeneous undirected graph, also
known as the random graph. It follows from the extension property characteristic to
the random graph (see Figure 5) that R has properties (△) and (∴) (see Definition 3.5)
and so is an MB-homogeneous graph by Proposition 3.6.
Example 4.4. Let D be the Fraı̈ssé limit of the class of directed graphs without loops
or 2-cycles, otherwise known as the generic digraph. It is well known that D satisfies
the following extension property (see [1]).
(DEP) For any finite and pairwise disjoint sets of vertices U, V, W of D,
there exists a vertex x of D such that there is an arc from x to every
element of U , an arc to x from every element of V , and x is independent of
every vertex in W . (see Figure 10 for a diagram of an example.)
Using this, we show that D is MB-homogeneous by demonstrating that it has the
1PMEP and 1PBEP in turn.
Suppose that A ⊆ B ∈ Age(D) with B r A = {b} and that f : A −
→ D is a
→
−
monomorphism. Decompose A into three disjoint sets b
= {a ∈ A : b −
→ a},
− = {a ∈ A : b ←
b←
− a} and bk = {a ∈ A : b k a}. The fact that f is injective means
→ f, b←
− f and bk f are pairwise disjoint subsets of V D. Using the DEP,
that the sets b−
→ f , an arc from all
select a vertex x ∈ V D such that x has an arc to all elements of b−
←
−
elements of b f and is independent of all elements of b k f . Define g : B −
→ D to be
the map such that bg = x and g|A = f ; due to our choice of x, this is a monomorphism.
So D has the 1PMEP.
Now suppose that X ⊆ Y ∈ Age(D) with Y r X = {y} and that f¯ : X → D is an
antimonomorphism. The fact that X is finite implies that im f¯ is finite, and so there
exists a vertex w ∈ V D such that w is independent of all elements in X f¯ by the DEP.
29
U
V
x
W
D
Figure 10: Example of directed extension property in D
Define ḡ : Y → D to be the function such that yḡ = w and ḡ|X = f¯; thanks to our
choice of w, this is an antimonomorphism as all non-arcs are preserved. So D has the
1PBEP and hence D is MB-homogeneous by Proposition 2.5.
Remark. It can be shown that the class of loopless digraphs with 2-cycles is a Fraı̈ssé
class and hence there exists a unique homogeneous generic digraph with 2-cycles D ∗ ,
with a slightly different extension property (see Ch. 4 of [26]). With this, we can use
a similar proof to above to show that D ∗ is MB-homogeneous.
We continue our investigation by focusing on a class of structures known as cores;
a structure M is a core if every endomorphism of M is an embedding. Every ℵ0 categorical structure has a core, and is homomorphically equivalent to a model-complete
core [3]. Cores play an important role in the theory of constraint satisfaction problems;
see Bodirsky’s habilitation thesis [4] for an introduction to the topic. Widely studied
examples of cores include the countable dense linear order without endpoints (Q, <), the
complete graph on countably many vertices K ℵ0 , and its complement K̄ ℵ0 . Proving
that a structure M is a core is a useful way to show that a structure is not MBhomogeneous, as this next result shows.
Lemma 4.5. Let M be a core such that there exists a finite partial monomorphism
of M that is not an isomorphism. Then M is not MH-homogeneous (and hence, not
MB-homogeneous).
Proof. Let h be a finite partial monomorphism of a core M that is not an isomorphism.
As any endomorphism of M is an embedding, we cannot extend h.
We can use this result to detail some homogeneous structures that are not MBhomogeneous.
Example 4.6. Let G be the countable homogeneous Kn -free graph for n ≥ 3. By
results of Mudrinski in [28], it is shown that G is a core for all such n. As there are
finite partial monomorphisms of G that are not isomorphisms (send any non-edge to
an edge, for instance) we have that G is not MB-homogeneous by Lemma 4.5.
30
Example 4.7. The oriented graph analogues of the homogeneous Kn -free graphs are
the Henson digraphs MT . These are the Fraı̈ssé limits of the class of all digraphs not
embedding elements of some set T of finite tournaments. We show that MT is a core
for any non-empty T containing tournaments of three or more vertices.
Suppose for a contradiction there exists a γ ∈ End(MT ) such that for some independent pair of vertices v, w we have that vγ = wγ. Select a tournament Y ∈ T with
the least number of vertices, and choose x, y ∈ Y such that x −
→ y. Create a oriented
graph Y ′ by removing the arc between x and y, adding an extra vertex x′ and drawing
an arc x′ −
→ y; see Figure 11 for an example.
Y′
Y
x
y
x
y
x′
Figure 11: Construction of Y ′ from Y in Example 4.7
Note that there is no tournament on any |Y |-set of vertices of Y ′ ; so Y ′ ∈ Age(MT ).
By homogeneity of MT , we find a copy Y ′′ of Y ′ with v, w in place of x, x′ respectively,
and a vertex u in place of y. The image of Y ′′ under γ is a oriented graph on |Y | vertices
with vγ → uγ, that preserves all arcs involving v and u. It follows that Y ′′ γ ∼
= Y ; this
is a contradiction as Y does not belong to the age of MT . So γ must be injective.
Now assume that there exists an independent pair of vertices v, w ∈ MT such that
vγ −
→ wγ. Select a Y ∈ T in the same fashion as before, choose two vertices x −
→ y of Y
and remove this arc to obtain a oriented graph Z that embeds in MT . Via homogeneity
of MT , we find an isomorphic copy of Z with v, w in place of x, y respectively. Hence
the image of Z under γ induces a copy of Y in MT ; a contradiction as Y does not
belong to the age of MT . So MT is a core for any such set of tournaments T ; therefore
MT is not MB-homogeneous by Lemma 4.5.
Example 4.8. Let S(3) be the myopic local order, defined as follows. Distribute ℵ0
many points densely around the unit circle such that for every point a there are no
→ b if and
points b, c such that arg(a, b) = arg(b, c) = arg(c, a) = 2π
3 . Draw an arc a −
only if arg(a, b) < 2π/3; note that this means that S(3) embeds no directed 3-cycles.
We show that this structure is a core.
Assume that there is an endomorphism γ of S(3) with aγ −
→ bγ for some independent pair of points a, b ∈ S(3). As this occurs, we have that both arg(a, b) and
arg(b, a) > 2π/3. As arg(a, b) = 2π − arg(b, a), it follows that arg(b, a) < 4π/3. From
this, there exists a point c such that arg(b, c) = arg(c, a) < 2π/3 and so b −
→ c and
c −
→ a. The endomorphism γ then creates a directed 3-cycle (or a loop) and this is
a contradiction. Now suppose that for some non-related pair a, b ∈ S(3), there is an
endomorphism γ such that aγ = bγ. As before, we can find a point c ∈ S(3) such
that b −
→ c and c −
→ a. As γ is an endomorphism it must preserve these relations
and so bγ −
→ cγ and cγ −
→ aγ = bγ. Hence we have a directed 2-cycle and this is
31
obviously false; so S(3) is a core. Applying Lemma 4.5 again implies that S(3) is not
MB-homogeneous.
Finally, we note that as Proposition 2.5 is an if and only if statement, it is enough
to prove that a structure M does not have the MEP or the BEP to show that M is
not MB-homogeneous.
Example 4.9. Let Dn be the generic In -free digraph for n ≥ 3; this is a homogeneous
digraph [10]. We show here that Dn does not have the BEP, and hence is not MBhomogeneous by Proposition 2.5 So let T be some tournament on n − 1 vertices, and
let U be the disjoint union of T with a vertex u. We note that T ⊆ U ∈ Age(Dn ).
Let f¯ : T −
→ Dn be an antimonomorphism sending T to an independent set of n − 1
vertices in Dn ; such a substructure exists by construction. Then as antimonomorphisms
preserve non-arcs, a potential image point for u in Dn must be independent of T f¯; this
cannot happen as then Dn would induce an independent n-set. So Dn does not have
the BEP and hence is not MB-homogeneous by Proposition 2.5.
Example 4.10. Let H be the complement of the homogeneous Kn -free graph for n ≥ 3.
We note that H contains R as a spanning subgraph (as it is algebraically closed) and
so it is MM-homogeneous as it satisfies the MEP [9]. However, as H does not embed
an independent n-set, it does not contain an infinite null graph. As H is not complete
(as it has a non-edge), it follows that H is not MB-homogeneous by Corollary 3.2.
Following this final example, we are able to present a complete classification of those
homogeneous graphs that are also MB-homogeneous.
Theorem 4.11. The only countably infinite homogeneous and MB-homogeneous graphs
are the following:
ℵ
• K ℵ0 and its complement K 0 ;
F
F
• i∈N Kiℵ0 and its complement i∈N Kiℵ0 ;
• the random graph R.
Proof. We check every item in the classification of countably infinite, undirected homogeneous graphs given in [19]. We showed that those graphs on the list are MBhomogeneous in Example 4.1, Proposition 3.4 and Example 4.3. Any other disconnected homogeneous graph must be the a countable union of complete graphs or a finite
union of infinite complete graphs; these are not MB-homogeneous by Proposition 3.4.
The only other countable homogeneous graphs are the Kn -free graphs and their complements; these are not MB-homogeneous by Example 4.6 and Example 4.10.
We end on some open questions concerning MB-homogeneous graphs in general. As
we have seen, it would be an arduous task to classify MB-homogeneous graphs up to
isomorphism; the idea of bimorphism equivalence represents the best hope to identify all
MB-homogeneous graphs. A positive answer to the following question would constitute
the best classification result possible for MB-homogeneous graphs, given the amount
and range of examples above.
Question 4.12. Is every countable MB-homogeneous graph bimorphism equivalent to
one of the five graphs in Theorem 4.11?
32
Related to this question is the following:
Question 4.13. Are there only countably many MB-homogeneous graphs up to bimorphism equivalence?
Finally, we notice that every example of an MB-homogeneous graph in this article
is also HE-homogeneous. This motivates our final question:
Question 4.14. Is there a countably infinite MB-homogeneous graph that is not HEhomogeneous? Conversely, is there a HE-homogeneous graph that is not MB-homogeneous?
References
[1] L. Agarwal. Reducts of the generic digraph. Annals of Pure and Applied Logic,
167(3):370–391, 2016.
[2] M. Bhattacharjee, H. D. Macpherson, R. G. Möller, and P. M. Neumann. Notes
on Infinite Permutation Groups. Number 1698 in Lecture Notes in Mathematics.
Springer, 1998.
[3] M. Bodirsky. The core of a countably categorical structure. In STACS 2005, pages
110–120. Springer, 2005.
[4] M. Bodirsky. Complexity classification in infinite-domain constraint satisfaction.
arXiv preprint arXiv:1201.0856, 2012.
[5] M. Bodirsky and J. Nešetřil. Constraint satisfaction with countable homogeneous
templates. Journal of Logic and Computation, 16(3):359–373, 2006.
[6] M. Bodirsky and M. Pinsker. Schaefer’s theorem for graphs. Journal of the ACM
(JACM), 62(3):19, 2015.
[7] P. J. Cameron. Oligomorphic permutation groups, volume 152. Cambridge University Press, 1990.
[8] P. J. Cameron. The random graph. In The Mathematics of Paul Erdös II, pages
333–351. Springer, 1997.
[9] P. J. Cameron and J. Nešetřil. Homomorphism-homogeneous relational structures.
Combinatorics, Probability & Computing, 15(1-2):91–103, 2006.
[10] G. L. Cherlin. The classification of countable homogeneous directed graphs and
countable homogeneous n-tournaments, volume 621. American Mathematical Soc.,
1998.
[11] R. Diestel. Graph theory. Springer-Verlag, 2000.
[12] I. Dolinka. The Bergman property for endomorphism monoids of some Fraı̈ssé
limits. In Forum Mathematicum, volume 26, pages 357–376, 2014.
[13] I. Dolinka, R. D. Gray, J. D. McPhee, J. D. Mitchell, and M. Quick. Automorphism
groups of countable algebraically closed graphs and endomorphisms of the random
graph. Math. Proc. Cambridge Philos. Soc., 160(3):437–462, 2016.
33
[14] R. Fraı̈ssé. Sur certaines relations qui généralisent l’ordre des nombres rationnels.
Comptes Rendus de l’Acádemie des Sciences de Paris, 237(11):540–542, 1953.
[15] R. Frucht. Graphs of degree three with a given abstract group. Canadian J. Math,
1:365–378, 1949.
[16] D. Hartman, J. Hubička, and D. Mašulović. Homomorphism-homogeneous Lcolored graphs. European J. Combin., 35:313–323, 2014.
[17] W. Hodges. Model Theory, volume 42. Cambridge University Press, 1993.
[18] A. Kechris. Classical descriptive set theory, volume 156. Springer Science &
Business Media, 2012.
[19] A. H. Lachlan and R. E. Woodrow. Countable ultrahomogeneous undirected
graphs. Transactions of the American Mathematical Society, pages 51–94, 1980.
[20] D. Lockett and J. K. Truss. Generic endomorphisms of homogeneous structures.
Groups and Model Theory, Contemporary Mathematics, page 217, 2012.
[21] D. C. Lockett and J. K. Truss. Some more notions of homomorphism–homogeneity.
Discrete Mathematics, 336:69–79, 2014.
[22] H. D. Macpherson. A survey of homogeneous structures. Discrete Mathematics,
311(15):1599–1634, 2011.
[23] D. Mašulović. Homomorphism-homogeneous partially ordered sets.
24(4):215–226, 2007.
Order,
[24] D. Mašulović. Some classes of finite homomorphism-homogeneous point-line geometries. Combinatorica, 33(5):573–590, 2013.
[25] D. Mašulović and M. Pech.
Oligomorphic transformation monoids and
homomorphism-homogeneous structures. Fundamenta Mathematicae, 212(1):17–
34, 2011.
[26] J. D. McPhee. Endomorphisms of Fraı̈ssé limits and automorphism groups of
algebraically closed relational structures. PhD thesis, University of St Andrews,
2012.
[27] J. Meakin. Groups and semigroups: connections and contrasts. London Mathematical Society Lecture Note Series, 340:357, 2007.
[28] N. Mudrinski. Notes on endomorphisms of Henson graphs and their complements.
Ars Combinatoria, 96:173–183, 2010.
[29] J. R. Munkres. Topology, volume 2. Prentice Hall Upper Saddle River, 2000.
[30] G. E. Reyes. Local definability theory. Annals of Mathematical Logic, 1(1):95–137,
1970.
[31] M. Rusinov and P. Schweitzer. Homomorphism–homogeneous graphs. Journal of
Graph Theory, 65(3):253–262, 2010.
34
[32] J. H. Schmerl. Countable homogeneous partially ordered sets. Algebra Universalis,
9(1):317–321, 1979.
[33] B. Steinberg. A theory of transformation monoids: combinatorics and representation theory. The Electronic Journal of Combinatorics, 17(1):R164, 2010.
35
| 4math.GR
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.